Question

我正在尝试使用带有cuda和cufft库的opencv优化代码。每次执行快速傅立叶变换时，都必须从GpuMat下载cv :: Mat，然后执行cufft。（请参见下面的代码），然后再次下载fft的结果。有什么办法可以优化这个吗？我想知道是否可以通过任何方式直接通过GpuMat，而不必下载它。

std::vector<cv::cuda::GpuMat> ReconClass::FFT2(std::vector<cv::cuda::GpuMat>& mat, int height, int width)
{
    cufftHandle plan;
    cufftComplex* data, * datao, * devdata, * devdatao;

    size_t arraySize = sizeof(cufftComplex) * mat[0].size().area();
    cudaMallocHost((void**)& data, arraySize);
    cudaMallocHost((void**)& datao, arraySize);

    cudaMalloc((void**)& devdata, arraySize);
    cudaMalloc((void**)& devdatao, arraySize);

    cv::Mat iReal;
    cv::Mat iImag;
    mat[0].download(iReal);
    mat[1].download(iImag);

    for (int i = 0; i < height; i++)
    {
        for (int j = 0; j < width; j++)
        {
            data[i * width + j].x = iReal.at<float>(i, j);
            data[i * width + j].y = iImag.at<float>(i, j);
        }
    }
    cudaMemcpy(devdata, data, arraySize, cudaMemcpyHostToDevice);

    cufftPlan2d(&plan, height, width, CUFFT_C2C);

    if (!plan)
        std::cout << "the cufftPlan2d plan returned is null" << std::endl;

    cufftExecC2C(plan, (cufftComplex*)devdata, (cufftComplex*)devdatao, CUFFT_FORWARD);

    cudaMemcpy(datao, devdatao, arraySize, cudaMemcpyDeviceToHost);

    cv::Mat realRecon(height, width, CV_32F);
    cv::Mat imagRecon(height, width, CV_32F);

    for (int i = 0; i < height; i++)
    {
        for (int j = 0; j < width; j++)
        {
            realRecon.at<float>(i, j) = datao[i * width + j].x;
            imagRecon.at<float>(i, j) = datao[i * width + j].y;
        }
    }

    cv::cuda::GpuMat mat1, mat2;
    mat1.upload(realRecon);
    mat2.upload(imagRecon);

    std::vector<cv::cuda::GpuMat> re = { mat1 , mat2 };

    cufftDestroy(plan);
    cudaFreeHost(data);
    cudaFreeHost(datao);
    cudaFree(devdata);
    cudaFree(devdatao);

    return re;
}

Answer 1

我能够避免将其复制到CPU并将其复制回设备。此外，进行原位FFT也有助于提高性能。我在下面粘贴了我的代码。

void Dataransfer2Cuda(const cv::InputArray _dReal, const cv::InputArray _dImag, float2* zCufftcomplex)
{
    const cv::cuda::GpuMat Real = _dReal.getGpuMat();
    const cv::cuda::GpuMat Imag = _dImag.getGpuMat();

    dim3 cthreads(32, 32);
    dim3 cblocks(
        static_cast<int>(std::ceil(Real.size().width /
            static_cast<double>(cthreads.x))),
        static_cast<int>(std::ceil(Real.size().height /
            static_cast<double>(cthreads.y))));

    Kernel_DataTransfer2Cuda << <cblocks, cthreads >> > (Real, Imag, zCufftcomplex);

    if (cudaSuccess != cudaGetLastError())
        std::cout << "Dataransfer2Cuda(): gave an error" << std::endl;

    return;
}

void DataransferFromCuda(const float2* zCufftcomplex, cv::OutputArray _outputReal, cv::OutputArray _outputImag, std::size_t iWidth, std::size_t iHeight)
{

    _outputReal.create(iHeight, iWidth, CV_32F);
    _outputImag.create(iHeight, iWidth, CV_32F);

    cv::cuda::GpuMat outputReal = _outputReal.getGpuMat();
    cv::cuda::GpuMat outputImag = _outputImag.getGpuMat();


    dim3 cthreads(32, 32);
    dim3 cblocks(
        static_cast<int>(std::ceil(outputReal.size().width /
            static_cast<double>(cthreads.x))),
        static_cast<int>(std::ceil(outputReal.size().height /
            static_cast<double>(cthreads.y))));

    Kernel_DataTransferFromCuda << <cblocks, cthreads >> > (zCufftcomplex, outputReal, outputImag);

    if (cudaSuccess != cudaGetLastError())
        std::cout << "DataransferFromCuda(): gave an error" << std::endl;

    return;
}

std::vector<cv::cuda::GpuMat> ReconClass::FFT2(std::vector<cv::cuda::GpuMat>& mat, int height, int width)
{
    cufftHandle plan;
    cufftComplex* devdata;

    size_t arraySize = sizeof(cufftComplex) * mat[0].size().area();

    cudaMalloc((void**)& devdata, arraySize);

    Dataransfer2Cuda(mat[0], mat[1], devdata);

    cufftPlan2d(&plan, height, width, CUFFT_C2C);

    if (!plan)
        std::cout << "the cufftPlan2d plan returned is null" << std::endl;

    cufftExecC2C(plan, (cufftComplex*)devdata, (cufftComplex*)devdata, CUFFT_FORWARD);

    cv::cuda::GpuMat mat1, mat2;
    DataransferFromCuda(devdata, mat1, mat2, width, height);
    std::vector<cv::cuda::GpuMat> re = { mat1 , mat2 };

    cufftDestroy(plan);
    cudaFree(devdata);

    return re;
}

将GpuMat直接传递给cufftExecC2C函数以进行快速傅立叶变换

1 个答案: