WebRTC Encoded Transform

1. Introduction

The [WEBRTC-NV-USE-CASES] document describes several functions that can only be achieved by access to media (requirements N20-N22), including, but not limited to:

• Funny Hats

• Machine Learning

• Virtual Reality Gaming

These use cases further require that processing can be done in worker threads (requirement N23-N24).

Furthermore, the "trusted JavaScript cloud conferencing" use case requires such processing to be done on encoded media, not just the raw media.

This specification gives an interface inspired by [WEB-CODECS] to provide access to such functionality while retaining the setup flow of RTCPeerConnection.

This iteration of the specification provides access to encoded media, which is the output of the encoder part of a codec and the input to the decoder part of a codec.

2. Terminology

3. Specification

The Streams definition doesn’t use WebIDL much, but the WebRTC spec does. This specification shows the IDL extensions for WebRTC.

It uses an additional API on RTCRtpSender and RTCRtpReceiver to insert the processing into the pipeline.

// New dictionary
dictionary RTCInsertableStreams {
    ReadableStream readable;
    WritableStream writable;
};
typedef (SFrameTransform or RTCRtpScriptTransform) RTCRtpTransform;
// New methods for RTCRtpSender and RTCRtpReceiver
partial interface RTCRtpSender {
    attribute RTCRtpTransform? transform;
};
partial interface RTCRtpReceiver {
    attribute RTCRtpTransform? transform;
};

3.1. Extension operation

At the time when a codec is initialized as part of the encoder, and the corresponding flag is set in the RTCPeerConnection's RTCConfiguration argument, ensure that the codec is disabled and produces no output.

3.1.1. Stream creation

At construction of each RTCRtpSender or RTCRtpReceiver, run the following steps:

2. Initialize this.[[transform]] to null.

3. Initialize this.[[readable]] to a new ReadableStream.

4. Set up this.[[readable]]. this.[[readable]] is provided frames using the readEncodedData algorithm given this as parameter.

5. Set this.[[readable]].[[owner]] to this.

6. Initialize this.[[writable]] to a new WritableStream.

7. Set up this.[[writable]] with its writeAlgorithm set to writeEncodedData given this as parameter and its sizeAlgorithm to an algorithm that returns 0.

   Chunk size is set to 0 to explictly disable streams backpressure on the write side.

8. Set this.[[writable]].[[owner]] to this.

9. Initialize this.[[pipeToController]] to null.

10. Initialize this.[[lastReceivedFrameCounter]] to 0.

11. Initialize this.[[lastEnqueuedFrameCounter]] to 0.

12. Queue a task to run the following steps:

    1. If this.[[pipeToController]] is not null, abort these steps.

    2. Set this.[[pipeToController]] to a new AbortController.

    3. Call pipeTo with this.[[readable]], this.[[writable]], preventClose equal to true, preventAbort equal to true, preventCancel equal to true and this.[[pipeToController]].signal.

3.1.2. Stream processing

The readEncodedData algorithm is given a rtcObject as parameter. It is defined by running the following steps:

1. Wait for a frame to be produced by rtcObject’s encoder if it is a RTCRtpSender or rtcObject’s packetizer if it is a RTCRtpReceiver.

2. Increment rtcObject.[[lastEnqueuedFrameCounter]] by 1.

3. Let frame be the newly produced frame.

4. Set frame.[[owner]] to rtcObject.

5. Set frame.[[counter]] to rtcObject.[[lastEnqueuedFrameCounter]].

6. Enqueue frame in rtcObject.[[readable]].

The writeEncodedData algorithm is given a rtcObject as parameter and a frame as input. It is defined by running the following steps:

1. If frame.[[owner]] is not equal to rtcObject, abort these steps and return a promise resolved with undefined. A processor cannot create frames, or move frames between streams.

2. If frame.[[counter]] is equal or smaller than rtcObject.[[lastReceivedFrameCounter]], abort these steps and return a promise resolved with undefined. A processor cannot reorder frames, although it may delay them or drop them.

3. Set rtcObject.[[lastReceivedFrameCounter]] to frame[[counter]].

4. Enqueue the frame for processing as if it came directly from the encoded data source, by running one of the following steps:

   • If rtcObject is a RTCRtpSender, enqueue it to rtcObject’s packetizer, to be processed in parallel.

