From f9e65c53a29b89c167e713afa6d5eb92709047af Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Thu, 2 Apr 2020 21:23:07 -0400 Subject: updating config files --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 187 +++++++++++++++++++++++ 1 file changed, 187 insertions(+) create mode 100644 configs/tested-cfgs/SM75_RTX2060/gpgpusim.config (limited to 'configs/tested-cfgs/SM75_RTX2060/gpgpusim.config') diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config new file mode 100644 index 0000000..9d3992a --- /dev/null +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -0,0 +1,187 @@ +# This config models the Turing RTX 2060 +# For more info about turing architecture: +# https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/technologies/turing-architecture/NVIDIA-Turing-Architecture-Whitepaper.pdf +# "RTX on—The NVIDIA Turing GPU", IEEE MICRO 2020 + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# Device Limits +-gpgpu_stack_size_limit 1024 +-gpgpu_heap_size_limit 8388608 +-gpgpu_runtime_sync_depth_limit 2 +-gpgpu_runtime_pending_launch_count_limit 2048 +-gpgpu_kernel_launch_latency 5000 + +# Compute Capability +-gpgpu_compute_capability_major 7 +-gpgpu_compute_capability_minor 5 + +# PTX execution-driven +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 +# SASS trace-driven mode execution +#-trace_driven_mode 1 + +# high level architecture configuration +-gpgpu_n_clusters 30 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# volta clock domains +#-gpgpu_clock_domains ::: +-gpgpu_clock_domains 1365.0:1365.0:1365.0:3500.0 +# boost mode +# -gpgpu_clock_domains 1680.0:1680.0:1680.0:3500.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_registers_per_block 65536 +-gpgpu_occupancy_sm_number 75 + +# 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_INT,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_INT,OC_EX_SFU,OC_EX_MEM,EX_WB,ID_OC_TENSOR_CORE,OC_EX_TENSOR_CORE +## Turing has 4 SP SIMD units, 4 INT units, 4 SFU units, 8 Tensor core units +## We need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,0,4,4,4,4,0,4,4,4,8,4,4 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_int_units 4 +-gpgpu_tensor_core_avail 1 +-gpgpu_num_tensor_core_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +-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 +-ptx_opcode_latency_tesnor 64 +-ptx_opcode_initiation_tensor 64 + +# Trung has sub core model, in which each scheduler has its own register file and EUs +# i.e. schedulers are isolated +-sub_core_model 1 +# disable specialized operand collectors and use generic operand collectors instead +-enable_specialized_operand_collector 0 +-gpgpu_operand_collector_num_units_gen 8 +-gpgpu_operand_collector_num_in_ports_gen 8 +-gpgpu_operand_collector_num_out_ports_gen 8 +# turing has 8 banks dual-port, 4 schedulers, two banks per scheduler +# we increase #banks to 16 to mitigate the effect of Regisrer File Cache (RFC) which we do not implement in the current version +-gpgpu_num_reg_banks 16 +-gpgpu_reg_file_port_throughput 2 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Turing, a warp scheduler can issue 1 inst per cycle +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# Turing has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +-adaptive_cache_config 0 +-l1_banks 4 +-gpgpu_cache:dl1 S:1:128:512,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_sizeDefault 65536 +-gpgpu_shmem_per_block 65536 +-gmem_skip_L1D 0 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 20 +-smem_latency 20 +-gpgpu_flush_l1_cache 1 + +# 32 sets, each 128 bytes 32-way for each memory sub partition (96 KB per memory sub partition). This gives us 6MB L2 cache +-gpgpu_cache:dl2 S:32:128:32,L:B:m:L:L,A:192:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 1 +-memory_partition_indexing 0 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +-inst_fetch_throughput 4 +# 128 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +-gpgpu_tex_cache:l1 N:4:128:256,L:R:m:N:L,T:512:8,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 +-perfect_inst_const_cache 1 + +# interconnection +#-network_mode 1 +#-inter_config_file config_turing_islip.icnt +# use built-in local xbar +-network_mode 2 +-inct_in_buffer_limit 512 +-inct_out_buffer_limit 512 +-inct_subnets 2 +-arbiter_algo 1 + +# memory partition latency config +-rop_latency 160 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# Turing has GDDR6 +# http://monitorinsider.com/GDDR6.html +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 2 +-gpgpu_dram_burst_length 16 +-dram_data_command_freq_ratio 4 # GDDR6 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, scaled to 3500MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=4:RRD=10:RCD=20:RAS=50:RP=20:RC=62: + CL=20:WL=8:CDLR=9:WR=20:nbkgrp=4:CCDL=4:RTPL=4" + +# 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 + +# 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 Volta +-power_simulation_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + -- cgit v1.3 From fe8dd7eadbb67b9e917c92c9c540812db35fce37 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Thu, 2 Apr 2020 21:41:01 -0400 Subject: updating the turing config --- configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config | 2 +- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 18 +++++++++--------- 2 files changed, 10 insertions(+), 10 deletions(-) (limited to 'configs/tested-cfgs/SM75_RTX2060/gpgpusim.config') diff --git a/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config b/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config index 1951a3f..77617d6 100644 --- a/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config +++ b/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config @@ -40,7 +40,7 @@ -gpgpu_shader_registers 65536 -gpgpu_occupancy_sm_number 62 -# This implies a maximum of 32 warps/SM +# This implies a maximum of 64 warps/SM -gpgpu_shader_core_pipeline 2048:32 -gpgpu_shader_cta 16 -gpgpu_simd_model 1 diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index 9d3992a..e8329dd 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -1,12 +1,12 @@ # This config models the Turing RTX 2060 # For more info about turing architecture: -# https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/technologies/turing-architecture/NVIDIA-Turing-Architecture-Whitepaper.pdf -# "RTX on—The NVIDIA Turing GPU", IEEE MICRO 2020 +# 1- https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/technologies/turing-architecture/NVIDIA-Turing-Architecture-Whitepaper.pdf +# 2- "RTX on—The NVIDIA Turing GPU", IEEE MICRO 2020 # functional simulator specification -gpgpu_ptx_instruction_classification 0 -gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 +-gpgpu_ptx_force_max_capability 75 # Device Limits -gpgpu_stack_size_limit 1024 @@ -33,17 +33,17 @@ # volta clock domains #-gpgpu_clock_domains ::: --gpgpu_clock_domains 1365.0:1365.0:1365.0:3500.0 +-gpgpu_clock_domains 1365.0:1365.0:1365.0:7000.0 # boost mode -# -gpgpu_clock_domains 1680.0:1680.0:1680.0:3500.0 +# -gpgpu_clock_domains 1680.0:1680.0:1680.0:7000.0 # shader core pipeline config -gpgpu_shader_registers 65536 -gpgpu_registers_per_block 65536 -gpgpu_occupancy_sm_number 75 -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 +# This implies a maximum of 32 warps/SM +-gpgpu_shader_core_pipeline 1024:32 -gpgpu_shader_cta 32 -gpgpu_simd_model 1 @@ -89,7 +89,7 @@ -gpgpu_shmem_num_banks 32 -gpgpu_shmem_limited_broadcast 0 -gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 +-gpgpu_coalesce_arch 75 ## In Turing, a warp scheduler can issue 1 inst per cycle -gpgpu_max_insn_issue_per_warp 1 @@ -156,7 +156,7 @@ -gpgpu_n_mem_per_ctrlr 1 -gpgpu_dram_buswidth 2 -gpgpu_dram_burst_length 16 --dram_data_command_freq_ratio 4 # GDDR6 is QDR +-dram_data_command_freq_ratio 2 # GDDR6 is configured as DDR in Turing -gpgpu_mem_address_mask 1 -gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS -- cgit v1.3 From 52204ff08a9c9a21a99fee3f976d2a419c014fec Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Mon, 18 May 2020 21:49:34 -0400 Subject: fixing some failing apps --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 6 +++--- configs/tested-cfgs/SM7_QV100_SASS/gpgpusim.config | 4 ++-- src/gpgpu-sim/gpu-sim.cc | 3 ++- src/gpgpu-sim/shader.cc | 2 +- 4 files changed, 8 insertions(+), 7 deletions(-) (limited to 'configs/tested-cfgs/SM75_RTX2060/gpgpusim.config') diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index e8329dd..b89971e 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -33,9 +33,9 @@ # volta clock domains #-gpgpu_clock_domains ::: --gpgpu_clock_domains 1365.0:1365.0:1365.0:7000.0 +-gpgpu_clock_domains 1365.0:1365.0:1365.0:3500.0 # boost mode -# -gpgpu_clock_domains 1680.0:1680.0:1680.0:7000.0 +# -gpgpu_clock_domains 1680.0:1680.0:1680.0:3500.0 # shader core pipeline config -gpgpu_shader_registers 65536 @@ -156,7 +156,7 @@ -gpgpu_n_mem_per_ctrlr 1 -gpgpu_dram_buswidth 2 -gpgpu_dram_burst_length 16 --dram_data_command_freq_ratio 2 # GDDR6 is configured as DDR in Turing +-dram_data_command_freq_ratio 4 -gpgpu_mem_address_mask 1 -gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS diff --git a/configs/tested-cfgs/SM7_QV100_SASS/gpgpusim.config b/configs/tested-cfgs/SM7_QV100_SASS/gpgpusim.config index 0255f76..0df3eec 100644 --- a/configs/tested-cfgs/SM7_QV100_SASS/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100_SASS/gpgpusim.config @@ -74,8 +74,8 @@ -ptx_opcode_initiation_dp 4,4,4,4,130 -ptx_opcode_latency_sfu 100 -ptx_opcode_initiation_sfu 8 --ptx_opcode_latency_tesnor 6 --ptx_opcode_initiation_tensor 2 +-ptx_opcode_latency_tesnor 8 +-ptx_opcode_initiation_tensor 4 # Volta has sub core model, in which each scheduler has its own register file and EUs # i.e. schedulers are isolated diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index e5b9c9d..cd5fa56 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1828,7 +1828,8 @@ void shader_core_ctx::dump_warp_state( FILE *fout ) const void gpgpu_sim::perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ) { if (m_memory_config->m_perf_sim_memcpy) { - assert (dst_start_addr % 32 == 0); + //if(!m_config.trace_driven_mode) //in trace-driven mode, CUDA runtime can start nre data structure at any position + // assert (dst_start_addr % 32 == 0); for ( unsigned counter = 0; counter < count; counter += 32 ) { const unsigned wr_addr = dst_start_addr + counter; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index ed0c25e..65ec113 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -3091,7 +3091,7 @@ unsigned int shader_core_config::max_cta( const kernel_info_t &k ) const case VOLTA: { //For Volta, we assign the remaining shared memory to L1 cache //For more info about adaptive cache, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x - assert(gpgpu_shmem_size == 98304); //Volta has 96 KB shared + //assert(gpgpu_shmem_size == 98304); //Volta has 96 KB shared //To Do: make it flexible and not tuned to 9KB share memory unsigned max_assoc = m_L1D_config.get_max_assoc(); -- cgit v1.3 From 84213ffa0416139c7711eb0784b8a46b0de8c538 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 22 May 2020 20:16:04 -0400 Subject: code reformatting --- configs/tested-cfgs/SM2_GTX480/gpgpusim.config | 12 +- .../tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config | 21 ++- configs/tested-cfgs/SM6_TITANX/gpgpusim.config | 35 ++-- configs/tested-cfgs/SM6_TITANX/gpgpusim_old.config | 192 ------------------- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 73 ++++---- configs/tested-cfgs/SM7_QV100/gpgpusim.config | 61 +++--- .../SM7_QV100_SASS/config_volta_islip.icnt | 74 -------- configs/tested-cfgs/SM7_QV100_SASS/gpgpusim.config | 206 --------------------- .../tested-cfgs/SM7_TITANV/config_volta_islip.icnt | 74 -------- configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 205 -------------------- src/abstract_hardware_model.h | 2 +- src/gpgpu-sim/addrdec.cc | 2 +- src/gpgpu-sim/gpu-sim.cc | 51 +++-- src/gpgpu-sim/icnt_wrapper.cc | 12 +- src/gpgpu-sim/shader.cc | 2 +- src/gpgpu-sim/shader.h | 7 + src/trace-driven/gpgpusim_trace_driven_main.cc | 3 +- src/trace-driven/kepler_opcode.h | 5 +- src/trace-driven/trace_driven.cc | 104 ++++------- src/trace-driven/trace_driven.h | 28 ++- 20 files changed, 225 insertions(+), 944 deletions(-) delete mode 100644 configs/tested-cfgs/SM6_TITANX/gpgpusim_old.config delete mode 100644 configs/tested-cfgs/SM7_QV100_SASS/config_volta_islip.icnt delete mode 100644 configs/tested-cfgs/SM7_QV100_SASS/gpgpusim.config delete mode 100644 configs/tested-cfgs/SM7_TITANV/config_volta_islip.icnt delete mode 100644 configs/tested-cfgs/SM7_TITANV/gpgpusim.config (limited to 'configs/tested-cfgs/SM75_RTX2060/gpgpusim.config') diff --git a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config index 4a7a3c3..5a12e2e 100644 --- a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config @@ -63,10 +63,10 @@ -gpgpu_shmem_size 49152 -gpgpu_shmem_sizeDefault 49152 -icnt_flit_size 40 --gmem_skip_L1D 0 +-gpgpu_gmem_skip_L1D 0 -gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 35 --smem_latency 26 +-gpgpu_l1_latency 35 +-gpgpu_smem_latency 26 -gpgpu_flush_l1_cache 1 # The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected @@ -77,8 +77,8 @@ -gpgpu_cache:dl2 S:64:128:8,L:B:m:L:L,A:256:4,4:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 1 --memory_partition_indexing 0 +-gpgpu_perf_sim_memcpy 1 +-gpgpu_memory_partition_indexing 0 -gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,S:2:32,4 -gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,T:128:4,128:2 @@ -104,7 +104,7 @@ -inter_config_file config_fermi_islip.icnt # memory partition latency config --rop_latency 120 +-gpgpu_l2_rop_latency 120 -dram_latency 100 # dram model config diff --git a/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config b/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config index 77617d6..b7c0edc 100644 --- a/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config +++ b/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config @@ -67,6 +67,11 @@ -ptx_opcode_initiation_sfu 2 -ptx_opcode_latency_sfu 200 +-trace_opcode_latency_initiation_int 4,1 +-trace_opcode_latency_initiation_sp 4,1 +-trace_opcode_latency_initiation_dp 20,2 +-trace_opcode_latency_initiation_sfu 200,2 + # enable operand collector -gpgpu_operand_collector_num_units_sp 12 -gpgpu_operand_collector_num_units_sfu 6 @@ -112,10 +117,10 @@ -gpgpu_shmem_size_PrefShared 49152 # By default, L1 cache is disabled in Kepler P102 and only enabled for local memory # requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 --gmem_skip_L1D 1 +-gpgpu_gmem_skip_L1D 1 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 82 +-gpgpu_l1_latency 82 -smem_latency 24 -gpgpu_flush_l1_cache 1 @@ -123,12 +128,12 @@ -gpgpu_cache:dl2 S:32:128:16,L:B:m:L:L,A:256:64,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 --perf_sim_memcpy 1 --memory_partition_indexing 0 +-gpgpu_perf_sim_memcpy 1 +-gpgpu_memory_partition_indexing 0 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 --inst_fetch_throughput 8 +-gpgpu_inst_fetch_throughput 8 # 48 KB Tex -gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 # 12 KB Const @@ -139,7 +144,7 @@ -inter_config_file config_kepler_islip.icnt # memory partition latency config --rop_latency 120 +-gpgpu_l2_rop_latency 120 -dram_latency 100 # dram model config @@ -164,8 +169,8 @@ -dram_bnk_indexing_policy 0 -dram_bnkgrp_indexing_policy 1 -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 +#-dram_seperate_write_queue_enable 1 +#-dram_write_queue_size 64:56:32 # stat collection -gpgpu_memlatency_stat 14 diff --git a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config index d54b7d4..adbf66c 100644 --- a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -22,6 +22,8 @@ # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 +# SASS trace-driven mode execution +#-trace_driven_mode 1 # high level architecture configuration -gpgpu_n_clusters 28 @@ -62,13 +64,18 @@ -ptx_opcode_latency_dp 8,19,8,8,330 -ptx_opcode_initiation_dp 8,8,8,8,130 -ptx_opcode_initiation_sfu 4 --ptx_opcode_latency_sfu 8 +-ptx_opcode_latency_sfu 20 + +-trace_opcode_latency_initiation_int 4,1 +-trace_opcode_latency_initiation_sp 4,1 +-trace_opcode_latency_initiation_dp 20,8 +-trace_opcode_latency_initiation_sfu 20,4 # in sub_core_model, schedulers are isolated, each scheduler has its own register file and EUs --sub_core_model 1 +-gpgpu_sub_core_model 1 # enable operand collector # disable specialized operand collectors and use generic operand collectors instead --enable_specialized_operand_collector 0 +-gpgpu_enable_specialized_operand_collector 0 -gpgpu_operand_collector_num_units_gen 8 -gpgpu_operand_collector_num_in_ports_gen 8 -gpgpu_operand_collector_num_out_ports_gen 8 @@ -97,7 +104,7 @@ # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. # The defulat is to disable the L1 cache, unless cache modifieres are used --l1_banks 2 +-gpgpu_l1_banks 2 -gpgpu_cache:dl1 S:4:128:96,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_cache:dl1PrefL1 S:4:128:96,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_cache:dl1PrefShared S:4:128:96,L:L:s:N:L,A:256:8,16:0,32 @@ -107,36 +114,36 @@ -gpgpu_shmem_size_PrefShared 98304 # By default, L1 cache is disabled in Pascal P102. # requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 --gmem_skip_L1D 1 +-gpgpu_gmem_skip_L1D 1 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 82 --smem_latency 24 +-gpgpu_l1_latency 82 +-gpgpu_smem_latency 24 -gpgpu_flush_l1_cache 1 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache -gpgpu_cache:dl2 S:1:128:1024,L:B:m:L:L,A:256:64,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 --perf_sim_memcpy 1 --memory_partition_indexing 4 +-gpgpu_perf_sim_memcpy 1 +-gpgpu_memory_partition_indexing 4 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 --inst_fetch_throughput 8 +-gpgpu_inst_fetch_throughput 8 # 48 KB Tex # Note, TEX is deprected in Pascal, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod -gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 # 12 KB Const -gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 --perfect_inst_const_cache 1 +-gpgpu_perfect_inst_const_cache 1 # interconnection -network_mode 1 -inter_config_file config_pascal_islip.icnt # memory partition latency config --rop_latency 120 +-gpgpu_l2_rop_latency 120 -dram_latency 100 # dram model config @@ -161,8 +168,8 @@ -dram_bnk_indexing_policy 0 -dram_bnkgrp_indexing_policy 1 -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 +#-dram_seperate_write_queue_enable 1 +#-dram_write_queue_size 64:56:32 # stat collection -gpgpu_memlatency_stat 14 diff --git a/configs/tested-cfgs/SM6_TITANX/gpgpusim_old.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim_old.config deleted file mode 100644 index 17ad779..0000000 --- a/configs/tested-cfgs/SM6_TITANX/gpgpusim_old.config +++ /dev/null @@ -1,192 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 --gpgpu_ignore_resources_limitation 1 - -# Device Limits --gpgpu_stack_size_limit 1024 --gpgpu_heap_size_limit 8388608 --gpgpu_runtime_sync_depth_limit 2 --gpgpu_runtime_pending_launch_count_limit 2048 - -# Compute Capability --gpgpu_compute_capability_major 6 --gpgpu_compute_capability_minor 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration -# P102 has two semi-indp scheds per core, and two cores per cluster --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 --gpgpu_occupancy_sm_number 62 - -# This implies a maximum of 32 warps/SM --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 16 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_INT,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_INT,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 4 SFU units per SM. In this config, we split SM into two shader cores, each has 2 SPs and 2 SFUs -# There is no int unit in Pascal -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,1,0,2,1,2,1,0,2,1,5 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 1 - - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,8,8,8,130 --ptx_opcode_initiation_sfu 4 --ptx_opcode_latency_sfu 8 - - -# latencies and cache configs are adopted from: -# https://arxiv.org/pdf/1804.06826.pdf -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB -# Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres are used --gpgpu_cache:dl1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 49152 --gpgpu_shmem_sizeDefault 49152 --gpgpu_shmem_size_PrefL1 49152 --gpgpu_shmem_size_PrefShared 49152 -# By default, L1 cache is disabled in Pascal P102. -# requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 --gmem_skip_L1D 1 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 82 --smem_latency 24 --gpgpu_flush_l1_cache 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 --perf_sim_memcpy 1 --memory_partition_indexing 0 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# Note, TEX is deprected in Pascal, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_operand_collector_num_in_ports_sfu 2 --gpgpu_operand_collector_num_out_ports_sfu 2 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 -# Use Pascal Coalsce arhitetecture --gpgpu_coalesce_arch 61 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_pascal_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 64 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing, scaled to 2500MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: - CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" - --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal 102 has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Pascal 102 --power_simulation_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index b89971e..75b3c99 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -72,11 +72,31 @@ -ptx_opcode_latency_tesnor 64 -ptx_opcode_initiation_tensor 64 +-trace_opcode_latency_initiation_int 4,2 +-trace_opcode_latency_initiation_sp 4,2 +-trace_opcode_latency_initiation_dp 8,4 +-trace_opcode_latency_initiation_sfu 20,8 +-trace_opcode_latency_initiation_tensor 8,4 + +# Turing has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Greedy then oldest scheduler +-gpgpu_scheduler gto +## In Turing, a warp scheduler can issue 1 inst per cycle +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 75 + # Trung has sub core model, in which each scheduler has its own register file and EUs # i.e. schedulers are isolated --sub_core_model 1 +-gpgpu_sub_core_model 1 # disable specialized operand collectors and use generic operand collectors instead --enable_specialized_operand_collector 0 +-gpgpu_enable_specialized_operand_collector 0 -gpgpu_operand_collector_num_units_gen 