From 355de7f8d033a4b5082cb0816650d880f05d81c8 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 13 Sep 2017 16:03:39 -0400 Subject: Adding sperate dp_unit --- configs/GTX480/gpgpusim.config | 6 ++++-- configs/GeForceGTX750Ti/gpgpusim.config | 5 +++-- configs/QuadroFX5600/gpgpusim.config | 5 +++-- configs/QuadroFX5800/gpgpusim.config | 5 +++-- configs/TeslaC2050/gpgpusim.config | 3 ++- 5 files changed, 15 insertions(+), 9 deletions(-) (limited to 'configs') diff --git a/configs/GTX480/gpgpusim.config b/configs/GTX480/gpgpusim.config index 436cb41..7d8d91e 100644 --- a/configs/GTX480/gpgpusim.config +++ b/configs/GTX480/gpgpusim.config @@ -29,10 +29,12 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +#For Fermi, DP unit =0, DP inst is executed on SFU unit instead +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 -gpgpu_num_sp_units 2 -gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" diff --git a/configs/GeForceGTX750Ti/gpgpusim.config b/configs/GeForceGTX750Ti/gpgpusim.config index 8b030b6..bd6412c 100644 --- a/configs/GeForceGTX750Ti/gpgpusim.config +++ b/configs/GeForceGTX750Ti/gpgpusim.config @@ -28,10 +28,11 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 -gpgpu_num_sp_units 8 -gpgpu_num_sfu_units 32 +-gpgpu_num_dp_units 0 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" diff --git a/configs/QuadroFX5600/gpgpusim.config b/configs/QuadroFX5600/gpgpusim.config index cb87b65..82dc64a 100644 --- a/configs/QuadroFX5600/gpgpusim.config +++ b/configs/QuadroFX5600/gpgpusim.config @@ -17,10 +17,11 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,1,1,1,1,1,1 +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 -gpgpu_num_sp_units 1 -gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" diff --git a/configs/QuadroFX5800/gpgpusim.config b/configs/QuadroFX5800/gpgpusim.config index 82243c2..fdd87e0 100644 --- a/configs/QuadroFX5800/gpgpusim.config +++ b/configs/QuadroFX5800/gpgpusim.config @@ -16,10 +16,11 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,1,1,1,1,1,1 +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 -gpgpu_num_sp_units 1 -gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" diff --git a/configs/TeslaC2050/gpgpusim.config b/configs/TeslaC2050/gpgpusim.config index 442ab8b..4febbe5 100644 --- a/configs/TeslaC2050/gpgpusim.config +++ b/configs/TeslaC2050/gpgpusim.config @@ -33,9 +33,10 @@ # Pipeline widths and number of FUs # ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 -gpgpu_num_sp_units 2 -gpgpu_num_sfu_units 1 +-gpgpu_num_sfu_units 0 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" -- cgit v1.3 From 5b9e91d9b2b5d017a196eb4e02cc7215ec48f692 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 14 Sep 2017 20:19:49 -0400 Subject: Adding Pascal config files --- configs/Pascal-P100-HBM/config_fermi_islip.icnt | 70 +++++++++ configs/Pascal-P100-HBM/gpgpusim.config | 168 +++++++++++++++++++++ configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt | 70 +++++++++ configs/Pascal-P102-GDDR5X/gpgpusim.config | 162 ++++++++++++++++++++ 4 files changed, 470 insertions(+) create mode 100644 configs/Pascal-P100-HBM/config_fermi_islip.icnt create mode 100644 configs/Pascal-P100-HBM/gpgpusim.config create mode 100644 configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt create mode 100644 configs/Pascal-P102-GDDR5X/gpgpusim.config (limited to 'configs') diff --git a/configs/Pascal-P100-HBM/config_fermi_islip.icnt b/configs/Pascal-P100-HBM/config_fermi_islip.icnt new file mode 100644 index 0000000..a788090 --- /dev/null +++ b/configs/Pascal-P100-HBM/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 62; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config new file mode 100644 index 0000000..18aa82a --- /dev/null +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -0,0 +1,168 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 60 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 32 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Pscal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Nvidia_Tesla +-gpgpu_clock_domains 1480.0:2960.0:1480.0:715.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,2,2,1,2,2,2,1,6 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 2 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 2,2,2,2,130 +-ptx_opcode_latency_sfu 8 +-ptx_opcode_initiation_sfu 4 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_shmem_size 65536 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +# gpgpu_num_reg_banks should be increased to 32 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +# Use Fermi Coalsce arhitetecture which is the same as Pascal +-gpgpu_coalesce_arch 20 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +# DRAM latency should be lower compared to other configs, due to high-speed interposer connection +-dram_latency 60 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# GDDR5 timing +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=6:RCD=12:RAS=28:RP=12:RC=40: +# CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnkgrp_indexing_policy 1 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt new file mode 100644 index 0000000..602daee --- /dev/null +++ b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config new file mode 100644 index 0000000..4400454 --- /dev/null +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -0,0 +1,162 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:2834.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,1,4,1,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 1 + + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,8,8,8,130 +-ptx_opcode_latency_sfu 8 +-ptx_opcode_initiation_sfu 4 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +# gpgpu_num_reg_banks should be increased to 32 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +# Use Fermi Coalsce arhitetecture which is the same as Pascal +-gpgpu_coalesce_arch 20 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Pascal 102 has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 102 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + -- cgit v1.3 From 161f9cefeaf216f48f93e1192c817997cf875cac Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 26 Oct 2017 11:19:52 -0400 Subject: Changing the Titan X config file to use the last modifications --- configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt | 4 +- configs/Pascal-P102-GDDR5X/gpgpusim.config | 46 +++++++++++++--------- src/gpgpu-sim/gpu-sim.cc | 3 ++ src/gpgpu-sim/gpu-sim.h | 1 - 4 files changed, 32 insertions(+), 22 deletions(-) (limited to 'configs') diff --git a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt index 602daee..58e596d 100644 --- a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt +++ b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt @@ -1,6 +1,6 @@ //21*1 fly with 32 flits per packet under gpgpusim injection mode use_map = 0; -flit_size = 32; +flit_size = 40; // currently we do not use this, see subnets below network_count = 2; @@ -17,7 +17,7 @@ routing_function = dest_tag; // Flow control num_vcs = 1; -vc_buf_size = 8; +vc_buf_size = 32; wait_for_tail_credit = 0; diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 8b02680..36d13af 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -12,8 +12,9 @@ -gpgpu_ptx_save_converted_ptxplus 0 # high level architecture configuration +# P102 has two semi-indp scheds per core, and two cores per cluster -gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 +-gpgpu_n_cores_per_cluster 2 -gpgpu_n_mem 12 -gpgpu_n_sub_partition_per_mchannel 2 @@ -24,20 +25,20 @@ -gpgpu_clock_domains 1417.0:2834.0:1417.0:2500.0 # shader core pipeline config --gpgpu_shader_registers 65536 +-gpgpu_shader_registers 32768 -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 +# This implies a maximum of 32 warps/SM +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 16 -gpgpu_simd_model 1 # Pipeline widths and number of FUs # ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB ## Pascal GP102 has 4 SP SIMD units and 4 SFU units ## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,4,1,4,1,4,1,9 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 -gpgpu_num_dp_units 1 @@ -51,7 +52,6 @@ -ptx_opcode_initiation_fp 1,2,1,1,4 -ptx_opcode_latency_dp 8,19,8,8,330 -ptx_opcode_initiation_dp 8,8,8,8,130 --ptx_opcode_latency_sfu 8 -ptx_opcode_initiation_sfu 4 # ::,::::,::,:** @@ -60,13 +60,16 @@ # Pascal GP102 has 96KB Shared memory # Pascal GP102 has 64KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,S:128:8,8 --gpgpu_shmem_size 98304 +-gpgpu_cache:dl1 S:64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_shmem_size 49152 -gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:W:L,A:128:4,4:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:256:4,4:0,32 -gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 @@ -76,15 +79,14 @@ -gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 # enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units -gpgpu_operand_collector_num_units_sp 12 -gpgpu_operand_collector_num_units_sfu 6 -gpgpu_operand_collector_num_units_mem 8 -gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 4 --gpgpu_operand_collector_num_out_ports_sfu 4 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_operand_collector_num_in_ports_sfu 2 +-gpgpu_operand_collector_num_out_ports_sfu 2 -gpgpu_operand_collector_num_in_ports_mem 1 -gpgpu_operand_collector_num_out_ports_mem 1 -gpgpu_operand_collector_num_in_ports_dp 1 @@ -119,7 +121,7 @@ # the minimum DRAM latency (100 core cycles). I.e. # Total buffer space required = 100 x 924MHz / 700MHz = 132 -gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 +-gpgpu_dram_return_queue_size 192 # for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) # 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition @@ -136,8 +138,14 @@ -gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 32:28:16 + # Pascal 102 has four schedulers per core --gpgpu_num_sched_per_core 4 +-gpgpu_num_sched_per_core 2 # Two Level Scheduler with active and pending pools #-gpgpu_scheduler two_level_active:6:0:1 # Loose round robbin scheduler diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 7838875..6179d46 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -207,6 +207,9 @@ void memory_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-Write_Queue_Size", OPT_CSTR, &write_queue_size_opt, "Write_Queue_Size", "32:28:16"); + option_parser_register(opp, "-icnt_flit_size", OPT_UINT32, &icnt_flit_size, + "icnt_flit_size", + "32"); m_address_mapping.addrdec_setoption(opp); } diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 197350b..52c4643 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -213,7 +213,6 @@ struct memory_config { m_L2_config.init(&m_address_mapping); m_valid = true; - icnt_flit_size = 32; // Default 32 sscanf(write_queue_size_opt,"%d:%d:%d", &gpgpu_frfcfs_dram_write_queue_size,&write_high_watermark,&write_low_watermark); -- cgit v1.3 From c1bb3638d2c3a8f1b86e0e952e12bb2ed7f1c0dc Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 26 Oct 2017 11:29:45 -0400 Subject: adding sfu latency to config file --- configs/Pascal-P102-GDDR5X/gpgpusim.config | 1 + 1 file changed, 1 insertion(+) (limited to 'configs') diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 36d13af..5419c51 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -53,6 +53,7 @@ -ptx_opcode_latency_dp 8,19,8,8,330 -ptx_opcode_initiation_dp 8,8,8,8,130 -ptx_opcode_initiation_sfu 4 +-ptx_opcode_latency_sfu 8 # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo -- cgit v1.3 From f23021ad8663636e1103bd75a742480cb6238435 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 27 Oct 2017 22:35:52 -0400 Subject: add more statistics and chaging Pascal config --- configs/Pascal-P102-GDDR5X/gpgpusim.config | 12 +++++++----- src/abstract_hardware_model.cc | 2 +- src/gpgpu-sim/dram.cc | 3 ++- src/gpgpu-sim/dram_sched.cc | 2 ++ src/gpgpu-sim/gpu-cache.cc | 2 ++ src/gpgpu-sim/gpu-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.h | 2 +- src/gpgpu-sim/l2cache.cc | 2 +- src/gpgpu-sim/mem_latency_stat.cc | 17 +++++++++++++---- src/gpgpu-sim/mem_latency_stat.h | 4 ++++ src/gpgpu-sim/shader.cc | 8 ++++---- 11 files changed, 38 insertions(+), 18 deletions(-) (limited to 'configs') diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 5419c51..e830023 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -59,22 +59,24 @@ # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. # Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache +# Pascal GP102 has 24KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 S:64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_cache:dl1 N:32:128:6,L:L:m:N:H,S:128:8,16 -gpgpu_shmem_size 49152 -gmem_skip_L1D 1 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:256:4,4:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:128:4,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 +#-gpgpu_flush_l2_cache 1 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 # 48 KB Tex +# this is unused -gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 # 12 KB Const -gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 @@ -100,7 +102,7 @@ -gpgpu_shmem_limited_broadcast 0 -gpgpu_shmem_warp_parts 1 # Use Fermi Coalsce arhitetecture which is the same as Pascal --gpgpu_coalesce_arch 20 +-gpgpu_coalesce_arch 61 ## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units -gpgpu_max_insn_issue_per_warp 2 @@ -143,7 +145,7 @@ -dram_bnkgrp_indexing_policy 1 #-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 32:28:16 +#-Write_Queue_Size 64:56:32 # Pascal 102 has four schedulers per core -gpgpu_num_sched_per_core 2 diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 51265fd..d2a155c 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -314,7 +314,7 @@ void warp_inst_t::generate_mem_accesses() break; case global_space: case local_space: case param_space_local: - if( m_config->gpgpu_coalesce_arch == 13 || m_config->gpgpu_coalesce_arch == 20) { + if( m_config->gpgpu_coalesce_arch >= 13 && m_config->gpgpu_coalesce_arch <= 62) { if(isatomic()) memory_coalescing_arch_atomic(is_write, access_type); else diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index 7ed7b6f..de37f64 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -570,9 +570,10 @@ bool dram_t::issue_col_command(int j) bkgrp[grp]->RTPLc = m_config->tRTPL; issued = true; if(bk[j]->mrq->data->get_access_type() == L2_WR_ALLOC_R) - n_rd_L2_A++; + n_rd_L2_A++; else n_rd++; + bwutil += m_config->BL/m_config->data_command_freq_ratio; bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; bk[j]->n_access++; diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc index 7a140c5..ac4c827 100644 --- a/src/gpgpu-sim/dram_sched.cc +++ b/src/gpgpu-sim/dram_sched.cc @@ -220,6 +220,8 @@ void dram_t::scheduler_frfcfs() bk[b]->mrq = req; if (m_config->gpgpu_memlatency_stat) { mrq_latency = gpu_sim_cycle + gpu_tot_sim_cycle - bk[b]->mrq->timestamp; + m_stats->tot_mrq_latency += mrq_latency; + m_stats->tot_mrq_num++; bk[b]->mrq->timestamp = gpu_tot_sim_cycle + gpu_sim_cycle; m_stats->mrq_lat_table[LOGB2(mrq_latency)]++; if (mrq_latency > m_stats->max_mrq_latency) { diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index d199cca..32c2bb1 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -333,8 +333,10 @@ void tag_array::fill( unsigned index, unsigned time, mem_fetch* mf) void tag_array::flush() { for (unsigned i=0; i < m_config.get_num_lines(); i++) + if(m_lines[i]->is_modified_line()) { for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; + } } float tag_array::windowed_miss_rate( ) const diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 6179d46..0e06c5c 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1412,7 +1412,7 @@ void gpgpu_sim::cycle() if (mf) { unsigned response_size = mf->get_is_write()?mf->get_ctrl_size():mf->size(); if ( ::icnt_has_buffer( m_shader_config->mem2device(i), response_size ) ) { - if (!mf->get_is_write()) + //if (!mf->get_is_write()) mf->set_return_timestamp(gpu_sim_cycle+gpu_tot_sim_cycle); mf->set_status(IN_ICNT_TO_SHADER,gpu_sim_cycle+gpu_tot_sim_cycle); ::icnt_push( m_shader_config->mem2device(i), mf->get_tpc(), mf, response_size ); diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 52c4643..043fcee 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -199,7 +199,7 @@ struct memory_config { assert(nbkgrp>0 && "Number of bank groups cannot be zero"); tRCDWR = tRCD-(WL+1); tRTW = (CL+(BL/data_command_freq_ratio)+2-WL); - tWTR = (WL+(BL/data_command_freq_ratio)+tCDLR); + tWTR = (WL+(BL/data_command_freq_ratio)+tCDLR); tWTP = (WL+(BL/data_command_freq_ratio)+tWR); dram_atom_size = BL * busW * gpu_n_mem_per_ctrlr; // burst length x bus width x # chips per partition diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index cac59f1..8fbf448 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -628,6 +628,7 @@ std::vector memory_sub_partition::breakdown_request_to_sector_reques void memory_sub_partition::push( mem_fetch* m_req, unsigned long long cycle ) { if (m_req) { + m_stats->memlatstat_icnt2mem_pop(m_req); std::vector reqs; if(m_config->m_L2_config.m_cache_type == SECTOR) reqs = breakdown_request_to_sector_requests(m_req); @@ -637,7 +638,6 @@ void memory_sub_partition::push( mem_fetch* m_req, unsigned long long cycle ) for(unsigned i=0; imemlatstat_icnt2mem_pop(req); if( req->istexture() ) { m_icnt_L2_queue->push(req); req->set_status(IN_PARTITION_ICNT_TO_L2_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); diff --git a/src/gpgpu-sim/mem_latency_stat.cc b/src/gpgpu-sim/mem_latency_stat.cc index fde0eff..35d6d84 100644 --- a/src/gpgpu-sim/mem_latency_stat.cc +++ b/src/gpgpu-sim/mem_latency_stat.cc @@ -75,6 +75,10 @@ memory_stats_t::memory_stats_t( unsigned n_shader, const struct shader_core_conf max_mf_latency = 0; max_icnt2mem_latency = 0; max_icnt2sh_latency = 0; + tot_icnt2mem_latency = 0; + tot_icnt2sh_latency = 0; + tot_mrq_num = 0; + tot_mrq_latency = 0; memset(mrq_lat_table, 0, sizeof(unsigned)*32); memset(dq_lat_table, 0, sizeof(unsigned)*32); memset(mf_lat_table, 0, sizeof(unsigned)*32); @@ -158,6 +162,7 @@ void memory_stats_t::memlatstat_read_done(mem_fetch *mf) mf_max_lat_table[mf->get_tlx_addr().chip][mf->get_tlx_addr().bk] = mf_latency; unsigned icnt2sh_latency; icnt2sh_latency = (gpu_tot_sim_cycle+gpu_sim_cycle) - mf->get_return_timestamp(); + tot_icnt2sh_latency += icnt2sh_latency; icnt2sh_lat_table[LOGB2(icnt2sh_latency)]++; if (icnt2sh_latency > max_icnt2sh_latency) max_icnt2sh_latency = icnt2sh_latency; @@ -191,6 +196,7 @@ void memory_stats_t::memlatstat_icnt2mem_pop(mem_fetch *mf) if (m_memory_config->gpgpu_memlatency_stat) { unsigned icnt2mem_latency; icnt2mem_latency = (gpu_tot_sim_cycle+gpu_sim_cycle) - mf->get_timestamp(); + tot_icnt2mem_latency += icnt2mem_latency; icnt2mem_lat_table[LOGB2(icnt2mem_latency)]++; if (icnt2mem_latency > max_icnt2mem_latency) max_icnt2mem_latency = icnt2mem_latency; @@ -216,14 +222,17 @@ void memory_stats_t::memlatstat_print( unsigned n_mem, unsigned gpu_mem_n_bk ) unsigned max_bank_accesses, min_bank_accesses, max_chip_accesses, min_chip_accesses; if (m_memory_config->gpgpu_memlatency_stat) { + printf("maxmflatency = %d \n", max_mf_latency); + printf("max_icnt2mem_latency = %d \n", max_icnt2mem_latency); printf("maxmrqlatency = %d \n", max_mrq_latency); - printf("maxdqlatency = %d \n", max_dq_latency); - printf("maxmflatency = %d \n", max_mf_latency); + //printf("maxdqlatency = %d \n", max_dq_latency); + printf("max_icnt2sh_latency = %d \n", max_icnt2sh_latency); if (num_mfs) { printf("averagemflatency = %lld \n", mf_total_lat/num_mfs); + printf("avg_icnt2mem_latency = %lld \n", tot_icnt2mem_latency/num_mfs); + printf("avg_mrq_latency = %lld \n", tot_mrq_latency/tot_mrq_num); + printf("avg_icnt2sh_latency = %lld \n", tot_icnt2sh_latency/num_mfs); } - printf("max_icnt2mem_latency = %d \n", max_icnt2mem_latency); - printf("max_icnt2sh_latency = %d \n", max_icnt2sh_latency); printf("mrq_lat_table:"); for (i=0; i< 32; i++) { printf("%d \t", mrq_lat_table[i]); diff --git a/src/gpgpu-sim/mem_latency_stat.h b/src/gpgpu-sim/mem_latency_stat.h index 4968a3b..5b89202 100644 --- a/src/gpgpu-sim/mem_latency_stat.h +++ b/src/gpgpu-sim/mem_latency_stat.h @@ -56,6 +56,10 @@ public: unsigned max_dq_latency; unsigned max_mf_latency; unsigned max_icnt2mem_latency; + unsigned long long int tot_icnt2mem_latency; + unsigned long long int tot_icnt2sh_latency; + unsigned long long int tot_mrq_latency; + unsigned long long int tot_mrq_num; unsigned max_icnt2sh_latency; unsigned mrq_lat_table[32]; unsigned dq_lat_table[32]; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index c14e19f..bf482fb 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1458,7 +1458,7 @@ ldst_unit::process_cache_access( cache_t* cache, if( !write_sent ) delete mf; } else if ( status == RESERVATION_FAIL ) { - result = COAL_STALL; + result = BK_CONF; assert( !read_sent ); assert( !write_sent ); delete mf; @@ -1467,8 +1467,8 @@ ldst_unit::process_cache_access( cache_t* cache, //inst.clear_active( access.get_warp_mask() ); // threads in mf writeback when mf returns inst.accessq_pop_back(); } - if( !inst.accessq_empty() ) - result = BK_CONF; + if( !inst.accessq_empty() && result == NO_RC_FAIL) + result = COAL_STALL; return result; } @@ -1563,7 +1563,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea assert( CACHE_UNDEFINED != inst.cache_op ); stall_cond = process_memory_access_queue(m_L1D,inst); } - if( !inst.accessq_empty() ) + if( !inst.accessq_empty() && stall_cond == NO_RC_FAIL) stall_cond = COAL_STALL; if (stall_cond != NO_RC_FAIL) { stall_reason = stall_cond; -- cgit v1.3 From 110b88d9273d16f5eab1193402cb45557fa23a69 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 28 Oct 2017 10:18:42 -0400 Subject: Adding in the config Mahmoud made that requires no code changes to GPGPU-Sim --- .../gpgpu-sim-3.x/config_fermi_islip.icnt | 70 +++ configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 151 ++++++ .../gpgpu-sim-3.x/gpuwattch_gtx480.xml | 538 +++++++++++++++++++++ 3 files changed, 759 insertions(+) create mode 100644 configs/PascalTitanX/gpgpu-sim-3.x/config_fermi_islip.icnt create mode 100644 configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config create mode 100755 configs/PascalTitanX/gpgpu-sim-3.x/gpuwattch_gtx480.xml (limited to 'configs') diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/config_fermi_islip.icnt b/configs/PascalTitanX/gpgpu-sim-3.x/config_fermi_islip.icnt new file mode 100644 index 0000000..602daee --- /dev/null +++ b/configs/PascalTitanX/gpgpu-sim-3.x/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config new file mode 100644 index 0000000..782edf6 --- /dev/null +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -0,0 +1,151 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:2834.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 4 SFU units +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,1,4,4,1,8 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,8,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +# gpgpu_num_reg_banks should be increased to 32 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Pascal GP102 has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpuwattch_gtx480.xml b/configs/PascalTitanX/gpgpu-sim-3.x/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + -- cgit v1.3 From c56115c9df274708ada5964f4f46e2d08b1acd03 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Sun, 29 Oct 2017 20:53:30 -0400 Subject: Update gpgpusim.config --- configs/Pascal-P100-HBM/gpgpusim.config | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'configs') diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config index 5b038de..0f6761d 100644 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -58,8 +58,8 @@ -gpgpu_shmem_size 65536 -gmem_skip_L1D 0 -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:W:L,A:128:4,4:0,32 +# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache +-gpgpu_cache:dl2 S:32:128:16,L:B:m:W:L,A:128:4,4:0,32 -gpgpu_cache:dl2_texture_only 0 # 4 KB Inst. -- cgit v1.3 From ff6c395da7b70d9cdbb72a8cfc3721a7f27a2a25 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Mon, 30 Oct 2017 17:08:07 -0400 Subject: Update gpgpusim.config --- configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 23 +++++++++++++--------- 1 file changed, 14 insertions(+), 9 deletions(-) (limited to 'configs') diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 782edf6..2f09ab6 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -35,18 +35,23 @@ # ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB ## Pascal GP102 has 4 SP SIMD units and 4 SFU units ## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,1,4,4,1,8 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 1 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 -ptx_opcode_latency_int 4,13,4,5,145 -ptx_opcode_initiation_int 1,1,1,1,4 -ptx_opcode_latency_fp 4,13,4,5,39 -ptx_opcode_initiation_fp 1,2,1,1,4 -ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,8,4,4,130 +-ptx_opcode_initiation_dp 8,8,8,8,130 +-ptx_opcode_initiation_sfu 4 +-ptx_opcode_latency_sfu 8 # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo @@ -54,21 +59,21 @@ # Pascal GP102 has 96KB Shared memory # Pascal GP102 has 64KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 98304 -gmem_skip_L1D 1 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:W:L,A:128:8,4:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 # 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 # 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 # 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 # enable operand collector ## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units -- cgit v1.3 From 4aa3b5158f0e70e83cfe44dae79cd75110c52eb7 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Mon, 30 Oct 2017 17:08:45 -0400 Subject: Update gpgpusim.config --- configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'configs') diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 2f09ab6..7dfd6d5 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -36,8 +36,8 @@ ## Pascal GP102 has 4 SP SIMD units and 4 SFU units ## we need to scale the number of pipeline registers to be equal to the number of SP units -gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 -gpgpu_num_dp_units 1 # Instruction latencies and initiation intervals -- cgit v1.3 From 178abb228229332c4717c5fd565a83875c3d5c9e Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Mon, 30 Oct 2017 18:51:10 -0400 Subject: Update gpgpusim.config --- configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-) (limited to 'configs') diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 7dfd6d5..82dfa9a 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -32,10 +32,10 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB ## Pascal GP102 has 4 SP SIMD units and 4 SFU units ## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_pipeline_widths 4,1,4,1,4,1,4,1,9 -gpgpu_num_sp_units 4 -gpgpu_num_sfu_units 4 -gpgpu_num_dp_units 1 @@ -77,15 +77,18 @@ # enable operand collector ## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 -gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 -gpgpu_operand_collector_num_in_ports_sp 4 -gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_sfu 4 +-gpgpu_operand_collector_num_out_ports_sfu 4 -gpgpu_operand_collector_num_in_ports_mem 1 -gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 # gpgpu_num_reg_banks should be increased to 32 -gpgpu_num_reg_banks 32 -- cgit v1.3 From 0b1a646c0ddf5d4db64a2b6fe9dfa30789cf1cd4 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 30 Oct 2017 20:23:51 -0400 Subject: adding new stats and change the PascalP100-HBM config --- configs/Pascal-P100-HBM/config_fermi_islip.icnt | 9 +++++--- configs/Pascal-P100-HBM/gpgpusim.config | 25 ++++++++++++---------- configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt | 3 +++ src/gpgpu-sim/dram.cc | 13 +++++++---- src/gpgpu-sim/dram.h | 4 ++++ src/gpgpu-sim/dram_sched.cc | 21 +++++++++++++----- src/gpgpu-sim/gpu-sim.cc | 4 +++- src/gpgpu-sim/gpu-sim.h | 8 +++++++ 8 files changed, 63 insertions(+), 24 deletions(-) (limited to 'configs') diff --git a/configs/Pascal-P100-HBM/config_fermi_islip.icnt b/configs/Pascal-P100-HBM/config_fermi_islip.icnt index a788090..0a73c81 100644 --- a/configs/Pascal-P100-HBM/config_fermi_islip.icnt +++ b/configs/Pascal-P100-HBM/config_fermi_islip.icnt @@ -1,13 +1,13 @@ //21*1 fly with 32 flits per packet under gpgpusim injection mode use_map = 0; -flit_size = 32; +flit_size = 40; // currently we do not use this, see subnets below network_count = 2; // Topology topology = fly; -k = 62; +k = 60; n = 1; // Routing @@ -17,7 +17,10 @@ routing_function = dest_tag; // Flow control num_vcs = 1; -vc_buf_size = 8; +vc_buf_size = 64; +input_buffer_size = 64; +ejection_buffer_size = 64; +boundary_buffer_size = 64; wait_for_tail_credit = 0; diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config index 5b038de..1029194 100644 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -54,13 +54,17 @@ # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. # Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:128:8,8 +-gpgpu_cache:dl1 S:64:128:6,L:L:m:N:H,A:128:8,32:0,32 -gpgpu_shmem_size 65536 --gmem_skip_L1D 0 +-gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:W:L,A:128:4,4:0,32 +# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:W:L,A:256:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +#-gpgpu_flush_l2_cache 1 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 @@ -89,8 +93,7 @@ -gpgpu_shmem_num_banks 32 -gpgpu_shmem_limited_broadcast 0 -gpgpu_shmem_warp_parts 1 -# Use Fermi Coalscer arhitetecture for now! Need to be canged to pascal Coalscer --gpgpu_coalesce_arch 20 +-gpgpu_coalesce_arch 60 ## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units -gpgpu_max_insn_issue_per_warp 2 @@ -123,10 +126,6 @@ -gpgpu_mem_address_mask 1 -gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS -# GDDR5 timing -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=6:RCD=12:RAS=28:RP=12:RC=40: -# CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - # HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) # Timing for 1 GHZ # tRRDl and tWTR are missing, need to be added @@ -140,8 +139,12 @@ # HBM has dual bus interface, in which it can issue two col and row commands at a time -dual_bus_interface 1 # select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 -dram_bnkgrp_indexing_policy 1 +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + # Pascal has two schedulers per core -gpgpu_num_sched_per_core 2 # Two Level Scheduler with active and pending pools @@ -158,7 +161,7 @@ -visualizer_enabled 0 # power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 1 +-power_simulation_enabled 0 -gpuwattch_xml_file gpuwattch_gtx480.xml # tracing functionality diff --git a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt index 58e596d..94b2378 100644 --- a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt +++ b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt @@ -18,6 +18,9 @@ routing_function = dest_tag; num_vcs = 1; vc_buf_size = 32; +input_buffer_size = 32; +ejection_buffer_size = 32; +boundary_buffer_size = 32; wait_for_tail_credit = 0; diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index de37f64..a57508c 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -52,6 +52,10 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m //rowblp access_num=0; hits_num=0; + read_num=0; + write_num=0; + hits_read_num=0; + hits_write_num=0; banks_1time=0; banks_acess_total=0; banks_acess_total_after=0; @@ -718,6 +722,8 @@ void dram_t::print( FILE* simFile) const fprintf(simFile, "\n------------------------------------------------------------------------\n"); printf("\nRow_Buffer_Locality = %.6f", (float)hits_num / access_num); + printf("\nRow_Buffer_Locality_read = %.6f", (float)hits_read_num / read_num); + printf("\nRow_Buffer_Locality_write = %.6f", (float)hits_write_num / write_num); printf("\nBank_Level_Parallism = %.6f", (float)banks_1time / banks_acess_total); printf("\nBank_Level_Parallism_Col = %.6f", (float)banks_time_rw / banks_access_rw_total); printf("\nBank_Level_Parallism_Ready = %.6f", (float)banks_time_ready /banks_access_ready_total); @@ -735,7 +741,7 @@ void dram_t::print( FILE* simFile) const printf("RCDWRc_limit = %d \n", RCDWRc_limit); printf("WTRc_limit = %d \n", WTRc_limit); printf("RTWc_limit = %d \n", RTWc_limit); - printf("CCDLc_limit %d \n", CCDLc_limit); + printf("CCDLc_limit = %d \n", CCDLc_limit); printf("rwq = %d \n", rwq_limit); printf("CCDLc_limit_alone = %d \n", CCDLc_limit_alone); printf("WTRc_limit_alone = %d \n", WTRc_limit_alone); @@ -751,15 +757,14 @@ void dram_t::print( FILE* simFile) const printf("n_pre = %d \n", n_pre); printf("n_ref = %d \n", n_ref); printf("n_req = %d \n", n_req ); - printf("n_req4 = %d \n", n_req*4 ); printf("total_req = %d \n\n", n_rd+n_wr+n_rd_L2_A+n_wr_WB); printf("issued_total_row = %lu \n", issued_total_row); printf("issued_total_col = %lu \n", issued_total_col); printf("Row_Bus_Util = %.6f \n", (float)issued_total_row / n_cmd); printf("CoL_Bus_Util = %.6f \n", (float)issued_total_col / n_cmd); - printf("Either_Row_CoL_Bus_Util %.6f \n", (float)issued_total / n_cmd); - printf("Issued_on_Two_Bus_Simul_Util %.6f \n", (float)issued_two /n_cmd); + printf("Either_Row_CoL_Bus_Util = %.6f \n", (float)issued_total / n_cmd); + printf("Issued_on_Two_Bus_Simul_Util = %.6f \n", (float)issued_two /n_cmd); printf("issued_two_Eff = %.6f \n", (float)issued_two /issued_total); printf("queue_avg = %.6f \n\n", (float)ave_mrqs/n_cmd ); diff --git a/src/gpgpu-sim/dram.h b/src/gpgpu-sim/dram.h index 29731a7..0d4c0e7 100644 --- a/src/gpgpu-sim/dram.h +++ b/src/gpgpu-sim/dram.h @@ -198,7 +198,11 @@ private: //row locality, BLP and other statistics unsigned long access_num; + unsigned long read_num; + unsigned long write_num; unsigned long long hits_num; + unsigned long long hits_read_num; + unsigned long long hits_write_num; unsigned long long banks_1time; unsigned long long banks_acess_total; unsigned long long banks_acess_total_after; diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc index ac4c827..f754d36 100644 --- a/src/gpgpu-sim/dram_sched.cc +++ b/src/gpgpu-sim/dram_sched.cc @@ -142,14 +142,25 @@ dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row ) rowhit = true; } } - //rowblp - m_dram->access_num++; - if(rowhit) - m_dram->hits_num++; - std::list::iterator next = m_current_last_row[bank]->back(); dram_req_t *req = (*next); + //rowblp stats + m_dram->access_num++; + bool is_write = req->data->is_write(); + if(is_write) + m_dram->write_num++; + else + m_dram->read_num++; + + if(rowhit) { + m_dram->hits_num++; + if(is_write) + m_dram->hits_write_num++; + else + m_dram->hits_read_num++; + } + m_stats->concurrent_row_access[m_dram->id][bank]++; m_stats->row_access[m_dram->id][bank]++; m_current_last_row[bank]->pop_back(); diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 0e06c5c..11ac5df 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -207,6 +207,9 @@ void memory_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-Write_Queue_Size", OPT_CSTR, &write_queue_size_opt, "Write_Queue_Size", "32:28:16"); + option_parser_register(opp, "-Elimnate_rw_turnaround", OPT_BOOL, &elimnate_rw_turnaround, + "elimnate_rw_turnaround i.e set tWTR and tRTW = 0", + "0"); option_parser_register(opp, "-icnt_flit_size", OPT_UINT32, &icnt_flit_size, "icnt_flit_size", "32"); @@ -435,7 +438,6 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) option_parser_register(opp, "-gpgpu_flush_l2_cache", OPT_BOOL, &gpgpu_flush_l2_cache, "Flush L2 cache at the end of each kernel call", "0"); - option_parser_register(opp, "-gpgpu_deadlock_detect", OPT_BOOL, &gpu_deadlock_detect, "Stop the simulation at deadlock (1=on (default), 0=off)", "1"); diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 043fcee..c04648c 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -198,8 +198,14 @@ struct memory_config { bk_tag_length = i-1; assert(nbkgrp>0 && "Number of bank groups cannot be zero"); tRCDWR = tRCD-(WL+1); + if(elimnate_rw_turnaround) + { + tRTW = 0; + tWTR = 0; + } else { tRTW = (CL+(BL/data_command_freq_ratio)+2-WL); tWTR = (WL+(BL/data_command_freq_ratio)+tCDLR); + } tWTP = (WL+(BL/data_command_freq_ratio)+tWR); dram_atom_size = BL * busW * gpu_n_mem_per_ctrlr; // burst length x bus width x # chips per partition @@ -266,6 +272,8 @@ struct memory_config { unsigned nbk; + bool elimnate_rw_turnaround; + unsigned data_command_freq_ratio; // frequency ratio between DRAM data bus and command bus (2 for GDDR3, 4 for GDDR5) unsigned dram_atom_size; // number of bytes transferred per read or write command -- cgit v1.3 From 4df2f7d2499600fa0f0c5642c07fed6576aa5713 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Mon, 30 Oct 2017 22:36:16 -0400 Subject: fixing TITANX config --- configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'configs') diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 82dfa9a..4191eb0 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -64,16 +64,16 @@ -gmem_skip_L1D 1 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2 N:64:128:16,L:B:m:W:L,A:128:8,4:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 # 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,A:2:48,4 # 48 KB Tex -gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 # 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,A:2:64,4 # enable operand collector ## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units -- cgit v1.3 From dccac457bb2263602a6a244297168cae8fda8e47 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 15 Nov 2017 20:58:22 -0500 Subject: Update gpgpusim.config --- configs/Pascal-P102-GDDR5X/gpgpusim.config | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) (limited to 'configs') diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index e830023..257560e 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -22,7 +22,7 @@ #-gpgpu_clock_domains ::: # Pascal NVIDIA TITAN X clock domains are adopted from # https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:2834.0:1417.0:2500.0 +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 # shader core pipeline config -gpgpu_shader_registers 32768 @@ -113,7 +113,7 @@ -inter_config_file config_fermi_islip.icnt # memory partition latency config --rop_latency 120 +-rop_latency 0 -dram_latency 100 # dram model config @@ -123,8 +123,8 @@ # To allow 100% DRAM utility, there should at least be enough buffer to sustain # the minimum DRAM latency (100 core cycles). I.e. # Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 