From dcf7e6a8445a5d8e4383add44ca6f5d0d198ab02 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 17 Jul 2017 22:32:30 -0400 Subject: Improving GPU core model. This commits contains: 1- round robin inst issue for warp multiple schedulers 2- add sector mask in the memory request (to bused later for L2 sector cache) 3- Adding Fermi coalescer 4- Ensure different exen units are used in dual_issue mode 5- Report how many dual_issue happened 6- Adding oldest_first scheduler --- src/gpgpu-sim/gpu-sim.cc | 20 ++++++++++++++++---- 1 file changed, 16 insertions(+), 4 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 58a5d16..26432c4 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -285,6 +285,15 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shmem_warp_parts", OPT_INT32, &mem_warp_parts, "Number of portions a warp is divided into for shared memory bank conflict check ", "2"); + option_parser_register(opp, "-gpgpu_L1_warp_parts_cached", OPT_INT32, &L1_warp_parts_cached, + "Number of portions a warp is divided into when the request is cached", + "2"); + option_parser_register(opp, "-gpgpu_L1_warp_parts_cached", OPT_INT32, &L1_warp_parts_non_cached, + "Number of portions a warp is divided into when the request is not cached", + "4"); + option_parser_register(opp, "-gpgpu_shmem_warp_parts", OPT_INT32, &mem_warp_parts, + "Number of portions a warp is divided into for shared memory bank conflict check ", + "2"); option_parser_register(opp, "-gpgpu_warpdistro_shader", OPT_INT32, &gpgpu_warpdistro_shader, "Specify which shader core to collect the warp size distribution from", "-1"); @@ -293,7 +302,7 @@ void shader_core_config::reg_options(class OptionParser * opp) "0"); option_parser_register(opp, "-gpgpu_local_mem_map", OPT_BOOL, &gpgpu_local_mem_map, "Mapping from local memory space address to simulated GPU physical address space (default = enabled)", - "1"); + "1"); option_parser_register(opp, "-gpgpu_num_reg_banks", OPT_INT32, &gpgpu_num_reg_banks, "Number of register banks (default = 8)", "8"); @@ -337,14 +346,17 @@ void shader_core_config::reg_options(class OptionParser * opp) "number of collector unit in ports (default = 0)", "0"); option_parser_register(opp, "-gpgpu_coalesce_arch", OPT_INT32, &gpgpu_coalesce_arch, - "Coalescing arch (default = 13, anything else is off for now)", + "Coalescing arch (GT200 = 13, Fermi = 20)", "13"); option_parser_register(opp, "-gpgpu_num_sched_per_core", OPT_INT32, &gpgpu_num_sched_per_core, "Number of warp schedulers per core", "1"); option_parser_register(opp, "-gpgpu_max_insn_issue_per_warp", OPT_INT32, &gpgpu_max_insn_issue_per_warp, - "Max number of instructions that can be issued per warp in one cycle by scheduler", - "2"); + "Max number of instructions that can be issued per warp in one cycle by scheduler (either 1 or 2)", + "2"); + option_parser_register(opp, "-gpgpu_dual_issue_diff_exec_units", OPT_BOOL, &gpgpu_dual_issue_diff_exec_units, + "should dual issue use two different execution unit resources", + "1"); option_parser_register(opp, "-gpgpu_simt_core_sim_order", OPT_INT32, &simt_core_sim_order, "Select the simulation order of cores in a cluster (0=Fix, 1=Round-Robin)", "1"); -- cgit v1.3 From b806d3a4021a12638e1e67fcd67012dca4b1689e Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 17 Jul 2017 23:20:42 -0400 Subject: Fixing some typos and adding comments --- src/abstract_hardware_model.cc | 14 +++++++------- src/abstract_hardware_model.h | 2 +- src/gpgpu-sim/gpu-cache.cc | 2 +- src/gpgpu-sim/gpu-sim.cc | 12 ++++++------ src/gpgpu-sim/gpu-sim.h | 2 -- src/gpgpu-sim/shader.cc | 43 +++++++++++++++++++++--------------------- src/gpgpu-sim/shader.h | 4 ++-- 7 files changed, 39 insertions(+), 40 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index f75c525..64eb43c 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -315,11 +315,11 @@ void warp_inst_t::generate_mem_accesses() case global_space: case local_space: case param_space_local: if( m_config->gpgpu_coalesce_arch == 13 || m_config->gpgpu_coalesce_arch == 20) { - if(isatomic()) - memory_coalescing_arch_atomic(is_write, access_type); - else - memory_coalescing_arch(is_write, access_type); - } else abort(); + if(isatomic()) + memory_coalescing_arch_atomic(is_write, access_type); + else + memory_coalescing_arch(is_write, access_type); + } else abort(); break; @@ -358,7 +358,7 @@ void warp_inst_t::memory_coalescing_arch( bool is_write, mem_access_type access_ unsigned segment_size = 0; unsigned warp_parts; - //TO DO: need to double check how double number are coalesced! + //TO DO: need to double check how doubles are coalesced! if(data_size == 1) { //If it is byte data, then coalesce on the whole 32 threads, regardless the arch version @@ -367,7 +367,7 @@ void warp_inst_t::memory_coalescing_arch( bool is_write, mem_access_type access_ else if(m_config->gpgpu_coalesce_arch == 13) { //mem_warp_parts should equal 2 for arch=13 - //we still use the parameter mem_warp_parts for arch=13 to ensure it is backward compatibility with older gpgpu config files + //use the parameter mem_warp_parts for arch=13 to ensure it is backward compatibility with older gpgpu config files warp_parts = m_config->mem_warp_parts; } else if(m_config->gpgpu_coalesce_arch == 20) diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index cd185a1..910a9ed 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -685,7 +685,7 @@ public: bool wr, const active_mask_t &active_mask, const mem_access_byte_mask_t &byte_mask, - const mem_access_sector_mask_t §or_mask) + const mem_access_sector_mask_t §or_mask) : m_warp_mask(active_mask), m_byte_mask(byte_mask), m_sector_mask(sector_mask) { init(); diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index cfd0dc8..7af7db0 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -868,7 +868,7 @@ data_cache::wr_miss_wa( new_addr_type addr, false, // Now performing a read mf->get_access_warp_mask(), mf->get_access_byte_mask(), - mf->get_access_sector_mask()); + mf->get_access_sector_mask()); mem_fetch *n_mf = new mem_fetch( *ma, NULL, diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 26432c4..483a7b6 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -302,7 +302,7 @@ void shader_core_config::reg_options(class OptionParser * opp) "0"); option_parser_register(opp, "-gpgpu_local_mem_map", OPT_BOOL, &gpgpu_local_mem_map, "Mapping from local memory space address to simulated GPU physical address space (default = enabled)", - "1"); + "1"); option_parser_register(opp, "-gpgpu_num_reg_banks", OPT_INT32, &gpgpu_num_reg_banks, "Number of register banks (default = 8)", "8"); @@ -346,17 +346,17 @@ void shader_core_config::reg_options(class OptionParser * opp) "number of collector unit in ports (default = 0)", "0"); option_parser_register(opp, "-gpgpu_coalesce_arch", OPT_INT32, &gpgpu_coalesce_arch, - "Coalescing arch (GT200 = 13, Fermi = 20)", + "Coalescing arch (GT200 = 13, Fermi = 20)", "13"); option_parser_register(opp, "-gpgpu_num_sched_per_core", OPT_INT32, &gpgpu_num_sched_per_core, "Number of warp schedulers per core", "1"); option_parser_register(opp, "-gpgpu_max_insn_issue_per_warp", OPT_INT32, &gpgpu_max_insn_issue_per_warp, - "Max number of instructions that can be issued per warp in one cycle by scheduler (either 1 or 2)", - "2"); + "Max number of instructions that can be issued per warp in one cycle by scheduler (either 1 or 2)", + "2"); option_parser_register(opp, "-gpgpu_dual_issue_diff_exec_units", OPT_BOOL, &gpgpu_dual_issue_diff_exec_units, - "should dual issue use two different execution unit resources", - "1"); + "should dual issue use two different execution unit resources", + "1"); option_parser_register(opp, "-gpgpu_simt_core_sim_order", OPT_INT32, &simt_core_sim_order, "Select the simulation order of cores in a cluster (0=Fix, 1=Round-Robin)", "1"); diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 23e6144..7d92c66 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -314,8 +314,6 @@ public: unsigned num_shader() const { return m_shader_config.num_shader(); } unsigned num_cluster() const { return m_shader_config.n_simt_clusters; } unsigned get_max_concurrent_kernel() const { return max_concurrent_kernel; } - unsigned get_gpu_max_cycle_opt() const { return gpu_max_cycle_opt; } - unsigned get_gpu_max_insn_opt() const { return gpu_max_insn_opt; } private: void init_clock_domains(void ); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 015995e..d826b69 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -132,8 +132,8 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, CONCRETE_SCHEDULER_TWO_LEVEL_ACTIVE : sched_config.find("gto") != std::string::npos ? CONCRETE_SCHEDULER_GTO : - sched_config.find("old") != std::string::npos ? - CONCRETE_SCHEDULER_OLDEST_FIRST : + sched_config.find("old") != std::string::npos ? + CONCRETE_SCHEDULER_OLDEST_FIRST : sched_config.find("warp_limiting") != std::string::npos ? CONCRETE_SCHEDULER_WARP_LIMITING: NUM_CONCRETE_SCHEDULERS; @@ -187,16 +187,16 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, break; case CONCRETE_SCHEDULER_OLDEST_FIRST: schedulers.push_back( - new oldest_scheduler( m_stats, - this, - m_scoreboard, - m_simt_stack, - &m_warp, - &m_pipeline_reg[ID_OC_SP], - &m_pipeline_reg[ID_OC_SFU], - &m_pipeline_reg[ID_OC_MEM], - i - ) + new oldest_scheduler( m_stats, + this, + m_scoreboard, + m_simt_stack, + &m_warp, + &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_MEM], + i + ) ); break; case CONCRETE_SCHEDULER_WARP_LIMITING: @@ -750,12 +750,13 @@ void shader_core_ctx::issue_warp( register_set& pipe_reg_set, const warp_inst_t* void shader_core_ctx::issue(){ - unsigned j; - for (unsigned i = 0; i < schedulers.size(); i++) { - j = (Issue_Prio + i) % schedulers.size(); - schedulers[j]->cycle(); - } - Issue_Prio = (Issue_Prio+1)% schedulers.size(); + //Ensure fair round robin issu between schedulers + unsigned j; + for (unsigned i = 0; i < schedulers.size(); i++) { + j = (Issue_Prio + i) % schedulers.size(); + schedulers[j]->cycle(); + } + Issue_Prio = (Issue_Prio+1)% schedulers.size(); //really is issue; //for (unsigned i = 0; i < schedulers.size(); i++) { @@ -876,8 +877,8 @@ void scheduler_unit::cycle() unsigned checked=0; unsigned issued=0; exec_unit_type_t previous_issued_inst_exec_type = exec_unit_type_t::NONE; - unsigned max_issue = m_shader->m_config->gpgpu_max_insn_issue_per_warp; - bool diff_exec_units = m_shader->m_config->gpgpu_dual_issue_diff_exec_units; + unsigned max_issue = m_shader->m_config->gpgpu_max_insn_issue_per_warp; + bool diff_exec_units = m_shader->m_config->gpgpu_dual_issue_diff_exec_units; while( !warp(warp_id).waiting() && !warp(warp_id).ibuffer_empty() && (checked < max_issue) && (checked <= issued) && (issued < max_issue) ) { const warp_inst_t *pI = warp(warp_id).ibuffer_next_inst(); //Jin: handle cdp latency; @@ -1062,7 +1063,7 @@ void oldest_scheduler::order_warps() m_supervised_warps, m_last_supervised_issued, m_supervised_warps.size(), - ORDERED_PRIORITY_FUNC_ONLY, + ORDERED_PRIORITY_FUNC_ONLY, scheduler_unit::sort_warps_by_oldest_dynamic_id ); } diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 395c1ed..2fed420 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -316,7 +316,7 @@ enum concrete_scheduler CONCRETE_SCHEDULER_GTO, CONCRETE_SCHEDULER_TWO_LEVEL_ACTIVE, CONCRETE_SCHEDULER_WARP_LIMITING, - CONCRETE_SCHEDULER_OLDEST_FIRST, + CONCRETE_SCHEDULER_OLDEST_FIRST, NUM_CONCRETE_SCHEDULERS }; @@ -456,7 +456,7 @@ public: : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, mem_out, id ){} virtual ~oldest_scheduler () {} virtual void order_warps (); - virtual void done_adding_supervised_warps() { + virtual void done_adding_supervised_warps() { m_last_supervised_issued = m_supervised_warps.begin(); } -- cgit v1.3 From f678104dcc3e5c970b871244e18e38f97c0caaa5 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 12 Sep 2017 18:53:26 -0400 Subject: Adding HBM model --- src/gpgpu-sim/dram.cc | 597 +++++++++++++++++++++++++++++++++----------- src/gpgpu-sim/dram.h | 59 ++++- src/gpgpu-sim/dram_sched.cc | 11 +- src/gpgpu-sim/gpu-sim.cc | 9 + src/gpgpu-sim/gpu-sim.h | 4 + 5 files changed, 534 insertions(+), 146 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index a0e024b..352bd58 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -49,11 +49,41 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m m_stats = stats; m_config = config; + //rowblp + access_num=0; + hits_num=0; + banks_1time=0; + banks_acess_total=0; + banks_acess_total_after=0; + banks_time_ready=0; + banks_access_ready_total=0; + issued_two=0; + issued_total=0; + issued_total_row=0; + issued_total_col=0; + CCDc = 0; RRDc = 0; RTWc = 0; WTRc = 0; + wasted_bw_row=0; + wasted_bw_col=0; + util_bw=0; + idle_bw=0; + RCDc_limit=0; + CCDLc_limit=0; + CCDLc_limit_alone=0; + CCDc_limit=0; + WTRc_limit=0; + WTRc_limit_alone=0; + RCDWRc_limit=0; + RTWc_limit=0; + RTWc_limit_alone=0; + rwq_limit=0; + write_to_read_ratio_blp_rw_average=0; + bkgrp_parallsim_rw=0; + rw = READ; //read mode is default bkgrp = (bankgrp_t**) calloc(sizeof(bankgrp_t*), m_config->nbkgrp); @@ -74,12 +104,13 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m bk[i]->state = BANK_IDLE; bk[i]->bkgrpindex = i/(m_config->nbk/m_config->nbkgrp); } - prio = 0; + prio = 0; + rwq = new fifo_pipeline("rwq",m_config->CL,m_config->CL+1); mrqq = new fifo_pipeline("mrqq",0,2); returnq = new fifo_pipeline("dramreturnq",0,m_config->gpgpu_dram_return_queue_size==0?1024:m_config->gpgpu_dram_return_queue_size); m_frfcfs_scheduler = NULL; - if ( m_config->scheduler_type == DRAM_FRFCFS ) + if ( m_config->scheduler_type == DRAM_FRFCFS) m_frfcfs_scheduler = new frfcfs_scheduler(m_config,this,stats); n_cmd = 0; n_activity = 0; @@ -88,6 +119,8 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m n_pre = 0; n_rd = 0; n_wr = 0; + n_wr_WB=0; + n_rd_L2_A=0; n_req = 0; max_mrqs_temp = 0; bwutil = 0; @@ -113,11 +146,12 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m mrqq_Dist = StatCreate("mrqq_length",1, queue_limit()); else //queue length is unlimited; mrqq_Dist = StatCreate("mrqq_length",1,64); //track up to 64 entries + } bool dram_t::full() const { - if(m_config->scheduler_type == DRAM_FRFCFS ){ + if(m_config->scheduler_type == DRAM_FRFCFS){ if(m_config->gpgpu_frfcfs_dram_sched_queue_size == 0 ) return false; return m_frfcfs_scheduler->num_pending() >= m_config->gpgpu_frfcfs_dram_sched_queue_size; } @@ -127,7 +161,7 @@ bool dram_t::full() const unsigned dram_t::que_length() const { unsigned nreqs = 0; - if (m_config->scheduler_type == DRAM_FRFCFS ) { + if (m_config->scheduler_type == DRAM_FRFCFS) { nreqs = m_frfcfs_scheduler->num_pending(); } else { nreqs = mrqq->get_length(); @@ -146,7 +180,7 @@ unsigned int dram_t::queue_limit() const } -dram_req_t::dram_req_t( class mem_fetch *mf ) +dram_req_t::dram_req_t( class mem_fetch *mf, unsigned banks, unsigned dram_bnk_indexing_policy) { txbytes = 0; dqbytes = 0; @@ -154,7 +188,15 @@ dram_req_t::dram_req_t( class mem_fetch *mf ) const addrdec_t &tlx = mf->get_tlx_addr(); - bk = tlx.bk; + if(dram_bnk_indexing_policy == 0) { + int lbank = log2(banks); + bk = tlx.bk ^ (((1<get_data_size(); @@ -169,14 +211,15 @@ void dram_t::push( class mem_fetch *data ) { assert(id == data->get_tlx_addr().chip); // Ensure request is in correct memory partition - dram_req_t *mrq = new dram_req_t(data); + dram_req_t *mrq = new dram_req_t(data,m_config->nbk,m_config->dram_bnk_indexing_policy); + data->set_status(IN_PARTITION_MC_INTERFACE_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); - mrqq->push(mrq); + mrqq->push(mrq); // stats... n_req += 1; n_req_partial += 1; - if ( m_config->scheduler_type == DRAM_FRFCFS ) { + if ( m_config->scheduler_type == DRAM_FRFCFS) { unsigned nreqs = m_frfcfs_scheduler->num_pending(); if ( nreqs > max_mrqs_temp) max_mrqs_temp = nreqs; @@ -212,6 +255,7 @@ void dram_t::cycle() printf("\tDQ: BK%d Row:%03x Col:%03x", cmd->bk, cmd->row, cmd->col + cmd->dqbytes); #endif cmd->dqbytes += m_config->dram_atom_size; + if (cmd->dqbytes >= cmd->nbytes) { mem_fetch *data = cmd->data; data->set_status(IN_PARTITION_MC_RETURNQ,gpu_sim_cycle+gpu_tot_sim_cycle); @@ -240,7 +284,7 @@ void dram_t::cycle() printf("Error: Unknown DRAM scheduler type\n"); assert(0); } - if ( m_config->scheduler_type == DRAM_FRFCFS ) { + if ( m_config->scheduler_type == DRAM_FRFCFS) { unsigned nreqs = m_frfcfs_scheduler->num_pending(); if ( nreqs > max_mrqs) { max_mrqs = nreqs; @@ -258,130 +302,123 @@ void dram_t::cycle() unsigned k=m_config->nbk; bool issued = false; - // check if any bank is ready to issue a new read + //collect row buffer locality, BLP and other statistics + ///////////////////////////////////////////////////////////////////////// + unsigned int memory_Pending=0; for (unsigned i=0;inbk;i++) { - unsigned j = (i + prio) % m_config->nbk; - unsigned grp = j>>m_config->bk_tag_length; - if (bk[j]->mrq) { //if currently servicing a memory request - bk[j]->mrq->data->set_status(IN_PARTITION_DRAM,gpu_sim_cycle+gpu_tot_sim_cycle); - // correct row activated for a READ - if ( !issued && !CCDc && !bk[j]->RCDc && - !(bkgrp[grp]->CCDLc) && - (bk[j]->curr_row == bk[j]->mrq->row) && - (bk[j]->mrq->rw == READ) && (WTRc == 0 ) && - (bk[j]->state == BANK_ACTIVE) && - !rwq->full() ) { - if (rw==WRITE) { - rw=READ; - rwq->set_min_length(m_config->CL); - } - rwq->push(bk[j]->mrq); - bk[j]->mrq->txbytes += m_config->dram_atom_size; - CCDc = m_config->tCCD; - bkgrp[grp]->CCDLc = m_config->tCCDL; - RTWc = m_config->tRTW; - bk[j]->RTPc = m_config->BL/m_config->data_command_freq_ratio; - bkgrp[grp]->RTPLc = m_config->tRTPL; - issued = true; - n_rd++; - bwutil += m_config->BL/m_config->data_command_freq_ratio; - bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; - bk[j]->n_access++; -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tRD Bk:%d Row:%03x Col:%03x \n", - j, bk[j]->curr_row, - bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); -#endif - // transfer done - if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { - bk[j]->mrq = NULL; - } - } else - // correct row activated for a WRITE - if ( !issued && !CCDc && !bk[j]->RCDWRc && - !(bkgrp[grp]->CCDLc) && - (bk[j]->curr_row == bk[j]->mrq->row) && - (bk[j]->mrq->rw == WRITE) && (RTWc == 0 ) && - (bk[j]->state == BANK_ACTIVE) && - !rwq->full() ) { - if (rw==READ) { - rw=WRITE; - rwq->set_min_length(m_config->WL); - } - rwq->push(bk[j]->mrq); - - bk[j]->mrq->txbytes += m_config->dram_atom_size; - CCDc = m_config->tCCD; - bkgrp[grp]->CCDLc = m_config->tCCDL; - WTRc = m_config->tWTR; - bk[j]->WTPc = m_config->tWTP; - issued = true; - n_wr++; - bwutil += m_config->BL/m_config->data_command_freq_ratio; - bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tWR Bk:%d Row:%03x Col:%03x \n", - j, bk[j]->curr_row, - bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); -#endif - // transfer done - if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { - bk[j]->mrq = NULL; - } - } - - else - // bank is idle - if ( !issued && !RRDc && - (bk[j]->state == BANK_IDLE) && - !bk[j]->RPc && !bk[j]->RCc ) { -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tACT BK:%d NewRow:%03x From:%03x \n", - j,bk[j]->mrq->row,bk[j]->curr_row); -#endif - // activate the row with current memory request - bk[j]->curr_row = bk[j]->mrq->row; - bk[j]->state = BANK_ACTIVE; - RRDc = m_config->tRRD; - bk[j]->RCDc = m_config->tRCD; - bk[j]->RCDWRc = m_config->tRCDWR; - bk[j]->RASc = m_config->tRAS; - bk[j]->RCc = m_config->tRC; - prio = (j + 1) % m_config->nbk; - issued = true; - n_act_partial++; - n_act++; - } - - else - // different row activated - if ( (!issued) && - (bk[j]->curr_row != bk[j]->mrq->row) && - (bk[j]->state == BANK_ACTIVE) && - (!bk[j]->RASc && !bk[j]->WTPc && - !bk[j]->RTPc && - !bkgrp[grp]->RTPLc) ) { - // make the bank idle again - bk[j]->state = BANK_IDLE; - bk[j]->RPc = m_config->tRP; - prio = (j + 1) % m_config->nbk; - issued = true; - n_pre++; - n_pre_partial++; -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tPRE BK:%d Row:%03x \n", j,bk[j]->curr_row); -#endif - } - } else { - if (!CCDc && !RRDc && !RTWc && !WTRc && !bk[j]->RCDc && !bk[j]->RASc - && !bk[j]->RCc && !bk[j]->RPc && !bk[j]->RCDWRc) k--; - bk[j]->n_idle++; - } + if (bk[i]->mrq) + memory_Pending++; + } + banks_1time += memory_Pending; + if(memory_Pending >0) + banks_acess_total++; + + unsigned int memory_pending_rw=0; + unsigned read_BLP_RW=0; + unsigned write_BLP_RW=0; + std::bitset<8> bnkgrp_RW_found; + //bool memory_pending_rw_found=false; + for (unsigned j=0;jnbk;j++) { + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && + (bk[j]->state == BANK_ACTIVE)))) + { + memory_pending_rw++; + read_BLP_RW++; + bnkgrp_RW_found.set(grp); + } + else if + (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && + (bk[j]->state == BANK_ACTIVE)))) + { + memory_pending_rw++; + write_BLP_RW++; + bnkgrp_RW_found.set(grp); + } + } + banks_time_rw += memory_pending_rw; + bkgrp_parallsim_rw += bnkgrp_RW_found.count(); + if(memory_pending_rw >0) + { + write_to_read_ratio_blp_rw_average += (double)write_BLP_RW/(write_BLP_RW+read_BLP_RW); + banks_access_rw_total++; + } + + unsigned int memory_Pending_ready=0; + for (unsigned j=0;jnbk;j++) { + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq && ((!CCDc && !bk[j]->RCDc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && (WTRc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full()) + || + (!CCDc && !bk[j]->RCDWRc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && (RTWc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full()))) + { + memory_Pending_ready++; + } + } + banks_time_ready += memory_Pending_ready; + if(memory_Pending_ready >0) + banks_access_ready_total++; + /////////////////////////////////////////////////////////////////////////////////// + + bool issued_col_cmd = false; + bool issued_row_cmd = false; + + if(m_config->dual_bus_interface) + { + //dual bus interface + //issue one row command and one column command + for (unsigned i=0;inbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + issued_col_cmd = issue_col_command(j); + if(issued_col_cmd) break; + } + for (unsigned i=0;inbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + issued_row_cmd = issue_row_command(j); + if(issued_row_cmd) break; + } + for (unsigned i=0;inbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + if(!bk[j]->mrq) { + if (!CCDc && !RRDc && !RTWc && !WTRc && !bk[j]->RCDc && !bk[j]->RASc + && !bk[j]->RCc && !bk[j]->RPc && !bk[j]->RCDWRc) k--; + bk[j]->n_idle++; + } + } } + else + { + //single bus interface + //issue only one row/column command + for (unsigned i=0;inbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + if(!issued_col_cmd) + issued_col_cmd = issue_col_command(j); + + if(!issued_col_cmd && !issued_row_cmd) + issued_row_cmd = issue_row_command(j); + + if(!bk[j]->mrq) { + if (!CCDc && !RRDc && !RTWc && !WTRc && !bk[j]->RCDc && !bk[j]->RASc + && !bk[j]->RCc && !bk[j]->RPc && !bk[j]->RCDWRc) k--; + bk[j]->n_idle++; + } + + } + } + + issued = issued_row_cmd || issued_col_cmd; if (!issued) { n_nop++; n_nop_partial++; @@ -395,6 +432,80 @@ void dram_t::cycle() } n_cmd++; n_cmd_partial++; + if(issued) + { + issued_total++; + if(issued_col_cmd && issued_row_cmd) + issued_two++; + } + if(issued_col_cmd) issued_total_col++; + if(issued_row_cmd) issued_total_row++; + + unsigned int memory_pending_found=0; + for (unsigned i=0;inbk;i++) { + if (bk[i]->mrq) + memory_pending_found++; + } + if(memory_pending_found>0) + banks_acess_total_after++; + + bool memory_pending_rw_found=false; + for (unsigned j=0;jnbk;j++) { + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && + (bk[j]->state == BANK_ACTIVE)) + || + ( + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && + (bk[j]->state == BANK_ACTIVE)))) + memory_pending_rw_found=true; + } + + //Collect some statistics + //check the limitation, why BW is wasted? + if(issued_col_cmd || CCDc) + util_bw++; + else if (memory_pending_rw_found) + { + wasted_bw_col++; + for (unsigned j=0;jnbk;j++) { + unsigned grp = get_bankgrp_number(j); + //read + if (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && + (bk[j]->state == BANK_ACTIVE)))) + { + if(bk[j]->RCDc) RCDc_limit++; + if(bkgrp[grp]->CCDLc) CCDLc_limit++; + if(WTRc) WTRc_limit++; + if(CCDc) CCDc_limit++; + if(rwq->full()) rwq_limit++; + if(bkgrp[grp]->CCDLc && !WTRc) CCDLc_limit_alone++; + if(!bkgrp[grp]->CCDLc && WTRc) WTRc_limit_alone++; + } + //write + else if (bk[j]->mrq && ((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && + (bk[j]->state == BANK_ACTIVE))) + { + if(bk[j]->RCDWRc) RCDWRc_limit++; + if(bkgrp[grp]->CCDLc) CCDLc_limit++; + if(RTWc) RTWc_limit++; + if(CCDc) CCDc_limit++; + if(rwq->full()) rwq_limit++; + if(bkgrp[grp]->CCDLc && !RTWc) CCDLc_limit_alone++; + if(!bkgrp[grp]->CCDLc && RTWc) RTWc_limit_alone++; + } + } + } + else if (memory_pending_found) + wasted_bw_row++; + else if (!memory_pending_found) + idle_bw++; + else + assert(1); // decrements counters once for each time dram_issueCMD is called DEC2ZERO(RRDc); @@ -420,39 +531,231 @@ void dram_t::cycle() #endif } +bool dram_t::issue_col_command(int j) +{ + bool issued = false; + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq) { //if currently servicing a memory request + bk[j]->mrq->data->set_status(IN_PARTITION_DRAM,gpu_sim_cycle+gpu_tot_sim_cycle); + // correct row activated for a READ + if ( !issued && !CCDc && !bk[j]->RCDc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && (WTRc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full() ) { + if (rw==WRITE) { + rw=READ; + rwq->set_min_length(m_config->CL); + } + rwq->push(bk[j]->mrq); + bk[j]->mrq->txbytes += m_config->dram_atom_size; + CCDc = m_config->tCCD; + bkgrp[grp]->CCDLc = m_config->tCCDL; + RTWc = m_config->tRTW; + bk[j]->RTPc = m_config->BL/m_config->data_command_freq_ratio; + bkgrp[grp]->RTPLc = m_config->tRTPL; + issued = true; + if(bk[j]->mrq->data->get_access_type() == L2_WR_ALLOC_R) + n_rd_L2_A++; + else + n_rd++; + bwutil += m_config->BL/m_config->data_command_freq_ratio; + bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; + bk[j]->n_access++; + +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tRD Bk:%d Row:%03x Col:%03x \n", + j, bk[j]->curr_row, + bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); +#endif + // transfer done + if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { + bk[j]->mrq = NULL; + } + } else + // correct row activated for a WRITE + if ( !issued && !CCDc && !bk[j]->RCDWRc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && (RTWc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full() ) { + if (rw==READ) { + rw=WRITE; + rwq->set_min_length(m_config->WL); + } + rwq->push(bk[j]->mrq); + + bk[j]->mrq->txbytes += m_config->dram_atom_size; + CCDc = m_config->tCCD; + bkgrp[grp]->CCDLc = m_config->tCCDL; + WTRc = m_config->tWTR; + bk[j]->WTPc = m_config->tWTP; + issued = true; + + if(bk[j]->mrq->data->get_access_type() == L2_WRBK_ACC) + n_wr_WB++; + else + n_wr++; + bwutil += m_config->BL/m_config->data_command_freq_ratio; + bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tWR Bk:%d Row:%03x Col:%03x \n", + j, bk[j]->curr_row, + bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); +#endif + // transfer done + if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { + bk[j]->mrq = NULL; + } + } + + } + + return issued; +} + +bool dram_t::issue_row_command(int j) +{ + bool issued = false; + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq) { //if currently servicing a memory request + bk[j]->mrq->data->set_status(IN_PARTITION_DRAM,gpu_sim_cycle+gpu_tot_sim_cycle); + // bank is idle + //else + if ( !issued && !RRDc && + (bk[j]->state == BANK_IDLE) && + !bk[j]->RPc && !bk[j]->RCc) { // +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tACT BK:%d NewRow:%03x From:%03x \n", + j,bk[j]->mrq->row,bk[j]->curr_row); +#endif + // activate the row with current memory request + bk[j]->curr_row = bk[j]->mrq->row; + bk[j]->state = BANK_ACTIVE; + RRDc = m_config->tRRD; + bk[j]->RCDc = m_config->tRCD; + bk[j]->RCDWRc = m_config->tRCDWR; + bk[j]->RASc = m_config->tRAS; + bk[j]->RCc = m_config->tRC; + prio = (j + 1) % m_config->nbk; + issued = true; + n_act_partial++; + n_act++; + } + + else + // different row activated + if ( (!issued) && + (bk[j]->curr_row != bk[j]->mrq->row) && + (bk[j]->state == BANK_ACTIVE) && + (!bk[j]->RASc && !bk[j]->WTPc && + !bk[j]->RTPc && + !bkgrp[grp]->RTPLc) ) { + // make the bank idle again + bk[j]->state = BANK_IDLE; + bk[j]->RPc = m_config->tRP; + prio = (j + 1) % m_config->nbk; + issued = true; + n_pre++; + n_pre_partial++; +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tPRE BK:%d Row:%03x \n", j,bk[j]->curr_row); +#endif + } + } + return issued; +} + + //if mrq is being serviced by dram, gets popped after CL latency fulfilled -class mem_fetch* dram_t::return_queue_pop() +class mem_fetch* dram_t::return_queue_pop() { return returnq->pop(); } -class mem_fetch* dram_t::return_queue_top() +class mem_fetch* dram_t::return_queue_top() { return returnq->top(); } + void dram_t::print( FILE* simFile) const { unsigned i; fprintf(simFile,"DRAM[%d]: %d bks, busW=%d BL=%d CL=%d, ", id, m_config->nbk, m_config->busW, m_config->BL, m_config->CL ); fprintf(simFile,"tRRD=%d tCCD=%d, tRCD=%d tRAS=%d tRP=%d tRC=%d\n", - m_config->tCCD, m_config->tRRD, m_config->tRCD, m_config->tRAS, m_config->tRP, m_config->tRC ); - fprintf(simFile,"n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_req=%d n_rd=%d n_write=%d bw_util=%.4g\n", - n_cmd, n_nop, n_act, n_pre, n_req, n_rd, n_wr, + m_config->tRRD, m_config->tCCD, m_config->tRCD, m_config->tRAS, m_config->tRP, m_config->tRC ); + fprintf(simFile,"n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_ref_event=%d n_req=%d n_rd=%d n_rd_L2_A=%d n_write=%d n_wr_bk=%d bw_util=%.4g\n", + n_cmd, n_nop, n_act, n_pre, n_ref, n_req, n_rd, n_rd_L2_A, n_wr, n_wr_WB, (float)bwutil/n_cmd); fprintf(simFile,"n_activity=%d dram_eff=%.4g\n", n_activity, (float)bwutil/n_activity); for (i=0;inbk;i++) { fprintf(simFile, "bk%d: %da %di ",i,bk[i]->n_access,bk[i]->n_idle); } + fprintf(simFile, "\n"); + fprintf(simFile, "\n------------------------------------------------------------------------\n"); + + printf("\nRow_Buffer_Locality = %.6f", (float)hits_num / access_num); + printf("\nBank_Level_Parallism = %.6f", (float)banks_1time / banks_acess_total); + printf("\nBank_Level_Parallism_Col = %.6f", (float)banks_time_rw / banks_access_rw_total); + printf("\nBank_Level_Parallism_Ready = %.6f", (float)banks_time_ready /banks_access_ready_total); + printf("\nwrite_to_read_ratio_blp_rw_average = %.6f", write_to_read_ratio_blp_rw_average /banks_access_rw_total); + printf("\nGrpLevelPara = %.6f \n", (float)bkgrp_parallsim_rw /banks_access_rw_total); + + printf("\nbwutil = %.6f \n", (float)bwutil/n_cmd); + printf("total_CMD = %d \n", n_cmd); + printf("util_bw = %d \n", util_bw); + printf("Wasted_Col = %d \n", wasted_bw_col); + printf("Wasted_Row %d \n", wasted_bw_row); + printf("Idle = %d \n\n", idle_bw); + + printf("RCDc_limit = %d \n", RCDc_limit); + printf("RCDWRc_limit = %d \n", RCDWRc_limit); + printf("WTRc_limit = %d \n", WTRc_limit); + printf("RTWc_limit = %d \n", RTWc_limit); + printf("CCDLc_limit %d \n", CCDLc_limit); + printf("rwq = %d \n", rwq_limit); + printf("CCDLc_limit_alone = %d \n", CCDLc_limit_alone); + printf("WTRc_limit_alone = %d \n", WTRc_limit_alone); + printf("RTWc_limit_alone = %d \n", RTWc_limit_alone); + + printf("total_CMD = %d \n", n_cmd); + printf("n_nop = %d \n", n_nop); + printf("Read = %d \n", n_rd); + printf("Write = %d \n",n_wr); + printf("L2_Alloc = %d \n", n_rd_L2_A); + printf("L2_WB = %d \n", n_wr_WB); + printf("n_act = %d \n", n_act); + printf("n_pre = %d \n", n_pre); + printf("n_ref = %d \n", n_ref); + printf("n_req = %d \n", n_req ); + printf("n_req4 = %d \n", n_req*4 ); + printf("total_req = %d \n\n", n_rd+n_wr+n_rd_L2_A+n_wr_WB); + + printf("issued_total_row = %lu \n", issued_total_row); + printf("issued_total_col = %lu \n", issued_total_col); + printf("Row_Bus_Util = %.6f \n", (float)issued_total_row / n_cmd); + printf("CoL_Bus_Util = %.6f \n", (float)issued_total_col / n_cmd); + printf("Either_Row_CoL_Bus_Util %.6f \n", (float)issued_total / n_cmd); + printf("Issued_on_Two_Bus_Simul_Util %.6f \n", (float)issued_two /n_cmd); + printf("issued_two_Eff = %.6f \n", (float)issued_two /issued_total); + printf("queue_avg = %.6f \n\n", (float)ave_mrqs/n_cmd ); + fprintf(simFile, "\n"); fprintf(simFile, "dram_util_bins:"); for (i=0;i<10;i++) fprintf(simFile, " %d", dram_util_bins[i]); fprintf(simFile, "\ndram_eff_bins:"); for (i=0;i<10;i++) fprintf(simFile, " %d", dram_eff_bins[i]); fprintf(simFile, "\n"); - if(m_config->scheduler_type== DRAM_FRFCFS) + if(m_config->scheduler_type== DRAM_FRFCFS) fprintf(simFile, "mrqq: max=%d avg=%g\n", max_mrqs, (float)ave_mrqs/n_cmd); } @@ -476,8 +779,8 @@ void dram_t::visualize() const void dram_t::print_stat( FILE* simFile ) { - fprintf(simFile,"DRAM (%d): n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_req=%d n_rd=%d n_write=%d bw_util=%.4g ", - id, n_cmd, n_nop, n_act, n_pre, n_req, n_rd, n_wr, + fprintf(simFile,"DRAM (%d): n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_ref=%d n_req=%d n_rd=%d n_write=%d bw_util=%.4g ", + id, n_cmd, n_nop, n_act, n_pre, n_ref, n_req, n_rd, n_wr, (float)bwutil/n_cmd); fprintf(simFile, "mrqq: %d %.4g mrqsmax=%d ", max_mrqs, (float)ave_mrqs/n_cmd, max_mrqs_temp); fprintf(simFile, "\n"); @@ -516,6 +819,7 @@ void dram_t::visualizer_print( gzFile visualizer_file ) n_pre_partial = 0; n_req_partial = 0; + // dram access type classification for (unsigned j = 0; j < m_config->nbk; j++) { gzprintf(visualizer_file,"dramglobal_acc_r: %u %u %u\n", id, j, @@ -553,3 +857,16 @@ void dram_t::set_dram_power_stats( unsigned &cmd, wr = n_wr; req = n_req; } + +unsigned dram_t::get_bankgrp_number(unsigned i) +{ + if(m_config->dram_bnkgrp_indexing_policy == 0) { //higher bits + return i>>m_config->bk_tag_length; + } + else if (m_config->dram_bnkgrp_indexing_policy == 1) { //lower bits + return i&((m_config->nbkgrp-1)); + } + else { + assert(1); + } +} diff --git a/src/gpgpu-sim/dram.h b/src/gpgpu-sim/dram.h index a8bff14..0ec5166 100644 --- a/src/gpgpu-sim/dram.h +++ b/src/gpgpu-sim/dram.h @@ -31,9 +31,15 @@ #include "delayqueue.h" #include +#include +#include +#include +#include +#include #include #include #include +#include #define READ 'R' //define read and write states #define WRITE 'W' @@ -42,7 +48,7 @@ class dram_req_t { public: - dram_req_t( class mem_fetch *data ); + dram_req_t( class mem_fetch *data , unsigned banks, unsigned dram_bnk_indexing_policy); unsigned int row; unsigned int col; @@ -106,6 +112,7 @@ public: class mem_fetch* return_queue_pop(); class mem_fetch* return_queue_top(); + void push( class mem_fetch *data ); void cycle(); void dram_log (int task); @@ -123,17 +130,24 @@ public: unsigned &wr, unsigned &req) const; -private: - void scheduler_fifo(); - void scheduler_frfcfs(); + const struct memory_config *m_config; +private: bankgrp_t **bkgrp; bank_t **bk; unsigned int prio; + unsigned get_bankgrp_number(unsigned i); + + void scheduler_fifo(); + void scheduler_frfcfs(); + + bool issue_col_command(int j); + bool issue_row_command(int j); + unsigned int RRDc; unsigned int CCDc; unsigned int RTWc; //read to write penalty applies across banks @@ -146,7 +160,7 @@ private: fifo_pipeline *rwq; fifo_pipeline *mrqq; //buffer to hold packets when DRAM processing is over - //should be filled with dram clock and popped with l2or icnt clock + //should be filled with dram clock and popped with l2or icnt clock fifo_pipeline *returnq; unsigned int dram_util_bins[10]; @@ -158,11 +172,46 @@ private: unsigned int n_nop; unsigned int n_act; unsigned int n_pre; + unsigned int n_ref; unsigned int n_rd; + unsigned int n_rd_L2_A; unsigned int n_wr; + unsigned int n_wr_WB; unsigned int n_req; unsigned int max_mrqs_temp; + unsigned wasted_bw_row; + unsigned wasted_bw_col; + unsigned util_bw; + unsigned idle_bw; + unsigned RCDc_limit; + unsigned CCDLc_limit; + unsigned CCDLc_limit_alone; + unsigned CCDc_limit; + unsigned WTRc_limit; + unsigned WTRc_limit_alone; + unsigned RCDWRc_limit; + unsigned RTWc_limit; + unsigned RTWc_limit_alone; + unsigned rwq_limit; + + //row locality, BLP and other statistics + unsigned long access_num; + unsigned long long hits_num; + unsigned long long banks_1time; + unsigned long long banks_acess_total; + unsigned long long banks_acess_total_after; + unsigned long long banks_time_rw; + unsigned long long banks_access_rw_total; + unsigned long long banks_time_ready; + unsigned long long banks_access_ready_total; + unsigned long long issued_two; + unsigned long long issued_total; + unsigned long long issued_total_row; + unsigned long long issued_total_col; + double write_to_read_ratio_blp_rw_average; + unsigned long long bkgrp_parallsim_rw; + unsigned int bwutil; unsigned int max_mrqs; unsigned int ave_mrqs; diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc index 8303e86..008b5bb 100644 --- a/src/gpgpu-sim/dram_sched.cc +++ b/src/gpgpu-sim/dram_sched.cc @@ -78,6 +78,9 @@ void frfcfs_scheduler::data_collection(unsigned int bank) dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row ) { + //row + bool rowhit = true; + if ( m_last_row[bank] == NULL ) { if ( m_queue[bank].empty() ) return NULL; @@ -89,11 +92,17 @@ dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row ) assert( bin_ptr != m_bins[bank].end() ); // where did the request go??? m_last_row[bank] = &(bin_ptr->second); data_collection(bank); + rowhit = false; } else { m_last_row[bank] = &(bin_ptr->second); - + rowhit = true; } } + //rowblp + m_dram->access_num++; + if(rowhit) + m_dram->hits_num++; + std::list::iterator next = m_last_row[bank]->back(); dram_req_t *req = (*next); diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 483a7b6..8a69847 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -192,6 +192,15 @@ void memory_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-dram_latency", OPT_UINT32, &dram_latency, "DRAM latency (default 30)", "30"); + option_parser_register(opp, "-dual_bus_interface", OPT_UINT32, &dual_bus_interface, + "dual_bus_interface", + "0"); + option_parser_register(opp, "-dram_bnk_indexing_policy", OPT_UINT32, &dram_bnk_indexing_policy, + "dram_bnk_indexing_policy", + "0"); + option_parser_register(opp, "-dram_bnkgrp_indexing_policy", OPT_UINT32, &dram_bnkgrp_indexing_policy, + "dram_bnkgrp_indexing_policy", + "0"); m_address_mapping.addrdec_setoption(opp); } diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 7d92c66..f379a17 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -270,6 +270,10 @@ struct memory_config { linear_to_raw_address_translation m_address_mapping; unsigned icnt_flit_size; + + unsigned dram_bnk_indexing_policy; + unsigned dram_bnkgrp_indexing_policy; + bool dual_bus_interface; }; // global counters and flags (please try not to add to this list!!!) -- cgit v1.3 From 355de7f8d033a4b5082cb0816650d880f05d81c8 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 13 Sep 2017 16:03:39 -0400 Subject: Adding sperate dp_unit --- configs/GTX480/gpgpusim.config | 6 ++- configs/GeForceGTX750Ti/gpgpusim.config | 5 ++- configs/QuadroFX5600/gpgpusim.config | 5 ++- configs/QuadroFX5800/gpgpusim.config | 5 ++- configs/TeslaC2050/gpgpusim.config | 3 +- src/abstract_hardware_model.h | 2 + src/cuda-sim/cuda-sim.cc | 5 ++- src/gpgpu-sim/gpu-sim.cc | 12 ++++++ src/gpgpu-sim/shader.cc | 70 +++++++++++++++++++++++++++++---- src/gpgpu-sim/shader.h | 41 ++++++++++++++++--- 10 files changed, 132 insertions(+), 22 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/GTX480/gpgpusim.config b/configs/GTX480/gpgpusim.config index 436cb41..7d8d91e 100644 --- a/configs/GTX480/gpgpusim.config +++ b/configs/GTX480/gpgpusim.config @@ -29,10 +29,12 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 +# 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 +#For Fermi, DP unit =0, DP inst is executed on SFU unit instead +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 -gpgpu_num_sp_units 2 -gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" diff --git a/configs/GeForceGTX750Ti/gpgpusim.config b/configs/GeForceGTX750Ti/gpgpusim.config index 8b030b6..bd6412c 100644 --- a/configs/GeForceGTX750Ti/gpgpusim.config +++ b/configs/GeForceGTX750Ti/gpgpusim.config @@ -28,10 +28,11 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 +# 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 +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 -gpgpu_num_sp_units 8 -gpgpu_num_sfu_units 32 +-gpgpu_num_dp_units 0 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" diff --git a/configs/QuadroFX5600/gpgpusim.config b/configs/QuadroFX5600/gpgpusim.config index cb87b65..82dc64a 100644 --- a/configs/QuadroFX5600/gpgpusim.config +++ b/configs/QuadroFX5600/gpgpusim.config @@ -17,10 +17,11 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,1,1,1,1,1,1 +# 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 +-gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 -gpgpu_num_sp_units 1 -gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" diff --git a/configs/QuadroFX5800/gpgpusim.config b/configs/QuadroFX5800/gpgpusim.config index 82243c2..fdd87e0 100644 --- a/configs/QuadroFX5800/gpgpusim.config +++ b/configs/QuadroFX5800/gpgpusim.config @@ -16,10 +16,11 @@ -gpgpu_simd_model 1 # Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,1,1,1,1,1,1 +# 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 +-gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 -gpgpu_num_sp_units 1 -gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" diff --git a/configs/TeslaC2050/gpgpusim.config b/configs/TeslaC2050/gpgpusim.config index 442ab8b..4febbe5 100644 --- a/configs/TeslaC2050/gpgpusim.config +++ b/configs/TeslaC2050/gpgpusim.config @@ -33,9 +33,10 @@ # Pipeline widths and number of FUs # ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 -gpgpu_num_sp_units 2 -gpgpu_num_sfu_units 1 +-gpgpu_num_sfu_units 0 # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 910a9ed..7d7773c 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -76,6 +76,7 @@ enum uarch_op_t { NO_OP=-1, ALU_OP=1, SFU_OP, + DP_OP, ALU_SFU_OP, LOAD_OP, STORE_OP, @@ -131,6 +132,7 @@ typedef enum special_operations_t special_ops; // Required to identify for the p enum operation_pipeline_t { UNKOWN_OP, SP__OP, + DP__OP, SFU__OP, MEM__OP }; diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index d4ace76..c7c32fa 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -678,6 +678,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[0]; initiation_interval = dp_init[0]; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: @@ -699,6 +700,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[1]; initiation_interval = dp_init[1]; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: @@ -721,7 +723,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[2]; initiation_interval = dp_init[2]; - op = ALU_SFU_OP; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: @@ -744,6 +746,7 @@ void ptx_instruction::set_opcode_and_latency() case FF64_TYPE: latency = dp_latency[3]; initiation_interval = dp_init[3]; + op = DP_OP; break; case B32_TYPE: case U32_TYPE: diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 8a69847..c5930fc 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -321,6 +321,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_operand_collector_num_units_sp", OPT_INT32, &gpgpu_operand_collector_num_units_sp, "number of collector units (default = 4)", "4"); + option_parser_register(opp, "-gpgpu_operand_collector_num_units_dp", OPT_INT32, &gpgpu_operand_collector_num_units_dp, + "number of collector units (default = 0)", + "0"); option_parser_register(opp, "-gpgpu_operand_collector_num_units_sfu", OPT_INT32, &gpgpu_operand_collector_num_units_sfu, "number of collector units (default = 4)", "4"); @@ -333,6 +336,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_sp", OPT_INT32, &gpgpu_operand_collector_num_in_ports_sp, "number of collector unit in ports (default = 1)", "1"); + option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_dp", OPT_INT32, &gpgpu_operand_collector_num_in_ports_dp, + "number of collector unit in ports (default = 0)", + "0"); option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_sfu", OPT_INT32, &gpgpu_operand_collector_num_in_ports_sfu, "number of collector unit in ports (default = 1)", "1"); @@ -345,6 +351,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_sp", OPT_INT32, &gpgpu_operand_collector_num_out_ports_sp, "number of collector unit in ports (default = 1)", "1"); + option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_dp", OPT_INT32, &gpgpu_operand_collector_num_out_ports_dp, + "number of collector unit in ports (default = 0)", + "0"); option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_sfu", OPT_INT32, &gpgpu_operand_collector_num_out_ports_sfu, "number of collector unit in ports (default = 1)", "1"); @@ -376,6 +385,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_num_sp_units", OPT_INT32, &gpgpu_num_sp_units, "Number of SP units (default=1)", "1"); + option_parser_register(opp, "-gpgpu_num_dp_units", OPT_INT32, &gpgpu_num_dp_units, + "Number of DP units (default=0)", + "0"); option_parser_register(opp, "-gpgpu_num_sfu_units", OPT_INT32, &gpgpu_num_sfu_units, "Number of SF units (default=1)", "1"); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index bb2cf0e..5547a18 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -150,6 +150,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_simt_stack, &m_warp, &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_MEM], i @@ -164,6 +165,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_simt_stack, &m_warp, &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_MEM], i, @@ -179,6 +181,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_simt_stack, &m_warp, &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_MEM], i @@ -193,6 +196,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_simt_stack, &m_warp, &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_MEM], i @@ -207,6 +211,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_simt_stack, &m_warp, &m_pipeline_reg[ID_OC_SP], + &m_pipeline_reg[ID_OC_DP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_MEM], i, @@ -228,8 +233,9 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, } //op collector configuration - enum { SP_CUS, SFU_CUS, MEM_CUS, GEN_CUS }; + enum { SP_CUS, DP_CUS, SFU_CUS, MEM_CUS, GEN_CUS }; m_operand_collector.add_cu_set(SP_CUS, m_config->gpgpu_operand_collector_num_units_sp, m_config->gpgpu_operand_collector_num_out_ports_sp); + m_operand_collector.add_cu_set(DP_CUS, m_config->gpgpu_operand_collector_num_units_sp, m_config->gpgpu_operand_collector_num_out_ports_dp); m_operand_collector.add_cu_set(SFU_CUS, m_config->gpgpu_operand_collector_num_units_sfu, m_config->gpgpu_operand_collector_num_out_ports_sfu); m_operand_collector.add_cu_set(MEM_CUS, m_config->gpgpu_operand_collector_num_units_mem, m_config->gpgpu_operand_collector_num_out_ports_mem); m_operand_collector.add_cu_set(GEN_CUS, m_config->gpgpu_operand_collector_num_units_gen, m_config->gpgpu_operand_collector_num_out_ports_gen); @@ -246,6 +252,15 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, in_ports.clear(),out_ports.clear(),cu_sets.clear(); } + for (unsigned i = 0; i < m_config->gpgpu_operand_collector_num_in_ports_dp; i++) { + in_ports.push_back(&m_pipeline_reg[ID_OC_DP]); + out_ports.push_back(&m_pipeline_reg[OC_EX_DP]); + cu_sets.push_back((unsigned)DP_CUS); + cu_sets.push_back((unsigned)GEN_CUS); + m_operand_collector.add_port(in_ports,out_ports,cu_sets); + in_ports.clear(),out_ports.clear(),cu_sets.clear(); + } + for (unsigned i = 0; i < m_config->gpgpu_operand_collector_num_in_ports_sfu; i++) { in_ports.push_back(&m_pipeline_reg[ID_OC_SFU]); out_ports.push_back(&m_pipeline_reg[OC_EX_SFU]); @@ -280,7 +295,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_operand_collector.init( m_config->gpgpu_num_reg_banks, this ); // execute - m_num_function_units = m_config->gpgpu_num_sp_units + m_config->gpgpu_num_sfu_units + 1; // sp_unit, sfu, ldst_unit + m_num_function_units = m_config->gpgpu_num_sp_units + m_config->gpgpu_num_dp_units + m_config->gpgpu_num_sfu_units + 1; // sp_unit, sfu, ldst_unit //m_dispatch_port = new enum pipeline_stage_name_t[ m_num_function_units ]; //m_issue_port = new enum pipeline_stage_name_t[ m_num_function_units ]; @@ -292,12 +307,18 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_issue_port.push_back(OC_EX_SP); } + for (int k = 0; k < m_config->gpgpu_num_dp_units; k++) { + m_fu.push_back(new dp_unit( &m_pipeline_reg[EX_WB], m_config, this )); + m_dispatch_port.push_back(ID_OC_DP); + m_issue_port.push_back(OC_EX_DP); + } + for (int k = 0; k < m_config->gpgpu_num_sfu_units; k++) { m_fu.push_back(new sfu( &m_pipeline_reg[EX_WB], m_config, this )); m_dispatch_port.push_back(ID_OC_SFU); m_issue_port.push_back(OC_EX_SFU); } - + m_ldst_unit = new ldst_unit( m_icnt, m_mem_fetch_allocator, this, &m_operand_collector, m_scoreboard, config, mem_config, stats, shader_id, tpc_id ); m_fu.push_back(m_ldst_unit); m_dispatch_port.push_back(ID_OC_MEM); @@ -754,7 +775,7 @@ void shader_core_ctx::issue(){ unsigned j; for (unsigned i = 0; i < schedulers.size(); i++) { j = (Issue_Prio + i) % schedulers.size(); - schedulers[j]->cycle(); + schedulers[j]->cycle(); } Issue_Prio = (Issue_Prio+1)% schedulers.size(); @@ -879,6 +900,7 @@ void scheduler_unit::cycle() exec_unit_type_t previous_issued_inst_exec_type = exec_unit_type_t::NONE; unsigned max_issue = m_shader->m_config->gpgpu_max_insn_issue_per_warp; bool diff_exec_units = m_shader->m_config->gpgpu_dual_issue_diff_exec_units; + while( !warp(warp_id).waiting() && !warp(warp_id).ibuffer_empty() && (checked < max_issue) && (checked <= issued) && (issued < max_issue) ) { const warp_inst_t *pI = warp(warp_id).ibuffer_next_inst(); //Jin: handle cdp latency; @@ -920,9 +942,11 @@ void scheduler_unit::cycle() previous_issued_inst_exec_type = exec_unit_type_t::MEM; } } else { + bool sp_pipe_avail = m_sp_out->has_free(); bool sfu_pipe_avail = m_sfu_out->has_free(); - if( sp_pipe_avail && (pI->op != SFU_OP) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::SP)) { + bool dp_pipe_avail = m_dp_out->has_free(); + if( sp_pipe_avail && (pI->op != SFU_OP && pI->op != DP_OP) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::SP)) { //Jin: special for CDP api if(pI->m_is_cdp && !warp(warp_id).m_cdp_dummy) { @@ -948,7 +972,17 @@ void scheduler_unit::cycle() issued_inst=true; warp_inst_issued = true; previous_issued_inst_exec_type = exec_unit_type_t::SP; - } else if ( (pI->op == SFU_OP) || (pI->op == ALU_SFU_OP) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::SFU)) { + } else if ( (m_shader->m_config->gpgpu_num_dp_units != 0) && (pI->op == DP_OP) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::DP)) { + if( dp_pipe_avail ) { + m_shader->issue_warp(*m_dp_out,pI,active_mask,warp_id); + issued++; + issued_inst=true; + warp_inst_issued = true; + previous_issued_inst_exec_type = exec_unit_type_t::DP; + std::cout<<"DP inst is issued"<m_config->gpgpu_num_dp_units == 0 && pI->op == DP_OP) || (pI->op == SFU_OP) || (pI->op == ALU_SFU_OP)) && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::SFU)) { if( sfu_pipe_avail ) { m_shader->issue_warp(*m_sfu_out,pI,active_mask,warp_id); issued++; @@ -1139,11 +1173,12 @@ swl_scheduler::swl_scheduler ( shader_core_stats* stats, shader_core_ctx* shader Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* mem_out, int id, char* config_string ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, mem_out, id ) + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, mem_out, id ) { unsigned m_prioritization_readin; int ret = sscanf( config_string, @@ -1588,6 +1623,13 @@ void sp_unit::active_lanes_in_pipeline(){ m_core->incfuactivelanes_stat(active_count); m_core->incfumemactivelanes_stat(active_count); } +void dp_unit::active_lanes_in_pipeline(){ + unsigned active_count=pipelined_simd_unit::get_active_lanes_in_pipeline(); + assert(active_count<=m_core->get_config()->warp_size); + m_core->incspactivelanes_stat(active_count); + m_core->incfuactivelanes_stat(active_count); + m_core->incfumemactivelanes_stat(active_count); +} void sfu::active_lanes_in_pipeline(){ unsigned active_count=pipelined_simd_unit::get_active_lanes_in_pipeline(); @@ -1603,6 +1645,12 @@ sp_unit::sp_unit( register_set* result_port, const shader_core_config *config,sh m_name = "SP "; } +dp_unit::dp_unit( register_set* result_port, const shader_core_config *config,shader_core_ctx *core) + : pipelined_simd_unit(result_port,config,config->max_sfu_latency,core) +{ + m_name = "DP "; +} + void sp_unit :: issue(register_set& source_reg) { warp_inst_t** ready_reg = source_reg.get_ready(); @@ -1612,6 +1660,14 @@ void sp_unit :: issue(register_set& source_reg) pipelined_simd_unit::issue(source_reg); } +void dp_unit :: issue(register_set& source_reg) +{ + warp_inst_t** ready_reg = source_reg.get_ready(); + //m_core->incexecstat((*ready_reg)); + (*ready_reg)->op_pipe=DP__OP; + m_core->incsp_stat(m_core->get_config()->warp_size,(*ready_reg)->latency); + pipelined_simd_unit::issue(source_reg); +} pipelined_simd_unit::pipelined_simd_unit( register_set* result_port, const shader_core_config *config, unsigned max_latency,shader_core_ctx *core ) : simd_function_unit(config) diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 2fed420..544d8d8 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -326,12 +326,13 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* mem_out, int id) : m_supervised_warps(), m_stats(stats), m_shader(shader), m_scoreboard(scoreboard), m_simt_stack(simt), /*m_pipeline_reg(pipe_regs),*/ m_warp(warp), - m_sp_out(sp_out),m_sfu_out(sfu_out),m_mem_out(mem_out), m_id(id){} + m_sp_out(sp_out),m_dp_out(dp_out),m_sfu_out(sfu_out),m_mem_out(mem_out), m_id(id){} virtual ~scheduler_unit(){} virtual void add_supervised_warp_id(int i) { m_supervised_warps.push_back(&warp(i)); @@ -403,6 +404,7 @@ protected: //warp_inst_t** m_pipeline_reg; std::vector* m_warp; register_set* m_sp_out; + register_set* m_dp_out; register_set* m_sfu_out; register_set* m_mem_out; @@ -415,10 +417,11 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* mem_out, int id ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, mem_out, id ){} virtual ~lrr_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -432,10 +435,11 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* mem_out, int id ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, mem_out, id ){} virtual ~gto_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -450,10 +454,11 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* mem_out, int id ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, mem_out, id ){} virtual ~oldest_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -468,11 +473,12 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* mem_out, int id, char* config_str ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, mem_out, id ), + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out, mem_out, id ), m_pending_warps() { unsigned inner_level_readin; @@ -518,6 +524,7 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* mem_out, int id, @@ -1088,6 +1095,22 @@ public: virtual void issue( register_set& source_reg ); }; +class dp_unit : public pipelined_simd_unit +{ +public: + dp_unit( register_set* result_port, const shader_core_config *config, shader_core_ctx *core ); + virtual bool can_issue( const warp_inst_t &inst ) const + { + switch(inst.op) { + case DP_OP: break; + default: return false; + } + return pipelined_simd_unit::can_issue(inst); + } + virtual void active_lanes_in_pipeline(); + virtual void issue( register_set& source_reg ); +}; + class sp_unit : public pipelined_simd_unit { public: @@ -1226,9 +1249,11 @@ protected: enum pipeline_stage_name_t { ID_OC_SP=0, + ID_OC_DP, ID_OC_SFU, ID_OC_MEM, OC_EX_SP, + OC_EX_DP, OC_EX_SFU, OC_EX_MEM, EX_WB, @@ -1237,9 +1262,11 @@ enum pipeline_stage_name_t { const char* const pipeline_stage_name_decode[] = { "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", @@ -1328,21 +1355,25 @@ struct shader_core_config : public core_config //op collector int gpgpu_operand_collector_num_units_sp; + int gpgpu_operand_collector_num_units_dp; int gpgpu_operand_collector_num_units_sfu; int gpgpu_operand_collector_num_units_mem; int gpgpu_operand_collector_num_units_gen; unsigned int gpgpu_operand_collector_num_in_ports_sp; + unsigned int gpgpu_operand_collector_num_in_ports_dp; unsigned int gpgpu_operand_collector_num_in_ports_sfu; unsigned int gpgpu_operand_collector_num_in_ports_mem; unsigned int gpgpu_operand_collector_num_in_ports_gen; unsigned int gpgpu_operand_collector_num_out_ports_sp; + unsigned int gpgpu_operand_collector_num_out_ports_dp; unsigned int gpgpu_operand_collector_num_out_ports_sfu; unsigned int gpgpu_operand_collector_num_out_ports_mem; unsigned int gpgpu_operand_collector_num_out_ports_gen; int gpgpu_num_sp_units; + int gpgpu_num_dp_units; int gpgpu_num_sfu_units; int gpgpu_num_mem_units; -- cgit v1.3 From 0f8404321fbf71ec86297a65149a5c27ba5ae528 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 14 Sep 2017 20:18:59 -0400 Subject: adding some condig comments --- src/gpgpu-sim/gpu-sim.cc | 12 ++++++------ 1 file changed, 6 insertions(+), 6 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 c5930fc..b424d2c 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -193,13 +193,13 @@ void memory_config::reg_options(class OptionParser * opp) "DRAM latency (default 30)", "30"); option_parser_register(opp, "-dual_bus_interface", OPT_UINT32, &dual_bus_interface, - "dual_bus_interface", + "dual_bus_interface (default = 0) ", "0"); option_parser_register(opp, "-dram_bnk_indexing_policy", OPT_UINT32, &dram_bnk_indexing_policy, - "dram_bnk_indexing_policy", + "dram_bnk_indexing_policy (0 = normal indexing, 1 = Xoring with the higher bits) (Default = 0)", "0"); option_parser_register(opp, "-dram_bnkgrp_indexing_policy", OPT_UINT32, &dram_bnkgrp_indexing_policy, - "dram_bnkgrp_indexing_policy", + "dram_bnkgrp_indexing_policy (0 = take higher bits, 1 = take lower bits) (Default = 0)", "0"); m_address_mapping.addrdec_setoption(opp); @@ -373,15 +373,15 @@ void shader_core_config::reg_options(class OptionParser * opp) "Max number of instructions that can be issued per warp in one cycle by scheduler (either 1 or 2)", "2"); option_parser_register(opp, "-gpgpu_dual_issue_diff_exec_units", OPT_BOOL, &gpgpu_dual_issue_diff_exec_units, - "should dual issue use two different execution unit resources", + "should dual issue use two different execution unit resources (Default = 1)", "1"); option_parser_register(opp, "-gpgpu_simt_core_sim_order", OPT_INT32, &simt_core_sim_order, "Select the simulation order of cores in a cluster (0=Fix, 1=Round-Robin)", "1"); option_parser_register(opp, "-gpgpu_pipeline_widths", OPT_CSTR, &pipeline_widths_string, "Pipeline widths " - "ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB", - "1,1,1,1,1,1,1" ); + "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", + "1,1,1,1,1,1,1,1,1" ); option_parser_register(opp, "-gpgpu_num_sp_units", OPT_INT32, &gpgpu_num_sp_units, "Number of SP units (default=1)", "1"); -- cgit v1.3 From 07b375cc0c0da2b8a8ca7263f11ecca8ebaf8fe9 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 13 Oct 2017 22:20:57 -0400 Subject: ensure that first launch uses simt cluster 0 --- src/gpgpu-sim/gpu-sim.