8 -gpgpu_operand_collector_num_in_ports_gen 8 -gpgpu_operand_collector_num_out_ports_gen 8 @@ -85,65 +105,50 @@ -gpgpu_num_reg_banks 16 -gpgpu_reg_file_port_throughput 2 -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 75 - -## In Turing, a warp scheduler can issue 1 inst per cycle --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -# Turing has four schedulers per core --gpgpu_num_sched_per_core 4 -# Greedy then oldest scheduler --gpgpu_scheduler gto - # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo --adaptive_cache_config 0 --l1_banks 4 +-gpgpu_adaptive_cache_config 0 +-gpgpu_l1_banks 4 -gpgpu_cache:dl1 S:1:128:512,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 65536 -gpgpu_shmem_sizeDefault 65536 -gpgpu_shmem_per_block 65536 --gmem_skip_L1D 0 --icnt_flit_size 40 +-gpgpu_gmem_skip_L1D 0 -gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 20 --smem_latency 20 +-gpgpu_l1_latency 20 +-gpgpu_smem_latency 20 -gpgpu_flush_l1_cache 1 # 32 sets, each 128 bytes 32-way for each memory sub partition (96 KB per memory sub partition). This gives us 6MB L2 cache -gpgpu_cache:dl2 S:32:128:32,L:B:m:L:L,A:192:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 1 --memory_partition_indexing 0 +-gpgpu_perf_sim_memcpy 1 +-gpgpu_memory_partition_indexing 0 # 128 KB Inst. -gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 --inst_fetch_throughput 4 +-gpgpu_inst_fetch_throughput 4 # 128 KB Tex # Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod -gpgpu_tex_cache:l1 N:4:128:256,L:R:m:N:L,T:512:8,128:2 # 64 KB Const -gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 --perfect_inst_const_cache 1 +-gpgpu_perfect_inst_const_cache 1 # interconnection #-network_mode 1 #-inter_config_file config_turing_islip.icnt # use built-in local xbar -network_mode 2 --inct_in_buffer_limit 512 --inct_out_buffer_limit 512 --inct_subnets 2 --arbiter_algo 1 +-icnt_in_buffer_limit 512 +-icnt_out_buffer_limit 512 +-icnt_subnets 2 +-icnt_arbiter_algo 1 +-icnt_flit_size 40 # memory partition latency config --rop_latency 160 +-gpgpu_l2_rop_latency 160 -dram_latency 100 # dram model config @@ -168,8 +173,8 @@ -dram_bnk_indexing_policy 0 -dram_bnkgrp_indexing_policy 1 -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 +#-dram_seperate_write_queue_enable 1 +#-dram_write_queue_size 64:56:32 # stat collection -gpgpu_memlatency_stat 14 diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index 1ed4fb2..30b7d13 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -28,6 +28,8 @@ # PTX execution-driven -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 +# SASS trace-driven mode support +#-trace_driven_mode 1 # high level architecture configuration -gpgpu_n_clusters 80 @@ -77,11 +79,17 @@ -ptx_opcode_latency_tesnor 64 -ptx_opcode_initiation_tensor 64 +-trace_opcode_latency_initiation_int 4,2 +-trace_opcode_latency_initiation_sp 4,2 +-trace_opcode_latency_initiation_dp 8,4 +-trace_opcode_latency_initiation_sfu 20,8 +-trace_opcode_latency_initiation_tensor 8,4 + # Volta has sub core model, in which each scheduler has its own register file and EUs # i.e. schedulers are isolated --sub_core_model 1 +-gpgpu_sub_core_model 1 # disable specialized operand collectors and use generic operand collectors instead --enable_specialized_operand_collector 0 +-gpgpu_enable_specialized_operand_collector 0 -gpgpu_operand_collector_num_units_gen 8 -gpgpu_operand_collector_num_in_ports_gen 8 -gpgpu_operand_collector_num_out_ports_gen 8 @@ -96,6 +104,10 @@ -gpgpu_shmem_warp_parts 1 -gpgpu_coalesce_arch 60 +# Volta has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Greedy then oldest scheduler +-gpgpu_scheduler gto ## In Volta, a warp scheduler can issue 1 inst per cycle -gpgpu_max_insn_issue_per_warp 1 -gpgpu_dual_issue_diff_exec_units 1 @@ -108,49 +120,49 @@ # if the assigned shd mem = 0, then L1 cache = 128KB # For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x # disable this mode in case of multi kernels/apps execution --adaptive_cache_config 1 +-gpgpu_adaptive_cache_config 1 # Volta unified cache has four banks --l1_banks 4 +-gpgpu_l1_banks 4 -gpgpu_cache:dl1 S:1:128:256,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 98304 -gpgpu_shmem_sizeDefault 98304 -gpgpu_shmem_per_block 65536 --gmem_skip_L1D 0 --icnt_flit_size 40 +-gpgpu_gmem_skip_L1D 0 -gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 20 --smem_latency 20 +-gpgpu_l1_latency 20 +-gpgpu_smem_latency 20 -gpgpu_flush_l1_cache 1 # 32 sets, each 128 bytes 24-way for each memory sub partition (96 KB per memory sub partition). This gives us 6MB L2 cache -gpgpu_cache:dl2 S:32:128:24,L:B:m:L:L,A:192:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 1 --memory_partition_indexing 2 +-gpgpu_perf_sim_memcpy 1 +-gpgpu_memory_partition_indexing 2 # 128 KB Inst. -gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 --inst_fetch_throughput 4 +-gpgpu_inst_fetch_throughput 4 # 128 KB Tex # Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod -gpgpu_tex_cache:l1 N:4:128:256,L:R:m:N:L,T:512:8,128:2 # 64 KB Const -gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 --perfect_inst_const_cache 1 +-gpgpu_perfect_inst_const_cache 1 # interconnection #-network_mode 1 #-inter_config_file config_volta_islip.icnt # use built-in local xbar -network_mode 2 --inct_in_buffer_limit 512 --inct_out_buffer_limit 512 --inct_subnets 2 --arbiter_algo 1 +-icnt_in_buffer_limit 512 +-icnt_out_buffer_limit 512 +-icnt_subnets 2 +-icnt_flit_size 40 +-icnt_arbiter_algo 1 # memory partition latency config --rop_latency 160 +-gpgpu_l2_rop_latency 160 -dram_latency 100 # dram model config @@ -177,22 +189,13 @@ CL=12:WL=2:CDLR=3:WR=10: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 +-dram_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 - -# Volta 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 +#-dram_seperate_write_queue_enable 1 +#-dram_write_queue_size 64:56:32 # stat collection -gpgpu_memlatency_stat 14 diff --git a/configs/tested-cfgs/SM7_QV100_SASS/config_volta_islip.icnt b/configs/tested-cfgs/SM7_QV100_SASS/config_volta_islip.icnt deleted file mode 100644 index 5ad7ecd..0000000 --- a/configs/tested-cfgs/SM7_QV100_SASS/config_volta_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 = 144; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 256; -input_buffer_size = 256; -ejection_buffer_size = 256; -boundary_buffer_size = 256; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 2.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/tested-cfgs/SM7_QV100_SASS/gpgpusim.config b/configs/tested-cfgs/SM7_QV100_SASS/gpgpusim.config deleted file mode 100644 index 0df3eec..0000000 --- a/configs/tested-cfgs/SM7_QV100_SASS/gpgpusim.config +++ /dev/null @@ -1,206 +0,0 @@ -# This config models the Volta -# For more info about volta architecture: -# 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 - -# Device Limits --gpgpu_stack_size_limit 1024 --gpgpu_heap_size_limit 8388608 --gpgpu_runtime_sync_depth_limit 2 --gpgpu_runtime_pending_launch_count_limit 2048 --gpgpu_kernel_launch_latency 5000 - -# Compute Capability --gpgpu_compute_capability_major 7 --gpgpu_compute_capability_minor 0 - -# SASS trace-driven mode support --trace_driven_mode 1 --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 80 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 32 --gpgpu_n_sub_partition_per_mchannel 2 - -# volta clock domains -#-gpgpu_clock_domains ::: --gpgpu_clock_domains 1132.0:1132.0:1132.0:850.0 -# boost mode -# -gpgpu_clock_domains 1628.0:1628.0:1628.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 --gpgpu_registers_per_block 65536 --gpgpu_occupancy_sm_number 70 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_INT,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_INT,OC_EX_SFU,OC_EX_MEM,EX_WB,ID_OC_TENSOR_CORE,OC_EX_TENSOR_CORE -## Volta GV100 has 4 SP SIMD units, 4 SFU units, 4 DP units per core, 4 Tensor core units -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,4,4,4,4,4,4,4,8,4,4 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 --gpgpu_num_int_units 4 --gpgpu_tensor_core_avail 1 --gpgpu_num_tensor_core_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit --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 --ptx_opcode_latency_tesnor 8 --ptx_opcode_initiation_tensor 4 - -# Volta has sub core model, in which each scheduler has its own register file and EUs -# i.e. schedulers are isolated --sub_core_model 1 -# disable specialized operand collectors and use generic operand collectors instead --enable_specialized_operand_collector 0 --gpgpu_operand_collector_num_units_gen 8 --gpgpu_operand_collector_num_in_ports_gen 8 --gpgpu_operand_collector_num_out_ports_gen 8 -# volta has 8 banks, 4 schedulers, two banks per scheduler -# we increase #banks to 16 to mitigate the effect of Regisrer File Cache (RFC) which we do not implement in the current version --gpgpu_num_reg_banks 16 --gpgpu_reg_file_port_throughput 2 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Volta, a warp scheduler can issue 1 inst per cycle --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -# Volta has four schedulers per core --gpgpu_num_sched_per_core 4 -# Greedy then oldest scheduler --gpgpu_scheduler gto - -## L1/shared memory configuration -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Defualt config is 32KB DL1 and 96KB shared memory -# In Volta, we assign the remaining shared memory to L1 cache -# if the assigned shd mem = 0, then L1 cache = 128KB -# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x -# disable this mode in case of multi kernels/apps execution --adaptive_cache_config 1 -# Volta unified cache has four banks --l1_banks 4 --gpgpu_cache:dl1 S:1:128:256,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 98304 --gpgpu_shmem_sizeDefault 98304 --gpgpu_shmem_per_block 65536 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 20 --smem_latency 20 --gpgpu_flush_l1_cache 1 - -# 32 sets, each 128 bytes 24-way for each memory sub partition (96 KB per memory sub partition). This gives us 6MB L2 cache --gpgpu_cache:dl2 S:32:128:24,L:B:m:L:L,A:192:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 1 --memory_partition_indexing 2 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 --inst_fetch_throughput 4 -# 128 KB Tex -# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod --gpgpu_tex_cache:l1 N:4:128:256,L:R:m:N:L,T:512:8,128:2 -# 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 --perfect_inst_const_cache 1 - -# interconnection -#-network_mode 1 -#-inter_config_file config_volta_islip.icnt -# use built-in local xbar --network_mode 2 --inct_in_buffer_limit 512 --inct_out_buffer_limit 512 --inct_subnets 2 --arbiter_algo 1 - -# memory partition latency config --rop_latency 160 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, three stacks, 24 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.RBBBCCCB.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 850 MHZ, V100 HBM runs at 850 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=2:CDLR=3:WR=10: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 - -# 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 Volta --power_simulation_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/tested-cfgs/SM7_TITANV/config_volta_islip.icnt b/configs/tested-cfgs/SM7_TITANV/config_volta_islip.icnt deleted file mode 100644 index 615d0a9..0000000 --- a/configs/tested-cfgs/SM7_TITANV/config_volta_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 = 88; -n = 1; - -// Routing - -routing_function = dest_tag; - - -// Flow control - -num_vcs = 1; -vc_buf_size = 256; -input_buffer_size = 256; -ejection_buffer_size = 256; -boundary_buffer_size = 256; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 1; -output_speedup = 1; -internal_speedup = 2.