240 # for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) # 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -- cgit v1.3 From 3fe52ab24a75493ad201ef82837dced36272b8f5 Mon Sep 17 00:00:00 2001 From: Mahmoud Abdallah Date: Sat, 18 Nov 2017 23:56:06 -0500 Subject: improving P102 accuracy --- configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt | 8 ++++---- configs/Pascal-P102-GDDR5X/gpgpusim.config | 10 +++++----- 2 files changed, 9 insertions(+), 9 deletions(-) (limited to 'configs') diff --git a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt index 94b2378..714d933 100644 --- a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt +++ b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt @@ -17,10 +17,10 @@ routing_function = dest_tag; // Flow control num_vcs = 1; -vc_buf_size = 32; -input_buffer_size = 32; -ejection_buffer_size = 32; -boundary_buffer_size = 32; +vc_buf_size = 64; +input_buffer_size = 256; +ejection_buffer_size = 64; +boundary_buffer_size = 64; wait_for_tail_credit = 0; diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 257560e..cb69767 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -61,7 +61,7 @@ # Pascal GP102 has 96KB Shared memory # Pascal GP102 has 24KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 N:32:128:6,L:L:m:N:H,S:128:8,16 +-gpgpu_cache:dl1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 49152 -gmem_skip_L1D 1 -icnt_flit_size 40 @@ -71,7 +71,7 @@ -gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:128:4,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 -#-gpgpu_flush_l2_cache 1 +-perf_sim_memcpy 1 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 @@ -113,7 +113,7 @@ -inter_config_file config_fermi_islip.icnt # memory partition latency config --rop_latency 0 +-rop_latency 100 -dram_latency 100 # dram model config @@ -138,8 +138,8 @@ # Use the same GDDR5 timing from hynix H5GQ1H24AFR # disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: + CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" -dram_bnk_indexing_policy 0 -dram_bnkgrp_indexing_policy 1 -- cgit v1.3 From 68cc9491f40e41d6421a2e6b068b4302cedf80ec Mon Sep 17 00:00:00 2001 From: Mahmoud Abdallah Date: Sun, 19 Nov 2017 00:01:30 -0500 Subject: improving P100 accuracy --- configs/Pascal-P100-HBM/config_fermi_islip.icnt | 8 ++++---- configs/Pascal-P100-HBM/gpgpusim.config | 8 ++++---- configs/Pascal-P102-GDDR5X/gpgpusim.config | 2 +- 3 files changed, 9 insertions(+), 9 deletions(-) (limited to 'configs') diff --git a/configs/Pascal-P100-HBM/config_fermi_islip.icnt b/configs/Pascal-P100-HBM/config_fermi_islip.icnt index 0a73c81..e7c2c3b 100644 --- a/configs/Pascal-P100-HBM/config_fermi_islip.icnt +++ b/configs/Pascal-P100-HBM/config_fermi_islip.icnt @@ -17,10 +17,10 @@ routing_function = dest_tag; // Flow control num_vcs = 1; -vc_buf_size = 64; -input_buffer_size = 64; -ejection_buffer_size = 64; -boundary_buffer_size = 64; +vc_buf_size = 128; +input_buffer_size = 256; +ejection_buffer_size = 128; +boundary_buffer_size = 128; wait_for_tail_credit = 0; diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config index 1029194..533a865 100644 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -17,7 +17,7 @@ #-gpgpu_clock_domains ::: # Pascal NVIDIA GP100 clock domains are adopted from # https://en.wikipedia.org/wiki/Nvidia_Tesla --gpgpu_clock_domains 1480.0:2960.0:1480.0:715.0 +-gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 # shader core pipeline config -gpgpu_shader_registers 65536 @@ -54,7 +54,7 @@ # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. # Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:6,L:L:m:N:H,A:128:8,32:0,32 +-gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 65536 -gmem_skip_L1D 1 -icnt_flit_size 40 @@ -64,7 +64,7 @@ -gpgpu_cache:dl2 S:64:128:16,L:B:m:W:L,A:256:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 -#-gpgpu_flush_l2_cache 1 +-perf_sim_memcpy 0 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 @@ -116,7 +116,7 @@ # the minimum DRAM latency (100 core cycles). I.e. # Total buffer space required = 100 x 924MHz / 700MHz = 132 -gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 +-gpgpu_dram_return_queue_size 192 # for HBM, 32 channles, each (128 bits) 16 bytes width -gpgpu_n_mem_per_ctrlr 1 diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index cb69767..0c6c126 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -71,7 +71,7 @@ -gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:128:4,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 --perf_sim_memcpy 1 +-perf_sim_memcpy 0 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -- cgit v1.3 From 8719ea402730f6f429e32d63304f9af79bb83df0 Mon Sep 17 00:00:00 2001 From: Mahmoud Abdallah Date: Sun, 19 Nov 2017 00:29:29 -0500 Subject: improving the accuracy of titanX --- configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) (limited to 'configs') diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 4191eb0..4407870 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -21,7 +21,7 @@ #-gpgpu_clock_domains ::: # Pascal NVIDIA TITAN X clock domains are adopted from # https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:2834.0:1417.0:2500.0 +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 # shader core pipeline config -gpgpu_shader_registers 65536 @@ -105,7 +105,7 @@ -inter_config_file config_fermi_islip.icnt # memory partition latency config --rop_latency 120 +-rop_latency 100 -dram_latency 100 # dram model config @@ -130,8 +130,8 @@ # Use the same GDDR5 timing from hynix H5GQ1H24AFR # disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: + CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" # Pascal GP102 has four schedulers per core -gpgpu_num_sched_per_core 4 -- cgit v1.3 From edb99bbfb39ddd09fc7124cd481dca484a60c5b8 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 13 Mar 2018 12:31:49 -0400 Subject: adding the config file for P102 for gpgpu-sim 3.x --- configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 43 +++++++++------------- 1 file changed, 18 insertions(+), 25 deletions(-) (limited to 'configs') diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 4407870..28689ce 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -33,25 +33,21 @@ # Pipeline widths and number of FUs # ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 4 SFU units +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit ## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,4,1,4,1,4,1,9 +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 -gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 1 +-gpgpu_num_sfu_units 1 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +# SFU is 32-width in pascal, then dp units initiation is 1 cycle -ptx_opcode_latency_int 4,13,4,5,145 -ptx_opcode_initiation_int 1,1,1,1,4 -ptx_opcode_latency_fp 4,13,4,5,39 -ptx_opcode_initiation_fp 1,2,1,1,4 -ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,8,8,8,130 --ptx_opcode_initiation_sfu 4 --ptx_opcode_latency_sfu 8 +-ptx_opcode_initiation_dp 4,8,4,4,130 # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo @@ -59,36 +55,33 @@ # Pascal GP102 has 96KB Shared memory # Pascal GP102 has 64KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 98304 -gmem_skip_L1D 1 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 N:64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 # 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,A:2:48,4 +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 # 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 # 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,A:2:64,4 +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 # enable operand collector ## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 -gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 -gpgpu_operand_collector_num_in_ports_sp 4 -gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 4 --gpgpu_operand_collector_num_out_ports_sfu 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 -gpgpu_operand_collector_num_in_ports_mem 1 -gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 # gpgpu_num_reg_banks should be increased to 32 -gpgpu_num_reg_banks 32 @@ -105,7 +98,7 @@ -inter_config_file config_fermi_islip.icnt # memory partition latency config --rop_latency 100 +-rop_latency 120 -dram_latency 100 # dram model config @@ -130,10 +123,10 @@ # Use the same GDDR5 timing from hynix H5GQ1H24AFR # disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: - CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" -# Pascal GP102 has four schedulers per core +# Fermi has two schedulers per core -gpgpu_num_sched_per_core 4 # Two Level Scheduler with active and pending pools #-gpgpu_scheduler two_level_active:6:0:1 -- cgit v1.3 From f35ae9d24fa14f838352a95bc2c1445821f53449 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 13 Mar 2018 15:04:03 -0400 Subject: adding config file for P100 of gpgpsim3.x --- configs/Pascal-P100-HBM/gpgpusim.config | 6 +- .../gpgpu-sim-3.x/config_fermi_islip.icnt | 70 +++ configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config | 152 ++++++ .../PascalP100X/gpgpu-sim-3.x/gpuwattch_gtx480.xml | 538 +++++++++++++++++++++ 4 files changed, 762 insertions(+), 4 deletions(-) create mode 100644 configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt create mode 100644 configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config create mode 100755 configs/PascalP100X/gpgpu-sim-3.x/gpuwattch_gtx480.xml (limited to 'configs') diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config index 533a865..0c4744a 100644 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -56,7 +56,7 @@ # Pascal GP100 has 64KB Shared memory -gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 65536 --gmem_skip_L1D 1 +-gmem_skip_L1D 0 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 @@ -86,7 +86,6 @@ -gpgpu_operand_collector_num_out_ports_mem 1 -gpgpu_operand_collector_num_in_ports_dp 1 -gpgpu_operand_collector_num_out_ports_dp 1 -# gpgpu_num_reg_banks should be increased to 32 -gpgpu_num_reg_banks 32 # shared memory bankconflict detection @@ -105,8 +104,7 @@ # memory partition latency config -rop_latency 120 -# DRAM latency should be lower compared to other configs, due to high-speed interposer connection --dram_latency 60 +-dram_latency 100 # dram model config -gpgpu_dram_scheduler 1 diff --git a/configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt b/configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt new file mode 100644 index 0000000..602daee --- /dev/null +++ b/configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config new file mode 100644 index 0000000..8afdd7f --- /dev/null +++ b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config @@ -0,0 +1,152 @@ +# This config models the Pascal GP100 +# For more info about this card, see Nvidia White paper +# https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 60 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 32 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Pscal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Nvidia_Tesla +-gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,2,1,2,2,1,4 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 2,2,2,2,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +# Pascal GP100 has 48KB L1 cache +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 65536 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/PascalP100X/gpgpu-sim-3.x/gpuwattch_gtx480.xml b/configs/PascalP100X/gpgpu-sim-3.x/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/PascalP100X/gpgpu-sim-3.x/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + -- cgit v1.3 From a43bb3a68074f4b4be9b50b307dfddfb20386745 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 13 Mar 2018 15:38:55 -0400 Subject: fixing the inter nodes in P102 gpgpusim3.x --- configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'configs') diff --git a/configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt b/configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt index 602daee..d26c8d9 100644 --- a/configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt +++ b/configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt @@ -7,7 +7,7 @@ network_count = 2; // Topology topology = fly; -k = 52; +k = 60; n = 1; // Routing -- cgit v1.3 From bd2700c7a1b5be3e71ee2f0e9abab76719f2f462 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sun, 25 Mar 2018 16:41:51 -0400 Subject: Need to explicity set this or the stupid GPGPU-Sim defaults of 16k kicks in. --- configs/Pascal-P100-HBM/gpgpusim.config | 2 ++ configs/Pascal-P102-GDDR5X/gpgpusim.config | 2 ++ configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config | 2 ++ configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 2 ++ 4 files changed, 8 insertions(+) (limited to 'configs') diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config index 0c4744a..31fbecb 100644 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -56,6 +56,8 @@ # Pascal GP100 has 64KB Shared memory -gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 65536 +-gpgpu_shmem_size_PrefShared 65536 -gmem_skip_L1D 0 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 0c6c126..8e1e56f 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -63,6 +63,8 @@ # The defulat is to disable the L1 cache, unless cache modifieres is used -gpgpu_cache:dl1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 49152 +-gpgpu_shmem_size_PrefL1 49152 +-gpgpu_shmem_size_PrefShared 49152 -gmem_skip_L1D 1 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 diff --git a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config index 8afdd7f..b87d958 100644 --- a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config @@ -56,6 +56,8 @@ # Pascal GP100 has 48KB L1 cache -gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 65536 +-gpgpu_shmem_size_PrefShared 65536 -gmem_skip_L1D 0 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 28689ce..2df24e1 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -57,6 +57,8 @@ # The defulat is to disable the L1 cache, unless cache modifieres is used -gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 98304 +-gpgpu_shmem_size_PrefL1 98304 +-gpgpu_shmem_size_PrefShared 98304 -gmem_skip_L1D 1 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache -- cgit v1.3 From 4e91a60a48b07f41f4bfb4d59fa2355024a3914b Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sun, 25 Mar 2018 16:51:21 -0400 Subject: Need to make sure we don't kill the L1 too --- configs/Pascal-P100-HBM/gpgpusim.config | 2 ++ configs/Pascal-P102-GDDR5X/gpgpusim.config | 2 ++ configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config | 2 ++ configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 2 ++ src/gpgpu-sim/gpu-sim.cc | 2 +- 5 files changed, 9 insertions(+), 1 deletion(-) (limited to 'configs') diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config index 31fbecb..857a0ff 100644 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -55,6 +55,8 @@ # Note: Hashing set index function (H) only applies to a set size of 32 or 64. # Pascal GP100 has 64KB Shared memory -gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 65536 -gpgpu_shmem_size_PrefL1 65536 -gpgpu_shmem_size_PrefShared 65536 diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 8e1e56f..feb8c0b 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -62,6 +62,8 @@ # Pascal GP102 has 24KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used -gpgpu_cache:dl1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 49152 -gpgpu_shmem_size_PrefL1 49152 -gpgpu_shmem_size_PrefShared 49152 diff --git a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config index b87d958..a5e6736 100644 --- a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config @@ -55,6 +55,8 @@ # Pascal GP100 has 64KB Shared memory # Pascal GP100 has 48KB L1 cache -gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 65536 -gpgpu_shmem_size_PrefL1 65536 -gpgpu_shmem_size_PrefShared 65536 diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 2df24e1..f78bd02 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -56,6 +56,8 @@ # Pascal GP102 has 64KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used -gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 98304 -gpgpu_shmem_size_PrefL1 98304 -gpgpu_shmem_size_PrefShared 98304 diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 17f1714..bb448b3 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -245,7 +245,7 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); - option_parser_register(opp, "-gpgpu_cache:dl1PreShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, + option_parser_register(opp, "-gpgpu_cache:dl1PrefShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); -- cgit v1.3 From 1af82fbe97428654b06b7cd8d40c2d2ce4592aae Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 26 Mar 2018 11:35:25 -0400 Subject: Fixing resources limitation assertion for compute cab 61 --- configs/Pascal-P102-GDDR5X/gpgpusim.config | 1 + src/gpgpu-sim/gpu-sim.cc | 3 +++ src/gpgpu-sim/shader.cc | 4 ++++ src/gpgpu-sim/shader.h | 1 + 4 files changed, 9 insertions(+) (limited to 'configs') diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 0c6c126..b7bcdaf 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -6,6 +6,7 @@ -gpgpu_ptx_instruction_classification 0 -gpgpu_ptx_sim_mode 0 -gpgpu_ptx_force_max_capability 61 +-gpgpu_ignore_resources_limitation 1 # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 17f1714..48af41e 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -268,6 +268,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shader_registers", OPT_UINT32, &gpgpu_shader_registers, "Number of registers per shader core. Limits number of concurrent CTAs. (default 8192)", "8192"); + option_parser_register(opp, "-gpgpu_ignore_resources_limitation", OPT_BOOL, &gpgpu_ignore_resources_limitation, + "gpgpu_ignore_resources_limitation (default 0)", + "0"); option_parser_register(opp, "-gpgpu_shader_cta", OPT_UINT32, &max_cta_per_core, "Maximum number of concurrent CTAs in shader (default 8)", "8"); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index bf482fb..8173191 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -2614,6 +2614,10 @@ unsigned int shader_core_config::max_cta( const kernel_info_t &k ) const assert( result <= MAX_CTA_PER_SHADER ); if (result < 1) { printf ("GPGPU-Sim uArch: ERROR ** Kernel requires more resources than shader has.\n"); + if(gpgpu_ignore_resources_limitation) { + printf ("GPGPU-Sim uArch: gpgpu_ignore_resources_limitation is set, ignore the ERROR!\n"); + return 1; + } abort(); } diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 5b41c06..ae22eaa 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1386,6 +1386,7 @@ struct shader_core_config : public core_config unsigned gpgpu_num_reg_banks; bool gpgpu_reg_bank_use_warp_id; bool gpgpu_local_mem_map; + bool gpgpu_ignore_resources_limitation; unsigned max_sp_latency; unsigned max_sfu_latency; -- cgit v1.3 From a5578ff87d0516310d799a8e88e3288c1b44a429 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Tue, 27 Mar 2018 11:59:59 -0400 Subject: Update the write policy in P100 --- configs/Pascal-P100-HBM/gpgpusim.config | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'configs') diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config index 857a0ff..9a7259e 100644 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -65,7 +65,7 @@ -gpgpu_n_cluster_ejection_buffer_size 32 # 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:W:L,A:256:4,32:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:256:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 -perf_sim_memcpy 0 -- cgit v1.3 From f169276e3916d6ffbaef78b0b836c91f1e169044 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 15 May 2018 19:57:44 -0400 Subject: adding Volta model configs --- .../gpgpu-sim-3.x/config_fermi_islip.icnt | 70 ++++++++ .../gpgpu-sim-3.x/gpgpusim.config | 166 ++++++++++++++++++ .../gpgpu-sim-4.x/config_fermi_islip.icnt | 74 ++++++++ .../gpgpu-sim-4.x/gpgpusim.config | 186 +++++++++++++++++++++ 4 files changed, 496 insertions(+) create mode 100644 configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt create mode 100644 configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config create mode 100644 configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/config_fermi_islip.icnt create mode 100644 configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config (limited to 'configs') diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt b/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt new file mode 100644 index 0000000..d26c8d9 --- /dev/null +++ b/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 60; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config b/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config new file mode 100644 index 0000000..c0406e9 --- /dev/null +++ b/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config @@ -0,0 +1,166 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +-gpgpu_cache:dl1 S:64:128:8,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:64:128:16,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 1 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 S:64:128:24,L:B:m:F:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/config_fermi_islip.icnt b/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/config_fermi_islip.icnt new file mode 100644 index 0000000..616e9f3 --- /dev/null +++ b/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/config_fermi_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 128; +input_buffer_size = 256; +ejection_buffer_size = 128; +boundary_buffer_size = 128; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config b/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config new file mode 100644 index 0000000..3d8c161 --- /dev/null +++ b/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config @@ -0,0 +1,186 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +-gpgpu_cache:dl1 S:64:128:8,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:64:128:16,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 1 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 S:64:128:24,L:B:m:F:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 70 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + -- cgit v1.3 From e06cd8efa02bf54f4788b0655930c9c6c0e12311 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 15 May 2018 19:58:48 -0400 Subject: fixnig icnt volta --- configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'configs') diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt b/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt index d26c8d9..fac792a 100644 --- a/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt +++ b/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt @@ -7,7 +7,7 @@ network_count = 2; // Topology topology = fly; -k = 60; +k = 64; n = 1; // Routing -- cgit v1.3 From c2b43f1ff592563c1abf8bc1ac3230f2892b380a Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 15 May 2018 20:17:27 -0400 Subject: 60 coals mode --- configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'configs') diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config b/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config index 3d8c161..8969168 100644 --- a/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config +++ b/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config @@ -109,7 +109,7 @@ -gpgpu_shmem_num_banks 32 -gpgpu_shmem_limited_broadcast 0 -gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 70 +-gpgpu_coalesce_arch 60 ## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units -gpgpu_max_insn_issue_per_warp 1 -- cgit v1.3 From 8738aa4cde553dcf3b3315f4e7e61f1fa4faeadd Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 15 May 2018 20:29:42 -0400 Subject: fixing vache for volta --- .../Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config | 17 ++++++++--------- 1 file changed, 8 insertions(+), 9 deletions(-) (limited to 'configs') diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config b/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config index c0406e9..f1ae2dc 100644 --- a/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config @@ -43,10 +43,9 @@ # ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB ## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core ## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_pipeline_widths 4,4,1,4,4,1,9 -gpgpu_num_sp_units 4 -gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" @@ -64,25 +63,25 @@ # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. # Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:8,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:16,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 -gpgpu_shmem_size 65536 -gpgpu_shmem_size_PrefL1 1 -gpgpu_shmem_size_PrefShared 98304 -gmem_skip_L1D 0 # 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,L:B:m:F:L,A:256:4,32:0,32 +-gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 # 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +-gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 # 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 # 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 +-gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 # enable operand collector -gpgpu_operand_collector_num_units_sp 14 -- cgit v1.3 From 97fcd60a4fdef22d3ab89eb660ec4cdc07553f72 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Wed, 16 May 2018 18:20:14 -0400 Subject: Moving all the 3.x cfgs to a new subdirectory --- .../3.x-cfgs/PASCAL_P100/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/PASCAL_P100/gpgpusim.config | 156 ++++++ configs/3.x-cfgs/PASCAL_P100/gpuwattch_gtx480.xml | 538 +++++++++++++++++++++ .../3.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/PASCAL_TITANX/gpgpusim.config | 156 ++++++ .../3.x-cfgs/PASCAL_TITANX/gpuwattch_gtx480.xml | 538 +++++++++++++++++++++ .../3.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/VOLTA_TITANV/gpgpusim.config | 165 +++++++ .../gpgpu-sim-3.x/config_fermi_islip.icnt | 70 --- .../gpgpu-sim-3.x/gpgpusim.config | 165 ------- .../gpgpu-sim-3.x/config_fermi_islip.icnt | 70 --- configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config | 156 ------ .../PascalP100X/gpgpu-sim-3.x/gpuwattch_gtx480.xml | 538 --------------------- .../gpgpu-sim-3.x/config_fermi_islip.icnt | 70 --- configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 156 ------ .../gpgpu-sim-3.x/gpuwattch_gtx480.xml | 538 --------------------- 16 files changed, 1763 insertions(+), 1763 deletions(-) create mode 100644 configs/3.x-cfgs/PASCAL_P100/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/PASCAL_P100/gpgpusim.config create mode 100755 configs/3.x-cfgs/PASCAL_P100/gpuwattch_gtx480.xml create mode 100644 configs/3.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/PASCAL_TITANX/gpgpusim.config create mode 100755 configs/3.x-cfgs/PASCAL_TITANX/gpuwattch_gtx480.xml create mode 100644 configs/3.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/VOLTA_TITANV/gpgpusim.config delete mode 100644 configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt delete mode 100644 configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config delete mode 100644 configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt delete mode 100644 configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config delete mode 100755 configs/PascalP100X/gpgpu-sim-3.x/gpuwattch_gtx480.xml delete mode 100644 configs/PascalTitanX/gpgpu-sim-3.x/config_fermi_islip.icnt delete mode 100644 configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config delete mode 100755 configs/PascalTitanX/gpgpu-sim-3.x/gpuwattch_gtx480.xml (limited to 'configs') diff --git a/configs/3.x-cfgs/PASCAL_P100/config_fermi_islip.icnt b/configs/3.x-cfgs/PASCAL_P100/config_fermi_islip.icnt new file mode 100644 index 0000000..d26c8d9 --- /dev/null +++ b/configs/3.x-cfgs/PASCAL_P100/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 60; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/PASCAL_P100/gpgpusim.config b/configs/3.x-cfgs/PASCAL_P100/gpgpusim.config new file mode 100644 index 0000000..a5e6736 --- /dev/null +++ b/configs/3.x-cfgs/PASCAL_P100/gpgpusim.config @@ -0,0 +1,156 @@ +# This config models the Pascal GP100 +# For more info about this card, see Nvidia White paper +# https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 60 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 32 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Pscal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Nvidia_Tesla +-gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,2,1,2,2,1,4 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 2,2,2,2,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +# Pascal GP100 has 48KB L1 cache +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 65536 +-gpgpu_shmem_size_PrefShared 65536 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/3.x-cfgs/PASCAL_P100/gpuwattch_gtx480.xml b/configs/3.x-cfgs/PASCAL_P100/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/3.x-cfgs/PASCAL_P100/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt b/configs/3.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt new file mode 100644 index 0000000..602daee --- /dev/null +++ b/configs/3.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/PASCAL_TITANX/gpgpusim.config b/configs/3.x-cfgs/PASCAL_TITANX/gpgpusim.config new file mode 100644 index 0000000..f78bd02 --- /dev/null +++ b/configs/3.x-cfgs/PASCAL_TITANX/gpgpusim.config @@ -0,0 +1,156 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,8,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 98304 +-gpgpu_shmem_size_PrefL1 98304 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +# gpgpu_num_reg_banks should be increased to 32 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/3.x-cfgs/PASCAL_TITANX/gpuwattch_gtx480.xml b/configs/3.x-cfgs/PASCAL_TITANX/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/3.x-cfgs/PASCAL_TITANX/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt b/configs/3.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..fac792a --- /dev/null +++ b/configs/3.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/VOLTA_TITANV/gpgpusim.config b/configs/3.x-cfgs/VOLTA_TITANV/gpgpusim.config new file mode 100644 index 0000000..f1ae2dc --- /dev/null +++ b/configs/3.x-cfgs/VOLTA_TITANV/gpgpusim.config @@ -0,0 +1,165 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,1,4,4,1,9 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +-gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 1 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 + +# 128 KB Inst. +-gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt b/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt deleted file mode 100644 index fac792a..0000000 --- a/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config b/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config deleted file mode 100644 index f1ae2dc..0000000 --- a/configs/Nvidia-Titan-Volta/gpgpu-sim-3.x/gpgpusim.config +++ /dev/null @@ -1,165 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,1,4,4,1,9 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 --gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 --gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 1 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 0 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 - -# 128 KB Inst. --gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 14 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_units_mem 10 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt b/configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt deleted file mode 100644 index d26c8d9..0000000 --- a/configs/PascalP100X/gpgpu-sim-3.x/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 60; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config deleted file mode 100644 index a5e6736..0000000 --- a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config +++ /dev/null @@ -1,156 +0,0 @@ -# This config models the Pascal GP100 -# For more info about this card, see Nvidia White paper -# https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 32 --gpgpu_n_sub_partition_per_mchannel 1 - -# Pscal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Nvidia_Tesla --gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,2,1,2,2,1,4 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 2,2,2,2,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory -# Pascal GP100 has 48KB L1 cache --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 65536 --gpgpu_shmem_size_PrefShared 65536 --gmem_skip_L1D 0 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 14 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_units_mem 10 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/PascalP100X/gpgpu-sim-3.x/gpuwattch_gtx480.xml b/configs/PascalP100X/gpgpu-sim-3.x/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/PascalP100X/gpgpu-sim-3.x/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/config_fermi_islip.icnt b/configs/PascalTitanX/gpgpu-sim-3.x/config_fermi_islip.icnt deleted file mode 100644 index 602daee..0000000 --- a/configs/PascalTitanX/gpgpu-sim-3.x/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config deleted file mode 100644 index f78bd02..0000000 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ /dev/null @@ -1,156 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,8,4,4,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_shmem_size 98304 --gpgpu_shmem_size_PrefL1 98304 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 -# gpgpu_num_reg_banks should be increased to 32 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpuwattch_gtx480.xml b/configs/PascalTitanX/gpgpu-sim-3.x/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -- cgit v1.3 From 355c86895faf122deecb5de5ef17ff4ff9d654d9 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Wed, 16 May 2018 19:20:53 -0400 Subject: More sane naming convention --- configs/4.x-cfgs/GTX1080/config_fermi_islip.icnt | 70 +++ configs/4.x-cfgs/GTX1080/gpgpusim.config | 149 ++++++ configs/4.x-cfgs/GTX1080/gpuwattch_gtx1080Ti.xml | 538 ++++++++++++++++++++ configs/4.x-cfgs/GTX480/config_fermi_islip.icnt | 70 +++ configs/4.x-cfgs/GTX480/gpgpusim.config | 135 ++++++ configs/4.x-cfgs/GTX480/gpuwattch_gtx480.xml | 538 ++++++++++++++++++++ configs/4.x-cfgs/GTX750/config_fermi_islip.icnt | 70 +++ configs/4.x-cfgs/GTX750/gpgpusim.config | 131 +++++ configs/4.x-cfgs/GTX750/gpuwattch_gtx750Ti.xml | 539 +++++++++++++++++++++ .../4.x-cfgs/PASCAL_P100/config_fermi_islip.icnt | 73 +++ configs/4.x-cfgs/PASCAL_P100/gpgpusim.config | 173 +++++++ .../4.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt | 73 +++ configs/4.x-cfgs/PASCAL_TITANX/gpgpusim.config | 178 +++++++ configs/4.x-cfgs/QuadroFX5600/gpgpusim.config | 97 ++++ .../QuadroFX5600/gpuwattch_quadrofx5600.xml | 538 ++++++++++++++++++++ .../4.x-cfgs/QuadroFX5600/icnt_config_islip.icnt | 70 +++ .../4.x-cfgs/QuadroFX5800/config_quadro_islip.icnt | 69 +++ configs/4.x-cfgs/QuadroFX5800/gpgpusim.config | 87 ++++ .../4.x-cfgs/TeslaC2050/config_fermi_islip.icnt | 70 +++ configs/4.x-cfgs/TeslaC2050/gpgpusim.config | 133 +++++ .../4.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt | 74 +++ configs/4.x-cfgs/VOLTA_TITANV/gpgpusim.config | 186 +++++++ configs/GTX480/config_fermi_islip.icnt | 70 --- configs/GTX480/gpgpusim.config | 135 ------ configs/GTX480/gpuwattch_gtx480.xml | 538 -------------------- configs/GeForceGTX1080Ti/config_fermi_islip.icnt | 70 --- configs/GeForceGTX1080Ti/gpgpusim.config | 149 ------ configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml | 538 -------------------- configs/GeForceGTX750Ti/config_fermi_islip.icnt | 70 --- configs/GeForceGTX750Ti/gpgpusim.config | 131 ----- configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml | 539 --------------------- .../gpgpu-sim-4.x/config_fermi_islip.icnt | 74 --- .../gpgpu-sim-4.x/gpgpusim.config | 186 ------- configs/Pascal-P100-HBM/config_fermi_islip.icnt | 73 --- configs/Pascal-P100-HBM/gpgpusim.config | 173 ------- configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt | 73 --- configs/Pascal-P102-GDDR5X/gpgpusim.config | 178 ------- configs/QuadroFX5600/gpgpusim.config | 97 ---- configs/QuadroFX5600/gpuwattch_quadrofx5600.xml | 538 -------------------- configs/QuadroFX5600/icnt_config_islip.icnt | 70 --- configs/QuadroFX5800/config_quadro_islip.icnt | 69 --- configs/QuadroFX5800/gpgpusim.config | 87 ---- configs/TeslaC2050/config_fermi_islip.icnt | 70 --- configs/TeslaC2050/gpgpusim.config | 133 ----- 44 files changed, 4061 insertions(+), 4061 deletions(-) create mode 100644 configs/4.x-cfgs/GTX1080/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/GTX1080/gpgpusim.config create mode 100755 configs/4.x-cfgs/GTX1080/gpuwattch_gtx1080Ti.xml create mode 100644 configs/4.x-cfgs/GTX480/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/GTX480/gpgpusim.config create mode 100755 configs/4.x-cfgs/GTX480/gpuwattch_gtx480.xml create mode 100644 configs/4.x-cfgs/GTX750/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/GTX750/gpgpusim.config create mode 100755 configs/4.x-cfgs/GTX750/gpuwattch_gtx750Ti.xml create mode 100644 configs/4.x-cfgs/PASCAL_P100/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/PASCAL_P100/gpgpusim.config create mode 100644 configs/4.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/PASCAL_TITANX/gpgpusim.config create mode 100644 configs/4.x-cfgs/QuadroFX5600/gpgpusim.config create mode 100644 configs/4.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml create mode 100644 configs/4.x-cfgs/QuadroFX5600/icnt_config_islip.icnt create mode 100644 configs/4.x-cfgs/QuadroFX5800/config_quadro_islip.icnt create mode 100644 configs/4.x-cfgs/QuadroFX5800/gpgpusim.config create mode 100644 configs/4.x-cfgs/TeslaC2050/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/TeslaC2050/gpgpusim.config create mode 100644 configs/4.