void COPROC_NVIDIA::get(
vector<string>& warnings
) {
int cuda_ndevs, retval;
char buf[256];
#ifdef _WIN32
HMODULE cudalib = LoadLibrary("nvcuda.dll");
if (!cudalib) {
warnings.push_back("No NVIDIA library found");
return;
}
__cuDeviceGetCount = (CUDA_GDC)GetProcAddress( cudalib, "cuDeviceGetCount" );
__cuDriverGetVersion = (CUDA_GDV)GetProcAddress( cudalib, "cuDriverGetVersion" );
__cuInit = (CUDA_GDI)GetProcAddress( cudalib, "cuInit" );
__cuDeviceGet = (CUDA_GDG)GetProcAddress( cudalib, "cuDeviceGet" );
__cuDeviceGetAttribute = (CUDA_GDA)GetProcAddress( cudalib, "cuDeviceGetAttribute" );
__cuDeviceGetName = (CUDA_GDN)GetProcAddress( cudalib, "cuDeviceGetName" );
__cuDeviceTotalMem = (CUDA_GDM)GetProcAddress( cudalib, "cuDeviceTotalMem" );
__cuDeviceComputeCapability = (CUDA_GDCC)GetProcAddress( cudalib, "cuDeviceComputeCapability" );
__cuCtxCreate = (CUDA_CC)GetProcAddress( cudalib, "cuCtxCreate" );
__cuCtxDestroy = (CUDA_CD)GetProcAddress( cudalib, "cuCtxDestroy" );
__cuMemAlloc = (CUDA_MA)GetProcAddress( cudalib, "cuMemAlloc" );
__cuMemFree = (CUDA_MF)GetProcAddress( cudalib, "cuMemFree" );
__cuMemGetInfo = (CUDA_MGI)GetProcAddress( cudalib, "cuMemGetInfo" );
#ifndef SIM
NvAPI_Initialize();
NvAPI_ShortString ss;
NvU32 Version = 0;
NvAPI_SYS_GetDriverAndBranchVersion(&Version, ss);
#if 0
// NvAPI now provides an API for getting #cores :-)
// But not FLOPs per clock cycle :-(
// Anyway, don't use this for now because server code estimates FLOPS
// based on compute capability, so we may as well do the same
// See http://docs.nvidia.com/gameworks/content/gameworkslibrary/coresdk/nvapi/
//
NvPhysicalGpuHandle GPUHandle[NVAPI_MAX_PHYSICAL_GPUS];
NvU32 GpuCount, nc;
NvAPI_EnumPhysicalGPUs(GPUHandle, &GpuCount);
for (unsigned int i=0; i<GpuCount; i++) {
NvAPI_GPU_GetGpuCoreCount(GPUHandle[i], &nc);
}
#endif
#endif
#else
#ifdef __APPLE__
cudalib = dlopen("/usr/local/cuda/lib/libcuda.dylib", RTLD_NOW);
#else
cudalib = dlopen("libcuda.so", RTLD_NOW);
#endif
if (!cudalib) {
warnings.push_back("No NVIDIA library found");
return;
}
__cuDeviceGetCount = (int(*)(int*)) dlsym(cudalib, "cuDeviceGetCount");
__cuDriverGetVersion = (int(*)(int*)) dlsym( cudalib, "cuDriverGetVersion" );
__cuInit = (int(*)(unsigned int)) dlsym( cudalib, "cuInit" );
__cuDeviceGet = (int(*)(int*, int)) dlsym( cudalib, "cuDeviceGet" );
__cuDeviceGetAttribute = (int(*)(int*, int, int)) dlsym( cudalib, "cuDeviceGetAttribute" );
__cuDeviceGetName = (int(*)(char*, int, int)) dlsym( cudalib, "cuDeviceGetName" );
__cuDeviceTotalMem = (int(*)(size_t*, int)) dlsym( cudalib, "cuDeviceTotalMem" );
__cuDeviceComputeCapability = (int(*)(int*, int*, int)) dlsym( cudalib, "cuDeviceComputeCapability" );
__cuCtxCreate = (int(*)(void**, unsigned int, unsigned int)) dlsym( cudalib, "cuCtxCreate" );
__cuCtxDestroy = (int(*)(void*)) dlsym( cudalib, "cuCtxDestroy" );
__cuMemAlloc = (int(*)(unsigned int*, size_t)) dlsym( cudalib, "cuMemAlloc" );
__cuMemFree = (int(*)(unsigned int)) dlsym( cudalib, "cuMemFree" );
__cuMemGetInfo = (int(*)(size_t*, size_t*)) dlsym( cudalib, "cuMemGetInfo" );
dlclose(cudalib);
#endif
if (!__cuDriverGetVersion) {
warnings.push_back("cuDriverGetVersion() missing from NVIDIA library");
return;
}
if (!__cuInit) {
warnings.push_back("cuInit() missing from NVIDIA library");
return;
}
if (!__cuDeviceGetCount) {
warnings.push_back("cuDeviceGetCount() missing from NVIDIA library");
return;
}
if (!__cuDeviceGet) {
warnings.push_back("cuDeviceGet() missing from NVIDIA library");
return;
}
if (!__cuDeviceGetAttribute) {
warnings.push_back("cuDeviceGetAttribute() missing from NVIDIA library");
return;
}
if (!__cuDeviceTotalMem) {
