From a2ba2f57e8a24b9dd6ec6f2568accbbf439a9dca Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 19 May 2021 19:39:48 -0400 Subject: updating config files with right adaptive cache parameters --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 18 ++++++++++-------- configs/tested-cfgs/SM7_QV100/gpgpusim.config | 19 ++++++++++--------- configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 17 +++++++++-------- configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 11 +++++++---- 4 files changed, 36 insertions(+), 29 deletions(-) diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index d7573ab..9e50fa3 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -99,19 +99,21 @@ # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo --gpgpu_adaptive_cache_config 0 --gpgpu_shmem_option 0,8,16,32,64,100 --gpgpu_unified_l1d_size 128 +-gpgpu_adaptive_cache_config 1 +-gpgpu_shmem_option 32,64 +-gpgpu_unified_l1d_size 96 +# L1 cache configuration -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:512,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_l1_latency 20 +-gpgpu_gmem_skip_L1D 0 +-gpgpu_n_cluster_ejection_buffer_size 32 +-gpgpu_flush_l1_cache 1 +# shared memory configuration -gpgpu_shmem_size 65536 -gpgpu_shmem_sizeDefault 65536 -gpgpu_shmem_per_block 65536 --gpgpu_gmem_skip_L1D 0 --gpgpu_n_cluster_ejection_buffer_size 32 --gpgpu_l1_latency 20 -gpgpu_smem_latency 20 --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 us 3MB L2 cache -gpgpu_cache:dl2 S:64:128:16,L:B:m:L:P,A:192:4,32:0,32 diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index 59c7f43..3750de0 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -94,7 +94,7 @@ -gpgpu_shmem_num_banks 32 -gpgpu_shmem_limited_broadcast 0 -gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 +-gpgpu_coalesce_arch 70 # Volta has four schedulers per core -gpgpu_num_sched_per_core 4 @@ -113,20 +113,21 @@ # 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 -gpgpu_adaptive_cache_config 1 -# Volta unified cache has four banks +-gpgpu_shmem_option 0,8,16,32,64,96 +-gpgpu_unified_l1d_size 128 +# L1 cache configuration -gpgpu_l1_banks 4 -gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_l1_cache_write_ratio 25 +-gpgpu_l1_latency 20 +-gpgpu_gmem_skip_L1D 0 +-gpgpu_flush_l1_cache 1 +-gpgpu_n_cluster_ejection_buffer_size 32 +# shared memory configuration -gpgpu_shmem_size 98304 -gpgpu_shmem_sizeDefault 98304 -gpgpu_shmem_per_block 65536 --gpgpu_gmem_skip_L1D 0 --gpgpu_n_cluster_ejection_buffer_size 32 --gpgpu_l1_latency 20 -gpgpu_smem_latency 20 --gpgpu_flush_l1_cache 1 --gpgpu_l1_cache_write_ratio 25 --gpgpu_shmem_option 0,12,24,48,96 --gpgpu_unified_l1d_size 128 # 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:P,A:192:4,32:0,32 diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index 32245d7..e7f7305 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -114,20 +114,21 @@ # 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 -gpgpu_adaptive_cache_config 1 -# Volta unified cache has four banks +-gpgpu_shmem_option 0,8,16,32,64,96 +-gpgpu_unified_l1d_size 128 +# L1 cache configuration -gpgpu_l1_banks 4 -gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_l1_cache_write_ratio 25 +-gpgpu_gmem_skip_L1D 0 +-gpgpu_l1_latency 20 +-gpgpu_flush_l1_cache 1 +-gpgpu_n_cluster_ejection_buffer_size 32 +# shared memory configuration -gpgpu_shmem_size 98304 -gpgpu_shmem_sizeDefault 98304 -gpgpu_shmem_per_block 65536 --gpgpu_gmem_skip_L1D 0 --gpgpu_n_cluster_ejection_buffer_size 32 --gpgpu_l1_latency 20 -gpgpu_smem_latency 20 --gpgpu_flush_l1_cache 1 --gpgpu_l1_cache_write_ratio 25 --gpgpu_shmem_option 0,12,24,48,96 --gpgpu_unified_l1d_size 128 # 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:P,A:192:4,32:0,32 diff --git a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config index f5418ad..3c0db06 100644 --- a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -107,17 +107,20 @@ # For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#global-memory-8-x # disable this mode in case of multi kernels/apps execution -gpgpu_adaptive_cache_config 1 +-gpgpu_shmem_option 0,8,16,32,64,100 +-gpgpu_unified_l1d_size 128 # Ampere unified cache has four banks -gpgpu_l1_banks 4 -gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_gmem_skip_L1D 0 +-gpgpu_l1_latency 20 +-gpgpu_n_cluster_ejection_buffer_size 32 +-gpgpu_flush_l1_cache 1 +# shared memory configuration -gpgpu_shmem_size 102400 -gpgpu_shmem_sizeDefault 102400 -gpgpu_shmem_per_block 102400 --gpgpu_gmem_skip_L1D 0 --gpgpu_n_cluster_ejection_buffer_size 32 --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 3MB L2 cache -gpgpu_cache:dl2 S:32:128:24,L:B:m:L:P,A:192:4,32:0,32 -- cgit v1.3 From b63d19a55c320b0bfd3ba4c80fe6f47a11bba39b Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 19 May 2021 19:41:22 -0400 Subject: updating config files --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 1 + configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 1 + configs/tested-cfgs/TITAN_V/gpgpusim.config | 173 +++++++++++++++++++++++ configs/tested-cfgs/TITAN_V/trace.config | 18 +++ 4 files changed, 193 insertions(+) create mode 100644 configs/tested-cfgs/TITAN_V/gpgpusim.config create mode 100644 configs/tested-cfgs/TITAN_V/trace.config diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index 9e50fa3..856f5cf 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -105,6 +105,7 @@ # L1 cache configuration -gpgpu_l1_banks 4 -gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_l1_cache_write_ratio 25 -gpgpu_l1_latency 20 -gpgpu_gmem_skip_L1D 0 -gpgpu_n_cluster_ejection_buffer_size 32 diff --git a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config index 3c0db06..9123e20 100644 --- a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -112,6 +112,7 @@ # Ampere unified cache has four banks -gpgpu_l1_banks 4 -gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_l1_cache_write_ratio 25 -gpgpu_gmem_skip_L1D 0 -gpgpu_l1_latency 20 -gpgpu_n_cluster_ejection_buffer_size 32 diff --git a/configs/tested-cfgs/TITAN_V/gpgpusim.config b/configs/tested-cfgs/TITAN_V/gpgpusim.config new file mode 100644 index 0000000..8b5cb20 --- /dev/null +++ b/configs/tested-cfgs/TITAN_V/gpgpusim.config @@ -0,0 +1,173 @@ +# 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 6745 +-gpgpu_TB_launch_latency 0 + +# Compute Capability +-gpgpu_compute_capability_major 7 +-gpgpu_compute_capability_minor 0 + +# PTX execution-driven +-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 24 +-gpgpu_n_sub_partition_per_mchannel 2 + +# clock domains +#-gpgpu_clock_domains ::: +-gpgpu_clock_domains 1200:1200:1200:850 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_registers_per_block 65536 +-gpgpu_occupancy_sm_number 70 + +-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 +-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 4,4,4,4,21 +-ptx_opcode_initiation_int 2,2,2,2,2 +-ptx_opcode_latency_fp 4,4,4,4,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,8,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 21 +-ptx_opcode_initiation_sfu 8 +-ptx_opcode_latency_tesnor 35 +-ptx_opcode_initiation_tensor 32 + +# sub core model: in which each scheduler has its own register file and EUs +# i.e. schedulers are isolated +-gpgpu_sub_core_model 1 +# disable specialized operand collectors and use generic operand collectors instead +-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 +# register banks +-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 70 + +# warp scheduling +-gpgpu_num_sched_per_core 4 +-gpgpu_scheduler gto +# a warp scheduler issue mode +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +## L1/shared memory configuration +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# In adaptive cache, we