From e841e5f21b9d86910a6cc10de3af016912c43ce0 Mon Sep 17 00:00:00 2001 From: sspenst Date: Wed, 6 Jul 2016 15:08:11 -0700 Subject: Added the ability to load from sstarr memory after data has been stored in it --- src/gpgpu-sim/gpu-sim.cc | 3 +++ src/gpgpu-sim/shader.cc | 1 + src/gpgpu-sim/shader.h | 1 + 3 files changed, 5 insertions(+) (limited to 'src/gpgpu-sim') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index eafb909..e7ba8e5 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1049,6 +1049,9 @@ void shader_core_ctx::mem_instruction_stats(const warp_inst_t &inst) case shared_space: m_stats->gpgpu_n_shmem_insn += active_count; break; + case sstarr_space: + m_stats->gpgpu_n_sstarr_insn += active_count; + break; case const_space: m_stats->gpgpu_n_const_insn += active_count; break; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index ff2fac7..9c3f816 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -386,6 +386,7 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout, "gpgpu_n_load_insn = %d\n", gpgpu_n_load_insn); fprintf(fout, "gpgpu_n_store_insn = %d\n", gpgpu_n_store_insn); fprintf(fout, "gpgpu_n_shmem_insn = %d\n", gpgpu_n_shmem_insn); + fprintf(fout, "gpgpu_n_shmem_insn = %d\n", gpgpu_n_sstarr_insn); fprintf(fout, "gpgpu_n_tex_insn = %d\n", gpgpu_n_tex_insn); fprintf(fout, "gpgpu_n_const_mem_insn = %d\n", gpgpu_n_const_insn); fprintf(fout, "gpgpu_n_param_mem_insn = %d\n", gpgpu_n_param_insn); diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 38d09e9..3b9859f 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1372,6 +1372,7 @@ struct shader_core_stats_pod { unsigned gpgpu_n_load_insn; unsigned gpgpu_n_store_insn; unsigned gpgpu_n_shmem_insn; + unsigned gpgpu_n_sstarr_insn; unsigned gpgpu_n_tex_insn; unsigned gpgpu_n_const_insn; unsigned gpgpu_n_param_insn; -- cgit v1.3 From 68674d4ba230df0d3bf9f4e5b035f4cf9cfc185b Mon Sep 17 00:00:00 2001 From: negargoli93 Date: Sat, 12 May 2018 16:09:04 -0700 Subject: commit for eece527project --- src/abstract_hardware_model.h | 6 +- src/cuda-sim/cuda-sim.cc | 11 ++- src/cuda-sim/instructions.cc | 213 +++++++++++++++++++----------------------- src/cuda-sim/opcodes.def | 2 +- src/cuda-sim/ptx.l | 2 +- src/cuda-sim/ptx_ir.h | 25 +++++ src/gpgpu-sim/scoreboard.cc | 4 + src/gpgpusim_entrypoint.cc | 1 - 8 files changed, 136 insertions(+), 128 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index cdd9cf3..9dc58d4 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -750,7 +750,7 @@ public: }; // the maximum number of destination, source, or address uarch operands in a instruction -#define MAX_REG_OPERANDS 8 +#define MAX_REG_OPERANDS 32 struct dram_callback_t { dram_callback_t() { function=NULL; instruction=NULL; thread=NULL; } @@ -825,8 +825,8 @@ public: address_type reconvergence_pc; // -1 => not a branch, -2 => use function return address - unsigned out[4]; - unsigned in[4]; + unsigned out[8]; + unsigned in[8]; unsigned char is_vectorin; unsigned char is_vectorout; int pred; // predicate register number diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 54d8796..006738a 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -792,9 +792,9 @@ void ptx_instruction::set_opcode_and_latency() initiation_interval = dp_init[2]; op = SFU_OP; break; - case BSMAD_OP: - latency = int_precision/int_lane_width; - initiation_interval = int_init_precision/int_init_lane_width; + case MMA_OP: + latency = 64; + initiation_interval = 64; break; case SHFL_OP: latency = 32; @@ -1301,6 +1301,7 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) *((warp_inst_t*)pJ) = inst; // copy active mask information pI = pJ; } + if((pI->get_opcode()!=MMA_OP)||((pI->get_opcode()==MMA_OP)&&(lane_id==0))){ switch ( pI->get_opcode() ) { #define OP_DEF(OP,FUNC,STR,DST,CLASSIFICATION) case OP: FUNC(pI,this); op_classification = CLASSIFICATION; break; #define OP_W_DEF(OP,FUNC,STR,DST,CLASSIFICATION) case OP: FUNC(pI,get_core(),inst); op_classification = CLASSIFICATION; break; @@ -1308,7 +1309,7 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) #undef OP_DEF #undef OP_W_DEF default: printf( "Execution error: Invalid opcode (0x%x)\n", pI->get_opcode() ); break; - } + }} delete pJ; pI = pI_saved; @@ -1930,7 +1931,7 @@ void functionalCoreSim::executeWarp(unsigned i, bool &allAtBarrier, bool & someO { if(!m_warpAtBarrier[i] && m_liveThreadCount[i]!=0){ warp_inst_t inst =getExecuteWarp(i); - execute_warp_inst_t(inst,i); + execute_warp_inst_t(inst,i); if(inst.isatomic()) inst.do_atomic(true); if(inst.op==BARRIER_OP || inst.op==MEMORY_BARRIER_OP ) m_warpAtBarrier[i]=true; updateSIMTStack( i, &inst ); diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 493e307..7903343 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -771,6 +771,17 @@ void add_impl( const ptx_instruction *pI, ptx_thread_info *thread ) unsigned i_type = pI->get_type(); src1_data = thread->get_operand_value(src1, dst, i_type, thread, 1); src2_data = thread->get_operand_value(src2, dst, i_type, thread, 1); + //unsigned warpId_aa,warp_size_aa; + //warpId_aa = pI->warp_id(); + //warp_size_aa=32; + //dim3 t=thread->get_tid(); + //unsigned tid_aa=warp_size_aa*warpId_aa; + + ptx_thread_info *thread2; + thread2=thread; + src1_data = thread2->get_operand_value(src1, dst, i_type, thread2, 1); + src2_data = thread2->get_operand_value(src2, dst, i_type, thread2, 1); + unsigned rounding_mode = pI->rounding_mode(); int orig_rm = fegetround(); @@ -1483,135 +1494,102 @@ unsigned trunc(unsigned num, unsigned precision) { return num; } -void bsmad_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) +void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) { - // operands: - // 0 = output - // 1 = input precision - // 2 = output precision - // 3 = buffer0 - // 4 = buffer1 - // 5 = buffer2 - // 6 = buffer3 - // 7 = synapse value - // 8 = output value + int i,j,k,thrd; + int row,offset; + printf("mmaWorld\n"); + ptx_reg_t matrix_a[16][16]; + ptx_reg_t matrix_b[16][16]; + ptx_reg_t matrix_c[16][16]; + ptx_reg_t matrix_d[16][16]; + ptx_reg_t src_data; + ptx_thread_info *thread; - const operand_info &dst = pI->dst(); - const operand_info &src1 = pI->src1(); - const operand_info &src2 = pI->src2(); unsigned type = pI->get_type(); int tid = inst.warp_id_func() * core->get_warp_size(); - ptx_thread_info *thread = core->get_thread_info()[tid]; - const int ip = (thread->get_operand_value(src1, dst, type, thread, 1)).u32; - const int op = (thread->get_operand_value(src2, dst, type, thread, 1)).u32; - const int THREADS = inst.active_count(); - const int INBUFFERS = 4; - const int OUTBUFFERS = (((32/ip)*INBUFFERS) / (32/op)) + ((((32/ip)*INBUFFERS) % (32/op)) != 0); - if (OUTBUFFERS > THREADS) { - printf("GPGPU-Sim PTX: BSMAD ERROR - Number of output registers required (%d) is greater than the number available (%d)\n", OUTBUFFERS, THREADS); - abort(); - } - ptx_warp_info *warp_info = thread->m_warp_info; - warp_info->inc_done_threads(); + const operand_info &dst = pI->operand_lookup(0); + +//NOT WOR thread = core->get_thread_info()[tid]; +//NOT WOR const operand_info &src_a= pI->operand_lookup(1); +//NOT WOR src_data= (thread->get_operand_value(src_a, dst, type, thread, 1)); +//NOT WOR thread->set_operand_value(dst, src_data, type, thread, pI); + for (thrd=0; thrd < core->get_warp_size(); thrd++){ + row=thrd/2; + offset=8*(thrd%2); + thread = core->get_thread_info()[tid+thrd]; + printf("thread=%d:",thrd); + for(i=8;i<=31;i++){ + const operand_info &src_a= pI->operand_lookup(i); + src_data= (thread->get_operand_value(src_a, dst, type, thread, 1)); + printf("%f ",src_data.f32); + if(i<=15) + matrix_a[row][offset+(i)%8]=src_data; + else if((i>15)&&(i<=23)) + matrix_b[row][offset+(i)%8]=src_data; + else if(i>23) + matrix_c[row][offset+(i)%8]=src_data; + - // threads within the warp are executed sequentially by the simulator, store output in first four registers - if (warp_info->get_done_threads() <= OUTBUFFERS) { - unsigned buffer[THREADS][INBUFFERS]; - unsigned synapse[THREADS]; - unsigned output; - - // loop through all threads in the warp and get all data - for (unsigned i = 0, j = 0; i < core->get_warp_size(); i++) { - if (inst.active(i)) { - const operand_info &src3 = pI->operand_lookup(3); - const operand_info &src4 = pI->operand_lookup(4); - const operand_info &src5 = pI->operand_lookup(5); - const operand_info &src6 = pI->operand_lookup(6); - const operand_info &src7 = pI->operand_lookup(7); - const operand_info &src8 = pI->operand_lookup(8); - - thread = core->get_thread_info()[tid+i]; - // get buffer data and synapse data from each thread - buffer[j][0] = (thread->get_operand_value(src3, dst, type, thread, 1)).u32; - buffer[j][1] = (thread->get_operand_value(src4, dst, type, thread, 1)).u32; - buffer[j][2] = (thread->get_operand_value(src5, dst, type, thread, 1)).u32; - buffer[j][3] = (thread->get_operand_value(src6, dst, type, thread, 1)).u32; - synapse[j] = (thread->get_operand_value(src7, dst, type, thread, 1)).u32; - j++; - // get output data from the first 4 threads - if (j == warp_info->get_done_threads()) { - output = (thread->get_operand_value(src8, dst, type, thread, 1)).u32; - } - } } + printf("\n"); + } - // unpack registers, compute enough outputs to fill an output register - unsigned *unpacked_output = (unsigned*)calloc(32/op,sizeof(unsigned)); - unsigned buffer_data_start = (32/op)*(warp_info->get_done_threads()-1); - for (unsigned i = buffer_data_start; i < (32/op + buffer_data_start) && i < (32/ip)*INBUFFERS; i++) { - unsigned buf = i/(32/ip); - unsigned pos = i%(32/ip); - // sum values from the buffers - int sum = 0; - unsigned mask = (unsigned)(pow(2,ip)-1) << (pos*ip); - for (int j = 0; j < THREADS; j++) { - //sum += ((mask & buffer[j][buf]) >> (pos*ip)) * synapse[j]; - sum += trunc(((mask & buffer[j][buf]) >> (pos*ip)) * synapse[j], op); - } - // get the previous output - mask = (unsigned)(pow(2,op)-1) << (op*(i-buffer_data_start)); - int past_output = (mask & output) >> (op*(i-buffer_data_start)); - unpacked_output[i-buffer_data_start] = trunc(trunc(sum,op) + past_output,op); - // truncate sum, truncate (truncated sum + past_output) + printf("MATRIX_A\n"); + for (i=0;i<16;i++){ + for(j=0;j<16;j++){ + printf("%f ",matrix_a[i][j].f32); } - - // truncate output - /*for (unsigned i = 0; i < 32/op; i++) { - int mask = 1, latest_one = -1; - unsigned data = unpacked_output[i]; - for (unsigned j = 0; j < sizeof(unsigned)*8; j++) { - int bit = data & mask; - if (bit == 1) latest_one = j; - data >>= 1; - } - if (latest_one >= op) { - // round_up is 1 if the most significant truncated digit is a 1, otherwise it is 0 - int round_up = (unpacked_output[i] & (1 << (latest_one-op))) >> (latest_one-op); - unsigned shifted_output = unpacked_output[i] >> (latest_one-op+1); - // if shifted_output is a number like 1111, don't round up - if (shifted_output == (pow(2,op)-1)) round_up = 0; - unpacked_output[i] = shifted_output + round_up; - } - }*/ - - // pack the outputs into one register - unsigned mask = pow(2,op)-1; - unsigned output_data = 0; - for (int i = 0; i < 32/op; i++) { - output_data |= (unpacked_output[i] & mask) << (op*i); + printf("\n"); + } + printf("MATRIX_B\n"); + for (i=0;i<16;i++){ + for(j=0;j<16;j++){ + printf("%f ",matrix_b[i][j].f32); } - - // store the result in the correct thread's output register - for (unsigned i = 0, j = 0; i < core->get_warp_size(); i++) { - if (inst.active(i)) j++; - if (j == warp_info->get_done_threads()) { - thread = core->get_thread_info()[tid+i]; - ptx_reg_t data; - data.u32 = output_data; - thread->set_operand_value(dst, data, type, thread, pI); - break; + printf("\n"); + } + printf("MATRIX_C\n"); + for (i=0;i<16;i++){ + for(j=0;j<16;j++){ + printf("%f ",matrix_c[i][j].f32); + } + printf("\n"); + } + for (i=0;i<16;i++){ + for(j=0;j<16;j++){ + matrix_d[i][j].f32=0; + } + } + + for (i=0;i<16;i++){ + for(j=0;j<16;j++){ + for(k=0;k<16;k++){ + matrix_d[i][j].f32=matrix_d[i][j].f32+matrix_a[i][k].f32*matrix_b[j][k].f32; } + matrix_d[i][j].f32+=matrix_c[i][j].f32; } } - - // once the warp has finished, set the number of completed threads back to 0 for the next warp - if (warp_info->get_done_threads() == THREADS) { - warp_info->reset_done_threads(); + printf("MATRIX_D\n"); + for (i=0;i<16;i++){ + for(j=0;j<16;j++){ + printf("%f ",matrix_d[i][j].f32); + } + printf("\n"); + } + for (thrd=0; thrd < core->get_warp_size(); thrd++){ + thread = core->get_thread_info()[tid+thrd]; + row=thrd/2; + offset=8*(thrd%2); + for(i=0;i<8;i++){ + const operand_info &dst = pI->operand_lookup(i); + const symbol *r2; + r2=dst.get_symbol(); + printf("thrd=%d,i=%d,register%s, data=%f\n",thrd,i,(r2->name()).c_str(),matrix_d[row][offset+i].f32); + thread->set_operand_value(dst, matrix_d[row][offset+i], type, thread, pI); + } } - - // set the latency assuming 4 bits of each input get processed every cycle - // mutable latency variable??? - //pI->latency = (ip+3)/4; + } void call_impl( const ptx_instruction *pI, ptx_thread_info *thread ) @@ -4098,7 +4076,8 @@ void st_impl( const ptx_instruction *pI, ptx_thread_info *thread ) if (!vector_spec) { data = thread->get_operand_value(src1, dst, type, thread, 1); mem->write(addr,size/8,&data.s64,thread,pI); - } else { + printf("addr=%d data=%d\n",addr,data.s64); + } else { if (vector_spec == V2_TYPE) { ptx_reg_t* ptx_regs = new ptx_reg_t[2]; thread->get_vector_operand_values(src1, ptx_regs, 2); diff --git a/src/cuda-sim/opcodes.def b/src/cuda-sim/opcodes.def index 41f2f22..a3cc83f 100644 --- a/src/cuda-sim/opcodes.def +++ b/src/cuda-sim/opcodes.def @@ -52,7 +52,7 @@ OP_DEF(BRA_OP,bra_impl,"bra",0,3) OP_DEF(BRX_OP,brx_impl,"brx",0,3) OP_DEF(BREV_OP,brev_impl,"brev",1,1) OP_DEF(BRKPT_OP,brkpt_impl,"brkpt",1,9) -OP_W_DEF(BSMAD_OP,bsmad_impl,"bsmad",1,1) +OP_W_DEF(MMA_OP,mma_impl,"mma",1,1) OP_DEF(CALL_OP,call_impl,"call",1,3) OP_DEF(CALLP_OP,callp_impl,"callp",1,3) OP_DEF(CLZ_OP,clz_impl,"clz",1,1) diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l index 7620134..e07e339 100644 --- a/src/cuda-sim/ptx.l +++ b/src/cuda-sim/ptx.l @@ -68,7 +68,7 @@ bra TC; ptx_lval.int_value = BRA_OP; return OPCODE; brx TC; ptx_lval.int_value = BRX_OP; return OPCODE; brev TC; ptx_lval.int_value = BREV_OP; return OPCODE; brkpt TC; ptx_lval.int_value = BRKPT_OP; return OPCODE; -bsmad TC; ptx_lval.int_value = BSMAD_OP; return OPCODE; +mma TC; ptx_lval.int_value = MMA_OP; return OPCODE; call TC; BEGIN(NOT_OPCODE); ptx_lval.int_value = CALL_OP; return OPCODE; // blocking opcode token in case the callee has the same name as an opcode callp TC; BEGIN(NOT_OPCODE); ptx_lval.int_value = CALLP_OP; return OPCODE; clz TC; ptx_lval.int_value = CLZ_OP; return OPCODE; diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 4c10373..0601b97 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -948,6 +948,31 @@ public: assert( m_operands.size() > 3 ); return m_operands[3]; } + const operand_info &src4() const + { + assert( m_operands.size() > 4 ); + return m_operands[4]; + } + const operand_info &src5() const + { + assert( m_operands.size() > 5 ); + return m_operands[5]; + } + const operand_info &src6() const + { + assert( m_operands.size() > 6 ); + return m_operands[6]; + } + const operand_info &src7() const + { + assert( m_operands.size() > 7 ); + return m_operands[7]; + } + const operand_info &src8() const + { + assert( m_operands.size() > 8 ); + return m_operands[8]; + } const operand_info &operand_lookup( unsigned n ) const { diff --git a/src/gpgpu-sim/scoreboard.cc b/src/gpgpu-sim/scoreboard.cc index f412054..b538fdf 100644 --- a/src/gpgpu-sim/scoreboard.cc +++ b/src/gpgpu-sim/scoreboard.cc @@ -146,6 +146,10 @@ bool Scoreboard::checkCollision( unsigned wid, const class inst_t *inst ) const 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->ar1 > 0) inst_regs.insert(inst->ar1); if(inst->ar2 > 0) inst_regs.insert(inst->ar2); diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index 04845e7..a6d7eb4 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -134,7 +134,6 @@ void *gpgpu_sim_thread_concurrent(void*) gpgpu_cuda_ptx_sim_main_func(*kernel); g_the_gpu->finish_functional_sim(kernel); } - //performance simulation if( g_the_gpu->active() ) { g_the_gpu->cycle(); -- cgit v1.3 From fa0089a5d3a86ef348fae9a83a862f5219892bab Mon Sep 17 00:00:00 2001 From: aamir Date: Wed, 30 May 2018 23:07:44 -0700 Subject: adding code for wmma_ld_impl, error at decode space --- cuda-kernels/.tensor_core_ptx.swp | Bin 0 -> 16384 bytes cuda-kernels/Makefile | 3 +- cuda-kernels/tensor_core | Bin 48541 -> 2750968 bytes cuda-kernels/tensor_core_ptx | 171 +++++++++++ src/Makefile | 2 +- src/cuda-sim/.ptx.y.swp | Bin 0 -> 36864 bytes src/cuda-sim/Makefile | 2 +- src/cuda-sim/instructions.cc | 72 ++++- src/cuda-sim/ptx.y~ | 608 ++++++++++++++++++++++++++++++++++++++ src/cuda-sim/ptx_ir.h | 4 +- src/cuda-sim/ptx_sim.h | 9 + src/gpgpu-sim/Makefile | 2 +- src/intersim2/Makefile | 2 +- 13 files changed, 865 insertions(+), 10 deletions(-) create mode 100644 cuda-kernels/.tensor_core_ptx.swp create mode 100644 cuda-kernels/tensor_core_ptx create mode 100644 src/cuda-sim/.ptx.y.swp create mode 100644 src/cuda-sim/ptx.y~ (limited to 'src/gpgpu-sim') diff --git a/cuda-kernels/.tensor_core_ptx.swp b/cuda-kernels/.tensor_core_ptx.swp new file mode 100644 index 0000000..6d7bad4 Binary files /dev/null and b/cuda-kernels/.tensor_core_ptx.swp differ diff --git a/cuda-kernels/Makefile b/cuda-kernels/Makefile index 51a7760..673460f 100755 --- a/cuda-kernels/Makefile +++ b/cuda-kernels/Makefile @@ -1,5 +1,6 @@ all: tensor_core.cu - nvcc -arch=sm_70 -lcudart -g -o tensor_core tensor_core.cu + nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o tensor_core tensor_core.cu +# nvcc -arch=sm_70 -lcudart -g -o tensor_core tensor_core.cu .PHONY: clean: diff --git a/cuda-kernels/tensor_core b/cuda-kernels/tensor_core index b25f3d9..cb53851 100755 Binary files a/cuda-kernels/tensor_core and b/cuda-kernels/tensor_core differ diff --git a/cuda-kernels/tensor_core_ptx b/cuda-kernels/tensor_core_ptx new file mode 100644 index 0000000..36074cb --- /dev/null +++ b/cuda-kernels/tensor_core_ptx @@ -0,0 +1,171 @@ +// +// Generated by NVIDIA NVVM Compiler +// +// Compiler Build ID: CL-22781540 +// Cuda compilation tools, release 9.0, V9.0.176 +// Based on LLVM 3.4svn +// + +.version 6.0 +.target sm_70 +.address_size 64 + + // .globl _Z12wmma_exampleP6__halfS0_Pfiiiff +.extern .func (.param .b32 func_retval0) vprintf +( + .param .b64 vprintf_param_0, + .param .b64 vprintf_param_1 +) +; +.global .align 16 .b8 $str[9] = {99, 108, 111, 99, 107, 61, 37, 100, 0}; + +.visible .entry _Z12wmma_exampleP6__halfS0_Pfiiiff( + .param .u64 _Z12wmma_exampleP6__halfS0_Pfiiiff_param_0, + .param .u64 _Z12wmma_exampleP6__halfS0_Pfiiiff_param_1, + .param .u64 _Z12wmma_exampleP6__halfS0_Pfiiiff_param_2, + .param .u32 _Z12wmma_exampleP6__halfS0_Pfiiiff_param_3, + .param .u32 _Z12wmma_exampleP6__halfS0_Pfiiiff_param_4, + .param .u32 _Z12wmma_exampleP6__halfS0_Pfiiiff_param_5, + .param .f32 _Z12wmma_exampleP6__halfS0_Pfiiiff_param_6, + .param .f32 _Z12wmma_exampleP6__halfS0_Pfiiiff_param_7 +) +{ + .local .align 8 .b8 __local_depot0[8]; + .reg .b64 %SP; + .reg .b64 %SPL; + .reg .pred %p<6>; + .reg .f32 %f<34>; + .reg .b32 %r<38>; + .reg .b64 %rd<18>; + + + mov.u64 %rd17, __local_depot0; + cvta.local.u64 %SP, %rd17; + ld.param.u64 %rd1, [_Z12wmma_exampleP6__halfS0_Pfiiiff_param_0]; + ld.param.u64 %rd2, [_Z12wmma_exampleP6__halfS0_Pfiiiff_param_1]; + ld.param.u64 %rd3, [_Z12wmma_exampleP6__halfS0_Pfiiiff_param_2]; + ld.param.u32 %r4, [_Z12wmma_exampleP6__halfS0_Pfiiiff_param_3]; + ld.param.u32 %r7, [_Z12wmma_exampleP6__halfS0_Pfiiiff_param_4]; + ld.param.u32 %r5, [_Z12wmma_exampleP6__halfS0_Pfiiiff_param_5]; + // inline asm + mov.u32 %r6, %clock; + // inline asm + mov.u32 %r8, %ntid.x; + mov.u32 %r9, %ctaid.x; + mov.u32 %r10, %tid.x; + mad.lo.s32 %r11, %r8, %r9, %r10; + mov.u32 %r12, WARP_SZ; + div.u32 %r13, %r11, %r12; + mov.u32 %r14, %ntid.y; + mov.u32 %r15, %ctaid.y; + mov.u32 %r16, %tid.y; + mad.lo.s32 %r17, %r14, %r15, %r16; + shl.b32 %r2, %r13, 4; + shl.b32 %r3, %r17, 4; + setp.lt.s32 %p1, %r2, %r4; + setp.gt.s32 %p2, %r5, 0; + and.pred %p3, %p1, %p2; + setp.lt.s32 %p4, %r3, %r7; + and.pred %p5, %p3, %p4; + mov.f32 %f26, 0f00000000; + mov.f32 %f27, %f26; + mov.f32 %f28, %f26; + mov.f32 %f29, %f26; + mov.f32 %f30, %f26; + mov.f32 %f31, %f26; + mov.f32 %f32, %f26; + mov.f32 %f33, %f26; + @!%p5 bra BB0_2; + bra.uni BB0_1; + +BB0_1: + mul.wide.s32 %rd4, %r2, 2; + add.s64 %rd5, %rd1, %rd4; + wmma.load.a.sync.row.m16n16k16.f16 {%r18, %r19, %r20, %r21, %r22, %r23, %r24, %r25}, [%rd5], %r4; + mul.wide.s32 %rd6, %r3, 2; + add.s64 %rd7, %rd2, %rd6; + wmma.load.b.sync.col.m16n16k16.f16 {%r26, %r27, %r28, %r29, %r30, %r31, %r32, %r33}, [%rd7], %r5; + mov.f32 %f25, 0f00000000; + wmma.mma.sync.row.col.m16n16k16.f32.f32 {%f33, %f32, %f31, %f30, %f29, %f28, %f27, %f26}, {%r18, %r19, %r20, %r21, %r22, %r23, %r24, %r25}, {%r26, %r27, %r28, %r29, %r30, %r31, %r32, %r33}, {%f25, %f25, %f25, %f25, %f25, %f25, %f25, %f25}; + +BB0_2: + add.u64 %rd8, %SP, 0; + cvta.to.local.u64 %rd9, %rd8; + mul.lo.s32 %r35, %r3, %r4; + cvt.s64.s32 %rd10, %r35; + cvt.s64.s32 %rd11, %r2; + add.s64 %rd12, %rd10, %rd11; + shl.b64 %rd13, %rd12, 2; + add.s64 %rd14, %rd3, %rd13; + wmma.store.d.sync.col.m16n16k16.f32 [%rd14], {%f33, %f32, %f31, %f30, %f29, %f28, %f27, %f26}, %r4; + // inline asm + mov.u32 %r34, %clock; + // inline asm + sub.s32 %r36, %r34, %r6; + st.local.u32 [%rd9], %r36; + mov.u64 %rd15, $str; + cvta.global.u64 %rd16, %rd15; + // Callseq Start 0 + { + .reg .b32 temp_param_reg; + // } + .param .b64 param0; + st.param.b64 [param0+0], %rd16; + .param .b64 param1; + st.param.b64 [param1+0], %rd8; + .param .b32 retval0; + call.uni (retval0), + vprintf, + ( + param0, + param1 + ); + ld.param.b32 %r37, [retval0+0]; + + //{ + }// Callseq End 0 + ret; +} + + // .globl _Z17convertFp32ToFp16P6__halfPfi +.visible .entry _Z17convertFp32ToFp16P6__halfPfi( + .param .u64 _Z17convertFp32ToFp16P6__halfPfi_param_0, + .param .u64 _Z17convertFp32ToFp16P6__halfPfi_param_1, + .param .u32 _Z17convertFp32ToFp16P6__halfPfi_param_2 +) +{ + .reg .pred %p<2>; + .reg .b16 %rs<2>; + .reg .f32 %f<2>; + .reg .b32 %r<6>; + .reg .b64 %rd<9>; + + + ld.param.u64 %rd1, [_Z17convertFp32ToFp16P6__halfPfi_param_0]; + ld.param.u64 %rd2, [_Z17convertFp32ToFp16P6__halfPfi_param_1]; + ld.param.u32 %r2, [_Z17convertFp32ToFp16P6__halfPfi_param_2]; + mov.u32 %r3, %ntid.x; + mov.u32 %r4, %ctaid.x; + mov.u32 %r5, %tid.x; + mad.lo.s32 %r1, %r4, %r3, %r5; + setp.ge.s32 %p1, %r1, %r2; + @%p1 bra BB1_2; + + cvta.to.global.u64 %rd3, %rd2; + mul.wide.s32 %rd4, %r1, 4; + add.s64 %rd5, %rd3, %rd4; + ld.global.f32 %f1, [%rd5]; + // inline asm + { cvt.rn.f16.f32 %rs1, %f1;} + + // inline asm + cvta.to.global.u64 %rd6, %rd1; + mul.wide.s32 %rd7, %r1, 2; + add.s64 %rd8, %rd6, %rd7; + st.global.u16 [%rd8], %rs1; + +BB1_2: + ret; +} + + diff --git a/src/Makefile b/src/Makefile index 6001669..09194f3 100644 --- a/src/Makefile +++ b/src/Makefile @@ -46,7 +46,7 @@ ifeq ($(TRACE),1) endif ifneq ($(DEBUG),1) - OPTFLAGS += -O3 + OPTFLAGS += -O0 else CXXFLAGS += endif diff --git a/src/cuda-sim/.ptx.y.swp b/src/cuda-sim/.ptx.y.swp new file mode 100644 index 0000000..c8a83b5 Binary files /dev/null and b/src/cuda-sim/.ptx.y.swp differ diff --git a/src/cuda-sim/Makefile b/src/cuda-sim/Makefile index 999dad7..a65e8e1 100644 --- a/src/cuda-sim/Makefile +++ b/src/cuda-sim/Makefile @@ -42,7 +42,7 @@ include ../../version_detection.mk OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/cuda-sim -OPT := -O3 -g3 -Wall -Wno-unused-function -Wno-sign-compare +OPT := -O0 -g3 -Wall -Wno-unused-function -Wno-sign-compare ifeq ($(DEBUG),1) OPT := -g3 -Wall -Wno-unused-function -Wno-sign-compare endif diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 446cdbf..16f33c6 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -643,6 +643,33 @@ void ptx_thread_info::set_vector_operand_values( const operand_info &dst, m_last_set_operand_value = data1; } +void ptx_thread_info::set_wmma_vector_operand_values( const operand_info &dst, + const ptx_reg_t &data1, + const ptx_reg_t &data2, + const ptx_reg_t &data3, + const ptx_reg_t &data4, + const ptx_reg_t &data5, + const ptx_reg_t &data6, + const ptx_reg_t &data7, + const ptx_reg_t &data8 ) +{ + unsigned num_elements = dst.get_vect_nelem(); + if (num_elements > 7) { + set_reg(dst.vec_symbol(0), data1); + set_reg(dst.vec_symbol(1), data2); + set_reg(dst.vec_symbol(2), data3); + set_reg(dst.vec_symbol(3), data4); + set_reg(dst.vec_symbol(4), data5); + set_reg(dst.vec_symbol(5), data6); + set_reg(dst.vec_symbol(6), data7); + set_reg(dst.vec_symbol(7), data8); + } + else{ + printf("error:set_wmma_vector_operands"); + } + + m_last_set_operand_value = data1; +} #define my_abs(a) (((a)<0)?(-a):(a)) @@ -1493,9 +1520,6 @@ unsigned trunc(unsigned num, unsigned precision) { } return num; } -void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) -{ -} void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) { } @@ -2595,6 +2619,48 @@ void ldu_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { ld_exec(pI,thread); } +void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) +{ + const operand_info &dst = pI->dst(); + const operand_info &src1 = pI->src1(); + const operand_info &src2 = pI->src2(); + + unsigned type = pI->get_type(); + + int tid = inst.warp_id_func() * core->get_warp_size(); + int thrd; + ptx_thread_info *thread; + thread = core->get_thread_info()[tid]; + + ptx_reg_t src1_data = thread->get_operand_value(src1, dst, type, thread, 1); + + ptx_reg_t data; + memory_space_t space = pI->get_space(); + + memory_space *mem = NULL; + addr_t addr = src1_data.u32; + + decode_space(space,thread,src1,mem,addr); + + size_t size; + int t; + data.u64=0; + type_info_key::type_decode(type,size,t); + ptx_reg_t data1, data2, data3, data4; + ptx_reg_t data5, data6, data7, data8; + mem->read(addr,size/8,&data1.s64); + mem->read(addr+size/8,size/8,&data2.s64); + mem->read(addr+2*size/8,size/8,&data3.s64); + mem->read(addr+3*size/8,size/8,&data4.s64); + mem->read(addr+4*size/8,size/8,&data5.s64); + mem->read(addr+5*size/8,size/8,&data6.s64); + mem->read(addr+6*size/8,size/8,&data7.s64); + mem->read(addr+7*size/8,size/8,&data8.s64); + thread->set_wmma_vector_operand_values(dst,data1,data2,data3,data4,data5,data6,data7,data8); + + thread->m_last_effective_address = addr; + thread->m_last_memory_space = space; +} void lg2_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { diff --git a/src/cuda-sim/ptx.y~ b/src/cuda-sim/ptx.y~ new file mode 100644 index 0000000..0710ecd --- /dev/null +++ b/src/cuda-sim/ptx.y~ @@ -0,0 +1,608 @@ +/* +Copyright (c) 2009-2011, Tor M. Aamodt +The University of British Columbia +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +Redistributions of source code must retain the above copyright notice, this +list of conditions and the following disclaimer. +Redistributions in binary form must reproduce the above copyright notice, this +list of conditions and the following disclaimer in the documentation and/or +other materials provided with the distribution. +Neither the name of The University of British Columbia nor the names of its +contributors may be used to endorse or promote products derived from this +software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +%union { + double double_value; + float float_value; + int int_value; + char * string_value; + void * ptr_value; +} + +%token STRING +%token OPCODE +%token WMMA_DIRECTIVE +%token LAYOUT +%token CONFIGURATION +%token ALIGN_DIRECTIVE +%token BRANCHTARGETS_DIRECTIVE +%token BYTE_DIRECTIVE +%token CALLPROTOTYPE_DIRECTIVE +%token CALLTARGETS_DIRECTIVE +%token CONST_DIRECTIVE +%token CONSTPTR_DIRECTIVE +%token PTR_DIRECTIVE +%token ENTRY_DIRECTIVE +%token EXTERN_DIRECTIVE +%token WEAK_DIRECTIVE +%token FILE_DIRECTIVE +%token FUNC_DIRECTIVE +%token GLOBAL_DIRECTIVE +%token LOCAL_DIRECTIVE +%token LOC_DIRECTIVE +%token MAXNCTAPERSM_DIRECTIVE +%token MAXNNREG_DIRECTIVE +%token MAXNTID_DIRECTIVE +%token MINNCTAPERSM_DIRECTIVE +%token PARAM_DIRECTIVE +%token PRAGMA_DIRECTIVE +%token REG_DIRECTIVE +%token REQNTID_DIRECTIVE +%token SECTION_DIRECTIVE +%token SHARED_DIRECTIVE +%token SREG_DIRECTIVE +%token SSTARR_DIRECTIVE +%token STRUCT_DIRECTIVE +%token SURF_DIRECTIVE +%token TARGET_DIRECTIVE +%token TEX_DIRECTIVE +%token UNION_DIRECTIVE +%token VERSION_DIRECTIVE +%token ADDRESS_SIZE_DIRECTIVE +%token VISIBLE_DIRECTIVE +%token WEAK_DIRECTIVE +%token IDENTIFIER +%token INT_OPERAND +%token FLOAT_OPERAND +%token DOUBLE_OPERAND +%token S8_TYPE +%token S16_TYPE +%token S32_TYPE +%token S64_TYPE +%token U8_TYPE +%token U16_TYPE +%token U32_TYPE +%token U64_TYPE +%token F16_TYPE +%token F32_TYPE +%token F64_TYPE +%token FF64_TYPE +%token B8_TYPE +%token B16_TYPE +%token B32_TYPE +%token B64_TYPE +%token BB64_TYPE +%token BB128_TYPE +%token PRED_TYPE +%token TEXREF_TYPE +%token SAMPLERREF_TYPE +%token SURFREF_TYPE +%token V2_TYPE +%token V3_TYPE +%token V4_TYPE +%token COMMA +%token PRED +%token HALF_OPTION +%token EXTP_OPTION +%token EQ_OPTION +%token NE_OPTION +%token LT_OPTION +%token LE_OPTION +%token GT_OPTION +%token GE_OPTION +%token LO_OPTION +%token LS_OPTION +%token HI_OPTION +%token HS_OPTION +%token EQU_OPTION +%token NEU_OPTION +%token LTU_OPTION +%token LEU_OPTION +%token GTU_OPTION +%token GEU_OPTION +%token NUM_OPTION +%token NAN_OPTION +%token CF_OPTION +%token SF_OPTION +%token NSF_OPTION +%token LEFT_SQUARE_BRACKET +%token RIGHT_SQUARE_BRACKET +%token WIDE_OPTION +%token SPECIAL_REGISTER +%token MINUS +%token PLUS +%token COLON +%token SEMI_COLON +%token EXCLAMATION +%token PIPE +%token RIGHT_BRACE +%token LEFT_BRACE +%token EQUALS +%token PERIOD +%token BACKSLASH +%token DIMENSION_MODIFIER +%token RN_OPTION +%token RZ_OPTION +%token RM_OPTION +%token RP_OPTION +%token RNI_OPTION +%token RZI_OPTION +%token RMI_OPTION +%token RPI_OPTION +%token UNI_OPTION +%token GEOM_MODIFIER_1D +%token GEOM_MODIFIER_2D +%token GEOM_MODIFIER_3D +%token SAT_OPTION +%token FTZ_OPTION +%token NEG_OPTION +%token SYNC_OPTION +%token RED_OPTION +%token ARRIVE_OPTION +%token ATOMIC_POPC +%token ATOMIC_AND +%token ATOMIC_OR +%token ATOMIC_XOR +%token ATOMIC_CAS +%token ATOMIC_EXCH +%token ATOMIC_ADD +%token ATOMIC_INC +%token ATOMIC_DEC +%token ATOMIC_MIN +%token ATOMIC_MAX +%token LEFT_ANGLE_BRACKET +%token RIGHT_ANGLE_BRACKET +%token LEFT_PAREN +%token RIGHT_PAREN +%token APPROX_OPTION +%token FULL_OPTION +%token ANY_OPTION +%token ALL_OPTION +%token BALLOT_OPTION +%token GLOBAL_OPTION +%token CTA_OPTION +%token SYS_OPTION +%token EXIT_OPTION +%token ABS_OPTION +%token TO_OPTION +%token CA_OPTION; +%token CG_OPTION; +%token CS_OPTION; +%token LU_OPTION; +%token CV_OPTION; +%token WB_OPTION; +%token WT_OPTION; +%token NC_OPTION; +%token UP_OPTION; +%token DOWN_OPTION; +%token BFLY_OPTION; +%token IDX_OPTION; + +%type function_decl_header +%type function_decl + +%{ + #include "ptx_parser.h" + #include + #include + #include + void syntax_not_implemented(); + extern int g_func_decl; + int ptx_lex(void); + int ptx_error(const char *); +%} + +%% + +input: /* empty */ + | input directive_statement + | input function_defn + | input function_decl + ; + +function_defn: function_decl { set_symtab($1); func_header(".skip"); } statement_block { end_function(); } + | function_decl { set_symtab($1); } block_spec_list { func_header(".skip"); } statement_block { end_function(); } + ; + +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); } + | 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"); } + ; + +block_spec_list: block_spec + | block_spec_list block_spec + ; + +function_decl: function_decl_header LEFT_PAREN { start_function($1); func_header_info("(");} param_entry RIGHT_PAREN {func_header_info(")");} function_ident_param { $$ = reset_symtab(); } + | function_decl_header { start_function($1); } function_ident_param { $$ = reset_symtab(); } + | function_decl_header { start_function($1); add_function_name(""); g_func_decl=0; $$ = reset_symtab(); } + ; + +function_ident_param: IDENTIFIER { add_function_name($1); } LEFT_PAREN {func_header_info("(");} param_list RIGHT_PAREN { g_func_decl=0; func_header_info(")"); } + | IDENTIFIER { add_function_name($1); g_func_decl=0; } + ; + +function_decl_header: ENTRY_DIRECTIVE { $$ = 1; g_func_decl=1; func_header(".entry"); } + | VISIBLE_DIRECTIVE ENTRY_DIRECTIVE { $$ = 1; g_func_decl=1; func_header(".entry"); } + | WEAK_DIRECTIVE ENTRY_DIRECTIVE { $$ = 1; g_func_decl=1; func_header(".entry"); } + | FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; func_header(".func"); } + | VISIBLE_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; func_header(".func"); } + | WEAK_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; func_header(".func"); } + | EXTERN_DIRECTIVE FUNC_DIRECTIVE { $$ = 2; g_func_decl=1; func_header(".func"); } + | WEAK_DIRECTIVE FUNC_DIRECTIVE { $$ = 0; g_func_decl=1; func_header(".func"); } + ; + +param_list: /*empty*/ + | param_entry { add_directive(); } + | param_list COMMA {func_header_info(",");} param_entry { add_directive(); } + +param_entry: PARAM_DIRECTIVE { add_space_spec(param_space_unclassified,0); } variable_spec ptr_spec identifier_spec { add_function_arg(); } + | REG_DIRECTIVE { add_space_spec(reg_space,0); } variable_spec identifier_spec { add_function_arg(); } + +ptr_spec: /*empty*/ + | PTR_DIRECTIVE ptr_space_spec ptr_align_spec + | PTR_DIRECTIVE ptr_align_spec + +ptr_space_spec: GLOBAL_DIRECTIVE { add_ptr_spec(global_space); } + | LOCAL_DIRECTIVE { add_ptr_spec(local_space); } + | SHARED_DIRECTIVE { add_ptr_spec(shared_space); } + +ptr_align_spec: ALIGN_DIRECTIVE INT_OPERAND + +statement_block: LEFT_BRACE statement_list RIGHT_BRACE + +statement_list: directive_statement { add_directive(); } + | instruction_statement { add_instruction(); } + | statement_list directive_statement { add_directive(); } + | statement_list instruction_statement { add_instruction(); } + | statement_list {start_inst_group();} statement_block {end_inst_group();} + | {start_inst_group();} statement_block {end_inst_group();} + ; + +directive_statement: variable_declaration SEMI_COLON + | VERSION_DIRECTIVE DOUBLE_OPERAND { add_version_info($2, 0); } + | VERSION_DIRECTIVE DOUBLE_OPERAND PLUS { add_version_info($2,1); } + | ADDRESS_SIZE_DIRECTIVE INT_OPERAND {/*Do nothing*/} + | TARGET_DIRECTIVE IDENTIFIER COMMA IDENTIFIER { target_header2($2,$4); } + | TARGET_DIRECTIVE IDENTIFIER COMMA IDENTIFIER COMMA IDENTIFIER { target_header3($2,$4,$6); } + | TARGET_DIRECTIVE IDENTIFIER { target_header($2); } + | FILE_DIRECTIVE INT_OPERAND STRING { add_file($2,$3); } + | FILE_DIRECTIVE INT_OPERAND STRING COMMA INT_OPERAND COMMA INT_OPERAND { add_file($2,$3); } + | LOC_DIRECTIVE INT_OPERAND INT_OPERAND INT_OPERAND + | PRAGMA_DIRECTIVE STRING SEMI_COLON { add_pragma($2); } + | function_decl SEMI_COLON {/*Do nothing*/} + ; + +variable_declaration: variable_spec identifier_list { add_variables(); } + | variable_spec identifier_spec EQUALS initializer_list { add_variables(); } + | variable_spec identifier_spec EQUALS literal_operand { add_variables(); } + | CONSTPTR_DIRECTIVE IDENTIFIER COMMA IDENTIFIER COMMA INT_OPERAND { add_constptr($2, $4, $6); } + ; + +variable_spec: var_spec_list { set_variable_type(); } + +identifier_list: identifier_spec + | identifier_list COMMA identifier_spec; + +identifier_spec: IDENTIFIER { add_identifier($1,0,NON_ARRAY_IDENTIFIER); func_header_info($1);} + | IDENTIFIER LEFT_ANGLE_BRACKET INT_OPERAND RIGHT_ANGLE_BRACKET { func_header_info($1); func_header_info_int("<", $3); func_header_info(">"); + int i,lbase,l; + char *id = NULL; + lbase = strlen($1); + for( i=0; i < $3; i++ ) { + l = lbase + (int)log10(i+1)+10; + id = (char*) malloc(l); + snprintf(id,l,"%s%u",$1,i); + add_identifier(id,0,NON_ARRAY_IDENTIFIER); + } + free($1); + } + | IDENTIFIER LEFT_SQUARE_BRACKET RIGHT_SQUARE_BRACKET { add_identifier($1,0,ARRAY_IDENTIFIER_NO_DIM); func_header_info($1); func_header_info("["); func_header_info("]");} + | IDENTIFIER LEFT_SQUARE_BRACKET INT_OPERAND RIGHT_SQUARE_BRACKET { add_identifier($1,$3,ARRAY_IDENTIFIER); func_header_info($1); func_header_info_int("[",$3); func_header_info("]");} + ; + +var_spec_list: var_spec + | var_spec_list var_spec; + +var_spec: space_spec + | type_spec + | align_spec + | EXTERN_DIRECTIVE { add_extern_spec(); } + | WEAK_DIRECTIVE + ; + +align_spec: ALIGN_DIRECTIVE INT_OPERAND { add_alignment_spec($2); } + +space_spec: REG_DIRECTIVE { add_space_spec(reg_space,0); } + | SREG_DIRECTIVE { add_space_spec(reg_space,0); } + | addressable_spec + ; + +addressable_spec: CONST_DIRECTIVE { add_space_spec(const_space,$1); } + | GLOBAL_DIRECTIVE { add_space_spec(global_space,0); } + | LOCAL_DIRECTIVE { add_space_spec(local_space,0); } + | PARAM_DIRECTIVE { add_space_spec(param_space_unclassified,0); } + | SHARED_DIRECTIVE { add_space_spec(shared_space,0); } + | SSTARR_DIRECTIVE { add_space_spec(sstarr_space,0); } + | SURF_DIRECTIVE { add_space_spec(surf_space,0); } + | TEX_DIRECTIVE { add_space_spec(tex_space,0); } + ; + +type_spec: scalar_type + | vector_spec scalar_type + ; + +vector_spec: V2_TYPE { add_option(V2_TYPE); func_header_info(".v2");} + | V3_TYPE { add_option(V3_TYPE); func_header_info(".v3");} + | V4_TYPE { add_option(V4_TYPE); func_header_info(".v4");} + ; + +scalar_type: S8_TYPE { add_scalar_type_spec( S8_TYPE ); } + | S16_TYPE { add_scalar_type_spec( S16_TYPE ); } + | S32_TYPE { add_scalar_type_spec( S32_TYPE ); } + | S64_TYPE { add_scalar_type_spec( S64_TYPE ); } + | U8_TYPE { add_scalar_type_spec( U8_TYPE ); } + | U16_TYPE { add_scalar_type_spec( U16_TYPE ); } + | U32_TYPE { add_scalar_type_spec( U32_TYPE ); } + | U64_TYPE { add_scalar_type_spec( U64_TYPE ); } + | F16_TYPE { add_scalar_type_spec( F16_TYPE ); } + | F32_TYPE { add_scalar_type_spec( F32_TYPE ); } + | F64_TYPE { add_scalar_type_spec( F64_TYPE ); } + | FF64_TYPE { add_scalar_type_spec( FF64_TYPE ); } + | B8_TYPE { add_scalar_type_spec( B8_TYPE ); } + | B16_TYPE { add_scalar_type_spec( B16_TYPE ); } + | B32_TYPE { add_scalar_type_spec( B32_TYPE ); } + | B64_TYPE { add_scalar_type_spec( B64_TYPE ); } + | BB64_TYPE { add_scalar_type_spec( BB64_TYPE ); } + | BB128_TYPE { add_scalar_type_spec( BB128_TYPE ); } + | PRED_TYPE { add_scalar_type_spec( PRED_TYPE ); } + | TEXREF_TYPE { add_scalar_type_spec( TEXREF_TYPE ); } + | SAMPLERREF_TYPE { add_scalar_type_spec( SAMPLERREF_TYPE ); } + | SURFREF_TYPE { add_scalar_type_spec( SURFREF_TYPE ); } + ; + +initializer_list: LEFT_BRACE literal_list RIGHT_BRACE { add_array_initializer(); } + | LEFT_BRACE initializer_list RIGHT_BRACE { syntax_not_implemented(); } + +literal_list: literal_operand + | literal_list COMMA literal_operand; + +instruction_statement: instruction SEMI_COLON + | IDENTIFIER COLON { add_label($1); } + | pred_spec instruction SEMI_COLON; + +instruction: opcode_spec LEFT_PAREN operand RIGHT_PAREN { set_return(); } COMMA operand COMMA LEFT_PAREN operand_list RIGHT_PAREN + | opcode_spec operand COMMA LEFT_PAREN operand_list RIGHT_PAREN + | opcode_spec operand COMMA LEFT_PAREN RIGHT_PAREN + | opcode_spec operand_list + | opcode_spec + ; + +opcode_spec: OPCODE { add_opcode($1); } option_list + | OPCODE { add_opcode($1); } + +pred_spec: PRED IDENTIFIER { add_pred($2,0, -1); } + | PRED EXCLAMATION IDENTIFIER { add_pred($3,1, -1); } + | PRED IDENTIFIER LT_OPTION { add_pred($2,0,1); } + | PRED IDENTIFIER EQ_OPTION { add_pred($2,0,2); } + | PRED IDENTIFIER LE_OPTION { add_pred($2,0,3); } + | PRED IDENTIFIER NE_OPTION { add_pred($2,0,5); } + | PRED IDENTIFIER GE_OPTION { add_pred($2,0,6); } + | PRED IDENTIFIER EQU_OPTION { add_pred($2,0,10); } + | PRED IDENTIFIER GTU_OPTION { add_pred($2,0,12); } + | PRED IDENTIFIER NEU_OPTION { add_pred($2,0,13); } + | PRED IDENTIFIER CF_OPTION { add_pred($2,0,17); } + | PRED IDENTIFIER SF_OPTION { add_pred($2,0,19); } + | PRED IDENTIFIER NSF_OPTION { add_pred($2,0,28); } + ; + +option_list: option + | option option_list ; + +option: type_spec + | compare_spec + | addressable_spec + | rounding_mode + | wmma_spec + | SYNC_OPTION { add_option(SYNC_OPTION); } + | ARRIVE_OPTION { add_option(ARRIVE_OPTION); } + | RED_OPTION { add_option(RED_OPTION); } + | UNI_OPTION { add_option(UNI_OPTION); } + | WIDE_OPTION { add_option(WIDE_OPTION); } + | ANY_OPTION { add_option(ANY_OPTION); } + | ALL_OPTION { add_option(ALL_OPTION); } + | BALLOT_OPTION { add_option(BALLOT_OPTION); } + | GLOBAL_OPTION { add_option(GLOBAL_OPTION); } + | CTA_OPTION { add_option(CTA_OPTION); } + | SYS_OPTION { add_option(SYS_OPTION); } + | GEOM_MODIFIER_1D { add_option(GEOM_MODIFIER_1D); } + | GEOM_MODIFIER_2D { add_option(GEOM_MODIFIER_2D); } + | GEOM_MODIFIER_3D { add_option(GEOM_MODIFIER_3D); } + | SAT_OPTION { add_option(SAT_OPTION); } + | FTZ_OPTION { add_option(FTZ_OPTION); } + | NEG_OPTION { add_option(NEG_OPTION); } + | APPROX_OPTION { add_option(APPROX_OPTION); } + | FULL_OPTION { add_option(FULL_OPTION); } + | EXIT_OPTION { add_option(EXIT_OPTION); } + | ABS_OPTION { add_option(ABS_OPTION); } + | atomic_operation_spec ; + | TO_OPTION { add_option(TO_OPTION); } + | HALF_OPTION { add_option(HALF_OPTION); } + | EXTP_OPTION { add_option(EXTP_OPTION); } + | CA_OPTION { add_option(CA_OPTION); } + | CG_OPTION { add_option(CG_OPTION); } + | CS_OPTION { add_option(CS_OPTION); } + | LU_OPTION { add_option(LU_OPTION); } + | CV_OPTION { add_option(CV_OPTION); } + | WB_OPTION { add_option(WB_OPTION); } + | WT_OPTION { add_option(WT_OPTION); } + | NC_OPTION { add_option(NC_OPTION); } + | UP_OPTION { add_option(UP_OPTION); } + | DOWN_OPTION { add_option(DOWN_OPTION); } + | BFLY_OPTION { add_option(BFLY_OPTION); } + | IDX_OPTION { add_option(IDX_OPTION); } + ; + +atomic_operation_spec: ATOMIC_AND { add_option(ATOMIC_AND); } + | ATOMIC_POPC { add_option(ATOMIC_POPC); } + | ATOMIC_OR { add_option(ATOMIC_OR); } + | ATOMIC_XOR { add_option(ATOMIC_XOR); } + | ATOMIC_CAS { add_option(ATOMIC_CAS); } + | ATOMIC_EXCH { add_option(ATOMIC_EXCH); } + | ATOMIC_ADD { add_option(ATOMIC_ADD); } + | ATOMIC_INC { add_option(ATOMIC_INC); } + | ATOMIC_DEC { add_option(ATOMIC_DEC); } + | ATOMIC_MIN { add_option(ATOMIC_MIN); } + | ATOMIC_MAX { add_option(ATOMIC_MAX); } + ; + +rounding_mode: floating_point_rounding_mode + | integer_rounding_mode; + + +floating_point_rounding_mode: RN_OPTION { add_option(RN_OPTION); } + | RZ_OPTION { add_option(RZ_OPTION); } + | RM_OPTION { add_option(RM_OPTION); } + | RP_OPTION { add_option(RP_OPTION); } + ; + +integer_rounding_mode: RNI_OPTION { add_option(RNI_OPTION); } + | RZI_OPTION { add_option(RZI_OPTION); } + | RMI_OPTION { add_option(RMI_OPTION); } + | RPI_OPTION { add_option(RPI_OPTION); } + ; + +compare_spec:EQ_OPTION { add_option(EQ_OPTION); } + | NE_OPTION { add_option(NE_OPTION); } + | LT_OPTION { add_option(LT_OPTION); } + | LE_OPTION { add_option(LE_OPTION); } + | GT_OPTION { add_option(GT_OPTION); } + | GE_OPTION { add_option(GE_OPTION); } + | LO_OPTION { add_option(LO_OPTION); } + | LS_OPTION { add_option(LS_OPTION); } + | HI_OPTION { add_option(HI_OPTION); } + | HS_OPTION { add_option(HS_OPTION); } + | EQU_OPTION { add_option(EQU_OPTION); } + | NEU_OPTION { add_option(NEU_OPTION); } + | LTU_OPTION { add_option(LTU_OPTION); } + | LEU_OPTION { add_option(LEU_OPTION); } + | GTU_OPTION { add_option(GTU_OPTION); } + | GEU_OPTION { add_option(GEU_OPTION); } + | NUM_OPTION { add_option(NUM_OPTION); } + | NAN_OPTION { add_option(NAN_OPTION); } + ; + +wmma_spec: WMMA_DIRECTIVE LAYOUT CONFIGURATION{add_wmma_option($1);add_wmma_option($2);add_wmma_option($3);} + | WMMA_DIRECTIVE LAYOUT LAYOUT CONFIGURATION{add_wmma_option($1);add_wmma_option($2),add_wmma_option($3),add_wmma_option($4)} + | WMMA_DIRECTIVE LAYOUT CONFIGURATION ptr_space_spec{add_wmma_option($1);add_wmma_option($2),add_wmma_option($3),add_wmma_option($4)} + | WMMA_DIRECTIVE LAYOUT LAYOUT CONFIGURATION ptr_space_spec{add_wmma_option($1);add_wmma_option($2),add_wmma_option($3),add_wmma_option($4)} + ; + +operand_list: operand + | operand COMMA operand_list; + +operand: IDENTIFIER { add_scalar_operand( $1 ); } + | EXCLAMATION IDENTIFIER { add_neg_pred_operand( $2 ); } + | MINUS IDENTIFIER { add_scalar_operand( $2 ); change_operand_neg(); } + | memory_operand + | literal_operand + | builtin_operand + | vector_operand + | MINUS vector_operand { change_operand_neg(); } + | tex_operand + | IDENTIFIER PLUS INT_OPERAND { add_address_operand($1,$3); } + | IDENTIFIER LO_OPTION { add_scalar_operand( $1 ); change_operand_lohi(1);} + | MINUS IDENTIFIER LO_OPTION { add_scalar_operand( $2 ); change_operand_lohi(1); change_operand_neg();} + | IDENTIFIER HI_OPTION { add_scalar_operand( $1 ); change_operand_lohi(2);} + | MINUS IDENTIFIER HI_OPTION { add_scalar_operand( $2 ); change_operand_lohi(2); change_operand_neg();} + | IDENTIFIER PIPE IDENTIFIER { add_2vector_operand($1,$3); change_double_operand_type(-1);} + | IDENTIFIER PIPE IDENTIFIER LO_OPTION { add_2vector_operand($1,$3); change_double_operand_type(-1); change_operand_lohi(1);} + | IDENTIFIER PIPE IDENTIFIER HI_OPTION { add_2vector_operand($1,$3); change_double_operand_type(-1); change_operand_lohi(2);} + | IDENTIFIER BACKSLASH IDENTIFIER { add_2vector_operand($1,$3); change_double_operand_type(-3);} + | IDENTIFIER BACKSLASH IDENTIFIER LO_OPTION { add_2vector_operand($1,$3); change_double_operand_type(-3); change_operand_lohi(1);} + | IDENTIFIER BACKSLASH IDENTIFIER HI_OPTION { add_2vector_operand($1,$3); change_double_operand_type(-3); change_operand_lohi(2);} + ; + +vector_operand: LEFT_BRACE IDENTIFIER COMMA IDENTIFIER RIGHT_BRACE { add_2vector_operand($2,$4); } + | LEFT_BRACE IDENTIFIER COMMA IDENTIFIER COMMA IDENTIFIER RIGHT_BRACE { add_3vector_operand($2,$4,$6); } + | LEFT_BRACE IDENTIFIER COMMA IDENTIFIER COMMA IDENTIFIER COMMA IDENTIFIER RIGHT_BRACE { add_4vector_operand($2,$4,$6,$8); } + | LEFT_BRACE IDENTIFIER COMMA IDENTIFIER COMMA IDENTIFIER COMMA IDENTIFIER COMMA IDENTIFIER COMMA IDENTIFIER COMMA IDENTIFIER COMMA IDENTIFIER RIGHT_BRACE { add_8vector_operand($2,$4,$6,$8,$10,$12,$14,$16); } + | LEFT_BRACE IDENTIFIER RIGHT_BRACE { add_1vector_operand($2); } + ; + +tex_operand: LEFT_SQUARE_BRACKET IDENTIFIER COMMA { add_scalar_operand($2); } + vector_operand + RIGHT_SQUARE_BRACKET + ; + +builtin_operand: SPECIAL_REGISTER DIMENSION_MODIFIER { add_builtin_operand($1,$2); } + | SPECIAL_REGISTER { add_builtin_operand($1,-1); } + ; + +memory_operand : LEFT_SQUARE_BRACKET address_expression RIGHT_SQUARE_BRACKET { add_memory_operand(); } + | IDENTIFIER LEFT_SQUARE_BRACKET address_expression RIGHT_SQUARE_BRACKET { add_memory_operand(); change_memory_addr_space($1); } + | IDENTIFIER LEFT_SQUARE_BRACKET literal_operand RIGHT_SQUARE_BRACKET { change_memory_addr_space($1); } + | IDENTIFIER LEFT_SQUARE_BRACKET twin_operand RIGHT_SQUARE_BRACKET { change_memory_addr_space($1); add_memory_operand();} + | MINUS memory_operand { change_operand_neg(); } + ; + +twin_operand : IDENTIFIER PLUS IDENTIFIER { add_double_operand($1,$3); change_double_operand_type(1); } + | IDENTIFIER PLUS IDENTIFIER LO_OPTION { add_double_operand($1,$3); change_double_operand_type(1); change_operand_lohi(1); } + | IDENTIFIER PLUS IDENTIFIER HI_OPTION { add_double_operand($1,$3); change_double_operand_type(1); change_operand_lohi(2); } + | IDENTIFIER PLUS EQUALS IDENTIFIER { add_double_operand($1,$4); change_double_operand_type(2); } + | IDENTIFIER PLUS EQUALS IDENTIFIER LO_OPTION { add_double_operand($1,$4); change_double_operand_type(2); change_operand_lohi(1); } + | IDENTIFIER PLUS EQUALS IDENTIFIER HI_OPTION { add_double_operand($1,$4); change_double_operand_type(2); change_operand_lohi(2); } + | IDENTIFIER PLUS EQUALS INT_OPERAND { add_address_operand($1,$4); change_double_operand_type(3); } + ; + +literal_operand : INT_OPERAND { add_literal_int($1); } + | FLOAT_OPERAND { add_literal_float($1); } + | DOUBLE_OPERAND { add_literal_double($1); } + ; + +address_expression: IDENTIFIER { add_address_operand($1,0); } + | IDENTIFIER LO_OPTION { add_address_operand($1,0); change_operand_lohi(1);} + | IDENTIFIER HI_OPTION { add_address_operand($1,0); change_operand_lohi(2); } + | IDENTIFIER PLUS INT_OPERAND { add_address_operand($1,$3); } + | INT_OPERAND { add_address_operand2($1); } + ; + +%% + +extern int ptx_lineno; +extern const char *g_filename; + +void syntax_not_implemented() +{ + printf("Parse error (%s:%u): this syntax is not (yet) implemented:\n",g_filename,ptx_lineno); + ptx_error(NULL); + abort(); +} diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 833f175..16cc975 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1078,7 +1078,7 @@ public: int membar_level() const { return m_membar_level; } bool has_memory_read() const { - if( m_opcode == LD_OP || m_opcode == LDU_OP || m_opcode == TEX_OP ) + if( m_opcode == LD_OP || m_opcode == LDU_OP || m_opcode == TEX_OP|| m_opcode==MMA_LD_OP) return true; // Check PTXPlus operand type below // Source operands are memory operands @@ -1090,7 +1090,7 @@ public: return false; } bool has_memory_write() const { - if( m_opcode == ST_OP ) return true; + if( m_opcode == ST_OP || m_opcode==MMA_ST_OP ) return true; // Check PTXPlus operand type below // Destination operand is a memory operand ptx_instruction::const_iterator op=op_iter_begin(); diff --git a/src/cuda-sim/ptx_sim.h b/src/cuda-sim/ptx_sim.h index 05acf20..403ce5b 100644 --- a/src/cuda-sim/ptx_sim.h +++ b/src/cuda-sim/ptx_sim.h @@ -303,6 +303,15 @@ public: const ptx_reg_t &data2, const ptx_reg_t &data3, const ptx_reg_t &data4 ); + void set_wmma_vector_operand_values( const operand_info &dst, + const ptx_reg_t &data1, + const ptx_reg_t &data2, + const ptx_reg_t &data3, + const ptx_reg_t &data4, + const ptx_reg_t &data5, + const ptx_reg_t &data6, + const ptx_reg_t &data7, + const ptx_reg_t &data8 ); function_info *func_info() { diff --git a/src/gpgpu-sim/Makefile b/src/gpgpu-sim/Makefile index f10a8a4..4f77699 100644 --- a/src/gpgpu-sim/Makefile +++ b/src/gpgpu-sim/Makefile @@ -48,7 +48,7 @@ ifeq ($(GNUC_CPP0X), 1) endif ifneq ($(DEBUG),1) - OPTFLAGS += -O3 + OPTFLAGS += -O0 else CXXFLAGS += endif diff --git a/src/intersim2/Makefile b/src/intersim2/Makefile index bd42000..4ef21ac 100644 --- a/src/intersim2/Makefile +++ b/src/intersim2/Makefile @@ -44,7 +44,7 @@ endif CPPFLAGS += -Wall $(INCPATH) $(DEFINE) ifneq ($(DEBUG),1) -CPPFLAGS += -O3 +CPPFLAGS += -O0 endif CPPFLAGS += -g CPPFLAGS += -fPIC -- cgit v1.3 From 774e28201d239433357db4cbd00c0a62273773cc Mon Sep 17 00:00:00 2001 From: Aamir Raihan Date: Fri, 8 Jun 2018 13:56:54 -0700 Subject: regression passed --- src/Makefile | 2 +- src/cuda-sim/Makefile | 2 +- src/cuda-sim/instructions.cc | 11 ----------- src/cuda-sim/ptx.l | 4 ++-- src/gpgpu-sim/Makefile | 2 +- src/intersim2/Makefile | 2 +- 6 files changed, 6 insertions(+), 17 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/src/Makefile b/src/Makefile index 09194f3..6001669 100644 --- a/src/Makefile +++ b/src/Makefile @@ -46,7 +46,7 @@ ifeq ($(TRACE),1) endif ifneq ($(DEBUG),1) - OPTFLAGS += -O0 + OPTFLAGS += -O3 else CXXFLAGS += endif diff --git a/src/cuda-sim/Makefile b/src/cuda-sim/Makefile index d08a696..e8977a6 100644 --- a/src/cuda-sim/Makefile +++ b/src/cuda-sim/Makefile @@ -42,7 +42,7 @@ include ../../version_detection.mk OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/cuda-sim -OPT := -O0 -g3 -Wall -Wno-unused-function -Wno-sign-compare +OPT := -O3 -g3 -Wall -Wno-unused-function -Wno-sign-compare ifeq ($(DEBUG),1) OPT := -g3 -Wall -Wno-unused-function -Wno-sign-compare endif diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 59415f4..92c8529 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -835,17 +835,6 @@ void add_impl( const ptx_instruction *pI, ptx_thread_info *thread ) unsigned i_type = pI->get_type(); src1_data = thread->get_operand_value(src1, dst, i_type, thread, 1); src2_data = thread->get_operand_value(src2, dst, i_type, thread, 1); - //unsigned warpId_aa,warp_size_aa; - //warpId_aa = pI->warp_id(); - //warp_size_aa=32; - //dim3 t=thread->get_tid(); - //unsigned tid_aa=warp_size_aa*warpId_aa; - - ptx_thread_info *thread2; - thread2=thread; - src1_data = thread2->get_operand_value(src1, dst, i_type, thread2, 1); - src2_data = thread2->get_operand_value(src2, dst, i_type, thread2, 1); - unsigned rounding_mode = pI->rounding_mode(); int orig_rm = fegetround(); diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l index f67dddd..03d6838 100644 --- a/src/cuda-sim/ptx.l +++ b/src/cuda-sim/ptx.l @@ -68,7 +68,7 @@ bra TC; ptx_lval.int_value = BRA_OP; return OPCODE; brx TC; ptx_lval.int_value = BRX_OP; return OPCODE; brev TC; ptx_lval.int_value = BREV_OP; return OPCODE; brkpt TC; ptx_lval.int_value = BRKPT_OP; return OPCODE; -wmma\.mma TC; ptx_lval.int_value = MMA_OP; return OPCODE; +wmma TC; ptx_lval.int_value = MMA_OP; return OPCODE; wmma\.load TC; ptx_lval.int_value = MMA_LD_OP; return OPCODE; wmma\.store TC; ptx_lval.int_value = MMA_ST_OP; return OPCODE; @@ -160,7 +160,7 @@ breakaddr TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE; \.b\.sync TC; ptx_lval.int_value = LOAD_B; return WMMA_DIRECTIVE; \.c\.sync TC; ptx_lval.int_value = LOAD_C; return WMMA_DIRECTIVE; \.d\.sync TC; ptx_lval.int_value = STORE_D; return WMMA_DIRECTIVE; -\.sync TC;ptx_lval.int_value=MMA; return WMMA_DIRECTIVE; +\.mma\.sync TC;ptx_lval.int_value=MMA; return WMMA_DIRECTIVE; \.row TC; ptx_lval.int_value = ROW; return LAYOUT; \.col TC; ptx_lval.int_value = COL; return LAYOUT; \.m16n16k16 TC; ptx_lval.int_value = M16N16K16; return CONFIGURATION; diff --git a/src/gpgpu-sim/Makefile b/src/gpgpu-sim/Makefile index 4f77699..f10a8a4 100644 --- a/src/gpgpu-sim/Makefile +++ b/src/gpgpu-sim/Makefile @@ -48,7 +48,7 @@ ifeq ($(GNUC_CPP0X), 1) endif ifneq ($(DEBUG),1) - OPTFLAGS += -O0 + OPTFLAGS += -O3 else CXXFLAGS += endif diff --git a/src/intersim2/Makefile b/src/intersim2/Makefile index 4ef21ac..bd42000 100644 --- a/src/intersim2/Makefile +++ b/src/intersim2/Makefile @@ -44,7 +44,7 @@ endif CPPFLAGS += -Wall $(INCPATH) $(DEFINE) ifneq ($(DEBUG),1) -CPPFLAGS += -O0 +CPPFLAGS += -O3 endif CPPFLAGS += -g CPPFLAGS += -fPIC -- cgit v1.3 From 262663ac90d2aa801d6af1eb9bf8a75ee9a5bb18 Mon Sep 17 00:00:00 2001 From: negargoli93 Date: Wed, 20 Jun 2018 18:20:05 -0700 Subject: Tensor core timing model --- cuda-kernels/gpgpusim.config | 7 +++-- src/abstract_hardware_model.h | 2 ++ src/cuda-sim/cuda-sim.cc | 5 ++-- src/gpgpu-sim/gpu-sim.cc | 16 ++++++++-- src/gpgpu-sim/scoreboard.cc | 40 +++++++++++++++++++------ src/gpgpu-sim/shader.cc | 70 +++++++++++++++++++++++++++++++++++++++---- src/gpgpu-sim/shader.h | 47 ++++++++++++++++++++++++++--- 7 files changed, 163 insertions(+), 24 deletions(-) (limited to 'src/gpgpu-sim') 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* 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* 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* 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* 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* 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* 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* 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)); -- cgit v1.3 From d907c7d848be6ced2b7f2bd2df84b39e57dfbedc Mon Sep 17 00:00:00 2001 From: negargoli Date: Fri, 22 Jun 2018 11:09:18 -0700 Subject: No need to change the config file for adding tensor-core --- cuda-kernels/gpgpusim.config | 9 ++++----- src/gpgpu-sim/shader.cc | 8 ++++---- src/gpgpu-sim/shader.h | 16 +++++++++++----- 3 files changed, 19 insertions(+), 14 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/cuda-kernels/gpgpusim.config b/cuda-kernels/gpgpusim.config index 69a110f..272ad3d 100755 --- a/cuda-kernels/gpgpusim.config +++ b/cuda-kernels/gpgpusim.config @@ -33,10 +33,9 @@ # 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 4,1,1,1,4,1,1,1,,6 +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 -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 @@ -72,14 +71,14 @@ ## 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_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_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/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index fcac755..c01f867 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -218,7 +218,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, 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(TENSOR_CORE_CUS, config->gpgpu_operand_collector_num_units_tensor_core, 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); @@ -243,7 +243,7 @@ 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_tensor_core; i++) { + for (unsigned i = 0; i < 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); @@ -280,7 +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 + m_config->gpgpu_num_tensor_core_units + 1; // sp_unit, sfu, ldst_unit + m_num_function_units = m_config->gpgpu_num_sp_units + m_config->gpgpu_num_sfu_units + 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 ]; @@ -298,7 +298,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_issue_port.push_back(OC_EX_SFU); } - for (int k = 0; k < m_config->gpgpu_num_tensor_core_units; k++) { + for (int k = 0; k < 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); diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 2c4c43d..b7deae6 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1224,15 +1224,15 @@ 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, + ID_OC_TENSOR_CORE, + OC_EX_TENSOR_CORE, N_PIPELINE_STAGES -}; + }; const char* const pipeline_stage_name_decode[] = { "ID_OC_SP", @@ -1266,9 +1266,11 @@ struct shader_core_config : public core_config char* toks = new char[100]; char* tokd = toks; strcpy(toks,pipeline_widths_string); - + toks = strtok(toks,","); - for (unsigned i = 0; i < N_PIPELINE_STAGES; i++) { + pipe_widths[OC_EX_TENSOR_CORE]=1; + pipe_widths[ID_OC_TENSOR_CORE]=1; + for (unsigned i = 0; i < N_PIPELINE_STAGES-2; i++) { assert(toks); ntok = sscanf(toks,"%d", &pipe_widths[i]); assert(ntok == 1); @@ -1286,6 +1288,10 @@ struct shader_core_config : public core_config max_sfu_latency = 512; max_tensor_core_latency = 512; max_sp_latency = 32; + gpgpu_num_tensor_core_units=1; + gpgpu_operand_collector_num_units_tensor_core=24; + gpgpu_operand_collector_num_in_ports_tensor_core=1; + gpgpu_operand_collector_num_out_ports_tensor_core=1; m_L1I_config.init(m_L1I_config.m_config_string,FuncCachePreferNone); m_L1T_config.init(m_L1T_config.m_config_string,FuncCachePreferNone); m_L1C_config.init(m_L1C_config.m_config_string,FuncCachePreferNone); -- cgit v1.3 From e541026cfc0ee4be25e7093cb7ff3acfa3cbb6e7 Mon Sep 17 00:00:00 2001 From: negargoli93 Date: Mon, 16 Jul 2018 13:18:23 -0700 Subject: fix pipline for tensor_core and change config --- cuda-kernels/config_fermi_islip.icnt | 2 +- cuda-kernels/gpgpusim.config | 10 +++++----- src/cuda-sim/cuda-sim.cc | 8 ++++++-- src/gpgpu-sim/gpu-sim.cc | 2 +- src/gpgpu-sim/shader.cc | 10 +++++----- src/gpgpu-sim/shader.h | 18 +++++++++--------- 6 files changed, 27 insertions(+), 23 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/cuda-kernels/config_fermi_islip.icnt b/cuda-kernels/config_fermi_islip.icnt index a788090..3b8b496 100755 --- a/cuda-kernels/config_fermi_islip.icnt +++ b/cuda-kernels/config_fermi_islip.icnt @@ -7,7 +7,7 @@ network_count = 2; // Topology topology = fly; -k = 62; +k = 102; n = 1; // Routing diff --git a/cuda-kernels/gpgpusim.config b/cuda-kernels/gpgpusim.config index 272ad3d..2510d21 100755 --- a/cuda-kernels/gpgpusim.config +++ b/cuda-kernels/gpgpusim.config @@ -10,7 +10,7 @@ -gpgpu_ptx_save_converted_ptxplus 0 # high level architecture configuration --gpgpu_n_clusters 40 +-gpgpu_n_clusters 80 -gpgpu_n_cores_per_cluster 1 -gpgpu_n_mem 11 -gpgpu_n_sub_partition_per_mchannel 2 @@ -33,7 +33,7 @@ # 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 # Instruction latencies and initiation intervals @@ -72,15 +72,15 @@ -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_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_operand_collector_num_in_ports_mem 10 +-gpgpu_operand_collector_num_out_ports_mem 10 # gpgpu_num_reg_banks should be increased to 32, but it gives an error! -gpgpu_num_reg_banks 32 diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 6da0840..506bc95 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -648,8 +648,10 @@ void ptx_instruction::set_opcode_and_latency() if ( has_memory_write() ) op = STORE_OP; break; case LD_OP: op = LOAD_OP; break; + case MMA_LD_OP: op = LOAD_OP; break; case LDU_OP: op = LOAD_OP; break; case ST_OP: op = STORE_OP; break; + case MMA_ST_OP: op = STORE_OP; break; case BRA_OP: op = BRANCH_OP; break; case BREAKADDR_OP: op = BRANCH_OP; break; case TEX_OP: op = LOAD_OP; mem_op=TEX; break; @@ -897,9 +899,11 @@ void ptx_instruction::pre_decode() case WB_OPTION: cache_op = CACHE_WRITE_BACK; break; case WT_OPTION: cache_op = CACHE_WRITE_THROUGH; break; default: - if( m_opcode == LD_OP || m_opcode == LDU_OP ) + //if( m_opcode == LD_OP || m_opcode == LDU_OP ) + if( m_opcode == MMA_LD_OP || m_opcode == LD_OP || m_opcode == LDU_OP ) cache_op = CACHE_ALL; - else if( m_opcode == ST_OP ) + //else if( m_opcode == ST_OP ) + else if( m_opcode == ST_OP || m_opcode == ST_OP ) cache_op = CACHE_WRITE_BACK; else if( m_opcode == ATOM_OP ) cache_op = CACHE_GLOBAL; diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index cc23051..3e064c7 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1541,7 +1541,7 @@ void gpgpu_sim::cycle() } } - if (!(gpu_sim_cycle % 20000)) { + if (!(gpu_sim_cycle % 50000)) { // deadlock detection if (m_config.gpu_deadlock_detect && gpu_sim_insn == last_gpu_sim_insn) { gpu_deadlock = true; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index c01f867..226e7f0 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1387,7 +1387,7 @@ ldst_unit::process_cache_access( cache_t* cache, assert( !read_sent ); inst.accessq_pop_back(); if ( inst.is_load() ) { - for ( unsigned r=0; r < 4; r++) + for ( unsigned r=0; r < 8; r++) if (inst.out[r] > 0) m_pending_writes[inst.warp_id()][inst.out[r]]--; } @@ -1489,7 +1489,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea inst.accessq_pop_back(); //inst.clear_active( access.get_warp_mask() ); if( inst.is_load() ) { - for( unsigned r=0; r < 4; r++) + for( unsigned r=0; r < 8; r++) if(inst.out[r] > 0) assert( m_pending_writes[inst.warp_id()][inst.out[r]] > 0 ); } else if( inst.is_store() ) @@ -1767,7 +1767,7 @@ void ldst_unit:: issue( register_set ®_set ) if (inst->is_load() and inst->space.get_type() != shared_space) { unsigned warp_id = inst->warp_id(); unsigned n_accesses = inst->accessq_count(); - for (unsigned r = 0; r < 4; r++) { + for (unsigned r = 0; r < 8; r++) { unsigned reg_id = inst->out[r]; if (reg_id > 0) { m_pending_writes[warp_id][reg_id] += n_accesses; @@ -1789,7 +1789,7 @@ void ldst_unit::writeback() if( !m_next_wb.empty() ) { if( m_operand_collector->writeback(m_next_wb) ) { bool insn_completed = false; - for( unsigned r=0; r < 4; r++ ) { + for( unsigned r=0; r < 8; r++ ) { if( m_next_wb.out[r] > 0 ) { if( m_next_wb.space.get_type() != shared_space ) { assert( m_pending_writes[m_next_wb.warp_id()][m_next_wb.out[r]] > 0 ); @@ -1991,7 +1991,7 @@ void ldst_unit::cycle() //} bool pending_requests=false; - for( unsigned r=0; r<4; r++ ) { + for( unsigned r=0; r<8; r++ ) { unsigned reg_id = pipe_reg.out[r]; if( reg_id > 0 ) { if( m_pending_writes[warp_id].find(reg_id) != m_pending_writes[warp_id].end() ) { diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index b7deae6..90a3134 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1224,13 +1224,13 @@ 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, - ID_OC_TENSOR_CORE, - OC_EX_TENSOR_CORE, N_PIPELINE_STAGES }; @@ -1268,9 +1268,9 @@ struct shader_core_config : public core_config strcpy(toks,pipeline_widths_string); toks = strtok(toks,","); - pipe_widths[OC_EX_TENSOR_CORE]=1; - pipe_widths[ID_OC_TENSOR_CORE]=1; - for (unsigned i = 0; i < N_PIPELINE_STAGES-2; i++) { + // pipe_widths[OC_EX_TENSOR_CORE]=1; + // pipe_widths[ID_OC_TENSOR_CORE]=1; + for (unsigned i = 0; i < N_PIPELINE_STAGES; i++) { assert(toks); ntok = sscanf(toks,"%d", &pipe_widths[i]); assert(ntok == 1); @@ -1286,12 +1286,12 @@ 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_tensor_core_latency = 64; max_sp_latency = 32; - gpgpu_num_tensor_core_units=1; + gpgpu_num_tensor_core_units=8; gpgpu_operand_collector_num_units_tensor_core=24; - gpgpu_operand_collector_num_in_ports_tensor_core=1; - gpgpu_operand_collector_num_out_ports_tensor_core=1; + gpgpu_operand_collector_num_in_ports_tensor_core=8; + gpgpu_operand_collector_num_out_ports_tensor_core=8; m_L1I_config.init(m_L1I_config.m_config_string,FuncCachePreferNone); m_L1T_config.init(m_L1T_config.m_config_string,FuncCachePreferNone); m_L1C_config.init(m_L1C_config.m_config_string,FuncCachePreferNone); -- cgit v1.3 From 9e7cd8867d76fb99eadfadfa09947ff057d012d3 Mon Sep 17 00:00:00 2001 From: negargoli93 Date: Thu, 16 Aug 2018 15:20:07 -0700 Subject: Timing model for VCORE --- src/abstract_hardware_model.h | 2 ++ src/cuda-sim/cuda-sim.cc | 28 ++++++++++++++++--- src/cuda-sim/ptx_ir.h | 4 +-- src/gpgpu-sim/gpu-sim.cc | 16 +++++++++-- src/gpgpu-sim/shader.cc | 62 ++++++++++++++++++++++++++++++++++++++++--- src/gpgpu-sim/shader.h | 57 ++++++++++++++++++++++++++++++++++----- 6 files changed, 152 insertions(+), 17 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 7fe5d82..781509f 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -78,6 +78,7 @@ enum uarch_op_t { ALU_OP=1, SFU_OP, TENSOR_CORE_OP, + VP_CORE_OP, ALU_SFU_OP, LOAD_OP, STORE_OP, @@ -135,6 +136,7 @@ enum operation_pipeline_t { SP__OP, SFU__OP, TENSOR_CORE__OP, + VP_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 2fe5667..1ad12ee 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -542,7 +542,7 @@ void ptx_instruction::set_mul_div_or_other_archop(){ sp_op=INT_DIV_OP; break; default: - if((op==ALU_OP)||(op==TENSOR_CORE_OP)) + if((op==ALU_OP)||(op==VP_CORE_OP)) sp_op=INT__OP; break; } @@ -649,9 +649,11 @@ void ptx_instruction::set_opcode_and_latency() break; case LD_OP: op = LOAD_OP; break; case MMA_LD_OP: op = LOAD_OP; break; + case VP_LD_OP: op = LOAD_OP; break; case LDU_OP: op = LOAD_OP; break; case ST_OP: op = STORE_OP; break; case MMA_ST_OP: op = STORE_OP; break; + case VP_ST_OP: op = STORE_OP; break; case BRA_OP: op = BRANCH_OP; break; case BREAKADDR_OP: op = BRANCH_OP; break; case TEX_OP: op = LOAD_OP; mem_op=TEX; break; @@ -799,6 +801,26 @@ void ptx_instruction::set_opcode_and_latency() initiation_interval = 64; op=TENSOR_CORE_OP; break; + case VP_MMA_OP: + if(get_wmma_type()==VP_MMA4) + { + latency = 5; + initiation_interval = 5; + } + if(get_wmma_type()==VP_MMA8) + { + latency = 5; + initiation_interval = 5; + } + if(get_wmma_type()==VP_MMA16) + { + latency = 5; + initiation_interval = 5; + } + op=VP_CORE_OP; + op=VP_CORE_OP; + op=VP_CORE_OP; + break; case SHFL_OP: latency = 32; initiation_interval = 15; @@ -900,10 +922,10 @@ void ptx_instruction::pre_decode() case WT_OPTION: cache_op = CACHE_WRITE_THROUGH; break; default: //if( m_opcode == LD_OP || m_opcode == LDU_OP ) - if( m_opcode == MMA_LD_OP || m_opcode == LD_OP || m_opcode == LDU_OP ) + if( m_opcode ==VP_LD_OP || m_opcode == MMA_LD_OP || m_opcode == LD_OP || m_opcode == LDU_OP ) cache_op = CACHE_ALL; //else if( m_opcode == ST_OP ) - else if( m_opcode == ST_OP || m_opcode == ST_OP ) + else if( m_opcode == VP_ST_OP ||m_opcode == MMA_ST_OP || m_opcode == ST_OP ) cache_op = CACHE_WRITE_BACK; else if( m_opcode == ATOM_OP ) cache_op = CACHE_GLOBAL; diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index cb4556e..e025013 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1093,7 +1093,7 @@ public: int membar_level() const { return m_membar_level; } bool has_memory_read() const { - if( m_opcode == LD_OP || m_opcode == LDU_OP || m_opcode == TEX_OP|| m_opcode==MMA_LD_OP) + if( m_opcode == LD_OP || m_opcode == LDU_OP || m_opcode == TEX_OP|| m_opcode==MMA_LD_OP || m_opcode==VP_LD_OP) return true; // Check PTXPlus operand type below // Source operands are memory operands @@ -1105,7 +1105,7 @@ public: return false; } bool has_memory_write() const { - if( m_opcode == ST_OP || m_opcode==MMA_ST_OP ) return true; + if( m_opcode == ST_OP || m_opcode==MMA_ST_OP || m_opcode==VP_ST_OP ) return true; // Check PTXPlus operand type below // Destination operand is a memory operand ptx_instruction::const_iterator op=op_iter_begin(); diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 3e064c7..7a797b5 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -309,6 +309,9 @@ void shader_core_config::reg_options(class OptionParser * opp) 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_vp_core", OPT_INT32, &gpgpu_operand_collector_num_units_vp_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"); @@ -324,6 +327,9 @@ void shader_core_config::reg_options(class OptionParser * opp) 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_vp_core", OPT_INT32, &gpgpu_operand_collector_num_in_ports_vp_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"); @@ -339,6 +345,9 @@ void shader_core_config::reg_options(class OptionParser * opp) 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_vp_core", OPT_INT32, &gpgpu_operand_collector_num_out_ports_vp_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"); @@ -359,8 +368,8 @@ 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_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" ); + "ID_OC_SP,ID_OC_SFU,ID_OC_TENSOR_CORE,ID_OC_VP_CORE,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_TENSOR_CORE,OC_EX_TENSOR_CORE,OC_EX_MEM,EX_WB", + "1,1,1,1,1,1,1,1,1,1,1" ); option_parser_register(opp, "-gpgpu_num_sp_units", OPT_INT32, &gpgpu_num_sp_units, "Number of SP units (default=1)", "1"); @@ -370,6 +379,9 @@ void shader_core_config::reg_options(class OptionParser * opp) 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_vp_core_units", OPT_INT32, &gpgpu_num_vp_core_units, + "Number of vp_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/shader.cc b/src/gpgpu-sim/shader.cc index 226e7f0..6f11ad9 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -149,6 +149,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_TENSOR_CORE], + &m_pipeline_reg[ID_OC_VP_CORE], &m_pipeline_reg[ID_OC_MEM], i ) @@ -164,6 +165,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_TENSOR_CORE], + &m_pipeline_reg[ID_OC_VP_CORE], &m_pipeline_reg[ID_OC_MEM], i, config->gpgpu_scheduler_string @@ -180,6 +182,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_TENSOR_CORE], + &m_pipeline_reg[ID_OC_VP_CORE], &m_pipeline_reg[ID_OC_MEM], i ) @@ -195,6 +198,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, &m_pipeline_reg[ID_OC_SP], &m_pipeline_reg[ID_OC_SFU], &m_pipeline_reg[ID_OC_TENSOR_CORE], + &m_pipeline_reg[ID_OC_VP_CORE], &m_pipeline_reg[ID_OC_MEM], i, config->gpgpu_scheduler_string @@ -215,10 +219,11 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, } //op collector configuration - enum { SP_CUS, SFU_CUS, TENSOR_CORE_CUS, MEM_CUS, GEN_CUS }; + enum { SP_CUS, SFU_CUS, TENSOR_CORE_CUS, VP_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, config->gpgpu_operand_collector_num_units_tensor_core, config->gpgpu_operand_collector_num_out_ports_tensor_core); + m_operand_collector.add_cu_set(VP_CORE_CUS, config->gpgpu_operand_collector_num_units_vp_core, config->gpgpu_operand_collector_num_out_ports_vp_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); @@ -252,6 +257,14 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, in_ports.clear(),out_ports.clear(),cu_sets.clear(); } + for (unsigned i = 0; i < config->gpgpu_operand_collector_num_in_ports_vp_core; i++) { + in_ports.push_back(&m_pipeline_reg[ID_OC_VP_CORE]); + out_ports.push_back(&m_pipeline_reg[OC_EX_VP_CORE]); + cu_sets.push_back((unsigned)VP_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]); @@ -267,10 +280,12 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, 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_VP_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_VP_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); @@ -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 + config->gpgpu_num_tensor_core_units + 1; // sp_unit, sfu, ldst_unit + m_num_function_units = m_config->gpgpu_num_sp_units + m_config->gpgpu_num_sfu_units + config->gpgpu_num_tensor_core_units + config->gpgpu_num_vp_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 ]; @@ -304,6 +319,12 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_issue_port.push_back(OC_EX_TENSOR_CORE); } + for (int k = 0; k < config->gpgpu_num_vp_core_units; k++) { + m_fu.push_back(new vp_core( &m_pipeline_reg[EX_WB], m_config, this )); + m_dispatch_port.push_back(ID_OC_VP_CORE); + m_issue_port.push_back(OC_EX_VP_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); m_dispatch_port.push_back(ID_OC_MEM); @@ -910,7 +931,8 @@ void scheduler_unit::cycle() bool sp_pipe_avail = m_sp_out->has_free(); bool sfu_pipe_avail = m_sfu_out->has_free(); bool tensor_core_pipe_avail = m_tensor_core_out->has_free(); - if( sp_pipe_avail && (pI->op != SFU_OP) && (pI->op != TENSOR_CORE_OP)) { + bool vp_core_pipe_avail = m_vp_core_out->has_free(); + if( sp_pipe_avail && (pI->op != SFU_OP) && (pI->op != TENSOR_CORE_OP) && (pI->op !=VP_CORE_OP)) { //Jin: special for CDP api if(pI->m_is_cdp && !warp(warp_id).m_cdp_dummy) { @@ -950,6 +972,14 @@ void scheduler_unit::cycle() issued_inst=true; warp_inst_issued = true; } + } + else if ( (pI->op == VP_CORE_OP) ) { + if( vp_core_pipe_avail ) { + m_shader->issue_warp(*m_vp_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", @@ -1116,10 +1146,11 @@ swl_scheduler::swl_scheduler ( shader_core_stats* stats, shader_core_ctx* shader register_set* sp_out, register_set* sfu_out, register_set* tensor_core_out, + register_set* vp_core_out, register_set* mem_out, int id, char* config_string ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, mem_out, id ) + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, vp_core_out, mem_out, id ) { unsigned m_prioritization_readin; int ret = sscanf( config_string, @@ -1548,6 +1579,12 @@ tensor_core:: tensor_core( register_set* result_port, const shader_core_config m_name = "TENSOR_CORE"; } +vp_core:: vp_core( register_set* result_port, const shader_core_config *config,shader_core_ctx *core ) + : pipelined_simd_unit(result_port,config,config->max_vp_core_latency,core) +{ + m_name = "VP_CORE"; +} + void sfu::issue( register_set& source_reg ) { @@ -1569,6 +1606,16 @@ void tensor_core::issue( register_set& source_reg ) pipelined_simd_unit::issue(source_reg); } +void vp_core::issue( register_set& source_reg ) +{ + warp_inst_t** ready_reg = source_reg.get_ready(); + //m_core->incexecstat((*ready_reg)); + + (*ready_reg)->op_pipe= VP_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(); @@ -1599,6 +1646,13 @@ void tensor_core::active_lanes_in_pipeline(){ m_core->incfumemactivelanes_stat(active_count); } +void vp_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) diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 90a3134..d292d56 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -319,11 +319,12 @@ public: register_set* sp_out, register_set* sfu_out, register_set* tensor_core_out, + register_set* vp_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_tensor_core_out(tensor_core_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_vp_core_out(vp_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)); @@ -397,6 +398,7 @@ protected: register_set* m_sp_out; register_set* m_sfu_out; register_set* m_tensor_core_out; + register_set* m_vp_core_out; register_set* m_mem_out; int m_id; @@ -410,9 +412,10 @@ public: register_set* sp_out, register_set* sfu_out, register_set* tensor_core_out, + register_set* vp_core_out, register_set* mem_out, int id ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, vp_core_out, mem_out, id ){} virtual ~lrr_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -428,9 +431,10 @@ public: register_set* sp_out, register_set* sfu_out, register_set* tensor_core_out, + register_set* vp_core_out, register_set* mem_out, int id ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out,vp_core_out, mem_out, id ){} virtual ~gto_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -448,10 +452,11 @@ public: register_set* sp_out, register_set* sfu_out, register_set* tensor_core_out, + register_set* vp_core_out, register_set* mem_out, int id, char* config_str ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, mem_out, id ), + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, vp_core_out, mem_out, id ), m_pending_warps() { unsigned inner_level_readin; @@ -499,6 +504,7 @@ public: register_set* sp_out, register_set* sfu_out, register_set* tensor_core_out, + register_set* vp_core_out, register_set* mem_out, int id, char* config_string ); @@ -1083,6 +1089,22 @@ public: virtual void active_lanes_in_pipeline(); virtual void issue( register_set& source_reg ); }; +class vp_core : public pipelined_simd_unit +{ +public: + vp_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 VP_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 @@ -1225,10 +1247,12 @@ enum pipeline_stage_name_t { ID_OC_SP=0, ID_OC_SFU, ID_OC_TENSOR_CORE, + ID_OC_VP_CORE, ID_OC_MEM, OC_EX_SP, OC_EX_SFU, OC_EX_TENSOR_CORE, + OC_EX_VP_CORE, OC_EX_MEM, EX_WB, N_PIPELINE_STAGES @@ -1238,10 +1262,12 @@ const char* const pipeline_stage_name_decode[] = { "ID_OC_SP", "ID_OC_SFU", "ID_OC_TENSOR_CORE", + "ID_OC_VP_CORE", "ID_OC_MEM", "OC_EX_SP", "OC_EX_SFU", "OC_EX_TENSOR_CORE", + "OC_EX_VP_CORE", "OC_EX_MEM", "EX_WB", "N_PIPELINE_STAGES" @@ -1286,13 +1312,21 @@ 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 = 64; max_sp_latency = 32; + + max_tensor_core_latency = 64; gpgpu_num_tensor_core_units=8; gpgpu_operand_collector_num_units_tensor_core=24; gpgpu_operand_collector_num_in_ports_tensor_core=8; gpgpu_operand_collector_num_out_ports_tensor_core=8; - m_L1I_config.init(m_L1I_config.m_config_string,FuncCachePreferNone); + + max_vp_core_latency = 64; + gpgpu_num_vp_core_units=8; + gpgpu_operand_collector_num_units_vp_core=24; + gpgpu_operand_collector_num_in_ports_vp_core=8; + gpgpu_operand_collector_num_out_ports_vp_core=8; + + m_L1I_config.init(m_L1I_config.m_config_string,FuncCachePreferNone); m_L1T_config.init(m_L1T_config.m_config_string,FuncCachePreferNone); m_L1C_config.init(m_L1C_config.m_config_string,FuncCachePreferNone); m_L1D_config.init(m_L1D_config.m_config_string,FuncCachePreferNone); @@ -1339,24 +1373,28 @@ struct shader_core_config : public core_config 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_vp_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_vp_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_vp_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_vp_core_units; int gpgpu_num_mem_units; //Shader core resources @@ -1370,6 +1408,7 @@ struct shader_core_config : public core_config unsigned max_sp_latency; unsigned max_sfu_latency; unsigned max_tensor_core_latency; + unsigned max_vp_core_latency; unsigned n_simt_cores_per_cluster; unsigned n_simt_clusters; @@ -1408,6 +1447,7 @@ struct shader_core_stats_pod { unsigned *m_num_sp_acesses; unsigned *m_num_sfu_acesses; unsigned *m_num_tensor_core_acesses; + unsigned *m_num_vp_core_acesses; unsigned *m_num_trans_acesses; unsigned *m_num_mem_acesses; unsigned *m_num_sp_committed; @@ -1415,6 +1455,7 @@ struct shader_core_stats_pod { unsigned *m_num_tlb_accesses; unsigned *m_num_sfu_committed; unsigned *m_num_tensor_core_committed; + unsigned *m_num_vp_core_committed; unsigned *m_num_mem_committed; unsigned *m_read_regfile_acesses; unsigned *m_write_regfile_acesses; @@ -1424,6 +1465,7 @@ struct shader_core_stats_pod { unsigned *m_active_sp_lanes; unsigned *m_active_sfu_lanes; unsigned *m_active_tensor_core_lanes; + unsigned *m_active_vp_core_lanes; unsigned *m_active_fu_lanes; unsigned *m_active_fu_mem_lanes; unsigned *m_n_diverge; // number of divergence occurring in this shader @@ -1496,6 +1538,7 @@ public: 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_vp_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)); @@ -1504,10 +1547,12 @@ public: 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_vp_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_vp_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)); -- cgit v1.3 From 7a77d951e6a900d61436df12826bb677aeaee6e6 Mon Sep 17 00:00:00 2001 From: aamir Date: Sun, 9 Sep 2018 15:10:06 -0700 Subject: minor changes for generating mem transaction in timing model. NOTE NOT COMPLETED --- cuda-kernels/gpgpusim.config | 2 +- cutlass-example/gpgpusim.config | 2 +- src/abstract_hardware_model.cc | 2 +- src/abstract_hardware_model.h | 8 +++-- src/cuda-sim/cuda-sim.cc | 65 ++++++++++++++++++++++++----------------- src/cuda-sim/instructions.cc | 64 +++++++++++++++++++++++++++++----------- src/gpgpu-sim/shader.cc | 2 +- src/gpgpu-sim/shader.h | 8 +++++ 8 files changed, 104 insertions(+), 49 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/cuda-kernels/gpgpusim.config b/cuda-kernels/gpgpusim.config index 2510d21..3daa539 100755 --- a/cuda-kernels/gpgpusim.config +++ b/cuda-kernels/gpgpusim.config @@ -33,7 +33,7 @@ # 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 4,1,1,1,4,1,1,1,6 +-gpgpu_pipeline_widths 4,1,1,1,1,4,1,1,1,1,6 -gpgpu_num_sp_units 4 -gpgpu_num_sfu_units 1 # Instruction latencies and initiation intervals diff --git a/cutlass-example/gpgpusim.config b/cutlass-example/gpgpusim.config index 2510d21..3daa539 100644 --- a/cutlass-example/gpgpusim.config +++ b/cutlass-example/gpgpusim.config @@ -33,7 +33,7 @@ # 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 4,1,1,1,4,1,1,1,6 +-gpgpu_pipeline_widths 4,1,1,1,1,4,1,1,1,1,6 -gpgpu_num_sp_units 4 -gpgpu_num_sfu_units 1 # Instruction latencies and initiation intervals diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index d668de7..b24a77e 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -184,7 +184,7 @@ void warp_inst_t::generate_mem_accesses() { if( empty() || op == MEMORY_BARRIER_OP || m_mem_accesses_created ) return; - if ( !((op == LOAD_OP) || (op == STORE_OP)) ) + if (!((op == LOAD_OP) || (op==TENSOR_CORE_LOAD_OP) || (op==VP_LOAD_OP) || (op == STORE_OP)||(op==TENSOR_CORE_STORE_OP)||(op==VP_STORE_OP))) return; if( m_warp_active_mask.count() == 0 ) return; // predicated off diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 781509f..9c418fa 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -81,6 +81,10 @@ enum uarch_op_t { VP_CORE_OP, ALU_SFU_OP, LOAD_OP, + VP_LOAD_OP, + TENSOR_CORE_LOAD_OP, + TENSOR_CORE_STORE_OP, + VP_STORE_OP, STORE_OP, BRANCH_OP, BARRIER_OP, @@ -801,8 +805,8 @@ public: { fprintf(fp," [inst @ pc=0x%04x] ", pc ); } - bool is_load() const { return (op == LOAD_OP || memory_op == memory_load); } - bool is_store() const { return (op == STORE_OP || memory_op == memory_store); } + bool is_load() const { return (op == LOAD_OP ||op==TENSOR_CORE_LOAD_OP||op==VP_LOAD_OP || memory_op == memory_load); } + bool is_store() const { return (op == STORE_OP ||op==TENSOR_CORE_STORE_OP||op==VP_STORE_OP || memory_op == memory_store); } unsigned get_num_operands() const {return num_operands;} unsigned get_num_regs() const {return num_regs;} void set_num_regs(unsigned num) {num_regs=num;} diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index bcf64c4..8f684e2 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -648,12 +648,12 @@ void ptx_instruction::set_opcode_and_latency() if ( has_memory_write() ) op = STORE_OP; break; case LD_OP: op = LOAD_OP; break; - case MMA_LD_OP: op = LOAD_OP; break; - case VP_LD_OP: op = LOAD_OP; break; + case MMA_LD_OP: op = TENSOR_CORE_LOAD_OP; break; + case VP_LD_OP: op = VP_LOAD_OP; break; case LDU_OP: op = LOAD_OP; break; case ST_OP: op = STORE_OP; break; - case MMA_ST_OP: op = STORE_OP; break; - case VP_ST_OP: op = STORE_OP; break; + case MMA_ST_OP: op = TENSOR_CORE_STORE_OP; break; + case VP_ST_OP: op = VP_STORE_OP; break; case BRA_OP: op = BRANCH_OP; break; case BREAKADDR_OP: op = BRANCH_OP; break; case TEX_OP: op = LOAD_OP; mem_op=TEX; break; @@ -1334,6 +1334,7 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) skip = !pred_lookup(pI->get_pred_mod(), pred_value.pred & 0x000F); } } + int inst_opcode=pI->get_opcode(); if( skip ) { inst.set_not_active(lane_id); @@ -1346,17 +1347,21 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) pI = pJ; } - int inst_opcode=pI->get_opcode(); + if(((inst_opcode==MMA_OP||inst_opcode==MMA_LD_OP||inst_opcode==MMA_ST_OP||inst_opcode==VP_MMA_OP||inst_opcode==VP_LD_OP||inst_opcode==VP_ST_OP))){ + if(inst.active_count()!=MAX_WARP_SIZE) + while(1); + } if(((inst_opcode!=MMA_OP)&&(inst_opcode!=MMA_LD_OP)&&(inst_opcode!=MMA_ST_OP)&&(inst_opcode!=VP_LD_OP)&&(inst_opcode!=VP_ST_OP)&&(inst_opcode!=VP_MMA_OP))||((inst_opcode==MMA_OP||inst_opcode==MMA_LD_OP||inst_opcode==MMA_ST_OP||inst_opcode==VP_MMA_OP||inst_opcode==VP_LD_OP||inst_opcode==VP_ST_OP)&&(lane_id==0))){ - switch ( inst_opcode ) { -#define OP_DEF(OP,FUNC,STR,DST,CLASSIFICATION) case OP: FUNC(pI,this); op_classification = CLASSIFICATION; break; -#define OP_W_DEF(OP,FUNC,STR,DST,CLASSIFICATION) case OP: FUNC(pI,get_core(),inst); op_classification = CLASSIFICATION; break; -#include "opcodes.def" -#undef OP_DEF -#undef OP_W_DEF - default: printf( "Execution error: Invalid opcode (0x%x)\n", pI->get_opcode() ); break; - }} + switch ( inst_opcode ) { + #define OP_DEF(OP,FUNC,STR,DST,CLASSIFICATION) case OP: FUNC(pI,this); op_classification = CLASSIFICATION; break; + #define OP_W_DEF(OP,FUNC,STR,DST,CLASSIFICATION) case OP: FUNC(pI,get_core(),inst); op_classification = CLASSIFICATION; break; + #include "opcodes.def" + #undef OP_DEF + #undef OP_W_DEF + default: printf( "Execution error: Invalid opcode (0x%x)\n", pI->get_opcode() ); break; + } + } delete pJ; pI = pI_saved; @@ -1398,13 +1403,17 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) _memory_op_t insn_memory_op = no_memory_op; unsigned insn_data_size = 0; if ( (pI->has_memory_read() || pI->has_memory_write()) ) { - insn_memaddr = last_eaddr(); - insn_space = last_space(); - unsigned to_type = pI->get_type(); - insn_data_size = datatype2size(to_type); - insn_memory_op = pI->has_memory_read() ? memory_load : memory_store; - } - + //if(!((inst_opcode==MMA_LD_OP||inst_opcode==MMA_ST_OP||inst_opcode==VP_LD_OP||inst_opcode==VP_ST_OP))) + //if(!((inst_opcode==MMA_LD_OP||inst_opcode==VP_LD_OP))) + //{ + insn_memaddr = last_eaddr(); + insn_space = last_space(); + unsigned to_type = pI->get_type(); + insn_data_size = datatype2size(to_type); + insn_memory_op = pI->has_memory_read() ? memory_load : memory_store; + //} + } + if ( pI->get_opcode() == BAR_OP && pI->barrier_op() == RED_OPTION) { inst.add_callback( lane_id, last_callback().function, last_callback().instruction, this,false /*not atomic*/); } @@ -1476,12 +1485,16 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) // "Return values" if(!skip) { - inst.space = insn_space; - inst.set_addr(lane_id, insn_memaddr); - inst.data_size = insn_data_size; // simpleAtomicIntrinsics - assert( inst.memory_op == insn_memory_op ); - } - + //if(!((inst_opcode==MMA_LD_OP||inst_opcode==MMA_ST_OP||inst_opcode==VP_LD_OP||inst_opcode==VP_ST_OP))) + //if(!((inst_opcode==MMA_LD_OP||inst_opcode==VP_LD_OP))) + //{ + inst.space = insn_space; + inst.set_addr(lane_id, insn_memaddr); + inst.data_size = insn_data_size; // simpleAtomicIntrinsics + assert( inst.memory_op == insn_memory_op ); + //} + } + } catch ( int x ) { printf("GPGPU-Sim PTX: ERROR (%d) executing intruction (%s:%u)\n", x, pI->source_file(), pI->source_line() ); printf("GPGPU-Sim PTX: '%s'\n", pI->get_source() ); diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 39b8ba5..77e8e71 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -3196,6 +3196,7 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) unsigned wmma_type = pI->get_wmma_type(); unsigned wmma_layout = pI->get_wmma_layout(0); int stride; + _memory_op_t insn_memory_op = pI->has_memory_read() ? memory_load : memory_store; for (thrd=0; thrd < core->get_warp_size(); thrd++) { thread = core->get_thread_info()[tid+thrd]; odd=thrd%2; @@ -3224,6 +3225,7 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) if(g_debug_instruction) printf("mma_st: thrd=%d,addr=%x, fp(size=%d), stride=%d\n",thrd,addr_reg.u32,size,src2_data.u32); addr_t new_addr = addr+thread_group_offset(thrd,wmma_type,wmma_layout,type,stride)*size/8; + addr_t push_addr; ptx_reg_t nw_v[8]; for(k=0;k<8;k++){ @@ -3235,7 +3237,10 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) for(k=0;k<8;k++){ if(type==F32_TYPE){ - mem->write(new_addr+4*acc_float_offset(k,wmma_layout,stride),size/8,&v[k].s64,thread,pI); + //mem->write(new_addr+4*acc_float_offset(k,wmma_layout,stride),size/8,&v[k].s64,thread,pI); + push_addr=new_addr+4*acc_float_offset(k,wmma_layout,stride); + mem->write(push_addr,size/8,&v[k].s64,thread,pI); + if(g_debug_instruction){ printf("wmma:store:thread%d=%x,%x,%x,%x,%x,%x,%x,%x\n",thrd,v[0].s64,v[1].s64,v[2].s64,v[3].s64,v[4].s64,v[5].s64,v[6].s64,v[7].s64); float temp; @@ -3250,10 +3255,16 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) } else if(type==F16_TYPE){ - if(wmma_layout==ROW) - mem->write(new_addr+k*2,size/8,&nw_v[k].s64,thread,pI); - else if(wmma_layout==COL) - mem->write(new_addr+k*2*stride,size/8,&nw_v[k].s64,thread,pI); + if(wmma_layout==ROW){ + //mem->write(new_addr+k*2,size/8,&nw_v[k].s64,thread,pI); + push_addr=new_addr+k*2; + mem->write(push_addr,size/8,&nw_v[k].s64,thread,pI); + } + else if(wmma_layout==COL){ + //mem->write(new_addr+k*2*stride,size/8,&nw_v[k].s64,thread,pI); + push_addr=new_addr+k*2*stride; + mem->write(push_addr,size/8,&nw_v[k].s64,thread,pI); + } if(g_debug_instruction) printf("wmma:store:thread%d=%x,%x,%x,%x,%x,%x,%x,%x\n",thrd,nw_v[0].s64,nw_v[1].s64,nw_v[2].s64,nw_v[3].s64,nw_v[4].s64,nw_v[5].s64,nw_v[6].s64,nw_v[7].s64); } @@ -3427,6 +3438,7 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) int tid = inst.warp_id_func()*core->get_warp_size(); int thrd,stride; ptx_thread_info *thread; + _memory_op_t insn_memory_op = pI->has_memory_read() ? memory_load : memory_store; for (thrd=0; thrd < core->get_warp_size(); thrd++){ thread = core->get_thread_info()[tid+thrd]; @@ -3452,13 +3464,18 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) printf("mma_ld: thrd=%d,addr=%x, fpsize=%d, stride=%d\n",thrd,src1_data.u32,size,src2_data.u32); addr_t new_addr = addr+thread_group_offset(thrd,wmma_type,wmma_layout,type,stride)*size/8; - + addr_t fetch_addr; if(wmma_type==LOAD_A){ for(i=0;i<16;i++){ - if(wmma_layout==ROW) - mem->read(new_addr+2*i,size/8,&data[i].s64); + if(wmma_layout==ROW){ + //mem->read(new_addr+2*i,size/8,&data[i].s64); + fetch_addr=new_addr+2*i; + mem->read(fetch_addr,size/8,&data[i].s64); + } else if(wmma_layout==COL){ - mem->read(new_addr+2*(i%4)+2*stride*4*(i/4),size/8,&data[i].s64); + //mem->read(new_addr+2*(i%4)+2*stride*4*(i/4),size/8,&data[i].s64); + fetch_addr=new_addr+2*(i%4)+2*stride*4*(i/4); + mem->read(fetch_addr,size/8,&data[i].s64); } else{ printf("mma_ld:wrong_layout_type\n"); @@ -3468,10 +3485,15 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) } else if(wmma_type==LOAD_B){ for(i=0;i<16;i++){ - if(wmma_layout==COL) - mem->read(new_addr+2*i,size/8,&data[i].s64); + if(wmma_layout==COL){ + //mem->read(new_addr+2*i,size/8,&data[i].s64); + fetch_addr=new_addr+2*i; + mem->read(fetch_addr,size/8,&data[i].s64); + } else if(wmma_layout==ROW){ - mem->read(new_addr+2*(i%4)+2*stride*4*(i/4),size/8,&data[i].s64); + //mem->read(new_addr+2*(i%4)+2*stride*4*(i/4),size/8,&data[i].s64); + fetch_addr=new_addr+2*(i%4)+2*stride*4*(i/4); + mem->read(fetch_addr,size/8,&data[i].s64); } else{ printf("mma_ld:wrong_layout_type\n"); @@ -3482,17 +3504,25 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) else if(wmma_type==LOAD_C){ for(i=0;i<8;i++){ if(type==F16_TYPE){ - if(wmma_layout==ROW) - mem->read(new_addr+2*i,size/8,&data[i].s64); - else if(wmma_layout==COL) - mem->read(new_addr+2*stride*i,size/8,&data[i].s64); + if(wmma_layout==ROW){ + //mem->read(new_addr+2*i,size/8,&data[i].s64); + fetch_addr=new_addr+2*i; + mem->read(fetch_addr,size/8,&data[i].s64); + } + else if(wmma_layout==COL){ + //mem->read(new_addr+2*stride*i,size/8,&data[i].s64); + fetch_addr=new_addr+2*stride*i; + mem->read(fetch_addr,size/8,&data[i].s64); + } else{ printf("mma_ld:wrong_type\n"); abort(); } } else if(type==F32_TYPE){ - mem->read(new_addr+4*acc_float_offset(i,wmma_layout,stride),size/8,&data[i].s64); + //mem->read(new_addr+4*acc_float_offset(i,wmma_layout,stride),size/8,&data[i].s64); + fetch_addr=new_addr+4*acc_float_offset(i,wmma_layout,stride); + mem->read(fetch_addr,size/8,&data[i].s64); } else{ printf("wrong type"); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 6f11ad9..5e80fb1 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -920,7 +920,7 @@ void scheduler_unit::cycle() ready_inst = true; const active_mask_t &active_mask = m_simt_stack[warp_id]->get_active_mask(); assert( warp(warp_id).inst_in_pipeline() ); - if ( (pI->op == LOAD_OP) || (pI->op == STORE_OP) || (pI->op == MEMORY_BARRIER_OP) ) { + if ( (pI->op == LOAD_OP)||(pI->op ==TENSOR_CORE_LOAD_OP)||(pI->op ==VP_LOAD_OP)|| (pI->op == STORE_OP)|| (pI->op==TENSOR_CORE_STORE_OP) ||(pI->op==VP_STORE_OP) || (pI->op == MEMORY_BARRIER_OP) ) { if( m_mem_out->has_free() ) { m_shader->issue_warp(*m_mem_out,pI,active_mask,warp_id); issued++; diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index d292d56..d9558b0 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1116,7 +1116,11 @@ public: switch(inst.op) { case SFU_OP: return false; case LOAD_OP: return false; + case TENSOR_CORE_LOAD_OP: return false; + case VP_LOAD_OP: return false; case STORE_OP: return false; + case TENSOR_CORE_STORE_OP: return false; + case VP_STORE_OP: return false; case MEMORY_BARRIER_OP: return false; default: break; } @@ -1158,7 +1162,11 @@ public: { switch(inst.op) { case LOAD_OP: break; + case TENSOR_CORE_LOAD_OP: break; + case VP_LOAD_OP: break; case STORE_OP: break; + case TENSOR_CORE_STORE_OP: break; + case VP_STORE_OP: break; case MEMORY_BARRIER_OP: break; default: return false; } -- cgit v1.3 From 3949357047a621a06b2e7fb4fd6099cce1469d27 Mon Sep 17 00:00:00 2001 From: aamir Date: Sun, 16 Sep 2018 20:07:33 -0700 Subject: print for mem transaction --- src/abstract_hardware_model.cc | 1 + src/abstract_hardware_model.h | 13 ++++++++++++- src/cuda-sim/instructions.cc | 15 ++++++++++++++- src/gpgpu-sim/shader.cc | 6 +++++- src/gpgpu-sim/traffic_breakdown.cc | 1 + 5 files changed, 33 insertions(+), 3 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 83e76fe..72ece0b 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -352,6 +352,7 @@ void warp_inst_t::generate_mem_accesses() ptx_file_line_stats_add_uncoalesced_gmem( pc, m_accessq.size() - starting_queue_size ); } m_mem_accesses_created=true; + print_m_accessq(); } void warp_inst_t::memory_coalescing_arch_13( bool is_write, mem_access_type access_type ) diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 9c418fa..d628745 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -933,7 +933,18 @@ public: for(unsigned i=0; i::iterator it; + for (it = m_accessq.begin(); it != m_accessq.end(); ++it){ + printf("MEM_TXN_GEN:%s:%x, Size:%d \n",mem_access_type_str(it->get_type()), it->get_addr(),it->get_size()); + } + } + } struct transaction_info { std::bitset<4> chunks; // bitmask: 32-byte chunks accessed mem_access_byte_mask_t bytes; diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 42c63ca..52d89f2 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -3354,31 +3354,41 @@ void vp_ld_impl(const ptx_instruction *pI, core_t *core, warp_inst_t &inst) printf("vp_ld: thrx=%d,addr=%x, base_addr=%x, size=%d, stride=%d\n",thrd,new_addr,addr,size,src2_data.u32); if(wmma_type==LOAD_A||wmma_type==LOAD_C){ + printf("lda/c:"); for(i=0;i<8;i++){ if(wmma_layout==ROW){ //mem->read(new_addr+4*i,size/8,&data[i].s64); mem->read(new_addr+4*i,size/8,&data[i].s64); + printf("%x ", new_addr+4*i); mem_txn_addr[num_mem_txn++]=new_addr+4*i; } else if(wmma_layout==COL){ //mem->read(new_addr+4*stride*i,size/8,&data[i].s64); + printf("%x ", new_addr+4*stride*i); mem->read(new_addr+4*stride*i,size/8,&data[i].s64); mem_txn_addr[num_mem_txn++]=new_addr+4*stride*i; } } + } + else if(wmma_type==LOAD_B4){ + printf("ldb4:"); if(wmma_layout==ROW){ mem->read(new_addr,size/8,&data[0].s64); + printf("%x ",new_addr); mem_txn_addr[num_mem_txn++]=new_addr; } else if(wmma_layout==COL){ } } else if(wmma_type==LOAD_B8){ + printf("ldb8:"); if(wmma_layout==ROW){ mem->read(new_addr,size/8,&data[0].s64); mem->read(new_addr+4,size/8,&data[1].s64); + printf("%x ",new_addr,new_addr+4); + mem_txn_addr[num_mem_txn++]=new_addr; mem_txn_addr[num_mem_txn++]=new_addr+4; } @@ -3387,12 +3397,13 @@ void vp_ld_impl(const ptx_instruction *pI, core_t *core, warp_inst_t &inst) } } else if(wmma_type==LOAD_B16){ - printf("LOADB16_MODE"); + printf("ldb16:"); if(wmma_layout==ROW){ mem->read(new_addr,size/8,&data[0].s64); mem->read(new_addr+4,size/8,&data[1].s64); mem->read(new_addr+8,size/8,&data[2].s64); mem->read(new_addr+12,size/8,&data[3].s64); + printf("%x ",new_addr,new_addr+4,new_addr+8,new_addr+12); mem_txn_addr[num_mem_txn++]=new_addr; mem_txn_addr[num_mem_txn++]=new_addr+4; mem_txn_addr[num_mem_txn++]=new_addr+8; @@ -3405,6 +3416,8 @@ void vp_ld_impl(const ptx_instruction *pI, core_t *core, warp_inst_t &inst) printf("wrong vp_load type\n");; abort(); } + printf("\n"); + //generate timing memory request inst.space = space; inst.set_addr(thrd, (new_addr_type *)mem_txn_addr , num_mem_txn); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 5e80fb1..e745f03 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -741,7 +741,10 @@ void shader_core_ctx::func_exec_inst( warp_inst_t &inst ) { execute_warp_inst_t(inst); if( inst.is_load() || inst.is_store() ) - inst.generate_mem_accesses(); + { + inst.generate_mem_accesses(); + //inst.print_m_accessq(); + } } void shader_core_ctx::issue_warp( register_set& pipe_reg_set, const warp_inst_t* next_inst, const active_mask_t &active_mask, unsigned warp_id ) @@ -1512,6 +1515,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea // bypass L1 cache unsigned control_size = inst.is_store() ? WRITE_PACKET_SIZE : READ_PACKET_SIZE; unsigned size = access.get_size() + control_size; + //printf("Interconnect:Addr: %x, size=%d\n",access.get_addr(),size); if( m_icnt->full(size, inst.is_store() || inst.isatomic()) ) { stall_cond = ICNT_RC_FAIL; } else { diff --git a/src/gpgpu-sim/traffic_breakdown.cc b/src/gpgpu-sim/traffic_breakdown.cc index 32f0d30..587067f 100644 --- a/src/gpgpu-sim/traffic_breakdown.cc +++ b/src/gpgpu-sim/traffic_breakdown.cc @@ -46,6 +46,7 @@ std::string traffic_breakdown::classify_memfetch(class mem_fetch * mf) break; default: assert(0 && "Unknown traffic type"); } + printf("%s:Icnt:%s:Request: %x,Size%d,DataSize:%d,CntrlSize%d:\n",m_network_name.c_str(),traffic_name.c_str(),mf->get_addr(),mf->size(),mf->get_data_size(),mf->get_ctrl_size()); return traffic_name; } -- cgit v1.3 From 09e6092ace5213a5d5a49bf80b052802c06a4268 Mon Sep 17 00:00:00 2001 From: aamir Date: Sat, 22 Sep 2018 15:28:40 -0700 Subject: debug prints --- cuda-kernels/Makefile | 1 + cuda-kernels/genericMatrixMultiply.cu | 6 +- cuda-kernels/gpgpusim.config | 2 +- cuda-kernels/scripts | 118 +++++++++++++++++++++++++++++----- cuda-kernels/v16p_kernel.cu | 4 +- src/Makefile | 2 +- src/cuda-sim/Makefile | 2 +- src/cuda-sim/cuda-sim.cc | 8 ++- src/gpgpu-sim/Makefile | 2 +- src/gpgpu-sim/gpu-sim.cc | 4 +- src/gpgpu-sim/shader.cc | 16 +++-- src/gpgpu-sim/shader.h | 2 +- 12 files changed, 131 insertions(+), 36 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/cuda-kernels/Makefile b/cuda-kernels/Makefile index 8effd11..b33519f 100755 --- a/cuda-kernels/Makefile +++ b/cuda-kernels/Makefile @@ -7,4 +7,5 @@ clean: rm _cuob* rm gpgpusim_power* rm gpgpu_inst_stats.txt + rm gpgpusim_visualizer* # nvcc -arch=sm_70 --gpu-architecture=compute_50 --gpu-code=compute_50 -lcudart -g -o tensor_core tensor_core.cu diff --git a/cuda-kernels/genericMatrixMultiply.cu b/cuda-kernels/genericMatrixMultiply.cu index 6a5d33f..8b96483 100644 --- a/cuda-kernels/genericMatrixMultiply.cu +++ b/cuda-kernels/genericMatrixMultiply.cu @@ -40,9 +40,9 @@ void cudaErrCheck_(cudaError_t stat, const char *file, int line) { using namespace nvcuda; // Must be multiples of 16 for wmma code to work -#define MATRIX_M (32) -#define MATRIX_N (32) -#define MATRIX_K (32) +#define MATRIX_M (64) +#define MATRIX_N (64) +#define MATRIX_K (64) diff --git a/cuda-kernels/gpgpusim.config b/cuda-kernels/gpgpusim.config index 3daa539..5f1be25 100755 --- a/cuda-kernels/gpgpusim.config +++ b/cuda-kernels/gpgpusim.config @@ -138,7 +138,7 @@ -gpgpu_memlatency_stat 14 -gpgpu_runtime_stat 500 -enable_ptx_file_line_stats 1 --visualizer_enabled 0 +-visualizer_enabled 1 # power model configs -power_simulation_enabled 1 diff --git a/cuda-kernels/scripts b/cuda-kernels/scripts index b9a3b50..a24fb72 100755 --- a/cuda-kernels/scripts +++ b/cuda-kernels/scripts @@ -10,11 +10,16 @@ nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp4_256 ./vp4_128>vp4_log128 ./vp4_256>vp4_log256 -grep "kernel_name.