From e87f15a76c5b3225911850fcdb4074c16682ae50 Mon Sep 17 00:00:00 2001 From: Akshay Jain Date: Wed, 21 Mar 2018 13:40:06 -0400 Subject: Change 124 by jain156@akshayj-lt1 on 2016/12/08 20:17:53 Added Memory divergence stats Data shows same divergence characteristics across PTX and PTX+ --- src/gpgpu-sim/shader.cc | 9 +++++++++ 1 file changed, 9 insertions(+) (limited to 'src/gpgpu-sim/shader.cc') diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index bf482fb..72e0266 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -438,6 +438,9 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout,"gpgpu_n_mem_texture = %d\n", gpgpu_n_mem_texture); fprintf(fout,"gpgpu_n_mem_const = %d\n", gpgpu_n_mem_const); + fprintf(fout,"gpgpu_n_times_gmem_accesses_by_warps = %lld\n", gpgpu_n_times_gmem_accesses_by_warps); + fprintf(fout,"gpgpu_n_total_gmem_accesses_by_warps = %lld\n", gpgpu_n_total_gmem_accesses_by_warps); + fprintf(fout, "gpgpu_n_load_insn = %d\n", gpgpu_n_load_insn); fprintf(fout, "gpgpu_n_store_insn = %d\n", gpgpu_n_store_insn); fprintf(fout, "gpgpu_n_shmem_insn = %d\n", gpgpu_n_shmem_insn); @@ -740,9 +743,15 @@ void shader_core_ctx::fetch() void shader_core_ctx::func_exec_inst( warp_inst_t &inst ) { + unsigned starting_queue_size; execute_warp_inst_t(inst); if( inst.is_load() || inst.is_store() ) + starting_queue_size = inst.accessq_count(); inst.generate_mem_accesses(); + if ( inst.space.get_type() == global_space ) { + m_stats->gpgpu_n_times_gmem_accesses_by_warps++; + m_stats->gpgpu_n_total_gmem_accesses_by_warps += inst.accessq_count() - starting_queue_size; + } } void shader_core_ctx::issue_warp( register_set& pipe_reg_set, const warp_inst_t* next_inst, const active_mask_t &active_mask, unsigned warp_id ) -- cgit v1.3 From 160db7d46af433d12288505005ebd8c41be59922 Mon Sep 17 00:00:00 2001 From: Akshay Jain Date: Thu, 22 Mar 2018 04:34:40 -0400 Subject: Change 253 by jain156@akshayj-lt1 on 2017/05/30 00:40:03 Replaced the mem div stats with mem div histogram --- src/gpgpu-sim/shader.cc | 8 ++++---- src/gpgpu-sim/shader.h | 9 ++++++--- 2 files changed, 10 insertions(+), 7 deletions(-) (limited to 'src/gpgpu-sim/shader.cc') diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 72e0266..e7204c8 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -438,8 +438,9 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout,"gpgpu_n_mem_texture = %d\n", gpgpu_n_mem_texture); fprintf(fout,"gpgpu_n_mem_const = %d\n", gpgpu_n_mem_const); - fprintf(fout,"gpgpu_n_times_gmem_accesses_by_warps = %lld\n", gpgpu_n_times_gmem_accesses_by_warps); - fprintf(fout,"gpgpu_n_total_gmem_accesses_by_warps = %lld\n", gpgpu_n_total_gmem_accesses_by_warps); + fprintf(fout,"gpgpu_mem_divergence_hist "); + gpgpu_mem_divergence_hist->fprint(fout); + fprintf(fout,"\n"); fprintf(fout, "gpgpu_n_load_insn = %d\n", gpgpu_n_load_insn); fprintf(fout, "gpgpu_n_store_insn = %d\n", gpgpu_n_store_insn); @@ -749,8 +750,7 @@ void shader_core_ctx::func_exec_inst( warp_inst_t &inst ) starting_queue_size = inst.accessq_count(); inst.generate_mem_accesses(); if ( inst.space.get_type() == global_space ) { - m_stats->gpgpu_n_times_gmem_accesses_by_warps++; - m_stats->gpgpu_n_total_gmem_accesses_by_warps += inst.accessq_count() - starting_queue_size; + m_stats->gpgpu_mem_divergence_hist->add2bin(inst.accessq_count() - starting_queue_size); } } diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 2475677..45658ac 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -53,6 +53,8 @@ #include "stats.h" #include "gpu-cache.h" #include "traffic_breakdown.h" +#include "histogram.h" + @@ -1471,9 +1473,8 @@ struct shader_core_stats_pod { int gpgpu_n_mem_write_global; int gpgpu_n_mem_read_inst; - //warps memory divergence count - unsigned long long gpgpu_n_times_gmem_accesses_by_warps; - unsigned long long gpgpu_n_total_gmem_accesses_by_warps; + //warps combined memory divergence histogram + linear_histogram* gpgpu_mem_divergence_hist; int gpgpu_n_mem_l2_writeback; int gpgpu_n_mem_l1_write_allocate; @@ -1543,6 +1544,7 @@ public: m_incoming_traffic_stats = new traffic_breakdown("memtocore"); gpgpu_n_shmem_bank_access = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); + gpgpu_mem_divergence_hist = new linear_histogram(1, "", config->warp_size+1); m_shader_dynamic_warp_issue_distro.resize( config->num_shader() ); m_shader_warp_slot_issue_distro.resize( config->num_shader() ); @@ -1557,6 +1559,7 @@ public: free(m_n_diverge); free(shader_cycle_distro); free(last_shader_cycle_distro); + free(gpgpu_mem_divergence_hist); } void new_grid() -- cgit v1.3 From 525f177a1d4f1fc309d22197bc8af26ee8d4c454 Mon Sep 17 00:00:00 2001 From: Akshay Jain Date: Thu, 22 Mar 2018 04:40:44 -0400 Subject: Change 283 by jain156@akshayj-lt1 on 2017/07/03 19:52:55 Adding the global load and store divergence statistics. The previous histogram has been modified to look at all memory requests, but that is not meaningful - so it can as well be removed. --- src/gpgpu-sim/shader.cc | 15 +++++++++++++-- src/gpgpu-sim/shader.h | 8 +++++++- 2 files changed, 20 insertions(+), 3 deletions(-) (limited to 'src/gpgpu-sim/shader.cc') diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index e7204c8..db6a532 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -441,6 +441,12 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout,"gpgpu_mem_divergence_hist "); gpgpu_mem_divergence_hist->fprint(fout); fprintf(fout,"\n"); + fprintf(fout,"gpgpu_gmem_ld_divergence_hist "); + gpgpu_gmem_ld_divergence_hist->fprint(fout); + fprintf(fout,"\n"); + fprintf(fout,"gpgpu_gmem_st_divergence_hist "); + gpgpu_gmem_st_divergence_hist->fprint(fout); + fprintf(fout,"\n"); fprintf(fout, "gpgpu_n_load_insn = %d\n", gpgpu_n_load_insn); fprintf(fout, "gpgpu_n_store_insn = %d\n", gpgpu_n_store_insn); @@ -746,12 +752,17 @@ void shader_core_ctx::func_exec_inst( warp_inst_t &inst ) { unsigned starting_queue_size; execute_warp_inst_t(inst); - if( inst.is_load() || inst.is_store() ) + if( inst.is_load() || inst.is_store() ) { starting_queue_size = inst.accessq_count(); inst.generate_mem_accesses(); if ( inst.space.get_type() == global_space ) { - m_stats->gpgpu_mem_divergence_hist->add2bin(inst.accessq_count() - starting_queue_size); + if (inst.is_load()) + m_stats->gpgpu_gmem_ld_divergence_hist->add2bin(inst.accessq_count() - starting_queue_size); + else if (inst.is_store()) + m_stats->gpgpu_gmem_st_divergence_hist->add2bin(inst.accessq_count() - starting_queue_size); } + m_stats->gpgpu_mem_divergence_hist->add2bin(inst.accessq_count() - starting_queue_size); + } } void shader_core_ctx::issue_warp( register_set& pipe_reg_set, const warp_inst_t* next_inst, const active_mask_t &active_mask, unsigned warp_id ) diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 45658ac..da7eb32 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1475,7 +1475,9 @@ struct shader_core_stats_pod { //warps combined memory divergence histogram linear_histogram* gpgpu_mem_divergence_hist; - + linear_histogram* gpgpu_gmem_ld_divergence_hist; + linear_histogram* gpgpu_gmem_st_divergence_hist; + int gpgpu_n_mem_l2_writeback; int gpgpu_n_mem_l1_write_allocate; int gpgpu_n_mem_l2_write_allocate; @@ -1545,6 +1547,8 @@ public: gpgpu_n_shmem_bank_access = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); gpgpu_mem_divergence_hist = new linear_histogram(1, "", config->warp_size+1); + gpgpu_gmem_ld_divergence_hist = new linear_histogram(1, "", config->warp_size+1); + gpgpu_gmem_st_divergence_hist = new linear_histogram(1, "", config->warp_size+1); m_shader_dynamic_warp_issue_distro.resize( config->num_shader() ); m_shader_warp_slot_issue_distro.resize( config->num_shader() ); @@ -1560,6 +1564,8 @@ public: free(shader_cycle_distro); free(last_shader_cycle_distro); free(gpgpu_mem_divergence_hist); + free(gpgpu_gmem_ld_divergence_hist); + free(gpgpu_gmem_st_divergence_hist); } void new_grid() -- cgit v1.3 From b43b94569734d9f27b2e4bbb22272aa4243fbd56 Mon Sep 17 00:00:00 2001 From: Akshay Jain Date: Thu, 22 Mar 2018 04:45:59 -0400 Subject: Change 332 by jain156@akshayj-lt1 on 2018/02/01 05:20:33 Added shmem_divergence_hist and warp_inst_classification --- src/abstract_hardware_model.h | 7 +++++++ src/cuda-sim/cuda-sim.cc | 4 +++- src/gpgpu-sim/shader.cc | 13 +++++++++++++ src/gpgpu-sim/shader.h | 7 +++++++ 4 files changed, 30 insertions(+), 1 