diff options
| -rwxr-xr-x | cuda-kernels/gpgpusim.config | 7 | ||||
| -rw-r--r-- | src/abstract_hardware_model.h | 2 | ||||
| -rw-r--r-- | src/cuda-sim/cuda-sim.cc | 5 | ||||
| -rw-r--r-- | src/gpgpu-sim/gpu-sim.cc | 16 | ||||
| -rw-r--r-- | src/gpgpu-sim/scoreboard.cc | 40 | ||||
| -rw-r--r-- | src/gpgpu-sim/shader.cc | 70 | ||||
| -rw-r--r-- | src/gpgpu-sim/shader.h | 47 |
7 files changed, 163 insertions, 24 deletions
diff --git a/cuda-kernels/gpgpusim.config b/cuda-kernels/gpgpusim.config index 306d7f9..69a110f 100755 --- a/cuda-kernels/gpgpusim.config +++ b/cuda-kernels/gpgpusim.config @@ -33,10 +33,10 @@ # 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_pipeline_widths 4,1,1,1,4,1,1,1,,6 -gpgpu_num_sp_units 4 -gpgpu_num_sfu_units 1 - +-gpgpu_num_tensor_core_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 @@ -72,11 +72,14 @@ ## 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_tensor_core 24 -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_tensor_core 1 +-gpgpu_operand_collector_num_out_ports_tensor_core 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! diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index e00c941..e2e116e 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -77,6 +77,7 @@ enum uarch_op_t { NO_OP=-1, ALU_OP=1, SFU_OP, + TENSOR_CORE_OP, ALU_SFU_OP, LOAD_OP, STORE_OP, @@ -133,6 +134,7 @@ enum operation_pipeline_t { UNKOWN_OP, SP__OP, SFU__OP, + TENSOR_CORE__OP, MEM__OP }; typedef enum operation_pipeline_t operation_pipeline; diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 7552acf..6da0840 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -520,7 +520,7 @@ void ptx_instruction::set_mul_div_or_other_archop(){ sp_op=FP_EXP_OP; break; default: - if(op==ALU_OP) + if((op==ALU_OP)||(op==TENSOR_CORE_OP)) sp_op=FP__OP; break; @@ -542,7 +542,7 @@ void ptx_instruction::set_mul_div_or_other_archop(){ sp_op=INT_DIV_OP; break; default: - if(op==ALU_OP) + if((op==ALU_OP)||(op==TENSOR_CORE_OP)) sp_op=INT__OP; break; } @@ -795,6 +795,7 @@ void ptx_instruction::set_opcode_and_latency() case MMA_OP: latency = 64; initiation_interval = 64; + op=TENSOR_CORE_OP; break; case SHFL_OP: latency = 32; diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 3bd1892..cc23051 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -306,6 +306,9 @@ void shader_core_config::reg_options(class OptionParser * opp) 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"); + option_parser_register(opp, "-gpgpu_operand_collector_num_units_tensor_core", OPT_INT32, &gpgpu_operand_collector_num_units_tensor_core, + "number of collector units (default = 4)", + "4"); option_parser_register(opp, "-gpgpu_operand_collector_num_units_mem", OPT_INT32, &gpgpu_operand_collector_num_units_mem, "number of collector units (default = 2)", "2"); @@ -318,6 +321,9 @@ void shader_core_config::reg_options(class OptionParser * opp) 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"); + option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_tensor_core", OPT_INT32, &gpgpu_operand_collector_num_in_ports_tensor_core, + "number of collector unit in ports (default = 1)", + "1"); option_parser_register(opp, "-gpgpu_operand_collector_num_in_ports_mem", OPT_INT32, &gpgpu_operand_collector_num_in_ports_mem, "number of collector unit in ports (default = 1)", "1"); @@ -330,6 +336,9 @@ void shader_core_config::reg_options(class OptionParser * opp) 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"); + option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_tensor_core", OPT_INT32, &gpgpu_operand_collector_num_out_ports_tensor_core, + "number of collector unit in ports (default = 1)", + "1"); option_parser_register(opp, "-gpgpu_operand_collector_num_out_ports_mem", OPT_INT32, &gpgpu_operand_collector_num_out_ports_mem, "number of collector unit in ports (default = 1)", "1"); @@ -350,14 +359,17 @@ void shader_core_config::reg_options(class