aboutsummaryrefslogtreecommitdiff
path: root/libcuda/cuda_runtime_api.cc
diff options
context:
space:
mode:
Diffstat (limited to 'libcuda/cuda_runtime_api.cc')
-rw-r--r--libcuda/cuda_runtime_api.cc369
1 files changed, 278 insertions, 91 deletions
diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc
index cbe8a11..eda8d8e 100644
--- a/libcuda/cuda_runtime_api.cc
+++ b/libcuda/cuda_runtime_api.cc
@@ -145,6 +145,10 @@
#include <mach-o/dyld.h>
#endif
+std::map<void *,void **> pinned_memory; //support for pinned memories added
+std::map<void *, size_t> pinned_memory_size;
+int no_of_ptx=0;
+
extern void synchronize();
extern void exit_simulation();
@@ -474,6 +478,8 @@ __host__ cudaError_t CUDARTAPI cudaMallocHost(void **ptr, size_t size)
GPGPUSim_Context();
*ptr = malloc(size);
if ( *ptr ) {
+ //track pinned memory size allocated in the host so that same amount of memory is also allocated in GPU.
+ pinned_memory_size[*ptr]=size;
return g_last_cudaError = cudaSuccess;
} else {
return g_last_cudaError = cudaErrorMemoryAllocation;
@@ -764,6 +770,16 @@ __host__ cudaError_t CUDARTAPI cudaMemset(void *mem, int c, size_t count)
return g_last_cudaError = cudaSuccess;
}
+//memset operation is done but i think its not async?
+__host__ cudaError_t CUDARTAPI cudaMemsetAsync(void *mem, int c, size_t count, cudaStream_t stream=0)
+{
+ printf("GPGPU-Sim PTX: WARNING: Asynchronous memset not supported (%s)\n", __my_func__);
+ CUctx_st *context = GPGPUSim_Context();
+ gpgpu_t *gpu = context->get_device()->get_gpgpu();
+ gpu->gpu_memset((size_t)mem, c, count);
+ return g_last_cudaError = cudaSuccess;
+}
+
__host__ cudaError_t CUDARTAPI cudaMemset2D(void *mem, size_t pitch, int c, size_t width, size_t height)
{
cuda_not_implemented(__my_func__,__LINE__);
@@ -816,6 +832,58 @@ __host__ cudaError_t CUDARTAPI cudaGetDeviceProperties(struct cudaDeviceProp *pr
}
}
+#if (CUDART_VERSION > 5000)
+__host__ cudaError_t CUDARTAPI cudaDeviceGetAttribute(int *value, enum cudaDeviceAttr attr, int device)
+{
+ const struct cudaDeviceProp *prop;
+ _cuda_device_id *dev = GPGPUSim_Init();
+ if (device <= dev->num_devices() ) {
+ prop = dev->get_prop();
+ switch (attr) {
+ case 5:
+ *value= prop->maxGridSize[0];
+ break;
+ case 6:
+ *value= prop->maxGridSize[1];
+ break;
+ case 7:
+ *value= prop->maxGridSize[2];
+ break;
+ case 10:
+ *value= prop->warpSize;
+ break;
+ case 12:
+ *value= prop->regsPerBlock;
+ break;
+ case 14:
+ *value= prop->textureAlignment ;
+ break;
+ case 16:
+ *value= prop->multiProcessorCount ;
+ break;
+ case 39:
+ *value= dev->get_gpgpu()->threads_per_core();
+ break;
+ case 75:
+ *value= 8 ;
+ break;
+ case 76:
+ *value= 3 ;
+ break;
+ case 78:
+ *value= 0 ; //TODO: as of now, we dont support stream priorities.
