CUDA Host Function

Submits a user-defined function that receives the stream's CUDA stream handle, so that custom CUDA kernels or external library CUDA calls can run in the same order as VPI work (including Tegra backends). More...

Typedefs
typedef void(*	VPICUDAHostFunction) (cudaStream_t cudaStream, void *userData)
	Pointer to a CUDA host function invoked with the stream's CUDA stream. More...

Functions
VPIStatus	vpiSubmitCUDAHostFunction (VPIStream stream, VPICUDAHostFunction hostFunc, void *userData)
	Submits a CUDA host function to run on the stream in submission order. More...

Detailed Description

Submits a user-defined function that receives the stream's CUDA stream handle, so that custom CUDA kernels or external library CUDA calls can run in the same order as VPI work (including Tegra backends).

Typedef Documentation

VPICUDAHostFunction

typedef void(* VPICUDAHostFunction) (cudaStream_t cudaStream, void *userData)

#include <vpi/CUDAInterop.h>

Pointer to a CUDA host function invoked with the stream's CUDA stream.

The callback runs on VPI's stream worker thread after all prior commands (VPI algorithms, Tegra backends, and previous callbacks) have been submitted. Launch custom CUDA kernels or call external libraries (e.g. cuBLAS, cuFFT) using cudaStream so that they execute in submission order with VPI work.

Parameters

[in]	cudaStream	The CUDA stream for this VPIStream's CUDA backend.
[in]	userData	Callback data passed to the submission function.

Definition at line 93 of file CUDAInterop.h.

Function Documentation

vpiSubmitCUDAHostFunction()

VPIStatus vpiSubmitCUDAHostFunction	(	VPIStream	stream,
		VPICUDAHostFunction	hostFunc,
		void *	userData
	)

#include <vpi/CUDAInterop.h>

Submits a CUDA host function to run on the stream in submission order.

The callback is executed after all prior commands on the stream (including VPI Tegra and CUDA work) have been submitted. The callback receives the stream's CUDA stream so the user can launch custom kernels or call external CUDA libraries on it. Subsequent VPI submissions run after the callback, preserving order.

The stream must have the CUDA backend enabled (e.g. created with vpiStreamCreate or vpiStreamCreateWrapperCUDA with CUDA enabled). This call implicitly flushes the stream before enqueueing the callback.

Parameters

[in]	stream	Stream that will execute the callback (mandatory).
[in]	hostFunc	Callback to run with the CUDA stream (mandatory).
[in]	userData	Pointer passed unchanged to hostFunc.

Note

userData must remain valid until the callback has executed (for example, after vpiStreamSync).

Return values

VPI_ERROR_INVALID_ARGUMENT	stream or hostFunc is NULL.
VPI_ERROR_INVALID_OPERATION	Stream does not have CUDA backend enabled.
VPI_ERROR_INVALID_OPERATION	VPI not built with CUDA support.
VPI_ERROR_INVALID_CONTEXT	Current context was destroyed.
VPI_SUCCESS	Callback enqueued successfully.