openh264 自适应量化功能源码分析

发布于:2024-06-07 ⋅ 阅读:(171) ⋅ 点赞:(0)

openh264

  • OpenH264是一个开源的H.264/AVC视频编解码器,由Cisco公司发起并贡献了最初的代码基础。它提供了一个用于视频编码和解码的库,支持H.264视频压缩标准,广泛应用于视频会议、流媒体和视频存储等领域。
  • OpenH264是实现H.264编解码功能的一个很好的选择,特别是对于需要免费解决方案的场景。
  • OpenH264的一些关键特性:
  • 开源:OpenH264遵循BSD风格的开源许可证,允许任何人自由使用、修改和分发。

  • 跨平台:支持多种操作系统,包括Windows、Linux、Mac OS X等。

  • 高性能:提供了高效的编码和解码算法,能够处理高分辨率视频。

  • 硬件加速:支持多种硬件加速技术,如Intel Quick Sync Video、NVIDIA CUDA等。

  • 编码配置灵活:允许用户根据需要配置编码参数,如分辨率、帧率、比特率等。

  • 解码能力:除了编码功能外,OpenH264也提供了解码能力,能够解码H.264编码的视频流。

  • API接口:提供了一套API接口,方便开发者集成到自己的应用程序中。

  • 社区支持:作为一个开源项目,OpenH264得到了活跃的社区支持,不断有新的功能和改进被加入。

自适应量化功能源码文件位置

  • openh264/codec/processing/adaptivequantization/AdaptiveQuantization.cpp

自适应量化功能源码文件流程

在这里插入图片描述

  • 说明
    • 可以看到目前 openh264 选择关闭自适应量化算法,需要重构。
    • 实现自适应量化的核心功能就是 Process 函数。

自适应量化功能原理分析

  1. 功能:实现视频编码自适应量化处理,涉及到图像的运动和纹理分析,以及量化参数的动态调整等。
  2. 过程
  • 初始化一些变量,包括图像宽度、高度、宏块(MB)宽度和高度,以及总宏块数等。
  • 声明了一些用于存储运动纹理单元SMotionTextureUnit和SVAACalcResult(Variable Adaptive Analysis)计算结果的指针;其中通过VAACalcSadSsdBgd_c/VAACalcSadBgd_c函数计算
    SVAACalcResult结构体中变量。
  • 初始化了一些用于计算量化步长和量化参数的变量。
  • 获取源图像和参考图像的Y分量的指针以及它们的跨度(stride)。
  • 进入运动分析部分,计算宏块的运动残差方差和纹理方差。
  • 如果SVAACalcResult计算结果中的指针与传入的图像指针相同;
    • 则使用SVAACalcResult结果;
    • 对每个宏块,根据运动和纹理分析结果,计算运动指数iAverageMotionIndex和纹理指数iAverageTextureIndex,并累加到平均值中。
  • 否则,
    • 调用m_pfVar函数进行运动和纹理分析;
    • 对每个宏块,根据运动和纹理分析结果,计算运动指数iAverageMotionIndex和纹理指数iAverageTextureIndex,并累加到平均值中。
  • 计算平均运动指数和纹理指数,并进行一些条件判断和调整。
  • 双层 for 循环处理每个宏块;
    • 根据运动和纹理指数映射到量化参数(QP),计算iMotionTextureIndexToDeltaQp;
    • 根自适应量化模式(AQ_QUALITY_MODE或AQ_BITRATE_MODE),调整量化参数iMotionTextureIndexToDeltaQp。
  • 将计算出的量化参数映射存储到m_sAdaptiveQuantParam结构体中iAverMotionTextureIndexToDeltaQp。
  • 设置返回值为成功(RET_SUCCESS)。
  1. 相关源码
  • Process函数
EResult CAdaptiveQuantization::Process (int32_t iType, SPixMap* pSrcPixMap, SPixMap* pRefPixMap) {
  EResult eReturn = RET_INVALIDPARAM;

  int32_t iWidth     = pSrcPixMap->sRect.iRectWidth;
  int32_t iHeight    = pSrcPixMap->sRect.iRectHeight;
  int32_t iMbWidth  = iWidth  >> 4;
  int32_t iMbHeight = iHeight >> 4;
  int32_t iMbTotalNum    = iMbWidth * iMbHeight;

