From b430b36911b48228ed7eb77457cc378261151a13 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Thu, 25 Feb 2021 16:25:43 -0500 Subject: adding new RTX 3070 config --- .../SM86_RTX3070/config_ampere_islip.icnt | 74 ++++++++ configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 192 +++++++++++++++++++++ 2 files changed, 266 insertions(+) create mode 100644 configs/tested-cfgs/SM86_RTX3070/config_ampere_islip.icnt create mode 100644 configs/tested-cfgs/SM86_RTX3070/gpgpusim.config (limited to 'configs') diff --git a/configs/tested-cfgs/SM86_RTX3070/config_ampere_islip.icnt b/configs/tested-cfgs/SM86_RTX3070/config_ampere_islip.icnt new file mode 100644 index 0000000..6775d5d --- /dev/null +++ b/configs/tested-cfgs/SM86_RTX3070/config_ampere_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 = 78; +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/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config new file mode 100644 index 0000000..2010aa6 --- /dev/null +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -0,0 +1,192 @@ +# This config models the Ampere RTX 3070 +# For more info about Ampere architecture: +# https://developer.download.nvidia.com/video/gputechconf/gtc/2020/presentations/s21730-inside-the-nvidia-ampere-architecture.pdf +# https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/ampere/pdf/NVIDIA-ampere-GA102-GPU-Architecture-Whitepaper-V1.pdf +# https://en.wikipedia.org/wiki/GeForce_30_series +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 86 + +# 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 +-gpgpu_TB_launch_latency 0 + +# Compute Capability +-gpgpu_compute_capability_major 8 +-gpgpu_compute_capability_minor 6 + +# PTX execution-driven +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 46 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 16 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Ampere clock domains +#-gpgpu_clock_domains ::: +-gpgpu_clock_domains 1320.0:1320.0:1320.0:3500.0 +# boost mode +# -gpgpu_clock_domains 1780.0:1780.0:1780.0:3500.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_registers_per_block 65536 +-gpgpu_occupancy_sm_number 86 + +# 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 +## Ampere GA102 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 4,13,4,5,145,21 +-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 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 +-ptx_opcode_latency_tesnor 32 +-ptx_opcode_initiation_tensor 32 + +# Ampere 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 +# 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 +# Ampere has 24 double-ported banks, 4 schedulers, 6 banks per scheduler +-gpgpu_num_reg_banks 24 +-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 86 + +# Ampere has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Greedy then oldest scheduler +-gpgpu_scheduler gto +## In Ampere, a warp scheduler can issue 1 inst per cycle +-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 +# 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 +# disable this mode in case of multi kernels/apps execution +-gpgpu_adaptive_cache_config 1 +# Ampere unified cache has four banks +-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 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 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-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 +-gpgpu_inst_fetch_throughput 4 +# 128 KB Tex +# Note, TEX is deprecated, 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 +#-network_mode 1 +#-inter_config_file config_ampere_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 +-icnt_arbiter_algo 1 + +# memory partition latency config +-gpgpu_l2_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 + +# Ampere RTX3060 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 +-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 + +#-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 for Ampere +-power_simulation_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + -- cgit v1.3 From 09f10eb4c28b6cc76c7c7cc3181c340cf8ec2be5 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Thu, 25 Mar 2021 12:44:09 -0400 Subject: change the L1 cache policy to be on-miss based on recent ubench --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 2 +- configs/tested-cfgs/SM7_QV100/gpgpusim.config | 2 +- configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) (limited to 'configs') diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index 