linux + opencv + cuvid中使用cv::cuda::GpuMat類的一些坑

1.我最終成功實現了opencv中利用cuvid實現GPU視頻解碼：
核心代碼是：

cv::cuda::GpuMat d_frame;
    cv::Ptr<cv::cudacodec::VideoReader> d_reader = cv::cudacodec::createVideoReader(mp4_file_name);
for (;;)
{
    if (!d_reader->nextFrame(d_frame))  //BRGA格式
        break;
    gpu_frame_count++;
    cv::Mat frame2;
    d_frame.download(frame2);
    cv::imwrite("xxx.png", frame2);
}

2.GupMat類的參考地址是：
  https://docs.opencv.org/master/d0/d60/classcv_1_1cuda_1_1GpuMat.html

  源碼在： opencv-master/modules/core/include/opencv2/core/cuda.hpp

  GPUMat類的成員變量都是public的，就算沒有提供訪問的方法也沒關系。

  一些重要的成員變量和成員函數是：

class CV_EXPORTS_W GpuMat
{
public:

    /** @brief Performs data download from GpuMat (Blocking call)

    This function copies data from device memory to host memory. As being a blocking call, it is
    guaranteed that the copy operation is finished when this function returns.
    */
    CV_WRAP void download(OutputArray dst) const;

    /** @brief Performs data download from GpuMat (Non-Blocking call)

    This function copies data from device memory to host memory. As being a non-blocking call, this
    function may return even if the copy operation is not finished.

    The copy operation may be overlapped with operations in other non-default streams if \p stream is
    not the default stream and \p dst is HostMem allocated with HostMem::PAGE_LOCKED option.
    */
    CV_WRAP void download(OutputArray dst, Stream& stream) const;

    //! the number of rows and columns
    int rows, cols;

    //! a distance between successive rows in bytes; includes the gap if any
    CV_PROP size_t step;

    //! pointer to the data
    uchar* data;

    //! helper fields used in locateROI and adjustROI
    uchar* datastart;
    const uchar* dataend;

};

data是GPU內存中，存儲圖像數據的指針

datastart的地址與data相同

dataend指向圖像存儲空間的結束位置。（很可惜，這里是錯誤的）

rows 是圖片的高度

cols是圖片的寬度

channels() 返回4， 說明每個像素是四個字節，格式是BGRA

step是圖片每行的字節數。注意：這個值是按2的冪對齊的。我測試中使用的圖片，寬度是480，每像素四字節的話，一行應該是1920; 而此處的step值是2048， 每行多出來32像素，這些像素的alpha通道值為0。

因此，雖然看起來dataend-datastart是GPU內存所占空間大小，但實際的所占空間是：step*rows

 

3. GpuMat類使用dowmload()方法后，Mat類會去掉多余的對齊的像素
   具體怎么做到的呢？搜索了很久終於找到源碼原來在：opencv-master/modules/core/src/cuda/gpu_mat.cu

   download方法的源碼是：

void cv::cuda::GpuMat::download(OutputArray _dst) const
{
    CV_DbgAssert( !empty() );

    _dst.create(size(), type());
    Mat dst = _dst.getMat();

    CV_CUDEV_SAFE_CALL( cudaMemcpy2D(dst.data, dst.step, data, step, cols * elemSize(), rows, cudaMemcpyDeviceToHost) );
}

直接這樣拷貝也是可以的：
cudaMemcpy(host_data, d_frame.data,  d_frame.rows * d_frame.step , cudaMemcpyDeviceToHost);
但要注意：
#include <cuda_runtime.h>
cudaGetDeviceCount(&num_devices);
cudaSetDevice(cuda_device);
//調用各種函數來初始化cuda運行環境，否則一執行就崩潰