From 67a257442f46a9e8e02f63a9d71fcc7ca54f3f5e Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 28 Mar 2018 14:00:16 -0400 Subject: adding new counters for parition level para and L2 BW --- src/gpgpu-sim/gpu-sim.cc | 36 ++++++++++++++++++++++++++++++++++++ 1 file changed, 36 insertions(+) (limited to 'src') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 48af41e..228d4be 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -87,6 +87,12 @@ unsigned long long gpu_tot_sim_cycle = 0; // performance counter for stalls due to congestion. unsigned int gpu_stall_dramfull = 0; unsigned int gpu_stall_icnt2sh = 0; +unsigned long long partiton_reqs_in_parallel = 0; +unsigned long long partiton_reqs_in_parallel_total = 0; +unsigned long long partiton_reqs_in_parallel_util = 0; +unsigned long long partiton_reqs_in_parallel_util_total = 0; +unsigned long long partiton_replys_in_parallel = 0; +unsigned long long partiton_replys_in_parallel_total = 0; /* Clock Domains */ @@ -786,6 +792,9 @@ void gpgpu_sim::init() gpu_sim_insn = 0; last_gpu_sim_insn = 0; m_total_cta_launched=0; + partiton_reqs_in_parallel = 0; + partiton_replys_in_parallel = 0; + partiton_reqs_in_parallel_util = 0; reinit_clock_domains(); set_param_gpgpu_num_shaders(m_config.num_shader()); @@ -822,8 +831,14 @@ void gpgpu_sim::update_stats() { gpu_tot_sim_cycle += gpu_sim_cycle; gpu_tot_sim_insn += gpu_sim_insn; gpu_tot_issued_cta += m_total_cta_launched; + partiton_reqs_in_parallel_total += partiton_reqs_in_parallel; + partiton_replys_in_parallel_total += partiton_replys_in_parallel; + partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; gpu_sim_cycle = 0; + partiton_reqs_in_parallel = 0; + partiton_replys_in_parallel = 0; + partiton_reqs_in_parallel_util = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } @@ -1007,6 +1022,19 @@ void gpgpu_sim::gpu_print_stat() printf("gpu_stall_dramfull = %d\n", gpu_stall_dramfull); printf("gpu_stall_icnt2sh = %d\n", gpu_stall_icnt2sh ); + printf("partiton_reqs_in_parallel = %lld\n", partiton_reqs_in_parallel); + printf("partiton_reqs_in_parallel_total = %lld\n", partiton_reqs_in_parallel_total ); + printf("partiton_level_parallism = %12.4f\n", (float)partiton_reqs_in_parallel / gpu_sim_cycle); + printf("partiton_level_parallism_total = %12.4f\n", (float)(partiton_reqs_in_parallel+partiton_reqs_in_parallel_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("partiton_reqs_in_parallel_util = %lld\n", partiton_reqs_in_parallel_util); + printf("partiton_reqs_in_parallel_util_total = %lld\n", partiton_reqs_in_parallel_util_total ); + printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle); + printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("partiton_replys_in_parallel = %lld\n", partiton_replys_in_parallel); + printf("partiton_replys_in_parallel_total = %lld\n", partiton_replys_in_parallel_total ); + printf("L2_BW = %12.4f GB/Sec\n", ((float)(partiton_replys_in_parallel * 32) / (gpu_sim_cycle * m_config.icnt_period)) / 1000000000); + printf("L2_BW_total = %12.4f GB/Sec\n", ((float)((partiton_replys_in_parallel+partiton_replys_in_parallel_total) * 32) / ((gpu_tot_sim_cycle+gpu_sim_cycle) * m_config.icnt_period)) / 1000000000 ); + time_t curr_time; time(&curr_time); unsigned long long elapsed_time = MAX( curr_time - g_simulation_starttime, 1 ); @@ -1412,6 +1440,7 @@ void gpgpu_sim::cycle() for (unsigned i=0;in_simt_clusters;i++) m_cluster[i]->icnt_cycle(); } + unsigned partiton_replys_in_parallel_per_cycle = 0; if (clock_mask & ICNT) { // pop from memory controller to interconnect for (unsigned i=0;im_n_mem_sub_partition;i++) { @@ -1424,6 +1453,7 @@ void gpgpu_sim::cycle() mf->set_status(IN_ICNT_TO_SHADER,gpu_sim_cycle+gpu_tot_sim_cycle); ::icnt_push( m_shader_config->mem2device(i), mf->get_tpc(), mf, response_size ); m_memory_sub_partition[i]->pop(); + partiton_replys_in_parallel_per_cycle++; } else { gpu_stall_icnt2sh++; } @@ -1432,6 +1462,7 @@ void gpgpu_sim::cycle() } } } + partiton_replys_in_parallel += partiton_replys_in_parallel_per_cycle; if (clock_mask & DRAM) { for (unsigned i=0;im_n_mem;i++){ @@ -1444,6 +1475,7 @@ void gpgpu_sim::cycle() } // L2 operations follow L2 clock domain + unsigned partiton_reqs_in_parallel_per_cycle = 0; if (clock_mask & L2) { m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].clear(); for (unsigned i=0;im_n_mem_sub_partition;i++) { @@ -1455,11 +1487,15 @@ void gpgpu_sim::cycle() } else { mem_fetch* mf = (mem_fetch*) icnt_pop( m_shader_config->mem2device(i) ); m_memory_sub_partition[i]->push( mf, gpu_sim_cycle + gpu_tot_sim_cycle ); + partiton_reqs_in_parallel_per_cycle++; } m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle+gpu_tot_sim_cycle); m_memory_sub_partition[i]->accumulate_L2cache_stats(m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]); } } + partiton_reqs_in_parallel += partiton_reqs_in_parallel_per_cycle; + if(partiton_reqs_in_parallel_per_cycle > 0) + partiton_reqs_in_parallel_util += partiton_reqs_in_parallel_per_cycle; if (clock_mask & ICNT) { icnt_transfer(); -- cgit v1.3 From 93cdd2e217d7903ccf715a00b6dae8265bb38fca Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 28 Mar 2018 15:22:33 -0400 Subject: adding new stats counter --- src/gpgpu-sim/gpu-sim.cc | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) (limited to 'src') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index f216e7c..006d243 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -91,6 +91,8 @@ unsigned long long partiton_reqs_in_parallel = 0; unsigned long long partiton_reqs_in_parallel_total = 0; unsigned long long partiton_reqs_in_parallel_util = 0; unsigned long long partiton_reqs_in_parallel_util_total = 0; +unsigned long long gpu_sim_cycle_parition_util = 0; +unsigned long long gpu_tot_sim_cycle_parition_util = 0; unsigned long long partiton_replys_in_parallel = 0; unsigned long long partiton_replys_in_parallel_total = 0; @@ -795,6 +797,7 @@ void gpgpu_sim::init() partiton_reqs_in_parallel = 0; partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; + gpu_sim_cycle_parition_util = 0; reinit_clock_domains(); set_param_gpgpu_num_shaders(m_config.num_shader()); @@ -834,11 +837,14 @@ void gpgpu_sim::update_stats() { partiton_reqs_in_parallel_total += partiton_reqs_in_parallel; partiton_replys_in_parallel_total += partiton_replys_in_parallel; partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; + gpu_tot_sim_cycle_parition_util += gpu_sim_cycle_parition_util ; gpu_sim_cycle = 0; partiton_reqs_in_parallel = 0; partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; + gpu_sim_cycle_parition_util = 0; + gpu_tot_sim_cycle = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } @@ -1028,8 +1034,10 @@ void gpgpu_sim::gpu_print_stat() printf("partiton_level_parallism_total = %12.4f\n", (float)(partiton_reqs_in_parallel+partiton_reqs_in_parallel_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); printf("partiton_reqs_in_parallel_util = %lld\n", partiton_reqs_in_parallel_util); printf("partiton_reqs_in_parallel_util_total = %lld\n", partiton_reqs_in_parallel_util_total ); - printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle); - printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("gpu_sim_cycle_parition_util = %lld\n", gpu_sim_cycle_parition_util); + printf("gpu_tot_sim_cycle_parition_util = %lld\n", gpu_tot_sim_cycle_parition_util ); + printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle_parition_util); + printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_sim_cycle_parition_util+gpu_tot_sim_cycle_parition_util) ); printf("partiton_replys_in_parallel = %lld\n", partiton_replys_in_parallel); printf("partiton_replys_in_parallel_total = %lld\n", partiton_replys_in_parallel_total ); printf("L2_BW = %12.4f GB/Sec\n", ((float)(partiton_replys_in_parallel * 32) / (gpu_sim_cycle * m_config.icnt_period)) / 1000000000); @@ -1494,8 +1502,10 @@ void gpgpu_sim::cycle() } } partiton_reqs_in_parallel += partiton_reqs_in_parallel_per_cycle; - if(partiton_reqs_in_parallel_per_cycle > 0) + if(partiton_reqs_in_parallel_per_cycle > 0){ partiton_reqs_in_parallel_util += partiton_reqs_in_parallel_per_cycle; + gpu_sim_cycle_parition_util++; + } if (clock_mask & ICNT) { icnt_transfer(); -- cgit v1.3 From c22eb8a8f6e41921e4e1a5e705cca8d5bbed67e0 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Wed, 28 Mar 2018 15:29:46 -0400 Subject: gpuwaatch build does not parallelize nicely