6fe04ee..8be9a73 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -101,7 +101,7 @@ # ** Optional parameter - Required when mshr_type==Texture Fifo -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_cache:dl1 S:1:128:512,L:L:m:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 65536 -gpgpu_shmem_sizeDefault 65536 -gpgpu_shmem_per_block 65536 diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index c4818d1..18f5564 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -115,7 +115,7 @@ -gpgpu_adaptive_cache_config 1 # Volta unified cache has four banks -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:256,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 98304 -gpgpu_shmem_sizeDefault 98304 -gpgpu_shmem_per_block 65536 diff --git a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config index 2010aa6..11dbcaf 100644 --- a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -109,7 +109,7 @@ -gpgpu_adaptive_cache_config 1 # Ampere unified cache has four banks -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:256,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:256:8,16:0,32 -gpgpu_shmem_size 102400 -gpgpu_shmem_sizeDefault 102400 -gpgpu_shmem_per_block 102400 -- cgit v1.3 From 1ee03f0116511ac3c2d6ac7688d916191f4f0a6b Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Thu, 25 Mar 2021 12:54:14 -0400 Subject: change the L1 cache policy based on recent ubench --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 2 +- configs/tested-cfgs/SM7_QV100/gpgpusim.config | 2 +- configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) (limited to 'configs') diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index 8be9a73..6189dca 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -101,7 +101,7 @@ # ** Optional parameter - Required when mshr_type==Texture Fifo -gpgpu_adaptive_cache_config 0 -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:512,L:L:m:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1 S:1:128:512,L:L:m:N:L,A:512:8,16:0,32 -gpgpu_shmem_size 65536 -gpgpu_shmem_sizeDefault 65536 -gpgpu_shmem_per_block 65536 diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index 18f5564..bc5677c 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -115,7 +115,7 @@ -gpgpu_adaptive_cache_config 1 # Volta unified cache has four banks -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 -gpgpu_shmem_size 98304 -gpgpu_shmem_sizeDefault 98304 -gpgpu_shmem_per_block 65536 diff --git a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config index 11dbcaf..f5418ad 100644 --- a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -109,7 +109,7 @@ -gpgpu_adaptive_cache_config 1 # Ampere unified cache has four banks -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 -gpgpu_shmem_size 102400 -gpgpu_shmem_sizeDefault 102400 -gpgpu_shmem_per_block 102400 -- cgit v1.3 From f2a7d9ce6cd13977d97a0601d732551a5451ac71 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Sat, 15 May 2021 09:09:20 -0400 Subject: fixing streaming cache based on recent ubench --- configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 2 +- src/gpgpu-sim/gpu-cache.cc | 13 --------- src/gpgpu-sim/gpu-cache.h | 38 ++++++++++++++++---------- src/gpgpu-sim/shader.cc | 15 ++++++++++ 4 files changed, 39 insertions(+), 29 deletions(-) (limited to 'configs') diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index 3fa51ee..3af314c 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -116,7 +116,7 @@ -gpgpu_adaptive_cache_config 1 # Volta unified cache has four banks -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:1:128:256,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1 S:1:128:256,L:L:m:N:L,A:512:8,16:0,32 -gpgpu_shmem_size 98304 -gpgpu_shmem_sizeDefault 98304 -gpgpu_shmem_per_block 65536 diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 1c36d22..c6a125d 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -312,15 +312,6 @@ enum cache_request_status tag_array::probe(new_addr_type addr, unsigned &idx, abort(); // if an unreserved block exists, it is either invalid or // replaceable - if (probe_mode && m_config.is_streaming()) { - line_table::const_iterator i = - pending_lines.find(m_config.block_addr(addr)); - assert(mf); - if (!mf->is_write() && i != pending_lines.end()) { - if (i->second != mf->get_inst().get_uid()) return SECTOR_MISS; - } - } - return MISS; } @@ -1060,7 +1051,6 @@ void baseline_cache::fill(mem_fetch *mf, unsigned