OptionParser * opp) "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_SFU,ID_OC_TENSOR_CORE,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_TENSOR_CORE,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"); option_parser_register(opp, "-gpgpu_num_sfu_units", OPT_INT32, &gpgpu_num_sfu_units, "Number of SF units (default=1)", "1"); + option_parser_register(opp, "-gpgpu_num_tensor_core_units", OPT_INT32, &gpgpu_num_tensor_core_units, + "Number of tensor_core units (default=1)", + "1"); option_parser_register(opp, "-gpgpu_num_mem_units", OPT_INT32, &gpgpu_num_mem_units, "Number if ldst units (default=1) WARNING: not hooked up to anything", "1"); diff --git a/src/gpgpu-sim/scoreboard.cc b/src/gpgpu-sim/scoreboard.cc index b538fdf..4d1b43a 100644 --- a/src/gpgpu-sim/scoreboard.cc +++ b/src/gpgpu-sim/scoreboard.cc @@ -82,7 +82,7 @@ const bool Scoreboard::islongop (unsigned warp_id,unsigned regnum) { void Scoreboard::reserveRegisters(const class warp_inst_t* inst) { - for( unsigned r=0; r < 4; r++) { + for( unsigned r=0; r < 8; r++) { if(inst->out[r] > 0) { reserveRegister(inst->warp_id(), inst->out[r]); SHADER_DPRINTF( SCOREBOARD, @@ -100,7 +100,7 @@ void Scoreboard::reserveRegisters(const class warp_inst_t* inst) inst->space.get_type() == param_space_local || inst->space.get_type() == param_space_unclassified || inst->space.get_type() == tex_space)){ - for ( unsigned r=0; r<4; r++) { + for ( unsigned r=0; r<8; r++) { if(inst->out[r] > 0) { SHADER_DPRINTF( SCOREBOARD, "New longopreg marked - warp:%d, reg: %d\n", @@ -115,7 +115,7 @@ void Scoreboard::reserveRegisters(const class warp_inst_t* inst) // Release registers for an instruction void Scoreboard::releaseRegisters(const class warp_inst_t *inst) { - for( unsigned r=0; r < 4; r++) { + for( unsigned r=0; r < 8; r++) { if(inst->out[r] > 0) { SHADER_DPRINTF( SCOREBOARD, "Register Released - warp:%d, reg: %d\n", @@ -142,15 +142,37 @@ bool Scoreboard::checkCollision( unsigned wid, const class inst_t *inst ) const if(inst->out[1] > 0) inst_regs.insert(inst->out[1]); if(inst->out[2] > 0) inst_regs.insert(inst->out[2]); if(inst->out[3] > 0) inst_regs.insert(inst->out[3]); - if(inst->in[0] > 0) inst_regs.insert(inst->in[0]); + if(inst->out[4] > 0) inst_regs.insert(inst->out[4]); + if(inst->out[5] > 0) inst_regs.insert(inst->out[5]); + if(inst->out[6] > 0) inst_regs.insert(inst->out[6]); + if(inst->out[7] > 0) inst_regs.insert(inst->out[7]); + + if(inst->in[0] > 0) inst_regs.insert(inst->in[0]); if(inst->in[1] > 0) inst_regs.insert(inst->in[1]); if(inst->in[2] > 0) inst_regs.insert(inst->in[2]); if(inst->in[3] > 0) inst_regs.insert(inst->in[3]); - if(inst->in[3] > 0) inst_regs.insert(inst->in[4]); - if(inst->in[3] > 0) inst_regs.insert(inst->in[5]); - if(inst->in[3] > 0) inst_regs.insert(inst->in[6]); - if(inst->in[3] > 0) inst_regs.insert(inst->in[7]); - if(inst->pred > 0) inst_regs.insert(inst->pred); + if(inst->in[4] > 0) inst_regs.insert(inst->in[4]); + if(inst->in[5] > 0) inst_regs.insert(inst->in[5]); + if(inst->in[6] > 0) inst_regs.insert(inst->in[6]); + if(inst->in[7] > 0) inst_regs.insert(inst->in[7]); + if(inst->in[8] > 0) inst_regs.insert(inst->in[8]); + if(inst->in[9] > 0) inst_regs.insert(inst->in[9]); + if(inst->in[10] > 0) inst_regs.insert(inst->in[10]); + if(inst->in[11] > 0) inst_regs.insert(inst->in[11]); + if(inst->in[12] > 0) inst_regs.insert(inst->in[12]); + if(inst->in[13] > 0) inst_regs.insert(inst->in[13]); + if(inst->in[14] > 0) inst_regs.insert(inst->in[14]); + if(inst->in[15] > 0) inst_regs.insert(inst->in[15]); + if(inst->in[16] > 0) inst_regs.insert(inst->in[16]); + if(inst->in[17] > 0) inst_regs.insert(inst->in[17]); + if(inst->in[18] > 0) inst_regs.insert(inst->in[18]); + if(inst->in[19] > 0) inst_regs.insert(inst->in[19]); + if(inst->in[20] > 0) inst_regs.insert(inst->in[20]); + if(inst->in[21] > 0) inst_regs.insert(inst->in[21]); + if(inst->in[22] > 