stagefright in android

1.Video Playback的流程

在Android上，预设的多媒体框架(multimedia framework)是OpenCORE。OpenCORE的优点是兼顾了跨平台的移植性，而且已经过多方验证，所以相对来说较為稳定；但是其缺点是过於庞大复杂，需要耗费相当多的时间去维护。从Android 2.0开始，Google引进了架构稍為简洁的Stagefright，并且有逐渐取代OpenCORE的趋势 (註1)。

[图1] Stagefright在Android多媒体架构中的位置。

[图2] Stagefright所涵盖的模组 (註2)。
以下我们就先来看看Stagefright是如何播放一个影片档。
Stagefright在Android中是以shared library的形式存在(libstagefright.so)，其中的module -- AwesomePlayer可用来播放video/audio (註3)。AwesomePlayer提供许多API，可以让上层的应用程式(Java/JNI)来呼叫，我们以一个简单的程式来说明video playback的流程。
在Java中，若要播放一个影片档，我们会这样写：
MediaPlayer mp = new MediaPlayer();
mp.setDataSource(PATH_TO_FILE); ...... (1)
mp.prepare(); ........................ (2)、(3)
mp.start(); .......................... (4)

在Stagefright中，则会看到相对应的处理；

(1) 将档案的绝对路径指定给mUri

status_t AwesomePlayer::setDataSource(const char* uri, ...)

{

  return setDataSource_l(uri, ...);

}

status_t AwesomePlayer::setDataSource_l(const char* uri, ...)

{

  mUri = uri;

}

(2)启动mQueue，作为event handler

status_t AwesomePlayer::prepare()

{

  return prepare_l();

}

status_t AwesomePlayer::prepare_l()

{

  prepareAsync_l();

  while (mFlags & PREPARING)

  {

    mPreparedCondition.wait(mLock);

  }

}

status_t AwesomePlayer::prepareAsync_l()

{

  mQueue.start();

  mFlags |= PREPARING;

  mAsyncPrepareEvent = new AwesomeEvent(this&AwesomePlayer::onPrepareAsyncEvent);

  mQueue.postEvent(mAsyncPrepareEvent);

}

(3) onPrepareAsyncEvent被触发

void AwesomePlayer::onPrepareAsyncEvent()

{

  finishSetDataSource_l();

  initVideoDecoder(); ...... (3.3)

  initAudioDecoder();

}

status_t AwesomePlayer::finishSetDataSource_l()

{

  dataSource = DataSource::CreateFromURI(mUri.string(), ...);

  sp<MediaExtractor> extractor =MediaExtractor::Create(dataSource); ..... (3.1)

  return setDataSource_l(extractor); ......................... (3.2)

}

(3.1) 解析mUri所指定的档案，并且根据其header来选择对应的extractor

sp<MediaExtractor> MediaExtractor::Create(const sp<DataSource> &source, ...)

{

  source->sniff(&tmp, ...);

  mime = tmp.string();

  if (!strcasecmp(mime, MEDIA_MIMETYPE_CONTAINER_MPEG4)

  {

    return new MPEG4Extractor(source);

  }

  else if (!strcasecmp(mime, MEDIA_MIMETYPE_AUDIO_MPEG))

  {

    return new MP3Extractor(source);

  }

  else if (!strcasecmp(mime, MEDIA_MIMETYPE_AUDIO_AMR_NB)

  {

    return new AMRExtractor(source);

  }

}

(3.2) 使用extractor对档案做A/V(Audio/Video)的分离 (mVideoTrack/mAudioTrack)

status_t AwesomePlayer::setDataSource_l(const sp<MediaExtractor> &extractor)

{

  for (size_t i = 0; i < extractor->countTracks(); ++i)

  {

    sp<MetaData> meta = extractor->getTrackMetaData(i);

    CHECK(meta->findCString(kKeyMIMEType, &mime));

    if(!haveVideo && !strncasecmp(mime, "video/", 6))

    {

      setVideoSource(extractor->getTrack(i));

      haveVideo = true;

    }

    else if (!haveAudio && !strncasecmp(mime, "audio/", 6))

    {

      setAudioSource(extractor->getTrack(i));

      haveAudio = true;

    }

  }

}

void AwesomePlayer::setVideoSource(sp<MediaSource> source)

{

  mVideoTrack = source;

}

(3.3) 根据mVideoTrack中的编码类型来选择video decoder (mVideoSource)

status_t AwesomePlayer::initVideoDecoder()

{

  mVideoSource = OMXCodec::Create(mClient.interface(),mVideoTrack->getFormat(),false,mVideoTrack);

}

(4) 将mVideoEvent放入mQueue中，开始解码播放，并交由mVideoRenderer来画出

status_t AwesomePlayer::play()

{

  return play_l();

}

status_t AwesomePlayer::play_l()

{

  postVideoEvent_l();