time) { m_tag_array->fill(e->second.m_cache_index, time, mf); else if (m_config.m_alloc_policy == ON_FILL) { m_tag_array->fill(e->second.m_block_addr, time, mf); - if (m_config.is_streaming()) m_tag_array->remove_pending_line(mf); } else abort(); bool has_atomic = false; @@ -1136,9 +1126,6 @@ void baseline_cache::send_read_request(new_addr_type addr, m_tag_array->access(block_addr, time, cache_index, wb, evicted, mf); m_mshrs.add(mshr_addr, mf); - if (m_config.is_streaming() && m_config.m_cache_type == SECTOR) { - m_tag_array->add_pending_line(mf); - } m_extra_mf_fields[mf] = extra_mf_fields( mshr_addr, mf->get_addr(), cache_index, mf->get_data_size(), m_config); mf->set_data_size(m_config.get_atom_sz()); diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 00c09ae..aa0a7e8 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -574,22 +574,26 @@ class cache_config { exit_parse_error(); } if (m_alloc_policy == STREAMING) { - // For streaming cache, we set the alloc policy to be on-fill to remove - // all line_alloc_fail stalls we set the MSHRs to be equal to max - // allocated cache lines. This is possible by moving TAG to be shared - // between cache line and MSHR enrty (i.e. for each cache line, there is - // an MSHR rntey associated with it) This is the easiest think we can - // think about to model (mimic) L1 streaming cache in Pascal and Volta - // Based on our microbenchmakrs, MSHRs entries have been increasing - // substantially in Pascal and Volta For more information about streaming - // cache, see: - // http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - // https://ieeexplore.ieee.org/document/8344474/ + /* + For streaming cache: + (1) we set the alloc policy to be on-fill to remove all line_alloc_fail stalls. + if the whole memory is allocated to the L1 cache, then make the allocation to be on_MISS + otherwise, make it ON_FILL to eliminate line allocation fails. + i.e. MSHR throughput is the same, independent on the L1 cache size/associativity + So, we set the allocation policy per kernel basis, see shader.cc, max_cta() function + + (2) We also set the MSHRs to be equal to max + allocated cache lines. This is possible by moving TAG to be shared + between cache line and MSHR enrty (i.e. for each cache line, there is + an MSHR rntey associated with it). This is the easiest think we can + think of to model (mimic) L1 streaming cache in Pascal and Volta + + For more information about streaming cache, see: + http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + https://ieeexplore.ieee.org/document/8344474/ + */ m_is_streaming = true; m_alloc_policy = ON_FILL; - m_mshr_entries = m_nset * m_assoc * MAX_DEFAULT_CACHE_SIZE_MULTIBLIER; - if (m_cache_type == SECTOR) m_mshr_entries *= SECTOR_CHUNCK_SIZE; - m_mshr_max_merge = MAX_WARP_PER_SM; } switch (mshr_type) { case 'F': @@ -638,7 +642,8 @@ class cache_config { } // detect invalid configuration - if (m_alloc_policy == ON_FILL and m_write_policy == WRITE_BACK) { + if ((m_alloc_policy == ON_FILL || m_alloc_policy == STREAMING) + and m_write_policy == WRITE_BACK) { // A writeback cache with allocate-on-fill policy will inevitably lead to // deadlock: The deadlock happens when an incoming cache-fill evicts a // dirty line, generating a writeback request. If the memory subsystem is @@ -750,6 +755,9 @@ class cache_config { } bool is_streaming() { return m_is_streaming; } FuncCache get_cache_status() { return cache_status; } + void set_allocation_policy(enum allocation_policy_t alloc) { + m_alloc_policy = alloc; + } char *m_config_string; char *m_config_stringPrefL1; char *m_config_stringPrefShared; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index c6e7b8f..0ad9547 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -3308,6 +3308,21 @@ unsigned int shader_core_config::max_cta(const kernel_info_t &k) const { m_L1D_config.get_total_size_inKB()); } + if(m_L1D_config.is_streaming()) { + //for streaming cache, if the whole memory is allocated + //to the L1 cache, then make the allocation to be on_MISS + //otherwise, make it ON_FILL to eliminate line allocation fails + //i.e. MSHR throughput is the same, independent on the L1 cache size/associativity + if(total_shmed == 0) { + m_L1D_config.set_allocation_policy(ON_MISS); + printf("GPGPU-Sim: Reconfigure L1 allocation to ON_MISS\n"); + } + else { + m_L1D_config.set_allocation_policy(ON_FILL); + printf("GPGPU-Sim: Reconfigure L1 allocation to ON_FILL\n"); + } + } + k.cache_config_set = true; } -- cgit v1.3 From