00001
00021 #if defined(__cplusplus)
00022 extern "C" {
00023 #endif
00024
00025 /*
00026 * Ugly hacks to avoid dependencies on the real nvml.h until it starts
00027 * getting included with the CUDA toolkit or a GDK that's got a known
00028 * install location, etc.
00029 */
00030 typedef enum wrap_nvmlReturn_enum {
00031 WRAPNVML_SUCCESS = 0
00032 } wrap_nvmlReturn_t;
00033
00034 typedef void * wrap_nvmlDevice_t;
00035
00036 /* our own version of the PCI info struct */
00037 typedef struct {
00038 char bus_id_str[16]; /* string form of bus info */
00039 unsigned int domain;
00040 unsigned int bus;
00041 unsigned int device;
00042 unsigned int pci_device_id; /* combined device and vendor id */
00043 unsigned int pci_subsystem_id;
00044 unsigned int res0; /* NVML internal use only */
00045 unsigned int res1;
00046 unsigned int res2;
00047 unsigned int res3;
00048 } wrap_nvmlPciInfo_t;
00049
00050
00051 /*
00052 * Handle to hold the function pointers for the entry points we need,
00053 * and the shared library itself.
00054 */
00055 typedef struct {
00056 void *nvml_dll;
00057 int nvml_gpucount;
00058 int cuda_gpucount;
00059 unsigned int *nvml_pci_domain_id;
00060 unsigned int *nvml_pci_bus_id;
00061 unsigned int *nvml_pci_device_id;
00062 int *nvml_cuda_device_id; /* map NVML dev to CUDA dev */
00063 int *cuda_nvml_device_id; /* map CUDA dev to NVML dev */
00064 wrap_nvmlDevice_t *devs;
00065 wrap_nvmlReturn_t (*nvmlInit)(void);
00066 wrap_nvmlReturn_t (*nvmlDeviceGetCount)(int *);
00067 wrap_nvmlReturn_t (*nvmlDeviceGetHandleByIndex)(int, wrap_nvmlDevice_t *);
00068 wrap_nvmlReturn_t (*nvmlDeviceGetPciInfo)(wrap_nvmlDevice_t, wrap_nvmlPciInfo_t *);
00069 wrap_nvmlReturn_t (*nvmlDeviceGetName)(wrap_nvmlDevice_t, char *, int);
00070 wrap_nvmlReturn_t (*nvmlDeviceGetTemperature)(wrap_nvmlDevice_t, int, unsigned int *);
00071 wrap_nvmlReturn_t (*nvmlDeviceGetFanSpeed)(wrap_nvmlDevice_t, unsigned int *);
00072 wrap_nvmlReturn_t (*nvmlDeviceGetPowerUsage)(wrap_nvmlDevice_t, unsigned int *);
00073 wrap_nvmlReturn_t (*nvmlDeviceGetCpuAffinity)(wrap_nvmlDevice_t, unsigned int cpuSetSize, unsigned long *cpuSet);
00074 wrap_nvmlReturn_t (*nvmlDeviceSetCpuAffinity)(wrap_nvmlDevice_t);
00075 wrap_nvmlReturn_t (*nvmlShutdown)(void);
00076 } wrap_nvml_handle;
00077
00078
00079 wrap_nvml_handle * wrap_nvml_create();
00080 int wrap_nvml_destroy(wrap_nvml_handle *nvmlh);
00081
00082 /*
00083 * Query the number of GPUs seen by NVML
00084 */
00085 int wrap_nvml_get_gpucount(wrap_nvml_handle *nvmlh, int *gpucount);
00086
00087 /*
00088 * Query the number of GPUs seen by CUDA
00089 */
00090 int wrap_cuda_get_gpucount(wrap_nvml_handle *nvmlh, int *gpucount);
00091
00092
00093 /*
00094 * query the name of the GPU model from the CUDA device ID
00095 *
00096 */
00097 int wrap_nvml_get_gpu_name(wrap_nvml_handle *nvmlh,
00098 int gpuindex,
00099 char *namebuf,
00100 int bufsize);
00101
00102 /*
00103 * Query the current GPU temperature (Celsius), from the CUDA device ID
00104 */
00105 int wrap_nvml_get_tempC(wrap_nvml_handle *nvmlh,
00106 int gpuindex, unsigned int *tempC);
00107
00108 /*
00109 * Query the current GPU fan speed (percent) from the CUDA device ID
00110 */
00111 int wrap_nvml_get_fanpcnt(wrap_nvml_handle *nvmlh,
00112 int gpuindex, unsigned int *fanpcnt);
00113
00114 /*
00115 * Query the current GPU power usage in millwatts from the CUDA device ID
00116 *
00117 * This feature is only available on recent GPU generations and may be
00118 * limited in some cases only to Tesla series GPUs.
00119 * If the query is run on an unsupported GPU, this routine will return -1.
00120 */
00121 int wrap_nvml_get_power_usage(wrap_nvml_handle *nvmlh,
00122 int gpuindex,
00123 unsigned int *milliwatts);
00124
00125 /*
00126 * Query the current GPU's CPU affinity mask
00127 *
00128 * If the query is run on an unsupported GPU, this routine will return -1.
00129 */
00130 int wrap_nvml_get_cpu_affinity(wrap_nvml_handle *nvmlh,
00131 int gpuindex, unsigned int cpuSetSize,
00132 unsigned long *cpuSet);
00133
00134 /*
00135 * Set the CPU affinity mask for best access to the indexed GPU
00136 *
00137 * The routine returns -1 on failure.
00138 */
00139 int wrap_nvml_set_cpu_affinity(wrap_nvml_handle *nvmlh, int gpuindex);
00140
00141 #if defined(__cplusplus)
00142 }
00143 #endif
00144