diff options
| -rw-r--r-- | src/gpgpu-sim/gpu-cache.cc | 67 | ||||
| -rw-r--r-- | src/gpgpu-sim/gpu-cache.h | 69 | ||||
| -rw-r--r-- | src/gpgpu-sim/gpu-sim.cc | 8 | ||||
| -rw-r--r-- | src/gpgpu-sim/l2cache.cc | 38 | ||||
| -rw-r--r-- | src/gpgpu-sim/l2cache.h | 3 | ||||
| -rw-r--r-- | src/gpgpu-sim/mem_latency_stat.cc | 6 | ||||
| -rw-r--r-- | src/gpgpu-sim/mem_latency_stat.h | 8 |
7 files changed, 147 insertions, 52 deletions
diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index d050946..0ef6452 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -769,11 +769,7 @@ void cache_stats::print_stats(FILE *fout, const char *cache_name) const{ mem_access_type_str((enum mem_access_type)type), cache_request_status_str((enum cache_request_status)status), m_stats[type][status]); - fprintf(fout, "\t%s[%s][%s] = %u\n", - m_cache_name.c_str(), - mem_access_type_str((enum mem_access_type)type), - cache_request_status_str((enum cache_request_status)status), - m_stats_pw[type][status]); + if(status != RESERVATION_FAIL) total_access[type]+= m_stats[type][status]; } @@ -834,23 +830,6 @@ unsigned cache_stats::get_stats(enum mem_access_type *access_type, unsigned num_ return total; } -unsigned cache_stats::get_stats_pw(enum mem_access_type *access_type, unsigned num_access_type, enum cache_request_status *access_status, unsigned num_access_status) const{ - /// - /// Returns a sum of the stats corresponding to each "access_type" and "access_status" pair. - /// "access_type" is an array of "num_access_type" mem_access_types. - /// "access_status" is an array of "num_access_status" cache_request_statuses. - /// - unsigned total=0; - for(unsigned type =0; type < num_access_type; ++type){ - for(unsigned status=0; status < num_access_status; ++status){ - if(!check_valid((int)access_type[type], (int)access_status[status])) - assert(0 && "Unknown cache access type or access outcome"); - total += m_stats_pw[access_type[type]][access_status[status]]; - } - } - return total; -} - void cache_stats::get_sub_stats(struct cache_sub_stats &css) const{ /// /// Overwrites "css" with the appropriate statistics from this cache. @@ -881,11 +860,11 @@ void cache_stats::get_sub_stats(struct cache_sub_stats &css) const{ css = t_css; } -void cache_stats::get_sub_stats_pw(struct cache_sub_stats &css) const{ +void cache_stats::get_sub_stats_pw(struct cache_sub_stats_pw &css) const{ /// /// Overwrites "css" with the appropriate statistics from this cache. /// - struct cache_sub_stats t_css; + struct cache_sub_stats_pw t_css; t_css.clear(); for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { @@ -893,23 +872,43 @@ void cache_stats::get_sub_stats_pw(struct cache_sub_stats &css) const{ if(status == HIT || status == MISS || status == SECTOR_MISS || status == HIT_RESERVED) t_css.accesses += m_stats_pw[type][status]; - if(status == MISS || status == SECTOR_MISS) - t_css.misses += m_stats_pw[type][status]; + if(status == HIT){ + if(type == GLOBAL_ACC_R || type == CONST_ACC_R || type == INST_ACC_R){ + t_css.read_hits += m_stats_pw[type][status]; + } else if(type == GLOBAL_ACC_W){ + t_css.write_hits += m_stats_pw[type][status]; + } + } - if(status == HIT_RESERVED) - t_css.pending_hits += m_stats_pw[type][status]; + if(status == MISS || status == SECTOR_MISS){ + if(type == GLOBAL_ACC_R || type == CONST_ACC_R || type == INST_ACC_R){ + t_css.read_misses += m_stats_pw[type][status]; + } else if(type == GLOBAL_ACC_W){ + t_css.write_misses += m_stats_pw[type][status]; + } + } - if(status == RESERVATION_FAIL) - t_css.res_fails += m_stats_pw[type][status]; + if(status == HIT_RESERVED){ + if(type == GLOBAL_ACC_R || type == CONST_ACC_R || type == INST_ACC_R){ + t_css.read_pending_hits += m_stats_pw[type][status]; + } else if(type == GLOBAL_ACC_W){ + t_css.write_pending_hits += m_stats_pw[type][status]; + } + } + + if(status == RESERVATION_FAIL){ + if(type == GLOBAL_ACC_R || type == CONST_ACC_R || type == INST_ACC_R){ + t_css.read_res_fails += m_stats_pw[type][status]; + } else if(type == GLOBAL_ACC_W){ + t_css.write_res_fails += m_stats_pw[type][status]; + } + } } } - t_css.port_available_cycles = m_cache_port_available_cycles; - t_css.data_port_busy_cycles = m_cache_data_port_busy_cycles; - t_css.fill_port_busy_cycles = m_cache_fill_port_busy_cycles; - css = t_css; } + bool