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/VOLTA_TITANV/gpgpusim.config delete mode 100644 configs/GTX480/config_fermi_islip.icnt delete mode 100644 configs/GTX480/gpgpusim.config delete mode 100755 configs/GTX480/gpuwattch_gtx480.xml delete mode 100644 configs/GeForceGTX1080Ti/config_fermi_islip.icnt delete mode 100644 configs/GeForceGTX1080Ti/gpgpusim.config delete mode 100755 configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml delete mode 100644 configs/GeForceGTX750Ti/config_fermi_islip.icnt delete mode 100644 configs/GeForceGTX750Ti/gpgpusim.config delete mode 100755 configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml delete mode 100644 configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/config_fermi_islip.icnt delete mode 100644 configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config delete mode 100644 configs/Pascal-P100-HBM/config_fermi_islip.icnt delete mode 100644 configs/Pascal-P100-HBM/gpgpusim.config delete mode 100644 configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt delete mode 100644 configs/Pascal-P102-GDDR5X/gpgpusim.config delete mode 100644 configs/QuadroFX5600/gpgpusim.config delete mode 100644 configs/QuadroFX5600/gpuwattch_quadrofx5600.xml delete mode 100644 configs/QuadroFX5600/icnt_config_islip.icnt delete mode 100644 configs/QuadroFX5800/config_quadro_islip.icnt delete mode 100644 configs/QuadroFX5800/gpgpusim.config delete mode 100644 configs/TeslaC2050/config_fermi_islip.icnt delete mode 100644 configs/TeslaC2050/gpgpusim.config (limited to 'configs') diff --git a/configs/4.x-cfgs/GTX1080/config_fermi_islip.icnt b/configs/4.x-cfgs/GTX1080/config_fermi_islip.icnt new file mode 100644 index 0000000..2a69ddd --- /dev/null +++ b/configs/4.x-cfgs/GTX1080/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 50; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/GTX1080/gpgpusim.config b/configs/4.x-cfgs/GTX1080/gpgpusim.config new file mode 100644 index 0000000..47c2b6a --- /dev/null +++ b/configs/4.x-cfgs/GTX1080/gpgpusim.config @@ -0,0 +1,149 @@ +# This config models the Pascal GP102 (GeForceGTX 1080Ti) + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 11 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 1,2,1,1,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The default is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 +-gpgpu_cache:dl2_texture_only 0 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +# gpgpu_num_reg_banks should be increased to 32, but it gives an error! +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) +# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Pascal has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx1080Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/GTX1080/gpuwattch_gtx1080Ti.xml b/configs/4.x-cfgs/GTX1080/gpuwattch_gtx1080Ti.xml new file mode 100755 index 0000000..02619ff --- /dev/null +++ b/configs/4.x-cfgs/GTX1080/gpuwattch_gtx1080Ti.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/4.x-cfgs/GTX480/config_fermi_islip.icnt b/configs/4.x-cfgs/GTX480/config_fermi_islip.icnt new file mode 100644 index 0000000..7820e4e --- /dev/null +++ b/configs/4.x-cfgs/GTX480/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 27; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/GTX480/gpgpusim.config b/configs/4.x-cfgs/GTX480/gpgpusim.config new file mode 100644 index 0000000..03fcda1 --- /dev/null +++ b/configs/4.x-cfgs/GTX480/gpgpusim.config @@ -0,0 +1,135 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 15 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 700.0:700.0:700.0:924.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +#For Fermi, DP unit =0, DP inst is executed on SFU +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 116 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/GTX480/gpuwattch_gtx480.xml b/configs/4.x-cfgs/GTX480/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/4.x-cfgs/GTX480/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/4.x-cfgs/GTX750/config_fermi_islip.icnt b/configs/4.x-cfgs/GTX750/config_fermi_islip.icnt new file mode 100644 index 0000000..069ca02 --- /dev/null +++ b/configs/4.x-cfgs/GTX750/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 7; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/GTX750/gpgpusim.config b/configs/4.x-cfgs/GTX750/gpgpusim.config new file mode 100644 index 0000000..9366f93 --- /dev/null +++ b/configs/4.x-cfgs/GTX750/gpgpusim.config @@ -0,0 +1,131 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 52 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 5 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 2 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Maxwell clock domains +#-gpgpu_clock_domains ::: +# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. +-gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +-gpgpu_num_sp_units 8 +-gpgpu_num_sfu_units 32 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 6,12,13,13,210 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 6,12,6,6,374 +-ptx_opcode_initiation_fp 1,1,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 +-gmem_skip_L1D 1 +-gpgpu_shmem_size 65536 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache +-gpgpu_cache:dl2 N:1024:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 N:16:128:32,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 150 +-dram_latency 130 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 300 + +# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 32 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Maxwell has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. +# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx750Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/GTX750/gpuwattch_gtx750Ti.xml b/configs/4.x-cfgs/GTX750/gpuwattch_gtx750Ti.xml new file mode 100755 index 0000000..e2b2324 --- /dev/null +++ b/configs/4.x-cfgs/GTX750/gpuwattch_gtx750Ti.xml @@ -0,0 +1,539 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/4.x-cfgs/PASCAL_P100/config_fermi_islip.icnt b/configs/4.x-cfgs/PASCAL_P100/config_fermi_islip.icnt new file mode 100644 index 0000000..e7c2c3b --- /dev/null +++ b/configs/4.x-cfgs/PASCAL_P100/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 60; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 128; +input_buffer_size = 256; +ejection_buffer_size = 128; +boundary_buffer_size = 128; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/PASCAL_P100/gpgpusim.config b/configs/4.x-cfgs/PASCAL_P100/gpgpusim.config new file mode 100644 index 0000000..9a7259e --- /dev/null +++ b/configs/4.x-cfgs/PASCAL_P100/gpgpusim.config @@ -0,0 +1,173 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 60 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 32 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Pscal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Nvidia_Tesla +-gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,2,2,1,2,2,2,1,6 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 2 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 2,2,2,2,130 +-ptx_opcode_latency_sfu 8 +-ptx_opcode_initiation_sfu 4 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +-gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 65536 +-gpgpu_shmem_size_PrefShared 65536 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 + +# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 + +# 4 KB Inst. +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt b/configs/4.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt new file mode 100644 index 0000000..714d933 --- /dev/null +++ b/configs/4.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; +input_buffer_size = 256; +ejection_buffer_size = 64; +boundary_buffer_size = 64; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/PASCAL_TITANX/gpgpusim.config b/configs/4.x-cfgs/PASCAL_TITANX/gpgpusim.config new file mode 100644 index 0000000..3097d19 --- /dev/null +++ b/configs/4.x-cfgs/PASCAL_TITANX/gpgpusim.config @@ -0,0 +1,178 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 +-gpgpu_ignore_resources_limitation 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +# P102 has two semi-indp scheds per core, and two cores per cluster +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 32 warps/SM +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 16 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 4 SFU units +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 1 + + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,8,8,8,130 +-ptx_opcode_initiation_sfu 4 +-ptx_opcode_latency_sfu 8 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 24KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_shmem_size 49152 +-gpgpu_shmem_size_PrefL1 49152 +-gpgpu_shmem_size_PrefShared 49152 +-gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:128:4,16:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 +-perf_sim_memcpy 0 + +# 4 KB Inst. +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# this is unused +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_operand_collector_num_in_ports_sfu 2 +-gpgpu_operand_collector_num_out_ports_sfu 2 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +# gpgpu_num_reg_banks should be increased to 32 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +# Use Fermi Coalsce arhitetecture which is the same as Pascal +-gpgpu_coalesce_arch 61 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 100 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 240 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: + CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" + +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal 102 has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 102 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/QuadroFX5600/gpgpusim.config b/configs/4.x-cfgs/QuadroFX5600/gpgpusim.config new file mode 100644 index 0000000..e3cab18 --- /dev/null +++ b/configs/4.x-cfgs/QuadroFX5600/gpgpusim.config @@ -0,0 +1,97 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 12 + +# high level architecture configuration +-gpgpu_n_clusters 8 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 6 +-gpgpu_clock_domains 337.5:600.0:600.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) +-gpgpu_shader_core_pipeline 768:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +-gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +# TLB parameters +#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 +#-gpgpu_tlbl2_latency 45 + + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file icnt_config_islip.icnt + +# dram scheduler config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 55 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + + + +# Using cuobjdump to extract ptx/SASS +-gpgpu_ptx_use_cuobjdump 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 +-power_trace_enabled 0 +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_quadrofx5600.xml + +-steady_power_levels_enabled 1 +-steady_state_definition 8,4 diff --git a/configs/4.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml b/configs/4.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml new file mode 100644 index 0000000..2c5a6fc --- /dev/null +++ b/configs/4.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/4.x-cfgs/QuadroFX5600/icnt_config_islip.icnt b/configs/4.x-cfgs/QuadroFX5600/icnt_config_islip.icnt new file mode 100644 index 0000000..de3bcc8 --- /dev/null +++ b/configs/4.x-cfgs/QuadroFX5600/icnt_config_islip.icnt @@ -0,0 +1,70 @@ +//14*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 14; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/QuadroFX5800/config_quadro_islip.icnt b/configs/4.x-cfgs/QuadroFX5800/config_quadro_islip.icnt new file mode 100644 index 0000000..cfe9cac --- /dev/null +++ b/configs/4.x-cfgs/QuadroFX5800/config_quadro_islip.icnt @@ -0,0 +1,69 @@ +//18*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 18; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/QuadroFX5800/gpgpusim.config b/configs/4.x-cfgs/QuadroFX5800/gpgpusim.config new file mode 100644 index 0000000..56dbb17 --- /dev/null +++ b/configs/4.x-cfgs/QuadroFX5800/gpgpusim.config @@ -0,0 +1,87 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 13 + +# high level architecture configuration +-gpgpu_n_clusters 10 +-gpgpu_n_cores_per_cluster 3 +-gpgpu_n_mem 8 +-gpgpu_clock_domains 325.0:650.0:650.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +-gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file config_quadro_islip.icnt + +# dram scheduler config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 325MHz = 74 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 58 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-dram_data_command_freq_ratio 2 # GDDR3 is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/TeslaC2050/config_fermi_islip.icnt b/configs/4.x-cfgs/TeslaC2050/config_fermi_islip.icnt new file mode 100644 index 0000000..a11bd8e --- /dev/null +++ b/configs/4.x-cfgs/TeslaC2050/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//20*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 26; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; diff --git a/configs/4.x-cfgs/TeslaC2050/gpgpusim.config b/configs/4.x-cfgs/TeslaC2050/gpgpusim.config new file mode 100644 index 0000000..aa5f5f3 --- /dev/null +++ b/configs/4.x-cfgs/TeslaC2050/gpgpusim.config @@ -0,0 +1,133 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + +# Using cuobjdump to extract ptx/SASS +#-gpgpu_ptx_use_cuobjdump 1 # use default + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + + +# high level architecture configuration +-gpgpu_n_clusters 14 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 575.0:575.0:575.0:750.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 750MHz / 575MHz = 130 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 114 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt b/configs/4.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..616e9f3 --- /dev/null +++ b/configs/4.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 128; +input_buffer_size = 256; +ejection_buffer_size = 128; +boundary_buffer_size = 128; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/VOLTA_TITANV/gpgpusim.config b/configs/4.x-cfgs/VOLTA_TITANV/gpgpusim.config new file mode 100644 index 0000000..8969168 --- /dev/null +++ b/configs/4.x-cfgs/VOLTA_TITANV/gpgpusim.config @@ -0,0 +1,186 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +-gpgpu_cache:dl1 S:64:128:8,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:64:128:16,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 1 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 S:64:128:24,L:B:m:F:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/GTX480/config_fermi_islip.icnt b/configs/GTX480/config_fermi_islip.icnt deleted file mode 100644 index 7820e4e..0000000 --- a/configs/GTX480/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 27; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/GTX480/gpgpusim.config b/configs/GTX480/gpgpusim.config deleted file mode 100644 index 03fcda1..0000000 --- a/configs/GTX480/gpgpusim.config +++ /dev/null @@ -1,135 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 15 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 700.0:700.0:700.0:924.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -#For Fermi, DP unit =0, DP inst is executed on SFU --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 116 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/GTX480/gpuwattch_gtx480.xml b/configs/GTX480/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/GTX480/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/GeForceGTX1080Ti/config_fermi_islip.icnt b/configs/GeForceGTX1080Ti/config_fermi_islip.icnt deleted file mode 100644 index 2a69ddd..0000000 --- a/configs/GeForceGTX1080Ti/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 50; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/GeForceGTX1080Ti/gpgpusim.config b/configs/GeForceGTX1080Ti/gpgpusim.config deleted file mode 100644 index 47c2b6a..0000000 --- a/configs/GeForceGTX1080Ti/gpgpusim.config +++ /dev/null @@ -1,149 +0,0 @@ -# This config models the Pascal GP102 (GeForceGTX 1080Ti) - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 11 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 1,2,1,1,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache -# The default is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 --gpgpu_shmem_size 98304 --gmem_skip_L1D 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 --gpgpu_cache:dl2_texture_only 0 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 -# gpgpu_num_reg_banks should be increased to 32, but it gives an error! --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) -# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Pascal has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx1080Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml b/configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml deleted file mode 100755 index 02619ff..0000000 --- a/configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/GeForceGTX750Ti/config_fermi_islip.icnt b/configs/GeForceGTX750Ti/config_fermi_islip.icnt deleted file mode 100644 index 069ca02..0000000 --- a/configs/GeForceGTX750Ti/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 7; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/GeForceGTX750Ti/gpgpusim.config b/configs/GeForceGTX750Ti/gpgpusim.config deleted file mode 100644 index 9366f93..0000000 --- a/configs/GeForceGTX750Ti/gpgpusim.config +++ /dev/null @@ -1,131 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 52 - - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 5 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 2 --gpgpu_n_sub_partition_per_mchannel 1 - -# Maxwell clock domains -#-gpgpu_clock_domains ::: -# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. --gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 8 --gpgpu_num_sfu_units 32 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 6,12,13,13,210 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 6,12,6,6,374 --ptx_opcode_initiation_fp 1,1,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 --gmem_skip_L1D 1 --gpgpu_shmem_size 65536 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache --gpgpu_cache:dl2 N:1024:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:16:128:32,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 150 --dram_latency 130 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 300 - -# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 32 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Maxwell has four schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs -# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. -# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx750Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml b/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml deleted file mode 100755 index e2b2324..0000000 --- a/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml +++ /dev/null @@ -1,539 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/config_fermi_islip.icnt b/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/config_fermi_islip.icnt deleted file mode 100644 index 616e9f3..0000000 --- a/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/config_fermi_islip.icnt +++ /dev/null @@ -1,74 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 128; -input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config b/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config deleted file mode 100644 index 8969168..0000000 --- a/configs/Nvidia-Titan-Volta/gpgpu-sim-4.x/gpgpusim.config +++ /dev/null @@ -1,186 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 100 --ptx_opcode_initiation_sfu 8 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:8,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:16,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 1 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,L:B:m:F:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/Pascal-P100-HBM/config_fermi_islip.icnt b/configs/Pascal-P100-HBM/config_fermi_islip.icnt deleted file mode 100644 index e7c2c3b..0000000 --- a/configs/Pascal-P100-HBM/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 60; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 128; -input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config deleted file mode 100644 index 9a7259e..0000000 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ /dev/null @@ -1,173 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 32 --gpgpu_n_sub_partition_per_mchannel 1 - -# Pscal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Nvidia_Tesla --gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,2,2,1,2,2,2,1,6 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 2 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 2,2,2,2,130 --ptx_opcode_latency_sfu 8 --ptx_opcode_initiation_sfu 4 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 65536 --gpgpu_shmem_size_PrefShared 65536 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt deleted file mode 100644 index 714d933..0000000 --- a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 64; -input_buffer_size = 256; -ejection_buffer_size = 64; -boundary_buffer_size = 64; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config deleted file mode 100644 index 3097d19..0000000 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ /dev/null @@ -1,178 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 --gpgpu_ignore_resources_limitation 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration -# P102 has two semi-indp scheds per core, and two cores per cluster --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 - -# This implies a maximum of 32 warps/SM --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 16 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 4 SFU units -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 1 - - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,8,8,8,130 --ptx_opcode_initiation_sfu 4 --ptx_opcode_latency_sfu 8 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 24KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 49152 --gpgpu_shmem_size_PrefL1 49152 --gpgpu_shmem_size_PrefShared 49152 --gmem_skip_L1D 1 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:128:4,16:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 --perf_sim_memcpy 0 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# this is unused --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_operand_collector_num_in_ports_sfu 2 --gpgpu_operand_collector_num_out_ports_sfu 2 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 -# gpgpu_num_reg_banks should be increased to 32 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 -# Use Fermi Coalsce arhitetecture which is the same as Pascal --gpgpu_coalesce_arch 61 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 100 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 240 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: - CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" - --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal 102 has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 102 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/QuadroFX5600/gpgpusim.config b/configs/QuadroFX5600/gpgpusim.config deleted file mode 100644 index e3cab18..0000000 --- a/configs/QuadroFX5600/gpgpusim.config +++ /dev/null @@ -1,97 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 12 - -# high level architecture configuration --gpgpu_n_clusters 8 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 6 --gpgpu_clock_domains 337.5:600.0:600.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 -#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) --gpgpu_shader_core_pipeline 768:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour --gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - -# TLB parameters -#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 -#-gpgpu_tlbl2_latency 45 - - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file icnt_config_islip.icnt - -# dram scheduler config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 55 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz -# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - - - -# Using cuobjdump to extract ptx/SASS --gpgpu_ptx_use_cuobjdump 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 --power_trace_enabled 0 --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_quadrofx5600.xml - --steady_power_levels_enabled 1 --steady_state_definition 8,4 diff --git a/configs/QuadroFX5600/gpuwattch_quadrofx5600.xml b/configs/QuadroFX5600/gpuwattch_quadrofx5600.xml deleted file mode 100644 index 2c5a6fc..0000000 --- a/configs/QuadroFX5600/gpuwattch_quadrofx5600.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/QuadroFX5600/icnt_config_islip.icnt b/configs/QuadroFX5600/icnt_config_islip.icnt deleted file mode 100644 index de3bcc8..0000000 --- a/configs/QuadroFX5600/icnt_config_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//14*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 14; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/QuadroFX5800/config_quadro_islip.icnt b/configs/QuadroFX5800/config_quadro_islip.icnt deleted file mode 100644 index cfe9cac..0000000 --- a/configs/QuadroFX5800/config_quadro_islip.icnt +++ /dev/null @@ -1,69 +0,0 @@ -//18*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 18; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/QuadroFX5800/gpgpusim.config b/configs/QuadroFX5800/gpgpusim.config deleted file mode 100644 index 56dbb17..0000000 --- a/configs/QuadroFX5800/gpgpusim.config +++ /dev/null @@ -1,87 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 13 - -# high level architecture configuration --gpgpu_n_clusters 10 --gpgpu_n_cores_per_cluster 3 --gpgpu_n_mem 8 --gpgpu_clock_domains 325.0:650.0:650.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo --gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file config_quadro_islip.icnt - -# dram scheduler config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 325MHz = 74 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 58 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --dram_data_command_freq_ratio 2 # GDDR3 is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/TeslaC2050/config_fermi_islip.icnt b/configs/TeslaC2050/config_fermi_islip.icnt deleted file mode 100644 index a11bd8e..0000000 --- a/configs/TeslaC2050/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//20*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 26; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; diff --git a/configs/TeslaC2050/gpgpusim.config b/configs/TeslaC2050/gpgpusim.config deleted file mode 100644 index aa5f5f3..0000000 --- a/configs/TeslaC2050/gpgpusim.config +++ /dev/null @@ -1,133 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - -# Using cuobjdump to extract ptx/SASS -#-gpgpu_ptx_use_cuobjdump 1 # use default - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - - -# high level architecture configuration --gpgpu_n_clusters 14 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 575.0:575.0:575.0:750.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 750MHz / 575MHz = 130 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 114 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 -- cgit v1.3 From 06845c73e513082262dff97ab3828c6c67ee2af2 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 19 May 2018 13:06:25 -0400 Subject: Config rename --- .../3.x-cfgs/PASCAL_P100/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/PASCAL_P100/gpgpusim.config | 156 ------ configs/3.x-cfgs/PASCAL_P100/gpuwattch_gtx480.xml | 538 -------------------- .../3.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/PASCAL_TITANX/gpgpusim.config | 156 ------ .../3.x-cfgs/PASCAL_TITANX/gpuwattch_gtx480.xml | 538 -------------------- configs/3.x-cfgs/SM6_P100/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/SM6_P100/gpgpusim.config | 156 ++++++ configs/3.x-cfgs/SM6_P100/gpuwattch_gtx480.xml | 538 ++++++++++++++++++++ .../3.x-cfgs/SM6_TITANX/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/SM6_TITANX/gpgpusim.config | 156 ++++++ configs/3.x-cfgs/SM6_TITANX/gpuwattch_gtx480.xml | 538 ++++++++++++++++++++ .../3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/SM7_TITANV/gpgpusim.config | 165 +++++++ .../3.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/VOLTA_TITANV/gpgpusim.config | 165 ------- configs/4.x-cfgs/GTX1080/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/GTX1080/gpgpusim.config | 149 ------ configs/4.x-cfgs/GTX1080/gpuwattch_gtx1080Ti.xml | 538 -------------------- configs/4.x-cfgs/GTX480/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/GTX480/gpgpusim.config | 135 ------ configs/4.x-cfgs/GTX480/gpuwattch_gtx480.xml | 538 -------------------- configs/4.x-cfgs/GTX750/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/GTX750/gpgpusim.config | 131 ----- configs/4.x-cfgs/GTX750/gpuwattch_gtx750Ti.xml | 539 --------------------- .../4.x-cfgs/PASCAL_P100/config_fermi_islip.icnt | 73 --- configs/4.x-cfgs/PASCAL_P100/gpgpusim.config | 173 ------- .../4.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt | 73 --- configs/4.x-cfgs/PASCAL_TITANX/gpgpusim.config | 178 ------- configs/4.x-cfgs/QuadroFX5600/gpgpusim.config | 97 ---- .../QuadroFX5600/gpuwattch_quadrofx5600.xml | 538 -------------------- .../4.x-cfgs/QuadroFX5600/icnt_config_islip.icnt | 70 --- .../4.x-cfgs/QuadroFX5800/config_quadro_islip.icnt | 69 --- configs/4.x-cfgs/QuadroFX5800/gpgpusim.config | 87 ---- configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config | 97 ++++ .../SM1_QFX5600/gpuwattch_quadrofx5600.xml | 538 ++++++++++++++++++++ .../4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt | 70 +++ .../4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt | 69 +++ configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config | 87 ++++ configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt | 70 +++ configs/4.x-cfgs/SM2_C2050/gpgpusim.config | 133 +++++ .../4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt | 70 +++ configs/4.x-cfgs/SM2_GTX480/gpgpusim.config | 135 ++++++ configs/4.x-cfgs/SM2_GTX480/gpuwattch_gtx480.xml | 538 ++++++++++++++++++++ .../4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt | 70 +++ configs/4.x-cfgs/SM5_GTX750/gpgpusim.config | 131 +++++ configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml | 539 +++++++++++++++++++++ .../4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt | 70 +++ configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config | 149 ++++++ .../4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml | 538 ++++++++++++++++++++ configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt | 73 +++ configs/4.x-cfgs/SM6_P100/gpgpusim.config | 173 +++++++ .../4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt | 73 +++ configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 178 +++++++ .../4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 74 +++ configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 186 +++++++ .../4.x-cfgs/TeslaC2050/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/TeslaC2050/gpgpusim.config | 133 ----- .../4.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt | 74 --- configs/4.x-cfgs/VOLTA_TITANV/gpgpusim.config | 186 ------- 60 files changed, 5824 insertions(+), 5824 deletions(-) delete mode 100644 configs/3.x-cfgs/PASCAL_P100/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/PASCAL_P100/gpgpusim.config delete mode 100755 configs/3.x-cfgs/PASCAL_P100/gpuwattch_gtx480.xml delete mode 100644 configs/3.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/PASCAL_TITANX/gpgpusim.config delete mode 100755 configs/3.x-cfgs/PASCAL_TITANX/gpuwattch_gtx480.xml create mode 100644 configs/3.x-cfgs/SM6_P100/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/SM6_P100/gpgpusim.config create mode 100755 configs/3.x-cfgs/SM6_P100/gpuwattch_gtx480.xml create mode 100644 configs/3.x-cfgs/SM6_TITANX/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/SM6_TITANX/gpgpusim.config create mode 100755 configs/3.x-cfgs/SM6_TITANX/gpuwattch_gtx480.xml create mode 100644 configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/SM7_TITANV/gpgpusim.config delete mode 100644 configs/3.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/VOLTA_TITANV/gpgpusim.config delete mode 100644 configs/4.x-cfgs/GTX1080/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/GTX1080/gpgpusim.config delete mode 100755 configs/4.x-cfgs/GTX1080/gpuwattch_gtx1080Ti.xml delete mode 100644 configs/4.x-cfgs/GTX480/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/GTX480/gpgpusim.config delete mode 100755 configs/4.x-cfgs/GTX480/gpuwattch_gtx480.xml delete mode 100644 configs/4.x-cfgs/GTX750/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/GTX750/gpgpusim.config delete mode 100755 configs/4.x-cfgs/GTX750/gpuwattch_gtx750Ti.xml delete mode 100644 configs/4.x-cfgs/PASCAL_P100/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/PASCAL_P100/gpgpusim.config delete mode 100644 configs/4.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/PASCAL_TITANX/gpgpusim.config delete mode 100644 configs/4.x-cfgs/QuadroFX5600/gpgpusim.config delete mode 100644 configs/4.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml delete mode 100644 configs/4.x-cfgs/QuadroFX5600/icnt_config_islip.icnt delete mode 100644 configs/4.x-cfgs/QuadroFX5800/config_quadro_islip.icnt delete mode 100644 configs/4.x-cfgs/QuadroFX5800/gpgpusim.config create mode 100644 configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config create mode 100644 configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml create mode 100644 configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt create mode 100644 configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt create mode 100644 configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config create mode 100644 configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM2_C2050/gpgpusim.config create mode 100644 configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM2_GTX480/gpgpusim.config create mode 100755 configs/4.x-cfgs/SM2_GTX480/gpuwattch_gtx480.xml create mode 100644 configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM5_GTX750/gpgpusim.config create mode 100755 configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml create mode 100644 configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config create mode 100755 configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml create mode 100644 configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM6_P100/gpgpusim.config create mode 100644 configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM6_TITANX/gpgpusim.config create mode 100644 configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM7_TITANV/gpgpusim.config delete mode 100644 configs/4.x-cfgs/TeslaC2050/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/TeslaC2050/gpgpusim.config delete mode 100644 configs/4.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/VOLTA_TITANV/gpgpusim.config (limited to 'configs') diff --git a/configs/3.x-cfgs/PASCAL_P100/config_fermi_islip.icnt b/configs/3.x-cfgs/PASCAL_P100/config_fermi_islip.icnt deleted file mode 100644 index d26c8d9..0000000 --- a/configs/3.x-cfgs/PASCAL_P100/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 60; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/PASCAL_P100/gpgpusim.config b/configs/3.x-cfgs/PASCAL_P100/gpgpusim.config deleted file mode 100644 index a5e6736..0000000 --- a/configs/3.x-cfgs/PASCAL_P100/gpgpusim.config +++ /dev/null @@ -1,156 +0,0 @@ -# This config models the Pascal GP100 -# For more info about this card, see Nvidia White paper -# https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 32 --gpgpu_n_sub_partition_per_mchannel 1 - -# Pscal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Nvidia_Tesla --gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,2,1,2,2,1,4 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 2,2,2,2,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory -# Pascal GP100 has 48KB L1 cache --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 65536 --gpgpu_shmem_size_PrefShared 65536 --gmem_skip_L1D 0 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 14 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_units_mem 10 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/3.x-cfgs/PASCAL_P100/gpuwattch_gtx480.xml b/configs/3.x-cfgs/PASCAL_P100/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/3.x-cfgs/PASCAL_P100/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/3.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt b/configs/3.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt deleted file mode 100644 index 602daee..0000000 --- a/configs/3.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/PASCAL_TITANX/gpgpusim.config b/configs/3.x-cfgs/PASCAL_TITANX/gpgpusim.config deleted file mode 100644 index f78bd02..0000000 --- a/configs/3.x-cfgs/PASCAL_TITANX/gpgpusim.config +++ /dev/null @@ -1,156 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,8,4,4,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_shmem_size 98304 --gpgpu_shmem_size_PrefL1 98304 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 -# gpgpu_num_reg_banks should be increased to 32 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/3.x-cfgs/PASCAL_TITANX/gpuwattch_gtx480.xml b/configs/3.x-cfgs/PASCAL_TITANX/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/3.x-cfgs/PASCAL_TITANX/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/3.x-cfgs/SM6_P100/config_fermi_islip.icnt b/configs/3.x-cfgs/SM6_P100/config_fermi_islip.icnt new file mode 100644 index 0000000..d26c8d9 --- /dev/null +++ b/configs/3.x-cfgs/SM6_P100/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 60; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/SM6_P100/gpgpusim.config b/configs/3.x-cfgs/SM6_P100/gpgpusim.config new file mode 100644 index 0000000..a5e6736 --- /dev/null +++ b/configs/3.x-cfgs/SM6_P100/gpgpusim.config @@ -0,0 +1,156 @@ +# This config models the Pascal GP100 +# For more info about this card, see Nvidia White paper +# https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 60 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 32 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Pscal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Nvidia_Tesla +-gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,2,1,2,2,1,4 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 2,2,2,2,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +# Pascal GP100 has 48KB L1 cache +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 65536 +-gpgpu_shmem_size_PrefShared 65536 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/3.x-cfgs/SM6_P100/gpuwattch_gtx480.xml b/configs/3.x-cfgs/SM6_P100/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/3.x-cfgs/SM6_P100/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/SM6_TITANX/config_fermi_islip.icnt b/configs/3.x-cfgs/SM6_TITANX/config_fermi_islip.icnt new file mode 100644 index 0000000..602daee --- /dev/null +++ b/configs/3.x-cfgs/SM6_TITANX/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/3.x-cfgs/SM6_TITANX/gpgpusim.config new file mode 100644 index 0000000..f78bd02 --- /dev/null +++ b/configs/3.x-cfgs/SM6_TITANX/gpgpusim.config @@ -0,0 +1,156 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,8,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 98304 +-gpgpu_shmem_size_PrefL1 98304 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +# gpgpu_num_reg_banks should be increased to 32 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/3.x-cfgs/SM6_TITANX/gpuwattch_gtx480.xml b/configs/3.x-cfgs/SM6_TITANX/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/3.x-cfgs/SM6_TITANX/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..fac792a --- /dev/null +++ b/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config new file mode 100644 index 0000000..f1ae2dc --- /dev/null +++ b/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config @@ -0,0 +1,165 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,1,4,4,1,9 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +-gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 1 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 + +# 128 KB Inst. +-gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/3.