Repository navigation
feat: encode H.265 and VP9 and decode VP9 with VA-API on Linux, when turned on - #352
Merged
Merged
Conversation
devopvoid
force-pushed
the
feat/linux-vaapi-hevc-encoder
branch
from
October 11, 2026 18:42
483dc4c to
189614c
Compare
The VA-API encoder factory offers H.265 Main as well, where the driver encodes it (Intel, AMD), but only with the system property webrtc.hardware.vaapiExperimental set: it has not run on such a GPU yet. Without it, VA-API encodes H.264 alone, as before. VaapiH265Encoder works as the H.264 one does: synchronous, one reference frame, P-frames only, the driver writing the parameter sets and slice headers. It takes the block sizes and coding tools the driver says it takes (VAConfigAttribEncHEVCBlockSizes and Features), as FFmpeg's VA-API encoder does, and FFmpeg's defaults where the driver says nothing. The frame size is kept to whole coding blocks, so the coded picture needs no cropping. The encode entrypoint is found per profile, since Intel drivers often encode H.265 with the low-power one only. There is no software H.265 to fall back to: a failure is an error, upon which WebRTC switches to the next negotiated codec. The H.264 encoder is left as it is. The property is read like the VP9 one, through a shared reader. A new test sends H.265 from whichever hardware encoder there is to a Java decoder that checks every key frame is a whole access unit, so that an encoder can be tested on a GPU that does not decode H.265.
The VA-API encoder factory offers VP9 profile 0 where the driver encodes it (Intel), with the same switch as hardware VP9 encoding on Windows, the system property webrtc.hardware.vp9Encoder: it has not run on such a GPU yet. Without it, or where the driver has no VP9 encoder, libvpx encodes VP9 as before, and takes over from a hardware encoder that fails. VaapiVp9Encoder works as the H.265 one does: synchronous, one reference frame, inter frames only, the driver writing the frame header. Its parameters follow FFmpeg's VA-API VP9 encoder without B-frames: surfaces aligned to 64, key frames only where asked for, every frame refreshing all reference slots and the next one referring to slot 0 as its last frame. The output goes through EncoderOutputProcessor like that of the Windows VP9 encoder. What the H.265 and VP9 encoders share, the configuration, context and surfaces, uploading frames, the parameter buffers, rate control and reading the output, moves into VaapiEncodeSession; the H.264 encoder keeps its own copy. The VP9 hardware test runs on Linux too.
The hardware decoder factory decodes VP9 profile 0 with VA-API on the GPUs of Intel and AMD, after NVDEC, where the driver decodes it, but only with the system property webrtc.hardware.vaapiExperimental set: it has not run on such a GPU yet. VA-API leaves parsing to the application. Vp9FrameHeaderParser parses the uncompressed headers per the VP9 specification, keeping the loop filter deltas, segmentation and reference sizes that carry over, and works out the quantizer scales and loop filter levels of each segment with libvpx's quantizer lookup, which WebRTC links. Its output matched FFmpeg's trace_headers on 186 frames of libvpx streams with hidden frames, error resilience, segmentation, tiles, an odd size and lossless. VaapiVp9Decoder decodes synchronously into ten surfaces, eight reference slots and the frame being decoded, splits superframes, and copies frames out as NV12. Profiles other than 0, spatial layers, inter frames that change the size and any failure go to libvpx.
devopvoid
force-pushed
the
feat/linux-vaapi-hevc-encoder
branch
from
October 11, 2026 20:22
189614c to
8157048
Compare
SendableMetatype
added a commit
to EduGeyser/webrtc-java
that referenced
this pull request
Oct 15, 2026
Brings in the work upstream added from October 11 to 15: hardware video codecs on every platform, through Media Foundation, VideoToolbox, VA-API and NVDEC, with WebRTC built with H.265 (devopvoid#346 to devopvoid#348, devopvoid#352, devopvoid#353, devopvoid#356, devopvoid#359, devopvoid#371 to devopvoid#377); HEVC and AV1 decoding in the media player, with dav1d and Vulkan Video (devopvoid#331, devopvoid#345); encoded frame injectors (devopvoid#365); receiver source changes and audio packet infos (devopvoid#366); audio codec factories (devopvoid#340); network thread observers called on a callback thread (devopvoid#364); screen capture with ScreenCaptureKit (devopvoid#354); Windows audio devices that mix more than two channels (devopvoid#358); autoninja (devopvoid#357); and the release channels of the docs site (devopvoid#361). Fork-side adaptations: - The factory keeps the CreatePeerConnectionFactory overload without a worker thread, and still switches device capture on the network thread. devopvoid#355 copies WebRTC's factory setup and has the audio module learn a worker thread, which only matters for a worker thread distinct from the network thread, and WebRTC is removing those. The audio codec factories and hardware codec switches of devopvoid#340 and devopvoid#359 go to that overload. - WebRTC stays on a commit on main, so the move to branch 8059 (devopvoid#360) and its fixed clang package are left out. - The data channels variant builds WebRTC without H.265 and without the extra targets, leaves out the audio codec factory bindings and addTransceiverOfType, and skips the new tests that need audio or video. - dav1d is built against the sysroot on native Linux builds as well, as FFmpeg is. CI installs Meson for the full variant only, checks out the dav1d submodule, and adds its commit to the FFmpeg cache key. The fork's Windows SDK step stays, without upstream's second one. - The release workflow records the docs' version in the tagged commit, as upstream's now does. The docs site still builds from main, since this fork publishes none. - The Maven files stay deleted: Jetty is bumped in the version catalog, and the docs lockfile takes upstream's bumps on top of the fork's newer VitePress. - The media module's README builds with Gradle.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #348.
Adds VA-API codecs beyond H.264 for the GPUs of Intel and AMD on Linux. None of them has run on such a GPU yet, so each is off by default:
webrtc.hardware.vaapiExperimentalwebrtc.hardware.vp9Encoder(same switch as on Windows)webrtc.hardware.vaapiExperimentalEncoders
VaapiH265EncoderandVaapiVp9Encoder: synchronous, one reference frame, inter frames only, the driver writing the headers. Parameters follow FFmpeg's VA-API encoders. H.265 takes the block sizes and coding tools the driver reports; the encode entrypoint is found per profile, since Intel often encodes H.265 with the low-power one only.VaapiEncodeSessionholds what both share: configuration, context, surfaces, frame upload, rate control, reading the output. The H.264 encoder is left as it is.Decoder
Vp9FrameHeaderParserparses the uncompressed VP9 headers per the specification. It keeps the state that carries over between frames, and works out per-segment quantizer scales and loop filter levels with libvpx's quantizer lookup, which WebRTC already links.VaapiVp9Decoderdecodes into ten surfaces, splits superframes, and copies frames out as NV12. Other profiles, spatial layers, inter frames that change the size, and any failure go to libvpx.Testing
trace_headersfield by field on 186 frames of libvpx streams: realtime with cyclic-refresh segmentation, altref superframes with tiles, error resilience, an odd size, lossless. The "show existing frame" path has not been exercised.HardwareVideoEncoderIntegrationTestandHardwareVideoDecoderIntegrationTestpass. A new test checks that H.265 key frames from any hardware encoder are whole access units, and the VP9 hardware encoder test now also runs on Linux.