}

void AwesomePlayer::postVideoEvent_l(int64_t delayUs)

{

  mQueue.postEventWithDelay(mVideoEvent, delayUs);

}

void AwesomePlayer::onVideoEvent()

{

  mVideoSource->read(&mVideoBuffer, &options);

  [Check Timestamp]

  mVideoRenderer->render(mVideoBuffer);

  postVideoEvent_l();

}

2.选择Video Decoder

我们来看一看Stagefright是如何根据影片档的类型来选择适合的Video Decoder

(1) Video Decoder是在onPrepareAsyncEvent中的initVideoDecoder被決定的。OMXCodec::Create()会回传Video Decoder给mVideoSource。

status_t AwesomePlayer::initVideoDecoder()

{

  mVideoSource = OMXCodec::Create(mClient.interface(),mVideoTrack->getFormat(),false,mVideoTrack);

}

sp<MediaSource> OMXCodec::Create(&omx, &meta, createEncoder, &source,matchComponentName)

{

  meta->findCString(kKeyMIMEType, &mime);

  findMatchingCodecs(mime, ..., &matchingCodecs); ........ (2)

  for (size_t i = 0; i < matchingCodecs.size(); ++i)

  {

    componentName = matchingCodecs[i].string();

    softwareCodec =InstantiateSoftwareCodec(componentName, ...); ..... (3)

    if (softwareCodec != NULL) return softwareCodec;

      err = omx->allocateNode(componentName, ..., &node); ... (4)

    if (err == OK)

    {

      codec = new OMXCodec(..., componentName, ...); ...... (5)

      return codec;

    }

   }

}

(2) 根据mVideoTrack的MIME从kDecoderInfo挑出合适的components

void OMXCodec::findMatchingCodecs(mime, ..., matchingCodecs)

{

  for (int index = 0;; ++index)

  {

    componentName = GetCodec(kDecoderInfo,sizeof(kDecoderInfo)/sizeof(kDecoderInfo[0]),mime,index);

    matchingCodecs->push(String8(componentName));

  }

}

static const CodecInfo kDecoderInfo[] =

{

  ...

  { MEDIA_MIMETYPE_VIDEO_MPEG4, "OMX.qcom.video.decoder.mpeg4" },

  { MEDIA_MIMETYPE_VIDEO_MPEG4, "OMX.TI.Video.Decoder" },

  { MEDIA_MIMETYPE_VIDEO_MPEG4, "M4vH263Decoder" },

  ...

}

GetCodec会依据MIME从kDecoderInfo挑出所有的component name，然后存到matchingCodecs中。

(3) 根据matchingCodecs中component的顺序，我们会先去检查其是否为software decoder

static sp<MediaSource> InstantiateSoftwareCodec(name, ...)
{
  FactoryInfo kFactoryInfo[] =
  {
    ...
    FACTORY_REF(M4vH263Decoder)
    ...
  };
  for (i = 0; i < sizeof(kFactoryInfo)/sizeof(kFactoryInfo[0]); ++i)
  {
    if (!strcmp(name, kFactoryInfo[i].name))
      return (*kFactoryInfo[i].CreateFunc)(source);
  }
}

所有的software decoder都会被列在kFactoryInfo中，我们藉由传进来的name来对应到适合的decoder。

(4) 如果該component不是software decoder，则试著去配置对应的OMX component

status_t OMX::allocateNode(name, ..., node)

{

  mMaster->makeComponentInstance(name,&OMXNodeInstance::kCallbacks,instance,handle);

}

OMX_ERRORTYPE OMXMaster::makeComponentInstance(name, ...)

{

  plugin->makeComponentInstance(name, ...);

}

OMX_ERRORTYPE OMXPVCodecsPlugin::makeComponentInstance(name, ...)

{

  return OMX_MasterGetHandle(..., name, ...);

}

OMX_ERRORTYPE OMX_MasterGetHandle(...)

{

  return OMX_GetHandle(...);

}

(5) 若该component為OMX deocder，则回传；否则继续检查下一个component

 

3.Video Buffer传输流程

下面介绍Stagefright中是如何和OMX Video decoder传递buffer

(1) OMXCodec会在一开始的时候透过read函式来传送未解码的data给decoder，并且要求decoder将解码后的data传回來

status_t OMXCodec::read(...)

{

  if(mInitialBufferSubmit)

  {

    mInitialBufferSubmit = false;

    drainInputBuffers(); //----- OMX_EmptyThisBuffer(清空InputBuffer)

    fillOutputBuffers(); //----- OMX_FillThisBuffer(填充OutputBuffer)

  }

  ...