warnings.push_back("cuDeviceTotalMem() missing from NVIDIA library");
return;
}
if (!__cuDeviceComputeCapability) {
warnings.push_back("cuDeviceComputeCapability() missing from NVIDIA library");
return;
}
if (!__cuMemAlloc) {
warnings.push_back("cuMemAlloc() missing from NVIDIA library");
return;
}
if (!__cuMemFree) {
warnings.push_back("cuMemFree() missing from NVIDIA library");
return;
}
#ifdef __APPLE__
// If system is just booting, CUDA driver may not be ready yet
for (int retryCount=0; retryCount<45; retryCount++) {
#endif
retval = (*__cuInit)(0);
#ifdef __APPLE__
if (!retval) break;
if (TickCount() > (120*60)) break; // Don't retry if system has been up for over 2 minutes
boinc_sleep(1.);
continue;
}
#endif
if (retval) {
sprintf(buf, "NVIDIA drivers present but no GPUs found");
warnings.push_back(buf);
return;
}
retval = (*__cuDriverGetVersion)(&cuda_version);
if (retval) {
sprintf(buf, "cuDriverGetVersion() returned %d", retval);
warnings.push_back(buf);
return;
}
have_cuda = true;
retval = (*__cuDeviceGetCount)(&cuda_ndevs);
if (retval) {
sprintf(buf, "cuDeviceGetCount() returned %d", retval);
warnings.push_back(buf);
return;
}
sprintf(buf, "NVIDIA library reports %d GPU%s", cuda_ndevs, (cuda_ndevs==1)?"":"s");
warnings.push_back(buf);
int j, itemp;
size_t global_mem = 0;
COPROC_NVIDIA cc;
string s;
for (j=0; j<cuda_ndevs; j++) {
memset(&cc.prop, 0, sizeof(cc.prop));
CUdevice device;
retval = (*__cuDeviceGet)(&device, j);
if (retval) {
sprintf(buf, "cuDeviceGet(%d) returned %d", j, retval);
warnings.push_back(buf);
return;
}
(*__cuDeviceGetName)(cc.prop.name, 256, device);
if (retval) {
sprintf(buf, "cuDeviceGetName(%d) returned %d", j, retval);
warnings.push_back(buf);
return;
}
(*__cuDeviceComputeCapability)(&cc.prop.major, &cc.prop.minor, device);
(*__cuDeviceTotalMem)(&global_mem, device);
cc.prop.totalGlobalMem = (double) global_mem;
(*__cuDeviceGetAttribute)(&itemp, CU_DEVICE_ATTRIBUTE_SHARED_MEMORY_PER_BLOCK, device);
cc.prop.sharedMemPerBlock = (double) itemp;
(*__cuDeviceGetAttribute)(&cc.prop.regsPerBlock, CU_DEVICE_ATTRIBUTE_REGISTERS_PER_BLOCK, device);
(*__cuDeviceGetAttribute)(&cc.prop.warpSize, CU_DEVICE_ATTRIBUTE_WARP_SIZE, device);
(*__cuDeviceGetAttribute)(&itemp, CU_DEVICE_ATTRIBUTE_MAX_PITCH, device);
cc.prop.memPitch = (double) itemp;
retval = (*__cuDeviceGetAttribute)(&cc.prop.maxThreadsPerBlock, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, device);
retval = (*__cuDeviceGetAttribute)(&cc.prop.maxThreadsDim[0], CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X, device);
(*__cuDeviceGetAttribute)(&cc.prop.maxThreadsDim[1], CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y, device);
(*__cuDeviceGetAttribute)(&cc.prop.maxThreadsDim[2], CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z, device);
(*__cuDeviceGetAttribute)(&cc.prop.maxGridSize[0], CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X, device);
(*__cuDeviceGetAttribute)(&cc.prop.maxGridSize[1], CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y, device);
(*__cuDeviceGetAttribute)(&cc.prop.maxGridSize[2], CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z, device);
(*__cuDeviceGetAttribute)(&cc.prop.clockRate, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, device);
(*__cuDeviceGetAttribute)(&itemp, CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY, device);
cc.prop.totalConstMem = (double) itemp;
(*__cuDeviceGetAttribute)(&itemp, CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT, device);
cc.prop.textureAlignment = (double) itemp;
(*__cuDeviceGetAttribute)(&cc.prop.deviceOverlap, CU_DEVICE_ATTRIBUTE_GPU_OVERLAP, device);
