00001
00002
00003
00004
00005
00006
00007
00016 #ifndef HWLOC_CUDA_H
00017 #define HWLOC_CUDA_H
00018
00019 #include "hwloc.h"
00020 #include "hwloc/autogen/config.h"
00021 #include "hwloc/helper.h"
00022 #ifdef HWLOC_LINUX_SYS
00023 #include "hwloc/linux.h"
00024 #endif
00025
00026 #include <cuda.h>
00027
00028
00029 #ifdef __cplusplus
00030 extern "C" {
00031 #endif
00032
00033
00046 static __hwloc_inline int
00047 hwloc_cuda_get_device_pci_ids(hwloc_topology_t topology __hwloc_attribute_unused,
00048 CUdevice cudevice, int *domain, int *bus, int *dev)
00049 {
00050 CUresult cres;
00051
00052 #if CUDA_VERSION >= 4000
00053 cres = cuDeviceGetAttribute(domain, CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID, cudevice);
00054 if (cres != CUDA_SUCCESS) {
00055 errno = ENOSYS;
00056 return -1;
00057 }
00058 #else
00059 *domain = 0;
00060 #endif
00061 cres = cuDeviceGetAttribute(bus, CU_DEVICE_ATTRIBUTE_PCI_BUS_ID, cudevice);
00062 if (cres != CUDA_SUCCESS) {
00063 errno = ENOSYS;
00064 return -1;
00065 }
00066 cres = cuDeviceGetAttribute(dev, CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID, cudevice);
00067 if (cres != CUDA_SUCCESS) {
00068 errno = ENOSYS;
00069 return -1;
00070 }
00071
00072 return 0;
00073 }
00074
00091 static __hwloc_inline int
00092 hwloc_cuda_get_device_cpuset(hwloc_topology_t topology __hwloc_attribute_unused,
00093 CUdevice cudevice, hwloc_cpuset_t set)
00094 {
00095 #ifdef HWLOC_LINUX_SYS
00096
00097 #define HWLOC_CUDA_DEVICE_SYSFS_PATH_MAX 128
00098 char path[HWLOC_CUDA_DEVICE_SYSFS_PATH_MAX];
00099 int domainid, busid, deviceid;
00100
00101 if (hwloc_cuda_get_device_pci_ids(topology, cudevice, &domainid, &busid, &deviceid))
00102 return -1;
00103
00104 if (!hwloc_topology_is_thissystem(topology)) {
00105 errno = EINVAL;
00106 return -1;
00107 }
00108
00109 sprintf(path, "/sys/bus/pci/devices/%04x:%02x:%02x.0/local_cpus", domainid, busid, deviceid);
00110 if (hwloc_linux_read_path_as_cpumask(path, set) < 0
00111 || hwloc_bitmap_iszero(set))
00112 hwloc_bitmap_copy(set, hwloc_topology_get_complete_cpuset(topology));
00113 #else
00114
00115 hwloc_bitmap_copy(set, hwloc_topology_get_complete_cpuset(topology));
00116 #endif
00117 return 0;
00118 }
00119
00130 static __hwloc_inline hwloc_obj_t
00131 hwloc_cuda_get_device_pcidev(hwloc_topology_t topology, CUdevice cudevice)
00132 {
00133 int domain, bus, dev;
00134
00135 if (hwloc_cuda_get_device_pci_ids(topology, cudevice, &domain, &bus, &dev))
00136 return NULL;
00137
00138 return hwloc_get_pcidev_by_busid(topology, domain, bus, dev, 0);
00139 }
00140
00156 static __hwloc_inline hwloc_obj_t
00157 hwloc_cuda_get_device_osdev(hwloc_topology_t topology, CUdevice cudevice)
00158 {
00159 hwloc_obj_t osdev = NULL;
00160 int domain, bus, dev;
00161
00162 if (hwloc_cuda_get_device_pci_ids(topology, cudevice, &domain, &bus, &dev))
00163 return NULL;
00164
00165 osdev = NULL;
00166 while ((osdev = hwloc_get_next_osdev(topology, osdev)) != NULL) {
00167 hwloc_obj_t pcidev = osdev->parent;
00168 if (strncmp(osdev->name, "cuda", 4))
00169 continue;
00170 if (pcidev
00171 && pcidev->type == HWLOC_OBJ_PCI_DEVICE
00172 && (int) pcidev->attr->pcidev.domain == domain
00173 && (int) pcidev->attr->pcidev.bus == bus
00174 && (int) pcidev->attr->pcidev.dev == dev
00175 && pcidev->attr->pcidev.func == 0)
00176 return osdev;
00177
00178 }
00179
00180 return NULL;
00181 }
00182
00198 static __hwloc_inline hwloc_obj_t
00199 hwloc_cuda_get_device_osdev_by_index(hwloc_topology_t topology, unsigned idx)
00200 {
00201 hwloc_obj_t osdev = NULL;
00202 while ((osdev = hwloc_get_next_osdev(topology, osdev)) != NULL) {
00203 if (HWLOC_OBJ_OSDEV_COPROC == osdev->attr->osdev.type
00204 && osdev->name
00205 && !strncmp("cuda", osdev->name, 4)
00206 && atoi(osdev->name + 4) == (int) idx)
00207 return osdev;
00208 }
00209 return NULL;
00210 }
00211
00215 #ifdef __cplusplus
00216 }
00217 #endif
00218
00219
00220 #endif