   • If rtcObject is a RTCRtpReceiver, enqueue it to rtcObject’s decoder, to be processed in parallel.

5. Return a promise resolved with undefined.

On sender side, as part of readEncodedData, frames produced by rtcObject’s encoder MUST be enqueued in rtcObject.[[readable]] in the encoder’s output order. As writeEncodedData ensures that the transform cannot reorder frames, the encoder’s output order is also the order followed by packetizers to generate RTP packets and assign RTP packet sequence numbers.

On receiver side, as part of readEncodedData, frames produced by rtcObject’s packetizer MUST be enqueued in rtcObject.[[readable]] in the same encoder’s output order. To ensure the order is respected, the depacketizer will typically use RTP packet sequence numbers to reorder RTP packets as needed before enqueuing frames in rtcObject.[[readable]]. As writeEncodedData ensures that the transform cannot reorder frames, this will be the order expected by rtcObject’s decoder.

3.2. Extension attribute

A RTCRtpTransform has two private slots called [[readable]] and [[writable]].

Each RTCRtpTransform has an association steps set, which is empty by default.

The transform getter steps are:

1. Return this.[[transform]].

The transform setter steps are:

2. Let transform be the argument to the setter.

3. Let checkedTransform set to transform if it is not null or to an identity transform stream otherwise.

4. Let reader be the result of getting a reader for checkedTransform.[[readable]].

5. Let writer be the result of getting a writer for checkedTransform.[[writable]].

6. Initialize newPipeToController to a new AbortController.

7. If this.[[pipeToController]] is not null, run the following steps:

   1. Add the chain transform algorithm to this.[[pipeToController]].signal.

   2. signal abort this.[[pipeToController]].signal.

8. Else, run the chain transform algorithm steps.

9. Set this.[[pipeToController]] to newPipeToController.

10. Set this.[[transform]] to transform.

11. Run the steps in the set of association steps of transform with this.

The chain transform algorithm steps are defined as:

1. If newPipeToController is aborted, abort these steps.

2. Release reader.

3. Release writer.

4. Assert that newPipeToController is the same object as rtcObject.[[pipeToController]].

5. Call pipeTo with rtcObject.[[readable]], checkedTransform.[[writable]], preventClose equal to false, preventAbort equal to false, preventCancel equal to true and newPipeToController.signal.

6. Call pipeTo with checkedTransform.[[readable]], rtcObject.[[writable]], preventClose equal to true, preventAbort equal to true, preventCancel equal to false and newPipeToController.signal.

This algorithm is defined so that transforms can be updated dynamically. There is no guarantee on which frame will happen the switch from the previous transform to the new transform.

If a web application sets the transform synchronously at creation of the RTCRtpSender (for instance when calling addTrack), the transform will receive the first frame generated by the RTCRtpSender's encoder. Similarly, if a web application sets the transform synchronously at creation of the RTCRtpReceiver (for instance when calling addTrack, or at track event handler), the transform will receive the first full frame generated by the RTCRtpReceiver's packetizer.

4. SFrameTransform

The API presented in this section represents a preliminary proposal based on protocol proposals that have not yet been adopted by an IETF WG. As a result, both the API and underlying protocol are likely to change significantly going forward.

enum SFrameTransformRole {
    "encrypt",
    "decrypt"
};
dictionary SFrameTransformOptions {
    SFrameTransformRole role = "encrypt";
};
typedef [EnforceRange] unsigned long long SmallCryptoKeyID;
typedef (SmallCryptoKeyID or bigint) CryptoKeyID;
[Exposed=(Window,DedicatedWorker)]
interface SFrameTransform {
    constructor(optional SFrameTransformOptions options = {});
    Promise<undefined> setEncryptionKey(CryptoKey key, optional CryptoKeyID keyID);
    attribute EventHandler onerror;
};
SFrameTransform includes GenericTransformStream;
enum SFrameTransformErrorEventType {
    "authentication",
    "keyID",
    "syntax"
};
[Exposed=(Window,DedicatedWorker)]
interface SFrameTransformErrorEvent : Event {
    constructor(DOMString type, SFrameTransformErrorEventInit eventInitDict);
    readonly attribute SFrameTransformErrorEventType errorType;
    readonly attribute CryptoKeyID? keyID;
    readonly attribute any frame;
};
dictionary SFrameTransformErrorEventInit : EventInit {
    required SFrameTransformErrorEventType errorType;
    required any frame;
    CryptoKeyID? keyID;
};

The new SFrameTransform(options) constructor steps are:

1. Let transformAlgorithm be an algorithm which takes a frame as input and runs the SFrame transform algorithm with this and frame.