*vp_example" -A 589 vp4_log16 >vp4_16_summary -grep "kernel_name.*vp_example" -A 589 vp4_log32 >vp4_32_summary -grep "kernel_name.*vp_example" -A 589 vp4_log64 >vp4_64_summary -grep "kernel_name.*vp_example" -A 589 vp4_log128 >vp4_128_summary -grep "kernel_name.*vp_example" -A 589 vp4_log256 >vp4_256_summary +#grep "kernel_name.*convertInt32To" -A 589 vp4_log16 >vp4_16_summary +#grep "kernel_name.*convertInt32To" -A 589 vp4_log32 >vp4_32_summary +#grep "kernel_name.*convertInt32To" -A 589 vp4_log64 >vp4_64_summary +#grep "kernel_name.*convertInt32To" -A 589 vp4_log128 >vp4_128_summary +#grep "kernel_name.*convertInt32To" -A 589 vp4_log256 >vp4_256_summary +###grep "kernel_name.*vp_example" -A 589 vp4_log16 >>vp4_16_summary +###grep "kernel_name.*vp_example" -A 589 vp4_log32 >>vp4_32_summary +###grep "kernel_name.*vp_example" -A 589 vp4_log64 >>vp4_64_summary +#grep "kernel_name.*vp_example" -A 589 vp4_log128 >>vp4_128_summary +#grep "kernel_name.*vp_example" -A 589 vp4_log256 >>vp4_256_summary nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp8_16 v8p_genericMatrixMultiply.cu -DSIZE=16 @@ -28,11 +33,16 @@ nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp8_256 ./vp8_128>vp8_log128 ./vp8_256>vp8_log256 -grep "kernel_name.*vp_example" -A 589 vp8_log16 >vp8_16_summary -grep "kernel_name.*vp_example" -A 589 vp8_log32 >vp8_32_summary -grep "kernel_name.*vp_example" -A 589 vp8_log64 >vp8_64_summary -grep "kernel_name.*vp_example" -A 589 vp8_log128 >vp8_128_summary -grep "kernel_name.*vp_example" -A 589 vp8_log256 >vp8_256_summary +#grep "kernel_name.*convertInt32To" -A 589 vp8_log16 >vp8_16_summary +#grep "kernel_name.*convertInt32To" -A 589 vp8_log32 >vp8_32_summary +#grep "kernel_name.*convertInt32To" -A 589 vp8_log64 >vp8_64_summary +#grep "kernel_name.*convertInt32To" -A 589 vp8_log128 >vp8_128_summary +#grep "kernel_name.*convertInt32To" -A 589 vp8_log256 >vp8_256_summary +##grep "kernel_name.*vp_example" -A 589 vp8_log16 >>vp8_16_summary +##grep "kernel_name.*vp_example" -A 589 vp8_log32 >>vp8_32_summary +##grep "kernel_name.*vp_example" -A 589 vp8_log64 >>vp8_64_summary +#grep "kernel_name.*vp_example" -A 589 vp8_log128 >>vp8_128_summary +#grep "kernel_name.*vp_example" -A 589 vp8_log256 >>vp8_256_summary nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp16_16 v16p_genericMatrixMultiply.cu -DSIZE=16 nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp16_32 v16p_genericMatrixMultiply.cu -DSIZE=32 @@ -45,11 +55,85 @@ nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp16_256 ./vp16_128>vp16_log128 ./vp16_256>vp16_log256 -grep "kernel_name.*vp_example" -A 589 vp16_log16 >vp16_16_summary -grep "kernel_name.*vp_example" -A 589 vp16_log32 >vp16_32_summary -grep "kernel_name.*vp_example" -A 589 vp16_log64 >vp16_64_summary -grep "kernel_name.*vp_example" -A 589 vp16_log128 >vp16_128_summary -grep "kernel_name.*vp_example" -A 589 vp16_log256 >vp16_256_summary +#grep "kernel_name.*convertInt32To" -A 589 vp16_log16 >vp16_16_summary +#grep "kernel_name.*convertInt32To" -A 589 vp16_log32 >vp16_32_summary +#grep "kernel_name.*convertInt32To" -A 589 vp16_log64 >vp16_64_summary +#grep "kernel_name.*convertInt32To" -A 589 vp16_log128 >vp16_128_summary +#grep "kernel_name.*convertInt32To" -A 589 vp16_log256 >vp16_256_summary +##grep "kernel_name.*vp_example" -A 589 vp16_log16 >>vp16_16_summary +##grep "kernel_name.*vp_example" -A 589 vp16_log32 >>vp16_32_summary +##grep "kernel_name.*vp_example" -A 589 vp16_log64 >>vp16_64_summary +#grep "kernel_name.*vp_example" -A 589 vp16_log128 >>vp16_128_summary +#grep "kernel_name.*vp_example" -A 589 vp16_log256 >>vp16_256_summary -mv vp* VPlog/ -make clean +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp4_log16 > kernel_log_vp4_log16 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp4_log32 > kernel_log_vp4_log32 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp4_log64 > kernel_log_vp4_log64 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp4_log128 > kernel_log_vp4_log128 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp4_log256 > kernel_log_vp4_log256 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp8_log16 > kernel_log_vp8_log16 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp8_log32 > kernel_log_vp8_log32 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp8_log64 > kernel_log_vp8_log64 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp8_log128 > kernel_log_vp8_log128 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp8_log256 > kernel_log_vp8_log256 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp16_log16 > kernel_log_vp16_log16 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp16_log32 > kernel_log_vp16_log32 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp16_log64 > kernel_log_vp16_log64 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp16_log128 > kernel_log_vp16_log128 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp16_log256 > kernel_log_vp16_log256 +### +###grep memtocore:Icnt kernel_log_vp4_log16>memtocore_vp4_log16 +###grep coretomem:Icnt kernel_log_vp4_log16>coretomem_vp4_log16 +###grep MEM_TXN kernel_log_vp4_log16> gen_mem_vp4_log16 +###grep memtocore:Icnt kernel_log_vp4_log32>memtocore_vp4_log32 +###grep coretomem:Icnt kernel_log_vp4_log32>coretomem_vp4_log32 +###grep MEM_TXN kernel_log_vp4_log32> gen_mem_vp4_log32 +###grep memtocore:Icnt kernel_log_vp4_log64>memtocore_vp4_log64 +###grep coretomem:Icnt kernel_log_vp4_log64>coretomem_vp4_log64 +###grep MEM_TXN kernel_log_vp4_log64> gen_mem_vp4_log64 +####grep memtocore:Icnt kernel_log_vp4_log128>memtocore_vp4_log128 +####grep coretomem:Icnt kernel_log_vp4_log128>coretomem_vp4_log128 +####grep MEM_TXN kernel_log_vp4_log128> gen_mem_vp4_log128 +####grep memtocore:Icnt kernel_log_vp4_log256>memtocore_vp4_log256 +####grep coretomem:Icnt kernel_log_vp4_log256>coretomem_vp4_log256 +####grep MEM_TXN kernel_log_vp4_log256> gen_mem_vp4_log256 +### +###grep memtocore:Icnt kernel_log_vp8_log16>memtocore_vp8_log16 +###grep coretomem:Icnt kernel_log_vp8_log16>coretomem_vp8_log16 +###grep MEM_TXN kernel_log_vp8_log16> gen_mem_vp8_log16 +###grep memtocore:Icnt kernel_log_vp8_log32>memtocore_vp8_log32 +###grep coretomem:Icnt kernel_log_vp8_log32>coretomem_vp8_log32 +###grep MEM_TXN kernel_log_vp8_log32> gen_mem_vp8_log32 +###grep memtocore:Icnt kernel_log_vp8_log64>memtocore_vp8_log64 +###grep coretomem:Icnt kernel_log_vp8_log64>coretomem_vp8_log64 +###grep MEM_TXN kernel_log_vp8_log64> gen_mem_vp8_log64 +####grep memtocore:Icnt kernel_log_vp8_log128>memtocore_vp8_log128 +####grep coretomem:Icnt kernel_log_vp8_log128>coretomem_vp8_log128 +####grep MEM_TXN kernel_log_vp8_log128> gen_mem_vp8_log128 +####grep memtocore:Icnt kernel_log_vp8_log256>memtocore_vp8_log256 +####grep coretomem:Icnt kernel_log_vp8_log256>coretomem_vp8_log256 +####grep MEM_TXN kernel_log_vp8_log256> gen_mem_vp8_log256 +### +###grep memtocore:Icnt kernel_log_vp16_log16>memtocore_vp16_log16 +###grep coretomem:Icnt kernel_log_vp16_log16>coretomem_vp16_log16 +###grep MEM_TXN kernel_log_vp16_log16> gen_mem_vp16_log16 +###grep memtocore:Icnt kernel_log_vp16_log32>memtocore_vp16_log32 +###grep coretomem:Icnt kernel_log_vp16_log32>coretomem_vp16_log32 +###grep MEM_TXN kernel_log_vp16_log32> gen_mem_vp16_log32 +###grep memtocore:Icnt kernel_log_vp16_log64>memtocore_vp16_log64 +###grep coretomem:Icnt kernel_log_vp16_log64>coretomem_vp16_log64 +###grep MEM_TXN kernel_log_vp16_log64> gen_mem_vp16_log64 +####grep memtocore:Icnt kernel_log_vp16_log128>memtocore_vp16_log128 +####grep coretomem:Icnt kernel_log_vp16_log128>coretomem_vp16_log128 +####grep MEM_TXN kernel_log_vp16_log128> gen_mem_vp16_log128 +####grep memtocore:Icnt kernel_log_vp16_log256>memtocore_vp16_log256 +####grep coretomem:Icnt kernel_log_vp16_log256>coretomem_vp16_log256 +####grep MEM_TXN kernel_log_vp16_log256> gen_mem_vp16_log256 +### +###mv kernel_* MemTraffic/ +###mv memtocore* MemTraffic/ +###mv coretomem* MemTraffic/ +###mv gen_mem* MemTraffic/ +### +###mv vp* VPlog/ +###make clean diff --git a/cuda-kernels/v16p_kernel.cu b/cuda-kernels/v16p_kernel.cu index 864597d..d84328f 100644 --- a/cuda-kernels/v16p_kernel.cu +++ b/cuda-kernels/v16p_kernel.cu @@ -31,7 +31,7 @@ const int WMMA_M = 16; const int WMMA_N = 16; const int WMMA_K = 16; -__global__ void v4p_example(int *a_int32, int *b_int4, int *c,int *d_int32, int M, int N, int K) { +__global__ void v16p_example(int *a_int32, int *b_int4, int *c,int *d_int32, int M, int N, int K) { int registers_a[8]; int registers_b[8]; @@ -294,7 +294,7 @@ int main(int argc, char* argv[]) { //AAMIR //AAMIR printf("Running with wmma...\n"); cudaErrCheck(cudaEventRecord(startWMMA)); - v4p_example <<< 1, 32>>> (a_int32, b_int16, c_int32, d_int32, MATRIX_M, MATRIX_N, MATRIX_K); + v16p_example <<< 1, 32>>> (a_int32, b_int16, c_int32, d_int32, MATRIX_M, MATRIX_N, MATRIX_K); cudaErrCheck(cudaEventRecord(stopWMMA)); cudaErrCheck(cudaEventSynchronize(stopWMMA)); diff --git a/src/Makefile b/src/Makefile index 6001669..09194f3 100644 --- a/src/Makefile +++ b/src/Makefile @@ -46,7 +46,7 @@ ifeq ($(TRACE),1) endif ifneq ($(DEBUG),1) - OPTFLAGS += -O3 + OPTFLAGS += -O0 else CXXFLAGS += endif diff --git a/src/cuda-sim/Makefile b/src/cuda-sim/Makefile index e8977a6..d08a696 100644 --- a/src/cuda-sim/Makefile +++ b/src/cuda-sim/Makefile @@ -42,7 +42,7 @@ include ../../version_detection.mk OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/cuda-sim -OPT := -O3 -g3 -Wall -Wno-unused-function -Wno-sign-compare +OPT := -O0 -g3 -Wall -Wno-unused-function -Wno-sign-compare ifeq ($(DEBUG),1) OPT := -g3 -Wall -Wno-unused-function -Wno-sign-compare endif diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 9735a0e..dd7edd9 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -809,7 +809,7 @@ void ptx_instruction::set_opcode_and_latency() } if(get_wmma_type()==VP_MMA8) { - latency = 128; + latency =128; initiation_interval = 128; } if(get_wmma_type()==VP_MMA16) @@ -1295,12 +1295,14 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) addr_t pc = next_instr(); assert( pc == inst.pc ); // make sure timing model and functional model are in sync const ptx_instruction *pI = m_func_info->get_instruction(pc); - #if 0 + //#if 0 if(lane_id==0){ + //printf("EXECUTION_FLOW:LINE_NUM:%d\n",pI->source_line()); printf("EXECUTION_FLOW:LINE_NUM:%d\n",pI->source_line()); printf("EXECUTION_FLOW:INST:%s\n",pI->get_source()); + printf("EXECUTION_FLOW:PC%d\n",pc); } - #endif + //#endif set_npc( pc + pI->inst_size() ); diff --git a/src/gpgpu-sim/Makefile b/src/gpgpu-sim/Makefile index f10a8a4..4f77699 100644 --- a/src/gpgpu-sim/Makefile +++ b/src/gpgpu-sim/Makefile @@ -48,7 +48,7 @@ ifeq ($(GNUC_CPP0X), 1) endif ifneq ($(DEBUG),1) - OPTFLAGS += -O3 + OPTFLAGS += -O0 else CXXFLAGS += endif diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 7a797b5..eba6f54 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1469,7 +1469,9 @@ void gpgpu_sim::cycle() if( g_single_step && ((gpu_sim_cycle+gpu_tot_sim_cycle) >= g_single_step) ) { asm("int $03"); } - gpu_sim_cycle++; + printf("gpu_sim_cycle=%d\n",gpu_sim_cycle); + gpu_sim_cycle++; + if( g_interactive_debugger_enabled ) gpgpu_debug(); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index e745f03..eeb59bb 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1321,10 +1321,16 @@ void ldst_unit::get_L1T_sub_stats(struct cache_sub_stats &css) const{ void shader_core_ctx::warp_inst_complete(const warp_inst_t &inst) { - #if 0 - printf("[warp_inst_complete] uid=%u core=%u warp=%u pc=%#x @ time=%llu issued@%llu\n", - inst.get_uid(), m_sid, inst.warp_id(), inst.pc, gpu_tot_sim_cycle + gpu_sim_cycle, inst.get_issue_cycle()); - #endif + //#if 0 + //printf("[warp_inst_complete] uid=%u core=%u warp=%u pc=%#x @ time=%llu issued@%llu\n", + //const ptx_instruction *pI = m_func_info->get_instruction(inst.pc); + printf("[warp_inst_complete] uid=%u core=%u warp=%u pc=%d @ time=%llu \n", + inst.get_uid(), m_sid, inst.warp_id(), inst.pc, gpu_sim_cycle); + //printf("^instruction:%s",(pI->m_source).c_str()); + //inst.get_uid(), m_sid, inst.warp_id(), inst.pc, gpu_tot_sim_cycle + gpu_sim_cycle, inst.get_issue_cycle()); + //#endif + + if(inst.op_pipe==SP__OP) m_stats->m_num_sp_committed[m_sid]++; else if(inst.op_pipe==SFU__OP) @@ -1510,7 +1516,6 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea if (m_core->get_config()->gmem_skip_L1D) bypassL1D = true; } - if( bypassL1D ) { // bypass L1 cache unsigned control_size = inst.is_store() ? WRITE_PACKET_SIZE : READ_PACKET_SIZE; @@ -3472,6 +3477,7 @@ void simt_core_cluster::icnt_inject_request_packet(class mem_fetch *mf) case CONST_ACC_R: m_stats->gpgpu_n_mem_const++; break; case TEXTURE_ACC_R: m_stats->gpgpu_n_mem_texture++; break; case GLOBAL_ACC_R: m_stats->gpgpu_n_mem_read_global++; break; + //case GLOBAL_ACC_R: m_stats->gpgpu_n_mem_read_global++; printf("read_global%d\n",m_stats->gpgpu_n_mem_read_global); break; case GLOBAL_ACC_W: m_stats->gpgpu_n_mem_write_global++; break; case LOCAL_ACC_R: m_stats->gpgpu_n_mem_read_local++; break; case LOCAL_ACC_W: m_stats->gpgpu_n_mem_write_local++; break; diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index d9558b0..99abb63 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1328,7 +1328,7 @@ struct shader_core_config : public core_config gpgpu_operand_collector_num_in_ports_tensor_core=8; gpgpu_operand_collector_num_out_ports_tensor_core=8; - max_vp_core_latency = 64; + max_vp_core_latency = 512; gpgpu_num_vp_core_units=8; gpgpu_operand_collector_num_units_vp_core=24; gpgpu_operand_collector_num_in_ports_vp_core=8; -- cgit v1.3 From aeb78a3dcd5613404162514e98d18ee9dbf31569 Mon Sep 17 00:00:00 2001 From: Md Aamir Raihan Date: Fri, 26 Oct 2018 11:39:26 -0700 Subject: made the changes compatible with old config files --- src/cuda-sim/cuda-sim.cc | 21 ++++++++------------- src/gpgpu-sim/gpu-sim.cc | 5 ++++- src/gpgpu-sim/shader.h | 23 +++++++++++++++-------- 3 files changed, 27 insertions(+), 22 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 7587db7..86a1d45 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -69,9 +69,9 @@ unsigned cdp_latency[5]; void ptx_opcocde_latency_options (option_parser_t opp) { option_parser_register(opp, "-ptx_opcode_latency_int", OPT_CSTR, &opcode_latency_int, - "Opcode latencies for integers " - "Default 1,1,19,25,145,1,4", - "1,1,19,25,145,1,4"); + "Opcode latencies for integers " + "Default 1,1,19,25,145", + "1,1,19,25,145"); option_parser_register(opp, "-ptx_opcode_latency_fp", OPT_CSTR, &opcode_latency_fp, "Opcode latencies for single precision floating points " "Default 1,1,1,1,30", @@ -81,9 +81,9 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Default 8,8,8,8,335", "8,8,8,8,335"); option_parser_register(opp, "-ptx_opcode_initiation_int", OPT_CSTR, &opcode_initiation_int, - "Opcode initiation intervals for integers " - "Default 1,1,4,4,32,1,1", - "1,1,4,4,32,1"); + "Opcode initiation intervals for integers " + "Default 1,1,4,4,32", + "1,1,4,4,32"); option_parser_register(opp, "-ptx_opcode_initiation_fp", OPT_CSTR, &opcode_initiation_fp, "Opcode initiation intervals for single precision floating points " "Default 1,1,1,1,5", @@ -590,13 +590,9 @@ void ptx_instruction::set_bar_type() void ptx_instruction::set_opcode_and_latency() { unsigned int_latency[5]; - unsigned int_precision; - unsigned int_lane_width; unsigned fp_latency[5]; unsigned dp_latency[5]; unsigned int_init[5]; - unsigned int_init_precision; - unsigned int_init_lane_width; unsigned fp_init[5]; unsigned dp_init[5]; /* @@ -605,11 +601,10 @@ void ptx_instruction::set_opcode_and_latency() * [2] MUL * [3] MAD * [4] DIV - * [5] BSMAD */ sscanf(opcode_latency_int, "%u,%u,%u,%u,%u,%u,%u", &int_latency[0],&int_latency[1],&int_latency[2], - &int_latency[3],&int_latency[4],&int_precision,&int_lane_width); + &int_latency[3],&int_latency[4]); sscanf(opcode_latency_fp, "%u,%u,%u,%u,%u", &fp_latency[0],&fp_latency[1],&fp_latency[2], &fp_latency[3],&fp_latency[4]); @@ -618,7 +613,7 @@ void ptx_instruction::set_opcode_and_latency() &dp_latency[3],&dp_latency[4]); sscanf(opcode_initiation_int, "%u,%u,%u,%u,%u,%u, %u", &int_init[0],&int_init[1],&int_init[2], - &int_init[3],&int_init[4],&int_init_precision,&int_init_lane_width); + &int_init[3],&int_init[4]); sscanf(opcode_initiation_fp, "%u,%u,%u,%u,%u", &fp_init[0],&fp_init[1],&fp_init[2], &fp_init[3],&fp_init[4]); diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index d4b636b..63ba759 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -369,8 +369,11 @@ 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_TENSOR_CORE,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_TENSOR_CORE,OC_EX_TENSOR_CORE,OC_EX_MEM,EX_WB", + "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", "1,1,1,1,1,1,1,1,1,1" ); + option_parser_register(opp, "-gpgpu_tensor_core_avail", OPT_INT32, &gpgpu_tensor_core_avail, + "Tensor Core Available (default=0)", + "0"); option_parser_register(opp, "-gpgpu_num_sp_units", OPT_INT32, &gpgpu_num_sp_units, "Number of SP units (default=1)", "1"); diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 8ad6514..97e438f 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1228,26 +1228,26 @@ 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, + ID_OC_TENSOR_CORE, + OC_EX_TENSOR_CORE, N_PIPELINE_STAGES }; 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", + "ID_OC_TENSOR_CORE", + "OC_EX_TENSOR_CORE", "N_PIPELINE_STAGES" }; @@ -1272,16 +1272,22 @@ struct shader_core_config : public core_config strcpy(toks,pipeline_widths_string); toks = strtok(toks,","); - // pipe_widths[OC_EX_TENSOR_CORE]=1; - // pipe_widths[ID_OC_TENSOR_CORE]=1; - for (unsigned i = 0; i < N_PIPELINE_STAGES; i++) { + + /* Removing the tensorcore pipeline while reading the config files if the tensor core is not available. + If we won't remove it, old regression will be broken. + So to support the legacy config files it's best to handle in this way. + */ + int num_config_to_read=N_PIPELINE_STAGES-2*(!gpgpu_tensor_core_avail); + + for (unsigned i = 0; i MAX_THREAD_PER_SM) { printf("GPGPU-Sim uArch: Error ** increase MAX_THREAD_PER_SM in abstract_hardware_model.h from %u to %u\n", MAX_THREAD_PER_SM, n_thread_per_shader); @@ -1361,6 +1367,7 @@ struct shader_core_config : public core_config unsigned int gpgpu_operand_collector_num_out_ports_gen; int gpgpu_num_sp_units; + int gpgpu_tensor_core_avail; int gpgpu_num_sfu_units; int gpgpu_num_tensor_core_units; int gpgpu_num_mem_units; -- cgit v1.3 From 8db50a319911b803a386a7d7eca10afbea4d7e75 Mon Sep 17 00:00:00 2001 From: aamir Date: Fri, 2 Nov 2018 21:37:31 -0700 Subject: added incount and outcount, addressed all of the pull request comments --- libcuda/cuda_runtime_api.cc | 8 +++++--- src/abstract_hardware_model.h | 6 ++++++ src/cuda-sim/cuda-sim.cc | 22 ++++++++++++++++++---- src/cuda-sim/instructions.cc | 14 ++++++++++---- src/cuda-sim/ptx_parser.cc | 2 +- src/gpgpu-sim/scoreboard.cc | 44 ++++++++----------------------------------- src/gpgpu-sim/shader.cc | 18 ++++++++++++------ 7 files changed, 60 insertions(+), 54 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index a79e740..d67fd85 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -986,7 +986,6 @@ __host__ cudaError_t CUDARTAPI cudaLaunchKernel ( const char* hostFun, dim3 grid struct CUstream_st *s = (struct CUstream_st *)stream; g_cuda_launch_stack.push_back( kernel_config(gridDim,blockDim,sharedMem,s) ); - //printf("cudaLaunchKernel:sizeof(Arg[0])=%d)\n ",sizeof(args[0])); kernel_config &config = g_cuda_launch_stack.back(); config.set_arg(args[0],432,0);//standard interface for cutlass library #TODO Implementing a generalized kernel @@ -1006,12 +1005,15 @@ __host__ cudaError_t CUDARTAPI cudaLaunchKernel ( const char* hostFun, dim3 grid dim3 blockDim1 = config1.block_dim(); printf("GPGPU-Sim PTX: pushing kernel \'%s\' to stream %u, gridDim= (%u,%u,%u) blockDim = (%u,%u,%u) \n", kname.c_str(), stream1?stream1->get_uid():0, gridDim1.x,gridDim1.y,gridDim1.z,blockDim1.x,blockDim1.y,blockDim1.z ); + + /*Kernel is hardcoded to enable the cutlass library*/ + std::string cutlass("cutlass"); + assert(kname.find(cutlass) != std::string::npos); + stream_operation op(grid,g_ptx_sim_mode,stream1); g_stream_manager->push(op); g_cuda_launch_stack.pop_back(); return g_last_cudaError = cudaSuccess; - - } /******************************************************************************* diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index f561f34..71d3d89 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -36,6 +36,10 @@ class kernel_info_t; #define MAX_CTA_PER_SHADER 32 #define MAX_BARRIERS_PER_CTA 16 +//After expanding the vector input and output operands +#define MAX_INPUT_VALUES 24 +#define MAX_OUTPUT_VALUES 8 + enum _memory_space_t { undefined_space=0, reg_space, @@ -830,7 +834,9 @@ public: address_type reconvergence_pc; // -1 => not a branch, -2 => use function return address unsigned out[8]; + unsigned outcount; unsigned in[24]; + unsigned incount; unsigned char is_vectorin; unsigned char is_vectorout; int pred; // predicate register number diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 86a1d45..23c5ad5 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -602,7 +602,7 @@ void ptx_instruction::set_opcode_and_latency() * [3] MAD * [4] DIV */ - sscanf(opcode_latency_int, "%u,%u,%u,%u,%u,%u,%u", + sscanf(opcode_latency_int, "%u,%u,%u,%u,%u", &int_latency[0],&int_latency[1],&int_latency[2], &int_latency[3],&int_latency[4]); sscanf(opcode_latency_fp, "%u,%u,%u,%u,%u", @@ -850,12 +850,14 @@ void ptx_instruction::pre_decode() { pc = m_PC; isize = m_inst_size; - for(unsigned i=0; i<8; i++) { + for(unsigned i=0; i0) + outcount++; + + for(int i=0;i0) + incount++; // Get predicate if(has_pred()) { @@ -1316,7 +1327,10 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) if(((inst_opcode==MMA_OP||inst_opcode==MMA_LD_OP||inst_opcode==MMA_ST_OP))){ if(inst.active_count()!=MAX_WARP_SIZE) - while(1); + { + printf("Tensor Core operation are warp synchronous operation. All the threads needs to be active."); + assert(0); + } } if(((inst_opcode!=MMA_OP)&&(inst_opcode!=MMA_LD_OP)&&(inst_opcode!=MMA_ST_OP))||((inst_opcode==MMA_OP||inst_opcode==MMA_LD_OP||inst_opcode==MMA_ST_OP)&&(lane_id==0))){ diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 2677c40..06cad54 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -1726,7 +1726,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) unsigned b_layout = pI->get_wmma_layout(1); unsigned type = pI->get_type(); unsigned type2 = pI->get_type2(); - int tid = inst.warp_id_func() * core->get_warp_size(); + int tid = inst.warp_id() * core->get_warp_size(); const operand_info &dst = pI->operand_lookup(0); unsigned thread_group_index; @@ -2962,7 +2962,7 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) const operand_info &src = pI->operand_lookup(1); const operand_info &src1 = pI->operand_lookup(0); const operand_info &src2 = pI->operand_lookup(2); - int tid = inst.warp_id_func()*core->get_warp_size(); + int tid = inst.warp_id()*core->get_warp_size(); unsigned type = pI->get_type(); unsigned wmma_type = pI->get_wmma_type(); unsigned wmma_layout = pI->get_wmma_layout(0); @@ -3069,7 +3069,7 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) unsigned type = pI->get_type(); unsigned wmma_type = pI->get_wmma_type(); unsigned wmma_layout = pI->get_wmma_layout(0); - int tid = inst.warp_id_func()*core->get_warp_size(); + int tid = inst.warp_id()*core->get_warp_size(); int thrd,stride; ptx_thread_info *thread; _memory_op_t insn_memory_op = pI->has_memory_read() ? memory_load : memory_store; @@ -4469,7 +4469,13 @@ void set_impl( const ptx_instruction *pI, ptx_thread_info *thread ) void shfl_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) { unsigned i_type = pI->get_type(); - int tid = inst.warp_id_func() * core->get_warp_size(); + int tid; + + if(core->get_gpu()->is_functional_sim()) + tid = inst.warp_id_func() * core->get_warp_size(); + else + tid = inst.warp_id() * core->get_warp_size(); + ptx_thread_info *thread = core->get_thread_info()[tid]; ptx_warp_info *warp_info = thread->m_warp_info; int lane = warp_info->get_done_threads(); diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index d5324d0..9671ab7 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -39,7 +39,7 @@ void set_ptx_warp_size(const struct core_config * warp_size) g_shader_core_config=warp_size; } -static bool g_debug_ir_generation=true; +static bool g_debug_ir_generation=false; const char *g_filename; unsigned g_max_regs_per_thread = 0; diff --git a/src/gpgpu-sim/scoreboard.cc b/src/gpgpu-sim/scoreboard.cc index 4d1b43a..ebec891 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 < 8; r++) { + for( unsigned r=0; r < MAX_OUTPUT_VALUES; 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<8; r++) { + for ( unsigned r=0; rout[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 < 8; r++) { + for( unsigned r=0; r < MAX_OUTPUT_VALUES; r++) { if(inst->out[r] > 0) { SHADER_DPRINTF( SCOREBOARD, "Register Released - warp:%d, reg: %d\n", @@ -138,39 +138,11 @@ bool Scoreboard::checkCollision( unsigned wid, const class inst_t *inst ) const // Get list of all input and output registers std::set inst_regs; - if(inst->out[0] > 0) inst_regs.insert(inst->out[0]); - 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->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[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]); + for(int iii=0;iiioutcount;iii++) + inst_regs.insert(inst->out[iii]); + + for(int jjj=0;jjjincount;jjj++) + inst_regs.insert(inst->in[jjj]); if(inst->pred > 0) inst_regs.insert(inst->pred); if(inst->ar1 > 0) inst_regs.insert(inst->ar1); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 23f255d..f9cfa58 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -425,7 +425,7 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout, "gpgpu_n_load_insn = %d\n", gpgpu_n_load_insn); fprintf(fout, "gpgpu_n_store_insn = %d\n", gpgpu_n_store_insn); fprintf(fout, "gpgpu_n_shmem_insn = %d\n", gpgpu_n_shmem_insn); - fprintf(fout, "gpgpu_n_shmem_insn = %d\n", gpgpu_n_sstarr_insn); + fprintf(fout, "gpgpu_n_sstarr_insn = %d\n", gpgpu_n_sstarr_insn); fprintf(fout, "gpgpu_n_tex_insn = %d\n", gpgpu_n_tex_insn); fprintf(fout, "gpgpu_n_const_mem_insn = %d\n", gpgpu_n_const_insn); fprintf(fout, "gpgpu_n_param_mem_insn = %d\n", gpgpu_n_param_insn); @@ -1290,6 +1290,12 @@ void ldst_unit::get_L1T_sub_stats(struct cache_sub_stats &css) const{ void shader_core_ctx::warp_inst_complete(const warp_inst_t &inst) { + + #if 0 + printf("[warp_inst_complete] uid=%u core=%u warp=%u pc=%#x @ time=%llu issued@%llu\n", + inst.get_uid(), m_sid, inst.warp_id(), inst.pc, gpu_tot_sim_cycle + gpu_sim_cycle, inst.get_issue_cycle()); + #endif + if(inst.op_pipe==SP__OP) m_stats->m_num_sp_committed[m_sid]++; else if(inst.op_pipe==SFU__OP) @@ -1386,7 +1392,7 @@ ldst_unit::process_cache_access( cache_t* cache, assert( !read_sent ); inst.accessq_pop_back(); if ( inst.is_load() ) { - for ( unsigned r=0; r < 8; r++) + for ( unsigned r=0; r < MAX_OUTPUT_VALUES; r++) if (inst.out[r] > 0) m_pending_writes[inst.warp_id()][inst.out[r]]--; } @@ -1488,7 +1494,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea inst.accessq_pop_back(); //inst.clear_active( access.get_warp_mask() ); if( inst.is_load() ) { - for( unsigned r=0; r < 8; r++) + for( unsigned r=0; r < MAX_OUTPUT_VALUES; r++) if(inst.out[r] > 0) assert( m_pending_writes[inst.warp_id()][inst.out[r]] > 0 ); } else if( inst.is_store() ) @@ -1765,7 +1771,7 @@ void ldst_unit:: issue( register_set ®_set ) if (inst->is_load() and inst->space.get_type() != shared_space) { unsigned warp_id = inst->warp_id(); unsigned n_accesses = inst->accessq_count(); - for (unsigned r = 0; r < 8; r++) { + for (unsigned r = 0; r < MAX_OUTPUT_VALUES; r++) { unsigned reg_id = inst->out[r]; if (reg_id > 0) { m_pending_writes[warp_id][reg_id] += n_accesses; @@ -1787,7 +1793,7 @@ void ldst_unit::writeback() if( !m_next_wb.empty() ) { if( m_operand_collector->writeback(m_next_wb) ) { bool insn_completed = false; - for( unsigned r=0; r < 8; r++ ) { + for( unsigned r=0; r < MAX_OUTPUT_VALUES; r++ ) { if( m_next_wb.out[r] > 0 ) { if( m_next_wb.space.get_type() != shared_space ) { assert( m_pending_writes[m_next_wb.warp_id()][m_next_wb.out[r]] > 0 ); @@ -1989,7 +1995,7 @@ void ldst_unit::cycle() //} bool pending_requests=false; - for( unsigned r=0; r<8; r++ ) { + for( unsigned r=0; r 0 ) { if( m_pending_writes[warp_id].find(reg_id) != m_pending_writes[warp_id].end() ) { -- cgit v1.3 From eae030c9d1d607d1c14e4ade99cb5caea6403efd Mon Sep 17 00:00:00 2001 From: aamir Date: Sat, 3 Nov 2018 17:03:58 -0700 Subject: merged with memory subsytem. Regression is passing but tensorcore kernel is stuck in deadlock --- .gitignore | 1 + .travis.yml | 31 +- CHANGES | 30 + Jenkinsfile | 126 ++++ README | 10 +- aerialvision/organizedata.py | 2 +- configs/GTX480/config_fermi_islip.icnt | 70 -- configs/GTX480/gpgpusim.config | 134 ---- configs/GTX480/gpuwattch_gtx480.xml | 538 -------------- configs/GeForceGTX1080Ti/config_fermi_islip.icnt | 70 -- configs/GeForceGTX1080Ti/gpgpusim.config | 150 ---- configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml | 538 -------------- configs/GeForceGTX750Ti/config_fermi_islip.icnt | 70 -- configs/GeForceGTX750Ti/gpgpusim.config | 131 ---- configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml | 539 -------------- configs/Pascal_TITANX/config_fermi_islip.icnt | 70 -- configs/Pascal_TITANX/gpgpusim.config | 156 ---- configs/QuadroFX5600/gpgpusim.config | 99 --- configs/QuadroFX5600/gpuwattch_quadrofx5600.xml | 538 -------------- configs/QuadroFX5600/icnt_config_islip.icnt | 70 -- configs/QuadroFX5800/config_quadro_islip.icnt | 69 -- configs/QuadroFX5800/gpgpusim.config | 88 --- configs/TeslaC2050/config_fermi_islip.icnt | 70 -- configs/TeslaC2050/gpgpusim.config | 133 ---- .../deprecated-cfgs/GTX480/config_fermi_islip.icnt | 70 ++ configs/deprecated-cfgs/GTX480/gpgpusim.config | 134 ++++ .../deprecated-cfgs/GTX480/gpuwattch_gtx480.xml | 538 ++++++++++++++ .../GeForceGTX750Ti/config_fermi_islip.icnt | 70 ++ .../GeForceGTX750Ti/gpgpusim.config | 131 ++++ .../GeForceGTX750Ti/gpuwattch_gtx750Ti.xml | 539 ++++++++++++++ .../deprecated-cfgs/QuadroFX5600/gpgpusim.config | 99 +++ .../QuadroFX5600/gpuwattch_quadrofx5600.xml | 538 ++++++++++++++ .../QuadroFX5600/icnt_config_islip.icnt | 70 ++ .../QuadroFX5800/config_quadro_islip.icnt | 69 ++ .../deprecated-cfgs/QuadroFX5800/gpgpusim.config | 88 +++ .../SM6_GTX1080/config_fermi_islip.icnt | 70 ++ .../deprecated-cfgs/SM6_GTX1080/gpgpusim.config | 150 ++++ .../SM6_GTX1080/gpuwattch_gtx1080Ti.xml | 538 ++++++++++++++ .../SM6_P100/config_fermi_islip.icnt | 70 ++ configs/deprecated-cfgs/SM6_P100/gpgpusim.config | 156 ++++ .../deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml | 538 ++++++++++++++ .../SM6_TITANX/config_fermi_islip.icnt | 70 ++ configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config | 157 +++++ .../SM6_TITANX/gpuwattch_gtx480.xml | 538 ++++++++++++++ .../SM7_TITANV/config_fermi_islip.icnt | 70 ++ configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config | 165 +++++ .../TeslaC2050/config_fermi_islip.icnt | 70 ++ configs/deprecated-cfgs/TeslaC2050/gpgpusim.config | 133 ++++ .../tested-cfgs/SM2_GTX480/config_fermi_islip.icnt | 70 ++ configs/tested-cfgs/SM2_GTX480/gpgpusim.config | 153 ++++ .../tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml | 538 ++++++++++++++ .../tested-cfgs/SM6_TITANX/config_fermi_islip.icnt | 73 ++ configs/tested-cfgs/SM6_TITANX/gpgpusim.config | 185 +++++ .../tested-cfgs/SM7_TITANV/config_fermi_islip.icnt | 74 ++ configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 191 +++++ cuobjdump_to_ptxplus/ptx_parser.h | 4 + libcuda/cuda_runtime_api.cc | 49 +- nightly.jenkinsfile | 75 ++ setup_environment | 2 +- src/abstract_hardware_model.cc | 49 +- src/abstract_hardware_model.h | 33 +- src/cuda-sim/cuda-sim.cc | 47 +- src/cuda-sim/instructions.cc | 1 + src/cuda-sim/ptx.y | 3 +- src/cuda-sim/ptx_ir.cc | 7 +- src/cuda-sim/ptx_ir.h | 7 + src/cuda-sim/ptx_parser.cc | 4 + src/cuda-sim/ptx_parser.h | 1 + src/gpgpu-sim/addrdec.cc | 71 ++ src/gpgpu-sim/addrdec.h | 9 + src/gpgpu-sim/delayqueue.h | 1 + src/gpgpu-sim/dram.cc | 637 +++++++++++++---- src/gpgpu-sim/dram.h | 77 +- src/gpgpu-sim/dram_sched.cc | 121 +++- src/gpgpu-sim/dram_sched.h | 12 + src/gpgpu-sim/gpu-cache.cc | 782 +++++++++++++++++---- src/gpgpu-sim/gpu-cache.h | 644 +++++++++++++++-- src/gpgpu-sim/gpu-sim.cc | 125 +++- src/gpgpu-sim/gpu-sim.h | 57 +- src/gpgpu-sim/l2cache.cc | 166 ++++- src/gpgpu-sim/l2cache.h | 18 + src/gpgpu-sim/l2cache_trace.h | 16 + src/gpgpu-sim/mem_fetch.cc | 7 +- src/gpgpu-sim/mem_fetch.h | 13 +- src/gpgpu-sim/mem_latency_stat.cc | 19 +- src/gpgpu-sim/mem_latency_stat.h | 4 + src/gpgpu-sim/shader.cc | 431 ++++++++++-- src/gpgpu-sim/shader.h | 93 ++- src/gpgpusim_entrypoint.cc | 1 + src/gpuwattch/makefile | 4 +- src/intersim2/Makefile | 2 +- src/intersim2/interconnect_interface.cpp | 9 +- src/trace_streams.tup | 2 + version | 2 +- 94 files changed, 9589 insertions(+), 4104 deletions(-) create mode 100644 Jenkinsfile delete mode 100644 configs/GTX480/config_fermi_islip.icnt delete mode 100644 configs/GTX480/gpgpusim.config delete mode 100755 configs/GTX480/gpuwattch_gtx480.xml delete mode 100644 configs/GeForceGTX1080Ti/config_fermi_islip.icnt delete mode 100644 configs/GeForceGTX1080Ti/gpgpusim.config delete mode 100755 configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml delete mode 100644 configs/GeForceGTX750Ti/config_fermi_islip.icnt delete mode 100644 configs/GeForceGTX750Ti/gpgpusim.config delete mode 100755 configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml delete mode 100644 configs/Pascal_TITANX/config_fermi_islip.icnt delete mode 100644 configs/Pascal_TITANX/gpgpusim.config delete mode 100644 configs/QuadroFX5600/gpgpusim.config delete mode 100644 configs/QuadroFX5600/gpuwattch_quadrofx5600.xml delete mode 100644 configs/QuadroFX5600/icnt_config_islip.icnt delete mode 100644 configs/QuadroFX5800/config_quadro_islip.icnt delete mode 100644 configs/QuadroFX5800/gpgpusim.config delete mode 100644 configs/TeslaC2050/config_fermi_islip.icnt delete mode 100644 configs/TeslaC2050/gpgpusim.config create mode 100644 configs/deprecated-cfgs/GTX480/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/GTX480/gpgpusim.config create mode 100755 configs/deprecated-cfgs/GTX480/gpuwattch_gtx480.xml create mode 100644 configs/deprecated-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/GeForceGTX750Ti/gpgpusim.config create mode 100755 configs/deprecated-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml create mode 100644 configs/deprecated-cfgs/QuadroFX5600/gpgpusim.config create mode 100644 configs/deprecated-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml create mode 100644 configs/deprecated-cfgs/QuadroFX5600/icnt_config_islip.icnt create mode 100644 configs/deprecated-cfgs/QuadroFX5800/config_quadro_islip.icnt create mode 100644 configs/deprecated-cfgs/QuadroFX5800/gpgpusim.config create mode 100644 configs/deprecated-cfgs/SM6_GTX1080/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/SM6_GTX1080/gpgpusim.config create mode 100755 configs/deprecated-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml create mode 100644 configs/deprecated-cfgs/SM6_P100/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/SM6_P100/gpgpusim.config create mode 100755 configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml create mode 100644 configs/deprecated-cfgs/SM6_TITANX/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config create mode 100755 configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml create mode 100644 configs/deprecated-cfgs/SM7_TITANV/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config create mode 100644 configs/deprecated-cfgs/TeslaC2050/config_fermi_islip.icnt create mode 100644 configs/deprecated-cfgs/TeslaC2050/gpgpusim.config create mode 100644 configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt create mode 100644 configs/tested-cfgs/SM2_GTX480/gpgpusim.config create mode 100755 configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml create mode 100644 configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt create mode 100644 configs/tested-cfgs/SM6_TITANX/gpgpusim.config create mode 100644 configs/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt create mode 100644 configs/tested-cfgs/SM7_TITANV/gpgpusim.config create mode 100644 nightly.jenkinsfile (limited to 'src/gpgpu-sim') diff --git a/.gitignore b/.gitignore index d21cc8f..6c6ed5f 100644 --- a/.gitignore +++ b/.gitignore @@ -32,4 +32,5 @@ cuobjdump_to_ptxplus/sass_parser.hh cuobjdump_to_ptxplus/sass_parser.output build/* +tags *.swp diff --git a/.travis.yml b/.travis.yml index 7a12a1c..b38f468 100644 --- a/.travis.yml +++ b/.travis.yml @@ -1,7 +1,7 @@ sudo: required before_install: - - docker pull tgrogers/gpgpu-sim_regress:latest + - docker pull tgrogers/gpgpu-sim_regress:volta_update language: cpp @@ -10,34 +10,13 @@ matrix: include: - services: docker env: - - CONFIG=configs.gtx480.yml + - CONFIG=GTX480 - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ -# This config is just taking far too long... -# - services: docker -# env: -# - CONFIG=configs.gtx750ti.yml -# - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ -# - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - services: docker env: - - CONFIG=configs.quadro5600.yml - - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - services: docker - env: - - CONFIG=configs.quadro5800.yml - - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - services: docker - env: - - CONFIG=configs.teslac2050.yml - - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ - - services: docker - env: - - CONFIG=configs.gtx1080ti.yml - - CUDA_INSTALL_PATH=/usr/local/cuda-4.2/ + - CONFIG=TITANV + - CUDA_INSTALL_PATH=/usr/local/cuda-9.1/ - PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-9.1/ -script: docker run -v `pwd`:/home/runner/gpgpu-sim_distribution:rw tgrogers/gpgpu-sim_regress:latest /bin/bash -c "./start_torque.sh; chown -R runner /home/runner/gpgpu-sim_distribution; su - runner -c 'export CUDA_INSTALL_PATH=$CUDA_INSTALL_PATH && export PTXAS_CUDA_INSTALL_PATH=$PTXAS_CUDA_INSTALL_PATH && source /home/runner/gpgpu-sim_distribution/setup_environment && make -j -C /home/runner/gpgpu-sim_distribution && cd /home/runner/gpgpu-sim_simulations/ && git pull && /home/runner/gpgpu-sim_simulations/util/job_launching/run_simulations.py -c /home/runner/gpgpu-sim_simulations/util/job_launching/regression_recipies/rodinia_2.0-ft/$CONFIG -N regress && /home/runner/gpgpu-sim_simulations/util/job_launching/monitor_func_test.py -v -N regress'" +script: docker run -v `pwd`:/home/runner/gpgpu-sim_distribution:rw tgrogers/gpgpu-sim_regress:volta_update /bin/bash -c "./start_torque.sh; chown -R runner /home/runner/gpgpu-sim_distribution; su - runner -c 'export CUDA_INSTALL_PATH=$CUDA_INSTALL_PATH && export PTXAS_CUDA_INSTALL_PATH=$PTXAS_CUDA_INSTALL_PATH && source /home/runner/gpgpu-sim_distribution/setup_environment && make -j -C /home/runner/gpgpu-sim_distribution && cd /home/runner/gpgpu-sim_simulations/ && git pull && /home/runner/gpgpu-sim_simulations/util/job_launching/run_simulations.py -C $CONFIG -B rodinia_2.0-ft -N regress && /home/runner/gpgpu-sim_simulations/util/job_launching/monitor_func_test.py -v -N regress'" diff --git a/CHANGES b/CHANGES index dbf5b39..60be885 100644 --- a/CHANGES +++ b/CHANGES @@ -1,4 +1,34 @@ LOG: +Version 4.0.0 (development branch) versus 3.2.3 +-Front-End: +1- Support .nc cache modifier and __ldg function to access the read-only L1D cache +2- Partially-support some SASS_60 in the PTXP_PLUS (not completed yet) +-GPU Core: +1- Fermi/Pascal coalescer: coalescing on 32-bytes sectors. +2- Adding separate dp unit pipeline. +3- diff dual issue: allow scheduler to issue diff insts at a time +4- Fair memory issue from multiple schedulers. +-Cache System: +1- Sector L1/L2 cache +2- Fetch-on-write and lazy-fetch-on-read write allocation policy. +3- Improving the L1 cache throughput (streaming L1 cache) +4- Performance model for CUDA memory copy. +5- Support memory partition indexing to reduce partition camping (POLY, XOR and PAE (ISCA’18) Indexing) +6- Adaptive cache configuration +-Memory: +1- Performance Model for HBM (mainly the dual-bus interface) +2- Separate Read/Write buffers. +3- Advanced bank indexing function. +-Statistics: +1- Adding more detailed cache statistics to define and analyze cache bottlenecks. +2- Adding more detailed memory statistics (BLP, RBL, etc) to define and analyze memory bottlenecks. +3- Addig new system stats: gpu occupancy, L2BW, etc +-Configs: +Adding the Pascal and Volta config files that has been correlated against real hardware. +See the correlation website here: +https://engineering.purdue.edu/tgrogers/group/correlator.html + + Version 3.2.3+edits (development branch) versus 3.2.3 - Support for running regression tests using Travis - Support added for CUDA dynamic parallelism (courtesy of Jin Wang from Georgia Tech) diff --git a/Jenkinsfile b/Jenkinsfile new file mode 100644 index 0000000..1969aea --- /dev/null +++ b/Jenkinsfile @@ -0,0 +1,126 @@ +pipeline { + agent { + label "purdue-cluster" + } + + options { + disableConcurrentBuilds() + } + + stages { + stage('simulator-build') { + steps { + parallel "4.2": { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/4.2_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + make -j' + }, "9.1" : { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + make -j' + } + } + } + stage('simulations-build'){ + steps{ + sh 'rm -rf gpgpu-sim_simulations' + sh 'git clone git@github.rcac.purdue.edu:TimRogersGroup/gpgpu-sim_simulations.git && \ + cd gpgpu-sim_simulations && \ + git checkout purdue-cluster && \ + git pull && \ + ln -s /home/tgrogers-raid/a/common/data_dirs benchmarks/' + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/4.2_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + cd gpgpu-sim_simulations && \ + source ./benchmarks/src/setup_environment && \ + make -j -C ./benchmarks/src rodinia_2.0-ft sdk-4.2 && \ + make -C ./benchmarks/src data' + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + cd gpgpu-sim_simulations && \ + source ./benchmarks/src/setup_environment && \ + make -j -C ./benchmarks/src/ rodinia_2.0-ft sdk-4.2 && \ + make -C ./benchmarks/src data' + } + } + stage('regress'){ + steps { + parallel "4.2-rodinia": { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/4.2_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + ./gpgpu-sim_simulations/util/job_launching/run_simulations.py -B rodinia_2.0-ft -C GTX480,GTX480-PTXPLUS -N regress-$$ && \ + PLOTDIR="jenkins/${JOB_NAME}/${BUILD_NUMBER}/4.2-rodinia" && ssh tgrogers@dynamo.ecn.purdue.edu mkdir -p /home/dynamo/a/tgrogers/website/gpgpu-sim-plots/$PLOTDIR && \ + ./gpgpu-sim_simulations/util/job_launching/monitor_func_test.py -v -N regress-$$ -s stats-per-app-4.2.csv && \ + ./gpgpu-sim_simulations/util/plotting/plot-get-stats.py -c stats-per-app-4.2.csv -p tgrogers@dynamo.ecn.purdue.edu:~/website/gpgpu-sim-plots/$PLOTDIR -w https://engineering.purdue.edu/tgrogers/gpgpu-sim-plots/$PLOTDIR -n $PLOTDIR ' + }, "9.1-functest": { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + ./gpgpu-sim_simulations/util/job_launching/run_simulations.py -B rodinia_2.0-ft,sdk-4.2 -C TITANX,TITANX-L1ON -N regress-$$ && \ + PLOTDIR="jenkins/${JOB_NAME}/${BUILD_NUMBER}/9.1-rodinia" && ssh tgrogers@dynamo.ecn.purdue.edu mkdir -p /home/dynamo/a/tgrogers/website/gpgpu-sim-plots/$PLOTDIR && \ + ./gpgpu-sim_simulations/util/job_launching/monitor_func_test.py -v -s stats-per-app-9.1.csv -N regress-$$ && \ + ./gpgpu-sim_simulations/util/plotting/plot-get-stats.py -c stats-per-app-9.1.csv -p tgrogers@dynamo.ecn.purdue.edu:~/website/gpgpu-sim-plots/$PLOTDIR -w https://engineering.purdue.edu/tgrogers/gpgpu-sim-plots/$PLOTDIR -n $PLOTDIR' + } + } + } + stage('4.2-correlate'){ + steps { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/4.2_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + PLOTDIR="jenkins/${JOB_NAME}" &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B rodinia_2.0-ft -C GTX480,GTX480-PTXPLUS > stats-per-kernel-4.2.csv &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh stats-per-kernel-4.2.csv $PLOTDIR ${BUILD_NUMBER}' + } + } + stage('9.1-correlate'){ + steps { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + PLOTDIR="jenkins/${JOB_NAME}" &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B rodinia_2.0-ft,sdk-4.2 -C TITANX,TITANX-L1ON > stats-per-kernel-9.1.csv &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh stats-per-kernel-9.1.csv $PLOTDIR ${BUILD_NUMBER}' + } + } + stage('archive-and-delta') { + steps { + sh 'rm -rf gpgpu-sim-results-repo' + sh 'git clone git@github.com:purdue-aalp/gpgpu-sim-results-repo.git' + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/4.2_env_setup.sh &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B rodinia_2.0-ft -C GTX480-PTXPLUS > stats-per-kernel-4.2-ptxplus.csv &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B rodinia_2.0-ft -C GTX480 > stats-per-kernel-4.2-ptx.csv' + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B rodinia_2.0-ft,sdk-4.2 -C TITANX > stats-per-kernel-9.1-titanx.csv' + sh './gpgpu-sim_simulations/util/plotting/merge-stats.py -c ./gpgpu-sim-results-repo/jenkins/quick-regress/AALP/gpgpu-sim_distribution/dev-purdue-integration/stats-per-app-4.2.csv,./stats-per-app-4.2.csv -R > per-app-merge-4.2.csv' + sh './gpgpu-sim_simulations/util/plotting/merge-stats.py -c ./gpgpu-sim-results-repo/jenkins/quick-regress/AALP/gpgpu-sim_distribution/dev-purdue-integration/stats-per-app-9.1.csv,./stats-per-app-9.1.csv -R > per-app-merge-9.1.csv' + sh 'PLOTDIR="jenkins/${JOB_NAME}" &&\ + ./gpgpu-sim_simulations/util/plotting/plot-get-stats.py -c per-app-merge-4.2.csv -p tgrogers@dynamo.ecn.purdue.edu:~/website/gpgpu-sim-plots/$PLOTDIR/deltas -w https://engineering.purdue.edu/tgrogers/gpgpu-sim-plots/$PLOTDIR -n $PLOTDIR/deltas &&\ + ./gpgpu-sim_simulations/util/plotting/plot-get-stats.py -c per-app-merge-9.1.csv -p tgrogers@dynamo.ecn.purdue.edu:~/website/gpgpu-sim-plots/$PLOTDIR/deltas -w https://engineering.purdue.edu/tgrogers/gpgpu-sim-plots/$PLOTDIR/deltas -n $PLOTDIR/deltas &&\ + ./gpgpu-sim_simulations/util/plotting/merge-stats.py -c ./gpgpu-sim-results-repo/jenkins/quick-regress/AALP/gpgpu-sim_distribution/dev-purdue-integration/stats-per-kernel-4.2-ptx.csv,./stats-per-kernel-4.2-ptx.csv -R > per-kernel-merge-4.2-ptx.csv &&\ + ./gpgpu-sim_simulations/util/plotting/merge-stats.py -c ./gpgpu-sim-results-repo/jenkins/quick-regress/AALP/gpgpu-sim_distribution/dev-purdue-integration/stats-per-kernel-4.2-ptxplus.csv,./stats-per-kernel-4.2-ptxplus.csv -R > per-kernel-merge-4.2-ptxplus.csv &&\ + ./gpgpu-sim_simulations/util/plotting/merge-stats.py -c ./gpgpu-sim-results-repo/jenkins/quick-regress/AALP/gpgpu-sim_distribution/dev-purdue-integration/stats-per-kernel-9.1-titanx.csv,./stats-per-kernel-9.1-titanx.csv -R > per-kernel-merge-9.1-titanx.csv &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh per-kernel-merge-4.2-ptx.csv $PLOTDIR ${BUILD_NUMBER} &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh per-kernel-merge-4.2-ptxplus.csv $PLOTDIR ${BUILD_NUMBER} &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh per-kernel-merge-9.1-titanx.csv $PLOTDIR ${BUILD_NUMBER} &&\ + mkdir -p ./jenkins/quick-regress/${JOB_NAME}/ && cp stats-per-*.csv ./jenkins/quick-regress/${JOB_NAME}/ &&\ + cd ./gpgpu-sim-results-repo &&\ + git diff --quiet && git diff --staged --quiet || git commit -am "Jenkins automated checkin ${BUILD_NUMBER}" &&\ + git push' + } + } + } + post { + success { + emailext body: "See ${BUILD_URL}", + recipientProviders: [[$class: 'CulpritsRecipientProvider'], + [$class: 'RequesterRecipientProvider']], + subject: "[AALP Jenkins] Build #${BUILD_NUMBER} - Success!", + to: 'tgrogers@purdue.edu' + } + failure { + emailext body: "See ${BUILD_URL}", + recipientProviders: [[$class: 'CulpritsRecipientProvider'], + [$class: 'RequesterRecipientProvider']], + subject: "[AALP Jenkins] Build #${BUILD_NUMBER} - ${currentBuild.result}", + to: 'tgrogers@purdue.edu' + } + } +} diff --git a/README b/README index 6e2d734..7eaae5d 100644 --- a/README +++ b/README @@ -5,8 +5,8 @@ AerialVision and a configurable and extensible energy model called GPUWattch. GPGPU-Sim and GPUWattch have been rigorously validated with performance and power measurements of real hardware GPUs. -This version of GPGPU-Sim has been tested with CUDA version 2.3, 3.1, 4.0, -5.0, 5.5, 6.0 and 7.5. +This version of GPGPU-Sim has been tested with CUDA version 4.2, +5.0, 5.5, 6.0 and 7.5, 8.0, 9.0, 9.1 Please see the copyright notice in the file COPYRIGHT distributed with this release in the same directory as this file. @@ -18,6 +18,12 @@ Analyzing CUDA Workloads Using a Detailed GPU Simulator, in IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), Boston, MA, April 19-21, 2009. +If you use the memory system in GPGPU-Sim, or the Volta/Pascal models, +please cite: +Mahmoud Khairy, Jain Akshay, Tor Aamodt, Timothy G Rogers, +Exploring Modern GPU Memory System Design Challenges through Accurate Modeling, arXiv:1810.07269, +https://arxiv.org/abs/1810.07269 + If you use the GPUWattch energy model in your research, please cite: Jingwen Leng, Tayler Hetherington, Ahmed ElTantawy, Syed Gilani, Nam Sung Kim, diff --git a/aerialvision/organizedata.py b/aerialvision/organizedata.py index 090b90f..ea947cd 100644 --- a/aerialvision/organizedata.py +++ b/aerialvision/organizedata.py @@ -97,7 +97,7 @@ def organizedata(fileVars): 'sparse':OrganizeSparse, # Vector data with 2D index (used by DRAM access stats) 'custom':0 } - data_type_char = {int:'I', float:'f'} + data_type_char = {int:'L', float:'d'} print "Organizing data into internal format..." diff --git a/configs/GTX480/config_fermi_islip.icnt b/configs/GTX480/config_fermi_islip.icnt deleted file mode 100644 index 7820e4e..0000000 --- a/configs/GTX480/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 27; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/GTX480/gpgpusim.config b/configs/GTX480/gpgpusim.config deleted file mode 100644 index ee90c12..0000000 --- a/configs/GTX480/gpgpusim.config +++ /dev/null @@ -1,134 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 15 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 700.0:700.0:700.0:924.