From 8c81c1d04e8d20b08f122a12ce090b4f926adb4c Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 22 Aug 2018 10:19:52 -0400 Subject: adding lazy-fetch-on-read and invalidate operation to cache --- src/gpgpu-sim/l2cache.cc | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) (limited to 'src/gpgpu-sim/l2cache.cc') diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index b1465a8..359d3c8 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -428,7 +428,7 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) m_icnt_L2_queue->pop(); } } else if ( status != RESERVATION_FAIL ) { - if(mf->is_write() && m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE && !was_writeallocate_sent(events)) { + if(mf->is_write() && (m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE || m_config->m_L2_config.m_write_alloc_policy == LAZY_FETCH_ON_READ) && !was_writeallocate_sent(events)) { mf->set_reply(); mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); m_L2_icnt_queue->push(mf); @@ -568,7 +568,15 @@ unsigned memory_sub_partition::flushL2() if (!m_config->m_L2_config.disabled()) { m_L2cache->flush(); } - return 0; // L2 is read only in this version + return 0; //TODO: write the flushed data to the main memory +} + +unsigned memory_sub_partition::invalidateL2() +{ + if (!m_config->m_L2_config.disabled()) { + m_L2cache->invalidate(); + } + return 0; } bool memory_sub_partition::busy() const -- cgit v1.3 From 5e7bd910c07c066d5d1cc4b12f8aa7abefcdb411 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 27 Aug 2018 16:13:13 -0400 Subject: improving code quality --- src/gpgpu-sim/gpu-cache.cc | 72 +++++++++++++++++++++++----------------------- src/gpgpu-sim/gpu-cache.h | 8 ++++++ src/gpgpu-sim/l2cache.cc | 9 +++--- src/gpgpu-sim/mem_fetch.h | 10 +++++-- 4 files changed, 57 insertions(+), 42 deletions(-) (limited to 'src/gpgpu-sim/l2cache.cc') diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 9d81de9..f181f20 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -916,8 +916,8 @@ void baseline_cache::cycle(){ void baseline_cache::fill(mem_fetch *mf, unsigned time){ if(m_config.m_mshr_type == SECTOR_ASSOC) { - assert(mf->original_mf); - extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf->original_mf); + assert(mf->get_original_mf()); + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf->get_original_mf()); assert( e != m_extra_mf_fields.end() ); e->second.pending_read--; @@ -927,7 +927,7 @@ void baseline_cache::fill(mem_fetch *mf, unsigned time){ return; } else { mem_fetch *temp = mf; - mf = mf->original_mf; + mf = mf->get_original_mf(); delete temp; } } @@ -1294,45 +1294,45 @@ data_cache::wr_miss_wa_lazy_fetch_on_read( new_addr_type addr, new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); - //if the request writes to the whole cache line/sector, then, write and set cache line Modified. - //and no need to send read request to memory or reserve mshr + //if the request writes to the whole cache line/sector, then, write and set cache line Modified. + //and no need to send read request to memory or reserve mshr - if(miss_queue_full(0)) { - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); - return RESERVATION_FAIL; // cannot handle request this cycle - } + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + return RESERVATION_FAIL; // cannot handle request this cycle + } - bool wb = false; - evicted_block_info evicted; + bool wb = false; + evicted_block_info evicted; - cache_request_status m_status = m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); - assert(m_status != HIT); - cache_block_t* block = m_tag_array->get_block(cache_index); - block->set_status(MODIFIED, mf->get_access_sector_mask()); - if(m_status == HIT_RESERVED) { - block->set_ignore_on_fill(true, mf->get_access_sector_mask()); - block->set_modified_on_fill(true, mf->get_access_sector_mask()); - } + cache_request_status m_status = m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); + assert(m_status != HIT); + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); + if(m_status == HIT_RESERVED) { + block->set_ignore_on_fill(true, mf->get_access_sector_mask()); + block->set_modified_on_fill(true, mf->get_access_sector_mask()); + } - if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) - { - block->set_m_readable(true, mf->get_access_sector_mask()); - } else - { - block->set_m_readable(false, mf->get_access_sector_mask()); - } + if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) + { + block->set_m_readable(true, mf->get_access_sector_mask()); + } else + { + block->set_m_readable(false, mf->get_access_sector_mask()); + } - if( m_status != RESERVATION_FAIL ){ - // If evicted block is modified and not a write-through - // (already modified lower level) - if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) { - mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, - m_wrbk_type,evicted.m_modified_size,true); - send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); - } - return MISS; + if( m_status != RESERVATION_FAIL ){ + // If evicted block is modified and not a write-through + // (already modified lower level) + if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) { + mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, + m_wrbk_type,evicted.m_modified_size,true); + send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); } - return RESERVATION_FAIL; + return