2. Set this.[[transform]] to a new TransformStream.

3. Set up this.[[transform]] with transformAlgorithm set to transformAlgorithm.

4. Let options be the method’s first argument.

5. Set this.[[role]] to options["role"].

6. Set this.[[readable]] to this.[[transform]].[[readable]].

7. Set this.[[writable]] to this.[[transform]].[[writable]].

4.1. Algorithm

The SFrame transform algorithm, given sframe as a SFrameTransform object and frame, runs these steps:

1. Let role be sframe.[[role]].

2. If frame.[[owner]] is a RTCRtpSender, set role to 'encrypt'.

3. If frame.[[owner]] is a RTCRtpReceiver, set role to 'decrypt'.

4. Let data be undefined.

5. If frame is a BufferSource, set data to frame.

6. If frame is a RTCEncodedAudioFrame, set data to frame.data

7. If frame is a RTCEncodedVideoFrame, set data to frame.data

8. If data is undefined, abort these steps.

9. Let buffer be the result of running the SFrame algorithm with data and role as parameters. This algorithm is defined by the SFrame specification and returns an ArrayBuffer.

10. If the SFrame algorithm exits abruptly with an error, queue a task to run the following sub steps:

    1. If the processing fails on decryption side due to data not following the SFrame format, fire an event named error at sframe, using the SFrameTransformErrorEvent interface with its errorType attribute set to syntax and its frame attribute set to frame.

    2. If the processing fails on decryption side due to the key identifier parsed in data being unknown, fire an event named error at sframe, using the SFrameTransformErrorEvent interface with its errorType attribute set to keyID, its frame attribute set to frame and its keyID attribute set to the keyID value parsed in the SFrame header.

    3. If the processing fails on decryption side due to validation of the authentication tag, fire an event named error at sframe, using the SFrameTransformErrorEvent interface with its errorType attribute set to authentication and its frame attribute set to frame.

    4. Abort these steps.

11. If frame is a BufferSource, set frame to buffer.

12. If frame is a RTCEncodedAudioFrame, set frame.data to buffer.

13. If frame is a RTCEncodedVideoFrame, set frame.data to buffer.

14. Enqueue frame in sframe.[[transform]].

4.2. Methods

1. Let promise be a new promise.

2. If keyID is a bigint which cannot be represented as a integer between 0 and 2⁶⁴-1 inclusive, reject promise with a RangeError exception.

3. Otherwise, in parallel, run the following steps:

   1. Set key with its optional keyID as key material to use for the SFrame transform algorithm, as defined by the SFrame specification.

   2. If setting the key material fails, reject promise with an InvalidModificationError exception and abort these steps.

   3. Resolve promise with undefined.

4. Return promise.

5. RTCRtpScriptTransform

// New enum for video frame types. Will eventually re-use the equivalent defined
// by WebCodecs.
enum RTCEncodedVideoFrameType {
    "empty",
    "key",
    "delta",
};
dictionary RTCEncodedVideoFrameMetadata {
    long long frameId;
    sequence<long long> dependencies;
    unsigned short width;
    unsigned short height;
    long spatialIndex;
    long temporalIndex;
    long synchronizationSource;
    octet payloadType;
    sequence<long> contributingSources;
};
// New interfaces to define encoded video and audio frames. Will eventually
// re-use or extend the equivalent defined in WebCodecs.
[Exposed=(Window,DedicatedWorker)]
interface RTCEncodedVideoFrame {
    readonly attribute RTCEncodedVideoFrameType type;
    readonly attribute unsigned long timestamp; // RTP timestamp.
    attribute ArrayBuffer data;
    RTCEncodedVideoFrameMetadata getMetadata();
};
dictionary RTCEncodedAudioFrameMetadata {
    long synchronizationSource;
    octet payloadType;
    sequence<long> contributingSources;
};
[Exposed=(Window,DedicatedWorker)]
interface RTCEncodedAudioFrame {
    readonly attribute unsigned long timestamp; // RTP timestamp.
    attribute ArrayBuffer data;
    RTCEncodedAudioFrameMetadata getMetadata();
};
// New interfaces to expose JavaScript-based transforms.
[Exposed=DedicatedWorker]
interface RTCTransformEvent : Event {
    readonly attribute RTCRtpScriptTransformer transformer;
};
partial interface DedicatedWorkerGlobalScope {
    attribute EventHandler onrtctransform;
};
[Exposed=DedicatedWorker]
interface RTCRtpScriptTransformer {
    readonly attribute ReadableStream readable;
    readonly attribute WritableStream writable;
    readonly attribute any options;
    Promise<unsigned long long> generateKeyFrame(optional DOMString rid);
    Promise<undefined> sendKeyFrameRequest();
};
[Exposed=Window]
interface RTCRtpScriptTransform {
    constructor(Worker worker, optional any options, optional sequence<object> transfer);
};