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (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 eac92b4..470fcf4 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -671,7 +671,7 @@ gpgpu_sim::gpgpu_sim( const gpgpu_sim_config &config ) m_running_kernels.resize( config.max_concurrent_kernel, NULL ); m_last_issued_kernel = 0; - m_last_cluster_issue = 0; + m_last_cluster_issue = m_shader_config->n_simt_clusters-1; // this causes first launch to use simt cluster 0 *average_pipeline_duty_cycle=0; *active_sms=0; -- cgit v1.3 From 50b7ac49a78948f61fa685d717de90feaa277b9b Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 25 Oct 2017 13:15:44 -0400 Subject: The commits includes: 1- REEAD/WERITE buffer for DRAM 2- Fixing FETCH_ON_WRITE cahce policy bug --- src/gpgpu-sim/dram.cc | 11 +++++- src/gpgpu-sim/dram.h | 2 +- src/gpgpu-sim/dram_sched.cc | 95 ++++++++++++++++++++++++++++++++++----------- src/gpgpu-sim/dram_sched.h | 12 ++++++ src/gpgpu-sim/gpu-cache.cc | 34 ++++++++++++---- src/gpgpu-sim/gpu-sim.cc | 7 +++- src/gpgpu-sim/gpu-sim.h | 9 +++++ src/gpgpu-sim/l2cache.cc | 29 ++++++++------ src/gpgpu-sim/mem_fetch.cc | 4 +- src/gpgpu-sim/mem_fetch.h | 4 +- 10 files changed, 160 insertions(+), 47 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index 5c1ddab..7ed7b6f 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -149,11 +149,18 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m } -bool dram_t::full() const +bool dram_t::full(bool is_write) const { if(m_config->scheduler_type == DRAM_FRFCFS){ if(m_config->gpgpu_frfcfs_dram_sched_queue_size == 0 ) return false; - return m_frfcfs_scheduler->num_pending() >= m_config->gpgpu_frfcfs_dram_sched_queue_size; + if(m_config->seperate_write_queue_enabled){ + if(is_write) + return m_frfcfs_scheduler->num_write_pending() >= m_config->gpgpu_frfcfs_dram_write_queue_size; + else + return m_frfcfs_scheduler->num_pending() >= m_config->gpgpu_frfcfs_dram_sched_queue_size; + } + else + return m_frfcfs_scheduler->num_pending() >= m_config->gpgpu_frfcfs_dram_sched_queue_size; } else return mrqq->full(); } diff --git a/src/gpgpu-sim/dram.h b/src/gpgpu-sim/dram.h index 331b4f1..29731a7 100644 --- a/src/gpgpu-sim/dram.h +++ b/src/gpgpu-sim/dram.h @@ -101,7 +101,7 @@ public: dram_t( unsigned int parition_id, const struct memory_config *config, class memory_stats_t *stats, class memory_partition_unit *mp ); - bool full() const; + bool full(bool is_write) const; void print( FILE* simFile ) const; void visualize() const; void print_stat( FILE* simFile ); diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc index 008b5bb..7a140c5 100644 --- a/src/gpgpu-sim/dram_sched.cc +++ b/src/gpgpu-sim/dram_sched.cc @@ -36,6 +36,7 @@ frfcfs_scheduler::frfcfs_scheduler( const memory_config *config, dram_t *dm, mem m_config = config; m_stats = stats; m_num_pending = 0; + m_num_write_pending = 0; m_dram = dm; m_queue = new std::list[m_config->nbk]; m_bins = new std::map::iterator> >[ m_config->nbk ]; @@ -49,15 +50,36 @@ frfcfs_scheduler::frfcfs_scheduler( const memory_config *config, dram_t *dm, mem curr_row_service_time[i] = 0; row_service_timestamp[i] = 0; } + if(m_config->seperate_write_queue_enabled) { + m_write_queue = new std::list[m_config->nbk]; + m_write_bins = new std::map::iterator> >[ m_config->nbk ]; + m_last_write_row = new std::list::iterator>*[ m_config->nbk ]; + + for ( unsigned i=0; i < m_config->nbk; i++ ) { + m_write_queue[i].clear(); + m_write_bins[i].clear(); + m_last_write_row[i] = NULL; + } + } + m_mode = READ_MODE; } void frfcfs_scheduler::add_req( dram_req_t *req ) { - m_num_pending++; - m_queue[req->bk].push_front(req); - std::list::iterator ptr = m_queue[req->bk].begin(); - m_bins[req->bk][req->row].push_front( ptr ); //newest reqs to the front + if(m_config->seperate_write_queue_enabled && req->data->is_write()) { + assert(m_num_write_pending < m_config->gpgpu_frfcfs_dram_write_queue_size); + m_num_write_pending++; + m_write_queue[req->bk].push_front(req); + std::list::iterator ptr = m_write_queue[req->bk].begin(); + m_write_bins[req->bk][req->row].push_front( ptr ); //newest reqs to the front + } else { + assert(m_num_pending < m_config->gpgpu_frfcfs_dram_sched_queue_size); + m_num_pending++; + m_queue[req->bk].push_front(req); + std::list::iterator ptr = m_queue[req->bk].begin(); + m_bins[req->bk][req->row].push_front( ptr ); //newest reqs to the front + } } void frfcfs_scheduler::data_collection(unsigned int bank) @@ -80,21 +102,43 @@ dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row ) { //row bool rowhit = true; + std::list *m_current_queue = m_queue; + std::map::iterator> > *m_current_bins = m_bins ; + std::list::iterator> **m_current_last_row = m_last_row; + + if(m_config->seperate_write_queue_enabled) { + if(m_mode == READ_MODE && + ((m_num_write_pending >= m_config->write_high_watermark ) + || (m_queue[bank].empty() && !m_write_queue[bank].empty()))) { + m_mode = WRITE_MODE; + } + else if(m_mode == WRITE_MODE && + (( m_num_write_pending < m_config->write_low_watermark ) + || (!m_queue[bank].empty() && m_write_queue[bank].empty()))){ + m_mode = READ_MODE; + } + } + + if(m_mode == WRITE_MODE) { + m_current_queue = m_write_queue; + m_current_bins = m_write_bins ; + m_current_last_row = m_last_write_row; + } - if ( m_last_row[bank] == NULL ) { - if ( m_queue[bank].empty() ) + if ( m_current_last_row[bank] == NULL ) { + if ( m_current_queue[bank].empty() ) return NULL; - std::map::iterator> >::iterator bin_ptr = m_bins[bank].find( curr_row ); - if ( bin_ptr == m_bins[bank].end()) { - dram_req_t *req = m_queue[bank].back(); - bin_ptr = m_bins[bank].find( req->row ); - assert( bin_ptr != m_bins[bank].end() ); // where did the request go??? - m_last_row[bank] = &(bin_ptr->second); + std::map::iterator> >::iterator bin_ptr = m_current_bins[bank].find( curr_row ); + if ( bin_ptr == m_current_bins[bank].end()) { + dram_req_t *req = m_current_queue[bank].back(); + bin_ptr = m_current_bins[bank].find( req->row ); + assert( bin_ptr != m_current_bins[bank].end() ); // where did the request go??? + m_current_last_row[bank] = &(bin_ptr->second); data_collection(bank); rowhit = false; } else { - m_last_row[bank] = &(bin_ptr->second); + m_current_last_row[bank] = &(bin_ptr->second); rowhit = true; } } @@ -103,25 +147,32 @@ dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row ) if(rowhit) m_dram->hits_num++; - std::list::iterator next = m_last_row[bank]->back(); + std::list::iterator next = m_current_last_row[bank]->back(); dram_req_t *req = (*next); m_stats->concurrent_row_access[m_dram->id][bank]++; m_stats->row_access[m_dram->id][bank]++; - m_last_row[bank]->pop_back(); + m_current_last_row[bank]->pop_back(); - m_queue[bank].erase(next); - if ( m_last_row[bank]->empty() ) { - m_bins[bank].erase( req->row ); - m_last_row[bank] = NULL; + m_current_queue[bank].erase(next); + if ( m_current_last_row[bank]->empty() ) { + m_current_bins[bank].erase( req->row ); + m_current_last_row[bank] = NULL; } #ifdef DEBUG_FAST_IDEAL_SCHED if ( req ) printf("%08u : DRAM(%u) scheduling memory request to bank=%u, row=%u\n", (unsigned)gpu_sim_cycle, m_dram->id, req->bk, req->row ); #endif - assert( req != NULL && m_num_pending != 0 ); - m_num_pending--; + + if(m_config->seperate_write_queue_enabled && req->data->is_write()) { + assert( req != NULL && m_num_write_pending != 0 ); + m_num_write_pending--; + } + else { + assert( req != NULL && m_num_pending != 0 ); + m_num_pending--; + } return req; } @@ -138,7 +189,7 @@ void dram_t::scheduler_frfcfs() { unsigned mrq_latency; frfcfs_scheduler *sched = m_frfcfs_scheduler; - while ( !mrqq->empty() && (!m_config->gpgpu_frfcfs_dram_sched_queue_size || sched->num_pending() < m_config->gpgpu_frfcfs_dram_sched_queue_size)) { + while ( !mrqq->empty() ) { dram_req_t *req = mrqq->pop(); // Power stats diff --git a/src/gpgpu-sim/dram_sched.h b/src/gpgpu-sim/dram_sched.h index 3860f5b..63f5831 100644 --- a/src/gpgpu-sim/dram_sched.h +++ b/src/gpgpu-sim/dram_sched.h @@ -35,6 +35,11 @@ #include #include +enum memory_mode { + READ_MODE = 0, + WRITE_MODE +}; + class frfcfs_scheduler { public: frfcfs_scheduler( const memory_config *config, dram_t *dm, memory_stats_t *stats ); @@ -43,17 +48,24 @@ public: dram_req_t *schedule( unsigned bank, unsigned curr_row ); void print( FILE *fp ); unsigned num_pending() const { return m_num_pending;} + unsigned num_write_pending() const { return m_num_write_pending;} private: const memory_config *m_config; dram_t *m_dram; unsigned m_num_pending; + unsigned m_num_write_pending; std::list *m_queue; std::map::iterator> > *m_bins; std::list::iterator> **m_last_row; unsigned *curr_row_service_time; //one set of variables for each bank. unsigned *row_service_timestamp; //tracks when scheduler began servicing current row + std::list *m_write_queue; + std::map::iterator> > *m_write_bins; + std::list::iterator> **m_last_write_row; + + enum memory_mode m_mode; memory_stats_t *m_stats; }; diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index eadc094..d199cca 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -945,10 +945,9 @@ void baseline_cache::send_read_request(new_addr_type addr, new_addr_type block_a mf->set_addr( block_addr ); m_miss_queue.push_back(mf); mf->set_status(m_miss_queue_status,time); - if(wa) - events.push_back(cache_event(WRITE_ALLOCATE_SENT)); - else + if(!wa) events.push_back(cache_event(READ_REQUEST_SENT)); + do_miss = true; } else if(mshr_hit && !mshr_avail) @@ -1087,6 +1086,8 @@ data_cache::wr_miss_wa_naive( new_addr_type addr, send_read_request(addr, block_addr, cache_index, n_mf, time, do_miss, wb, evicted, events, false, true); + events.push_back(cache_event(WRITE_ALLOCATE_SENT)); + if( do_miss ){ // If evicted block is modified and not a write-through // (already modified lower level) @@ -1111,7 +1112,7 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, { new_addr_type block_addr = m_config.block_addr(addr); - new_addr_type mshr_addr = m_config.block_addr(mf->get_addr()); + new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) { @@ -1147,10 +1148,23 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, } else { - if(miss_queue_full(1)) { - m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); - return RESERVATION_FAIL; - } + bool mshr_hit = m_mshrs.probe(mshr_addr); + bool mshr_avail = !m_mshrs.full(mshr_addr); + if(miss_queue_full(1) + || (!(mshr_hit && mshr_avail) + && !(!mshr_hit && mshr_avail && (m_miss_queue.size() < m_config.m_miss_queue_size)))) { + //check what is the exactly the failure reason + if(miss_queue_full(1) ) + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + else if(mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL); + else if (!mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL); + else + assert(0); + + return RESERVATION_FAIL; + } //prevent Write - Read - Write in pending mshr @@ -1177,8 +1191,10 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, mf->get_sid(), mf->get_tpc(), mf->get_mem_config(), + NULL, mf); + new_addr_type block_addr = m_config.block_addr(addr); bool do_miss = false; bool wb = false; @@ -1191,6 +1207,8 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, cache_block_t* block = m_tag_array->get_block(cache_index); block->set_modified_on_fill(true, mf->get_access_sector_mask()); + events.push_back(cache_event(WRITE_ALLOCATE_SENT)); + if( do_miss ){ // If evicted block is modified and not a write-through // (already modified lower level) diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 470fcf4..7838875 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -201,7 +201,12 @@ void memory_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-dram_bnkgrp_indexing_policy", OPT_UINT32, &dram_bnkgrp_indexing_policy, "dram_bnkgrp_indexing_policy (0 = take higher bits, 1 = take lower bits) (Default = 0)", "0"); - + option_parser_register(opp, "-Seperate_Write_Queue_Enable", OPT_BOOL, &seperate_write_queue_enabled, + "Seperate_Write_Queue_Enable", + "0"); + option_parser_register(opp, "-Write_Queue_Size", OPT_CSTR, &write_queue_size_opt, + "Write_Queue_Size", + "32:28:16"); m_address_mapping.addrdec_setoption(opp); } diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index f379a17..197350b 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -214,6 +214,9 @@ struct memory_config { m_valid = true; icnt_flit_size = 32; // Default 32 + + sscanf(write_queue_size_opt,"%d:%d:%d", + &gpgpu_frfcfs_dram_write_queue_size,&write_high_watermark,&write_low_watermark); } void reg_options(class OptionParser * opp); @@ -274,6 +277,12 @@ struct memory_config { unsigned dram_bnk_indexing_policy; unsigned dram_bnkgrp_indexing_policy; bool dual_bus_interface; + + bool seperate_write_queue_enabled; + char *write_queue_size_opt; + unsigned gpgpu_frfcfs_dram_write_queue_size; + unsigned write_high_watermark; + unsigned write_low_watermark; }; // 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 f7323c5..cac59f1 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -93,7 +93,9 @@ memory_partition_unit::arbitration_metadata::arbitration_metadata(const struct m m_private_credit_limit = 1; m_shared_credit_limit = config->gpgpu_frfcfs_dram_sched_queue_size + config->gpgpu_dram_return_queue_size - - (config->m_n_sub_partition_per_memory_channel - 1); + - (config->m_n_sub_partition_per_memory_channel - 1); + if(config->seperate_write_queue_enabled ) + m_shared_credit_limit += config->gpgpu_frfcfs_dram_write_queue_size; if (config->gpgpu_frfcfs_dram_sched_queue_size == 0 or config->gpgpu_dram_return_queue_size == 0) { @@ -220,7 +222,8 @@ void memory_partition_unit::dram_cycle() m_dram->cycle(); m_dram->dram_log(SAMPLELOG); - if( !m_dram->full() ) { + // 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(); @@ -228,6 +231,9 @@ void memory_partition_unit::dram_cycle() 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; @@ -239,12 +245,13 @@ void memory_partition_unit::dram_cycle() break; // the DRAM should only accept one request per 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() ) { - mem_fetch* mf = m_dram_latency_queue.front().req; - m_dram_latency_queue.pop_front(); + + 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); } } @@ -343,12 +350,12 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); m_L2_icnt_queue->push(mf); }else{ - if(m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE && mf->original_mf) + if(m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE) { - assert(mf->original_mf); - mf->original_mf->set_reply(); - mf->original_mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); - m_L2_icnt_queue->push(mf->original_mf); + 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); } m_request_tracker.erase(mf); delete mf; diff --git a/src/gpgpu-sim/mem_fetch.cc b/src/gpgpu-sim/mem_fetch.cc index b8e918f..c05a693 100644 --- a/src/gpgpu-sim/mem_fetch.cc +++ b/src/gpgpu-sim/mem_fetch.cc @@ -40,7 +40,8 @@ mem_fetch::mem_fetch( const mem_access_t &access, unsigned sid, unsigned tpc, const class memory_config *config, - mem_fetch *m_original_mf) + mem_fetch *m_original_mf, + mem_fetch *m_original_wr_mf) { m_request_uid = sm_next_mf_request_uid++; m_access = access; @@ -63,6 +64,7 @@ mem_fetch::mem_fetch( const mem_access_t &access, m_mem_config = config; icnt_flit_size = config->icnt_flit_size; original_mf = m_original_mf; + original_wr_mf = m_original_wr_mf; } mem_fetch::~mem_fetch() diff --git a/src/gpgpu-sim/mem_fetch.h b/src/gpgpu-sim/mem_fetch.h index 76e7419..278cf32 100644 --- a/src/gpgpu-sim/mem_fetch.h +++ b/src/gpgpu-sim/mem_fetch.h @@ -56,7 +56,8 @@ public: unsigned sid, unsigned tpc, const class memory_config *config, - mem_fetch *original_mf = NULL); + mem_fetch *original_mf = NULL, + mem_fetch *original_wr_mf = NULL); ~mem_fetch(); void set_status( enum mem_fetch_status status, unsigned long long cycle ); @@ -115,6 +116,7 @@ public: unsigned get_num_flits(bool simt_to_mem); mem_fetch* original_mf; + mem_fetch* original_wr_mf; private: // request source information unsigned m_request_uid; -- cgit v1.3 From 161f9cefeaf216f48f93e1192c817997cf875cac Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 26 Oct 2017 11:19:52 -0400 Subject: Changing the Titan X config file to use the last modifications --- configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt | 4 +- configs/Pascal-P102-GDDR5X/gpgpusim.config | 46 +++++++++++++--------- src/gpgpu-sim/gpu-sim.cc | 3 ++ src/gpgpu-sim/gpu-sim.h | 1 - 4 files changed, 32 insertions(+), 22 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt index 602daee..58e596d 100644 --- a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt +++ b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt @@ -1,6 +1,6 @@ //21*1 fly with 32 flits per packet under gpgpusim injection mode use_map = 0; -flit_size = 32; +flit_size = 40; // currently we do not use this, see subnets below network_count = 2; @@ -17,7 +17,7 @@ routing_function = dest_tag; // Flow control num_vcs = 1; -vc_buf_size = 8; +vc_buf_size = 32; wait_for_tail_credit = 0; diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 8b02680..36d13af 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -12,8 +12,9 @@ -gpgpu_ptx_save_converted_ptxplus 0 # high level architecture configuration +# P102 has two semi-indp scheds per core, and two cores per cluster -gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 +-gpgpu_n_cores_per_cluster 2 -gpgpu_n_mem 12 -gpgpu_n_sub_partition_per_mchannel 2 @@ -24,20 +25,20 @@ -gpgpu_clock_domains 1417.0:2834.0:1417.0:2500.0 # shader core pipeline config --gpgpu_shader_registers 65536 +-gpgpu_shader_registers 32768 -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 +# This implies a maximum of 32 warps/SM +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 16 -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 GP102 has 4 SP SIMD units and 4 SFU units ## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,4,1,4,1,4,1,9 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 -gpgpu_num_dp_units 1 @@ -51,7 +52,6 @@ -ptx_opcode_initiation_fp 1,2,1,1,4 -ptx_opcode_latency_dp 8,19,8,8,330 -ptx_opcode_initiation_dp 8,8,8,8,130 --ptx_opcode_latency_sfu 8 -ptx_opcode_initiation_sfu 4 # ::,::::,::,:** @@ -60,13 +60,16 @@ # Pascal GP102 has 96KB Shared memory # Pascal GP102 has 64KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,S:128:8,8 --gpgpu_shmem_size 98304 +-gpgpu_cache:dl1 S:64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_shmem_size 49152 -gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 # 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:W:L,A:128:4,4:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:256:4,4:0,32 -gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 @@ -76,15 +79,14 @@ -gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 # enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units -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 4 --gpgpu_operand_collector_num_out_ports_sfu 4 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_operand_collector_num_in_ports_sfu 2 +-gpgpu_operand_collector_num_out_ports_sfu 2 -gpgpu_operand_collector_num_in_ports_mem 1 -gpgpu_operand_collector_num_out_ports_mem 1 -gpgpu_operand_collector_num_in_ports_dp 1 @@ -119,7 +121,7 @@ # 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 116 +-gpgpu_dram_return_queue_size 192 # for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) # 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition @@ -136,8 +138,14 @@ -gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 32:28:16 + # Pascal 102 has four schedulers per core --gpgpu_num_sched_per_core 4 +-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 diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 7838875..6179d46 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -207,6 +207,9 @@ void memory_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-Write_Queue_Size", OPT_CSTR, &write_queue_size_opt, "Write_Queue_Size", "32:28:16"); + option_parser_register(opp, "-icnt_flit_size", OPT_UINT32, &icnt_flit_size, + "icnt_flit_size", + "32"); m_address_mapping.addrdec_setoption(opp); } diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 197350b..52c4643 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -213,7 +213,6 @@ struct memory_config { m_L2_config.init(&m_address_mapping); m_valid = true; - icnt_flit_size = 32; // Default 32 sscanf(write_queue_size_opt,"%d:%d:%d", &gpgpu_frfcfs_dram_write_queue_size,&write_high_watermark,&write_low_watermark); -- cgit v1.3 From f23021ad8663636e1103bd75a742480cb6238435 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 27 Oct 2017 22:35:52 -0400 Subject: add more statistics and chaging Pascal config --- configs/Pascal-P102-GDDR5X/gpgpusim.config | 12 +++++++----- src/abstract_hardware_model.cc | 2 +- src/gpgpu-sim/dram.cc | 3 ++- src/gpgpu-sim/dram_sched.cc | 2 ++ src/gpgpu-sim/gpu-cache.cc | 2 ++ src/gpgpu-sim/gpu-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.h | 2 +- src/gpgpu-sim/l2cache.cc | 2 +- src/gpgpu-sim/mem_latency_stat.cc | 17 +++++++++++++---- src/gpgpu-sim/mem_latency_stat.h | 4 ++++ src/gpgpu-sim/shader.cc | 8 ++++---- 11 files changed, 38 insertions(+), 18 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 5419c51..e830023 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -59,22 +59,24 @@ # ** 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 64KB L1 cache +# Pascal GP102 has 24KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 S:64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_cache:dl1 N:32:128:6,L:L:m:N:H,S:128:8,16 -gpgpu_shmem_size 49152 -gmem_skip_L1D 1 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 # 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:F:L,A:256:4,4:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:F:L,A:128:4,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 +#-gpgpu_flush_l2_cache 1 # 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 # 12 KB Const -gpgpu_const_cache:l1 N:128:64:2,L:R:f:N:L,S:2:64,4 @@ -100,7 +102,7 @@ -gpgpu_shmem_limited_broadcast 0 -gpgpu_shmem_warp_parts 1 # Use Fermi Coalsce arhitetecture which is the same as Pascal --gpgpu_coalesce_arch 20 +-gpgpu_coalesce_arch 61 ## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units -gpgpu_max_insn_issue_per_warp 2 @@ -143,7 +145,7 @@ -dram_bnkgrp_indexing_policy 1 #-Seperate_Write_Queue_Enable 1 -#-Write_Queue_Size 32:28:16 +#-Write_Queue_Size 64:56:32 # Pascal 102 has four schedulers per core -gpgpu_num_sched_per_core 2 diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 51265fd..d2a155c 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -314,7 +314,7 @@ void warp_inst_t::generate_mem_accesses() break; case global_space: case local_space: case param_space_local: - if( m_config->gpgpu_coalesce_arch == 13 || m_config->gpgpu_coalesce_arch == 20) { + if( m_config->gpgpu_coalesce_arch >= 13 && m_config->gpgpu_coalesce_arch <= 62) { if(isatomic()) memory_coalescing_arch_atomic(is_write, access_type); else diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index 7ed7b6f..de37f64 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -570,9 +570,10 @@ bool dram_t::issue_col_command(int j) bkgrp[grp]->RTPLc = m_config->tRTPL; issued = true; if(bk[j]->mrq->data->get_access_type() == L2_WR_ALLOC_R) - n_rd_L2_A++; + n_rd_L2_A++; else n_rd++; + bwutil += m_config->BL/m_config->data_command_freq_ratio; bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; bk[j]->n_access++; diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc index 7a140c5..ac4c827 100644 --- a/src/gpgpu-sim/dram_sched.cc +++ b/src/gpgpu-sim/dram_sched.cc @@ -220,6 +220,8 @@ void dram_t::scheduler_frfcfs() bk[b]->mrq = req; if (m_config->gpgpu_memlatency_stat) { mrq_latency = gpu_sim_cycle + gpu_tot_sim_cycle - bk[b]->mrq->timestamp; + m_stats->tot_mrq_latency += mrq_latency; + m_stats->tot_mrq_num++; bk[b]->mrq->timestamp = gpu_tot_sim_cycle + gpu_sim_cycle; m_stats->mrq_lat_table[LOGB2(mrq_latency)]++; if (mrq_latency > m_stats->max_mrq_latency) { diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index d199cca..32c2bb1 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -333,8 +333,10 @@ void tag_array::fill( unsigned index, unsigned time, mem_fetch* mf) void tag_array::flush() { for (unsigned i=0; i < m_config.get_num_lines(); i++) + if(m_lines[i]->is_modified_line()) { for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; + } } float tag_array::windowed_miss_rate( ) const diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 6179d46..0e06c5c 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1412,7 +1412,7 @@ void gpgpu_sim::cycle() if (mf) { unsigned response_size = mf->get_is_write()?mf->get_ctrl_size():mf->size(); if ( ::icnt_has_buffer( m_shader_config->mem2device(i), response_size ) ) { - if (!mf->get_is_write()) + //if (!mf->get_is_write()) mf->set_return_timestamp(gpu_sim_cycle+gpu_tot_sim_cycle); mf->set_status(IN_ICNT_TO_SHADER,gpu_sim_cycle+gpu_tot_sim_cycle); ::icnt_push( m_shader_config->mem2device(i), mf->get_tpc(), mf, response_size ); diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 52c4643..043fcee 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -199,7 +199,7 @@ struct memory_config { assert(nbkgrp>0 && "Number of bank groups cannot be zero"); tRCDWR = tRCD-(WL+1); tRTW = (CL+(BL/data_command_freq_ratio)+2-WL); - tWTR = (WL+(BL/data_command_freq_ratio)+tCDLR); + tWTR = (WL+(BL/data_command_freq_ratio)+tCDLR); tWTP = (WL+(BL/data_command_freq_ratio)+tWR); dram_atom_size = BL * busW * gpu_n_mem_per_ctrlr; // burst length x bus width x # chips per partition diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index cac59f1..8fbf448 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -628,6 +628,7 @@ std::vector memory_sub_partition::breakdown_request_to_sector_reques void memory_sub_partition::push( mem_fetch* m_req, unsigned long long cycle ) { if (m_req) { + m_stats->memlatstat_icnt2mem_pop(m_req); std::vector reqs; if(m_config->m_L2_config.m_cache_type == SECTOR) reqs = breakdown_request_to_sector_requests(m_req); @@ -637,7 +638,6 @@ void memory_sub_partition::push( mem_fetch* m_req, unsigned long long cycle ) for(unsigned i=0; imemlatstat_icnt2mem_pop(req); if( req->istexture() ) { m_icnt_L2_queue->push(req); req->set_status(IN_PARTITION_ICNT_TO_L2_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); diff --git a/src/gpgpu-sim/mem_latency_stat.cc b/src/gpgpu-sim/mem_latency_stat.cc index fde0eff..35d6d84 100644 --- a/src/gpgpu-sim/mem_latency_stat.cc +++ b/src/gpgpu-sim/mem_latency_stat.cc @@ -75,6 +75,10 @@ memory_stats_t::memory_stats_t( unsigned n_shader, const struct shader_core_conf max_mf_latency = 0; max_icnt2mem_latency = 0; max_icnt2sh_latency = 0; + tot_icnt2mem_latency = 0; + tot_icnt2sh_latency = 0; + tot_mrq_num = 0; + tot_mrq_latency = 0; memset(mrq_lat_table, 0, sizeof(unsigned)*32); memset(dq_lat_table, 0, sizeof(unsigned)*32); memset(mf_lat_table, 0, sizeof(unsigned)*32); @@ -158,6 +162,7 @@ void memory_stats_t::memlatstat_read_done(mem_fetch *mf) mf_max_lat_table[mf->get_tlx_addr().chip][mf->get_tlx_addr().bk] = mf_latency; unsigned icnt2sh_latency; icnt2sh_latency = (gpu_tot_sim_cycle+gpu_sim_cycle) - mf->get_return_timestamp(); + tot_icnt2sh_latency += icnt2sh_latency; icnt2sh_lat_table[LOGB2(icnt2sh_latency)]++; if (icnt2sh_latency > max_icnt2sh_latency) max_icnt2sh_latency = icnt2sh_latency; @@ -191,6 +196,7 @@ void memory_stats_t::memlatstat_icnt2mem_pop(mem_fetch *mf) if (m_memory_config->gpgpu_memlatency_stat) { unsigned icnt2mem_latency; icnt2mem_latency = (gpu_tot_sim_cycle+gpu_sim_cycle) - mf->get_timestamp(); + tot_icnt2mem_latency += icnt2mem_latency; icnt2mem_lat_table[LOGB2(icnt2mem_latency)]++; if (icnt2mem_latency > max_icnt2mem_latency) max_icnt2mem_latency = icnt2mem_latency; @@ -216,14 +222,17 @@ void memory_stats_t::memlatstat_print( unsigned n_mem, unsigned gpu_mem_n_bk ) unsigned max_bank_accesses, min_bank_accesses, max_chip_accesses, min_chip_accesses; if (m_memory_config->gpgpu_memlatency_stat) { + printf("maxmflatency = %d \n", max_mf_latency); + printf("max_icnt2mem_latency = %d \n", max_icnt2mem_latency); printf("maxmrqlatency = %d \n", max_mrq_latency); - printf("maxdqlatency = %d \n", max_dq_latency); - printf("maxmflatency = %d \n", max_mf_latency); + //printf("maxdqlatency = %d \n", max_dq_latency); + printf("max_icnt2sh_latency = %d \n", max_icnt2sh_latency); if (num_mfs) { printf("averagemflatency = %lld \n", mf_total_lat/num_mfs); + printf("avg_icnt2mem_latency = %lld \n", tot_icnt2mem_latency/num_mfs); + printf("avg_mrq_latency = %lld \n", tot_mrq_latency/tot_mrq_num); + printf("avg_icnt2sh_latency = %lld \n", tot_icnt2sh_latency/num_mfs); } - printf("max_icnt2mem_latency = %d \n", max_icnt2mem_latency); - printf("max_icnt2sh_latency = %d \n", max_icnt2sh_latency); printf("mrq_lat_table:"); for (i=0; i< 32; i++) { printf("%d \t", mrq_lat_table[i]); diff --git a/src/gpgpu-sim/mem_latency_stat.h b/src/gpgpu-sim/mem_latency_stat.h index 4968a3b..5b89202 100644 --- a/src/gpgpu-sim/mem_latency_stat.h +++ b/src/gpgpu-sim/mem_latency_stat.h @@ -56,6 +56,10 @@ public: unsigned max_dq_latency; unsigned max_mf_latency; unsigned max_icnt2mem_latency; + unsigned long long int tot_icnt2mem_latency; + unsigned long long int tot_icnt2sh_latency; + unsigned long long int tot_mrq_latency; + unsigned long long int tot_mrq_num; unsigned max_icnt2sh_latency; unsigned mrq_lat_table[32]; unsigned dq_lat_table[32]; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index c14e19f..bf482fb 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1458,7 +1458,7 @@ ldst_unit::process_cache_access( cache_t* cache, if( !write_sent ) delete mf; } else if ( status == RESERVATION_FAIL ) { - result = COAL_STALL; + result = BK_CONF; assert( !read_sent ); assert( !write_sent ); delete mf; @@ -1467,8 +1467,8 @@ ldst_unit::process_cache_access( cache_t* cache, //inst.clear_active( access.get_warp_mask() ); // threads in mf writeback when mf returns inst.accessq_pop_back(); } - if( !inst.accessq_empty() ) - result = BK_CONF; + if( !inst.accessq_empty() && result == NO_RC_FAIL) + result = COAL_STALL; return result; } @@ -1563,7 +1563,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea assert( CACHE_UNDEFINED != inst.cache_op ); stall_cond = process_memory_access_queue(m_L1D,inst); } - if( !inst.accessq_empty() ) + if( !inst.accessq_empty() && stall_cond == NO_RC_FAIL) stall_cond = COAL_STALL; if (stall_cond != NO_RC_FAIL) { stall_reason = stall_cond; -- cgit v1.3 From 0b1a646c0ddf5d4db64a2b6fe9dfa30789cf1cd4 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 30 Oct 2017 20:23:51 -0400 Subject: adding new stats and change the PascalP100-HBM config --- configs/Pascal-P100-HBM/config_fermi_islip.icnt | 9 +++++--- configs/Pascal-P100-HBM/gpgpusim.config | 25 ++++++++++++---------- configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt | 3 +++ src/gpgpu-sim/dram.cc | 13 +++++++---- src/gpgpu-sim/dram.h | 4 ++++ src/gpgpu-sim/dram_sched.cc | 21 +++++++++++++----- src/gpgpu-sim/gpu-sim.cc | 4 +++- src/gpgpu-sim/gpu-sim.h | 8 +++++++ 8 files changed, 63 insertions(+), 24 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/Pascal-P100-HBM/config_fermi_islip.icnt b/configs/Pascal-P100-HBM/config_fermi_islip.icnt index a788090..0a73c81 100644 --- a/configs/Pascal-P100-HBM/config_fermi_islip.icnt +++ b/configs/Pascal-P100-HBM/config_fermi_islip.icnt @@ -1,13 +1,13 @@ //21*1 fly with 32 flits per packet under gpgpusim injection mode use_map = 0; -flit_size = 32; +flit_size = 40; // currently we do not use this, see subnets below network_count = 2; // Topology topology = fly; -k = 62; +k = 60; n = 1; // Routing @@ -17,7 +17,10 @@ routing_function = dest_tag; // Flow control num_vcs = 1; -vc_buf_size = 8; +vc_buf_size = 64; +input_buffer_size = 64; +ejection_buffer_size = 64; +boundary_buffer_size = 64; wait_for_tail_credit = 0; diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config index 5b038de..1029194 100644 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -54,13 +54,17 @@ # ** 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 N:64:128:6,L:L:m:N:H,S:128:8,8 +-gpgpu_cache:dl1 S:64:128:6,L:L:m:N:H,A:128:8,32:0,32 -gpgpu_shmem_size 65536 --gmem_skip_L1D 0 +-gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 -# 64 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:W:L,A:128:4,4:0,32 +# 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:W:L,A:256:4,32:0,32 -gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +#-gpgpu_flush_l2_cache 1 # 4 KB Inst. -gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 @@ -89,8 +93,7 @@ -gpgpu_shmem_num_banks 32 -gpgpu_shmem_limited_broadcast 0 -gpgpu_shmem_warp_parts 1 -# Use Fermi Coalscer arhitetecture for now! Need to be canged to pascal Coalscer --gpgpu_coalesce_arch 20 +-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 @@ -123,10 +126,6 @@ -gpgpu_mem_address_mask 1 -gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS -# GDDR5 timing -#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=6:RCD=12:RAS=28:RP=12:RC=40: -# CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - # 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 @@ -140,8 +139,12 @@ # 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 @@ -158,7 +161,7 @@ -visualizer_enabled 0 # power model configs, disable it untill we create a real energy model for Pascal 100 --power_simulation_enabled 1 +-power_simulation_enabled 0 -gpuwattch_xml_file gpuwattch_gtx480.xml # tracing functionality diff --git a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt index 58e596d..94b2378 100644 --- a/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt +++ b/configs/Pascal-P102-GDDR5X/config_fermi_islip.icnt @@ -18,6 +18,9 @@ routing_function = dest_tag; num_vcs = 1; vc_buf_size = 32; +input_buffer_size = 32; +ejection_buffer_size = 32; +boundary_buffer_size = 32; wait_for_tail_credit = 0; diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index de37f64..a57508c 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -52,6 +52,10 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m //rowblp access_num=0; hits_num=0; + read_num=0; + write_num=0; + hits_read_num=0; + hits_write_num=0; banks_1time=0; banks_acess_total=0; banks_acess_total_after=0; @@ -718,6 +722,8 @@ void dram_t::print( FILE* simFile) const fprintf(simFile, "\n------------------------------------------------------------------------\n"); printf("\nRow_Buffer_Locality = %.6f", (float)hits_num / access_num); + printf("\nRow_Buffer_Locality_read = %.6f", (float)hits_read_num / read_num); + printf("\nRow_Buffer_Locality_write = %.6f", (float)hits_write_num / write_num); printf("\nBank_Level_Parallism = %.6f", (float)banks_1time / banks_acess_total); printf("\nBank_Level_Parallism_Col = %.6f", (float)banks_time_rw / banks_access_rw_total); printf("\nBank_Level_Parallism_Ready = %.6f", (float)banks_time_ready /banks_access_ready_total); @@ -735,7 +741,7 @@ void dram_t::print( FILE* simFile) const printf("RCDWRc_limit = %d \n", RCDWRc_limit); printf("WTRc_limit = %d \n", WTRc_limit); printf("RTWc_limit = %d \n", RTWc_limit); - printf("CCDLc_limit %d \n", CCDLc_limit); + printf("CCDLc_limit = %d \n", CCDLc_limit); printf("rwq = %d \n", rwq_limit); printf("CCDLc_limit_alone = %d \n", CCDLc_limit_alone); printf("WTRc_limit_alone = %d \n", WTRc_limit_alone); @@ -751,15 +757,14 @@ void dram_t::print( FILE* simFile) const printf("n_pre = %d \n", n_pre); printf("n_ref = %d \n", n_ref); printf("n_req = %d \n", n_req ); - printf("n_req4 = %d \n", n_req*4 ); printf("total_req = %d \n\n", n_rd+n_wr+n_rd_L2_A+n_wr_WB); printf("issued_total_row = %lu \n", issued_total_row); printf("issued_total_col = %lu \n", issued_total_col); printf("Row_Bus_Util = %.6f \n", (float)issued_total_row / n_cmd); printf("CoL_Bus_Util = %.6f \n", (float)issued_total_col / n_cmd); - printf("Either_Row_CoL_Bus_Util %.6f \n", (float)issued_total / n_cmd); - printf("Issued_on_Two_Bus_Simul_Util %.6f \n", (float)issued_two /n_cmd); + printf("Either_Row_CoL_Bus_Util = %.6f \n", (float)issued_total / n_cmd); + printf("Issued_on_Two_Bus_Simul_Util = %.6f \n", (float)issued_two /n_cmd); printf("issued_two_Eff = %.6f \n", (float)issued_two /issued_total); printf("queue_avg = %.6f \n\n", (float)ave_mrqs/n_cmd ); diff --git a/src/gpgpu-sim/dram.h b/src/gpgpu-sim/dram.h index 29731a7..0d4c0e7 100644 --- a/src/gpgpu-sim/dram.h +++ b/src/gpgpu-sim/dram.h @@ -198,7 +198,11 @@ private: //row locality, BLP and other statistics unsigned long access_num; + unsigned long read_num; + unsigned long write_num; unsigned long long hits_num; + unsigned long long hits_read_num; + unsigned long long hits_write_num; unsigned long long banks_1time; unsigned long long banks_acess_total; unsigned long long banks_acess_total_after; diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc index ac4c827..f754d36 100644 --- a/src/gpgpu-sim/dram_sched.cc +++ b/src/gpgpu-sim/dram_sched.cc @@ -142,14 +142,25 @@ dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row ) rowhit = true; } } - //rowblp - m_dram->access_num++; - if(rowhit) - m_dram->hits_num++; - std::list::iterator next = m_current_last_row[bank]->back(); dram_req_t *req = (*next); + //rowblp stats + m_dram->access_num++; + bool is_write = req->data->is_write(); + if(is_write) + m_dram->write_num++; + else + m_dram->read_num++; + + if(rowhit) { + m_dram->hits_num++; + if(is_write) + m_dram->hits_write_num++; + else + m_dram->hits_read_num++; + } + m_stats->concurrent_row_access[m_dram->id][bank]++; m_stats->row_access[m_dram->id][bank]++; m_current_last_row[bank]->pop_back(); diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 0e06c5c..11ac5df 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -207,6 +207,9 @@ void memory_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-Write_Queue_Size", OPT_CSTR, &write_queue_size_opt, "Write_Queue_Size", "32:28:16"); + option_parser_register(opp, "-Elimnate_rw_turnaround", OPT_BOOL, &elimnate_rw_turnaround, + "elimnate_rw_turnaround i.e set tWTR and tRTW = 0", + "0"); option_parser_register(opp, "-icnt_flit_size", OPT_UINT32, &icnt_flit_size, "icnt_flit_size", "32"); @@ -435,7 +438,6 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) option_parser_register(opp, "-gpgpu_flush_l2_cache", OPT_BOOL, &gpgpu_flush_l2_cache, "Flush L2 cache at the end of each kernel call", "0"); - option_parser_register(opp, "-gpgpu_deadlock_detect", OPT_BOOL, &gpu_deadlock_detect, "Stop the simulation at deadlock (1=on (default), 0=off)", "1"); diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 043fcee..c04648c 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -198,8 +198,14 @@ struct memory_config { bk_tag_length = i-1; assert(nbkgrp>0 && "Number of bank groups cannot be zero"); tRCDWR = tRCD-(WL+1); + if(elimnate_rw_turnaround) + { + tRTW = 0; + tWTR = 0; + } else { tRTW = (CL+(BL/data_command_freq_ratio)+2-WL); tWTR = (WL+(BL/data_command_freq_ratio)+tCDLR); + } tWTP = (WL+(BL/data_command_freq_ratio)+tWR); dram_atom_size = BL * busW * gpu_n_mem_per_ctrlr; // burst length x bus width x # chips per partition @@ -266,6 +272,8 @@ struct memory_config { unsigned nbk; + bool elimnate_rw_turnaround; + unsigned data_command_freq_ratio; // frequency ratio between DRAM data bus and command bus (2 for GDDR3, 4 for GDDR5) unsigned dram_atom_size; // number of bytes transferred per read or write command -- cgit v1.3 From 9233f6f9eeea537187deb64add77a320442aa621 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 18 Nov 2017 15:48:26 -0500 Subject: vectoradd is successfully filling the l2 --- src/cuda-sim/cuda-sim.cc | 4 ++++ src/gpgpu-sim/gpu-cache.cc | 23 +++++++++++++++++------ src/gpgpu-sim/gpu-cache.h | 11 +++++++++++ src/gpgpu-sim/gpu-sim.cc | 17 +++++++++++++++++ src/gpgpu-sim/gpu-sim.h | 2 ++ src/gpgpu-sim/l2cache.cc | 9 +++++++++ src/gpgpu-sim/l2cache.h | 6 ++++++ 7 files changed, 66 insertions(+), 6 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index a668db1..685ae53 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -401,6 +401,10 @@ void gpgpu_t::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t coun char *src_data = (char*)src; for (unsigned n=0; n < count; n ++ ) m_global_mem->write(dst_start_addr+n,1, src_data+n,NULL,NULL); + + // Copy into the performance model. + extern gpgpu_sim* g_the_gpu; + g_the_gpu->memcpy_to_gpu(dst_start_addr, src, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 32c2bb1..37fc5ea 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -190,11 +190,17 @@ void tag_array::init( int core_id, int type_id ) m_type_id = type_id; } + enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx, mem_fetch* mf) const { + mem_access_sector_mask_t mask = mf->get_access_sector_mask(); + return probe(addr, idx, mask); +} + + +enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx, mem_access_sector_mask_t mask) const { //assert( m_config.m_write_policy == READ_ONLY ); unsigned set_index = m_config.set_index(addr); new_addr_type tag = m_config.tag(addr); - mem_access_sector_mask_t mask = mf->get_access_sector_mask(); unsigned invalid_line = (unsigned)-1; unsigned valid_line = (unsigned)-1; @@ -310,18 +316,23 @@ enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, void tag_array::fill( new_addr_type addr, unsigned time, mem_fetch* mf) { - assert( m_config.m_alloc_policy == ON_FILL ); + fill(addr, time, mf->get_access_sector_mask()); +} + +void tag_array::fill( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ) +{ + //assert( m_config.m_alloc_policy == ON_FILL ); unsigned idx; - enum cache_request_status status = probe(addr,idx,mf); + enum cache_request_status status = probe(addr,idx,mask); assert(status==MISS||status==SECTOR_MISS); // MSHR should have prevented redundant memory request if(status==MISS) - m_lines[idx]->allocate( m_config.tag(addr), m_config.block_addr(addr), time, mf->get_access_sector_mask() ); + m_lines[idx]->allocate( m_config.tag(addr), m_config.block_addr(addr), time, mask ); else if (status==SECTOR_MISS) { assert(m_config.m_cache_type == SECTOR); - ((sector_cache_block*)m_lines[idx])->allocate_sector( time, mf->get_access_sector_mask() ); + ((sector_cache_block*)m_lines[idx])->allocate_sector( time, mask ); } - m_lines[idx]->fill(time, mf->get_access_sector_mask()); + m_lines[idx]->fill(time, mask); } void tag_array::fill( unsigned index, unsigned time, mem_fetch* mf) diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 3e1691a..3713126 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -690,11 +690,13 @@ public: ~tag_array(); enum cache_request_status probe( new_addr_type addr, unsigned &idx, mem_fetch* mf ) const; + enum cache_request_status probe( new_addr_type addr, unsigned &idx, mem_access_sector_mask_t mask ) const; enum cache_request_status access( new_addr_type addr, unsigned time, unsigned &idx, mem_fetch* mf ); enum cache_request_status access( new_addr_type addr, unsigned time, unsigned &idx, bool &wb, evicted_block_info &evicted, mem_fetch* mf ); void fill( new_addr_type addr, unsigned time, mem_fetch* mf ); void fill( unsigned idx, unsigned time, mem_fetch* mf ); + void fill( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ); unsigned size() const { return m_config.get_num_lines();} cache_block_t* get_block(unsigned idx) { return m_lines[idx];} @@ -969,6 +971,15 @@ public: bool data_port_free() const { return m_bandwidth_management.data_port_free(); } bool fill_port_free() const { return m_bandwidth_management.fill_port_free(); } + // This is a gapping hole we are poking in the system to quickly handle + // filling the cache on cudamemcopies. We don't care about anything other than + // L2 state after the memcopy - so just force the tag array to act as though + // something is read or written without doing anything else. + void force_tag_access( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ) + { + m_tag_array->fill( addr, time, mask ); + } + protected: // Constructor that can be used by derived classes with custom tag arrays baseline_cache( const char *name, diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 11ac5df..263cbad 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1595,6 +1595,23 @@ void shader_core_ctx::dump_warp_state( FILE *fout ) const m_warp[w].print(fout); } + +void gpgpu_sim::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t count ) +{ + assert (dst_start_addr % 32 == 0); + // Right now - I am just going to assume you write the whole last cache line... +// assert (count % 128 == 0); + for ( unsigned counter = 0; counter < count; counter += 32 ) { + const size_t wr_addr = dst_start_addr + counter; + addrdec_t raw_addr; + mem_access_sector_mask_t mask; + mask.set(wr_addr % 128 / 32); + m_memory_config->m_address_mapping.addrdec_tlx( wr_addr, &raw_addr ); + const unsigned partition_id = raw_addr.sub_partition / m_memory_config->m_n_sub_partition_per_memory_channel; + m_memory_partition_unit[ partition_id ]->handle_memcpy_to_gpu( wr_addr, raw_addr.sub_partition, mask ); + } +} + void gpgpu_sim::dump_pipeline( int mask, int s, int m ) const { /* diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index c04648c..6382adf 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -425,6 +425,8 @@ public: void gpu_print_stat(); void dump_pipeline( int mask, int s, int m ) const; + void memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t count ); + //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 8fbf448..1df7858 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -74,6 +74,15 @@ memory_partition_unit::memory_partition_unit( unsigned partition_id, } } +void memory_partition_unit::handle_memcpy_to_gpu( size_t addr, unsigned global_subpart_id, mem_access_sector_mask_t mask ) +{ + unsigned p = global_sub_partition_id_to_local_id(global_subpart_id); + std::string mystring = + mask.to_string(); + MEMPART_DPRINTF("Copy Engine Request Received For Address=%zu, local_subpart=%u, sector_mask=%s \n", addr, p, mystring.c_str()); + m_sub_partition[p]->force_l2_tag_update(addr,gpu_sim_cycle+gpu_tot_sim_cycle, mask); +} + memory_partition_unit::~memory_partition_unit() { delete m_dram; diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 2cc0e76..b613a94 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -72,6 +72,7 @@ public: void print_stat( FILE *fp ) { m_dram->print_stat(fp); } void visualize() const { m_dram->visualize(); } void print( FILE *fp ) const; + void handle_memcpy_to_gpu( size_t dst_start_addr, unsigned subpart_id, mem_access_sector_mask_t mask ); class memory_sub_partition * get_sub_partition(int sub_partition_id) { @@ -178,6 +179,11 @@ public: void accumulate_L2cache_stats(class cache_stats &l2_stats) const; void get_L2cache_sub_stats(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, time, mask ); + } + private: // data unsigned m_id; //< the global sub partition ID -- cgit v1.3 From d8766e4eb7551afcc8c9ca168449bcd20974af60 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 18 Nov 2017 16:24:58 -0500 Subject: Making the perf sim copy optional, getting rid of an assert that will happen with the new hack and incrementing the cycle so that cudamemcopies take some time (if we don't do this the LRU in the cache does not work) --- src/gpgpu-sim/gpu-cache.cc | 2 +- src/gpgpu-sim/gpu-sim.cc | 27 ++++++++++++++++----------- src/gpgpu-sim/gpu-sim.h | 1 + 3 files changed, 18 insertions(+), 12 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 37fc5ea..e1e41a5 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -324,7 +324,7 @@ void tag_array::fill( new_addr_type addr, unsigned time, mem_access_sector_mask_ //assert( m_config.m_alloc_policy == ON_FILL ); unsigned idx; enum cache_request_status status = probe(addr,idx,mask); - assert(status==MISS||status==SECTOR_MISS); // MSHR should have prevented redundant memory request + //assert(status==MISS||status==SECTOR_MISS); // MSHR should have prevented redundant memory request if(status==MISS) m_lines[idx]->allocate( m_config.tag(addr), m_config.block_addr(addr), time, mask ); else if (status==SECTOR_MISS) { diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 263cbad..438769f 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -140,6 +140,8 @@ void power_config::reg_options(class OptionParser * opp) 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", "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, @@ -1598,17 +1600,20 @@ void shader_core_ctx::dump_warp_state( FILE *fout ) const void gpgpu_sim::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t count ) { - assert (dst_start_addr % 32 == 0); - // Right now - I am just going to assume you write the whole last cache line... -// assert (count % 128 == 0); - for ( unsigned counter = 0; counter < count; counter += 32 ) { - const size_t wr_addr = dst_start_addr + counter; - addrdec_t raw_addr; - mem_access_sector_mask_t mask; - mask.set(wr_addr % 128 / 32); - m_memory_config->m_address_mapping.addrdec_tlx( wr_addr, &raw_addr ); - const unsigned partition_id = raw_addr.sub_partition / m_memory_config->m_n_sub_partition_per_memory_channel; - m_memory_partition_unit[ partition_id ]->handle_memcpy_to_gpu( wr_addr, raw_addr.sub_partition, mask ); + if (m_memory_config->m_perf_sim_memcpy) { + assert (dst_start_addr % 32 == 0); + // Right now - I am just going to assume you write the whole last cache line... + // assert (count % 128 == 0); + for ( unsigned counter = 0; counter < count; counter += 32 ) { + const size_t wr_addr = dst_start_addr + counter; + addrdec_t raw_addr; + mem_access_sector_mask_t mask; + mask.set(wr_addr % 128 / 32); + m_memory_config->m_address_mapping.addrdec_tlx( wr_addr, &raw_addr ); + const unsigned partition_id = raw_addr.sub_partition / m_memory_config->m_n_sub_partition_per_memory_channel; + m_memory_partition_unit[ partition_id ]->handle_memcpy_to_gpu( wr_addr, raw_addr.sub_partition, mask ); + gpu_sim_cycle += 1; + } } } diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 6382adf..81f13cb 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -290,6 +290,7 @@ struct memory_config { unsigned gpgpu_frfcfs_dram_write_queue_size; unsigned write_high_watermark; unsigned write_low_watermark; + bool m_perf_sim_memcpy; }; // global counters and flags (please try not to add to this list!!!) -- cgit v1.3 From b9420953855042566ea3371882aef89f66a3f055 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 18 Nov 2017 17:38:24 -0500 Subject: fixing a stupid inheritance bug --- src/cuda-sim/cuda-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.h | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 685ae53..5516b1e 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -404,7 +404,7 @@ void gpgpu_t::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t coun // Copy into the performance model. extern gpgpu_sim* g_the_gpu; - g_the_gpu->memcpy_to_gpu(dst_start_addr, src, count); + g_the_gpu->perf_memcpy_to_gpu(dst_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 438769f..0c2ea1c 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1598,7 +1598,7 @@ void shader_core_ctx::dump_warp_state( FILE *fout ) const } -void gpgpu_sim::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t count ) +void gpgpu_sim::perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ) { if (m_memory_config->m_perf_sim_memcpy) { assert (dst_start_addr % 32 == 0); diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 81f13cb..1778008 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -426,7 +426,7 @@ public: void gpu_print_stat(); void dump_pipeline( int mask, int s, int m ) const; - void memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t count ); + void perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ); //The next three functions added to be used by the functional simulation function -- cgit v1.3 From 05182e14ea348b9d57fc14aeae1ff3f684512262 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 18 Nov 2017 19:30:53 -0500 Subject: 64-address used on the CPU side, but GPGPU-Sim Truncates it to 32-bit.... truncating the address here fixes the issue and we start propoerly hitting in the L2 --- src/gpgpu-sim/gpu-sim.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (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 0c2ea1c..0267c31 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1605,7 +1605,7 @@ void gpgpu_sim::perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ) // Right now - I am just going to assume you write the whole last cache line... // assert (count % 128 == 0); for ( unsigned counter = 0; counter < count; counter += 32 ) { - const size_t wr_addr = dst_start_addr + counter; + const unsigned wr_addr = dst_start_addr + counter; addrdec_t raw_addr; mem_access_sector_mask_t mask; mask.set(wr_addr % 128 / 32); -- cgit v1.3 From 21528f301886ffdba5e921091658446d23c9c377 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 18 Nov 2017 21:00:17 -0500 Subject: fixing the cycle issues with using the cudamemcpies --- src/gpgpu-sim/gpu-sim.cc | 4 +--- src/gpgpu-sim/l2cache.cc | 5 +++-- src/gpgpu-sim/l2cache.h | 10 +++++++++- 3 files changed, 13 insertions(+), 6 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 0267c31..8dc80d2 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1602,8 +1602,7 @@ void gpgpu_sim::perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ) { if (m_memory_config->m_perf_sim_memcpy) { assert (dst_start_addr % 32 == 0); - // Right now - I am just going to assume you write the whole last cache line... - // assert (count % 128 == 0); + for ( unsigned counter = 0; counter < count; counter += 32 ) { const unsigned wr_addr = dst_start_addr + counter; addrdec_t raw_addr; @@ -1612,7 +1611,6 @@ void gpgpu_sim::perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ) m_memory_config->m_address_mapping.addrdec_tlx( wr_addr, &raw_addr ); const unsigned partition_id = raw_addr.sub_partition / m_memory_config->m_n_sub_partition_per_memory_channel; m_memory_partition_unit[ partition_id ]->handle_memcpy_to_gpu( wr_addr, raw_addr.sub_partition, mask ); - gpu_sim_cycle += 1; } } } diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index fb0d588..b1465a8 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -315,6 +315,7 @@ memory_sub_partition::memory_sub_partition( unsigned sub_partition_id, m_id = sub_partition_id; m_config=config; m_stats=stats; + m_memcpy_cycle_offset = 0; assert(m_id < m_config->m_n_mem_sub_partition); @@ -378,7 +379,7 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) if ( !m_config->m_L2_config.disabled() && m_L2cache->waiting_for_fill(mf) ) { if (m_L2cache->fill_port_free()) { mf->set_status(IN_PARTITION_L2_FILL_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); - m_L2cache->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle); + m_L2cache->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle+m_memcpy_cycle_offset); m_dram_L2_queue->pop(); } } else if ( !m_L2_icnt_queue->full() ) { @@ -404,7 +405,7 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) bool port_free = m_L2cache->data_port_free(); if ( !output_full && port_free ) { std::list events; - enum cache_request_status status = m_L2cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle,events); + enum cache_request_status status = m_L2cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle+m_memcpy_cycle_offset,events); bool write_sent = was_write_sent(events); bool read_sent = was_read_sent(events); MEM_SUBPART_DPRINTF("Probing L2 cache Address=%llx, status=%u\n", mf->get_addr(), status); diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index b613a94..2d13918 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -181,7 +181,8 @@ public: void force_l2_tag_update(new_addr_type addr, unsigned time, mem_access_sector_mask_t mask) { - m_L2cache->force_tag_access( addr, time, mask ); + m_L2cache->force_tag_access( addr, m_memcpy_cycle_offset + time, mask ); + m_memcpy_cycle_offset += 1; } private: @@ -216,6 +217,13 @@ private: friend class L2interface; std::vector breakdown_request_to_sector_requests(mem_fetch* mf); + + // This is a cycle offset that has to be applied to the l2 accesses to account for + // the cudamemcpy read/writes. We want GPGPU-Sim to only count cycles for kernel execution + // but we want cudamemcpy to go through the L2. Everytime an access is made from cudamemcpy + // this counter is incremented, and when the l2 is accessed (in both cudamemcpyies and otherwise) + // this value is added to the gpgpu-sim cycle counters. + unsigned m_memcpy_cycle_offset; }; class L2interface : public mem_fetch_interface { -- cgit v1.3 From db4011bce0d88b44584eccf3854f3d82876cdc77 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 18 Nov 2017 21:03:48 -0500 Subject: copy to the L2 by default --- src/gpgpu-sim/gpu-sim.