+ break;
+ default:
+ printf("ERROR: implement the attribute numbered %d \n",attr);
+ abort();
+ }
+ return g_last_cudaError = cudaSuccess;
+ } else {
+ return g_last_cudaError = cudaErrorInvalidDevice;
+ }
+}
+#endif
+
__host__ cudaError_t CUDARTAPI cudaChooseDevice(int *device, const struct cudaDeviceProp *prop)
{
_cuda_device_id *dev = GPGPUSim_Init();
@@ -969,7 +1037,16 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun )
printf("\nGPGPU-Sim PTX: cudaLaunch for 0x%p (mode=%s) on stream %u\n", 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);
+ //do dynamic PDOM analysis for performance simulation scenario
std::string kname = grid->name();
+ function_info *kernel_func_info = grid->entry();
+ if (kernel_func_info->is_pdom_set()) {
+ printf("GPGPU-Sim PTX: PDOM analysis already done for %s \n", kname.c_str() );
+ } else {
+ printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", kname.c_str() );
+ kernel_func_info->do_pdom();
+ kernel_func_info->set_pdom();
+ }
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",
@@ -999,8 +1076,17 @@ __host__ cudaError_t CUDARTAPI cudaStreamCreate(cudaStream_t *stream)
return g_last_cudaError = cudaSuccess;
}
-__host__ __device__ cudaError_t CUDARTAPI cudaStreamCreateWithFlags(cudaStream_t *stream, unsigned int flags) {
- return cudaStreamCreate(stream);
+//TODO: introduce priorities
+__host__ cudaError_t CUDARTAPI cudaStreamCreateWithPriority(cudaStream_t *stream, unsigned int flags, int priority) {
+ return cudaStreamCreate(stream);
+}
+
+__host__ cudaError_t CUDARTAPI cudaDeviceGetStreamPriorityRange(int* leastPriority, int* greatestPriority) {
+ return cudaSuccess;
+}
+
+__host__ __device__ cudaError_t CUDARTAPI cudaStreamCreateWithFlags(cudaStream_t *pStream, unsigned int flags) {
+ return cudaStreamCreate(pStream);
}
__host__ cudaError_t CUDARTAPI cudaStreamDestroy(cudaStream_t stream)
@@ -1341,6 +1427,26 @@ std::string get_app_binary(){
return self_exe_path;
}
+//above func gives abs path whereas this give just the name of application.
+char* get_app_binary_name(std::string abs_path){
+ char *self_exe_path;
+#ifdef __APPLE__
+ //TODO: get apple device and check the result.
+ printf("WARNING: not tested for Apple-mac devices \n");
+ abort();
+#else
+ char* buf = strdup(abs_path.c_str());
+ char *token = strtok(buf, "/");
+ while(token !=NULL){
+ self_exe_path = token;
+ token = strtok(NULL,"/");
+ }
+#endif
+ self_exe_path = strtok(self_exe_path, ".");
+ printf("self exe links to: %s\n", self_exe_path);
+ return self_exe_path;
+}
+
//! Call cuobjdump to extract everything (-elf -sass -ptx)
/*!
* This Function extract the whole PTX (for all the files) using cuobjdump
@@ -1350,105 +1456,154 @@ std::string get_app_binary(){
* enabled
* */
void extract_code_using_cuobjdump(){
- CUctx_st *context = GPGPUSim_Context();
- char command[1000];
+ CUctx_st *context = GPGPUSim_Context();
+ unsigned forced_max_capability = context->get_device()->get_gpgpu()->get_config().get_forced_max_capability();
- std::string app_binary = get_app_binary();
+ //prevent the dumping by cuobjdump everytime we execute the code!
+ const char *override_cuobjdump = getenv("CUOBJDUMP_SIM_FILE");
+ char command[1000], ptx_file[1000];
+ std::string app_binary = get_app_binary();
+ //Running cuobjdump using dynamic link to current process
+ snprintf(command,1000,"md5sum %s ", app_binary.c_str());
+ printf("Running md5sum using \"%s\"\n", command);
+ system(command);
+ // Running cuobjdump using dynamic link to current process
+ // Needs the option '-all' to extract PTX from CDP-enabled binary
+ extern bool g_cdp_enabled;
- char fname[1024];
+ //dump ptx for all individial ptx files into sepearte files which is later used by ptxas.
+ char fname2[1024];
+ snprintf(fname2,1024,"_cuobjdump_list_ptx_XXXXXX");
+ int fd2=mkstemp(fname2);
+ close(fd2);
+ snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -lptx -arch=sm_%u %s > %s", forced_max_capability, app_binary.c_str(), fname2);
+ int result = system(command);
+ if( result != 0 ) {
+ printf("WARNING: Failed to execute cuobjdump to get list of ptx files \n");
+ exit(0);
+ } else {
+ /*
+ as we got list of ptx files, we need to extract one by one into seperate files so that ptxas can understand it.
+ In this way, the duplicate definitions in a single embedded file can be prevented.
+ No of lines in the file is equal to no of ptx fileis available.
+ */
+ FILE *fp = fopen(fname2,"r");
+ if (fp==NULL) {
+ printf("WARNING: cuobjdump file error! Could not open file %s \n", fname2);
+ exit(0);
+ } else {
+ for (char c = getc(fp); c != EOF; c = getc(fp))
+ if (c == '\n')
+ no_of_ptx = no_of_ptx + 1;
+ fclose(fp);
+ }
+ }
+ if(!g_cdp_enabled) {
+ //based on the list above, dump ptx files individually. Format of dumped ptx file is prog_name.unique_no.sm_<>.ptx
+ for (int index=1; index<= no_of_ptx; index++){
+ snprintf(ptx_file, 1000, "%s.%d.sm_%u.ptx", get_app_binary_name(app_binary), index, forced_max_capability);
+ printf("Extracting specific PTX file named %s \n",ptx_file);
+ snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -arch=sm_%u -xptx %s %s", forced_max_capability, ptx_file, app_binary.c_str());
+ if (system(command)!=0) {
+ printf("ERROR: command: %s failed \n",command);
+ exit(0);
+ }
+ }
+ }
+ //TODO: redundant to dump twice. how can it be prevented?
+ //dump only for specific arch
+ char fname[1024];
+ if ((override_cuobjdump == NULL) || (strlen(override_cuobjdump)==0)) {
snprintf(fname,1024,"_cuobjdump_complete_output_XXXXXX");
int fd=mkstemp(fname);
close(fd);
- // Running cuobjdump using dynamic link to current process
- snprintf(command,1000,"md5sum %s ", app_binary.c_str());
- printf("Running md5sum using \"%s\"\n", command);
- system(command);
- // Running cuobjdump using dynamic link to current process
- // Needs the option '-all' to extract PTX from CDP-enabled binary
- extern bool g_cdp_enabled;
if(!g_cdp_enabled)
- snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass %s > %s", app_binary.c_str(), fname);
+ snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -arch=sm_%u %s > %s", forced_max_capability, app_binary.c_str(), fname);
else
- snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -all %s > %s", app_binary.c_str(), fname);
+ snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -all %s > %s", app_binary.c_str(), fname);
bool parse_output = true;
- int result = system(command);
+ result = system(command);
if(result) {
- if (context->get_device()->get_gpgpu()->get_config().experimental_lib_support() && (result == 65280)) {
- // Some CUDA application may exclusively use kernels provided by CUDA
- // libraries (e.g. CUBLAS). Skipping cuobjdump extraction from the
- // executable for this case.
- // 65280 is the return code from cuobjdump denoting the specific error (tested on CUDA 4.0/4.1/4.2)
- printf("WARNING: Failed to execute: %s\n", command);
- printf(" Executable binary does not contain any GPU kernel.\n");
- parse_output = false;
- } else {
- printf("ERROR: Failed to execute: %s\n", command);
- exit(1);
- }
- }
+ if (context->get_device()->get_gpgpu()->get_config().experimental_lib_support() && (result == 65280)) {
+ // Some CUDA application may exclusively use kernels provided by CUDA
+ // libraries (e.g. CUBLAS). Skipping cuobjdump extraction from the
+ // executable for this case.
+ // 65280 is the return code from cuobjdump denoting the specific error (tested on CUDA 4.0/4.1/4.2)
+ printf("WARNING: Failed to execute: %s\n", command);
+ printf(" Executable binary does not contain any GPU kernel.\n");
+ parse_output = false;
+ } else {
+ printf("ERROR: Failed to execute: %s\n", command);
+ exit(1);
+ }
+ }
- if (parse_output) {
- printf("Parsing file %s\n", fname);
- cuobjdump_in = fopen(fname, "r");
+ if (parse_output) {
+ printf("Parsing file %s\n", fname);
+ cuobjdump_in = fopen(fname, "r");
- cuobjdump_parse();
- fclose(cuobjdump_in);
- printf("Done parsing!!!\n");
- } else {
- printf("Parsing skipped for %s\n", fname);
- }
+ cuobjdump_parse();
+ fclose(cuobjdump_in);
+ printf("Done parsing!!!\n");
+ } else {
+ printf("Parsing skipped for %s\n", fname);
+ }
- if (context->get_device()->get_gpgpu()->get_config().experimental_lib_support()){
- //Experimental library support
- //Currently only for cufft
+ if (context->get_device()->get_gpgpu()->get_config().experimental_lib_support()){
+ //Experimental library support