5.1. Operations

The new RTCRtpScriptTransform(worker, options, transfer) constructor steps are:

1. Set t1 to an identity transform stream.

2. Set t2 to an identity transform stream.

3. Set this.[[writable]] to t1.[[writable]].

4. Set this.[[readable]] to t2.[[readable]].

5. Let serializedOptions be the result of StructuredSerializeWithTransfer(options, transfer).

6. Let serializedReadable be the result of StructuredSerializeWithTransfer(t1.[[readable]], « t1.[[readable]] »).

7. Let serializedWritable be the result of StructuredSerializeWithTransfer(t2.[[writable]], « t2.[[writable]] »).

8. Queue a task on the DOM manipulation task source worker’s global scope to run the following steps:

   1. Let transformerOptions be the result of StructuredDeserialize(serializedOptions, the current Realm).

   2. Let readable be the result of StructuredDeserialize(serializedReadable, the current Realm).

   3. Let writable be the result of StructuredDeserialize(serializedWritable, the current Realm).

   4. Let transformer be a new RTCRtpScriptTransformer.

   5. Set transformer.[[options]] to transformerOptions.

   6. Set transformer.[[readable]] to readable.

   7. Set transformer.[[writable]] to writable.

   8. Let event be the result of creating an event with RTCTransformEvent.

   9. Set event.type attribute to "rtctransform".

   10. Set event.transformer to transformer.

   11. Dispatch event on worker’s global scope.

// FIXME: Describe error handling (worker closing flag true at RTCRtpScriptTransform creation time. And worker being terminated while transform is processing data).

Each RTCRtpScriptTransform has the following set of association steps, given rtcObject:

1. Let transform be the RTCRtpScriptTransform object that owns the association steps.

2. Let encoder be rtcObject’s encoder if rtcObject is a RTCRtpSender or undefined otherwise.

3. Let depacketizer be rtcObject’s depacketizer if rtcObject is a RTCRtpReceiver or undefined otherwise.

4. Queue a task on the DOM manipulation task source worker’s global scope to run the following steps:

   1. Let transformer be the RTCRtpScriptTransformer object associated to transform.

   2. Set transformer.[[encoder]] to encoder.

   3. Set transformer.[[depacketizer]] to depacketizer.

The generateKeyFrame(rid) method steps are:

1. Let promise be a new promise.

2. Run the generate key frame algorithm with promise, this.[[encoder]] and rid.

3. Return promise.

The sendKeyFrameRequest() method steps are:

1. Let promise be a new promise.

2. Run the send request key frame algorithm with promise and this.[[depacketizer]].

3. Return promise.

5.2. Attributes

A RTCRtpScriptTransformer has the following private slots called [[depacketizer]], [[encoder]], [[options]], [[readable]] and [[writable]]. In addition, a RTCRtpScriptTransformer is always associated with its parent RTCRtpScriptTransform transform. This allows algorithms to go from an RTCRtpScriptTransformer object to its RTCRtpScriptTransform parent and vice versa.