adaptively assign the remaining shared memory to L1 cache +-gpgpu_adaptive_cache_config 1 +-gpgpu_l1_banks 4 +-gpgpu_cache:dl1 S:4:128:64,L:L:m:N:L,A:512:64,16:0,32 +-gpgpu_shmem_size 98304 +-gpgpu_shmem_sizeDefault 98304 +-gpgpu_shmem_per_block 49152 +-gpgpu_gmem_skip_L1D 0 +-gpgpu_n_cluster_ejection_buffer_size 32 +-gpgpu_l1_latency 33 +-gpgpu_smem_latency 27 +-gpgpu_flush_l1_cache 1 + +# L2 cache +-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 +-gpgpu_memory_partition_indexing 0 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +-gpgpu_inst_fetch_throughput 4 +# 128 KB Tex +# Note, TEX is deprected since 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 +-gpgpu_perfect_inst_const_cache 1 + +# interconnection +# 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 +-icnt_arbiter_algo 1 + +# memory partition latency config +-gpgpu_l2_rop_latency 177 +-dram_latency 103 + +# dram sched config +-gpgpu_dram_scheduler 1 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# dram model config +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Mem timing +-gpgpu_dram_timing_opt nbk=16:CCD=1:RRD=4:RCD=12:RAS=29:RP=12:RC=40:CL=12:WL=2:CDLR=3:WR=11:nbkgrp=4:CCDL=2:RTPL=4 +-dram_dual_bus_interface 1 + +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-dram_seperate_write_queue_enable 1 +#-dram_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 +-power_simulation_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/tested-cfgs/TITAN_V/trace.config b/configs/tested-cfgs/TITAN_V/trace.config new file mode 100644 index 0000000..6e193f7 --- /dev/null +++ b/configs/tested-cfgs/TITAN_V/trace.config @@ -0,0 +1,18 @@ +-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 21,8 +-trace_opcode_latency_initiation_tensor 2,2 + +#execute branch insts on spec unit 1 +#,,,,, +-specialized_unit_1 1,4,4,4,4,BRA +-trace_opcode_latency_initiation_spec_op_1 4,4 + +#TEX unit, make fixed latency for all tex insts +-specialized_unit_2 1,4,200,4,4,TEX +-trace_opcode_latency_initiation_spec_op_2 200,4 + +#tensor unit +-specialized_unit_3 1,4,2,4,4,TENSOR +-trace_opcode_latency_initiation_spec_op_3 2,2 -- cgit v1.3 From e3d186bbeade78dec776989ccec2a0c0aea27fb4 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 19 May 2021 19:43:29 -0400 Subject: chaning @sets to 4 based on recent ubenchs --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 2 +- configs/tested-cfgs/SM7_QV100/gpgpusim.config | 2 +- configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 2 +- configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 2 +- 4 files changed, 4 insertions(+), 4 deletions(-) diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index 856f5cf..a63d50f 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -104,7 +104,7 @@ -gpgpu_unified_l1d_size 96 # L1 cache configuration -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_cache:dl1 S:4:128:64,L:L:m:N:L,A:512:8,16:0,32 -gpgpu_l1_cache_write_ratio 25 -gpgpu_l1_latency 20 -gpgpu_gmem_skip_L1D 0 diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index 3750de0..47bf1c8 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -117,7 +117,7 @@ -gpgpu_unified_l1d_size 128 # L1 cache configuration -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_cache:dl1 S:4:128:64,L:L:m:N:L,A:512:8,16:0,32 -gpgpu_l1_cache_write_ratio 25 -gpgpu_l1_latency 20 -gpgpu_gmem_skip_L1D 0 diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index e7f7305..3db64b3 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -118,7 +118,7 @@ -gpgpu_unified_l1d_size 128 # L1 cache configuration -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_cache:dl1 S:4:128:64,L:L:m:N:L,A:512:8,16:0,32 -gpgpu_l1_cache_write_ratio 25 -gpgpu_gmem_skip_L1D 0 -gpgpu_l1_latency 20 diff --git a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config index 9123e20..c70cfe8 100644 --- a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -111,7 +111,7 @@ -gpgpu_unified_l1d_size 128 # Ampere unified cache has four banks -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_cache:dl1 S:4:128:64,L:L:m:N:L,A:512:8,16:0,32 -gpgpu_l1_cache_write_ratio 25 -gpgpu_gmem_skip_L1D 0 -gpgpu_l1_latency 20 -- cgit v1.3 From 24ffab25f41d76b94fd2012a8897312a73a7165f Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 19 May 2021 21:17:40 -0400 Subject: moving shmem option to the base class and change the code to accept turing config --- src/abstract_hardware_model.h | 1 + src/gpgpu-sim/gpu-cache.h | 3 +-- src/gpgpu-sim/shader.cc | 46 +++++++++---------------------------------- src/gpgpu-sim/shader.h | 26 ++++++++++++++++++++++++ 4 files changed, 37 insertions(+), 39 deletions(-) diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 17a1cec..b33c50b 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -374,6 +374,7 @@ class core_config { unsigned mem_warp_parts; mutable unsigned gpgpu_shmem_size; char *gpgpu_shmem_option; + std::vector shmem_opt_list; unsigned gpgpu_shmem_sizeDefault; unsigned gpgpu_shmem_sizePrefL1; unsigned gpgpu_shmem_sizePrefShared; diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index d801528..26ed621 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -577,6 +577,7 @@ class cache_config { } // set * assoc * cacheline size. Then convert Byte to KB + // gpgpu_unified_cache_size is in KB while original_sz is in B unsigned original_size = m_nset * m_assoc * m_line_sz / 1024; if (m_unified_cache_size > 0) { max_cache_multiplier = m_unified_cache_size / original_size; @@ -785,12 +786,10 @@ class cache_config { } unsigned get_max_num_lines() const { assert(m_valid); - // gpgpu_unified_cache_size is in KB while original_sz is in B return max_cache_multiplier * m_nset * original_m_assoc; } unsigned get_max_assoc() const { assert(m_valid); - // gpgpu_unified_cache_size is in KB while original_sz is in B return max_cache_multiplier * original_m_assoc; } void print(FILE *fp) const { diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 75fbe16..bc747d6 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -3334,56 +3334,28 @@ unsigned int shader_core_config::max_cta(const kernel_info_t &k) const { if (adaptive_cache_config && !k.cache_config_set) { // For more info about adaptive cache, see // https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x - std::vector shmem_list; - for (unsigned i = 0; i < strlen(gpgpu_shmem_option); i++) { - char option[4]; - int j = 0; - while (gpgpu_shmem_option[i] != ',' && i < strlen(gpgpu_shmem_option)) { - if (gpgpu_shmem_option[i] == ' ') { - // skip spaces - i++; - } else { - if (!isdigit(gpgpu_shmem_option[i])) { - // check for non digits, which should not be here - assert(0 && "invalid config: -gpgpu_shmem_option"); - } - option[j] = gpgpu_shmem_option[i]; - j++; - i++; - } - } - // convert KB -> B - shmem_list.push_back((unsigned)atoi(option) * 1024); - } - unsigned total_shmem = kernel_info->smem * result; // Unified cache config is in KB. Converting to B unsigned total_unified = m_L1D_config.m_unified_cache_size * 1024; - std::sort(shmem_list.begin(), shmem_list.end()); - assert(total_shmem >= 0 && total_shmem <= shmem_list.back()); + assert(total_shmem >= 0 && total_shmem <= shmem_opt_list.back()); switch (adaptive_cache_config) { case FIXED: break; case ADAPTIVE_CACHE: { - // For more info about adaptive cache, see bool l1d_configured = false; unsigned max_assoc = m_L1D_config.get_max_assoc(); - if (total_shmem == 0) { - m_L1D_config.set_assoc(max_assoc); - l1d_configured = true; - } else { - for (std::vector::iterator it = shmem_list.begin(); - it < shmem_list.end() - 1; it++) { - if (total_shmem > *it && total_shmem <= *(it + 1)) { - float l1_ratio = 1 - (float) *(it + 1) / total_unified; - m_L1D_config.set_assoc(max_assoc * l1_ratio); - l1d_configured = true; - break; - } + for (std::vector::const_iterator it = shmem_opt_list.begin(); + it < shmem_opt_list.end(); it++) { + if (total_shmem <= *it) { + float l1_ratio = 1 - ((float) *(it) / total_unified); + m_L1D_config.set_assoc(max_assoc * l1_ratio); + l1d_configured = true; + break; } } + assert(l1d_configured && "no shared memory option found"); break; } diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index a7a2c02..42bbdcb 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1495,6 +1495,32 @@ class shader_core_config : public core_config { } else break; // we only accept continuous specialized_units, i.e., 1,2,3,4 } + + //parse gpgpu_shmem_option for adpative cache config + if(adaptive_cache_config) { + for (unsigned i = 0; i < strlen(gpgpu_shmem_option); i++) { + char option[4]; + int j = 0; + while (gpgpu_shmem_option[i] != ',' && i < strlen(gpgpu_shmem_option)) { + if (gpgpu_shmem_option[i] == ' ') { + // skip spaces + i++; + } else { + if (!isdigit(gpgpu_shmem_option[i])) { + // check for non digits, which should not be here + assert(0 && "invalid config: -gpgpu_shmem_option"); + } + option[j] = gpgpu_shmem_option[i]; + j++; + i++; + } + } + // convert KB -> B + shmem_opt_list.push_back((unsigned)atoi(option) * 1024); + } + std::sort(shmem_opt_list.begin(), shmem_opt_list.end()); + } + } void reg_options(class OptionParser *opp); unsigned max_cta(const kernel_info_t &k) const; -- cgit v1.3 From fedcde3789f7921647caee184c0fa104403c848d Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 19 May 2021 21:42:29 -0400 Subject: moving the unified size from the base class config to l1 config --- src/gpgpu-sim/gpu-cache.h | 30 ++++++++++++++++-------------- src/gpgpu-sim/shader.cc | 3 ++- 2 files changed, 18 insertions(+), 15 deletions(-) diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 26ed621..8bd62da 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -575,15 +575,6 @@ class cache_config { } exit_parse_error(); } - - // set * assoc * cacheline size. Then convert Byte to KB - // gpgpu_unified_cache_size is in KB while original_sz is in B - unsigned original_size = m_nset * m_assoc * m_line_sz / 1024; - if (m_unified_cache_size > 0) { - max_cache_multiplier = m_unified_cache_size / original_size; - } else { - max_cache_multiplier = MAX_DEFAULT_CACHE_SIZE_MULTIBLIER; - } switch (ct) { case 'N': @@ -694,7 +685,6 @@ class cache_config { m_sector_sz_log2 = LOGB2(SECTOR_SIZE); original_m_assoc = m_assoc; - // For more details about difference between FETCH_ON_WRITE and WRITE // VALIDAE policies Read: Jouppi, Norman P. "Cache write policies and // performance". ISCA 93. WRITE_ALLOCATE is the old write policy in @@ -786,11 +776,11 @@ class cache_config { } unsigned get_max_num_lines() const { assert(m_valid); - return max_cache_multiplier * m_nset * original_m_assoc; + return get_max_cache_multiplier() * m_nset * original_m_assoc; } unsigned get_max_assoc() const { assert(m_valid); - return max_cache_multiplier * original_m_assoc; + return get_max_cache_multiplier() * original_m_assoc; } void print(FILE *fp) const { fprintf(fp, "Size = %d B (%d Set x %d-way x %d byte line)\n", @@ -799,6 +789,8 @@ class cache_config { virtual unsigned set_index(new_addr_type addr) const; + virtual unsigned get_max_cache_multiplier() const { return MAX_DEFAULT_CACHE_SIZE_MULTIBLIER;} + unsigned hash_function(new_addr_type addr, unsigned m_nset, unsigned m_line_sz_log2, unsigned m_nset_log2, unsigned m_index_function) const; @@ -840,7 +832,6 @@ class cache_config { char *m_config_stringPrefL1; char *m_config_stringPrefShared; FuncCache cache_status; - unsigned m_unified_cache_size; unsigned m_wr_percent; write_allocate_policy_t get_write_allocate_policy() { return m_write_alloc_policy; @@ -867,7 +858,6 @@ class cache_config { unsigned m_sector_sz_log2; unsigned original_m_assoc; bool m_is_streaming; - unsigned max_cache_multiplier; enum replacement_policy_t