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt b/configs/3.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt deleted file mode 100644 index fac792a..0000000 --- a/configs/3.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/VOLTA_TITANV/gpgpusim.config b/configs/3.x-cfgs/VOLTA_TITANV/gpgpusim.config deleted file mode 100644 index f1ae2dc..0000000 --- a/configs/3.x-cfgs/VOLTA_TITANV/gpgpusim.config +++ /dev/null @@ -1,165 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,1,4,4,1,9 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 --gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 --gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 1 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 0 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 - -# 128 KB Inst. --gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 14 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_units_mem 10 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/GTX1080/config_fermi_islip.icnt b/configs/4.x-cfgs/GTX1080/config_fermi_islip.icnt deleted file mode 100644 index 2a69ddd..0000000 --- a/configs/4.x-cfgs/GTX1080/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 50; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/GTX1080/gpgpusim.config b/configs/4.x-cfgs/GTX1080/gpgpusim.config deleted file mode 100644 index 47c2b6a..0000000 --- a/configs/4.x-cfgs/GTX1080/gpgpusim.config +++ /dev/null @@ -1,149 +0,0 @@ -# This config models the Pascal GP102 (GeForceGTX 1080Ti) - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 11 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 1,2,1,1,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache -# The default is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 --gpgpu_shmem_size 98304 --gmem_skip_L1D 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 --gpgpu_cache:dl2_texture_only 0 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 -# gpgpu_num_reg_banks should be increased to 32, but it gives an error! --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) -# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Pascal has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx1080Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/GTX1080/gpuwattch_gtx1080Ti.xml b/configs/4.x-cfgs/GTX1080/gpuwattch_gtx1080Ti.xml deleted file mode 100755 index 02619ff..0000000 --- a/configs/4.x-cfgs/GTX1080/gpuwattch_gtx1080Ti.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/4.x-cfgs/GTX480/config_fermi_islip.icnt b/configs/4.x-cfgs/GTX480/config_fermi_islip.icnt deleted file mode 100644 index 7820e4e..0000000 --- a/configs/4.x-cfgs/GTX480/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 27; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/GTX480/gpgpusim.config b/configs/4.x-cfgs/GTX480/gpgpusim.config deleted file mode 100644 index 03fcda1..0000000 --- a/configs/4.x-cfgs/GTX480/gpgpusim.config +++ /dev/null @@ -1,135 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 15 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 700.0:700.0:700.0:924.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -#For Fermi, DP unit =0, DP inst is executed on SFU --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 116 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/GTX480/gpuwattch_gtx480.xml b/configs/4.x-cfgs/GTX480/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/4.x-cfgs/GTX480/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/4.x-cfgs/GTX750/config_fermi_islip.icnt b/configs/4.x-cfgs/GTX750/config_fermi_islip.icnt deleted file mode 100644 index 069ca02..0000000 --- a/configs/4.x-cfgs/GTX750/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 7; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/GTX750/gpgpusim.config b/configs/4.x-cfgs/GTX750/gpgpusim.config deleted file mode 100644 index 9366f93..0000000 --- a/configs/4.x-cfgs/GTX750/gpgpusim.config +++ /dev/null @@ -1,131 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 52 - - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 5 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 2 --gpgpu_n_sub_partition_per_mchannel 1 - -# Maxwell clock domains -#-gpgpu_clock_domains ::: -# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. --gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 8 --gpgpu_num_sfu_units 32 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 6,12,13,13,210 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 6,12,6,6,374 --ptx_opcode_initiation_fp 1,1,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 --gmem_skip_L1D 1 --gpgpu_shmem_size 65536 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache --gpgpu_cache:dl2 N:1024:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:16:128:32,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 150 --dram_latency 130 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 300 - -# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 32 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Maxwell has four schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs -# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. -# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx750Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/GTX750/gpuwattch_gtx750Ti.xml b/configs/4.x-cfgs/GTX750/gpuwattch_gtx750Ti.xml deleted file mode 100755 index e2b2324..0000000 --- a/configs/4.x-cfgs/GTX750/gpuwattch_gtx750Ti.xml +++ /dev/null @@ -1,539 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/4.x-cfgs/PASCAL_P100/config_fermi_islip.icnt b/configs/4.x-cfgs/PASCAL_P100/config_fermi_islip.icnt deleted file mode 100644 index e7c2c3b..0000000 --- a/configs/4.x-cfgs/PASCAL_P100/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 60; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 128; -input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/PASCAL_P100/gpgpusim.config b/configs/4.x-cfgs/PASCAL_P100/gpgpusim.config deleted file mode 100644 index 9a7259e..0000000 --- a/configs/4.x-cfgs/PASCAL_P100/gpgpusim.config +++ /dev/null @@ -1,173 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 32 --gpgpu_n_sub_partition_per_mchannel 1 - -# Pscal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Nvidia_Tesla --gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,2,2,1,2,2,2,1,6 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 2 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 2,2,2,2,130 --ptx_opcode_latency_sfu 8 --ptx_opcode_initiation_sfu 4 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 65536 --gpgpu_shmem_size_PrefShared 65536 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt b/configs/4.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt deleted file mode 100644 index 714d933..0000000 --- a/configs/4.x-cfgs/PASCAL_TITANX/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 64; -input_buffer_size = 256; -ejection_buffer_size = 64; -boundary_buffer_size = 64; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/PASCAL_TITANX/gpgpusim.config b/configs/4.x-cfgs/PASCAL_TITANX/gpgpusim.config deleted file mode 100644 index 3097d19..0000000 --- a/configs/4.x-cfgs/PASCAL_TITANX/gpgpusim.config +++ /dev/null @@ -1,178 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 --gpgpu_ignore_resources_limitation 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration -# P102 has two semi-indp scheds per core, and two cores per cluster --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 - -# This implies a maximum of 32 warps/SM --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 16 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 4 SFU units -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 1 - - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,8,8,8,130 --ptx_opcode_initiation_sfu 4 --ptx_opcode_latency_sfu 8 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 24KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 49152 --gpgpu_shmem_size_PrefL1 49152 --gpgpu_shmem_size_PrefShared 49152 --gmem_skip_L1D 1 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:128:4,16:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 --perf_sim_memcpy 0 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# this is unused --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_operand_collector_num_in_ports_sfu 2 --gpgpu_operand_collector_num_out_ports_sfu 2 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 -# gpgpu_num_reg_banks should be increased to 32 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 -# Use Fermi Coalsce arhitetecture which is the same as Pascal --gpgpu_coalesce_arch 61 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 100 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 240 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: - CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" - --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal 102 has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 102 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/QuadroFX5600/gpgpusim.config b/configs/4.x-cfgs/QuadroFX5600/gpgpusim.config deleted file mode 100644 index e3cab18..0000000 --- a/configs/4.x-cfgs/QuadroFX5600/gpgpusim.config +++ /dev/null @@ -1,97 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 12 - -# high level architecture configuration --gpgpu_n_clusters 8 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 6 --gpgpu_clock_domains 337.5:600.0:600.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 -#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) --gpgpu_shader_core_pipeline 768:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour --gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - -# TLB parameters -#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 -#-gpgpu_tlbl2_latency 45 - - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file icnt_config_islip.icnt - -# dram scheduler config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 55 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz -# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - - - -# Using cuobjdump to extract ptx/SASS --gpgpu_ptx_use_cuobjdump 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 --power_trace_enabled 0 --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_quadrofx5600.xml - --steady_power_levels_enabled 1 --steady_state_definition 8,4 diff --git a/configs/4.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml b/configs/4.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml deleted file mode 100644 index 2c5a6fc..0000000 --- a/configs/4.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/4.x-cfgs/QuadroFX5600/icnt_config_islip.icnt b/configs/4.x-cfgs/QuadroFX5600/icnt_config_islip.icnt deleted file mode 100644 index de3bcc8..0000000 --- a/configs/4.x-cfgs/QuadroFX5600/icnt_config_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//14*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 14; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/QuadroFX5800/config_quadro_islip.icnt b/configs/4.x-cfgs/QuadroFX5800/config_quadro_islip.icnt deleted file mode 100644 index cfe9cac..0000000 --- a/configs/4.x-cfgs/QuadroFX5800/config_quadro_islip.icnt +++ /dev/null @@ -1,69 +0,0 @@ -//18*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 18; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/QuadroFX5800/gpgpusim.config b/configs/4.x-cfgs/QuadroFX5800/gpgpusim.config deleted file mode 100644 index 56dbb17..0000000 --- a/configs/4.x-cfgs/QuadroFX5800/gpgpusim.config +++ /dev/null @@ -1,87 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 13 - -# high level architecture configuration --gpgpu_n_clusters 10 --gpgpu_n_cores_per_cluster 3 --gpgpu_n_mem 8 --gpgpu_clock_domains 325.0:650.0:650.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo --gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file config_quadro_islip.icnt - -# dram scheduler config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 325MHz = 74 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 58 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --dram_data_command_freq_ratio 2 # GDDR3 is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config b/configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config new file mode 100644 index 0000000..e3cab18 --- /dev/null +++ b/configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config @@ -0,0 +1,97 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 12 + +# high level architecture configuration +-gpgpu_n_clusters 8 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 6 +-gpgpu_clock_domains 337.5:600.0:600.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) +-gpgpu_shader_core_pipeline 768:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +-gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +# TLB parameters +#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 +#-gpgpu_tlbl2_latency 45 + + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file icnt_config_islip.icnt + +# dram scheduler config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 55 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + + + +# Using cuobjdump to extract ptx/SASS +-gpgpu_ptx_use_cuobjdump 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 +-power_trace_enabled 0 +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_quadrofx5600.xml + +-steady_power_levels_enabled 1 +-steady_state_definition 8,4 diff --git a/configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml b/configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml new file mode 100644 index 0000000..2c5a6fc --- /dev/null +++ b/configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt b/configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt new file mode 100644 index 0000000..de3bcc8 --- /dev/null +++ b/configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt @@ -0,0 +1,70 @@ +//14*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 14; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt b/configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt new file mode 100644 index 0000000..cfe9cac --- /dev/null +++ b/configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt @@ -0,0 +1,69 @@ +//18*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 18; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config b/configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config new file mode 100644 index 0000000..56dbb17 --- /dev/null +++ b/configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config @@ -0,0 +1,87 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 13 + +# high level architecture configuration +-gpgpu_n_clusters 10 +-gpgpu_n_cores_per_cluster 3 +-gpgpu_n_mem 8 +-gpgpu_clock_domains 325.0:650.0:650.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +-gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file config_quadro_islip.icnt + +# dram scheduler config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 325MHz = 74 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 58 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-dram_data_command_freq_ratio 2 # GDDR3 is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt b/configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt new file mode 100644 index 0000000..a11bd8e --- /dev/null +++ b/configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//20*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 26; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; diff --git a/configs/4.x-cfgs/SM2_C2050/gpgpusim.config b/configs/4.x-cfgs/SM2_C2050/gpgpusim.config new file mode 100644 index 0000000..aa5f5f3 --- /dev/null +++ b/configs/4.x-cfgs/SM2_C2050/gpgpusim.config @@ -0,0 +1,133 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + +# Using cuobjdump to extract ptx/SASS +#-gpgpu_ptx_use_cuobjdump 1 # use default + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + + +# high level architecture configuration +-gpgpu_n_clusters 14 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 575.0:575.0:575.0:750.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 750MHz / 575MHz = 130 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 114 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt b/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt new file mode 100644 index 0000000..7820e4e --- /dev/null +++ b/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 27; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config new file mode 100644 index 0000000..03fcda1 --- /dev/null +++ b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config @@ -0,0 +1,135 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 15 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 700.0:700.0:700.0:924.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +#For Fermi, DP unit =0, DP inst is executed on SFU +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 116 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM2_GTX480/gpuwattch_gtx480.xml b/configs/4.x-cfgs/SM2_GTX480/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/4.x-cfgs/SM2_GTX480/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt b/configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt new file mode 100644 index 0000000..069ca02 --- /dev/null +++ b/configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 7; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM5_GTX750/gpgpusim.config b/configs/4.x-cfgs/SM5_GTX750/gpgpusim.config new file mode 100644 index 0000000..9366f93 --- /dev/null +++ b/configs/4.x-cfgs/SM5_GTX750/gpgpusim.config @@ -0,0 +1,131 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 52 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 5 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 2 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Maxwell clock domains +#-gpgpu_clock_domains ::: +# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. +-gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +-gpgpu_num_sp_units 8 +-gpgpu_num_sfu_units 32 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 6,12,13,13,210 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 6,12,6,6,374 +-ptx_opcode_initiation_fp 1,1,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 +-gmem_skip_L1D 1 +-gpgpu_shmem_size 65536 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache +-gpgpu_cache:dl2 N:1024:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 N:16:128:32,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 150 +-dram_latency 130 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 300 + +# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 32 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Maxwell has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. +# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx750Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml b/configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml new file mode 100755 index 0000000..e2b2324 --- /dev/null +++ b/configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml @@ -0,0 +1,539 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt new file mode 100644 index 0000000..2a69ddd --- /dev/null +++ b/configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 50; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config b/configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config new file mode 100644 index 0000000..47c2b6a --- /dev/null +++ b/configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config @@ -0,0 +1,149 @@ +# This config models the Pascal GP102 (GeForceGTX 1080Ti) + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 11 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 1,2,1,1,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The default is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 +-gpgpu_cache:dl2_texture_only 0 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +# gpgpu_num_reg_banks should be increased to 32, but it gives an error! +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) +# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Pascal has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx1080Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml b/configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml new file mode 100755 index 0000000..02619ff --- /dev/null +++ b/configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt new file mode 100644 index 0000000..e7c2c3b --- /dev/null +++ b/configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 60; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 128; +input_buffer_size = 256; +ejection_buffer_size = 128; +boundary_buffer_size = 128; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM6_P100/gpgpusim.config b/configs/4.x-cfgs/SM6_P100/gpgpusim.config new file mode 100644 index 0000000..9a7259e --- /dev/null +++ b/configs/4.x-cfgs/SM6_P100/gpgpusim.config @@ -0,0 +1,173 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 60 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 32 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Pscal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Nvidia_Tesla +-gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,2,2,1,2,2,2,1,6 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 2 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 2,2,2,2,130 +-ptx_opcode_latency_sfu 8 +-ptx_opcode_initiation_sfu 4 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +-gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 65536 +-gpgpu_shmem_size_PrefShared 65536 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 + +# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 + +# 4 KB Inst. +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt new file mode 100644 index 0000000..714d933 --- /dev/null +++ b/configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; +input_buffer_size = 256; +ejection_buffer_size = 64; +boundary_buffer_size = 64; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config new file mode 100644 index 0000000..3097d19 --- /dev/null +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -0,0 +1,178 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 +-gpgpu_ignore_resources_limitation 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +# P102 has two semi-indp scheds per core, and two cores per cluster +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 32 warps/SM +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 16 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 4 SFU units +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 1 + + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,8,8,8,130 +-ptx_opcode_initiation_sfu 4 +-ptx_opcode_latency_sfu 8 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 24KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_shmem_size 49152 +-gpgpu_shmem_size_PrefL1 49152 +-gpgpu_shmem_size_PrefShared 49152 +-gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:128:4,16:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 +-perf_sim_memcpy 0 + +# 4 KB Inst. +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# this is unused +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_operand_collector_num_in_ports_sfu 2 +-gpgpu_operand_collector_num_out_ports_sfu 2 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +# gpgpu_num_reg_banks should be increased to 32 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +# Use Fermi Coalsce arhitetecture which is the same as Pascal +-gpgpu_coalesce_arch 61 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 100 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 240 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: + CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" + +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal 102 has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 102 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..616e9f3 --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 128; +input_buffer_size = 256; +ejection_buffer_size = 128; +boundary_buffer_size = 128; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config new file mode 100644 index 0000000..8969168 --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -0,0 +1,186 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +-gpgpu_cache:dl1 S:64:128:8,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:64:128:16,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 1 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 S:64:128:24,L:B:m:F:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/TeslaC2050/config_fermi_islip.icnt b/configs/4.x-cfgs/TeslaC2050/config_fermi_islip.icnt deleted file mode 100644 index a11bd8e..0000000 --- a/configs/4.x-cfgs/TeslaC2050/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//20*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 26; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; diff --git a/configs/4.x-cfgs/TeslaC2050/gpgpusim.config b/configs/4.x-cfgs/TeslaC2050/gpgpusim.config deleted file mode 100644 index aa5f5f3..0000000 --- a/configs/4.x-cfgs/TeslaC2050/gpgpusim.config +++ /dev/null @@ -1,133 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - -# Using cuobjdump to extract ptx/SASS -#-gpgpu_ptx_use_cuobjdump 1 # use default - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - - -# high level architecture configuration --gpgpu_n_clusters 14 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 575.0:575.0:575.0:750.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 750MHz / 575MHz = 130 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 114 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt b/configs/4.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt deleted file mode 100644 index 616e9f3..0000000 --- a/configs/4.x-cfgs/VOLTA_TITANV/config_fermi_islip.icnt +++ /dev/null @@ -1,74 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 128; -input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/VOLTA_TITANV/gpgpusim.config b/configs/4.x-cfgs/VOLTA_TITANV/gpgpusim.config deleted file mode 100644 index 8969168..0000000 --- a/configs/4.x-cfgs/VOLTA_TITANV/gpgpusim.config +++ /dev/null @@ -1,186 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 100 --ptx_opcode_initiation_sfu 8 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:8,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:16,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 1 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,L:B:m:F:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - -- cgit v1.3 From 1e1c08286a505418d0e3ad1ee819e15881e9cb43 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 15 Jun 2018 16:46:14 -0400 Subject: memory partition indexing --- configs/4.x-cfgs/SM6_P100/gpgpusim.config | 1 + src/gpgpu-sim/addrdec.cc | 48 +++++++++++++++++++++++++++++++ src/gpgpu-sim/addrdec.h | 8 ++++++ src/gpgpu-sim/dram.cc | 2 +- src/gpgpu-sim/gpu-cache.cc | 36 ++++++++++++++++++++++- src/gpgpu-sim/gpu-cache.h | 7 +++-- 6 files changed, 98 insertions(+), 4 deletions(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_P100/gpgpusim.config b/configs/4.x-cfgs/SM6_P100/gpgpusim.config index 9a7259e..a43c0bb 100644 --- a/configs/4.x-cfgs/SM6_P100/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_P100/gpgpusim.config @@ -69,6 +69,7 @@ -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 -perf_sim_memcpy 0 +-memory_partition_indexing 2 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 diff --git a/src/gpgpu-sim/addrdec.cc b/src/gpgpu-sim/addrdec.cc index 422576d..cfd90ec 100644 --- a/src/gpgpu-sim/addrdec.cc +++ b/src/gpgpu-sim/addrdec.cc @@ -62,6 +62,9 @@ void linear_to_raw_address_translation::addrdec_setoption(option_parser_t opp) option_parser_register(opp, "-gpgpu_mem_address_mask", OPT_INT32, &gpgpu_mem_address_mask, "0 = old addressing mask, 1 = new addressing mask, 2 = new add. mask + flipped bank sel and chip sel bits", "0"); + option_parser_register(opp, "-memory_partition_indexing", OPT_UINT32, &memory_partition_indexing, + "0 = no indexing, 1 = bitwise xoring, 2 = IPoly, 3 = custom indexing", + "0"); } new_addr_type linear_to_raw_address_translation::partition_address( new_addr_type addr ) const @@ -103,6 +106,51 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, addrdec_ tlx->burst= addrdec_packbits(addrdec_mask[BURST], rest_of_addr, addrdec_mkhigh[BURST], addrdec_mklow[BURST]); } + switch(memory_partition_indexing){ + case CONSECUTIVE: + //Do nothing + break; + case BITWISE_PERMUTATION: + assert(!gap); + tlx->chip = (tlx->chip) ^ (tlx->row & (m_n_channel-1)); + assert(tlx->chip < m_n_channel); + break; + case IPOLY: + /* + * Set Indexing function from "Pseudo-randomly interleaved memory." + * Rau, B. R et al. + * ISCA 1991 + * + * equations are adopted from: + * "Sacat: streaming-aware conflict-avoiding thrashing-resistant gpgpu cache management scheme." + * Khairy et al. + * IEEE TPDS 2017. + */ + if(m_n_channel == 32) { + std::bitset<64> a(tlx->row); + std::bitset<5> chip(tlx->chip); + chip[0] = a[13]^a[12]^a[11]^a[10]^a[9]^a[6]^a[5]^a[3]^a[0]^chip[0]; + chip[1] = a[14]^a[13]^a[12]^a[11]^a[10]^a[7]^a[6]^a[4]^a[1]^chip[1]; + chip[2] = a[14]^a[10]^a[9]^a[8]^a[7]^a[6]^a[3]^a[2]^a[0]^chip[2]; + chip[3] = a[11]^a[10]^a[9]^a[8]^a[7]^a[4]^a[3]^a[1]^chip[3]; + chip[4] = a[12]^a[11]^a[10]^a[9]^a[8]^a[5]^a[4]^a[2]^chip[4]; + tlx->chip = chip.to_ulong(); + + } + else{ /* Else incorrect number of channels for the hashing function */ + assert("\nGPGPU-Sim memory_partition_indexing error: The number of channels should be " + "32 for the hashing IPOLY index function.\n" && 0); + } + assert(tlx->chip < m_n_channel); + break; + case CUSTOM: + /* No custom set function implemented */ + break; + default: + assert("\nUndefined set index function.\n" && 0); + break; + } + // combine the chip address and the lower bits of DRAM bank address to form the subpartition ID unsigned sub_partition_addr_mask = m_n_sub_partition_in_channel - 1; tlx->sub_partition = tlx->chip * m_n_sub_partition_in_channel diff --git a/src/gpgpu-sim/addrdec.h b/src/gpgpu-sim/addrdec.h index fd9af8d..a18ff63 100644 --- a/src/gpgpu-sim/addrdec.h +++ b/src/gpgpu-sim/addrdec.h @@ -35,6 +35,13 @@ #include "../abstract_hardware_model.h" +enum partition_index_function{ + CONSECUTIVE = 0, + BITWISE_PERMUTATION, + IPOLY, + CUSTOM +}; + struct addrdec_t { void print( FILE *fp ) const; @@ -72,6 +79,7 @@ private: const char *addrdec_option; int gpgpu_mem_address_mask; + partition_index_function memory_partition_indexing; bool run_test; int ADDR_CHIP_S; diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index a57508c..92aa819 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -204,7 +204,7 @@ dram_req_t::dram_req_t( class mem_fetch *mf, unsigned banks, unsigned dram_bnk_i } else if(dram_bnk_indexing_policy == 1) { int lbank = log2(banks); - bk = tlx.bk ^ (((1< a(addr); + std::bitset<6> index; + index[0] = a[25]^a[24]^a[23]^a[22]^a[21]^a[18]^a[17]^a[15]^a[12]^a[7]; //10 + index[1] = a[26]^a[25]^a[24]^a[23]^a[22]^a[19]^a[18]^a[16]^a[13]^a[8]; //10 + index[2] = a[26]^a[22]^a[21]^a[20]^a[19]^a[18]^a[15]^a[14]^a[12]^a[9]; //10 + index[3] = a[23]^a[22]^a[21]^a[20]^a[19]^a[16]^a[15]^a[13]^a[10]; //9 + index[4] = a[24]^a[23]^a[22]^a[21]^a[20]^a[17]^a[16]^a[14]^a[11]; //9 + + if(m_nset == 64) + index[5] = a[12]; + + set_index = index.to_ulong(); + + }else{ /* Else incorrect number of sets for the hashing function */ + assert("\nGPGPU-Sim cache configuration error: The number of sets should be " + "32 or 64 for the hashing set index function.\n" && 0); + } + break; + case CUSTOM_SET_FUNCTION: /* No custom set function implemented */ break; @@ -104,6 +134,10 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{ case LINEAR_SET_FUNCTION: set_index = (addr >> m_line_sz_log2) & (m_nset-1); break; + + default: + assert("\nUndefined set index function.