0) inst_regs.insert(inst->in[22]); + if(inst->in[23] > 0) inst_regs.insert(inst->in[23]); + + if(inst->pred > 0) inst_regs.insert(inst->pred); if(inst->ar1 > 0) inst_regs.insert(inst->ar1); if(inst->ar2 > 0) inst_regs.insert(inst->ar2); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 92cdb5b..fcac755 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -148,6 +148,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_warp, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], i ) @@ -162,6 +163,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_warp, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], i, config->gpgpu_scheduler_string @@ -177,6 +179,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_warp, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], i ) @@ -191,6 +194,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_warp, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], i, config->gpgpu_scheduler_string @@ -211,9 +215,10 @@ 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, SFU_CUS, TENSOR_CORE_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(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(TENSOR_CORE_CUS, m_config->gpgpu_operand_collector_num_units_tensor_core, m_config->gpgpu_operand_collector_num_out_ports_tensor_core); 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); @@ -237,7 +242,17 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, 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_tensor_core; i++) { + in_ports.push_back(&m_pipeline_reg[ID_OC_TENSOR_CORE]); + out_ports.push_back(&m_pipeline_reg[OC_EX_TENSOR_CORE]); + cu_sets.push_back((unsigned)TENSOR_CORE_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_mem; i++) { in_ports.push_back(&m_pipeline_reg[ID_OC_MEM]); out_ports.push_back(&m_pipeline_reg[OC_EX_MEM]); @@ -251,9 +266,11 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, for (unsigned i = 0; i < m_config->gpgpu_operand_collector_num_in_ports_gen; i++) { in_ports.push_back(&m_pipeline_reg[ID_OC_SP]); in_ports.push_back(&m_pipeline_reg[ID_OC_SFU]); + in_ports.push_back(&m_pipeline_reg[ID_OC_TENSOR_CORE]); in_ports.push_back(&m_pipeline_reg[ID_OC_MEM]); out_ports.push_back(&m_pipeline_reg[OC_EX_SP]); out_ports.push_back(&m_pipeline_reg[OC_EX_SFU]); + out_ports.push_back(&m_pipeline_reg[OC_EX_TENSOR_CORE]); out_ports.push_back(&m_pipeline_reg[OC_EX_MEM]); cu_sets.push_back((unsigned)GEN_CUS); m_operand_collector.add_port(in_ports,out_ports,cu_sets); @@ -263,7 +280,7 @@ 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_sfu_units + m_config->gpgpu_num_tensor_core_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 ]; @@ -280,6 +297,12 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_dispatch_port.push_back(ID_OC_SFU); m_issue_port.push_back(OC_EX_SFU); } + + for (int k = 0; k < m_config->gpgpu_num_tensor_core_units; k++) { + m_fu.push_back(new tensor_core( &m_pipeline_reg[EX_WB], m_config, this )); + m_dispatch_port.push_back(ID_OC_TENSOR_CORE); + m_issue_port.push_back(OC_EX_TENSOR_CORE); + } 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); @@ -886,7 +909,8 @@ void scheduler_unit::cycle() } 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) ) { + bool tensor_core_pipe_avail = m_tensor_core_out->has_free(); + if( sp_pipe_avail && (pI->op != SFU_OP) && (pI->op != TENSOR_CORE_OP)) { //Jin: special for CDP api if(pI->m_is_cdp && !warp(warp_id).m_cdp_dummy) { @@ -918,6 +942,14 @@ void scheduler_unit::cycle() issued_inst=true; warp_inst_issued = true; } + } + else if ( (pI->op == TENSOR_CORE_OP) ) { + if( tensor_core_pipe_avail ) { + m_shader->issue_warp(*m_tensor_core_out,pI,active_mask,warp_id); + issued++; + issued_inst=true; + warp_inst_issued = true; + } } } } else { SCHED_DPRINTF( "Warp (warp_id %u, dynamic_warp_id %u) fails scoreboard\n", @@ -1083,10 +1115,11 @@ swl_scheduler::swl_scheduler ( shader_core_stats* stats, shader_core_ctx* shader std::vector<shd_warp_t>* warp, register_set* sp_out, register_set* sfu_out, + register_set* tensor_core_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, sfu_out, tensor_core_out, mem_out, id ) { unsigned m_prioritization_readin; int ret = sscanf( config_string, @@ -1509,6 +1542,13 @@ sfu:: sfu( register_set* result_port, const shader_core_config *config,shader_c m_name = "SFU"; } +tensor_core:: tensor_core( register_set* result_port, const shader_core_config *config,shader_core_ctx *core ) + : pipelined_simd_unit(result_port,config,config->max_tensor_core_latency,core) +{ + m_name = "TENSOR_CORE"; +} + + void sfu::issue( register_set& source_reg ) { warp_inst_t** ready_reg = source_reg.get_ready(); @@ -1519,6 +1559,17 @@ void sfu::issue( register_set& source_reg ) pipelined_simd_unit::issue(source_reg); } +void tensor_core::issue( register_set& source_reg ) +{ + warp_inst_t** ready_reg = source_reg.get_ready(); + //m_core->incexecstat((*ready_reg)); + + (*ready_reg)->op_pipe= TENSOR_CORE__OP; + m_core->incsfu_stat(m_core->get_config()->warp_size,(*ready_reg)->latency); + pipelined_simd_unit::issue(source_reg); +} + + void ldst_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); @@ -1540,6 +1591,15 @@ void sfu::active_lanes_in_pipeline(){ m_core->incfumemactivelanes_stat(active_count); } +void tensor_core::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->incsfuactivelanes_stat(active_count); + m_core->incfuactivelanes_stat(active_count); + m_core->incfumemactivelanes_stat(active_count); +} + + sp_unit::sp_unit( register_set* result_port, const shader_core_config *config,shader_core_ctx *core) : pipelined_simd_unit(result_port,config,config->max_sp_latency,core) { @@ -2237,7 +2297,7 @@ void shader_core_ctx::incexecstat(warp_inst_t *&inst) switch(inst->sp_op){ case INT__OP: - incialu_stat(inst->active_count(),25); + incialu_stat(inst->active_count(),32); break; case INT_MUL_OP: incimul_stat(inst->active_count(),7.2); diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index db2af01..2c4c43d 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -318,11 +318,12 @@ public: std::vector<shd_warp_t>* warp, register_set* sp_out, register_set* sfu_out, + register_set* tensor_core_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_sfu_out(sfu_out),m_tensor_core_out(tensor_core_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)); @@ -395,6 +396,7 @@ protected: std::vector<shd_warp_t>* m_warp; register_set* m_sp_out; register_set* m_sfu_out; + register_set* m_tensor_core_out; register_set* m_mem_out; int m_id; @@ -407,9 +409,10 @@ public: std::vector<shd_warp_t>* warp, register_set* sp_out, register_set* sfu_out, + register_set* tensor_core_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, sfu_out, tensor_core_out, mem_out, id ){} virtual ~lrr_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -424,9 +427,10 @@ public: std::vector<shd_warp_t>* warp, register_set* sp_out, register_set* sfu_out, + register_set* tensor_core_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, sfu_out, tensor_core_out, mem_out, id ){} virtual ~gto_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -443,10 +447,11 @@ public: std::vector<shd_warp_t>* warp, register_set* sp_out, register_set* sfu_out, + register_set* tensor_core_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, sfu_out, tensor_core_out, mem_out, id ), m_pending_warps() { unsigned inner_level_readin; @@ -493,6 +498,7 @@ public: std::vector<shd_warp_t>* warp, register_set* sp_out, register_set* sfu_out, + register_set* tensor_core_out, register_set* mem_out, int id, char* config_string ); @@ -1062,6 +1068,23 @@ public: virtual void issue( register_set& source_reg ); }; +class tensor_core : public pipelined_simd_unit +{ +public: + tensor_core( 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 