cache_stats::check_valid(int type, int status) const{ /// /// Verify a valid access_type/access_status diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index c1061ef..5449db7 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -968,6 +968,66 @@ struct cache_sub_stats{ void print_port_stats(FILE *fout, const char *cache_name) const; }; + +// Used for collecting AerialVision per-window statistics +struct cache_sub_stats_pw{ + unsigned accesses; + unsigned write_misses; + unsigned write_hits; + unsigned write_pending_hits; + unsigned write_res_fails; + + unsigned read_misses; + unsigned read_hits; + unsigned read_pending_hits; + unsigned read_res_fails; + + cache_sub_stats_pw(){ + clear(); + } + void clear(){ + accesses = 0; + write_misses = 0; + write_hits = 0; + write_pending_hits = 0; + write_res_fails = 0; + read_misses = 0; + read_hits = 0; + read_pending_hits = 0; + read_res_fails = 0; + } + cache_sub_stats_pw &operator+=(const cache_sub_stats_pw &css){ + /// + /// Overloading += operator to easily accumulate stats + /// + accesses += css.accesses; + write_misses += css.write_misses; + read_misses += css.read_misses; + write_pending_hits += css.write_pending_hits; + read_pending_hits += css.read_pending_hits; + write_res_fails += css.write_res_fails; + read_res_fails += css.read_res_fails; + return *this; + } + + cache_sub_stats_pw operator+(const cache_sub_stats_pw &cs){ + /// + /// Overloading + operator to easily accumulate stats + /// + cache_sub_stats_pw ret; + ret.accesses = accesses + cs.accesses; + ret.write_misses = write_misses + cs.write_misses; + ret.read_misses = read_misses + cs.read_misses; + ret.write_pending_hits = write_pending_hits + cs.write_pending_hits; + ret.read_pending_hits = read_pending_hits + cs.read_pending_hits; + ret.write_res_fails = write_res_fails + cs.write_res_fails; + ret.read_res_fails = read_res_fails + cs.read_res_fails; + return ret; + } + +}; + + /// /// Cache_stats /// Used to record statistics for each cache. @@ -996,8 +1056,7 @@ public: void get_sub_stats(struct cache_sub_stats &css) const; // Get per-window cache stats for AerialVision - unsigned get_stats_pw(enum mem_access_type *access_type, unsigned num_access_type, enum cache_request_status *access_status, unsigned num_access_status) const; - void get_sub_stats_pw(struct cache_sub_stats &css) const; + void get_sub_stats_pw(struct cache_sub_stats_pw &css) const; void sample_cache_port_utility(bool data_port_busy, bool fill_port_busy); private: @@ -1092,8 +1151,12 @@ public: void get_sub_stats(struct cache_sub_stats &css) const { m_stats.get_sub_stats(css); } + // Clear per-window stats for AerialVision support + void clear_pw(){ + m_stats.clear_pw(); + } // Per-window sub stats for AerialVision support - void get_sub_stats_pw(struct cache_sub_stats &css) const { + void get_sub_stats_pw(struct cache_sub_stats_pw &css) const { m_stats.get_sub_stats_pw(css); } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index c253367..56b9681 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -917,6 +917,7 @@ void gpgpu_sim::update_stats() { void gpgpu_sim::print_stats() { ptx_file_line_stats_write_file(); + visualizer_printstat(); gpu_print_stat(); if (g_network_mode) { @@ -1148,9 +1149,6 @@ void gpgpu_sim::gpu_print_stat() cache_stats l2_stats; struct cache_sub_stats l2_css; struct cache_sub_stats total_l2_css; - cache_stats l2_stats_pw; - struct cache_sub_stats l2_css_pw; - struct cache_sub_stats total_l2_css_pw; l2_stats.clear(); l2_css.clear(); total_l2_css.clear(); @@ -1159,15 +1157,11 @@ void gpgpu_sim::gpu_print_stat() for (unsigned i=0;i<m_memory_config->m_n_mem_sub_partition;i++){ m_memory_sub_partition[i]->accumulate_L2cache_stats(l2_stats); m_memory_sub_partition[i]->get_L2cache_sub_stats(l2_css); - m_memory_sub_partition[i]->get_L2cache_sub_stats_pw(l2_css_pw); fprintf( stdout, "L2_cache_bank[%d]: Access = %u, Miss = %u, Miss_rate = %.3lf, Pending_hits = %u, Reservation_fails = %u\n", i, l2_css.accesses, l2_css.misses, (double)l2_css.misses / (double)l2_css.accesses, l2_css.pending_hits, l2_css.res_fails); - fprintf( stdout, "L2_cache_bank[%d]: Access = %u, Miss = %u, Miss_rate = %.3lf, Pending_hits = %u, Reservation_fails = %u\n", - i, l2_css_pw.accesses, l2_css_pw.misses, (double)l2_css_pw.misses / (double)l2_css_pw.accesses, l2_css_pw.pending_hits, l2_css_pw.res_fails); total_l2_css += l2_css; - total_l2_css_pw += l2_css_pw; } if (!m_memory_config->m_L2_config.disabled() && m_memory_config->m_L2_config.get_num_lines()) { //L2c_print_cache_stat(); diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index a662446..0237525 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -519,14 +519,23 @@ void memory_sub_partition::print( FILE *fp ) const void memory_stats_t::visualizer_print( gzFile visualizer_file ) { - // gzprintf(visualizer_file, "Ltwowritemiss: %d\n", L2_write_miss); - // gzprintf(visualizer_file, "Ltwowritehit: %d\n", L2_write_access-L2_write_miss); - // gzprintf(visualizer_file, "Ltworeadmiss: %d\n", L2_read_miss); - // gzprintf(visualizer_file, "Ltworeadhit: %d\n", L2_read_access-L2_read_miss); + gzprintf(visualizer_file, "Ltwowritemiss: %d\n", L2_write_miss); + gzprintf(visualizer_file, "Ltwowritehit: %d\n", L2_write_hit); + gzprintf(visualizer_file, "Ltworeadmiss: %d\n", L2_read_miss); + gzprintf(visualizer_file, "Ltworeadhit: %d\n", L2_read_hit); + clear_L2_stats_pw(); + if (num_mfs) gzprintf(visualizer_file, "averagemflatency: %lld\n", mf_total_lat/num_mfs); } +void memory_stats_t::clear_L2_stats_pw(){ + L2_write_miss = 0; + L2_write_hit = 0; + L2_read_miss = 0; + L2_read_hit = 0; +} + void gpgpu_sim::print_dram_stats(FILE *fout) const { unsigned cmd=0; @@ -718,13 +727,28 @@ void memory_sub_partition::get_L2cache_sub_stats(struct cache_sub_stats &css) co } } -void memory_sub_partition::get_L2cache_sub_stats_pw(struct cache_sub_stats &css) const{ +void memory_sub_partition::get_L2cache_sub_stats_pw(struct cache_sub_stats_pw &css) const{ if (!m_config->m_L2_config.disabled()) { m_L2cache->get_sub_stats_pw(css); } } + +void memory_sub_partition::clear_L2cache_stats_pw() { + if (!m_config->m_L2_config.disabled()) { + m_L2cache->clear_pw(); + } +} + void memory_sub_partition::visualizer_print( gzFile visualizer_file ) { - // TODO: Add visualizer stats for L2 cache -} + // TODO: Add visualizer stats for L2 cache + cache_sub_stats_pw temp_sub_stats; + get_L2cache_sub_stats_pw(temp_sub_stats); + m_stats->L2_read_miss += temp_sub_stats.read_misses; + m_stats->L2_write_miss += temp_sub_stats.write_misses; + m_stats->L2_read_hit += temp_sub_stats.read_hits; + m_stats->L2_write_hit += temp_sub_stats.write_hits; + + clear_L2cache_stats_pw(); +} diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 57f3e94..beafdd3 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -181,7 +181,8 @@ public: void get_L2cache_sub_stats(struct cache_sub_stats &css) const; // Support for getting per-window L2 stats for AerialVision - void get_L2cache_sub_stats_pw(struct cache_sub_stats &css) const; + void get_L2cache_sub_stats_pw(struct cache_sub_stats_pw &css) const; + void clear_L2cache_stats_pw(); void force_l2_tag_update(new_addr_type addr, unsigned time, mem_access_sector_mask_t mask) { diff --git a/src/gpgpu-sim/mem_latency_stat.cc b/src/gpgpu-sim/mem_latency_stat.cc index c5452b9..49abae4 100644 --- a/src/gpgpu-sim/mem_latency_stat.cc +++ b/src/gpgpu-sim/mem_latency_stat.cc @@ -129,6 +129,12 @@ memory_stats_t::memory_stats_t( unsigned n_shader, const struct shader_core_conf } } + // AerialVision L2 stats + L2_read_miss = 0; + L2_write_miss = 0; + L2_read_hit = 0; + L2_write_hit = 0; + L2_cbtoL2length = (unsigned int*) calloc(mem_config->m_n_mem, sizeof(unsigned int)); L2_cbtoL2writelength = (unsigned int*) calloc(mem_config->m_n_mem, sizeof(unsigned int)); L2_L2tocblength = (unsigned int*) calloc(mem_config->m_n_mem, sizeof(unsigned int)); diff --git a/src/gpgpu-sim/mem_latency_stat.h b/src/gpgpu-sim/mem_latency_stat.h index 5b89202..982b9ae 100644 --- a/src/gpgpu-sim/mem_latency_stat.h +++ b/src/gpgpu-sim/mem_latency_stat.h @@ -47,6 +47,9 @@ public: void visualizer_print( gzFile visualizer_file ); + // Reset local L2 stats that are aggregated each sampling window + void clear_L2_stats_pw(); + unsigned m_n_shader; const struct shader_core_config *m_shader_config; @@ -84,6 +87,11 @@ public: unsigned ***mem_access_type_stats; // dram access type classification + // AerialVision L2 stats + unsigned L2_read_miss; + unsigned L2_write_miss; + unsigned L2_read_hit; + unsigned L2_write_hit; // L2 cache stats unsigned int *L2_cbtoL2length; |