MISS; + } + return RESERVATION_FAIL; } /// No write-allocate miss: Simply send write request to lower level memory diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 96b9a6d..dee43f2 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -355,15 +355,18 @@ struct sector_cache_block : public cache_block_t { return m_status[sidx]; } + virtual void set_status(enum cache_block_state status, mem_access_sector_mask_t sector_mask) { unsigned sidx = get_sector_index(sector_mask); m_status[sidx] = status; } + virtual unsigned get_last_access_time() { return m_line_last_access_time; } + virtual void set_last_access_time(unsigned time, mem_access_sector_mask_t sector_mask) { unsigned sidx = get_sector_index(sector_mask); @@ -371,25 +374,30 @@ struct sector_cache_block : public cache_block_t { m_last_sector_access_time[sidx] = time; m_line_last_access_time = time; } + virtual unsigned get_alloc_time() { return m_line_alloc_time; } + virtual void set_ignore_on_fill(bool m_ignore, mem_access_sector_mask_t sector_mask) { unsigned sidx = get_sector_index(sector_mask); m_ignore_on_fill_status[sidx] = m_ignore; } + virtual void set_modified_on_fill(bool m_modified, mem_access_sector_mask_t sector_mask) { unsigned sidx = get_sector_index(sector_mask); m_set_modified_on_fill[sidx] = m_modified; } + virtual void set_m_readable(bool readable, mem_access_sector_mask_t sector_mask) { unsigned sidx = get_sector_index(sector_mask); m_readable[sidx] = readable; } + virtual bool is_readable(mem_access_sector_mask_t sector_mask) { unsigned sidx = get_sector_index(sector_mask); return m_readable[sidx]; diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index 359d3c8..f42610f 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -362,10 +362,11 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) }else{ if(m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE) { - assert(mf->original_wr_mf); - mf->original_wr_mf->set_reply(); - mf->original_wr_mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); - m_L2_icnt_queue->push(mf->original_wr_mf); + mem_fetch* original_wr_mf = mf->get_original_wr_mf(); + assert(original_wr_mf); + original_wr_mf->set_reply(); + original_wr_mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); + m_L2_icnt_queue->push(original_wr_mf); } m_request_tracker.erase(mf); delete mf; diff --git a/src/gpgpu-sim/mem_fetch.h b/src/gpgpu-sim/mem_fetch.h index 278cf32..da67d49 100644 --- a/src/gpgpu-sim/mem_fetch.h +++ b/src/gpgpu-sim/mem_fetch.h @@ -115,8 +115,10 @@ public: const memory_config *get_mem_config(){return m_mem_config;} unsigned get_num_flits(bool simt_to_mem); - mem_fetch* original_mf; - mem_fetch* original_wr_mf; + + mem_fetch* get_original_mf() { return original_mf; } + mem_fetch* get_original_wr_mf() { return original_wr_mf; } + private: // request source information unsigned m_request_uid; @@ -148,6 +150,10 @@ private: const class memory_config *m_mem_config; unsigned icnt_flit_size; + + mem_fetch* original_mf; //this pointer is set up when a request is divided into sector requests at L2 cache (if the req size > L2 sector size), so the pointer refers to the original request + mem_fetch* original_wr_mf; //this pointer refers to the original write req, when fetch-on-write policy is used + }; #endif -- cgit v1.3 From f852a7108e691045dd3910065836a817babcde8c Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 7 Sep 2018 22:01:31 -0400 Subject: adding streamin cache + fixing TEX cache + adding l1 latency and smem latency --- configs/4.x-cfgs/SM2_GTX480/gpgpusim.config | 2 +- configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt | 73 --------- configs/4.x-cfgs/SM6_P100/gpgpusim.config | 174 --------------------- .../SM6_P100_64SMs/config_fermi_islip.icnt | 73 --------- configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config | 174 --------------------- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 27 ++-- src/abstract_hardware_model.h | 1 + src/gpgpu-sim/gpu-cache.cc | 20 ++- src/gpgpu-sim/gpu-cache.h | 36 ++++- src/gpgpu-sim/gpu-sim.cc | 6 + src/gpgpu-sim/l2cache.cc | 8 +- src/gpgpu-sim/shader.cc | 126 ++++++++++++++- src/gpgpu-sim/shader.h | 6 + 13 files changed, 207 insertions(+), 519 deletions(-) delete mode 100644 configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM6_P100/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config (limited to 'src/gpgpu-sim/l2cache.cc') diff --git a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config index 7f8da49..35341f7 100644 --- a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config @@ -67,7 +67,7 @@ -memory_partition_indexing 0 -gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,S:2:32,4 --gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,T:128:4,128:2 -gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,S:2:32,4 # enable operand collector diff --git a/configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt deleted file mode 100644 index e7c2c3b..0000000 --- a/configs/4.x-cfgs/SM6_P100/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//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 = 60; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 128; -input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; - -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 = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// 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/4.x-cfgs/SM6_P100/gpgpusim.config