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/abstract_hardware_model.h | 18 +++++++++++++----- 1 file changed, 13 insertions(+), 5 deletions(-) (limited to 'src/abstract_hardware_model.h') diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index aaa4b00..cd185a1 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -332,12 +332,15 @@ struct core_config { unsigned gpgpu_shmem_sizeDefault; unsigned gpgpu_shmem_sizePrefL1; unsigned gpgpu_shmem_sizePrefShared; + unsigned L1_warp_parts_cached; + unsigned L1_warp_parts_non_cached; // texture and constant cache line sizes (used to determine number of memory accesses) unsigned gpgpu_cache_texl1_linesize; unsigned gpgpu_cache_constl1_linesize; unsigned gpgpu_max_insn_issue_per_warp; + bool gmem_skip_L1D; // on = global memory access always skip the L1 cache }; // bounded stack that implements simt reconvergence using pdom mechanism from MICRO'07 paper @@ -615,6 +618,8 @@ private: const unsigned MAX_MEMORY_ACCESS_SIZE = 128; typedef std::bitset mem_access_byte_mask_t; +const unsigned SECTOR_CHUNCK_SIZE = 4; +typedef std::bitset mem_access_sector_mask_t; #define NO_PARTIAL_WRITE (mem_access_byte_mask_t()) #define MEM_ACCESS_TYPE_TUP_DEF \ @@ -679,8 +684,9 @@ public: unsigned size, bool wr, const active_mask_t &active_mask, - const mem_access_byte_mask_t &byte_mask ) - : m_warp_mask(active_mask), m_byte_mask(byte_mask) + const mem_access_byte_mask_t &byte_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(); m_type = type; @@ -696,6 +702,7 @@ public: bool is_write() const { return m_write; } enum mem_access_type get_type() const { return m_type; } mem_access_byte_mask_t get_byte_mask() const { return m_byte_mask; } + mem_access_sector_mask_t get_sector_mask() const { return m_sector_mask; } void print(FILE *fp) const { @@ -729,6 +736,7 @@ private: mem_access_type m_type; active_mask_t m_warp_mask; mem_access_byte_mask_t m_byte_mask; + mem_access_sector_mask_t m_sector_mask; static unsigned sm_next_access_uid; }; @@ -938,9 +946,9 @@ public: }; void generate_mem_accesses(); - void memory_coalescing_arch_13( bool is_write, mem_access_type access_type ); - void memory_coalescing_arch_13_atomic( bool is_write, mem_access_type access_type ); - void memory_coalescing_arch_13_reduce_and_send( bool is_write, mem_access_type access_type, const transaction_info &info, new_addr_type addr, unsigned segment_size ); + void memory_coalescing_arch( bool is_write, mem_access_type access_type ); + void memory_coalescing_arch_atomic( bool is_write, mem_access_type access_type ); + void memory_coalescing_arch_reduce_and_send( bool is_write, mem_access_type access_type, const transaction_info &info, new_addr_type addr, unsigned segment_size ); void add_callback( unsigned lane_id, void (*function)(const class inst_t*, class ptx_thread_info*), -- 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/abstract_hardware_model.h') 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 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/abstract_hardware_model.h') 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 b4cf66026f66291293c6e11e9c07976f773732fa Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Tue, 3 Oct 2017 11:36:52 -0400 Subject: Add lonestar tick support Signed-off-by: Mengchi Zhang --- cuobjdump_to_ptxplus/ptx_parser.h | 4 ++++ libcuda/cuda_runtime_api.cc | 49 +++++++++++++++++++++++++-------------- src/abstract_hardware_model.h | 10 +++++++- src/cuda-sim/cuda-sim.cc | 5 ++-- src/cuda-sim/ptx.y | 3 ++- src/cuda-sim/ptx_ir.cc | 1 + src/cuda-sim/ptx_ir.h | 7 ++++++ src/cuda-sim/ptx_parser.cc | 4 ++++ src/cuda-sim/ptx_parser.h | 1 + 9 files changed, 62 insertions(+), 22 deletions(-) (limited to 'src/abstract_hardware_model.h') diff --git a/cuobjdump_to_ptxplus/ptx_parser.h b/cuobjdump_to_ptxplus/ptx_parser.h index 418a733..729eaec 100644 --- a/cuobjdump_to_ptxplus/ptx_parser.h +++ b/cuobjdump_to_ptxplus/ptx_parser.h @@ -373,4 +373,8 @@ void func_header_info_int(const char* s, int i) g_headerList->getListEnd().addOperand(buff); } } + +void maxnt_id(int x, int y, int z) { + +} #endif //_PTX_PARSER_H_ diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index d8e7832..4f5e0fb 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -336,23 +336,19 @@ class _cuda_device_id *GPGPUSim_Init() prop->minor = 2; prop->totalGlobalMem = 0x80000000 /* 2 GB */; prop->memPitch = 0; + if(prop->major >= 2) { + prop->maxThreadsPerBlock = 1024; + prop->maxThreadsDim[0] = 1024; + prop->maxThreadsDim[1] = 1024; + } + else + { + prop->maxThreadsPerBlock = 512; + prop->maxThreadsDim[0] = 512; + prop->maxThreadsDim[1] = 512; + } - if(prop->major >= 2) { - prop->maxThreadsPerBlock = 1024; - prop->maxThreadsDim[0] = 1024; - prop->maxThreadsDim[1] = 1024; - } - else - { - prop->maxThreadsPerBlock = 512; - prop->maxThreadsDim[0] = 512; - prop->maxThreadsDim[1] = 512; - } - - prop->maxThreadsDim[2] = 64; - prop->maxGridSize[0] = 0x40000000; - prop->maxGridSize[1] = 0x40000000; - prop->maxGridSize[2] = 0x40000000; + prop->maxThreadsDim[2] = 64; prop->maxGridSize[0] = 0x40000000; prop->maxGridSize[1] = 0x40000000; prop->maxGridSize[2] = 0x40000000; @@ -366,7 +362,7 @@ class _cuda_device_id *GPGPUSim_Init() prop->multiProcessorCount = the_gpu->get_config().num_shader(); #endif #if (CUDART_VERSION >= 4000) - prop->maxThreadsPerMultiProcessor = the_gpu->threads_per_core(); + prop->maxThreadsPerMultiProcessor = the_gpu->threads_per_core(); #endif the_gpu->set_prop(prop); the_device = new _cuda_device_id(the_gpu); @@ -2112,6 +2108,20 @@ cudaError_t CUDARTAPI cudaSetDeviceFlags( int flags ) return g_last_cudaError = cudaErrorUnknown; } +size_t getMaxThreadsPerBlock(struct cudaFuncAttributes *attr) { + _cuda_device_id *dev = GPGPUSim_Init(); + struct cudaDeviceProp prop; + + prop = *dev->get_prop(); + + size_t max = prop.maxThreadsPerBlock; + + if ((prop.regsPerBlock / attr->numRegs) < max) + max = prop.regsPerBlock / attr->numRegs; + + return max; +} + cudaError_t CUDARTAPI cudaFuncGetAttributes(struct cudaFuncAttributes *attr, const char *hostFun ) { CUctx_st *context = GPGPUSim_Context(); @@ -2122,7 +2132,10 @@ cudaError_t CUDARTAPI cudaFuncGetAttributes(struct cudaFuncAttributes *attr, con attr->constSizeBytes = kinfo->cmem; attr->localSizeBytes = kinfo->lmem; attr->numRegs = kinfo->regs; - attr->maxThreadsPerBlock = 0; // from pragmas? + if(kinfo->maxthreads > 0) + attr->maxThreadsPerBlock = kinfo->maxthreads; + else + attr->maxThreadsPerBlock = getMaxThreadsPerBlock(attr); #if CUDART_VERSION >= 3000 attr->ptxVersion = kinfo->ptx_version; attr->binaryVersion = kinfo->sm_target; diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index cec75f9..1b764e2 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -517,7 +517,14 @@ public: const struct textureReference* get_texref(const std::string &texname) const { std::map::const_iterator t=m_NameToTextureRef.find(texname); - assert( t != m_NameToTextureRef.end() ); + if( t == m_NameToTextureRef.end() ) { + // search for :: prefixed names + std::string temp("::" + texname); + t=m_NameToTextureRef.find(temp); + } + + assert(t != m_NameToTextureRef.end()); + return t->second; } const struct cudaArray* get_texarray( const struct textureReference *texref ) const @@ -568,6 +575,7 @@ struct gpgpu_ptx_sim_info int cmem; int gmem; int regs; + unsigned maxthreads; unsigned ptx_version; unsigned sm_target; }; diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 2f166aa..f143d35 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1159,13 +1159,14 @@ void function_info::finalize( memory_space *param_mem ) // copy the parameter over word-by-word so that parameter that crosses a memory page can be copied over //Jin: copy parameter using aligned rules const size_t word_size = 4; - param_address = (param_address + size - 1) / size * size; //aligned with size + //param_address = (param_address + size - 1) / size * size; //aligned with size TODO: align not correct for (size_t idx = 0; idx < size; idx += word_size) { const char *pdata = reinterpret_cast(param_value.pdata) + idx; // cast to char * for ptr arithmetic param_mem->write(param_address + idx, word_size, pdata,NULL,NULL); + param_mem->read(param_address + idx, word_size, &data); } unsigned offset = p.get_offset(); - assert(offset == param_address); + //assert(offset == param_address); param->set_address(param_address); param_address += size; } diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y index e00aa4b..c0c58a6 100644 --- a/src/cuda-sim/ptx.y +++ b/src/cuda-sim/ptx.y @@ -228,7 +228,8 @@ function_defn: function_decl { set_symtab($1); func_header(".skip"); } statement block_spec: MAXNTID_DIRECTIVE INT_OPERAND COMMA INT_OPERAND COMMA INT_OPERAND {func_header_info_int(".maxntid", $2); func_header_info_int(",", $4); - func_header_info_int(",", $6); } + func_header_info_int(",", $6); + maxnt_id($2, $4, $6);} | MINNCTAPERSM_DIRECTIVE INT_OPERAND { func_header_info_int(".minnctapersm", $2); printf("GPGPU-Sim: Warning: .minnctapersm ignored. \n"); } | MAXNCTAPERSM_DIRECTIVE INT_OPERAND { func_header_info_int(".maxnctapersm", $2); printf("GPGPU-Sim: Warning: .maxnctapersm ignored. \n"); } ; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 8ebdcf8..ee36957 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -222,6 +222,7 @@ bool symbol_table::add_function_decl( const char *name, int entry_point, functio } else { *func_info = new function_info(entry_point); (*func_info)->set_name(name); + (*func_info)->set_maxnt_id(0); m_function_info_lookup[key] = *func_info; } diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 9ad1571..36ef3d5 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1245,6 +1245,7 @@ public: const struct gpgpu_ptx_sim_info* get_kernel_info () const { + assert (m_kernel_info.maxthreads == maxnt_id); return &m_kernel_info; } @@ -1252,6 +1253,8 @@ public: m_kernel_info = info; m_kernel_info.ptx_version = 10*get_ptx_version().ver(); m_kernel_info.sm_target = get_ptx_version().target(); + // THIS DEPENDS ON ptxas being called after the PTX is parsed. + m_kernel_info.maxthreads = maxnt_id; } symbol_table *get_symtab() { @@ -1275,7 +1278,11 @@ public: } bool is_entry_point() const { return m_entry_point; } + void set_maxnt_id(unsigned maxthreads) { maxnt_id = maxthreads;} + unsigned get_maxnt_id() { return maxnt_id;} + private: + unsigned maxnt_id; unsigned m_uid; unsigned m_local_mem_framesize; bool m_entry_point; diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index baa3bcd..e5731a8 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -969,6 +969,10 @@ void target_header3(char* a, char* b, char* c) g_global_symbol_table->set_sm_target(a,b,c); } +void maxnt_id(int x, int y, int z) { + g_func_info->set_maxnt_id(x * y * z); +} + void func_header(const char* a) {} //intentional dummy function void func_header_info(const char* a) {} //intentional dummy function void func_header_info_int(const char* a, int b) {} //intentional dummy function diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h index 32f3903..13042e1 100644 --- a/src/cuda-sim/ptx_parser.h +++ b/src/cuda-sim/ptx_parser.h @@ -93,6 +93,7 @@ void change_double_operand_type( int addr_type ); void change_operand_neg( ); void set_immediate_operand_type( ); void version_header(double a); +void maxnt_id(int x, int y, int z); //Jin: handle instructino group for cdp void start_inst_group(); -- cgit v1.3 From 3082f63e86be2e115e4a071069dda8fe9452d366 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 6 Sep 2018 16:33:18 -0400 Subject: Add .nc option Signed-off-by: Mengchi Zhang --- src/abstract_hardware_model.h | 1 + src/cuda-sim/cuda-sim.cc | 1 + src/cuda-sim/ptx_ir.cc | 1 + src/gpgpu-sim/shader.cc | 2 +- 4 files changed, 4 insertions(+), 1 deletion(-) (limited to 'src/abstract_hardware_model.h') diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 1b764e2..6df7b89 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -664,6 +664,7 @@ enum cache_operator_type { CACHE_ALL, // .ca CACHE_LAST_USE, // .lu CACHE_VOLATILE, // .cv + CACHE_L1, // .nc // loads and stores CACHE_STREAMING, // .cs diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 9246613..93bbc1d 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -898,6 +898,7 @@ void ptx_instruction::pre_decode() switch( m_cache_option ) { case CA_OPTION: cache_op = CACHE_ALL; break; + case NC_OPTION: cache_op = CACHE_L1; break; case CG_OPTION: cache_op = CACHE_GLOBAL; break; case CS_OPTION: cache_op = CACHE_STREAMING; break; case LU_OPTION: cache_op = CACHE_LAST_USE; break; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index ee36957..ba6d7ed 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -1211,6 +1211,7 @@ ptx_instruction::ptx_instruction( int opcode, case EXTP_OPTION: break; case NC_OPTION: + m_cache_option = last_ptx_inst_option; break; case UP_OPTION: case DOWN_OPTION: diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index d2f40a1..533c6f9 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1537,7 +1537,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea bypassL1D = true; } else if (inst.space.is_global()) { // global memory access // skip L1 cache if the option is enabled - if (m_core->get_config()->gmem_skip_L1D) + if (m_core->get_config()->gmem_skip_L1D && (CACHE_L1 != inst.cache_op)) bypassL1D = true; } -- 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/abstract_hardware_model.h') 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 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/abstract_hardware_model.h') 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