|
|
@ -43,16 +43,12 @@ |
|
|
|
|
|
|
|
|
|
|
|
int opt_platform_id = -1; |
|
|
|
int opt_platform_id = -1; |
|
|
|
|
|
|
|
|
|
|
|
int clDevicesNum(void) { |
|
|
|
bool get_opencl_platform(int preferred_platform_id, cl_platform_id *platform) { |
|
|
|
cl_int status; |
|
|
|
cl_int status; |
|
|
|
char pbuff[256]; |
|
|
|
|
|
|
|
cl_uint numDevices; |
|
|
|
|
|
|
|
cl_uint numPlatforms; |
|
|
|
cl_uint numPlatforms; |
|
|
|
int most_devices = -1; |
|
|
|
|
|
|
|
cl_platform_id *platforms = NULL; |
|
|
|
cl_platform_id *platforms = NULL; |
|
|
|
cl_platform_id platform = NULL; |
|
|
|
unsigned int i; |
|
|
|
unsigned int i, mdplatform = 0; |
|
|
|
bool ret = false; |
|
|
|
int ret = -1; |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
status = clGetPlatformIDs(0, NULL, &numPlatforms); |
|
|
|
status = clGetPlatformIDs(0, NULL, &numPlatforms); |
|
|
|
/* If this fails, assume no GPUs. */ |
|
|
|
/* If this fails, assume no GPUs. */ |
|
|
@ -66,6 +62,11 @@ int clDevicesNum(void) { |
|
|
|
goto out; |
|
|
|
goto out; |
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (preferred_platform_id >= (int)numPlatforms) { |
|
|
|
|
|
|
|
applog(LOG_ERR, "Specified platform that does not exist"); |
|
|
|
|
|
|
|
goto out; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
platforms = (cl_platform_id *)malloc(numPlatforms*sizeof(cl_platform_id)); |
|
|
|
platforms = (cl_platform_id *)malloc(numPlatforms*sizeof(cl_platform_id)); |
|
|
|
status = clGetPlatformIDs(numPlatforms, platforms, NULL); |
|
|
|
status = clGetPlatformIDs(numPlatforms, platforms, NULL); |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
if (status != CL_SUCCESS) { |
|
|
@ -74,32 +75,49 @@ int clDevicesNum(void) { |
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
for (i = 0; i < numPlatforms; i++) { |
|
|
|
for (i = 0; i < numPlatforms; i++) { |
|
|
|
if (opt_platform_id >= 0 && (int)i != opt_platform_id) |
|
|
|
if (preferred_platform_id >= 0 && (int)i != preferred_platform_id) |
|
|
|
continue; |
|
|
|
continue; |
|
|
|
|
|
|
|
|
|
|
|
status = clGetPlatformInfo( platforms[i], CL_PLATFORM_VENDOR, sizeof(pbuff), pbuff, NULL); |
|
|
|
*platform = platforms[i]; |
|
|
|
|
|
|
|
ret = true; |
|
|
|
|
|
|
|
break; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
out: |
|
|
|
|
|
|
|
if (platforms) free(platforms); |
|
|
|
|
|
|
|
return ret; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
int clDevicesNum(void) { |
|
|
|
|
|
|
|
cl_int status; |
|
|
|
|
|
|
|
char pbuff[256]; |
|
|
|
|
|
|
|
cl_uint numDevices; |
|
|
|
|
|
|
|
cl_platform_id platform = NULL; |
|
|
|
|
|
|
|
int ret = -1; |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (!get_opencl_platform(opt_platform_id, &platform)) { |
|
|
|
|
|
|
|
goto out; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
status = clGetPlatformInfo(platform, CL_PLATFORM_VENDOR, sizeof(pbuff), pbuff, NULL); |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
applog(LOG_ERR, "Error %d: Getting Platform Info. (clGetPlatformInfo)", status); |
|
|
|
applog(LOG_ERR, "Error %d: Getting Platform Info. (clGetPlatformInfo)", status); |
|
|
|
goto out; |
|
|
|
goto out; |
|
|
|
} |
|
|
|
} |
|
|
|
platform = platforms[i]; |
|
|
|
|
|
|
|
applog(LOG_INFO, "CL Platform %d vendor: %s", i, pbuff); |
|
|
|
applog(LOG_INFO, "CL Platform vendor: %s", pbuff); |
|
|
|
status = clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(pbuff), pbuff, NULL); |
|
|
|
status = clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(pbuff), pbuff, NULL); |
|
|
|
if (status == CL_SUCCESS) |
|
|
|
if (status == CL_SUCCESS) |
|
|
|
applog(LOG_INFO, "CL Platform %d name: %s", i, pbuff); |
|
|
|
applog(LOG_INFO, "CL Platform name: %s", pbuff); |
|
|
|
status = clGetPlatformInfo(platform, CL_PLATFORM_VERSION, sizeof(pbuff), pbuff, NULL); |
|
|
|