\n" && 0); + break; } // Linear function selected or custom set index function not implemented diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 76b07a2..d2b7757 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -464,8 +464,10 @@ enum mshr_config_t { }; enum set_index_function{ - FERMI_HASH_SET_FUNCTION = 0, - LINEAR_SET_FUNCTION, + LINEAR_SET_FUNCTION = 0, + BITWISE_XORING_FUNCTION, + HASH_IPOLY_FUNCTION, + FERMI_HASH_SET_FUNCTION, CUSTOM_SET_FUNCTION }; @@ -587,6 +589,7 @@ public: switch(sif){ case 'H': m_set_index_function = FERMI_HASH_SET_FUNCTION; break; + case 'P': m_set_index_function = HASH_IPOLY_FUNCTION; break; case 'C': m_set_index_function = CUSTOM_SET_FUNCTION; break; case 'L': m_set_index_function = LINEAR_SET_FUNCTION; break; default: exit_parse_error(); -- cgit v1.3 From 8c81c1d04e8d20b08f122a12ce090b4f926adb4c Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 22 Aug 2018 10:19:52 -0400 Subject: adding lazy-fetch-on-read and invalidate operation to cache --- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 2 +- configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 2 +- src/gpgpu-sim/gpu-cache.cc | 110 ++++++++++++++++------------ src/gpgpu-sim/gpu-cache.h | 22 ++++-- src/gpgpu-sim/gpu-sim.cc | 6 +- src/gpgpu-sim/l2cache.cc | 12 ++- src/gpgpu-sim/l2cache.h | 1 + src/gpgpu-sim/shader.cc | 16 ++++ src/gpgpu-sim/shader.h | 3 + 9 files changed, 115 insertions(+), 59 deletions(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index 3097d19..7368882 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -73,7 +73,7 @@ -gpgpu_n_cluster_ejection_buffer_size 32 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:128:4,16:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:128:4,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 -perf_sim_memcpy 0 diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config index 8969168..da98547 100644 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -78,7 +78,7 @@ -gpgpu_n_cluster_ejection_buffer_size 32 # 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,L:B:m:F:L,A:256:4,32:0,32 +-gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:256:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 -perf_sim_memcpy 0 diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 75ec00a..9d81de9 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -382,6 +382,8 @@ void tag_array::fill( unsigned index, unsigned time, mem_fetch* mf) m_lines[index]->fill(time, mf->get_access_sector_mask()); } + +//TODO: we need write back the flushed data to the upper level void tag_array::flush() { for (unsigned i=0; i < m_config.get_num_lines(); i++) @@ -391,6 +393,13 @@ void tag_array::flush() } } +void tag_array::invalidate() +{ + for (unsigned i=0; i < m_config.get_num_lines(); i++) + for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) + m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; +} + float tag_array::windowed_miss_rate( ) const { unsigned n_access = m_access - m_prev_snapshot_access; @@ -1162,52 +1171,7 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, unsigned time, std::list &events, enum cache_request_status status ) { - - new_addr_type block_addr = m_config.block_addr(addr); - new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); - - - //if the request writes to the whole cache line/sector, then, write and set cache line Modified. - //and no need to send read request to memory or reserve mshr - - if(miss_queue_full(0)) { - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); - return RESERVATION_FAIL; // cannot handle request this cycle - } - - bool wb = false; - evicted_block_info evicted; - - cache_request_status m_status = m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); - assert(m_status != HIT); - cache_block_t* block = m_tag_array->get_block(cache_index); - block->set_status(MODIFIED, mf->get_access_sector_mask()); - if(m_status == HIT_RESERVED) { - block->set_ignore_on_fill(true, mf->get_access_sector_mask()); - block->set_modified_on_fill(true, mf->get_access_sector_mask()); - } - - if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) - { - block->set_m_readable(true, mf->get_access_sector_mask()); - } else - { - block->set_m_readable(false, mf->get_access_sector_mask()); - } - - if( m_status != RESERVATION_FAIL ){ - // If evicted block is modified and not a write-through - // (already modified lower level) - if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) { - mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, - m_wrbk_type,evicted.m_modified_size,true); - send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); - } - return MISS; - } - return RESERVATION_FAIL; - - /*new_addr_type block_addr = m_config.block_addr(addr); + new_addr_type block_addr = m_config.block_addr(addr); new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) @@ -1316,7 +1280,59 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, return MISS; } return RESERVATION_FAIL; - }*/ + } +} + +enum cache_request_status +data_cache::wr_miss_wa_lazy_fetch_on_read( new_addr_type addr, + unsigned cache_index, mem_fetch *mf, + unsigned time, std::list &events, + enum cache_request_status status ) +{ + + new_addr_type block_addr = m_config.block_addr(addr); + new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); + + + //if the request writes to the whole cache line/sector, then, write and set cache line Modified. + //and no need to send read request to memory or reserve mshr + + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + return RESERVATION_FAIL; // cannot handle request this cycle + } + + bool wb = false; + evicted_block_info evicted; + + cache_request_status m_status = m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); + assert(m_status != HIT); + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); + if(m_status == HIT_RESERVED) { + block->set_ignore_on_fill(true, mf->get_access_sector_mask()); + block->set_modified_on_fill(true, mf->get_access_sector_mask()); + } + + if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) + { + block->set_m_readable(true, mf->get_access_sector_mask()); + } else + { + block->set_m_readable(false, mf->get_access_sector_mask()); + } + + if( m_status != RESERVATION_FAIL ){ + // If evicted block is modified and not a write-through + // (already modified lower level) + if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) { + mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, + m_wrbk_type,evicted.m_modified_size,true); + send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); + } + return MISS; + } + return RESERVATION_FAIL; } /// No write-allocate miss: Simply send write request to lower level memory diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index d2b7757..96b9a6d 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -454,7 +454,8 @@ enum allocation_policy_t { enum write_allocate_policy_t { NO_WRITE_ALLOCATE, WRITE_ALLOCATE, - FETCH_ON_WRITE + FETCH_ON_WRITE, + LAZY_FETCH_ON_READ }; enum mshr_config_t { @@ -555,6 +556,7 @@ public: case 'N': m_write_alloc_policy = NO_WRITE_ALLOCATE; break; case 'W': m_write_alloc_policy = WRITE_ALLOCATE; break; case 'F': m_write_alloc_policy = FETCH_ON_WRITE; break; + case 'L': m_write_alloc_policy = LAZY_FETCH_ON_READ; break; default: exit_parse_error(); } @@ -572,9 +574,9 @@ public: assert(0 && "Invalid cache configuration: Writeback cache cannot allocate new line on fill. "); } - if(m_write_alloc_policy == FETCH_ON_WRITE && m_alloc_policy == ON_FILL) + if((m_write_alloc_policy == FETCH_ON_WRITE || m_write_alloc_policy == LAZY_FETCH_ON_READ )&& m_alloc_policy == ON_FILL) { - assert(0 && "Invalid cache configuration: FETCH_ON_WRITE cannot work properly with ON_FILL policy. Cache must be ON_MISS. "); + assert(0 && "Invalid cache configuration: FETCH_ON_WRITE and LAZY_FETCH_ON_READ cannot work properly with ON_FILL policy. Cache must be ON_MISS. "); } if(m_cache_type == SECTOR) { @@ -742,7 +744,8 @@ public: unsigned size() const { return m_config.get_num_lines();} cache_block_t* get_block(unsigned idx) { return m_lines[idx];} - void flush(); // flash invalidate all entries + void flush(); // flush all written entries + void invalidate(); // invalidate all entries void new_window(); void print( FILE *stream, unsigned &total_access, unsigned &total_misses ) const; @@ -994,6 +997,7 @@ public: mem_fetch *next_access(){return m_mshrs.next_access();} // flash invalidate all entries in cache void flush(){m_tag_array->flush();} + void invalidate(){m_tag_array->invalidate();} void print(FILE *fp, unsigned &accesses, unsigned &misses) const; void display_state( FILE *fp ) const; @@ -1179,6 +1183,7 @@ public: case NO_WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_no_wa; break; case WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_wa_naive; break; case FETCH_ON_WRITE: m_wr_miss = &data_cache::wr_miss_wa_fetch_on_write; break; + case LAZY_FETCH_ON_READ: m_wr_miss = &data_cache::wr_miss_wa_lazy_fetch_on_read; break; default: assert(0 && "Error: Must set valid cache write miss policy\n"); break; // Need to set a write miss function @@ -1299,7 +1304,14 @@ protected: mem_fetch *mf, unsigned time, std::list &events, - enum cache_request_status status ); // write-allocate with read-fetch-only + enum cache_request_status status ); // write-allocate with fetch-on-every-write + enum cache_request_status + wr_miss_wa_lazy_fetch_on_read( new_addr_type addr, + unsigned cache_index, + mem_fetch *mf, + unsigned time, + std::list &events, + enum cache_request_status status ); // write-allocate with read-fetch-only enum cache_request_status wr_miss_wa_write_validate( new_addr_type addr, unsigned cache_index, diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index c5d4464..d48de25 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1548,12 +1548,12 @@ void gpgpu_sim::cycle() issue_block2core(); - // Depending on configuration, flush the caches once all of threads are completed. + // Depending on configuration, invalidate the caches once all of threads are completed. int all_threads_complete = 1; if (m_config.gpgpu_flush_l1_cache) { for (unsigned i=0;in_simt_clusters;i++) { if (m_cluster[i]->get_not_completed() == 0) - m_cluster[i]->cache_flush(); + m_cluster[i]->cache_invalidate(); else all_threads_complete = 0 ; } @@ -1575,7 +1575,7 @@ void gpgpu_sim::cycle() int dlc = 0; for (unsigned i=0;im_n_mem;i++) { dlc = m_memory_sub_partition[i]->flushL2(); - assert (dlc == 0); // need to model actual writes to DRAM here + assert (dlc == 0); // TODO: need to model actual writes to DRAM here printf("Dirty lines flushed from L2 %d is %d\n", i, dlc ); } } diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index b1465a8..359d3c8 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -428,7 +428,7 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) m_icnt_L2_queue->pop(); } } else if ( status != RESERVATION_FAIL ) { - if(mf->is_write() && m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE && !was_writeallocate_sent(events)) { + if(mf->is_write() && (m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE || m_config->m_L2_config.m_write_alloc_policy == LAZY_FETCH_ON_READ) && !was_writeallocate_sent(events)) { mf->set_reply(); mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); m_L2_icnt_queue->push(mf); @@ -568,7 +568,15 @@ unsigned memory_sub_partition::flushL2() if (!m_config->m_L2_config.disabled()) { m_L2cache->flush(); } - return 0; // L2 is read only in this version + return 0; //TODO: write the flushed data to the main memory +} + +unsigned memory_sub_partition::invalidateL2() +{ + if (!m_config->m_L2_config.disabled()) { + m_L2cache->invalidate(); + } + return 0; } bool memory_sub_partition::busy() const diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 2d13918..18c0a8b 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -162,6 +162,7 @@ public: void set_done( mem_fetch *mf ); unsigned flushL2(); + unsigned invalidateL2(); // interface to L2_dram_queue bool L2_dram_queue_empty() const; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index d2f40a1..b660b8c 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1593,6 +1593,11 @@ void ldst_unit::flush(){ m_L1D->flush(); } +void ldst_unit::invalidate(){ + // Flush L1D cache + m_L1D->invalidate(); +} + simd_function_unit::simd_function_unit( const shader_core_config *config ) { m_config=config; @@ -2644,6 +2649,11 @@ void shader_core_ctx::cache_flush() m_ldst_unit->flush(); } +void shader_core_ctx::cache_invalidate() +{ + m_ldst_unit->invalidate(); +} + // modifiers std::list opndcoll_rfu_t::arbiter_t::allocate_reads() { @@ -3461,6 +3471,12 @@ void simt_core_cluster::cache_flush() m_core[i]->cache_flush(); } +void simt_core_cluster::cache_invalidate() +{ + for( unsigned i=0; i < m_config->n_simt_cores_per_cluster; i++ ) + m_core[i]->cache_invalidate(); +} + bool simt_core_cluster::icnt_injection_buffer_full(unsigned size, bool write) { unsigned request_size = size; diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index ae22eaa..cc441b3 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1155,6 +1155,7 @@ public: void fill( mem_fetch *mf ); void flush(); + void invalidate(); void writeback(); // accessors @@ -1655,6 +1656,7 @@ public: void issue_block2core( class kernel_info_t &kernel ); void cache_flush(); + void cache_invalidate(); void accept_fetch_response( mem_fetch *mf ); void accept_ldst_unit_response( class mem_fetch * mf ); void broadcast_barrier_reduction(unsigned cta_id, unsigned bar_id,warp_set_t warps); @@ -1947,6 +1949,7 @@ public: void reinit(); unsigned issue_block2core(); void cache_flush(); + void cache_invalidate(); bool icnt_injection_buffer_full(unsigned size, bool write); void icnt_inject_request_packet(class mem_fetch *mf); -- cgit v1.3 From 891fd3e19541d80106686d0e73a5dcb6d0a8536e Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 22 Aug 2018 10:29:24 -0400 Subject: Update gpgpusim.config --- configs/4.x-cfgs/SM6_P100/gpgpusim.config | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_P100/gpgpusim.config b/configs/4.x-cfgs/SM6_P100/gpgpusim.config index a43c0bb..a4e745d 100644 --- a/configs/4.x-cfgs/SM6_P100/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_P100/gpgpusim.config @@ -65,7 +65,7 @@ -gpgpu_n_cluster_ejection_buffer_size 32 # 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:256:4,32:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 -perf_sim_memcpy 0 -- cgit v1.3 From 944f6dbf23d792dde360d3a4f2334de3b541de52 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 27 Aug 2018 20:28:24 -0400 Subject: fixing ead/write buffer and new configs files --- configs/3.x-cfgs/GTX480/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/GTX480/gpgpusim.config | 133 +++++ configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml | 538 ++++++++++++++++++++ .../GeForceGTX750Ti/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config | 130 +++++ .../GeForceGTX750Ti/gpuwattch_gtx750Ti.xml | 539 +++++++++++++++++++++ configs/3.x-cfgs/QuadroFX5600/gpgpusim.config | 96 ++++ .../QuadroFX5600/gpuwattch_quadrofx5600.xml | 538 ++++++++++++++++++++ .../3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt | 70 +++ .../3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt | 69 +++ configs/3.x-cfgs/QuadroFX5800/gpgpusim.config | 86 ++++ .../3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config | 149 ++++++ .../3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml | 538 ++++++++++++++++++++ .../3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/SM7_TITANV/gpgpusim.config | 165 ------- .../3.x-cfgs/TeslaC2050/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/TeslaC2050/gpgpusim.config | 132 +++++ configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config | 97 ---- .../SM1_QFX5600/gpuwattch_quadrofx5600.xml | 538 -------------------- .../4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt | 70 --- .../4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt | 69 --- configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config | 87 ---- configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/SM2_C2050/gpgpusim.config | 133 ----- .../4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt | 4 +- configs/4.x-cfgs/SM2_GTX480/gpgpusim.config | 23 +- .../4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/SM5_GTX750/gpgpusim.config | 131 ----- configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml | 539 --------------------- .../4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config | 149 ------ .../4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml | 538 -------------------- .../SM6_P100_64SMs/config_fermi_islip.icnt | 73 +++ configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config | 174 +++++++ .../4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 74 --- configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 186 ------- src/gpgpu-sim/addrdec.cc | 17 +- src/gpgpu-sim/dram.cc | 12 +- src/gpgpu-sim/dram_sched.cc | 6 +- src/gpgpu-sim/gpu-sim.cc | 3 +- src/gpgpu-sim/shader.cc | 4 +- 42 files changed, 3596 insertions(+), 3074 deletions(-) create mode 100644 configs/3.x-cfgs/GTX480/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/GTX480/gpgpusim.config create mode 100755 configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml create mode 100644 configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config create mode 100755 configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml create mode 100644 configs/3.x-cfgs/QuadroFX5600/gpgpusim.config create mode 100644 configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml create mode 100644 configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt create mode 100644 configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt create mode 100644 configs/3.x-cfgs/QuadroFX5800/gpgpusim.config create mode 100644 configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config create mode 100755 configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml delete mode 100644 configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/SM7_TITANV/gpgpusim.config create mode 100644 configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/TeslaC2050/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml delete mode 100644 configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt delete mode 100644 configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt delete mode 100644 configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM2_C2050/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM5_GTX750/gpgpusim.config delete mode 100755 configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml delete mode 100644 configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config delete mode 100755 configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml create mode 100644 configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM7_TITANV/gpgpusim.config (limited to 'configs') diff --git a/configs/3.x-cfgs/GTX480/config_fermi_islip.icnt b/configs/3.x-cfgs/GTX480/config_fermi_islip.icnt new file mode 100644 index 0000000..7820e4e --- /dev/null +++ b/configs/3.x-cfgs/GTX480/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 27; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/GTX480/gpgpusim.config b/configs/3.x-cfgs/GTX480/gpgpusim.config new file mode 100644 index 0000000..436cb41 --- /dev/null +++ b/configs/3.x-cfgs/GTX480/gpgpusim.config @@ -0,0 +1,133 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 15 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 700.0:700.0:700.0:924.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 116 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml b/configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt b/configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt new file mode 100644 index 0000000..069ca02 --- /dev/null +++ b/configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 7; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config b/configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config new file mode 100644 index 0000000..8b030b6 --- /dev/null +++ b/configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config @@ -0,0 +1,130 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 52 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 5 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 2 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Maxwell clock domains +#-gpgpu_clock_domains ::: +# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. +-gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 8 +-gpgpu_num_sfu_units 32 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 6,12,13,13,210 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 6,12,6,6,374 +-ptx_opcode_initiation_fp 1,1,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gmem_skip_L1D 1 +-gpgpu_shmem_size 65536 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache +-gpgpu_cache:dl2 1024:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 16:128:32,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 150 +-dram_latency 130 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 300 + +# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 32 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Maxwell has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. +# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx750Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml b/configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml new file mode 100755 index 0000000..e2b2324 --- /dev/null +++ b/configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml @@ -0,0 +1,539 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/QuadroFX5600/gpgpusim.config b/configs/3.x-cfgs/QuadroFX5600/gpgpusim.config new file mode 100644 index 0000000..cb87b65 --- /dev/null +++ b/configs/3.x-cfgs/QuadroFX5600/gpgpusim.config @@ -0,0 +1,96 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 12 + +# high level architecture configuration +-gpgpu_n_clusters 8 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 6 +-gpgpu_clock_domains 337.5:600.0:600.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) +-gpgpu_shader_core_pipeline 768:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,1,1,1,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +-gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +# TLB parameters +#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 +#-gpgpu_tlbl2_latency 45 + + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file icnt_config_islip.icnt + +# dram scheduler config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 55 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + + + +# Using cuobjdump to extract ptx/SASS +-gpgpu_ptx_use_cuobjdump 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 +-power_trace_enabled 0 +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_quadrofx5600.xml + +-steady_power_levels_enabled 1 +-steady_state_definition 8,4 diff --git a/configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml b/configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml new file mode 100644 index 0000000..2c5a6fc --- /dev/null +++ b/configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt b/configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt new file mode 100644 index 0000000..de3bcc8 --- /dev/null +++ b/configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt @@ -0,0 +1,70 @@ +//14*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 14; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt b/configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt new file mode 100644 index 0000000..cfe9cac --- /dev/null +++ b/configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt @@ -0,0 +1,69 @@ +//18*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 18; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/QuadroFX5800/gpgpusim.config b/configs/3.x-cfgs/QuadroFX5800/gpgpusim.config new file mode 100644 index 0000000..82243c2 --- /dev/null +++ b/configs/3.x-cfgs/QuadroFX5800/gpgpusim.config @@ -0,0 +1,86 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 13 + +# high level architecture configuration +-gpgpu_n_clusters 10 +-gpgpu_n_cores_per_cluster 3 +-gpgpu_n_mem 8 +-gpgpu_clock_domains 325.0:650.0:650.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,1,1,1,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +-gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file config_quadro_islip.icnt + +# dram scheduler config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 325MHz = 74 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 58 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-dram_data_command_freq_ratio 2 # GDDR3 is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt b/configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt new file mode 100644 index 0000000..2a69ddd --- /dev/null +++ b/configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 50; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config b/configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config new file mode 100644 index 0000000..47c2b6a --- /dev/null +++ b/configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config @@ -0,0 +1,149 @@ +# This config models the Pascal GP102 (GeForceGTX 1080Ti) + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 11 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 1,2,1,1,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The default is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 +-gpgpu_cache:dl2_texture_only 0 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +# gpgpu_num_reg_banks should be increased to 32, but it gives an error! +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) +# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Pascal has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx1080Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml b/configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml new file mode 100755 index 0000000..02619ff --- /dev/null +++ b/configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt deleted file mode 100644 index fac792a..0000000 --- a/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config deleted file mode 100644 index f1ae2dc..0000000 --- a/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config +++ /dev/null @@ -1,165 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,1,4,4,1,9 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 --gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 --gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 1 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 0 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 - -# 128 KB Inst. --gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 14 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_units_mem 10 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt b/configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt new file mode 100644 index 0000000..a11bd8e --- /dev/null +++ b/configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//20*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 26; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; diff --git a/configs/3.x-cfgs/TeslaC2050/gpgpusim.config b/configs/3.x-cfgs/TeslaC2050/gpgpusim.config new file mode 100644 index 0000000..442ab8b --- /dev/null +++ b/configs/3.x-cfgs/TeslaC2050/gpgpusim.config @@ -0,0 +1,132 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + +# Using cuobjdump to extract ptx/SASS +#-gpgpu_ptx_use_cuobjdump 1 # use default + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + + +# high level architecture configuration +-gpgpu_n_clusters 14 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 575.0:575.0:575.0:750.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 750MHz / 575MHz = 130 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 114 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config b/configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config deleted file mode 100644 index e3cab18..0000000 --- a/configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config +++ /dev/null @@ -1,97 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 12 - -# high level architecture configuration --gpgpu_n_clusters 8 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 6 --gpgpu_clock_domains 337.5:600.0:600.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 -#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) --gpgpu_shader_core_pipeline 768:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour --gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - -# TLB parameters -#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 -#-gpgpu_tlbl2_latency 45 - - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file icnt_config_islip.icnt - -# dram scheduler config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 55 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz -# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - - - -# Using cuobjdump to extract ptx/SASS --gpgpu_ptx_use_cuobjdump 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 --power_trace_enabled 0 --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_quadrofx5600.xml - --steady_power_levels_enabled 1 --steady_state_definition 8,4 diff --git a/configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml b/configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml deleted file mode 100644 index 2c5a6fc..0000000 --- a/configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt b/configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt deleted file mode 100644 index de3bcc8..0000000 --- a/configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//14*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 14; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt b/configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt deleted file mode 100644 index cfe9cac..0000000 --- a/configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt +++ /dev/null @@ -1,69 +0,0 @@ -//18*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 18; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config b/configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config deleted file mode 100644 index 56dbb17..0000000 --- a/configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config +++ /dev/null @@ -1,87 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 13 - -# high level architecture configuration --gpgpu_n_clusters 10 --gpgpu_n_cores_per_cluster 3 --gpgpu_n_mem 8 --gpgpu_clock_domains 325.0:650.0:650.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo --gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file config_quadro_islip.icnt - -# dram scheduler config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 325MHz = 74 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 58 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --dram_data_command_freq_ratio 2 # GDDR3 is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt b/configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt deleted file mode 100644 index a11bd8e..0000000 --- a/configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//20*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 26; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; diff --git a/configs/4.x-cfgs/SM2_C2050/gpgpusim.config b/configs/4.x-cfgs/SM2_C2050/gpgpusim.config deleted file mode 100644 index aa5f5f3..0000000 --- a/configs/4.x-cfgs/SM2_C2050/gpgpusim.config +++ /dev/null @@ -1,133 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - -# Using cuobjdump to extract ptx/SASS -#-gpgpu_ptx_use_cuobjdump 1 # use default - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - - -# high level architecture configuration --gpgpu_n_clusters 14 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 575.0:575.0:575.0:750.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 750MHz / 575MHz = 130 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 114 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt b/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt index 7820e4e..d372b26 100644 --- a/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt +++ b/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt @@ -1,6 +1,6 @@ //21*1 fly with 32 flits per packet under gpgpusim injection mode use_map = 0; -flit_size = 32; +flit_size = 40; // currently we do not use this, see subnets below network_count = 2; @@ -17,7 +17,7 @@ routing_function = dest_tag; // Flow control num_vcs = 1; -vc_buf_size = 8; +vc_buf_size = 64; wait_for_tail_credit = 0; diff --git a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config index 03fcda1..7f8da49 100644 --- a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config @@ -50,20 +50,25 @@ # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,S:128:8,8 -gpgpu_shmem_size 49152 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 # The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:32:8,8 #-gpgpu_shmem_size 16384 # 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2 S:64:128:8,L:B:m:L:L,A:256:4,4:0,32 -gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 +-memory_partition_indexing 0 --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,S:2:32,4 -gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,S:2:32,4 # enable operand collector -gpgpu_operand_collector_num_units_sp 6 @@ -76,6 +81,7 @@ -gpgpu_shmem_num_banks 32 -gpgpu_shmem_limited_broadcast 0 -gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 20 -gpgpu_max_insn_issue_per_warp 1 @@ -110,6 +116,13 @@ -gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + # Fermi has two schedulers per core -gpgpu_num_sched_per_core 2 # Two Level Scheduler with active and pending pools diff --git a/configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt b/configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt deleted file mode 100644 index 069ca02..0000000 --- a/configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 7; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM5_GTX750/gpgpusim.config b/configs/4.x-cfgs/SM5_GTX750/gpgpusim.config deleted file mode 100644 index 9366f93..0000000 --- a/configs/4.x-cfgs/SM5_GTX750/gpgpusim.config +++ /dev/null @@ -1,131 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 52 - - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 5 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 2 --gpgpu_n_sub_partition_per_mchannel 1 - -# Maxwell clock domains -#-gpgpu_clock_domains ::: -# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. --gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 8 --gpgpu_num_sfu_units 32 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 6,12,13,13,210 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 6,12,6,6,374 --ptx_opcode_initiation_fp 1,1,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 --gmem_skip_L1D 1 --gpgpu_shmem_size 65536 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache --gpgpu_cache:dl2 N:1024:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:16:128:32,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 150 --dram_latency 130 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 300 - -# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 32 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Maxwell has four schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs -# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. -# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx750Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml b/configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml deleted file mode 100755 index e2b2324..0000000 --- a/configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml +++ /dev/null @@ -1,539 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt deleted file mode 100644 index 2a69ddd..0000000 --- a/configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 50; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config b/configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config deleted file mode 100644 index 47c2b6a..0000000 --- a/configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config +++ /dev/null @@ -1,149 +0,0 @@ -# This config models the Pascal GP102 (GeForceGTX 1080Ti) - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 11 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 1,2,1,1,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache -# The default is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 --gpgpu_shmem_size 98304 --gmem_skip_L1D 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 --gpgpu_cache:dl2_texture_only 0 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 -# gpgpu_num_reg_banks should be increased to 32, but it gives an error! --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) -# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Pascal has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx1080Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml b/configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml deleted file mode 100755 index 02619ff..0000000 --- a/configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt new file mode 100644 index 0000000..81153b0 --- /dev/null +++ b/configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 128; +input_buffer_size = 256; +ejection_buffer_size = 128; +boundary_buffer_size = 128; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config b/configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config new file mode 100644 index 0000000..edcd919 --- /dev/null +++ b/configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config @@ -0,0 +1,174 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 60 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 32 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 32 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Pscal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Nvidia_Tesla +-gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,2,2,1,2,2,2,1,6 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 2 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 2,2,2,2,130 +-ptx_opcode_latency_sfu 8 +-ptx_opcode_initiation_sfu 4 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +-gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 65536 +-gpgpu_shmem_size_PrefShared 65536 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 + +# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 +-memory_partition_indexing 2 + +# 4 KB Inst. +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt deleted file mode 100644 index 616e9f3..0000000 --- a/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt +++ /dev/null @@ -1,74 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 128; -input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config deleted file mode 100644 index da98547..0000000 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ /dev/null @@ -1,186 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 100 --ptx_opcode_initiation_sfu 8 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:8,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:16,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 1 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/src/gpgpu-sim/addrdec.cc b/src/gpgpu-sim/addrdec.cc index cfd90ec..e7e27b7 100644 --- a/src/gpgpu-sim/addrdec.cc +++ b/src/gpgpu-sim/addrdec.cc @@ -132,7 +132,7 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, addrdec_ chip[0] = a[13]^a[12]^a[11]^a[10]^a[9]^a[6]^a[5]^a[3]^a[0]^chip[0]; chip[1] = a[14]^a[13]^a[12]^a[11]^a[10]^a[7]^a[6]^a[4]^a[1]^chip[1]; chip[2] = a[14]^a[10]^a[9]^a[8]^a[7]^a[6]^a[3]^a[2]^a[0]^chip[2]; - chip[3] = a[11]^a[10]^a[9]^a[8]^a[7]^a[4]^a[3]^a[1]^chip[3]; + chip[3] = a[11]^a[10]^a[9]^a[8]^a[7]^a[4]^a[3]^a[1]^chip[3]; chip[4] = a[12]^a[11]^a[10]^a[9]^a[8]^a[5]^a[4]^a[2]^chip[4]; tlx->chip = chip.to_ulong(); @@ -144,8 +144,21 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, addrdec_ assert(tlx->chip < m_n_channel); break; case CUSTOM: - /* No custom set function implemented */ + { + //random selected bits + //do you custom hashing function here, similar to + //Liu, Yuxi, et al. "Get Out of the Valley: Power-Efficient Address Mapping for GPUs." ISCA 2018 + std::bitset<64> b(tlx->row); + std::bitset<5> chip(tlx->chip); + chip[0] = b[13]^b[10]^b[9]^b[5]^b[0]^chip[0]; + chip[1] = b[12]^b[11]^b[6]^b[1]^chip[1]; + chip[2] = b[14]^b[9]^b[8]^b[7]^b[2]^chip[2]; + chip[3] = b[11]^b[10]^b[8]^b[3]^chip[3]; + chip[4] = b[12]^b[9]^b[8]^b[5]^b[4]^chip[4]; + tlx->chip = chip.to_ulong(); + assert(tlx->chip < m_n_channel); break; + } default: assert("\nUndefined set index function.