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 --- src/gpgpu-sim/gpu-sim.cc | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') 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 ); } } -- 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 'src/gpgpu-sim/gpu-sim.cc') 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 'src/gpgpu-sim/gpu-sim.cc') 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