十七:H264VideoStreamParser详解
很多人要做实时H264 RTP传输,那么如何充分利用live555来做呢?
大家可以看到现有的H264VideoFileServerMediaSubsession中,sink使用了H264VideoRTPSink,source使用了H264VideoStreamFramer,然而这个连接是很复杂的,在这两个节点间要插入了很多其它的节点,其实际情况是这样的:ByteStreamFileSource-->H264VideoStreamParser-->H264VideoStreamFramer-->H264FUAFragmenter-->H264VideoRTPSink.哇!真的这么复杂吗?一点没错!
当然你可以不用理它们的来龙去脉,你只需自己实现一个source,能采集图像并进行h264编码的source(当然你可以用CPU也可以用DSP进行编码),然后用它替代ByteStreamFileSource,就成了,比如你这个source可以叫做H264ByteStreamSource.当然为了提高效率,采集和编码部分应放在另一个线程中执行.
然而,我还是很想了解H264VideoStreamParser到底是什么,Parser到底有什么用?它做了什么?它与H264VideoStreamFramer是如何配合的?它们之间有内存copy发生吗?
先设想一个问题:
H264VideoStreamFramer是什么角色?跟据H264VideoFileServerMediaSubsession的代码,H264VideoStreamFramer是真正代表source的,Sink所面对的Source就是它.但是它又连接了一个ByteStreamFileSource.look一下这部分代码:
FramedSource* H264VideoFileServerMediaSubsession:: createNewStreamSource(unsigned /*clientSessionId*/, unsigned& estBitrate) { estBitrate = 500; // kbps, estimate // Create the video source: ByteStreamFileSource* fileSource = ByteStreamFileSource::createNew(envir(), fFileName); if (fileSource == NULL) return NULL; fFileSize = fileSource->fileSize(); // Create a framer for the Video Elementary Stream: return H264VideoStreamFramer::createNew(envir(), fileSource); }
是吧?我没有忽悠吧?
ByteStreamFileSource是从文件取得数据的,它不管是到底什么媒体格式,它只是读文件.所以很明显H264VideoStreamFramer利用ByteStreamFileSource从文件取得数据,然后H264VideoStreamFramer再对数据进行分析.比如找出每个NALU,然后传给Sink.但是H264VideoStreamFramer没有自己去分析,而是利用了Parser,所以那一串中就多了一个H264VideoStreamParser.
H264VideoStreamParser拥有两个source指针,一个是FramedSource* fInputSource,另一个是H264VideoStreamFramer* fUsingSource.可以看出,H264VideoStreamParser把fInputSource和fUsingSource串了起来,那么fInputSource就是ByteStreamFileSource.
我们想像一下H264VideoStreamParser的所作所为:H264VideoStreamFramer把自己的缓冲(其实是sink的)传给H264VideoStreamParser,每当H264VideoStreamFramer要获取一个NALU时,就跟H264VideoStreamParser要,H264VideoStreamParser就从ByteStreamFileSource读一坨数据,然后进行分析,如果取得了一个NALU,就传给H264VideoStreamFramer.唉,H264VideoStreamFramer真是个不劳而获的坏家伙!
看一下实际的流程:
//Sink调用Source(H264VideoStreamFramer)的GetNextFrame()获取数据, //H264VideoStreamFramer从MPEGVideoStreamFramer派生,所以下面的函数会被调用: void MPEGVideoStreamFramer::doGetNextFrame() { fParser->registerReadInterest(fTo, fMaxSize); continueReadProcessing(); } void MPEGVideoStreamFramer::continueReadProcessing(void* clientData, unsigned char* /*ptr*/, unsigned /*size*/, struct timeval /*presentationTime*/) { MPEGVideoStreamFramer* framer = (MPEGVideoStreamFramer*) clientData; framer->continueReadProcessing(); }上两个是过渡,最终在这里执行:
void MPEGVideoStreamFramer::continueReadProcessing() { //调用Parser的parser()分析出一个NALU.如果得到了一个NALU,则 //用afterGetting(this)返回给Sink. unsigned acquiredFrameSize = fParser->parse(); if (acquiredFrameSize > 0) { // We were able to acquire a frame from the input. // It has already been copied to the reader's space. fFrameSize = acquiredFrameSize; fNumTruncatedBytes = fParser->numTruncatedBytes(); // "fPresentationTime" should have already been computed. // Compute "fDurationInMicroseconds" now: fDurationInMicroseconds = (fFrameRate == 0.0 || ((int) fPictureCount) < 0) ? 0 : (unsigned) ((fPictureCount * 1000000) / fFrameRate); fPictureCount = 0; // Call our own 'after getting' function. Because we're not a 'leaf' // source, we can call this directly, without risking infinite recursion. afterGetting(this); } else { //执行到此处并不代表parser()中没有取得数据!! // We were unable to parse a complete frame from the input, because: // - we had to read more data from the source stream, or // - the source stream has ended. } }
void StreamParser::ensureValidBytes1(unsigned numBytesNeeded) { // We need to read some more bytes from the input source. // First, clarify how much data to ask for: unsigned maxInputFrameSize = fInputSource->maxFrameSize(); if (maxInputFrameSize > numBytesNeeded) numBytesNeeded = maxInputFrameSize; // First, check whether these new bytes would overflow the current // bank. If so, start using a new bank now. if (fCurParserIndex + numBytesNeeded > BANK_SIZE) { // Swap banks, but save any still-needed bytes from the old bank: unsigned numBytesToSave = fTotNumValidBytes - fSavedParserIndex; unsigned char const* from = &curBank()[fSavedParserIndex]; fCurBankNum = (fCurBankNum + 1) % 2; fCurBank = fBank[fCurBankNum]; memmove(curBank(), from, numBytesToSave); fCurParserIndex = fCurParserIndex - fSavedParserIndex; fSavedParserIndex = 0; fTotNumValidBytes = numBytesToSave; } // ASSERT: fCurParserIndex + numBytesNeeded > fTotNumValidBytes // && fCurParserIndex + numBytesNeeded <= BANK_SIZE if (fCurParserIndex + numBytesNeeded > BANK_SIZE) { // If this happens, it means that we have too much saved parser state. // To fix this, increase BANK_SIZE as appropriate. fInputSource->envir() << "StreamParser internal error (" << fCurParserIndex << "+ " << numBytesNeeded << " > " << BANK_SIZE << ")\n"; fInputSource->envir().internalError(); } // Try to read as many new bytes as will fit in the current bank: unsigned maxNumBytesToRead = BANK_SIZE - fTotNumValidBytes; fInputSource->getNextFrame(&curBank()[fTotNumValidBytes], maxNumBytesToRead, afterGettingBytes, this, onInputClosure, this); throw NO_MORE_BUFFERED_INPUT; }可以看到一个奇怪的现象:这个函数没有返回值,但最终抛出了一个异常,而且只要执行这个函数,就会抛出这个异常.
catch (int /*e*/) { #ifdef DEBUG fprintf(stderr, "H264VideoStreamParser::parse() EXCEPTION (This is normal behavior - *not* an error)\n"); #endif return 0; // the parsing got interrupted }可见parser()此时返回0,parser()返回0就执行到MPEGVideoStreamFramer::continueReadProcessing()中的else{}部分了,回去看看吧,其实啥也没做.也就是说,第一次调用parser()时,它只是从ByteStreamFileSource获取数据,那么这个parser()获取数据后什么也不做,但实际上对NALU分析和处里在这次Parser()的调用中已经完成了,不是在它本身完成的,而是在它引起了parser()的嵌套调用中完成.好迷糊,理顺一下过程就知道了: