winCE下使用一组buffer实现音频回环

 

 

最近有做项目，在WinCE下从一个音频设备获取数据，通过另一个音频设备输出。要求是尽量小的延迟，声音不要有停顿，也不能有丢失，在高的采样率的情况下，尽量少的占用cpu。东西很简单，但要做好也挺麻烦，特别是有了上面的要求。

先简单介绍一下思路：

使用wavein系列API，从一个音频驱动中读数据，然后使用waveout系列的API，把读到的数据写到另外一个音频驱动中。

 

如何实现比较好的性能。

１、Buffer尽量小，以减小延迟。为了保持声音连续和稳定，可以多开几个buffer。

２、尽量不要搬动数据，当采样率高了以后，数据量其实是挺大的，做一次复制操作就会消耗一定的CPU。经过N次的实验，终于成功让两个驱动使用了一组buffer,避免了数据的复制。

３、读和写一定要使用两个线程，中间使用消息等方式沟通。避免声音的停顿和丢失。

 

几个重要的变量和空间：

HWAVEIN m_hWaveIn; //for waveInOpen HWAVEOUT m_hWaveOut; //for waveOutOpen PBYTE m_pBuffer[WAV_IN_BUFFER_COUNT]; //the buffer used by wave drivers PWAVEHDR m_pWaveInHdr[WAV_IN_BUFFER_COUNT]; //the point of wavhdr struct, for input PWAVEHDR m_pWaveOutHdr[WAV_OUT_BUFFER_COUNT]; //the point of wavhdr struct, for output DWORD m_dwBufferRealSize[WAV_IN_BUFFER_COUNT]; //the actual size of the each buffer read from wavin driver.　　DWORD m_dwValidBlockStart; //the index of the first valid buffer block, valid means the buffer has data.DWORD m_dwValidBlocks; //the counts of the valid buffer blocks

初始化：

先分配buffer空间，打开输入输出设备。Prepare好head。顺序如下。需注意的是写数据时，需要使用线程回调，这样就可以和读数据线程互不影响。

 

if(!AllocSpace()) goto END; for (id = 0; id < waveInGetNumDevs(); id++) { WAVEINCAPS stCaps; memset(&stCaps, 0, sizeof(WAVEINCAPS)); waveInGetDevCaps(id, &stCaps, sizeof(WAVEINCAPS)); if (_tcsstr(stCaps.szPname, _T("DEV1")) != NULL) { break;; } } //open the wave in device result = waveInOpen(&m_hWaveIn,id,&m_WaveFormatEx,(DWORD)WaveInProc,NULL,CALLBACK_FUNCTION); if ( result != MMSYSERR_NOERROR ) { goto END; } //add the wave in buffer for( i = 0; i<WAV_IN_BUFFER_COUNT; i++ ) { m_pWaveInHdr[i]->lpData = (LPSTR)m_pBuffer[i]; m_pWaveInHdr[i]->dwBufferLength = INP_BUFFER_SIZE; m_pWaveInHdr[i]->dwBytesRecorded = 0; m_pWaveInHdr[i]->dwUser = 0; m_pWaveInHdr[i]->dwFlags = 0; m_pWaveInHdr[i]->dwLoops = 1; m_pWaveInHdr[i]->lpNext = NULL; m_pWaveInHdr[i]->reserved = 0; waveInPrepareHeader(m_hWaveIn,m_pWaveInHdr[i],sizeof(WAVEHDR)); waveInAddBuffer (m_hWaveIn, m_pWaveInHdr[i], sizeof (WAVEHDR)) ; } for( i = 0; i<WAV_OUT_BUFFER_COUNT; i++ ) { m_pWaveOutHdr[i]->lpData = NULL; m_pWaveOutHdr[i]->dwBufferLength = INP_BUFFER_SIZE; m_pWaveOutHdr[i]->dwUser = 0; m_pWaveOutHdr[i]->lpNext = NULL; } // Start the write thread as write callback m_writeThread = CreateThread(NULL, 0, WriteThreadProc, (void*)this, 0, &m_writeThreadId); if( m_writeThread == NULL ) goto END; for (id = 0; id < waveOutGetNumDevs(); id++) { WAVEOUTCAPS stCaps; memset(&stCaps, 0, sizeof(WAVEINCAPS)); waveOutGetDevCaps(id, &stCaps, sizeof(WAVEINCAPS)); if (_tcsstr(stCaps.szPname, _T("Blue")) == NULL) { break;; } } //open the wave out device, here use thread as call backresult = waveOutOpen(&m_hWaveOut, id, &m_WaveFormatEx, m_writeThreadId, (DWORD)this, CALLBACK_THREAD);