f7833519471ce92619bd1e4807ec07eb55aed76e Mon Sep 17 00:00:00 2001 From: JRPAN <25518778+JRPan@users.noreply.github.com> Date: Mon, 17 May 2021 17:35:06 -0400 Subject: new configs - adaptive cache and cache write ratio --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 2 ++ configs/tested-cfgs/SM7_QV100/gpgpusim.config | 6 ++++++ configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 3 +++ src/abstract_hardware_model.h | 2 ++ src/gpgpu-sim/gpu-cache.h | 5 +++++ src/gpgpu-sim/gpu-sim.cc | 7 +++++++ 6 files changed, 25 insertions(+) (limited to 'configs') diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index 6189dca..e006085 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -110,6 +110,8 @@ -gpgpu_l1_latency 20 -gpgpu_smem_latency 20 -gpgpu_flush_l1_cache 1 +-gpgpu_shmem_option 0,8,16,32,64,100 +-gpgpu_unified_l1d_size 128 # 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 bc5677c..043fce6 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -124,6 +124,9 @@ -gpgpu_l1_latency 20 -gpgpu_smem_latency 20 -gpgpu_flush_l1_cache 1 +-gpgpu_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 @@ -203,3 +206,6 @@ #-trace_components WARP_SCHEDULER,SCOREBOARD #-trace_sampling_core 0 +-gpgpu_cache_write_ratio 25 +-gpgpu_shmem_option 0,12,24,48,96 +-gpgpu_unified_l1d_size 128 \ No newline at end of file diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index 3fa51ee..1f0c15f 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -125,6 +125,9 @@ -gpgpu_l1_latency 20 -gpgpu_smem_latency 20 -gpgpu_flush_l1_cache 1 +-gpgpu_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/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 982e416..e796571 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -373,6 +373,8 @@ class core_config { } unsigned mem_warp_parts; mutable unsigned gpgpu_shmem_size; + char *gpgpu_shmem_option; + unsigned gpgpu_unified_l1d_size; 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 00c09ae..ccc935b 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -491,6 +491,7 @@ class cache_config { m_data_port_width = 0; m_set_index_function = LINEAR_SET_FUNCTION; m_is_streaming = false; + m_wr_percent = 0; } void init(char *config, FuncCache status) { cache_status = status; @@ -754,6 +755,10 @@ class cache_config { char *m_config_stringPrefL1; char *m_config_stringPrefShared; FuncCache cache_status; + unsigned m_wr_percent; + write_allocate_policy_t get_write_allocate_policy() { + return m_write_alloc_policy; + } protected: void exit_parse_error() { diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index fd36e00..bd09cdb 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -249,6 +249,7 @@ void shader_core_config::reg_options(class OptionParser *opp) { " {::,:::,::, | none}", "none"); + option_parser_register(opp,"-gpgpu_cache_write_ratio",OPT_UINT32,&m_L1D_config.m_wr_percent,"L1D write ratio","0"); option_parser_register(opp, "-gpgpu_l1_banks", OPT_UINT32, &m_L1D_config.l1_banks, "The number of L1 cache banks", "1"); @@ -326,6 +327,12 @@ 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, "-gpgpu_shmem_option", OPT_CSTR, &gpgpu_shmem_option, + "Option list of shared memory sizes", "0"); + option_parser_register( + opp, "-gpgpu_unified_l1d_size", OPT_UINT32, &gpgpu_unified_l1d_size, + "Size of unified data cache(L1D + shared memory) in KB", "0"); option_parser_register(opp, "-gpgpu_adaptive_cache_config", OPT_BOOL, &adaptive_cache_config, "adaptive_cache_config", "0"); option_parser_register( -- cgit v1.3 From 4a762a933a054b5124fa46a12789ea98f5e2411d Mon Sep 17 00:00:00 2001 From: JRPAN <25518778+JRPan@users.noreply.github.com> Date: Wed, 19 May 2021 15:22:31 -0400 Subject: formatting again --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 4 ++-- configs/tested-cfgs/SM7_QV100/gpgpusim.config | 8 ++------ configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 2 +- src/gpgpu-sim/gpu-sim.cc | 2 +- src/gpgpu-sim/shader.cc | 1 + 5 files changed, 7 insertions(+), 10 deletions(-) (limited to 'configs') diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index e006085..d7573ab 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -100,6 +100,8 @@ # ::,::::,::,:** # ** 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_l1_banks 4 -gpgpu_cache:dl1 S:1:128:512,L:L:m:N:L,A:512:8,16:0,32 -gpgpu_shmem_size 65536 @@ -110,8 +112,6 @@ -gpgpu_l1_latency 20 -gpgpu_smem_latency 20 -gpgpu_flush_l1_cache 1 --gpgpu_shmem_option 0,8,16,32,64,100 --gpgpu_unified_l1d_size 128 # 