ExoPlayer 自适应流切换分析

2023-10-23 17:58:19 浏览数 (3)

一、前言

自适应流切换属于多路流切换的方式中的一种,ExoPlayer作为MediaCodec使用的集大成者,不仅具备通过MergingMediaSource实现不同流的组合切换,同样也具备基于MGEG-DASH、HLS、smoothing-stream 协议的的自适应流切换。当然,在项目中每种方案的选型都要充分考虑团队条件。

主要区别如下:

  • MergingMediaSource 方式更适合团队人力有限,后台服务支持有限的情况,不需要在资源传输和编码上做更多的考虑,普通的CDN部署就可以,相比更加节省成本。而自适应流相对要求比较专业,对服务器的部署、资源分片、资源编码也是有一定要求的。
  • MergingMediaSource 方式可实现不同编码的流合并,而自适应流方面部分协议如HLS有较严格的要求,主要要求是ts分片的编码尽可能保持一致,这样做的目的是为尽可能实现MediaCodec的重复利用。当然,MergingMediaSource方式如果每路流的Format差别不大,视频解码器完全可以通过PPS、SPS或flush buffer 的方式实现MediaCodec利用,音频解码器也是可以通过输入特定字节特征实现MediaCodec复用。
  • 在ExoPlayer中,MergingMediaSource 中的同一类型(视频类型、音频类型、字幕类型等)的数据,由于缺乏必要的码率参数, 无法将相似Format的Track数据合并为一组,因此使用的FixedTrackSelection对同一类型的资源,自然而然也不支持多路流的自动切换。而自适应流完成可以实现Format分组,最终创建AdaptiveTrackSelection 动态管理各路流。

二、基础知识点

前言的内容对于初学ExoPlayer开发者而言还是有些抽象,下面我们梳理一下ExoPlayer的关键类,方便理解本篇内容。

  • Renderer渲染器:负责解码器的Format支持能力检测、解码器的注册、解码器的销毁、解码器的复用、采样数据读取、数据渲染或输出、丢帧、跳帧以及音画同步等工作。ExoPlayer支持Renderer的拆解、组合、关闭和启用,也支持自定义的解码器接入,比如通过SimpleDecoder实现FFMPEG对视频和音频的解码渲染。
  • MediaClock媒体时钟:负责音画同步、播放进度管理等工作。在ExoPlayer中国存在两种时钟,一种是独立时钟StandaloneMediaClock,另一种是通过音频Renderer实现的Audio Master模式的时钟。
  • Extractor 解封装器:负责将媒体资源中的每一路流的Moov信息、采样表、Format、采样数据(如SPS、PPS、各种帧数据)拆解出来,同时会对一些数据,便于Track和Format的选择以及码流切换。ExoPlayer内置了大量的解封装器,同样也支持自定义的Extractor来实现特定目的。当然,自适应流Format的解析一般是通过MediaSource去解析的,只有视频容器需要通过Extractor去解析。
  • Decoder解码器:负责解码采样数据,其中MediaCodec具备使用硬解和软解的能力的同时,而且支持渲染。
  • TrackSelector:多路流切换的核心类,负责通过Format检测实现TrackSelection的分组、以及TrackGroup和Renderer、TrackGroup和TrackSelection的配对工作。
  • DataSource 数据源:负责提供数据。
  • MediaSource 媒体源:在ExoPlayer中,得益于对从DataSource中抽象出了MediaSource,使得ExoPlayer在多路流管理方面更加灵活方便。
  • SeekPoint:在ExoPlayer中,SeekPoint 往往是IDR帧即将开始的位置。
  • FixedTrackSelection:固定流选择器,播放过程中TrackSelection保持固定,普通协议默认的TrackSelection,但也因为这个原因,使得其适用于MergingMediaSource方式的多路流切换。
  • AdptiveTrackSelection: 自适应流选择器,可以根据Bandwidth实现动态选择分片。当然,可以通过一些策略,实现用户自行的切换,类似bilibili的码流切换。
  • TrackGroup : 同一类型资源的Track Format 分组。
  • MapTrackInfo : 这个类实际上是一个Renderer和TrackGroup相关信息的集合类,主要保存Renderer能力信息和TrackGroup信息,某种程度上可以看到数据格式和Renderer全局信息。
  • selection分组原理:TrackSelector会通过SELECTION_ELIGIBILITY_FIXED、SELECTION_ELIGIBILITY_ADAPTIVE对资源进行分组,当然前提是各个Format之间能够相互兼容,具体兼容逻辑参考VideoTrackInfo类和AudioTrackInfo类。
  • Bandwidth:ExoPlayer中对网速检测的重要工具,检测结果用于AdaptiveTrackSelection进行分片选择。
  • DefaultMediaSourceFactory 用于实现DataSource转换为MediaSource的路由工厂,通过mineType和uri后缀识别出创建那种MediaSource。

三、自适应流切换分析

3.1 原理图

在不同网速时自动切换为兼容当前bitrate的媒体流,匹配条件一般在自适应流的清单文件中就已经提前设定了,保证当前网络的bitrate大于清单协议中媒体流的最低bandWidth,就可以切换到指定的媒体流Track。

通过原理图我们可以了解到以下信息:

  • 默认情况下,自适应流的切换不需要查找SeekPoint,而是通过选择下一个分片实现。
  • 默认情况下,自适应流通过网速检测实现了分片切换。
  • 从图上可知,每个分片的的播放时间和I帧的开始位置也需要做到严格对齐。

注意:之所以强调默认情况,一个重要的原因是ExopPlayer具备高度可扩展性,我们可以通过修改部分代码实现其他行为。

3.2 核心逻辑

核心逻辑主要为:

  • 清单文件解析
  • 建立Renderer与TrackGroup、Selection之间的映射关系
  • 开始分片加载
  • 网速带宽检测 与 AdaptiveTrackSelection 选择合适的分片
  • 解码器复用或重启
  • 完成切换

3.2.1 自适应流清单文件解析

ExoPlayer中支持DASH、HLS、Smoothing-Stream协议,我们这里以HLS和DASH协议进行流程分析,毕竟目前使用Smoothing-Stream也就微软自己为主。接下来先看看HLS和DASH的清单文件,方便我们后续测试。

3.2.1.1 hls 协议清单文件

代码语言:javascript复制
#EXTM3U

#EXT-X-MEDIA:TYPE=AUDIO,GROUP-ID="bipbop_audio",LANGUAGE="eng",NAME="BipBop Audio 1",AUTOSELECT=YES,DEFAULT=YES
#EXT-X-MEDIA:TYPE=AUDIO,GROUP-ID="bipbop_audio",LANGUAGE="eng",NAME="BipBop Audio 2",AUTOSELECT=NO,DEFAULT=NO,URI="alternate_audio_aac/prog_index.m3u8"


#EXT-X-MEDIA:TYPE=SUBTITLES,GROUP-ID="subs",NAME="English",DEFAULT=YES,AUTOSELECT=YES,FORCED=NO,LANGUAGE="en",CHARACTERISTICS="public.accessibility.transcribes-spoken-dialog, public.accessibility.describes-music-and-sound",URI="subtitles/eng/prog_index.m3u8"
#EXT-X-MEDIA:TYPE=SUBTITLES,GROUP-ID="subs",NAME="English (Forced)",DEFAULT=NO,AUTOSELECT=NO,FORCED=YES,LANGUAGE="en",URI="subtitles/eng_forced/prog_index.m3u8"
#EXT-X-MEDIA:TYPE=SUBTITLES,GROUP-ID="subs",NAME="Français",DEFAULT=NO,AUTOSELECT=YES,FORCED=NO,LANGUAGE="fr",CHARACTERISTICS="public.accessibility.transcribes-spoken-dialog, public.accessibility.describes-music-and-sound",URI="subtitles/fra/prog_index.m3u8"
#EXT-X-MEDIA:TYPE=SUBTITLES,GROUP-ID="subs",NAME="Français (Forced)",DEFAULT=NO,AUTOSELECT=NO,FORCED=YES,LANGUAGE="fr",URI="subtitles/fra_forced/prog_index.m3u8"
#EXT-X-MEDIA:TYPE=SUBTITLES,GROUP-ID="subs",NAME="Español",DEFAULT=NO,AUTOSELECT=YES,FORCED=NO,LANGUAGE="es",CHARACTERISTICS="public.accessibility.transcribes-spoken-dialog, public.accessibility.describes-music-and-sound",URI="subtitles/spa/prog_index.m3u8"
#EXT-X-MEDIA:TYPE=SUBTITLES,GROUP-ID="subs",NAME="Español (Forced)",DEFAULT=NO,AUTOSELECT=NO,FORCED=YES,LANGUAGE="es",URI="subtitles/spa_forced/prog_index.m3u8"
#EXT-X-MEDIA:TYPE=SUBTITLES,GROUP-ID="subs",NAME="日本語",DEFAULT=NO,AUTOSELECT=YES,FORCED=NO,LANGUAGE="ja",CHARACTERISTICS="public.accessibility.transcribes-spoken-dialog, public.accessibility.describes-music-and-sound",URI="subtitles/jpn/prog_index.m3u8"
#EXT-X-MEDIA:TYPE=SUBTITLES,GROUP-ID="subs",NAME="日本語 (Forced)",DEFAULT=NO,AUTOSELECT=NO,FORCED=YES,LANGUAGE="ja",URI="subtitles/jpn_forced/prog_index.m3u8"


#EXT-X-STREAM-INF:BANDWIDTH=263851,CODECS="mp4a.40.2, avc1.4d400d",RESOLUTION=416x234,AUDIO="bipbop_audio",SUBTITLES="subs"
gear1/prog_index.m3u8
#EXT-X-I-FRAME-STREAM-INF:BANDWIDTH=28451,CODECS="avc1.4d400d",URI="gear1/iframe_index.m3u8"

#EXT-X-STREAM-INF:BANDWIDTH=577610,CODECS="mp4a.40.2, avc1.4d401e",RESOLUTION=640x360,AUDIO="bipbop_audio",SUBTITLES="subs"
gear2/prog_index.m3u8
#EXT-X-I-FRAME-STREAM-INF:BANDWIDTH=181534,CODECS="avc1.4d401e",URI="gear2/iframe_index.m3u8"

#EXT-X-STREAM-INF:BANDWIDTH=915905,CODECS="mp4a.40.2, avc1.4d401f",RESOLUTION=960x540,AUDIO="bipbop_audio",SUBTITLES="subs"
gear3/prog_index.m3u8
#EXT-X-I-FRAME-STREAM-INF:BANDWIDTH=297056,CODECS="avc1.4d401f",URI="gear3/iframe_index.m3u8"

#EXT-X-STREAM-INF:BANDWIDTH=1030138,CODECS="mp4a.40.2, avc1.4d401f",RESOLUTION=1280x720,AUDIO="bipbop_audio",SUBTITLES="subs"
gear4/prog_index.m3u8
#EXT-X-I-FRAME-STREAM-INF:BANDWIDTH=339492,CODECS="avc1.4d401f",URI="gear4/iframe_index.m3u8"

