Skip to content

Latest commit

 

History

History
167 lines (118 loc) · 7.59 KB

MediaPipelineDesign.md

File metadata and controls

167 lines (118 loc) · 7.59 KB

Media Pipeline

Overview

LibMedia turns demuxed media data into decoded audio and video, then feeds it to sinks that are driven by playback time. The pipeline is pull-based: downstream nodes ask upstream nodes whether data is available, then pull data when the reported status allows it.

The current pipeline is made of three categories of nodes:

  • Producers materialize raw audio/video data.
  • Sinks consume that raw audio/video data.
  • Processors are both producers and sinks, pulling data from an input producer to transform it before a downstream sink makes use of it.

Audio and video use separate producer interfaces because the data they carry is different, but their control flow is intentionally similar.

Producer Interface

Producers expose three core operations:

  • peek() reports the current status and, when data is available, exposes the head of the producer's output without removing it.
  • consume() removes the head that a prior peek() exposed.
  • set_wake_handler(...) is used by the downstream node to wake itself when a previously-observed status would result in it going to sleep.

Downstream nodes peek() when they need new data, then either read the head and consume() it or go to sleep. Peeking without consuming lets a node decide whether to take the head, rather than being forced to accept data it may have no room for.

PipelineStatus

PipelineStatus is the shared status language between nodes:

  • Pending: Data is not available yet, but may become available later.
  • HaveData: A peek exposes data.
  • Blocked: Data is not available yet, i.e. we have run out of buffered data from over the network.
  • Suspended: The upstream node has released its decoding resources and will not produce data until it receives a seek.
  • EndOfStream: We have reached the end of the input data.
  • Error: The upstream node encountered an error and cannot continue without a seek.

HaveData carries data. All other statuses require handling by downstream code.

Pending, Blocked, Suspended, EndOfStream and Error are waiting statuses. A downstream node that normally polls for upstream input seeing these should sleep until its wake handler fires.

EndOfStream and Error are terminal statuses. They resolve seeks even though they are still "waiting" statuses for helper APIs, because no more data is required to decide the seek has completed.

Wake Handlers

Wake handlers do not carry state. A wake means the downstream node should peek() again.

Nodes should dispatch a wake when the downstream node may be waiting, and the status has changed to one that resolves a seek: HaveData, EndOfStream, or Error.

Some nodes (i.e. AudioMixer) may also wake when moving away from a previously dispatched EndOfStream, so stale EOS state can be recovered after a seek or after new input becomes available.

Seeking

Seeking is initiated downstream and forwarded upstream. For downstream nodes, a seek completes when the upstream node reports a status that resolves the new position. Usually that means HaveData, but EndOfStream or Error can also complete the wait.

Sinks report pipeline status changes to their users. A user such as the playback manager can interpret a seek-resolving status as seek completion.

If a node forwards a seek upstream, it can clear or invalidate its cached data, since the upstream node is required to emit data at or before the target timestamp.

However, if a node uses its cached data to resolve seeks, it must also resolve seeks past the last frame in that cached data if its last known status is EndOfStream. This ensures that a seek past EOS will not invalidate in-flight frames by bumping a seek generation.

Decoded Data Producers

Decoded data producers run demux and decode video/audio data into queues in order to absorb variance in processing for each individual coded frame.

When a downstream consume() removes an item from the queue, the producer wakes its decoder thread so it can refill the queue. When the queue is full and no downstream activity occurs for the idle timeout, the producer auto-suspends: it drops its decoder, its queue, and any frame pool buffers it can free, then wakes the consumer with a Suspended status.

peek() and consume() defer the idle timeout, but they never resume a suspended producer. When a downstream node needs new data, it must explicitly seek() the suspended node to resume it, so that the pipeline may insert pre-roll data at any stage where it is needed (i.e. for audio stretching).

Playback Manager

PlaybackManager creates, modifies and drives the end-to-end playback pipeline. When media data is received, it will spawn a thread to determine what tracks are present, and connect the necessary nodes for playback to begin.

Behavior is defined through state handlers derived from PlaybackStateHandler, with each overriding methods for various functions and signals to determine when to transition to another state. Upon entering or exiting, a state handler will drive changes to the pipeline, like pausing/resuming or seeking.

The media time is derived from the audio sink when audio is used, otherwise GenericTimeProvider calculates time from the monotonic clock.

Audio Pipeline

The usual audio path is:

DecodedAudioProducer
    -> AudioMixer
    -> AudioTimeStretchProcessor
    -> AudioPlaybackSink
    -> PlaybackStream

AudioPlaybackSink owns a small queue of AudioBlocks pulled from upstream, which it fills on a separate thread to avoid blocking the PlaybackStream data callback. It starts pulling input when its start() method is called, so that all inputs can be connected first. When its input reaches EndOfStream, it continues producing silence to allow the time to advance in case there is video data still available.

AudioMixer sequences/mixes audio from multiple inputs into a single output. It emits silence for gaps, and pauses output when any input indicates data is not ready yet. When all inputs are at EndOfStream, the output becomes EndOfStream as well.

AudioTimeStretchProcessor stretches audio without affecting pitch.

Video Pipeline

The usual video path is:

DecodedVideoProducer -> DisplayingVideoSink

DisplayingVideoSink keeps the current frame and the next frame. It advances when playback time passes the current frame's display interval.

Video seeks are comparatively expensive, so the sink can satisfy some seeks locally when the requested timestamp is already covered by the current or next frame interval. This avoids unnecessary upstream seeks during scrubbing. The same optimization is less important for audio because it decodes exponentially faster.

Threading

The media pipeline crosses several threads:

  • The main thread coordinates playback state and serializes cross-thread communication through its event loop
  • Decoded data producers manage decoding threads to amortize the cost to decode each frame
  • The audio sink runs a processing thread to amortize the cost of mixing/stretching each audio block
  • The playback stream data callback is invoked from an (often system-managed) thread

Data should only be consumed via peek() and consume() from a single thread, to allow producers to use lock-free structures internally. Control methods like seek() should similarly only initiate from a single thread, or be otherwise serialized.

Avoid synchronous callback dispatch while holding locks unless the callee is known not to reenter the same pipeline. Deferred dispatch is often used to keep lock ordering simple.

Wake handlers must be removed before disconnecting or destroying a node. An upstream node must never invoke a handler after downstream has cleared it.