deletion(-) (limited to 'src/gpgpu-sim/shader.cc') diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index cec75f9..9ba31d9 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -798,6 +798,7 @@ public: arch_reg.dst[i] = -1; } isize=0; + op_classification = 0; } bool valid() const { return m_decoded; } virtual void print_insn( FILE *fp ) const @@ -816,6 +817,7 @@ public: address_type pc; // program counter address of instruction unsigned isize; // size of instruction in bytes op_type op; // opcode (uarch visible) + int op_classification; // classification of opcode for statistics purpopses barrier_type bar_type; reduction_type red_type; @@ -1023,6 +1025,11 @@ public: return cycles > 0; } + unsigned get_cycles() + { + return cycles; + } + void print( FILE *fout ) const; unsigned get_uid() const { return m_uid; } diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 17a7798..169c491 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1331,8 +1331,10 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) if ( gpgpu_ptx_instruction_classification ) { init_inst_classification_stat(); - if (op_classification) + if (op_classification) { StatAddSample( g_inst_classification_stat[g_ptx_kernel_count], op_classification); + inst.op_classification = op_classification; + } if (pI->get_space().get_type()) StatAddSample( g_inst_mem_classification_stat[g_ptx_kernel_count], ( int )pI->get_space().get_type()); StatAddSample( g_inst_op_classification_stat[g_ptx_kernel_count], (int) pI->get_opcode() ); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index db6a532..905139c 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -447,6 +447,12 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout,"gpgpu_gmem_st_divergence_hist "); gpgpu_gmem_st_divergence_hist->fprint(fout); fprintf(fout,"\n"); + fprintf(fout,"gpgpu_shmem_divergence_hist "); + gpgpu_shmem_divergence_hist->fprint(fout); + fprintf(fout,"\n"); + fprintf(fout,"warp_inst_classification "); + warp_inst_classification->fprint(fout); + fprintf(fout,"\n"); fprintf(fout, "gpgpu_n_load_insn = %d\n", gpgpu_n_load_insn); fprintf(fout, "gpgpu_n_store_insn = %d\n", gpgpu_n_store_insn); @@ -752,6 +758,9 @@ void shader_core_ctx::func_exec_inst( warp_inst_t &inst ) { unsigned starting_queue_size; execute_warp_inst_t(inst); + if (inst.op_classification) { + m_stats->warp_inst_classification->add2bin(inst.op_classification); + } if( inst.is_load() || inst.is_store() ) { starting_queue_size = inst.accessq_count(); inst.generate_mem_accesses(); @@ -761,6 +770,10 @@ void shader_core_ctx::func_exec_inst( warp_inst_t &inst ) else if (inst.is_store()) m_stats->gpgpu_gmem_st_divergence_hist->add2bin(inst.accessq_count() - starting_queue_size); } + else if ( inst.space.get_type() == shared_space ) { + m_stats->gpgpu_shmem_divergence_hist->add2bin(inst.get_cycles()); + } + m_stats->gpgpu_mem_divergence_hist->add2bin(inst.accessq_count() - starting_queue_size); } } diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index da7eb32..9fd1ecb 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1477,6 +1477,9 @@ struct shader_core_stats_pod { linear_histogram* gpgpu_mem_divergence_hist; linear_histogram* gpgpu_gmem_ld_divergence_hist; linear_histogram* gpgpu_gmem_st_divergence_hist; + linear_histogram* gpgpu_shmem_divergence_hist; + + linear_histogram* warp_inst_classification; int gpgpu_n_mem_l2_writeback; int gpgpu_n_mem_l1_write_allocate; @@ -1549,6 +1552,9 @@ public: gpgpu_mem_divergence_hist = new linear_histogram(1, "", config->warp_size+1); gpgpu_gmem_ld_divergence_hist = new linear_histogram(1, "", config->warp_size+1); gpgpu_gmem_st_divergence_hist = new linear_histogram(1, "", config->warp_size+1); + gpgpu_shmem_divergence_hist = new linear_histogram(1, "", config->warp_size+1); + + warp_inst_classification = new linear_histogram(1, "", 12); m_shader_dynamic_warp_issue_distro.resize( config->num_shader() ); m_shader_warp_slot_issue_distro.resize( config->num_shader() ); @@ -1566,6 +1572,7 @@ public: free(gpgpu_mem_divergence_hist); free(gpgpu_gmem_ld_divergence_hist); free(gpgpu_gmem_st_divergence_hist); + free(warp_inst_classification); } void new_grid() -- cgit v1.3