From 6c1fb702e17b00fd7de72ac7dd4a31584d5978b9 Mon Sep 17 00:00:00 2001 From: sspenst Date: Fri, 8 Jul 2016 14:51:44 -0700 Subject: Made gridDim and blockDim global variables so that they can be accessed from sst_impl --- libcuda/cuda_runtime_api.cc | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index e8a0e91..3eff4af 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -180,6 +180,9 @@ cudaError_t g_last_cudaError = cudaSuccess; extern stream_manager *g_stream_manager; +dim3 gridDim = (dim3){1,1,1}; +dim3 blockDim = (dim3){1,1,1}; + void register_ptx_function( const char *name, function_info *impl ) { // no longer need this @@ -959,8 +962,8 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); kernel_info_t *grid = gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); std::string kname = grid->name(); - dim3 gridDim = config.grid_dim(); - dim3 blockDim = config.block_dim(); + gridDim = config.grid_dim(); + blockDim = config.block_dim(); printf("GPGPU-Sim PTX: pushing kernel \'%s\' to stream %u, gridDim= (%u,%u,%u) blockDim = (%u,%u,%u) \n", kname.c_str(), stream?stream->get_uid():0, gridDim.x,gridDim.y,gridDim.z,blockDim.x,blockDim.y,blockDim.z ); stream_operation op(grid,g_ptx_sim_mode,stream); -- cgit v1.3 From adc311951d67b0685ebf2fab4ce6410f96f0039a Mon Sep 17 00:00:00 2001 From: sspenst Date: Mon, 11 Jul 2016 11:16:46 -0700 Subject: Reverted the previous commit to add a cleaner way of getting NUM_THREADS. Now, sst_impl doesn't functionally execute on the last indexed element of an array, but instead on the actual last thread that executes --- libcuda/cuda_runtime_api.cc | 7 ++----- src/abstract_hardware_model.h | 3 --- src/cuda-sim/instructions.cc | 13 ++++++++----- src/cuda-sim/ptx_sim.cc | 16 ++++++++++++++++ src/cuda-sim/ptx_sim.h | 4 ++++ 5 files changed, 30 insertions(+), 13 deletions(-) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 3eff4af..e8a0e91 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -180,9 +180,6 @@ cudaError_t g_last_cudaError = cudaSuccess; extern stream_manager *g_stream_manager; -dim3 gridDim = (dim3){1,1,1}; -dim3 blockDim = (dim3){1,1,1}; - void register_ptx_function( const char *name, function_info *impl ) { // no longer need this @@ -962,8 +959,8 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); kernel_info_t *grid = gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); std::string kname = grid->name(); - gridDim = config.grid_dim(); - blockDim = config.block_dim(); + dim3 gridDim = config.grid_dim(); + dim3 blockDim = config.block_dim(); printf("GPGPU-Sim PTX: pushing kernel \'%s\' to stream %u, gridDim= (%u,%u,%u) blockDim = (%u,%u,%u) \n", kname.c_str(), stream?stream->get_uid():0, gridDim.x,gridDim.y,gridDim.z,blockDim.x,blockDim.y,blockDim.z ); stream_operation op(grid,g_ptx_sim_mode,stream); diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 46c3279..6ed9b8e 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -162,9 +162,6 @@ struct dim3 { }; #endif -extern dim3 gridDim; -extern dim3 blockDim; - void increment_x_then_y_then_z( dim3 &i, const dim3 &bound); class kernel_info_t { diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index fd3b1fa..b5a3db4 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -3754,6 +3754,7 @@ void sst_impl( const ptx_instruction *pI, ptx_thread_info *thread ) memory_space_t space = pI->get_space(); memory_space *mem = NULL; addr_t addr = src2_data.u32 * 4; // this assumes sstarr memory starts at address 0 + ptx_cta_info *cta_info = thread->m_cta_info; decode_space(space,thread,src1,mem,addr); @@ -3765,18 +3766,19 @@ void sst_impl( const ptx_instruction *pI, ptx_thread_info *thread ) mem->write(addr,size/8,&src3_data.s64,thread,pI); // sync threads - cpI->set_bar_id(dst_data.u32); + cpI->set_bar_id(16); // use 16 for sst because bar uses an int from 0-15 thread->m_last_effective_address = addr; thread->m_last_memory_space = space; thread->m_last_dram_callback.function = bar_callback; thread->m_last_dram_callback.instruction = cpI; - int NUM_THREADS = blockDim.x * blockDim.y * blockDim.z; - if (src2_data.s64 == NUM_THREADS-1) { - // pick only one thread to load all of the data back from sstarr memory + // the last thread that executes loads all of the data back from sstarr memory + int NUM_THREADS = cta_info->num_threads(); + cta_info->inc_bar_threads(); + if (NUM_THREADS == cta_info->get_bar_threads()) { unsigned offset = 0; - addr -= (NUM_THREADS-1)*4; + addr = 0; ptx_reg_t data; float sstarr_fdata[NUM_THREADS]; signed long long sstarr_ldata[NUM_THREADS]; @@ -3812,6 +3814,7 @@ void sst_impl( const ptx_instruction *pI, ptx_thread_info *thread ) offset++; } + cta_info->reset_bar_threads(); thread->m_last_effective_address = addr+(NUM_THREADS-1)*4; thread->m_last_memory_space = space; } diff --git a/src/cuda-sim/ptx_sim.cc b/src/cuda-sim/ptx_sim.cc index 511e8d6..f48115b 100644 --- a/src/cuda-sim/ptx_sim.cc +++ b/src/cuda-sim/ptx_sim.cc @@ -44,6 +44,7 @@ ptx_cta_info::ptx_cta_info( unsigned sm_idx ) m_sm_idx = sm_idx; m_uid = g_ptx_cta_info_uid++; + m_bar_threads = 0; } void ptx_cta_info::add_thread( ptx_thread_info *thd ) @@ -128,6 +129,21 @@ unsigned ptx_cta_info::get_sm_idx() const return m_sm_idx; } +unsigned ptx_cta_info::get_bar_threads() const +{ + return m_bar_threads; +} + +void ptx_cta_info::inc_bar_threads() +{ + m_bar_threads++; +} + +void ptx_cta_info::reset_bar_threads() +{ + m_bar_threads = 0; +} + unsigned g_ptx_thread_info_uid_next=1; unsigned g_ptx_thread_info_delete_count=0; diff --git a/src/cuda-sim/ptx_sim.h b/src/cuda-sim/ptx_sim.h index c66b68c..4e748e9 100644 --- a/src/cuda-sim/ptx_sim.h +++ b/src/cuda-sim/ptx_sim.h @@ -158,8 +158,12 @@ public: void register_thread_exit( ptx_thread_info *thd ); void register_deleted_thread( ptx_thread_info *thd ); unsigned get_sm_idx() const; + unsigned get_bar_threads() const; + void inc_bar_threads(); + void reset_bar_threads(); private: + unsigned m_bar_threads; unsigned long long m_uid; unsigned m_sm_idx; std::set m_threads_in_cta; -- cgit v1.3 From b3ad8abea43b7d1e8887f57d6e30c5a40cf752a6 Mon Sep 17 00:00:00 2001 From: aamir Date: Sat, 21 Jul 2018 19:21:30 -0700 Subject: merging the changes of cutlass on negar tensorcore branch --- .gitignore | 6 +- cutlass-example | 1 + libcuda/cuda_runtime_api.cc | 42 +++++++++++++- src/abstract_hardware_model.h | 2 +- src/cuda-sim/cuda-sim.cc | 7 +++ src/cuda-sim/instructions.cc | 125 ++++++++++++++++++++++++++++++++---------- 6 files changed, 151 insertions(+), 32 deletions(-) create mode 160000 cutlass-example (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/.gitignore b/.gitignore index 53fadb5..b214862 100644 --- a/.gitignore +++ b/.gitignore @@ -1,3 +1,7 @@ +*~ +*.swp +*.swo + src/intersim2/lex.yy.c src/intersim2/y.tab.c src/intersim2/y.tab.h @@ -26,4 +30,4 @@ cuobjdump_to_ptxplus/sass_parser.cc cuobjdump_to_ptxplus/sass_parser.hh cuobjdump_to_ptxplus/sass_parser.output -build/* \ No newline at end of file +build/* diff --git a/cutlass-example b/cutlass-example new file mode 160000 index 0000000..e69115d --- /dev/null +++ b/cutlass-example @@ -0,0 +1 @@ +Subproject commit e69115d598bef1e5372c9134322972b2ea450d9f diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index cbe8a11..aa43ffa 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -952,7 +952,9 @@ __host__ cudaError_t CUDARTAPI cudaSetupArgument(const void *arg, size_t size, s gpgpusim_ptx_assert( !g_cuda_launch_stack.empty(), "empty launch stack" ); kernel_config &config = g_cuda_launch_stack.back(); config.set_arg(arg,size,offset); - + #if 0 + printf("cudaSetupArgument:size%d,offset%d,sizeof(Arg[0])=%d)\n ",size,offset,sizeof(arg)); + #endif return g_last_cudaError = cudaSuccess; } @@ -980,6 +982,44 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) return g_last_cudaError = cudaSuccess; } +__host__ cudaError_t CUDARTAPI cudaLaunchKernel ( const char* hostFun, dim3 gridDim, dim3 blockDim, const void** args, size_t sharedMem, cudaStream_t stream ) +{ + 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])); + //gpgpusim_ptx_assert( !g_cuda_launch_stack.empty(), "empty launch stack" ); + kernel_config &config = g_cuda_launch_stack.back(); + config.set_arg(args[0],432,0); + printf("cudaLaunchParameter\n"); + for(int i=0;i<108;i++){ + printf("cudaLaunchParameter:%d:%08x\n",i,*(*((int **)args)+i)); + } + + CUctx_st* context = GPGPUSim_Context(); + char *mode = getenv("PTX_SIM_MODE_FUNC"); + if( mode ) + sscanf(mode,"%u", &g_ptx_sim_mode); + gpgpusim_ptx_assert( !g_cuda_launch_stack.empty(), "empty launch stack" ); + kernel_config config1 = g_cuda_launch_stack.back(); + struct CUstream_st *stream1 = config1.get_stream(); + printf("\nGPGPU-Sim PTX: cudaLaunch for 0x%p (mode=%s) on stream %u\n", hostFun, + g_ptx_sim_mode?"functional simulation":"performance simulation", stream1?stream1->get_uid():0 ); + kernel_info_t *grid = gpgpu_cuda_ptx_sim_init_grid(hostFun,config1.get_args(),config1.grid_dim(),config1.block_dim(),context); + std::string kname = grid->name(); + dim3 gridDim1 = config1.grid_dim(); + 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 ); + 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 e2e116e..7fe5d82 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -386,7 +386,7 @@ protected: std::deque m_stack; }; -#define GLOBAL_HEAP_START 0x703E20000 +#define GLOBAL_HEAP_START 0xC0000000 // start allocating from this address (lower values used for allocating globals in .ptx file) #define SHARED_MEM_SIZE_MAX (64*1024) #define LOCAL_MEM_SIZE_MAX (8*1024) diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 506bc95..8284ad5 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1273,6 +1273,13 @@ 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(lane_id==0){ + printf("EXECUTION_FLOW:LINE_NUM:%d\n",pI->source_line()); + printf("EXECUTION_FLOW:INST:%s\n",pI->get_source()); + } + #endif + set_npc( pc + pI->inst_size() ); diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index cebce1e..7af157f 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -48,7 +48,7 @@ using half_float::half; unsigned ptx_instruction::g_num_ptx_inst_uid=0; -bool g_debug_instruction = 0; +bool g_debug_instruction = 1; const char *g_opcode_string[NUM_OPCODES] = { @@ -239,9 +239,9 @@ ptx_reg_t ptx_thread_info::get_operand_value( const operand_info &op, operand_in } else if ( op.is_sstarr() ) { result.u64 = op.get_symbol()->get_address() + op.get_addr_offset(); } else { - const char *name = op.name().c_str(); - printf("GPGPU-Sim PTX: ERROR ** get_operand_value : unknown memory operand type for %s\n", name ); - abort(); + const char *name = op.name().c_str(); + printf("GPGPU-Sim PTX: ERROR ** get_operand_value : unknown memory operand type for %s\n", name ); + abort(); } } else if ( op.is_literal() ) { @@ -260,10 +260,18 @@ ptx_reg_t ptx_thread_info::get_operand_value( const operand_info &op, operand_in result.u64 = op.get_symbol()->get_address(); } else if ( op.is_function_address() ) { result.u64 = (size_t)op.get_symbol()->get_pc(); - } else { + }else { const char *name = op.name().c_str(); - printf("GPGPU-Sim PTX: ERROR ** get_operand_value : unknown operand type for %s\n", name ); - assert(0); + const symbol *sym2 = op.get_symbol(); + const type_info *type2 = sym2->type(); + const type_info_key &info2 = type2->get_key(); + if ( info2.is_param_kernel() ) { + result.u64 = sym2->get_address()+ op.get_addr_offset(); + } + else{ + printf("GPGPU-Sim PTX: ERROR ** get_operand_value : unknown operand type for %s\n", name ); + assert(0); + } } if(op.get_operand_lohi() == 1) @@ -1551,7 +1559,44 @@ void bfe_impl( const ptx_instruction *pI, ptx_thread_info *thread ) thread->set_operand_value(dst,d, i_type, thread, pI); } -void bfi_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { inst_not_implemented(pI); } +void bfi_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { + int i,max; + ptx_reg_t src1_data, src2_data; + ptx_reg_t src3_data, src4_data, data; + + const operand_info &dst = pI->dst(); //get operand info of sources and destination + const operand_info &src1 = pI->src1(); //use them to determine that they are of type 'register' + const operand_info &src2 = pI->src2(); + const operand_info &src3 = pI->src3(); + const operand_info &src4 = pI->src4(); + + 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); + src3_data = thread->get_operand_value(src3, dst, i_type, thread, 1); + src4_data = thread->get_operand_value(src4, dst, i_type, thread, 1); + + switch ( i_type ) { + case B32_TYPE: + max = 32; + break; + case B64_TYPE: + max = 64; + break; + default: + printf("Execution error: type mismatch with instruction\n"); + assert(0); + break; + } + data=src2_data; + unsigned pos = src3_data.u32 & 0xFF; + unsigned len = src4_data.u32 & 0xFF; + for(i=0;iset_operand_value(dst, data, i_type, thread, pI); +} void bfind_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { inst_not_implemented(pI); } void bra_impl( const ptx_instruction *pI, ptx_thread_info *thread ) @@ -2180,16 +2225,19 @@ ptx_reg_t f2x( ptx_reg_t x, unsigned from_width, unsigned to_width, int to_sign, } } else { switch ( to_width ) { - case 16: //assert(0); break; + case 16: mytemp=half(x.f32); myfloat=mytemp; y.f16 =mytemp; - //y.f16 = half(x.f32); + #if 0 printf("f2x: %f\n",myfloat); + #endif break; case 32: - y.f32=float(x.f16); - + y.f32=float(x.f16); + #if 0 + printf("f2xnew:%x:my%f:%f\n",x.f16,y.f32,half_float::detail::half2float(x.u16)); + #endif break; // handled by f2f case 64: y.f64 = x.f32; @@ -2813,7 +2861,12 @@ void decode_space( memory_space_t &space, ptx_thread_info *thread, const operand space = param_space_kernel; else if( ti.is_param_local() ) { space = param_space_local; - } else { + } + //mov r1, param-label + else if (ti.is_reg() ){ + space = param_space_kernel; + } + else { printf("GPGPU-Sim PTX: ERROR ** cannot resolve .param space for '%s'\n", s->name().c_str() ); abort(); } @@ -2908,6 +2961,7 @@ void ldu_impl( const ptx_instruction *pI, ptx_thread_info *thread ) void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) { size_t size; + unsigned smid; int t; int thrd,odd,inx,k; ptx_thread_info *thread; @@ -2936,7 +2990,12 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) memory_space *mem = NULL; addr_t addr = addr_reg.u32; - + + smid = thread->get_hw_sid(); + if( whichspace(addr) == shared_space ) { + addr= generic_to_shared(smid,addr); + space = shared_space; + } decode_space(space,thread,src1,mem,addr); type_info_key::type_decode(type,size,t); @@ -2961,7 +3020,7 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) int l; printf("thread=%d:",thrd); for(l=0;l<8;l++){ - temp=v[0].f32; + temp=v[l].f32; printf("%f",temp); } printf("\n"); @@ -2988,6 +3047,7 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) { size_t size; int t,i; + unsigned smid; const operand_info &dst = pI->dst(); const operand_info &src1 = pI->src1(); const operand_info &src2 = pI->src2(); @@ -3008,9 +3068,16 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) memory_space *mem = NULL; addr_t addr = src1_data.u32; - decode_space(space,thread,src1,mem,addr); - type_info_key::type_decode(type,size,t); + smid = thread->get_hw_sid(); + if( whichspace(addr) == shared_space ) { + addr= generic_to_shared(smid,addr); + space = shared_space; + } + + decode_space(space,thread,src1,mem,addr); + type_info_key::type_decode(type,size,t); + ptx_reg_t data[16]; if(g_debug_instruction) printf("mma_ld: thrd=%d,addr=%d, fpsize=%d, stride=%d\n",thrd,src1_data.u32,size,src2_data.u32); @@ -3070,13 +3137,13 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) } if(g_debug_instruction){ if(type==F16_TYPE){ - printf("\nthread%d= ",thrd); + printf("\nmma_ld:thread%d= ",thrd); for(i=0;i<16;i++){ printf("%x ",data[i].u64); } printf("\n"); - printf("\nthread%d= ",thrd); + printf("\nmma_ld:thread%d= ",thrd); float temp; for(i=0;i<16;i++){ temp=data[i].f16; @@ -3085,12 +3152,12 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) printf("\n"); } else{ - printf("\nthread%d= ",thrd); + printf("\nmma_ld:thread%d= ",thrd); for(i=0;i<8;i++){ printf("%f ",data[i].f32); } printf("\n"); - printf("\nthread%d= ",thrd); + printf("\nmma_ld:thread%d= ",thrd); for(i=0;i<8;i++){ printf("%x ",data[i].u64); } @@ -3119,15 +3186,15 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) else thread->set_wmma_vector_operand_values(dst,nw_data[0],nw_data[1],nw_data[2],nw_data[3],nw_data[4],nw_data[5],nw_data[6],nw_data[7]); if(g_debug_instruction){ - printf("wmma_load:data[0].s64=%x,data[1].s64=%x,new_data[0].s64=%x\n",data[0].u64,data[1].u64,nw_data[0].u64); - printf("wmma_load:data[2].s64=%x,data[3].s64=%x,new_data[1].s64=%x\n",data[2].u64,data[3].u64,nw_data[1].u64); - printf("wmma_load:data[4].s64=%x,data[5].s64=%x,new_data[2].s64=%x\n",data[4].u64,data[5].u64,nw_data[2].u64); - printf("wmma_load:data[6].s64=%x,data[7].s64=%x,new_data[3].s64=%x\n",data[6].u64,data[7].u64,nw_data[3].u64); + printf("mma_ld:data[0].s64=%x,data[1].s64=%x,new_data[0].s64=%x\n",data[0].u64,data[1].u64,nw_data[0].u64); + printf("mma_ld:data[2].s64=%x,data[3].s64=%x,new_data[1].s64=%x\n",data[2].u64,data[3].u64,nw_data[1].u64); + printf("mma_ld:data[4].s64=%x,data[5].s64=%x,new_data[2].s64=%x\n",data[4].u64,data[5].u64,nw_data[2].u64); + printf("mma_ld:data[6].s64=%x,data[7].s64=%x,new_data[3].s64=%x\n",data[6].u64,data[7].u64,nw_data[3].u64); if(wmma_type!=LOAD_C){ - printf("wmma_load:data[8].s64=%x,data[9].s64=%x,new_data[4].s64=%x\n",data[8].u64,data[9].u64,nw_data[4].s64); - printf("wmma_load:data[10].s64=%x,data[11].s64=%x,new_data[5].s64=%x\n",data[10].u64,data[11].u64,nw_data[5].u64); - printf("wmma_load:data[12].s64=%x,data[13].s64=%x,new_data[6].s64=%x\n",data[12].u64,data[13].u64,nw_data[6].u64); - printf("wmma_load:data[14].s64=%x,data[15].s64=%x,new_data[7].s64=%x\n",data[14].u64,data[15].u64,nw_data[3].u64); + printf("mma_ld:data[8].s64=%x,data[9].s64=%x,new_data[4].s64=%x\n",data[8].u64,data[9].u64,nw_data[4].s64); + printf("mma_ld:data[10].s64=%x,data[11].s64=%x,new_data[5].s64=%x\n",data[10].u64,data[11].u64,nw_data[5].u64); + printf("mma_ld:data[12].s64=%x,data[13].s64=%x,new_data[6].s64=%x\n",data[12].u64,data[13].u64,nw_data[6].u64); + printf("mma_ld:data[14].s64=%x,data[15].s64=%x,new_data[7].s64=%x\n",data[14].u64,data[15].u64,nw_data[3].u64); } } } -- cgit v1.3