1.我最终成功实现了opencv中利用cuvid实现GPU视频解码：

核心代码是：

1 cv::cuda::GpuMat d_frame; 2     cv::Ptr
     
       d_reader = cv::cudacodec::createVideoReader(mp4_file_name); 3 for (;;) 4 { 5     if (!d_reader->nextFrame(d_frame))  //BRGA格式 6         break; 7     gpu_frame_count++; 8     cv::Mat frame2; 9     d_frame.download(frame2);10     cv::imwrite("xxx.png", frame2);11 }
     

2.GupMat类的参考地址是：

  https://docs.opencv.org/master/d0/d60/classcv_1_1cuda_1_1GpuMat.html

  源码在： opencv-master/modules/core/include/opencv2/core/cuda.hpp

  GPUMat类的成员变量都是public的，就算没有提供访问的方法也没关系。

  一些重要的成员变量和成员函数是：

1 class CV_EXPORTS_W GpuMat 2 { 3 public: 4  5     /** @brief Performs data download from GpuMat (Blocking call) 6  7     This function copies data from device memory to host memory. As being a blocking call, it is 8     guaranteed that the copy operation is finished when this function returns. 9     */10     CV_WRAP void download(OutputArray dst) const;11 12     /** @brief Performs data download from GpuMat (Non-Blocking call)13 14     This function copies data from device memory to host memory. As being a non-blocking call, this15     function may return even if the copy operation is not finished.16 17     The copy operation may be overlapped with operations in other non-default streams if \p stream is18     not the default stream and \p dst is HostMem allocated with HostMem::PAGE_LOCKED option.19     */20     CV_WRAP void download(OutputArray dst, Stream& stream) const;21 22     //! the number of rows and columns23     int rows, cols;24 25     //! a distance between successive rows in bytes; includes the gap if any26     CV_PROP size_t step;27 28     //! pointer to the data29     uchar* data;30 31     //! helper fields used in locateROI and adjustROI32     uchar* datastart;33     const uchar* dataend;34 35 };

data是GPU内存中，存储图像数据的指针

datastart的地址与data相同

dataend指向图像存储空间的结束位置。（很可惜，这里是错误的）

rows 是图片的高度

cols是图片的宽度

channels() 返回4， 说明每个像素是四个字节，格式是BGRA

step是图片每行的字节数。注意：这个值是按2的幂对齐的。我测试中使用的图片，宽度是480，每像素四字节的话，一行应该是1920; 而此处的step值是2048， 每行多出来32像素，这些像素的alpha通道值为0。

因此，虽然看起来dataend-datastart是GPU内存所占空间大小，但实际的所占空间是：step*rows

 

3. GpuMat类使用dowmload()方法后，Mat类会去掉多余的对齐的像素

   具体怎么做到的呢？搜索了很久终于找到源码原来在：opencv-master/modules/core/src/cuda/gpu_mat.cu

   download方法的源码是：

1 void cv::cuda::GpuMat::download(OutputArray _dst) const2 {3     CV_DbgAssert( !empty() );4 5     _dst.create(size(), type());6     Mat dst = _dst.getMat();7 8     CV_CUDEV_SAFE_CALL( cudaMemcpy2D(dst.data, dst.step, data, step, cols * elemSize(), rows, cudaMemcpyDeviceToHost) );9 }

直接这样拷贝也是可以的：

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);

//调用各种函数来初始化cuda运行环境，否则一执行就崩溃