\n" && 0); break; diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index 92aa819..ac63327 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -730,13 +730,15 @@ void dram_t::print( FILE* simFile) const printf("\nwrite_to_read_ratio_blp_rw_average = %.6f", write_to_read_ratio_blp_rw_average /banks_access_rw_total); printf("\nGrpLevelPara = %.6f \n", (float)bkgrp_parallsim_rw /banks_access_rw_total); - printf("\nbwutil = %.6f \n", (float)bwutil/n_cmd); + printf("\nBW Util details:\n"); + printf("bwutil = %.6f \n", (float)bwutil/n_cmd); printf("total_CMD = %d \n", n_cmd); printf("util_bw = %d \n", util_bw); printf("Wasted_Col = %d \n", wasted_bw_col); - printf("Wasted_Row %d \n", wasted_bw_row); - printf("Idle = %d \n\n", idle_bw); + printf("Wasted_Row = %d \n", wasted_bw_row); + printf("Idle = %d \n", idle_bw); + printf("\nBW Util Bottlenecks: \n"); printf("RCDc_limit = %d \n", RCDc_limit); printf("RCDWRc_limit = %d \n", RCDWRc_limit); printf("WTRc_limit = %d \n", WTRc_limit); @@ -747,6 +749,7 @@ void dram_t::print( FILE* simFile) const printf("WTRc_limit_alone = %d \n", WTRc_limit_alone); printf("RTWc_limit_alone = %d \n", RTWc_limit_alone); + printf("\nCommands details: \n"); printf("total_CMD = %d \n", n_cmd); printf("n_nop = %d \n", n_nop); printf("Read = %d \n", n_rd); @@ -757,8 +760,9 @@ void dram_t::print( FILE* simFile) const printf("n_pre = %d \n", n_pre); printf("n_ref = %d \n", n_ref); printf("n_req = %d \n", n_req ); - printf("total_req = %d \n\n", n_rd+n_wr+n_rd_L2_A+n_wr_WB); + printf("total_req = %d \n", n_rd+n_wr+n_rd_L2_A+n_wr_WB); + printf("\nDual Bus Interface Util: \n"); printf("issued_total_row = %lu \n", issued_total_row); printf("issued_total_col = %lu \n", issued_total_col); printf("Row_Bus_Util = %.6f \n", (float)issued_total_row / n_cmd); diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc index f754d36..ff50050 100644 --- a/src/gpgpu-sim/dram_sched.cc +++ b/src/gpgpu-sim/dram_sched.cc @@ -109,12 +109,14 @@ dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row ) if(m_config->seperate_write_queue_enabled) { if(m_mode == READ_MODE && ((m_num_write_pending >= m_config->write_high_watermark ) - || (m_queue[bank].empty() && !m_write_queue[bank].empty()))) { + // || (m_queue[bank].empty() && !m_write_queue[bank].empty()) + )) { m_mode = WRITE_MODE; } else if(m_mode == WRITE_MODE && (( m_num_write_pending < m_config->write_low_watermark ) - || (!m_queue[bank].empty() && m_write_queue[bank].empty()))){ + // || (!m_queue[bank].empty() && m_write_queue[bank].empty()) + )){ m_mode = READ_MODE; } } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index d48de25..ea2dfba 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1495,7 +1495,8 @@ void gpgpu_sim::cycle() } else { mem_fetch* mf = (mem_fetch*) icnt_pop( m_shader_config->mem2device(i) ); m_memory_sub_partition[i]->push( mf, gpu_sim_cycle + gpu_tot_sim_cycle ); - partiton_reqs_in_parallel_per_cycle++; + if(mf) + partiton_reqs_in_parallel_per_cycle++; } m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle+gpu_tot_sim_cycle); m_memory_sub_partition[i]->accumulate_L2cache_stats(m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 4828875..b59e5d2 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -451,7 +451,7 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout, "gpgpu_n_intrawarp_mshr_merge = %d\n", gpgpu_n_intrawarp_mshr_merge); fprintf(fout, "gpgpu_n_cmem_portconflict = %d\n", gpgpu_n_cmem_portconflict); - fprintf(fout, "gpgpu_stall_shd_mem[c_mem][bk_conf] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][BK_CONF]); + fprintf(fout, "gpgpu_stall_shd_mem[c_mem][resource_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][BK_CONF]); //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][MSHR_RC_FAIL]); //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][ICNT_RC_FAIL]); //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][DATA_PORT_STALL]); @@ -459,7 +459,7 @@ void shader_core_stats::print( FILE* fout ) const //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][ICNT_RC_FAIL]); //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][DATA_PORT_STALL]); fprintf(fout, "gpgpu_stall_shd_mem[s_mem][bk_conf] = %d\n", gpu_stall_shd_mem_breakdown[S_MEM][BK_CONF]); - fprintf(fout, "gpgpu_stall_shd_mem[gl_mem][bk_conf] = %d\n", + fprintf(fout, "gpgpu_stall_shd_mem[gl_mem][resource_stall] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][BK_CONF] + gpu_stall_shd_mem_breakdown[G_MEM_ST][BK_CONF] + gpu_stall_shd_mem_breakdown[L_MEM_LD][BK_CONF] + -- cgit v1.3 From f852a7108e691045dd3910065836a817babcde8c Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 7 Sep 2018 22:01:31 -0400 Subject: adding streamin cache + fixing TEX cache + adding l1 latency and smem latency --- configs/4.x-cfgs/SM2_GTX480/gpgpusim.config | 2 +- configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt | 73 --------- configs/4.x-cfgs/SM6_P100/gpgpusim.config | 174 --------------------- .../SM6_P100_64SMs/config_fermi_islip.icnt | 73 --------- configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config | 174 --------------------- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 27 ++-- src/abstract_hardware_model.h | 1 + src/gpgpu-sim/gpu-cache.cc | 20 ++- src/gpgpu-sim/gpu-cache.h | 36 ++++- src/gpgpu-sim/gpu-sim.cc | 6 + src/gpgpu-sim/l2cache.cc | 8 +- src/gpgpu-sim/shader.cc | 126 ++++++++++++++- src/gpgpu-sim/shader.h | 6 + 13 files changed, 207 insertions(+), 519 deletions(-) delete mode 100644 configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM6_P100/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config (limited to 'configs') diff --git a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config index 7f8da49..35341f7 100644 --- a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config @@ -67,7 +67,7 @@ -memory_partition_indexing 0 -gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,S:2:32,4 --gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,T:128:4,128:2 -gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,S:2:32,4 # enable operand collector diff --git a/configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt deleted file mode 100644 index e7c2c3b..0000000 --- a/configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 60; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 128; -input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM6_P100/gpgpusim.config b/configs/4.x-cfgs/SM6_P100/gpgpusim.config deleted file mode 100644 index a4e745d..0000000 --- a/configs/4.x-cfgs/SM6_P100/gpgpusim.config +++ /dev/null @@ -1,174 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 32 --gpgpu_n_sub_partition_per_mchannel 1 - -# Pscal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Nvidia_Tesla --gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,2,2,1,2,2,2,1,6 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 2 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 2,2,2,2,130 --ptx_opcode_latency_sfu 8 --ptx_opcode_initiation_sfu 4 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 65536 --gpgpu_shmem_size_PrefShared 65536 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 --memory_partition_indexing 2 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt deleted file mode 100644 index 81153b0..0000000 --- a/configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 128; -input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config b/configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config deleted file mode 100644 index edcd919..0000000 --- a/configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config +++ /dev/null @@ -1,174 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 32 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 32 --gpgpu_n_sub_partition_per_mchannel 1 - -# Pscal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Nvidia_Tesla --gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,2,2,1,2,2,2,1,6 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 2 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 2,2,2,2,130 --ptx_opcode_latency_sfu 8 --ptx_opcode_initiation_sfu 4 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 65536 --gpgpu_shmem_size_PrefShared 65536 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 --memory_partition_indexing 2 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index 7368882..9ea7202 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -56,24 +56,29 @@ -ptx_opcode_initiation_sfu 4 -ptx_opcode_latency_sfu 8 + +# latencies and cache configs are adopted from: +# https://arxiv.org/pdf/1804.06826.pdf # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 24KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +# Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB +# Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres are used +-gpgpu_cache:dl1 S:48:128:4,L:L:s:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:48:128:4,L:L:s:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:48:128:4,L:L:s:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 49152 -gpgpu_shmem_size_PrefL1 49152 -gpgpu_shmem_size_PrefShared 49152 -gmem_skip_L1D 1 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 82 +-smem_latency 24 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:128:4,16:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 -perf_sim_memcpy 0 @@ -81,8 +86,7 @@ # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 # 48 KB Tex -# this is unused --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 # 12 KB Const -gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 @@ -99,7 +103,6 @@ -gpgpu_operand_collector_num_out_ports_mem 1 -gpgpu_operand_collector_num_in_ports_dp 1 -gpgpu_operand_collector_num_out_ports_dp 1 -# gpgpu_num_reg_banks should be increased to 32 -gpgpu_num_reg_banks 32 # shared memory bankconflict detection @@ -118,7 +121,7 @@ -inter_config_file config_fermi_islip.icnt # memory partition latency config --rop_latency 100 +-rop_latency 120 -dram_latency 100 # dram model config @@ -128,7 +131,7 @@ # To allow 100% DRAM utility, there should at least be enough buffer to sustain # the minimum DRAM latency (100 core cycles). I.e. # Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_frfcfs_dram_sched_queue_size 64 -gpgpu_dram_return_queue_size 240 # for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 1b764e2..a70b077 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -392,6 +392,7 @@ protected: #define LOCAL_MEM_SIZE_MAX (8*1024) #define MAX_STREAMING_MULTIPROCESSORS 64 #define MAX_THREAD_PER_SM 2048 +#define MAX_WARP_PER_SM 64 #define TOTAL_LOCAL_MEM_PER_SM (MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX) #define TOTAL_SHARED_MEM (MAX_STREAMING_MULTIPROCESSORS*SHARED_MEM_SIZE_MAX) #define TOTAL_LOCAL_MEM (MAX_STREAMING_MULTIPROCESSORS*MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX) diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index f181f20..a11853a 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -71,6 +71,7 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{ switch(m_set_index_function){ case FERMI_HASH_SET_FUNCTION: + case BITWISE_XORING_FUNCTION: /* * Set Indexing function from "A Detailed GPU Cache Model Based on Reuse Distance Theory" * Cedric Nugteren et al. @@ -1581,7 +1582,7 @@ enum cache_request_status tex_cache::access( new_addr_type addr, mem_fetch *mf, if ( status == MISS ) { // we need to send a memory request... unsigned rob_index = m_rob.push( rob_entry(cache_index, mf, block_addr) ); - m_extra_mf_fields[mf] = extra_mf_fields(rob_index); + m_extra_mf_fields[mf] = extra_mf_fields(rob_index, m_config); mf->set_data_size(m_config.get_line_sz()); m_tags.fill(cache_index,time,mf); // mark block as valid m_request_fifo.push(mf); @@ -1636,6 +1637,23 @@ void tex_cache::cycle(){ /// Place returning cache block into reorder buffer void tex_cache::fill( mem_fetch *mf, unsigned time ) { + if(m_config.m_mshr_type == SECTOR_TEX_FIFO) { + assert(mf->get_original_mf()); + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf->get_original_mf()); + assert( e != m_extra_mf_fields.end() ); + e->second.pending_read--; + + if(e->second.pending_read > 0) { + //wait for the other requests to come back + delete mf; + return; + } else { + mem_fetch *temp = mf; + mf = mf->get_original_mf(); + delete temp; + } + } + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf); assert( e != m_extra_mf_fields.end() ); assert( e->second.m_valid ); diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index dee43f2..4ed382c 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -455,7 +455,8 @@ enum write_policy_t { enum allocation_policy_t { ON_MISS, - ON_FILL + ON_FILL, + STREAMING }; @@ -467,8 +468,9 @@ enum write_allocate_policy_t { }; enum mshr_config_t { - TEX_FIFO, + TEX_FIFO, // Tex cache ASSOC, // normal cache + SECTOR_TEX_FIFO, //Tex cache sends requests to high-level sector cache SECTOR_ASSOC // normal cache sends requests to high-level sector cache }; @@ -485,6 +487,12 @@ enum cache_type{ SECTOR }; +#define MAX_WARP_PER_SHADER 64 +#define INCT_TOTAL_BUFFER 64 +#define L2_TOTAL 64 +#define MAX_WARP_PER_SHADER 64 +#define MAX_WARP_PER_SHADER 64 + class cache_config { public: cache_config() @@ -544,10 +552,27 @@ public: switch (ap) { case 'm': m_alloc_policy = ON_MISS; break; case 'f': m_alloc_policy = ON_FILL; break; + case 's': m_alloc_policy = STREAMING; break; default: exit_parse_error(); } + if(m_alloc_policy == STREAMING) { + //For streaming cache, we set the alloc policy to be on-fill to remove all line_alloc_fail stalls + //we set the MSHRs to be equal to the cache line. This is possible by moving TAG to be shared between cache line and MSHR enrty (i.e. for each cache line, there is an MSHR rntey associated with it) + // This is the easiest think we can think about to model (mimics) L1 streaming cache in Pascal and Volta + //Based on our microbenchmakrs, MSHRs entries have been increasing substantially in Pascal and Volta + //For more information about streaming cache, see: + // http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + // https://ieeexplore.ieee.org/document/8344474/ + + m_alloc_policy = ON_FILL; + m_mshr_entries = m_nset*m_assoc; + if(m_cache_type == SECTOR) + m_mshr_entries *= SECTOR_CHUNCK_SIZE; + m_mshr_max_merge = MAX_WARP_PER_SM; + } switch (mshr_type) { case 'F': m_mshr_type = TEX_FIFO; assert(ntok==14); break; + case 'T': m_mshr_type = SECTOR_TEX_FIFO; assert(ntok==14); break; case 'A': m_mshr_type = ASSOC; break; case 'S' : m_mshr_type = SECTOR_ASSOC; break; default: exit_parse_error(); @@ -722,6 +747,7 @@ class l1d_cache_config : public cache_config{ public: l1d_cache_config() : cache_config(){} virtual unsigned set_index(new_addr_type addr) const; + unsigned l1_latency; }; class l2_cache_config : public cache_config { @@ -1442,7 +1468,7 @@ public: m_result_fifo(config.m_result_fifo_entries) { m_name = name; - assert(config.m_mshr_type == TEX_FIFO); + assert(config.m_mshr_type == TEX_FIFO || config.m_mshr_type == SECTOR_TEX_FIFO ); assert(config.m_write_policy == READ_ONLY); assert(config.m_alloc_policy == ON_MISS); m_memport=memport; @@ -1595,13 +1621,15 @@ private: struct extra_mf_fields { extra_mf_fields() { m_valid = false;} - extra_mf_fields( unsigned i ) + extra_mf_fields( unsigned i, const cache_config &m_config ) { m_valid = true; m_rob_index = i; + pending_read = m_config.m_mshr_type == SECTOR_TEX_FIFO? m_config.m_line_sz/SECTOR_SIZE : 0; } bool m_valid; unsigned m_rob_index; + unsigned pending_read; }; cache_stats m_stats; diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index ea2dfba..08d4525 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -250,6 +250,12 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); + option_parser_register(opp, "-l1_latency", OPT_UINT32, &m_L1D_config.l1_latency, + "L1 Hit Latency", + "0"); + option_parser_register(opp, "-smem_latency", OPT_UINT32, &smem_latency, + "smem Latency", + "3"); option_parser_register(opp, "-gpgpu_cache:dl1PrefL1", OPT_CSTR, &m_L1D_config.m_config_stringPrefL1, "per-shader L1 data cache config " " {::,:::,::, | none}", diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index f42610f..25da107 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -609,7 +609,7 @@ std::vector memory_sub_partition::breakdown_request_to_sector_reques } else { printf("Invalid sector received, address = 0x%06x, sector mask = %s, data size = %d", - mf->get_addr(), mf->get_access_sector_mask(), mf->get_data_size(), mf->get_data_size()); + mf->get_addr(), mf->get_access_sector_mask(), mf->get_data_size()); assert(0 && "Undefined sector mask is received"); } @@ -640,7 +640,11 @@ std::vector memory_sub_partition::breakdown_request_to_sector_reques result.push_back(n_mf); byte_sector_mask <<= SECTOR_SIZE; } - } else assert(0 && "Undefined data size is received"); + } else { + printf("Invalid sector received, address = 0x%06x, sector mask = %d, byte mask = , data size = %d", + mf->get_addr(), mf->get_access_sector_mask().count(), mf->get_data_size()); + assert(0 && "Undefined data size is received"); + } return result; } diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index b59e5d2..51689e3 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1488,6 +1488,111 @@ mem_stage_stall_type ldst_unit::process_memory_access_queue( cache_t *cache, war return process_cache_access( cache, mf->get_addr(), inst, events, mf, status ); } +mem_stage_stall_type ldst_unit::process_memory_access_queue_l1cache( l1_cache *cache, warp_inst_t &inst ) +{ + mem_stage_stall_type result = NO_RC_FAIL; + if( inst.accessq_empty() ) + return result; + + mem_fetch *mf = m_mf_allocator->alloc(inst,inst.accessq_back()); + + if(m_config->m_L1D_config.l1_latency > 0) + { + if((l1_latency_queue[m_config->m_L1D_config.l1_latency-1]) == NULL) + { + l1_latency_queue[m_config->m_L1D_config.l1_latency-1] = mf; + + if( mf->get_inst().is_store() ) { + unsigned inc_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)? + (mf->get_data_size()/SECTOR_SIZE) : 1; + + for(unsigned i=0; i< inc_ack; ++i) + m_core->inc_store_req( inst.warp_id() ); + } + + inst.accessq_pop_back(); + } + else + { + result = BK_CONF; + delete mf; + } + if( !inst.accessq_empty() && result !=BK_CONF) + result = COAL_STALL; + return result; + } + else + { + std::list events; + enum cache_request_status status = cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle,events); + return process_cache_access( cache, mf->get_addr(), inst, events, mf, status ); + } +} + +void ldst_unit::L1_latency_queue_cycle() +{ + //std::deque< std::pair >::iterator it = m_latency_queue.begin(); + if((l1_latency_queue[0]) != NULL) + { + mem_fetch* mf_next = l1_latency_queue[0]; + std::list events; + enum cache_request_status status = m_L1D->access(mf_next->get_addr(),mf_next,gpu_sim_cycle+gpu_tot_sim_cycle,events); + + bool write_sent = was_write_sent(events); + bool read_sent = was_read_sent(events); + + if ( status == HIT ) { + assert( !read_sent ); + l1_latency_queue[0] = NULL; + if ( mf_next->get_inst().is_load() ) { + for ( unsigned r=0; r < 4; r++) + if (mf_next->get_inst().out[r] > 0) + { + assert(m_pending_writes[mf_next->get_inst().warp_id()][mf_next->get_inst().out[r]]>0); + unsigned still_pending = --m_pending_writes[mf_next->get_inst().warp_id()][mf_next->get_inst().out[r]]; + if(!still_pending) + { + m_pending_writes[mf_next->get_inst().warp_id()].erase(mf_next->get_inst().out[r]); + m_scoreboard->releaseRegister(mf_next->get_inst().warp_id(),mf_next->get_inst().out[r]); + m_core->warp_inst_complete(mf_next->get_inst()); + } + } + } + + //For write hit in WB policy + if(mf_next->get_inst().is_store() && !write_sent) + { + unsigned dec_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)? + (mf_next->get_data_size()/SECTOR_SIZE) : 1; + + mf_next->set_reply(); + + for(unsigned i=0; i< dec_ack; ++i) + m_core->store_ack(mf_next); + } + + if( !write_sent ) + delete mf_next; + + } else if ( status == RESERVATION_FAIL ) { + assert( !read_sent ); + assert( !write_sent ); + } else { + assert( status == MISS || status == HIT_RESERVED ); + l1_latency_queue[0] = NULL; + } + } + + for( unsigned stage = 0; stagem_L1D_config.l1_latency-1; ++stage) + if( l1_latency_queue[stage] == NULL) { + l1_latency_queue[stage] = l1_latency_queue[stage+1] ; + l1_latency_queue[stage+1] = NULL; + } + +} + + + bool ldst_unit::constant_cycle( warp_inst_t &inst, mem_stage_stall_type &rc_fail, mem_stage_access_type &fail_type) { if( inst.empty() || ((inst.space.get_type() != const_space) && (inst.space.get_type() != param_space_kernel)) ) @@ -1561,7 +1666,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea } } else { assert( CACHE_UNDEFINED != inst.cache_op ); - stall_cond = process_memory_access_queue(m_L1D,inst); + stall_cond = process_memory_access_queue_l1cache(m_L1D,inst); } if( !inst.accessq_empty() && stall_cond == NO_RC_FAIL) stall_cond = COAL_STALL; @@ -1771,8 +1876,9 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt, const memory_config *mem_config, shader_core_stats *stats, unsigned sid, - unsigned tpc ) : pipelined_simd_unit(NULL,config,3,core), m_next_wb(config) + unsigned tpc ) : pipelined_simd_unit(NULL,config,config->smem_latency,core), m_next_wb(config) { + assert(config->smem_latency > 1); init( icnt, mf_allocator, core, @@ -1793,6 +1899,12 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt, m_icnt, m_mf_allocator, IN_L1D_MISS_QUEUE ); + + if(m_config->m_L1D_config.l1_latency > 0) + { + for(int i=0; im_L1D_config.l1_latency; i++ ) + l1_latency_queue.push_back((mem_fetch*)NULL); + } } } @@ -2019,7 +2131,11 @@ void ldst_unit::cycle() m_L1T->cycle(); m_L1C->cycle(); - if( m_L1D ) m_L1D->cycle(); + if( m_L1D ) { + m_L1D->cycle(); + if(m_config->m_L1D_config.l1_latency > 0) + L1_latency_queue_cycle(); + } warp_inst_t &pipe_reg = *m_dispatch_reg; enum mem_stage_stall_type rc_fail = NO_RC_FAIL; @@ -2042,9 +2158,9 @@ void ldst_unit::cycle() unsigned warp_id = pipe_reg.warp_id(); if( pipe_reg.is_load() ) { if( pipe_reg.space.get_type() == shared_space ) { - if( m_pipeline_reg[2]->empty() ) { + if( m_pipeline_reg[m_config->smem_latency-1]->empty() ) { // new shared memory request - move_warp(m_pipeline_reg[2],m_dispatch_reg); + move_warp(m_pipeline_reg[m_config->smem_latency-1],m_dispatch_reg); m_dispatch_reg->clear(); } } else { diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index cc441b3..e07096e 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1220,6 +1220,7 @@ protected: mem_fetch *mf, enum cache_request_status status ); mem_stage_stall_type process_memory_access_queue( cache_t *cache, warp_inst_t &inst ); + mem_stage_stall_type process_memory_access_queue_l1cache( l1_cache *cache, warp_inst_t &inst ); const memory_config *m_memory_config; class mem_fetch_interface *m_icnt; @@ -1248,6 +1249,9 @@ protected: // for debugging unsigned long long m_last_inst_gpu_sim_cycle; unsigned long long m_last_inst_gpu_tot_sim_cycle; + + std::deque l1_latency_queue; + void L1_latency_queue_cycle(); }; enum pipeline_stage_name_t { @@ -1399,6 +1403,8 @@ struct shader_core_config : public core_config int simt_core_sim_order; + unsigned smem_latency; + unsigned mem2device(unsigned memid) const { return memid + n_simt_clusters; } //Jin: concurrent kernel on sm -- cgit v1.3 From 46817ba62f475ba1202c31a8939ea1a1de81a410 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Tue, 11 Sep 2018 21:46:07 -0400 Subject: Adding in the occupancy nymber --- configs/3.x-cfgs/SM6_TITANX/gpgpusim.config | 1 + 1 file changed, 1 insertion(+) (limited to 'configs') diff --git a/configs/3.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/3.x-cfgs/SM6_TITANX/gpgpusim.config index f78bd02..28912a3 100644 --- a/configs/3.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/3.x-cfgs/SM6_TITANX/gpgpusim.config @@ -25,6 +25,7 @@ # shader core pipeline config -gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 61 # This implies a maximum of 64 warps/SM -gpgpu_shader_core_pipeline 2048:32 -- cgit v1.3 From 96bbd0063b479ad51f3d296f5e87a56151536d86 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 12 Sep 2018 21:03:00 -0400 Subject: change L1 indexing to linear for TITANX --- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index 9ea7202..8b47f98 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -65,9 +65,9 @@ # Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB # Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres are used --gpgpu_cache:dl1 S:48:128:4,L:L:s:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:48:128:4,L:L:s:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:48:128:4,L:L:s:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1 S:48:128:4,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:48:128:4,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:48:128:4,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 49152 -gpgpu_shmem_size_PrefL1 49152 -gpgpu_shmem_size_PrefShared 49152 -- cgit v1.3 From dd0a4eef2ae526984ac3e7dc93a0d1cf92a9a6a8 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 18 Sep 2018 10:52:10 -0400 Subject: removing some comments from config files --- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index 9ea7202..4c0586d 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -109,7 +109,7 @@ -gpgpu_shmem_num_banks 32 -gpgpu_shmem_limited_broadcast 0 -gpgpu_shmem_warp_parts 1 -# Use Fermi Coalsce arhitetecture which is the same as Pascal +# Use Pascal Coalsce arhitetecture -gpgpu_coalesce_arch 61 ## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units @@ -144,8 +144,7 @@ -gpgpu_mem_address_mask 1 -gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +# Use the same GDDR5 timing -gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" -- cgit v1.3 From 9463cf92f121f8cc18a7e601c5838c3911d303b0 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 18 Sep 2018 12:40:36 -0400 Subject: double internal speedup for TITANX router --- configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt index 714d933..dec4789 100644 --- a/configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt +++ b/configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt @@ -35,9 +35,9 @@ routing_delay = 0; vc_alloc_delay = 1; sw_alloc_delay = 1; -input_speedup = 2; +input_speedup = 1; output_speedup = 1; -internal_speedup = 1.0; +internal_speedup = 2.0; // Traffic, GPGPU-Sim does not use this -- cgit v1.3 From 79a890fc4c22135a368504a5112a0f0a9982659e Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 20 Sep 2018 16:48:23 -0400 Subject: fixing some config file --- configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt | 4 ++-- configs/4.x-cfgs/SM2_GTX480/gpgpusim.config | 9 ++++++--- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 6 +++--- 3 files changed, 11 insertions(+), 8 deletions(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt b/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt index d372b26..c399db9 100644 --- a/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt +++ b/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt @@ -32,9 +32,9 @@ routing_delay = 0; vc_alloc_delay = 1; sw_alloc_delay = 1; -input_speedup = 2; +input_speedup = 1; output_speedup = 1; -internal_speedup = 1.0; +internal_speedup = 2.0; // Traffic, GPGPU-Sim does not use this diff --git a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config index 35341f7..c96432b 100644 --- a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config @@ -50,13 +50,16 @@ # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,S:128:8,8 +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,S:64:8,8 -gpgpu_shmem_size 49152 -icnt_flit_size 40 +-gmem_skip_L1D 0 -gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 35 +-smem_latency 26 # The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:32:8,8 +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:64:8,8 #-gpgpu_shmem_size 16384 # 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache @@ -100,7 +103,7 @@ # To allow 100% DRAM utility, there should at least be enough buffer to sustain # the minimum DRAM latency (100 core cycles). I.e. # Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_frfcfs_dram_sched_queue_size 64 -gpgpu_dram_return_queue_size 116 # for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index 8e93723..45a87cd 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -65,9 +65,9 @@ # Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB # Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres are used --gpgpu_cache:dl1 S:48:128:4,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:48:128:4,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:48:128:4,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 49152 -gpgpu_shmem_size_PrefL1 49152 -gpgpu_shmem_size_PrefShared 49152 -- cgit v1.3 From 7e414028b3754d40ab92d10205dc412316c4ab29 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Sun, 7 Oct 2018 13:37:24 -0400 Subject: adding Volta V100 config --- .../4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 74 ++++++++ configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 187 +++++++++++++++++++++ 2 files changed, 261 insertions(+) create mode 100644 configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM7_TITANV/gpgpusim.config (limited to 'configs') diff --git a/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..cea3acf --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 256; +input_buffer_size = 256; +ejection_buffer_size = 128; +boundary_buffer_size = 128; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config new file mode 100644 index 0000000..de3558a --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -0,0 +1,187 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Volta NVIDIA GV100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Defualt config is 64KB DL1 and 64KB shared memory +-gpgpu_cache:dl1 S:4:128:128,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:4:128:192,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 32768 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 28 +-smem_latency 19 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Volta V100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + -- cgit v1.3 From fdbb4aef32a36ff625612c82c58039699af2b23b Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 9 Oct 2018 12:07:25 -0400 Subject: flushing L1 cache --- configs/4.x-cfgs/SM2_GTX480/gpgpusim.config | 1 + configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 1 + configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 1 + 3 files changed, 3 insertions(+) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config index c96432b..b965aff 100644 --- a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config @@ -57,6 +57,7 @@ -gpgpu_n_cluster_ejection_buffer_size 32 -l1_latency 35 -smem_latency 26 +-gpgpu_flush_l1_cache 1 # The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected #-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:64:8,8 diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index 45a87cd..3723163 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -76,6 +76,7 @@ -gpgpu_n_cluster_ejection_buffer_size 32 -l1_latency 82 -smem_latency 24 +-gpgpu_flush_l1_cache 1 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache -gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config index de3558a..cc4c931 100644 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -77,6 +77,7 @@ -gpgpu_n_cluster_ejection_buffer_size 32 -l1_latency 28 -smem_latency 19 +-gpgpu_flush_l1_cache 1 # 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache -gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 -- cgit v1.3 From 0e46a261dfeba9c19d1637f46277986d7eb1b9d8 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 9 Oct 2018 17:12:57 -0400 Subject: adding adaptive volta cache config --- configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 19 ++++++++++-------- src/abstract_hardware_model.cc | 2 ++ src/abstract_hardware_model.h | 2 ++ src/gpgpu-sim/gpu-cache.cc | 9 ++++----- src/gpgpu-sim/gpu-cache.h | 31 +++++++++++++++++++++++++---- src/gpgpu-sim/gpu-sim.cc | 3 +++ src/gpgpu-sim/shader.cc | 29 +++++++++++++++++++++++++++ src/gpgpu-sim/shader.h | 2 ++ 8 files changed, 80 insertions(+), 17 deletions(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config index cc4c931..a7056db 100644 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -65,19 +65,21 @@ # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo -# Defualt config is 64KB DL1 and 64KB shared memory --gpgpu_cache:dl1 S:4:128:128,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:4:128:192,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 32768 --gpgpu_shmem_size_PrefShared 98304 +# Defualt config is 32KB DL1 and 96KB shared memory +# In Volta, we assign the remaining shared memory to L1 cache +# if the assigned shd mem = 0, then L1 cache = 128KB +# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x +# disable this mode in case of multi kernels/apps execution +-adpative_volta_cache_config 1 +-gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 98304 -gmem_skip_L1D 0 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 -l1_latency 28 -smem_latency 19 -gpgpu_flush_l1_cache 1 +-adpative_volta_cache_config 1 # 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache -gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 @@ -87,7 +89,8 @@ # 128 KB Inst. -gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -# 48 KB Tex +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod -gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 # 64 KB Const -gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index d2a155c..39ed81c 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -591,6 +591,8 @@ kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info * //Jin: launch latency management m_launch_latency = g_kernel_launch_latency; + + volta_cache_config_set=false; } kernel_info_t::~kernel_info_t() diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index e708fa7..6fa2ba0 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -301,6 +301,8 @@ public: unsigned long long start_cycle; unsigned long long end_cycle; unsigned m_launch_latency; + + mutable bool volta_cache_config_set; }; struct core_config { diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index a11853a..8ae4702 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -29,7 +29,6 @@ #include "stat-tool.h" #include -#define MAX_DEFAULT_CACHE_SIZE_MULTIBLIER 4 // used to allocate memory that is large enough to adapt the changes in cache size across kernels const char * cache_request_status_str(enum cache_request_status status) @@ -165,7 +164,7 @@ unsigned l2_cache_config::set_index(new_addr_type addr) const{ tag_array::~tag_array() { - unsigned cache_lines_num = MAX_DEFAULT_CACHE_SIZE_MULTIBLIER*m_config.get_num_lines(); + unsigned cache_lines_num = m_config.get_max_num_lines(); for(unsigned i=0; iis_modified_line()) { for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; @@ -396,7 +395,7 @@ void tag_array::flush() void tag_array::invalidate() { - for (unsigned i=0; i < m_config.get_num_lines(); i++) + for (unsigned i=0; i < m_config.get_max_num_lines(); i++) for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; } diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 4ed382c..9174d7f 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -38,6 +38,8 @@ #include "addrdec.h" #include +#define MAX_DEFAULT_CACHE_SIZE_MULTIBLIER 4 + enum cache_block_state { INVALID=0, RESERVED, @@ -557,15 +559,15 @@ public: } if(m_alloc_policy == STREAMING) { //For streaming cache, we set the alloc policy to be on-fill to remove all line_alloc_fail stalls - //we set the MSHRs to be equal to the cache line. This is possible by moving TAG to be shared between cache line and MSHR enrty (i.e. for each cache line, there is an MSHR rntey associated with it) - // This is the easiest think we can think about to model (mimics) L1 streaming cache in Pascal and Volta + //we set the MSHRs to be equal to max allocated cache lines. This is possible by moving TAG to be shared between cache line and MSHR enrty (i.e. for each cache line, there is an MSHR rntey associated with it) + //This is the easiest think we can think about to model (mimic) L1 streaming cache in Pascal and Volta //Based on our microbenchmakrs, MSHRs entries have been increasing substantially in Pascal and Volta //For more information about streaming cache, see: // http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf // https://ieeexplore.ieee.org/document/8344474/ m_alloc_policy = ON_FILL; - m_mshr_entries = m_nset*m_assoc; + m_mshr_entries = m_nset*m_assoc*MAX_DEFAULT_CACHE_SIZE_MULTIBLIER; if(m_cache_type == SECTOR) m_mshr_entries *= SECTOR_CHUNCK_SIZE; m_mshr_max_merge = MAX_WARP_PER_SM; @@ -581,6 +583,7 @@ public: m_nset_log2 = LOGB2(m_nset); m_valid = true; m_atom_sz = (m_cache_type == SECTOR)? SECTOR_SIZE : m_line_sz; + original_m_assoc = m_assoc; //For more details about difference between FETCH_ON_WRITE and WRITE VALIDAE policies //Read: Jouppi, Norman P. "Cache write policies and performance". ISCA 93. @@ -646,7 +649,11 @@ public: assert( m_valid ); return m_nset * m_assoc; } - + unsigned get_max_num_lines() const + { + assert( m_valid ); + return MAX_DEFAULT_CACHE_SIZE_MULTIBLIER * m_nset * original_m_assoc; + } void print( FILE *fp ) const { fprintf( fp, "Size = %d B (%d Set x %d-way x %d byte line)\n", @@ -687,6 +694,21 @@ public: { return m_mshr_type; } + void set_assoc(unsigned n) + { + //set new assoc. L1 cache dynamically resized in Volta + m_assoc = n; + } + unsigned get_nset() const + { + assert( m_valid ); + return m_nset; + } + unsigned get_total_size_inKB() const + { + assert( m_valid ); + return (m_assoc*m_nset*m_line_sz)/1024; + } FuncCache get_cache_status() {return cache_status;} char *m_config_string; char *m_config_stringPrefL1; @@ -708,6 +730,7 @@ protected: unsigned m_nset_log2; unsigned m_assoc; unsigned m_atom_sz; + unsigned original_m_assoc; enum replacement_policy_t m_replacement_policy; // 'L' = LRU, 'F' = FIFO enum write_policy_t m_write_policy; // 'T' = write through, 'B' = write back, 'R' = read only diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 08d4525..080cbac 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -307,6 +307,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shmem_size", OPT_UINT32, &gpgpu_shmem_size, "Size of shared memory per shader core (default 16kB)", "16384"); + option_parser_register(opp, "-adpative_volta_cache_config", OPT_BOOL, &adpative_volta_cache_config, + "adpative_volta_cache_config", + "0"); option_parser_register(opp, "-gpgpu_shmem_size", OPT_UINT32, &gpgpu_shmem_sizeDefault, "Size of shared memory per shader core (default 16kB)", "16384"); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 0e2e1c2..2cc8a2e 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -2744,6 +2744,35 @@ unsigned int shader_core_config::max_cta( const kernel_info_t &k ) const abort(); } + if(adpative_volta_cache_config && !k.volta_cache_config_set) { + //For Volta, we assign the remaining shared memory to L1 cache + //For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x + unsigned total_shmed = kernel_info->smem * result; + assert(total_shmed >=0 && total_shmed <= gpgpu_shmem_size); + assert(gpgpu_shmem_size == 98304); //Volta has 96 KB shared + assert(m_L1D_config.get_nset() == 4); //Volta L1 has four sets + if(total_shmed < gpgpu_shmem_size){ + if(total_shmed == 0) + m_L1D_config.set_assoc(256); //L1 is 128KB ans shd=0 + else if(total_shmed > 0 && total_shmed <= 8192) + m_L1D_config.set_assoc(240); //L1 is 120KB ans shd=8KB + else if(total_shmed > 8192 && total_shmed <= 16384) + m_L1D_config.set_assoc(224); //L1 is 112KB ans shd=16KB + else if(total_shmed > 16384 && total_shmed <= 32768) + m_L1D_config.set_assoc(192); //L1 is 96KB ans shd=32KB + else if(total_shmed > 32768 && total_shmed <= 65536) + m_L1D_config.set_assoc(128); //L1 is 64KB ans shd=64KB + else if(total_shmed > 65536 && total_shmed <= gpgpu_shmem_size) + m_L1D_config.set_assoc(64); //L1 is 32KB and shd=96KB + else + assert(0); + + printf ("GPGPU-Sim: Reconfigure L1 cache in Volta Archi to %uKB\n", m_L1D_config.get_total_size_inKB()); + } + + k.volta_cache_config_set = true; + } + return result; } diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index e07096e..92b4159 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1409,6 +1409,8 @@ struct shader_core_config : public core_config //Jin: concurrent kernel on sm bool gpgpu_concurrent_kernel_sm; + + bool adpative_volta_cache_config; }; struct shader_core_stats_pod { -- cgit v1.3 From dff0779550852ea3eba72c8b113feb3c4978823f Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 10 Oct 2018 17:48:52 -0400 Subject: adding some comment to config files --- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 3 +++ 1 file changed, 3 insertions(+) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index 3723163..cc7419c 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -71,6 +71,8 @@ -gpgpu_shmem_size 49152 -gpgpu_shmem_size_PrefL1 49152 -gpgpu_shmem_size_PrefShared 49152 +# By default, L1 cache is disabled in Pascal P102. +# requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 -gmem_skip_L1D 1 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 @@ -87,6 +89,7 @@ # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 # 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod -gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 # 12 KB Const -gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 -- cgit v1.3 From 9321b08f408ca7f0af9f0fad14a03f1b23a7b670 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 10 Oct 2018 17:49:28 -0400 Subject: adding TITANV 3.x config files --- .../3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 70 +++++++++ configs/3.x-cfgs/SM7_TITANV/gpgpusim.config | 165 +++++++++++++++++++++ 2 files changed, 235 insertions(+) create mode 100644 configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/SM7_TITANV/gpgpusim.config (limited to 'configs') diff --git a/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..fac792a --- /dev/null +++ b/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config new file mode 100644 index 0000000..738c881 --- /dev/null +++ b/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config @@ -0,0 +1,165 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,1,4,4,1,9 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Volta GV100 has 64KB Shared memory +-gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 1 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 + +# 128 KB Inst. +-gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + -- cgit v1.3 From 99634600280d64ec2a346c1d729d0e7a94a95337 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 10 Oct 2018 17:53:56 -0400 Subject: fixing freq --- configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config index a7056db..b4576dc 100644 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -27,7 +27,7 @@ #-gpgpu_clock_domains ::: # Volta NVIDIA GV100 clock domains are adopted from # https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 +-gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 # boost mode # -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 -- cgit v1.3 From c58ee08f0e7ec7c7ee6794d8ce2616ed1b0c10f2 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 10 Oct 2018 17:56:50 -0400 Subject: fixing freq of TITANV --- configs/3.x-cfgs/SM7_TITANV/gpgpusim.config | 4 ++-- configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 4 ++-- 2 files changed, 4 insertions(+), 4 deletions(-) (limited to 'configs') diff --git a/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config index 738c881..aefb04a 100644 --- a/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config @@ -27,9 +27,9 @@ #-gpgpu_clock_domains ::: # Pascal NVIDIA GP100 clock domains are adopted from # https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 +-gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 # boost mode -# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 # shader core pipeline config -gpgpu_shader_registers 65536 diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config index b4576dc..41757f8 100644 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -27,9 +27,9 @@ #-gpgpu_clock_domains ::: # Volta NVIDIA GV100 clock domains are adopted from # https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 +-gpgpu_clock_domains 1200.0:2000.0:1200.0:850.0 # boost mode -# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 # shader core pipeline config -gpgpu_shader_registers 65536 -- cgit v1.3 From b1c00ed4369c648c2059cb6ecfdab48126ce574b Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 10 Oct 2018 18:13:23 -0400 Subject: adding new config files --- .../SM6_TITANX_1Xintes/config_fermi_islip.icnt | 73 ++++++++ .../4.x-cfgs/SM6_TITANX_1Xintes/gpgpusim.config | 184 ++++++++++++++++++++ .../SM7_TITANV_1Xicnt/config_fermi_islip.icnt | 74 ++++++++ configs/4.x-cfgs/SM7_TITANV_1Xicnt/gpgpusim.config | 191 +++++++++++++++++++++ .../config_fermi_islip.icnt | 74 ++++++++ .../SM7_TITANV_1Xicnt_1Xintes/gpgpusim.config | 191 +++++++++++++++++++++ 6 files changed, 787 insertions(+) create mode 100644 configs/4.x-cfgs/SM6_TITANX_1Xintes/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM6_TITANX_1Xintes/gpgpusim.config create mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt/gpgpusim.config create mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/gpgpusim.config (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX_1Xintes/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_TITANX_1Xintes/config_fermi_islip.icnt new file mode 100644 index 0000000..85ac0d7 --- /dev/null +++ b/configs/4.x-cfgs/SM6_TITANX_1Xintes/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; +input_buffer_size = 256; +ejection_buffer_size = 64; +boundary_buffer_size = 64; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM6_TITANX_1Xintes/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX_1Xintes/gpgpusim.config new file mode 100644 index 0000000..cc7419c --- /dev/null +++ b/configs/4.x-cfgs/SM6_TITANX_1Xintes/gpgpusim.config @@ -0,0 +1,184 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 +-gpgpu_ignore_resources_limitation 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +# P102 has two semi-indp scheds per core, and two cores per cluster +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 32 warps/SM +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 16 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 4 SFU units +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 1 + + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,8,8,8,130 +-ptx_opcode_initiation_sfu 4 +-ptx_opcode_latency_sfu 8 + + +# latencies and cache configs are adopted from: +# https://arxiv.org/pdf/1804.06826.pdf +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB +# Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres are used +-gpgpu_cache:dl1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 49152 +-gpgpu_shmem_size_PrefL1 49152 +-gpgpu_shmem_size_PrefShared 49152 +# By default, L1 cache is disabled in Pascal P102. +# requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 +-gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 82 +-smem_latency 24 +-gpgpu_flush_l1_cache 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 +-perf_sim_memcpy 0 + +# 4 KB Inst. +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_operand_collector_num_in_ports_sfu 2 +-gpgpu_operand_collector_num_out_ports_sfu 2 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +# Use Pascal Coalsce arhitetecture +-gpgpu_coalesce_arch 61 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 240 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: + CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" + +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal 102 has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 102 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt new file mode 100644 index 0000000..cea3acf --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 256; +input_buffer_size = 256; +ejection_buffer_size = 128; +boundary_buffer_size = 128; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV_1Xicnt/gpgpusim.config new file mode 100644 index 0000000..a25e329 --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV_1Xicnt/gpgpusim.config @@ -0,0 +1,191 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Volta NVIDIA GV100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Defualt config is 32KB DL1 and 96KB shared memory +# In Volta, we assign the remaining shared memory to L1 cache +# if the assigned shd mem = 0, then L1 cache = 128KB +# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x +# disable this mode in case of multi kernels/apps execution +-adpative_volta_cache_config 1 +-gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 28 +-smem_latency 19 +-gpgpu_flush_l1_cache 1 +-adpative_volta_cache_config 1 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Volta V100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt new file mode 100644 index 0000000..dba0e76 --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 256; +input_buffer_size = 256; +ejection_buffer_size = 128; +boundary_buffer_size = 128; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/gpgpusim.config new file mode 100644 index 0000000..a25e329 --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/gpgpusim.config @@ -0,0 +1,191 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Volta NVIDIA GV100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Defualt config is 32KB DL1 and 96KB shared memory +# In Volta, we assign the remaining shared memory to L1 cache +# if the assigned shd mem = 0, then L1 cache = 128KB +# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x +# disable this mode in case of multi kernels/apps execution +-adpative_volta_cache_config 1 +-gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 28 +-smem_latency 19 +-gpgpu_flush_l1_cache 1 +-adpative_volta_cache_config 1 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Volta V100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + -- cgit v1.3 From 2ff5a7fc23b35490d155807191650de2cf073e71 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 10 Oct 2018 20:19:13 -0400 Subject: fixing TITANV inct buffer error --- configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt index cea3acf..2f25889 100644 --- a/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt +++ b/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt @@ -20,8 +20,8 @@ routing_function = dest_tag; num_vcs = 1; vc_buf_size = 256; input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; +ejection_buffer_size = 256; +boundary_buffer_size = 256; wait_for_tail_credit = 0; -- cgit v1.3 From 8a728aa50ea27765e90fc346731aa5e083f29929 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 10 Oct 2018 20:23:19 -0400 Subject: fixing TITANV buffer error --- configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt | 4 ++-- configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt | 4 ++-- 2 files changed, 4 insertions(+), 4 deletions(-) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt index cea3acf..2f25889 100644 --- a/configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt +++ b/configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt @@ -20,8 +20,8 @@ routing_function = dest_tag; num_vcs = 1; vc_buf_size = 256; input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; +ejection_buffer_size = 256; +boundary_buffer_size = 256; wait_for_tail_credit = 0; diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt index dba0e76..fca9ffe 100644 --- a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt +++ b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt @@ -20,8 +20,8 @@ routing_function = dest_tag; num_vcs = 1; vc_buf_size = 256; input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; +ejection_buffer_size = 256; +boundary_buffer_size = 256; wait_for_tail_credit = 0; -- cgit v1.3 From fcaf5f1217e7349e2239f51a7bce1fad13df920c Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 10 Oct 2018 20:46:45 -0400 Subject: adding new config files --- .../SM6_TITANX_1Xintes_LB/config_fermi_islip.icnt | 73 ++++++++ .../4.x-cfgs/SM6_TITANX_1Xintes_LB/gpgpusim.config | 184 ++++++++++++++++++++ .../config_fermi_islip.icnt | 74 ++++++++ .../SM7_TITANV_1Xicnt_1Xintes_LB/gpgpusim.config | 191 +++++++++++++++++++++ 4 files changed, 522 insertions(+) create mode 100644 configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/gpgpusim.config create mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/gpgpusim.config (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/config_fermi_islip.icnt new file mode 100644 index 0000000..d4e3467 --- /dev/null +++ b/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 32; +input_buffer_size = 32; +ejection_buffer_size = 32; +boundary_buffer_size = 32; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/gpgpusim.config new file mode 100644 index 0000000..cc7419c --- /dev/null +++ b/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/gpgpusim.config @@ -0,0 +1,184 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 +-gpgpu_ignore_resources_limitation 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +# P102 has two semi-indp scheds per core, and two cores per cluster +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 32 warps/SM +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 16 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 4 SFU units +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 1 + + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,8,8,8,130 +-ptx_opcode_initiation_sfu 4 +-ptx_opcode_latency_sfu 8 + + +# latencies and cache configs are adopted from: +# https://arxiv.org/pdf/1804.06826.pdf +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB +# Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres are used +-gpgpu_cache:dl1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 49152 +-gpgpu_shmem_size_PrefL1 49152 +-gpgpu_shmem_size_PrefShared 49152 +# By default, L1 cache is disabled in Pascal P102. +# requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 +-gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 82 +-smem_latency 24 +-gpgpu_flush_l1_cache 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 +-perf_sim_memcpy 0 + +# 4 KB Inst. +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_operand_collector_num_in_ports_sfu 2 +-gpgpu_operand_collector_num_out_ports_sfu 2 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +# Use Pascal Coalsce arhitetecture +-gpgpu_coalesce_arch 61 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 240 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: + CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" + +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal 102 has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 102 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/config_fermi_islip.icnt new file mode 100644 index 0000000..c6badd3 --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/config_fermi_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 32; +input_buffer_size = 32; +ejection_buffer_size = 32; +boundary_buffer_size = 32; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/gpgpusim.config new file mode 100644 index 0000000..a25e329 --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/gpgpusim.config @@ -0,0 +1,191 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Volta NVIDIA GV100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Defualt config is 32KB DL1 and 96KB shared memory +# In Volta, we assign the remaining shared memory to L1 cache +# if the assigned shd mem = 0, then L1 cache = 128KB +# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x +# disable this mode in case of multi kernels/apps execution +-adpative_volta_cache_config 1 +-gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 28 +-smem_latency 19 +-gpgpu_flush_l1_cache 1 +-adpative_volta_cache_config 1 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Volta V100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + -- cgit v1.3 From 6f345dd40ed6c73eefec2a52858ce62f72a6fa25 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 10 Oct 2018 20:58:14 -0400 Subject: adding some new inct config files and delete oterhs --- configs/4.x-cfgs/SM6_TITANX/1Xintes.icnt | 73 ++++++++ configs/4.x-cfgs/SM6_TITANX/1Xintes_LB.icnt | 73 ++++++++ .../4.x-cfgs/SM6_TITANX_1Xintes/gpgpusim.config | 184 -------------------- .../SM6_TITANX_1Xintes_LB/config_fermi_islip.icnt | 73 -------- .../4.x-cfgs/SM6_TITANX_1Xintes_LB/gpgpusim.config | 184 -------------------- configs/4.x-cfgs/SM7_TITANV/.icnt | 74 ++++++++ configs/4.x-cfgs/SM7_TITANV/1Xintes.icnt | 74 ++++++++ configs/4.x-cfgs/SM7_TITANV/1Xintes_LB.icnt | 74 ++++++++ .../SM7_TITANV_1Xicnt/config_fermi_islip.icnt | 74 -------- configs/4.x-cfgs/SM7_TITANV_1Xicnt/gpgpusim.config | 191 --------------------- .../config_fermi_islip.icnt | 74 -------- .../SM7_TITANV_1Xicnt_1Xintes/gpgpusim.config | 191 --------------------- .../config_fermi_islip.icnt | 74 -------- .../SM7_TITANV_1Xicnt_1Xintes_LB/gpgpusim.config | 191 --------------------- 14 files changed, 368 insertions(+), 1236 deletions(-) create mode 100644 configs/4.x-cfgs/SM6_TITANX/1Xintes.icnt create mode 100644 configs/4.x-cfgs/SM6_TITANX/1Xintes_LB.icnt delete mode 100644 configs/4.x-cfgs/SM6_TITANX_1Xintes/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/gpgpusim.config create mode 100644 configs/4.x-cfgs/SM7_TITANV/.icnt create mode 100644 configs/4.x-cfgs/SM7_TITANV/1Xintes.icnt create mode 100644 configs/4.x-cfgs/SM7_TITANV/1Xintes_LB.icnt delete mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/gpgpusim.config (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX/1Xintes.icnt b/configs/4.x-cfgs/SM6_TITANX/1Xintes.icnt new file mode 100644 index 0000000..85ac0d7 --- /dev/null +++ b/configs/4.x-cfgs/SM6_TITANX/1Xintes.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; +input_buffer_size = 256; +ejection_buffer_size = 64; +boundary_buffer_size = 64; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM6_TITANX/1Xintes_LB.icnt b/configs/4.x-cfgs/SM6_TITANX/1Xintes_LB.icnt new file mode 100644 index 0000000..d4e3467 --- /dev/null +++ b/configs/4.x-cfgs/SM6_TITANX/1Xintes_LB.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 32; +input_buffer_size = 32; +ejection_buffer_size = 32; +boundary_buffer_size = 32; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM6_TITANX_1Xintes/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX_1Xintes/gpgpusim.config deleted file mode 100644 index cc7419c..0000000 --- a/configs/4.x-cfgs/SM6_TITANX_1Xintes/gpgpusim.config +++ /dev/null @@ -1,184 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 --gpgpu_ignore_resources_limitation 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration -# P102 has two semi-indp scheds per core, and two cores per cluster --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 - -# This implies a maximum of 32 warps/SM --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 16 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 4 SFU units -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 1 - - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,8,8,8,130 --ptx_opcode_initiation_sfu 4 --ptx_opcode_latency_sfu 8 - - -# latencies and cache configs are adopted from: -# https://arxiv.org/pdf/1804.06826.pdf -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB -# Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres are used --gpgpu_cache:dl1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 49152 --gpgpu_shmem_size_PrefL1 49152 --gpgpu_shmem_size_PrefShared 49152 -# By default, L1 cache is disabled in Pascal P102. -# requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 --gmem_skip_L1D 1 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 82 --smem_latency 24 --gpgpu_flush_l1_cache 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 --perf_sim_memcpy 0 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_operand_collector_num_in_ports_sfu 2 --gpgpu_operand_collector_num_out_ports_sfu 2 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 -# Use Pascal Coalsce arhitetecture --gpgpu_coalesce_arch 61 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 240 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: - CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" - --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal 102 has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 102 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/config_fermi_islip.icnt deleted file mode 100644 index d4e3467..0000000 --- a/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 32; -input_buffer_size = 32; -ejection_buffer_size = 32; -boundary_buffer_size = 32; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/gpgpusim.config deleted file mode 100644 index cc7419c..0000000 --- a/configs/4.x-cfgs/SM6_TITANX_1Xintes_LB/gpgpusim.config +++ /dev/null @@ -1,184 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 --gpgpu_ignore_resources_limitation 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration -# P102 has two semi-indp scheds per core, and two cores per cluster --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 - -# This implies a maximum of 32 warps/SM --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 16 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 4 SFU units -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 1 - - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,8,8,8,130 --ptx_opcode_initiation_sfu 4 --ptx_opcode_latency_sfu 8 - - -# latencies and cache configs are adopted from: -# https://arxiv.org/pdf/1804.06826.pdf -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB -# Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres are used --gpgpu_cache:dl1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 49152 --gpgpu_shmem_size_PrefL1 49152 --gpgpu_shmem_size_PrefShared 49152 -# By default, L1 cache is disabled in Pascal P102. -# requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 --gmem_skip_L1D 1 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 82 --smem_latency 24 --gpgpu_flush_l1_cache 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 --perf_sim_memcpy 0 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_operand_collector_num_in_ports_sfu 2 --gpgpu_operand_collector_num_out_ports_sfu 2 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 -# Use Pascal Coalsce arhitetecture --gpgpu_coalesce_arch 61 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 240 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: - CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" - --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal 102 has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 102 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM7_TITANV/.icnt b/configs/4.x-cfgs/SM7_TITANV/.icnt new file mode 100644 index 0000000..2f25889 --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV/.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 256; +input_buffer_size = 256; +ejection_buffer_size = 256; +boundary_buffer_size = 256; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM7_TITANV/1Xintes.icnt b/configs/4.x-cfgs/SM7_TITANV/1Xintes.icnt new file mode 100644 index 0000000..fca9ffe --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV/1Xintes.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 256; +input_buffer_size = 256; +ejection_buffer_size = 256; +boundary_buffer_size = 256; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM7_TITANV/1Xintes_LB.icnt b/configs/4.x-cfgs/SM7_TITANV/1Xintes_LB.icnt new file mode 100644 index 0000000..c6badd3 --- /dev/null +++ b/configs/4.x-cfgs/SM7_TITANV/1Xintes_LB.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 32; +input_buffer_size = 32; +ejection_buffer_size = 32; +boundary_buffer_size = 32; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt deleted file mode 100644 index 2f25889..0000000 --- a/configs/4.x-cfgs/SM7_TITANV_1Xicnt/config_fermi_islip.icnt +++ /dev/null @@ -1,74 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 256; -input_buffer_size = 256; -ejection_buffer_size = 256; -boundary_buffer_size = 256; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 2.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV_1Xicnt/gpgpusim.config deleted file mode 100644 index a25e329..0000000 --- a/configs/4.x-cfgs/SM7_TITANV_1Xicnt/gpgpusim.config +++ /dev/null @@ -1,191 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Volta NVIDIA GV100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 100 --ptx_opcode_initiation_sfu 8 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Defualt config is 32KB DL1 and 96KB shared memory -# In Volta, we assign the remaining shared memory to L1 cache -# if the assigned shd mem = 0, then L1 cache = 128KB -# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x -# disable this mode in case of multi kernels/apps execution --adpative_volta_cache_config 1 --gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 98304 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 28 --smem_latency 19 --gpgpu_flush_l1_cache 1 --adpative_volta_cache_config 1 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Volta V100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt deleted file mode 100644 index fca9ffe..0000000 --- a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/config_fermi_islip.icnt +++ /dev/null @@ -1,74 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 256; -input_buffer_size = 256; -ejection_buffer_size = 256; -boundary_buffer_size = 256; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/gpgpusim.config deleted file mode 100644 index a25e329..0000000 --- a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes/gpgpusim.config +++ /dev/null @@ -1,191 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Volta NVIDIA GV100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 100 --ptx_opcode_initiation_sfu 8 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Defualt config is 32KB DL1 and 96KB shared memory -# In Volta, we assign the remaining shared memory to L1 cache -# if the assigned shd mem = 0, then L1 cache = 128KB -# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x -# disable this mode in case of multi kernels/apps execution --adpative_volta_cache_config 1 --gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 98304 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 28 --smem_latency 19 --gpgpu_flush_l1_cache 1 --adpative_volta_cache_config 1 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Volta V100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/config_fermi_islip.icnt deleted file mode 100644 index c6badd3..0000000 --- a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/config_fermi_islip.icnt +++ /dev/null @@ -1,74 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 32; -input_buffer_size = 32; -ejection_buffer_size = 32; -boundary_buffer_size = 32; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/gpgpusim.config deleted file mode 100644 index a25e329..0000000 --- a/configs/4.x-cfgs/SM7_TITANV_1Xicnt_1Xintes_LB/gpgpusim.config +++ /dev/null @@ -1,191 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Volta NVIDIA GV100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 100 --ptx_opcode_initiation_sfu 8 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Defualt config is 32KB DL1 and 96KB shared memory -# In Volta, we assign the remaining shared memory to L1 cache -# if the assigned shd mem = 0, then L1 cache = 128KB -# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x -# disable this mode in case of multi kernels/apps execution --adpative_volta_cache_config 1 --gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 98304 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 28 --smem_latency 19 --gpgpu_flush_l1_cache 1 --adpative_volta_cache_config 1 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Volta V100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - -- cgit v1.3 From 3d0f3691c6ccff4a912fc0fb751291a96832cd13 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 10 Oct 2018 21:04:24 -0400 Subject: removing unnecassery files --- .../SM6_TITANX_1Xintes/config_fermi_islip.icnt | 73 ---------------------- 1 file changed, 73 deletions(-) delete mode 100644 configs/4.x-cfgs/SM6_TITANX_1Xintes/config_fermi_islip.icnt (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX_1Xintes/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_TITANX_1Xintes/config_fermi_islip.icnt deleted file mode 100644 index 85ac0d7..0000000 --- a/configs/4.x-cfgs/SM6_TITANX_1Xintes/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 64; -input_buffer_size = 256; -ejection_buffer_size = 64; -boundary_buffer_size = 64; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - -- cgit v1.3 From a6b9171a158d29d5d1ad415d087f483feb1af965 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 11 Oct 2018 12:58:35 -0400 Subject: enable memcpu-perf by deafault --- configs/4.x-cfgs/SM2_GTX480/gpgpusim.config | 2 +- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 2 +- configs/4.x-cfgs/SM7_TITANV/.icnt | 74 ----------------------------- configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 3 +- 4 files changed, 3 insertions(+), 78 deletions(-) delete mode 100644 configs/4.x-cfgs/SM7_TITANV/.icnt (limited to 'configs') diff --git a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config index b965aff..05663c2 100644 --- a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config @@ -67,7 +67,7 @@ -gpgpu_cache:dl2 S:64:128:8,L:B:m:L:L,A:256:4,4:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 +-perf_sim_memcpy 1 -memory_partition_indexing 0 -gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,S:2:32,4 diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index cc7419c..1b4e6e3 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -84,7 +84,7 @@ -gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 --perf_sim_memcpy 0 +-perf_sim_memcpy 1 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 diff --git a/configs/4.x-cfgs/SM7_TITANV/.icnt b/configs/4.x-cfgs/SM7_TITANV/.icnt deleted file mode 100644 index 2f25889..0000000 --- a/configs/4.x-cfgs/SM7_TITANV/.icnt +++ /dev/null @@ -1,74 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 256; -input_buffer_size = 256; -ejection_buffer_size = 256; -boundary_buffer_size = 256; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 2.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config index 41757f8..6fe441b 100644 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -79,13 +79,12 @@ -l1_latency 28 -smem_latency 19 -gpgpu_flush_l1_cache 1 --adpative_volta_cache_config 1 # 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache -gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 +-perf_sim_memcpy 1 # 128 KB Inst. -gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -- cgit v1.3 From 0f435a57674d5839a413bec5292eca66a6389dff Mon Sep 17 00:00:00 2001 From: tgrogers Date: Thu, 11 Oct 2018 23:28:08 -0400 Subject: must specify occupancy number --- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 1 + configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 1 + 2 files changed, 2 insertions(+) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index 1b4e6e3..3842508 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -27,6 +27,7 @@ # shader core pipeline config -gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 62 # This implies a maximum of 32 warps/SM -gpgpu_shader_core_pipeline 1024:32 diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config index 6fe441b..c8351da 100644 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -33,6 +33,7 @@ # shader core pipeline config -gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 70 # This implies a maximum of 64 warps/SM -gpgpu_shader_core_pipeline 2048:32 -- cgit v1.3 From 74c939d205c0eff341f80f5867183dfa3e70c114 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Mon, 15 Oct 2018 20:16:06 -0400 Subject: Specify the register sm occpancy number for 480 --- configs/4.x-cfgs/SM2_GTX480/gpgpusim.config | 1 + 1 file changed, 1 insertion(+) (limited to 'configs') diff --git a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config index 05663c2..4096b09 100644 --- a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config @@ -22,6 +22,7 @@ # shader core pipeline config -gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 # This implies a maximum of 48 warps/SM -gpgpu_shader_core_pipeline 1536:32 -- cgit v1.3 From b37069409322cac50161e96f198f689bb6de88a8 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 16 Oct 2018 16:57:25 -0400 Subject: delete unnecssary config files --- configs/4.x-cfgs/SM6_TITANX/1Xintes.icnt | 73 ---------------------------- configs/4.x-cfgs/SM6_TITANX/1Xintes_LB.icnt | 73 ---------------------------- configs/4.x-cfgs/SM7_TITANV/1Xintes.icnt | 74 ----------------------------- configs/4.x-cfgs/SM7_TITANV/1Xintes_LB.icnt | 74 ----------------------------- 4 files changed, 294 deletions(-) delete mode 100644 configs/4.x-cfgs/SM6_TITANX/1Xintes.icnt delete mode 100644 configs/4.x-cfgs/SM6_TITANX/1Xintes_LB.icnt delete mode 100644 configs/4.x-cfgs/SM7_TITANV/1Xintes.icnt delete mode 100644 configs/4.x-cfgs/SM7_TITANV/1Xintes_LB.icnt (limited to 'configs') diff --git a/configs/4.x-cfgs/SM6_TITANX/1Xintes.icnt b/configs/4.x-cfgs/SM6_TITANX/1Xintes.icnt deleted file mode 100644 index 85ac0d7..0000000 --- a/configs/4.x-cfgs/SM6_TITANX/1Xintes.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 64; -input_buffer_size = 256; -ejection_buffer_size = 64; -boundary_buffer_size = 64; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM6_TITANX/1Xintes_LB.icnt b/configs/4.x-cfgs/SM6_TITANX/1Xintes_LB.icnt deleted file mode 100644 index d4e3467..0000000 --- a/configs/4.x-cfgs/SM6_TITANX/1Xintes_LB.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 32; -input_buffer_size = 32; -ejection_buffer_size = 32; -boundary_buffer_size = 32; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM7_TITANV/1Xintes.icnt b/configs/4.x-cfgs/SM7_TITANV/1Xintes.icnt deleted file mode 100644 index fca9ffe..0000000 --- a/configs/4.x-cfgs/SM7_TITANV/1Xintes.icnt +++ /dev/null @@ -1,74 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 256; -input_buffer_size = 256; -ejection_buffer_size = 256; -boundary_buffer_size = 256; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM7_TITANV/1Xintes_LB.icnt b/configs/4.x-cfgs/SM7_TITANV/1Xintes_LB.icnt deleted file mode 100644 index c6badd3..0000000 --- a/configs/4.x-cfgs/SM7_TITANV/1Xintes_LB.icnt +++ /dev/null @@ -1,74 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 32; -input_buffer_size = 32; -ejection_buffer_size = 32; -boundary_buffer_size = 32; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - -- cgit v1.3 From e41955fb2b2041c67f5a4a5627d4870c4cda13aa Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 16 Oct 2018 17:28:41 -0400 Subject: remove pascal titianx --- configs/Pascal_TITANX/config_fermi_islip.icnt | 70 ------------ configs/Pascal_TITANX/gpgpusim.config | 156 -------------------------- 2 files changed, 226 deletions(-) delete mode 100644 configs/Pascal_TITANX/config_fermi_islip.icnt delete mode 100644 configs/Pascal_TITANX/gpgpusim.config (limited to 'configs') diff --git a/configs/Pascal_TITANX/config_fermi_islip.icnt b/configs/Pascal_TITANX/config_fermi_islip.icnt deleted file mode 100644 index 602daee..0000000 --- a/configs/Pascal_TITANX/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/Pascal_TITANX/gpgpusim.config b/configs/Pascal_TITANX/gpgpusim.config deleted file mode 100644 index f78bd02..0000000 --- a/configs/Pascal_TITANX/gpgpusim.config +++ /dev/null @@ -1,156 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,8,4,4,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_shmem_size 98304 --gpgpu_shmem_size_PrefL1 98304 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 -# gpgpu_num_reg_banks should be increased to 32 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - -- cgit v1.3 From ab0d5d829d3f825736c5f91a365e755b2c2752c9 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Tue, 23 Oct 2018 11:14:22 -0400 Subject: Moving configs to folders with better names --- configs/3.x-cfgs/GTX480/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/GTX480/gpgpusim.config | 134 ----- configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml | 538 -------------------- .../GeForceGTX750Ti/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config | 131 ----- .../GeForceGTX750Ti/gpuwattch_gtx750Ti.xml | 539 --------------------- configs/3.x-cfgs/QuadroFX5600/gpgpusim.config | 99 ---- .../QuadroFX5600/gpuwattch_quadrofx5600.xml | 538 -------------------- .../3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt | 70 --- .../3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt | 69 --- configs/3.x-cfgs/QuadroFX5800/gpgpusim.config | 88 ---- .../3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config | 150 ------ .../3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml | 538 -------------------- configs/3.x-cfgs/SM6_P100/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/SM6_P100/gpgpusim.config | 156 ------ configs/3.x-cfgs/SM6_P100/gpuwattch_gtx480.xml | 538 -------------------- .../3.x-cfgs/SM6_TITANX/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/SM6_TITANX/gpgpusim.config | 157 ------ configs/3.x-cfgs/SM6_TITANX/gpuwattch_gtx480.xml | 538 -------------------- .../3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/SM7_TITANV/gpgpusim.config | 165 ------- .../3.x-cfgs/TeslaC2050/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/TeslaC2050/gpgpusim.config | 133 ----- .../4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/SM2_GTX480/gpgpusim.config | 153 ------ configs/4.x-cfgs/SM2_GTX480/gpuwattch_gtx480.xml | 538 -------------------- .../4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt | 73 --- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 185 ------- .../4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 74 --- configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 191 -------- .../deprecated-cfgs/GTX480/config_fermi_islip.icnt | 70 +++ configs/deprecated-cfgs/GTX480/gpgpusim.config | 134 +++++ .../deprecated-cfgs/GTX480/gpuwattch_gtx480.xml | 538 ++++++++++++++++++++ .../GeForceGTX750Ti/config_fermi_islip.icnt | 70 +++ .../GeForceGTX750Ti/gpgpusim.config | 131 +++++ .../GeForceGTX750Ti/gpuwattch_gtx750Ti.xml | 539 +++++++++++++++++++++ .../deprecated-cfgs/QuadroFX5600/gpgpusim.config | 99 ++++ .../QuadroFX5600/gpuwattch_quadrofx5600.xml | 538 ++++++++++++++++++++ .../QuadroFX5600/icnt_config_islip.icnt | 70 +++ .../QuadroFX5800/config_quadro_islip.icnt | 69 +++ .../deprecated-cfgs/QuadroFX5800/gpgpusim.config | 88 ++++ .../SM6_GTX1080/config_fermi_islip.icnt | 70 +++ .../deprecated-cfgs/SM6_GTX1080/gpgpusim.config | 150 ++++++ .../SM6_GTX1080/gpuwattch_gtx1080Ti.xml | 538 ++++++++++++++++++++ .../SM6_P100/config_fermi_islip.icnt | 70 +++ configs/deprecated-cfgs/SM6_P100/gpgpusim.config | 156 ++++++ .../deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml | 538 ++++++++++++++++++++ .../SM6_TITANX/config_fermi_islip.icnt | 70 +++ configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config | 157 ++++++ .../SM6_TITANX/gpuwattch_gtx480.xml | 538 ++++++++++++++++++++ .../SM7_TITANV/config_fermi_islip.icnt | 70 +++ configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config | 165 +++++++ .../TeslaC2050/config_fermi_islip.icnt | 70 +++ configs/deprecated-cfgs/TeslaC2050/gpgpusim.config | 133 +++++ .../tested-cfgs/SM2_GTX480/config_fermi_islip.icnt | 70 +++ configs/tested-cfgs/SM2_GTX480/gpgpusim.config | 153 ++++++ .../tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml | 538 ++++++++++++++++++++ .../tested-cfgs/SM6_TITANX/config_fermi_islip.icnt | 73 +++ configs/tested-cfgs/SM6_TITANX/gpgpusim.config | 185 +++++++ .../tested-cfgs/SM7_TITANV/config_fermi_islip.icnt | 74 +++ configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 191 ++++++++ 62 files changed, 6355 insertions(+), 6355 deletions(-) delete mode 100644 configs/3.x-cfgs/GTX480/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/GTX480/gpgpusim.config delete mode 100755 configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml delete mode 100644 configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config delete mode 100755 configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml delete mode 100644 configs/3.x-cfgs/QuadroFX5600/gpgpusim.config delete mode 100644 configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml delete mode 100644 configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt delete mode 100644 configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt delete mode 100644 configs/3.x-cfgs/QuadroFX5800/gpgpusim.config delete mode 100644 configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config delete mode 100755 configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml delete mode 100644 configs/3.x-cfgs/SM6_P100/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/SM6_P100/gpgpusim.config delete mode 100755 configs/3.x-cfgs/SM6_P100/gpuwattch_gtx480.xml delete mode 100644 configs/3.x-cfgs/SM6_TITANX/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/SM6_TITANX/gpgpusim.config delete mode 100755 configs/3.x-cfgs/SM6_TITANX/gpuwattch_gtx480.xml delete mode 100644 configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/SM7_TITANV/gpgpusim.config delete mode 100644 configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/TeslaC2050/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM2_GTX480/gpgpusim.config delete mode 100755 configs/4.x-cfgs/SM2_GTX480/gpuwattch_gtx480.xml delete mode 100644 configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM6_TITANX/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM7_TITANV/gpgpusim.config create mode 100644 configs/deprecated-cfgs/GTX480/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/GTX480/gpgpusim.config create mode 100755 configs/deprecated-cfgs/GTX480/gpuwattch_gtx480.xml create mode 100644 configs/deprecated-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/GeForceGTX750Ti/gpgpusim.config create mode 100755 configs/deprecated-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml create mode 100644 configs/deprecated-cfgs/QuadroFX5600/gpgpusim.config create mode 100644 configs/deprecated-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml create mode 100644 configs/deprecated-cfgs/QuadroFX5600/icnt_config_islip.icnt create mode 100644 configs/deprecated-cfgs/QuadroFX5800/config_quadro_islip.icnt create mode 100644 configs/deprecated-cfgs/QuadroFX5800/gpgpusim.config create mode 100644 configs/deprecated-cfgs/SM6_GTX1080/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/SM6_GTX1080/gpgpusim.config create mode 100755 configs/deprecated-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml create mode 100644 configs/deprecated-cfgs/SM6_P100/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/SM6_P100/gpgpusim.config create mode 100755 configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml create mode 100644 configs/deprecated-cfgs/SM6_TITANX/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config create mode 100755 configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml create mode 100644 configs/deprecated-cfgs/SM7_TITANV/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config create mode 100644 configs/deprecated-cfgs/TeslaC2050/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/TeslaC2050/gpgpusim.config create mode 100644 configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt create mode 100644 configs/tested-cfgs/SM2_GTX480/gpgpusim.config create mode 100755 configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml create mode 100644 configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt create mode 100644 configs/tested-cfgs/SM6_TITANX/gpgpusim.config create mode 100644 configs/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt create mode 100644 configs/tested-cfgs/SM7_TITANV/gpgpusim.config (limited to 'configs') diff --git a/configs/3.x-cfgs/GTX480/config_fermi_islip.icnt b/configs/3.x-cfgs/GTX480/config_fermi_islip.icnt deleted file mode 100644 index 7820e4e..0000000 --- a/configs/3.x-cfgs/GTX480/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 27; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/GTX480/gpgpusim.config b/configs/3.x-cfgs/GTX480/gpgpusim.config deleted file mode 100644 index ee90c12..0000000 --- a/configs/3.x-cfgs/GTX480/gpgpusim.config +++ /dev/null @@ -1,134 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 15 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 700.0:700.0:700.0:924.