+ //Currently only for cufft
- std::stringstream cmd;
- cmd << "ldd " << app_binary << " | grep $CUDA_INSTALL_PATH | awk \'{print $3}\' > _tempfile_.txt";
- int result = system(cmd.str().c_str());
- if(result){
- std::cout << "Failed to execute: " << cmd << std::endl;
- exit(1);
- }
- std::ifstream libsf;
- libsf.open("_tempfile_.txt");
- if(!libsf.is_open()) {
- std::cout << "Failed to open: _tempfile_.txt" << std::endl;
- exit(1);
- }
+ std::stringstream cmd;
+ cmd << "ldd " << app_binary << " | grep $CUDA_INSTALL_PATH | awk \'{print $3}\' > _tempfile_.txt";
+ int result = system(cmd.str().c_str());
+ if(result){
+ std::cout << "Failed to execute: " << cmd << std::endl;
+ exit(1);
+ }
+ std::ifstream libsf;
+ libsf.open("_tempfile_.txt");
+ if(!libsf.is_open()) {
+ std::cout << "Failed to open: _tempfile_.txt" << std::endl;
+ exit(1);
+ }
- //Save the original section list
- std::list<cuobjdumpSection*> tmpsl = cuobjdumpSectionList;
- cuobjdumpSectionList.clear();
+ //Save the original section list
+ std::list<cuobjdumpSection*> tmpsl = cuobjdumpSectionList;
+ cuobjdumpSectionList.clear();
- std::string line;
- std::getline(libsf, line);
- std::cout << "DOING: " << line << std::endl;
- int cnt=1;
- while(libsf.good()){
- std::stringstream libcodfn;
- libcodfn << "_cuobjdump_complete_lib_" << cnt << "_";
- cmd.str(""); //resetting
- cmd << "$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass ";
- cmd << line;
- cmd << " > ";
- cmd << libcodfn.str();
- std::cout << "Running cuobjdump on " << line << std::endl;
- std::cout << "Using command: " << cmd.str() << std::endl;
- result = system(cmd.str().c_str());
- if(result) {printf("ERROR: Failed to execute: %s\n", command); exit(1);}
- std::cout << "Done" << std::endl;
+ std::string line;
+ std::getline(libsf, line);
+ std::cout << "DOING: " << line << std::endl;
+ int cnt=1;
+ while(libsf.good()){
+ std::stringstream libcodfn;
+ libcodfn << "_cuobjdump_complete_lib_" << cnt << "_";
+ cmd.str(""); //resetting
+ cmd << "$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass ";
+ cmd << line;
+ cmd << " > ";
+ cmd << libcodfn.str();
+ std::cout << "Running cuobjdump on " << line << std::endl;
+ std::cout << "Using command: " << cmd.str() << std::endl;
+ result = system(cmd.str().c_str());
+ if(result) {printf("ERROR: Failed to execute: %s\n", command); exit(1);}
+ std::cout << "Done" << std::endl;
- std::cout << "Trying to parse " << libcodfn << std::endl;
- cuobjdump_in = fopen(libcodfn.str().c_str(), "r");
- cuobjdump_parse();
- fclose(cuobjdump_in);
- std::getline(libsf, line);
- }
- libSectionList = cuobjdumpSectionList;
+ std::cout << "Trying to parse " << libcodfn << std::endl;
+ cuobjdump_in = fopen(libcodfn.str().c_str(), "r");
+ cuobjdump_parse();
+ fclose(cuobjdump_in);
+ std::getline(libsf, line);
+ }
+ libSectionList = cuobjdumpSectionList;
- //Restore the original section list
- cuobjdumpSectionList = tmpsl;
- }
+ //Restore the original section list
+ cuobjdumpSectionList = tmpsl;
+ }
+ } else {
+ printf("GPGPU-Sim PTX: overriding cuobjdump with '%s' (CUOBJDUMP_SIM_FILE is set)\n", override_cuobjdump);
+ snprintf(fname,1024, "%s",override_cuobjdump);
+ }
}
//! Read file into char*
@@ -1680,8 +1835,11 @@ cuobjdumpPTXSection* findPTXSection(const std::string identifier){
void cuobjdumpInit(){
CUctx_st *context = GPGPUSim_Context();
extract_code_using_cuobjdump(); //extract all the output of cuobjdump to _cuobjdump_*.*
- cuobjdumpSectionList = pruneSectionList(cuobjdumpSectionList, context);
- cuobjdumpSectionList = mergeSections(cuobjdumpSectionList);
+ const char* pre_load = getenv("CUOBJDUMP_SIM_FILE");
+ if (pre_load ==NULL || strlen(pre_load)==0){
+ cuobjdumpSectionList = pruneSectionList(cuobjdumpSectionList, context);
+ cuobjdumpSectionList = mergeSections(cuobjdumpSectionList);
+ }
}
std::map<int, std::string> fatbinmap;
@@ -1707,6 +1865,8 @@ void cuobjdumpParseBinary(unsigned int handle){
return;
}
+ //Why to search for capability value if we can directly find it in device info?