 

 

 

 

WIM_DATA消息的处理:

这里先讲一下读与写怎么同步。有两个变量，dwValidBlocks，dwValidBlockStart。初始都为０。这两个变量表示了从dwValidBlockStart开始的dwValidBlocks块数据，是新的，有效地。

dwValidBlocks有两个线程同时维护，当读线程读一个block了以后，dwValidBlocks＋＋，写线程写完一个block后dwValidBlocks――。

dwValidBlockStart平时由读线程维护，每读掉一个数据dwValidBlockStart＋＋。但是如果写线程没有及时写掉数据，造成所有block都满了，则由读线程更新掉最远一次的数据，即dwValidBlockStart＋＋。

 

这里需要注意的是WIM_DATA的响应中，没有必要使用waveInUnprepare,waveInPrepare。因为没有什么参数发生改变，直接扔回去就可以了。

 //----------------------------------------------------------------------//Decription://On WIM_DATA//------------------------------------------------------------------------void CSoundBase::OnWIM_DATA(WPARAM wParam, LPARAM lParam){if(m_bRecording == TRUE){ DWORD dwBytesRecorded = ((PWAVEHDR)wParam)->dwBytesRecorded; if(m_hThreadSyn[0]) { //record how much the read data size is of the buffer block just got; m_dwBufferRealSize[m_dwReadIndex] = dwBytesRecorded; DEBUGMSG(0, (TEXT("get %d/r/n"),m_dwReadIndex)); m_dwReadIndex++; if(m_dwReadIndex == WAV_IN_BUFFER_COUNT) m_dwReadIndex = 0; EnterCriticalSection(&m_critSecRtp); //DEBUGMSG(1, (TEXT("%d/r/n"),dwLastCount)); m_dwValidBlocks++ ; if( m_dwValidBlocks == WAV_IN_BUFFER_COUNT ) { m_dwValidBlockStart++; if( m_dwValidBlockStart == WAV_IN_BUFFER_COUNT ) m_dwValidBlockStart = 0; m_dwValidBlocks = WAV_IN_BUFFER_COUNT-1; } LeaveCriticalSection(&m_critSecRtp);//RETURN the buffer to the wavein driverwaveInAddBuffer(m_hWaveIn,(LPWAVEHDR)wParam, sizeof(WAVEHDR));//tell the write thread: the data is ready,you can write now. if( m_bDataBegining == FALSE && m_dwValidBlocks == WAV_OUT_BUFFER_COUNT) { SetEvent(m_hThreadSyn[0]); m_bDataBegining = TRUE; } }}}

 

 

写线程回调处理:

写线程做的事情就是不停的把buffer里面的数据写到输出的音频驱动中。这里需要提的是，写线程不是一有数据就开始写，而是等到有WAV_OUT_BUFFER_COUNT个block后，再开始写。一下子把这几个buffer一起扔下去，这样做的好处是保证一直有buffer在被驱动处理，而不会出现由于数据来不及扔下去造成声音停顿或当中有ｐｏｐ音出现。其实WAV_OUT_BUFFER_COUNT只要=2，应该就可以了。

对于MM_WOM_DONE的处理，需要waveOutUnprepareHeader，修改完那个pWaveOutHdr的结构以后，再扔回去即可。

 