status = clGetPlatformInfo(platform, CL_PLATFORM_VERSION, sizeof(pbuff), pbuff, NULL); |
|
|
|
if (status == CL_SUCCESS) |
|
|
|
if (status == CL_SUCCESS) |
|
|
|
applog(LOG_INFO, "CL Platform %d version: %s", i, pbuff); |
|
|
|
applog(LOG_INFO, "CL Platform version: %s", pbuff); |
|
|
|
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &numDevices); |
|
|
|
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &numDevices); |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
applog(LOG_INFO, "Error %d: Getting Device IDs (num)", status); |
|
|
|
applog(LOG_INFO, "Error %d: Getting Device IDs (num)", status); |
|
|
|
continue; |
|
|
|
goto out; |
|
|
|
} |
|
|
|
|
|
|
|
applog(LOG_INFO, "Platform %d devices: %d", i, numDevices); |
|
|
|
|
|
|
|
if ((int)numDevices > most_devices) { |
|
|
|
|
|
|
|
most_devices = numDevices; |
|
|
|
|
|
|
|
mdplatform = i; |
|
|
|
|
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
applog(LOG_INFO, "Platform devices: %d", numDevices); |
|
|
|
if (numDevices) { |
|
|
|
if (numDevices) { |
|
|
|
unsigned int j; |
|
|
|
unsigned int j; |
|
|
|
cl_device_id *devices = (cl_device_id *)malloc(numDevices*sizeof(cl_device_id)); |
|
|
|
cl_device_id *devices = (cl_device_id *)malloc(numDevices*sizeof(cl_device_id)); |
|
|
@ -111,17 +129,31 @@ int clDevicesNum(void) { |
|
|
|
} |
|
|
|
} |
|
|
|
free(devices); |
|
|
|
free(devices); |
|
|
|
} |
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (opt_platform_id < 0) |
|
|
|
|
|
|
|
opt_platform_id = mdplatform;; |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
ret = most_devices; |
|
|
|
ret = numDevices; |
|
|
|
out: |
|
|
|
out: |
|
|
|
if (platforms) free(platforms); |
|
|
|
|
|
|
|
return ret; |
|
|
|
return ret; |
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
static cl_int create_opencl_context(cl_context *context, cl_platform_id *platform) |
|
|
|
|
|
|
|
{ |
|
|
|
|
|
|
|
cl_context_properties cps[3] = { CL_CONTEXT_PLATFORM, (cl_context_properties)*platform, 0 }; |
|
|
|
|
|
|
|
cl_int status; |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
*context = clCreateContextFromType(cps, CL_DEVICE_TYPE_GPU, NULL, NULL, &status); |
|
|
|
|
|
|
|
return status; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
static cl_int create_opencl_command_queue(cl_command_queue *command_queue, cl_context *context, cl_device_id *device, cl_command_queue_properties cq_properties) |
|
|
|
|
|
|
|
{ |
|
|
|
|
|
|
|
cl_int status; |
|
|
|
|
|
|
|
*command_queue = clCreateCommandQueue(*context, *device, |
|
|
|
|
|
|
|
cq_properties, &status); |
|
|
|
|
|
|
|
if (status != CL_SUCCESS) /* Try again without OOE enable */ |
|
|
|
|
|
|
|
*command_queue = clCreateCommandQueue(*context, *device, 0, &status); |
|
|
|
|
|
|
|
return status; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
static float get_opencl_version(cl_device_id device) |
|
|
|
static float get_opencl_version(cl_device_id device) |
|
|
|
{ |
|
|
|
{ |
|
|
|
/* Check for OpenCL >= 1.0 support, needed for global offset parameter usage. */ |
|
|
|
/* Check for OpenCL >= 1.0 support, needed for global offset parameter usage. */ |
|
|
@ -144,65 +176,41 @@ static float get_opencl_version(cl_device_id device) |
|
|
|
return version; |
|
|
|
return version; |
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
static bool get_opencl_bit_align_support(cl_device_id *device) |
|
|
|
|
|
|
|
{ |
|
|
|
|
|
|
|
char extensions[1024]; |
|
|
|
|
|
|
|
const char * camo = "cl_amd_media_ops"; |
|
|
|
|
|
|
|
char *find; |
|
|
|
|
|
|
|
cl_int status; |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
status = clGetDeviceInfo(*device, CL_DEVICE_EXTENSIONS, 1024, (void *)extensions, NULL); |
|
|
|
|
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
|
|
|
|
return false; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