b/configs/4.x-cfgs/SM6_P100/gpgpusim.config deleted file mode 100644 index a4e745d..0000000 --- a/configs/4.x-cfgs/SM6_P100/gpgpusim.config +++ /dev/null @@ -1,174 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 32 --gpgpu_n_sub_partition_per_mchannel 1 - -# Pscal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Nvidia_Tesla --gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# 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_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,2,2,1,2,2,2,1,6 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 2 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 2,2,2,2,130 --ptx_opcode_latency_sfu 8 --ptx_opcode_initiation_sfu 4 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 65536 --gpgpu_shmem_size_PrefShared 65536 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 --memory_partition_indexing 2 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# 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 Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt deleted file mode 100644 index 81153b0..0000000 --- a/configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt +++ /dev/null @@ -1,73 +0,0 @@ -//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 = 64; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 128; -input_buffer_size = 256; -ejection_buffer_size = 128; -boundary_buffer_size = 128; - -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 = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// 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/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config b/configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config deleted file mode 100644 index edcd919..0000000 --- a/configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config +++ /dev/null @@ -1,174 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 60 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 32 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 32 --gpgpu_n_sub_partition_per_mchannel 1 - -# Pscal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Nvidia_Tesla --gpgpu_clock_domains 1480.0:1480.0:1480.0:715.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 - -# 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_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 2,2,2,1,2,2,2,1,6 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 2 --gpgpu_num_dp_units 2 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 2,2,2,2,130 --ptx_opcode_latency_sfu 8 --ptx_opcode_initiation_sfu 4 - - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP100 has 64KB Shared memory --gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 65536 --gpgpu_shmem_size_PrefShared 65536 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 32 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 4MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 --perf_sim_memcpy 0 --memory_partition_indexing 2 - -# 4 KB Inst. --gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 12 --gpgpu_operand_collector_num_units_sfu 6 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_units_dp 6 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 --gpgpu_operand_collector_num_in_ports_dp 1 --gpgpu_operand_collector_num_out_ports_dp 1 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 --gpgpu_coalesce_arch 60 - -## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units --gpgpu_max_insn_issue_per_warp 2 --gpgpu_dual_issue_diff_exec_units 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 192 - -# for HBM, 32 channles, each (128 bits) 16 bytes width --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 16 --gpgpu_dram_burst_length 2 --dram_data_command_freq_ratio 2 # HBM is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS - -# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) -# Timing for 1 GHZ -# tRRDl and tWTR are missing, need to be added -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: -# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" - -# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ --gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: - CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" - -# HBM has dual bus interface, in which it can issue two col and row commands at a time --dual_bus_interface 1 -# select lower bits for bnkgrp to increase bnkgrp parallelism --dram_bnk_indexing_policy 0 --dram_bnkgrp_indexing_policy 1 - -#-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 64:56:32 - -# Pascal has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# 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 Pascal 100 --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index 7368882..9ea7202 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -56,24 +56,29 @@ -ptx_opcode_initiation_sfu 4 -ptx_opcode_latency_sfu 8 + +# latencies and cache configs are adopted from: +# https://arxiv.org/pdf/1804.06826.pdf # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 24KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +# Pascal GP102 has 96KB Shared memory divided over 2 cores, each has 48KB +# Pascal GP102 has 2 banks L1 cache, where each is 24KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres are used +-gpgpu_cache:dl1 