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 --gpgpu_occupancy_sm_number 20 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 116 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/GTX480/gpuwattch_gtx480.xml b/configs/GTX480/gpuwattch_gtx480.xml deleted file mode 100755 index 304e0fd..0000000 --- a/configs/GTX480/gpuwattch_gtx480.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/GeForceGTX1080Ti/config_fermi_islip.icnt b/configs/GeForceGTX1080Ti/config_fermi_islip.icnt deleted file mode 100644 index 2a69ddd..0000000 --- a/configs/GeForceGTX1080Ti/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 50; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/GeForceGTX1080Ti/gpgpusim.config b/configs/GeForceGTX1080Ti/gpgpusim.config deleted file mode 100644 index fb044c6..0000000 --- a/configs/GeForceGTX1080Ti/gpgpusim.config +++ /dev/null @@ -1,150 +0,0 @@ -# This config models the Pascal GP102 (GeForceGTX 1080Ti) - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 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 1 --gpgpu_n_mem 11 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 --gpgpu_occupancy_sm_number 60 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 32 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB -## Pascal GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,2,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 1,2,1,1,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache -# The default is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 --gpgpu_shmem_size 98304 --gmem_skip_L1D 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 --gpgpu_cache:dl2_texture_only 0 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 -# gpgpu_num_reg_banks should be increased to 32, but it gives an error! --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) -# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Pascal has four schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx1080Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml b/configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml deleted file mode 100755 index 02619ff..0000000 --- a/configs/GeForceGTX1080Ti/gpuwattch_gtx1080Ti.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/GeForceGTX750Ti/config_fermi_islip.icnt b/configs/GeForceGTX750Ti/config_fermi_islip.icnt deleted file mode 100644 index 069ca02..0000000 --- a/configs/GeForceGTX750Ti/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 7; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/GeForceGTX750Ti/gpgpusim.config b/configs/GeForceGTX750Ti/gpgpusim.config deleted file mode 100644 index c675aab..0000000 --- a/configs/GeForceGTX750Ti/gpgpusim.config +++ /dev/null @@ -1,131 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 52 - - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 5 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 2 --gpgpu_n_sub_partition_per_mchannel 1 - -# Maxwell clock domains -#-gpgpu_clock_domains ::: -# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. --gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 - -# shader core pipeline config --gpgpu_shader_registers 65536 --gpgpu_occupancy_sm_number 52 - -# This implies a maximum of 64 warps/SM --gpgpu_shader_core_pipeline 2048:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 --gpgpu_num_sp_units 8 --gpgpu_num_sfu_units 32 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 6,12,13,13,210 --ptx_opcode_initiation_int 1,1,1,1,4 --ptx_opcode_latency_fp 6,12,6,6,374 --ptx_opcode_initiation_fp 1,1,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - --gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 --gmem_skip_L1D 1 --gpgpu_shmem_size 65536 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache --gpgpu_cache:dl2 1024:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 16:128:32,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 150 --dram_latency 130 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 300 - -# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 32 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Maxwell has four schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs -# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. -# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. --power_simulation_enabled 0 --gpuwattch_xml_file gpuwattch_gtx750Ti.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml b/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml deleted file mode 100755 index e2b2324..0000000 --- a/configs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml +++ /dev/null @@ -1,539 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/Pascal_TITANX/config_fermi_islip.icnt b/configs/Pascal_TITANX/config_fermi_islip.icnt deleted file mode 100644 index 602daee..0000000 --- a/configs/Pascal_TITANX/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//21*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we do not use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 52; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/Pascal_TITANX/gpgpusim.config b/configs/Pascal_TITANX/gpgpusim.config deleted file mode 100644 index f78bd02..0000000 --- a/configs/Pascal_TITANX/gpgpusim.config +++ /dev/null @@ -1,156 +0,0 @@ -# This config models the Pascal GP102 (NVIDIA TITAN X) -# For more info about this card, see Nvidia White paper -# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf - -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 61 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# high level architecture configuration --gpgpu_n_clusters 28 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 12 --gpgpu_n_sub_partition_per_mchannel 2 - -# Pascal clock domains -#-gpgpu_clock_domains ::: -# Pascal NVIDIA TITAN X clock domains are adopted from -# https://en.wikipedia.org/wiki/GeForce_10_series --gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 - -# 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 GP102 has 4 SP SIMD units and 1 SFU unit -## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,1,1,4,1,1,6 --gpgpu_num_sp_units 4 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" -# SFU is 32-width in pascal, then dp units initiation is 1 cycle --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 --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 4,8,4,4,130 - -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. -# Pascal GP102 has 96KB Shared memory -# Pascal GP102 has 64KB L1 cache -# The defulat is to disable the L1 cache, unless cache modifieres is used --gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 --gpgpu_shmem_size 98304 --gpgpu_shmem_size_PrefL1 98304 --gpgpu_shmem_size_PrefShared 98304 --gmem_skip_L1D 1 - -# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache --gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 --gpgpu_cache:dl2_texture_only 0 --gpgpu_dram_partition_queues 32:32:32:32 - -# 4 KB Inst. --gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 -# 48 KB Tex --gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 -# 12 KB Const --gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 - -# enable operand collector -## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units --gpgpu_operand_collector_num_units_sp 20 --gpgpu_operand_collector_num_units_sfu 4 --gpgpu_operand_collector_num_units_mem 8 --gpgpu_operand_collector_num_in_ports_sp 4 --gpgpu_operand_collector_num_out_ports_sp 4 --gpgpu_operand_collector_num_in_ports_sfu 1 --gpgpu_operand_collector_num_out_ports_sfu 1 --gpgpu_operand_collector_num_in_ports_mem 1 --gpgpu_operand_collector_num_out_ports_mem 1 -# gpgpu_num_reg_banks should be increased to 32 --gpgpu_num_reg_banks 32 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - -## In Pascal, a warp scheduler can issue 2 insts per cycle --gpgpu_max_insn_issue_per_warp 2 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 924MHz / 700MHz = 132 --gpgpu_frfcfs_dram_sched_queue_size 64 --gpgpu_dram_return_queue_size 116 - -# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) -# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition -# the atom size of GDDR5X (the smallest read request) is 32 bytes --gpgpu_n_mem_per_ctrlr 1 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5X is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS - -# Use the same GDDR5 timing from hynix H5GQ1H24AFR -# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 4 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# power model configs --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_gtx480.xml - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 - diff --git a/configs/QuadroFX5600/gpgpusim.config b/configs/QuadroFX5600/gpgpusim.config deleted file mode 100644 index 6f836ee..0000000 --- a/configs/QuadroFX5600/gpgpusim.config +++ /dev/null @@ -1,99 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 12 - -# high level architecture configuration --gpgpu_n_clusters 8 --gpgpu_n_cores_per_cluster 2 --gpgpu_n_mem 6 --gpgpu_clock_domains 337.5:600.0:600.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 --gpgpu_occupancy_sm_number 12 - --gpgpu_occupancy_sm_number 12 -#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) --gpgpu_shader_core_pipeline 768:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,1,1,1,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour --gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - -# TLB parameters -#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 -#-gpgpu_tlbl2_latency 45 - - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file icnt_config_islip.icnt - -# dram scheduler config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 55 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz -# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - - - -# Using cuobjdump to extract ptx/SASS --gpgpu_ptx_use_cuobjdump 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 --power_trace_enabled 0 --power_simulation_enabled 1 --gpuwattch_xml_file gpuwattch_quadrofx5600.xml - --steady_power_levels_enabled 1 --steady_state_definition 8,4 diff --git a/configs/QuadroFX5600/gpuwattch_quadrofx5600.xml b/configs/QuadroFX5600/gpuwattch_quadrofx5600.xml deleted file mode 100644 index 2c5a6fc..0000000 --- a/configs/QuadroFX5600/gpuwattch_quadrofx5600.xml +++ /dev/null @@ -1,538 +0,0 @@ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - diff --git a/configs/QuadroFX5600/icnt_config_islip.icnt b/configs/QuadroFX5600/icnt_config_islip.icnt deleted file mode 100644 index de3bcc8..0000000 --- a/configs/QuadroFX5600/icnt_config_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//14*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 14; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/QuadroFX5800/config_quadro_islip.icnt b/configs/QuadroFX5800/config_quadro_islip.icnt deleted file mode 100644 index cfe9cac..0000000 --- a/configs/QuadroFX5800/config_quadro_islip.icnt +++ /dev/null @@ -1,69 +0,0 @@ -//18*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 18; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; - diff --git a/configs/QuadroFX5800/gpgpusim.config b/configs/QuadroFX5800/gpgpusim.config deleted file mode 100644 index fef1110..0000000 --- a/configs/QuadroFX5800/gpgpusim.config +++ /dev/null @@ -1,88 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 13 - -# high level architecture configuration --gpgpu_n_clusters 10 --gpgpu_n_cores_per_cluster 3 --gpgpu_n_mem 8 --gpgpu_clock_domains 325.0:650.0:650.0:800.0 - -# shader core pipeline config --gpgpu_shader_registers 16384 --gpgpu_occupancy_sm_number 13 - --gpgpu_shader_core_pipeline 1024:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 1,1,1,1,1,1,1 --gpgpu_num_sp_units 1 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 1,1,19,25,145 --ptx_opcode_initiation_int 1,1,4,4,32 --ptx_opcode_latency_fp 1,1,1,1,30 --ptx_opcode_initiation_fp 1,1,1,1,5 --ptx_opcode_latency_dp 8,8,8,8,335 --ptx_opcode_initiation_dp 8,8,8,8,130 - -# memory stage behaviour -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo --gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 --gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 --gpgpu_cache:dl2_texture_only 1 - --gpgpu_shmem_warp_parts 2 - -# interconnection --network_mode 1 --inter_config_file config_quadro_islip.icnt - -# dram scheduler config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (30 core cycles). I.e. -# Total buffer space required = 30 x 800MHz / 325MHz = 74 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 58 - -# dram model config --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 4 --dram_data_command_freq_ratio 2 # GDDR3 is DDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS -# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz --gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 - --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/TeslaC2050/config_fermi_islip.icnt b/configs/TeslaC2050/config_fermi_islip.icnt deleted file mode 100644 index a11bd8e..0000000 --- a/configs/TeslaC2050/config_fermi_islip.icnt +++ /dev/null @@ -1,70 +0,0 @@ -//20*1 fly with 32 flits per packet under gpgpusim injection mode -use_map = 0; -flit_size = 32; - -// currently we donot use this, see subnets below -network_count = 2; - -// Topology -topology = fly; -k = 26; -n = 1; - -// Routing - -routing_function = dest_tag; - -// Flow control - -num_vcs = 1; -vc_buf_size = 8; - -wait_for_tail_credit = 0; - -// Router architecture - -vc_allocator = islip; //separable_input_first; -sw_allocator = islip; //separable_input_first; -alloc_iters = 1; - -credit_delay = 0; -routing_delay = 0; -vc_alloc_delay = 1; -sw_alloc_delay = 1; - -input_speedup = 2; -output_speedup = 1; -internal_speedup = 1.0; - -// Traffic, GPGPU-Sim does not use this - -traffic = uniform; -packet_size ={{1,2,3,4},{10,20}}; -packet_size_rate={{1,1,1,1},{2,1}}; - -// Simulation - Don't change - -sim_type = gpgpusim; -//sim_type = latency; -injection_rate = 0.1; - - -subnets = 2; - -// Always use read and write no matter following line -//use_read_write = 1; - - -read_request_subnet = 0; -read_reply_subnet = 1; -write_request_subnet = 0; -write_reply_subnet = 1; - -read_request_begin_vc = 0; -read_request_end_vc = 0; -write_request_begin_vc = 0; -write_request_end_vc = 0; -read_reply_begin_vc = 0; -read_reply_end_vc = 0; -write_reply_begin_vc = 0; -write_reply_end_vc = 0; diff --git a/configs/TeslaC2050/gpgpusim.config b/configs/TeslaC2050/gpgpusim.config deleted file mode 100644 index 6ac2c12..0000000 --- a/configs/TeslaC2050/gpgpusim.config +++ /dev/null @@ -1,133 +0,0 @@ -# functional simulator specification --gpgpu_ptx_instruction_classification 0 --gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 - -# Using cuobjdump to extract ptx/SASS -#-gpgpu_ptx_use_cuobjdump 1 # use default - -# SASS execution (only supported with CUDA >= 4.0) --gpgpu_ptx_convert_to_ptxplus 0 --gpgpu_ptx_save_converted_ptxplus 0 - - -# high level architecture configuration --gpgpu_n_clusters 14 --gpgpu_n_cores_per_cluster 1 --gpgpu_n_mem 6 --gpgpu_n_sub_partition_per_mchannel 2 - -# Fermi clock domains -#-gpgpu_clock_domains ::: -# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided -# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 --gpgpu_clock_domains 575.0:575.0:575.0:750.0 - -# shader core pipeline config --gpgpu_shader_registers 32768 --gpgpu_occupancy_sm_number 20 - -# This implies a maximum of 48 warps/SM --gpgpu_shader_core_pipeline 1536:32 --gpgpu_shader_cta 8 --gpgpu_simd_model 1 - -# Pipeline widths and number of FUs -# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB --gpgpu_pipeline_widths 2,1,1,2,1,1,2 --gpgpu_num_sp_units 2 --gpgpu_num_sfu_units 1 - -# Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 --ptx_opcode_latency_fp 4,13,4,5,39 --ptx_opcode_initiation_fp 1,2,1,1,4 --ptx_opcode_latency_dp 8,19,8,8,330 --ptx_opcode_initiation_dp 8,16,8,8,130 - - -# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb -# ::,::::,::,:** -# ** Optional parameter - Required when mshr_type==Texture Fifo -# Note: Hashing set index function (H) only applies to a set size of 32 or 64. --gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 --gpgpu_shmem_size 49152 - -# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected -#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 -#-gpgpu_shmem_size 16384 - -# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache --gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 --gpgpu_cache:dl2_texture_only 0 - --gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 --gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 --gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 - -# enable operand collector --gpgpu_operand_collector_num_units_sp 6 --gpgpu_operand_collector_num_units_sfu 8 --gpgpu_operand_collector_num_in_ports_sp 2 --gpgpu_operand_collector_num_out_ports_sp 2 --gpgpu_num_reg_banks 16 - -# shared memory bankconflict detection --gpgpu_shmem_num_banks 32 --gpgpu_shmem_limited_broadcast 0 --gpgpu_shmem_warp_parts 1 - --gpgpu_max_insn_issue_per_warp 1 - -# interconnection --network_mode 1 --inter_config_file config_fermi_islip.icnt - -# memory partition latency config --rop_latency 120 --dram_latency 100 - -# dram model config --gpgpu_dram_scheduler 1 -# The DRAM return queue and the scheduler queue together should provide buffer -# to sustain the memory level parallelism to tolerate DRAM latency -# To allow 100% DRAM utility, there should at least be enough buffer to sustain -# the minimum DRAM latency (100 core cycles). I.e. -# Total buffer space required = 100 x 750MHz / 575MHz = 130 --gpgpu_frfcfs_dram_sched_queue_size 16 --gpgpu_dram_return_queue_size 114 - -# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition --gpgpu_n_mem_per_ctrlr 2 --gpgpu_dram_buswidth 4 --gpgpu_dram_burst_length 8 --dram_data_command_freq_ratio 4 # GDDR5 is QDR --gpgpu_mem_address_mask 1 --gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS - -# GDDR5 timing from hynix H5GQ1H24AFR -# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 --gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: - CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" - -# Fermi has two schedulers per core --gpgpu_num_sched_per_core 2 -# Two Level Scheduler with active and pending pools -#-gpgpu_scheduler two_level_active:6:0:1 -# Loose round robbin scheduler -#-gpgpu_scheduler lrr -# Greedy then oldest scheduler --gpgpu_scheduler gto - -# stat collection --gpgpu_memlatency_stat 14 --gpgpu_runtime_stat 500 --enable_ptx_file_line_stats 1 --visualizer_enabled 0 - -# tracing functionality -#-trace_enabled 1 -#-trace_components WARP_SCHEDULER,SCOREBOARD -#-trace_sampling_core 0 diff --git a/configs/deprecated-cfgs/GTX480/config_fermi_islip.icnt b/configs/deprecated-cfgs/GTX480/config_fermi_islip.icnt new file mode 100644 index 0000000..7820e4e --- /dev/null +++ b/configs/deprecated-cfgs/GTX480/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 27; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/GTX480/gpgpusim.config b/configs/deprecated-cfgs/GTX480/gpgpusim.config new file mode 100644 index 0000000..ee90c12 --- /dev/null +++ b/configs/deprecated-cfgs/GTX480/gpgpusim.config @@ -0,0 +1,134 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 15 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 700.0:700.0:700.0:924.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 116 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/deprecated-cfgs/GTX480/gpuwattch_gtx480.xml b/configs/deprecated-cfgs/GTX480/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/deprecated-cfgs/GTX480/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt b/configs/deprecated-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt new file mode 100644 index 0000000..069ca02 --- /dev/null +++ b/configs/deprecated-cfgs/GeForceGTX750Ti/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 7; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/GeForceGTX750Ti/gpgpusim.config b/configs/deprecated-cfgs/GeForceGTX750Ti/gpgpusim.config new file mode 100644 index 0000000..c675aab --- /dev/null +++ b/configs/deprecated-cfgs/GeForceGTX750Ti/gpgpusim.config @@ -0,0 +1,131 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 52 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 5 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 2 +-gpgpu_n_sub_partition_per_mchannel 1 + +# Maxwell clock domains +#-gpgpu_clock_domains ::: +# In Maxwell, each pipeline has 32 execution units, so the Core clock is just the reported number. +-gpgpu_clock_domains 1137.0:1137.0:1137.0:2700.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 52 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 8 +-gpgpu_num_sfu_units 32 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 6,12,13,13,210 +-ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_fp 6,12,6,6,374 +-ptx_opcode_initiation_fp 1,1,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gmem_skip_L1D 1 +-gpgpu_shmem_size 65536 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 1024 sets, each 128 bytes 8-way for each memory sub partition. This gives 2MB L2 cache +-gpgpu_cache:dl2 1024:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 16:128:32,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 150 +-dram_latency 130 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 300 + +# for Maxwell, bus width is 128bits, this is 64 bytes (32 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 32 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Maxwell has four schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +# power modeling is currently disabled for GTX750Ti. The gpuwattch_gtx750ti.xml file present is only a placeholder and has NOT been tested for accuracy. +# proper modeling of Maxwell power is planned, but should be considered completely unsupported at present. +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx750Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/deprecated-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml b/configs/deprecated-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml new file mode 100755 index 0000000..e2b2324 --- /dev/null +++ b/configs/deprecated-cfgs/GeForceGTX750Ti/gpuwattch_gtx750Ti.xml @@ -0,0 +1,539 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/QuadroFX5600/gpgpusim.config b/configs/deprecated-cfgs/QuadroFX5600/gpgpusim.config new file mode 100644 index 0000000..6f836ee --- /dev/null +++ b/configs/deprecated-cfgs/QuadroFX5600/gpgpusim.config @@ -0,0 +1,99 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 12 + +# high level architecture configuration +-gpgpu_n_clusters 8 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 6 +-gpgpu_clock_domains 337.5:600.0:600.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +-gpgpu_occupancy_sm_number 12 + +-gpgpu_occupancy_sm_number 12 +#8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) +-gpgpu_shader_core_pipeline 768:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,1,1,1,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +-gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +# TLB parameters +#-gpgpu_cache:tlbl1 1:524288:16:1:L:R:m,A:32:8,8 +#-gpgpu_tlbl2_latency 45 + + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file icnt_config_islip.icnt + +# dram scheduler config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 337.5MHz = 71 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 55 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +# {nbk:tCCD:tRRD:tRCD:tRAS:tRP:tRC:CL:WL:tCDLR:tWR:nbkgrp:tCCDL:tRTPL} +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + + + +# Using cuobjdump to extract ptx/SASS +-gpgpu_ptx_use_cuobjdump 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 +-power_trace_enabled 0 +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_quadrofx5600.xml + +-steady_power_levels_enabled 1 +-steady_state_definition 8,4 diff --git a/configs/deprecated-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml b/configs/deprecated-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml new file mode 100644 index 0000000..2c5a6fc --- /dev/null +++ b/configs/deprecated-cfgs/QuadroFX5600/gpuwattch_quadrofx5600.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/QuadroFX5600/icnt_config_islip.icnt b/configs/deprecated-cfgs/QuadroFX5600/icnt_config_islip.icnt new file mode 100644 index 0000000..de3bcc8 --- /dev/null +++ b/configs/deprecated-cfgs/QuadroFX5600/icnt_config_islip.icnt @@ -0,0 +1,70 @@ +//14*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 14; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/QuadroFX5800/config_quadro_islip.icnt b/configs/deprecated-cfgs/QuadroFX5800/config_quadro_islip.icnt new file mode 100644 index 0000000..cfe9cac --- /dev/null +++ b/configs/deprecated-cfgs/QuadroFX5800/config_quadro_islip.icnt @@ -0,0 +1,69 @@ +//18*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 18; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/QuadroFX5800/gpgpusim.config b/configs/deprecated-cfgs/QuadroFX5800/gpgpusim.config new file mode 100644 index 0000000..fef1110 --- /dev/null +++ b/configs/deprecated-cfgs/QuadroFX5800/gpgpusim.config @@ -0,0 +1,88 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 13 + +# high level architecture configuration +-gpgpu_n_clusters 10 +-gpgpu_n_cores_per_cluster 3 +-gpgpu_n_mem 8 +-gpgpu_clock_domains 325.0:650.0:650.0:800.0 + +# shader core pipeline config +-gpgpu_shader_registers 16384 +-gpgpu_occupancy_sm_number 13 + +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 1,1,1,1,1,1,1 +-gpgpu_num_sp_units 1 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 1,1,19,25,145 +-ptx_opcode_initiation_int 1,1,4,4,32 +-ptx_opcode_latency_fp 1,1,1,1,30 +-ptx_opcode_initiation_fp 1,1,1,1,5 +-ptx_opcode_latency_dp 8,8,8,8,335 +-ptx_opcode_initiation_dp 8,8,8,8,130 + +# memory stage behaviour +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +-gpgpu_cache:il1 4:256:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 8:128:5,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 +-gpgpu_cache:dl2 16:256:8,L:B:m:W:L,A:16:4,4 +-gpgpu_cache:dl2_texture_only 1 + +-gpgpu_shmem_warp_parts 2 + +# interconnection +-network_mode 1 +-inter_config_file config_quadro_islip.icnt + +# dram scheduler config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (30 core cycles). I.e. +# Total buffer space required = 30 x 800MHz / 325MHz = 74 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 58 + +# dram model config +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 4 +-dram_data_command_freq_ratio 2 # GDDR3 is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RRBBBCCC.CCCSSSSS +# GDDR3 timing from Samsung K4J52324QH-HC12 @ 800MHz +-gpgpu_dram_timing_opt nbk=8:CCD=2:RRD=8:RCD=12:RAS=25:RP=10:RC=35:CL=10:WL=7:CDLR=6:WR=11 + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 + +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/deprecated-cfgs/SM6_GTX1080/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM6_GTX1080/config_fermi_islip.icnt new file mode 100644 index 0000000..2a69ddd --- /dev/null +++ b/configs/deprecated-cfgs/SM6_GTX1080/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 50; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/SM6_GTX1080/gpgpusim.config b/configs/deprecated-cfgs/SM6_GTX1080/gpgpusim.config new file mode 100644 index 0000000..fb044c6 --- /dev/null +++ b/configs/deprecated-cfgs/SM6_GTX1080/gpgpusim.config @@ -0,0 +1,150 @@ +# This config models the Pascal GP102 (GeForceGTX 1080Ti) + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 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 1 +-gpgpu_n_mem 11 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1481.0:2962.0:1481.0:2750.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 60 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 1,2,1,1,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The default is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,8 +-gpgpu_shmem_size 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:1024:1024,4:0,32 # used to be 128:4 +-gpgpu_cache:dl2_texture_only 0 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +# gpgpu_num_reg_banks should be increased to 32, but it gives an error! +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA GeForceGTX 1080Ti, bus width is 352bits (11 DRAM chips x 32 bits) +# 11 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Pascal has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx1080Ti.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/deprecated-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml b/configs/deprecated-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml new file mode 100755 index 0000000..02619ff --- /dev/null +++ b/configs/deprecated-cfgs/SM6_GTX1080/gpuwattch_gtx1080Ti.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/SM6_P100/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM6_P100/config_fermi_islip.icnt new file mode 100644 index 0000000..d26c8d9 --- /dev/null +++ b/configs/deprecated-cfgs/SM6_P100/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 60; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/SM6_P100/gpgpusim.config b/configs/deprecated-cfgs/SM6_P100/gpgpusim.config new file mode 100644 index 0000000..a5e6736 --- /dev/null +++ b/configs/deprecated-cfgs/SM6_P100/gpgpusim.config @@ -0,0 +1,156 @@ +# This config models the Pascal GP100 +# For more info about this card, see Nvidia White paper +# https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf + +# 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 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 2,2,1,2,2,1,4 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-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 + +# ::,::::,::,:** +# ** 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 +# Pascal GP100 has 48KB L1 cache +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 65536 +-gpgpu_shmem_size_PrefShared 65536 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 0 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml b/configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/deprecated-cfgs/SM6_P100/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/SM6_TITANX/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM6_TITANX/config_fermi_islip.icnt new file mode 100644 index 0000000..602daee --- /dev/null +++ b/configs/deprecated-cfgs/SM6_TITANX/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config b/configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config new file mode 100644 index 0000000..28912a3 --- /dev/null +++ b/configs/deprecated-cfgs/SM6_TITANX/gpgpusim.config @@ -0,0 +1,157 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 61 + +# 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 GP102 has 4 SP SIMD units and 1 SFU unit +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,1,1,4,1,1,6 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# SFU is 32-width in pascal, then dp units initiation is 1 cycle +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,1,1,1,4 +-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 4,8,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Pascal GP102 has 96KB Shared memory +# Pascal GP102 has 64KB L1 cache +# The defulat is to disable the L1 cache, unless cache modifieres is used +-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_shmem_size 98304 +-gpgpu_shmem_size_PrefL1 98304 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 64:128:16,L:B:m:W:L,A:128:8,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 32:32:32:32 + +# 4 KB Inst. +-gpgpu_cache:il1 8:128:4,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 12 KB Const +-gpgpu_const_cache:l1 128:64:2,L:R:f:N:L,A:2:64,4 + +# enable operand collector +## larger operand collectors and reg_banks are needed for the 4 warp schedulers and 4 SIMD units +-gpgpu_operand_collector_num_units_sp 20 +-gpgpu_operand_collector_num_units_sfu 4 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +# gpgpu_num_reg_banks should be increased to 32 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 2 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing from hynix H5GQ1H24AFR +# disable bank groups for now, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=1:CCDL=0:RTPL=0" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 + diff --git a/configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml b/configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/deprecated-cfgs/SM6_TITANX/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/deprecated-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/deprecated-cfgs/SM7_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..fac792a --- /dev/null +++ b/configs/deprecated-cfgs/SM7_TITANV/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 64; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; + diff --git a/configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config b/configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config new file mode 100644 index 0000000..aefb04a --- /dev/null +++ b/configs/deprecated-cfgs/SM7_TITANV/gpgpusim.config @@ -0,0 +1,165 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA GP100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:1200.0:1200.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 + +# This implies a maximum of 64 warps/SM +-gpgpu_shader_core_pipeline 2048:32 +-gpgpu_shader_cta 32 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,1,4,4,1,9 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +# Volta GV100 has 64KB Shared memory +-gpgpu_cache:dl1 64:128:8,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefL1 64:128:16,L:L:m:N:H,A:256:8,16:0 +-gpgpu_cache:dl1PrefShared 32:128:6,L:L:m:N:H,A:256:8,16:0 +-gpgpu_shmem_size 65536 +-gpgpu_shmem_size_PrefL1 1 +-gpgpu_shmem_size_PrefShared 98304 +-gmem_skip_L1D 0 + +# 64 sets, each 128 bytes 24-way for each memory sub partition (192 KB per memory sub partition). This gives 4.5MB L2 cache +-gpgpu_cache:dl2 64:128:24,L:B:m:W:L,A:256:4,32:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 + +# 128 KB Inst. +-gpgpu_cache:il1 64:128:16,L:R:f:N:L,A:2:48,4 +# 48 KB Tex +-gpgpu_tex_cache:l1 16:128:24,L:R:m:N:L,F:128:4,128:2 +# 64 KB Const +-gpgpu_const_cache:l1 128:64:8,L:R:f:N:L,A:2:64,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 14 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_units_mem 10 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +## In Pascal, a warp scheduler can issue 2 insts per cycle +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Pascal P100 HBM runs at 715 MHZ +-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=3:RCD=10:RAS=24:RP=10:RC=34: + CL=10:WL=2:CDLR=3:WR=9:nbkgrp=4:CCDL=2:RTPL=3" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_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/deprecated-cfgs/TeslaC2050/config_fermi_islip.icnt b/configs/deprecated-cfgs/TeslaC2050/config_fermi_islip.icnt new file mode 100644 index 0000000..a11bd8e --- /dev/null +++ b/configs/deprecated-cfgs/TeslaC2050/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//20*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 32; + +// currently we donot use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 26; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 8; + +wait_for_tail_credit = 0; + +// Router architecture + +vc_allocator = islip; //separable_input_first; +sw_allocator = islip; //separable_input_first; +alloc_iters = 1; + +credit_delay = 0; +routing_delay = 0; +vc_alloc_delay = 1; +sw_alloc_delay = 1; + +input_speedup = 2; +output_speedup = 1; +internal_speedup = 1.0; + +// Traffic, GPGPU-Sim does not use this + +traffic = uniform; +packet_size ={{1,2,3,4},{10,20}}; +packet_size_rate={{1,1,1,1},{2,1}}; + +// Simulation - Don't change + +sim_type = gpgpusim; +//sim_type = latency; +injection_rate = 0.1; + + +subnets = 2; + +// Always use read and write no matter following line +//use_read_write = 1; + + +read_request_subnet = 0; +read_reply_subnet = 1; +write_request_subnet = 0; +write_reply_subnet = 1; + +read_request_begin_vc = 0; +read_request_end_vc = 0; +write_request_begin_vc = 0; +write_request_end_vc = 0; +read_reply_begin_vc = 0; +read_reply_end_vc = 0; +write_reply_begin_vc = 0; +write_reply_end_vc = 0; diff --git a/configs/deprecated-cfgs/TeslaC2050/gpgpusim.config b/configs/deprecated-cfgs/TeslaC2050/gpgpusim.config new file mode 100644 index 0000000..6ac2c12 --- /dev/null +++ b/configs/deprecated-cfgs/TeslaC2050/gpgpusim.config @@ -0,0 +1,133 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + +# Using cuobjdump to extract ptx/SASS +#-gpgpu_ptx_use_cuobjdump 1 # use default + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + + +# high level architecture configuration +-gpgpu_n_clusters 14 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 575.0:575.0:575.0:750.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_MEM,EX_WB +-gpgpu_pipeline_widths 2,1,1,2,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 32:128:4,L:L:m:N:H,A:32:8,8 +-gpgpu_shmem_size 49152 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:32:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 64:128:8,L:B:m:W:L,A:32:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 + +-gpgpu_cache:il1 4:128:4,L:R:f:N:L,A:2:32,4 +-gpgpu_tex_cache:l1 4:128:24,L:R:m:N:L,F:128:4,128:2 +-gpgpu_const_cache:l1 64:64:2,L:R:f:N:L,A:2:32,4 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 750MHz / 575MHz = 130 +-gpgpu_frfcfs_dram_sched_queue_size 16 +-gpgpu_dram_return_queue_size 114 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt b/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt new file mode 100644 index 0000000..c399db9 --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/config_fermi_islip.icnt @@ -0,0 +1,70 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 27; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; + +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 = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// 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/tested-cfgs/SM2_GTX480/gpgpusim.config b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config new file mode 100644 index 0000000..4096b09 --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config @@ -0,0 +1,153 @@ +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 20 + + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 15 +-gpgpu_n_cores_per_cluster 1 +-gpgpu_n_mem 6 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Fermi clock domains +#-gpgpu_clock_domains ::: +# In Fermi, each pipeline has 16 execution units, so the Core clock needs to be divided +# by 2. (GPGPU-Sim simulates a warp (32 threads) in a single cycle). 1400/2 = 700 +-gpgpu_clock_domains 700.0:700.0:700.0:924.