(*__cuDeviceGetAttribute)(&cc.prop.multiProcessorCount, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, device);
(*__cuDeviceGetAttribute)(&cc.pci_info.bus_id, CU_DEVICE_ATTRIBUTE_PCI_BUS_ID, device);
(*__cuDeviceGetAttribute)(&cc.pci_info.device_id, CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID, device);
(*__cuDeviceGetAttribute)(&cc.pci_info.domain_id, CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID, device);
if (cc.prop.major <= 0) continue; // major == 0 means emulation
if (cc.prop.major > 100) continue; // e.g. 9999 is an error
#if defined(_WIN32) && !defined(SIM)
cc.display_driver_version = Version;
#elif defined(__APPLE__)
cc.display_driver_version = NSVersionOfRunTimeLibrary("cuda");
#else
cc.display_driver_version = nvidia_driver_version();
#endif
cc.have_cuda = true;
cc.cuda_version = cuda_version;
cc.device_num = j;
cc.set_peak_flops();
get_available_nvidia_ram(cc, warnings);
nvidia_gpus.push_back(cc);
}
if (!nvidia_gpus.size()) {
warnings.push_back("No CUDA-capable NVIDIA GPUs found");
}
}
void COPROC_NVIDIA::get(
bool use_all, // if false, use only those equivalent to most capable
vector<string>& warnings,
vector<int>& ignore_devs
) {
int cuda_ndevs, retval;
char buf[256];
#ifdef _WIN32
HMODULE cudalib = LoadLibrary("nvcuda.dll");
if (!cudalib) {
warnings.push_back("No NVIDIA library found");
return;
}
__cuDeviceGetCount = (CUDA_GDC)GetProcAddress( cudalib, "cuDeviceGetCount" );
__cuDriverGetVersion = (CUDA_GDV)GetProcAddress( cudalib, "cuDriverGetVersion" );
__cuInit = (CUDA_GDI)GetProcAddress( cudalib, "cuInit" );
__cuDeviceGet = (CUDA_GDG)GetProcAddress( cudalib, "cuDeviceGet" );
__cuDeviceGetAttribute = (CUDA_GDA)GetProcAddress( cudalib, "cuDeviceGetAttribute" );
__cuDeviceGetName = (CUDA_GDN)GetProcAddress( cudalib, "cuDeviceGetName" );
__cuDeviceTotalMem = (CUDA_GDM)GetProcAddress( cudalib, "cuDeviceTotalMem" );
__cuDeviceComputeCapability = (CUDA_GDCC)GetProcAddress( cudalib, "cuDeviceComputeCapability" );
__cuCtxCreate = (CUDA_CC)GetProcAddress( cudalib, "cuCtxCreate" );
__cuCtxDestroy = (CUDA_CD)GetProcAddress( cudalib, "cuCtxDestroy" );
__cuMemAlloc = (CUDA_MA)GetProcAddress( cudalib, "cuMemAlloc" );
__cuMemFree = (CUDA_MF)GetProcAddress( cudalib, "cuMemFree" );
__cuMemGetInfo = (CUDA_MGI)GetProcAddress( cudalib, "cuMemGetInfo" );
#ifndef SIM
NvAPI_Status nvapiStatus;
NV_DISPLAY_DRIVER_VERSION Version;
memset(&Version, 0, sizeof(Version));
Version.version = NV_DISPLAY_DRIVER_VERSION_VER;
NvAPI_Initialize();
nvapiStatus = NvAPI_GetDisplayDriverVersion(NULL, &Version);
#endif
#else
#ifdef __APPLE__
cudalib = dlopen("/usr/local/cuda/lib/libcuda.dylib", RTLD_NOW);
#else
cudalib = dlopen("libcuda.so", RTLD_NOW);
#endif
if (!cudalib) {
warnings.push_back("No NVIDIA library found");
return;
}
__cuDeviceGetCount = (int(*)(int*)) dlsym(cudalib, "cuDeviceGetCount");
__cuDriverGetVersion = (int(*)(int*)) dlsym( cudalib, "cuDriverGetVersion" );
__cuInit = (int(*)(unsigned int)) dlsym( cudalib, "cuInit" );
__cuDeviceGet = (int(*)(int*, int)) dlsym( cudalib, "cuDeviceGet" );
__cuDeviceGetAttribute = (int(*)(int*, int, int)) dlsym( cudalib, "cuDeviceGetAttribute" );
__cuDeviceGetName = (int(*)(char*, int, int)) dlsym( cudalib, "cuDeviceGetName" );
__cuDeviceTotalMem = (int(*)(size_t*, int)) dlsym( cudalib, "cuDeviceTotalMem" );
__cuDeviceComputeCapability = (int(*)(int*, int*, int)) dlsym( cudalib, "cuDeviceComputeCapability" );
__cuCtxCreate = (int(*)(void**, unsigned int, unsigned int)) dlsym( cudalib, "cuCtxCreate" );
__cuCtxDestroy = (int(*)(void*)) dlsym( cudalib, "cuCtxDestroy" );
__cuMemAlloc = (int(*)(unsigned int*, size_t)) dlsym( cudalib, "cuMemAlloc" );
__cuMemFree = (int(*)(unsigned int)) dlsym( cudalib, "cuMemFree" );
__cuMemGetInfo = (int(*)(size_t*, size_t*)) dlsym( cudalib, "cuMemGetInfo" );
#endif