}

void OMXCodec::drainInputBuffers()

{

  Vector<BufferInfo> *buffers = &mPortBuffers[kPortIndexInput];

  for (i = 0; i < buffers->size(); ++i)

  {

    drainInputBuffer(&buffers->editItemAt(i));

  }

}

void OMXCodec::drainInputBuffer(BufferInfo *info)

{

  mOMX->emptyBuffer(...);

}

void OMXCodec::fillOutputBuffers()

{

  Vector<BufferInfo> *buffers = &mPortBuffers[kPortIndexOutput];

  for (i = 0; i < buffers->size(); ++i)

  {

    fillOutputBuffer(&buffers->editItemAt(i));

  }

}

void OMXCodec::fillOutputBuffer(BufferInfo *info)

{

  mOMX->fillBuffer(...);

}

(2) Decoder从input port读取资料后，开始进行解码，并且回传EmptyBufferDone通知OMXCodec

void OMXCodec::on_message(const omx_message &msg)

{

  switch (msg.type)

  {

    case omx_message::EMPTY_BUFFER_DONE:

    {

      IOMX::buffer_id buffer = msg.u.extended_buffer_data.buffer;

      drainInputBuffer(&buffers->editItemAt(i));

    }

  }

}

OMXCodec收到EMPTY_BUFFER_DONE之后，继续传送下一个未解码的资料给decoder。

(3) Decoder将解码完的资料送到output port，并回传FillBufferDone通知OMXCodec

void OMXCodec::on_message(const omx_message &msg)

{

  switch (msg.type)

  {

    case omx_message::FILL_BUFFER_DONE:

    { 

      IOMX::buffer_id buffer = msg.u.extended_buffer_data.buffer;

      fillOutputBuffer(info);

      mFilledBuffers.push_back(i);

      mBufferFilled.signal();

    }

  }

}

OMXCodec收到FILL_BUFFER_DONE之后，将解码后的资料放入mFilledBuffers，发出mBufferFilled信号，并且要求decoder继续送出资料。

(4) read函式在后段等待mBufferFilled信号。当mFilledBuffers被填入资料后，read函式将其指定给buffer指标，并回传给AwesomePlayer

status_t OMXCodec::read(MediaBuffer **buffer, ...)

{

  ...

  while (mFilledBuffers.empty())

  {

    mBufferFilled.wait(mLock);

  }

  BufferInfo *info = &mPortBuffers[kPortIndexOutput].editItemAt(index);

  info->mMediaBuffer->add_ref();

  *buffer = info->mMediaBuffer;

}

 

4.Video Rendering

AwesomePlayer::onVideoEvent除了透过OMXCodec::read取得解码后的资料外，还必须将这些资料(mVideoBuffer)传给Video renderer，以便画到UIScreen上去。

(1) 要将mVideoBuffer中的资料绘制出来之前，必须先建立mVideoRenderer

void AwesomePlayer::onVideoEvent()

{

...

if (mVideoRenderer == NULL)

{

    initRenderer_l();

}

...

}

void AwesomePlayer::initRenderer_l()

{

  if (!strncmp("OMX.", component, 4))

  {

    mVideoRenderer = new AwesomeRemoteRenderer(mClient.interface()->createRenderer(mISurface,component,...)); .......... (2)

  }

  else

  {

    mVideoRenderer = new AwesomeLocalRenderer(...,component,mISurface); ............................ (3)

  }

}

(2) 如果video decoder是OMX component，則建立一個AwesomeRemoteRenderer作為mVideoRenderer
从上段的程式码(1)来看，AwesomeRemoteRenderer的本质是由OMX::createRenderer所创建的。createRenderer会先建立一个hardware renderer -- SharedVideoRenderer (libstagefrighthw.so)；若失败，则建立software renderer -- SoftwareRenderer (surface)。
sp<IOMXRenderer> OMX::createRenderer(...)
{
  VideoRenderer *impl = NULL;
  libHandle = dlopen("libstagefrighthw.so", RTLD_NOW);
  if (libHandle)
  {
    CreateRendererFunc func = dlsym(libHandle, ...);
    impl = (*func)(...); <----------------- Hardware Renderer
  }
  if (!impl)
  {
    impl = new SoftwareRenderer(...); <---- Software Renderer
  }
}

(3) 如果video decoder是software component，則建立一个AwesomeLocalRenderer作為mVideoRenderer
AwesomeLocalRenderer的constructor会呼叫本身的init函式，其所做的事和OMX::createRenderer一模一样。

void AwesomeLocalRenderer::init(...)