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 + +# This implies a maximum of 48 warps/SM +-gpgpu_shader_core_pipeline 1536:32 +-gpgpu_shader_cta 8 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_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 +-gpgpu_pipeline_widths 2,0,1,1,2,0,1,1,2 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 1 +-gpgpu_num_dp_units 0 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 1,2,2,1,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 1,2,1,1,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 8,16,8,8,130 + + +# In Fermi, the cache and shared memory can be configured to 16kb:48kb(default) or 48kb:16kb +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# Note: Hashing set index function (H) only applies to a set size of 32 or 64. +-gpgpu_cache:dl1 N:32:128:4,L:L:m:N:H,S:64:8,8 +-gpgpu_shmem_size 49152 +-icnt_flit_size 40 +-gmem_skip_L1D 0 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 35 +-smem_latency 26 +-gpgpu_flush_l1_cache 1 + +# The alternative configuration for fermi in case cudaFuncCachePreferL1 is selected +#-gpgpu_cache:dl1 N:64:128:6,L:L:m:N:H,S:64:8,8 +#-gpgpu_shmem_size 16384 + +# 64 sets, each 128 bytes 8-way for each memory sub partition. This gives 786KB L2 cache +-gpgpu_cache:dl2 S:64:128:8,L:B:m:L:L,A:256:4,4:0,32 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 1 +-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,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 +-gpgpu_operand_collector_num_units_sp 6 +-gpgpu_operand_collector_num_units_sfu 8 +-gpgpu_operand_collector_num_in_ports_sp 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_num_reg_banks 16 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 20 + +-gpgpu_max_insn_issue_per_warp 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 116 + +# for Fermi, bus width is 384bits, this is 8 bytes (4 bytes at each DRAM chip) per memory partition +-gpgpu_n_mem_per_ctrlr 2 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5 is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.BBBCCCCB.CCSSSSSS + +# GDDR5 timing from hynix H5GQ1H24AFR +# to disable bank groups, set nbkgrp to 1 and tCCDL and tRTPL to 0 +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=6:RCD=12:RAS=28:RP=12:RC=40: + CL=12:WL=4:CDLR=5:WR=12:nbkgrp=4:CCDL=3:RTPL=2" + +# select lower bits for bnkgrp to increase bnkgrp parallelism +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Fermi has two schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs +-power_simulation_enabled 1 +-gpuwattch_xml_file gpuwattch_gtx480.xml + +# tracing functionality +#-trace_enabled 1 +#-trace_components WARP_SCHEDULER,SCOREBOARD +#-trace_sampling_core 0 diff --git a/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml b/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml new file mode 100755 index 0000000..304e0fd --- /dev/null +++ b/configs/tested-cfgs/SM2_GTX480/gpuwattch_gtx480.xml @@ -0,0 +1,538 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt b/configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt new file mode 100644 index 0000000..dec4789 --- /dev/null +++ b/configs/tested-cfgs/SM6_TITANX/config_fermi_islip.icnt @@ -0,0 +1,73 @@ +//21*1 fly with 32 flits per packet under gpgpusim injection mode +use_map = 0; +flit_size = 40; + +// currently we do not use this, see subnets below +network_count = 2; + +// Topology +topology = fly; +k = 52; +n = 1; + +// Routing + +routing_function = dest_tag; + +// Flow control + +num_vcs = 1; +vc_buf_size = 64; +input_buffer_size = 256; +ejection_buffer_size = 64; +boundary_buffer_size = 64; + +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 = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// 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/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config new file mode 100644 index 0000000..3842508 --- /dev/null +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -0,0 +1,185 @@ +# This config models the Pascal GP102 (NVIDIA TITAN X) +# For more info about this card, see Nvidia White paper +# http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 61 +-gpgpu_ignore_resources_limitation 1 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +# P102 has two semi-indp scheds per core, and two cores per cluster +-gpgpu_n_clusters 28 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 12 +-gpgpu_n_sub_partition_per_mchannel 2 + +# Pascal clock domains +#-gpgpu_clock_domains ::: +# Pascal NVIDIA TITAN X clock domains are adopted from +# https://en.wikipedia.org/wiki/GeForce_10_series +-gpgpu_clock_domains 1417.0:1417.0:1417.0:2500.0 + +# shader core pipeline config +-gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 62 + +# This implies a maximum of 32 warps/SM +-gpgpu_shader_core_pipeline 1024:32 +-gpgpu_shader_cta 16 +-gpgpu_simd_model 1 + +# Pipeline widths and number of FUs +# ID_OC_SP,ID_OC_DP,ID_OC_SFU,ID_OC_MEM,OC_EX_SP,OC_EX_DP,OC_EX_SFU,OC_EX_MEM,EX_WB +## Pascal GP102 has 4 SP SIMD units and 4 SFU units +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 2,1,2,1,2,1,2,1,5 +-gpgpu_num_sp_units 2 +-gpgpu_num_sfu_units 2 +-gpgpu_num_dp_units 1 + + +# 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 8,8,8,8,130 +-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 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:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:4:128:48,L:L:s:N:L,A:256:8,16:0,32 +-gpgpu_shmem_size 49152 +-gpgpu_shmem_size_PrefL1 49152 +-gpgpu_shmem_size_PrefShared 49152 +# By default, L1 cache is disabled in Pascal P102. +# requests with .nc modifier or __ldg mehtod will be cached in L1 cache even with gmem_skip_L1D=1 +-gmem_skip_L1D 1 +-icnt_flit_size 40 +-gpgpu_n_cluster_ejection_buffer_size 32 +-l1_latency 82 +-smem_latency 24 +-gpgpu_flush_l1_cache 1 + +# 64 sets, each 128 bytes 16-way for each memory sub partition (128 KB per memory sub partition). This gives 3MB L2 cache +-gpgpu_cache:dl2 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 1 + +# 4 KB Inst. +-gpgpu_cache:il1 N:8:128:4,L:R:f:N:L,S:2:48,4 +# 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 +# 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 2 +-gpgpu_operand_collector_num_out_ports_sp 2 +-gpgpu_operand_collector_num_in_ports_sfu 2 +-gpgpu_operand_collector_num_out_ports_sfu 2 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-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 +# Use Pascal Coalsce arhitetecture +-gpgpu_coalesce_arch 61 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 2 +-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 240 + +# for NVIDIA TITAN X, bus width is 384bits (12 DRAM chips x 32 bits) +# 12 memory paritions, 4 bytes (1 DRAM chip) per memory partition +# the atom size of GDDR5X (the smallest read request) is 32 bytes +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 4 +-gpgpu_dram_burst_length 8 +-dram_data_command_freq_ratio 4 # GDDR5X is QDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBCCCC.BCCSSSSS + +# Use the same GDDR5 timing +-gpgpu_dram_timing_opt "nbk=16:CCD=2:RRD=8:RCD=16:RAS=37:RP=16:RC=52: + CL=16:WL=6:CDLR=7:WR=16:nbkgrp=4:CCDL=4:RTPL=3" + +-dram_bnk_indexing_policy 0 +-dram_bnkgrp_indexing_policy 1 + +#-Seperate_Write_Queue_Enable 1 +#-Write_Queue_Size 64:56:32 + +# Pascal 102 has four schedulers per core +-gpgpu_num_sched_per_core 2 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, disable it untill we create a real energy model for Pascal 102 +-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/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt b/configs/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt new file mode 100644 index 0000000..2f25889 --- /dev/null +++ b/configs/tested-cfgs/SM7_TITANV/config_fermi_islip.icnt @@ -0,0 +1,74 @@ +//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 = 256; +input_buffer_size = 256; +ejection_buffer_size = 256; +boundary_buffer_size = 256; + +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 = 1; +output_speedup = 1; +internal_speedup = 2.0; + +// 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/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config new file mode 100644 index 0000000..c8351da --- /dev/null +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -0,0 +1,191 @@ +# This config models the Volta Titan X +# For more info about this card: +# http://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf +# https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8344474&tag=1# +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +# https://www.hotchips.org/wp-content/uploads/hc_archives/hc29/HC29.21-Monday-Pub/HC29.21.10-GPU-Gaming-Pub/HC29.21.132-Volta-Choquette-NVIDIA-Final3.pdf +# https://devblogs.nvidia.com/inside-volta/ +# http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf + +# functional simulator specification +-gpgpu_ptx_instruction_classification 0 +-gpgpu_ptx_sim_mode 0 +-gpgpu_ptx_force_max_capability 70 + +# SASS execution (only supported with CUDA >= 4.0) +-gpgpu_ptx_convert_to_ptxplus 0 +-gpgpu_ptx_save_converted_ptxplus 0 + +# high level architecture configuration +-gpgpu_n_clusters 40 +-gpgpu_n_cores_per_cluster 2 +-gpgpu_n_mem 24 +-gpgpu_n_sub_partition_per_mchannel 1 + +# volta clock domains +#-gpgpu_clock_domains ::: +# Volta NVIDIA GV100 clock domains are adopted from +# https://en.wikipedia.org/wiki/Volta_(microarchitecture) +-gpgpu_clock_domains 1200.0:2000.0:1200.0:850.0 +# boost mode +# -gpgpu_clock_domains 1455.0:2000.0:1455.0:850.0 + +# shader core pipeline config +-gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 70 + +# 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 +## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core +## we need to scale the number of pipeline registers to be equal to the number of SP units +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_num_sp_units 4 +-gpgpu_num_sfu_units 4 +-gpgpu_num_dp_units 4 + +# Instruction latencies and initiation intervals +# "ADD,MAX,MUL,MAD,DIV" +# All Div operations are executed on SFU unit +# Throughput (initiation latency) are adopted from +# http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf +-ptx_opcode_latency_int 4,13,4,5,145 +-ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_fp 4,13,4,5,39 +-ptx_opcode_initiation_fp 2,2,2,2,4 +-ptx_opcode_latency_dp 8,19,8,8,330 +-ptx_opcode_initiation_dp 4,4,4,4,130 +-ptx_opcode_latency_sfu 100 +-ptx_opcode_initiation_sfu 8 + + +# ::,::::,::,:** +# ** Optional parameter - Required when mshr_type==Texture Fifo +# 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 + +# 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 +-gpgpu_cache:dl2_texture_only 0 +-gpgpu_dram_partition_queues 64:64:64:64 +-perf_sim_memcpy 1 + +# 128 KB Inst. +-gpgpu_cache:il1 N:64:128:16,L:R:f:N:L,S:2:48,4 +# 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 + +# enable operand collector +-gpgpu_operand_collector_num_units_sp 12 +-gpgpu_operand_collector_num_units_sfu 6 +-gpgpu_operand_collector_num_units_mem 8 +-gpgpu_operand_collector_num_units_dp 6 +-gpgpu_operand_collector_num_in_ports_sp 4 +-gpgpu_operand_collector_num_out_ports_sp 4 +-gpgpu_operand_collector_num_in_ports_sfu 1 +-gpgpu_operand_collector_num_out_ports_sfu 1 +-gpgpu_operand_collector_num_in_ports_mem 1 +-gpgpu_operand_collector_num_out_ports_mem 1 +-gpgpu_operand_collector_num_in_ports_dp 1 +-gpgpu_operand_collector_num_out_ports_dp 1 +-gpgpu_num_reg_banks 32 + +# shared memory bankconflict detection +-gpgpu_shmem_num_banks 32 +-gpgpu_shmem_limited_broadcast 0 +-gpgpu_shmem_warp_parts 1 +-gpgpu_coalesce_arch 60 + +## In Pascal, a warp scheduler can issue 2 insts per cycle using 2 diff execution units +-gpgpu_max_insn_issue_per_warp 1 +-gpgpu_dual_issue_diff_exec_units 1 + +# interconnection +-network_mode 1 +-inter_config_file config_fermi_islip.icnt + +# memory partition latency config +-rop_latency 120 +-dram_latency 100 + +# dram model config +-gpgpu_dram_scheduler 1 +# The DRAM return queue and the scheduler queue together should provide buffer +# to sustain the memory level parallelism to tolerate DRAM latency +# To allow 100% DRAM utility, there should at least be enough buffer to sustain +# the minimum DRAM latency (100 core cycles). I.e. +# Total buffer space required = 100 x 924MHz / 700MHz = 132 +-gpgpu_frfcfs_dram_sched_queue_size 64 +-gpgpu_dram_return_queue_size 192 + +# for HBM, 32 channles, each (128 bits) 16 bytes width +-gpgpu_n_mem_per_ctrlr 1 +-gpgpu_dram_buswidth 16 +-gpgpu_dram_burst_length 2 +-dram_data_command_freq_ratio 2 # HBM is DDR +-gpgpu_mem_address_mask 1 +-gpgpu_mem_addr_mapping dramid@8;00000000.00000000.00000000.00000000.0000RRRR.RRRRRRRR.RBBBBCCC.CCCSSSSS + +# HBM timing are adopted from hynix JESD235 standered and nVidia HPCA 2017 paper (http://www.cs.utah.edu/~nil/pubs/hpca17.pdf) +# Timing for 1 GHZ +# tRRDl and tWTR are missing, need to be added +#-gpgpu_dram_timing_opt "nbk=16:CCD=1:RRD=4:RCD=14:RAS=33:RP=14:RC=47: +# CL=14:WL=2:CDLR=3:WR=12:nbkgrp=4:CCDL=2:RTPL=4" + +# Timing for 715 MHZ, Tesla Volta V100 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 4 +# Two Level Scheduler with active and pending pools +#-gpgpu_scheduler two_level_active:6:0:1 +# Loose round robbin scheduler +#-gpgpu_scheduler lrr +# Greedy then oldest scheduler +-gpgpu_scheduler gto + +# stat collection +-gpgpu_memlatency_stat 14 +-gpgpu_runtime_stat 500 +-enable_ptx_file_line_stats 1 +-visualizer_enabled 0 + +# power model configs, 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/cuobjdump_to_ptxplus/ptx_parser.h b/cuobjdump_to_ptxplus/ptx_parser.h index a534e92..ee7a942 100644 --- a/cuobjdump_to_ptxplus/ptx_parser.h +++ b/cuobjdump_to_ptxplus/ptx_parser.h @@ -376,4 +376,8 @@ 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 d67fd85..61af0ee 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -338,10 +338,19 @@ struct _cuda_device_id *GPGPUSim_Init() prop->minor = 2; prop->totalGlobalMem = 0x80000000 /* 2 GB */; prop->memPitch = 0; - prop->maxThreadsPerBlock = 512; - prop->maxThreadsDim[0] = 512; - prop->maxThreadsDim[1] = 512; - prop->maxThreadsDim[2] = 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; @@ -353,6 +362,9 @@ struct _cuda_device_id *GPGPUSim_Init() prop->clockRate = the_gpu->shader_clock(); #if (CUDART_VERSION >= 2010) prop->multiProcessorCount = the_gpu->get_config().num_shader(); +#endif +#if (CUDART_VERSION >= 4000) + prop->maxThreadsPerMultiProcessor = the_gpu->threads_per_core(); #endif the_gpu->set_prop(prop); the_device = new _cuda_device_id(the_gpu); @@ -2167,8 +2179,28 @@ cudaError_t CUDARTAPI cudaSetValidDevices(int *device_arr, int len) cudaError_t CUDARTAPI cudaSetDeviceFlags( int flags ) { - cuda_not_implemented(__my_func__,__LINE__); - return g_last_cudaError = cudaErrorUnknown; + // This flag is implicitly always on (unless you are using the driver API). It is safe for GPGPU-Sim to + // just ignore it. + if ( cudaDeviceMapHost == flags ) { + return g_last_cudaError = cudaSuccess; + } else { + cuda_not_implemented(__my_func__,__LINE__); + 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 ) @@ -2181,7 +2213,10 @@ 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/nightly.jenkinsfile b/nightly.jenkinsfile new file mode 100644 index 0000000..5221b3b --- /dev/null +++ b/nightly.jenkinsfile @@ -0,0 +1,75 @@ +pipeline { + agent { + label "purdue-cluster" + } + + options { + disableConcurrentBuilds() + overrideIndexTriggers(true) + } + + triggers { + pollSCM('0 1 * * *') + } + + stages { + stage('nightly-simulator-build') { + steps { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + make -j' + } + } + stage('nightly-simulations-build'){ + steps{ + sh 'rm -rf gpgpu-sim_simulations' + sh 'git clone git@github.rcac.purdue.edu:TimRogersGroup/gpgpu-sim_simulations.git && \ + cd gpgpu-sim_simulations && \ + git checkout purdue-cluster && \ + git pull && \ + ln -s /home/tgrogers-raid/a/common/data_dirs benchmarks/' + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + cd gpgpu-sim_simulations && \ + source ./benchmarks/src/setup_environment && \ + make -i -j -C ./benchmarks/src/ all && \ + make -C ./benchmarks/src data' + } + } + stage('nightly-2B-insn-run'){ + steps { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + ./gpgpu-sim_simulations/util/job_launching/run_simulations.py -B `cat ./gpgpu-sim_simulations/util/job_launching/apps/all-apps.list` -C TITANX-2B,TITANX-L1ON-2B,P100-2B,TITANV-2B -N nightly-$$ && \ + PLOTDIR="jenkins/${JOB_NAME}/${BUILD_NUMBER}/getstats" && ssh tgrogers@dynamo.ecn.purdue.edu mkdir -p /home/dynamo/a/tgrogers/website/gpgpu-sim-plots/$PLOTDIR && \ + ./gpgpu-sim_simulations/util/job_launching/monitor_func_test.py -I -S 1800 -v -s stats-$$.csv -N nightly-$$ && \ + ./gpgpu-sim_simulations/util/plotting/plot-get-stats.py -c stats-$$.csv -p tgrogers@dynamo.ecn.purdue.edu:~/website/gpgpu-sim-plots/$PLOTDIR -w https://engineering.purdue.edu/tgrogers/gpgpu-sim-plots/$PLOTDIR -n $PLOTDIR' + } + } + stage('nightly-correlate'){ + steps { + sh 'source /home/tgrogers-raid/a/common/gpgpu-sim-setup/9.1_env_setup.sh &&\ + source `pwd`/setup_environment &&\ + PLOTDIR="jenkins/${JOB_NAME}" &&\ + ./gpgpu-sim_simulations/util/job_launching/get_stats.py -R -K -k -B `cat ./gpgpu-sim_simulations/util/job_launching/apps/all-apps.list` -C TITANX-2B,TITANX-L1ON-2B,P100-2B,TITANV-2B > nightly-stats-per-kernel-9.1.csv &&\ + ./gpgpu-sim_simulations/util/plotting/correlate_and_publish.sh nightly-stats-per-kernel-9.1.csv $PLOTDIR ${BUILD_NUMBER}' + } + } + } + post { + success { + emailext body: "See ${BUILD_URL}", + recipientProviders: [[$class: 'CulpritsRecipientProvider'], + [$class: 'RequesterRecipientProvider']], + subject: "[AALP Jenkins] Build #${BUILD_NUMBER} - Success!", + to: 'tgrogers@purdue.edu' + } + failure { + emailext body: "See ${BUILD_URL}", + recipientProviders: [[$class: 'CulpritsRecipientProvider'], + [$class: 'RequesterRecipientProvider']], + subject: "[AALP Jenkins] Build #${BUILD_NUMBER} - ${currentBuild.result}", + to: 'tgrogers@purdue.edu' + } + } +} diff --git a/setup_environment b/setup_environment index f1af978..9578942 100644 --- a/setup_environment +++ b/setup_environment @@ -1,6 +1,6 @@ # see README before running this -ps -p $$ | awk '/bash/ || / sh/ || /zsh/ {exit 1;}' && echo "ERROR ** source setup_environment must be run in a bash, zsh or sh shell; see README" && exit +ps -p $$ | awk '/bash/ || / sh/ || /zsh/ {exit 1;}' && echo "WARNING ** source setup_environment must be run in a bash, zsh or sh shell; see README" export GPGPUSIM_SETUP_ENVIRONMENT_WAS_RUN= export GPGPUSIM_ROOT="$( cd "$( dirname "$BASH_SOURCE" )" && pwd )" diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index f7f1016..acb376a 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -316,12 +316,12 @@ void warp_inst_t::generate_mem_accesses() break; case global_space: case local_space: case param_space_local: - if( m_config->gpgpu_coalesce_arch == 13 ) { - if(isatomic()) - memory_coalescing_arch_13_atomic(is_write, access_type); - else - memory_coalescing_arch_13(is_write, access_type); - } else abort(); + if( m_config->gpgpu_coalesce_arch >= 13 && m_config->gpgpu_coalesce_arch <= 62) { + if(isatomic()) + memory_coalescing_arch_atomic(is_write, access_type); + else + memory_coalescing_arch(is_write, access_type); + } else abort(); break; @@ -345,7 +345,7 @@ void warp_inst_t::generate_mem_accesses() byte_mask.set(idx+i); } for( a=accesses.begin(); a != accesses.end(); ++a ) - m_accessq.push_back( mem_access_t(access_type,a->first,cache_block_size,is_write,a->second,byte_mask) ); + m_accessq.push_back( mem_access_t(access_type,a->first,cache_block_size,is_write,a->second, byte_mask, mem_access_sector_mask_t())); } if ( space.get_type() == global_space ) { @@ -354,15 +354,32 @@ void warp_inst_t::generate_mem_accesses() m_mem_accesses_created=true; } -void warp_inst_t::memory_coalescing_arch_13( bool is_write, mem_access_type access_type ) +void warp_inst_t::memory_coalescing_arch( bool is_write, mem_access_type access_type ) { // see the CUDA manual where it discusses coalescing rules before reading this unsigned segment_size = 0; unsigned warp_parts = m_config->mem_warp_parts; + bool sector_segment_size = false; + + if(m_config->gpgpu_coalesce_arch >= 20 && m_config->gpgpu_coalesce_arch < 39) + { + //Fermi and Kepler, L1 is normal and L2 is sector + if(m_config->gmem_skip_L1D || cache_op == CACHE_GLOBAL) + sector_segment_size = true; + else + sector_segment_size = false; + } + else if(m_config->gpgpu_coalesce_arch >= 40) + { + //Maxwell and Pascal, L1 and L2 are sectors + //all requests should be 32 bytes + sector_segment_size = true; + } + switch( data_size ) { case 1: segment_size = 32; break; - case 2: segment_size = 64; break; - case 4: case 8: case 16: segment_size = 128; break; + case 2: segment_size = sector_segment_size? 32 : 64; break; + case 4: case 8: case 16: segment_size = sector_segment_size? 32 : 128; break; } unsigned subwarp_size = m_config->warp_size / warp_parts; @@ -413,13 +430,13 @@ void warp_inst_t::memory_coalescing_arch_13( bool is_write, mem_access_type acce new_addr_type addr = t->first; const transaction_info &info = t->second; - memory_coalescing_arch_13_reduce_and_send(is_write, access_type, info, addr, segment_size); + memory_coalescing_arch_reduce_and_send(is_write, access_type, info, addr, segment_size); } } } -void warp_inst_t::memory_coalescing_arch_13_atomic( bool is_write, mem_access_type access_type ) +void warp_inst_t::memory_coalescing_arch_atomic( bool is_write, mem_access_type access_type ) { assert(space.get_type() == global_space); // Atomics allowed only for global memory @@ -488,13 +505,13 @@ void warp_inst_t::memory_coalescing_arch_13_atomic( bool is_write, mem_access_ty for(t=transaction_list.begin(); t!=transaction_list.end(); t++) { // For each transaction const transaction_info &info = *t; - memory_coalescing_arch_13_reduce_and_send(is_write, access_type, info, addr, segment_size); + memory_coalescing_arch_reduce_and_send(is_write, access_type, info, addr, segment_size); } } } } -void warp_inst_t::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 warp_inst_t::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 ) { assert( (addr & (segment_size-1)) == 0 ); @@ -543,7 +560,7 @@ void warp_inst_t::memory_coalescing_arch_13_reduce_and_send( bool is_write, mem_ assert(lower_half_used && upper_half_used); } } - m_accessq.push_back( mem_access_t(access_type,addr,size,is_write,info.active,info.bytes) ); + m_accessq.push_back( mem_access_t(access_type,addr,size,is_write,info.active,info.bytes, info.chunks) ); } void warp_inst_t::completed( unsigned long long cycle ) const @@ -577,6 +594,8 @@ 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 71d3d89..a612bac 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -82,6 +82,7 @@ enum uarch_op_t { ALU_OP=1, SFU_OP, TENSOR_CORE_OP, + DP_OP, ALU_SFU_OP, LOAD_OP, TENSOR_CORE_LOAD_OP, @@ -139,6 +140,7 @@ 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, TENSOR_CORE__OP, MEM__OP @@ -308,6 +310,8 @@ public: unsigned long long start_cycle; unsigned long long end_cycle; unsigned m_launch_latency; + + mutable bool volta_cache_config_set; }; struct core_config { @@ -347,6 +351,7 @@ struct core_config { 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 @@ -398,6 +403,7 @@ 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) @@ -523,7 +529,14 @@ 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 @@ -574,6 +587,7 @@ struct gpgpu_ptx_sim_info int cmem; int gmem; int regs; + unsigned maxthreads; unsigned ptx_version; unsigned sm_target; }; @@ -625,6 +639,9 @@ private: const unsigned MAX_MEMORY_ACCESS_SIZE = 128; typedef std::bitset mem_access_byte_mask_t; +const unsigned SECTOR_CHUNCK_SIZE = 4; //four sectors +const unsigned SECTOR_SIZE = 32 ; //sector is 32 bytes width +typedef std::bitset mem_access_sector_mask_t; #define NO_PARTIAL_WRITE (mem_access_byte_mask_t()) #define MEM_ACCESS_TYPE_TUP_DEF \ @@ -660,6 +677,7 @@ enum cache_operator_type { CACHE_ALL, // .ca CACHE_LAST_USE, // .lu CACHE_VOLATILE, // .cv + CACHE_L1, // .nc // loads and stores CACHE_STREAMING, // .cs @@ -689,8 +707,9 @@ 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; @@ -706,6 +725,7 @@ 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 { @@ -739,6 +759,7 @@ 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; }; @@ -961,9 +982,9 @@ 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*), diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index d5c5c3d..87a0b9c 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -62,8 +62,8 @@ addr_t g_debug_pc = 0xBEEF1518; unsigned g_ptx_sim_num_insn = 0; unsigned gpgpu_param_num_shaders = 0; -char *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp; -char *opcode_initiation_int, *opcode_initiation_fp, *opcode_initiation_dp; +char *opcode_latency_int, *opcode_latency_fp, *opcode_latency_dp,*opcode_latency_sfu; +char *opcode_initiation_int, *opcode_initiation_fp, *opcode_initiation_dp,*opcode_initiation_sfu; char *cdp_latency_str; unsigned cdp_latency[5]; @@ -80,6 +80,10 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Opcode latencies for double precision floating points " "Default 8,8,8,8,335", "8,8,8,8,335"); + option_parser_register(opp, "-ptx_opcode_latency_sfu", OPT_CSTR, &opcode_latency_sfu, + "Opcode latencies for SFU instructions" + "Default 8", + "8"); option_parser_register(opp, "-ptx_opcode_initiation_int", OPT_CSTR, &opcode_initiation_int, "Opcode initiation intervals for integers " "Default 1,1,4,4,32", @@ -92,6 +96,10 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Opcode initiation intervals for double precision floating points " "Default 8,8,8,8,130", "8,8,8,8,130"); + option_parser_register(opp, "-ptx_opcode_initiation_sfu", OPT_CSTR, &opcode_initiation_sfu, + "Opcode initiation intervals for sfu instructions" + "Default 8", + "8"); option_parser_register(opp, "-cdp_latency", OPT_CSTR, &cdp_latency_str, "CDP API latency write(dst_start_addr+n,1, src_data+n,NULL,NULL); + + // Copy into the performance model. + extern gpgpu_sim* g_the_gpu; + g_the_gpu->perf_memcpy_to_gpu(dst_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); @@ -408,6 +420,10 @@ void gpgpu_t::memcpy_from_gpu( void *dst, size_t src_start_addr, size_t count ) unsigned char *dst_data = (unsigned char*)dst; for (unsigned n=0; n < count; n ++ ) m_global_mem->read(src_start_addr+n,1,dst_data+n); + + // Copy into the performance model. + extern gpgpu_sim* g_the_gpu; + g_the_gpu->perf_memcpy_to_gpu(src_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); @@ -592,9 +608,11 @@ void ptx_instruction::set_opcode_and_latency() unsigned int_latency[5]; unsigned fp_latency[5]; unsigned dp_latency[5]; + unsigned sfu_latency; unsigned int_init[5]; unsigned fp_init[5]; unsigned dp_init[5]; + unsigned sfu_init; /* * [0] ADD,SUB * [1] MAX,Min @@ -611,7 +629,9 @@ void ptx_instruction::set_opcode_and_latency() sscanf(opcode_latency_dp, "%u,%u,%u,%u,%u", &dp_latency[0],&dp_latency[1],&dp_latency[2], &dp_latency[3],&dp_latency[4]); - sscanf(opcode_initiation_int, "%u,%u,%u,%u,%u,%u, %u", + sscanf(opcode_latency_sfu, "%u", + &sfu_latency); + sscanf(opcode_initiation_int, "%u,%u,%u,%u,%u", &int_init[0],&int_init[1],&int_init[2], &int_init[3],&int_init[4]); sscanf(opcode_initiation_fp, "%u,%u,%u,%u,%u", @@ -620,8 +640,10 @@ void ptx_instruction::set_opcode_and_latency() sscanf(opcode_initiation_dp, "%u,%u,%u,%u,%u", &dp_init[0],&dp_init[1],&dp_init[2], &dp_init[3],&dp_init[4]); + sscanf(opcode_initiation_sfu, "%u", + &sfu_init); sscanf(cdp_latency_str, "%u,%u,%u,%u,%u", - &cdp_latency[0],&cdp_latency[1],&cdp_latency[2], + &cdp_latency[0],&cdp_latency[1],&cdp_latency[2], &cdp_latency[3],&cdp_latency[4]); if(!m_operands.empty()){ @@ -684,6 +706,7 @@ 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: @@ -705,6 +728,7 @@ 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: @@ -721,13 +745,12 @@ void ptx_instruction::set_opcode_and_latency() case F32_TYPE: latency = fp_latency[2]; initiation_interval = fp_init[2]; - op = ALU_SFU_OP; break; case F64_TYPE: 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: @@ -750,6 +773,7 @@ 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: @@ -785,8 +809,8 @@ void ptx_instruction::set_opcode_and_latency() break; case SQRT_OP: case SIN_OP: case COS_OP: case EX2_OP: case LG2_OP: case RSQRT_OP: case RCP_OP: //Using double to approximate those - latency = dp_latency[2]; - initiation_interval = dp_init[2]; + latency = sfu_latency; + initiation_interval = sfu_init; op = SFU_OP; break; case MMA_OP: @@ -796,7 +820,7 @@ void ptx_instruction::set_opcode_and_latency() break; case SHFL_OP: latency = 32; - initiation_interval = 15; + initiation_interval = 4; break; default: break; @@ -889,6 +913,7 @@ 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; @@ -1169,13 +1194,13 @@ 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); } 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/instructions.cc b/src/cuda-sim/instructions.cc index 4614f25..973eeab 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -33,6 +33,7 @@ #include "ptx.tab.h" #include #include +#include #include #include "cuda-math.h" #include "../abstract_hardware_model.h" diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y index 8b901a3..45392fb 100644 --- a/src/cuda-sim/ptx.y +++ b/src/cuda-sim/ptx.y @@ -237,7 +237,8 @@ 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 d12c741..1813f8c 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; } @@ -1234,16 +1235,16 @@ ptx_instruction::ptx_instruction( int opcode, m_inst_size = 4; // bytes break; case EXTP_OPTION: - break; + break; case NC_OPTION: - break; + m_cache_option = last_ptx_inst_option; + break; case UP_OPTION: case DOWN_OPTION: case BFLY_OPTION: case IDX_OPTION: m_shfl_op = last_ptx_inst_option; break; - case PRMT_F4E_MODE: case PRMT_B4E_MODE: case PRMT_RC8_MODE: diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 0767379..4c36cfc 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1333,6 +1333,7 @@ public: const struct gpgpu_ptx_sim_info* get_kernel_info () const { + assert (m_kernel_info.maxthreads == maxnt_id); return &m_kernel_info; } @@ -1340,6 +1341,8 @@ 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() { @@ -1363,7 +1366,11 @@ 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 9671ab7..cf40365 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -1012,6 +1012,10 @@ 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 8094b43..7b6e3a2 100644 --- a/src/cuda-sim/ptx_parser.h +++ b/src/cuda-sim/ptx_parser.h @@ -95,6 +95,7 @@ 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(); diff --git a/src/gpgpu-sim/addrdec.cc b/src/gpgpu-sim/addrdec.cc index 422576d..8651869 100644 --- a/src/gpgpu-sim/addrdec.cc +++ b/src/gpgpu-sim/addrdec.cc @@ -62,6 +62,9 @@ void linear_to_raw_address_translation::addrdec_setoption(option_parser_t opp) option_parser_register(opp, "-gpgpu_mem_address_mask", OPT_INT32, &gpgpu_mem_address_mask, "0 = old addressing mask, 1 = new addressing mask, 2 = new add. mask + flipped bank sel and chip sel bits", "0"); + option_parser_register(opp, "-memory_partition_indexing", OPT_UINT32, &memory_partition_indexing, + "0 = no indexing, 1 = bitwise xoring, 2 = IPoly, 3 = custom indexing", + "0"); } new_addr_type linear_to_raw_address_translation::partition_address( new_addr_type addr ) const @@ -103,6 +106,74 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, addrdec_ tlx->burst= addrdec_packbits(addrdec_mask[BURST], rest_of_addr, addrdec_mkhigh[BURST], addrdec_mklow[BURST]); } + switch(memory_partition_indexing){ + case CONSECUTIVE: + //Do nothing + break; + case BITWISE_PERMUTATION: + { + assert(!gap); + tlx->chip = (tlx->chip) ^ (tlx->row & (m_n_channel-1)); + assert(tlx->chip < m_n_channel); + break; + } + case IPOLY: + { + /* + * Set Indexing function from "Pseudo-randomly interleaved memory." + * Rau, B. R et al. + * ISCA 1991 + * + * equations are adopted from: + * "Sacat: streaming-aware conflict-avoiding thrashing-resistant gpgpu cache management scheme." + * Khairy et al. + * IEEE TPDS 2017. + */ + if(m_n_channel == 32) { + std::bitset<64> a(tlx->row); + std::bitset<5> chip(tlx->chip); + chip[0] = a[13]^a[12]^a[11]^a[10]^a[9]^a[6]^a[5]^a[3]^a[0]^chip[0]; + chip[1] = a[14]^a[13]^a[12]^a[11]^a[10]^a[7]^a[6]^a[4]^a[1]^chip[1]; + chip[2] = a[14]^a[10]^a[9]^a[8]^a[7]^a[6]^a[3]^a[2]^a[0]^chip[2]; + chip[3] = a[11]^a[10]^a[9]^a[8]^a[7]^a[4]^a[3]^a[1]^chip[3]; + chip[4] = a[12]^a[11]^a[10]^a[9]^a[8]^a[5]^a[4]^a[2]^chip[4]; + tlx->chip = chip.to_ulong(); + + } + else{ /* Else incorrect number of channels for the hashing function */ + assert("\nGPGPU-Sim memory_partition_indexing error: The number of channels should be " + "32 for the hashing IPOLY index function.\n" && 0); + } + assert(tlx->chip < m_n_channel); + break; + } + case PAE: + { + //Page Address Entropy + //random selected bits from the page and bank bits + //similar to + //Liu, Yuxi, et al. "Get Out of the Valley: Power-Efficient Address Mapping for GPUs." ISCA 2018 + std::bitset<64> a(tlx->row); + std::bitset<5> chip(tlx->chip); + std::bitset<4> b(tlx->bk); + chip[0] = a[13]^a[10]^a[9]^a[5]^a[0]^b[3]^b[0]^chip[0]; + chip[1] = a[12]^a[11]^a[6]^a[1]^b[3]^b[2]^b[1]^chip[1]; + chip[2] = a[14]^a[9]^a[8]^a[7]^a[2]^b[1]^chip[2]; + chip[3] = a[11]^a[10]^a[8]^a[3]^b[2]^b[3]^chip[3]; + chip[4] = a[12]^a[9]^a[8]^a[5]^a[4]^b[1]^b[0]^chip[4]; + tlx->chip = chip.to_ulong(); + assert(tlx->chip < m_n_channel); + break; + } + case CUSTOM: + /* No custom set function implemented */ + //Do you custom index here + break; + default: + assert("\nUndefined set index function.\n" && 0); + break; + } + // combine the chip address and the lower bits of DRAM bank address to form the subpartition ID unsigned sub_partition_addr_mask = m_n_sub_partition_in_channel - 1; tlx->sub_partition = tlx->chip * m_n_sub_partition_in_channel diff --git a/src/gpgpu-sim/addrdec.h b/src/gpgpu-sim/addrdec.h index fd9af8d..bdc5fec 100644 --- a/src/gpgpu-sim/addrdec.h +++ b/src/gpgpu-sim/addrdec.h @@ -35,6 +35,14 @@ #include "../abstract_hardware_model.h" +enum partition_index_function{ + CONSECUTIVE = 0, + BITWISE_PERMUTATION, + IPOLY, + PAE, + CUSTOM +}; + struct addrdec_t { void print( FILE *fp ) const; @@ -72,6 +80,7 @@ private: const char *addrdec_option; int gpgpu_mem_address_mask; + partition_index_function memory_partition_indexing; bool run_test; int ADDR_CHIP_S; diff --git a/src/gpgpu-sim/delayqueue.h b/src/gpgpu-sim/delayqueue.h index b25f143..0caa5d4 100644 --- a/src/gpgpu-sim/delayqueue.h +++ b/src/gpgpu-sim/delayqueue.h @@ -161,6 +161,7 @@ public: } bool full() const { return (m_max_len && m_length >= m_max_len); } + bool is_avilable_size(unsigned size) const { return (m_max_len && m_length+size-1 >= m_max_len); } bool empty() const { return m_head == NULL; } unsigned get_n_element() const { return m_n_element; } unsigned get_length() const { return m_length; } diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc index a0e024b..6c11b43 100644 --- a/src/gpgpu-sim/dram.cc +++ b/src/gpgpu-sim/dram.cc @@ -49,11 +49,45 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m m_stats = stats; m_config = config; + //rowblp + access_num=0; + hits_num=0; + read_num=0; + write_num=0; + hits_read_num=0; + hits_write_num=0; + banks_1time=0; + banks_acess_total=0; + banks_acess_total_after=0; + banks_time_ready=0; + banks_access_ready_total=0; + issued_two=0; + issued_total=0; + issued_total_row=0; + issued_total_col=0; + CCDc = 0; RRDc = 0; RTWc = 0; WTRc = 0; + wasted_bw_row=0; + wasted_bw_col=0; + util_bw=0; + idle_bw=0; + RCDc_limit=0; + CCDLc_limit=0; + CCDLc_limit_alone=0; + CCDc_limit=0; + WTRc_limit=0; + WTRc_limit_alone=0; + RCDWRc_limit=0; + RTWc_limit=0; + RTWc_limit_alone=0; + rwq_limit=0; + write_to_read_ratio_blp_rw_average=0; + bkgrp_parallsim_rw=0; + rw = READ; //read mode is default bkgrp = (bankgrp_t**) calloc(sizeof(bankgrp_t*), m_config->nbkgrp); @@ -74,12 +108,13 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m bk[i]->state = BANK_IDLE; bk[i]->bkgrpindex = i/(m_config->nbk/m_config->nbkgrp); } - prio = 0; + prio = 0; + rwq = new fifo_pipeline("rwq",m_config->CL,m_config->CL+1); mrqq = new fifo_pipeline("mrqq",0,2); returnq = new fifo_pipeline("dramreturnq",0,m_config->gpgpu_dram_return_queue_size==0?1024:m_config->gpgpu_dram_return_queue_size); m_frfcfs_scheduler = NULL; - if ( m_config->scheduler_type == DRAM_FRFCFS ) + if ( m_config->scheduler_type == DRAM_FRFCFS) m_frfcfs_scheduler = new frfcfs_scheduler(m_config,this,stats); n_cmd = 0; n_activity = 0; @@ -88,6 +123,8 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m n_pre = 0; n_rd = 0; n_wr = 0; + n_wr_WB=0; + n_rd_L2_A=0; n_req = 0; max_mrqs_temp = 0; bwutil = 0; @@ -113,13 +150,21 @@ dram_t::dram_t( unsigned int partition_id, const struct memory_config *config, m mrqq_Dist = StatCreate("mrqq_length",1, queue_limit()); else //queue length is unlimited; mrqq_Dist = StatCreate("mrqq_length",1,64); //track up to 64 entries + } -bool dram_t::full() const +bool dram_t::full(bool is_write) const { - if(m_config->scheduler_type == DRAM_FRFCFS ){ + if(m_config->scheduler_type == DRAM_FRFCFS){ if(m_config->gpgpu_frfcfs_dram_sched_queue_size == 0 ) return false; - return m_frfcfs_scheduler->num_pending() >= m_config->gpgpu_frfcfs_dram_sched_queue_size; + if(m_config->seperate_write_queue_enabled){ + if(is_write) + return m_frfcfs_scheduler->num_write_pending() >= m_config->gpgpu_frfcfs_dram_write_queue_size; + else + return m_frfcfs_scheduler->num_pending() >= m_config->gpgpu_frfcfs_dram_sched_queue_size; + } + else + return m_frfcfs_scheduler->num_pending() >= m_config->gpgpu_frfcfs_dram_sched_queue_size; } else return mrqq->full(); } @@ -127,7 +172,7 @@ bool dram_t::full() const unsigned dram_t::que_length() const { unsigned nreqs = 0; - if (m_config->scheduler_type == DRAM_FRFCFS ) { + if (m_config->scheduler_type == DRAM_FRFCFS) { nreqs = m_frfcfs_scheduler->num_pending(); } else { nreqs = mrqq->get_length(); @@ -146,7 +191,7 @@ unsigned int dram_t::queue_limit() const } -dram_req_t::dram_req_t( class mem_fetch *mf ) +dram_req_t::dram_req_t( class mem_fetch *mf, unsigned banks, unsigned dram_bnk_indexing_policy) { txbytes = 0; dqbytes = 0; @@ -154,7 +199,29 @@ dram_req_t::dram_req_t( class mem_fetch *mf ) const addrdec_t &tlx = mf->get_tlx_addr(); - bk = tlx.bk; + switch(dram_bnk_indexing_policy){ + case LINEAR_BK_INDEX: + { + bk = tlx.bk; + break; + } + case BITWISE_XORING_BK_INDEX: + { + //xoring bank bits with lower bits of the page + int lbank = log2(banks); + bk = tlx.bk ^ (tlx.row & ((1<get_data_size(); @@ -169,14 +236,15 @@ void dram_t::push( class mem_fetch *data ) { assert(id == data->get_tlx_addr().chip); // Ensure request is in correct memory partition - dram_req_t *mrq = new dram_req_t(data); + dram_req_t *mrq = new dram_req_t(data,m_config->nbk,m_config->dram_bnk_indexing_policy); + data->set_status(IN_PARTITION_MC_INTERFACE_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); - mrqq->push(mrq); + mrqq->push(mrq); // stats... n_req += 1; n_req_partial += 1; - if ( m_config->scheduler_type == DRAM_FRFCFS ) { + if ( m_config->scheduler_type == DRAM_FRFCFS) { unsigned nreqs = m_frfcfs_scheduler->num_pending(); if ( nreqs > max_mrqs_temp) max_mrqs_temp = nreqs; @@ -212,6 +280,7 @@ void dram_t::cycle() printf("\tDQ: BK%d Row:%03x Col:%03x", cmd->bk, cmd->row, cmd->col + cmd->dqbytes); #endif cmd->dqbytes += m_config->dram_atom_size; + if (cmd->dqbytes >= cmd->nbytes) { mem_fetch *data = cmd->data; data->set_status(IN_PARTITION_MC_RETURNQ,gpu_sim_cycle+gpu_tot_sim_cycle); @@ -240,7 +309,7 @@ void dram_t::cycle() printf("Error: Unknown DRAM scheduler type\n"); assert(0); } - if ( m_config->scheduler_type == DRAM_FRFCFS ) { + if ( m_config->scheduler_type == DRAM_FRFCFS) { unsigned nreqs = m_frfcfs_scheduler->num_pending(); if ( nreqs > max_mrqs) { max_mrqs = nreqs; @@ -258,130 +327,123 @@ void dram_t::cycle() unsigned k=m_config->nbk; bool issued = false; - // check if any bank is ready to issue a new read + //collect row buffer locality, BLP and other statistics + ///////////////////////////////////////////////////////////////////////// + unsigned int memory_pending=0; for (unsigned i=0;inbk;i++) { - unsigned j = (i + prio) % m_config->nbk; - unsigned grp = j>>m_config->bk_tag_length; - if (bk[j]->mrq) { //if currently servicing a memory request - bk[j]->mrq->data->set_status(IN_PARTITION_DRAM,gpu_sim_cycle+gpu_tot_sim_cycle); - // correct row activated for a READ - if ( !issued && !CCDc && !bk[j]->RCDc && - !(bkgrp[grp]->CCDLc) && - (bk[j]->curr_row == bk[j]->mrq->row) && - (bk[j]->mrq->rw == READ) && (WTRc == 0 ) && - (bk[j]->state == BANK_ACTIVE) && - !rwq->full() ) { - if (rw==WRITE) { - rw=READ; - rwq->set_min_length(m_config->CL); - } - rwq->push(bk[j]->mrq); - bk[j]->mrq->txbytes += m_config->dram_atom_size; - CCDc = m_config->tCCD; - bkgrp[grp]->CCDLc = m_config->tCCDL; - RTWc = m_config->tRTW; - bk[j]->RTPc = m_config->BL/m_config->data_command_freq_ratio; - bkgrp[grp]->RTPLc = m_config->tRTPL; - issued = true; - n_rd++; - bwutil += m_config->BL/m_config->data_command_freq_ratio; - bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; - bk[j]->n_access++; -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tRD Bk:%d Row:%03x Col:%03x \n", - j, bk[j]->curr_row, - bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); -#endif - // transfer done - if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { - bk[j]->mrq = NULL; - } - } else - // correct row activated for a WRITE - if ( !issued && !CCDc && !bk[j]->RCDWRc && - !(bkgrp[grp]->CCDLc) && - (bk[j]->curr_row == bk[j]->mrq->row) && - (bk[j]->mrq->rw == WRITE) && (RTWc == 0 ) && - (bk[j]->state == BANK_ACTIVE) && - !rwq->full() ) { - if (rw==READ) { - rw=WRITE; - rwq->set_min_length(m_config->WL); - } - rwq->push(bk[j]->mrq); - - bk[j]->mrq->txbytes += m_config->dram_atom_size; - CCDc = m_config->tCCD; - bkgrp[grp]->CCDLc = m_config->tCCDL; - WTRc = m_config->tWTR; - bk[j]->WTPc = m_config->tWTP; - issued = true; - n_wr++; - bwutil += m_config->BL/m_config->data_command_freq_ratio; - bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tWR Bk:%d Row:%03x Col:%03x \n", - j, bk[j]->curr_row, - bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); -#endif - // transfer done - if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { - bk[j]->mrq = NULL; - } - } - - else - // bank is idle - if ( !issued && !RRDc && - (bk[j]->state == BANK_IDLE) && - !bk[j]->RPc && !bk[j]->RCc ) { -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tACT BK:%d NewRow:%03x From:%03x \n", - j,bk[j]->mrq->row,bk[j]->curr_row); -#endif - // activate the row with current memory request - bk[j]->curr_row = bk[j]->mrq->row; - bk[j]->state = BANK_ACTIVE; - RRDc = m_config->tRRD; - bk[j]->RCDc = m_config->tRCD; - bk[j]->RCDWRc = m_config->tRCDWR; - bk[j]->RASc = m_config->tRAS; - bk[j]->RCc = m_config->tRC; - prio = (j + 1) % m_config->nbk; - issued = true; - n_act_partial++; - n_act++; - } - - else - // different row activated - if ( (!issued) && - (bk[j]->curr_row != bk[j]->mrq->row) && - (bk[j]->state == BANK_ACTIVE) && - (!bk[j]->RASc && !bk[j]->WTPc && - !bk[j]->RTPc && - !bkgrp[grp]->RTPLc) ) { - // make the bank idle again - bk[j]->state = BANK_IDLE; - bk[j]->RPc = m_config->tRP; - prio = (j + 1) % m_config->nbk; - issued = true; - n_pre++; - n_pre_partial++; -#ifdef DRAM_VERIFY - PRINT_CYCLE=1; - printf("\tPRE BK:%d Row:%03x \n", j,bk[j]->curr_row); -#endif - } - } else { - if (!CCDc && !RRDc && !RTWc && !WTRc && !bk[j]->RCDc && !bk[j]->RASc - && !bk[j]->RCc && !bk[j]->RPc && !bk[j]->RCDWRc) k--; - bk[j]->n_idle++; - } + if (bk[i]->mrq) + memory_pending++; } + banks_1time += memory_pending; + if(memory_pending >0) + banks_acess_total++; + + unsigned int memory_pending_rw=0; + unsigned read_blp_rw=0; + unsigned write_blp_rw=0; + std::bitset<8> bnkgrp_rw_found; //assume max we have 8 bank groups + + for (unsigned j=0;jnbk;j++) { + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && + (bk[j]->state == BANK_ACTIVE)))) + { + memory_pending_rw++; + read_blp_rw++; + bnkgrp_rw_found.set(grp); + } + else if + (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && + (bk[j]->state == BANK_ACTIVE)))) + { + memory_pending_rw++; + write_blp_rw++; + bnkgrp_rw_found.set(grp); + } + } + banks_time_rw += memory_pending_rw; + bkgrp_parallsim_rw += bnkgrp_rw_found.count(); + if(memory_pending_rw >0) + { + write_to_read_ratio_blp_rw_average += (double)write_blp_rw/(write_blp_rw+read_blp_rw); + banks_access_rw_total++; + } + + unsigned int memory_Pending_ready=0; + for (unsigned j=0;jnbk;j++) { + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq && ((!CCDc && !bk[j]->RCDc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && (WTRc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full()) + || + (!CCDc && !bk[j]->RCDWRc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && (RTWc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full()))) + { + memory_Pending_ready++; + } + } + banks_time_ready += memory_Pending_ready; + if(memory_Pending_ready >0) + banks_access_ready_total++; + /////////////////////////////////////////////////////////////////////////////////// + + bool issued_col_cmd = false; + bool issued_row_cmd = false; + + if(m_config->dual_bus_interface) + { + //dual bus interface + //issue one row command and one column command + for (unsigned i=0;inbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + issued_col_cmd = issue_col_command(j); + if(issued_col_cmd) break; + } + for (unsigned i=0;inbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + issued_row_cmd = issue_row_command(j); + if(issued_row_cmd) break; + } + for (unsigned i=0;inbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + if(!bk[j]->mrq) { + if (!CCDc && !RRDc && !RTWc && !WTRc && !bk[j]->RCDc && !bk[j]->RASc + && !bk[j]->RCc && !bk[j]->RPc && !bk[j]->RCDWRc) k--; + bk[j]->n_idle++; + } + } + } + else + { + //single bus interface + //issue only one row/column command + for (unsigned i=0;inbk;i++) { + unsigned j = (i + prio) % m_config->nbk; + if(!issued_col_cmd) + issued_col_cmd = issue_col_command(j); + + if(!issued_col_cmd && !issued_row_cmd) + issued_row_cmd = issue_row_command(j); + + if(!bk[j]->mrq) { + if (!CCDc && !RRDc && !RTWc && !WTRc && !bk[j]->RCDc && !bk[j]->RASc + && !bk[j]->RCc && !bk[j]->RPc && !bk[j]->RCDWRc) k--; + bk[j]->n_idle++; + } + + } + } + + issued = issued_row_cmd || issued_col_cmd; if (!issued) { n_nop++; n_nop_partial++; @@ -395,6 +457,85 @@ void dram_t::cycle() } n_cmd++; n_cmd_partial++; + if(issued) + { + issued_total++; + if(issued_col_cmd && issued_row_cmd) + issued_two++; + } + if(issued_col_cmd) issued_total_col++; + if(issued_row_cmd) issued_total_row++; + + + //Collect some statistics + //check the limitation, see where BW is wasted? + ///////////////////////////////////////////////////////// + unsigned int memory_pending_found=0; + for (unsigned i=0;inbk;i++) { + if (bk[i]->mrq) + memory_pending_found++; + } + if(memory_pending_found>0) + banks_acess_total_after++; + + bool memory_pending_rw_found=false; + for (unsigned j=0;jnbk;j++) { + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && + (bk[j]->state == BANK_ACTIVE)) + || + ( + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && + (bk[j]->state == BANK_ACTIVE)))) + memory_pending_rw_found=true; + } + + + if(issued_col_cmd || CCDc) + util_bw++; + else if (memory_pending_rw_found) + { + wasted_bw_col++; + for (unsigned j=0;jnbk;j++) { + unsigned grp = get_bankgrp_number(j); + //read + if (bk[j]->mrq && (((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && + (bk[j]->state == BANK_ACTIVE)))) + { + if(bk[j]->RCDc) RCDc_limit++; + if(bkgrp[grp]->CCDLc) CCDLc_limit++; + if(WTRc) WTRc_limit++; + if(CCDc) CCDc_limit++; + if(rwq->full()) rwq_limit++; + if(bkgrp[grp]->CCDLc && !WTRc) CCDLc_limit_alone++; + if(!bkgrp[grp]->CCDLc && WTRc) WTRc_limit_alone++; + } + //write + else if (bk[j]->mrq && ((bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && + (bk[j]->state == BANK_ACTIVE))) + { + if(bk[j]->RCDWRc) RCDWRc_limit++; + if(bkgrp[grp]->CCDLc) CCDLc_limit++; + if(RTWc) RTWc_limit++; + if(CCDc) CCDc_limit++; + if(rwq->full()) rwq_limit++; + if(bkgrp[grp]->CCDLc && !RTWc) CCDLc_limit_alone++; + if(!bkgrp[grp]->CCDLc && RTWc) RTWc_limit_alone++; + } + } + } + else if (memory_pending_found) + wasted_bw_row++; + else if (!memory_pending_found) + idle_bw++; + else + assert(1); + + ///////////////////////////////////////////////////////// // decrements counters once for each time dram_issueCMD is called DEC2ZERO(RRDc); @@ -420,39 +561,237 @@ void dram_t::cycle() #endif } +bool dram_t::issue_col_command(int j) +{ + bool issued = false; + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq) { //if currently servicing a memory request + bk[j]->mrq->data->set_status(IN_PARTITION_DRAM,gpu_sim_cycle+gpu_tot_sim_cycle); + // correct row activated for a READ + if ( !issued && !CCDc && !bk[j]->RCDc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == READ) && (WTRc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full() ) { + if (rw==WRITE) { + rw=READ; + rwq->set_min_length(m_config->CL); + } + rwq->push(bk[j]->mrq); + bk[j]->mrq->txbytes += m_config->dram_atom_size; + CCDc = m_config->tCCD; + bkgrp[grp]->CCDLc = m_config->tCCDL; + RTWc = m_config->tRTW; + bk[j]->RTPc = m_config->BL/m_config->data_command_freq_ratio; + bkgrp[grp]->RTPLc = m_config->tRTPL; + issued = true; + if(bk[j]->mrq->data->get_access_type() == L2_WR_ALLOC_R) + n_rd_L2_A++; + else + n_rd++; + + bwutil += m_config->BL/m_config->data_command_freq_ratio; + bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; + bk[j]->n_access++; + +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tRD Bk:%d Row:%03x Col:%03x \n", + j, bk[j]->curr_row, + bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); +#endif + // transfer done + if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { + bk[j]->mrq = NULL; + } + } else + // correct row activated for a WRITE + if ( !issued && !CCDc && !bk[j]->RCDWRc && + !(bkgrp[grp]->CCDLc) && + (bk[j]->curr_row == bk[j]->mrq->row) && + (bk[j]->mrq->rw == WRITE) && (RTWc == 0 ) && + (bk[j]->state == BANK_ACTIVE) && + !rwq->full() ) { + if (rw==READ) { + rw=WRITE; + rwq->set_min_length(m_config->WL); + } + rwq->push(bk[j]->mrq); + + bk[j]->mrq->txbytes += m_config->dram_atom_size; + CCDc = m_config->tCCD; + bkgrp[grp]->CCDLc = m_config->tCCDL; + WTRc = m_config->tWTR; + bk[j]->WTPc = m_config->tWTP; + issued = true; + + if(bk[j]->mrq->data->get_access_type() == L2_WRBK_ACC) + n_wr_WB++; + else + n_wr++; + bwutil += m_config->BL/m_config->data_command_freq_ratio; + bwutil_partial += m_config->BL/m_config->data_command_freq_ratio; +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tWR Bk:%d Row:%03x Col:%03x \n", + j, bk[j]->curr_row, + bk[j]->mrq->col + bk[j]->mrq->txbytes - m_config->dram_atom_size); +#endif + // transfer done + if ( !