if (!__cuDriverGetVersion) {
warnings.push_back("cuDriverGetVersion() missing from NVIDIA library");
return;
}
if (!__cuInit) {
warnings.push_back("cuInit() missing from NVIDIA library");
return;
}
if (!__cuDeviceGetCount) {
warnings.push_back("cuDeviceGetCount() missing from NVIDIA library");
return;
}
if (!__cuDeviceGet) {
warnings.push_back("cuDeviceGet() missing from NVIDIA library");
return;
}
if (!__cuDeviceGetAttribute) {
warnings.push_back("cuDeviceGetAttribute() missing from NVIDIA library");
return;
}
if (!__cuDeviceTotalMem) {
warnings.push_back("cuDeviceTotalMem() missing from NVIDIA library");
return;
}
if (!__cuDeviceComputeCapability) {
warnings.push_back("cuDeviceComputeCapability() missing from NVIDIA library");
return;
}
if (!__cuCtxCreate) {
warnings.push_back("cuCtxCreate() missing from NVIDIA library");
return;
}
if (!__cuCtxDestroy) {
warnings.push_back("cuCtxDestroy() missing from NVIDIA library");
return;
}
if (!__cuMemAlloc) {
warnings.push_back("cuMemAlloc() missing from NVIDIA library");
return;
}
if (!__cuMemFree) {
warnings.push_back("cuMemFree() missing from NVIDIA library");
return;
}
if (!__cuMemGetInfo) {
warnings.push_back("cuMemGetInfo() missing from NVIDIA library");
return;
}
retval = (*__cuInit)(0);
if (retval) {
sprintf(buf, "NVIDIA drivers present but no GPUs found");
warnings.push_back(buf);
return;
}
retval = (*__cuDriverGetVersion)(&cuda_version);
if (retval) {
sprintf(buf, "cuDriverGetVersion() returned %d", retval);
warnings.push_back(buf);
return;
}
retval = (*__cuDeviceGetCount)(&cuda_ndevs);
if (retval) {
sprintf(buf, "cuDeviceGetCount() returned %d", retval);
warnings.push_back(buf);
return;
}
sprintf(buf, "NVIDIA library reports %d GPU%s", cuda_ndevs, (cuda_ndevs==1)?"":"s");
warnings.push_back(buf);
int j, itemp;
unsigned int i;
size_t global_mem;
COPROC_NVIDIA cc;
string s;
for (j=0; j<cuda_ndevs; j++) {
memset(&cc.prop, 0, sizeof(cc.prop));
CUdevice device;
retval = (*__cuDeviceGet)(&device, j);
if (retval) {
sprintf(buf, "cuDeviceGet(%d) returned %d", j, retval);
warnings.push_back(buf);
return;
}
(*__cuDeviceGetName)(cc.prop.name, 256, device);
if (retval) {
sprintf(buf, "cuDeviceGetName(%d) returned %d", j, retval);
warnings.push_back(buf);
return;
}
(*__cuDeviceComputeCapability)(&cc.prop.major, &cc.prop.minor, device);
(*__cuDeviceTotalMem)(&global_mem, device);
cc.prop.totalGlobalMem = (double) global_mem;
(*__cuDeviceGetAttribute)(&itemp, CU_DEVICE_ATTRIBUTE_SHARED_MEMORY_PER_BLOCK, device);
cc.prop.sharedMemPerBlock = (double) itemp;
(*__cuDeviceGetAttribute)(&cc.prop.regsPerBlock, CU_DEVICE_ATTRIBUTE_REGISTERS_PER_BLOCK, device);
(*__cuDeviceGetAttribute)(&cc.prop.warpSize, CU_DEVICE_ATTRIBUTE_WARP_SIZE, device);
(*__cuDeviceGetAttribute)(&itemp, CU_DEVICE_ATTRIBUTE_MAX_PITCH, device);
cc.prop.memPitch = (double) itemp;
retval = (*__cuDeviceGetAttribute)(&cc.prop.maxThreadsPerBlock, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, device);
retval = (*__cuDeviceGetAttribute)(&cc.prop.maxThreadsDim[0], CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X, device);
(*__cuDeviceGetAttribute)(&cc.prop.maxThreadsDim[1], CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y, device);
(*__cuDeviceGetAttribute)(&cc.prop.maxThreadsDim[2], CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z, device);
(*__cuDeviceGetAttribute)(&cc.prop.maxGridSize[0], CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X, device);