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config deleted file mode 100644 index 0339b0d..0000000 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ /dev/null @@ -1,205 +0,0 @@ -# This config models the Volta Titan V -# For more info about volta architecture: -# 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 - - -# Device Limits --gpgpu_stack_size_limit 1024 --gpgpu_heap_size_limit 8388608 --gpgpu_runtime_sync_depth_limit 2 --gpgpu_runtime_pending_launch_count_limit 2048 - -# Compute Capability --gpgpu_compute_capability_major 7 --gpgpu_compute_capability_minor 0 - -# 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 2 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Volta NVIDIA TITANV clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:1455.0:1455.0:850.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 --gpgpu_registers_per_block 65536 --gpgpu_occupancy_sm_number 70 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_DP,ID_OC_INT,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_INT,OC_EX_SFU,OC_EX_MEM,EX_WB,ID_OC_TENSOR_CORE,OC_EX_TENSOR_CORE -## Volta TITANV has 4 SP SIMD units, 4 INT units, 4 SFU units, 4 DP units per core, 4 Tensor core units -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,4,4,4,4,4,4,4,8,4,4 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 --gpgpu_num_int_units 4 --gpgpu_tensor_core_avail 1 --gpgpu_num_tensor_core_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 --ptx_opcode_latency_tesnor 64 --ptx_opcode_initiation_tensor 64 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Defualt config is 32KB DL1 and 96KB shared memory -# In Volta, we assign the remaining shared memory to L1 cache -# if the assigned shd mem = 0, then L1 cache = 128KB -# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x -# disable this mode in case of multi kernels/apps execution --adaptive_cache_config 1 -# Volta unified cache has four banks --l1_banks 4 -#-mem_unit_ports 4 --gpgpu_cache:dl1 S:1:128:256,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 98304 --gpgpu_shmem_sizeDefault 98304 --gpgpu_shmem_per_block 65536 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 --l1_latency 20 --smem_latency 20 --gpgpu_flush_l1_cache 1 - -# 32 sets, each 128 bytes 24-way for each memory sub partition (96 KB per memory sub partition). This gives us 4.5MB L2 cache --gpgpu_cache:dl2 S:32:128:24,L:B:m:L:L,A:192:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 1 --memory_partition_indexing 4 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -# 48 KB Tex -# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod --gpgpu_tex_cache:l1 N:4:128:256,L:R:m:N:L,T:512:8,128:2 -# 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 - -# Volta has sub core model, in which each scheduler has its own register file and EUs -# i.e. schedulers are isolated --sub_core_model 1 -# disable specialized operand collectors and use generic operand collectors instead --enable_specialized_operand_collector 0 --gpgpu_operand_collector_num_units_gen 8 --gpgpu_operand_collector_num_in_ports_gen 8 --gpgpu_operand_collector_num_out_ports_gen 8 -# volta has 8 banks, 4 schedulers, two banks per scheduler --gpgpu_num_reg_banks 8 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Volta, a warp scheduler can issue 1 inst per cycle --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_volta_islip.icnt -# for local xbar, use: -# "-network_mode 2 -inct_in_buffer_limit 512 -inct_out_buffer_limit 512 -inct_subnets 2" - -# memory partition latency config --rop_latency 160 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, three stacks, 24 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.RBBBCCCB.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 850 MHZ, TITANV HBM runs at 850 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=2:CDLR=3:WR=10: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 - -# Volta has four schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs, disable it untill we create a real energy model for Volta --power_simulation_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 135b03b..6c19e2d 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -70,7 +70,7 @@ enum FuncCache enum AdaptiveCache { FIXED = 0, - VOLTA = 1 + ADAPTIVE_VOLTA = 1 }; #ifdef __cplusplus diff --git a/src/gpgpu-sim/addrdec.cc b/src/gpgpu-sim/addrdec.cc index c34cb32..d430568 100644 --- a/src/gpgpu-sim/addrdec.cc +++ b/src/gpgpu-sim/addrdec.cc @@ -63,7 +63,7 @@ void linear_to_raw_address_translation::addrdec_setoption(option_parser_t opp) option_parser_register(opp, "-gpgpu_mem_address_mask", OPT_INT32, &gpgpu_mem_address_mask, "0 = old addressing mask, 1 = new addressing mask, 2 = new add. mask + flipped bank sel and chip sel bits", "0"); - option_parser_register(opp, "-memory_partition_indexing", OPT_UINT32, &memory_partition_indexing, + option_parser_register(opp, "-gpgpu_memory_partition_indexing", OPT_UINT32, &memory_partition_indexing, "0 = no indexing, 1 = bitwise xoring, 2 = IPoly, 3 = custom indexing", "0"); } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index cd5fa56..641ddbc 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -137,9 +137,9 @@ void power_config::reg_options(class OptionParser * opp) void memory_config::reg_options(class OptionParser * opp) { - option_parser_register(opp, "-perf_sim_memcpy", OPT_BOOL, &m_perf_sim_memcpy, + option_parser_register(opp, "-gpgpu_perf_sim_memcpy", OPT_BOOL, &m_perf_sim_memcpy, "Fill the L2 cache on memcpy", "1"); - option_parser_register(opp, "-simple_dram_model", OPT_BOOL, &simple_dram_model, + option_parser_register(opp, "-gpgpu_simple_dram_model", OPT_BOOL, &simple_dram_model, "simple_dram_model with fixed latency and BW", "0"); option_parser_register(opp, "-gpgpu_dram_scheduler", OPT_INT32, &scheduler_type, "0 = fifo, 1 = FR-FCFS (defaul)", "1"); @@ -187,13 +187,13 @@ void memory_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_dram_timing_opt", OPT_CSTR, &gpgpu_dram_timing_opt, "DRAM timing parameters = {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL}", "4:2:8:12:21:13:34:9:4:5:13:1:0:0"); - option_parser_register(opp, "-rop_latency", OPT_UINT32, &rop_latency, + option_parser_register(opp, "-gpgpu_l2_rop_latency", OPT_UINT32, &rop_latency, "ROP queue latency (default 85)", "85"); option_parser_register(opp, "-dram_latency", OPT_UINT32, &dram_latency, "DRAM latency (default 30)", "30"); - option_parser_register(opp, "-dual_bus_interface", OPT_UINT32, &dual_bus_interface, + option_parser_register(opp, "-dram_dual_bus_interface", OPT_UINT32, &dual_bus_interface, "dual_bus_interface (default = 0) ", "0"); option_parser_register(opp, "-dram_bnk_indexing_policy", OPT_UINT32, &dram_bnk_indexing_policy, @@ -202,13 +202,13 @@ void memory_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-dram_bnkgrp_indexing_policy", OPT_UINT32, &dram_bnkgrp_indexing_policy, "dram_bnkgrp_indexing_policy (0 = take higher bits, 1 = take lower bits) (Default = 0)", "0"); - option_parser_register(opp, "-Seperate_Write_Queue_Enable", OPT_BOOL, &seperate_write_queue_enabled, + option_parser_register(opp, "-dram_seperate_write_queue_enable", OPT_BOOL, &seperate_write_queue_enabled, "Seperate_Write_Queue_Enable", "0"); - option_parser_register(opp, "-Write_Queue_Size", OPT_CSTR, &write_queue_size_opt, + option_parser_register(opp, "-dram_write_queue_size", OPT_CSTR, &write_queue_size_opt, "Write_Queue_Size", "32:28:16"); - option_parser_register(opp, "-Elimnate_rw_turnaround", OPT_BOOL, &elimnate_rw_turnaround, + option_parser_register(opp, "-dram_elimnate_rw_turnaround", OPT_BOOL, &elimnate_rw_turnaround, "elimnate_rw_turnaround i.e set tWTR and tRTW = 0", "0"); option_parser_register(opp, "-icnt_flit_size", OPT_UINT32, &icnt_flit_size, @@ -240,13 +240,13 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); - option_parser_register(opp, "-l1_banks", OPT_UINT32, &m_L1D_config.l1_banks, + option_parser_register(opp, "-gpgpu_l1_banks", OPT_UINT32, &m_L1D_config.l1_banks, "The number of L1 cache banks", "1"); - option_parser_register(opp, "-l1_latency", OPT_UINT32, &m_L1D_config.l1_latency, + option_parser_register(opp, "-gpgpu_l1_latency", OPT_UINT32, &m_L1D_config.l1_latency, "L1 Hit Latency", "1"); - option_parser_register(opp, "-smem_latency", OPT_UINT32, &smem_latency, + option_parser_register(opp, "-gpgpu_smem_latency", OPT_UINT32, &smem_latency, "smem Latency", "3"); option_parser_register(opp, "-gpgpu_cache:dl1PrefL1", OPT_CSTR, &m_L1D_config.m_config_stringPrefL1, @@ -257,7 +257,7 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); - option_parser_register(opp, "-gmem_skip_L1D", OPT_BOOL, &gmem_skip_L1D, + option_parser_register(opp, "-gpgpu_gmem_skip_L1D", OPT_BOOL, &gmem_skip_L1D, "global memory access skip L1D cache (implements -Xptxas -dlcm=cg, default=no skip)", "0"); @@ -306,7 +306,7 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shmem_size", OPT_UINT32, &gpgpu_shmem_size, "Size of shared memory per shader core (default 16kB)", "16384"); - option_parser_register(opp, "-adaptive_cache_config", OPT_UINT32, &adaptive_cache_config, + option_parser_register(opp, "-gpgpu_adaptive_cache_config", OPT_UINT32, &adaptive_cache_config, "adaptive_cache_config", "0"); option_parser_register(opp, "-gpgpu_shmem_sizeDefault", OPT_UINT32, &gpgpu_shmem_sizeDefault, @@ -327,7 +327,7 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shmem_warp_parts", OPT_INT32, &mem_warp_parts, "Number of portions a warp is divided into for shared memory bank conflict check ", "2"); - option_parser_register(opp, "-mem_unit_ports", OPT_INT32, &mem_unit_ports, + option_parser_register(opp, "-gpgpu_mem_unit_ports", OPT_INT32, &mem_unit_ports, "The number of memory transactions allowed per core cycle", "1"); option_parser_register(opp, "-gpgpu_shmem_warp_parts", OPT_INT32, &mem_warp_parts, @@ -348,10 +348,10 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_reg_bank_use_warp_id", OPT_BOOL, &gpgpu_reg_bank_use_warp_id, "Use warp ID in mapping registers to banks (default = off)", "0"); - option_parser_register(opp, "-sub_core_model", OPT_BOOL, &sub_core_model, + option_parser_register(opp, "-gpgpu_sub_core_model", OPT_BOOL, &sub_core_model, "Sub Core Volta/Pascal model (default = off)", "0"); - option_parser_register(opp, "-enable_specialized_operand_collector", OPT_BOOL, &enable_specialized_operand_collector, + option_parser_register(opp, "-gpgpu_enable_specialized_operand_collector", OPT_BOOL, &enable_specialized_operand_collector, "enable_specialized_operand_collector", "1"); option_parser_register(opp, "-gpgpu_operand_collector_num_units_sp", OPT_INT32, &gpgpu_operand_collector_num_units_sp, @@ -467,15 +467,32 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_concurrent_kernel_sm", OPT_BOOL, &gpgpu_concurrent_kernel_sm, "Support concurrent kernels on a SM (default = disabled)", "0"); - option_parser_register(opp, "-perfect_inst_const_cache", OPT_BOOL, &perfect_inst_const_cache, + option_parser_register(opp, "-gpgpu_perfect_inst_const_cache", OPT_BOOL, &perfect_inst_const_cache, "perfect inst and const cache mode, so all inst and const hits in the cache(default = disabled)", "0"); - option_parser_register(opp, "-inst_fetch_throughput", OPT_INT32, &inst_fetch_throughput, + option_parser_register(opp, "-gpgpu_inst_fetch_throughput", OPT_INT32, &inst_fetch_throughput, "the number of fetched intruction per warp each cycle", "1"); option_parser_register(opp, "-gpgpu_reg_file_port_throughput", OPT_INT32, ®_file_port_throughput, "the number ports of the register file", "1"); + + //used for trace-driven mode + option_parser_register(opp, "-trace_opcode_latency_initiation_int", OPT_CSTR, &trace_opcode_latency_initiation_int, + "Opcode latencies and initiation for integers in trace driven mode ", + "4,1"); + option_parser_register(opp, "-trace_opcode_latency_initiation_sp", OPT_CSTR, &trace_opcode_latency_initiation_sp, + "Opcode latencies and initiation for sp in trace driven mode ", + "4,1"); + option_parser_register(opp, "-trace_opcode_latency_initiation_dp", OPT_CSTR, &trace_opcode_latency_initiation_dp, + "Opcode latencies and initiation for dp in trace driven mode ", + "4,1"); + option_parser_register(opp, "-trace_opcode_latency_initiation_sfu", OPT_CSTR, &trace_opcode_latency_initiation_sfu, + "Opcode latencies and initiation for sfu in trace driven mode ", + "4,1"); + option_parser_register(opp, "-trace_opcode_latency_initiation_tensor", OPT_CSTR, &trace_opcode_latency_initiation_tensor, + "Opcode latencies and initiation for tensor in trace driven mode ", + "4,1"); } void gpgpu_sim_config::reg_options(option_parser_t opp) diff --git a/src/gpgpu-sim/icnt_wrapper.cc b/src/gpgpu-sim/icnt_wrapper.cc index e449bf1..6e3e596 100644 --- a/src/gpgpu-sim/icnt_wrapper.cc +++ b/src/gpgpu-sim/icnt_wrapper.cc @@ -179,12 +179,12 @@ void icnt_reg_options( class OptionParser * opp ) //parameters for local xbar - option_parser_register(opp, "-inct_in_buffer_limit", OPT_UINT32, &g_inct_config.in_buffer_limit, "in_buffer_limit", "64"); - option_parser_register(opp, "-inct_out_buffer_limit", OPT_UINT32, &g_inct_config.out_buffer_limit, "out_buffer_limit", "64"); - option_parser_register(opp, "-inct_subnets", OPT_UINT32, &g_inct_config.subnets, "subnets", "2"); - option_parser_register(opp, "-arbiter_algo", OPT_UINT32, &g_inct_config.arbiter_algo, "arbiter_algo", "1"); - option_parser_register(opp, "-inct_verbose", OPT_UINT32, &g_inct_config.verbose, "inct_verbose", "0"); - option_parser_register(opp, "-inct_grant_cycles", OPT_UINT32, &g_inct_config.grant_cycles, "grant_cycles", "1"); + option_parser_register(opp, "-icnt_in_buffer_limit", OPT_UINT32, &g_inct_config.in_buffer_limit, "in_buffer_limit", "64"); + option_parser_register(opp, "-icnt_out_buffer_limit", OPT_UINT32, &g_inct_config.out_buffer_limit, "out_buffer_limit", "64"); + option_parser_register(opp, "-icnt_subnets", OPT_UINT32, &g_inct_config.subnets, "subnets", "2"); + option_parser_register(opp, "-icnt_arbiter_algo", OPT_UINT32, &g_inct_config.arbiter_algo, "arbiter_algo", "1"); + option_parser_register(opp, "-icnt_verbose", OPT_UINT32, &g_inct_config.verbose, "inct_verbose", "0"); + option_parser_register(opp, "-icnt_grant_cycles", OPT_UINT32, &g_inct_config.grant_cycles, "grant_cycles", "1"); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 65ec113..900ec90 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -3088,7 +3088,7 @@ unsigned int shader_core_config::max_cta( const kernel_info_t &k ) const switch (adaptive_cache_config) { case FIXED: break; - case VOLTA: { + case ADAPTIVE_VOLTA: { //For Volta, we assign the remaining shared memory to L1 cache //For more info about adaptive cache, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x //assert(gpgpu_shmem_size == 98304); //Volta has 96 KB shared diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 665e3a5..ca85903 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1527,6 +1527,13 @@ class shader_core_config : public core_config unsigned inst_fetch_throughput; unsigned reg_file_port_throughput; + char* trace_opcode_latency_initiation_int; + char* trace_opcode_latency_initiation_sp; + char* trace_opcode_latency_initiation_dp; + char* trace_opcode_latency_initiation_sfu; + char* trace_opcode_latency_initiation_tensor; + + }; struct shader_core_stats_pod { diff --git a/src/trace-driven/gpgpusim_trace_driven_main.cc b/src/trace-driven/gpgpusim_trace_driven_main.cc index bedac4c..5e07ace 100644 --- a/src/trace-driven/gpgpusim_trace_driven_main.cc +++ b/src/trace-driven/gpgpusim_trace_driven_main.cc @@ -48,6 +48,7 @@ int main ( int argc, const char **argv ) //prints stats trace_parser tracer(m_gpgpu_sim->get_config().get_traces_filename(), m_gpgpu_sim, m_gpgpu_context); + trace_config config(m_gpgpu_sim); std::vector commandlist = tracer.parse_kernellist_file(); bool first_kernel=true; @@ -69,7 +70,7 @@ int main ( int argc, const char **argv ) first_kernel = false; continue; } - kernel_info = tracer.parse_kernel_info(commandlist[i]); + kernel_info = tracer.parse_kernel_info(commandlist[i], &config); m_gpgpu_sim->launch(kernel_info); } diff --git a/src/trace-driven/kepler_opcode.h b/src/trace-driven/kepler_opcode.h index f2bbc90..4aa8e0f 100644 --- a/src/trace-driven/kepler_opcode.h +++ b/src/trace-driven/kepler_opcode.h @@ -99,12 +99,13 @@ static const std::unordered_map Kepler_OpcodeMap = { //Load/Store Instructions //For now, we ignore constant loads, consider it as ALU_OP, TO DO {"LDC", OpcodeChar(OP_LDC, ALU_OP)}, - {"LD", OpcodeChar(OP_LD, LOAD_OP)}, + //in Kepler, LD is load global so set it to LDG + {"LD", OpcodeChar(OP_LDG, LOAD_OP)}, {"LDG", OpcodeChar(OP_LDG, LOAD_OP)}, {"LDL", OpcodeChar(OP_LDL, LOAD_OP)}, {"LDS", OpcodeChar(OP_LDS, LOAD_OP)}, {"LDSLK", OpcodeChar(OP_LDSLK, LOAD_OP)}, - {"ST", OpcodeChar(OP_ST, STORE_OP)}, + {"ST", OpcodeChar(OP_STG, STORE_OP)}, {"STL", OpcodeChar(OP_STL, STORE_OP)}, {"STS", OpcodeChar(OP_STS, STORE_OP)}, {"STSCUL", OpcodeChar(OP_STSCUL, STORE_OP)}, diff --git a/src/trace-driven/trace_driven.cc b/src/trace-driven/trace_driven.cc index 35e953e..22c527e 100644 --- a/src/trace-driven/trace_driven.cc +++ b/src/trace-driven/trace_driven.cc @@ -98,7 +98,7 @@ void trace_parser::parse_memcpy_info(const std::string& memcpy_command, size_t& ss>>std::dec>>count; } -trace_kernel_info_t* trace_parser::parse_kernel_info(const std::string& kerneltraces_filepath) { +trace_kernel_info_t* trace_parser::parse_kernel_info(const std::string& kerneltraces_filepath, trace_config* config) { ifs.open(kerneltraces_filepath.c_str()); @@ -167,7 +167,7 @@ trace_kernel_info_t* trace_parser::parse_kernel_info(const std::string& kerneltr dim3 blockDim(tb_dim_x, tb_dim_y, tb_dim_z); trace_function_info* function_info = new trace_function_info(info, m_gpgpu_context); function_info->set_name(kernel_name.c_str()); - trace_kernel_info_t* kernel_info = new trace_kernel_info_t(gridDim, blockDim, binary_verion, function_info, &ifs, m_gpgpu_sim, m_gpgpu_context); + trace_kernel_info_t* kernel_info = new trace_kernel_info_t(gridDim, blockDim, binary_verion, function_info, &ifs, m_gpgpu_sim, m_gpgpu_context, config); return kernel_info; } @@ -211,11 +211,12 @@ address_type trace_shd_warp_t::get_pc(){ return warp_traces[trace_pc].pc; } -trace_kernel_info_t::trace_kernel_info_t(dim3 gridDim, dim3 blockDim, unsigned m_binary_verion, trace_function_info* m_function_info, std::ifstream* inputstream, gpgpu_sim * gpgpu_sim, gpgpu_context* gpgpu_context):kernel_info_t(gridDim, blockDim, m_function_info) { +trace_kernel_info_t::trace_kernel_info_t(dim3 gridDim, dim3 blockDim, unsigned m_binary_verion, trace_function_info* m_function_info, std::ifstream* inputstream, gpgpu_sim * gpgpu_sim, gpgpu_context* gpgpu_context, class trace_config* config):kernel_info_t(gridDim, blockDim, m_function_info) { ifs = inputstream; m_gpgpu_sim = gpgpu_sim; m_gpgpu_context = gpgpu_context; binary_verion = m_binary_verion; + m_tconfig = config; //resolve the binary version if(m_binary_verion == VOLTA_BINART_VERSION) @@ -285,8 +286,8 @@ bool trace_kernel_info_t::get_next_threadblock_traces(std::vectorgetShaderCoreConfig(), m_gpgpu_context); - inst.parse_from_string(line, OpcodeMap, binary_verion); + trace_warp_inst_t inst(m_gpgpu_sim->getShaderCoreConfig(), m_gpgpu_context, m_tconfig); + inst.parse_from_string(line, OpcodeMap); threadblock_traces[warp_id]->push_back(inst); } } @@ -323,7 +324,7 @@ unsigned trace_warp_inst_t::get_datawidth_from_opcode(const std::vector* OpcodeMap, unsigned binary_verion){ +bool trace_warp_inst_t::parse_from_string(std::string trace, const std::unordered_map* OpcodeMap){ std::stringstream ss; ss.str(trace); @@ -473,7 +474,7 @@ bool trace_warp_inst_t::parse_from_string(std::string trace, const std::unordere //remove redundant registers //fill latency and initl - set_latency(op); + m_tconfig->set_latency(op, latency, initiation_interval); //fill addresses if(mem_width > 0) { @@ -548,10 +549,7 @@ bool trace_warp_inst_t::parse_from_string(std::string trace, const std::unordere //right now, we consider all loads are shared. assert(mem_width>0); data_size = get_datawidth_from_opcode(opcode_tokens); - if(binary_verion == KEPLER_BINART_VERSION) - space.set_type(global_space); - else - space.set_type(shared_space); + space.set_type(shared_space); if(m_opcode == OP_LD) memory_op = memory_load; else @@ -584,59 +582,29 @@ bool trace_warp_inst_t::parse_from_string(std::string trace, const std::unordere return true; } -void trace_warp_inst_t::set_latency(unsigned category) +trace_config::trace_config(gpgpu_sim* m_gpgpu_sim){ + + this->m_gpgpu_sim=m_gpgpu_sim; + parse_config(); +} + +void trace_config::parse_config() { - unsigned int_latency[5]; - unsigned fp_latency[5]; - unsigned dp_latency[5]; - unsigned sfu_latency; - unsigned tensor_latency; - unsigned int_init[5]; - unsigned fp_init[5]; - unsigned dp_init[5]; - unsigned sfu_init; - unsigned tensor_init; - - /* - * [0] ADD,SUB - * [1] MAX,Min - * [2] MUL - * [3] MAD - * [4] DIV - */ - sscanf(m_gpgpu_context->func_sim->opcode_latency_int, "%u,%u,%u,%u,%u", - &int_latency[0],&int_latency[1],&int_latency[2], - &int_latency[3],&int_latency[4]); - sscanf(m_gpgpu_context->func_sim->opcode_latency_fp, "%u,%u,%u,%u,%u", - &fp_latency[0],&fp_latency[1],&fp_latency[2], - &fp_latency[3],&fp_latency[4]); - sscanf(m_gpgpu_context->func_sim->opcode_latency_dp, "%u,%u,%u,%u,%u", - &dp_latency[0],&dp_latency[1],&dp_latency[2], - &dp_latency[3],&dp_latency[4]); - sscanf(m_gpgpu_context->func_sim->opcode_latency_sfu, "%u", - &sfu_latency); - sscanf(m_gpgpu_context->func_sim->opcode_latency_tensor, "%u", - &tensor_latency); - sscanf(m_gpgpu_context->func_sim->opcode_initiation_int, "%u,%u,%u,%u,%u", - &int_init[0],&int_init[1],&int_init[2], - &int_init[3],&int_init[4]); - sscanf(m_gpgpu_context->func_sim->opcode_initiation_fp, "%u,%u,%u,%u,%u", - &fp_init[0],&fp_init[1],&fp_init[2], - &fp_init[3],&fp_init[4]); - sscanf(m_gpgpu_context->func_sim->opcode_initiation_dp, "%u,%u,%u,%u,%u", - &dp_init[0],&dp_init[1],&dp_init[2], - &dp_init[3],&dp_init[4]); - sscanf(m_gpgpu_context->func_sim->opcode_initiation_sfu, "%u", - &sfu_init); - sscanf(m_gpgpu_context->func_sim->opcode_initiation_tensor, "%u", - &tensor_init); - sscanf(m_gpgpu_context->func_sim->cdp_latency_str, "%u,%u,%u,%u,%u", - &m_gpgpu_context->func_sim->cdp_latency[0], - &m_gpgpu_context->func_sim->cdp_latency[1], - &m_gpgpu_context->func_sim->cdp_latency[2], - &m_gpgpu_context->func_sim->cdp_latency[3], - &m_gpgpu_context->func_sim->cdp_latency[4]); + sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_int, "%u,%u", + &int_latency,&int_init); + sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_sp, "%u,%u", + &fp_latency,&fp_init); + sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_dp, "%u,%u", + &dp_latency,&dp_init); + sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_sfu, "%u,%u", + &sfu_latency,&sfu_init); + sscanf(m_gpgpu_sim->getShaderCoreConfig()->trace_opcode_latency_initiation_tensor, "%u,%u", + &tensor_latency,&tensor_init); + +} +void