Invalid Resource Handle on cuLaunchKernel

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I’m working on cuda 10.2 and I’m new in cude dev.
My purpose is to convert a float16 memory (mapped from Directx12 texture) in float32 memory.
(I base my code on sample vectorAdd.)

int numDevice = 0;
cudaGetDeviceCount(&numDevice);
bAvailable = numDevice > 0;
if (bAvailable)
{
	cudaSetDevice(0);

	char* ptx;
	size_t ptxSize;
	const char*kernel_file = "D:/Dev/Optro/features/debug_imGui/OptroView/Optroview/OptroViewCore/interop/float16_to_float32.cu";
/*compileFileToPTX and loadPTX are inspired from sample nvrtc_helper.h*/
	if (compileFileToPTX(kernel_file, 0, nullptr, &ptx, &ptxSize, true))
	{
		m_cudaModule = loadPTX(ptx, 0, nullptr);
		if(m_cudaModule)
			checkCudaErrors(cuModuleGetFunction(&m_kernelFun, m_cudaModule, "convertFloat16ToFloat32"));
	}
}

Variable m_cudaExternalMemoryPtr is the float16 memory mapped from DirectX texture.
Variable m_cudaExternalFloat32MemoryPtr is the float32 memory.
In the following code m_cudaExternalMemoryPtr is not null

if(m_cudaExternalFloat32MemoryPtr)
{
	if(m_memoryFloat32Size != 2 * m_memorySize)
	{
		cudaFree(m_cudaExternalFloat32MemoryPtr);
		m_cudaExternalFloat32MemoryPtr = nullptr;
		m_memoryFloat32Size = 0;
	}
}
	
if (m_cudaExternalFloat32MemoryPtr == nullptr)
{
	m_memoryFloat32Size = 2 * m_memorySize;
	cudaMalloc(&m_cudaExternalFloat32MemoryPtr, m_memoryFloat32Size);
}

size_t numElements = m_memoryFloat32Size / sizeof(float);
// Define block and grid dimensions
dim3 blockDim(256, 1, 1);
dim3 gridDim((numElements + blockDim.x - 1) / blockDim.x, 1, 1);

//int teated = 0;
CUdeviceptr d_NumElem;
checkCudaErrors(cuMemAlloc(&d_NumElem, sizeof(size_t)))
checkCudaErrors(cuMemcpyHtoD(d_NumElem, &numElements, sizeof(size_t)));

void* arr[] = { m_cudaExternalMemoryPtr, m_cudaExternalFloat32MemoryPtr,  &d_NumElem };
checkCudaErrors(cuLaunchKernel(a_kernelFun, gridDim.x, gridDim.y,
	gridDim.z, /* grid dim */
	blockDim.x, blockDim.y,
	blockDim.z, /* block dim */
	0, 0,            /* shared mem, stream */
	&arr[0],         /* arguments */
	0))

The kernel code is:

#include <cuda_fp16.h>

extern "C" __global__ void convertFloat16ToFloat32(const half* input, float* output, size_t numElements) {
    int tid = blockIdx.x * blockDim.x + threadIdx.x;
    if (tid < numElements) {
        output[tid] = __half2float(input[tid]);
    }
}

I have an invalid resource handle error on cuLaunchKernel.

Do you have an Idea ?
Thanks,
Sorry for my english

2

invalid resource handle often means that a resource you are using, such as a stream or memory handle, is not valid, or not valid for the device you are trying to do work on (many types of memory and all streams have an inherent device association). It’s quite possible there is some connection to the DX12 interop you may be doing.

There is not enough information here for me to be able to spot an issue. I usually don’t work on such problems in much detail unless a short and complete reproducing example code is given.

AFAIK the checkCudaErrors() helper function provided in NVIDIA sample codes is not designed for use with the driver API, although that may not have any relevance to what you are reporting.