diff options
| -rw-r--r-- | debug_tools/WatchYourStep/ptxjitplus/launchkernels | 3 | ||||
| -rw-r--r-- | debug_tools/WatchYourStep/ptxjitplus/ptxjitplus.cpp | 10 | ||||
| -rw-r--r-- | libcuda/cuda_runtime_api.cc | 2 | ||||
| -rw-r--r-- | src/cuda-sim/cuda-sim.cc | 37 | ||||
| -rw-r--r-- | src/cuda-sim/ptx_ir.h | 2 |
5 files changed, 44 insertions, 10 deletions
diff --git a/debug_tools/WatchYourStep/ptxjitplus/launchkernels b/debug_tools/WatchYourStep/ptxjitplus/launchkernels new file mode 100644 index 0000000..d2fd015 --- /dev/null +++ b/debug_tools/WatchYourStep/ptxjitplus/launchkernels @@ -0,0 +1,3 @@ +#Launches kernels from $1 to $2 +#Note must source this script +for num in $(eval echo {$1..$2}); do export WYS_LAUNCH_NUM=$num; echo Launching kernel $num...; ./ptxjitplus; done diff --git a/debug_tools/WatchYourStep/ptxjitplus/ptxjitplus.cpp b/debug_tools/WatchYourStep/ptxjitplus/ptxjitplus.cpp index df0ff8d..1f094ba 100644 --- a/debug_tools/WatchYourStep/ptxjitplus/ptxjitplus.cpp +++ b/debug_tools/WatchYourStep/ptxjitplus/ptxjitplus.cpp @@ -81,18 +81,16 @@ void ptxJIT(int argc, char **argv, CUmodule *phModule, CUfunction *phKernel, CUl if (sizeof(void *)==4) { // Load the PTX from the string myPtx32 - printf("Loading myPtx32[] program\n"); - // PTX May also be loaded from file, as per below. - myErr = cuLinkAddData(*lState, CU_JIT_INPUT_PTX, (void *)myPtx32, strlen(myPtx32)+1, 0, 0, 0, 0); + printf("Loading myPtx32[] program...\n"); + printf("WARNING: 32-bit execution is untested"); } else { // Load the PTX from the string myPtx (64-bit) printf("Loading myPtx[] program\n"); - //myErr = cuLinkAddData(*lState, CU_JIT_INPUT_PTX, (void *)myPtx64, strlen(myPtx64)+1, 0, 0, 0, 0); - // PTX May also be loaded from file, as per below. - myErr = cuLinkAddFile(*lState, CU_JIT_INPUT_PTX, "myPtx64.ptx",0,0,0); } + std::string ptx_file (std::string("../data/ptx.config") + wys_launch_num); + myErr = cuLinkAddFile(*lState, CU_JIT_INPUT_PTX, ptx_file.c_str(),0,0,0); if (myErr != CUDA_SUCCESS) { diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 410f15f..b67ea85 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -2624,7 +2624,7 @@ kernel_info_t *gpgpu_cuda_ptx_sim_init_grid( const char *hostFun, fflush(stdout); if(g_debug_execution >= 3){ - entry->debug_param(g_mallocPtr_Size, result->get_param_memory(), (gpgpu_t *) context->get_device()->get_gpgpu(), gridDim, blockDim); + entry->ptx_jit_config(g_mallocPtr_Size, result->get_param_memory(), (gpgpu_t *) context->get_device()->get_gpgpu(), gridDim, blockDim); } return result; diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 3003c4c..fc4f82a 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1226,7 +1226,7 @@ void function_info::list_param( FILE *fout ) const fflush(fout); } -void function_info::debug_param(std::map<unsigned long long, size_t> mallocPtr_Size, memory_space *param_mem, gpgpu_t* gpu, dim3 gridDim, dim3 blockDim) const +void function_info::ptx_jit_config(std::map<unsigned long long, size_t> mallocPtr_Size, memory_space *param_mem, gpgpu_t* gpu, dim3 gridDim, dim3 blockDim) const { static unsigned long counter = 0; std::vector< std::pair<size_t, unsigned char*> > param_data; @@ -1236,7 +1236,7 @@ void function_info::debug_param(std::map<unsigned long long, size_t> mallocPtr_S assert(gpgpusim_path!=NULL); std::string command = std::string("mkdir ") + gpgpusim_path + "/debug_tools/WatchYourStep/data"; system(command.c_str()); - std::string filename(std::string(gpgpusim_path) + "/debug_tools/WatchYourStep/data/params.config" + std::to_string(counter++)); + std::string filename(std::string(gpgpusim_path) + "/debug_tools/WatchYourStep/data/params.config" + std::to_string(counter)); for( std::map<unsigned,param_info>::const_iterator i=m_ptx_kernel_param_info.begin(); i!=m_ptx_kernel_param_info.end(); i++ ) { const param_info &p = i->second; @@ -1280,6 +1280,39 @@ void function_info::debug_param(std::map<unsigned long long, size_t> mallocPtr_S } fflush(fout); fclose(fout); + + //ptx config + char buff[1024]; + std::string ptx_config_fn(std::string(gpgpusim_path) + "/debug_tools/WatchYourStep/data/ptx.config" + std::to_string(counter)); + snprintf(buff, 1024, "grep -rn \".entry %s\" *.ptx | cut -d \":\" -f 1-2 > %s", get_name().c_str(), ptx_config_fn.c_str()); + system(buff); + FILE *fin = fopen(ptx_config_fn.c_str(), "r"); + char ptx_source[256]; + unsigned line_number; + int numscanned = fscanf(fin, "%[^:]:%u", ptx_source, &line_number); + assert(numscanned == 2); + fclose(fin); + snprintf(buff, 1024, "grep -rn \".version\" %s | cut -d \":\" -f 1 | xargs -I \"{}\" awk \"NR>={}&&NR<={}+2\" %s > %s", ptx_source, ptx_source, ptx_config_fn.c_str()); + system(buff); + fin = fopen(ptx_source, "r"); + assert(fin!=NULL); + fout = fopen(ptx_config_fn.c_str(), "a"); + assert(fout!=NULL); + for (unsigned i = 0; i<line_number; i++){ + fgets(buff, 1024, fin); + assert(!feof(fin)); + } + fprintf(fout, "\n\n"); + do{ + fprintf(fout, "%s", buff); + fgets(buff, 1024, fin); + assert(!feof(fin)); + } while(strstr(buff, "entry")==NULL); + + fclose(fin); + fflush(fout); + fclose(fout); + counter++; } template<int activate_level> diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index b29621c..e0b5e96 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1257,7 +1257,7 @@ public: void finalize( memory_space *param_mem ); void param_to_shared( memory_space *shared_mem, symbol_table *symtab ); void list_param( FILE *fout ) const; - void debug_param(std::map<unsigned long long, size_t> mallocPtr_Size, memory_space *param_mem, gpgpu_t* gpu, dim3 gridDim, dim3 blockDim) const; + void ptx_jit_config(std::map<unsigned long long, size_t> mallocPtr_Size, memory_space *param_mem, gpgpu_t* gpu, dim3 gridDim, dim3 blockDim) const; const struct gpgpu_ptx_sim_info* get_kernel_info () const { |