trace_config::set_latency(unsigned category, unsigned& latency, unsigned& initiation_interval) +{ initiation_interval = latency = 1; switch(category){ @@ -645,16 +613,16 @@ void trace_warp_inst_t::set_latency(unsigned category) case BRANCH_OP: case CALL_OPS: case RET_OPS: - latency = int_latency[0]; - initiation_interval = int_init[0]; + latency = int_latency; + initiation_interval = int_init; break; case SP_OP: - latency = fp_latency[0]; - initiation_interval = fp_init[0]; + latency = fp_latency; + initiation_interval = fp_init; break; case DP_OP: - latency = dp_latency[0]; - initiation_interval = dp_init[0]; + latency = dp_latency; + initiation_interval = dp_init; break; case SFU_OP: latency = sfu_latency; diff --git a/src/trace-driven/trace_driven.h b/src/trace-driven/trace_driven.h index 9539e6d..2888f86 100644 --- a/src/trace-driven/trace_driven.h +++ b/src/trace-driven/trace_driven.h @@ -37,18 +37,20 @@ public: trace_warp_inst_t() { m_gpgpu_context=NULL; m_opcode=0; + m_tconfig=NULL; } - trace_warp_inst_t(const class core_config *config, gpgpu_context* gpgpu_context ):warp_inst_t(config) { + trace_warp_inst_t(const class core_config *config, gpgpu_context* gpgpu_context, class trace_config* tconfig ):warp_inst_t(config) { m_gpgpu_context = gpgpu_context; m_opcode=0; + m_tconfig=tconfig; } - bool parse_from_string(std::string trace, const std::unordered_map* OpcodeMap, unsigned binary_verion); + bool parse_from_string(std::string trace, const std::unordered_map* OpcodeMap); private: - void set_latency(unsigned cat); gpgpu_context* m_gpgpu_context; + class trace_config* m_tconfig; unsigned m_opcode; bool check_opcode_contain(const std::vector& opcode, std::string param); unsigned get_datawidth_from_opcode(const std::vector& opcode); @@ -56,7 +58,7 @@ private: class trace_kernel_info_t: public kernel_info_t { public: - trace_kernel_info_t(dim3 gridDim, dim3 blockDim, unsigned m_binary_verion, trace_function_info* m_function_info, std::ifstream* inputstream, gpgpu_sim * gpgpu_sim, gpgpu_context* gpgpu_context); + trace_kernel_info_t(dim3 gridDim, dim3 blockDim, unsigned m_binary_verion, trace_function_info* m_function_info, std::ifstream* inputstream, gpgpu_sim * gpgpu_sim, gpgpu_context* gpgpu_context, class trace_config* config); bool get_next_threadblock_traces(std::vector*> threadblock_traces); @@ -64,19 +66,35 @@ private: std::ifstream* ifs; gpgpu_sim * m_gpgpu_sim; gpgpu_context* m_gpgpu_context; + trace_config* m_tconfig; unsigned binary_verion; const std::unordered_map* OpcodeMap; }; +class trace_config { +public: + trace_config(gpgpu_sim * m_gpgpu_sim); + + void set_latency(unsigned category, unsigned& latency, unsigned& initiation_interval); + void parse_config(); + + +private: + + unsigned int_latency, fp_latency, dp_latency, sfu_latency, tensor_latency; + unsigned int_init, fp_init, dp_init, sfu_init, tensor_init; + gpgpu_sim* m_gpgpu_sim; + +}; class trace_parser { public: trace_parser(const char* kernellist_filepath, gpgpu_sim * m_gpgpu_sim, gpgpu_context* m_gpgpu_context); std::vector parse_kernellist_file(); - trace_kernel_info_t* parse_kernel_info(const std::string& kerneltraces_filepath); + trace_kernel_info_t* parse_kernel_info(const std::string& kerneltraces_filepath, trace_config* config); void parse_memcpy_info(const std::string& memcpy_command, size_t& add, size_t& count); void kernel_finalizer(trace_kernel_info_t* kernel_info); -- cgit v1.3 From 90a36a59f5619790b7f6d80375f69d05a75c0a82 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Sat, 23 May 2020 21:08:12 -0400 Subject: adding back titianC --- .../tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config | 4 +- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 4 +- .../tested-cfgs/SM7_TITANV/config_volta_islip.icnt | 74 ++++++++++++++++++++++ configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 18 ++++-- 4 files changed, 89 insertions(+), 11 deletions(-) create mode 100644 configs/tested-cfgs/SM7_TITANV/config_volta_islip.icnt (limited to 'configs/tested-cfgs/SM75_RTX2060/gpgpusim.config') diff --git a/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config b/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config index b7c0edc..9d0862a 100644 --- a/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config +++ b/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config @@ -58,8 +58,8 @@ # "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_int 4,13,4,5,145,32 +-ptx_opcode_initiation_int 1,1,1,1,4,4 -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 diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index 75b3c99..5d23d1a 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -61,8 +61,8 @@ # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" # All Div operations are executed on SFU unit --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_int 4,13,4,5,145,32 +-ptx_opcode_initiation_int 2,2,2,2,8,4 -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 diff --git a/configs/tested-cfgs/SM7_TITANV/config_volta_islip.icnt b/configs/tested-cfgs/SM7_TITANV/config_volta_islip.icnt new file mode 100644 index 0000000..615d0a9 --- /dev/null +++ b/configs/tested-cfgs/SM7_TITANV/config_volta_islip.icnt @@ -0,0 +1,74 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 88; +n = 1; + +// Routing + +routing_function = dest_tag; + + +// Flow control + +num_vcs = 1; +vc_buf_size = 256; +input_buffer_size = 256; +ejection_buffer_size = 256; +boundary_buffer_size = 256; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index eece246..b5f88ce 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -70,8 +70,8 @@ # 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_int 4,13,4,5,145,32 +-ptx_opcode_initiation_int 2,2,2,2,8,4 -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 @@ -124,7 +124,6 @@ -gpgpu_adaptive_cache_config 1 # Volta unified cache has four banks -gpgpu_l1_banks 4 -#-mem_unit_ports 4 -gpgpu_cache:dl1 S:1:128:256,L:L:s:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 98304 -gpgpu_shmem_sizeDefault 98304 @@ -153,11 +152,15 @@ -gpgpu_perfect_inst_const_cache 1 # interconnection --network_mode 1 --inter_config_file config_volta_islip.icnt +#-network_mode 1 +#-inter_config_file config_volta_islip.icnt +# use built-in local xbar +-network_mode 2 +-icnt_in_buffer_limit 512 +-icnt_out_buffer_limit 512 +-icnt_subnets 2 -icnt_flit_size 40 -# for local xbar, use: -# "-network_mode 2 -inct_in_buffer_limit 512 -inct_out_buffer_limit 512 -inct_subnets 2" +-icnt_arbiter_algo 1 # memory partition latency config -gpgpu_l2_rop_latency 160 @@ -209,3 +212,4 @@ #-trace_components WARP_SCHEDULER,SCOREBOARD #-trace_sampling_core 0 + -- cgit v1.3 From 6b72554af7018a8dc42e607f6983a070fe5e5a42 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 28 May 2020 09:05:56 -0400 Subject: splitting execution-driven and trace-driven config files --- configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config | 5 ----- configs/tested-cfgs/SM3_KEPLER_TITAN/trace.config | 4 ++++ configs/tested-cfgs/SM6_TITANX/gpgpusim.config | 7 ------- configs/tested-cfgs/SM6_TITANX/trace.config | 4 ++++ configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 8 -------- configs/tested-cfgs/SM75_RTX2060/trace.config | 5 +++++ configs/tested-cfgs/SM7_QV100/gpgpusim.config | 8 -------- configs/tested-cfgs/SM7_QV100/trace.config | 6 ++++++ configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 8 -------- configs/tested-cfgs/SM7_TITANV/trace.config | 6 ++++++ 10 files changed, 25 insertions(+), 36 deletions(-) create mode 100644 configs/tested-cfgs/SM3_KEPLER_TITAN/trace.config create mode 100644 configs/tested-cfgs/SM6_TITANX/trace.config create mode 100644 configs/tested-cfgs/SM75_RTX2060/trace.config create mode 100644 configs/tested-cfgs/SM7_QV100/trace.config create mode 100644 configs/tested-cfgs/SM7_TITANV/trace.config (limited to 'configs/tested-cfgs/SM75_RTX2060/gpgpusim.config') diff --git a/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config b/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config index 9d0862a..b173dd0 100644 --- a/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config +++ b/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config @@ -67,11 +67,6 @@ -ptx_opcode_initiation_sfu 2 -ptx_opcode_latency_sfu 200 --trace_opcode_latency_initiation_int 4,1 --trace_opcode_latency_initiation_sp 4,1 --trace_opcode_latency_initiation_dp 20,2 --trace_opcode_latency_initiation_sfu 200,2 - # enable operand collector -gpgpu_operand_collector_num_units_sp 12 -gpgpu_operand_collector_num_units_sfu 6 diff --git a/configs/tested-cfgs/SM3_KEPLER_TITAN/trace.config b/configs/tested-cfgs/SM3_KEPLER_TITAN/trace.config new file mode 100644 index 0000000..4c80036 --- /dev/null +++ b/configs/tested-cfgs/SM3_KEPLER_TITAN/trace.config @@ -0,0 +1,4 @@ +-trace_opcode_latency_initiation_int 4,1 +-trace_opcode_latency_initiation_sp 4,1 +-trace_opcode_latency_initiation_dp 20,2 +-trace_opcode_latency_initiation_sfu 200,2 diff --git a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config index a686238..ce6f745 100644 --- a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -22,8 +22,6 @@ # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 -# SASS trace-driven mode execution -#-trace_driven_mode 1 # high level architecture configuration -gpgpu_n_clusters 28 @@ -66,11 +64,6 @@ -ptx_opcode_initiation_sfu 4 -ptx_opcode_latency_sfu 20 --trace_opcode_latency_initiation_int 4,1 --trace_opcode_latency_initiation_sp 4,1 --trace_opcode_latency_initiation_dp 20,8 --trace_opcode_latency_initiation_sfu 20,4 - # in sub_core_model, schedulers are isolated, each scheduler has its own register file and EUs -gpgpu_sub_core_model 1 # enable operand collector diff --git a/configs/tested-cfgs/SM6_TITANX/trace.config b/configs/tested-cfgs/SM6_TITANX/trace.config new file mode 100644 index 0000000..88bcdc0 --- /dev/null +++ b/configs/tested-cfgs/SM6_TITANX/trace.config @@ -0,0 +1,4 @@ +-trace_opcode_latency_initiation_int 4,1 +-trace_opcode_latency_initiation_sp 4,1 +-trace_opcode_latency_initiation_dp 20,8 +-trace_opcode_latency_initiation_sfu 20,4 diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index 5d23d1a..8a4be23 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -22,8 +22,6 @@ # PTX execution-driven -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 -# SASS trace-driven mode execution -#-trace_driven_mode 1 # high level architecture configuration -gpgpu_n_clusters 30 @@ -72,12 +70,6 @@ -ptx_opcode_latency_tesnor 64 -ptx_opcode_initiation_tensor 64 --trace_opcode_latency_initiation_int 4,2 --trace_opcode_latency_initiation_sp 4,2 --trace_opcode_latency_initiation_dp 8,4 --trace_opcode_latency_initiation_sfu 20,8 --trace_opcode_latency_initiation_tensor 8,4 - # Turing has four schedulers per core -gpgpu_num_sched_per_core 4 # Greedy then oldest scheduler diff --git a/configs/tested-cfgs/SM75_RTX2060/trace.config b/configs/tested-cfgs/SM75_RTX2060/trace.config new file mode 100644 index 0000000..41987cf --- /dev/null +++ b/configs/tested-cfgs/SM75_RTX2060/trace.config @@ -0,0 +1,5 @@ +-trace_opcode_latency_initiation_int 4,2 +-trace_opcode_latency_initiation_sp 4,2 +-trace_opcode_latency_initiation_dp 8,4 +-trace_opcode_latency_initiation_sfu 20,8 +-trace_opcode_latency_initiation_tensor 8,4 diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index 5856e5d..c31c060 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -28,8 +28,6 @@ # PTX execution-driven -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 -# SASS trace-driven mode support -#-trace_driven_mode 1 # high level architecture configuration -gpgpu_n_clusters 80 @@ -79,12 +77,6 @@ -ptx_opcode_latency_tesnor 64 -ptx_opcode_initiation_tensor 64 --trace_opcode_latency_initiation_int 4,2 --trace_opcode_latency_initiation_sp 4,2 --trace_opcode_latency_initiation_dp 8,4 --trace_opcode_latency_initiation_sfu 20,8 --trace_opcode_latency_initiation_tensor 8,4 - # Volta has sub core model, in which each scheduler has its own register file and EUs # i.e. schedulers are isolated -gpgpu_sub_core_model 1 diff --git a/configs/tested-cfgs/SM7_QV100/trace.config b/configs/tested-cfgs/SM7_QV100/trace.config new file mode 100644 index 0000000..04ac009 --- /dev/null +++ b/configs/tested-cfgs/SM7_QV100/trace.config @@ -0,0 +1,6 @@ +-trace_opcode_latency_initiation_int 4,2 +-trace_opcode_latency_initiation_sp 4,2 +-trace_opcode_latency_initiation_dp 8,4 +-trace_opcode_latency_initiation_sfu 20,8 +-trace_opcode_latency_initiation_tensor 8,4 + diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index b5f88ce..ef28dd8 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -26,8 +26,6 @@ # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 -gpgpu_ptx_save_converted_ptxplus 0 -# SASS trace-driven mode support -#-trace_driven_mode 1 # high level architecture configuration -gpgpu_n_clusters 40 @@ -81,12 +79,6 @@ -ptx_opcode_latency_tesnor 64 -ptx_opcode_initiation_tensor 64 --trace_opcode_latency_initiation_int 4,2 --trace_opcode_latency_initiation_sp 4,2 --trace_opcode_latency_initiation_dp 8,4 --trace_opcode_latency_initiation_sfu 20,8 --trace_opcode_latency_initiation_tensor 8,4 - # Volta has sub core model, in which each scheduler has its own register file and EUs # i.e. schedulers are isolated -gpgpu_sub_core_model 1 diff --git a/configs/tested-cfgs/SM7_TITANV/trace.config b/configs/tested-cfgs/SM7_TITANV/trace.config new file mode 100644 index 0000000..04ac009 --- /dev/null +++ b/configs/tested-cfgs/SM7_TITANV/trace.config @@ -0,0 +1,6 @@ +-trace_opcode_latency_initiation_int 4,2 +-trace_opcode_latency_initiation_sp 4,2 +-trace_opcode_latency_initiation_dp 8,4 +-trace_opcode_latency_initiation_sfu 20,8 +-trace_opcode_latency_initiation_tensor 8,4 + -- cgit v1.3 From 3e580ee62a9cc8010930f692d8a6201a31ed77e0 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 1 Jun 2020 14:56:30 -0400 Subject: moving all ipoly equstions to one file --- .../tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config | 2 +- configs/tested-cfgs/SM6_TITANX/gpgpusim.config | 2 +- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 4 +- configs/tested-cfgs/SM7_QV100/gpgpusim.config | 2 +- configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 2 +- src/gpgpu-sim/addrdec.cc | 149 ++++++--------------- src/gpgpu-sim/addrdec.h | 2 + src/gpgpu-sim/dram.cc | 24 +--- src/gpgpu-sim/gpu-cache.cc | 97 ++++++-------- src/gpgpu-sim/gpu-cache.h | 26 ++-- src/gpgpu-sim/gpu-sim.cc | 6 + 11 files changed, 116 insertions(+), 200 deletions(-) (limited to 'configs/tested-cfgs/SM75_RTX2060/gpgpusim.config') diff --git a/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config b/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config index b173dd0..c83159f 100644 --- a/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config +++ b/configs/tested-cfgs/SM3_KEPLER_TITAN/gpgpusim.config @@ -120,7 +120,7 @@ -gpgpu_flush_l1_cache 1 # 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 1.5MB L2 cache --gpgpu_cache:dl2 S:32:128:16,L:B:m:L:L,A:256:64,16:0,32 +-gpgpu_cache:dl2 S:32:128:16,L:B:m:L:P,A:256:64,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 -gpgpu_perf_sim_memcpy 1 diff --git a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config index ce6f745..5b243a5 100644 --- a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -115,7 +115,7 @@ -gpgpu_flush_l1_cache 1 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:1:128:1024,L:B:m:L:L,A:256:64,16:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:P,A:256:64,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 -gpgpu_perf_sim_memcpy 1 diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index 8a4be23..6fe04ee 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -111,8 +111,8 @@ -gpgpu_smem_latency 20 -gpgpu_flush_l1_cache 1 -# 32 sets, each 128 bytes 32-way for each memory sub partition (96 KB per memory sub partition). This gives us 6MB L2 cache --gpgpu_cache:dl2 S:32:128:32,L:B:m:L:L,A:192:4,32:0,32 +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives us 3MB L2 cache +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:P,A:192:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 -gpgpu_perf_sim_memcpy 1 diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index c31c060..c4818d1 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -126,7 +126,7 @@ -gpgpu_flush_l1_cache 1 # 32 sets, each 128 bytes 24-way for each memory sub partition (96 KB per memory sub partition). This gives us 6MB L2 cache --gpgpu_cache:dl2 S:32:128:24,L:B:m:L:L,A:192:4,32:0,32 +-gpgpu_cache:dl2 S:32:128:24,L:B:m:L:P,A:192:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 -gpgpu_perf_sim_memcpy 1 diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index ef28dd8..64edc67 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -127,7 +127,7 @@ -gpgpu_flush_l1_cache 1 # 32 sets, each 128 bytes 24-way for each memory sub partition (96 KB per memory sub partition). This gives us 4.5MB L2 cache --gpgpu_cache:dl2 S:32:128:24,L:B:m:L:L,A:192:4,32:0,32 +-gpgpu_cache:dl2 S:32:128:24,L:B:m:L:P,A:192:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 64:64:64:64 -gpgpu_perf_sim_memcpy 1 diff --git a/src/gpgpu-sim/addrdec.cc b/src/gpgpu-sim/addrdec.cc index 91ba47f..c01b8fa 100644 --- a/src/gpgpu-sim/addrdec.cc +++ b/src/gpgpu-sim/addrdec.cc @@ -31,9 +31,12 @@ #include #include "../option_parser.h" #include "gpu-sim.h" +#include "hashing.h" static long int powli(long int x, long int y); static unsigned int LOGB2_32(unsigned int v); +static unsigned next_powerOf2(unsigned n); + static new_addr_type addrdec_packbits(new_addr_type mask, new_addr_type val, unsigned char high, unsigned char low); static void addrdec_getmasklimit(new_addr_type mask, unsigned char *high, @@ -133,121 +136,29 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, break; case BITWISE_PERMUTATION: { assert(!gap); - tlx->chip = (tlx->chip) ^ (rest_of_addr_high_bits & (m_n_channel - 1)); + tlx->chip = + bitwise_hash_function(rest_of_addr_high_bits, tlx->chip, m_n_channel); assert(tlx->chip < m_n_channel); break; } case IPOLY: { - /* - * Set Indexing function from "Pseudo-randomly interleaved memory." - * Rau, B. R et al. - * ISCA 1991 - * http://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=348DEA37A3E440473B3C075EAABC63B6?doi=10.1.1.12.7149&rep=rep1&type=pdf - * - * equations are corresponding to IPOLY(37) and are adopted from: - * "Sacat: streaming-aware conflict-avoiding thrashing-resistant gpgpu - * cache management scheme." Khairy et al. IEEE TPDS 2017. - * - * equations for 32 banks are corresponding to IPOLY(37) - * equations for 64 banks are corresponding to IPOLY(67) - * To see all the IPOLY equations for all the degrees, see - * http://wireless-systems.ece.gatech.edu/6604/handouts/Peterson's%20Table.pdf - * - * We generate these equations using GF(2) arithmetic: - * http://www.ee.unb.ca/cgi-bin/tervo/calc.pl?num=&den=&f=d&e=1&m=1 - * - * We go through all the strides 128 (10000000), 256 (100000000),... and - * do modular arithmetic in GF(2) Then, we create the H-matrix and group - * each bit together, for more info read the ISCA 1991 paper - * - * IPOLY hashing guarantees conflict-free for all 2^n strides which widely - * exit in GPGPU applications and also show good performance for other - * strides. - */ - assert(!gap); - if (m_n_channel == 32 && m_n_sub_partition_in_channel == 1) { - std::bitset<64> a(rest_of_addr_high_bits); - std::bitset<5> chip(tlx->chip); - chip[0] = a[13] ^ a[12] ^ a[11] ^ a[10] ^ a[9] ^ a[6] ^ a[5] ^ a[3] ^ - a[0] ^ chip[0]; - chip[1] = a[14] ^ a[13] ^ a[12] ^ a[11] ^ a[10] ^ a[7] ^ a[6] ^ a[4] ^ - a[1] ^ chip[1]; - chip[2] = a[14] ^ a[10] ^ a[9] ^ a[8] ^ a[7] ^ a[6] ^ a[3] ^ a[2] ^ - a[0] ^ chip[2]; - chip[3] = - a[11] ^ a[10] ^ a[9] ^ a[8] ^ a[7] ^ a[4] ^ a[3] ^ a[1] ^ chip[3]; - chip[4] = - a[12] ^ a[11] ^ a[10] ^ a[9] ^ a[8] ^ a[5] ^ a[4] ^ a[2] ^ chip[4]; - tlx->chip = chip.to_ulong(); - break; - } else if (m_n_channel == 16 && m_n_sub_partition_in_channel == 2) { - std::bitset<64> a(rest_of_addr_high_bits); - std::bitset<4> chip(tlx->chip); - std::bitset<32> bk(tlx->bk); - 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]; - tlx->chip = chip.to_ulong(); - unsigned par_id = - a[12] ^ a[11] ^ a[10] ^ a[9] ^ a[8] ^ a[5] ^ a[4] ^ a[2] ^ bk[0]; - tlx->sub_partition = tlx->chip * m_n_sub_partition_in_channel + par_id; - assert(tlx->chip < m_n_channel); - assert(tlx->sub_partition < m_n_channel * m_n_sub_partition_in_channel); - return; - break; - } else if (m_n_channel == 32 && m_n_sub_partition_in_channel == 2) { - std::bitset<64> a(rest_of_addr_high_bits); - std::bitset<5> chip(tlx->chip); - std::bitset<32> bk(tlx->bk); - chip[0] = a[18] ^ a[17] ^ a[16] ^ a[15] ^ a[12] ^ a[10] ^ a[6] ^ a[5] ^ - a[0] ^ chip[0]; - chip[1] = a[15] ^ a[13] ^ a[12] ^ a[11] ^ a[10] ^ a[7] ^ a[5] ^ a[1] ^ - a[0] ^ chip[1]; - chip[2] = a[16] ^ a[14] ^ a[13] ^ a[12] ^ a[11] ^ a[8] ^ a[6] ^ a[2] ^ - a[1] ^ chip[2]; - chip[3] = a[17] ^ a[15] ^ a[14] ^ a[13] ^ a[12] ^ a[9] ^ a[7] ^ a[3] ^ - a[2] ^ chip[3]; - chip[4] = a[18] ^ a[16] ^ a[15] ^ a[14] ^ a[13] ^ a[10] ^ a[8] ^ a[4] ^ - a[3] ^ chip[4]; - tlx->chip = chip.to_ulong(); - unsigned par_id = - a[17] ^ a[16] ^ a[15] ^ a[14] ^ a[11] ^ a[9] ^ a[5] ^ a[4] ^ bk[0]; - tlx->sub_partition = tlx->chip * m_n_sub_partition_in_channel + par_id; - assert(tlx->chip < m_n_channel); - assert(tlx->sub_partition < m_n_channel * m_n_sub_partition_in_channel); - return; - break; - } else { /* Else incorrect number of channels for the hashing function */ - assert( - "\nGPGPU-Sim memory_partition_indexing error: The number of " - "channels should be " - "32 or 64 for the hashing IPOLY index function.