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (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 8dc80d2..17f1714 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -141,7 +141,7 @@ void power_config::reg_options(class OptionParser * opp) 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", "0"); + "Fill the L2 cache on memcpy", "1"); 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, -- cgit v1.3 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/cuda-sim/cuda-sim.cc | 44 ++++++++++++++++++++------------------------ src/cuda-sim/cuda-sim.h | 2 ++ src/cuda-sim/opcodes.def | 2 ++ src/gpgpu-sim/gpu-sim.cc | 7 ++++--- 4 files changed, 28 insertions(+), 27 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 2f166aa..17a7798 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -52,8 +52,10 @@ int gpgpu_ptx_instruction_classification; void ** g_inst_classification_stat = NULL; +void ** g_inst_mem_classification_stat = NULL; void ** g_inst_op_classification_stat= NULL; int g_ptx_kernel_count = -1; // used for classification stat collection purposes +int g_ptx_kernel_count_prev = -1; // used for classification stat collection purposes int g_debug_execution = 0; int g_debug_thread_uid = 0; addr_t g_debug_pc = 0xBEEF1518; @@ -1240,12 +1242,16 @@ void init_inst_classification_stat() #define MAX_CLASS_KER 1024 char kernelname[MAX_CLASS_KER] =""; if (!g_inst_classification_stat) g_inst_classification_stat = (void**)calloc(MAX_CLASS_KER, sizeof(void*)); - snprintf(kernelname, MAX_CLASS_KER, "Kernel %d Classification\n",g_ptx_kernel_count ); + snprintf(kernelname, MAX_CLASS_KER, "Kernel %d INST Classification",g_ptx_kernel_count ); assert( g_ptx_kernel_count < MAX_CLASS_KER ) ; // a static limit on number of kernels increase it if it fails! g_inst_classification_stat[g_ptx_kernel_count] = StatCreate(kernelname,1,20); + if (!g_inst_mem_classification_stat) g_inst_mem_classification_stat = (void**)calloc(MAX_CLASS_KER, sizeof(void*)); + snprintf(kernelname, MAX_CLASS_KER, "Kernel %d MEM Classification",g_ptx_kernel_count ); + g_inst_mem_classification_stat[g_ptx_kernel_count] = StatCreate(kernelname,1,20); if (!g_inst_op_classification_stat) g_inst_op_classification_stat = (void**)calloc(MAX_CLASS_KER, sizeof(void*)); - snprintf(kernelname, MAX_CLASS_KER, "Kernel %d OP Classification\n",g_ptx_kernel_count ); + snprintf(kernelname, MAX_CLASS_KER, "Kernel %d OP Classification",g_ptx_kernel_count ); g_inst_op_classification_stat[g_ptx_kernel_count] = StatCreate(kernelname,1,100); + g_ptx_kernel_count_prev++; } static unsigned get_tex_datasize( const ptx_instruction *pI, ptx_thread_info *thread ) @@ -1323,6 +1329,15 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) delete pJ; pI = pI_saved; + if ( gpgpu_ptx_instruction_classification ) { + init_inst_classification_stat(); + if (op_classification) + StatAddSample( g_inst_classification_stat[g_ptx_kernel_count], op_classification); + if (pI->get_space().get_type()) + StatAddSample( g_inst_mem_classification_stat[g_ptx_kernel_count], ( int )pI->get_space().get_type()); + StatAddSample( g_inst_op_classification_stat[g_ptx_kernel_count], (int) pI->get_opcode() ); + } + // Run exit instruction if exit option included if(pI->is_exit()) exit_impl(pI,this); @@ -1408,27 +1423,6 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) if(!(this->m_functionalSimulationMode)) ptx_file_line_stats_add_exec_count(pI); - if ( gpgpu_ptx_instruction_classification ) { - init_inst_classification_stat(); - unsigned space_type=0; - switch ( pI->get_space().get_type() ) { - case global_space: space_type = 10; break; - case local_space: space_type = 11; break; - case tex_space: space_type = 12; break; - case surf_space: space_type = 13; break; - case param_space_kernel: - case param_space_local: - space_type = 14; break; - case shared_space: space_type = 15; break; - case const_space: space_type = 16; break; - default: - space_type = 0 ; - break; - } - StatAddSample( g_inst_classification_stat[g_ptx_kernel_count], op_classification); - if (space_type) StatAddSample( g_inst_classification_stat[g_ptx_kernel_count], ( int )space_type); - StatAddSample( g_inst_op_classification_stat[g_ptx_kernel_count], (int) pI->get_opcode() ); - } if ( (g_ptx_sim_num_insn % 100000) == 0 ) { dim3 ctaid = get_ctaid(); dim3 tid = get_tid(); @@ -1848,8 +1842,10 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) //******PRINTING******* printf( "GPGPU-Sim: Done functional simulation (%u instructions simulated).\n", g_ptx_sim_num_insn ); + fflush(stdout); if ( gpgpu_ptx_instruction_classification ) { - StatDisp( g_inst_classification_stat[g_ptx_kernel_count]); + StatDisp ( g_inst_classification_stat[g_ptx_kernel_count]); + StatDisp ( g_inst_mem_classification_stat[g_ptx_kernel_count]); StatDisp ( g_inst_op_classification_stat[g_ptx_kernel_count]); } diff --git a/src/cuda-sim/cuda-sim.h b/src/cuda-sim/cuda-sim.h index 958daba..ef9549f 100644 --- a/src/cuda-sim/cuda-sim.h +++ b/src/cuda-sim/cuda-sim.h @@ -44,8 +44,10 @@ extern int g_ptx_sim_mode; extern int g_debug_execution; extern int g_debug_thread_uid; extern void ** g_inst_classification_stat; +extern void ** g_inst_mem_classification_stat; extern void ** g_inst_op_classification_stat; extern int g_ptx_kernel_count; // used for classification stat collection purposes +extern int g_ptx_kernel_count_prev; // used for classification stat collection purposes void ptx_opcocde_latency_options (option_parser_t opp); extern class kernel_info_t *gpgpu_opencl_ptx_sim_init_grid(class function_info *entry, diff --git a/src/cuda-sim/opcodes.def b/src/cuda-sim/opcodes.def index e1b1422..ccf64d8 100644 --- a/src/cuda-sim/opcodes.def +++ b/src/cuda-sim/opcodes.def @@ -35,7 +35,9 @@ SFU 4 Mem(except Tex) 5 Tex 6 Nop 7 +Breakpoint 9 Other 10 +Scalar video 11 */ OP_DEF(ABS_OP,abs_impl,"abs",1,1) OP_DEF(ADD_OP,add_impl,"add",1,1) 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 From 4e91a60a48b07f41f4bfb4d59fa2355024a3914b Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sun, 25 Mar 2018 16:51:21 -0400 Subject: Need to make sure we don't kill the L1 too --- configs/Pascal-P100-HBM/gpgpusim.config | 2 ++ configs/Pascal-P102-GDDR5X/gpgpusim.config | 2 ++ configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config | 2 ++ configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 2 ++ src/gpgpu-sim/gpu-sim.cc | 2 +- 5 files changed, 9 insertions(+), 1 deletion(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config index 31fbecb..857a0ff 100644 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -55,6 +55,8 @@ # 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 diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 8e1e56f..feb8c0b 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -62,6 +62,8 @@ # 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 -gpgpu_shmem_size 49152 -gpgpu_shmem_size_PrefL1 49152 -gpgpu_shmem_size_PrefShared 49152 diff --git a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config index b87d958..a5e6736 100644 --- a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config @@ -55,6 +55,8 @@ # Pascal GP100 has 64KB Shared memory # Pascal GP100 has 48KB L1 cache -gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 65536 -gpgpu_shmem_size_PrefL1 65536 -gpgpu_shmem_size_PrefShared 65536 diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 2df24e1..f78bd02 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -56,6 +56,8 @@ # Pascal GP102 has 64KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used -gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 98304 -gpgpu_shmem_size_PrefL1 98304 -gpgpu_shmem_size_PrefShared 98304 diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 17f1714..bb448b3 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -245,7 +245,7 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); - option_parser_register(opp, "-gpgpu_cache:dl1PreShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, + option_parser_register(opp, "-gpgpu_cache:dl1PrefShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); -- cgit v1.3 From 1af82fbe97428654b06b7cd8d40c2d2ce4592aae Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 26 Mar 2018 11:35:25 -0400 Subject: Fixing resources limitation assertion for compute cab 61 --- configs/Pascal-P102-GDDR5X/gpgpusim.config | 1 + src/gpgpu-sim/gpu-sim.cc | 3 +++ src/gpgpu-sim/shader.cc | 4 ++++ src/gpgpu-sim/shader.h | 1 + 4 files changed, 9 insertions(+) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 0c6c126..b7bcdaf 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -6,6 +6,7 @@ -gpgpu_ptx_instruction_classification 0 -gpgpu_ptx_sim_mode 0 -gpgpu_ptx_force_max_capability 61 +-gpgpu_ignore_resources_limitation 1 # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 17f1714..48af41e 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -268,6 +268,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shader_registers", OPT_UINT32, &gpgpu_shader_registers, "Number of registers per shader core. Limits number of concurrent CTAs. (default 8192)", "8192"); + option_parser_register(opp, "-gpgpu_ignore_resources_limitation", OPT_BOOL, &gpgpu_ignore_resources_limitation, + "gpgpu_ignore_resources_limitation (default 0)", + "0"); option_parser_register(opp, "-gpgpu_shader_cta", OPT_UINT32, &max_cta_per_core, "Maximum number of concurrent CTAs in shader (default 8)", "8"); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index bf482fb..8173191 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -2614,6 +2614,10 @@ unsigned int shader_core_config::max_cta( const kernel_info_t &k ) const assert( result <= MAX_CTA_PER_SHADER ); if (result < 1) { printf ("GPGPU-Sim uArch: ERROR ** Kernel requires more resources than shader has.\n"); + if(gpgpu_ignore_resources_limitation) { + printf ("GPGPU-Sim uArch: gpgpu_ignore_resources_limitation is set, ignore the ERROR!\n"); + return 1; + } abort(); } diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 5b41c06..ae22eaa 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1386,6 +1386,7 @@ struct shader_core_config : public core_config unsigned gpgpu_num_reg_banks; bool gpgpu_reg_bank_use_warp_id; bool gpgpu_local_mem_map; + bool gpgpu_ignore_resources_limitation; unsigned max_sp_latency; unsigned max_sfu_latency; -- cgit v1.3 From 6b7d14e294d8cb73e2281b7e18a4083e3271a85e Mon Sep 17 00:00:00 2001 From: Nathan Conrad Date: Mon, 26 Mar 2018 19:20:24 -0400 Subject: Provide portable (non-x86) breakpoint method which should work on all linuxes. Tested on PowerPC. --- src/gpgpu-sim/gpu-sim.cc | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (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 58a5d16..3829861 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -32,6 +32,7 @@ #include #include #include +#include #include "zlib.h" @@ -1440,7 +1441,7 @@ void gpgpu_sim::cycle() if( g_single_step && ((gpu_sim_cycle+gpu_tot_sim_cycle) >= g_single_step) ) { - asm("int $03"); + raise(SIGTRAP); // Debug breakpoint } gpu_sim_cycle++; if( g_interactive_debugger_enabled ) -- cgit v1.3 From 67a257442f46a9e8e02f63a9d71fcc7ca54f3f5e Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 28 Mar 2018 14:00:16 -0400 Subject: adding new counters for parition level para and L2 BW --- src/gpgpu-sim/gpu-sim.cc | 36 ++++++++++++++++++++++++++++++++++++ 1 file changed, 36 insertions(+) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 48af41e..228d4be 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -87,6 +87,12 @@ unsigned long long gpu_tot_sim_cycle = 0; // performance counter for stalls due to congestion. unsigned int gpu_stall_dramfull = 0; unsigned int gpu_stall_icnt2sh = 0; +unsigned long long partiton_reqs_in_parallel = 0; +unsigned long long partiton_reqs_in_parallel_total = 0; +unsigned long long partiton_reqs_in_parallel_util = 0; +unsigned long long partiton_reqs_in_parallel_util_total = 0; +unsigned long long partiton_replys_in_parallel = 0; +unsigned long long partiton_replys_in_parallel_total = 0; /* Clock Domains */ @@ -786,6 +792,9 @@ void gpgpu_sim::init() gpu_sim_insn = 0; last_gpu_sim_insn = 0; m_total_cta_launched=0; + partiton_reqs_in_parallel = 0; + partiton_replys_in_parallel = 0; + partiton_reqs_in_parallel_util = 0; reinit_clock_domains(); set_param_gpgpu_num_shaders(m_config.num_shader()); @@ -822,8 +831,14 @@ void gpgpu_sim::update_stats() { gpu_tot_sim_cycle += gpu_sim_cycle; gpu_tot_sim_insn += gpu_sim_insn; gpu_tot_issued_cta += m_total_cta_launched; + partiton_reqs_in_parallel_total += partiton_reqs_in_parallel; + partiton_replys_in_parallel_total += partiton_replys_in_parallel; + partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; gpu_sim_cycle = 0; + partiton_reqs_in_parallel = 0; + partiton_replys_in_parallel = 0; + partiton_reqs_in_parallel_util = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } @@ -1007,6 +1022,19 @@ void gpgpu_sim::gpu_print_stat() printf("gpu_stall_dramfull = %d\n", gpu_stall_dramfull); printf("gpu_stall_icnt2sh = %d\n", gpu_stall_icnt2sh ); + printf("partiton_reqs_in_parallel = %lld\n", partiton_reqs_in_parallel); + printf("partiton_reqs_in_parallel_total = %lld\n", partiton_reqs_in_parallel_total ); + printf("partiton_level_parallism = %12.4f\n", (float)partiton_reqs_in_parallel / gpu_sim_cycle); + printf("partiton_level_parallism_total = %12.4f\n", (float)(partiton_reqs_in_parallel+partiton_reqs_in_parallel_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("partiton_reqs_in_parallel_util = %lld\n", partiton_reqs_in_parallel_util); + printf("partiton_reqs_in_parallel_util_total = %lld\n", partiton_reqs_in_parallel_util_total ); + printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle); + printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("partiton_replys_in_parallel = %lld\n", partiton_replys_in_parallel); + printf("partiton_replys_in_parallel_total = %lld\n", partiton_replys_in_parallel_total ); + printf("L2_BW = %12.4f GB/Sec\n", ((float)(partiton_replys_in_parallel * 32) / (gpu_sim_cycle * m_config.icnt_period)) / 1000000000); + printf("L2_BW_total = %12.4f GB/Sec\n", ((float)((partiton_replys_in_parallel+partiton_replys_in_parallel_total) * 32) / ((gpu_tot_sim_cycle+gpu_sim_cycle) * m_config.icnt_period)) / 1000000000 ); + time_t curr_time; time(&curr_time); unsigned long long elapsed_time = MAX( curr_time - g_simulation_starttime, 1 ); @@ -1412,6 +1440,7 @@ void gpgpu_sim::cycle() for (unsigned i=0;in_simt_clusters;i++) m_cluster[i]->icnt_cycle(); } + unsigned partiton_replys_in_parallel_per_cycle = 0; if (clock_mask & ICNT) { // pop from memory controller to interconnect for (unsigned i=0;im_n_mem_sub_partition;i++) { @@ -1424,6 +1453,7 @@ void gpgpu_sim::cycle() mf->set_status(IN_ICNT_TO_SHADER,gpu_sim_cycle+gpu_tot_sim_cycle); ::icnt_push( m_shader_config->mem2device(i), mf->get_tpc(), mf, response_size ); m_memory_sub_partition[i]->pop(); + partiton_replys_in_parallel_per_cycle++; } else { gpu_stall_icnt2sh++; } @@ -1432,6 +1462,7 @@ void gpgpu_sim::cycle() } } } + partiton_replys_in_parallel += partiton_replys_in_parallel_per_cycle; if (clock_mask & DRAM) { for (unsigned i=0;im_n_mem;i++){ @@ -1444,6 +1475,7 @@ void gpgpu_sim::cycle() } // L2 operations follow L2 clock domain + unsigned partiton_reqs_in_parallel_per_cycle = 0; if (clock_mask & L2) { m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].clear(); for (unsigned i=0;im_n_mem_sub_partition;i++) { @@ -1455,11 +1487,15 @@ void gpgpu_sim::cycle() } else { mem_fetch* mf = (mem_fetch*) icnt_pop( m_shader_config->mem2device(i) ); m_memory_sub_partition[i]->push( mf, gpu_sim_cycle + gpu_tot_sim_cycle ); + partiton_reqs_in_parallel_per_cycle++; } m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle+gpu_tot_sim_cycle); m_memory_sub_partition[i]->accumulate_L2cache_stats(m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]); } } + partiton_reqs_in_parallel += partiton_reqs_in_parallel_per_cycle; + if(partiton_reqs_in_parallel_per_cycle > 0) + partiton_reqs_in_parallel_util += partiton_reqs_in_parallel_per_cycle; if (clock_mask & ICNT) { icnt_transfer(); -- cgit v1.3 From 93cdd2e217d7903ccf715a00b6dae8265bb38fca Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 28 Mar 2018 15:22:33 -0400 Subject: adding new stats counter --- src/gpgpu-sim/gpu-sim.cc | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index f216e7c..006d243 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -91,6 +91,8 @@ unsigned long long partiton_reqs_in_parallel = 0; unsigned long long partiton_reqs_in_parallel_total = 0; unsigned long long partiton_reqs_in_parallel_util = 0; unsigned long long partiton_reqs_in_parallel_util_total = 0; +unsigned long long gpu_sim_cycle_parition_util = 0; +unsigned long long gpu_tot_sim_cycle_parition_util = 0; unsigned long long partiton_replys_in_parallel = 0; unsigned long long partiton_replys_in_parallel_total = 0; @@ -795,6 +797,7 @@ void gpgpu_sim::init() partiton_reqs_in_parallel = 0; partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; + gpu_sim_cycle_parition_util = 0; reinit_clock_domains(); set_param_gpgpu_num_shaders(m_config.num_shader()); @@ -834,11 +837,14 @@ void gpgpu_sim::update_stats() { partiton_reqs_in_parallel_total += partiton_reqs_in_parallel; partiton_replys_in_parallel_total += partiton_replys_in_parallel; partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; + gpu_tot_sim_cycle_parition_util += gpu_sim_cycle_parition_util ; gpu_sim_cycle = 0; partiton_reqs_in_parallel = 0; partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; + gpu_sim_cycle_parition_util = 0; + gpu_tot_sim_cycle = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } @@ -1028,8 +1034,10 @@ void gpgpu_sim::gpu_print_stat() printf("partiton_level_parallism_total = %12.4f\n", (float)(partiton_reqs_in_parallel+partiton_reqs_in_parallel_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); printf("partiton_reqs_in_parallel_util = %lld\n", partiton_reqs_in_parallel_util); printf("partiton_reqs_in_parallel_util_total = %lld\n", partiton_reqs_in_parallel_util_total ); - printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle); - printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("gpu_sim_cycle_parition_util = %lld\n", gpu_sim_cycle_parition_util); + printf("gpu_tot_sim_cycle_parition_util = %lld\n", gpu_tot_sim_cycle_parition_util ); + printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle_parition_util); + printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_sim_cycle_parition_util+gpu_tot_sim_cycle_parition_util) ); printf("partiton_replys_in_parallel = %lld\n", partiton_replys_in_parallel); printf("partiton_replys_in_parallel_total = %lld\n", partiton_replys_in_parallel_total ); printf("L2_BW = %12.4f GB/Sec\n", ((float)(partiton_replys_in_parallel * 32) / (gpu_sim_cycle * m_config.icnt_period)) / 1000000000); @@ -1494,8 +1502,10 @@ void gpgpu_sim::cycle() } } partiton_reqs_in_parallel += partiton_reqs_in_parallel_per_cycle; - if(partiton_reqs_in_parallel_per_cycle > 0) + if(partiton_reqs_in_parallel_per_cycle > 0){ partiton_reqs_in_parallel_util += partiton_reqs_in_parallel_per_cycle; + gpu_sim_cycle_parition_util++; + } if (clock_mask & ICNT) { icnt_transfer(); -- cgit v1.3 From 21ac2041365dc8521a6926244b96e85fd19cc2d3 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Wed, 28 Mar 2018 22:52:12 -0400 Subject: Fixing a minor typo in an old config --- src/gpgpu-sim/gpu-sim.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 58a5d16..d452888 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -223,7 +223,7 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); - option_parser_register(opp, "-gpgpu_cache:dl1PreShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, + option_parser_register(opp, "-gpgpu_cache:dl1PrefShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); -- cgit v1.3 From 3b9cc9285b5394320873e42edee03d8f23d6d17e Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 30 Mar 2018 11:27:57 -0400 Subject: fixing gpu-tot-cycle bug --- src/gpgpu-sim/gpu-sim.cc | 1 - 1 file changed, 1 deletion(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 006d243..7835496 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -844,7 +844,6 @@ void gpgpu_sim::update_stats() { partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; gpu_sim_cycle_parition_util = 0; - gpu_tot_sim_cycle = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } -- cgit v1.3 From 89db73061e043c26df22c7f18d9adb106d8078ac Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 31 Mar 2018 17:56:50 -0400 Subject: Getting rid of our constant, annoying prints. Running workloads of any size causes ridiculous output file sizes --- src/cuda-sim/cuda-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.cc | 12 +++++++----- src/gpgpu-sim/shader.cc | 14 ++++++++------ src/trace_streams.tup | 1 + 4 files changed, 17 insertions(+), 12 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index d4ace76..a34b99b 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1408,7 +1408,7 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) if ( (g_ptx_sim_num_insn % 100000) == 0 ) { dim3 ctaid = get_ctaid(); dim3 tid = get_tid(); - printf("GPGPU-Sim PTX: %u instructions simulated : ctaid=(%u,%u,%u) tid=(%u,%u,%u)\n", + DPRINTF(LIVENESS, "GPGPU-Sim PTX: %u instructions simulated : ctaid=(%u,%u,%u) tid=(%u,%u,%u)\n", g_ptx_sim_num_insn, ctaid.x,ctaid.y,ctaid.z,tid.x,tid.y,tid.z ); fflush(stdout); } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index d452888..0ef267d 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -36,6 +36,7 @@ #include "shader.h" +#include "shader_trace.h" #include "dram.h" #include "mem_fetch.h" @@ -1175,8 +1176,8 @@ bool shader_core_ctx::occupy_shader_resource_1block(kernel_info_t & k, bool occu m_occupied_regs += (padded_cta_size * ((kernel_info->regs+3)&~3)); m_occupied_ctas++; - printf("GPGPU-Sim uArch: Shader %d occupied %d threads, %d shared mem, %d registers, %d ctas\n", - m_sid, m_occupied_n_threads, m_occupied_shmem, m_occupied_regs, m_occupied_ctas); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: Occupied %d threads, %d shared mem, %d registers, %d ctas\n", + m_occupied_n_threads, m_occupied_shmem, m_occupied_regs, m_occupied_ctas); } return true; @@ -1301,8 +1302,8 @@ void shader_core_ctx::issue_block2core( kernel_info_t &kernel ) m_n_active_cta++; shader_CTA_count_log(m_sid, 1); - printf("GPGPU-Sim uArch: core:%3d, cta:%2u, start_tid:%4u, end_tid:%4u, initialized @(%lld,%lld)\n", - m_sid, free_cta_hw_id, start_thread, end_thread, gpu_sim_cycle, gpu_tot_sim_cycle ); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: cta:%2u, start_tid:%4u, end_tid:%4u, initialized @(%lld,%lld)\n", + free_cta_hw_id, start_thread, end_thread, gpu_sim_cycle, gpu_tot_sim_cycle ); } @@ -1499,7 +1500,8 @@ void gpgpu_sim::cycle() hrs = elapsed_time/3600 - 24*days; minutes = elapsed_time/60 - 60*(hrs + 24*days); sec = elapsed_time - 60*(minutes + 60*(hrs + 24*days)); - printf("GPGPU-Sim uArch: cycles simulated: %lld inst.: %lld (ipc=%4.1f) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", + + DPRINTF(LIVENESS, "GPGPU-Sim uArch: cycles simulated: %lld inst.: %lld (ipc=%4.1f) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", gpu_tot_sim_cycle + gpu_sim_cycle, gpu_tot_sim_insn + gpu_sim_insn, (double)gpu_sim_insn/(double)gpu_sim_cycle, (unsigned)((gpu_tot_sim_insn+gpu_sim_insn) / elapsed_time), diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index d17e51d..4640d65 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1969,12 +1969,12 @@ void shader_core_ctx::register_cta_thread_exit( unsigned cta_num, kernel_info_t m_barriers.deallocate_barrier(cta_num); shader_CTA_count_unlog(m_sid, 1); - printf("GPGPU-Sim uArch: Shader %d finished CTA #%d (%lld,%lld), %u CTAs running\n", m_sid, cta_num, gpu_sim_cycle, gpu_tot_sim_cycle, - m_n_active_cta ); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: Finished CTA #%d (%lld,%lld), %u CTAs running\n", + cta_num, gpu_sim_cycle, gpu_tot_sim_cycle, m_n_active_cta); if( m_n_active_cta == 0 ) { - printf("GPGPU-Sim uArch: Shader %u empty (last released kernel %u \'%s\').\n", m_sid, kernel->get_uid(), - kernel->name().c_str() ); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: Empty (last released kernel %u \'%s\').\n", + kernel->get_uid(), kernel->name().c_str()); fflush(stdout); //Shader can only be empty when no more cta are dispatched @@ -1989,8 +1989,10 @@ void shader_core_ctx::register_cta_thread_exit( unsigned cta_num, kernel_info_t kernel->dec_running(); if( !m_gpu->kernel_more_cta_left(kernel) ) { if( !kernel->running() ) { - printf("GPGPU-Sim uArch: GPU detected kernel %u \'%s\' finished on shader %u.\n", kernel->get_uid(), - kernel->name().c_str(), m_sid ); + SHADER_DPRINTF(LIVENESS, + "GPGPU-Sim uArch: GPU detected kernel %u \'%s\' finished on shader %u.