TENSOR_CORE_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: @@ -1201,9 +1224,11 @@ protected: enum pipeline_stage_name_t { ID_OC_SP=0, ID_OC_SFU, + ID_OC_TENSOR_CORE, ID_OC_MEM, OC_EX_SP, OC_EX_SFU, + OC_EX_TENSOR_CORE, OC_EX_MEM, EX_WB, N_PIPELINE_STAGES @@ -1212,9 +1237,11 @@ enum pipeline_stage_name_t { const char* const pipeline_stage_name_decode[] = { "ID_OC_SP", "ID_OC_SFU", + "ID_OC_TENSOR_CORE", "ID_OC_MEM", "OC_EX_SP", "OC_EX_SFU", + "OC_EX_TENSOR_CORE", "OC_EX_MEM", "EX_WB", "N_PIPELINE_STAGES" @@ -1257,6 +1284,7 @@ struct shader_core_config : public core_config max_warps_per_shader = n_thread_per_shader/warp_size; assert( !(n_thread_per_shader % warp_size) ); max_sfu_latency = 512; + max_tensor_core_latency = 512; max_sp_latency = 32; m_L1I_config.init(m_L1I_config.m_config_string,FuncCachePreferNone); m_L1T_config.init(m_L1T_config.m_config_string,FuncCachePreferNone); @@ -1304,21 +1332,25 @@ struct shader_core_config : public core_config //op collector int gpgpu_operand_collector_num_units_sp; int gpgpu_operand_collector_num_units_sfu; + int gpgpu_operand_collector_num_units_tensor_core; 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_sfu; + unsigned int gpgpu_operand_collector_num_in_ports_tensor_core; 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_sfu; + unsigned int gpgpu_operand_collector_num_out_ports_tensor_core; 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_sfu_units; + int gpgpu_num_tensor_core_units; int gpgpu_num_mem_units; //Shader core resources @@ -1331,6 +1363,7 @@ struct shader_core_config : public core_config unsigned max_sp_latency; unsigned max_sfu_latency; + unsigned max_tensor_core_latency; unsigned n_simt_cores_per_cluster; unsigned n_simt_clusters; @@ -1368,12 +1401,14 @@ struct shader_core_stats_pod { unsigned *m_num_fpdiv_acesses; unsigned *m_num_sp_acesses; unsigned *m_num_sfu_acesses; + unsigned *m_num_tensor_core_acesses; unsigned *m_num_trans_acesses; unsigned *m_num_mem_acesses; unsigned *m_num_sp_committed; unsigned *m_num_tlb_hits; unsigned *m_num_tlb_accesses; unsigned *m_num_sfu_committed; + unsigned *m_num_tensor_core_committed; unsigned *m_num_mem_committed; unsigned *m_read_regfile_acesses; unsigned *m_write_regfile_acesses; @@ -1382,6 +1417,7 @@ struct shader_core_stats_pod { unsigned *m_num_imul32_acesses; unsigned *m_active_sp_lanes; unsigned *m_active_sfu_lanes; + unsigned *m_active_tensor_core_lanes; unsigned *m_active_fu_lanes; unsigned *m_active_fu_mem_lanes; unsigned *m_n_diverge; // number of divergence occurring in this shader @@ -1453,6 +1489,7 @@ public: m_num_fpdiv_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_num_sp_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_num_sfu_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); + m_num_tensor_core_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_num_trans_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_num_mem_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_num_sp_committed= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); @@ -1460,9 +1497,11 @@ public: m_num_tlb_accesses=(unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_active_sp_lanes= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_active_sfu_lanes= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); + m_active_tensor_core_lanes= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_active_fu_lanes= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_active_fu_mem_lanes= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_num_sfu_committed= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); + m_num_tensor_core_committed= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_num_mem_committed= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_read_regfile_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); m_write_regfile_acesses= (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); |