S:48:128:4,L:L:s:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:48:128:4,L:L:s:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:48:128:4,L:L:s:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 49152 -gpgpu_shmem_size_PrefL1 49152 -gpgpu_shmem_size_PrefShared 49152 -gmem_skip_L1D 1 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 82 +-smem_latency 24 # 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:128:4,16:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:256:64,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 -perf_sim_memcpy 0 @@ -81,8 +86,7 @@ # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 # 48 KB Tex -# this is unused --gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 # 12 KB Const -gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 @@ -99,7 +103,6 @@ -gpgpu_operand_collector_num_out_ports_mem 1 -gpgpu_operand_collector_num_in_ports_dp 1 -gpgpu_operand_collector_num_out_ports_dp 1 -# gpgpu_num_reg_banks should be increased to 32 -gpgpu_num_reg_banks 32 # shared memory bankconflict detection @@ -118,7 +121,7 @@ -inter_config_file config_fermi_islip.icnt # memory partition latency config --rop_latency 100 +-rop_latency 120 -dram_latency 100 # dram model config @@ -128,7 +131,7 @@ # To allow 100% DRAM utility, there should at least be enough buffer to sustain # the minimum DRAM latency (100 core cycles). I.e. # Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_frfcfs_dram_sched_queue_size 64 -gpgpu_dram_return_queue_size 240 # for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 1b764e2..a70b077 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -392,6 +392,7 @@ protected: #define LOCAL_MEM_SIZE_MAX (8*1024) #define MAX_STREAMING_MULTIPROCESSORS 64 #define MAX_THREAD_PER_SM 2048 +#define MAX_WARP_PER_SM 64 #define TOTAL_LOCAL_MEM_PER_SM (MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX) #define TOTAL_SHARED_MEM (MAX_STREAMING_MULTIPROCESSORS*SHARED_MEM_SIZE_MAX) #define TOTAL_LOCAL_MEM (MAX_STREAMING_MULTIPROCESSORS*MAX_THREAD_PER_SM*LOCAL_MEM_SIZE_MAX) diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index f181f20..a11853a 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -71,6 +71,7 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{ switch(m_set_index_function){ case FERMI_HASH_SET_FUNCTION: + case BITWISE_XORING_FUNCTION: /* * Set Indexing function from "A Detailed GPU Cache Model Based on Reuse Distance Theory" * Cedric Nugteren et al. @@ -1581,7 +1582,7 @@ enum cache_request_status tex_cache::access( new_addr_type addr, mem_fetch *mf, if ( status == MISS ) { // we need to send a memory request... unsigned rob_index = m_rob.push( rob_entry(cache_index, mf, block_addr) ); - m_extra_mf_fields[mf] = extra_mf_fields(rob_index); + m_extra_mf_fields[mf] = extra_mf_fields(rob_index, m_config); mf->set_data_size(m_config.get_line_sz()); m_tags.fill(cache_index,time,mf); // mark block as valid m_request_fifo.push(mf); @@ -1636,6 +1637,23 @@ void tex_cache::cycle(){ /// Place returning cache block into reorder buffer void tex_cache::fill( mem_fetch *mf, unsigned time ) { + if(m_config.m_mshr_type == SECTOR_TEX_FIFO) { + assert(mf->get_original_mf()); + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf->get_original_mf()); + assert( e != m_extra_mf_fields.end() ); + e->second.pending_read--; + + if(e->second.pending_read > 0) { + //wait for the other requests to come back + delete mf; + return; + } else { + mem_fetch *temp = mf; + mf = mf->get_original_mf(); + delete temp; + } + } + extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf); assert( e != m_extra_mf_fields.end() ); assert( e->second.m_valid ); diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index dee43f2..4ed382c 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -455,7 +455,8 @@ enum write_policy_t { enum allocation_policy_t { ON_MISS, - ON_FILL + ON_FILL, + STREAMING }; @@ -467,8 +468,9 @@ enum write_allocate_policy_t { }; enum mshr_config_t { - TEX_FIFO, + TEX_FIFO, // Tex cache ASSOC, // normal cache + SECTOR_TEX_FIFO, //Tex cache sends requests to high-level sector cache SECTOR_ASSOC // normal cache sends requests to high-level sector cache }; @@ -485,6 +487,12 @@ enum cache_type{ SECTOR }; +#define MAX_WARP_PER_SHADER 64 +#define INCT_TOTAL_BUFFER 64 +#define L2_TOTAL 64 +#define MAX_WARP_PER_SHADER 64 +#define MAX_WARP_PER_SHADER 64 + class cache_config { public: cache_config() @@ -544,10 +552,27 @@ public: switch (ap) { case 'm': m_alloc_policy = ON_MISS; break; case 'f': m_alloc_policy = ON_FILL; break; + case 's': m_alloc_policy = STREAMING; break; default: 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 the cache line. 