without -p --- src/gpuwattch/makefile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'src') diff --git a/src/gpuwattch/makefile b/src/gpuwattch/makefile index ab718cc..354c9ec 100644 --- a/src/gpuwattch/makefile +++ b/src/gpuwattch/makefile @@ -11,10 +11,10 @@ opt: $(TAR).mk obj_opt @$(MAKE) TAG=opt -C . -f $(TAR).mk obj_dbg: - mkdir $@ + mkdir -p $@ obj_opt: - mkdir $@ + mkdir -p $@ depend: @$(MAKE) TAG=opt -C . -f $(TAR).mk depend -- cgit v1.3 From 21ac2041365dc8521a6926244b96e85fd19cc2d3 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Wed, 28 Mar 2018 22:52:12 -0400 Subject: Fixing a minor typo in an old config --- src/gpgpu-sim/gpu-sim.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 58a5d16..d452888 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -223,7 +223,7 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); - option_parser_register(opp, "-gpgpu_cache:dl1PreShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, + option_parser_register(opp, "-gpgpu_cache:dl1PrefShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); -- cgit v1.3 From 3b9cc9285b5394320873e42edee03d8f23d6d17e Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 30 Mar 2018 11:27:57 -0400 Subject: fixing gpu-tot-cycle bug --- src/gpgpu-sim/gpu-sim.cc | 1 - 1 file changed, 1 deletion(-) (limited to 'src') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 006d243..7835496 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -844,7 +844,6 @@ void gpgpu_sim::update_stats() { partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; gpu_sim_cycle_parition_util = 0; - gpu_tot_sim_cycle = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } -- cgit v1.3 From 6aba79083ec47a78ad2a9572a03cd52f42141def Mon Sep 17 00:00:00 2001 From: tgrogers Date: Fri, 30 Mar 2018 16:10:27 -0400 Subject: When CDP was added - we started printing garbage stats after every memcpy. I am stopping this by gaurding against it in the same way we do for updating the stat counts (checking if we have executed any cycles before printing). I am not sure what this does to CDP stat prints - as I am not sure what CDP kernels so to sim_cycles.... However EVERY other app except CDP is spewing junk to the output and it really messes up any per-kernel stat collection scripts. --- src/gpgpusim_entrypoint.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src') diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index ad4587a..dac9fd8 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -156,8 +156,8 @@ void *gpgpu_sim_thread_concurrent(void*) printf("GPGPU-Sim: ** STOP simulation thread (no work) **\n"); fflush(stdout); } - g_the_gpu->print_stats(); if(sim_cycles) { + g_the_gpu->print_stats(); g_the_gpu->update_stats(); print_simulation_time(); } -- cgit v1.3 From 9b586f9a61cac16b147e6dfea62abecc0a2a8eb7 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Fri, 30 Mar 2018 16:20:36 -0400 Subject: Moving the print to avoid spurrious statistics --- src/gpgpusim_entrypoint.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src') diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index ad4587a..dac9fd8 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -156,8 +156,8 @@ void *gpgpu_sim_thread_concurrent(void*) printf("GPGPU-Sim: ** STOP simulation thread (no work) **\n"); fflush(stdout); } - g_the_gpu->print_stats(); if(sim_cycles) { + g_the_gpu->print_stats(); g_the_gpu->update_stats(); print_simulation_time(); } -- cgit v1.3 From 89db73061e043c26df22c7f18d9adb106d8078ac Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 31 Mar 2018 17:56:50 -0400 Subject: Getting rid of our constant, annoying prints. Running workloads of any size causes ridiculous output file sizes --- src/cuda-sim/cuda-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.cc | 12 +++++++----- src/gpgpu-sim/shader.cc | 14 ++++++++------ src/trace_streams.tup | 1 + 4 files changed, 17 insertions(+), 12 deletions(-) (limited to 'src') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index d4ace76..a34b99b 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1408,7 +1408,7 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) if ( (g_ptx_sim_num_insn % 100000) == 0 ) { dim3 ctaid = get_ctaid(); dim3 