find = strstr(extensions, camo); |
|
|
|
|
|
|
|
return !!find; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
_clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *algorithm) |
|
|
|
_clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *algorithm) |
|
|
|
{ |
|
|
|
{ |
|
|
|
_clState *clState = (_clState *)calloc(1, sizeof(_clState)); |
|
|
|
_clState *clState = (_clState *)calloc(1, sizeof(_clState)); |
|
|
|
struct cgpu_info *cgpu = &gpus[gpu]; |
|
|
|
struct cgpu_info *cgpu = &gpus[gpu]; |
|
|
|
cl_platform_id platform = NULL; |
|
|
|
cl_platform_id platform = NULL; |
|
|
|
char pbuff[256], vbuff[255]; |
|
|
|
char pbuff[256]; |
|
|
|
build_kernel_data *build_data = (build_kernel_data *) alloca(sizeof(struct _build_kernel_data)); |
|
|
|
build_kernel_data *build_data = (build_kernel_data *) alloca(sizeof(struct _build_kernel_data)); |
|
|
|
cl_platform_id* platforms; |
|
|
|
|
|
|
|
cl_uint preferred_vwidth; |
|
|
|
cl_uint preferred_vwidth; |
|
|
|
cl_device_id *devices; |
|
|
|
cl_device_id *devices; |
|
|
|
cl_uint numPlatforms; |
|
|
|
|
|
|
|
cl_uint numDevices; |
|
|
|
cl_uint numDevices; |
|
|
|
cl_int status; |
|
|
|
cl_int status; |
|
|
|
|
|
|
|
|
|
|
|
status = clGetPlatformIDs(0, NULL, &numPlatforms); |
|
|
|
if (!get_opencl_platform(opt_platform_id, &platform)) { |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
|
|
|
|
applog(LOG_ERR, "Error %d: Getting Platforms. (clGetPlatformsIDs)", status); |
|
|
|
|
|
|
|
return NULL; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
platforms = (cl_platform_id *)alloca(numPlatforms*sizeof(cl_platform_id)); |
|
|
|
|
|
|
|
status = clGetPlatformIDs(numPlatforms, platforms, NULL); |
|
|
|
|
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
|
|
|
|
applog(LOG_ERR, "Error %d: Getting Platform Ids. (clGetPlatformsIDs)", status); |
|
|
|
|
|
|
|
return NULL; |
|
|
|
return NULL; |
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
if (opt_platform_id >= (int)numPlatforms) { |
|
|
|
numDevices = clDevicesNum(); |
|
|
|
applog(LOG_ERR, "Specified platform that does not exist"); |
|
|
|
|
|
|
|
return NULL; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
status = clGetPlatformInfo(platforms[opt_platform_id], CL_PLATFORM_VENDOR, sizeof(pbuff), pbuff, NULL); |
|
|
|
|
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
|
|
|
|
applog(LOG_ERR, "Error %d: Getting Platform Info. (clGetPlatformInfo)", status); |
|
|
|
|
|
|
|
return NULL; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
platform = platforms[opt_platform_id]; |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (platform == NULL) { |
|
|
|
|
|
|
|
perror("NULL platform found!\n"); |
|
|
|
|
|
|
|
return NULL; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
applog(LOG_INFO, "CL Platform vendor: %s", pbuff); |
|
|
|
|
|
|
|
status = clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(pbuff), pbuff, NULL); |
|
|
|
|
|
|
|
if (status == CL_SUCCESS) |
|
|
|
|
|
|
|
applog(LOG_INFO, "CL Platform name: %s", pbuff); |
|
|
|
|
|
|
|
status = clGetPlatformInfo(platform, CL_PLATFORM_VERSION, sizeof(vbuff), vbuff, NULL); |
|
|
|
|
|
|
|
if (status == CL_SUCCESS) |
|
|
|
|
|
|
|
applog(LOG_INFO, "CL Platform version: %s", vbuff); |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &numDevices); |
|
|
|
if (numDevices <= 0 ) return NULL; |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
|
|
|
|
applog(LOG_ERR, "Error %d: Getting Device IDs (num)", status); |
|
|
|
|
|
|
|
return NULL; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (numDevices > 0 ) { |
|
|
|
|
|
|
|
devices = (cl_device_id *)alloca(numDevices*sizeof(cl_device_id)); |
|
|
|
devices = (cl_device_id *)alloca(numDevices*sizeof(cl_device_id)); |
|
|
|
|
|
|
|
|
|
|
|
/* Now, get the device list data */ |
|
|
|
/* Now, get the device list data */ |
|
|
@ -224,58 +232,31 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *alg |
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
applog(LOG_INFO, "\t%i\t%s", i, pbuff); |
|
|
|
applog(LOG_INFO, "\t%i\t%s", i, pbuff); |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (gpu < numDevices) { |
|
|
|
|
|
|
|
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL); |
|
|
|
|
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
|
|
|
|
applog(LOG_ERR, "Error %d: Getting Device Info", status); |
|
|
|
|
|
|
|
return NULL; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (i == gpu) { |
|
|
|
applog(LOG_INFO, "Selected %i: %s", gpu, pbuff); |
|
|
|
applog(LOG_INFO, "Selected %i: %s", gpu, pbuff); |
|
|
|
strncpy(name, pbuff, nameSize); |
|
|
|
strncpy(name, pbuff, nameSize); |
|
|
|
} else { |
|
|
|
} |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (gpu >= numDevices) { |
|
|
|
applog(LOG_ERR, "Invalid GPU %i", gpu); |
|
|
|
applog(LOG_ERR, "Invalid GPU %i", gpu); |
|
|
|
return NULL; |
|
|
|
return NULL; |
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
} else return NULL; |
|
|
|
status = create_opencl_context(&clState->context, &platform); |
|
|
|
|
|
|
|
|
|
|
|
cl_context_properties cps[3] = { CL_CONTEXT_PLATFORM, (cl_context_properties)platform, 0 }; |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
clState->context = clCreateContextFromType(cps, CL_DEVICE_TYPE_GPU, NULL, NULL, &status); |
|
|
|
|
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
applog(LOG_ERR, "Error %d: Creating Context. (clCreateContextFromType)", status); |
|
|
|
applog(LOG_ERR, "Error %d: Creating Context. (clCreateContextFromType)", status); |
|
|
|
return NULL; |
|
|
|
return NULL; |
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
/////////////////////////////////////////////////////////////////
|
|
|
|
status = create_opencl_command_queue(&clState->commandQueue, &clState->context, &devices[gpu], cgpu->algorithm.cq_properties); |
|
|
|
// Create an OpenCL command queue
|
|
|
|
|
|
|
|
/////////////////////////////////////////////////////////////////
|
|
|
|
|
|
|
|
clState->commandQueue = clCreateCommandQueue(clState->context, devices[gpu], |
|
|
|
|
|
|
|
cgpu->algorithm.cq_properties, &status); |
|
|
|
|
|
|
|
if (status != CL_SUCCESS) /* Try again without OOE enable */ |
|
|
|
|
|
|
|
clState->commandQueue = clCreateCommandQueue(clState->context, devices[gpu], 0 , &status); |
|
|
|
|
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
applog(LOG_ERR, "Error %d: Creating Command Queue. (clCreateCommandQueue)", status); |
|
|
|
applog(LOG_ERR, "Error %d: Creating Command Queue. (clCreateCommandQueue)", status); |
|
|
|
return NULL; |
|
|
|
return NULL; |
|
|
|
} |
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
/* Check for BFI INT support. Hopefully people don't mix devices with
|
|
|
|
clState->hasBitAlign = get_opencl_bit_align_support(&devices[gpu]); |
|
|
|
* and without it! */ |
|
|
|
|
|
|
|
char extensions[1024]; |
|
|
|
|
|
|
|
const char * camo = "cl_amd_media_ops"; |
|
|
|
|
|
|
|
char *find; |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_EXTENSIONS, 1024, (void *)extensions, NULL); |
|
|
|
|
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
|
|
|
|
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_EXTENSIONS", status); |
|
|
|
|
|
|
|
return NULL; |
|
|
|
|
|
|
|
} |
|
|
|
|
|
|
|
find = strstr(extensions, camo); |
|
|
|
|
|
|
|
if (find) |
|
|
|
|
|
|
|
clState->hasBitAlign = true; |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT, sizeof(cl_uint), (void *)&preferred_vwidth, NULL); |
|
|
|
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT, sizeof(cl_uint), (void *)&preferred_vwidth, NULL); |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|
if (status != CL_SUCCESS) { |
|
|
|