func deviceGetByPCIBusId(busid string) (handle, error) {
var dev C.int
id := C.CString(busid)
r := C.cudaDeviceGetByPCIBusId(&dev, id)
C.free(unsafe.Pointer(id))
return handle{dev}, errorString(r)
}
func NewDevice(busID string) (*Device, error) {
var (
dev C.int
prop C.struct_cudaDeviceProp
)
id := C.CString(busID)
if err := cudaErr(C.cudaDeviceGetByPCIBusId(&dev, id)); err != nil {
return nil, err
}
C.free(unsafe.Pointer(id))
if err := cudaErr(C.cudaGetDeviceProperties(&prop, dev)); err != nil {
return nil, err
}
arch := fmt.Sprintf("%d.%d", prop.major, prop.minor)
cores, ok := archToCoresPerSM[arch]
if !ok {
return nil, fmt.Errorf("unsupported CUDA arch: %s", arch)
}
// Destroy the active CUDA context
cudaErr(C.cudaDeviceReset())
return &Device{
handle: dev,
Family: archToFamily[arch[:1]],
Arch: arch,
Cores: cores * uint(prop.multiProcessorCount),
Memory: MemoryInfo{
ECC: bool(prop.ECCEnabled != 0),
Global: uint(prop.totalGlobalMem / (1024 * 1024)),
Shared: uint(prop.sharedMemPerMultiprocessor / 1024),
Constant: uint(prop.totalConstMem / 1024),
L2Cache: uint(prop.l2CacheSize / 1024),
Bandwidth: 2 * uint((prop.memoryClockRate/1000)*(prop.memoryBusWidth/8)) / 1000,
},
}, nil
}