(*__cuDeviceGetAttribute)(&cc.prop.maxGridSize[1], CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y, device);
(*__cuDeviceGetAttribute)(&cc.prop.maxGridSize[2], CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z, device);
(*__cuDeviceGetAttribute)(&cc.prop.clockRate, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, device);
(*__cuDeviceGetAttribute)(&itemp, CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY, device);
cc.prop.totalConstMem = (double) itemp;
(*__cuDeviceGetAttribute)(&itemp, CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT, device);
cc.prop.textureAlignment = (double) itemp;
(*__cuDeviceGetAttribute)(&cc.prop.deviceOverlap, CU_DEVICE_ATTRIBUTE_GPU_OVERLAP, device);
(*__cuDeviceGetAttribute)(&cc.prop.multiProcessorCount, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, device);
(*__cuDeviceGetAttribute)(&cc.pci_info.bus_id, CU_DEVICE_ATTRIBUTE_PCI_BUS_ID, device);
(*__cuDeviceGetAttribute)(&cc.pci_info.device_id, CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID, device);
(*__cuDeviceGetAttribute)(&cc.pci_info.domain_id, CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID, device);
if (cc.prop.major <= 0) continue; // major == 0 means emulation
if (cc.prop.major > 100) continue; // e.g. 9999 is an error
#if defined(_WIN32) && !defined(SIM)
cc.display_driver_version = Version.drvVersion;
#elif defined(__APPLE__)
cc.display_driver_version = NSVersionOfRunTimeLibrary("cuda");
#else
cc.display_driver_version = 0;
#endif
cc.have_cuda = true;
cc.cuda_version = cuda_version;
cc.device_num = j;
cc.set_peak_flops();
cc.get_available_ram();
nvidia_gpus.push_back(cc);
}
if (!nvidia_gpus.size()) {
warnings.push_back("No CUDA-capable NVIDIA GPUs found");
return;
}
// identify the most capable non-ignored instance
//
bool first = true;
for (i=0; i<nvidia_gpus.size(); i++) {
if (in_vector(nvidia_gpus[i].device_num, ignore_devs)) continue;
if (first) {
*this = nvidia_gpus[i];
first = false;
} else if (nvidia_compare(nvidia_gpus[i], *this, false) > 0) {
*this = nvidia_gpus[i];
}
}
// see which other instances are equivalent,
// and set "count", "device_nums", and "pci_infos"
//
count = 0;
for (i=0; i<nvidia_gpus.size(); i++) {
if (in_vector(nvidia_gpus[i].device_num, ignore_devs)) {
nvidia_gpus[i].is_used = COPROC_IGNORED;
} else if (use_all || !nvidia_compare(nvidia_gpus[i], *this, true)) {
device_nums[count] = nvidia_gpus[i].device_num;
pci_infos[count] = nvidia_gpus[i].pci_info;
count++;
nvidia_gpus[i].is_used = COPROC_USED;
} else {
nvidia_gpus[i].is_used = COPROC_UNUSED;
}
}
}
void COPROCS::get_opencl(
vector<string>& warnings
) {
cl_int ciErrNum;
cl_platform_id platforms[MAX_OPENCL_PLATFORMS];
cl_uint num_platforms, platform_index, num_devices, device_index;
cl_device_id devices[MAX_COPROC_INSTANCES];
char platform_version[256];
char platform_vendor[256];
char buf[256];
OPENCL_DEVICE_PROP prop;
int current_CUDA_index;
int current_CAL_index;
int min_CAL_target;
int num_CAL_devices = (int)ati_gpus.size();
vector<int>devnums_pci_slot_sort;
vector<OPENCL_DEVICE_PROP>::iterator it;
#ifdef _WIN32
opencl_lib = LoadLibrary("OpenCL.dll");
if (!opencl_lib) {
warnings.push_back("No OpenCL library found");
return;
}
__clGetPlatformIDs = (CL_PLATFORMIDS)GetProcAddress( opencl_lib, "clGetPlatformIDs" );
__clGetPlatformInfo = (CL_PLATFORMINFO)GetProcAddress( opencl_lib, "clGetPlatformInfo" );
__clGetDeviceIDs = (CL_DEVICEIDS)GetProcAddress( opencl_lib, "clGetDeviceIDs" );
__clGetDeviceInfo = (CL_INFO)GetProcAddress( opencl_lib, "clGetDeviceInfo" );
#else
#ifdef __APPLE__
opencl_lib = dlopen("/System/Library/Frameworks/OpenCL.framework/Versions/Current/OpenCL", RTLD_NOW);
#else
//TODO: Is this correct?
