ffmpeg hevc编码失败

Question

我用ffmpeg对yuv数据进行h265编码，但是编码后的图像总是不正确，如下图：

但是，使用以下命令可以正确编码： ffmpeg -f rawvideo -s 480x256 -pix_fmt yuv420p -i origin.yuv -c:v hevc -f hevc -x265-params keyint=1:crf=18 out.h265 ，下图：

这是我的代码：

void H265ImageCodec::InitCPUEncoder() {
  avcodec_register_all();
  AVCodec* encoder = avcodec_find_encoder(AV_CODEC_ID_H265);
  CHECK(encoder) << "Can not find encoder with h265.";
  // context
  encode_context_ = avcodec_alloc_context3(encoder);
  CHECK(encode_context_) << "Could not allocate video codec context.";
  encode_context_->codec_id = AV_CODEC_ID_H265;
  encode_context_->profile = FF_PROFILE_HEVC_MAIN;
  encode_context_->codec_type = AVMEDIA_TYPE_VIDEO;
  encode_context_->width = width_; // it's 480
  encode_context_->height = height_; // it's 256
  encode_context_->bit_rate = 384 * 1024;
  encode_context_->pix_fmt = AVPixelFormat::AV_PIX_FMT_YUV420P;
  encode_context_->time_base = (AVRational){1, 25};
  encode_context_->framerate = (AVRational){25, 1};
  AVDictionary* options = NULL;
  av_dict_set(&options, "preset", "ultrafast", 0);
  av_dict_set(&options, "tune", "zero-latency", 0);
  av_opt_set(encode_context_->priv_data, "x265-params", "keyint=1:crf=18",
             0);  // crf: Quality-controlled variable bitrate
  avcodec_open2(encode_context_, encoder, &options);

  encode_frame_ = av_frame_alloc();
  encode_frame_->format = encode_context_->pix_fmt;
  encode_frame_->width = encode_context_->width;
  encode_frame_->height = encode_context_->height;
  av_frame_get_buffer(encode_frame_, 0);
  // packet init
  encode_packet_ = av_packet_alloc();
}

std::string H265ImageCodec::EncodeImage(std::string_view raw_image) {
  av_packet_unref(encode_packet_);
  av_frame_make_writable(encode_frame_);
  const int64 y_size = width_ * height_;
  int64 offset = 0;
  memcpy(encode_frame_->data[0], raw_image.data() + offset, y_size);
  offset += y_size;
  memcpy(encode_frame_->data[1], raw_image.data() + offset, y_size / 4);
  offset += y_size / 4;
  memcpy(encode_frame_->data[2], raw_image.data() + offset, y_size / 4);
  avcodec_send_frame(encode_context_, encode_frame_);
  int ret = avcodec_receive_packet(encode_context_, encode_packet_);
  CHECK_EQ(ret, 0) << "receive encode packet ret: " << ret;
  std::string h265_frame(reinterpret_cast<char*>(encode_packet_->data),
                         encode_packet_->size);
  return h265_frame;
}

知道是什么原因造成的吗？

Answer 1

如评论所述，问题是encode_frame_中的 U 和 V 缓冲区encode_frame_内存中不连续。

执行encode_frame_ = av_frame_alloc()时的步骤如下：

• encode_frame_->linesize[0] = 480
  该值等于宽度，所以Y通道在内存中是连续的。
• encode_frame_->linesize[1] = 256 （不等于 480/2）。
• encode_frame_->linesize[2] = 256 （不等于 480/2）。
  U 和 V 通道的行在内存中不连续。

内存中目标 U 通道的图示：

         <----------- 256 bytes ----------->
         <------- 240 elements ------->
    ^    uuuuuuuuuuuuuuuuuuuuuuuuuuuuuu xxxx
    |    uuuuuuuuuuuuuuuuuuuuuuuuuuuuuu xxxx
128 rows uuuuuuuuuuuuuuuuuuuuuuuuuuuuuu xxxx
    |    uuuuuuuuuuuuuuuuuuuuuuuuuuuuuu xxxx
    V    uuuuuuuuuuuuuuuuuuuuuuuuuuuuuu xxxx

---

为了检查，我们可以打印linesize ：

printf("encode_frame_->linesize[0] = %d\n", encode_frame_->linesize[0]);  //480
printf("encode_frame_->linesize[1] = %d\n", encode_frame_->linesize[1]);  //256 (not 240)
printf("encode_frame_->linesize[2] = %d\n", encode_frame_->linesize[2]);  //256 (not 240)

---

受cudaMemcpy2D的启发，我们可以实现函数memcpy2D ：

//memcpy from src to dst with optional source "pitch" and destination "pitch".
//The "pitch" is the step in bytes between two rows.
//The function interface is based on cudaMemcpy2D.
static void memcpy2D(void*         dst,
                     size_t        dpitch,
                     const void*   src,
                     size_t        spitch,
                     size_t        width,
                     size_t        height)
{
    const unsigned char* I = (unsigned char*)src;
    unsigned char* J = (unsigned char*)dst;

    for (size_t y = 0; y < height; y++)
    {
        const unsigned char* I0 = I + y*spitch; //Pointer to the beggining of the source row
        unsigned char* J0 = J + y*dpitch; //Pointer to the beggining of the destination row
        memcpy(J0, I0, width);  //Copy width bytes from row I0 to row J0
    }
}

---

使用memcpy2D而不是memcpy将数据复制到内存中可能不连续的目标帧：

//Copy Y channel:
memcpy2D(encode_frame_->data[0],      //void* dst,
         encode_frame_->linesize[0],  //size_t        dpitch,
         raw_image.data() + offset,   //const void* src,
         width_,                      //size_t        spitch,
         width_,                      //size_t        width,
         height_);                    //size_t        height)

offset += y_size;

//Copy U channel:
memcpy2D(encode_frame_->data[1],      //void* dst,
         encode_frame_->linesize[1],  //size_t        dpitch,
         raw_image.data() + offset,   //const void* src,
         width_/2,                    //size_t        spitch,
         width_/2,                    //size_t        width,
         height_/2);                  //size_t        height)

offset += y_size / 4;

//Copy V channel:
memcpy2D(encode_frame_->data[2],      //void* dst,
         encode_frame_->linesize[2],  //size_t        dpitch,
         raw_image.data() + offset,   //const void* src,
         width_/2,                    //size_t        spitch,
         width_/2,                    //size_t        width,
         height_/2);                  //size_t        height)