\n", kernel->get_uid(), + kernel->name().c_str(), m_sid); + if(m_kernel == kernel) m_kernel = NULL; m_gpu->set_kernel_done( kernel ); diff --git a/src/trace_streams.tup b/src/trace_streams.tup index c41690e..fd68a16 100644 --- a/src/trace_streams.tup +++ b/src/trace_streams.tup @@ -29,5 +29,6 @@ TS_TUP_BEGIN( trace_streams_type ) TS_TUP( WARP_SCHEDULER ), TS_TUP( SCOREBOARD ), TS_TUP( MEMORY_PARTITION_UNIT ), + TS_TUP( LIVENESS ), TS_TUP( NUM_TRACE_STREAMS ) TS_TUP_END( trace_streams_type ) -- cgit v1.3 From 60a32420caa0e79c7d5133ff67a55c02333b1dc2 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 28 Mar 2018 14:00:16 -0400 Subject: adding new counters for parition level para and L2 BW --- src/gpgpu-sim/gpu-sim.cc | 36 ++++++++++++++++++++++++++++++++++++ 1 file changed, 36 insertions(+) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 0ef267d..95951a8 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -88,6 +88,12 @@ unsigned long long gpu_tot_sim_cycle = 0; // performance counter for stalls due to congestion. unsigned int gpu_stall_dramfull = 0; unsigned int gpu_stall_icnt2sh = 0; +unsigned long long partiton_reqs_in_parallel = 0; +unsigned long long partiton_reqs_in_parallel_total = 0; +unsigned long long partiton_reqs_in_parallel_util = 0; +unsigned long long partiton_reqs_in_parallel_util_total = 0; +unsigned long long partiton_replys_in_parallel = 0; +unsigned long long partiton_replys_in_parallel_total = 0; /* Clock Domains */ @@ -745,6 +751,9 @@ void gpgpu_sim::init() gpu_sim_insn = 0; last_gpu_sim_insn = 0; m_total_cta_launched=0; + partiton_reqs_in_parallel = 0; + partiton_replys_in_parallel = 0; + partiton_reqs_in_parallel_util = 0; reinit_clock_domains(); set_param_gpgpu_num_shaders(m_config.num_shader()); @@ -781,8 +790,14 @@ void gpgpu_sim::update_stats() { gpu_tot_sim_cycle += gpu_sim_cycle; gpu_tot_sim_insn += gpu_sim_insn; gpu_tot_issued_cta += m_total_cta_launched; + partiton_reqs_in_parallel_total += partiton_reqs_in_parallel; + partiton_replys_in_parallel_total += partiton_replys_in_parallel; + partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; gpu_sim_cycle = 0; + partiton_reqs_in_parallel = 0; + partiton_replys_in_parallel = 0; + partiton_reqs_in_parallel_util = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } @@ -966,6 +981,19 @@ void gpgpu_sim::gpu_print_stat() printf("gpu_stall_dramfull = %d\n", gpu_stall_dramfull); printf("gpu_stall_icnt2sh = %d\n", gpu_stall_icnt2sh ); + printf("partiton_reqs_in_parallel = %lld\n", partiton_reqs_in_parallel); + printf("partiton_reqs_in_parallel_total = %lld\n", partiton_reqs_in_parallel_total ); + printf("partiton_level_parallism = %12.4f\n", (float)partiton_reqs_in_parallel / gpu_sim_cycle); + printf("partiton_level_parallism_total = %12.4f\n", (float)(partiton_reqs_in_parallel+partiton_reqs_in_parallel_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("partiton_reqs_in_parallel_util = %lld\n", partiton_reqs_in_parallel_util); + printf("partiton_reqs_in_parallel_util_total = %lld\n", partiton_reqs_in_parallel_util_total ); + printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle); + printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("partiton_replys_in_parallel = %lld\n", partiton_replys_in_parallel); + printf("partiton_replys_in_parallel_total = %lld\n", partiton_replys_in_parallel_total ); + printf("L2_BW = %12.4f GB/Sec\n", ((float)(partiton_replys_in_parallel * 32) / (gpu_sim_cycle * m_config.icnt_period)) / 1000000000); + printf("L2_BW_total = %12.4f GB/Sec\n", ((float)((partiton_replys_in_parallel+partiton_replys_in_parallel_total) * 32) / ((gpu_tot_sim_cycle+gpu_sim_cycle) * m_config.icnt_period)) / 1000000000 ); + time_t curr_time; time(&curr_time); unsigned long long elapsed_time = MAX( curr_time - g_simulation_starttime, 1 ); @@ -1367,6 +1395,7 @@ void gpgpu_sim::cycle() for (unsigned i=0;in_simt_clusters;i++) m_cluster[i]->icnt_cycle(); } + unsigned partiton_replys_in_parallel_per_cycle = 0; if (clock_mask & ICNT) { // pop from memory controller to interconnect for (unsigned i=0;im_n_mem_sub_partition;i++) { @@ -1379,6 +1408,7 @@ void gpgpu_sim::cycle() mf->set_status(IN_ICNT_TO_SHADER,gpu_sim_cycle+gpu_tot_sim_cycle); ::icnt_push( m_shader_config->mem2device(i), mf->get_tpc(), mf, response_size ); m_memory_sub_partition[i]->pop(); + partiton_replys_in_parallel_per_cycle++; } else { gpu_stall_icnt2sh++; } @@ -1387,6 +1417,7 @@ void gpgpu_sim::cycle() } } } + partiton_replys_in_parallel += partiton_replys_in_parallel_per_cycle; if (clock_mask & DRAM) { for (unsigned i=0;im_n_mem;i++){ @@ -1399,6 +1430,7 @@ void gpgpu_sim::cycle() } // L2 operations follow L2 clock domain + unsigned partiton_reqs_in_parallel_per_cycle = 0; if (clock_mask & L2) { m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].clear(); for (unsigned i=0;im_n_mem_sub_partition;i++) { @@ -1409,11 +1441,15 @@ void gpgpu_sim::cycle() } else { mem_fetch* mf = (mem_fetch*) icnt_pop( m_shader_config->mem2device(i) ); m_memory_sub_partition[i]->push( mf, gpu_sim_cycle + gpu_tot_sim_cycle ); + partiton_reqs_in_parallel_per_cycle++; } m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle+gpu_tot_sim_cycle); m_memory_sub_partition[i]->accumulate_L2cache_stats(m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]); } } + partiton_reqs_in_parallel += partiton_reqs_in_parallel_per_cycle; + if(partiton_reqs_in_parallel_per_cycle > 0) + partiton_reqs_in_parallel_util += partiton_reqs_in_parallel_per_cycle; if (clock_mask & ICNT) { icnt_transfer(); -- cgit v1.3 From 038df10118351d44f8950065ddeb698ae3e390a1 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 28 Mar 2018 15:22:33 -0400 Subject: adding new stats counter --- src/gpgpu-sim/gpu-sim.cc | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 95951a8..da35dc3 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -92,6 +92,8 @@ unsigned long long partiton_reqs_in_parallel = 0; unsigned long long partiton_reqs_in_parallel_total = 0; unsigned long long partiton_reqs_in_parallel_util = 0; unsigned long long partiton_reqs_in_parallel_util_total = 0; +unsigned long long gpu_sim_cycle_parition_util = 0; +unsigned long long gpu_tot_sim_cycle_parition_util = 0; unsigned long long partiton_replys_in_parallel = 0; unsigned long long partiton_replys_in_parallel_total = 0; @@ -754,6 +756,7 @@ void gpgpu_sim::init() partiton_reqs_in_parallel = 0; partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; + gpu_sim_cycle_parition_util = 0; reinit_clock_domains(); set_param_gpgpu_num_shaders(m_config.num_shader()); @@ -793,11 +796,14 @@ void gpgpu_sim::update_stats() { partiton_reqs_in_parallel_total += partiton_reqs_in_parallel; partiton_replys_in_parallel_total += partiton_replys_in_parallel; partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; + gpu_tot_sim_cycle_parition_util += gpu_sim_cycle_parition_util ; gpu_sim_cycle = 0; partiton_reqs_in_parallel = 0; partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; + gpu_sim_cycle_parition_util = 0; + gpu_tot_sim_cycle = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } @@ -987,8 +993,10 @@ void gpgpu_sim::gpu_print_stat() printf("partiton_level_parallism_total = %12.4f\n", (float)(partiton_reqs_in_parallel+partiton_reqs_in_parallel_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); printf("partiton_reqs_in_parallel_util = %lld\n", partiton_reqs_in_parallel_util); printf("partiton_reqs_in_parallel_util_total = %lld\n", partiton_reqs_in_parallel_util_total ); - printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle); - printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("gpu_sim_cycle_parition_util = %lld\n", gpu_sim_cycle_parition_util); + printf("gpu_tot_sim_cycle_parition_util = %lld\n", gpu_tot_sim_cycle_parition_util ); + printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle_parition_util); + printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_sim_cycle_parition_util+gpu_tot_sim_cycle_parition_util) ); printf("partiton_replys_in_parallel = %lld\n", partiton_replys_in_parallel); printf("partiton_replys_in_parallel_total = %lld\n", partiton_replys_in_parallel_total ); printf("L2_BW = %12.4f GB/Sec\n", ((float)(partiton_replys_in_parallel * 32) / (gpu_sim_cycle * m_config.icnt_period)) / 1000000000); @@ -1448,8 +1456,10 @@ void gpgpu_sim::cycle() } } partiton_reqs_in_parallel += partiton_reqs_in_parallel_per_cycle; - if(partiton_reqs_in_parallel_per_cycle > 0) + if(partiton_reqs_in_parallel_per_cycle > 0){ partiton_reqs_in_parallel_util += partiton_reqs_in_parallel_per_cycle; + gpu_sim_cycle_parition_util++; + } if (clock_mask & ICNT) { icnt_transfer(); -- cgit v1.3 From cfab001e7c65541eaf8f7573bbf4a2e5d69f296e Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 30 Mar 2018 11:27:57 -0400 Subject: fixing gpu-tot-cycle bug --- src/gpgpu-sim/gpu-sim.cc | 1 - 1 file changed, 1 deletion(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index da35dc3..550c550 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -803,7 +803,6 @@ void gpgpu_sim::update_stats() { partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; gpu_sim_cycle_parition_util = 0; - gpu_tot_sim_cycle = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } -- cgit v1.3 From 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 --- configs/4.x-cfgs/SM6_TITANX/gpgpusim.config | 2 +- configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 2 +- src/gpgpu-sim/gpu-cache.cc | 110 ++++++++++++++++------------ src/gpgpu-sim/gpu-cache.h | 22 ++++-- src/gpgpu-sim/gpu-sim.cc | 6 +- src/gpgpu-sim/l2cache.cc | 12 ++- src/gpgpu-sim/l2cache.h | 1 + src/gpgpu-sim/shader.cc | 16 ++++ src/gpgpu-sim/shader.h | 3 + 9 files changed, 115 insertions(+), 59 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config index 3097d19..7368882 100644 --- a/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/4.x-cfgs/SM6_TITANX/gpgpusim.config @@ -73,7 +73,7 @@ -gpgpu_n_cluster_ejection_buffer_size 32 # 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:F:L,A:128:4,16:0,32 +-gpgpu_cache:dl2 S:64:128:16,L:B:m:L:L,A:128:4,16:0,32 -gpgpu_cache:dl2_texture_only 0 -gpgpu_dram_partition_queues 32:32:32:32 -perf_sim_memcpy 0 diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config index 8969168..da98547 100644 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -78,7 +78,7 @@ -gpgpu_n_cluster_ejection_buffer_size 32 # 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,L:B:m:F:L,A:256:4,32:0,32 +-gpgpu_cache:dl2 S:64:128:24,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 diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 75ec00a..9d81de9 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -382,6 +382,8 @@ void tag_array::fill( unsigned index, unsigned time, mem_fetch* mf) m_lines[index]->fill(time, mf->get_access_sector_mask()); } + +//TODO: we need write back the flushed data to the upper level void tag_array::flush() { for (unsigned i=0; i < m_config.get_num_lines(); i++) @@ -391,6 +393,13 @@ void tag_array::flush() } } +void tag_array::invalidate() +{ + for (unsigned i=0; i < m_config.get_num_lines(); i++) + for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) + m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; +} + float tag_array::windowed_miss_rate( ) const { unsigned n_access = m_access - m_prev_snapshot_access; @@ -1162,52 +1171,7 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, unsigned time, std::list &events, enum cache_request_status status ) { - - new_addr_type block_addr = m_config.block_addr(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(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; - - 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( 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; - } - return RESERVATION_FAIL; - - /*new_addr_type block_addr = m_config.block_addr(addr); + new_addr_type block_addr = m_config.block_addr(addr); new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) @@ -1316,7 +1280,59 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, return MISS; } return RESERVATION_FAIL; - }*/ + } +} + +enum cache_request_status +data_cache::wr_miss_wa_lazy_fetch_on_read( new_addr_type addr, + unsigned cache_index, mem_fetch *mf, + unsigned time, std::list &events, + enum cache_request_status status ) +{ + + new_addr_type block_addr = m_config.block_addr(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(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; + + 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( 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; + } + 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 d2b7757..96b9a6d 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -454,7 +454,8 @@ enum allocation_policy_t { enum write_allocate_policy_t { NO_WRITE_ALLOCATE, WRITE_ALLOCATE, - FETCH_ON_WRITE + FETCH_ON_WRITE, + LAZY_FETCH_ON_READ }; enum mshr_config_t { @@ -555,6 +556,7 @@ public: case 'N': m_write_alloc_policy = NO_WRITE_ALLOCATE; break; case 'W': m_write_alloc_policy = WRITE_ALLOCATE; break; case 'F': m_write_alloc_policy = FETCH_ON_WRITE; break; + case 'L': m_write_alloc_policy = LAZY_FETCH_ON_READ; break; default: exit_parse_error(); } @@ -572,9 +574,9 @@ public: assert(0 && "Invalid cache configuration: Writeback cache cannot allocate new line on fill. "); } - if(m_write_alloc_policy == FETCH_ON_WRITE && m_alloc_policy == ON_FILL) + if((m_write_alloc_policy == FETCH_ON_WRITE || m_write_alloc_policy == LAZY_FETCH_ON_READ )&& m_alloc_policy == ON_FILL) { - assert(0 && "Invalid cache configuration: FETCH_ON_WRITE cannot work properly with ON_FILL policy. Cache must be ON_MISS. "); + assert(0 && "Invalid cache configuration: FETCH_ON_WRITE and LAZY_FETCH_ON_READ cannot work properly with ON_FILL policy. Cache must be ON_MISS. "); } if(m_cache_type == SECTOR) { @@ -742,7 +744,8 @@ public: unsigned size() const { return m_config.get_num_lines();} cache_block_t* get_block(unsigned idx) { return m_lines[idx];} - void flush(); // flash invalidate all entries + void flush(); // flush all written entries + void invalidate(); // invalidate all entries void new_window(); void print( FILE *stream, unsigned &total_access, unsigned &total_misses ) const; @@ -994,6 +997,7 @@ public: mem_fetch *next_access(){return m_mshrs.next_access();} // flash invalidate all entries in cache void flush(){m_tag_array->flush();} + void invalidate(){m_tag_array->invalidate();} void print(FILE *fp, unsigned &accesses, unsigned &misses) const; void display_state( FILE *fp ) const; @@ -1179,6 +1183,7 @@ public: case NO_WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_no_wa; break; case WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_wa_naive; break; case FETCH_ON_WRITE: m_wr_miss = &data_cache::wr_miss_wa_fetch_on_write; break; + case LAZY_FETCH_ON_READ: m_wr_miss = &data_cache::wr_miss_wa_lazy_fetch_on_read; break; default: assert(0 && "Error: Must set valid cache write miss policy\n"); break; // Need to set a write miss function @@ -1299,7 +1304,14 @@ protected: mem_fetch *mf, unsigned time, std::list &events, - enum cache_request_status status ); // write-allocate with read-fetch-only + enum cache_request_status status ); // write-allocate with fetch-on-every-write + enum cache_request_status + wr_miss_wa_lazy_fetch_on_read( new_addr_type addr, + unsigned cache_index, + mem_fetch *mf, + unsigned time, + std::list &events, + enum cache_request_status status ); // write-allocate with read-fetch-only enum cache_request_status wr_miss_wa_write_validate( new_addr_type addr, unsigned cache_index, diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index c5d4464..d48de25 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1548,12 +1548,12 @@ void gpgpu_sim::cycle() issue_block2core(); - // Depending on configuration, flush the caches once all of threads are completed. + // Depending on configuration, invalidate the caches once all of threads are completed. int all_threads_complete = 1; if (m_config.gpgpu_flush_l1_cache) { for (unsigned i=0;in_simt_clusters;i++) { if (m_cluster[i]->get_not_completed() == 0) - m_cluster[i]->cache_flush(); + m_cluster[i]->cache_invalidate(); else all_threads_complete = 0 ; } @@ -1575,7 +1575,7 @@ void gpgpu_sim::cycle() int dlc = 0; for (unsigned i=0;im_n_mem;i++) { dlc = m_memory_sub_partition[i]->flushL2(); - assert (dlc == 0); // need to model actual writes to DRAM here + assert (dlc == 0); // TODO: need to model actual writes to DRAM here printf("Dirty lines flushed from L2 %d is %d\n", i, dlc ); } } 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 diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 2d13918..18c0a8b 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -162,6 +162,7 @@ public: void set_done( mem_fetch *mf ); unsigned flushL2(); + unsigned invalidateL2(); // interface to L2_dram_queue bool L2_dram_queue_empty() const; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index d2f40a1..b660b8c 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1593,6 +1593,11 @@ void ldst_unit::flush(){ m_L1D->flush(); } +void ldst_unit::invalidate(){ + // Flush L1D cache + m_L1D->invalidate(); +} + simd_function_unit::simd_function_unit( const shader_core_config *config ) { m_config=config; @@ -2644,6 +2649,11 @@ void shader_core_ctx::cache_flush() m_ldst_unit->flush(); } +void shader_core_ctx::cache_invalidate() +{ + m_ldst_unit->invalidate(); +} + // modifiers std::list opndcoll_rfu_t::arbiter_t::allocate_reads() { @@ -3461,6 +3471,12 @@ void simt_core_cluster::cache_flush() m_core[i]->cache_flush(); } +void simt_core_cluster::cache_invalidate() +{ + for( unsigned i=0; i < m_config->n_simt_cores_per_cluster; i++ ) + m_core[i]->cache_invalidate(); +} + bool simt_core_cluster::icnt_injection_buffer_full(unsigned size, bool write) { unsigned request_size = size; diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index ae22eaa..cc441b3 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1155,6 +1155,7 @@ public: void fill( mem_fetch *mf ); void flush(); + void invalidate(); void writeback(); // accessors @@ -1655,6 +1656,7 @@ public: void issue_block2core( class kernel_info_t &kernel ); void cache_flush(); + void cache_invalidate(); void accept_fetch_response( mem_fetch *mf ); void accept_ldst_unit_response( class mem_fetch * mf ); void broadcast_barrier_reduction(unsigned cta_id, unsigned bar_id,warp_set_t warps); @@ -1947,6 +1949,7 @@ public: void reinit(); unsigned issue_block2core(); void cache_flush(); + void cache_invalidate(); bool icnt_injection_buffer_full(unsigned size, bool write); void icnt_inject_request_packet(class mem_fetch *mf); -- cgit v1.3 From 944f6dbf23d792dde360d3a4f2334de3b541de52 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 27 Aug 2018 20:28:24 -0400 Subject: fixing ead/write buffer and new configs files --- configs/3.x-cfgs/GTX480/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/GTX480/gpgpusim.config | 133 +++++ configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml | 538 ++++++++++++++++++++ .../GeForceGTX750Ti/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config | 130 +++++ .../GeForceGTX750Ti/gpuwattch_gtx750Ti.xml | 539 +++++++++++++++++++++ configs/3.x-cfgs/QuadroFX5600/gpgpusim.config | 96 ++++ .../QuadroFX5600/gpuwattch_quadrofx5600.xml | 538 ++++++++++++++++++++ .../3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt | 70 +++ .../3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt | 69 +++ configs/3.x-cfgs/QuadroFX5800/gpgpusim.config | 86 ++++ .../3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config | 149 ++++++ .../3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml | 538 ++++++++++++++++++++ .../3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 70 --- configs/3.x-cfgs/SM7_TITANV/gpgpusim.config | 165 ------- .../3.x-cfgs/TeslaC2050/config_fermi_islip.icnt | 70 +++ configs/3.x-cfgs/TeslaC2050/gpgpusim.config | 132 +++++ configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config | 97 ---- .../SM1_QFX5600/gpuwattch_quadrofx5600.xml | 538 -------------------- .../4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt | 70 --- .../4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt | 69 --- configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config | 87 ---- configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/SM2_C2050/gpgpusim.config | 133 ----- .../4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt | 4 +- configs/4.x-cfgs/SM2_GTX480/gpgpusim.config | 23 +- .../4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/SM5_GTX750/gpgpusim.config | 131 ----- configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml | 539 --------------------- .../4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt | 70 --- configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config | 149 ------ .../4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml | 538 -------------------- .../SM6_P100_64SMs/config_fermi_islip.icnt | 73 +++ configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config | 174 +++++++ .../4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt | 74 --- configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 186 ------- src/gpgpu-sim/addrdec.cc | 17 +- src/gpgpu-sim/dram.cc | 12 +- src/gpgpu-sim/dram_sched.cc | 6 +- src/gpgpu-sim/gpu-sim.cc | 3 +- src/gpgpu-sim/shader.cc | 4 +- 42 files changed, 3596 insertions(+), 3074 deletions(-) create mode 100644 configs/3.x-cfgs/GTX480/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/GTX480/gpgpusim.config create mode 100755 configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml create mode 100644 configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config create mode 100755 configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml create mode 100644 configs/3.x-cfgs/QuadroFX5600/gpgpusim.config create mode 100644 configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml create mode 100644 configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt create mode 100644 configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt create mode 100644 configs/3.x-cfgs/QuadroFX5800/gpgpusim.config create mode 100644 configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config create mode 100755 configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml delete mode 100644 configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt delete mode 100644 configs/3.x-cfgs/SM7_TITANV/gpgpusim.config create mode 100644 configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt create mode 100644 configs/3.x-cfgs/TeslaC2050/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml delete mode 100644 configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt delete mode 100644 configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt delete mode 100644 configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM2_C2050/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM5_GTX750/gpgpusim.config delete mode 100755 configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml delete mode 100644 configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config delete mode 100755 configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml create mode 100644 configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt create mode 100644 configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config delete mode 100644 configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt delete mode 100644 configs/4.x-cfgs/SM7_TITANV/gpgpusim.config (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/3.x-cfgs/GTX480/config_fermi_islip.icnt b/configs/3.x-cfgs/GTX480/config_fermi_islip.icnt new file mode 100644 index 0000000..7820e4e --- /dev/null +++ b/configs/3.x-cfgs/GTX480/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 27; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +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/3.x-cfgs/GTX480/gpgpusim.config b/configs/3.x-cfgs/GTX480/gpgpusim.config new file mode 100644 index 0000000..436cb41 --- /dev/null +++ b/configs/3.x-cfgs/GTX480/gpgpusim.config @@ -0,0 +1,133 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + + +# 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 15 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 700.0:700.0:700.0:924.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-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 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 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 16 +-gpgpu_dram_return_queue_size 116 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi 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 +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml b/configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/3.x-cfgs/GTX480/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt b/configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt new file mode 100644 index 0000000..069ca02 --- /dev/null +++ b/configs/3.x-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 7; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +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/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config b/configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config new file mode 100644 index 0000000..8b030b6 --- /dev/null +++ b/configs/3.x-cfgs/GeForceGTX750Ti/gpgpusim.config @@ -0,0 +1,130 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 52 + + +# 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 5 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 2 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Maxwell clock domains +#-gpgpu_clock_domains ::: +# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. +-gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 + +# 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 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 8 +-gpgpu_num_sfu_units 32 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 6,12,13,13,210 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 6,12,6,6,374 +-ptx_opcode_initiation_fp 1,1,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gmem_skip_L1D 1 +-gpgpu_shmem_size 65536 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache +-gpgpu_cache:dl2 1024:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 16:128:32,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 150 +-dram_latency 130 + +# 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 16 +-gpgpu_dram_return_queue_size 300 + +# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 32 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Maxwell has four schedulers per core +-gpgpu_num_sched_per_core 4 +# 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 +# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. +# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx750Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml b/configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml new file mode 100755 index 0000000..e2b2324 --- /dev/null +++ b/configs/3.x-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml @@ -0,0 +1,539 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/QuadroFX5600/gpgpusim.config b/configs/3.x-cfgs/QuadroFX5600/gpgpusim.config new file mode 100644 index 0000000..cb87b65 --- /dev/null +++ b/configs/3.x-cfgs/QuadroFX5600/gpgpusim.config @@ -0,0 +1,96 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 12 + +# high level architecture configuration +-gpgpu_n_clusters 8 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 6 +-gpgpu_clock_domains 337.5:600.0:600.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) +-gpgpu_shader_core_pipeline 768:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,1,1,1,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +-gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +# TLB parameters +#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 +#-gpgpu_tlbl2_latency 45 + + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file icnt_config_islip.icnt + +# dram scheduler 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 (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 55 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + + + +# Using cuobjdump to extract ptx/SASS +-gpgpu_ptx_use_cuobjdump 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 +-power_trace_enabled 0 +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_quadrofx5600.xml + +-steady_power_levels_enabled 1 +-steady_state_definition 8,4 diff --git a/configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml b/configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml new file mode 100644 index 0000000..2c5a6fc --- /dev/null +++ b/configs/3.x-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt b/configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt new file mode 100644 index 0000000..de3bcc8 --- /dev/null +++ b/configs/3.x-cfgs/QuadroFX5600/icnt_config_islip.icnt @@ -0,0 +1,70 @@ +//14*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 14; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +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/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt b/configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt new file mode 100644 index 0000000..cfe9cac --- /dev/null +++ b/configs/3.x-cfgs/QuadroFX5800/config_quadro_islip.icnt @@ -0,0 +1,69 @@ +//18*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 18; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +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/3.x-cfgs/QuadroFX5800/gpgpusim.config b/configs/3.x-cfgs/QuadroFX5800/gpgpusim.config new file mode 100644 index 0000000..82243c2 --- /dev/null +++ b/configs/3.x-cfgs/QuadroFX5800/gpgpusim.config @@ -0,0 +1,86 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 13 + +# high level architecture configuration +-gpgpu_n_clusters 10 +-gpgpu_n_cores_per_cluster 3 +-gpgpu_n_mem 8 +-gpgpu_clock_domains 325.0:650.0:650.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,1,1,1,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +-gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file config_quadro_islip.icnt + +# dram scheduler 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 (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 325MHz = 74 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 58 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-dram_data_command_freq_ratio 2 # GDDR3 is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt b/configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt new file mode 100644 index 0000000..2a69ddd --- /dev/null +++ b/configs/3.