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 043fce6..59c7f43 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -124,7 +124,7 @@ -gpgpu_l1_latency 20 -gpgpu_smem_latency 20 -gpgpu_flush_l1_cache 1 --gpgpu_cache_write_ratio 25 +-gpgpu_l1_cache_write_ratio 25 -gpgpu_shmem_option 0,12,24,48,96 -gpgpu_unified_l1d_size 128 @@ -204,8 +204,4 @@ # tracing functionality #-trace_enabled 1 #-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - --gpgpu_cache_write_ratio 25 --gpgpu_shmem_option 0,12,24,48,96 --gpgpu_unified_l1d_size 128 \ No newline at end of file +#-trace_sampling_core 0 \ No newline at end of file diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index 1f0c15f..3e080bc 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -125,7 +125,7 @@ -gpgpu_l1_latency 20 -gpgpu_smem_latency 20 -gpgpu_flush_l1_cache 1 --gpgpu_cache_write_ratio 25 +-gpgpu_l1_cache_write_ratio 25 -gpgpu_shmem_option 0,12,24,48,96 -gpgpu_unified_l1d_size 128 diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index bd09cdb..a2aa929 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -249,7 +249,7 @@ void shader_core_config::reg_options(class OptionParser *opp) { " {::,:::,::, | none}", "none"); - option_parser_register(opp,"-gpgpu_cache_write_ratio",OPT_UINT32,&m_L1D_config.m_wr_percent,"L1D write ratio","0"); + option_parser_register(opp,"-gpgpu_l1_cache_write_ratio",OPT_UINT32,&m_L1D_config.m_wr_percent,"L1D write ratio","0"); option_parser_register(opp, "-gpgpu_l1_banks", OPT_UINT32, &m_L1D_config.l1_banks, "The number of L1 cache banks", "1"); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index b2adb4f..141c700 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -3326,6 +3326,7 @@ unsigned int shader_core_config::max_cta(const kernel_info_t &k) const { // For more info about adaptive cache, see bool l1d_configured = false; unsigned l1_defined = m_L1D_config.get_original_sz() / 1024; + assert(gpgpu_unified_l1d_size % l1_defined == 0); unsigned max_assoc = m_L1D_config.get_original_assoc() * gpgpu_unified_l1d_size / l1_defined; -- cgit v1.3 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(-) (limited to 'configs') 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 (limited to 'configs') 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(-) (limited to 'configs') 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 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(-) (limited to 'configs') 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 (limited to 'configs') 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 From 778962ed40707369c97a03a3864cc1ee6c7470b6 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 26 May 2021 16:37:39 -0400 Subject: updating the configs based on the tuner output --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 118 +++++++++++------------ configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 108 +++++++++------------ 2 files changed, 100 insertions(+), 126 deletions(-) (limited to 'configs') diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index f715f3a..f35af1b 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -1,8 +1,3 @@ -# This config models the Turing RTX 2060 -# For more info about turing architecture: -# 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 @@ -13,7 +8,8 @@ -gpgpu_heap_size_limit 8388608 -gpgpu_runtime_sync_depth_limit 2 -gpgpu_runtime_pending_launch_count_limit 2048 --gpgpu_kernel_launch_latency 5000 +-gpgpu_kernel_launch_latency 7571 +-gpgpu_TB_launch_latency 0 # Compute Capability -gpgpu_compute_capability_major 7 @@ -27,31 +23,27 @@ -gpgpu_n_clusters 30 -gpgpu_n_cores_per_cluster 1 -gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 +-gpgpu_n_sub_partition_per_mchannel 2 -# volta clock domains +# 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 +-gpgpu_clock_domains 1365:1365:1365:3500.5 # shader core pipeline config -gpgpu_shader_registers 65536 -gpgpu_registers_per_block 65536 -gpgpu_occupancy_sm_number 75 -# This implies a maximum of 32 warps/SM --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 32 +-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,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_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 @@ -59,32 +51,18 @@ # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" # All Div operations are executed on SFU unit --ptx_opcode_latency_int 4,13,4,5,145,32 --ptx_opcode_initiation_int 2,2,2,2,8,4 --ptx_opcode_latency_fp 4,13,4,5,39 +-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 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 100 +-ptx_opcode_latency_dp 54,54,54,54,330 +-ptx_opcode_initiation_dp 64,64,64,64,130 +-ptx_opcode_latency_sfu 21 -ptx_opcode_initiation_sfu 8 -ptx_opcode_latency_tesnor 64 -ptx_opcode_initiation_tensor 64 -# 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 +# 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 @@ -92,31 +70,46 @@ -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 +# register banks +-gpgpu_num_reg_banks 8 -gpgpu_reg_file_port_throughput 2 +# 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 +# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x -gpgpu_adaptive_cache_config 1 --gpgpu_shmem_option 32,64 --gpgpu_unified_l1d_size 96 +-gpgpu_shmem_option 0,8,16,32,64,64 +-gpgpu_unified_l1d_size 64 # L1 cache configuration -gpgpu_l1_banks 4 --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_cache:dl1 S:4:128:128,L:T:m:L:L,A:256:32,16:0,32 +-gpgpu_l1_latency 32 -gpgpu_gmem_skip_L1D 0 --gpgpu_n_cluster_ejection_buffer_size 32 -gpgpu_flush_l1_cache 1 -# shared memory configuration +-gpgpu_n_cluster_ejection_buffer_size 32 +-gpgpu_l1_cache_write_ratio 25 + +# shared memory configuration -gpgpu_shmem_size 65536 -gpgpu_shmem_sizeDefault 65536 --gpgpu_shmem_per_block 65536 --gpgpu_smem_latency 20 +-gpgpu_shmem_per_block 49152 +-gpgpu_smem_latency 30 +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 75 -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives us 3MB L2 cache +# 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 @@ -127,34 +120,31 @@ -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 in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +# 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 -#-network_mode 1 -#-inter_config_file config_turing_islip.icnt # use built-in local xbar -network_mode 2 -icnt_in_buffer_limit 512 -icnt_out_buffer_limit 512 -icnt_subnets 2 --icnt_arbiter_algo 1 -icnt_flit_size 40 +-icnt_arbiter_algo 1 # memory partition latency config --gpgpu_l2_rop_latency 160 --dram_latency 100 +-gpgpu_l2_rop_latency 194 +-dram_latency 96 -# dram model config +# dram sched 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 +# dram model config -gpgpu_n_mem_per_ctrlr 1 -gpgpu_dram_buswidth 2 -gpgpu_dram_burst_length 16 @@ -162,9 +152,9 @@ -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" +# Mem timing +-gpgpu_dram_timing_opt nbk=16:CCD=4:RRD=12:RCD=24:RAS=55:RP=24:RC=78:CL=24:WL=8:CDLR=10:WR=24:nbkgrp=4:CCDL=6:RTPL=4 +-dram_dual_bus_interface 0 # select lower bits for bnkgrp to increase bnkgrp parallelism -dram_bnk_indexing_policy 0 @@ -179,7 +169,7 @@ -enable_ptx_file_line_stats 1 -visualizer_enabled 0 -# power model configs, disable it untill we create a real energy model for Volta +# power model configs, disable it untill we create a real energy model -power_simulation_enabled 0 # tracing functionality diff --git a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config index 02cdb9e..a68703f 100644 --- a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -1,19 +1,14 @@ -# This config models the Ampere RTX 3070 -# For more info about Ampere architecture: -# https://developer.download.nvidia.com/video/gputechconf/gtc/2020/presentations/s21730-inside-the-nvidia-ampere-architecture.pdf -# https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/ampere/pdf/NVIDIA-ampere-GA102-GPU-Architecture-Whitepaper-V1.pdf -# https://en.wikipedia.org/wiki/GeForce_30_series # functional simulator specification -gpgpu_ptx_instruction_classification 0 -gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 86 +-gpgpu_ptx_force_max_capability 86 # 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 +-gpgpu_kernel_launch_latency 7872 -gpgpu_TB_launch_latency 0 # Compute Capability @@ -30,26 +25,21 @@ -gpgpu_n_mem 16 -gpgpu_n_sub_partition_per_mchannel 2 -# Ampere clock domains +# clock domains #-gpgpu_clock_domains ::: --gpgpu_clock_domains 1320.0:1320.0:1320.0:3500.0 -# boost mode -# -gpgpu_clock_domains 1780.0:1780.0:1780.0:3500.0 +-gpgpu_clock_domains 1132:1132:1132:3500.5 # shader core pipeline config -gpgpu_shader_registers 65536 -gpgpu_registers_per_block 65536 -gpgpu_occupancy_sm_number 86 -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_core_pipeline 1536:32 -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 -## Ampere GA102 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 @@ -61,18 +51,18 @@ # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" # All Div operations are executed on SFU unit --ptx_opcode_latency_int 4,13,4,5,145,21 --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 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 100 +-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 1,1,1,1,2 +-ptx_opcode_latency_dp 55,55,55,55,330 +-ptx_opcode_initiation_dp 64,64,64,64,130 +-ptx_opcode_latency_sfu 21 -ptx_opcode_initiation_sfu 8 --ptx_opcode_latency_tesnor 32 --ptx_opcode_initiation_tensor 32 +-ptx_opcode_latency_tesnor 64 +-ptx_opcode_initiation_tensor 64 -# Ampere has sub core model, in which each scheduler has its own register file and EUs +# 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 @@ -80,50 +70,47 @@ -gpgpu_operand_collector_num_units_gen 8 -gpgpu_operand_collector_num_in_ports_gen 8 -gpgpu_operand_collector_num_out_ports_gen 8 -# Ampere has 24 double-ported banks, 4 schedulers, 6 banks per scheduler --gpgpu_num_reg_banks 24 +# register banks +-gpgpu_num_reg_banks 8 -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 86 - -# Ampere has four schedulers per core +# warp scheduling -gpgpu_num_sched_per_core 4 -# Greedy then oldest scheduler -gpgpu_scheduler gto -## In Ampere, a warp scheduler can issue 1 inst per cycle +# 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 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 -# disable this mode in case of multi kernels/apps execution +# In adaptive cache, we adaptively assign the remaining shared memory to L1 cache +# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x -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 +# L1 cache configuration -gpgpu_l1_banks 4 --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_cache:dl1 S:4:128:256,L:T:m:L:L,A:384:48,16:0,32 +-gpgpu_l1_latency 39 -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_n_cluster_ejection_buffer_size 32 +-gpgpu_l1_cache_write_ratio 25 + +# shared memory configuration -gpgpu_shmem_size 102400 -gpgpu_shmem_sizeDefault 102400 --gpgpu_shmem_per_block 102400 --gpgpu_smem_latency 20 +-gpgpu_shmem_per_block 49152 +-gpgpu_smem_latency 29 +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 86 -# 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 +# 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 @@ -133,15 +120,13 @@ -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 deprecated, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod +# 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 -#-network_mode 1 -#-inter_config_file config_ampere_islip.icnt # use built-in local xbar -network_mode 2 -icnt_in_buffer_limit 512 @@ -151,16 +136,15 @@ -icnt_arbiter_algo 1 # memory partition latency config --gpgpu_l2_rop_latency 160 --dram_latency 100 +-gpgpu_l2_rop_latency 187 +-dram_latency 254 -# dram model config +# dram sched config -gpgpu_dram_scheduler 1 -gpgpu_frfcfs_dram_sched_queue_size 64 -gpgpu_dram_return_queue_size 192 -# Ampere RTX3060 has GDDR6 -# http://monitorinsider.com/GDDR6.html +# dram model config -gpgpu_n_mem_per_ctrlr 1 -gpgpu_dram_buswidth 2 -gpgpu_dram_burst_length 16 @@ -168,9 +152,9 @@ -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" +# Mem timing +-gpgpu_dram_timing_opt nbk=16:CCD=4:RRD=12:RCD=24:RAS=55:RP=24:RC=78:CL=24:WL=8:CDLR=10:WR=24:nbkgrp=4:CCDL=6:RTPL=4 +-dram_dual_bus_interface 0 # select lower bits for bnkgrp to increase bnkgrp parallelism -dram_bnk_indexing_policy 0 @@ -185,7 +169,7 @@ -enable_ptx_file_line_stats 1 -visualizer_enabled 0 -# power model configs, disable it untill we create a real energy model for Ampere +# power model configs, disable it untill we create a real energy model -power_simulation_enabled 0 # tracing functionality -- cgit v1.3 From 3eea0140bc19dc1822d40e29d1aa55643894c6d3 Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Wed, 26 May 2021 19:44:28 -0400 Subject: changing kernel latency --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 2 +- configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) (limited to 'configs') diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index f35af1b..a994370 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -8,7 +8,7 @@ -gpgpu_heap_size_limit 8388608 -gpgpu_runtime_sync_depth_limit 2 -gpgpu_runtime_pending_launch_count_limit 2048 --gpgpu_kernel_launch_latency 7571 +-gpgpu_kernel_launch_latency 5000 -gpgpu_TB_launch_latency 0 # Compute Capability diff --git a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config index a68703f..fda3851 