+ #if 0
unsigned max_capability = 0;
for ( std::list<cuobjdumpSection*>::iterator iter = cuobjdumpSectionList.begin();
iter != cuobjdumpSectionList.end();
@@ -1715,12 +1875,17 @@ void cuobjdumpParseBinary(unsigned int handle){
if (capability > max_capability) max_capability = capability;
}
if (max_capability > 20) printf("WARNING: No guarantee that PTX will be parsed for SM version %u\n", max_capability);
+ #endif
+ unsigned max_capability = context->get_device()->get_gpgpu()->get_config().get_forced_max_capability();
- cuobjdumpPTXSection* ptx = findPTXSection(fname);
+ cuobjdumpPTXSection* ptx = NULL;
+ const char* pre_load = getenv("CUOBJDUMP_SIM_FILE");
+ if(pre_load==NULL || strlen(pre_load)==0)
+ ptx = findPTXSection(fname);
symbol_table *symtab;
char *ptxcode;
const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE");
- if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL) {
+ if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL or strlen(getenv("PTX_SIM_USE_PTX_FILE"))==0) {
ptxcode = readfile(ptx->getPTXfilename());
} else {
printf("GPGPU-Sim PTX: overriding embedded ptx with '%s' (PTX_SIM_USE_PTX_FILE is set)\n", override_ptx_name);
@@ -1740,7 +1905,8 @@ void cuobjdumpParseBinary(unsigned int handle){
delete[] ptxplus_str;
} else {
symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxcode, handle);
- printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle);
+ //if CUOBJDUMP_SIM_FILE is not set, ptx is NULL. So comment below.
+ //printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle);
context->add_binary(symtab, handle);
gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability );
}
@@ -1857,6 +2023,7 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin )
return (void**)fat_cubin_handle;
}
#endif
+ return 0;
}
void __cudaUnregisterFatBinary(void **fatCubinHandle)
@@ -2072,6 +2239,9 @@ cudaError_t cudaGLUnregisterBufferObject(GLuint bufferObj)
cudaError_t CUDARTAPI cudaHostAlloc(void **pHost, size_t bytes, unsigned int flags)
{
*pHost = malloc(bytes);
+ //need to track the size allocated so that cudaHostGetDevicePointer() can function properly.
+ //TODO: vary this function behavior based on flags value (following nvidia documentation)
+ pinned_memory_size[*pHost]=bytes;
if( *pHost )
return g_last_cudaError = cudaSuccess;
else
@@ -2080,8 +2250,25 @@ cudaError_t CUDARTAPI cudaHostAlloc(void **pHost, size_t bytes, unsigned int fl
cudaError_t CUDARTAPI cudaHostGetDevicePointer(void **pDevice, void *pHost, unsigned int flags)
{
- cuda_not_implemented(__my_func__,__LINE__);
- return g_last_cudaError = cudaErrorUnknown;
+ //only cpu memory allocation happens in cudaHostAlloc. Linking with device pointer to pinned memory happens here.
+ //TODO: once kernel is executed, the contents in global pointer of GPU must be copied back to CPU host pointer!
+ flags=0;
+ CUctx_st* context = GPGPUSim_Context();
+ gpgpu_t *gpu = context->get_device()->get_gpgpu();
+ std::map<void *, size_t>::const_iterator i = pinned_memory_size.find(pHost);
+ assert(i != pinned_memory_size.end());
+ size_t size = i->second;
+ *pDevice = gpu->gpu_malloc(size);
+ if(g_debug_execution >= 3)
+ printf("GPGPU-Sim PTX: cudaMallocing %zu bytes starting at 0x%llx..\n",size, (unsigned long long) *pDevice);
+ if ( *pDevice ) {
+ pinned_memory[pHost]=pDevice;
+ //Copy contents in cpu to gpu
+ gpu->memcpy_to_gpu((size_t)*pDevice,pHost,size);
+ return g_last_cudaError = cudaSuccess;
+ } else {
+ return g_last_cudaError = cudaErrorMemoryAllocation;
+ }
}
cudaError_t CUDARTAPI cudaSetValidDevices(int *device_arr, int len)
generated by cgit v1.3 (git 2.53.0) at 2026-07-19 17:02:41 +0000