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 50; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +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/3.x-cfgs/SM6_GTX1080/gpgpusim.config b/configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config new file mode 100644 index 0000000..47c2b6a --- /dev/null +++ b/configs/3.x-cfgs/SM6_GTX1080/gpgpusim.config @@ -0,0 +1,149 @@ +# This config models the Pascal GP102 (GeForceGTX 1080Ti) + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + +# 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 1 +-gpgpu_n_mem 11 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 + +# 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_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,2,8 +-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 1,2,1,1,130 + +# ::,::::,::,:** +# ** 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 64KB L1 cache +# The default is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 1 + +# 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 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 +-gpgpu_cache:dl2_texture_only 0 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-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_num_reg_banks should be increased to 32, but it gives an error! +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# 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 116 + +# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) +# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Pascal has four 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 +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx1080Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml b/configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml new file mode 100755 index 0000000..02619ff --- /dev/null +++ b/configs/3.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt deleted file mode 100644 index fac792a..0000000 --- a/configs/3.x-cfgs/SM7_TITANV/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// 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 = 8; - -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/3.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config deleted file mode 100644 index f1ae2dc..0000000 --- a/configs/3.x-cfgs/SM7_TITANV/gpgpusim.config +++ /dev/null @@ -1,165 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# 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 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 - -# 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 4,4,1,4,4,1,9 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 - -# ::,::::,::,:** -# ** 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 64:128:8,L:L:m:N:H,A:256:8,16:0 --gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 --gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 1 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 0 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 64:64:64:64 - -# 128 KB Inst. --gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 64 KB Const --gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 14 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_units_mem 10 --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_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 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 116 - -# 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" - -# Fermi 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 --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/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt b/configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt new file mode 100644 index 0000000..a11bd8e --- /dev/null +++ b/configs/3.x-cfgs/TeslaC2050/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//20*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 26; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +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/3.x-cfgs/TeslaC2050/gpgpusim.config b/configs/3.x-cfgs/TeslaC2050/gpgpusim.config new file mode 100644 index 0000000..442ab8b --- /dev/null +++ b/configs/3.x-cfgs/TeslaC2050/gpgpusim.config @@ -0,0 +1,132 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + +# Using cuobjdump to extract ptx/SASS +#-gpgpu_ptx_use_cuobjdump 1 # use default + +# 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 14 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 575.0:575.0:575.0:750.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-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 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 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 750MHz / 575MHz = 130 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 114 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi 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 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config b/configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config deleted file mode 100644 index e3cab18..0000000 --- a/configs/4.x-cfgs/SM1_QFX5600/gpgpusim.config +++ /dev/null @@ -1,97 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 12 - -# high level architecture configuration --gpgpu_n_clusters 8 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 6 --gpgpu_clock_domains 337.5:600.0:600.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 -#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) --gpgpu_shader_core_pipeline 768:32 --gpgpu_shader_cta 8 --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 --gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour --gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - -# TLB parameters -#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 -#-gpgpu_tlbl2_latency 45 - - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file icnt_config_islip.icnt - -# dram scheduler 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 (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 55 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz -# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - - - -# Using cuobjdump to extract ptx/SASS --gpgpu_ptx_use_cuobjdump 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 --power_trace_enabled 0 --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_quadrofx5600.xml - --steady_power_levels_enabled 1 --steady_state_definition 8,4 diff --git a/configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml b/configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml deleted file mode 100644 index 2c5a6fc..0000000 --- a/configs/4.x-cfgs/SM1_QFX5600/gpuwattch_quadrofx5600.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt b/configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt deleted file mode 100644 index de3bcc8..0000000 --- a/configs/4.x-cfgs/SM1_QFX5600/icnt_config_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//14*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 14; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -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/SM1_QFX5800/config_quadro_islip.icnt b/configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt deleted file mode 100644 index cfe9cac..0000000 --- a/configs/4.x-cfgs/SM1_QFX5800/config_quadro_islip.icnt +++ /dev/null @@ -1,69 +0,0 @@ -//18*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 18; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -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/SM1_QFX5800/gpgpusim.config b/configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config deleted file mode 100644 index 56dbb17..0000000 --- a/configs/4.x-cfgs/SM1_QFX5800/gpgpusim.config +++ /dev/null @@ -1,87 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 13 - -# high level architecture configuration --gpgpu_n_clusters 10 --gpgpu_n_cores_per_cluster 3 --gpgpu_n_mem 8 --gpgpu_clock_domains 325.0:650.0:650.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 8 --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 --gpgpu_pipeline_widths 1,0,1,1,1,0,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo --gpgpu_cache:il1 N:4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 N:16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file config_quadro_islip.icnt - -# dram scheduler 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 (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 325MHz = 74 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 58 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --dram_data_command_freq_ratio 2 # GDDR3 is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt b/configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt deleted file mode 100644 index a11bd8e..0000000 --- a/configs/4.x-cfgs/SM2_C2050/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//20*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 26; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -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/SM2_C2050/gpgpusim.config b/configs/4.x-cfgs/SM2_C2050/gpgpusim.config deleted file mode 100644 index aa5f5f3..0000000 --- a/configs/4.x-cfgs/SM2_C2050/gpgpusim.config +++ /dev/null @@ -1,133 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - -# Using cuobjdump to extract ptx/SASS -#-gpgpu_ptx_use_cuobjdump 1 # use default - -# 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 14 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 575.0:575.0:575.0:750.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --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 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 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 750MHz / 575MHz = 130 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 114 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Fermi 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 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt b/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt index 7820e4e..d372b26 100644 --- a/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt +++ b/configs/4.x-cfgs/SM2_GTX480/config_fermi_islip.icnt @@ -1,6 +1,6 @@ //21*1 fly with 32 flits per packet under gpgpusim injection mode use_map = 0; -flit_size = 32; +flit_size = 40; // currently we do not use this, see subnets below network_count = 2; @@ -17,7 +17,7 @@ routing_function = dest_tag; // Flow control num_vcs = 1; -vc_buf_size = 8; +vc_buf_size = 64; wait_for_tail_credit = 0; diff --git a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config index 03fcda1..7f8da49 100644 --- a/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/4.x-cfgs/SM2_GTX480/gpgpusim.config @@ -50,20 +50,25 @@ # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo # Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,S:128:8,8 -gpgpu_shmem_size 49152 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 # The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:32:8,8 #-gpgpu_shmem_size 16384 # 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 N:64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2 S:64:128:8,L:B:m:L:L,A:256:4,4:0,32 -gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 0 +-memory_partition_indexing 0 --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 +-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_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,S:2:32,4 # enable operand collector -gpgpu_operand_collector_num_units_sp 6 @@ -76,6 +81,7 @@ -gpgpu_shmem_num_banks 32 -gpgpu_shmem_limited_broadcast 0 -gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 20 -gpgpu_max_insn_issue_per_warp 1 @@ -110,6 +116,13 @@ -gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" +# 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 + # Fermi has two schedulers per core -gpgpu_num_sched_per_core 2 # Two Level Scheduler with active and pending pools diff --git a/configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt b/configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt deleted file mode 100644 index 069ca02..0000000 --- a/configs/4.x-cfgs/SM5_GTX750/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 7; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -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/SM5_GTX750/gpgpusim.config b/configs/4.x-cfgs/SM5_GTX750/gpgpusim.config deleted file mode 100644 index 9366f93..0000000 --- a/configs/4.x-cfgs/SM5_GTX750/gpgpusim.config +++ /dev/null @@ -1,131 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 52 - - -# 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 5 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 2 --gpgpu_n_sub_partition_per_mchannel 1 - -# Maxwell clock domains -#-gpgpu_clock_domains ::: -# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. --gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 - -# 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 8 --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 --gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 --gpgpu_num_sp_units 8 --gpgpu_num_sfu_units 32 --gpgpu_num_dp_units 0 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 6,12,13,13,210 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 6,12,6,6,374 --ptx_opcode_initiation_fp 1,1,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - --gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,A:32:8,8 --gmem_skip_L1D 1 --gpgpu_shmem_size 65536 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache --gpgpu_cache:dl2 N:1024:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 N:4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 N:16:128:32,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 N:64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 150 --dram_latency 130 - -# 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 16 --gpgpu_dram_return_queue_size 300 - -# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 32 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Maxwell has four schedulers per core --gpgpu_num_sched_per_core 4 -# 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 -# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. -# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx750Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml b/configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml deleted file mode 100755 index e2b2324..0000000 --- a/configs/4.x-cfgs/SM5_GTX750/gpuwattch_gtx750Ti.xml +++ /dev/null @@ -1,539 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt b/configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt deleted file mode 100644 index 2a69ddd..0000000 --- a/configs/4.x-cfgs/SM6_GTX1080/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 50; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -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_GTX1080/gpgpusim.config b/configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config deleted file mode 100644 index 47c2b6a..0000000 --- a/configs/4.x-cfgs/SM6_GTX1080/gpgpusim.config +++ /dev/null @@ -1,149 +0,0 @@ -# This config models the Pascal GP102 (GeForceGTX 1080Ti) - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - -# 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 1 --gpgpu_n_mem 11 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 - -# 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_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,2,8 --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 1,2,1,1,130 - -# ::,::::,::,:** -# ** 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 64KB L1 cache -# The default is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 --gpgpu_shmem_size 98304 --gmem_skip_L1D 1 - -# 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 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 --gpgpu_cache:dl2_texture_only 0 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --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_num_reg_banks should be increased to 32, but it gives an error! --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# 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 116 - -# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) -# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Pascal has four 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 --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx1080Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml b/configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml deleted file mode 100755 index 02619ff..0000000 --- a/configs/4.x-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 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 new file mode 100644 index 0000000..81153b0 --- /dev/null +++ b/configs/4.x-cfgs/SM6_P100_64SMs/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//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 new file mode 100644 index 0000000..edcd919 --- /dev/null +++ b/configs/4.x-cfgs/SM6_P100_64SMs/gpgpusim.config @@ -0,0 +1,174 @@ +# 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/SM7_TITANV/config_fermi_islip.icnt b/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt deleted file mode 100644 index 616e9f3..0000000 --- a/configs/4.x-cfgs/SM7_TITANV/config_fermi_islip.icnt +++ /dev/null @@ -1,74 +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/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config deleted file mode 100644 index da98547..0000000 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ /dev/null @@ -1,186 +0,0 @@ -# This config models the Volta Titan X -# For more info about this card: -# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf -# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) -# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf -# https://devblogs.nvidia.com/inside-volta/ -# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 70 - -# 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 40 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 24 --gpgpu_n_sub_partition_per_mchannel 1 - -# volta clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA GP100 clock domains are adopted from -# https://en.wikipedia.org/wiki/Volta_(microarchitecture) --gpgpu_clock_domains 1200.0:1200.0:2000.0:850.0 -# boost mode -# -gpgpu_clock_domains 1455.0:1455.0:2000.0:850.0 - -# 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 4,4,4,1,4,4,4,1,12 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 4 --gpgpu_num_dp_units 4 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from -# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 2,2,2,2,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 4,4,4,4,130 --ptx_opcode_latency_sfu 100 --ptx_opcode_initiation_sfu 8 - - -# ::,::::,::,:** -# ** 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:64:128:8,L:L:f:N:H,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:64:128:16,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 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 1 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 0 --icnt_flit_size 40 --gpgpu_n_cluster_ejection_buffer_size 32 - -# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache --gpgpu_cache:dl2 S:64:128:24,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 - -# 128 KB Inst. --gpgpu_cache:il1 N:64:128:16,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 -# 64 KB Const --gpgpu_const_cache:l1 N:128:64:8,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 1 --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/src/gpgpu-sim/addrdec.cc b/src/gpgpu-sim/addrdec.cc index cfd90ec..e7e27b7 100644 --- a/src/gpgpu-sim/addrdec.cc +++ b/src/gpgpu-sim/addrdec.cc @@ -132,7 +132,7 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, addrdec_ chip[0] = a[13]^a[12]^a[11]^a[10]^a[9]^a[6]^a[5]^a[3]^a[0]^chip[0]; chip[1] = a[14]^a[13]^a[12]^a[11]^a[10]^a[7]^a[6]^a[4]^a[1]^chip[1]; chip[2] = a[14]^a[10]^a[9]^a[8]^a[7]^a[6]^a[3]^a[2]^a[0]^chip[2]; - chip[3] = a[11]^a[10]^a[9]^a[8]^a[7]^a[4]^a[3]^a[1]^chip[3]; + chip[3] = a[11]^a[10]^a[9]^a[8]^a[7]^a[4]^a[3]^a[1]^chip[3]; chip[4] = a[12]^a[11]^a[10]^a[9]^a[8]^a[5]^a[4]^a[2]^chip[4]; tlx->chip = chip.to_ulong(); @@ -144,8 +144,21 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, addrdec_ assert(tlx->chip < m_n_channel); break; case CUSTOM: - /* No custom set function implemented */ + { + //random selected bits + //do you custom hashing function here, similar to + //Liu, Yuxi, et al. "Get Out of the Valley: Power-Efficient Address Mapping for GPUs." ISCA 2018 + std::bitset<64> b(tlx->row); + std::bitset<5> chip(tlx->chip); + chip[0] = b[13]^b[10]^b[9]^b[5]^b[0]^chip[0]; + chip[1] = b[12]^b[11]^b[6]^b[1]^chip[1]; + chip[2] = b[14]^b[9]^b[8]^b[7]^b[2]^chip[2]; + chip[3] = b[11]^b[10]^b[8]^b[3]^chip[3]; + chip[4] = b[12]^b[9]^b[8]^b[5]^b[4]^chip[4]; + tlx->chip = chip.to_ulong(); + assert(tlx->chip < m_n_channel); break; + } default: assert("\nUndefined set index function.\n" && 0); break; diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index 92aa819..ac63327 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -730,13 +730,15 @@ void dram_t::print( FILE* simFile) const printf("\nwrite_to_read_ratio_blp_rw_average = %.6f", write_to_read_ratio_blp_rw_average /banks_access_rw_total); printf("\nGrpLevelPara = %.6f \n", (float)bkgrp_parallsim_rw /banks_access_rw_total); - printf("\nbwutil = %.6f \n", (float)bwutil/n_cmd); + printf("\nBW Util details:\n"); + printf("bwutil = %.6f \n", (float)bwutil/n_cmd); printf("total_CMD = %d \n", n_cmd); printf("util_bw = %d \n", util_bw); printf("Wasted_Col = %d \n", wasted_bw_col); - printf("Wasted_Row %d \n", wasted_bw_row); - printf("Idle = %d \n\n", idle_bw); + printf("Wasted_Row = %d \n", wasted_bw_row); + printf("Idle = %d \n", idle_bw); + printf("\nBW Util Bottlenecks: \n"); printf("RCDc_limit = %d \n", RCDc_limit); printf("RCDWRc_limit = %d \n", RCDWRc_limit); printf("WTRc_limit = %d \n", WTRc_limit); @@ -747,6 +749,7 @@ void dram_t::print( FILE* simFile) const printf("WTRc_limit_alone = %d \n", WTRc_limit_alone); printf("RTWc_limit_alone = %d \n", RTWc_limit_alone); + printf("\nCommands details: \n"); printf("total_CMD = %d \n", n_cmd); printf("n_nop = %d \n", n_nop); printf("Read = %d \n", n_rd); @@ -757,8 +760,9 @@ void dram_t::print( FILE* simFile) const printf("n_pre = %d \n", n_pre); printf("n_ref = %d \n", n_ref); printf("n_req = %d \n", n_req ); - printf("total_req = %d \n\n", n_rd+n_wr+n_rd_L2_A+n_wr_WB); + printf("total_req = %d \n", n_rd+n_wr+n_rd_L2_A+n_wr_WB); + printf("\nDual Bus Interface Util: \n"); printf("issued_total_row = %lu \n", issued_total_row); printf("issued_total_col = %lu \n", issued_total_col); printf("Row_Bus_Util = %.6f \n", (float)issued_total_row / n_cmd); diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc index f754d36..ff50050 100644 --- a/src/gpgpu-sim/dram_sched.cc +++ b/src/gpgpu-sim/dram_sched.cc @@ -109,12 +109,14 @@ dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row ) if(m_config->seperate_write_queue_enabled) { if(m_mode == READ_MODE && ((m_num_write_pending >= m_config->write_high_watermark ) - || (m_queue[bank].empty() && !m_write_queue[bank].empty()))) { + // || (m_queue[bank].empty() && !m_write_queue[bank].empty()) + )) { m_mode = WRITE_MODE; } else if(m_mode == WRITE_MODE && (( m_num_write_pending < m_config->write_low_watermark ) - || (!m_queue[bank].empty() && m_write_queue[bank].empty()))){ + // || (!m_queue[bank].empty() && m_write_queue[bank].empty()) + )){ m_mode = READ_MODE; } } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index d48de25..ea2dfba 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1495,7 +1495,8 @@ void gpgpu_sim::cycle() } else { mem_fetch* mf = (mem_fetch*) icnt_pop( m_shader_config->mem2device(i) ); m_memory_sub_partition[i]->push( mf, gpu_sim_cycle + gpu_tot_sim_cycle ); - partiton_reqs_in_parallel_per_cycle++; + if(mf) + partiton_reqs_in_parallel_per_cycle++; } m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle+gpu_tot_sim_cycle); m_memory_sub_partition[i]->accumulate_L2cache_stats(m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 4828875..b59e5d2 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -451,7 +451,7 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout, "gpgpu_n_intrawarp_mshr_merge = %d\n", gpgpu_n_intrawarp_mshr_merge); fprintf(fout, "gpgpu_n_cmem_portconflict = %d\n", gpgpu_n_cmem_portconflict); - fprintf(fout, "gpgpu_stall_shd_mem[c_mem][bk_conf] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][BK_CONF]); + fprintf(fout, "gpgpu_stall_shd_mem[c_mem][resource_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][BK_CONF]); //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][MSHR_RC_FAIL]); //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][ICNT_RC_FAIL]); //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][DATA_PORT_STALL]); @@ -459,7 +459,7 @@ void shader_core_stats::print( FILE* fout ) const //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][ICNT_RC_FAIL]); //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][DATA_PORT_STALL]); fprintf(fout, "gpgpu_stall_shd_mem[s_mem][bk_conf] = %d\n", gpu_stall_shd_mem_breakdown[S_MEM][BK_CONF]); - fprintf(fout, "gpgpu_stall_shd_mem[gl_mem][bk_conf] = %d\n", + fprintf(fout, "gpgpu_stall_shd_mem[gl_mem][resource_stall] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][BK_CONF] + gpu_stall_shd_mem_breakdown[G_MEM_ST][BK_CONF] + gpu_stall_shd_mem_breakdown[L_MEM_LD][BK_CONF] + -- 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/gpu-sim.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 afb28556c4f08e4daca17525e7e377d9aaec0ee5 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sun, 7 Oct 2018 21:41:35 -0400 Subject: Unrolling Aksahy's stats - as they seem to be really screwing things up. --- configs/3.x-cfgs/GTX480/gpgpusim.config | 2 +- src/abstract_hardware_model.h | 7 ----- src/cuda-sim/cuda-sim.cc | 46 +++++++++++++++++---------------- src/cuda-sim/cuda-sim.h | 2 -- src/cuda-sim/opcodes.def | 2 -- src/gpgpu-sim/gpu-sim.cc | 10 +++---- src/gpgpu-sim/gpu-sim.h | 3 --- src/gpgpu-sim/l2cache.cc | 20 +++----------- src/gpgpu-sim/l2cache.h | 8 ++---- src/gpgpu-sim/shader.cc | 35 +------------------------ src/gpgpu-sim/shader.h | 22 +--------------- src/gpgpusim_entrypoint.cc | 3 --- src/stream_manager.cc | 2 +- 13 files changed, 38 insertions(+), 124 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/3.x-cfgs/GTX480/gpgpusim.config b/configs/3.x-cfgs/GTX480/gpgpusim.config index eb25bc3..436cb41 100644 --- a/configs/3.x-cfgs/GTX480/gpgpusim.config +++ b/configs/3.x-cfgs/GTX480/gpgpusim.config @@ -1,5 +1,5 @@ # functional simulator specification --gpgpu_ptx_instruction_classification 1 +-gpgpu_ptx_instruction_classification 0 -gpgpu_ptx_sim_mode 0 -gpgpu_ptx_force_max_capability 20 diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 35f289c..e708fa7 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -808,7 +808,6 @@ public: arch_reg.dst[i] = -1; } isize=0; - op_classification = 0; } bool valid() const { return m_decoded; } virtual void print_insn( FILE *fp ) const @@ -827,7 +826,6 @@ public: address_type pc; // program counter address of instruction unsigned isize; // size of instruction in bytes op_type op; // opcode (uarch visible) - int op_classification; // classification of opcode for statistics purpopses barrier_type bar_type; reduction_type red_type; @@ -1035,11 +1033,6 @@ public: return cycles > 0; } - unsigned get_cycles() - { - return cycles; - } - void print( FILE *fout ) const; unsigned get_uid() const { return m_uid; } diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 02cd395..93bbc1d 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -52,10 +52,8 @@ int gpgpu_ptx_instruction_classification; void ** g_inst_classification_stat = NULL; -void ** g_inst_mem_classification_stat = NULL; void ** g_inst_op_classification_stat= NULL; int g_ptx_kernel_count = -1; // used for classification stat collection purposes -int g_ptx_kernel_count_prev = -1; // used for classification stat collection purposes int g_debug_execution = 0; int g_debug_thread_uid = 0; addr_t g_debug_pc = 0xBEEF1518; @@ -1243,16 +1241,12 @@ void init_inst_classification_stat() #define MAX_CLASS_KER 1024 char kernelname[MAX_CLASS_KER] =""; if (!g_inst_classification_stat) g_inst_classification_stat = (void**)calloc(MAX_CLASS_KER, sizeof(void*)); - snprintf(kernelname, MAX_CLASS_KER, "Kernel %d INST Classification",g_ptx_kernel_count ); + snprintf(kernelname, MAX_CLASS_KER, "Kernel %d Classification\n",g_ptx_kernel_count ); assert( g_ptx_kernel_count < MAX_CLASS_KER ) ; // a static limit on number of kernels increase it if it fails! g_inst_classification_stat[g_ptx_kernel_count] = StatCreate(kernelname,1,20); - if (!g_inst_mem_classification_stat) g_inst_mem_classification_stat = (void**)calloc(MAX_CLASS_KER, sizeof(void*)); - snprintf(kernelname, MAX_CLASS_KER, "Kernel %d MEM Classification",g_ptx_kernel_count ); - g_inst_mem_classification_stat[g_ptx_kernel_count] = StatCreate(kernelname,1,20); if (!g_inst_op_classification_stat) g_inst_op_classification_stat = (void**)calloc(MAX_CLASS_KER, sizeof(void*)); - snprintf(kernelname, MAX_CLASS_KER, "Kernel %d OP Classification",g_ptx_kernel_count ); + snprintf(kernelname, MAX_CLASS_KER, "Kernel %d OP Classification\n",g_ptx_kernel_count ); g_inst_op_classification_stat[g_ptx_kernel_count] = StatCreate(kernelname,1,100); - g_ptx_kernel_count_prev++; } static unsigned get_tex_datasize( const ptx_instruction *pI, ptx_thread_info *thread ) @@ -1330,17 +1324,6 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) delete pJ; pI = pI_saved; - if ( gpgpu_ptx_instruction_classification ) { - init_inst_classification_stat(); - if (op_classification) { - StatAddSample( g_inst_classification_stat[g_ptx_kernel_count], op_classification); - inst.op_classification = op_classification; - } - if (pI->get_space().get_type()) - StatAddSample( g_inst_mem_classification_stat[g_ptx_kernel_count], ( int )pI->get_space().get_type()); - StatAddSample( g_inst_op_classification_stat[g_ptx_kernel_count], (int) pI->get_opcode() ); - } - // Run exit instruction if exit option included if(pI->is_exit()) exit_impl(pI,this); @@ -1426,6 +1409,27 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) if(!