opencl_lib = dlopen("libOpenCL.so", RTLD_NOW);
#endif
if (!opencl_lib) {
warnings.push_back("No OpenCL library found");
return;
}
__clGetPlatformIDs = (cl_int(*)(cl_uint, cl_platform_id*, cl_uint*)) dlsym( opencl_lib, "clGetPlatformIDs" );
__clGetPlatformInfo = (cl_int(*)(cl_platform_id, cl_platform_info, size_t, void*, size_t*)) dlsym( opencl_lib, "clGetPlatformInfo" );
__clGetDeviceIDs = (cl_int(*)(cl_platform_id, cl_device_type, cl_uint, cl_device_id*, cl_uint*)) dlsym( opencl_lib, "clGetDeviceIDs" );
__clGetDeviceInfo = (cl_int(*)(cl_device_id, cl_device_info, size_t, void*, size_t*)) dlsym( opencl_lib, "clGetDeviceInfo" );
#endif
if (!__clGetPlatformIDs) {
warnings.push_back("clGetPlatformIDs() missing from OpenCL library");
return;
}
if (!__clGetPlatformInfo) {
warnings.push_back("clGetPlatformInfo() missing from OpenCL library");
return;
}
if (!__clGetDeviceIDs) {
warnings.push_back("clGetDeviceIDs() missing from OpenCL library");
return;
}
if (!__clGetDeviceInfo) {
warnings.push_back("clGetDeviceInfo() missing from OpenCL library");
return;
}
ciErrNum = (*__clGetPlatformIDs)(MAX_OPENCL_PLATFORMS, platforms, &num_platforms);
if ((ciErrNum != CL_SUCCESS) || (num_platforms == 0)) {
warnings.push_back("clGetPlatformIDs() failed to return any OpenCL platforms");
return;
}
if (nvidia_gpus.size()) {
for (int i=0; i<(int)nvidia_gpus.size(); ++i) {
devnums_pci_slot_sort.push_back(i);
}
#ifdef __APPLE__
std::stable_sort(
devnums_pci_slot_sort.begin(),
devnums_pci_slot_sort.end(),
compare_pci_slots
);
#endif
}
for (platform_index=0; platform_index<num_platforms; ++platform_index) {
ciErrNum = (*__clGetPlatformInfo)(
platforms[platform_index], CL_PLATFORM_VERSION,
sizeof(platform_version), &platform_version, NULL
);
if (ciErrNum != CL_SUCCESS) {
snprintf(buf, sizeof(buf),
"Couldn't get PLATFORM_VERSION for platform #%d; error %d",
platform_index, ciErrNum
);
warnings.push_back(buf);
continue;
}
ciErrNum = (*__clGetPlatformInfo)(
platforms[platform_index], CL_PLATFORM_VENDOR,
sizeof(platform_vendor), &platform_vendor, NULL
);
if (ciErrNum != CL_SUCCESS) {
snprintf(buf, sizeof(buf),
"Couldn't get PLATFORM_VENDOR for platform #%d; error %d",
platform_index, ciErrNum
);
warnings.push_back(buf);
}
//////////// CPU //////////////
ciErrNum = (*__clGetDeviceIDs)(
platforms[platform_index], (CL_DEVICE_TYPE_CPU),
MAX_COPROC_INSTANCES, devices, &num_devices
);
if ((ciErrNum != CL_SUCCESS) && (num_devices != 0)) {
num_devices = 0; // No devices
if (ciErrNum != CL_DEVICE_NOT_FOUND) {
snprintf(buf, sizeof(buf),
"Couldn't get CPU Device IDs for platform #%d: error %d",
platform_index, ciErrNum
);
warnings.push_back(buf);
}
}
for (device_index=0; device_index<num_devices; ++device_index) {
memset(&prop, 0, sizeof(prop));
prop.device_id = devices[device_index];
strncpy(
prop.opencl_platform_version, platform_version,
sizeof(prop.opencl_platform_version)-1
);
ciErrNum = get_opencl_info(prop, device_index, warnings);
if (ciErrNum != CL_SUCCESS) continue;
prop.is_used = COPROC_UNUSED;
prop.get_device_version_int();
