nanovdb::util::cuda Namespace Reference

Classes
class	Timer
class	unique_ptr
	Simple (naive) implementation of a unique device pointer using stream ordered memory allocation and deallocation. More...

Functions
bool	memoryPoolsSupported (int device)
	Returns true if device supports stream-ordered memory pools, i.e. cudaMallocAsync/cudaFreeAsync. Queried once per process for all devices and cached; out-of-range device ids return false.
cudaError_t	mallocAsync (void **d_ptr, size_t size, cudaStream_t)
	Wrapper forced to synchronous cudaMalloc; see the mode comment above. The trailing stream argument is accepted for signature compatibility and ignored.
cudaError_t	freeAsync (void *d_ptr, cudaStream_t)
	Wrapper forced to synchronous cudaFree; see the mode comment above. The trailing stream argument is accepted for signature compatibility and ignored.
int	ptrToDevice (void *ptr)
	Returns the device ID associated with the specified pointer.
int	currentDevice ()
	Returns the ID of the current device.
int	deviceCount ()
	Returns the number of devices with compute capability greater or equal to 1.0 that are available for execution.
void	printDevInfo (int device, const char *preMsg=nullptr, std::FILE *file=stderr)
	Print information about a specific device.
size_t	blocksPerGrid (size_t numItems, size_t threadsPerBlock)
	Computes the number of blocks per grid given the problem size and number of threads per block.
cudaError_t	memAdvise (const void *devPtr, size_t count, cudaMemoryAdvise advice, int device)
	Compatbility wrapper for cudaMemAdvise/cudaMemAdvise.
cudaError_t	memPrefetchAsync (const void *devPtr, size_t count, int dstDevice, cudaStream_t stream)
	Compatbility wrapper for cudaMemPrefetchAsync/cudaMemPrefetchAsync.

Function Documentation

blocksPerGrid()

size_t blocksPerGrid ( size_t numItems, size_t threadsPerBlock )

inline

Computes the number of blocks per grid given the problem size and number of threads per block.

Parameters

numItems	Problem size
threadsPerBlock	Number of threads per block (second CUDA launch parameter)

Returns

number of blocks per grid (first CUDA launch parameter)

Note

CUDA launch parameters: kernel<<< blocksPerGrid, threadsPerBlock, sharedMemSize, streamID>>>

currentDevice()

int currentDevice ( )

inline

Returns the ID of the current device.

deviceCount()

int deviceCount ( )

inline

Returns the number of devices with compute capability greater or equal to 1.0 that are available for execution.

freeAsync()

cudaError_t freeAsync ( void * d_ptr, cudaStream_t  )

inline

Wrapper forced to synchronous cudaFree; see the mode comment above. The trailing stream argument is accepted for signature compatibility and ignored.

Parameters

d_ptr

Device pointer that will be freed

Returns

Cuda error code

mallocAsync()

cudaError_t mallocAsync ( void ** d_ptr, size_t size, cudaStream_t  )

inline

Wrapper forced to synchronous cudaMalloc; see the mode comment above. The trailing stream argument is accepted for signature compatibility and ignored.

Parameters

d_ptr	Device pointer to allocated device memory
size	Number of bytes to allocate

Returns

Cuda error code

memAdvise()

cudaError_t memAdvise ( const void * devPtr, size_t count, cudaMemoryAdvise advice, int device )

inline

Compatbility wrapper for cudaMemAdvise/cudaMemAdvise.

memoryPoolsSupported()

bool memoryPoolsSupported ( int device )

inline

Returns true if device supports stream-ordered memory pools, i.e. cudaMallocAsync/cudaFreeAsync. Queried once per process for all devices and cached; out-of-range device ids return false.

memPrefetchAsync()

cudaError_t memPrefetchAsync ( const void * devPtr, size_t count, int dstDevice, cudaStream_t stream )

inline

Compatbility wrapper for cudaMemPrefetchAsync/cudaMemPrefetchAsync.

printDevInfo()

void printDevInfo ( int device, const char * preMsg = nullptr, std::FILE * file = stderr )

inline

Print information about a specific device.

Parameters

device	device ID for which information will be printed
preMsg	optional message printed before the device information
file	Optional file stream to print to, e.g. stderr or stdout

ptrToDevice()

int ptrToDevice ( void * ptr )

inline

Returns the device ID associated with the specified pointer.

Note

If ptr points to host memory (only) the return ID is either cudaInvalidDeviceId = -2 or cudaCpuDeviceId = -1