#EXT-X-STREAM-INF:BANDWIDTH=1924009,CODECS="mp4a.40.2, avc1.4d401f",RESOLUTION=1920x1080,AUDIO="bipbop_audio",SUBTITLES="subs"
gear5/prog_index.m3u8
#EXT-X-I-FRAME-STREAM-INF:BANDWIDTH=669554,CODECS="avc1.4d401f",URI="gear5/iframe_index.m3u8"

#EXT-X-STREAM-INF:BANDWIDTH=41457,CODECS="mp4a.40.2",AUDIO="bipbop_audio",SUBTITLES="subs"
gear0/prog_index.m3u8

3.2.1.2 DASH协议清单文件

代码语言:javascript复制
<?xml version="1.0" encoding="UTF-8"?>
<!--Generated with https://github.com/google/shaka-packager version 97fc982-release-->
<MPD xmlns="urn:mpeg:dash:schema:mpd:2011" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xlink="http://www.w3.org/1999/xlink" xsi:schemaLocation="urn:mpeg:dash:schema:mpd:2011 DASH-MPD.xsd" xmlns:cenc="urn:mpeg:cenc:2013" minBufferTime="PT2S" type="static" profiles="urn:mpeg:dash:profile:isoff-on-demand:2011" mediaPresentationDuration="PT734S">
  <Period id="0">
    <AdaptationSet id="0" contentType="audio" lang="en">
      <Representation id="0" bandwidth="131596" codecs="mp4a.40.2" mimeType="audio/mp4" audioSamplingRate="44100">
        <AudioChannelConfiguration schemeIdUri="urn:mpeg:dash:23003:3:audio_channel_configuration:2011" value="2"/>
        <BaseURL>tears_audio_eng.mp4</BaseURL>
        <SegmentBase indexRange="745-1664" timescale="44100">
          <Initialization range="0-744"/>
        </SegmentBase>
      </Representation>
    </AdaptationSet>
    <AdaptationSet id="1" contentType="video" maxWidth="1920" maxHeight="856" frameRate="12288/512" par="38:17">
      <Representation id="1" bandwidth="769255" codecs="avc1.42c01e" mimeType="video/mp4" sar="852:857" width="320" height="142">
        <BaseURL>tears_h264_baseline_240p_800.mp4</BaseURL>
        <SegmentBase indexRange="827-1602" timescale="12288">
          <Initialization range="0-826"/>
        </SegmentBase>
      </Representation>
      <Representation id="2" bandwidth="1774254" codecs="avc1.4d401f" mimeType="video/mp4" sar="2242:2249" width="854" height="380">
        <BaseURL>tears_h264_main_480p_2000.mp4</BaseURL>
        <SegmentBase indexRange="829-1604" timescale="12288">
          <Initialization range="0-828"/>
        </SegmentBase>
      </Representation>
      <Representation id="3" bandwidth="7203938" codecs="avc1.4d4028" mimeType="video/mp4" sar="855:857" width="1280" height="570">
        <BaseURL>tears_h264_main_720p_8000.mp4</BaseURL>
        <SegmentBase indexRange="830-1605" timescale="12288">
          <Initialization range="0-829"/>
        </SegmentBase>
      </Representation>
      <Representation id="4" bandwidth="18316946" codecs="avc1.64002a" mimeType="video/mp4" sar="856:857" width="1920" height="856">
        <BaseURL>tears_h264_high_1080p_20000.mp4</BaseURL>
        <SegmentBase indexRange="832-1607" timescale="12288">
          <Initialization range="0-831"/>
        </SegmentBase>
      </Representation>
    </AdaptationSet>
  </Period>
</MPD>

两种协议的共同点中我们可以明确的发现又很多共同点:

  • bandWidth: 网络带宽,也就是下载的速率,但是在清单文件中一般表示的是支持该数据流的最低下载速率。
  • mimeType : 资源类型
  • codecs: 资源编码类型
  • width: 视频宽度
  • height: 视频高度

和其他协议的资源不同的是,由于使用清单文件的原因,基本可以实现在解封装之前就能获取到必要的Format信息。

解析时清单文件时,如果使用的是HLS协议,ExoPlayer内部利用HlsPlaylistParser类作为清单文件解析工具,如果是DASH则使用DashManifestParser解析清单,依次类推,smoothing-stream 使用SsmanifestParser进行将进行清单文件解析。

解析的流程主要如下

  • 使用DefaultMediaSourceFactory创建对应的自适应流MediaSource,如HlsMediaSource、DashMediaSource、SsMediaSource。
  • 自适应流MediaSource创建Parser解析工具,当然HlsMediaSource比较特殊,不直接持有Parser,而是通过DefaultHlsPlaylistTracker时线对Parser的调度。
  • 创建Loader和ParsingLoadable,ParsingLoadable类似于Runnable,属于加载任务,在ParsingLoadable中可以实现边加载边解析。
  • 对每组Track Format进行分组,使用TrackGroup保存分组信息。

3.2.2 Renderer与TrackGroup、Selection选择

在ExoPlayer中,DefaultTrackSelector主要负责这项工作,核心逻辑如下

代码语言:javascript复制
  @Override
  protected final Pair<@NullableType RendererConfiguration[], @NullableType ExoTrackSelection[]>
      selectTracks(
          MappedTrackInfo mappedTrackInfo,
          @Capabilities int[][][] rendererFormatSupports,
          @AdaptiveSupport int[] rendererMixedMimeTypeAdaptationSupport,
          MediaPeriodId mediaPeriodId,
          Timeline timeline)
          throws ExoPlaybackException {
    Parameters parameters;
    synchronized (lock) {
      parameters = this.parameters;
      if (parameters.constrainAudioChannelCountToDeviceCapabilities
          && Util.SDK_INT >= 32
          && spatializer != null) {
        // Initialize the spatializer now so we can get a reference to the playback looper with
        // Looper.myLooper().
        spatializer.ensureInitialized(this, checkStateNotNull(Looper.myLooper()));
      }
    }
    int rendererCount = mappedTrackInfo.getRendererCount();

