CudaMemCpy在复制向量时返回cudaErrorInvalidValue <cv::Point3f>

Question

CudaMemCpy returns cudaErrorInvalidValue on copying vector onto the device. I have tried giving "&input", "&input[0]",... I always get the same error but don't understand why?

Can you copy a vector using cudaMemcpy or do I need to copy the contents of that vector in a new array first?

void computeDepthChangeMap(unsigned char* depthChangeMap, size_t size, std::vector<cv::Point3f>* input, float dcf, int width, int height)                                           {
    unsigned char* dev_depthChangeMap = 0;
    float* dev_dcf = 0;
    int* dev_wdt = 0;
    int arraySize = size;
    cv::Point3f* dev_input = 0;
    cudaError_t cudaStatus;

    cudaStatus = cudaSetDevice(0);
    cudaStatus = cudaMalloc((void**)&dev_depthChangeMap, size);
    cudaStatus = cudaMalloc((void**)&dev_input, size);
    cudaStatus = cudaMalloc((void**)&dev_dcf, sizeof(float));
    cudaStatus = cudaMalloc((void**)&dev_wdt, sizeof(int));

    cudaStatus = cudaMemcpy(dev_depthChangeMap, depthChangeMap, size, cudaMemcpyHostToDevice);
    cudaStatus = cudaMemcpy(dev_wdt, &width, sizeof(int), cudaMemcpyHostToDevice);
    cudaStatus = cudaMemcpy(dev_dcf, &dcf, sizeof(float), cudaMemcpyHostToDevice);
    cudaStatus = cudaMemcpy(dev_input, &input[0], sizeof(cv::Point3f)*size, cudaMemcpyHostToDevice);

    //cuaStatus returns cudaErrorInvalidValue >> PROBLEM HERE << 

    dim3 threadsPerBlock(8, 8); //init x, y
    dim3 numBlocks(width / threadsPerBlock.x, height / threadsPerBlock.y);

    addKernel <<<numBlocks, threadsPerBlock >>>(dev_depthChangeMap, dev_dcf, dev_input, dev_wdt);


    cudaStatus = cudaGetLastError();   
    cudaStatus = cudaDeviceSynchronize();
    cudaStatus = cudaMemcpy(depthChangeMap, dev_depthChangeMap, size, cudaMemcpyDeviceToHost);
}

__global__ void addKernel(unsigned char* dev_depthChangeMap, float* dcf, cv::Point3f* inp, int* wdt)
{
    register int row_idx = (blockIdx.x * blockDim.x) + threadIdx.x;
    register int col_idx = (blockIdx.y * blockDim.y) + threadIdx.y;
    register int idx = row_idx * (*wdt) + col_idx;

    register float depth = inp[idx].z;
    register float depthR = inp[idx + 1].z;
    register float depthD = inp[idx + *wdt].z;

    //and so on

}

Answer 1

Yes, you can copy from std::vector using cudaMemcpy .

You don't have your sizes set up correctly:

void computeDepthChangeMap(unsigned char* depthChangeMap, size_t size, std::vector<cv::Point3f>* input, float dcf, int width, int height)                                           {

...
cudaStatus = cudaMalloc((void**)&dev_input, size);
                                            ^^^^

cudaStatus = cudaMemcpy(dev_input, &input[0], sizeof(cv::Point3f)*size, cudaMemcpyHostToDevice);
                                                     ^^^^^^^^^^^^^^^^^

These size parameters should all be in bytes . You can't copy data of length sizeof(cv::Point3f)*size bytes into an allocation of length size bytes.

Also, it seems that your function parameter is a pointer to a vector:

std::vector<cv::Point3f>* input,

based on the code you have shown, this is probably not what you want. You probably either want to pass the vector by value :

std::vector<cv::Point3f> input,

or more likely, by reference :

std::vector<cv::Point3f> &input,

Since you haven't shown how you intend to call this function, it's not possible to be entirely sure what is best here.