DWORD WINAPI CSoundBase::WriteThreadProc(LPVOID lpParameter){ MSG msg; WAVEHDR *pWaveHdr = NULL; CSoundBase* pSound = (CSoundBase*)lpParameter; //SetThreadPriority( pSound->m_writeThread, THREAD_PRIORITY_TIME_CRITICAL); DWORD dwRet; DWORD dwBytesWritten = 0; UINT i = 0; MMRESULT result; // Open the existing wave file in order to add wave data#ifdef _RECORD_TO_FILE_pSound->m_hSaveFile = CreateFile(pSound->m_szSavePath, GENERIC_WRITE, FILE_SHARE_READ, NULL, OPEN_EXISTING, 0, NULL);if( pSound->m_hSaveFile == INVALID_HANDLE_VALUE ) {goto EXIT; }//Set the file pointer to the end.SetFilePointer(pSound->m_hSaveFile,0,NULL,FILE_END);#endif dwRet = WaitForMultipleObjects(2, pSound->m_hThreadSyn,FALSE, INFINITE ); if( dwRet == WAIT_OBJECT_0 +1 ) goto EXIT; else if( dwRet == WAIT_OBJECT_0 ){//write the data blocks to the driver together for( i = 0; i< WAV_OUT_BUFFER_COUNT; i++) { pSound->m_pWaveOutHdr[i]->lpData = (LPSTR)pSound->m_pBuffer[i]; pSound->m_pWaveOutHdr[i]->dwBufferLength = pSound->m_dwBufferRealSize[i]; result = waveOutPrepareHeader(pSound->m_hWaveOut, pSound->m_pWaveOutHdr[i], sizeof(WAVEHDR)); } for( i = 0; i< WAV_OUT_BUFFER_COUNT; i++) { result = waveOutWrite(pSound->m_hWaveOut, pSound->m_pWaveOutHdr[i], sizeof(WAVEHDR)); } while(GetMessage(&msg, 0, 0, 0)) { switch (msg.message) { /* audio data playback by waveout device in system has been done */ case MM_WOM_DONE: pWaveHdr = (WAVEHDR*)msg.lParam; waveOutUnprepareHeader(pSound->m_hWaveOut, pWaveHdr, sizeof(WAVEHDR)); EnterCriticalSection(&pSound->m_critSecRtp); if( pSound->m_dwValidBlocks ) { pSound->m_dwValidBlockStart++; if( pSound->m_dwValidBlockStart == WAV_IN_BUFFER_COUNT ) pSound->m_dwValidBlockStart = 0; pSound->m_dwValidBlocks --; pWaveHdr->lpData = (LPSTR)pSound->m_pBuffer[pSound->m_dwValidBlockStart]; DEBUGMSG(1, (TEXT("write %d/r/n"),pSound->m_dwValidBlocks )); } //DEBUGMSG(1, (TEXT("writing block .../r/n") )); LeaveCriticalSection(&pSound->m_critSecRtp); result = waveOutPrepareHeader(pSound->m_hWaveOut, pWaveHdr, sizeof(WAVEHDR)); result = waveOutWrite(pSound->m_hWaveOut, pWaveHdr, sizeof(WAVEHDR)); break; default: break; } } }EXIT: DEBUGMSG(1, (TEXT("thread exit/r/n")));#ifdef _RECORD_TO_FILE_ if(pSound->m_hSaveFile != NULL){CloseHandle(pSound->m_hSaveFile);pSound->m_hSaveFile = NULL;}if (0 != pSound->m_dwDataLength){//Write the data length to the file.pSound->WriteWaveFileHeader(pSound->m_szSavePath,&pSound->m_WaveFormatEx,pSound->m_dwDataLength,FALSE);}#endif CloseHandle(pSound->m_writeThread); pSound->m_writeThread = NULL; return 0;}

 

 

小结

 

 

 

在写这个程序的过程中，花时间最多的有两个问题。一个是如何直接利用一组buffer，在应用层不搬动任何数据，实现音频回环。使用wav API的时候，很容易造成deadlock。另外一个，就是如何使声音不间断的播放。没有使用线程，经常会卡。写数据时缓存没有连续给到驱动，也会小卡，最后看了很多人的blog后，终于解决了问题。其实无需两个设备，只要一个codec，配置好wavein和waveout的设备ＩＤ，就可以用speaker听到mic的呻吟了。

程序小改一下就可以变译运行，但是还有问题，其实线程那部分还很挫，退不出来，而且还有很多其他小问题。没时间修改这个demo，也不能把现在用得代码贴出来，请谅解。

 

最后感谢不下雨（Norains）大虾，代码是在他的CSoundBase实现录音与播放上修改过来的。希望不要介意。

 

参考：

http://blog.csdn.net/norains/archive/2007/01/10/1479648.aspx　CSoundBase实现录音与播放

http://www.planet-source-code.com/vb/scripts/ShowCode.asp?txtCodeId=4422&lngWId=3　 Playing Audio in Windows using waveOut Interface