OPENCL_CPU_PROP c;
strlcpy(c.platform_vendor, platform_vendor, sizeof(c.platform_vendor));
c.opencl_prop = prop;
cpu_opencls.push_back(c);
}
//////////// GPUs //////////////
ciErrNum = (*__clGetDeviceIDs)(
platforms[platform_index], (CL_DEVICE_TYPE_GPU),
MAX_COPROC_INSTANCES, devices, &num_devices
);
if (ciErrNum == CL_DEVICE_NOT_FOUND) continue; // No devices
if (num_devices == 0) continue; // No devices
if (ciErrNum != CL_SUCCESS) {
snprintf(buf, sizeof(buf),
"Couldn't get Device IDs for platform #%d: error %d",
platform_index, ciErrNum
);
warnings.push_back(buf);
continue;
}
// Mac OpenCL does not recognize all NVIDIA GPUs returned by CUDA
// Fortunately, CUDA and OpenCL return the same GPU model name on
// the Mac, so we can use this to match OpenCL devices with CUDA.
//
current_CUDA_index = 0;
// ATI/AMD OpenCL does not always recognize all GPUs returned by CAL.
// This is complicated for several reasons:
// * CAL returns only an enum (CALtargetEnum) for the GPU's family,
// not specific model information.
// * OpenCL return only the GPU family name
// * Which GPUs support OpenCL varies with different versions of the
// AMD Catalyst drivers.
//
// To deal with this, we make some (probably imperfect) assumptions:
// * AMD drivers eliminate OpenCL support for older GPU families first.
// * Lower values of CALtargetEnum represent older GPU families.
// * All ATI/AMD GPUs reported by OpenCL are also reported by CAL (on
// systems where CAL is available) though the converse may not be true.
//
current_CAL_index = 0;
min_CAL_target = 0;
if (is_AMD(platform_vendor) && (num_CAL_devices > 0)) {
while (1) {
int numToMatch = 0;
for (int i=0; i<num_CAL_devices; ++i) {
if ((int)ati_gpus[i].attribs.target >= min_CAL_target) {
++numToMatch;
}
}
if (numToMatch == (int)num_devices) break;
if (numToMatch < (int)num_devices) {
warnings.push_back(
"Could not match ATI OpenCL and CAL GPUs: ignoring CAL."
);
// If we can't match ATI OpenCL and CAL GPUs, ignore CAL
// and keep OpenCL because AMD has deprecated CAL.
ati_gpus.clear();
ati.have_cal = false;
num_CAL_devices = 0;
break;
}
++min_CAL_target;
}
}
for (device_index=0; device_index<num_devices; ++device_index) {
memset(&prop, 0, sizeof(prop));
prop.device_id = devices[device_index];
strncpy(
prop.opencl_platform_version, platform_version,
sizeof(prop.opencl_platform_version)-1
);
//TODO: Should we store the platform(s) for each GPU found?
//TODO: Must we check if multiple platforms found the same GPU and merge the records?
ciErrNum = get_opencl_info(prop, device_index, warnings);
if (ciErrNum != CL_SUCCESS) continue;
prop.is_used = COPROC_UNUSED;
prop.get_device_version_int();
//////////// NVIDIA //////////////
if (is_NVIDIA(prop.vendor)) {
if (nvidia.have_cuda) {
// Mac OpenCL does not recognize all NVIDIA GPUs returned by
// CUDA but we assume that OpenCL and CUDA return devices
// with identical model name strings and that OpenCL returns
// devices in order of acending PCI slot.
//
// On other systems, assume OpenCL and CUDA return devices
// in the same order.
//
while (1) {
if (current_CUDA_index >= (int)(nvidia_gpus.size())) {
snprintf(buf, sizeof(buf),
"OpenCL NVIDIA index #%d does not match any CUDA device",
device_index
);
warnings.push_back(buf);
return; // Should never happen
}
if (!strcmp(prop.name,
nvidia_gpus[devnums_pci_slot_sort[current_CUDA_index]].prop.name)
) {
break; // We have a match
}
// This CUDA GPU is not recognized by OpenCL,
// so try the next
//
++current_CUDA_index;
}
prop.device_num = devnums_pci_slot_sort[current_CUDA_index];
} else {
prop.device_num = (int)(nvidia_opencls.size());
}
prop.opencl_device_index = device_index;
if (nvidia.have_cuda) {
prop.peak_flops = nvidia_gpus[prop.device_num].peak_flops;
} else {
COPROC_NVIDIA c;
c.opencl_prop = prop;
c.set_peak_flops();
prop.peak_flops = c.peak_flops;
}
if (nvidia_gpus.size()) {
// Assumes OpenCL device_num and CUDA device_num now match
//
prop.opencl_available_ram = nvidia_gpus[prop.device_num].available_ram;
} else {
prop.opencl_available_ram = prop.global_mem_size;
}
// Build nvidia_opencls vector in device_num order
for (it=nvidia_opencls.begin(); it<nvidia_opencls.end(); it++) {
if (it->device_num > prop.device_num) break;
}
nvidia_opencls.insert(it, prop);
++current_CUDA_index;
}
//////////// AMD / ATI //////////////
if (is_AMD(prop.vendor)) {
prop.opencl_device_index = device_index;
if (ati.have_cal) {
// AMD OpenCL does not recognize all AMD GPUs returned by
// CAL but we assume that OpenCL and CAL return devices in
// the same order. See additional comments earlier in
// this source file for more details.
//
while (1) {
if (current_CAL_index >= num_CAL_devices) {
snprintf(buf, sizeof(buf),
"OpenCL ATI device #%d does not match any CAL device",
device_index
);
warnings.push_back(buf);
return; // Should never happen
}
if ((int)ati_gpus[current_CAL_index].attribs.target >= min_CAL_target) {
break; // We have a match
}
// This CAL GPU is not recognized by OpenCL,
// so try the next
//
++current_CAL_index;
}
prop.device_num = current_CAL_index++;
// Always use GPU model name from CAL if
// available for ATI / AMD GPUs because
// (we believe) it is more user-friendly.
//
safe_strcpy(prop.name, ati_gpus[prop.device_num].name);
// Work around a bug in OpenCL which returns only
// 1/2 of total global RAM size: use the value from CAL.
// This bug applies only to ATI GPUs, not to NVIDIA
// See also further workaround code for Macs.
//
prop.global_mem_size = ati_gpus[prop.device_num].attribs.localRAM * MEGA;
prop.peak_flops = ati_gpus[prop.device_num].peak_flops;
} else { // ! ati.have_cal
prop.device_num = (int)(ati_opencls.size());
COPROC_ATI c;
c.opencl_prop = prop;
c.set_peak_flops();
prop.peak_flops = c.peak_flops;
}
if (ati_gpus.size()) {
prop.opencl_available_ram = ati_gpus[prop.device_num].available_ram;
} else {
prop.opencl_available_ram = prop.global_mem_size;
}
ati_opencls.push_back(prop);
}
//////////// INTEL GPU //////////////
//
if (is_intel(prop.vendor)) {
prop.device_num = (int)(intel_gpu_opencls.size());
prop.opencl_device_index = device_index;
COPROC_INTEL c;
c.opencl_prop = prop;
c.is_used = COPROC_UNUSED;
c.available_ram = prop.global_mem_size;
safe_strcpy(c.name, prop.name);
safe_strcpy(c.version, prop.opencl_driver_version);
c.set_peak_flops();
prop.peak_flops = c.peak_flops;
prop.opencl_available_ram = prop.global_mem_size;
intel_gpu_opencls.push_back(prop);
// At present Intel GPUs only support OpenCL
// and do not have a native GPGPU framework,
// so treat each detected Intel OpenCL GPU device as
// a native device.
//
intel_gpus.push_back(c);
}
}
}
#ifdef __APPLE__
// Work around a bug in OpenCL which returns only
// 1/2 of total global RAM size.
// This bug applies only to ATI GPUs, not to NVIDIA
// This has already been fixed on latest Catalyst
// drivers, but Mac does not use Catalyst drivers.
if (ati_opencls.size() > 0) {
opencl_get_ati_mem_size_from_opengl(warnings);
}
#endif
if ((nvidia_opencls.size() == 0) &&
(ati_opencls.size() == 0) &&
(intel_gpu_opencls.size() == 0)
) {
warnings.push_back(
"OpenCL library present but no OpenCL-capable GPUs found"
);
}
}