func (h handle) getDeviceProperties() (*deviceProp, error) {
var props C.struct_cudaDeviceProp
r := C.cudaGetDeviceProperties(&props, h.dev)
p := &deviceProp{
major: int(props.major),
minor: int(props.minor),
multiProcessorCount: uint(props.multiProcessorCount),
ECCEnabled: bool(props.ECCEnabled != 0),
totalGlobalMem: uint(props.totalGlobalMem),
sharedMemPerMultiprocessor: uint(props.sharedMemPerMultiprocessor),
totalConstMem: uint(props.totalConstMem),
l2CacheSize: uint(props.l2CacheSize),
memoryClockRate: uint(props.memoryClockRate),
memoryBusWidth: uint(props.memoryBusWidth),
}
return p, 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
}
// Returns the device properties
func GetDeviceProperties(device int) *DeviceProp {
var prop C.struct_cudaDeviceProp
err := Error(C.cudaGetDeviceProperties(&prop, C.int(device)))
if err != Success {
panic(err)
}
devProp := new(DeviceProp)
devProp.Name = C.GoString(&prop.name[0])
devProp.TotalGlobalMem = uint(prop.totalGlobalMem)
devProp.SharedMemPerBlock = uint(prop.sharedMemPerBlock)
devProp.RegsPerBlock = int(prop.regsPerBlock)
devProp.WarpSize = int(prop.warpSize)
devProp.MemPitch = int(prop.memPitch)
devProp.MaxThreadsPerBlock = int(prop.maxThreadsPerBlock)
devProp.MaxThreadsDim[0] = int(prop.maxThreadsDim[0])
devProp.MaxThreadsDim[1] = int(prop.maxThreadsDim[1])
devProp.MaxThreadsDim[2] = int(prop.maxThreadsDim[2])
devProp.MaxGridSize[0] = int(prop.maxGridSize[0])
devProp.MaxGridSize[1] = int(prop.maxGridSize[1])
devProp.MaxGridSize[2] = int(prop.maxGridSize[2])
devProp.TotalConstMem = uint(prop.totalConstMem)
devProp.Major = int(prop.major)
devProp.Minor = int(prop.minor)
devProp.ClockRate = int(prop.clockRate)
devProp.TextureAlignment = int(prop.textureAlignment)
devProp.DeviceOverlap = int(prop.deviceOverlap)
devProp.MultiProcessorCount = int(prop.multiProcessorCount)
devProp.KernelExecTimeoutEnabled = int(prop.kernelExecTimeoutEnabled)
devProp.Integrated = int(prop.integrated)
devProp.CanMapHostMemory = int(prop.canMapHostMemory)
devProp.ComputeMode = int(prop.computeMode)
devProp.ConcurrentKernels = int(prop.concurrentKernels)
devProp.ECCEnabled = int(prop.ECCEnabled)
devProp.PciBusID = int(prop.pciBusID)
devProp.PciDeviceID = int(prop.pciDeviceID)
devProp.TccDriver = int(prop.tccDriver)
return devProp
}