(bk[j]->mrq->txbytes < bk[j]->mrq->nbytes) ) { + bk[j]->mrq = NULL; + } + } + + } + + return issued; +} + +bool dram_t::issue_row_command(int j) +{ + bool issued = false; + unsigned grp = get_bankgrp_number(j); + if (bk[j]->mrq) { //if currently servicing a memory request + bk[j]->mrq->data->set_status(IN_PARTITION_DRAM,gpu_sim_cycle+gpu_tot_sim_cycle); + // bank is idle + //else + if ( !issued && !RRDc && + (bk[j]->state == BANK_IDLE) && + !bk[j]->RPc && !bk[j]->RCc) { // +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tACT BK:%d NewRow:%03x From:%03x \n", + j,bk[j]->mrq->row,bk[j]->curr_row); +#endif + // activate the row with current memory request + bk[j]->curr_row = bk[j]->mrq->row; + bk[j]->state = BANK_ACTIVE; + RRDc = m_config->tRRD; + bk[j]->RCDc = m_config->tRCD; + bk[j]->RCDWRc = m_config->tRCDWR; + bk[j]->RASc = m_config->tRAS; + bk[j]->RCc = m_config->tRC; + prio = (j + 1) % m_config->nbk; + issued = true; + n_act_partial++; + n_act++; + } + + else + // different row activated + if ( (!issued) && + (bk[j]->curr_row != bk[j]->mrq->row) && + (bk[j]->state == BANK_ACTIVE) && + (!bk[j]->RASc && !bk[j]->WTPc && + !bk[j]->RTPc && + !bkgrp[grp]->RTPLc) ) { + // make the bank idle again + bk[j]->state = BANK_IDLE; + bk[j]->RPc = m_config->tRP; + prio = (j + 1) % m_config->nbk; + issued = true; + n_pre++; + n_pre_partial++; +#ifdef DRAM_VERIFY + PRINT_CYCLE=1; + printf("\tPRE BK:%d Row:%03x \n", j,bk[j]->curr_row); +#endif + } + } + return issued; +} + + //if mrq is being serviced by dram, gets popped after CL latency fulfilled -class mem_fetch* dram_t::return_queue_pop() +class mem_fetch* dram_t::return_queue_pop() { return returnq->pop(); } -class mem_fetch* dram_t::return_queue_top() +class mem_fetch* dram_t::return_queue_top() { return returnq->top(); } + void dram_t::print( FILE* simFile) const { unsigned i; fprintf(simFile,"DRAM[%d]: %d bks, busW=%d BL=%d CL=%d, ", id, m_config->nbk, m_config->busW, m_config->BL, m_config->CL ); fprintf(simFile,"tRRD=%d tCCD=%d, tRCD=%d tRAS=%d tRP=%d tRC=%d\n", - m_config->tCCD, m_config->tRRD, m_config->tRCD, m_config->tRAS, m_config->tRP, m_config->tRC ); - fprintf(simFile,"n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_req=%d n_rd=%d n_write=%d bw_util=%.4g\n", - n_cmd, n_nop, n_act, n_pre, n_req, n_rd, n_wr, + m_config->tRRD, m_config->tCCD, m_config->tRCD, m_config->tRAS, m_config->tRP, m_config->tRC ); + fprintf(simFile,"n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_ref_event=%d n_req=%d n_rd=%d n_rd_L2_A=%d n_write=%d n_wr_bk=%d bw_util=%.4g\n", + n_cmd, n_nop, n_act, n_pre, n_ref, n_req, n_rd, n_rd_L2_A, n_wr, n_wr_WB, (float)bwutil/n_cmd); fprintf(simFile,"n_activity=%d dram_eff=%.4g\n", n_activity, (float)bwutil/n_activity); for (i=0;inbk;i++) { fprintf(simFile, "bk%d: %da %di ",i,bk[i]->n_access,bk[i]->n_idle); } + fprintf(simFile, "\n"); + fprintf(simFile, "\n------------------------------------------------------------------------\n"); + + printf("\nRow_Buffer_Locality = %.6f", (float)hits_num / access_num); + printf("\nRow_Buffer_Locality_read = %.6f", (float)hits_read_num / read_num); + printf("\nRow_Buffer_Locality_write = %.6f", (float)hits_write_num / write_num); + printf("\nBank_Level_Parallism = %.6f", (float)banks_1time / banks_acess_total); + printf("\nBank_Level_Parallism_Col = %.6f", (float)banks_time_rw / banks_access_rw_total); + printf("\nBank_Level_Parallism_Ready = %.6f", (float)banks_time_ready /banks_access_ready_total); + printf("\nwrite_to_read_ratio_blp_rw_average = %.6f", write_to_read_ratio_blp_rw_average /banks_access_rw_total); + printf("\nGrpLevelPara = %.6f \n", (float)bkgrp_parallsim_rw /banks_access_rw_total); + + printf("\nBW Util details:\n"); + printf("bwutil = %.6f \n", (float)bwutil/n_cmd); + printf("total_CMD = %d \n", n_cmd); + printf("util_bw = %d \n", util_bw); + printf("Wasted_Col = %d \n", wasted_bw_col); + printf("Wasted_Row = %d \n", wasted_bw_row); + printf("Idle = %d \n", idle_bw); + + printf("\nBW Util Bottlenecks: \n"); + printf("RCDc_limit = %d \n", RCDc_limit); + printf("RCDWRc_limit = %d \n", RCDWRc_limit); + printf("WTRc_limit = %d \n", WTRc_limit); + printf("RTWc_limit = %d \n", RTWc_limit); + printf("CCDLc_limit = %d \n", CCDLc_limit); + printf("rwq = %d \n", rwq_limit); + printf("CCDLc_limit_alone = %d \n", CCDLc_limit_alone); + printf("WTRc_limit_alone = %d \n", WTRc_limit_alone); + printf("RTWc_limit_alone = %d \n", RTWc_limit_alone); + + printf("\nCommands details: \n"); + printf("total_CMD = %d \n", n_cmd); + printf("n_nop = %d \n", n_nop); + printf("Read = %d \n", n_rd); + printf("Write = %d \n",n_wr); + printf("L2_Alloc = %d \n", n_rd_L2_A); + printf("L2_WB = %d \n", n_wr_WB); + printf("n_act = %d \n", n_act); + printf("n_pre = %d \n", n_pre); + printf("n_ref = %d \n", n_ref); + printf("n_req = %d \n", n_req ); + printf("total_req = %d \n", n_rd+n_wr+n_rd_L2_A+n_wr_WB); + + printf("\nDual Bus Interface Util: \n"); + printf("issued_total_row = %lu \n", issued_total_row); + printf("issued_total_col = %lu \n", issued_total_col); + printf("Row_Bus_Util = %.6f \n", (float)issued_total_row / n_cmd); + printf("CoL_Bus_Util = %.6f \n", (float)issued_total_col / n_cmd); + printf("Either_Row_CoL_Bus_Util = %.6f \n", (float)issued_total / n_cmd); + printf("Issued_on_Two_Bus_Simul_Util = %.6f \n", (float)issued_two /n_cmd); + printf("issued_two_Eff = %.6f \n", (float)issued_two /issued_total); + printf("queue_avg = %.6f \n\n", (float)ave_mrqs/n_cmd ); + fprintf(simFile, "\n"); fprintf(simFile, "dram_util_bins:"); for (i=0;i<10;i++) fprintf(simFile, " %d", dram_util_bins[i]); fprintf(simFile, "\ndram_eff_bins:"); for (i=0;i<10;i++) fprintf(simFile, " %d", dram_eff_bins[i]); fprintf(simFile, "\n"); - if(m_config->scheduler_type== DRAM_FRFCFS) + if(m_config->scheduler_type== DRAM_FRFCFS) fprintf(simFile, "mrqq: max=%d avg=%g\n", max_mrqs, (float)ave_mrqs/n_cmd); } @@ -476,8 +815,8 @@ void dram_t::visualize() const void dram_t::print_stat( FILE* simFile ) { - fprintf(simFile,"DRAM (%d): n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_req=%d n_rd=%d n_write=%d bw_util=%.4g ", - id, n_cmd, n_nop, n_act, n_pre, n_req, n_rd, n_wr, + fprintf(simFile,"DRAM (%d): n_cmd=%d n_nop=%d n_act=%d n_pre=%d n_ref=%d n_req=%d n_rd=%d n_write=%d bw_util=%.4g ", + id, n_cmd, n_nop, n_act, n_pre, n_ref, n_req, n_rd, n_wr, (float)bwutil/n_cmd); fprintf(simFile, "mrqq: %d %.4g mrqsmax=%d ", max_mrqs, (float)ave_mrqs/n_cmd, max_mrqs_temp); fprintf(simFile, "\n"); @@ -516,6 +855,7 @@ void dram_t::visualizer_print( gzFile visualizer_file ) n_pre_partial = 0; n_req_partial = 0; + // dram access type classification for (unsigned j = 0; j < m_config->nbk; j++) { gzprintf(visualizer_file,"dramglobal_acc_r: %u %u %u\n", id, j, @@ -553,3 +893,16 @@ void dram_t::set_dram_power_stats( unsigned &cmd, wr = n_wr; req = n_req; } + +unsigned dram_t::get_bankgrp_number(unsigned i) +{ + if(m_config->dram_bnkgrp_indexing_policy == HIGHER_BITS) { //higher bits + return i>>m_config->bk_tag_length; + } + else if (m_config->dram_bnkgrp_indexing_policy == LOWER_BITS) { //lower bits + return i&((m_config->nbkgrp-1)); + } + else { + assert(1); + } +} diff --git a/src/gpgpu-sim/dram.h b/src/gpgpu-sim/dram.h index 15c63e7..bee5b7b 100644 --- a/src/gpgpu-sim/dram.h +++ b/src/gpgpu-sim/dram.h @@ -31,9 +31,15 @@ #include "delayqueue.h" #include +#include +#include +#include +#include +#include #include #include #include +#include #define READ 'R' //define read and write states #define WRITE 'W' @@ -42,7 +48,7 @@ class dram_req_t { public: - dram_req_t( class mem_fetch *data ); + dram_req_t( class mem_fetch *data , unsigned banks, unsigned dram_bnk_indexing_policy); unsigned int row; unsigned int col; @@ -87,6 +93,17 @@ struct bank_t unsigned int bkgrpindex; }; +enum bank_index_function{ + LINEAR_BK_INDEX = 0, + BITWISE_XORING_BK_INDEX, + CUSTOM_BK_INDEX +}; + +enum bank_grp_bits_position{ + HIGHER_BITS = 0, + LOWER_BITS +}; + class mem_fetch; class dram_t @@ -95,7 +112,7 @@ public: dram_t( unsigned int parition_id, const struct memory_config *config, class memory_stats_t *stats, class memory_partition_unit *mp ); - bool full() const; + bool full(bool is_write) const; void print( FILE* simFile ) const; void visualize() const; void print_stat( FILE* simFile ); @@ -106,6 +123,7 @@ public: class mem_fetch* return_queue_pop(); class mem_fetch* return_queue_top(); + void push( class mem_fetch *data ); void cycle(); void dram_log (int task); @@ -123,17 +141,24 @@ public: unsigned &wr, unsigned &req) const; -private: - void scheduler_fifo(); - void scheduler_frfcfs(); + const struct memory_config *m_config; +private: bankgrp_t **bkgrp; bank_t **bk; unsigned int prio; + unsigned get_bankgrp_number(unsigned i); + + void scheduler_fifo(); + void scheduler_frfcfs(); + + bool issue_col_command(int j); + bool issue_row_command(int j); + unsigned int RRDc; unsigned int CCDc; unsigned int RTWc; //read to write penalty applies across banks @@ -146,7 +171,7 @@ private: fifo_pipeline *rwq; fifo_pipeline *mrqq; //buffer to hold packets when DRAM processing is over - //should be filled with dram clock and popped with l2or icnt clock + //should be filled with dram clock and popped with l2or icnt clock fifo_pipeline *returnq; unsigned int dram_util_bins[10]; @@ -158,11 +183,51 @@ private: unsigned int n_nop; unsigned int n_act; unsigned int n_pre; + unsigned int n_ref; unsigned int n_rd; + unsigned int n_rd_L2_A; unsigned int n_wr; + unsigned int n_wr_WB; unsigned int n_req; unsigned int max_mrqs_temp; + //some statistics to collect to see where BW is wasted? + unsigned wasted_bw_row; + unsigned wasted_bw_col; + unsigned util_bw; + unsigned idle_bw; + unsigned RCDc_limit; + unsigned CCDLc_limit; + unsigned CCDLc_limit_alone; + unsigned CCDc_limit; + unsigned WTRc_limit; + unsigned WTRc_limit_alone; + unsigned RCDWRc_limit; + unsigned RTWc_limit; + unsigned RTWc_limit_alone; + unsigned rwq_limit; + + //row locality, BLP and other statistics + unsigned long access_num; + unsigned long read_num; + unsigned long write_num; + unsigned long long hits_num; + unsigned long long hits_read_num; + unsigned long long hits_write_num; + unsigned long long banks_1time; + unsigned long long banks_acess_total; + unsigned long long banks_acess_total_after; + unsigned long long banks_time_rw; + unsigned long long banks_access_rw_total; + unsigned long long banks_time_ready; + unsigned long long banks_access_ready_total; + unsigned long long issued_two; + unsigned long long issued_total; + unsigned long long issued_total_row; + unsigned long long issued_total_col; + double write_to_read_ratio_blp_rw_average; + unsigned long long bkgrp_parallsim_rw; + unsigned int bwutil; unsigned int max_mrqs; unsigned int ave_mrqs; diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc index 8303e86..ff50050 100644 --- a/src/gpgpu-sim/dram_sched.cc +++ b/src/gpgpu-sim/dram_sched.cc @@ -36,6 +36,7 @@ frfcfs_scheduler::frfcfs_scheduler( const memory_config *config, dram_t *dm, mem m_config = config; m_stats = stats; m_num_pending = 0; + m_num_write_pending = 0; m_dram = dm; m_queue = new std::list[m_config->nbk]; m_bins = new std::map::iterator> >[ m_config->nbk ]; @@ -49,15 +50,36 @@ frfcfs_scheduler::frfcfs_scheduler( const memory_config *config, dram_t *dm, mem curr_row_service_time[i] = 0; row_service_timestamp[i] = 0; } + if(m_config->seperate_write_queue_enabled) { + m_write_queue = new std::list[m_config->nbk]; + m_write_bins = new std::map::iterator> >[ m_config->nbk ]; + m_last_write_row = new std::list::iterator>*[ m_config->nbk ]; + + for ( unsigned i=0; i < m_config->nbk; i++ ) { + m_write_queue[i].clear(); + m_write_bins[i].clear(); + m_last_write_row[i] = NULL; + } + } + m_mode = READ_MODE; } void frfcfs_scheduler::add_req( dram_req_t *req ) { - m_num_pending++; - m_queue[req->bk].push_front(req); - std::list::iterator ptr = m_queue[req->bk].begin(); - m_bins[req->bk][req->row].push_front( ptr ); //newest reqs to the front + if(m_config->seperate_write_queue_enabled && req->data->is_write()) { + assert(m_num_write_pending < m_config->gpgpu_frfcfs_dram_write_queue_size); + m_num_write_pending++; + m_write_queue[req->bk].push_front(req); + std::list::iterator ptr = m_write_queue[req->bk].begin(); + m_write_bins[req->bk][req->row].push_front( ptr ); //newest reqs to the front + } else { + assert(m_num_pending < m_config->gpgpu_frfcfs_dram_sched_queue_size); + m_num_pending++; + m_queue[req->bk].push_front(req); + std::list::iterator ptr = m_queue[req->bk].begin(); + m_bins[req->bk][req->row].push_front( ptr ); //newest reqs to the front + } } void frfcfs_scheduler::data_collection(unsigned int bank) @@ -78,41 +100,92 @@ void frfcfs_scheduler::data_collection(unsigned int bank) dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row ) { - if ( m_last_row[bank] == NULL ) { - if ( m_queue[bank].empty() ) + //row + bool rowhit = true; + std::list *m_current_queue = m_queue; + std::map::iterator> > *m_current_bins = m_bins ; + std::list::iterator> **m_current_last_row = m_last_row; + + if(m_config->seperate_write_queue_enabled) { + if(m_mode == READ_MODE && + ((m_num_write_pending >= m_config->write_high_watermark ) + // || (m_queue[bank].empty() && !m_write_queue[bank].empty()) + )) { + m_mode = WRITE_MODE; + } + else if(m_mode == WRITE_MODE && + (( m_num_write_pending < m_config->write_low_watermark ) + // || (!m_queue[bank].empty() && m_write_queue[bank].empty()) + )){ + m_mode = READ_MODE; + } + } + + if(m_mode == WRITE_MODE) { + m_current_queue = m_write_queue; + m_current_bins = m_write_bins ; + m_current_last_row = m_last_write_row; + } + + if ( m_current_last_row[bank] == NULL ) { + if ( m_current_queue[bank].empty() ) return NULL; - std::map::iterator> >::iterator bin_ptr = m_bins[bank].find( curr_row ); - if ( bin_ptr == m_bins[bank].end()) { - dram_req_t *req = m_queue[bank].back(); - bin_ptr = m_bins[bank].find( req->row ); - assert( bin_ptr != m_bins[bank].end() ); // where did the request go??? - m_last_row[bank] = &(bin_ptr->second); + std::map::iterator> >::iterator bin_ptr = m_current_bins[bank].find( curr_row ); + if ( bin_ptr == m_current_bins[bank].end()) { + dram_req_t *req = m_current_queue[bank].back(); + bin_ptr = m_current_bins[bank].find( req->row ); + assert( bin_ptr != m_current_bins[bank].end() ); // where did the request go??? + m_current_last_row[bank] = &(bin_ptr->second); data_collection(bank); + rowhit = false; } else { - m_last_row[bank] = &(bin_ptr->second); - + m_current_last_row[bank] = &(bin_ptr->second); + rowhit = true; } } - std::list::iterator next = m_last_row[bank]->back(); + std::list::iterator next = m_current_last_row[bank]->back(); dram_req_t *req = (*next); + //rowblp stats + m_dram->access_num++; + bool is_write = req->data->is_write(); + if(is_write) + m_dram->write_num++; + else + m_dram->read_num++; + + if(rowhit) { + m_dram->hits_num++; + if(is_write) + m_dram->hits_write_num++; + else + m_dram->hits_read_num++; + } + m_stats->concurrent_row_access[m_dram->id][bank]++; m_stats->row_access[m_dram->id][bank]++; - m_last_row[bank]->pop_back(); + m_current_last_row[bank]->pop_back(); - m_queue[bank].erase(next); - if ( m_last_row[bank]->empty() ) { - m_bins[bank].erase( req->row ); - m_last_row[bank] = NULL; + m_current_queue[bank].erase(next); + if ( m_current_last_row[bank]->empty() ) { + m_current_bins[bank].erase( req->row ); + m_current_last_row[bank] = NULL; } #ifdef DEBUG_FAST_IDEAL_SCHED if ( req ) printf("%08u : DRAM(%u) scheduling memory request to bank=%u, row=%u\n", (unsigned)gpu_sim_cycle, m_dram->id, req->bk, req->row ); #endif - assert( req != NULL && m_num_pending != 0 ); - m_num_pending--; + + if(m_config->seperate_write_queue_enabled && req->data->is_write()) { + assert( req != NULL && m_num_write_pending != 0 ); + m_num_write_pending--; + } + else { + assert( req != NULL && m_num_pending != 0 ); + m_num_pending--; + } return req; } @@ -129,7 +202,7 @@ void dram_t::scheduler_frfcfs() { unsigned mrq_latency; frfcfs_scheduler *sched = m_frfcfs_scheduler; - while ( !mrqq->empty() && (!m_config->gpgpu_frfcfs_dram_sched_queue_size || sched->num_pending() < m_config->gpgpu_frfcfs_dram_sched_queue_size)) { + while ( !mrqq->empty() ) { dram_req_t *req = mrqq->pop(); // Power stats @@ -160,6 +233,8 @@ void dram_t::scheduler_frfcfs() bk[b]->mrq = req; if (m_config->gpgpu_memlatency_stat) { mrq_latency = gpu_sim_cycle + gpu_tot_sim_cycle - bk[b]->mrq->timestamp; + m_stats->tot_mrq_latency += mrq_latency; + m_stats->tot_mrq_num++; bk[b]->mrq->timestamp = gpu_tot_sim_cycle + gpu_sim_cycle; m_stats->mrq_lat_table[LOGB2(mrq_latency)]++; if (mrq_latency > m_stats->max_mrq_latency) { diff --git a/src/gpgpu-sim/dram_sched.h b/src/gpgpu-sim/dram_sched.h index 3860f5b..63f5831 100644 --- a/src/gpgpu-sim/dram_sched.h +++ b/src/gpgpu-sim/dram_sched.h @@ -35,6 +35,11 @@ #include #include +enum memory_mode { + READ_MODE = 0, + WRITE_MODE +}; + class frfcfs_scheduler { public: frfcfs_scheduler( const memory_config *config, dram_t *dm, memory_stats_t *stats ); @@ -43,17 +48,24 @@ public: dram_req_t *schedule( unsigned bank, unsigned curr_row ); void print( FILE *fp ); unsigned num_pending() const { return m_num_pending;} + unsigned num_write_pending() const { return m_num_write_pending;} private: const memory_config *m_config; dram_t *m_dram; unsigned m_num_pending; + unsigned m_num_write_pending; std::list *m_queue; std::map::iterator> > *m_bins; std::list::iterator> **m_last_row; unsigned *curr_row_service_time; //one set of variables for each bank. unsigned *row_service_timestamp; //tracks when scheduler began servicing current row + std::list *m_write_queue; + std::map::iterator> > *m_write_bins; + std::list::iterator> **m_last_write_row; + + enum memory_mode m_mode; memory_stats_t *m_stats; }; diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc index 8886398..ba81440 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) @@ -38,7 +37,8 @@ const char * cache_request_status_str(enum cache_request_status status) "HIT", "HIT_RESERVED", "MISS", - "RESERVATION_FAIL" + "RESERVATION_FAIL", + "SECTOR_MISS" }; assert(sizeof(static_cache_request_status_str) / sizeof(const char*) == NUM_CACHE_REQUEST_STATUS); @@ -47,6 +47,22 @@ const char * cache_request_status_str(enum cache_request_status status) return static_cache_request_status_str[status]; } +const char * cache_fail_status_str(enum cache_reservation_fail_reason status) +{ + static const char * static_cache_reservation_fail_reason_str[] = { + "LINE_ALLOC_FAIL", + "MISS_QUEUE_FULL", + "MSHR_ENRTY_FAIL", + "MSHR_MERGE_ENRTY_FAIL", + "MSHR_RW_PENDING" + }; + + assert(sizeof(static_cache_reservation_fail_reason_str) / sizeof(const char*) == NUM_CACHE_RESERVATION_FAIL_STATUS); + assert(status < NUM_CACHE_RESERVATION_FAIL_STATUS); + + return static_cache_reservation_fail_reason_str[status]; +} + unsigned l1d_cache_config::set_index(new_addr_type addr) const{ unsigned set_index = m_nset; // Default to linear set index function unsigned lower_xor = 0; @@ -54,10 +70,11 @@ 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. - * ISCA 2014 + * HPCA 2014 */ if(m_nset == 32 || m_nset == 64){ // Lower xor value is bits 7-11 @@ -80,6 +97,36 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{ } break; + case HASH_IPOLY_FUNCTION: + /* + * Set Indexing function from "Pseudo-randomly interleaved memory." + * Rau, B. R et al. + * ISCA 1991 + * + * "Sacat: streaming-aware conflict-avoiding thrashing-resistant gpgpu cache management scheme." + * Khairy et al. + * IEEE TPDS 2017. + */ + if(m_nset == 32 || m_nset == 64){ + std::bitset<64> a(addr); + std::bitset<6> index; + index[0] = a[25]^a[24]^a[23]^a[22]^a[21]^a[18]^a[17]^a[15]^a[12]^a[7]; //10 + index[1] = a[26]^a[25]^a[24]^a[23]^a[22]^a[19]^a[18]^a[16]^a[13]^a[8]; //10 + index[2] = a[26]^a[22]^a[21]^a[20]^a[19]^a[18]^a[15]^a[14]^a[12]^a[9]; //10 + index[3] = a[23]^a[22]^a[21]^a[20]^a[19]^a[16]^a[15]^a[13]^a[10]; //9 + index[4] = a[24]^a[23]^a[22]^a[21]^a[20]^a[17]^a[16]^a[14]^a[11]; //9 + + if(m_nset == 64) + index[5] = a[12]; + + set_index = index.to_ulong(); + + }else{ /* Else incorrect number of sets for the hashing function */ + assert("\nGPGPU-Sim cache configuration error: The number of sets should be " + "32 or 64 for the hashing set index function.\n" && 0); + } + break; + case CUSTOM_SET_FUNCTION: /* No custom set function implemented */ break; @@ -87,6 +134,10 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{ case LINEAR_SET_FUNCTION: set_index = (addr >> m_line_sz_log2) & (m_nset-1); break; + + default: + assert("\nUndefined set index function.\n" && 0); + break; } // Linear function selected or custom set index function not implemented @@ -113,13 +164,16 @@ unsigned l2_cache_config::set_index(new_addr_type addr) const{ tag_array::~tag_array() { + unsigned cache_lines_num = m_config.get_max_num_lines(); + for(unsigned i=0; iget_addr()); + line_table::const_iterator i = pending_lines.find(addr); + if ( i == pending_lines.end() ) { + pending_lines[addr] = mf->get_inst().get_uid(); + } +} + +void tag_array::remove_pending_line(mem_fetch *mf){ + assert(mf); + new_addr_type addr = m_config.block_addr(mf->get_addr()); + line_table::const_iterator i = pending_lines.find(addr); + if ( i != pending_lines.end() ) { + pending_lines.erase(addr); + } } -enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx ) const { +enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx, mem_fetch* mf, bool probe_mode) const { + mem_access_sector_mask_t mask = mf->get_access_sector_mask(); + return probe(addr, idx, mask, probe_mode, mf); +} + + +enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx, mem_access_sector_mask_t mask, bool probe_mode, mem_fetch* mf) const { //assert( m_config.m_write_policy == READ_ONLY ); unsigned set_index = m_config.set_index(addr); new_addr_type tag = m_config.tag(addr); @@ -169,35 +263,45 @@ enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx ) // check for hit or pending hit for (unsigned way=0; waym_tag == tag) { - if ( line->m_status == RESERVED ) { + if ( line->get_status(mask) == RESERVED ) { idx = index; return HIT_RESERVED; - } else if ( line->m_status == VALID ) { + } else if ( line->get_status(mask) == VALID ) { idx = index; return HIT; - } else if ( line->m_status == MODIFIED ) { + } else if ( line->get_status(mask) == MODIFIED) { + if(line->is_readable(mask)) { + idx = index; + return HIT; + } + else { + idx = index; + return SECTOR_MISS; + } + + } else if ( line->is_valid_line() && line->get_status(mask) == INVALID ) { idx = index; - return HIT; - } else { - assert( line->m_status == INVALID ); + return SECTOR_MISS; + }else { + assert( line->get_status(mask) == INVALID ); } } - if (line->m_status != RESERVED) { + if (!line->is_reserved_line()) { all_reserved = false; - if (line->m_status == INVALID) { + if (line->is_invalid_line()) { invalid_line = index; } else { // valid line : keep track of most appropriate replacement candidate if ( m_config.m_replacement_policy == LRU ) { - if ( line->m_last_access_time < valid_timestamp ) { - valid_timestamp = line->m_last_access_time; + if ( line->get_last_access_time() < valid_timestamp ) { + valid_timestamp = line->get_last_access_time(); valid_line = index; } } else if ( m_config.m_replacement_policy == FIFO ) { - if ( line->m_alloc_time < valid_timestamp ) { - valid_timestamp = line->m_alloc_time; + if ( line->get_alloc_time() < valid_timestamp ) { + valid_timestamp = line->get_alloc_time(); valid_line = index; } } @@ -215,40 +319,59 @@ enum cache_request_status tag_array::probe( new_addr_type addr, unsigned &idx ) idx = valid_line; } else abort(); // if an unreserved block exists, it is either invalid or replaceable + + if(probe_mode && m_config.is_streaming()){ + line_table::const_iterator i = pending_lines.find(m_config.block_addr(addr)); + assert(mf); + if ( !mf->is_write() && i != pending_lines.end() ) { + if(i->second != mf->get_inst().get_uid()) + return SECTOR_MISS; + } + } + return MISS; } -enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, unsigned &idx ) +enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, unsigned &idx, mem_fetch* mf) { bool wb=false; - cache_block_t evicted; - enum cache_request_status result = access(addr,time,idx,wb,evicted); + evicted_block_info evicted; + enum cache_request_status result = access(addr,time,idx,wb,evicted,mf); assert(!wb); return result; } -enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, unsigned &idx, bool &wb, cache_block_t &evicted ) +enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, unsigned &idx, bool &wb, evicted_block_info &evicted, mem_fetch* mf ) { m_access++; + is_used = true; shader_cache_access_log(m_core_id, m_type_id, 0); // log accesses to cache - enum cache_request_status status = probe(addr,idx); + enum cache_request_status status = probe(addr,idx,mf); switch (status) { case HIT_RESERVED: m_pending_hit++; case HIT: - m_lines[idx].m_last_access_time=time; + m_lines[idx]->set_last_access_time(time, mf->get_access_sector_mask()); break; case MISS: m_miss++; shader_cache_access_log(m_core_id, m_type_id, 1); // log cache misses if ( m_config.m_alloc_policy == ON_MISS ) { - if( m_lines[idx].m_status == MODIFIED ) { + if( m_lines[idx]->is_modified_line()) { wb = true; - evicted = m_lines[idx]; + evicted.set_info(m_lines[idx]->m_block_addr, m_lines[idx]->get_modified_size()); } - m_lines[idx].allocate( m_config.tag(addr), m_config.block_addr(addr), time ); + m_lines[idx]->allocate( m_config.tag(addr), m_config.block_addr(addr), time, mf->get_access_sector_mask()); } break; + case SECTOR_MISS: + assert(m_config.m_cache_type == SECTOR); + m_sector_miss++; + shader_cache_access_log(m_core_id, m_type_id, 1); // log cache misses + if ( m_config.m_alloc_policy == ON_MISS ) { + ((sector_cache_block*)m_lines[idx])->allocate_sector( time, mf->get_access_sector_mask() ); + } + break; case RESERVATION_FAIL: m_res_fail++; shader_cache_access_log(m_core_id, m_type_id, 1); // log cache misses @@ -261,37 +384,70 @@ enum cache_request_status tag_array::access( new_addr_type addr, unsigned time, return status; } -void tag_array::fill( new_addr_type addr, unsigned time ) +void tag_array::fill( new_addr_type addr, unsigned time, mem_fetch* mf) +{ + fill(addr, time, mf->get_access_sector_mask()); +} + +void tag_array::fill( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ) { - assert( m_config.m_alloc_policy == ON_FILL ); + //assert( m_config.m_alloc_policy == ON_FILL ); unsigned idx; - enum cache_request_status status = probe(addr,idx); - assert(status==MISS); // MSHR should have prevented redundant memory request - m_lines[idx].allocate( m_config.tag(addr), m_config.block_addr(addr), time ); - m_lines[idx].fill(time); + enum cache_request_status status = probe(addr,idx,mask); + //assert(status==MISS||status==SECTOR_MISS); // MSHR should have prevented redundant memory request + if(status==MISS) + m_lines[idx]->allocate( m_config.tag(addr), m_config.block_addr(addr), time, mask ); + else if (status==SECTOR_MISS) { + assert(m_config.m_cache_type == SECTOR); + ((sector_cache_block*)m_lines[idx])->allocate_sector( time, mask ); + } + + m_lines[idx]->fill(time, mask); } -void tag_array::fill( unsigned index, unsigned time ) +void tag_array::fill( unsigned index, unsigned time, mem_fetch* mf) { assert( m_config.m_alloc_policy == ON_MISS ); - m_lines[index].fill(time); + m_lines[index]->fill(time, mf->get_access_sector_mask()); } + +//TODO: we need write back the flushed data to the upper level void tag_array::flush() { + if(!is_used) + return; + for (unsigned i=0; i < m_config.get_num_lines(); i++) - m_lines[i].m_status = INVALID; + if(m_lines[i]->is_modified_line()) { + for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) + m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; + } + + is_used = false; +} + +void tag_array::invalidate() +{ + if(!is_used) + return; + + for (unsigned i=0; i < m_config.get_num_lines(); i++) + for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++) + m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ; + + is_used = false; } float tag_array::windowed_miss_rate( ) const { unsigned n_access = m_access - m_prev_snapshot_access; - unsigned n_miss = m_miss - m_prev_snapshot_miss; + unsigned n_miss = (m_miss+m_sector_miss) - m_prev_snapshot_miss; // unsigned n_pending_hit = m_pending_hit - m_prev_snapshot_pending_hit; float missrate = 0.0f; if (n_access != 0) - missrate = (float) n_miss / n_access; + missrate = (float) (n_miss+m_sector_miss) / n_access; return missrate; } @@ -299,23 +455,24 @@ void tag_array::new_window() { m_prev_snapshot_access = m_access; m_prev_snapshot_miss = m_miss; + m_prev_snapshot_miss = m_miss + m_sector_miss; m_prev_snapshot_pending_hit = m_pending_hit; } void tag_array::print( FILE *stream, unsigned &total_access, unsigned &total_misses ) const { m_config.print(stream); - fprintf( stream, "\t\tAccess = %d, Miss = %d (%.3g), PendingHit = %d (%.3g)\n", - m_access, m_miss, (float) m_miss / m_access, + fprintf( stream, "\t\tAccess = %d, Miss = %d, Sector_Miss = %d, Total_Miss = %d (%.3g), PendingHit = %d (%.3g)\n", + m_access, m_miss, m_sector_miss, (m_miss+m_sector_miss), (float) (m_miss+m_sector_miss) / m_access, m_pending_hit, (float) m_pending_hit / m_access); - total_misses+=m_miss; + total_misses+=(m_miss+m_sector_miss); total_access+=m_access; } void tag_array::get_stats(unsigned &total_access, unsigned &total_misses, unsigned &total_hit_res, unsigned &total_res_fail) const{ // Update statistics from the tag array total_access = m_access; - total_misses = m_miss; + total_misses = (m_miss+m_sector_miss); total_hit_res = m_pending_hit; total_res_fail = m_res_fail; } @@ -324,16 +481,17 @@ void tag_array::get_stats(unsigned &total_access, unsigned &total_misses, unsign bool was_write_sent( const std::list &events ) { for( std::list::const_iterator e=events.begin(); e!=events.end(); e++ ) { - if( *e == WRITE_REQUEST_SENT ) + if( (*e).m_cache_event_type == WRITE_REQUEST_SENT ) return true; } return false; } -bool was_writeback_sent( const std::list &events ) +bool was_writeback_sent( const std::list &events, cache_event& wb_event) { for( std::list::const_iterator e=events.begin(); e!=events.end(); e++ ) { - if( *e == WRITE_BACK_REQUEST_SENT ) + if( (*e).m_cache_event_type == WRITE_BACK_REQUEST_SENT ) + wb_event = *e; return true; } return false; @@ -342,7 +500,16 @@ bool was_writeback_sent( const std::list &events ) bool was_read_sent( const std::list &events ) { for( std::list::const_iterator e=events.begin(); e!=events.end(); e++ ) { - if( *e == READ_REQUEST_SENT ) + if( (*e).m_cache_event_type == READ_REQUEST_SENT ) + return true; + } + return false; +} + +bool was_writeallocate_sent( const std::list &events ) +{ + for( std::list::const_iterator e=events.begin(); e!=events.end(); e++ ) { + if( (*e).m_cache_event_type == WRITE_ALLOCATE_SENT ) return true; } return false; @@ -375,11 +542,27 @@ void mshr_table::add( new_addr_type block_addr, mem_fetch *mf ){ } } +/// check is_read_after_write_pending +bool mshr_table::is_read_after_write_pending( new_addr_type block_addr){ + std::list my_list = m_data[block_addr].m_list; + bool write_found = false; + for (std::list::iterator it=my_list.begin(); it != my_list.end(); ++it) + { + if((*it)->is_write()) //Pending Write Request + write_found = true; + else if(write_found) //Pending Read Request and we found previous Write + return true; + } + + return false; + +} + /// Accept a new cache fill response: mark entry ready for processing void mshr_table::mark_ready( new_addr_type block_addr, bool &has_atomic ){ assert( !busy() ); table::iterator a = m_data.find(block_addr); - assert( a != m_data.end() ); // don't remove same request twice + assert( a != m_data.end() ); m_current_response.push_back( block_addr ); has_atomic = a->second.m_has_atomic; assert( m_current_response.size() <= m_data.size() ); @@ -417,9 +600,11 @@ void mshr_table::display( FILE *fp ) const{ /***************************************************************** Caches *****************************************************************/ cache_stats::cache_stats(){ m_stats.resize(NUM_MEM_ACCESS_TYPE); + m_fail_stats.resize(NUM_MEM_ACCESS_TYPE); for(unsigned i=0; i 0) - fprintf(fout, "\t%s[%s][%s] = %u\n", - m_cache_name.c_str(), - mem_access_type_str((enum mem_access_type)type), - "TOTAL_ACCESS", - total_access[type]); + if(total_access[type] > 0) + fprintf(fout, "\t%s[%s][%s] = %u\n", + m_cache_name.c_str(), + mem_access_type_str((enum mem_access_type)type), + "TOTAL_ACCESS", + total_access[type]); } } +void cache_stats::print_fail_stats(FILE *fout, const char *cache_name) const{ + std::string m_cache_name = cache_name; + for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { + for (unsigned fail = 0; fail < NUM_CACHE_RESERVATION_FAIL_STATUS; ++fail) { + if(m_fail_stats[type][fail] > 0){ + fprintf(fout, "\t%s[%s][%s] = %u\n", + m_cache_name.c_str(), + mem_access_type_str((enum mem_access_type)type), + cache_fail_status_str((enum cache_reservation_fail_reason)fail), + m_fail_stats[type][fail]); + } + } + } +} + void cache_sub_stats::print_port_stats(FILE *fout, const char *cache_name) const { float data_port_util = 0.0f; @@ -580,10 +813,10 @@ void cache_stats::get_sub_stats(struct cache_sub_stats &css) const{ for (unsigned type = 0; type < NUM_MEM_ACCESS_TYPE; ++type) { for (unsigned status = 0; status < NUM_CACHE_REQUEST_STATUS; ++status) { - if(status == HIT || status == MISS || status == HIT_RESERVED) + if(status == HIT || status == MISS || status == SECTOR_MISS || status == HIT_RESERVED) t_css.accesses += m_stats[type][status]; - if(status == MISS) + if(status == MISS || status == SECTOR_MISS) t_css.misses += m_stats[type][status]; if(status == HIT_RESERVED) @@ -611,6 +844,16 @@ bool cache_stats::check_valid(int type, int status) const{ return false; } +bool cache_stats::check_fail_valid(int type, int fail) const{ + /// + /// Verify a valid access_type/access_status + /// + if((type >= 0) && (type < NUM_MEM_ACCESS_TYPE) && (fail >= 0) && (fail < NUM_CACHE_RESERVATION_FAIL_STATUS)) + return true; + else + return false; +} + void cache_stats::sample_cache_port_utility(bool data_port_busy, bool fill_port_busy) { m_cache_port_available_cycles += 1; @@ -639,15 +882,18 @@ void baseline_cache::bandwidth_management::use_data_port(mem_fetch *mf, enum cac unsigned data_cycles = data_size / port_width + ((data_size % port_width > 0)? 1 : 0); m_data_port_occupied_cycles += data_cycles; } break; - case HIT_RESERVED: + case HIT_RESERVED: case MISS: { // the data array is accessed to read out the entire line for write-back - if (was_writeback_sent(events)) { - unsigned data_cycles = m_config.m_line_sz / port_width; + // in case of sector cache we need to write bank only the modified sectors + cache_event ev(WRITE_BACK_REQUEST_SENT); + if (was_writeback_sent(events, ev)) { + unsigned data_cycles = ev.m_evicted_block.m_modified_size / port_width; m_data_port_occupied_cycles += data_cycles; } } break; - case RESERVATION_FAIL: + case SECTOR_MISS: + case RESERVATION_FAIL: // Does not consume any port bandwidth break; default: @@ -660,7 +906,7 @@ void baseline_cache::bandwidth_management::use_data_port(mem_fetch *mf, enum cac void baseline_cache::bandwidth_management::use_fill_port(mem_fetch *mf) { // assume filling the entire line with the returned request - unsigned fill_cycles = m_config.m_line_sz / m_config.m_data_port_width; + unsigned fill_cycles = m_config.get_atom_sz() / m_config.m_data_port_width; m_fill_port_occupied_cycles += fill_cycles; } @@ -707,21 +953,43 @@ void baseline_cache::cycle(){ /// Interface for response from lower memory level (model bandwidth restictions in caller) void baseline_cache::fill(mem_fetch *mf, unsigned time){ + + if(m_config.m_mshr_type == SECTOR_ASSOC) { + 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 ); mf->set_data_size( e->second.m_data_size ); + mf->set_addr( e->second.m_addr ); if ( m_config.m_alloc_policy == ON_MISS ) - m_tag_array->fill(e->second.m_cache_index,time); - else if ( m_config.m_alloc_policy == ON_FILL ) - m_tag_array->fill(e->second.m_block_addr,time); + m_tag_array->fill(e->second.m_cache_index,time,mf); + else if ( m_config.m_alloc_policy == ON_FILL ) { + m_tag_array->fill(e->second.m_block_addr,time,mf); + if(m_config.is_streaming()) + m_tag_array->remove_pending_line(mf); + } else abort(); bool has_atomic = false; m_mshrs.mark_ready(e->second.m_block_addr, has_atomic); if (has_atomic) { assert(m_config.m_alloc_policy == ON_MISS); - cache_block_t &block = m_tag_array->get_block(e->second.m_cache_index); - block.m_status = MODIFIED; // mark line as dirty for atomic operation + cache_block_t* block = m_tag_array->get_block(e->second.m_cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); // mark line as dirty for atomic operation } m_extra_mf_fields.erase(mf); m_bandwidth_management.use_fill_port(mf); @@ -749,45 +1017,59 @@ void baseline_cache::send_read_request(new_addr_type addr, new_addr_type block_a unsigned time, bool &do_miss, std::list &events, bool read_only, bool wa){ bool wb=false; - cache_block_t e; + evicted_block_info e; send_read_request(addr, block_addr, cache_index, mf, time, do_miss, wb, e, events, read_only, wa); } /// Read miss handler. Check MSHR hit or MSHR available void baseline_cache::send_read_request(new_addr_type addr, new_addr_type block_addr, unsigned cache_index, mem_fetch *mf, - unsigned time, bool &do_miss, bool &wb, cache_block_t &evicted, std::list &events, bool read_only, bool wa){ + unsigned time, bool &do_miss, bool &wb, evicted_block_info &evicted, std::list &events, bool read_only, bool wa){ - bool mshr_hit = m_mshrs.probe(block_addr); - bool mshr_avail = !m_mshrs.full(block_addr); + new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); + bool mshr_hit = m_mshrs.probe(mshr_addr); + bool mshr_avail = !m_mshrs.full(mshr_addr); if ( mshr_hit && mshr_avail ) { if(read_only) - m_tag_array->access(block_addr,time,cache_index); + m_tag_array->access(block_addr,time,cache_index,mf); else - m_tag_array->access(block_addr,time,cache_index,wb,evicted); + m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); - m_mshrs.add(block_addr,mf); + m_mshrs.add(mshr_addr,mf); do_miss = true; + } else if ( !mshr_hit && mshr_avail && (m_miss_queue.size() < m_config.m_miss_queue_size) ) { if(read_only) - m_tag_array->access(block_addr,time,cache_index); + m_tag_array->access(block_addr,time,cache_index,mf); else - m_tag_array->access(block_addr,time,cache_index,wb,evicted); - - m_mshrs.add(block_addr,mf); - m_extra_mf_fields[mf] = extra_mf_fields(block_addr,cache_index, mf->get_data_size()); - mf->set_data_size( m_config.get_line_sz() ); + m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); + + m_mshrs.add(mshr_addr,mf); + if(m_config.is_streaming() && m_config.m_cache_type == SECTOR){ + m_tag_array->add_pending_line(mf); + } + m_extra_mf_fields[mf] = extra_mf_fields(mshr_addr,mf->get_addr(),cache_index, mf->get_data_size(), m_config); + mf->set_data_size( m_config.get_atom_sz() ); + mf->set_addr( mshr_addr ); m_miss_queue.push_back(mf); mf->set_status(m_miss_queue_status,time); if(!wa) - events.push_back(READ_REQUEST_SENT); + events.push_back(cache_event(READ_REQUEST_SENT)); + do_miss = true; } + else if(mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL); + else if (!mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL); + else + assert(0); } /// Sends write request to lower level memory (write or writeback) void data_cache::send_write_request(mem_fetch *mf, cache_event request, unsigned time, std::list &events){ - events.push_back(request); + + events.push_back(request); m_miss_queue.push_back(mf); mf->set_status(m_miss_queue_status,time); } @@ -798,40 +1080,44 @@ void data_cache::send_write_request(mem_fetch *mf, cache_event request, unsigned /// Write-back hit: Mark block as modified cache_request_status data_cache::wr_hit_wb(new_addr_type addr, unsigned cache_index, mem_fetch *mf, unsigned time, std::list &events, enum cache_request_status status ){ new_addr_type block_addr = m_config.block_addr(addr); - m_tag_array->access(block_addr,time,cache_index); // update LRU state - cache_block_t &block = m_tag_array->get_block(cache_index); - block.m_status = MODIFIED; + m_tag_array->access(block_addr,time,cache_index,mf); // update LRU state + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); return HIT; } /// Write-through hit: Directly send request to lower level memory cache_request_status data_cache::wr_hit_wt(new_addr_type addr, unsigned cache_index, mem_fetch *mf, unsigned time, std::list &events, enum cache_request_status status ){ - if(miss_queue_full(0)) + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); return RESERVATION_FAIL; // cannot handle request this cycle + } new_addr_type block_addr = m_config.block_addr(addr); - m_tag_array->access(block_addr,time,cache_index); // update LRU state - cache_block_t &block = m_tag_array->get_block(cache_index); - block.m_status = MODIFIED; + m_tag_array->access(block_addr,time,cache_index,mf); // update LRU state + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); // generate a write-through - send_write_request(mf, WRITE_REQUEST_SENT, time, events); + send_write_request(mf, cache_event(WRITE_REQUEST_SENT), time, events); return HIT; } /// Write-evict hit: Send request to lower level memory and invalidate corresponding block cache_request_status data_cache::wr_hit_we(new_addr_type addr, unsigned cache_index, mem_fetch *mf, unsigned time, std::list &events, enum cache_request_status status ){ - if(miss_queue_full(0)) + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); return RESERVATION_FAIL; // cannot handle request this cycle + } // generate a write-through/evict - cache_block_t &block = m_tag_array->get_block(cache_index); - send_write_request(mf, WRITE_REQUEST_SENT, time, events); + cache_block_t* block = m_tag_array->get_block(cache_index); + send_write_request(mf, cache_event(WRITE_REQUEST_SENT), time, events); // Invalidate block - block.m_status = INVALID; + block->set_status(INVALID, mf->get_access_sector_mask()); return HIT; } @@ -850,34 +1136,46 @@ enum cache_request_status data_cache::wr_hit_global_we_local_wb(new_addr_type ad /// Write-allocate miss: Send write request to lower level memory // and send a read request for the same block enum cache_request_status -data_cache::wr_miss_wa( new_addr_type addr, +data_cache::wr_miss_wa_naive( new_addr_type addr, unsigned cache_index, mem_fetch *mf, unsigned time, std::list &events, enum cache_request_status status ) { new_addr_type block_addr = m_config.block_addr(addr); + new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); // Write allocate, maximum 3 requests (write miss, read request, write back request) // Conservatively ensure the worst-case request can be handled this cycle - bool mshr_hit = m_mshrs.probe(block_addr); - bool mshr_avail = !m_mshrs.full(block_addr); + bool mshr_hit = m_mshrs.probe(mshr_addr); + bool mshr_avail = !m_mshrs.full(mshr_addr); if(miss_queue_full(2) || (!(mshr_hit && mshr_avail) - && !(!mshr_hit && mshr_avail - && (m_miss_queue.size() < m_config.m_miss_queue_size)))) + && !(!mshr_hit && mshr_avail && (m_miss_queue.size() < m_config.m_miss_queue_size)))) { + //check what is the exactly the failure reason + if(miss_queue_full(2) ) + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + else if(mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL); + else if (!mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL); + else + assert(0); + return RESERVATION_FAIL; + } - send_write_request(mf, WRITE_REQUEST_SENT, time, events); + send_write_request(mf, cache_event(WRITE_REQUEST_SENT), time, events); // Tries to send write allocate request, returns true on success and false on failure //if(!send_write_allocate(mf, addr, block_addr, cache_index, time, events)) // return RESERVATION_FAIL; const mem_access_t *ma = new mem_access_t( m_wr_alloc_type, mf->get_addr(), - mf->get_data_size(), + m_config.get_atom_sz(), false, // Now performing a read mf->get_access_warp_mask(), - mf->get_access_byte_mask() ); + mf->get_access_byte_mask(), + mf->get_access_sector_mask()); mem_fetch *n_mf = new mem_fetch( *ma, NULL, @@ -889,20 +1187,22 @@ data_cache::wr_miss_wa( new_addr_type addr, bool do_miss = false; bool wb = false; - cache_block_t evicted; + evicted_block_info evicted; // Send read request resulting from write miss send_read_request(addr, block_addr, cache_index, n_mf, time, do_miss, wb, evicted, events, false, true); + events.push_back(cache_event(WRITE_ALLOCATE_SENT)); + if( do_miss ){ // If evicted block is modified and not a write-through // (already modified lower level) if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) { + assert(status == MISS); //SECTOR_MISS and HIT_RESERVED should not send write back mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, - m_wrbk_type,m_config.get_line_sz(),true); - m_miss_queue.push_back(wb); - wb->set_status(m_miss_queue_status,time); + m_wrbk_type,evicted.m_modified_size,true); + send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); } return MISS; } @@ -910,6 +1210,177 @@ data_cache::wr_miss_wa( new_addr_type addr, return RESERVATION_FAIL; } + +enum cache_request_status +data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr, + unsigned cache_index, mem_fetch *mf, + unsigned time, std::list &events, + enum cache_request_status status ) +{ + new_addr_type block_addr = m_config.block_addr(addr); + new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); + + if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) + { + //if the request writes to the whole cache line/sector, then, write and set cache line Modified. + //and no need to send read request to memory or reserve mshr + + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + return RESERVATION_FAIL; // cannot handle request this cycle + } + + bool wb = false; + evicted_block_info evicted; + + cache_request_status status = m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); + assert(status != HIT); + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); + if(status == HIT_RESERVED) + block->set_ignore_on_fill(true, mf->get_access_sector_mask()); + + if( status != RESERVATION_FAIL ){ + // If evicted block is modified and not a write-through + // (already modified lower level) + if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) { + mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, + m_wrbk_type,evicted.m_modified_size,true); + send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); + } + return MISS; + } + return RESERVATION_FAIL; + } + else + { + bool mshr_hit = m_mshrs.probe(mshr_addr); + bool mshr_avail = !m_mshrs.full(mshr_addr); + if(miss_queue_full(1) + || (!(mshr_hit && mshr_avail) + && !(!mshr_hit && mshr_avail && (m_miss_queue.size() < m_config.m_miss_queue_size)))) { + //check what is the exactly the failure reason + if(miss_queue_full(1) ) + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + else if(mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_MERGE_ENRTY_FAIL); + else if (!mshr_hit && !mshr_avail) + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_ENRTY_FAIL); + else + assert(0); + + return RESERVATION_FAIL; + } + + + //prevent Write - Read - Write in pending mshr + //allowing another write will override the value of the first write, and the pending read request will read incorrect result from the second write + if(m_mshrs.probe(mshr_addr) && m_mshrs.is_read_after_write_pending(mshr_addr) && mf->is_write()) + { + //assert(0); + m_stats.inc_fail_stats(mf->get_access_type(), MSHR_RW_PENDING); + return RESERVATION_FAIL; + } + + const mem_access_t *ma = new mem_access_t( m_wr_alloc_type, + mf->get_addr(), + m_config.get_atom_sz(), + false, // Now performing a read + mf->get_access_warp_mask(), + mf->get_access_byte_mask(), + mf->get_access_sector_mask()); + + mem_fetch *n_mf = new mem_fetch( *ma, + NULL, + mf->get_ctrl_size(), + mf->get_wid(), + mf->get_sid(), + mf->get_tpc(), + mf->get_mem_config(), + NULL, + mf); + + + new_addr_type block_addr = m_config.block_addr(addr); + bool do_miss = false; + bool wb = false; + evicted_block_info evicted; + send_read_request( addr, + block_addr, + cache_index, + n_mf, time, do_miss, wb, evicted, events, false, true); + + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_modified_on_fill(true, mf->get_access_sector_mask()); + + events.push_back(cache_event(WRITE_ALLOCATE_SENT)); + + if( do_miss ){ + // If evicted block is modified and not a write-through + // (already modified lower level) + if(wb && (m_config.m_write_policy != WRITE_THROUGH) ){ + mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, + m_wrbk_type,evicted.m_modified_size,true); + send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); + } + return MISS; + } + return RESERVATION_FAIL; + } +} + +enum cache_request_status +data_cache::wr_miss_wa_lazy_fetch_on_read( new_addr_type addr, + unsigned cache_index, mem_fetch *mf, + unsigned time, std::list &events, + enum cache_request_status status ) +{ + + new_addr_type block_addr = m_config.block_addr(addr); + new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr()); + + + //if the request writes to the whole cache line/sector, then, write and set cache line Modified. + //and no need to send read request to memory or reserve mshr + + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + return RESERVATION_FAIL; // cannot handle request this cycle + } + + bool wb = false; + evicted_block_info evicted; + + cache_request_status m_status = m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf); + assert(m_status != HIT); + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()); + if(m_status == HIT_RESERVED) { + block->set_ignore_on_fill(true, mf->get_access_sector_mask()); + block->set_modified_on_fill(true, mf->get_access_sector_mask()); + } + + if(mf->get_access_byte_mask().count() == m_config.get_atom_sz()) + { + block->set_m_readable(true, mf->get_access_sector_mask()); + } else + { + block->set_m_readable(false, mf->get_access_sector_mask()); + } + + if( m_status != RESERVATION_FAIL ){ + // If evicted block is modified and not a write-through + // (already modified lower level) + if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) { + mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, + m_wrbk_type,evicted.m_modified_size,true); + send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events); + } + return MISS; + } + return RESERVATION_FAIL; +} + /// No write-allocate miss: Simply send write request to lower level memory enum cache_request_status data_cache::wr_miss_no_wa( new_addr_type addr, @@ -919,11 +1390,14 @@ data_cache::wr_miss_no_wa( new_addr_type addr, std::list &events, enum cache_request_status status ) { - if(miss_queue_full(0)) - return RESERVATION_FAIL; // cannot handle request this cycle + if(miss_queue_full(0)) { + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); + return RESERVATION_FAIL; // cannot handle request this cycle + } + // on miss, generate write through (no write buffering -- too many threads for that) - send_write_request(mf, WRITE_REQUEST_SENT, time, events); + send_write_request(mf, cache_event(WRITE_REQUEST_SENT), time, events); return MISS; } @@ -941,13 +1415,13 @@ data_cache::rd_hit_base( new_addr_type addr, enum cache_request_status status ) { new_addr_type block_addr = m_config.block_addr(addr); - m_tag_array->access(block_addr,time,cache_index); + m_tag_array->access(block_addr,time,cache_index,mf); // Atomics treated as global read/write requests - Perform read, mark line as // MODIFIED if(mf->isatomic()){ assert(mf->get_access_type() == GLOBAL_ACC_R); - cache_block_t &block = m_tag_array->get_block(cache_index); - block.m_status = MODIFIED; // mark line as dirty + cache_block_t* block = m_tag_array->get_block(cache_index); + block->set_status(MODIFIED, mf->get_access_sector_mask()) ; // mark line as dirty } return HIT; } @@ -963,15 +1437,17 @@ data_cache::rd_miss_base( new_addr_type addr, unsigned time, std::list &events, enum cache_request_status status ){ - if(miss_queue_full(1)) + if(miss_queue_full(1)) { // cannot handle request this cycle // (might need to generate two requests) + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); return RESERVATION_FAIL; + } new_addr_type block_addr = m_config.block_addr(addr); bool do_miss = false; bool wb = false; - cache_block_t evicted; + evicted_block_info evicted; send_read_request( addr, block_addr, cache_index, @@ -982,12 +1458,12 @@ data_cache::rd_miss_base( new_addr_type addr, // (already modified lower level) if(wb && (m_config.m_write_policy != WRITE_THROUGH) ){ mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr, - m_wrbk_type,m_config.get_line_sz(),true); + m_wrbk_type,evicted.m_modified_size,true); send_write_request(wb, WRITE_BACK_REQUEST_SENT, time, events); } return MISS; } - return RESERVATION_FAIL; + return RESERVATION_FAIL; } /// Access cache for read_only_cache: returns RESERVATION_FAIL if @@ -998,16 +1474,16 @@ read_only_cache::access( new_addr_type addr, unsigned time, std::list &events ) { - assert( mf->get_data_size() <= m_config.get_line_sz()); + assert( mf->get_data_size() <= m_config.get_atom_sz()); assert(m_config.m_write_policy == READ_ONLY); assert(!mf->get_is_write()); new_addr_type block_addr = m_config.block_addr(addr); unsigned cache_index = (unsigned)-1; - enum cache_request_status status = m_tag_array->probe(block_addr,cache_index); + enum cache_request_status status = m_tag_array->probe(block_addr,cache_index,mf); enum cache_request_status cache_status = RESERVATION_FAIL; if ( status == HIT ) { - cache_status = m_tag_array->access(block_addr,time,cache_index); // update LRU state + cache_status = m_tag_array->access(block_addr,time,cache_index,mf); // update LRU state }else if ( status != RESERVATION_FAIL ) { if(!miss_queue_full(0)){ bool do_miss=false; @@ -1018,7 +1494,10 @@ read_only_cache::access( new_addr_type addr, cache_status = RESERVATION_FAIL; }else{ cache_status = RESERVATION_FAIL; + m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL); } + }else { + m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL); } m_stats.inc_stats(mf->get_access_type(), m_stats.select_stats_status(status, cache_status)); @@ -1047,10 +1526,13 @@ data_cache::process_tag_probe( bool wr, access_status = (this->*m_wr_hit)( addr, cache_index, mf, time, events, probe_status ); - }else if ( probe_status != RESERVATION_FAIL ) { + }else if ( (probe_status != RESERVATION_FAIL) || (probe_status == RESERVATION_FAIL && m_config.m_write_alloc_policy == NO_WRITE_ALLOCATE) ) { access_status = (this->*m_wr_miss)( addr, cache_index, mf, time, events, probe_status ); + }else { + //the only reason for reservation fail here is LINE_ALLOC_FAIL (i.e all lines are reserved) + m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL); } }else{ // Read if(probe_status == HIT){ @@ -1061,6 +1543,9 @@ data_cache::process_tag_probe( bool wr, access_status = (this->*m_rd_miss)( addr, cache_index, mf, time, events, probe_status ); + }else { + //the only reason for reservation fail here is LINE_ALLOC_FAIL (i.e all lines are reserved) + m_stats.inc_fail_stats(mf->get_access_type(), LINE_ALLOC_FAIL); } } @@ -1080,12 +1565,12 @@ data_cache::access( new_addr_type addr, std::list &events ) { - assert( mf->get_data_size() <= m_config.get_line_sz()); + assert( mf->get_data_size() <= m_config.get_atom_sz()); bool wr = mf->get_is_write(); new_addr_type block_addr = m_config.block_addr(addr); unsigned cache_index = (unsigned)-1; enum cache_request_status probe_status - = m_tag_array->probe( block_addr, cache_index ); + = m_tag_array->probe( block_addr, cache_index, mf, true); enum cache_request_status access_status = process_tag_probe( wr, probe_status, addr, cache_index, mf, time, events ); m_stats.inc_stats(mf->get_access_type(), @@ -1134,7 +1619,7 @@ enum cache_request_status tex_cache::access( new_addr_type addr, mem_fetch *mf, // at this point, we will accept the request : access tags and immediately allocate line new_addr_type block_addr = m_config.block_addr(addr); unsigned cache_index = (unsigned)-1; - enum cache_request_status status = m_tags.access(block_addr,time,cache_index); + enum cache_request_status status = m_tags.access(block_addr,time,cache_index,mf); enum cache_request_status cache_status = RESERVATION_FAIL; assert( status != RESERVATION_FAIL ); assert( status != HIT_RESERVED ); // as far as tags are concerned: HIT or MISS @@ -1142,12 +1627,12 @@ 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); // mark block as valid + m_tags.fill(cache_index,time,mf); // mark block as valid m_request_fifo.push(mf); mf->set_status(m_request_queue_status,time); - events.push_back(READ_REQUEST_SENT); + events.push_back(cache_event(READ_REQUEST_SENT)); cache_status = MISS; } else { // the value *will* *be* in the cache already @@ -1174,7 +1659,7 @@ void tex_cache::cycle(){ unsigned rob_index = m_rob.next_pop_index(); const rob_entry &r = m_rob.peek(rob_index); assert( r.m_request == e.m_request ); - assert( r.m_block_addr == m_config.block_addr(e.m_request->get_addr()) ); + //assert( r.m_block_addr == m_config.block_addr(e.m_request->get_addr()) ); if ( r.m_ready ) { assert( r.m_index == e.m_cache_index ); m_cache[r.m_index].m_valid = true; @@ -1197,6 +1682,23 @@ 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 7535a1d..e663cf6 100644 --- a/src/gpgpu-sim/gpu-cache.h +++ b/src/gpgpu-sim/gpu-cache.h @@ -36,9 +36,12 @@ #include "../tr1_hash_map.h" #include "addrdec.h" +#include + +#define MAX_DEFAULT_CACHE_SIZE_MULTIBLIER 4 enum cache_block_state { - INVALID, + INVALID=0, RESERVED, VALID, MODIFIED @@ -49,13 +52,51 @@ enum cache_request_status { HIT_RESERVED, MISS, RESERVATION_FAIL, + SECTOR_MISS, NUM_CACHE_REQUEST_STATUS }; -enum cache_event { +enum cache_reservation_fail_reason { + LINE_ALLOC_FAIL= 0,// all line are reserved + MISS_QUEUE_FULL, // MISS queue (i.e. interconnect or DRAM) is full + MSHR_ENRTY_FAIL, + MSHR_MERGE_ENRTY_FAIL, + MSHR_RW_PENDING, + NUM_CACHE_RESERVATION_FAIL_STATUS +}; + +enum cache_event_type { WRITE_BACK_REQUEST_SENT, READ_REQUEST_SENT, - WRITE_REQUEST_SENT + WRITE_REQUEST_SENT, + WRITE_ALLOCATE_SENT +}; + +struct evicted_block_info { + new_addr_type m_block_addr; + unsigned m_modified_size; + evicted_block_info() { + m_block_addr = 0; + m_modified_size = 0; + } + void set_info(new_addr_type block_addr, unsigned modified_size){ + m_block_addr = block_addr; + m_modified_size = modified_size; + } +}; + +struct cache_event { + enum cache_event_type m_cache_event_type; + evicted_block_info m_evicted_block; //if it was write_back event, fill the the evicted block info + + cache_event(enum cache_event_type m_cache_event){ + m_cache_event_type = m_cache_event; + } + + cache_event(enum cache_event_type cache_event, evicted_block_info evicted_block){ + m_cache_event_type = cache_event; + m_evicted_block = evicted_block; + } }; const char * cache_request_status_str(enum cache_request_status status); @@ -65,33 +106,340 @@ struct cache_block_t { { m_tag=0; m_block_addr=0; - m_alloc_time=0; - m_fill_time=0; - m_last_access_time=0; - m_status=INVALID; } - void allocate( new_addr_type tag, new_addr_type block_addr, unsigned time ) + + virtual void allocate( new_addr_type tag, new_addr_type block_addr, unsigned time, mem_access_sector_mask_t sector_mask) = 0; + virtual void fill( unsigned time, mem_access_sector_mask_t sector_mask) = 0; + + virtual bool is_invalid_line() = 0; + virtual bool is_valid_line() = 0; + virtual bool is_reserved_line() = 0; + virtual bool is_modified_line() = 0; + + virtual enum cache_block_state get_status( mem_access_sector_mask_t sector_mask) = 0; + virtual void set_status(enum cache_block_state m_status, mem_access_sector_mask_t sector_mask) = 0; + + virtual unsigned get_last_access_time() = 0; + virtual void set_last_access_time(unsigned time, mem_access_sector_mask_t sector_mask) = 0; + virtual unsigned get_alloc_time() = 0; + virtual void set_ignore_on_fill(bool m_ignore, mem_access_sector_mask_t sector_mask) = 0; + virtual void set_modified_on_fill(bool m_modified, mem_access_sector_mask_t sector_mask) = 0; + virtual unsigned get_modified_size() = 0; + virtual void set_m_readable(bool readable, mem_access_sector_mask_t sector_mask)=0; + virtual bool is_readable(mem_access_sector_mask_t sector_mask)=0; + virtual void print_status()=0; + virtual ~cache_block_t() {} + + + new_addr_type m_tag; + new_addr_type m_block_addr; + +}; + +struct line_cache_block: public cache_block_t { + line_cache_block() + { + m_alloc_time=0; + m_fill_time=0; + m_last_access_time=0; + m_status=INVALID; + m_ignore_on_fill_status = false; + m_set_modified_on_fill = false; + m_readable = true; + } + void allocate( new_addr_type tag, new_addr_type block_addr, unsigned time, mem_access_sector_mask_t sector_mask) + { + m_tag=tag; + m_block_addr=block_addr; + m_alloc_time=time; + m_last_access_time=time; + m_fill_time=0; + m_status=RESERVED; + m_ignore_on_fill_status = false; + m_set_modified_on_fill = false; + } + void fill( unsigned time, mem_access_sector_mask_t sector_mask ) + { + //if(!m_ignore_on_fill_status) + // assert( m_status == RESERVED ); + + m_status = m_set_modified_on_fill? MODIFIED : VALID; + + m_fill_time=time; + } + virtual bool is_invalid_line() + { + return m_status == INVALID; + } + virtual bool is_valid_line() + { + return m_status == VALID; + } + virtual bool is_reserved_line() + { + return m_status == RESERVED; + } + virtual bool is_modified_line() + { + return m_status == MODIFIED; + } + + virtual enum cache_block_state get_status(mem_access_sector_mask_t sector_mask) + { + return m_status; + } + virtual void set_status(enum cache_block_state status, mem_access_sector_mask_t sector_mask) + { + m_status = status; + } + virtual unsigned get_last_access_time() + { + return m_last_access_time; + } + virtual void set_last_access_time(unsigned time, mem_access_sector_mask_t sector_mask) + { + m_last_access_time = time; + } + virtual unsigned get_alloc_time() + { + return m_alloc_time; + } + virtual void set_ignore_on_fill(bool m_ignore, mem_access_sector_mask_t sector_mask) + { + m_ignore_on_fill_status = m_ignore; + } + virtual void set_modified_on_fill(bool m_modified, mem_access_sector_mask_t sector_mask) + { + m_set_modified_on_fill = m_modified; + } + virtual unsigned get_modified_size() + { + return SECTOR_CHUNCK_SIZE * SECTOR_SIZE; //i.e. cache line size + } + virtual void set_m_readable(bool readable, mem_access_sector_mask_t sector_mask) + { + m_readable = readable; + } + virtual bool is_readable(mem_access_sector_mask_t sector_mask) { + return m_readable; + } + virtual void print_status() { + printf("m_block_addr is %llu, status = %u\n", m_block_addr, m_status); + } + + +private: + unsigned m_alloc_time; + unsigned m_last_access_time; + unsigned m_fill_time; + cache_block_state m_status; + bool m_ignore_on_fill_status; + bool m_set_modified_on_fill; + bool m_readable; +}; + +struct sector_cache_block : public cache_block_t { + sector_cache_block() { - m_tag=tag; - m_block_addr=block_addr; - m_alloc_time=time; - m_last_access_time=time; - m_fill_time=0; - m_status=RESERVED; + init(); } - void fill( unsigned time ) + + void init() { + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + m_sector_alloc_time[i]= 0; + m_sector_fill_time[i]= 0; + m_last_sector_access_time[i]= 0; + m_status[i]= INVALID; + m_ignore_on_fill_status[i] = false; + m_set_modified_on_fill[i] = false; + m_readable[i] = true; + } + m_line_alloc_time=0; + m_line_last_access_time=0; + m_line_fill_time=0; + } + + virtual void allocate( new_addr_type tag, new_addr_type block_addr, unsigned time, mem_access_sector_mask_t sector_mask ) { - assert( m_status == RESERVED ); - m_status=VALID; - m_fill_time=time; + allocate_line( tag, block_addr, time, sector_mask ); } - new_addr_type m_tag; - new_addr_type m_block_addr; - unsigned m_alloc_time; - unsigned m_last_access_time; - unsigned m_fill_time; - cache_block_state m_status; + void allocate_line( new_addr_type tag, new_addr_type block_addr, unsigned time, mem_access_sector_mask_t sector_mask ) + { + //allocate a new line + //assert(m_block_addr != 0 && m_block_addr != block_addr); + init(); + m_tag=tag; + m_block_addr=block_addr; + + unsigned sidx = get_sector_index(sector_mask); + + //set sector stats + m_sector_alloc_time[sidx]=time; + m_last_sector_access_time[sidx]=time; + m_sector_fill_time[sidx]=0; + m_status[sidx]=RESERVED; + m_ignore_on_fill_status[sidx] = false; + m_set_modified_on_fill[sidx] = false; + + //set line stats + m_line_alloc_time=time; //only set this for the first allocated sector + m_line_last_access_time=time; + m_line_fill_time=0; + } + + void allocate_sector(unsigned time, mem_access_sector_mask_t sector_mask ) + { + //allocate invalid sector of this allocated valid line + assert(is_valid_line()); + unsigned sidx = get_sector_index(sector_mask); + + //set sector stats + m_sector_alloc_time[sidx]=time; + m_last_sector_access_time[sidx]=time; + m_sector_fill_time[sidx]=0; + if(m_status[sidx]==MODIFIED) //this should be the case only for fetch-on-write policy //TO DO + m_set_modified_on_fill[sidx] = true; + else + m_set_modified_on_fill[sidx] = false; + + m_status[sidx]=RESERVED; + m_ignore_on_fill_status[sidx] = false; + //m_set_modified_on_fill[sidx] = false; + m_readable[sidx] = true; + + //set line stats + m_line_last_access_time=time; + m_line_fill_time=0; + } + + virtual void fill( unsigned time, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + + // if(!m_ignore_on_fill_status[sidx]) + // assert( m_status[sidx] == RESERVED ); + + m_status[sidx] = m_set_modified_on_fill[sidx]? MODIFIED : VALID; + + m_sector_fill_time[sidx]=time; + m_line_fill_time=time; + } + virtual bool is_invalid_line() { + //all the sectors should be invalid + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + if (m_status[i] != INVALID) + return false; + } + return true; + } + virtual bool is_valid_line() { return !(is_invalid_line()); } + virtual bool is_reserved_line() { + //if any of the sector is reserved, then the line is reserved + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + if (m_status[i] == RESERVED) + return true; + } + return false; + } + virtual bool is_modified_line() { + //if any of the sector is modified, then the line is modified + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + if (m_status[i] == MODIFIED) + return true; + } + return false; + } + + virtual enum cache_block_state get_status(mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + + return m_status[sidx]; + } + + virtual void set_status(enum cache_block_state status, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + m_status[sidx] = status; + } + + virtual unsigned get_last_access_time() + { + return m_line_last_access_time; + } + + virtual void set_last_access_time(unsigned time, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + + m_last_sector_access_time[sidx] = time; + m_line_last_access_time = time; + } + + virtual unsigned get_alloc_time() + { + return m_line_alloc_time; + } + + virtual void set_ignore_on_fill(bool m_ignore, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + m_ignore_on_fill_status[sidx] = m_ignore; + } + + virtual void set_modified_on_fill(bool m_modified, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + m_set_modified_on_fill[sidx] = m_modified; + } + + virtual void set_m_readable(bool readable, mem_access_sector_mask_t sector_mask) + { + unsigned sidx = get_sector_index(sector_mask); + m_readable[sidx] = readable; + } + + virtual bool is_readable(mem_access_sector_mask_t sector_mask) { + unsigned sidx = get_sector_index(sector_mask); + return m_readable[sidx]; + } + + virtual unsigned get_modified_size() + { + unsigned modified=0; + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + if (m_status[i] == MODIFIED) + modified++; + } + return modified * SECTOR_SIZE; + } + + virtual void print_status() { + printf("m_block_addr is %llu, status = %u %u %u %u\n", m_block_addr, m_status[0], m_status[1], m_status[2], m_status[3]); + } + + +private: + unsigned m_sector_alloc_time[SECTOR_CHUNCK_SIZE]; + unsigned m_last_sector_access_time[SECTOR_CHUNCK_SIZE]; + unsigned m_sector_fill_time[SECTOR_CHUNCK_SIZE]; + unsigned m_line_alloc_time; + unsigned m_line_last_access_time; + unsigned m_line_fill_time; + cache_block_state m_status[SECTOR_CHUNCK_SIZE]; + bool m_ignore_on_fill_status[SECTOR_CHUNCK_SIZE]; + bool m_set_modified_on_fill[SECTOR_CHUNCK_SIZE]; + bool m_readable[SECTOR_CHUNCK_SIZE]; + + unsigned get_sector_index(mem_access_sector_mask_t sector_mask) + { + assert(sector_mask.count() == 1); + for(unsigned i =0; i< SECTOR_CHUNCK_SIZE; ++i) { + if(sector_mask.to_ulong() & (1< line_table; + line_table pending_lines; }; class mshr_table { public: - mshr_table( unsigned num_entries, unsigned max_merged ) + mshr_table( unsigned num_entries, unsigned max_merged) : m_num_entries(num_entries), m_max_merged(max_merged) #if (tr1_hash_map_ismap == 0) @@ -414,6 +879,8 @@ public: /// Returns next ready access mem_fetch *next_access(); void display( FILE *fp ) const; + // Returns true if there is a pending read after write + bool is_read_after_write_pending(new_addr_type block_addr); void check_mshr_parameters( unsigned num_entries, unsigned max_merged ) { @@ -433,7 +900,9 @@ private: mshr_entry() : m_has_atomic(false) { } }; typedef tr1_hash_map table; + typedef tr1_hash_map line_table; table m_data; + line_table pending_lines; // it may take several cycles to process the merged requests bool m_current_response_ready; @@ -510,12 +979,14 @@ public: cache_stats(); void clear(); void inc_stats(int access_type, int access_outcome); + void inc_fail_stats(int access_type, int fail_outcome); enum cache_request_status select_stats_status(enum cache_request_status probe, enum cache_request_status access) const; - unsigned &operator()(int access_type, int access_outcome); - unsigned operator()(int access_type, int access_outcome) const; + unsigned &operator()(int access_type, int access_outcome, bool fail_outcome); + unsigned operator()(int access_type, int access_outcome, bool fail_outcome) const; cache_stats operator+(const cache_stats &cs); cache_stats &operator+=(const cache_stats &cs); void print_stats(FILE *fout, const char *cache_name = "Cache_stats") const; + void print_fail_stats(FILE *fout, const char *cache_name = "Cache_fail_stats") const; unsigned get_stats(enum mem_access_type *access_type, unsigned num_access_type, enum cache_request_status *access_status, unsigned num_access_status) const; void get_sub_stats(struct cache_sub_stats &css) const; @@ -523,8 +994,10 @@ public: void sample_cache_port_utility(bool data_port_busy, bool fill_port_busy); private: bool check_valid(int type, int status) const; + bool check_fail_valid(int type, int fail) const; std::vector< std::vector > m_stats; + std::vector< std::vector > m_fail_stats; unsigned long long m_cache_port_available_cycles; unsigned long long m_cache_data_port_busy_cycles; @@ -543,6 +1016,7 @@ public: bool was_write_sent( const std::list &events ); bool was_read_sent( const std::list &events ); +bool was_writeallocate_sent( const std::list &events ); /// Baseline cache /// Implements common functions for read_only_cache and data_cache @@ -552,7 +1026,7 @@ public: baseline_cache( const char *name, cache_config &config, int core_id, int type_id, mem_fetch_interface *memport, enum mem_fetch_status status ) : m_config(config), m_tag_array(new tag_array(config,core_id,type_id)), - m_mshrs(config.m_mshr_entries,config.m_mshr_max_merge), + m_mshrs(config.m_mshr_entries,config.m_mshr_max_merge), m_bandwidth_management(config) { init( name, config, memport, status ); @@ -564,7 +1038,7 @@ public: enum mem_fetch_status status ) { m_name = name; - assert(config.m_mshr_type == ASSOC); + assert(config.m_mshr_type == ASSOC || config.m_mshr_type == SECTOR_ASSOC); m_memport=memport; m_miss_queue_status = status; } @@ -594,6 +1068,7 @@ public: mem_fetch *next_access(){return m_mshrs.next_access();} // flash invalidate all entries in cache void flush(){m_tag_array->flush();} + void invalidate(){m_tag_array->invalidate();} void print(FILE *fp, unsigned &accesses, unsigned &misses) const; void display_state( FILE *fp ) const; @@ -612,6 +1087,15 @@ public: bool data_port_free() const { return m_bandwidth_management.data_port_free(); } bool fill_port_free() const { return m_bandwidth_management.fill_port_free(); } + // This is a gapping hole we are poking in the system to quickly handle + // filling the cache on cudamemcopies. We don't care about anything other than + // L2 state after the memcopy - so just force the tag array to act as though + // something is read or written without doing anything else. + void force_tag_access( new_addr_type addr, unsigned time, mem_access_sector_mask_t mask ) + { + m_tag_array->fill( addr, time, mask ); + } + protected: // Constructor that can be used by derived classes with custom tag arrays baseline_cache( const char *name, @@ -633,24 +1117,31 @@ protected: std::string m_name; cache_config &m_config; tag_array* m_tag_array; - mshr_table m_mshrs; + mshr_table m_mshrs; std::list m_miss_queue; enum mem_fetch_status m_miss_queue_status; mem_fetch_interface *m_memport; struct extra_mf_fields { extra_mf_fields() { m_valid = false;} - extra_mf_fields( new_addr_type a, unsigned i, unsigned d ) + extra_mf_fields( new_addr_type a, new_addr_type ad, unsigned i, unsigned d, const cache_config& m_config) { m_valid = true; m_block_addr = a; + m_addr = ad; m_cache_index = i; m_data_size = d; + pending_read = m_config.m_mshr_type == SECTOR_ASSOC? m_config.m_line_sz/SECTOR_SIZE : 0; + } bool m_valid; new_addr_type m_block_addr; + new_addr_type m_addr; unsigned m_cache_index; unsigned m_data_size; + //this variable is used when a load request generates multiple load transactions + //For example, a read request from non-sector L1 request sends a request to sector L2 + unsigned pending_read; }; typedef std::map extra_mf_fields_lookup; @@ -668,7 +1159,7 @@ protected: unsigned time, bool &do_miss, std::list &events, bool read_only, bool wa); /// Read miss handler. Check MSHR hit or MSHR available void send_read_request(new_addr_type addr, new_addr_type block_addr, unsigned cache_index, mem_fetch *mf, - unsigned time, bool &do_miss, bool &wb, cache_block_t &evicted, std::list &events, bool read_only, bool wa); + unsigned time, bool &do_miss, bool &wb, evicted_block_info &evicted, std::list &events, bool read_only, bool wa); /// Sub-class containing all metadata for port bandwidth management class bandwidth_management @@ -760,8 +1251,10 @@ public: // Set write miss function switch(m_config.m_write_alloc_policy){ - case WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_wa; break; case NO_WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_no_wa; break; + case WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_wa_naive; break; + case FETCH_ON_WRITE: m_wr_miss = &data_cache::wr_miss_wa_fetch_on_write; break; + case LAZY_FETCH_ON_READ: m_wr_miss = &data_cache::wr_miss_wa_lazy_fetch_on_read; break; default: assert(0 && "Error: Must set valid cache write miss policy\n"); break; // Need to set a write miss function @@ -870,12 +1363,33 @@ protected: /// Sends read request, and possible write-back request, // to lower level memory for a write miss with write-allocate enum cache_request_status - wr_miss_wa( new_addr_type addr, - unsigned cache_index, - mem_fetch *mf, - unsigned time, - std::list &events, - enum cache_request_status status ); // write-allocate + wr_miss_wa_naive( new_addr_type addr, + unsigned cache_index, + mem_fetch *mf, + unsigned time, + std::list &events, + enum cache_request_status status ); // write-allocate-send-write-and-read-request + enum cache_request_status + wr_miss_wa_fetch_on_write( new_addr_type addr, + unsigned cache_index, + mem_fetch *mf, + unsigned time, + std::list &events, + enum cache_request_status status ); // write-allocate with fetch-on-every-write + enum cache_request_status + wr_miss_wa_lazy_fetch_on_read( new_addr_type addr, + unsigned cache_index, + mem_fetch *mf, + unsigned time, + std::list &events, + enum cache_request_status status ); // write-allocate with read-fetch-only + enum cache_request_status + wr_miss_wa_write_validate( new_addr_type addr, + unsigned cache_index, + mem_fetch *mf, + unsigned time, + std::list &events, + enum cache_request_status status ); // write-allocate that writes with no read fetch enum cache_request_status wr_miss_no_wa( new_addr_type addr, unsigned cache_index, @@ -991,7 +1505,7 @@ 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; @@ -1144,13 +1658,15 @@ 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 63ba759..a8be4d2 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -150,6 +150,8 @@ void power_config::reg_options(class OptionParser * opp) void memory_config::reg_options(class OptionParser * opp) { + option_parser_register(opp, "-perf_sim_memcpy", OPT_BOOL, &m_perf_sim_memcpy, + "Fill the L2 cache on memcpy", "1"); option_parser_register(opp, "-gpgpu_dram_scheduler", OPT_INT32, &scheduler_type, "0 = fifo, 1 = FR-FCFS (defaul)", "1"); option_parser_register(opp, "-gpgpu_dram_partition_queues", OPT_CSTR, &gpgpu_L2_queue_config, @@ -202,7 +204,27 @@ void memory_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-dram_latency", OPT_UINT32, &dram_latency, "DRAM latency (default 30)", "30"); - + option_parser_register(opp, "-dual_bus_interface", OPT_UINT32, &dual_bus_interface, + "dual_bus_interface (default = 0) ", + "0"); + option_parser_register(opp, "-dram_bnk_indexing_policy", OPT_UINT32, &dram_bnk_indexing_policy, + "dram_bnk_indexing_policy (0 = normal indexing, 1 = Xoring with the higher bits) (Default = 0)", + "0"); + option_parser_register(opp, "-dram_bnkgrp_indexing_policy", OPT_UINT32, &dram_bnkgrp_indexing_policy, + "dram_bnkgrp_indexing_policy (0 = take higher bits, 1 = take lower bits) (Default = 0)", + "0"); + option_parser_register(opp, "-Seperate_Write_Queue_Enable", OPT_BOOL, &seperate_write_queue_enabled, + "Seperate_Write_Queue_Enable", + "0"); + option_parser_register(opp, "-Write_Queue_Size", OPT_CSTR, &write_queue_size_opt, + "Write_Queue_Size", + "32:28:16"); + option_parser_register(opp, "-Elimnate_rw_turnaround", OPT_BOOL, &elimnate_rw_turnaround, + "elimnate_rw_turnaround i.e set tWTR and tRTW = 0", + "0"); + option_parser_register(opp, "-icnt_flit_size", OPT_UINT32, &icnt_flit_size, + "icnt_flit_size", + "32"); m_address_mapping.addrdec_setoption(opp); } @@ -229,6 +251,12 @@ 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}", @@ -256,6 +284,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shader_registers", OPT_UINT32, &gpgpu_shader_registers, "Number of registers per shader core. Limits number of concurrent CTAs. (default 8192)", "8192"); + option_parser_register(opp, "-gpgpu_ignore_resources_limitation", OPT_BOOL, &gpgpu_ignore_resources_limitation, + "gpgpu_ignore_resources_limitation (default 0)", + "0"); option_parser_register(opp, "-gpgpu_shader_cta", OPT_UINT32, &max_cta_per_core, "Maximum number of concurrent CTAs in shader (default 8)", "8"); @@ -277,6 +308,9 @@ 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"); @@ -295,6 +329,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shmem_warp_parts", OPT_INT32, &mem_warp_parts, "Number of portions a warp is divided into for shared memory bank conflict check ", "2"); + option_parser_register(opp, "-gpgpu_shmem_warp_parts", OPT_INT32, &mem_warp_parts, + "Number of portions a warp is divided into for shared memory bank conflict check ", + "2"); option_parser_register(opp, "-gpgpu_warpdistro_shader", OPT_INT32, &gpgpu_warpdistro_shader, "Specify which shader core to collect the warp size distribution from", "-1"); @@ -303,7 +340,7 @@ 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"); @@ -313,6 +350,9 @@ 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"); @@ -328,6 +368,9 @@ 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"); @@ -343,6 +386,9 @@ 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"); @@ -356,27 +402,33 @@ void shader_core_config::reg_options(class OptionParser * opp) "number of collector unit in ports (default = 0)", "0"); option_parser_register(opp, "-gpgpu_coalesce_arch", OPT_INT32, &gpgpu_coalesce_arch, - "Coalescing arch (default = 13, anything else is off for now)", + "Coalescing arch (GT200 = 13, Fermi = 20)", "13"); option_parser_register(opp, "-gpgpu_num_sched_per_core", OPT_INT32, &gpgpu_num_sched_per_core, "Number of warp schedulers per core", "1"); option_parser_register(opp, "-gpgpu_max_insn_issue_per_warp", OPT_INT32, &gpgpu_max_insn_issue_per_warp, - "Max number of instructions that can be issued per warp in one cycle by scheduler", - "2"); + "Max number of instructions that can be issued per warp in one cycle by scheduler (either 1 or 2)", + "2"); + option_parser_register(opp, "-gpgpu_dual_issue_diff_exec_units", OPT_BOOL, &gpgpu_dual_issue_diff_exec_units, + "should dual issue use two different execution unit resources (Default = 1)", + "1"); option_parser_register(opp, "-gpgpu_simt_core_sim_order", OPT_INT32, &simt_core_sim_order, "Select the simulation order of cores in a cluster (0=Fix, 1=Round-Robin)", "1"); option_parser_register(opp, "-gpgpu_pipeline_widths", OPT_CSTR, &pipeline_widths_string, "Pipeline widths " - "ID_OC_SP,ID_OC_SFU,ID_OC_TENSOR_CORE,ID_OC_MEM,OC_EX_SP,OC_EX_SFU,OC_EX_TENSOR_CORE,OC_EX_MEM,EX_WB", - "1,1,1,1,1,1,1,1,1,1" ); + "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,ID_OC_TENSOR_CORE,OC_EX_TENSOR_CORE", + "1,1,1,1,1,1,1,1,1,1,1" ); option_parser_register(opp, "-gpgpu_tensor_core_avail", OPT_INT32, &gpgpu_tensor_core_avail, "Tensor Core Available (default=0)", "0"); 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"); @@ -425,7 +477,6 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) option_parser_register(opp, "-gpgpu_flush_l2_cache", OPT_BOOL, &gpgpu_flush_l2_cache, "Flush L2 cache at the end of each kernel call", "0"); - option_parser_register(opp, "-gpgpu_deadlock_detect", OPT_BOOL, &gpu_deadlock_detect, "Stop the simulation at deadlock (1=on (default), 0=off)", "1"); @@ -669,7 +720,7 @@ gpgpu_sim::gpgpu_sim( const gpgpu_sim_config &config ) m_running_kernels.resize( config.max_concurrent_kernel, NULL ); m_last_issued_kernel = 0; - m_last_cluster_issue = 0; + m_last_cluster_issue = m_shader_config->n_simt_clusters-1; // this causes first launch to use simt cluster 0 *average_pipeline_duty_cycle=0; *active_sms=0; @@ -813,6 +864,7 @@ void gpgpu_sim::update_stats() { partiton_replys_in_parallel_total += partiton_replys_in_parallel; partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; gpu_tot_sim_cycle_parition_util += gpu_sim_cycle_parition_util ; + gpu_tot_occupancy += gpu_occupancy; gpu_sim_cycle = 0; partiton_reqs_in_parallel = 0; @@ -821,6 +873,7 @@ void gpgpu_sim::update_stats() { gpu_sim_cycle_parition_util = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; + gpu_occupancy = occupancy_stats(); } void gpgpu_sim::print_stats() @@ -994,6 +1047,9 @@ void gpgpu_sim::gpu_print_stat() printf("gpu_tot_sim_insn = %lld\n", gpu_tot_sim_insn+gpu_sim_insn); printf("gpu_tot_ipc = %12.4f\n", (float)(gpu_tot_sim_insn+gpu_sim_insn) / (gpu_tot_sim_cycle+gpu_sim_cycle)); printf("gpu_tot_issued_cta = %lld\n", gpu_tot_issued_cta + m_total_cta_launched); + printf("gpu_occupancy = %.4f\% \n", gpu_occupancy.get_occ_fraction() * 100); + printf("gpu_tot_occupancy = %.4f\% \n", (gpu_occupancy + gpu_tot_occupancy).get_occ_fraction() * 100); + extern unsigned long long g_max_total_param_size; fprintf(statfout, "max_total_param_size = %llu\n", g_max_total_param_size); @@ -1032,6 +1088,8 @@ void gpgpu_sim::gpu_print_stat() } printf("\nTotal_core_cache_stats:\n"); core_cache_stats.print_stats(stdout, "Total_core_cache_stats_breakdown"); + printf("\nTotal_core_cache_fail_stats:\n"); + core_cache_stats.print_fail_stats(stdout, "Total_core_cache_fail_stats_breakdown"); shader_print_scheduler_stat( stdout, false ); m_shader_stats->print(stdout); @@ -1076,6 +1134,8 @@ void gpgpu_sim::gpu_print_stat() printf("L2_total_cache_reservation_fails = %u\n", total_l2_css.res_fails); printf("L2_total_cache_breakdown:\n"); l2_stats.print_stats(stdout, "L2_cache_stats_breakdown"); + printf("L2_total_cache_reservation_fail_breakdown:\n"); + l2_stats.print_fail_stats(stdout, "L2_cache_stats_fail_breakdown"); total_l2_css.print_port_stats(stdout, "L2_cache"); } } @@ -1429,7 +1489,7 @@ void gpgpu_sim::cycle() if (mf) { unsigned response_size = mf->get_is_write()?mf->get_ctrl_size():mf->size(); if ( ::icnt_has_buffer( m_shader_config->mem2device(i), response_size ) ) { - if (!mf->get_is_write()) + //if (!mf->get_is_write()) mf->set_return_timestamp(gpu_sim_cycle+gpu_tot_sim_cycle); mf->set_status(IN_ICNT_TO_SHADER,gpu_sim_cycle+gpu_tot_sim_cycle); ::icnt_push( m_shader_config->mem2device(i), mf->get_tpc(), mf, response_size ); @@ -1462,12 +1522,14 @@ void gpgpu_sim::cycle() for (unsigned i=0;im_n_mem_sub_partition;i++) { //move memory request from interconnect into memory partition (if not backed up) //Note:This needs to be called in DRAM clock domain if there is no L2 cache in the system - if ( m_memory_sub_partition[i]->full() ) { + //In the worst case, we may need to push SECTOR_CHUNCK_SIZE requests, so ensure you have enough buffer for them + if ( m_memory_sub_partition[i]->full(SECTOR_CHUNCK_SIZE) ) { gpu_stall_dramfull++; } else { mem_fetch* mf = (mem_fetch*) icnt_pop( m_shader_config->mem2device(i) ); m_memory_sub_partition[i]->push( mf, gpu_sim_cycle + gpu_tot_sim_cycle ); - partiton_reqs_in_parallel_per_cycle++; + if(mf) + partiton_reqs_in_parallel_per_cycle++; } m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle+gpu_tot_sim_cycle); m_memory_sub_partition[i]->accumulate_L2cache_stats(m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]); @@ -1494,6 +1556,8 @@ void gpgpu_sim::cycle() // Update core icnt/cache stats for GPUWattch m_cluster[i]->get_icnt_stats(m_power_stats->pwr_mem_stat->n_simt_to_mem[CURRENT_STAT_IDX][i], m_power_stats->pwr_mem_stat->n_mem_to_simt[CURRENT_STAT_IDX][i]); m_cluster[i]->get_cache_stats(m_power_stats->pwr_mem_stat->core_cache_stats[CURRENT_STAT_IDX]); + m_cluster[i]->get_current_occupancy(gpu_occupancy.aggregate_warp_slot_filled, gpu_occupancy.aggregate_theoretical_warp_slots); + } float temp=0; for (unsigned i=0;inum_shader();i++){ @@ -1521,12 +1585,12 @@ void gpgpu_sim::cycle() issue_block2core(); - // Depending on configuration, flush the caches once all of threads are completed. + // Depending on configuration, invalidate the caches once all of threads are completed. int all_threads_complete = 1; if (m_config.gpgpu_flush_l1_cache) { for (unsigned i=0;in_simt_clusters;i++) { if (m_cluster[i]->get_not_completed() == 0) - m_cluster[i]->cache_flush(); + m_cluster[i]->cache_invalidate(); else all_threads_complete = 0 ; } @@ -1548,7 +1612,7 @@ void gpgpu_sim::cycle() int dlc = 0; for (unsigned i=0;im_n_mem;i++) { dlc = m_memory_sub_partition[i]->flushL2(); - assert (dlc == 0); // need to model actual writes to DRAM here + assert (dlc == 0); // TODO: need to model actual writes to DRAM here printf("Dirty lines flushed from L2 %d is %d\n", i, dlc ); } } @@ -1560,15 +1624,20 @@ void gpgpu_sim::cycle() time_t curr_time; time(&curr_time); unsigned long long elapsed_time = MAX(curr_time - g_simulation_starttime, 1); - if ( (elapsed_time - last_liveness_message_time) >= m_config.liveness_message_freq ) { + if ( (elapsed_time - last_liveness_message_time) >= m_config.liveness_message_freq && DTRACE(LIVENESS) ) { days = elapsed_time/(3600*24); hrs = elapsed_time/3600 - 24*days; minutes = elapsed_time/60 - 60*(hrs + 24*days); sec = elapsed_time - 60*(minutes + 60*(hrs + 24*days)); - - DPRINTF(LIVENESS, "GPGPU-Sim uArch: cycles simulated: %lld inst.: %lld (ipc=%4.1f) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", - gpu_tot_sim_cycle + gpu_sim_cycle, gpu_tot_sim_insn + gpu_sim_insn, + + unsigned long long active = 0, total = 0; + for (unsigned i=0;in_simt_clusters;i++) { + m_cluster[i]->get_current_occupancy(active, total); + } + DPRINTF(LIVENESS, "uArch: inst.: %lld (ipc=%4.1f, occ=%0.4f\% [%llu / %llu]) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", + gpu_tot_sim_insn + gpu_sim_insn, (double)gpu_sim_insn/(double)gpu_sim_cycle, + float(active)/float(total) * 100, active, total, (unsigned)((gpu_tot_sim_insn+gpu_sim_insn) / elapsed_time), (unsigned)days,(unsigned)hrs,(unsigned)minutes,(unsigned)sec, ctime(&curr_time)); @@ -1620,6 +1689,24 @@ void shader_core_ctx::dump_warp_state( FILE *fout ) const m_warp[w].print(fout); } + +void gpgpu_sim::perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ) +{ + if (m_memory_config->m_perf_sim_memcpy) { + assert (dst_start_addr % 32 == 0); + + for ( unsigned counter = 0; counter < count; counter += 32 ) { + const unsigned wr_addr = dst_start_addr + counter; + addrdec_t raw_addr; + mem_access_sector_mask_t mask; + mask.set(wr_addr % 128 / 32); + m_memory_config->m_address_mapping.addrdec_tlx( wr_addr, &raw_addr ); + const unsigned partition_id = raw_addr.sub_partition / m_memory_config->m_n_sub_partition_per_memory_channel; + m_memory_partition_unit[ partition_id ]->handle_memcpy_to_gpu( wr_addr, raw_addr.sub_partition, mask ); + } + } +} + void gpgpu_sim::dump_pipeline( int mask, int s, int m ) const { /* diff --git a/src/gpgpu-sim/gpu-sim.h b/src/gpgpu-sim/gpu-sim.h index 7d92c66..1bae1fa 100644 --- a/src/gpgpu-sim/gpu-sim.h +++ b/src/gpgpu-sim/gpu-sim.h @@ -198,8 +198,14 @@ struct memory_config { bk_tag_length = i-1; assert(nbkgrp>0 && "Number of bank groups cannot be zero"); tRCDWR = tRCD-(WL+1); + if(elimnate_rw_turnaround) + { + tRTW = 0; + tWTR = 0; + } else { tRTW = (CL+(BL/data_command_freq_ratio)+2-WL); - tWTR = (WL+(BL/data_command_freq_ratio)+tCDLR); + tWTR = (WL+(BL/data_command_freq_ratio)+tCDLR); + } tWTP = (WL+(BL/data_command_freq_ratio)+tWR); dram_atom_size = BL * busW * gpu_n_mem_per_ctrlr; // burst length x bus width x # chips per partition @@ -213,7 +219,9 @@ struct memory_config { m_L2_config.init(&m_address_mapping); m_valid = true; - icnt_flit_size = 32; // Default 32 + + sscanf(write_queue_size_opt,"%d:%d:%d", + &gpgpu_frfcfs_dram_write_queue_size,&write_high_watermark,&write_low_watermark); } void reg_options(class OptionParser * opp); @@ -264,12 +272,25 @@ struct memory_config { unsigned nbk; + bool elimnate_rw_turnaround; + unsigned data_command_freq_ratio; // frequency ratio between DRAM data bus and command bus (2 for GDDR3, 4 for GDDR5) unsigned dram_atom_size; // number of bytes transferred per read or write command linear_to_raw_address_translation m_address_mapping; unsigned icnt_flit_size; + + unsigned dram_bnk_indexing_policy; + unsigned dram_bnkgrp_indexing_policy; + bool dual_bus_interface; + + bool seperate_write_queue_enabled; + char *write_queue_size_opt; + unsigned gpgpu_frfcfs_dram_write_queue_size; + unsigned write_high_watermark; + unsigned write_low_watermark; + bool m_perf_sim_memcpy; }; // global counters and flags (please try not to add to this list!!!) @@ -362,6 +383,32 @@ private: friend class gpgpu_sim; }; +struct occupancy_stats { + occupancy_stats() : aggregate_warp_slot_filled(0), aggregate_theoretical_warp_slots(0){} + occupancy_stats( unsigned long long wsf, unsigned long long tws ) + : aggregate_warp_slot_filled(wsf), aggregate_theoretical_warp_slots(tws){} + + unsigned long long aggregate_warp_slot_filled; + unsigned long long aggregate_theoretical_warp_slots; + + float get_occ_fraction() const { + return float(aggregate_warp_slot_filled) / float(aggregate_theoretical_warp_slots); + } + + occupancy_stats& operator+=(const occupancy_stats& rhs) { + aggregate_warp_slot_filled += rhs.aggregate_warp_slot_filled; + aggregate_theoretical_warp_slots += rhs.aggregate_theoretical_warp_slots; + return *this; + } + + occupancy_stats operator+(const occupancy_stats& rhs) const{ + return occupancy_stats( aggregate_warp_slot_filled + rhs.aggregate_warp_slot_filled, + aggregate_theoretical_warp_slots + rhs.aggregate_theoretical_warp_slots + ); + } +}; + + class gpgpu_sim : public gpgpu_t { public: gpgpu_sim( const gpgpu_sim_config &config ); @@ -405,6 +452,8 @@ public: void gpu_print_stat(); void dump_pipeline( int mask, int s, int m ) const; + void perf_memcpy_to_gpu( size_t dst_start_addr, size_t count ); + //The next three functions added to be used by the functional simulation function //! Get shader core configuration @@ -498,6 +547,9 @@ public: unsigned long long gpu_tot_sim_insn; unsigned long long gpu_sim_insn_last_update; unsigned gpu_sim_insn_last_update_sid; + occupancy_stats gpu_occupancy; + occupancy_stats gpu_tot_occupancy; + FuncCache get_cache_config(std::string kernel_name); void set_cache_config(std::string kernel_name, FuncCache cacheConfig ); @@ -526,4 +578,5 @@ public: } }; + #endif diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc index c5fc44e..25da107 100644 --- a/src/gpgpu-sim/l2cache.cc +++ b/src/gpgpu-sim/l2cache.cc @@ -74,6 +74,15 @@ memory_partition_unit::memory_partition_unit( unsigned partition_id, } } +void memory_partition_unit::handle_memcpy_to_gpu( size_t addr, unsigned global_subpart_id, mem_access_sector_mask_t mask ) +{ + unsigned p = global_sub_partition_id_to_local_id(global_subpart_id); + std::string mystring = + mask.to_string(); + MEMPART_DPRINTF("Copy Engine Request Received For Address=%llx, local_subpart=%u, global_subpart=%u, sector_mask=%s \n", addr, p, global_subpart_id, mystring.c_str()); + m_sub_partition[p]->force_l2_tag_update(addr,gpu_sim_cycle+gpu_tot_sim_cycle, mask); +} + memory_partition_unit::~memory_partition_unit() { delete m_dram; @@ -93,7 +102,9 @@ memory_partition_unit::arbitration_metadata::arbitration_metadata(const struct m m_private_credit_limit = 1; m_shared_credit_limit = config->gpgpu_frfcfs_dram_sched_queue_size + config->gpgpu_dram_return_queue_size - - (config->m_n_sub_partition_per_memory_channel - 1); + - (config->m_n_sub_partition_per_memory_channel - 1); + if(config->seperate_write_queue_enabled ) + m_shared_credit_limit += config->gpgpu_frfcfs_dram_write_queue_size; if (config->gpgpu_frfcfs_dram_sched_queue_size == 0 or config->gpgpu_dram_return_queue_size == 0) { @@ -220,7 +231,8 @@ void memory_partition_unit::dram_cycle() m_dram->cycle(); m_dram->dram_log(SAMPLELOG); - if( !m_dram->full() ) { + // mem_fetch *mf = m_sub_partition[spid]->L2_dram_queue_top(); + //if( !m_dram->full(mf->is_write()) ) { // L2->DRAM queue to DRAM latency queue // Arbitrate among multiple L2 subpartitions int last_issued_partition = m_arbitration_metadata.last_borrower(); @@ -228,6 +240,9 @@ void memory_partition_unit::dram_cycle() int spid = (p + last_issued_partition + 1) % m_config->m_n_sub_partition_per_memory_channel; if (!m_sub_partition[spid]->L2_dram_queue_empty() && can_issue_to_dram(spid)) { mem_fetch *mf = m_sub_partition[spid]->L2_dram_queue_top(); + if(m_dram->full(mf->is_write()) ) + break; + m_sub_partition[spid]->L2_dram_queue_pop(); MEMPART_DPRINTF("Issue mem_fetch request %p from sub partition %d to dram\n", mf, spid); dram_delay_t d; @@ -239,12 +254,13 @@ void memory_partition_unit::dram_cycle() break; // the DRAM should only accept one request per cycle } } - } + //} // DRAM latency queue - if( !m_dram_latency_queue.empty() && ( (gpu_sim_cycle+gpu_tot_sim_cycle) >= m_dram_latency_queue.front().ready_cycle ) && !m_dram->full() ) { - mem_fetch* mf = m_dram_latency_queue.front().req; - m_dram_latency_queue.pop_front(); + + if( !m_dram_latency_queue.empty() && ( (gpu_sim_cycle+gpu_tot_sim_cycle) >= m_dram_latency_queue.front().ready_cycle ) && !m_dram->full(m_dram_latency_queue.front().req->is_write()) ) { + mem_fetch* mf = m_dram_latency_queue.front().req; + m_dram_latency_queue.pop_front(); m_dram->push(mf); } } @@ -299,6 +315,7 @@ memory_sub_partition::memory_sub_partition( unsigned sub_partition_id, m_id = sub_partition_id; m_config=config; m_stats=stats; + m_memcpy_cycle_offset = 0; assert(m_id < m_config->m_n_mem_sub_partition); @@ -343,6 +360,14 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); m_L2_icnt_queue->push(mf); }else{ + if(m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE) + { + mem_fetch* original_wr_mf = mf->get_original_wr_mf(); + assert(original_wr_mf); + original_wr_mf->set_reply(); + original_wr_mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); + m_L2_icnt_queue->push(original_wr_mf); + } m_request_tracker.erase(mf); delete mf; } @@ -355,10 +380,11 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) if ( !m_config->m_L2_config.disabled() && m_L2cache->waiting_for_fill(mf) ) { if (m_L2cache->fill_port_free()) { mf->set_status(IN_PARTITION_L2_FILL_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); - m_L2cache->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle); + m_L2cache->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle+m_memcpy_cycle_offset); m_dram_L2_queue->pop(); } } else if ( !m_L2_icnt_queue->full() ) { + if(mf->is_write() && mf->get_type() == WRITE_ACK) mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); m_L2_icnt_queue->push(mf); m_dram_L2_queue->pop(); @@ -380,9 +406,10 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) bool port_free = m_L2cache->data_port_free(); if ( !output_full && port_free ) { std::list events; - enum cache_request_status status = m_L2cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle,events); + enum cache_request_status status = m_L2cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle+m_memcpy_cycle_offset,events); bool write_sent = was_write_sent(events); bool read_sent = was_read_sent(events); + MEM_SUBPART_DPRINTF("Probing L2 cache Address=%llx, status=%u\n", mf->get_addr(), status); if ( status == HIT ) { if( !write_sent ) { @@ -402,6 +429,11 @@ void memory_sub_partition::cache_cycle( unsigned cycle ) m_icnt_L2_queue->pop(); } } else if ( status != RESERVATION_FAIL ) { + if(mf->is_write() && (m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE || m_config->m_L2_config.m_write_alloc_policy == LAZY_FETCH_ON_READ) && !was_writeallocate_sent(events)) { + mf->set_reply(); + mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); + m_L2_icnt_queue->push(mf); + } // L2 cache accepted request m_icnt_L2_queue->pop(); } else { @@ -432,6 +464,11 @@ bool memory_sub_partition::full() const return m_icnt_L2_queue->full(); } +bool