HDR stage 2 (zero-copy 10-bit) + Apple Log capture - #84
Merged
Conversation
Stage 2 of HDR: the zero-copy mappers stop refusing 10-bit hardware surfaces. VAAPI P010 imports as R16 + GR1616 dmabuf planes; D3D11VA gets a keyed-mutex shared P010 texture pair, with the texture cache keyed on the surface format so an SDR<->HDR switch mid-connection recreates it instead of copying into a mismatched texture. The draw site needs nothing — it already picks techniques and scale factors from the mapped format. macOS keeps its BGRA behaviour (VideoToolbox converts 10-bit itself, display-referred) with the limitation documented. NV12 behaviour is unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015dWUcqAkNNuNELnMPS4RdT
PROTOCOL.md promotes "log" from reserved to specified — Apple Log 1 only (BT.2020 primaries/matrix, VUI transfer unspecified since H.273 has no Apple Log code point; the JSON field is the sole label and the receiver shows the flat image for LUT grading) — and reserves "log2", which uses Apple-Gamut primaries and must never be sent as "log". The STATE snapshot's colour fields (hdr/color + HEVC-only codecs), missed in the stage-1 spec update, are now documented. README gains the Apple Log feature bullet; PERFORMANCE.md covers the x422 capture quirk and the 10-bit zero-copy path with a placeholder for a real bench pair. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015dWUcqAkNNuNELnMPS4RdT
The decoder's BT.2020 branch becomes an explicit transfer switch: HLG and PQ map as before, and anything else — Apple Log's VUI says transfer unspecified, since H.273 has no code point for it — keeps the BT.2020 matrix and 10-bit ranges but renders as ordinary SDR, which is precisely the flat log image the user's Apply LUT filter expects. (This already happened by accident; now it's explicit, commented, and harness-proven byte-identical for SDR/HLG.) The GPU-pipeline draw site stops keying techniques on pixel format alone — 10-bit + HLG/PQ takes the HLG chain and extended canvas as before, while 10-bit with any other transfer takes new Log techniques: same BT.2020 limited-range matrix, no EOTF, no primaries conversion, no brightness scaling, clipped to the SDR canvas. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015dWUcqAkNNuNELnMPS4RdT
StreamColor gains .log (Apple Log 1 only; Apple Log 2 with Apple-Gamut primaries is reserved for a future "log2"). Log capture requires iOS 17 and a lens with an x422 format advertising .appleLog — gated at runtime per format, never by model — and delivers x422 straight into the Main10 HEVC encode (VideoToolbox converts internally), tagged BT.2020 primaries/matrix with the transfer deliberately unset so the VUI reads unspecified, the only correct value for Apple Log. The Options toggle becomes a Standard / HDR (HLG) / Apple Log picker (Log row only where capturable), persisted under "streamColor" with a one-time migration from "hdrEnabled". The issue-#81 attach-first pixel format guard covers all 10-bit formats; degrade is always to SDR, never sideways between HLG and Log (a Log stream that silently turned HLG would grade wrong). STATE announces "color":"log" (HLG keeps its "hdr" flag byte-for-byte); SDR behaviour is unchanged throughout. Release-Bump: minor Release-Beta: true Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015dWUcqAkNNuNELnMPS4RdT
MyNamesEMurray
enabled auto-merge
July 26, 2026 00:12
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
2 participants
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.
What & why
Two features, both built on the stage-1 groundwork:
Stage 2 — zero-copy 10-bit. The GPU pipeline's zero-copy mappers stop refusing 10-bit hardware surfaces: VAAPI P010 imports as R16 + GR1616 dmabuf planes (Linux), D3D11VA gets a keyed-mutex shared P010 texture with the cache keyed on surface format so an SDR↔HDR switch mid-connection recreates it (Windows). The draw site needed nothing — it already keys techniques and scale factors off the mapped-format enum stage 1 introduced. macOS keeps its BGRA path (VideoToolbox converts 10-bit itself), documented as a known limitation. HDR + GPU pipeline should now log
zero-copy active (D3D11 shared P010)instead of falling back.Apple Log (roadmap P3, unlocked by 10-bit HEVC). The app captures
.appleLog(iOS 17+, runtime-gated per format — Log formats are x422, not HLG's x420) and encodes HEVC Main10 tagged BT.2020 primaries/matrix with the transfer deliberately unset: H.273 has no Apple Log code point, so VUI transfer=unspecified is the only correct value, and the wire'sVIDEO_CONFIG {"color":"log"}(spec'd since stage 1) is the label. The plugin shows the untouched flat log image — correct BT.2020-matrix decode, then hands off — for the LUT-in-OBS grading workflow; the async path's accidental-but-correct behaviour is now explicit, and the GPU pipeline gains dedicated Log shader techniques (it would previously have HLG-linearized log frames). The Options HDR toggle becomes a Standard / HDR (HLG) / Apple Log picker with one-time migration from the old key; degrade is always to SDR, never sideways between HLG and Log. Apple Log 2 (Apple-Gamut primaries) is explicitly reserved as a future"log2"— never conflated with"log".Docs: PROTOCOL.md (
"log"specified,"log2"reserved, STATE colour fields documented), README, PERFORMANCE.md.How it was tested
h264-decoder.c, three generated samples): SDR and HLG outputs byte-identical before/after; the new log-like sample (ffprobe-verified VUI: bt2020/unspecified/bt2020nc) decodes toframes=18 format=I010 trc=default— the flat-image path.-Wall -Wextra -Werror(Linux, compiles the VAAPI arm for real); the Windows arm is header-verified (gs_texture_create_p010declared in the_WIN32block of graphics.h with the same shape as the NV12 creator) but needs a real-hardware run.syntax-check.sh; full type check in CI's macOS job.tools/bench-report.py) since this container has no GPU to quote numbers from, per PERFORMANCE.md; (2) Apple Log on an iPhone 15 Pro+ — the key unknown is x422 appearing inavailableVideoPixelFormatTypes(if not, Log degrades safely to SDR rather than crashing — the [TestFlight] Crash on iPhone16_1, iOS 27.0 (build 1047) #81 guard covers it); (3) a LUT sanity pass in OBS on the log image.🤖 Generated with Claude Code
https://claude.ai/code/session_015dWUcqAkNNuNELnMPS4RdT
Generated by Claude Code