m_replacement_policy; // 'L' = LRU, 'F' = FIFO enum write_policy_t @@ -922,6 +912,18 @@ class l1d_cache_config : public cache_config { unsigned l1_banks_byte_interleaving; unsigned l1_banks_byte_interleaving_log2; unsigned l1_banks_hashing_function; + unsigned m_unified_cache_size; + virtual unsigned get_max_cache_multiplier() const { + // set * assoc * cacheline size. Then convert Byte to KB + // gpgpu_unified_cache_size is in KB while original_sz is in B + if (m_unified_cache_size > 0) { + unsigned original_size = m_nset * original_m_assoc * m_line_sz / 1024; + assert(m_unified_cache_size % original_size == 0); + return m_unified_cache_size / original_size; + } else { + return MAX_DEFAULT_CACHE_SIZE_MULTIBLIER; + } + } }; class l2_cache_config : public cache_config { diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index bc747d6..7f27b7b 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -3335,10 +3335,11 @@ unsigned int shader_core_config::max_cta(const kernel_info_t &k) const { // For more info about adaptive cache, see // https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x unsigned total_shmem = kernel_info->smem * result; + assert(total_shmem >= 0 && total_shmem <= shmem_opt_list.back()); + // Unified cache config is in KB. Converting to B unsigned total_unified = m_L1D_config.m_unified_cache_size * 1024; - assert(total_shmem >= 0 && total_shmem <= shmem_opt_list.back()); switch (adaptive_cache_config) { case FIXED: break; -- cgit v1.3 From 8aee56d7401af9a91a1de3adae1b61329e0d30e5 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 19 May 2021 22:10:53 -0400 Subject: rename set_dirty_byte_mask --- src/gpgpu-sim/gpu-cache.h | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 8bd62da..91cde7e 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -210,13 +210,13 @@ struct line_cache_block : public cache_block_t { m_status = status; } virtual void set_byte_mask(mem_fetch *mf) { - m_byte_mask = m_byte_mask | mf->get_access_byte_mask(); + m_dirty_byte_mask = m_dirty_byte_mask | mf->get_access_byte_mask(); } virtual void set_byte_mask(mem_access_byte_mask_t byte_mask) { - m_byte_mask = m_byte_mask | byte_mask; + m_dirty_byte_mask = m_dirty_byte_mask | byte_mask; } virtual mem_access_byte_mask_t get_dirty_byte_mask() { - return m_byte_mask; + return m_dirty_byte_mask; } virtual mem_access_sector_mask_t get_dirty_sector_mask() { mem_access_sector_mask_t sector_mask; @@ -270,7 +270,7 @@ struct line_cache_block : public cache_block_t { bool m_set_readable_on_fill; bool m_set_byte_mask_on_fill; bool m_readable; - mem_access_byte_mask_t m_byte_mask; + mem_access_byte_mask_t m_dirty_byte_mask; }; struct sector_cache_block : public cache_block_t { @@ -290,7 +290,7 @@ struct sector_cache_block : public cache_block_t { m_line_alloc_time = 0; m_line_last_access_time = 0; m_line_fill_time = 0; - m_byte_mask.reset(); + m_dirty_byte_mask.reset(); } virtual void allocate(new_addr_type tag, new_addr_type block_addr, @@ -405,13 +405,13 @@ struct sector_cache_block : public cache_block_t { } virtual void set_byte_mask(mem_fetch *mf) { - m_byte_mask = m_byte_mask | mf->get_access_byte_mask(); + m_dirty_byte_mask = m_dirty_byte_mask | mf->get_access_byte_mask(); } virtual void set_byte_mask(mem_access_byte_mask_t byte_mask) { - m_byte_mask = m_byte_mask | byte_mask; + m_dirty_byte_mask = m_dirty_byte_mask | byte_mask; } virtual mem_access_byte_mask_t get_dirty_byte_mask() { - return m_byte_mask; + return m_dirty_byte_mask; } virtual mem_access_sector_mask_t get_dirty_sector_mask() { mem_access_sector_mask_t sector_mask; @@ -492,7 +492,7 @@ struct sector_cache_block : public cache_block_t { bool m_set_readable_on_fill[SECTOR_CHUNCK_SIZE]; bool m_set_byte_mask_on_fill; bool m_readable[SECTOR_CHUNCK_SIZE]; - mem_access_byte_mask_t m_byte_mask; + mem_access_byte_mask_t m_dirty_byte_mask; unsigned get_sector_index(mem_access_sector_mask_t