(this->m_functionalSimulationMode)) ptx_file_line_stats_add_exec_count(pI); + if ( gpgpu_ptx_instruction_classification ) { + init_inst_classification_stat(); + unsigned space_type=0; + switch ( pI->get_space().get_type() ) { + case global_space: space_type = 10; break; + case local_space: space_type = 11; break; + case tex_space: space_type = 12; break; + case surf_space: space_type = 13; break; + case param_space_kernel: + case param_space_local: + space_type = 14; break; + case shared_space: space_type = 15; break; + case const_space: space_type = 16; break; + default: + space_type = 0 ; + break; + } + StatAddSample( g_inst_classification_stat[g_ptx_kernel_count], op_classification); + if (space_type) StatAddSample( g_inst_classification_stat[g_ptx_kernel_count], ( int )space_type); + StatAddSample( g_inst_op_classification_stat[g_ptx_kernel_count], (int) pI->get_opcode() ); + } if ( (g_ptx_sim_num_insn % 100000) == 0 ) { dim3 ctaid = get_ctaid(); dim3 tid = get_tid(); @@ -1845,10 +1849,8 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) //******PRINTING******* printf( "GPGPU-Sim: Done functional simulation (%u instructions simulated).\n", g_ptx_sim_num_insn ); - fflush(stdout); if ( gpgpu_ptx_instruction_classification ) { - StatDisp ( g_inst_classification_stat[g_ptx_kernel_count]); - StatDisp ( g_inst_mem_classification_stat[g_ptx_kernel_count]); + StatDisp( g_inst_classification_stat[g_ptx_kernel_count]); StatDisp ( g_inst_op_classification_stat[g_ptx_kernel_count]); } diff --git a/src/cuda-sim/cuda-sim.h b/src/cuda-sim/cuda-sim.h index ef9549f..958daba 100644 --- a/src/cuda-sim/cuda-sim.h +++ b/src/cuda-sim/cuda-sim.h @@ -44,10 +44,8 @@ extern int g_ptx_sim_mode; extern int g_debug_execution; extern int g_debug_thread_uid; extern void ** g_inst_classification_stat; -extern void ** g_inst_mem_classification_stat; extern void ** g_inst_op_classification_stat; extern int g_ptx_kernel_count; // used for classification stat collection purposes -extern int g_ptx_kernel_count_prev; // used for classification stat collection purposes void ptx_opcocde_latency_options (option_parser_t opp); extern class kernel_info_t *gpgpu_opencl_ptx_sim_init_grid(class function_info *entry, diff --git a/src/cuda-sim/opcodes.def b/src/cuda-sim/opcodes.def index ccf64d8..e1b1422 100644 --- a/src/cuda-sim/opcodes.def +++ b/src/cuda-sim/opcodes.def @@ -35,9 +35,7 @@ SFU 4 Mem(except Tex) 5 Tex 6 Nop 7 -Breakpoint 9 Other 10 -Scalar video 11 */ OP_DEF(ABS_OP,abs_impl,"abs",1,1) OP_DEF(ADD_OP,add_impl,"add",1,1) diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index e437c63..08d4525 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -463,7 +463,7 @@ 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 1024 kernels now)", + "if enabled will classify ptx instruction types per kernel (Max 255 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)", @@ -686,7 +686,7 @@ 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, this); + m_memory_partition_unit[i] = new memory_partition_unit(i, m_memory_config, m_memory_stats); 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); @@ -1069,7 +1069,6 @@ 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 ); @@ -1122,9 +1121,8 @@ 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_prev]); - StatDisp( g_inst_mem_classification_stat[g_ptx_kernel_count_prev]); - StatDisp( g_inst_op_classification_stat[g_ptx_kernel_count_prev]); + StatDisp( g_inst_classification_stat[g_ptx_kernel_count]); + StatDisp( g_inst_op_classification_stat[g_ptx_kernel_count]); } #ifdef GPGPUSIM_POWER_MODEL diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index f9b5dad..1778008 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -36,7 +36,6 @@ #include #include #include -#include #include @@ -429,8 +428,6 @@ 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 596e07c..25da107 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -62,16 +62,15 @@ 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, - class gpgpu_sim *gpu) -: m_id(partition_id), m_config(config), m_stats(stats), m_arbitration_metadata(config), m_gpu(gpu) + class memory_stats_t *stats ) +: m_id(partition_id), m_config(config), m_stats(stats), m_arbitration_metadata(config) { 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_gpu); + m_sub_partition[p] = new memory_sub_partition(sub_partition_id, m_config, stats); } } @@ -311,13 +310,11 @@ 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 gpgpu_sim *gpu) + class memory_stats_t *stats ) { 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); @@ -414,15 +411,6 @@ 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 685b1d3..18c0a8b 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 , class gpgpu_sim *gpu); + memory_partition_unit( unsigned partition_id, const struct memory_config *config, class memory_stats_t *stats ); ~memory_partition_unit(); bool busy() const; @@ -93,8 +93,6 @@ public: unsigned get_mpid() const { return m_id; } - gpgpu_sim *m_gpu; - private: unsigned m_id; @@ -147,7 +145,7 @@ private: class memory_sub_partition { public: - 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 sub_partition_id, const struct memory_config *config, class memory_stats_t *stats ); ~memory_sub_partition(); unsigned get_id() const { return m_id; } @@ -188,8 +186,6 @@ public: m_memcpy_cycle_offset += 1; } - gpgpu_sim *m_gpu; - private: // data unsigned m_id; //< the global sub partition ID diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 80ac07e..0e2e1c2 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -438,22 +438,6 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout,"gpgpu_n_mem_texture = %d\n", gpgpu_n_mem_texture); fprintf(fout,"gpgpu_n_mem_const = %d\n", gpgpu_n_mem_const); - fprintf(fout,"gpgpu_mem_divergence_hist "); - gpgpu_mem_divergence_hist->fprint(fout); - fprintf(fout,"\n"); - fprintf(fout,"gpgpu_gmem_ld_divergence_hist "); - gpgpu_gmem_ld_divergence_hist->fprint(fout); - fprintf(fout,"\n"); - fprintf(fout,"gpgpu_gmem_st_divergence_hist "); - gpgpu_gmem_st_divergence_hist->fprint(fout); - fprintf(fout,"\n"); - fprintf(fout,"gpgpu_shmem_divergence_hist "); - gpgpu_shmem_divergence_hist->fprint(fout); - fprintf(fout,"\n"); - fprintf(fout,"warp_inst_classification "); - warp_inst_classification->fprint(fout); - fprintf(fout,"\n"); - fprintf(fout, "gpgpu_n_load_insn = %d\n", gpgpu_n_load_insn); fprintf(fout, "gpgpu_n_store_insn = %d\n", gpgpu_n_store_insn); fprintf(fout, "gpgpu_n_shmem_insn = %d\n", gpgpu_n_shmem_insn); @@ -756,26 +740,9 @@ void shader_core_ctx::fetch() void shader_core_ctx::func_exec_inst( warp_inst_t &inst ) { - unsigned starting_queue_size; execute_warp_inst_t(inst); - if (inst.op_classification) { - m_stats->warp_inst_classification->add2bin(inst.op_classification); - } - if( inst.is_load() || inst.is_store() ) { - starting_queue_size = inst.accessq_count(); + if( inst.is_load() || inst.is_store() ) inst.generate_mem_accesses(); - if ( inst.space.get_type() == global_space ) { - if (inst.is_load()) - m_stats->gpgpu_gmem_ld_divergence_hist->add2bin(inst.accessq_count() - starting_queue_size); - else if (inst.is_store()) - m_stats->gpgpu_gmem_st_divergence_hist->add2bin(inst.accessq_count() - starting_queue_size); - } - else if ( inst.space.get_type() == shared_space ) { - m_stats->gpgpu_shmem_divergence_hist->add2bin(inst.get_cycles()); - } - - m_stats->gpgpu_mem_divergence_hist->add2bin(inst.accessq_count() - starting_queue_size); - } } void shader_core_ctx::issue_warp( register_set& pipe_reg_set, const warp_inst_t* next_inst, const active_mask_t &active_mask, unsigned warp_id ) diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 6a40aee..e07096e 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -53,8 +53,6 @@ #include "stats.h" #include "gpu-cache.h" #include "traffic_breakdown.h" -#include "histogram.h" - @@ -1480,15 +1478,7 @@ struct shader_core_stats_pod { int gpgpu_n_mem_read_global; int gpgpu_n_mem_write_global; int gpgpu_n_mem_read_inst; - - //warps combined memory divergence histogram - linear_histogram* gpgpu_mem_divergence_hist; - linear_histogram* gpgpu_gmem_ld_divergence_hist; - linear_histogram* gpgpu_gmem_st_divergence_hist; - linear_histogram* gpgpu_shmem_divergence_hist; - - linear_histogram* warp_inst_classification; - + int gpgpu_n_mem_l2_writeback; int gpgpu_n_mem_l1_write_allocate; int gpgpu_n_mem_l2_write_allocate; @@ -1557,12 +1547,6 @@ public: m_incoming_traffic_stats = new traffic_breakdown("memtocore"); gpgpu_n_shmem_bank_access = (unsigned *)calloc(config->num_shader(), sizeof(unsigned)); - gpgpu_mem_divergence_hist = new linear_histogram(1, "", config->warp_size+1); - gpgpu_gmem_ld_divergence_hist = new linear_histogram(1, "", config->warp_size+1); - gpgpu_gmem_st_divergence_hist = new linear_histogram(1, "", config->warp_size+1); - gpgpu_shmem_divergence_hist = new linear_histogram(1, "", config->warp_size+1); - - warp_inst_classification = new linear_histogram(1, "", 12); m_shader_dynamic_warp_issue_distro.resize( config->num_shader() ); m_shader_warp_slot_issue_distro.resize( config->num_shader() ); @@ -1577,10 +1561,6 @@ public: free(m_n_diverge); free(shader_cycle_distro); free(last_shader_cycle_distro); - free(gpgpu_mem_divergence_hist); - free(gpgpu_gmem_ld_divergence_hist); - free(gpgpu_gmem_st_divergence_hist); - free(warp_inst_classification); } void new_grid() diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index a18e956..52e2f5e 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -163,8 +163,6 @@ void *gpgpu_sim_thread_concurrent(void*) printf("GPGPU-Sim: ** STOP simulation thread (no work) **\n"); fflush(stdout); } - //g_the_gpu->print_stats(); - if(sim_cycles) { g_the_gpu->print_stats(); g_the_gpu->update_stats(); @@ -272,7 +270,6 @@ 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) ); diff --git a/src/stream_manager.cc b/src/stream_manager.cc index 21115c6..3b6cbd5 100644 --- a/src/stream_manager.cc +++ b/src/stream_manager.cc @@ -230,7 +230,7 @@ bool stream_manager::operation( bool * sim) { bool check=check_finished_kernel(); pthread_mutex_lock(&m_lock); - if(check)m_gpu->print_stats(); +// if(check)m_gpu->print_stats(); stream_operation op =front(); if(!op.do_operation( m_gpu )) //not ready to execute { -- cgit v1.3 From a43799f779a2cf23728659733649506a2d5420df Mon Sep 17 00:00:00 2001 From: tgrogers Date: Tue, 9 Oct 2018 09:55:54 -0400 Subject: Adding in an occupancy metric to match the nvprof metric --- src/gpgpu-sim/gpu-cache.h | 2 +- src/gpgpu-sim/gpu-sim.cc | 20 ++++++++++++++++---- src/gpgpu-sim/gpu-sim.h | 30 ++++++++++++++++++++++++++++++ src/gpgpu-sim/shader.cc | 27 ++++++++++++++++++++++++++- src/gpgpu-sim/shader.h | 3 +++ 5 files changed, 76 insertions(+), 6 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 0b77f47..70d3790 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -414,7 +414,7 @@ struct sector_cache_block : public cache_block_t { } virtual void print_status() { - printf("m_block_addr is %u, status = %u %u %u %u\n", m_block_addr, m_status[0], m_status[1], m_status[2], m_status[3]); + printf("m_block_addr is %llu, status = %u %u %u %u\n", m_block_addr, m_status[0], m_status[1], m_status[2], m_status[3]); } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 08d4525..9e0801c 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -845,6 +845,7 @@ void gpgpu_sim::update_stats() { partiton_replys_in_parallel_total += partiton_replys_in_parallel; partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; gpu_tot_sim_cycle_parition_util += gpu_sim_cycle_parition_util ; + gpu_tot_occupancy += gpu_occupancy; gpu_sim_cycle = 0; partiton_reqs_in_parallel = 0; @@ -853,6 +854,7 @@ void gpgpu_sim::update_stats() { gpu_sim_cycle_parition_util = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; + gpu_occupancy = occupancy_stats(); } void gpgpu_sim::print_stats() @@ -1026,6 +1028,9 @@ void gpgpu_sim::gpu_print_stat() printf("gpu_tot_sim_insn = %lld\n", gpu_tot_sim_insn+gpu_sim_insn); printf("gpu_tot_ipc = %12.4f\n", (float)(gpu_tot_sim_insn+gpu_sim_insn) / (gpu_tot_sim_cycle+gpu_sim_cycle)); printf("gpu_tot_issued_cta = %lld\n", gpu_tot_issued_cta + m_total_cta_launched); + printf("gpu_occupancy = %.4f\% \n", gpu_occupancy.get_occ_fraction() * 100); + printf("gpu_tot_occupancy = %.4f\% \n", (gpu_occupancy + gpu_tot_occupancy).get_occ_fraction() * 100); + extern unsigned long long g_max_total_param_size; fprintf(statfout, "max_total_param_size = %llu\n", g_max_total_param_size); @@ -1529,6 +1534,8 @@ void gpgpu_sim::cycle() // Update core icnt/cache stats for GPUWattch m_cluster[i]->get_icnt_stats(m_power_stats->pwr_mem_stat->n_simt_to_mem[CURRENT_STAT_IDX][i], m_power_stats->pwr_mem_stat->n_mem_to_simt[CURRENT_STAT_IDX][i]); m_cluster[i]->get_cache_stats(m_power_stats->pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX]); + m_cluster[i]->get_current_occupancy(gpu_occupancy.aggregate_warp_slot_filled, gpu_occupancy.aggregate_theoretical_warp_slots); + } float temp=0; for (unsigned i=0;inum_shader();i++){ @@ -1594,15 +1601,20 @@ void gpgpu_sim::cycle() time_t curr_time; time(&curr_time); unsigned long long elapsed_time = MAX(curr_time - g_simulation_starttime, 1); - if ( (elapsed_time - last_liveness_message_time) >= m_config.liveness_message_freq ) { + if ( (elapsed_time - last_liveness_message_time) >= m_config.liveness_message_freq && DTRACE(LIVENESS) ) { days = elapsed_time/(3600*24); hrs = elapsed_time/3600 - 24*days; minutes = elapsed_time/60 - 60*(hrs + 24*days); sec = elapsed_time - 60*(minutes + 60*(hrs + 24*days)); - - DPRINTF(LIVENESS, "GPGPU-Sim uArch: cycles simulated: %lld inst.: %lld (ipc=%4.1f) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", - gpu_tot_sim_cycle + gpu_sim_cycle, gpu_tot_sim_insn + gpu_sim_insn, + + unsigned long long active = 0, total = 0; + for (unsigned i=0;in_simt_clusters;i++) { + m_cluster[i]->get_current_occupancy(active, total); + } + DPRINTF(LIVENESS, "uArch: inst.: %lld (ipc=%4.1f, occ=%0.4f\% [%llu / %llu]) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", + gpu_tot_sim_insn + gpu_sim_insn, (double)gpu_sim_insn/(double)gpu_sim_cycle, + float(active)/float(total) * 100, active, total, (unsigned)((gpu_tot_sim_insn+gpu_sim_insn) / elapsed_time), (unsigned)days,(unsigned)hrs,(unsigned)minutes,(unsigned)sec, ctime(&curr_time)); diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 1778008..1bae1fa 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -383,6 +383,32 @@ private: friend class gpgpu_sim; }; +struct occupancy_stats { + occupancy_stats() : aggregate_warp_slot_filled(0), aggregate_theoretical_warp_slots(0){} + occupancy_stats( unsigned long long wsf, unsigned long long tws ) + : aggregate_warp_slot_filled(wsf), aggregate_theoretical_warp_slots(tws){} + + unsigned long long aggregate_warp_slot_filled; + unsigned long long aggregate_theoretical_warp_slots; + + float get_occ_fraction() const { + return float(aggregate_warp_slot_filled) / float(aggregate_theoretical_warp_slots); + } + + occupancy_stats& operator+=(const occupancy_stats& rhs) { + aggregate_warp_slot_filled += rhs.aggregate_warp_slot_filled; + aggregate_theoretical_warp_slots += rhs.aggregate_theoretical_warp_slots; + return *this; + } + + occupancy_stats operator+(const occupancy_stats& rhs) const{ + return occupancy_stats( aggregate_warp_slot_filled + rhs.aggregate_warp_slot_filled, + aggregate_theoretical_warp_slots + rhs.aggregate_theoretical_warp_slots + ); + } +}; + + class gpgpu_sim : public gpgpu_t { public: gpgpu_sim( const gpgpu_sim_config &config ); @@ -521,6 +547,9 @@ public: unsigned long long gpu_tot_sim_insn; unsigned long long gpu_sim_insn_last_update; unsigned gpu_sim_insn_last_update_sid; + occupancy_stats gpu_occupancy; + occupancy_stats gpu_tot_occupancy; + FuncCache get_cache_config(std::string kernel_name); void set_cache_config(std::string kernel_name, FuncCache cacheConfig ); @@ -549,4 +578,5 @@ public: } }; + #endif diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 0e2e1c2..bb8e4b6 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -74,7 +74,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, shader_core_stats *stats ) : core_t( gpu, NULL, config->warp_size, config->n_thread_per_shader ), m_barriers( this, config->max_warps_per_shader, config->max_cta_per_core, config->max_barriers_per_cta, config->warp_size ), - m_dynamic_warp_id(0) + m_dynamic_warp_id(0), m_active_warps(0) { m_cluster = cluster; m_config = config; @@ -357,6 +357,7 @@ void shader_core_ctx::reinit(unsigned start_thread, unsigned end_thread, bool re m_occupied_ctas = 0; m_occupied_hwtid.reset(); m_occupied_cta_to_hwtid.clear(); + m_active_warps = 0; } for (unsigned i = start_thread; iget_pdom_stack_top_info(pc,rpc); } +float shader_core_ctx::get_current_occupancy( unsigned long long & active, unsigned long long & total ) const +{ + // To match the achieved_occupancy in nvprof, only SMs that are active are counted toward the occupancy. + if ( m_active_warps > 0 ) { + total += m_warp.size(); + active += m_active_warps; + return float(active) / float(total); + } else { + return 0; + } +} + void shader_core_stats::print( FILE* fout ) const { unsigned long long thread_icount_uarch=0; @@ -692,6 +706,8 @@ void shader_core_ctx::fetch() } if( did_exit ) m_warp[warp_id].set_done_exit(); + --m_active_warps; + assert(m_active_warps >= 0); } // this code fetches instructions from the i-cache or generates memory requests @@ -3524,6 +3540,15 @@ void simt_core_cluster::print_not_completed( FILE *fp ) const } } + +float simt_core_cluster::get_current_occupancy( unsigned long long& active, unsigned long long& total ) const { + float aggregate = 0.f; + for( unsigned i=0; i < m_config->n_simt_cores_per_cluster; i++ ) { + aggregate+=m_core[i]->get_current_occupancy( active, total ); + } + return aggregate / m_config->n_simt_cores_per_cluster; +} + unsigned simt_core_cluster::get_n_active_cta() const { unsigned n=0; diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index e07096e..f7b13e2 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1691,6 +1691,7 @@ public: // accessors virtual bool warp_waiting_at_barrier( unsigned warp_id ) const; void get_pdom_stack_top_info( unsigned tid, unsigned *pc, unsigned *rpc ) const; + float get_current_occupancy( unsigned long long & active, unsigned long long & total ) const; // used by pipeline timing model components: // modifiers @@ -1897,6 +1898,7 @@ public: std::vector m_pipeline_reg; Scoreboard *m_scoreboard; opndcoll_rfu_t m_operand_collector; + int m_active_warps; //schedule std::vector schedulers; @@ -1986,6 +1988,7 @@ public: void get_L1T_sub_stats(struct cache_sub_stats &css) const; void get_icnt_stats(long &n_simt_to_mem, long &n_mem_to_simt) const; + float get_current_occupancy( unsigned long long& active, unsigned long long & total ) const; private: unsigned m_cluster_id; -- cgit v1.3 From 0e46a261dfeba9c19d1637f46277986d7eb1b9d8 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Tue, 9 Oct 2018 17:12:57 -0400 Subject: adding adaptive volta cache config --- configs/4.x-cfgs/SM7_TITANV/gpgpusim.config | 19 ++++++++++-------- src/abstract_hardware_model.cc | 2 ++ src/abstract_hardware_model.h | 2 ++ src/gpgpu-sim/gpu-cache.cc | 9 ++++----- src/gpgpu-sim/gpu-cache.h | 31 +++++++++++++++++++++++++---- src/gpgpu-sim/gpu-sim.cc | 3 +++ src/gpgpu-sim/shader.cc | 29 +++++++++++++++++++++++++++ src/gpgpu-sim/shader.h | 2 ++ 8 files changed, 80 insertions(+), 17 deletions(-) (limited to 'src/gpgpu-sim/gpu-sim.cc') diff --git a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config index cc4c931..a7056db 100644 --- a/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/4.x-cfgs/SM7_TITANV/gpgpusim.config @@ -65,19 +65,21 @@ # ::,::::,::,:** # ** Optional parameter - Required when mshr_type==Texture Fifo -# Defualt config is 64KB DL1 and 64KB shared memory --gpgpu_cache:dl1 S:4:128:128,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefL1 S:4:128:192,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_cache:dl1PrefShared S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 --gpgpu_shmem_size 65536 --gpgpu_shmem_size_PrefL1 32768 --gpgpu_shmem_size_PrefShared 98304 +# Defualt config is 32KB DL1 and 96KB shared memory +# In Volta, we assign the remaining shared memory to L1 cache +# if the assigned shd mem = 0, then L1 cache = 128KB +# For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x +# disable this mode in case of multi kernels/apps execution +-adpative_volta_cache_config 1 +-gpgpu_cache:dl1 S:4:128:64,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 98304 -gmem_skip_L1D 0 -icnt_flit_size 40 -gpgpu_n_cluster_ejection_buffer_size 32 -l1_latency 28 -smem_latency 19 -gpgpu_flush_l1_cache 1 +-adpative_volta_cache_config 1 # 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache -gpgpu_cache:dl2 S:64:128:24,L:B:m:L:L,A:384:4,32:0,32 @@ -87,7 +89,8 @@ # 128 KB Inst. -gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 -# 48 KB Tex +# 48 KB Tex +# Note, TEX is deprected in Volta, It is used for legacy apps only. Use L1D cache instead with .nc modifier or __ldg mehtod -gpgpu_tex_cache:l1 N:16:128:24,L:R:m:N:L,T:128:4,128:2 # 64 KB Const -gpgpu_const_cache:l1 N:128:64:8,L:R:f:N:L,S:2:64,4 diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index d2a155c..39ed81c 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -591,6 +591,8 @@ kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info * //Jin: launch latency management m_launch_latency = g_kernel_launch_latency; + + volta_cache_config_set=false; } kernel_info_t::~kernel_info_t() diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index e708fa7..6fa2ba0 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -301,6 +301,8 @@ public: unsigned long long start_cycle; unsigned long long end_cycle; unsigned m_launch_latency; + + mutable bool volta_cache_config_set; }; struct core_config { diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index a11853a..8ae4702 100644 --- a/src/gpgpu-sim/gpu-cache.cc +++ b/src/gpgpu-sim/gpu-cache.cc @@ -29,7 +29,6 @@ #include "stat-tool.h" #include -#define MAX_DEFAULT_CACHE_SIZE_MULTIBLIER 4 // used to allocate memory that is large enough to adapt the changes in cache size across kernels const char * cache_request_status_str(enum cache_request_status status) @@ -165,7 +164,7 @@ unsigned l2_cache_config::set_index(new_addr_type addr) const{ tag_array::~tag_array() { - unsigned cache_lines_num = MAX_DEFAULT_CACHE_SIZE_MULTIBLIER*m_config.get_num_lines(); + unsigned cache_lines_num = m_config.get_max_num_lines(); for(unsigned i=0; iis_modified_line()) { for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; @@ -396,7 +395,7 @@ void tag_array::flush() void tag_array::invalidate() { - for (unsigned i=0; i < m_config.get_num_lines(); i++) + for (unsigned i=0; i < m_config.get_max_num_lines(); i++) for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; } diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h index 4ed382c..9174d7f 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -38,6 +38,8 @@ #include "addrdec.h" #include +#define MAX_DEFAULT_CACHE_SIZE_MULTIBLIER 4 + enum cache_block_state { INVALID=0, RESERVED, @@ -557,15 +559,15 @@ public: } 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 + //we set the MSHRs to be equal to max allocated cache lines. 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 (mimic) 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; + m_mshr_entries = m_nset*m_assoc*MAX_DEFAULT_CACHE_SIZE_MULTIBLIER; if(m_cache_type == SECTOR) m_mshr_entries *= SECTOR_CHUNCK_SIZE; m_mshr_max_merge = MAX_WARP_PER_SM; @@ -581,6 +583,7 @@ public: m_nset_log2 = LOGB2(m_nset); m_valid = true; m_atom_sz = (m_cache_type == SECTOR)? SECTOR_SIZE : m_line_sz; + original_m_assoc = m_assoc; //For more details about difference between FETCH_ON_WRITE and WRITE VALIDAE policies //Read: Jouppi, Norman P. "Cache write policies and performance". ISCA 93. @@ -646,7 +649,11 @@ public: assert( m_valid ); return m_nset * m_assoc; } - + unsigned get_max_num_lines() const + { + assert( m_valid ); + return MAX_DEFAULT_CACHE_SIZE_MULTIBLIER * m_nset * original_m_assoc; + } void print( FILE *fp ) const { fprintf( fp, "Size = %d B (%d Set x %d-way x %d byte line)\n", @@ -687,6 +694,21 @@ public: { return m_mshr_type; } + void set_assoc(unsigned n) + { + //set new assoc. L1 cache dynamically resized in Volta + m_assoc = n; + } + unsigned get_nset() const + { + assert( m_valid ); + return m_nset; + } + unsigned get_total_size_inKB() const + { + assert( m_valid ); + return (m_assoc*m_nset*m_line_sz)/1024; + } FuncCache get_cache_status() {return cache_status;} char *m_config_string; char *m_config_stringPrefL1; @@ -708,6 +730,7 @@ protected: unsigned m_nset_log2; unsigned m_assoc; unsigned m_atom_sz; + unsigned original_m_assoc; enum replacement_policy_t m_replacement_policy; // 'L' = LRU, 'F' = FIFO enum write_policy_t m_write_policy; // 'T' = write through, 'B' = write back, 'R' = read only diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 08d4525..080cbac 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -307,6 +307,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shmem_size", OPT_UINT32, &gpgpu_shmem_size, "Size of shared memory per shader core (default 16kB)", "16384"); + option_parser_register(opp, "-adpative_volta_cache_config", OPT_BOOL, &adpative_volta_cache_config, + "adpative_volta_cache_config", + "0"); option_parser_register(opp, "-gpgpu_shmem_size", OPT_UINT32, &gpgpu_shmem_sizeDefault, "Size of shared memory per shader core (default 16kB)", "16384"); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 0e2e1c2..2cc8a2e 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -2744,6 +2744,35 @@ unsigned int shader_core_config::max_cta( const kernel_info_t &k ) const abort(); } + if(adpative_volta_cache_config && !k.volta_cache_config_set) { + //For Volta, we assign the remaining shared memory to L1 cache + //For more info, see https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#shared-memory-7-x + unsigned total_shmed = kernel_info->smem * result; + assert(total_shmed >=0 && total_shmed <= gpgpu_shmem_size); + assert(gpgpu_shmem_size == 98304); //Volta has 96 KB shared + assert(m_L1D_config.get_nset() == 4); //Volta L1 has four sets + if(total_shmed < gpgpu_shmem_size){ + if(total_shmed == 0) + m_L1D_config.set_assoc(256); //L1 is 128KB ans shd=0 + else if(total_shmed > 0 && total_shmed <= 8192) + m_L1D_config.set_assoc(240); //L1 is 120KB ans shd=8KB + else if(total_shmed > 8192 && total_shmed <= 16384) + m_L1D_config.set_assoc(224); //L1 is 112KB ans shd=16KB + else if(total_shmed > 16384 && total_shmed <= 32768) + m_L1D_config.set_assoc(192); //L1 is 96KB ans shd=32KB + else if(total_shmed > 32768 && total_shmed <= 65536) + m_L1D_config.set_assoc(128); //L1 is 64KB ans shd=64KB + else if(total_shmed > 65536 && total_shmed <= gpgpu_shmem_size) + m_L1D_config.set_assoc(64); //L1 is 32KB and shd=96KB + else + assert(0); + + printf ("GPGPU-Sim: Reconfigure L1 cache in Volta Archi to %uKB\n", m_L1D_config.get_total_size_inKB()); + } + + k.volta_cache_config_set = true; + } + return result; } diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index e07096e..92b4159 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1409,6 +1409,8 @@ struct shader_core_config : public core_config //Jin: concurrent kernel on sm bool gpgpu_concurrent_kernel_sm; + + bool adpative_volta_cache_config; }; struct shader_core_stats_pod { -- cgit v1.3