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 (mimics) 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/ + + m_alloc_policy = ON_FILL; + m_mshr_entries = m_nset*m_assoc; + if(m_cache_type == SECTOR) + m_mshr_entries *= SECTOR_CHUNCK_SIZE; + m_mshr_max_merge = MAX_WARP_PER_SM; + } switch (mshr_type) { case 'F': m_mshr_type = TEX_FIFO; assert(ntok==14); break; + case 'T': m_mshr_type = SECTOR_TEX_FIFO; assert(ntok==14); break; case 'A': m_mshr_type = ASSOC; break; case 'S' : m_mshr_type = SECTOR_ASSOC; break; default: exit_parse_error(); @@ -722,6 +747,7 @@ class l1d_cache_config : public cache_config{ public: l1d_cache_config() : cache_config(){} virtual unsigned set_index(new_addr_type addr) const; + unsigned l1_latency; }; class l2_cache_config : public cache_config { @@ -1442,7 +1468,7 @@ public: m_result_fifo(config.m_result_fifo_entries) { m_name = name; - assert(config.m_mshr_type == TEX_FIFO); + assert(config.m_mshr_type == TEX_FIFO || config.m_mshr_type == SECTOR_TEX_FIFO ); assert(config.m_write_policy == READ_ONLY); assert(config.m_alloc_policy == ON_MISS); m_memport=memport; @@ -1595,13 +1621,15 @@ private: struct extra_mf_fields { extra_mf_fields() { m_valid = false;} - extra_mf_fields( unsigned i ) + extra_mf_fields( unsigned i, const cache_config &m_config ) { m_valid = true; m_rob_index = i; + pending_read = m_config.m_mshr_type == SECTOR_TEX_FIFO? m_config.m_line_sz/SECTOR_SIZE : 0; } bool m_valid; unsigned m_rob_index; + unsigned pending_read; }; cache_stats m_stats; diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index ea2dfba..08d4525 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -250,6 +250,12 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); + option_parser_register(opp, "-l1_latency", OPT_UINT32, &m_L1D_config.l1_latency, + "L1 Hit Latency", + "0"); + option_parser_register(opp, "-smem_latency", OPT_UINT32, &smem_latency, + "smem Latency", + "3"); option_parser_register(opp, "-gpgpu_cache:dl1PrefL1", OPT_CSTR, &m_L1D_config.m_config_stringPrefL1, "per-shader L1 data cache config " " {::,:::,::, | none}", diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index f42610f..25da107 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -609,7 +609,7 @@ std::vector memory_sub_partition::breakdown_request_to_sector_reques } else { printf("Invalid sector received, address = 0x%06x, sector mask = %s, data size = %d", - mf->get_addr(), mf->get_access_sector_mask(), mf->get_data_size(), mf->get_data_size()); + mf->get_addr(), mf->get_access_sector_mask(), mf->get_data_size()); assert(0 && "Undefined sector mask is received"); } @@ -640,7 +640,11 @@ std::vector memory_sub_partition::breakdown_request_to_sector_reques result.push_back(n_mf); byte_sector_mask <<= SECTOR_SIZE; } - } else assert(0 && "Undefined data size is received"); + } else { + printf("Invalid sector received, address = 0x%06x, sector mask = %d, byte mask = , data size = %d", + mf->get_addr(), mf->get_access_sector_mask().count(), mf->get_data_size()); + assert(0 && "Undefined data size is received"); + } return result; } diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index b59e5d2..51689e3 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1488,6 +1488,111 @@ mem_stage_stall_type ldst_unit::process_memory_access_queue( cache_t *cache, war return process_cache_access( cache, mf->get_addr(), inst, events, mf, status ); } +mem_stage_stall_type ldst_unit::process_memory_access_queue_l1cache( l1_cache *cache, warp_inst_t &inst ) +{ + mem_stage_stall_type result = NO_RC_FAIL; + if( inst.accessq_empty() ) + return result; + + mem_fetch *mf = m_mf_allocator->alloc(inst,inst.accessq_back()); + + if(m_config->m_L1D_config.l1_latency > 0) + { + if((l1_latency_queue[m_config->m_L1D_config.l1_latency-1]) == NULL) + { + l1_latency_queue[m_config->m_L1D_config.l1_latency-1] = mf; + + if( mf->get_inst().is_store() ) { + unsigned inc_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)? + (mf->get_data_size()/SECTOR_SIZE) : 1; + + for(unsigned i=0; i< inc_ack; ++i) + m_core->inc_store_req( inst.warp_id() ); + } + + inst.accessq_pop_back(); + } + else + { + result = BK_CONF; + delete mf; + } + if( !inst.accessq_empty() && result !=BK_CONF) + result = COAL_STALL; + return result; + } + else + { + std::list events; + enum cache_request_status status = cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle,events); + return process_cache_access( cache, mf->get_addr(), inst, events, mf, status ); + } +} + +void ldst_unit::L1_latency_queue_cycle() +{ + //std::deque< std::pair >::iterator it = m_latency_queue.begin(); + if((l1_latency_queue[0]) != NULL) + { + mem_fetch* mf_next = l1_latency_queue[0]; + std::list events; + enum cache_request_status status = m_L1D->access(mf_next->get_addr(),mf_next,gpu_sim_cycle+gpu_tot_sim_cycle,events); + + bool write_sent = was_write_sent(events); + bool read_sent = was_read_sent(events); + + if ( status == HIT ) { + assert( !read_sent ); + l1_latency_queue[0] = NULL; + if ( mf_next->get_inst().is_load() ) { + for ( unsigned r=0; r < 4; r++) + if (mf_next->get_inst().out[r] > 0) + { + assert(m_pending_writes[mf_next->get_inst().warp_id()][mf_next->get_inst().out[r]]>0); + unsigned still_pending = --m_pending_writes[mf_next->get_inst().warp_id()][mf_next->get_inst().out[r]]; + if(!still_pending) + { + m_pending_writes[mf_next->get_inst().warp_id()].erase(mf_next->get_inst().out[r]); + m_scoreboard->releaseRegister(mf_next->get_inst().warp_id(),mf_next->get_inst().out[r]); + m_core->warp_inst_complete(mf_next->get_inst()); + } + } + } + + //For write hit in WB policy + if(mf_next->get_inst().is_store() && !write_sent) + { + unsigned dec_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)? + (mf_next->get_data_size()/SECTOR_SIZE) : 1; + + mf_next->set_reply(); + + for(unsigned i=0; i< dec_ack; ++i) + m_core->store_ack(mf_next); + } + + if( !write_sent ) + delete mf_next; + + } else if ( status == RESERVATION_FAIL ) { + assert( !read_sent ); + assert( !write_sent ); + } else { + assert( status == MISS || status == HIT_RESERVED ); + l1_latency_queue[0] = NULL; + } + } + + for( unsigned stage = 