通过解码器代码的研究,已经对HEVC的编解码技术有了一个初步的认识。现在我们就对照着编码器的代码进一步理解HEVC视频编码算法的各个技术细节。
编码器在整个HM解决方案中的工程名为TAppEncoder,入口点函数位于encmain.cpp文件中:
int main(int argc, char* argv[])
{
TAppEncTop cTAppEncTop;
// print information
fprintf( stdout, "\n" );
fprintf( stdout, "HM software: Encoder Version [%s]", NV_VERSION );
fprintf( stdout, NVM_ONOS );
fprintf( stdout, NVM_COMPILEDBY );
fprintf( stdout, NVM_BITS );
fprintf( stdout, "\n" );
// create application encoder class
cTAppEncTop.create();
// parse configuration
try
{
if(!cTAppEncTop.parseCfg( argc, argv ))
{
cTAppEncTop.destroy();
return 1;
}
}
catch (po::ParseFailure& e)
{
cerr << "Error parsing option \""<< e.arg <<"\" with argument \""<< e.val <<"\"." << endl;
return 1;
}
// starting time
double dResult;
long lBefore = clock();
// call encoding function
cTAppEncTop.encode();
// ending time
dResult = (double)(clock()-lBefore) / CLOCKS_PER_SEC;
printf("\n Total Time: %12.3f sec.\n", dResult);
// destroy application encoder class
cTAppEncTop.destroy();
return 0;
}
可以很清楚地看到,整个main函数非常简洁清晰,主要可以分为几大部分,分别是输入软件信息、创建编码器类的实例、解析配置文件、获取开始时间、编码数据、计算耗费时间和销毁编码器类的实例几大部分。我们主要关心的编码过程仅通过调用编码器实例的一个方法实现:
// call encoding function
cTAppEncTop.encode();
该函数的实现如下:
Void TAppEncTop::encode()
{
fstream bitstreamFile(m_pchBitstreamFile, fstream::binary | fstream::out);
if (!bitstreamFile)
{
fprintf(stderr, "\nfailed to open bitstream file `%s' for writing\n", m_pchBitstreamFile);
exit(EXIT_FAILURE);
}
TComPicYuv* pcPicYuvOrg = new TComPicYuv;
TComPicYuv* pcPicYuvRec = NULL;
// initialize internal class & member variables
xInitLibCfg();
xCreateLib();
xInitLib();
// main encoder loop
Int iNumEncoded = 0;
Bool bEos = false;
list<AccessUnit> outputAccessUnits; ///< list of access units to write out. is populated by the encoding process
// allocate original YUV buffer
pcPicYuvOrg->create( m_iSourceWidth, m_iSourceHeight, m_uiMaxCUWidth, m_uiMaxCUHeight, m_uiMaxCUDepth );
while ( !bEos )
{
// get buffers
xGetBuffer(pcPicYuvRec);
// read input YUV file
m_cTVideoIOYuvInputFile.read( pcPicYuvOrg, m_aiPad );
// increase number of received frames
m_iFrameRcvd++;
bEos = (m_iFrameRcvd == m_framesToBeEncoded);
Bool flush = 0;
// if end of file (which is only detected on a read failure) flush the encoder of any queued pictures
if (m_cTVideoIOYuvInputFile.isEof())
{
flush = true;
bEos = true;
m_iFrameRcvd--;
m_cTEncTop.setFramesToBeEncoded(m_iFrameRcvd);
}
// call encoding function for one frame
m_cTEncTop.encode( bEos, flush ? 0 : pcPicYuvOrg, m_cListPicYuvRec, outputAccessUnits, iNumEncoded );
// write bistream to file if necessary
if ( iNumEncoded > 0 )
{
xWriteOutput(bitstreamFile, iNumEncoded, outputAccessUnits);
outputAccessUnits.clear();
}
}
m_cTEncTop.printSummary();
// delete original YUV buffer
pcPicYuvOrg->destroy();
delete pcPicYuvOrg;
pcPicYuvOrg = NULL;
// delete used buffers in encoder class
m_cTEncTop.deletePicBuffer();
// delete buffers & classes
xDeleteBuffer();
xDestroyLib();
printRateSummary();
return;
}
该函数中首先调用pcPicYuvOrg->create( m_iSourceWidth, m_iSourceHeight, m_uiMaxCUWidth, m_uiMaxCUHeight, m_uiMaxCUDepth )分配YUV数据缓存,然后再while循环中逐帧读取YUV数据、设置当前以编码的帧数、编码当前帧、写出码流,随后做其他清理工作。核心功能实现在m_cTEncTop.encode( bEos, flush ? 0 : pcPicYuvOrg, m_cListPicYuvRec, outputAccessUnits, iNumEncoded )函数中。在该函数中调用m_cGOPEncoder.compressGOP(m_iPOCLast, m_iNumPicRcvd, m_cListPic, rcListPicYuvRecOut, accessUnitsOut)进行编码一个GOP的操作。这个函数奇长无比,用了接近1500行代码,看来实现了很多很多很多的功能。这个碉堡了的函数究竟做了些啥事儿呢?这个函数中大部分内容就是在为了编码当前slice做准备,以及编码完成之后一些辅助操作。实际编码过程的操作由以下函数m_pcSliceEncoder->compressSlice( pcPic )实现。
这又是一个碉堡了的函数,占了将近400行……代码就不贴了,会死人的……简单看下好了。
首先还是各种编码的配置,包括配置熵编码器、初始化CU编码器等。在完成了一长串的设置之后,在compressCU函数中实现对一个CU的编码:
m_pcCuEncoder->compressCU( pcCU );
具体的细节,且待下文。