    // 建立Renderers 与 TrackGroup 映射,注意,我们前面提到的MappedTrackInfo,保存有Renderer和TrackGroup的相关信息
    // 这里也会尝试对不同TrackGroup Format进行依据兼容性进行合并分组到definition中
    ExoTrackSelection.@NullableType Definition[] definitions =
        selectAllTracks(
            mappedTrackInfo,
            rendererFormatSupports,
            rendererMixedMimeTypeAdaptationSupport,
            parameters);

    //对筛选出来的defintions 二次过滤
    applyTrackSelectionOverrides(mappedTrackInfo, parameters, definitions);
    applyLegacyRendererOverrides(mappedTrackInfo, parameters, definitions);


    //三次过滤,对Renderer不能使用解除映关系
    // Disable renderers if needed.
    for (int i = 0; i < rendererCount; i  ) {
      @C.TrackType int rendererType = mappedTrackInfo.getRendererType(i);
      if (parameters.getRendererDisabled(i)
          || parameters.disabledTrackTypes.contains(rendererType)) {
        definitions[i] = null;
      }
    }

  //建立TrackGroup与Selection之间的映射关系,并且传入BandwidthMeter用于获取网速策略的数据
    @NullableType
    ExoTrackSelection[] rendererTrackSelections =
        trackSelectionFactory.createTrackSelections(
            definitions, getBandwidthMeter(), mediaPeriodId, timeline);

    // Initialize the renderer configurations to the default configuration for all renderers with
    // selections, and null otherwise.
//创建Renderer初始化配置
    @NullableType
    RendererConfiguration[] rendererConfigurations = new RendererConfiguration[rendererCount];
    for (int i = 0; i < rendererCount; i  ) {
      @C.TrackType int rendererType = mappedTrackInfo.getRendererType(i);
      boolean forceRendererDisabled =
          parameters.getRendererDisabled(i) || parameters.disabledTrackTypes.contains(rendererType);
      boolean rendererEnabled =
          !forceRendererDisabled
              && (mappedTrackInfo.getRendererType(i) == C.TRACK_TYPE_NONE
                  || rendererTrackSelections[i] != null);
      rendererConfigurations[i] = rendererEnabled ? RendererConfiguration.DEFAULT : null;
    }

    // Configure audio and video renderers to use tunneling if appropriate.
    if (parameters.tunnelingEnabled) {
     //这种属于隧道渲染,具体流程好像是MediaCodec不负责解码,可以将未解码的数据输出到驱动层,由驱动层处理,如ac3音频数据直接输出到AudioTrack中,这方面资料太少,后续在研究。
      maybeConfigureRenderersForTunneling(
          mappedTrackInfo, rendererFormatSupports, rendererConfigurations, rendererTrackSelections);
    }

    return Pair.create(rendererConfigurations, rendererTrackSelections);
  }

这部分逻辑相对很复杂,但是我们com.google.android.exoplayer2.trackselection.AdaptiveTrackSelection.Factory#createTrackSelections要做必要的分析。

在ExoPlayer中默认使用改工厂适配Selection,具体逻辑如下

代码语言:javascript复制
   @Override
    public final @NullableType ExoTrackSelection[] createTrackSelections(
        @NullableType Definition[] definitions,
        BandwidthMeter bandwidthMeter,
        MediaPeriodId mediaPeriodId,
        Timeline timeline) {
      ImmutableList<ImmutableList<AdaptationCheckpoint>> adaptationCheckpoints =
          getAdaptationCheckpoints(definitions);
      ExoTrackSelection[] selections = new ExoTrackSelection[definitions.length];
      for (int i = 0; i < definitions.length; i  ) {
        @Nullable Definition definition = definitions[i];
        if (definition == null || definition.tracks.length == 0) {
          continue;
        }
       //遍历所有的definition,对分组中tracks 数量为1的创建FixedTrackSelection,对存在多个的创建AdaptiveTrackSeletion
        selections[i] =
            definition.tracks.length == 1
                ? new FixedTrackSelection(
                    definition.group,
                    /* track= */ definition.tracks[0],
                    /* type= */ definition.type)
                : createAdaptiveTrackSelection(
                    definition.group,
                    definition.tracks,
                    definition.type,
                    bandwidthMeter,
                    adaptationCheckpoints.get(i));
      }
      return selections;
    }

至此,Renderer 、TrackGroup、selection的映射关系建立完成。

那么这里有个疑问,如果利用MergingMediaSource合并多路流并修改参数,能否也实现AdaptiveTrackSelection,进入试下自适应能力?答案是否定的,因为MergingMediaSource合并的是完整的资源,在使用过程中并不会调用TrackSelection相关方法,当然ExoPlayer也没有实现资源的动态分片。

3.2.3 分片加载

DASH、HLS、Smoothing-Stream 加载分片的时候,单个分片都是用各自的实现的ChunkSource类,但是对于存在多个分片情况,ExoPlayer利用ChunkSampleStream和HlsSampleStreamWrapper将分片队列管理起来,核心方法是continueLoading中的实现,大致相同,这里以HlsSampleStreamWrapper的为参考。

