From 2d8d4455aa710914e87c5611cbb71f9330cdbc73 Mon Sep 17 00:00:00 2001 From: Akshay Jain Date: Thu, 22 Mar 2018 03:53:50 -0400 Subject: Change 180 by jain156@akshayj-lt1 on 2017/03/30 11:48:07 Added Memory Access breakdown statistics. Divided INST stats into INST type and INST Mem accesses. --- src/gpgpu-sim/gpu-sim.cc | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 17f1714..121e079 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -445,7 +445,7 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) "1"); option_parser_register(opp, "-gpgpu_ptx_instruction_classification", OPT_INT32, &gpgpu_ptx_instruction_classification, - "if enabled will classify ptx instruction types per kernel (Max 255 kernels now)", + "if enabled will classify ptx instruction types per kernel (Max 1024 kernels now)", "0"); option_parser_register(opp, "-gpgpu_ptx_sim_mode", OPT_INT32, &g_ptx_sim_mode, "Select between Performance (default) or Functional simulation (1)", @@ -1076,8 +1076,9 @@ void gpgpu_sim::gpu_print_stat() insn_warp_occ_print(stdout); } if ( gpgpu_ptx_instruction_classification ) { - StatDisp( g_inst_classification_stat[g_ptx_kernel_count]); - StatDisp( g_inst_op_classification_stat[g_ptx_kernel_count]); + StatDisp( g_inst_classification_stat[g_ptx_kernel_count_prev]); + StatDisp( g_inst_mem_classification_stat[g_ptx_kernel_count_prev]); + StatDisp( g_inst_op_classification_stat[g_ptx_kernel_count_prev]); } #ifdef GPGPUSIM_POWER_MODEL -- cgit v1.3 From 777ab7fd6761a6250bc8e4f37994125a3f8d331b Mon Sep 17 00:00:00 2001 From: Akshay Jain Date: Thu, 22 Mar 2018 04:20:13 -0400 Subject: Change 252 by jain156@akshayj-lt1 on 2017/05/29 10:51:32 Checking in the data footprint stats as reported in the micro paper. The implementation uses gpgpu pointer, due to which I don't feel fine pushing to mainline. May be I can do the stats check and increment in l2cache.cc through a public function to avoid exposing gpgpu pointer to public --- src/gpgpu-sim/gpu-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.h | 3 +++ src/gpgpu-sim/l2cache.cc | 20 ++++++++++++++++---- src/gpgpu-sim/l2cache.h | 8 ++++++-- src/gpgpusim_entrypoint.cc | 1 + 5 files changed, 27 insertions(+), 7 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 121e079..9a28b55 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -668,7 +668,7 @@ gpgpu_sim::gpgpu_sim( const gpgpu_sim_config &config ) m_memory_partition_unit = new memory_partition_unit*[m_memory_config->m_n_mem]; m_memory_sub_partition = new memory_sub_partition*[m_memory_config->m_n_mem_sub_partition]; for (unsigned i=0;im_n_mem;i++) { - m_memory_partition_unit[i] = new memory_partition_unit(i, m_memory_config, m_memory_stats); + m_memory_partition_unit[i] = new memory_partition_unit(i, m_memory_config, m_memory_stats, this); for (unsigned p = 0; p < m_memory_config->m_n_sub_partition_per_memory_channel; p++) { unsigned submpid = i * m_memory_config->m_n_sub_partition_per_memory_channel + p; m_memory_sub_partition[submpid] = m_memory_partition_unit[i]->get_sub_partition(p); diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 1778008..f9b5dad 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -36,6 +36,7 @@ #include #include #include +#include #include @@ -428,6 +429,8 @@ public: void perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ); + std::unordered_set data_footprint_stats; + //The next three functions added to be used by the functional simulation function //! Get shader core configuration diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index b1465a8..a4aaf28 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -62,15 +62,16 @@ mem_fetch * partition_mf_allocator::alloc(new_addr_type addr, mem_access_type ty memory_partition_unit::memory_partition_unit( unsigned partition_id, const struct memory_config *config, - class memory_stats_t *stats ) -: m_id(partition_id), m_config(config), m_stats(stats), m_arbitration_metadata(config) + class memory_stats_t *stats, + class gpgpu_sim *gpu) +: m_id(partition_id), m_config(config), m_stats(stats), m_arbitration_metadata(config), m_gpu(gpu) { m_dram = new dram_t(m_id,m_config,m_stats,this); m_sub_partition = new memory_sub_partition*[m_config->m_n_sub_partition_per_memory_channel]; for (unsigned p = 0; p < m_config->m_n_sub_partition_per_memory_channel; p++) { unsigned sub_partition_id = m_id * m_config->m_n_sub_partition_per_memory_channel + p; - m_sub_partition[p] = new memory_sub_partition(sub_partition_id, m_config, stats); + m_sub_partition[p] = new memory_sub_partition(sub_partition_id, m_config, stats, m_gpu); } } @@ -310,11 +311,13 @@ void memory_partition_unit::print( FILE *fp ) const memory_sub_partition::memory_sub_partition( unsigned sub_partition_id, const struct memory_config *config, - class memory_stats_t *stats ) + class memory_stats_t *stats , + class gpgpu_sim *gpu) { m_id = sub_partition_id; m_config=config; m_stats=stats; + m_gpu=gpu; m_memcpy_cycle_offset = 0; assert(m_id < m_config->m_n_mem_sub_partition); @@ -410,6 +413,15 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) bool read_sent = was_read_sent(events); MEM_SUBPART_DPRINTF("Probing L2 cache Address=%llx, status=%u\n", mf->get_addr(), status); + if ( (mf->get_access_type() == GLOBAL_ACC_R) || + (mf->get_access_type() == GLOBAL_ACC_W) || + (mf->get_access_type() == LOCAL_ACC_R) || + (mf->get_access_type() == LOCAL_ACC_W) || + (mf->get_access_type() == CONST_ACC_R) ) { + if (!m_gpu->data_footprint_stats.count(mf->get_addr())) + m_gpu->data_footprint_stats.insert(mf->get_addr()); + } + if ( status == HIT ) { if( !write_sent ) { // L2 cache replies diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 2d13918..6aa7505 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -58,7 +58,7 @@ private: class memory_partition_unit { public: - memory_partition_unit( unsigned partition_id, const struct memory_config *config, class memory_stats_t *stats ); + memory_partition_unit( unsigned partition_id, const struct memory_config *config, class memory_stats_t *stats , class gpgpu_sim *gpu); ~memory_partition_unit(); bool busy() const; @@ -93,6 +93,8 @@ public: unsigned get_mpid() const { return m_id; } + gpgpu_sim *m_gpu; + private: unsigned m_id; @@ -145,7 +147,7 @@ private: class memory_sub_partition { public: - memory_sub_partition( unsigned sub_partition_id, const struct memory_config *config, class memory_stats_t *stats ); + memory_sub_partition( unsigned sub_partition_id, const struct memory_config *config, class memory_stats_t *stats, class gpgpu_sim *gpu); ~memory_sub_partition(); unsigned get_id() const { return m_id; } @@ -185,6 +187,8 @@ public: m_memcpy_cycle_offset += 1; } + gpgpu_sim *m_gpu; + private: // data unsigned m_id; //< the global sub partition ID diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index ad4587a..1ee980a 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -263,6 +263,7 @@ void print_simulation_time() s = difference - 60*(m + 60*(h + 24*d)); fflush(stderr); + printf("gpgpu_data_footprint = %u requests at L2\n", g_the_gpu->data_footprint_stats.size()); printf("\n\ngpgpu_simulation_time = %u days, %u hrs, %u min, %u sec (%u sec)\n", (unsigned)d, (unsigned)h, (unsigned)m, (unsigned)s, (unsigned)difference ); printf("gpgpu_simulation_rate = %u (inst/sec)\n", (unsigned)(g_the_gpu->gpu_tot_sim_insn / difference) ); -- cgit v1.3 From 0241bdee08d2bdf95b7f6f9518403c22f8f0bc53 Mon Sep 17 00:00:00 2001 From: Akshay Jain Date: Thu, 22 Mar 2018 04:37:36 -0400 Subject: Change 257 by jain156@akshayj-lt1 on 2017/06/03 16:32:52 Adding gpuwattch power stats to the output. Enabled inst and mem breakdown stats in config file. --- configs/GTX480/gpgpusim.config | 2 +- src/gpgpu-sim/gpu-sim.cc | 1 + 2 files changed, 2 insertions(+), 1 deletion(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/GTX480/gpgpusim.config b/configs/GTX480/gpgpusim.config index 03fcda1..bff5d16 100644 --- a/configs/GTX480/gpgpusim.config +++ b/configs/GTX480/gpgpusim.config @@ -1,5 +1,5 @@ # functional simulator specification --gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_instruction_classification 1 -gpgpu_ptx_sim_mode 0 -gpgpu_ptx_force_max_capability 20 diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 9a28b55..fb57a1e 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1024,6 +1024,7 @@ void gpgpu_sim::gpu_print_stat() shader_print_scheduler_stat( stdout, false ); m_shader_stats->print(stdout); + m_power_stats->print(stdout); #ifdef GPGPUSIM_POWER_MODEL if(m_config.g_power_simulation_enabled){ m_gpgpusim_wrapper->print_power_kernel_stats(gpu_sim_cycle, gpu_tot_sim_cycle, gpu_tot_sim_insn + gpu_sim_insn, kernel_info_str, true ); -- cgit v1.3