100644 --- a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -8,7 +8,7 @@ -gpgpu_heap_size_limit 8388608 -gpgpu_runtime_sync_depth_limit 2 -gpgpu_runtime_pending_launch_count_limit 2048 --gpgpu_kernel_launch_latency 7872 +-gpgpu_kernel_launch_latency 5000 -gpgpu_TB_launch_latency 0 # Compute Capability -- cgit v1.3 From 6ad461a95ac71e0597274c4f750ce03bb3a6871e Mon Sep 17 00:00:00 2001 From: "Mahmoud Khairy A. Abdallah" Date: Thu, 27 May 2021 15:38:26 -0400 Subject: fixing configs --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 8 ++++---- configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 2 +- 2 files changed, 5 insertions(+), 5 deletions(-) (limited to 'configs') diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index a994370..cc3152c 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -55,7 +55,7 @@ -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 54,54,54,54,330 +-ptx_opcode_latency_dp 64,64,64,64,330 -ptx_opcode_initiation_dp 64,64,64,64,130 -ptx_opcode_latency_sfu 21 -ptx_opcode_initiation_sfu 8 @@ -87,11 +87,11 @@ # In adaptive cache, we adaptively assign the remaining shared memory to L1 cache # For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x -gpgpu_adaptive_cache_config 1 --gpgpu_shmem_option 0,8,16,32,64,64 --gpgpu_unified_l1d_size 64 +-gpgpu_shmem_option 32,64 +-gpgpu_unified_l1d_size 96 # L1 cache configuration -gpgpu_l1_banks 4 --gpgpu_cache:dl1 S:4:128:128,L:T:m:L:L,A:256:32,16:0,32 +-gpgpu_cache:dl1 S:4:128:64,L:T:m:L:L,A:256:32,16:0,32 -gpgpu_l1_latency 32 -gpgpu_gmem_skip_L1D 0 -gpgpu_flush_l1_cache 1 diff --git a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config index fda3851..098cb1d 100644 --- a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -55,7 +55,7 @@ -ptx_opcode_initiation_int 2,2,2,2,2 -ptx_opcode_latency_fp 4,4,4,4,39 -ptx_opcode_initiation_fp 1,1,1,1,2 --ptx_opcode_latency_dp 55,55,55,55,330 +-ptx_opcode_latency_dp 64,64,64,64,330 -ptx_opcode_initiation_dp 64,64,64,64,130 -ptx_opcode_latency_sfu 21 -ptx_opcode_initiation_sfu 8 -- cgit v1.3 From a8256e50a6d25338f659da76ff9c3595132f54b2 Mon Sep 17 00:00:00 2001 From: JRPAN <25518778+JRPan@users.noreply.github.com> Date: Mon, 23 Aug 2021 13:06:13 -0400 Subject: Update volta to use lrr scheduler --- configs/tested-cfgs/SM7_QV100/gpgpusim.config | 2 +- configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) (limited to 'configs') diff --git a/configs/tested-cfgs/SM7_QV100/gpgpusim.config b/configs/tested-cfgs/SM7_QV100/gpgpusim.config index 5f22a42..425bc16 100644 --- a/configs/tested-cfgs/SM7_QV100/gpgpusim.config +++ b/configs/tested-cfgs/SM7_QV100/gpgpusim.config @@ -99,7 +99,7 @@ # Volta has four schedulers per core -gpgpu_num_sched_per_core 4 # Greedy then oldest scheduler --gpgpu_scheduler gto +-gpgpu_scheduler lrr ## 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 diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index c44563f..0c69c70 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -100,7 +100,7 @@ # Volta has four schedulers per core -gpgpu_num_sched_per_core 4 # Greedy then oldest scheduler --gpgpu_scheduler gto +-gpgpu_scheduler lrr ## 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 -- cgit v1.3 From 84c4f46fb78b529ab2447d7a676f5b3ac2d9c05f Mon Sep 17 00:00:00 2001 From: JRPAN <25518778+JRPan@users.noreply.github.com> Date: Mon, 23 Aug 2021 13:06:54 -0400 Subject: Ampere and Turing also lrr scheduler --- configs/tested-cfgs/SM75_RTX2060/gpgpusim.config | 2 +- configs/tested-cfgs/SM86_RTX3070/gpgpusim.config | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) (limited to 'configs') diff --git a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config index cc3152c..0ae91a5 100644 --- a/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config +++ b/configs/tested-cfgs/SM75_RTX2060/gpgpusim.config @@ -76,7 +76,7 @@ # warp scheduling -gpgpu_num_sched_per_core 4 --gpgpu_scheduler gto +-gpgpu_scheduler lrr # a warp scheduler issue mode -gpgpu_max_insn_issue_per_warp 1 -gpgpu_dual_issue_diff_exec_units 1 diff --git a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config index 098cb1d..8543781 100644 --- a/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config +++ b/configs/tested-cfgs/SM86_RTX3070/gpgpusim.config @@ -76,7 +76,7 @@ # warp scheduling -gpgpu_num_sched_per_core 4 --gpgpu_scheduler gto +-gpgpu_scheduler lrr # a warp scheduler issue mode -gpgpu_max_insn_issue_per_warp 1 -gpgpu_dual_issue_diff_exec_units 1 -- cgit v1.3