代码语言:javascript复制
 public boolean continueLoading(long positionUs) {
    if (loadingFinished || loader.isLoading() || loader.hasFatalError()) {
      return false;
    }

    boolean pendingReset = isPendingReset();
    //获取资源队列
    List<BaseMediaChunk> chunkQueue; 
   //获取上次加载资源时间位置
    long loadPositionUs;
    if (pendingReset) {
      chunkQueue = Collections.emptyList();
      loadPositionUs = pendingResetPositionUs;
    } else {
      chunkQueue = readOnlyMediaChunks;
      loadPositionUs = getLastMediaChunk().endTimeUs;
    }
//获取下一个分片
    chunkSource.getNextChunk(positionUs, loadPositionUs, chunkQueue, nextChunkHolder);
    boolean endOfStream = nextChunkHolder.endOfStream;
    @Nullable Chunk loadable = nextChunkHolder.chunk;
    nextChunkHolder.clear();

    if (endOfStream) {
    // 如果没有资源可以加载了,标记加载结束
      pendingResetPositionUs = C.TIME_UNSET;
      loadingFinished = true;
      return true;
    }

    if (loadable == null) {
      return false;
    }
    //下面逻辑是加载状态直接的判断
    loadingChunk = loadable;
    if (isMediaChunk(loadable)) {
      BaseMediaChunk mediaChunk = (BaseMediaChunk) loadable;
      if (pendingReset) {
        // Only set the queue start times if we're not seeking to a chunk boundary. If we are
        // seeking to a chunk boundary then we want the queue to pass through all of the samples in
        // the chunk. Doing this ensures we'll always output the keyframe at the start of the chunk,
        // even if its timestamp is slightly earlier than the advertised chunk start time.
        if (mediaChunk.startTimeUs != pendingResetPositionUs) {
          primarySampleQueue.setStartTimeUs(pendingResetPositionUs);
          for (SampleQueue embeddedSampleQueue : embeddedSampleQueues) {
            embeddedSampleQueue.setStartTimeUs(pendingResetPositionUs);
          }
        }
        pendingResetPositionUs = C.TIME_UNSET;
      }
      mediaChunk.init(chunkOutput);
      mediaChunks.add(mediaChunk);
    } else if (loadable instanceof InitializationChunk) {
      ((InitializationChunk) loadable).init(chunkOutput);
    }
    long elapsedRealtimeMs =
        loader.startLoading(
            loadable, this, loadErrorHandlingPolicy.getMinimumLoadableRetryCount(loadable.type));
    mediaSourceEventDispatcher.loadStarted(
        new LoadEventInfo(loadable.loadTaskId, loadable.dataSpec, elapsedRealtimeMs),
        loadable.type,
        primaryTrackType,
        loadable.trackFormat,
        loadable.trackSelectionReason,
        loadable.trackSelectionData,
        loadable.startTimeUs,
        loadable.endTimeUs);
    return true;
  }

getNextChunk是continueLoading中的核心方法,继续看代码。

代码语言:javascript复制
  public final void getNextChunk(
      long playbackPositionUs,
      long loadPositionUs,
      List<? extends MediaChunk> queue,
      ChunkHolder out) {
    if (fatalError != null) {
      return;
    }

    StreamElement streamElement = manifest.streamElements[streamElementIndex];
    if (streamElement.chunkCount == 0) {
      // There aren't any chunks for us to load.
      //没有分片可以加载了
      out.endOfStream = !manifest.isLive;
      return;
    }

    int chunkIndex;  //获取最后一次加载片段的索引
    if (queue.isEmpty()) {
      chunkIndex = streamElement.getChunkIndex(loadPositionUs);
    } else {
      chunkIndex =
          (int) (queue.get(queue.size() - 1).getNextChunkIndex() - currentManifestChunkOffset);
      if (chunkIndex < 0) {
      //这个片段不存在,说明加载片段不完整
        // This is before the first chunk in the current manifest.
        fatalError = new BehindLiveWindowException();
        return;
      }
    }

    if (chunkIndex >= streamElement.chunkCount) {
      // This is beyond the last chunk in the current manifest.
      //这个索引不合法,直接停止加载
      out.endOfStream = !manifest.isLive;
      return;
    }

    long bufferedDurationUs = loadPositionUs - playbackPositionUs;
    long timeToLiveEdgeUs = resolveTimeToLiveEdgeUs(playbackPositionUs);

    MediaChunkIterator[] chunkIterators = new MediaChunkIterator[trackSelection.length()];
    for (int i = 0; i < chunkIterators.length; i  ) {
      int trackIndex = trackSelection.getIndexInTrackGroup(i);
      chunkIterators[i] = new StreamElementIterator(streamElement, trackIndex, chunkIndex);
    }

    //最关键的地方,这里会根据网速,筛选出下一个分片
    trackSelection.updateSelectedTrack(
        playbackPositionUs, bufferedDurationUs, timeToLiveEdgeUs, queue, chunkIterators);

    long chunkStartTimeUs = streamElement.getStartTimeUs(chunkIndex);
    long chunkEndTimeUs = chunkStartTimeUs   streamElement.getChunkDurationUs(chunkIndex);
    long chunkSeekTimeUs = queue.isEmpty() ? loadPositionUs : C.TIME_UNSET;
    int currentAbsoluteChunkIndex = chunkIndex   currentManifestChunkOffset;

    //这里是重点,获取选下一个需要加载的分片
    int trackSelectionIndex = trackSelection.getSelectedIndex();

   //获取分片解封装器
    ChunkExtractor chunkExtractor = chunkExtractors[trackSelectionIndex];

    int manifestTrackIndex = trackSelection.getIndexInTrackGroup(trackSelectionIndex);
    Uri uri = streamElement.buildRequestUri(manifestTrackIndex, chunkIndex);
    //绑定要加载的片段信息
    out.chunk =
        newMediaChunk(
            trackSelection.getSelectedFormat(),
            dataSource,
            uri,
            currentAbsoluteChunkIndex,
            chunkStartTimeUs,
            chunkEndTimeUs,
            chunkSeekTimeUs,
            trackSelection.getSelectionReason(),
            trackSelection.getSelectionData(),
            chunkExtractor);
  }