\n" && - 0); - } - assert(tlx->chip < m_n_channel); - break; - } - case PAE: { - // Page Address Entropy - // random selected bits from the page and bank bits - // similar to - // Liu, Yuxi, et al. "Get Out of the Valley: Power-Efficient Address - assert(!gap); - std::bitset<64> a(tlx->row); - std::bitset<5> chip(tlx->chip); - std::bitset<4> b(tlx->bk); - chip[0] = a[13] ^ a[10] ^ a[9] ^ a[5] ^ a[0] ^ b[3] ^ b[0] ^ chip[0]; - chip[1] = a[12] ^ a[11] ^ a[6] ^ a[1] ^ b[3] ^ b[2] ^ b[1] ^ chip[1]; - chip[2] = a[14] ^ a[9] ^ a[8] ^ a[7] ^ a[2] ^ b[1] ^ chip[2]; - chip[3] = a[11] ^ a[10] ^ a[8] ^ a[3] ^ b[2] ^ b[3] ^ chip[3]; - chip[4] = a[12] ^ a[9] ^ a[8] ^ a[5] ^ a[4] ^ b[1] ^ b[0] ^ chip[4]; - tlx->chip = chip.to_ulong(); + // assert(!gap); + unsigned sub_partition_addr_mask = m_n_sub_partition_in_channel - 1; + unsigned sub_partition = tlx->chip * m_n_sub_partition_in_channel + + (tlx->bk & sub_partition_addr_mask); + sub_partition = ipoly_hash_function( + rest_of_addr_high_bits, sub_partition, + nextPowerOf2_m_n_channel * m_n_sub_partition_in_channel); + + if (gap) // if it is not 2^n partitions, then take modular + sub_partition = + sub_partition % (m_n_channel * m_n_sub_partition_in_channel); + + tlx->chip = sub_partition / m_n_channel; + tlx->sub_partition = sub_partition; assert(tlx->chip < m_n_channel); + assert(tlx->sub_partition < m_n_channel * m_n_sub_partition_in_channel); + return; break; } case RANDOM: { @@ -377,6 +288,8 @@ void linear_to_raw_address_translation::init( log2sub_partition = ::LOGB2_32(n_sub_partition_in_channel); m_n_channel = n_channel; m_n_sub_partition_in_channel = n_sub_partition_in_channel; + nextPowerOf2_m_n_channel = ::next_powerOf2(n_channel); + m_n_sub_partition_total = n_channel * n_sub_partition_in_channel; gap = (n_channel - ::powli(2, nchipbits)); if (gap) { @@ -663,6 +576,22 @@ static unsigned int LOGB2_32(unsigned int v) { return r; } +// compute power of two greater than or equal to n +// https://www.techiedelight.com/round-next-highest-power-2/ +unsigned next_powerOf2(unsigned n) { + // decrement n (to handle the case when n itself + // is a power of 2) + n = n - 1; + + // do till only one bit is left + while (n & n - 1) n = n & (n - 1); // unset rightmost bit + + // n is now a power of two (less than n) + + // return next power of 2 + return n << 1; +} + static new_addr_type addrdec_packbits(new_addr_type mask, new_addr_type val, unsigned char high, unsigned char low) { unsigned pos = 0; diff --git a/src/gpgpu-sim/addrdec.h b/src/gpgpu-sim/addrdec.h index dd0e5a0..d8db416 100644 --- a/src/gpgpu-sim/addrdec.h +++ b/src/gpgpu-sim/addrdec.h @@ -87,8 +87,10 @@ class linear_to_raw_address_translation { unsigned int gap; unsigned m_n_channel; int m_n_sub_partition_in_channel; + int m_n_sub_partition_total; unsigned log2channel; unsigned log2sub_partition; + unsigned nextPowerOf2_m_n_channel; }; #endif diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index 041cfce..ca47c46 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -31,6 +31,7 @@ #include "dram_sched.h" #include "gpu-misc.h" #include "gpu-sim.h" +#include "hashing.h" #include "l2cache.h" #include "mem_fetch.h" #include "mem_latency_stat.h" @@ -207,8 +208,8 @@ dram_req_t::dram_req_t(class mem_fetch *mf, unsigned banks, } case BITWISE_XORING_BK_INDEX: { // xoring bank bits with lower bits of the page - int lbank = LOGB2(banks); - bk = tlx.bk ^ (tlx.row & ((1 << lbank) - 1)); + bk = bitwise_hash_function(tlx.row, tlx.bk, banks); + assert(bk < banks); break; } case IPOLY_BK_INDEX: { @@ -216,22 +217,9 @@ dram_req_t::dram_req_t(class mem_fetch *mf, unsigned banks, * memory." Rau, B. R et al. ISCA 1991 * http://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=348DEA37A3E440473B3C075EAABC63B6?doi=10.1.1.12.7149&rep=rep1&type=pdf */ - if (banks == 16) { - std::bitset<64> a(tlx.row); - std::bitset<4> b(tlx.bk); - b[0] = a[11] ^ a[10] ^ a[9] ^ a[8] ^ a[6] ^ a[4] ^ a[3] ^ a[0] ^ b[0]; - b[1] = a[12] ^ a[8] ^ a[7] ^ a[6] ^ a[5] ^ a[3] ^ a[1] ^ a[0] ^ b[1]; - b[2] = a[9] ^ a[8] ^ a[7] ^ a[6] ^ a[4] ^ a[2] ^ a[1] ^ b[2]; - b[3] = a[10] ^ a[9] ^ a[8] ^ a[7] ^ a[5] ^ a[3] ^ a[2] ^ b[3]; - bk = b.to_ulong(); - assert(bk < banks); - } else { /* Else incorrect number of channels for the hashing function */ - assert( - "\nGPGPU-Sim memory_banking indexing error: The number of banks " - "should be " - "16 for the hashing IPOLY index function.\n" && - 0); - } + // xoring bank bits with lower bits of the page + bk = ipoly_hash_function(tlx.row, tlx.bk, banks); + assert(bk < banks); break; } case CUSTOM_BK_INDEX: diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index adce3a2..75c3691 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -29,6 +29,7 @@ #include "gpu-cache.h" #include #include "gpu-sim.h" +#include "hashing.h" #include "stat-tool.h" // used to allocate memory that is large enough to adapt the changes in cache @@ -62,24 +63,31 @@ unsigned l1d_cache_config::set_bank(new_addr_type addr) const { // For sector cache, we select one sector per bank (sector interleaving) // This is what was found in Volta (one sector per bank, sector interleaving) // otherwise, line interleaving - if (m_cache_type == SECTOR) - return (addr >> m_sector_sz_log2) & (l1_banks - 1); - else - return (addr >> m_line_sz_log2) & (l1_banks - 1); + return cache_config::hash_function(addr, l1_banks, l1_banks_byte_interleaving, + m_l1_banks_log2, + l1_banks_hashing_function); +} + +unsigned cache_config::set_index(new_addr_type addr) const { + return cache_config::hash_function(addr, m_nset, m_line_sz_log2, m_nset_log2, + m_set_index_function); } -unsigned l1d_cache_config::set_index(new_addr_type addr) const { - unsigned set_index = m_nset; // Default to linear set index function - unsigned lower_xor = 0; - unsigned upper_xor = 0; +unsigned cache_config::hash_function(new_addr_type addr, unsigned m_nset, + unsigned m_line_sz_log2, + unsigned m_nset_log2, + unsigned m_index_function) const { + unsigned set_index = 0; - switch (m_set_index_function) { - case FERMI_HASH_SET_FUNCTION: - case BITWISE_XORING_FUNCTION: + switch (m_index_function) { + case FERMI_HASH_SET_FUNCTION: { /* * Set Indexing function from "A Detailed GPU Cache Model Based on Reuse * Distance Theory" Cedric Nugteren et al. HPCA 2014 */ + unsigned lower_xor = 0; + unsigned upper_xor = 0; + if (m_nset == 32 || m_nset == 64) { // Lower xor value is bits 7-11 lower_xor = (addr >> m_line_sz_log2) & 0x1F; @@ -102,54 +110,34 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const { 0); } break; + } - case HASH_IPOLY_FUNCTION: - /* - * Set Indexing function from "Pseudo-randomly interleaved memory." - * Rau, B. R et al. - * ISCA 1991 - * - * "Sacat: streaming-aware conflict-avoiding thrashing-resistant gpgpu - * cache management scheme." Khairy et al. IEEE TPDS 2017. - */ - if (m_nset == 32 || m_nset == 64) { - std::bitset<64> a(addr); - std::bitset<6> index; - index[0] = a[25] ^ a[24] ^ a[23] ^ a[22] ^ a[21] ^ a[18] ^ a[17] ^ - a[15] ^ a[12] ^ a[7]; // 10 - index[1] = a[26] ^ a[25] ^ a[24] ^ a[23] ^ a[22] ^ a[19] ^ a[18] ^ - a[16] ^ a[13] ^ a[8]; // 10 - index[2] = a[26] ^ a[22] ^ a[21] ^ a[20] ^ a[19] ^ a[18] ^ a[15] ^ - a[14] ^ a[12] ^ a[9]; // 10 - index[3] = a[23] ^ a[22] ^ a[21] ^ a[20] ^ a[19] ^ a[16] ^ a[15] ^ - a[13] ^ a[10]; // 9 - index[4] = a[24] ^ a[23] ^ a[22] ^ a[21] ^ a[20] ^ a[17] ^ a[16] ^ - a[14] ^ a[11]; // 9 - - if (m_nset == 64) index[5] = a[12]; - - set_index = index.to_ulong(); - - } else { /* Else incorrect number of sets for the hashing function */ - assert( - "\nGPGPU-Sim cache configuration error: The number of sets should " - "be " - "32 or 64 for the hashing set index function.\n" && - 0); - } + case BITWISE_XORING_FUNCTION: { + new_addr_type higher_bits = addr >> (m_line_sz_log2 + m_nset_log2); + unsigned index = (addr >> m_line_sz_log2) & (m_nset - 1); + set_index = bitwise_hash_function(higher_bits, index, m_nset); break; - - case CUSTOM_SET_FUNCTION: + } + case HASH_IPOLY_FUNCTION: { + new_addr_type higher_bits = addr >> (m_line_sz_log2 + m_nset_log2); + unsigned index = (addr >> m_line_sz_log2) & (m_nset - 1); + set_index = ipoly_hash_function(higher_bits, index, m_nset); + break; + } + case CUSTOM_SET_FUNCTION: { /* No custom set function implemented */ break; + } - case LINEAR_SET_FUNCTION: + case LINEAR_SET_FUNCTION: { set_index = (addr >> m_line_sz_log2) & (m_nset - 1); break; + } - default: + default: { assert("\nUndefined set index function.\n" && 0); break; + } } // Linear function selected or custom set index function not implemented @@ -166,13 +154,14 @@ void l2_cache_config::init(linear_to_raw_address_translation *address_mapping) { } unsigned l2_cache_config::set_index(new_addr_type addr) const { - if (!m_address_mapping) { - return (addr >> m_line_sz_log2) & (m_nset - 1); - } else { + new_addr_type part_addr = addr; + + if (m_address_mapping) { // Calculate set index without memory partition bits to reduce set camping - new_addr_type part_addr = m_address_mapping->partition_address(addr); - return (part_addr >> m_line_sz_log2) & (m_nset - 1); + part_addr = m_address_mapping->partition_address(addr); } + + return cache_config::set_index(part_addr); } tag_array::~tag_array() { diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 2a37876..5c28b41 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -686,17 +686,11 @@ class cache_config { m_line_sz * m_nset * m_assoc, m_nset, m_assoc, m_line_sz); } - virtual unsigned set_index(new_addr_type addr) const { - if (m_set_index_function != LINEAR_SET_FUNCTION) { - printf( - "\nGPGPU-Sim cache configuration error: Hashing or " - "custom set index function selected in configuration " - "file for a cache that has not overloaded the set_index " - "function\n"); - abort(); - } - return (addr >> m_line_sz_log2) & (m_nset - 1); - } + virtual unsigned set_index(new_addr_type addr) const; + + unsigned hash_function(new_addr_type addr, unsigned m_nset, + unsigned m_line_sz_log2, unsigned m_nset_log2, + unsigned m_index_function) const; new_addr_type tag(new_addr_type addr) const { // For generality, the tag includes both index and tag. This allows for more @@ -793,10 +787,18 @@ class cache_config { class l1d_cache_config : public cache_config { public: l1d_cache_config() : cache_config() {} - virtual unsigned set_index(new_addr_type addr) const; unsigned set_bank(new_addr_type addr) const; + void init(char *config, FuncCache status) { + m_banks_byte_interleaving_log2 = LOGB2(l1_banks_byte_interleaving); + m_l1_banks_log2 = LOGB2(l1_banks); + cache_config::init(config, status); + } unsigned l1_latency; unsigned l1_banks; + unsigned m_l1_banks_log2; + unsigned l1_banks_byte_interleaving; + unsigned m_banks_byte_interleaving_log2; + unsigned l1_banks_hashing_function; }; class l2_cache_config : public cache_config { diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 03aebf3..1650688 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -252,6 +252,12 @@ void shader_core_config::reg_options(class OptionParser *opp) { option_parser_register(opp, "-gpgpu_l1_banks", OPT_UINT32, &m_L1D_config.l1_banks, "The number of L1 cache banks", "1"); + option_parser_register(opp, "-gpgpu_l1_banks_byte_interleaving", OPT_UINT32, + &m_L1D_config.l1_banks_byte_interleaving, + "l1 banks byte interleaving granularity", "32"); + option_parser_register(opp, "-gpgpu_l1_banks_hashing_function", OPT_UINT32, + &m_L1D_config.l1_banks_hashing_function, + "l1 banks hashing function", "0"); option_parser_register(opp, "-gpgpu_l1_latency", OPT_UINT32, &m_L1D_config.l1_latency, "L1 Hit Latency", "1"); option_parser_register(opp, "-gpgpu_smem_latency", OPT_UINT32, &smem_latency, -- cgit v1.3