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 --gpgpu_occupancy_sm_number 20 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 116 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml b/configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt b/configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt deleted file mode 100644 index 069ca02..0000000 --- a/configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 7; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config b/configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config deleted file mode 100644 index c675aab..0000000 --- a/configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config +++ /dev/null @@ -1,131 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 52 - - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 5 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 2 --gpgpu_n_sub_partition_per_mchannel 1 - -# Maxwell clock domains -#-gpgpu_clock_domains ::: -# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. --gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 --gpgpu_occupancy_sm_number 52 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 --gpgpu_num_sp_units 8 --gpgpu_num_sfu_units 32 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 6,12,13,13,210 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 6,12,6,6,374 --ptx_opcode_initiation_fp 1,1,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - --gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 --gmem_skip_L1D 1 --gpgpu_shmem_size 65536 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache --gpgpu_cache:dl2 1024:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 16:128:32,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 150 --dram_latency 130 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 300 - -# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 32 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Maxwell has four schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs -# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. -# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx750Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml b/configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml deleted file mode 100755 index e2b2324..0000000 --- a/configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml +++ /dev/null @@ -1,539 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/3.x-cfgs/QuadroFX5600/gpgpusim.config b/configs/3.x-cfgs/QuadroFX5600/gpgpusim.config deleted file mode 100644 index 6f836ee..0000000 --- a/configs/3.x-cfgs/QuadroFX5600/gpgpusim.config +++ /dev/null @@ -1,99 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 12 - -# high level architecture configuration --gpgpu_n_clusters 8 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 6 --gpgpu_clock_domains 337.5:600.0:600.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 --gpgpu_occupancy_sm_number 12 - --gpgpu_occupancy_sm_number 12 -#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) --gpgpu_shader_core_pipeline 768:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,1,1,1,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour --gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - -# TLB parameters -#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 -#-gpgpu_tlbl2_latency 45 - - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file icnt_config_islip.icnt - -# dram scheduler config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 55 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz -# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - - - -# Using cuobjdump to extract ptx/SASS --gpgpu_ptx_use_cuobjdump 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 --power_trace_enabled 0 --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_quadrofx5600.xml - --steady_power_levels_enabled 1 --steady_state_definition 8,4 diff --git a/configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml b/configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml deleted file mode 100644 index 2c5a6fc..0000000 --- a/configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt b/configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt deleted file mode 100644 index de3bcc8..0000000 --- a/configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//14*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 14; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt b/configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt deleted file mode 100644 index cfe9cac..0000000 --- a/configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt +++ /dev/null @@ -1,69 +0,0 @@ -//18*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 18; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/QuadroFX5800/gpgpusim.config b/configs/3.x-cfgs/QuadroFX5800/gpgpusim.config deleted file mode 100644 index fef1110..0000000 --- a/configs/3.x-cfgs/QuadroFX5800/gpgpusim.config +++ /dev/null @@ -1,88 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 13 - -# high level architecture configuration --gpgpu_n_clusters 10 --gpgpu_n_cores_per_cluster 3 --gpgpu_n_mem 8 --gpgpu_clock_domains 325.0:650.0:650.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 --gpgpu_occupancy_sm_number 13 - --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,1,1,1,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo --gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file config_quadro_islip.icnt - -# dram scheduler config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 325MHz = 74 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 58 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --dram_data_command_freq_ratio 2 # GDDR3 is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt b/configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt deleted file mode 100644 index 2a69ddd..0000000 --- a/configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 50; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config b/configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config deleted file mode 100644 index fb044c6..0000000 --- a/configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config +++ /dev/null @@ -1,150 +0,0 @@ -# This config models the Pascal GP102 (GeForceGTX 1080Ti) - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 11 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 --gpgpu_occupancy_sm_number 60 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 1,2,1,1,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache -# The default is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 --gpgpu_shmem_size 98304 --gmem_skip_L1D 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 --gpgpu_cache:dl2_texture_only 0 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 -# gpgpu_num_reg_banks should be increased to 32, but it gives an error! --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) -# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Pascal has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx1080Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml b/configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml deleted file mode 100755 index 02619ff..0000000 --- a/configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/3.x-cfgs/SM6_P100/config_fermi_islip.icnt b/configs/3.x-cfgs/SM6_P100/config_fermi_islip.icnt deleted file mode 100644 index d26c8d9..0000000 --- a/configs/3.x-cfgs/SM6_P100/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 60; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/SM6_P100/gpgpusim.config b/configs/3.x-cfgs/SM6_P100/gpgpusim.config deleted file mode 100644 index a5e6736..0000000 --- a/configs/3.x-cfgs/SM6_P100/gpgpusim.config +++ /dev/null @@ -1,156 +0,0 @@ -# This config models the Pascal GP100 -# For more info about this card, see Nvidia White paper -# https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 32 --gpgpu_n_sub_partition_per_mchannel 1 - -# Pscal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Nvidia_Tesla --gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,2,1,2,2,1,4 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 2,2,2,2,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory -# Pascal GP100 has 48KB L1 cache --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 65536 --gpgpu_shmem_size_PrefShared 65536 --gmem_skip_L1D 0 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 14 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_units_mem 10 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/3.x-cfgs/SM6_P100/gpuwattch_gtx480.xml b/configs/3.x-cfgs/SM6_P100/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/3.x-cfgs/SM6_P100/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/3.x-cfgs/SM6_TITANX/config_fermi_islip.icnt b/configs/3.x-cfgs/SM6_TITANX/config_fermi_islip.icnt deleted file mode 100644 index 602daee..0000000 --- a/configs/3.x-cfgs/SM6_TITANX/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/3.x-cfgs/SM6_TITANX/gpgpusim.config deleted file mode 100644 index 28912a3..0000000 --- a/configs/3.x-cfgs/SM6_TITANX/gpgpusim.config +++ /dev/null @@ -1,157 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 --gpgpu_occupancy_sm_number 61 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,8,4,4,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_shmem_size 98304 --gpgpu_shmem_size_PrefL1 98304 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 -# gpgpu_num_reg_banks should be increased to 32 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/3.x-cfgs/SM6_TITANX/gpuwattch_gtx480.xml b/configs/3.x-cfgs/SM6_TITANX/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/3.x-cfgs/SM6_TITANX/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt deleted file mode 100644 index fac792a..0000000 --- a/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config deleted file mode 100644 index aefb04a..0000000 --- a/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config +++ /dev/null @@ -1,165 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,1,4,4,1,9 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Volta GV100 has 64KB Shared memory --gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 --gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 --gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 1 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 0 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 - -# 128 KB Inst. --gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 14 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_units_mem 10 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt b/configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt deleted file mode 100644 index a11bd8e..0000000 --- a/configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//20*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 26; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; diff --git a/configs/3.x-cfgs/TeslaC2050/gpgpusim.config b/configs/3.x-cfgs/TeslaC2050/gpgpusim.config deleted file mode 100644 index 6ac2c12..0000000 --- a/configs/3.x-cfgs/TeslaC2050/gpgpusim.config +++ /dev/null @@ -1,133 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - -# Using cuobjdump to extract ptx/SASS -#-gpgpu_ptx_use_cuobjdump 1 # use default - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - - -# high level architecture configuration --gpgpu_n_clusters 14 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 575.0:575.0:575.0:750.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 --gpgpu_occupancy_sm_number 20 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 750MHz / 575MHz = 130 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 114 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt b/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt deleted file mode 100644 index c399db9..0000000 --- a/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 27; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 64; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 2.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config deleted file mode 100644 index 4096b09..0000000 --- a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config +++ /dev/null @@ -1,153 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 15 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 700.0:700.0:700.0:924.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 --gpgpu_occupancy_sm_number 20 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -#For Fermi, DP unit =0, DP inst is executed on SFU --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,S:64:8,8 --gpgpu_shmem_size 49152 --icnt_flit_size 40 --gmem_skip_L1D 0 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 35 --smem_latency 26 --gpgpu_flush_l1_cache 1 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:64:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 S:64:128:8,L:B:m:L:L,A:256:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 1 --memory_partition_indexing 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,S:2:32,4 --gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,T:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,S:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 20 - --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM2_GTX480/gpuwattch_gtx480.xml b/configs/4.x-cfgs/SM2_GTX480/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/4.x-cfgs/SM2_GTX480/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt deleted file mode 100644 index dec4789..0000000 --- a/configs/4.x-cfgs/SM6_TITANX/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 64; -input_buffer_size = 256; -ejection_buffer_size = 64; -boundary_buffer_size = 64; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 2.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config deleted file mode 100644 index 3842508..0000000 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ /dev/null @@ -1,185 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 --gpgpu_ignore_resources_limitation 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration -# P102 has two semi-indp scheds per core, and two cores per cluster --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 --gpgpu_occupancy_sm_number 62 - -# This implies a maximum of 32 warps/SM --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 16 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 4 SFU units -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 1 - - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,8,8,8,130 --ptx_opcode_initiation_sfu 4 --ptx_opcode_latency_sfu 8 - - -# latencies and cache configs are adopted from: -# https://arxiv.org/pdf/1804.06826.pdf -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB -# Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres are used --gpgpu_cache:dl1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 49152 --gpgpu_shmem_size_PrefL1 49152 --gpgpu_shmem_size_PrefShared 49152 -# By default, L1 cache is disabled in Pascal P102. -# requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 --gmem_skip_L1D 1 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 82 --smem_latency 24 --gpgpu_flush_l1_cache 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 --perf_sim_memcpy 1 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_operand_collector_num_in_ports_sfu 2 --gpgpu_operand_collector_num_out_ports_sfu 2 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 -# Use Pascal Coalsce arhitetecture --gpgpu_coalesce_arch 61 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 240 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: - CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" - --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal 102 has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 102 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt deleted file mode 100644 index 2f25889..0000000 --- a/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt +++ /dev/null @@ -1,74 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 40; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 256; -input_buffer_size = 256; -ejection_buffer_size = 256; -boundary_buffer_size = 256; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 2.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config deleted file mode 100644 index c8351da..0000000 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ /dev/null @@ -1,191 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Volta NVIDIA GV100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:2000.0:1200.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 --gpgpu_occupancy_sm_number 70 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 100 --ptx_opcode_initiation_sfu 8 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Defualt config is 32KB DL1 and 96KB shared memory -# In Volta, we assign the remaining shared memory to L1 cache -# if the assigned shd mem = 0, then L1 cache = 128KB -# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x -# disable this mode in case of multi kernels/apps execution --adpative_volta_cache_config 1 --gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 98304 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 28 --smem_latency 19 --gpgpu_flush_l1_cache 1 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 1 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Volta V100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/deprecated-cfgs/GTX480/config_fermi_islip.icnt b/configs/deprecated-cfgs/GTX480/config_fermi_islip.icnt new file mode 100644 index 0000000..7820e4e --- /dev/null +++ b/configs/deprecated-cfgs/GTX480/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 27; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/GTX480/gpgpusim.config b/configs/deprecated-cfgs/GTX480/gpgpusim.config new file mode 100644 index 0000000..ee90c12 --- /dev/null +++ b/configs/deprecated-cfgs/GTX480/gpgpusim.config @@ -0,0 +1,134 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 15 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 700.0:700.0:700.0:924.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 116 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/deprecated-cfgs/GTX480/gpuwattch_gtx480.xml b/configs/deprecated-cfgs/GTX480/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/deprecated-cfgs/GTX480/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt b/configs/deprecated-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt new file mode 100644 index 0000000..069ca02 --- /dev/null +++ b/configs/deprecated-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 7; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/GeForceGTX750Ti/gpgpusim.config b/configs/deprecated-cfgs/GeForceGTX750Ti/gpgpusim.config new file mode 100644 index 0000000..c675aab --- /dev/null +++ b/configs/deprecated-cfgs/GeForceGTX750Ti/gpgpusim.config @@ -0,0 +1,131 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 52 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 5 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 2 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Maxwell clock domains +#-gpgpu_clock_domains ::: +# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. +-gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 52 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 8 +-gpgpu_num_sfu_units 32 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 6,12,13,13,210 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 6,12,6,6,374 +-ptx_opcode_initiation_fp 1,1,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gmem_skip_L1D 1 +-gpgpu_shmem_size 65536 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache +-gpgpu_cache:dl2 1024:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 16:128:32,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 150 +-dram_latency 130 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 300 + +# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 32 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Maxwell has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. +# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx750Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/deprecated-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml b/configs/deprecated-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml new file mode 100755 index 0000000..e2b2324 --- /dev/null +++ b/configs/deprecated-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml @@ -0,0 +1,539 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/QuadroFX5600/gpgpusim.config b/configs/deprecated-cfgs/QuadroFX5600/gpgpusim.config new file mode 100644 index 0000000..6f836ee --- /dev/null +++ b/configs/deprecated-cfgs/QuadroFX5600/gpgpusim.config @@ -0,0 +1,99 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 12 + +# high level architecture configuration +-gpgpu_n_clusters 8 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 6 +-gpgpu_clock_domains 337.5:600.0:600.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +-gpgpu_occupancy_sm_number 12 + +-gpgpu_occupancy_sm_number 12 +#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) +-gpgpu_shader_core_pipeline 768:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,1,1,1,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +-gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +# TLB parameters +#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 +#-gpgpu_tlbl2_latency 45 + + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file icnt_config_islip.icnt + +# dram scheduler config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 55 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + + + +# Using cuobjdump to extract ptx/SASS +-gpgpu_ptx_use_cuobjdump 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 +-power_trace_enabled 0 +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_quadrofx5600.xml + +-steady_power_levels_enabled 1 +-steady_state_definition 8,4 diff --git a/configs/deprecated-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml b/configs/deprecated-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml new file mode 100644 index 0000000..2c5a6fc --- /dev/null +++ b/configs/deprecated-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/QuadroFX5600/icnt_config_islip.icnt b/configs/deprecated-cfgs/QuadroFX5600/icnt_config_islip.icnt new file mode 100644 index 0000000..de3bcc8 --- /dev/null +++ b/configs/deprecated-cfgs/QuadroFX5600/icnt_config_islip.icnt @@ -0,0 +1,70 @@ +//14*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 14; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/QuadroFX5800/config_quadro_islip.icnt b/configs/deprecated-cfgs/QuadroFX5800/config_quadro_islip.icnt new file mode 100644 index 0000000..cfe9cac --- /dev/null +++ b/configs/deprecated-cfgs/QuadroFX5800/config_quadro_islip.icnt @@ -0,0 +1,69 @@ +//18*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 18; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/QuadroFX5800/gpgpusim.config b/configs/deprecated-cfgs/QuadroFX5800/gpgpusim.config new file mode 100644 index 0000000..fef1110 --- /dev/null +++ b/configs/deprecated-cfgs/QuadroFX5800/gpgpusim.config @@ -0,0 +1,88 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 13 + +# high level architecture configuration +-gpgpu_n_clusters 10 +-gpgpu_n_cores_per_cluster 3 +-gpgpu_n_mem 8 +-gpgpu_clock_domains 325.0:650.0:650.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +-gpgpu_occupancy_sm_number 13 + +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,1,1,1,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +-gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file config_quadro_islip.icnt + +# dram scheduler config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 325MHz = 74 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 58 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-dram_data_command_freq_ratio 2 # GDDR3 is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/deprecated-cfgs/SM6_GTX1080/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM6_GTX1080/config_fermi_islip.icnt new file mode 100644 index 0000000..2a69ddd --- /dev/null +++ b/configs/deprecated-cfgs/SM6_GTX1080/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 50; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/SM6_GTX1080/gpgpusim.config b/configs/deprecated-cfgs/SM6_GTX1080/gpgpusim.config new file mode 100644 index 0000000..fb044c6 --- /dev/null +++ b/configs/deprecated-cfgs/SM6_GTX1080/gpgpusim.config @@ -0,0 +1,150 @@ +# This config models the Pascal GP102 (GeForceGTX 1080Ti) + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 60 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 11 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 60 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 1,2,1,1,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The default is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 +-gpgpu_cache:dl2_texture_only 0 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +# gpgpu_num_reg_banks should be increased to 32, but it gives an error! +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) +# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Pascal has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx1080Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/deprecated-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml b/configs/deprecated-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml new file mode 100755 index 0000000..02619ff --- /dev/null +++ b/configs/deprecated-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/SM6_P100/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM6_P100/config_fermi_islip.icnt new file mode 100644 index 0000000..d26c8d9 --- /dev/null +++ b/configs/deprecated-cfgs/SM6_P100/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 60; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/SM6_P100/gpgpusim.config b/configs/deprecated-cfgs/SM6_P100/gpgpusim.config new file mode 100644 index 0000000..a5e6736 --- /dev/null +++ b/configs/deprecated-cfgs/SM6_P100/gpgpusim.config @@ -0,0 +1,156 @@ +# This config models the Pascal GP100 +# For more info about this card, see Nvidia White paper +# https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 60 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 32 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Pscal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Nvidia_Tesla +-gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,2,1,2,2,1,4 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 2,2,2,2,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP100 has 64KB Shared memory +# Pascal GP100 has 48KB L1 cache +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 65536 +-gpgpu_shmem_size_PrefShared 65536 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml b/configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/SM6_TITANX/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM6_TITANX/config_fermi_islip.icnt new file mode 100644 index 0000000..602daee --- /dev/null +++ b/configs/deprecated-cfgs/SM6_TITANX/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config b/configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config new file mode 100644 index 0000000..28912a3 --- /dev/null +++ b/configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config @@ -0,0 +1,157 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 61 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,8,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 98304 +-gpgpu_shmem_size_PrefL1 98304 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +# gpgpu_num_reg_banks should be increased to 32 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml b/configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM7_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..fac792a --- /dev/null +++ b/configs/deprecated-cfgs/SM7_TITANV/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config b/configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config new file mode 100644 index 0000000..aefb04a --- /dev/null +++ b/configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config @@ -0,0 +1,165 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,1,4,4,1,9 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Volta GV100 has 64KB Shared memory +-gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 1 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 + +# 128 KB Inst. +-gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/deprecated-cfgs/TeslaC2050/config_fermi_islip.icnt b/configs/deprecated-cfgs/TeslaC2050/config_fermi_islip.icnt new file mode 100644 index 0000000..a11bd8e --- /dev/null +++ b/configs/deprecated-cfgs/TeslaC2050/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//20*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 26; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; diff --git a/configs/deprecated-cfgs/TeslaC2050/gpgpusim.config b/configs/deprecated-cfgs/TeslaC2050/gpgpusim.config new file mode 100644 index 0000000..6ac2c12 --- /dev/null +++ b/configs/deprecated-cfgs/TeslaC2050/gpgpusim.config @@ -0,0 +1,133 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + +# Using cuobjdump to extract ptx/SASS +#-gpgpu_ptx_use_cuobjdump 1 # use default + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + + +# high level architecture configuration +-gpgpu_n_clusters 14 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 575.0:575.0:575.0:750.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 750MHz / 575MHz = 130 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 114 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt b/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt new file mode 100644 index 0000000..c399db9 --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 27; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config new file mode 100644 index 0000000..4096b09 --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config @@ -0,0 +1,153 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 15 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 700.0:700.0:700.0:924.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +#For Fermi, DP unit =0, DP inst is executed on SFU +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,S:64:8,8 +-gpgpu_shmem_size 49152 +-icnt_flit_size 40 +-gmem_skip_L1D 0 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 35 +-smem_latency 26 +-gpgpu_flush_l1_cache 1 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:64:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 S:64:128:8,L:B:m:L:L,A:256:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 1 +-memory_partition_indexing 0 + +-gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,S:2:32,4 +-gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,T:128:4,128:2 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,S:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 20 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml b/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt b/configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt new file mode 100644 index 0000000..dec4789 --- /dev/null +++ b/configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; +input_buffer_size = 256; +ejection_buffer_size = 64; +boundary_buffer_size = 64; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config new file mode 100644 index 0000000..3842508 --- /dev/null +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -0,0 +1,185 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 +-gpgpu_ignore_resources_limitation 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +# P102 has two semi-indp scheds per core, and two cores per cluster +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 62 + +# This implies a maximum of 32 warps/SM +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 16 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 4 SFU units +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 1 + + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,8,8,8,130 +-ptx_opcode_initiation_sfu 4 +-ptx_opcode_latency_sfu 8 + + +# latencies and cache configs are adopted from: +# https://arxiv.org/pdf/1804.06826.pdf +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB +# Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres are used +-gpgpu_cache:dl1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 49152 +-gpgpu_shmem_size_PrefL1 49152 +-gpgpu_shmem_size_PrefShared 49152 +# By default, L1 cache is disabled in Pascal P102. +# requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 +-gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 82 +-smem_latency 24 +-gpgpu_flush_l1_cache 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 +-perf_sim_memcpy 1 + +# 4 KB Inst. +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_operand_collector_num_in_ports_sfu 2 +-gpgpu_operand_collector_num_out_ports_sfu 2 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +# Use Pascal Coalsce arhitetecture +-gpgpu_coalesce_arch 61 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 240 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: + CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" + +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal 102 has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 102 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..2f25889 --- /dev/null +++ b/configs/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 256; +input_buffer_size = 256; +ejection_buffer_size = 256; +boundary_buffer_size = 256; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config new file mode 100644 index 0000000..c8351da --- /dev/null +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -0,0 +1,191 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Volta NVIDIA GV100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:2000.0:1200.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 70 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Defualt config is 32KB DL1 and 96KB shared memory +# In Volta, we assign the remaining shared memory to L1 cache +# if the assigned shd mem = 0, then L1 cache = 128KB +# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x +# disable this mode in case of multi kernels/apps execution +-adpative_volta_cache_config 1 +-gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 28 +-smem_latency 19 +-gpgpu_flush_l1_cache 1 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 1 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Volta V100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# HBM has dual bus interface, in which it can issue two col and row commands at a time +-dual_bus_interface 1 +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 100 +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + -- cgit v1.3