3.2.4 网速检测和AdaptiveTrackSelection分片选择

网速检测使用的默认的DefaultBandWidthMeter进行测试,具体原理是监控数据的某一段时间的下载流量,计算出平均网速。AdaptiveTrack Selection#updateSelectedTrack中会利用选择测量数据,重新计算使用哪个队列的数据。

代码语言:javascript复制
int newSelectedIndex = determineIdealSelectedIndex(nowMs, chunkDurationUs);

核心逻辑就是通过bitrate去做比较

代码语言:javascript复制
 @Override
  public void updateSelectedTrack(
      long playbackPositionUs,
      long bufferedDurationUs,
      long availableDurationUs,
      List<? extends MediaChunk> queue,
      MediaChunkIterator[] mediaChunkIterators) {
    long nowMs = clock.elapsedRealtime();
    long chunkDurationUs = getNextChunkDurationUs(mediaChunkIterators, queue);

    // Make initial selection
    if (reason == C.SELECTION_REASON_UNKNOWN) {
      reason = C.SELECTION_REASON_INITIAL;
      //初始化或者还没有选择过则直接一次性选择,不走兼容流程
      selectedIndex = determineIdealSelectedIndex(nowMs, chunkDurationUs);
      return;
    }

    int previousSelectedIndex = selectedIndex;
    @C.SelectionReason int previousReason = reason;
    int formatIndexOfPreviousChunk =
        queue.isEmpty() ? C.INDEX_UNSET : indexOf(Iterables.getLast(queue).trackFormat);
    if (formatIndexOfPreviousChunk != C.INDEX_UNSET) {
      previousSelectedIndex = formatIndexOfPreviousChunk;
      previousReason = Iterables.getLast(queue).trackSelectionReason;
    }
   //选择匹配bitrate的format
    int newSelectedIndex = determineIdealSelectedIndex(nowMs, chunkDurationUs);
    //判断如果该format所在的Track不在黑名单中,则走兼容逻辑
    if (!isBlacklisted(previousSelectedIndex, nowMs)) {
      // Revert back to the previous selection if conditions are not suitable for switching.
      Format currentFormat = getFormat(previousSelectedIndex);
      Format selectedFormat = getFormat(newSelectedIndex);
      //注意:进入这个地方就是一个坑点,如果切码流失败,原因是这里通过一些条件又给重置回去了
      long minDurationForQualityIncreaseUs =
          minDurationForQualityIncreaseUs(availableDurationUs, chunkDurationUs);
      if (selectedFormat.bitrate > currentFormat.bitrate
          && bufferedDurationUs < minDurationForQualityIncreaseUs) {
        // The selected track is a higher quality, but we have insufficient buffer to safely switch
        // up. Defer switching up for now.
        //如果选择的码流大于当前的,但是buffering的数据不够去安全的切换,因此还是选择当前Track 
        newSelectedIndex = previousSelectedIndex;
      } else if (selectedFormat.bitrate < currentFormat.bitrate
          && bufferedDurationUs >= maxDurationForQualityDecreaseUs) {
        // The selected track is a lower quality, but we have sufficient buffer to defer switching
        // down for now.
        //选择的码流小于当前的,但是buffer数据是足够,不至于去切换
        newSelectedIndex = previousSelectedIndex;
      }
    }
    // If we adapted, update the trigger.
//触发选择条件
    reason =
        newSelectedIndex == previousSelectedIndex ? previousReason : C.SELECTION_REASON_ADAPTIVE;
    selectedIndex = newSelectedIndex;
  }

上面提到2个坑点,bufferedDurationUs 大小影响切换,因此在项目中有必要规避此问题。

核心点determineIdealSelectedIndex,这里的逻辑就是获取当前带宽,然后匹配Selection中的采样队列

代码语言:javascript复制
 private int determineIdealSelectedIndex(long nowMs, long chunkDurationUs) {
     //获取带宽数据
    long effectiveBitrate = getAllocatedBandwidth(chunkDurationUs);
    int lowestBitrateAllowedIndex = 0;
    for (int i = 0; i < length; i  ) {
      if (nowMs == Long.MIN_VALUE || !isBlacklisted(i, nowMs)) {
        Format format = getFormat(i);
       //获取小于当前网速的Format
        if (canSelectFormat(format, format.bitrate, effectiveBitrate)) {
          return i;
        } else {
          lowestBitrateAllowedIndex = i;
        }
      }
    }
    return lowestBitrateAllowedIndex;
  }

带宽数据获取

代码语言:javascript复制
private long getTotalAllocatableBandwidth(long chunkDurationUs) {
  //注意这里将带宽x0.7f,实际上比实际值偏小了,因此设置网速时一定要除以0.7f
    long cautiousBandwidthEstimate =
        (long) (bandwidthMeter.getBitrateEstimate() * bandwidthFraction);

    long timeToFirstByteEstimateUs = bandwidthMeter.getTimeToFirstByteEstimateUs();
    if (timeToFirstByteEstimateUs == C.TIME_UNSET || chunkDurationUs == C.TIME_UNSET) {
    //默认情况下回到这里,带宽除以当前播放的速度 (倍速)
      return (long) (cautiousBandwidthEstimate / playbackSpeed); 
    }
    float availableTimeToLoadUs =
        max(chunkDurationUs / playbackSpeed - timeToFirstByteEstimateUs, 0);
    return (long) (cautiousBandwidthEstimate * availableTimeToLoadUs / chunkDurationUs);
  }

3.2.5 解码器的复用和重启

由于每种采样队列的分片Format有一些差别,可能需要解码器检测到格式变化。这个时候解码器可能需要重启,当然重启是多路流切换的最基本的做法,但是这种往往会出现卡顿或者短暂的黑屏,体验反而比较差。ExoPlayer对于无论是MergingMediaSource方式的多路流切换还是自适应流的切换导致onInputFormatChanged被调用做了相当多的优化,从而实现解码器的重复利用。

核心逻辑如下:

代码语言:javascript复制
  protected DecoderReuseEvaluation onInputFormatChanged(FormatHolder formatHolder)
      throws ExoPlaybackException {
    waitingForFirstSampleInFormat = true;
    Format newFormat = checkNotNull(formatHolder.format);
    if (newFormat.sampleMimeType == null) {
      // If the new format is invalid, it is either a media bug or it is not intended to be played.
      // See also https://github.com/google/ExoPlayer/issues/8283.

      throw createRendererException(
          new IllegalArgumentException(),
          newFormat,
          PlaybackException.ERROR_CODE_DECODING_FORMAT_UNSUPPORTED);
    }
    setSourceDrmSession(formatHolder.drmSession);
    inputFormat = newFormat;

    if (bypassEnabled) {
     //等到队列数据被清空后初始化
      bypassDrainAndReinitialize = true;
      return null; // Need to drain batch buffer first.
    }

    //如果Codec已经被释放或者还没有创建的情况下,重新初始化
    if (codec == null) {
      availableCodecInfos = null;
      maybeInitCodecOrBypass();
      return null;
    }

    // We have an existing codec that we may need to reconfigure, re-initialize, or release to
    // switch to bypass. If the existing codec instance is kept then its operating rate and DRM
    // session may need to be updated.

    // Copy the current codec and codecInfo to local variables so they remain accessible if the
    // member variables are updated during the logic below.
    MediaCodecAdapter codec = this.codec;
    MediaCodecInfo codecInfo = this.codecInfo;

    Format oldFormat = codecInputFormat;
    if (drmNeedsCodecReinitialization(codecInfo, newFormat, codecDrmSession, sourceDrmSession)) {
      drainAndReinitializeCodec();
      return new DecoderReuseEvaluation(
          codecInfo.name,
          oldFormat,
          newFormat,
          REUSE_RESULT_NO,
          DISCARD_REASON_DRM_SESSION_CHANGED);
    }
    boolean drainAndUpdateCodecDrmSession = sourceDrmSession != codecDrmSession;
    Assertions.checkState(!drainAndUpdateCodecDrmSession || Util.SDK_INT >= 23);

   //对比新旧解码器的重用的条件,这里代码太多,不在深入讨论。
    DecoderReuseEvaluation evaluation = canReuseCodec(codecInfo, oldFormat, newFormat);
    @DecoderDiscardReasons int overridingDiscardReasons = 0;
    switch (evaluation.result) {
      case REUSE_RESULT_NO:
//不重用,直接重启
        drainAndReinitializeCodec();
        break;
      case REUSE_RESULT_YES_WITH_FLUSH:
//可以重复使用,但需要清除队列,这种情况下需要调用MediaCodec.flush
        if (!updateCodecOperatingRate(newFormat)) {
          overridingDiscardReasons |= DISCARD_REASON_OPERATING_RATE_CHANGED;
        } else {
          codecInputFormat = newFormat;
          if (drainAndUpdateCodecDrmSession) {
            if (!drainAndUpdateCodecDrmSessionV23()) {
              overridingDiscardReasons |= DISCARD_REASON_WORKAROUND;
            }
          } else if (!drainAndFlushCodec()) {
            overridingDiscardReasons |= DISCARD_REASON_WORKAROUND;
          }
        }
        break;
      case REUSE_RESULT_YES_WITH_RECONFIGURATION:
//可以重用,但是需要重新注册信息,如果是视频,需要重新注册SPS、PPS信息,具体信息在
//com.google.android.exoplayer2.Format.Builder#initializationData 中,如果音频一般是资源类型
        if (!updateCodecOperatingRate(newFormat)) {
          overridingDiscardReasons |= DISCARD_REASON_OPERATING_RATE_CHANGED;
        } else {
          codecReconfigured = true;
          codecReconfigurationState = RECONFIGURATION_STATE_WRITE_PENDING;
          codecNeedsAdaptationWorkaroundBuffer =
              codecAdaptationWorkaroundMode == ADAPTATION_WORKAROUND_MODE_ALWAYS
                  || (codecAdaptationWorkaroundMode == ADAPTATION_WORKAROUND_MODE_SAME_RESOLUTION
                      && newFormat.width == oldFormat.width
                      && newFormat.height == oldFormat.height);
          codecInputFormat = newFormat;
          if (drainAndUpdateCodecDrmSession && !drainAndUpdateCodecDrmSessionV23()) {
            overridingDiscardReasons |= DISCARD_REASON_WORKAROUND;
          }
        }
        break;
      case REUSE_RESULT_YES_WITHOUT_RECONFIGURATION:
    //直接重用解码器,无需注册任何信息
        if (!updateCodecOperatingRate(newFormat)) {
          overridingDiscardReasons |= DISCARD_REASON_OPERATING_RATE_CHANGED;
        } else {
          codecInputFormat = newFormat;
          if (drainAndUpdateCodecDrmSession && !drainAndUpdateCodecDrmSessionV23()) {
            overridingDiscardReasons |= DISCARD_REASON_WORKAROUND;
          }
        }
        break;
      default:
        throw new IllegalStateException(); // Never happens.
    }

    if (evaluation.result != REUSE_RESULT_NO
        && (this.codec != codec || codecDrainAction == DRAIN_ACTION_REINITIALIZE)) {
      // Initial evaluation indicated reuse was possible, but codec re-initialization was triggered.
      // The reasons are indicated by overridingDiscardReasons.
      return new DecoderReuseEvaluation(
          codecInfo.name, oldFormat, newFormat, REUSE_RESULT_NO, overridingDiscardReasons);
    }

    return evaluation;
  }

3.2.6 完成切换

通过上述必要的逻辑,就能实现分片的切换,当然,每个部分代码量实在太多包括,资源加载部分也是一个核心环节,这里就不在继续分析了。

但是如何验证切换完成了,实际上是有回调的,参考下面接口实现。

代码语言:javascript复制
com.google.android.exoplayer2.audio.AudioRendererEventListener#onAudioInputFormatChanged(com.google.android.exoplayer2.Format)
com.google.android.exoplayer2.video.VideoRendererEventListener#onVideoInputFormatChanged(com.google.android.exoplayer2.Format)

四、实验

4.1 实验目的

实现手动切换分片

4.2 实验方法

自动以AdaptiveTrackSelection#Factory或者自定义BandwidthMeter,这里我们选择后者,因为改动较小。

4.2.1实现QmBandwidthMeter

代码语言:javascript复制
  private long bitrateEstimate;
  private long specificBitrate = C.TIME_UNSET;
  @Override
  public synchronized long getBitrateEstimate() {
     //如果用户没有指定比特率,则使用默认的,如果指定了,则使用用户的设备
    if(specificBitrate == C.RATE_UNSET_INT) {
       return bitrateEstimate;
    }
    return specificBitrate;
  }

4.2.2 设置到Builder中

代码语言:javascript复制
  //创建带宽测量类
  bandwidthMeter = new QmBandwidthMeter
          .Builder(getApplicationContext())
          .build();
  //缓冲控制类
  DefaultLoadControl loadControl = new DefaultLoadControl.Builder()
          .setAllocator(new DefaultAllocator(true,C.DEFAULT_BUFFER_SEGMENT_SIZE / 2))
          .setBufferDurationsMs(
              10_000, //低于20秒的缓冲就开始加载数据,不一定会buffer
              20_000, //一次加载最多加载20秒的数据
              DefaultLoadControl.DEFAULT_BUFFER_FOR_PLAYBACK_MS,
              DefaultLoadControl.DEFAULT_BUFFER_FOR_PLAYBACK_AFTER_REBUFFER_MS)
          .build();
                
      ExoPlayer.Builder playerBuilder =
          new ExoPlayer.Builder(/* context= */ this)
              .setLoadControl(loadControl)
              .setBandwidthMeter(bandwidthMeter)
              .setMediaSourceFactory(createMediaSourceFactory());
      setRenderersFactory(
          playerBuilder, intent.getBooleanExtra(IntentUtil.PREFER_EXTENSION_DECODERS_EXTRA, false));
      player = playerBuilder.build();

注意:这里设置缓冲控制类的前文已经说过, 分片选择存在坑点,由于bufferedDurationUs值过大,可能造成降码流失效, AdaptiveTrackSelection可能会重置会原来的selectionIndex为上一个分片Track,导致降码流失败,有必要做以下兼容。

4.2.3 以切换为下面的Hls分片为例子,实现从1920x1080 -> 640x360的切换

链接地址:https://devstreaming-cdn.apple.com/videos/streaming/examples/bipbop_4x3/bipbop_4x3_variant.m3u8

代码语言:javascript复制
#EXT-X-STREAM-INF:BANDWIDTH=1924009,CODECS="mp4a.40.2, avc1.4d401f",RESOLUTION=1920x1080,AUDIO="bipbop_audio",SUBTITLES="subs"gear5/prog_index.m3u8
#EXT-X-STREAM-INF:BANDWIDTH=577610,CODECS="mp4a.40.2, avc1.4d401e",RESOLUTION=640x360,AUDIO="bipbop_audio",SUBTITLES="subs"gear2/prog_index.m3u8

在起播后5s后设置带宽

【1】起播时设置带宽1924009/0.7f

【2】起播10s后设置带宽577610/0.7f

代码语言:javascript复制
//起播时
bandwidthMeter.setSpecificBitrate((long) Math.ceil(1924009/0.7f)); 
//播放一段时间后
bandwidthMeter.setSpecificBitrate((long) Math.ceil(577610/0.7f));

注意: 这里除以0.7f上面也说了,是因为带宽冗余机制造成的。

验证方法,实现下面的回调

代码语言:javascript复制
com.google.android.exoplayer2.audio.AudioRendererEventListener#onAudioInputFormatChanged(com.google.android.exoplayer2.Format)
com.google.android.exoplayer2.video.VideoRendererEventListener#onVideoInputFormatChanged(com.google.android.exoplayer2.Format)

4.3 实验结果

符合预期,成功实现了降码流

五、总结

ExoPlayer不仅支持多路流合并方式切换,也支持自适应流切换,具备高度可定制化的能力,因此,对于体验要求较高的场景,可完全通过修改自适应流相关接口实现更加顺滑的多路流切换。

ExoPlayer自适应流切换如果要改造的为用户所能选择的方式,需要修改BandwidthMeter和AdaptiveTrackSelection的一些参数。还有就是分片长度、分片IDR帧对齐、编码格式一致(尽可能解码器复用)是需要考虑的因素,另外分片可能切换等待时间较长,也是需要注意的问题。

0 人点赞