  SMotionTextureUnit* pMotionTexture = NULL;
  SVAACalcResult*     pVaaCalcResults = NULL;
  int32_t   iMotionTextureIndexToDeltaQp = 0;
  int32_t iAverMotionTextureIndexToDeltaQp = 0;  // double to uint32
  int64_t iAverageMotionIndex = 0;      // double to float
  int64_t iAverageTextureIndex = 0;

  int64_t iQStep = 0;
  int64_t iLumaMotionDeltaQp = 0;
  int64_t iLumaTextureDeltaQp = 0;

  uint8_t* pRefFrameY = NULL, *pCurFrameY = NULL;
  int32_t iRefStride = 0, iCurStride = 0;

  uint8_t* pRefFrameTmp = NULL, *pCurFrameTmp = NULL;
  int32_t i = 0, j = 0;

  pRefFrameY = (uint8_t*)pRefPixMap->pPixel[0];
  pCurFrameY = (uint8_t*)pSrcPixMap->pPixel[0];

  iRefStride  = pRefPixMap->iStride[0];
  iCurStride  = pSrcPixMap->iStride[0];

  /// motion //
  //  motion MB residual variance
  iAverageMotionIndex = 0;
  iAverageTextureIndex = 0;
  pMotionTexture = m_sAdaptiveQuantParam.pMotionTextureUnit;
  pVaaCalcResults = m_sAdaptiveQuantParam.pCalcResult;

  if (pVaaCalcResults->pRefY == pRefFrameY && pVaaCalcResults->pCurY == pCurFrameY) {
    int32_t iMbIndex = 0;
    int32_t iSumDiff, iSQDiff, uiSum, iSQSum;
    for (j = 0; j < iMbHeight; j ++) {
      pRefFrameTmp  = pRefFrameY;
      pCurFrameTmp  = pCurFrameY;
      for (i = 0; i < iMbWidth; i++) {
        iSumDiff =  pVaaCalcResults->pSad8x8[iMbIndex][0];
        iSumDiff += pVaaCalcResults->pSad8x8[iMbIndex][1];
        iSumDiff += pVaaCalcResults->pSad8x8[iMbIndex][2];
        iSumDiff += pVaaCalcResults->pSad8x8[iMbIndex][3];

        iSQDiff = pVaaCalcResults->pSsd16x16[iMbIndex];
        uiSum = pVaaCalcResults->pSum16x16[iMbIndex];
        iSQSum = pVaaCalcResults->pSumOfSquare16x16[iMbIndex];

        iSumDiff = iSumDiff >> 8;
        pMotionTexture->uiMotionIndex = (iSQDiff >> 8) - (iSumDiff * iSumDiff);

        uiSum = uiSum >> 8;
        pMotionTexture->uiTextureIndex = (iSQSum >> 8) - (uiSum * uiSum);

        iAverageMotionIndex += pMotionTexture->uiMotionIndex;
        iAverageTextureIndex += pMotionTexture->uiTextureIndex;
        pMotionTexture++;
        ++iMbIndex;
        pRefFrameTmp += MB_WIDTH_LUMA;
        pCurFrameTmp += MB_WIDTH_LUMA;
      }
      pRefFrameY += (iRefStride) << 4;
      pCurFrameY += (iCurStride) << 4;
    }
  } else {
    for (j = 0; j < iMbHeight; j ++) {
      pRefFrameTmp  = pRefFrameY;
      pCurFrameTmp  = pCurFrameY;
      for (i = 0; i < iMbWidth; i++) {
        m_pfVar (pRefFrameTmp, iRefStride, pCurFrameTmp, iCurStride, pMotionTexture);
        iAverageMotionIndex += pMotionTexture->uiMotionIndex;
        iAverageTextureIndex += pMotionTexture->uiTextureIndex;
        pMotionTexture++;
        pRefFrameTmp += MB_WIDTH_LUMA;
        pCurFrameTmp += MB_WIDTH_LUMA;

      }
      pRefFrameY += (iRefStride) << 4;
      pCurFrameY += (iCurStride) << 4;
    }
  }
  iAverageMotionIndex = WELS_DIV_ROUND64 (iAverageMotionIndex * AQ_INT_MULTIPLY, iMbTotalNum);
  iAverageTextureIndex = WELS_DIV_ROUND64 (iAverageTextureIndex * AQ_INT_MULTIPLY, iMbTotalNum);
  if ((iAverageMotionIndex <= AQ_PESN) && (iAverageMotionIndex >= -AQ_PESN)) {
    iAverageMotionIndex = AQ_INT_MULTIPLY;
  }
  if ((iAverageTextureIndex <= AQ_PESN) && (iAverageTextureIndex >= -AQ_PESN)) {
    iAverageTextureIndex = AQ_INT_MULTIPLY;
  }
  //  motion mb residual map to QP
  //  texture mb original map to QP
  iAverMotionTextureIndexToDeltaQp = 0;
  iAverageMotionIndex = WELS_DIV_ROUND64 (AVERAGE_TIME_MOTION * iAverageMotionIndex, AQ_TIME_INT_MULTIPLY);