0; stagem_L1D_config.l1_latency-1; ++stage) + if( l1_latency_queue[stage] == NULL) { + l1_latency_queue[stage] = l1_latency_queue[stage+1] ; + l1_latency_queue[stage+1] = NULL; + } + +} + + + bool ldst_unit::constant_cycle( warp_inst_t &inst, mem_stage_stall_type &rc_fail, mem_stage_access_type &fail_type) { if( inst.empty() || ((inst.space.get_type() != const_space) && (inst.space.get_type() != param_space_kernel)) ) @@ -1561,7 +1666,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea } } else { assert( CACHE_UNDEFINED != inst.cache_op ); - stall_cond = process_memory_access_queue(m_L1D,inst); + stall_cond = process_memory_access_queue_l1cache(m_L1D,inst); } if( !inst.accessq_empty() && stall_cond == NO_RC_FAIL) stall_cond = COAL_STALL; @@ -1771,8 +1876,9 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt, const memory_config *mem_config, shader_core_stats *stats, unsigned sid, - unsigned tpc ) : pipelined_simd_unit(NULL,config,3,core), m_next_wb(config) + unsigned tpc ) : pipelined_simd_unit(NULL,config,config->smem_latency,core), m_next_wb(config) { + assert(config->smem_latency > 1); init( icnt, mf_allocator, core, @@ -1793,6 +1899,12 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt, m_icnt, m_mf_allocator, IN_L1D_MISS_QUEUE ); + + if(m_config->m_L1D_config.l1_latency > 0) + { + for(int i=0; im_L1D_config.l1_latency; i++ ) + l1_latency_queue.push_back((mem_fetch*)NULL); + } } } @@ -2019,7 +2131,11 @@ void ldst_unit::cycle() m_L1T->cycle(); m_L1C->cycle(); - if( m_L1D ) m_L1D->cycle(); + if( m_L1D ) { + m_L1D->cycle(); + if(m_config->m_L1D_config.l1_latency > 0) + L1_latency_queue_cycle(); + } warp_inst_t &pipe_reg = *m_dispatch_reg; enum mem_stage_stall_type rc_fail = NO_RC_FAIL; @@ -2042,9 +2158,9 @@ void ldst_unit::cycle() unsigned warp_id = pipe_reg.warp_id(); if( pipe_reg.is_load() ) { if( pipe_reg.space.get_type() == shared_space ) { - if( m_pipeline_reg[2]->empty() ) { + if( m_pipeline_reg[m_config->smem_latency-1]->empty() ) { // new shared memory request - move_warp(m_pipeline_reg[2],m_dispatch_reg); + move_warp(m_pipeline_reg[m_config->smem_latency-1],m_dispatch_reg); m_dispatch_reg->clear(); } } else { diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index cc441b3..e07096e 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1220,6 +1220,7 @@ protected: mem_fetch *mf, enum cache_request_status status ); mem_stage_stall_type process_memory_access_queue( cache_t *cache, warp_inst_t &inst ); + mem_stage_stall_type process_memory_access_queue_l1cache( l1_cache *cache, warp_inst_t &inst ); const memory_config *m_memory_config; class mem_fetch_interface *m_icnt; @@ -1248,6 +1249,9 @@ protected: // for debugging unsigned long long m_last_inst_gpu_sim_cycle; unsigned long long m_last_inst_gpu_tot_sim_cycle; + + std::deque l1_latency_queue; + void L1_latency_queue_cycle(); }; enum pipeline_stage_name_t { @@ -1399,6 +1403,8 @@ struct shader_core_config : public core_config int simt_core_sim_order; + unsigned smem_latency; + unsigned mem2device(unsigned memid) const { return memid + n_simt_clusters; } //Jin: concurrent kernel on sm -- cgit v1.3 From 13f07a2c820422db7e4e88e43d692dfe8e1b8cad Mon Sep 17 00:00:00 2001 From: Nick Date: Tue, 19 Feb 2019 18:40:45 -0500 Subject: Add initial infrastrucutre to support L2 (and other) cache statistics for AerialVision --- src/gpgpu-sim/gpu-cache.cc | 80 ++++++++++++++++++++++++++++++++++++++++++++++ src/gpgpu-sim/gpu-cache.h | 14 ++++++++ src/gpgpu-sim/gpu-sim.cc | 7 ++++ src/gpgpu-sim/l2cache.cc | 5 +++ src/gpgpu-sim/l2cache.h | 3 ++ 5 files changed, 109 insertions(+) (limited to 'src/gpgpu-sim/l2cache.cc') diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index ba81440..d050946 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -600,9 +600,11 @@ void mshr_table::display( FILE *fp ) const{ /***************************************************************** Caches *****************************************************************/ cache_stats::cache_stats(){ m_stats.resize(NUM_MEM_ACCESS_TYPE); + m_stats_pw.resize(NUM_MEM_ACCESS_TYPE); m_fail_stats.resize(NUM_MEM_ACCESS_TYPE); for(unsigned i=0; iget_access_type(), m_stats.select_stats_status(status, cache_status)); + m_stats.inc_stats_pw(mf->get_access_type(), m_stats.select_stats_status(status, cache_status)); return cache_status; } @@ -1575,6 +1652,8 @@ data_cache::access( new_addr_type addr, = process_tag_probe( wr, probe_status, addr, cache_index, mf, time, events ); m_stats.inc_stats(mf->get_access_type(), m_stats.select_stats_status(probe_status, access_status)); + m_stats.inc_stats_pw(mf->get_access_type(), + m_stats.select_stats_status(probe_status, access_status)); return access_status; } @@ -1639,6 +1718,7 @@ enum cache_request_status tex_cache::access( new_addr_type addr, mem_fetch *mf, cache_status = HIT_RESERVED; } m_stats.inc_stats(mf->get_access_type(), m_stats.select_stats_status(status, cache_status)); + m_stats.inc_stats_pw(mf->get_access_type(), m_stats.select_stats_status(status, cache_status)); return cache_status; } diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index e663cf6..c1061ef 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -978,7 +978,11 @@ class cache_stats { public: cache_stats(); void clear(); + // Clear AerialVision cache stats after each window + void clear_pw(); void inc_stats(int access_type, int access_outcome); + // Increment AerialVision