Shim layer of the cuda C-API. More...

#include <cuda.hpp>

Public Member Functions
	cudaAPI (cudaAPI const &)=delete

void	operator= (cudaAPI const &)=delete

Static Public Member Functions
static KVIKIO_EXPORT cudaAPI &	instance ()

static CUresult	cuda_memcpy_async (CUdeviceptr dst, CUdeviceptr src, std::size_t size, CUstream stream)
	Asynchronous memcpy that prefers `cuMemcpyBatchAsync` when supported. More...

Public Attributes
int	driver_version {0}

decltype(cuInit) *	Init {nullptr}

decltype(cuMemHostAlloc) *	MemHostAlloc {nullptr}

decltype(cuMemFreeHost) *	MemFreeHost {nullptr}

decltype(cuMemHostRegister) *	MemHostRegister {nullptr}

decltype(cuMemHostUnregister) *	MemHostUnregister {nullptr}

decltype(cuMemcpyHtoDAsync) *	MemcpyHtoDAsync {nullptr}

decltype(cuMemcpyDtoHAsync) *	MemcpyDtoHAsync {nullptr}

decltype(cuMemcpyAsync) *	MemcpyAsync {nullptr}

detail::AnyCallable	MemcpyBatchAsync {}

decltype(cuPointerGetAttribute) *	PointerGetAttribute {nullptr}

decltype(cuPointerGetAttributes) *	PointerGetAttributes {nullptr}

decltype(cuCtxPushCurrent) *	CtxPushCurrent {nullptr}

decltype(cuCtxPopCurrent) *	CtxPopCurrent {nullptr}

decltype(cuCtxGetCurrent) *	CtxGetCurrent {nullptr}

decltype(cuCtxGetDevice) *	CtxGetDevice {nullptr}

decltype(cuMemGetAddressRange) *	MemGetAddressRange {nullptr}

decltype(cuGetErrorName) *	GetErrorName {nullptr}

decltype(cuGetErrorString) *	GetErrorString {nullptr}

decltype(cuDeviceGet) *	DeviceGet {nullptr}

decltype(cuDeviceGetCount) *	DeviceGetCount {nullptr}

decltype(cuDeviceGetAttribute) *	DeviceGetAttribute {nullptr}

decltype(cuDevicePrimaryCtxRetain) *	DevicePrimaryCtxRetain {nullptr}

decltype(cuDevicePrimaryCtxRelease) *	DevicePrimaryCtxRelease {nullptr}

decltype(cuStreamSynchronize) *	StreamSynchronize {nullptr}

decltype(cuStreamCreate) *	StreamCreate {nullptr}

decltype(cuStreamDestroy) *	StreamDestroy {nullptr}

decltype(cuDriverGetVersion) *	DriverGetVersion {nullptr}

Detailed Description

Shim layer of the cuda C-API.

This is a singleton class that use dlopen on construction to load the C-API of cuda.

For example, cudaAPI::instance().MemHostAlloc() corresponds to calling cuMemHostAlloc()

Definition at line 78 of file cuda.hpp.

Member Function Documentation

◆ cuda_memcpy_async()

static CUresult kvikio::cudaAPI::cuda_memcpy_async	(	CUdeviceptr	dst,
		CUdeviceptr	src,
		std::size_t	size,
		CUstream	stream
	)

static

Asynchronous memcpy that prefers cuMemcpyBatchAsync when supported.

Dispatches to cuMemcpyBatchAsync with CU_MEMCPY_SRC_ACCESS_ORDER_STREAM on CUDA >= 12.8 when stream is non-default; otherwise falls back to cuMemcpyAsync. The fallback is mandatory on the default (NULL) stream, which cuMemcpyBatchAsync rejects.

Parameters

dst	Destination pointer (host or device under UVA).
src	Source pointer (host or device under UVA).
size	Number of bytes to copy.
stream	CUDA stream for ordering.

Returns: CUresult from the underlying driver call.

The documentation for this class was generated from the following file:

cuda.hpp

Public Member Functions

Static Public Member Functions

Public Attributes

Detailed Description

Member Function Documentation

◆ cuda_memcpy_async()