memory_sub_partition::full(unsigned size) const +{ + return m_icnt_L2_queue->is_avilable_size(size); +} + bool memory_sub_partition::L2_dram_queue_empty() const { return m_L2_dram_queue->empty(); @@ -532,7 +569,15 @@ unsigned memory_sub_partition::flushL2() if (!m_config->m_L2_config.disabled()) { m_L2cache->flush(); } - return 0; // L2 is read only in this version + return 0; //TODO: write the flushed data to the main memory +} + +unsigned memory_sub_partition::invalidateL2() +{ + if (!m_config->m_L2_config.disabled()) { + m_L2cache->invalidate(); + } + return 0; } bool memory_sub_partition::busy() const @@ -540,21 +585,94 @@ bool memory_sub_partition::busy() const return !m_request_tracker.empty(); } -void memory_sub_partition::push( mem_fetch* req, unsigned long long cycle ) -{ - if (req) { - m_request_tracker.insert(req); - m_stats->memlatstat_icnt2mem_pop(req); - if( req->istexture() ) { - m_icnt_L2_queue->push(req); - req->set_status(IN_PARTITION_ICNT_TO_L2_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); - } else { - rop_delay_t r; - r.req = req; - r.ready_cycle = cycle + m_config->rop_latency; - m_rop.push(r); - req->set_status(IN_PARTITION_ROP_DELAY,gpu_sim_cycle+gpu_tot_sim_cycle); - } +std::vector memory_sub_partition::breakdown_request_to_sector_requests(mem_fetch* mf) +{ + std::vector result; + + if(mf->get_data_size() == SECTOR_SIZE && mf->get_access_sector_mask().count() == 1) { + result.push_back(mf); + } else if (mf->get_data_size() == 128 || mf->get_data_size() == 64) { + //We only accept 32, 64 and 128 bytes reqs + unsigned start=0, end=0; + if(mf->get_data_size() == 128) { + start=0; end=3; + } else if (mf->get_data_size() == 64 && mf->get_access_sector_mask().to_string() == "1100") { + start=2; end=3; + } else if (mf->get_data_size() == 64 && mf->get_access_sector_mask().to_string() == "0011") { + start=0; end=1; + } else if (mf->get_data_size() == 64 && (mf->get_access_sector_mask().to_string() == "1111" || mf->get_access_sector_mask().to_string() == "0000")) { + if(mf->get_addr() % 128 == 0) { + start=0; end=1; + } else { + start=2; end=3; + } + } 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()); + assert(0 && "Undefined sector mask is received"); + } + + std::bitset byte_sector_mask; + byte_sector_mask.reset(); + for(unsigned k=start*SECTOR_SIZE; k< SECTOR_SIZE; ++k) + byte_sector_mask.set(k); + + for(unsigned j=start, i=0; j<= end ; ++j, ++i){ + + const mem_access_t *ma = new mem_access_t( mf->get_access_type(), + mf->get_addr() + SECTOR_SIZE*i, + SECTOR_SIZE, + mf->is_write(), + mf->get_access_warp_mask(), + mf->get_access_byte_mask() & byte_sector_mask, + std::bitset().set(j)); + + mem_fetch *n_mf = new mem_fetch( *ma, + NULL, + mf->get_ctrl_size(), + mf->get_wid(), + mf->get_sid(), + mf->get_tpc(), + mf->get_mem_config(), + mf); + + result.push_back(n_mf); + byte_sector_mask <<= SECTOR_SIZE; + } + } 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; +} + +void memory_sub_partition::push( mem_fetch* m_req, unsigned long long cycle ) +{ + if (m_req) { + m_stats->memlatstat_icnt2mem_pop(m_req); + std::vector reqs; + if(m_config->m_L2_config.m_cache_type == SECTOR) + reqs = breakdown_request_to_sector_requests(m_req); + else + reqs.push_back(m_req); + + for(unsigned i=0; iistexture() ) { + m_icnt_L2_queue->push(req); + req->set_status(IN_PARTITION_ICNT_TO_L2_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle); + } else { + rop_delay_t r; + r.req = req; + r.ready_cycle = cycle + m_config->rop_latency; + m_rop.push(r); + req->set_status(IN_PARTITION_ROP_DELAY,gpu_sim_cycle+gpu_tot_sim_cycle); + } + } } } diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h index 3df54b1..18c0a8b 100644 --- a/src/gpgpu-sim/l2cache.h +++ b/src/gpgpu-sim/l2cache.h @@ -72,6 +72,7 @@ public: void print_stat( FILE *fp ) { m_dram->print_stat(fp); } void visualize() const { m_dram->visualize(); } void print( FILE *fp ) const; + void handle_memcpy_to_gpu( size_t dst_start_addr, unsigned subpart_id, mem_access_sector_mask_t mask ); class memory_sub_partition * get_sub_partition(int sub_partition_id) { @@ -154,12 +155,14 @@ public: void cache_cycle( unsigned cycle ); bool full() const; + bool full(unsigned size) const; void push( class mem_fetch* mf, unsigned long long clock_cycle ); class mem_fetch* pop(); class mem_fetch* top(); void set_done( mem_fetch *mf ); unsigned flushL2(); + unsigned invalidateL2(); // interface to L2_dram_queue bool L2_dram_queue_empty() const; @@ -177,6 +180,12 @@ public: void accumulate_L2cache_stats(class cache_stats &l2_stats) const; void get_L2cache_sub_stats(struct cache_sub_stats &css) const; + void force_l2_tag_update(new_addr_type addr, unsigned time, mem_access_sector_mask_t mask) + { + m_L2cache->force_tag_access( addr, m_memcpy_cycle_offset + time, mask ); + m_memcpy_cycle_offset += 1; + } + private: // data unsigned m_id; //< the global sub partition ID @@ -207,6 +216,15 @@ private: std::set m_request_tracker; friend class L2interface; + + std::vector breakdown_request_to_sector_requests(mem_fetch* mf); + + // This is a cycle offset that has to be applied to the l2 accesses to account for + // the cudamemcpy read/writes. We want GPGPU-Sim to only count cycles for kernel execution + // but we want cudamemcpy to go through the L2. Everytime an access is made from cudamemcpy + // this counter is incremented, and when the l2 is accessed (in both cudamemcpyies and otherwise) + // this value is added to the gpgpu-sim cycle counters. + unsigned m_memcpy_cycle_offset; }; class L2interface : public mem_fetch_interface { diff --git a/src/gpgpu-sim/l2cache_trace.h b/src/gpgpu-sim/l2cache_trace.h index 3dac87d..2235cdc 100644 --- a/src/gpgpu-sim/l2cache_trace.h +++ b/src/gpgpu-sim/l2cache_trace.h @@ -34,6 +34,9 @@ #define MEMPART_PRINT_STR SIM_PRINT_STR " %d - " #define MEMPART_DTRACE(x) ( DTRACE(x) && (Trace::sampling_memory_partition == -1 || Trace::sampling_memory_partition == (int)get_mpid()) ) +#define MEM_SUBPART_PRINT_STR SIM_PRINT_STR " %d - " +#define MEM_SUBPART_DTRACE(x) ( DTRACE(x) && (Trace::sampling_memory_partition == -1 || Trace::sampling_memory_partition == (int)m_id) ) + // Intended to be called from inside components of a memory partition // Depends on a get_mpid() function #define MEMPART_DPRINTF(...) do {\ @@ -46,10 +49,23 @@ }\ } while (0) +#define MEM_SUBPART_DPRINTF(...) do {\ + if (MEM_SUBPART_DTRACE(MEMORY_PARTITION_UNIT)) {\ + printf( MEM_SUBPART_PRINT_STR,\ + gpu_sim_cycle + gpu_tot_sim_cycle,\ + Trace::trace_streams_str[Trace::MEMORY_SUBPARTITION_UNIT],\ + m_id );\ + printf(__VA_ARGS__);\ + }\ +} while (0) + #else #define MEMPART_DTRACE(x) (false) #define MEMPART_DPRINTF(x, ...) do {} while (0) +#define MEM_SUBPART_DTRACE(x) (false) +#define MEM_SUBPART_DPRINTF(x, ...) do {} while (0) + #endif diff --git a/src/gpgpu-sim/mem_fetch.cc b/src/gpgpu-sim/mem_fetch.cc index 729636d..a260a35 100644 --- a/src/gpgpu-sim/mem_fetch.cc +++ b/src/gpgpu-sim/mem_fetch.cc @@ -39,7 +39,10 @@ mem_fetch::mem_fetch( const mem_access_t &access, unsigned wid, unsigned sid, unsigned tpc, - const struct memory_config *config ) + const struct memory_config *config, + mem_fetch *m_original_mf, + mem_fetch *m_original_wr_mf) + { m_request_uid = sm_next_mf_request_uid++; m_access = access; @@ -61,6 +64,8 @@ mem_fetch::mem_fetch( const mem_access_t &access, m_status_change = gpu_sim_cycle + gpu_tot_sim_cycle; m_mem_config = config; icnt_flit_size = config->icnt_flit_size; + original_mf = m_original_mf; + original_wr_mf = m_original_wr_mf; } mem_fetch::~mem_fetch() diff --git a/src/gpgpu-sim/mem_fetch.h b/src/gpgpu-sim/mem_fetch.h index de98748..e5efffd 100644 --- a/src/gpgpu-sim/mem_fetch.h +++ b/src/gpgpu-sim/mem_fetch.h @@ -55,7 +55,9 @@ public: unsigned wid, unsigned sid, unsigned tpc, - const struct memory_config *config ); + const struct memory_config *config, + mem_fetch *original_mf = NULL, + mem_fetch *original_wr_mf = NULL); ~mem_fetch(); void set_status( enum mem_fetch_status status, unsigned long long cycle ); @@ -104,6 +106,7 @@ public: enum mem_access_type get_access_type() const { return m_access.get_type(); } const active_mask_t& get_access_warp_mask() const { return m_access.get_warp_mask(); } mem_access_byte_mask_t get_access_byte_mask() const { return m_access.get_byte_mask(); } + mem_access_sector_mask_t get_access_sector_mask() const { return m_access.get_sector_mask(); } address_type get_pc() const { return m_inst.empty()?-1:m_inst.pc; } const warp_inst_t &get_inst() { return m_inst; } @@ -112,6 +115,10 @@ public: const memory_config *get_mem_config(){return m_mem_config;} unsigned get_num_flits(bool simt_to_mem); + + mem_fetch* get_original_mf() { return original_mf; } + mem_fetch* get_original_wr_mf() { return original_wr_mf; } + private: // request source information unsigned m_request_uid; @@ -143,6 +150,10 @@ private: const struct memory_config *m_mem_config; unsigned icnt_flit_size; + + mem_fetch* original_mf; //this pointer is set up when a request is divided into sector requests at L2 cache (if the req size > L2 sector size), so the pointer refers to the original request + mem_fetch* original_wr_mf; //this pointer refers to the original write req, when fetch-on-write policy is used + }; #endif diff --git a/src/gpgpu-sim/mem_latency_stat.cc b/src/gpgpu-sim/mem_latency_stat.cc index fde0eff..c5452b9 100644 --- a/src/gpgpu-sim/mem_latency_stat.cc +++ b/src/gpgpu-sim/mem_latency_stat.cc @@ -75,6 +75,10 @@ memory_stats_t::memory_stats_t( unsigned n_shader, const struct shader_core_conf max_mf_latency = 0; max_icnt2mem_latency = 0; max_icnt2sh_latency = 0; + tot_icnt2mem_latency = 0; + tot_icnt2sh_latency = 0; + tot_mrq_num = 0; + tot_mrq_latency = 0; memset(mrq_lat_table, 0, sizeof(unsigned)*32); memset(dq_lat_table, 0, sizeof(unsigned)*32); memset(mf_lat_table, 0, sizeof(unsigned)*32); @@ -158,6 +162,7 @@ void memory_stats_t::memlatstat_read_done(mem_fetch *mf) mf_max_lat_table[mf->get_tlx_addr().chip][mf->get_tlx_addr().bk] = mf_latency; unsigned icnt2sh_latency; icnt2sh_latency = (gpu_tot_sim_cycle+gpu_sim_cycle) - mf->get_return_timestamp(); + tot_icnt2sh_latency += icnt2sh_latency; icnt2sh_lat_table[LOGB2(icnt2sh_latency)]++; if (icnt2sh_latency > max_icnt2sh_latency) max_icnt2sh_latency = icnt2sh_latency; @@ -191,6 +196,7 @@ void memory_stats_t::memlatstat_icnt2mem_pop(mem_fetch *mf) if (m_memory_config->gpgpu_memlatency_stat) { unsigned icnt2mem_latency; icnt2mem_latency = (gpu_tot_sim_cycle+gpu_sim_cycle) - mf->get_timestamp(); + tot_icnt2mem_latency += icnt2mem_latency; icnt2mem_lat_table[LOGB2(icnt2mem_latency)]++; if (icnt2mem_latency > max_icnt2mem_latency) max_icnt2mem_latency = icnt2mem_latency; @@ -216,14 +222,19 @@ void memory_stats_t::memlatstat_print( unsigned n_mem, unsigned gpu_mem_n_bk ) unsigned max_bank_accesses, min_bank_accesses, max_chip_accesses, min_chip_accesses; if (m_memory_config->gpgpu_memlatency_stat) { + printf("maxmflatency = %d \n", max_mf_latency); + printf("max_icnt2mem_latency = %d \n", max_icnt2mem_latency); printf("maxmrqlatency = %d \n", max_mrq_latency); - printf("maxdqlatency = %d \n", max_dq_latency); - printf("maxmflatency = %d \n", max_mf_latency); + //printf("maxdqlatency = %d \n", max_dq_latency); + printf("max_icnt2sh_latency = %d \n", max_icnt2sh_latency); if (num_mfs) { printf("averagemflatency = %lld \n", mf_total_lat/num_mfs); + printf("avg_icnt2mem_latency = %lld \n", tot_icnt2mem_latency/num_mfs); + if(tot_mrq_num) + printf("avg_mrq_latency = %lld \n", tot_mrq_latency/tot_mrq_num); + + printf("avg_icnt2sh_latency = %lld \n", tot_icnt2sh_latency/num_mfs); } - printf("max_icnt2mem_latency = %d \n", max_icnt2mem_latency); - printf("max_icnt2sh_latency = %d \n", max_icnt2sh_latency); printf("mrq_lat_table:"); for (i=0; i< 32; i++) { printf("%d \t", mrq_lat_table[i]); diff --git a/src/gpgpu-sim/mem_latency_stat.h b/src/gpgpu-sim/mem_latency_stat.h index 4968a3b..5b89202 100644 --- a/src/gpgpu-sim/mem_latency_stat.h +++ b/src/gpgpu-sim/mem_latency_stat.h @@ -56,6 +56,10 @@ public: unsigned max_dq_latency; unsigned max_mf_latency; unsigned max_icnt2mem_latency; + unsigned long long int tot_icnt2mem_latency; + unsigned long long int tot_icnt2sh_latency; + unsigned long long int tot_mrq_latency; + unsigned long long int tot_mrq_num; unsigned max_icnt2sh_latency; unsigned mrq_lat_table[32]; unsigned dq_lat_table[32]; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index f9cfa58..82f9181 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -74,13 +74,14 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, shader_core_stats *stats ) : core_t( gpu, NULL, config->warp_size, config->n_thread_per_shader ), m_barriers( this, config->max_warps_per_shader, config->max_cta_per_core, config->max_barriers_per_cta, config->warp_size ), - m_dynamic_warp_id(0) + m_dynamic_warp_id(0), m_active_warps(0) { m_cluster = cluster; m_config = config; m_memory_config = mem_config; m_stats = stats; unsigned warp_size=config->warp_size; + Issue_Prio = 0; m_sid = shader_id; m_tpc = tpc_id; @@ -131,6 +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("warp_limiting") != std::string::npos ? CONCRETE_SCHEDULER_WARP_LIMITING: NUM_CONCRETE_SCHEDULERS; @@ -147,6 +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_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], @@ -162,6 +166,7 @@ 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_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], @@ -178,6 +183,7 @@ 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_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], @@ -185,6 +191,22 @@ 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_DP], + &m_pipeline_reg[ID_OC_SFU], + &m_pipeline_reg[ID_OC_TENSOR_CORE], + &m_pipeline_reg[ID_OC_MEM], + i + ) + ); + break; case CONCRETE_SCHEDULER_WARP_LIMITING: schedulers.push_back( new swl_scheduler( m_stats, @@ -193,6 +215,7 @@ 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_TENSOR_CORE], &m_pipeline_reg[ID_OC_MEM], @@ -215,8 +238,9 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, } //op collector configuration - enum { SP_CUS, SFU_CUS, TENSOR_CORE_CUS, MEM_CUS, GEN_CUS }; + enum { SP_CUS, DP_CUS, SFU_CUS, MEM_CUS, TENSOR_CORE_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_dp, 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(TENSOR_CORE_CUS, config->gpgpu_operand_collector_num_units_tensor_core, 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); @@ -234,6 +258,15 @@ 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]); @@ -279,7 +312,7 @@ 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 + config->gpgpu_num_tensor_core_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 + 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 ]; @@ -291,6 +324,12 @@ 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); @@ -307,7 +346,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, m_fu.push_back(m_ldst_unit); m_dispatch_port.push_back(ID_OC_MEM); m_issue_port.push_back(OC_EX_MEM); - + assert(m_num_function_units == m_fu.size() and m_fu.size() == m_dispatch_port.size() and m_fu.size() == m_issue_port.size()); //there are as many result buses as the width of the EX_WB stage @@ -341,6 +380,7 @@ void shader_core_ctx::reinit(unsigned start_thread, unsigned end_thread, bool re m_occupied_ctas = 0; m_occupied_hwtid.reset(); m_occupied_cta_to_hwtid.clear(); + m_active_warps = 0; } for (unsigned i = start_thread; iget_pdom_stack_top_info(pc,rpc); } +float shader_core_ctx::get_current_occupancy( unsigned long long & active, unsigned long long & total ) const +{ + // To match the achieved_occupancy in nvprof, only SMs that are active are counted toward the occupancy. + if ( m_active_warps > 0 ) { + total += m_warp.size(); + active += m_active_warps; + return float(active) / float(total); + } else { + return 0; + } +} + void shader_core_stats::print( FILE* fout ) const { unsigned long long thread_icount_uarch=0; @@ -436,15 +489,15 @@ void shader_core_stats::print( FILE* fout ) const fprintf(fout, "gpgpu_n_intrawarp_mshr_merge = %d\n", gpgpu_n_intrawarp_mshr_merge); fprintf(fout, "gpgpu_n_cmem_portconflict = %d\n", gpgpu_n_cmem_portconflict); - fprintf(fout, "gpgpu_stall_shd_mem[c_mem][bk_conf] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][BK_CONF]); - fprintf(fout, "gpgpu_stall_shd_mem[c_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[c_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[c_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][DATA_PORT_STALL]); - fprintf(fout, "gpgpu_stall_shd_mem[t_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[t_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[t_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][DATA_PORT_STALL]); + fprintf(fout, "gpgpu_stall_shd_mem[c_mem][resource_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][BK_CONF]); + //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][DATA_PORT_STALL]); + //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][DATA_PORT_STALL]); fprintf(fout, "gpgpu_stall_shd_mem[s_mem][bk_conf] = %d\n", gpu_stall_shd_mem_breakdown[S_MEM][BK_CONF]); - fprintf(fout, "gpgpu_stall_shd_mem[gl_mem][bk_conf] = %d\n", + fprintf(fout, "gpgpu_stall_shd_mem[gl_mem][resource_stall] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][BK_CONF] + gpu_stall_shd_mem_breakdown[G_MEM_ST][BK_CONF] + gpu_stall_shd_mem_breakdown[L_MEM_LD][BK_CONF] + @@ -462,22 +515,22 @@ void shader_core_stats::print( FILE* fout ) const gpu_stall_shd_mem_breakdown[L_MEM_LD][DATA_PORT_STALL] + gpu_stall_shd_mem_breakdown[L_MEM_ST][DATA_PORT_STALL] ); // data port stall at data cache - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_CACHE_RSRV_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_CACHE_RSRV_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_CACHE_RSRV_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][MSHR_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_ICNT_RC_FAIL]); - fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_CACHE_RSRV_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_CACHE_RSRV_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_CACHE_RSRV_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_CACHE_RSRV_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][MSHR_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_ICNT_RC_FAIL]); + //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_CACHE_RSRV_FAIL]); fprintf(fout, "gpu_reg_bank_conflict_stalls = %d\n", gpu_reg_bank_conflict_stalls); @@ -488,6 +541,14 @@ void shader_core_stats::print( FILE* fout ) const for (unsigned i = 3; i < m_config->warp_size + 3; i++) fprintf(fout, "\tW%d:%d", i-2, shader_cycle_distro[i]); fprintf(fout, "\n"); + fprintf(fout, "single_issue_nums: "); + for (unsigned i = 0; i < m_config->gpgpu_num_sched_per_core; i++) + fprintf(fout, "WS%d:%d\t", i, single_issue_nums[i]); + fprintf(fout, "\n"); + fprintf(fout, "dual_issue_nums: "); + for (unsigned i = 0; i < m_config->gpgpu_num_sched_per_core; i++) + fprintf(fout, "WS%d:%d\t", i, dual_issue_nums[i]); + fprintf(fout, "\n"); m_outgoing_traffic_stats->print(fout); m_incoming_traffic_stats->print(fout); @@ -669,13 +730,15 @@ void shader_core_ctx::fetch() } if( did_exit ) m_warp[warp_id].set_done_exit(); + --m_active_warps; + assert(m_active_warps >= 0); } // this code fetches instructions from the i-cache or generates memory requests if( !m_warp[warp_id].functional_done() && !m_warp[warp_id].imiss_pending() && m_warp[warp_id].ibuffer_empty() ) { address_type pc = m_warp[warp_id].get_pc(); address_type ppc = pc + PROGRAM_MEM_START; - unsigned nbytes=16; + unsigned nbytes=16; unsigned offset_in_block = pc & (m_config->m_L1I_config.get_line_sz()-1); if( (offset_in_block+nbytes) > m_config->m_L1I_config.get_line_sz() ) nbytes = (m_config->m_L1I_config.get_line_sz()-offset_in_block); @@ -750,10 +813,19 @@ void shader_core_ctx::issue_warp( register_set& pipe_reg_set, const warp_inst_t* } void shader_core_ctx::issue(){ + + //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++) { - schedulers[i]->cycle(); - } + //for (unsigned i = 0; i < schedulers.size(); i++) { + // schedulers[i]->cycle(); + //} } shd_warp_t& scheduler_unit::warp(int i){ @@ -868,7 +940,10 @@ void scheduler_unit::cycle() unsigned warp_id = (*iter)->get_warp_id(); unsigned checked=0; unsigned issued=0; - unsigned max_issue = m_shader->m_config->gpgpu_max_insn_issue_per_warp; + 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; @@ -901,18 +976,21 @@ void scheduler_unit::cycle() ready_inst = true; const active_mask_t &active_mask = m_simt_stack[warp_id]->get_active_mask(); assert( warp(warp_id).inst_in_pipeline() ); - if ( (pI->op == LOAD_OP)||(pI->op ==TENSOR_CORE_LOAD_OP)|| (pI->op == STORE_OP)|| (pI->op==TENSOR_CORE_STORE_OP) || (pI->op == MEMORY_BARRIER_OP) ) { - if( m_mem_out->has_free() ) { - m_shader->issue_warp(*m_mem_out,pI,active_mask,warp_id); - issued++; - issued_inst=true; - warp_inst_issued = true; - } + if ( (pI->op == LOAD_OP) || (pI->op == STORE_OP) || (pI->op == MEMORY_BARRIER_OP)||(pI->op==TENSOR_CORE_LOAD_OP)||(pI->op==TENSOR_CORE_STORE_OP) ) { + if( m_mem_out->has_free() && (!diff_exec_units || previous_issued_inst_exec_type != exec_unit_type_t::MEM)) { + m_shader->issue_warp(*m_mem_out,pI,active_mask,warp_id); + issued++; + issued_inst=true; + warp_inst_issued = true; + 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(); bool tensor_core_pipe_avail = m_tensor_core_out->has_free(); - if( sp_pipe_avail && (pI->op != SFU_OP) && (pI->op != TENSOR_CORE_OP) ) { + bool dp_pipe_avail = m_dp_out->has_free(); + if( sp_pipe_avail && (pI->op != TENSOR_CORE_OP) && (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) { @@ -937,12 +1015,23 @@ void scheduler_unit::cycle() issued++; issued_inst=true; warp_inst_issued = true; - } else if ( (pI->op == SFU_OP) || (pI->op == ALU_SFU_OP) ) { + previous_issued_inst_exec_type = exec_unit_type_t::SP; + } 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; + } + } //If the DP units = 0 (like in Fermi archi), then change DP inst to SFU inst + else if ( ((m_shader->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++; issued_inst=true; warp_inst_issued = true; + previous_issued_inst_exec_type = exec_unit_type_t::SFU; } } else if ( (pI->op == TENSOR_CORE_OP) ) { @@ -952,12 +1041,12 @@ void scheduler_unit::cycle() issued_inst=true; warp_inst_issued = true; } - } - } - } else { + } + }//end of else + } else { SCHED_DPRINTF( "Warp (warp_id %u, dynamic_warp_id %u) fails scoreboard\n", (*iter)->get_warp_id(), (*iter)->get_dynamic_warp_id() ); - } + } } } else if( valid ) { // this case can happen after a return instruction in diverged warp @@ -988,6 +1077,14 @@ void scheduler_unit::cycle() m_last_supervised_issued = supervised_iter; } } + + if(issued == 1) + m_stats->single_issue_nums[m_id]++; + else if(issued > 1) + m_stats->dual_issue_nums[m_id]++; + else + abort(); //issued should be > 0 + break; } } @@ -1045,6 +1142,16 @@ void gto_scheduler::order_warps() scheduler_unit::sort_warps_by_oldest_dynamic_id ); } +void oldest_scheduler::order_warps() +{ + order_by_priority( m_next_cycle_prioritized_warps, + m_supervised_warps, + m_last_supervised_issued, + m_supervised_warps.size(), + ORDERED_PRIORITY_FUNC_ONLY, + scheduler_unit::sort_warps_by_oldest_dynamic_id ); +} + void two_level_active_scheduler::do_on_warp_issued( unsigned warp_id, unsigned num_issued, @@ -1117,12 +1224,13 @@ 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* tensor_core_out, register_set* mem_out, int id, char* config_string ) - : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, sfu_out, tensor_core_out, mem_out, id ) + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ) { unsigned m_prioritization_readin; int ret = sscanf( config_string, @@ -1386,8 +1494,14 @@ ldst_unit::process_cache_access( cache_t* cache, mem_stage_stall_type result = NO_RC_FAIL; bool write_sent = was_write_sent(events); bool read_sent = was_read_sent(events); - if( write_sent ) - m_core->inc_store_req( inst.warp_id() ); + if( write_sent ) { + 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() ); + + } if ( status == HIT ) { assert( !read_sent ); inst.accessq_pop_back(); @@ -1399,7 +1513,7 @@ ldst_unit::process_cache_access( cache_t* cache, if( !write_sent ) delete mf; } else if ( status == RESERVATION_FAIL ) { - result = COAL_STALL; + result = BK_CONF; assert( !read_sent ); assert( !write_sent ); delete mf; @@ -1408,8 +1522,8 @@ ldst_unit::process_cache_access( cache_t* cache, //inst.clear_active( access.get_warp_mask() ); // threads in mf writeback when mf returns inst.accessq_pop_back(); } - if( !inst.accessq_empty() ) - result = BK_CONF; + if( !inst.accessq_empty() && result == NO_RC_FAIL) + result = COAL_STALL; return result; } @@ -1429,6 +1543,111 @@ 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)) ) @@ -1478,7 +1697,7 @@ 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; } if( bypassL1D ) { @@ -1502,9 +1721,9 @@ 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() ) + if( !inst.accessq_empty() && stall_cond == NO_RC_FAIL) stall_cond = COAL_STALL; if (stall_cond != NO_RC_FAIL) { stall_reason = stall_cond; @@ -1534,6 +1753,11 @@ void ldst_unit::flush(){ m_L1D->flush(); } +void ldst_unit::invalidate(){ + // Flush L1D cache + m_L1D->invalidate(); +} + simd_function_unit::simd_function_unit( const shader_core_config *config ) { m_config=config; @@ -1586,6 +1810,13 @@ 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(); @@ -1610,6 +1841,12 @@ 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(); @@ -1619,6 +1856,14 @@ 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) @@ -1712,8 +1957,9 @@ 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, @@ -1734,6 +1980,12 @@ 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); + } } } @@ -1916,12 +2168,12 @@ void ldst_unit::cycle() if( !m_response_fifo.empty() ) { mem_fetch *mf = m_response_fifo.front(); - if (mf->istexture()) { + if (mf->get_access_type() == TEXTURE_ACC_R) { if (m_L1T->fill_port_free()) { m_L1T->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle); m_response_fifo.pop_front(); } - } else if (mf->isconst()) { + } else if (mf->get_access_type() == CONST_ACC_R) { if (m_L1C->fill_port_free()) { mf->set_status(IN_SHADER_FETCHED,gpu_sim_cycle+gpu_tot_sim_cycle); m_L1C->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle); @@ -1960,7 +2212,11 @@ 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; @@ -1983,9 +2239,9 @@ 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 { @@ -2562,9 +2818,42 @@ unsigned int shader_core_config::max_cta( const kernel_info_t &k ) const assert( result <= MAX_CTA_PER_SHADER ); if (result < 1) { printf ("GPGPU-Sim uArch: ERROR ** Kernel requires more resources than shader has.\n"); + if(gpgpu_ignore_resources_limitation) { + printf ("GPGPU-Sim uArch: gpgpu_ignore_resources_limitation is set, ignore the ERROR!\n"); + return 1; + } abort(); } + 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; } @@ -2586,6 +2875,11 @@ void shader_core_ctx::cache_flush() m_ldst_unit->flush(); } +void shader_core_ctx::cache_invalidate() +{ + m_ldst_unit->invalidate(); +} + // modifiers std::list opndcoll_rfu_t::arbiter_t::allocate_reads() { @@ -3340,6 +3634,15 @@ void simt_core_cluster::print_not_completed( FILE *fp ) const } } + +float simt_core_cluster::get_current_occupancy( unsigned long long& active, unsigned long long& total ) const { + float aggregate = 0.f; + for( unsigned i=0; i < m_config->n_simt_cores_per_cluster; i++ ) { + aggregate+=m_core[i]->get_current_occupancy( active, total ); + } + return aggregate / m_config->n_simt_cores_per_cluster; +} + unsigned simt_core_cluster::get_n_active_cta() const { unsigned n=0; @@ -3403,6 +3706,12 @@ void simt_core_cluster::cache_flush() m_core[i]->cache_flush(); } +void simt_core_cluster::cache_invalidate() +{ + for( unsigned i=0; i < m_config->n_simt_cores_per_cluster; i++ ) + m_core[i]->cache_invalidate(); +} + bool simt_core_cluster::icnt_injection_buffer_full(unsigned size, bool write) { unsigned request_size = size; diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 97e438f..437506c 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -70,6 +70,14 @@ #define WRITE_MASK_SIZE 8 +enum exec_unit_type_t +{ + NONE = 0, + SP = 1, + SFU = 2, + MEM = 3, + DP = 4 +}; class thread_ctx_t { public: @@ -308,6 +316,7 @@ enum concrete_scheduler CONCRETE_SCHEDULER_GTO, CONCRETE_SCHEDULER_TWO_LEVEL_ACTIVE, CONCRETE_SCHEDULER_WARP_LIMITING, + CONCRETE_SCHEDULER_OLDEST_FIRST, NUM_CONCRETE_SCHEDULERS }; @@ -317,13 +326,14 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_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_tensor_core_out(tensor_core_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_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 +405,7 @@ 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_tensor_core_out; register_set* m_mem_out; @@ -408,11 +419,12 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_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, tensor_core_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){} virtual ~lrr_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -426,11 +438,12 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_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, tensor_core_out, mem_out, id ){} + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){} virtual ~gto_scheduler () {} virtual void order_warps (); virtual void done_adding_supervised_warps() { @@ -439,6 +452,25 @@ public: }; +class oldest_scheduler : public scheduler_unit { +public: + oldest_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* tensor_core_out, + register_set* mem_out, + int id ) + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ){} + virtual ~oldest_scheduler () {} + virtual void order_warps (); + virtual void done_adding_supervised_warps() { + m_last_supervised_issued = m_supervised_warps.begin(); + } + +}; class two_level_active_scheduler : public scheduler_unit { public: @@ -446,12 +478,13 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_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, tensor_core_out, mem_out, id ), + : scheduler_unit ( stats, shader, scoreboard, simt, warp, sp_out, dp_out, sfu_out,tensor_core_out, mem_out, id ), m_pending_warps() { unsigned inner_level_readin; @@ -497,6 +530,7 @@ public: Scoreboard* scoreboard, simt_stack** simt, std::vector* warp, register_set* sp_out, + register_set* dp_out, register_set* sfu_out, register_set* tensor_core_out, register_set* mem_out, @@ -1060,6 +1094,23 @@ public: switch(inst.op) { case SFU_OP: break; case ALU_SFU_OP: break; + case DP_OP: break; //for compute <= 29 (i..e Fermi and GT200) + default: return false; + } + return pipelined_simd_unit::can_issue(inst); + } + virtual void active_lanes_in_pipeline(); + 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); @@ -1098,6 +1149,7 @@ public: case STORE_OP: return false; case TENSOR_CORE_STORE_OP: return false; case MEMORY_BARRIER_OP: return false; + case DP_OP: return false; default: break; } return pipelined_simd_unit::can_issue(inst); @@ -1129,6 +1181,7 @@ public: void fill( mem_fetch *mf ); void flush(); + void invalidate(); void writeback(); // accessors @@ -1195,6 +1248,7 @@ 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; @@ -1223,13 +1277,18 @@ 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 { 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, @@ -1240,9 +1299,11 @@ 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", @@ -1340,27 +1401,29 @@ struct shader_core_config : public core_config mutable cache_config m_L1C_config; mutable l1d_cache_config m_L1D_config; - bool gmem_skip_L1D; // on = global memory access always skip the L1 cache - bool gpgpu_dwf_reg_bankconflict; int gpgpu_num_sched_per_core; int gpgpu_max_insn_issue_per_warp; + bool gpgpu_dual_issue_diff_exec_units; //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_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_dp; 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_dp; 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; @@ -1368,6 +1431,7 @@ struct shader_core_config : public core_config int gpgpu_num_sp_units; int gpgpu_tensor_core_avail; + int gpgpu_num_dp_units; int gpgpu_num_sfu_units; int gpgpu_num_tensor_core_units; int gpgpu_num_mem_units; @@ -1379,6 +1443,7 @@ struct shader_core_config : public core_config unsigned gpgpu_num_reg_banks; bool gpgpu_reg_bank_use_warp_id; bool gpgpu_local_mem_map; + bool gpgpu_ignore_resources_limitation; unsigned max_sp_latency; unsigned max_sfu_latency; @@ -1391,10 +1456,14 @@ 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 bool gpgpu_concurrent_kernel_sm; + + bool adpative_volta_cache_config; }; struct shader_core_stats_pod { @@ -1457,6 +1526,8 @@ struct shader_core_stats_pod { unsigned *last_shader_cycle_distro; unsigned *num_warps_issuable; unsigned gpgpu_n_stall_shd_mem; + unsigned* single_issue_nums; + unsigned* dual_issue_nums; //memory access classification int gpgpu_n_mem_read_local; @@ -1528,6 +1599,8 @@ public: m_n_diverge = (unsigned*) calloc(config->num_shader(),sizeof(unsigned)); shader_cycle_distro = (unsigned*) calloc(config->warp_size+3, sizeof(unsigned)); last_shader_cycle_distro = (unsigned*) calloc(m_config->warp_size+3, sizeof(unsigned)); + single_issue_nums = (unsigned*) calloc(config->gpgpu_num_sched_per_core,sizeof(unsigned)); + dual_issue_nums = (unsigned*) calloc(config->gpgpu_num_sched_per_core, sizeof(unsigned)); n_simt_to_mem = (long *)calloc(config->num_shader(), sizeof(long)); n_mem_to_simt = (long *)calloc(config->num_shader(), sizeof(long)); @@ -1651,6 +1724,7 @@ public: void issue_block2core( class kernel_info_t &kernel ); void cache_flush(); + void cache_invalidate(); void accept_fetch_response( mem_fetch *mf ); void accept_ldst_unit_response( class mem_fetch * mf ); void broadcast_barrier_reduction(unsigned cta_id, unsigned bar_id,warp_set_t warps); @@ -1679,6 +1753,7 @@ public: // accessors virtual bool warp_waiting_at_barrier( unsigned warp_id ) const; void get_pdom_stack_top_info( unsigned tid, unsigned *pc, unsigned *rpc ) const; + float get_current_occupancy( unsigned long long & active, unsigned long long & total ) const; // used by pipeline timing model components: // modifiers @@ -1885,10 +1960,14 @@ public: std::vector m_pipeline_reg; Scoreboard *m_scoreboard; opndcoll_rfu_t m_operand_collector; + int m_active_warps; //schedule std::vector schedulers; + //issue + unsigned int Issue_Prio; + // execute unsigned m_num_function_units; std::vector m_dispatch_port; @@ -1940,6 +2019,7 @@ public: void reinit(); unsigned issue_block2core(); void cache_flush(); + void cache_invalidate(); bool icnt_injection_buffer_full(unsigned size, bool write); void icnt_inject_request_packet(class mem_fetch *mf); @@ -1970,6 +2050,7 @@ public: void get_L1T_sub_stats(struct cache_sub_stats &css) const; void get_icnt_stats(long &n_simt_to_mem, long &n_mem_to_simt) const; + float get_current_occupancy( unsigned long long& active, unsigned long long & total ) const; private: unsigned m_cluster_id; diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index ede9f20..52e2f5e 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -172,6 +172,7 @@ void *gpgpu_sim_thread_concurrent(void*) g_sim_active = false; pthread_mutex_unlock(&g_sim_lock); } while( !g_sim_done ); + printf("GPGPU-Sim: *** simulation thread exiting ***\n"); fflush(stdout); diff --git a/src/gpuwattch/makefile b/src/gpuwattch/makefile index ab718cc..354c9ec 100644 --- a/src/gpuwattch/makefile +++ b/src/gpuwattch/makefile @@ -11,10 +11,10 @@ opt: $(TAR).mk obj_opt @$(MAKE) TAG=opt -C . -f $(TAR).mk obj_dbg: - mkdir $@ + mkdir -p $@ obj_opt: - mkdir $@ + mkdir -p $@ depend: @$(MAKE) TAG=opt -C . -f $(TAR).mk depend diff --git a/src/intersim2/Makefile b/src/intersim2/Makefile index bd42000..7d10b3f 100644 --- a/src/intersim2/Makefile +++ b/src/intersim2/Makefile @@ -36,7 +36,7 @@ DEBUG ?= 0 LEX = flex YACC = bison -y DEFINE = #-DTRACK_STALLS -DTRACK_BUFFERS -DTRACK_FLOWS -DTRACK_CREDITS -INCPATH = -I. -Iarbiters -Iallocators -Irouters -Inetworks -Ipower +INCPATH = -I. -Iarbiters -Iallocators -Irouters -Inetworks -Ipower -I$(GPGPUSIM_ROOT)/src ifeq ($(CREATE_LIBRARY),1) INCPATH += -I$(GPGPUSIM_ROOT)/src/gpgpu-sim/ diff --git a/src/intersim2/interconnect_interface.cpp b/src/intersim2/interconnect_interface.cpp index 4386821..1e1a2d7 100644 --- a/src/intersim2/interconnect_interface.cpp +++ b/src/intersim2/interconnect_interface.cpp @@ -44,6 +44,7 @@ #include "booksim.hpp" #include "intersim_config.hpp" #include "network.hpp" +#include "trace.h" InterconnectInterface* InterconnectInterface::New(const char* const config_file) { @@ -147,6 +148,8 @@ void InterconnectInterface::Push(unsigned input_deviceID, unsigned output_device { // it should have free buffer assert(HasBuffer(input_deviceID, size)); + + DPRINTF(INTERCONNECT, "Sent %d bytes from %d to %d", size, input_deviceID, output_deviceID); int output_icntID = _node_map[output_deviceID]; int input_icntID = _node_map[input_deviceID]; @@ -178,7 +181,11 @@ void InterconnectInterface::Push(unsigned input_deviceID, unsigned output_device case WRITE_REQUEST: packet_type = Flit::WRITE_REQUEST ;break; case READ_REPLY: packet_type = Flit::READ_REPLY ;break; case WRITE_ACK: packet_type = Flit::WRITE_REPLY ;break; - default: assert (0); + default: + { + cout<<"Type "<get_type()<<" is undefined!"< Date: Sun, 4 Nov 2018 15:08:57 -0800 Subject: bug in the wmma_ld for load-c in fp16 mode --- src/cuda-sim/instructions.cc | 49 +++++++++++++++++++++++++++++++++++--------- src/gpgpu-sim/shader.cc | 2 +- 2 files changed, 40 insertions(+), 11 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 973eeab..4109c5f 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -1722,14 +1722,19 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) ptx_reg_t src_data; ptx_thread_info *thread; int stride; + unsigned wmma_type = pI->get_wmma_type(); unsigned a_layout = pI->get_wmma_layout(0); unsigned b_layout = pI->get_wmma_layout(1); unsigned type = pI->get_type(); unsigned type2 = pI->get_type2(); - int tid = inst.warp_id() * core->get_warp_size(); + int tid ; const operand_info &dst = pI->operand_lookup(0); + if(core->get_gpu()->is_functional_sim()) + tid= inst.warp_id_func()*core->get_warp_size(); + else + tid= inst.warp_id()*core->get_warp_size(); unsigned thread_group_index; float temp; half temp2; @@ -2963,11 +2968,17 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) const operand_info &src = pI->operand_lookup(1); const operand_info &src1 = pI->operand_lookup(0); const operand_info &src2 = pI->operand_lookup(2); - int tid = inst.warp_id()*core->get_warp_size(); + int tid ; unsigned type = pI->get_type(); unsigned wmma_type = pI->get_wmma_type(); unsigned wmma_layout = pI->get_wmma_layout(0); int stride; + + if(core->get_gpu()->is_functional_sim()) + tid= inst.warp_id_func()*core->get_warp_size(); + else + tid= inst.warp_id()*core->get_warp_size(); + _memory_op_t insn_memory_op = pI->has_memory_read() ? memory_load : memory_store; for (thrd=0; thrd < core->get_warp_size(); thrd++) { thread = core->get_thread_info()[tid+thrd]; @@ -3034,14 +3045,15 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) //mem->write(new_addr+k*2,size/8,&nw_v[k].s64,thread,pI); push_addr=new_addr+k*2; mem->write(push_addr,size/8,&nw_v[k].s64,thread,pI); + if(k%2==0) + mem_txn_addr[num_mem_txn++]=push_addr; } else if(wmma_layout==COL){ //mem->write(new_addr+k*2*stride,size/8,&nw_v[k].s64,thread,pI); push_addr=new_addr+k*2*stride; mem->write(push_addr,size/8,&nw_v[k].s64,thread,pI); - } - if(k%2==0) mem_txn_addr[num_mem_txn++]=push_addr; + } if(debug_tensorcore) printf("wmma:store:thread%d=%x,%x,%x,%x,%x,%x,%x,%x\n",thrd,nw_v[0].s64,nw_v[1].s64,nw_v[2].s64,nw_v[3].s64,nw_v[4].s64,nw_v[5].s64,nw_v[6].s64,nw_v[7].s64); @@ -3051,7 +3063,12 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) delete [] v; inst.space = space; inst.set_addr(thrd, (new_addr_type *)mem_txn_addr , num_mem_txn); - inst.data_size = 4; // 4 byte transaction + + if((type==F16_TYPE)&&(wmma_layout==COL))//check the profiling xls for details + inst.data_size = 2; // 2 byte transaction + else + inst.data_size = 4; // 4 byte transaction + assert( inst.memory_op == insn_memory_op ); //thread->m_last_effective_address = addr; //thread->m_last_memory_space = space; @@ -3070,9 +3087,16 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) unsigned type = pI->get_type(); unsigned wmma_type = pI->get_wmma_type(); unsigned wmma_layout = pI->get_wmma_layout(0); - int tid = inst.warp_id()*core->get_warp_size(); + int tid; int thrd,stride; ptx_thread_info *thread; + + + if(core->get_gpu()->is_functional_sim()) + tid= inst.warp_id_func()*core->get_warp_size(); + else + tid= inst.warp_id()*core->get_warp_size(); + _memory_op_t insn_memory_op = pI->has_memory_read() ? memory_load : memory_store; for (thrd=0; thrd < core->get_warp_size(); thrd++){ @@ -3151,18 +3175,19 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) //mem->read(new_addr+2*i,size/8,&data[i].s64); fetch_addr=new_addr+2*i; mem->read(fetch_addr,size/8,&data[i].s64); + if(i%2==0) + mem_txn_addr[num_mem_txn++]=fetch_addr; } else if(wmma_layout==COL){ //mem->read(new_addr+2*stride*i,size/8,&data[i].s64); fetch_addr=new_addr+2*stride*i; mem->read(fetch_addr,size/8,&data[i].s64); + mem_txn_addr[num_mem_txn++]=fetch_addr; } else{ printf("mma_ld:wrong_type\n"); abort(); } - if(i%2==0) - mem_txn_addr[num_mem_txn++]=fetch_addr; } else if(type==F32_TYPE){ //mem->read(new_addr+4*acc_float_offset(i,wmma_layout,stride),size/8,&data[i].s64); @@ -3183,7 +3208,11 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) //generate timing memory request inst.space = space; inst.set_addr(thrd, (new_addr_type *)mem_txn_addr , num_mem_txn); - inst.data_size = 4; // 4 byte transaction + + if((wmma_type==LOAD_C)&&(type==F16_TYPE)&&(wmma_layout==COL))//memory address is scattered, check the profiling xls for more detail. + inst.data_size = 2; // 2 byte transaction + else + inst.data_size = 4; // 4 byte transaction assert( inst.memory_op == insn_memory_op ); if(debug_tensorcore){ @@ -3197,7 +3226,7 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst ) printf("\nmma_ld:thread%d= ",thrd); float temp; for(i=0;i<16;i++){ - temp=data[i].f16; + temp=data[i].f16; printf("%.2f ",temp); } printf("\n"); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 82f9181..879fea7 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -238,7 +238,7 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu, } //op collector configuration - enum { SP_CUS, DP_CUS, SFU_CUS, MEM_CUS, TENSOR_CORE_CUS, GEN_CUS }; + enum { SP_CUS, DP_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(DP_CUS, m_config->gpgpu_operand_collector_num_units_dp, 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); -- cgit v1.3 From b150969498792d50583674947d7c240cd6a11a68 Mon Sep 17 00:00:00 2001 From: aamir Date: Mon, 5 Nov 2018 16:50:52 -0800 Subject: addressed the deadlock issue --- CHANGES | 16 ++++++++++++++-- README | 4 ++++ configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 6 +++--- src/gpgpu-sim/shader.cc | 6 +++--- src/gpgpu-sim/shader.h | 2 +- 5 files changed, 25 insertions(+), 9 deletions(-) (limited to 'src/gpgpu-sim') diff --git a/CHANGES b/CHANGES index 60be885..a208986 100644 --- a/CHANGES +++ b/CHANGES @@ -3,11 +3,20 @@ Version 4.0.0 (development branch) versus 3.2.3 -Front-End: 1- Support .nc cache modifier and __ldg function to access the read-only L1D cache 2- Partially-support some SASS_60 in the PTXP_PLUS (not completed yet) --GPU Core: +3- Added parsing support for wmma.load,wmma.mma and wmma.store ptx instructions +4- Implmented cudaLaunchKernel for CUTLASS library +-GPU Core Functional Simulation +1- Implemented bfi and prmt instruction +2- Implemented wmma.load and wmma.store supporting all the layout configuration for TITANV GPU +3- Implemented wmma.mma instructions supporting all of its 32 configuration for TITANV GPU +4- debug support for wmma instruction using debug_tensorcore flag +-GPU Core Performance Simulation: 1- Fermi/Pascal coalescer: coalescing on 32-bytes sectors. 2- Adding separate dp unit pipeline. 3- diff dual issue: allow scheduler to issue diff insts at a time 4- Fair memory issue from multiple schedulers. +5- Added tensorcore unit pipeline. +6- Corrected the bug in vector load instruction -Cache System: 1- Sector L1/L2 cache 2- Fetch-on-write and lazy-fetch-on-read write allocation policy. @@ -23,12 +32,15 @@ Version 4.0.0 (development branch) versus 3.2.3 1- Adding more detailed cache statistics to define and analyze cache bottlenecks. 2- Adding more detailed memory statistics (BLP, RBL, etc) to define and analyze memory bottlenecks. 3- Addig new system stats: gpu occupancy, L2BW, etc +-Library: +1 Enabled CUTLASS Library on GPGPU-Sim +-Regression: +1- Added TensorCore Regression Kernel -Configs: Adding the Pascal and Volta config files that has been correlated against real hardware. See the correlation website here: https://engineering.purdue.edu/tgrogers/group/correlator.html - Version 3.2.3+edits (development branch) versus 3.2.3 - Support for running regression tests using Travis - Support added for CUDA dynamic parallelism (courtesy of Jin Wang from Georgia Tech) diff --git a/README b/README index 7eaae5d..e0692e8 100644 --- a/README +++ b/README @@ -24,6 +24,10 @@ Mahmoud Khairy, Jain Akshay, Tor Aamodt, Timothy G Rogers, Exploring Modern GPU Memory System Design Challenges through Accurate Modeling, arXiv:1810.07269, https://arxiv.org/abs/1810.07269 +If you use the tensorcore in GPGPU-Sim or CUTLASS Library in your research +please cite: + add the arxiv link here + If you use the GPUWattch energy model in your research, please cite: Jingwen Leng, Tayler Hetherington, Ahmed ElTantawy, Syed Gilani, Nam Sung Kim, diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index c8351da..7532c01 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -12,6 +12,7 @@ -gpgpu_ptx_instruction_classification 0 -gpgpu_ptx_sim_mode 0 -gpgpu_ptx_force_max_capability 70 +-gpgpu_tensor_core_avail 1 # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 @@ -41,14 +42,13 @@ -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 +# 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,ID_OC_TENSOR_CORE,OC_EX_TENSOR_CORE ## Volta GV100 has 2 SP SIMD units, 2 SFU units, 2 DP units per core ## we need to scale the number of pipeline registers to be equal to the number of SP units --gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12 +-gpgpu_pipeline_widths 4,4,4,1,4,4,4,1,12,1,1 -gpgpu_num_sp_units 4 -gpgpu_num_sfu_units 4 -gpgpu_num_dp_units 4 - # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" # All Div operations are executed on SFU unit diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 0c0707c..cd2ed26 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1179,7 +1179,7 @@ void two_level_active_scheduler::order_warps() for ( std::vector< shd_warp_t* >::iterator iter = m_next_cycle_prioritized_warps.begin(); iter != m_next_cycle_prioritized_warps.end(); ) { bool waiting = (*iter)->waiting(); - for (int i=0; i<4; i++){ + for (int i=0; iibuffer_next_inst(); //Is the instruction waiting on a long operation? if ( inst && inst->in[i] > 0 && this->m_scoreboard->islongop((*iter)->get_warp_id(), inst->in[i])){ @@ -1599,7 +1599,7 @@ void ldst_unit::L1_latency_queue_cycle() assert( !read_sent ); l1_latency_queue[0] = NULL; if ( mf_next->get_inst().is_load() ) { - for ( unsigned r=0; r < 4; r++) + for ( unsigned r=0; r < MAX_OUTPUT_VALUES; 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); @@ -2146,7 +2146,7 @@ void ldst_unit::issue( register_set ®_set ) if (inst->is_load() and inst->space.get_type() != shared_space) { unsigned warp_id = inst->warp_id(); unsigned n_accesses = inst->accessq_count(); - for (unsigned r = 0; r < 4; r++) { + for (unsigned r = 0; r < MAX_OUTPUT_VALUES; r++) { unsigned reg_id = inst->out[r]; if (reg_id > 0) { m_pending_writes[warp_id][reg_id] += n_accesses; diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 437506c..3542120 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1360,7 +1360,7 @@ struct shader_core_config : public core_config max_sp_latency = 32; max_tensor_core_latency = 64; - gpgpu_num_tensor_core_units=8; + gpgpu_num_tensor_core_units=4;//It will be (#TENSORCORE INSIDE SM)/2 (One warp is allocated to 2 Tensor Core) gpgpu_operand_collector_num_units_tensor_core=24; gpgpu_operand_collector_num_in_ports_tensor_core=8; gpgpu_operand_collector_num_out_ports_tensor_core=8; -- cgit v1.3