cache stats + void inc_stats_pw(int access_type, int access_outcome); void inc_fail_stats(int access_type, int fail_outcome); enum cache_request_status select_stats_status(enum cache_request_status probe, enum cache_request_status access) const; unsigned &operator()(int access_type, int access_outcome, bool fail_outcome); @@ -991,12 +995,18 @@ public: unsigned get_stats(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(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 sample_cache_port_utility(bool data_port_busy, bool fill_port_busy); private: bool check_valid(int type, int status) const; bool check_fail_valid(int type, int fail) const; std::vector< std::vector > m_stats; + // AerialVision cache stats (per-window) + std::vector< std::vector > m_stats_pw; std::vector< std::vector > m_fail_stats; unsigned long long m_cache_port_available_cycles; @@ -1082,6 +1092,10 @@ public: void get_sub_stats(struct cache_sub_stats &css) const { m_stats.get_sub_stats(css); } + // Per-window sub stats for AerialVision support + void get_sub_stats_pw(struct cache_sub_stats &css) const { + m_stats.get_sub_stats_pw(css); + } // accessors for cache bandwidth availability bool data_port_free() const { return m_bandwidth_management.data_port_free(); } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index ec570bf..c253367 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1148,6 +1148,9 @@ 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(); @@ -1156,11 +1159,15 @@ void gpgpu_sim::gpu_print_stat() for (unsigned i=0;im_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 25da107..a662446 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -718,6 +718,11 @@ 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{ + if (!m_config->m_L2_config.disabled()) { + m_L2cache->get_sub_stats_pw(css); + } +} void memory_sub_partition::visualizer_print( gzFile visualizer_file ) { // TODO: Add visualizer stats for L2 cache diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 18c0a8b..57f3e94 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -180,6 +180,9 @@ public: void accumulate_L2cache_stats(class cache_stats &l2_stats) const; 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 force_l2_tag_update(new_addr_type addr, unsigned time, mem_access_sector_mask_t mask) { m_L2cache->force_tag_access( addr, m_memcpy_cycle_offset + time, mask ); -- cgit v1.3 From c1833f27cdd0ad7d41afc64db0ec693571080c05 Mon Sep 17 00:00:00 2001 From: Nick Date: Wed, 20 Feb 2019 13:46:20 -0500 Subject: Add full support for deprecated AerialVision L2 stats --- src/gpgpu-sim/gpu-cache.cc | 67 +++++++++++++++++++------------------ src/gpgpu-sim/gpu-cache.h | 69 +++++++++++++++++++++++++++++++++++++-- src/gpgpu-sim/gpu-sim.cc | 8 +---- src/gpgpu-sim/l2cache.cc | 38 +++++++++++++++++---- src/gpgpu-sim/l2cache.h | 3 +- src/gpgpu-sim/mem_latency_stat.cc | 6 ++++ src/gpgpu-sim/mem_latency_stat.h | 8 +++++ 7 files changed, 147 insertions(+), 52 deletions(-) (limited to 'src/gpgpu-sim/l2cache.cc') 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;im_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; -- cgit v1.3 From 884ae41931010f129c0e9ba353f34177a0ae3599 Mon Sep 17 00:00:00 2001 From: Nick Date: Wed, 20 Feb 2019 14:33:29 -0500 Subject: Remove old comment, remove printstat used for functionality tests --- src/gpgpu-sim/gpu-sim.cc | 1 - src/gpgpu-sim/l2cache.cc | 3 ++- 2 files changed, 2 insertions(+), 2 deletions(-) (limited to 'src/gpgpu-sim/l2cache.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 56b9681..ec570bf 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -917,7 +917,6 @@ 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) { diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index 0237525..0edc3b7 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -741,7 +741,8 @@ void memory_sub_partition::clear_L2cache_stats_pw() { void memory_sub_partition::visualizer_print( gzFile visualizer_file ) { - // TODO: Add visualizer stats for L2 cache + // Support for L2 AerialVision stats + // Per-sub-partition stats would be trivial to extend from this cache_sub_stats_pw temp_sub_stats; get_L2cache_sub_stats_pw(temp_sub_stats); -- cgit v1.3 From 21d937256fbca004c926531cfef1adefcedeef91 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 29 Jul 2019 21:15:59 -0400 Subject: adding simple dram model --- src/gpgpu-sim/gpu-sim.cc | 7 ++++- src/gpgpu-sim/gpu-sim.h | 1 + src/gpgpu-sim/l2cache.cc | 77 ++++++++++++++++++++++++++++++++++++++++++------ src/gpgpu-sim/l2cache.h | 1 + 4 files changed, 76 insertions(+), 10 deletions(-) (limited to 'src/gpgpu-sim/l2cache.