  if (m_sAdaptiveQuantParam.iAdaptiveQuantMode == AQ_QUALITY_MODE) {
    iAverageTextureIndex = WELS_DIV_ROUND64 (AVERAGE_TIME_TEXTURE_QUALITYMODE * iAverageTextureIndex, AQ_TIME_INT_MULTIPLY);
  } else {
    iAverageTextureIndex = WELS_DIV_ROUND64 (AVERAGE_TIME_TEXTURE_BITRATEMODE * iAverageTextureIndex, AQ_TIME_INT_MULTIPLY);
  }

  int64_t iAQ_EPSN = - ((int64_t)AQ_PESN * AQ_TIME_INT_MULTIPLY * AQ_QSTEP_INT_MULTIPLY / AQ_INT_MULTIPLY);
  pMotionTexture = m_sAdaptiveQuantParam.pMotionTextureUnit;
  for (j = 0; j < iMbHeight; j ++) {
    for (i = 0; i < iMbWidth; i++) {
      int64_t a = WELS_DIV_ROUND64 ((int64_t) (pMotionTexture->uiTextureIndex) * AQ_INT_MULTIPLY * AQ_TIME_INT_MULTIPLY,
                                    iAverageTextureIndex);
      iQStep = WELS_DIV_ROUND64 ((a - AQ_TIME_INT_MULTIPLY) * AQ_QSTEP_INT_MULTIPLY, (a + MODEL_ALPHA));
      iLumaTextureDeltaQp = MODEL_TIME * iQStep;// range +- 6

      iMotionTextureIndexToDeltaQp = ((int32_t) (iLumaTextureDeltaQp / (AQ_TIME_INT_MULTIPLY)));

      a = WELS_DIV_ROUND64 (((int64_t)pMotionTexture->uiMotionIndex) * AQ_INT_MULTIPLY * AQ_TIME_INT_MULTIPLY,
                            iAverageMotionIndex);
      iQStep = WELS_DIV_ROUND64 ((a - AQ_TIME_INT_MULTIPLY) * AQ_QSTEP_INT_MULTIPLY, (a + MODEL_ALPHA));
      iLumaMotionDeltaQp = MODEL_TIME * iQStep;// range +- 6

      if ((m_sAdaptiveQuantParam.iAdaptiveQuantMode == AQ_QUALITY_MODE && iLumaMotionDeltaQp < iAQ_EPSN)
          || (m_sAdaptiveQuantParam.iAdaptiveQuantMode == AQ_BITRATE_MODE)) {
        iMotionTextureIndexToDeltaQp += ((int32_t) (iLumaMotionDeltaQp / (AQ_TIME_INT_MULTIPLY)));
      }

      m_sAdaptiveQuantParam.pMotionTextureIndexToDeltaQp[j * iMbWidth + i] = (int8_t) (iMotionTextureIndexToDeltaQp /
          AQ_QSTEP_INT_MULTIPLY);
      iAverMotionTextureIndexToDeltaQp += iMotionTextureIndexToDeltaQp;
      pMotionTexture++;
    }
  }

  m_sAdaptiveQuantParam.iAverMotionTextureIndexToDeltaQp = iAverMotionTextureIndexToDeltaQp / iMbTotalNum;

  eReturn = RET_SUCCESS;

  return eReturn;
}
  • m_pfVar函数(指向SampleVariance16x16_c函数)
void SampleVariance16x16_c (uint8_t* pRefY, int32_t iRefStride, uint8_t* pSrcY, int32_t iSrcStride,
                            SMotionTextureUnit* pMotionTexture) {
  uint32_t uiCurSquare = 0,  uiSquare = 0;
  uint16_t uiCurSum = 0,  uiSum = 0;

  for (int32_t y = 0; y < MB_WIDTH_LUMA; y++) {
    for (int32_t x = 0; x < MB_WIDTH_LUMA; x++) {
      uint32_t uiDiff = WELS_ABS (pRefY[x] - pSrcY[x]);
      uiSum += uiDiff;
      uiSquare += uiDiff * uiDiff;

      uiCurSum += pSrcY[x];
      uiCurSquare += pSrcY[x] * pSrcY[x];
    }
    pRefY += iRefStride;
    pSrcY += iSrcStride;
  }

  uiSum = uiSum >> 8;
  pMotionTexture->uiMotionIndex = (uiSquare >> 8) - (uiSum * uiSum);

  uiCurSum = uiCurSum >> 8;
  pMotionTexture->uiTextureIndex = (uiCurSquare >> 8) - (uiCurSum * uiCurSum);
}

网站公告

今日签到

点亮在社区的每一天
去签到