sector_mask) { assert(sector_mask.count() == 1); -- cgit v1.3 From b466afea67e6d6faf49f01ecfe378257fbdb93af Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 19 May 2021 22:20:04 -0400 Subject: eliminate redundant code in gpu-cache.h --- src/gpgpu-sim/gpu-cache.h | 10 ---------- 1 file changed, 10 deletions(-) diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 91cde7e..6698d92 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -596,16 +596,6 @@ class cache_config { default: exit_parse_error(); } - switch (rp) { - case 'L': - m_replacement_policy = LRU; - break; - case 'F': - m_replacement_policy = FIFO; - break; - default: - exit_parse_error(); - } switch (wp) { case 'R': m_write_policy = READ_ONLY; -- cgit v1.3 From 7fac247e3e1c4326081c3ea4d46da6c5dc83ccb9 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 19 May 2021 22:20:56 -0400 Subject: change L1 cache config in Volta+ to be write-through and write-allocate based on recent ubench --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 2 +- configs/tested-cfgs/SM7_QV100/gpgpusim.config | 2 +- configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 2 +- configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 3 +-- 4 files changed, 4 insertions(+), 5 deletions(-) diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index a63d50f..f715f3a 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -104,7 +104,7 @@ -gpgpu_unified_l1d_size 96 # L1 cache configuration -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:4:128:64,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_cache:dl1 S:4:128:64,L:T:m:L:L,A:512:8,16:0,32 -gpgpu_l1_cache_write_ratio 25 -gpgpu_l1_latency 20 -gpgpu_gmem_skip_L1D 0 diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index 47bf1c8..5f22a42 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -117,7 +117,7 @@ -gpgpu_unified_l1d_size 128 # L1 cache configuration -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:4:128:64,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_cache:dl1 S:4:128:64,L:T:m:L:L,A:512:8,16:0,32 -gpgpu_l1_cache_write_ratio 25 -gpgpu_l1_latency 20 -gpgpu_gmem_skip_L1D 0 diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index 3db64b3..c44563f 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -118,7 +118,7 @@ -gpgpu_unified_l1d_size 128 # L1 cache configuration -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:4:128:64,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_cache:dl1 S:4:128:64,L:T:m:L:L,A:512:8,16:0,32 -gpgpu_l1_cache_write_ratio 25 -gpgpu_gmem_skip_L1D 0 -gpgpu_l1_latency 20 diff --git a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config index c70cfe8..02cdb9e 100644 --- a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -101,7 +101,6 @@ ## L1/shared memory configuration # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo -# Default config is 28KB DL1 and 100KB shared memory # In Ampere, 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#global-memory-8-x @@ -111,7 +110,7 @@ -gpgpu_unified_l1d_size 128 # Ampere unified cache has four banks -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:4:128:64,L:L:m:N:L,A:512:8,16:0,32 +-gpgpu_cache:dl1 S:4:128:64,L:T:m:L:L,A:512:8,16:0,32 -gpgpu_l1_cache_write_ratio 25 -gpgpu_gmem_skip_L1D 0 -gpgpu_l1_latency 20 -- cgit v1.3 From 0d33266ff6ca9b880dff40f6338c8a5cae696438 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 19 May 2021 22:25:37 -0400 Subject: oops delete this config, it should not be pushed --- configs/tested-cfgs/TITAN_V/gpgpusim.config | 173 ---------------------------- configs/tested-cfgs/TITAN_V/trace.config | 18 --- 2 files changed, 191 deletions(-) delete mode 100644 configs/tested-cfgs/TITAN_V/gpgpusim.config delete mode 100644 configs/tested-cfgs/TITAN_V/trace.config diff --git a/configs/tested-cfgs/TITAN_V/gpgpusim.config b/configs/tested-cfgs/TITAN_V/gpgpusim.config deleted file