{

  mLibHandle = dlopen("libstagefrighthw.so", RTLD_NOW);

  if(mLibHandle)

  {

    CreateRendererFunc func = dlsym(...);

    mTarget = (*func)(...); <---------------- Hardware Renderer

  }

  if(mTarget == NULL)

  {

    mTarget = new SoftwareRenderer(...); <--- Software Renderer

  }

}

(4) mVideoRenderer一经建立就可以开始将解码后的资料传給它

void AwesomePlayer::onVideoEvent()

{

  if (!mVideoBuffer)

  {

    mVideoSource->read(&mVideoBuffer, ...);

  }

  [Check Timestamp]

  if (mVideoRenderer == NULL)

  {

    initRenderer_l();

  }

  mVideoRenderer->render(mVideoBuffer); <----- Render Data

}

 

5.Audio Playback的流程

这篇文章将会开始audio处理的流程。Stagefright中关於audio的部分是交由AudioPlayer来处理，它是在AwesomePlayer::play_l中被建立的。

 

(1) 当上层应用程式要求播放影音时，AudioPlayer同时被建立出来，并且被啟动

status_t AwesomePlayer::play_l()

{

...

mAudioPlayer = new AudioPlayer(mAudioSink, ...);

mAudioPlayer->start(...);

...

}

(2) AudioPlayer在啟动的过程中会先去读取第一笔解码后的资料，并且开啟audio output

status_t AudioPlayer::start(...)
{
mSource->read(&mFirstBuffer);
  if(mAudioSink.get() != NULL)
  {
    mAudioSink->open(..., &AudioPlayer::AudioSinkCallback, ...);
    mAudioSink->start();
  }
  else
  {
    mAudioTrack = new AudioTrack(..., &AudioPlayer::AudioCallback, ...);
    mAudioTrack->start();
  }
}
从AudioPlayer::start的程式码来看，AudioPlayer似乎并没有将mFirstBuffer传给audio output。

(3) 开启audio output的同时，AudioPlayer会将callback函式設給它，之后每当callback函式被呼叫，AudioPlayer便去audio decoder读取解码后的资料

size_t AudioPlayer::AudioSinkCallback(audioSink, buffer, size, ...)
{
  return fillBuffer(buffer, size);
}
void AudioPlayer::AudioCallback(..., info)
{
  buffer = info;
  fillBuffer(buffer->raw, buffer->size);
}
size_t AudioPlayer::fillBuffer(data, size)
{
  mSource->read(&mInputBuffer, ...);
  memcpy(data, mInputBuffer->data(), ...);
}

解码后audio资料的读取就是由callback函式所驱动，但是callback函式又是怎麼由audio output去驱动的，目前从程式码上还看不出来。另外一方面，从上面的程式片段可以看出，fillBuffer将资料(mInputBuffer)复製到data之后，audio output应该会去取用data。
至于audio decoder的工作流程则和video decoder相同，看参见上述第三部分Video Buffer传输流程

 

6.Audio和Video的同步

讲完了audio和video的处理流程，接下来要看的是audio和video同步化(synchronization)的问题。OpenCORE的做法是设置一个主clock，而audio和video就分别以此作為输出的依据。而在Stagefright中，audio的输出是透过callback函式来驱动，video则根据audio的timestamp来做同步。以下是详细的说明：
(1) 当callback函式驱动AudioPlayer读取解码后的资料时，AudioPlayer会取得两个时间戳 -- mPositionTimeMediaUs和mPositionTimeRealUs

size_t AudioPlayer::fillBuffer(data, size)
{
  ...
  mSource->read(&mInputBuffer, ...);
  mInputBuffer->meta_data()->findInt64(kKeyTime, &mPositionTimeMediaUs);
  mPositionTimeRealUs = ((mNumFramesPlayed + size_done / mFrameSize) * 1000000) /mSampleRate;
  ...
}

mPositionTimeMediaUs是资料裡面所载明的时间戳(timestamp)；mPositionTimeRealUs则是播放此资料的实际时间(依据frame number及sample rate得出)。
(2) Stagefright中的video便依据从AudioPlayer得出来之两个时间戳的差值，作為播放的依据

void AwesomePlayer::onVideoEvent()
{
...
  mVideoSource->read(&mVideoBuffer, ...);
  mVideoBuffer->meta_data()->findInt64(kKeyTime, &timeUs);
  mAudioPlayer->getMediaTimeMapping(&realTimeUs, &mediaTimeUs);
  mTimeSourceDeltaUs = realTimeUs - mediaTimeUs;
  nowUs = ts->getRealTimeUs() - mTimeSourceDeltaUs;
  latenessUs = nowUs - timeUs;
...
}

AwesomePlayer從AudioPlayer取得realTimeUs(即mPositionTimeRealUs)和mediaTimeUs(即mPositionTimeMediaUs)，并算出其差值mTimeSourceDeltaUs。
(3) 最后我们将该video资料做排程

void AwesomePlayer::onVideoEvent()
{
  ...
  if (latenessUs > 40000)
  {
    mVideoBuffer->release();
    mVideoBuffer = NULL;
    postVideoEvent_l();
    return;
  }
  if (latenessUs < -10000)
  {
    postVideoEvent_l(10000);
    return;
  }
  mVideoRenderer->render(mVideoBuffer);
  ...
}