The options getter steps are:

1. Return this.[[options]].

The readable getter steps are:

1. Return this.[[readable]].

The writable getter steps are:

1. Return this.[[writable]].

5.3. KeyFrame Algorithms

The generate key frame algorithm, given promise, encoder and rid, is defined by running these steps:

1. If encoder is undefined, reject promise with InvalidStateError, abort these steps.

2. If encoder is not processing video frames, reject promise with InvalidStateError, abort these steps.

3. If rid is defined, validate its value. If invalid, reject promise with NotAllowedError and abort these steps.

4. In parallel, run the following steps:

   1. Gather a list of video encoders, named videoEncoders from encoder, ordered according negotiated RIDs if any.

   2. If rid is defined, remove from videoEncoders any video encoder that does not match rid.

   3. If rid is undefined, remove from videoEncoders all video encoders except the first one.

   4. If videoEncoders is empty, reject promise with NotFoundError and abort these steps. videoEncoders is expected to be empty if the corresponding RTCRtpSender is not active, or the corresponding RTCRtpSender track is ended.

   5. Let videoEncoder be the first encoder in videoEncoders.

   6. If rid is undefined, set rid to the RID value corresponding to videoEncoder.

   7. Create a pending key frame task called task with task.[[rid]] set to rid and task.[[promise]]| set to promise.

   8. If encoder.[[pendingKeyFrameTasks]] is undefined, initialize encoder.[[pendingKeyFrameTasks]] to an empty set.

   9. Let shouldTriggerKeyFrame be true if encoder.[[pendingKeyFrameTasks]] contains a task whose [[rid]] value is equal to rid, and false otherwise.

   10. Add task to encoder.[[pendingKeyFrameTasks]].

   11. If shouldTriggerKeyFrame is true, instruct videoEncoder to generate a key frame for the next provided video frame.

For any RTCRtpScriptTransformer named transformer, the following steps are run just before any frame is enqueued in transformer.[[readable]]:

1. Let encoder be transformer.[[encoder]].

2. If encoder or encoder.[[pendingKeyFrameTasks]] is undefined, abort these steps.

3. If frame is not a video key frame, abort these steps.

4. For each task in encoder.[[pendingKeyFrameTasks]], run the following steps:

   1. If frame was generated by a video encoder identified by task.[[rid]], run the following steps:

      1. Remove task from encoder.[[pendingKeyFrameTasks]].

      2. Resolve task.[[promise]] with frame’s timestamp.

By resolving the promises just before enqueuing the corresponding key frame in a RTCRtpScriptTransformer's readable, the resolution callbacks of the promises are always executed just before the corresponding key frame is exposed. If the promise is associated to several rid values, it will be resolved when the first key frame corresponding to one the rid value is enqueued.

The send request key frame algorithm, given promise and depacketizer, is defined by running these steps:

1. If depacketizer is undefined, reject promise with InvalidStateError, abort these steps.

2. If depacketizer is not processing video packets, reject promise with InvalidStateError, abort these steps.

3. In parallel, run the following steps:

   1. If sending a Full Intra Request (FIR) by depacketizer’s receiver is not deemed appropriate, resolve promise with undefined and abort these steps. Section 4.3.1 of [RFC5104] provides guidelines of how and when it is appropriate to sending a Full Intra Request.

   2. Generate a Full Intra Request (FIR) packet as defined in section 4.3.1 of [RFC5104] and send it through depacketizer’s receiver.

   3. Resolve promise with undefined.

6. RTCRtpSender extension

An additional API on RTCRtpSender is added to complement the generation of key frame added to RTCRtpScriptTransformer.

partial interface RTCRtpSender {
    Promise<undefined> generateKeyFrame(optional sequence <DOMString> rids);
};

6.1. Extension operation

The generateKeyFrame(rids) method steps are:

1. Let promise be a new promise.

2. In parallel, run the generate key frame algorithm with promise, this’s encoder and rids.

3. Return promise.

7. Privacy and security considerations

This API gives Javascript access to the content of media streams. This is also available from other sources, such as Canvas and WebAudio.

However, streams that are isolated (as specified in [WEBRTC-IDENTITY]) or tainted with another origin, cannot be accessed using this API, since that would break the isolation rule.

The API will allow access to some aspects of timing information that are otherwise unavailable, which allows some fingerprinting surface.

The API will give access to encoded media, which means that the JS application will have full control over what’s delivered to internal components like the packetizer or the decoder. This may require additional care with auditing how data is handled inside these components.

For instance, packetizers may expect to see data only from trusted encoders, and may not be audited for reception of data from untrusted sources.

8. Examples

See the explainer document.