mode 100644 index 8b5cb20..0000000 --- a/configs/tested-cfgs/TITAN_V/gpgpusim.config +++ /dev/null @@ -1,173 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 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 6745 --gpgpu_TB_launch_latency 0 - -# Compute Capability --gpgpu_compute_capability_major 7 --gpgpu_compute_capability_minor 0 - -# PTX execution-driven --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 24 --gpgpu_n_sub_partition_per_mchannel 2 - -# clock domains -#-gpgpu_clock_domains ::: --gpgpu_clock_domains 1200:1200:1200:850 - -# shader core pipeline config --gpgpu_shader_registers 65536 --gpgpu_registers_per_block 65536 --gpgpu_occupancy_sm_number 70 - --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 --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 4,4,4,4,21 --ptx_opcode_initiation_int 2,2,2,2,2 --ptx_opcode_latency_fp 4,4,4,4,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,8,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 21 --ptx_opcode_initiation_sfu 8 --ptx_opcode_latency_tesnor 35 --ptx_opcode_initiation_tensor 32 - -# sub core model: in which each scheduler has its own register file and EUs -# i.e. schedulers are isolated --gpgpu_sub_core_model 1 -# disable specialized operand collectors and use generic operand collectors instead --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 -# register banks --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 70 - -# warp scheduling --gpgpu_num_sched_per_core 4 --gpgpu_scheduler gto -# a warp scheduler issue mode --gpgpu_max_insn_issue_per_warp 1 --gpgpu_dual_issue_diff_exec_units 1 - -## L1/shared memory configuration -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# In adaptive cache, we adaptively assign the remaining shared memory to L1 cache --gpgpu_adaptive_cache_config 1 --gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:4:128:64,L:L:m:N:L,A:512:64,16:0,32 --gpgpu_shmem_size 98304 --gpgpu_shmem_sizeDefault 98304 --gpgpu_shmem_per_block 49152 --gpgpu_gmem_skip_L1D 0 --gpgpu_n_cluster_ejection_buffer_size 32 --gpgpu_l1_latency 33 --gpgpu_smem_latency 27 --gpgpu_flush_l1_cache 1 - -# L2 cache --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 --gpgpu_memory_partition_indexing 0 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 --gpgpu_inst_fetch_throughput 4 -# 128 KB Tex -# Note, TEX is deprected since 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 --gpgpu_perfect_inst_const_cache 1 - -# interconnection -# 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 --icnt_arbiter_algo 1 - -# memory partition latency config --gpgpu_l2_rop_latency 177 --dram_latency 103 - -# dram sched config --gpgpu_dram_scheduler 1 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# dram model config --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Mem timing --gpgpu_dram_timing_opt nbk=16:CCD=1:RRD=4:RCD=12:RAS=29:RP=12:RC=40:CL=12:WL=2:CDLR=3:WR=11:nbkgrp=4:CCDL=2:RTPL=4 --dram_dual_bus_interface 1 - -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-dram_seperate_write_queue_enable 1 -#-dram_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 --power_simulation_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/tested-cfgs/TITAN_V/trace.config b/configs/tested-cfgs/TITAN_V/trace.config deleted file mode 100644 index 6e193f7..0000000 --- a/configs/tested-cfgs/TITAN_V/trace.config +++ /dev/null @@ -1,18 +0,0 @@ --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 21,8 --trace_opcode_latency_initiation_tensor 2,2 - -#execute branch insts on spec unit 1 -#,,,,, --specialized_unit_1 1,4,4,4,4,BRA --trace_opcode_latency_initiation_spec_op_1 4,4 - -#TEX unit, make fixed latency for all tex insts --specialized_unit_2 1,4,200,4,4,TEX --trace_opcode_latency_initiation_spec_op_2 200,4 - -#tensor unit --specialized_unit_3 1,4,2,4,4,TENSOR --trace_opcode_latency_initiation_spec_op_3 2,2 -- cgit v1.3