tid = get_tid(); - printf("GPGPU-Sim PTX: %u instructions simulated : ctaid=(%u,%u,%u) tid=(%u,%u,%u)\n", + DPRINTF(LIVENESS, "GPGPU-Sim PTX: %u instructions simulated : ctaid=(%u,%u,%u) tid=(%u,%u,%u)\n", g_ptx_sim_num_insn, ctaid.x,ctaid.y,ctaid.z,tid.x,tid.y,tid.z ); fflush(stdout); } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index d452888..0ef267d 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -36,6 +36,7 @@ #include "shader.h" +#include "shader_trace.h" #include "dram.h" #include "mem_fetch.h" @@ -1175,8 +1176,8 @@ bool shader_core_ctx::occupy_shader_resource_1block(kernel_info_t & k, bool occu m_occupied_regs += (padded_cta_size * ((kernel_info->regs+3)&~3)); m_occupied_ctas++; - printf("GPGPU-Sim uArch: Shader %d occupied %d threads, %d shared mem, %d registers, %d ctas\n", - m_sid, m_occupied_n_threads, m_occupied_shmem, m_occupied_regs, m_occupied_ctas); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: Occupied %d threads, %d shared mem, %d registers, %d ctas\n", + m_occupied_n_threads, m_occupied_shmem, m_occupied_regs, m_occupied_ctas); } return true; @@ -1301,8 +1302,8 @@ void shader_core_ctx::issue_block2core( kernel_info_t &kernel ) m_n_active_cta++; shader_CTA_count_log(m_sid, 1); - printf("GPGPU-Sim uArch: core:%3d, cta:%2u, start_tid:%4u, end_tid:%4u, initialized @(%lld,%lld)\n", - m_sid, free_cta_hw_id, start_thread, end_thread, gpu_sim_cycle, gpu_tot_sim_cycle ); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: cta:%2u, start_tid:%4u, end_tid:%4u, initialized @(%lld,%lld)\n", + free_cta_hw_id, start_thread, end_thread, gpu_sim_cycle, gpu_tot_sim_cycle ); } @@ -1499,7 +1500,8 @@ void gpgpu_sim::cycle() hrs = elapsed_time/3600 - 24*days; minutes = elapsed_time/60 - 60*(hrs + 24*days); sec = elapsed_time - 60*(minutes + 60*(hrs + 24*days)); - printf("GPGPU-Sim uArch: cycles simulated: %lld inst.: %lld (ipc=%4.1f) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", + + DPRINTF(LIVENESS, "GPGPU-Sim uArch: cycles simulated: %lld inst.: %lld (ipc=%4.1f) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", gpu_tot_sim_cycle + gpu_sim_cycle, gpu_tot_sim_insn + gpu_sim_insn, (double)gpu_sim_insn/(double)gpu_sim_cycle, (unsigned)((gpu_tot_sim_insn+gpu_sim_insn) / elapsed_time), diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index d17e51d..4640d65 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1969,12 +1969,12 @@ void shader_core_ctx::register_cta_thread_exit( unsigned cta_num, kernel_info_t m_barriers.deallocate_barrier(cta_num); shader_CTA_count_unlog(m_sid, 1); - printf("GPGPU-Sim uArch: Shader %d finished CTA #%d (%lld,%lld), %u CTAs running\n", m_sid, cta_num, gpu_sim_cycle, gpu_tot_sim_cycle, - m_n_active_cta ); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: Finished CTA #%d (%lld,%lld), %u CTAs running\n", + cta_num, gpu_sim_cycle, gpu_tot_sim_cycle, m_n_active_cta); if( m_n_active_cta == 0 ) { - printf("GPGPU-Sim uArch: Shader %u empty (last released kernel %u \'%s\').\n", m_sid, kernel->get_uid(), - kernel->name().c_str() ); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: Empty (last released kernel %u \'%s\').\n", + kernel->get_uid(), kernel->name().c_str()); fflush(stdout); //Shader can only be empty when no more cta are dispatched @@ -1989,8 +1989,10 @@ void shader_core_ctx::register_cta_thread_exit( unsigned cta_num, kernel_info_t kernel->dec_running(); if( !m_gpu->kernel_more_cta_left(kernel) ) { if( !kernel->running() ) { - printf("GPGPU-Sim uArch: GPU detected kernel %u \'%s\' finished on shader %u.\n", kernel->get_uid(), - kernel->name().c_str(), m_sid ); + SHADER_DPRINTF(LIVENESS, + "GPGPU-Sim uArch: GPU detected kernel %u \'%s\' finished on shader %u.\n", kernel->get_uid(), + kernel->name().c_str(), m_sid); + if(m_kernel == kernel) m_kernel = NULL; m_gpu->set_kernel_done( kernel ); diff --git a/src/trace_streams.tup b/src/trace_streams.tup index c41690e..fd68a16 100644 --- a/src/trace_streams.tup +++ b/src/trace_streams.tup @@ -29,5 +29,6 @@ TS_TUP_BEGIN( trace_streams_type ) TS_TUP( WARP_SCHEDULER ), TS_TUP( SCOREBOARD ), TS_TUP( MEMORY_PARTITION_UNIT ), + TS_TUP( LIVENESS ), TS_TUP( NUM_TRACE_STREAMS ) TS_TUP_END( trace_streams_type ) -- cgit v1.3