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 343ff86..92d5366 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -156,6 +156,8 @@ void memory_config::reg_options(class OptionParser * opp) { option_parser_register(opp, "-perf_sim_memcpy", OPT_BOOL, &m_perf_sim_memcpy, "Fill the L2 cache on memcpy", "1"); + option_parser_register(opp, "-simple_dram_model", OPT_BOOL, &simple_dram_model, + "simple_dram_model with fixed latency and BW", "0"); option_parser_register(opp, "-gpgpu_dram_scheduler", OPT_INT32, &scheduler_type, "0 = fifo, 1 = FR-FCFS (defaul)", "1"); option_parser_register(opp, "-gpgpu_dram_partition_queues", OPT_CSTR, &gpgpu_L2_queue_config, @@ -1606,7 +1608,10 @@ void gpgpu_sim::cycle() if (clock_mask & DRAM) { for (unsigned i=0;im_n_mem;i++){ - m_memory_partition_unit[i]->dram_cycle(); // Issue the dram command (scheduler + delay model) + if(m_memory_config->simple_dram_model) + m_memory_partition_unit[i]->simple_dram_model_cycle(); + else + m_memory_partition_unit[i]->dram_cycle(); // Issue the dram command (scheduler + delay model) // Update performance counters for DRAM m_memory_partition_unit[i]->set_dram_power_stats(m_power_stats->pwr_mem_stat->n_cmd[CURRENT_STAT_IDX][i], m_power_stats->pwr_mem_stat->n_activity[CURRENT_STAT_IDX][i], m_power_stats->pwr_mem_stat->n_nop[CURRENT_STAT_IDX][i], m_power_stats->pwr_mem_stat->n_act[CURRENT_STAT_IDX][i], m_power_stats->pwr_mem_stat->n_pre[CURRENT_STAT_IDX][i], diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index e98e499..c14d0a7 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -290,6 +290,7 @@ struct memory_config { unsigned write_high_watermark; unsigned write_low_watermark; bool m_perf_sim_memcpy; + bool simple_dram_model; }; // global counters and flags (please try not to add to this list!!!) diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index 0edc3b7..f24a596 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -203,7 +203,67 @@ int memory_partition_unit::global_sub_partition_id_to_local_id(int global_sub_pa return (global_sub_partition_id - m_id * m_config->m_n_sub_partition_per_memory_channel); } -void memory_partition_unit::dram_cycle() +void memory_partition_unit::simple_dram_model_cycle() +{ + + // pop completed memory request from dram and push it to dram-to-L2 queue + // of the original sub partition + if (!m_dram_latency_queue.empty() && ( (gpu_sim_cycle+gpu_tot_sim_cycle) >= m_dram_latency_queue.front().ready_cycle )) { + mem_fetch* mf_return = m_dram_latency_queue.front().req; + if( mf_return->get_access_type() != L1_WRBK_ACC && mf_return->get_access_type() != L2_WRBK_ACC ) { + mf_return->set_reply(); + + unsigned dest_global_spid = mf_return->get_sub_partition_id(); + int dest_spid = global_sub_partition_id_to_local_id(dest_global_spid); + assert(m_sub_partition[dest_spid]->get_id() == dest_global_spid); + if (!m_sub_partition[dest_spid]->dram_L2_queue_full()) { + if( mf_return->get_access_type() == L1_WRBK_ACC ) { + m_sub_partition[dest_spid]->set_done(mf_return); + delete mf_return; + } else { + m_sub_partition[dest_spid]->dram_L2_queue_push(mf_return); + mf_return->set_status(IN_PARTITION_DRAM_TO_L2_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); + m_arbitration_metadata.return_credit(dest_spid); + MEMPART_DPRINTF("mem_fetch request %p return from dram to sub partition %d\n", mf_return, dest_spid); + } + m_dram_latency_queue.pop_front(); + } + + } else { + this->set_done(mf_return); + delete mf_return; + m_dram_latency_queue.pop_front(); + } + } + + // mem_fetch *mf = m_sub_partition[spid]->L2_dram_queue_top(); + //if( !m_dram->full(mf->is_write()) ) { + // L2->DRAM queue to DRAM latency queue + // Arbitrate among multiple L2 subpartitions + int last_issued_partition = m_arbitration_metadata.last_borrower(); + for (unsigned p = 0; p < m_config->m_n_sub_partition_per_memory_channel; p++) { + int spid = (p + last_issued_partition + 1) % m_config->m_n_sub_partition_per_memory_channel; + if (!m_sub_partition[spid]->L2_dram_queue_empty() && can_issue_to_dram(spid)) { + mem_fetch *mf = m_sub_partition[spid]->L2_dram_queue_top(); + if(m_dram->full(mf->is_write()) ) + break; + + m_sub_partition[spid]->L2_dram_queue_pop(); + MEMPART_DPRINTF("Issue mem_fetch request %p from sub partition %d to dram\n", mf, spid); + dram_delay_t d; + d.req = mf; + d.ready_cycle = gpu_sim_cycle+gpu_tot_sim_cycle + m_config->dram_latency; + m_dram_latency_queue.push_back(d); + mf->set_status(IN_PARTITION_DRAM_LATENCY_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); + m_arbitration_metadata.borrow_credit(spid); + break; // the DRAM should only accept one request per cycle + } + } + //} + +} + +void memory_partition_unit::dram_cycle() { // pop completed memory request from dram and push it to dram-to-L2 queue // of the original sub partition @@ -228,8 +288,8 @@ void memory_partition_unit::dram_cycle() m_dram->return_queue_pop(); } - m_dram->cycle(); - m_dram->dram_log(SAMPLELOG); + m_dram->cycle(); + m_dram->dram_log(SAMPLELOG); // mem_fetch *mf = m_sub_partition[spid]->L2_dram_queue_top(); //if( !m_dram->full(mf->is_write()) ) { @@ -257,12 +317,11 @@ void memory_partition_unit::dram_cycle() //} // DRAM latency queue - - if( !m_dram_latency_queue.empty() && ( (gpu_sim_cycle+gpu_tot_sim_cycle) >= m_dram_latency_queue.front().ready_cycle ) && !m_dram->full(m_dram_latency_queue.front().req->is_write()) ) { - mem_fetch* mf = m_dram_latency_queue.front().req; - m_dram_latency_queue.pop_front(); - m_dram->push(mf); - } + if( !m_dram_latency_queue.empty() && ( (gpu_sim_cycle+gpu_tot_sim_cycle) >= m_dram_latency_queue.front().ready_cycle ) && !m_dram->full(m_dram_latency_queue.front().req->is_write()) ) { + mem_fetch* mf = m_dram_latency_queue.front().req; + m_dram_latency_queue.pop_front(); + m_dram->push(mf); + } } void memory_partition_unit::set_done( mem_fetch *mf ) diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index beafdd3..9a51c0e 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -65,6 +65,7 @@ public: void cache_cycle( unsigned cycle ); void dram_cycle(); + void simple_dram_model_cycle(); void set_done( mem_fetch *mf ); -- cgit v1.3