以下是开发过程中遇到的C++的一些特性记录. tags: C/C++, Decode, GPU, Python category: Documents

当前目的开发c++硬解码并使用python调用其动态库实现一次抽一次帧.

以下是开发过程中遇到的C++的一些特性记录.

1. 二级指针初始化与释放

• 二级指针需要初始化两次, 分别是二级指针和一级指针

  int **a;
  a = (int *)malloc(num * sizeof(int));
  for (int32_t i = 0; i < num; i++)
  {
  	a[i] = (int)malloc(num_2 * sizeof(int));
  }
  // or
  int **a;
  a = new int*[num];
  for (int32_t i = 0; i < num; i++)
  {
  	a[i] = new int[num_2];
  }
  

• 释放指针, 需要先释放一级指针再释放二级指针, 如果先释放二级指针, 一级指针的地址会丢失, 导致内存泄漏.

  for(i = 0; i < 100; i++)
  {
  　delete []temp[i];
  }
  delete []temp;
  

2. C++结构体与python传值

• c++定义结构体对象

  struct VideoPreprocessInf
  {
      int64_t frame_index = 0;
      int32_t ori_w = 0;
      int32_t ori_h = 0;
      int32_t resize_w = 0;
      int32_t resize_h = 0;
      int32_t img_w = 0;
      int32_t img_h = 0;
      int32_t pad_w = 0;
      int32_t pad_h = 0;
      int32_t roi_start_w = 0;
      int32_t roi_start_h = 0;
      int32_t roi_w = 0;
      int32_t roi_h = 0;
      int32_t channel = 3;
      int32_t output_size = 0;
      bool do_crop_middle = false;
      bool do_padding = false;
      float scale = 1.f;
      double mean[3] = {0.0f, 0.0f, 0.0f};
      double standard[3] = {1.f, 1.f, 1.f};
  
  } video_preprocess_inf;
  

• python定义同结构体对象

  class VideoPreprocessInf(Structure):
      _fields_ = (
          ('frame_index', c_longlong),
          ('ori_w ', c_int),
          ('ori_h ', c_int),
          ('resize_w ', c_int),
          ('resize_h ', c_int),
          ('img_w ', c_int),
          ('img_h ', c_int),
          ('pad_w ', c_int),
          ('pad_h ', c_int),
          ('roi_start_w ', c_int),
          ('roi_start_h ', c_int),
          ('roi_w ', c_int),
          ('roi_h ', c_int),
          ('channel', c_int),
          ('output_size', c_int),
          ('do_crop_middle', c_bool),
          ('do_padding', c_bool),
          ('scale', c_float),
          ('mean', c_double * 3),
          ('standard', c_double * 3),
      )
  

1. python端初始化结构体示例

   • 之后就可以输入C++动态库函数

     inf_list = [0, 0, 0, resize_w, resize_h, 0, 0, 0, 0, 0, 0, 0, 0,
                     channel, 0, do_crop_middle, do_padding, 1.0, mean, standard]
     video_preprocess_inf = VideoPreprocessInf(*inf_list)
     

2. C++端初始化结构体

   • 全局结构体只需要在结构体结尾加上变量, 其他地方只要使用变量即可

     struct VideoPreprocessInf
     {
         int64_t frame_index = 0;
     	  ...
     } video_preprocess_inf;
     

   • 需要当成调用多次的时候初始化:

     struct VideoPreprocessInf
     {
         int64_t frame_index = 0;
     	  ...
     };
     
     //初始化
     VideoPreprocessInf vpi = {
     		.frame_index = 1,
     		...
     };
     // 或者
     VideoPreprocessInf vpi = {
     		frame_index:1,
     		...
     };
     // 或者给需要修改的元素各自赋值
     VideoPreprocessInf vpi;
     vpi.frame_index = 1;
     ...
     

3. C++头文件使用C风格的函数

extern "C"
{
	int xxxx();
}

4. struct和typedef struct的区别

struct和typedef struct彻底明白了

Difference between ‘struct’ and ’typedef struct’ in C++?

5. 实现python调用c++单步抽帧

• 使用全局指针, 函数中全局指针指向new class demuxer 和 decoder

  Demuxer *demuxer_p = nullptr;
  DlDecoder *decoder_p = nullptr;
  
  VideoInitInf init_demuxer_decoder(char *inputFilename, int gpu_id)
  { // 初始化数据流提取器和解码器, 使用全局指针指向它们
    demuxer_p = new Demuxer(inputFilename);
    decoder_p = new DlDecoder(gpu_id, 0xf, 0xf, true, demuxer_p->GetVideoCodec());
  	...
  }
  
  float *video_decoder(VideoPreprocessInf video_preprocess_inf, bool do_extract)
  { // 单步抽帧
  	demuxer_p->Demux(&pVideo, &nVideoBytes);
    decoder_p->Decode(pVideo, nVideoBytes, &output_frames, &output_frame_count, 0);
  	...
  }
  
  void release_demuxer_decoder()
  { // 如果指针指向内存地址, 则清除缓存, 重置指针
  	if (demuxer_p)
    {
      delete demuxer_p;
      demuxer_p = nullptr;
    }
    if (decoder_p)
    {
      delete decoder_p;
      decoder_p = nullptr;
    }
  	...
  }
  

6. C++ 11 Lambda表达式

遇到符号[=]

C++ 11 Lambda表达式

7. CUDA C编程

如何在 CUDA C/C++ 中实现数据传输的重叠

cudaMemcpyAsync等知识

7.1. python c++ 混合编程

华为开发者论坛

环境

https://github.com/pybind/pybind11

pip install pybind11
pip install pytest

直接使用pybind11文件夹, 直接放在项目文件夹根目录下, 和CMakeLists.txt还有cu.cc文件放一起.

git clone https://github.com/pybind/pybind11.git

7.1.1. cmakelists编译的过程中不停循环不终止问题

cmakelists循环执行_阿伟跑呀的博客-CSDN博客

7.2. C++ 传承python方法2

7.3. python 使用ctype与C++通信

C++编译 Cmakelists.txt

SET (LIBHELLO_SRC ${SRCS})
ADD_LIBRARY (AppDec SHARED ${LIBHELLO_SRC})

# 输出动态库
install(TARGETS AppDec
        LIBRARY DESTINATION ${CMAKE_BINARY_DIR}/bin
        )