static int
hwloc_cuda_discover(struct hwloc_backend *backend)
{
struct hwloc_topology *topology = backend->topology;
enum hwloc_type_filter_e filter;
cudaError_t cures;
int nb, i;
hwloc_topology_get_type_filter(topology, HWLOC_OBJ_OS_DEVICE, &filter);
if (filter == HWLOC_TYPE_FILTER_KEEP_NONE)
return 0;
cures = cudaGetDeviceCount(&nb);
if (cures)
return -1;
for (i = 0; i < nb; i++) {
int domain, bus, dev;
char cuda_name[32];
char number[32];
struct cudaDeviceProp prop;
hwloc_obj_t cuda_device, parent;
unsigned cores;
cuda_device = hwloc_alloc_setup_object(HWLOC_OBJ_OS_DEVICE, -1);
snprintf(cuda_name, sizeof(cuda_name), "cuda%d", i);
cuda_device->name = strdup(cuda_name);
cuda_device->depth = (unsigned) HWLOC_TYPE_DEPTH_UNKNOWN;
cuda_device->attr->osdev.type = HWLOC_OBJ_OSDEV_COPROC;
cuda_device->subtype = strdup("CUDA");
hwloc_obj_add_info(cuda_device, "Backend", "CUDA");
hwloc_obj_add_info(cuda_device, "GPUVendor", "NVIDIA Corporation");
cures = cudaGetDeviceProperties(&prop, i);
if (!cures)
hwloc_obj_add_info(cuda_device, "GPUModel", prop.name);
snprintf(number, sizeof(number), "%llu", ((unsigned long long) prop.totalGlobalMem) >> 10);
hwloc_obj_add_info(cuda_device, "CUDAGlobalMemorySize", number);
snprintf(number, sizeof(number), "%llu", ((unsigned long long) prop.l2CacheSize) >> 10);
hwloc_obj_add_info(cuda_device, "CUDAL2CacheSize", number);
snprintf(number, sizeof(number), "%d", prop.multiProcessorCount);
hwloc_obj_add_info(cuda_device, "CUDAMultiProcessors", number);
cores = hwloc_cuda_cores_per_MP(prop.major, prop.minor);
if (cores) {
snprintf(number, sizeof(number), "%u", cores);
hwloc_obj_add_info(cuda_device, "CUDACoresPerMP", number);
}
snprintf(number, sizeof(number), "%llu", ((unsigned long long) prop.sharedMemPerBlock) >> 10);
hwloc_obj_add_info(cuda_device, "CUDASharedMemorySizePerMP", number);
parent = NULL;
if (hwloc_cudart_get_device_pci_ids(NULL /* topology unused */, i, &domain, &bus, &dev) == 0) {
parent = hwloc_pci_belowroot_find_by_busid(topology, domain, bus, dev, 0);
if (!parent)
parent = hwloc_pci_find_busid_parent(topology, domain, bus, dev, 0);
}
if (!parent)
parent = hwloc_get_root_obj(topology);
hwloc_insert_object_by_parent(topology, parent, cuda_device);
}
return 0;
}
static int
hwloc_cuda_backend_notify_new_object(struct hwloc_backend *backend,
struct hwloc_obj *pcidev)
{
struct hwloc_topology *topology = backend->topology;
struct hwloc_cuda_backend_data_s *data = backend->private_data;
unsigned i;
if (!(hwloc_topology_get_flags(topology) & (HWLOC_TOPOLOGY_FLAG_IO_DEVICES|HWLOC_TOPOLOGY_FLAG_WHOLE_IO)))
return 0;
if (!hwloc_topology_is_thissystem(topology)) {
hwloc_debug("%s", "\nno CUDA detection (not thissystem)\n");
return 0;
}
if (HWLOC_OBJ_PCI_DEVICE != pcidev->type)
return 0;
if (data->nr_devices == (unsigned) -1) {
/* first call, lookup all devices */
hwloc_cuda_query_devices(data);
/* if it fails, data->nr_devices = 0 so we won't do anything below and in next callbacks */
}
if (!data->nr_devices)
/* found no devices */
return 0;
for(i=0; i<data->nr_devices; i++) {
struct hwloc_cuda_device_info_s *info = &data->devices[i];
char cuda_name[32];
char number[32];
struct cudaDeviceProp prop;
hwloc_obj_t cuda_device;
cudaError_t cures;
unsigned cores;
if (info->pcidomain != pcidev->attr->pcidev.domain)
continue;
if (info->pcibus != pcidev->attr->pcidev.bus)
continue;
if (info->pcidev != pcidev->attr->pcidev.dev)
continue;
if (info->pcifunc != pcidev->attr->pcidev.func)
continue;
cuda_device = hwloc_alloc_setup_object(HWLOC_OBJ_OS_DEVICE, -1);
snprintf(cuda_name, sizeof(cuda_name), "cuda%d", info->idx);
cuda_device->name = strdup(cuda_name);
cuda_device->depth = (unsigned) HWLOC_TYPE_DEPTH_UNKNOWN;
cuda_device->attr->osdev.type = HWLOC_OBJ_OSDEV_COPROC;
hwloc_obj_add_info(cuda_device, "CoProcType", "CUDA");
hwloc_obj_add_info(cuda_device, "Backend", "CUDA");
hwloc_obj_add_info(cuda_device, "GPUVendor", "NVIDIA Corporation");
cures = cudaGetDeviceProperties(&prop, info->idx);
if (!cures)
hwloc_obj_add_info(cuda_device, "GPUModel", prop.name);
snprintf(number, sizeof(number), "%llu", ((unsigned long long) prop.totalGlobalMem) >> 10);
hwloc_obj_add_info(cuda_device, "CUDAGlobalMemorySize", number);
snprintf(number, sizeof(number), "%llu", ((unsigned long long) prop.l2CacheSize) >> 10);
hwloc_obj_add_info(cuda_device, "CUDAL2CacheSize", number);
snprintf(number, sizeof(number), "%d", prop.multiProcessorCount);
hwloc_obj_add_info(cuda_device, "CUDAMultiProcessors", number);
cores = hwloc_cuda_cores_per_MP(prop.major, prop.minor);
if (cores) {
snprintf(number, sizeof(number), "%u", cores);
hwloc_obj_add_info(cuda_device, "CUDACoresPerMP", number);
}
snprintf(number, sizeof(number), "%llu", ((unsigned long long) prop.sharedMemPerBlock) >> 10);
hwloc_obj_add_info(cuda_device, "CUDASharedMemorySizePerMP", number);
hwloc_insert_object_by_parent(topology, pcidev, cuda_device);
return 1;
}
return 0;
}
