Repository navigation
#1795: Local2D layers honour their declared format (all backends + service zones) - #1809
Merged
Merged
Conversation
…ed (#1795) The over-flatten and the 2D-under backdrop flatten sampled every Local2D layer through the non-decoding SRV into a UNORM scratch, so the declared format was ignored (a true-linear UNORM layer rendered too dark). Both scratches are now TYPELESS with a UNORM SRV (every reader, and the bridge plane, still sees ENCODED bytes) and an _SRGB RTV for the flatten. The flatten samples through the view matching its target: format-honest into an _SRGB RTV, non-decoding under the legacy hatch. Measured (cube_handle_d3d11_win, DXR_LOCAL2D_PANEL=1, opaque checker authored 40/235): _SRGB panel 40/235; UNORM panel 110/246 (one encode); UNORM under DXR_COLOR_LEGACY_UNORM_ENCODED=1 40/235. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…l UNORM A/B knob for the Local2D colour contract (#1795), the same switch displayxr-common uses for its swapchains. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…1795) The over-flatten and the 2D-under backdrop flatten sampled every Local2D layer through the non-decoding bind (GL_SKIP_DECODE_EXT) into a GL_RGBA8 scratch, so the declared format was ignored (a true-linear UNORM layer rendered too dark). Both flattens now draw into a private GL_SRGB8_ALPHA8 target with GL_FRAMEBUFFER_SRGB on for the flatten pass only (previous enable restored), sampling each layer format-honestly (gl_bind_layer_source compose=true: an _SRGB source decodes, GL_RGBA8 is read as linear), and publish into the unchanged GL_RGBA8 scratches with glCopyImageSubData -- the compose target's #1610 pattern. Every reader (post-weave masked composite, #885 diag dump, DP set_background_2d) still sees ENCODED bytes. Legacy hatch, no glCopyImageSubData, or an incomplete sRGB attachment keep today's direct non-decoding flatten. A test pins the new publish as a raw copy, never a draw. Measured (cube_handle_gl_win, DXR_LOCAL2D_PANEL=1, opaque checker authored 40/235): (hardware A/B pending). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ed (#1795) The over-flatten and the 2D-under backdrop flatten sampled every Local2D layer through the non-decoding view into a UNORM scratch, so the declared format was ignored (a true-linear UNORM layer rendered too dark). Both scratches are now TYPELESS with an _SRGB RTV for the flatten; every reader still sees ENCODED bytes through the UNORM member (the masked composite resolves the typeless SRV to UNORM, the DP already views the backdrop as R8G8B8A8_UNORM, and the bridge / deposit planes stay UNORM chains fed by a same-family copy, #1663). The flatten samples through the view matching its target: format-honest into an _SRGB RTV (with _SRGB PSO twins, since D3D12 bakes the RTV format), non-decoding under the legacy hatch. Measured (cube_handle_d3d12_win, DXR_LOCAL2D_PANEL=1): (hardware A/B pending) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…clared (#1795) The over-flatten and the 2D-under backdrop flatten sampled every Local2D layer through the non-decoding view into a UNORM scratch, so the declared format was ignored (a true-linear UNORM layer rendered too dark). Both scratches are now B8G8R8A8_UNORM + MUTABLE_FORMAT with a {UNORM,_SRGB} format list (the #1799 compose-target shape): the UNORM view is what every reader samples, so they still see ENCODED bytes, and the flatten renders through an _SRGB view with a second flatten render pass + pipeline pair (vk_local2d_composite_init_srgb_flatten, opt-in, so the Linux/macOS/Android multi-compositor callers are unchanged). The source view follows the target: format-honest (true view) into an _SRGB attachment, non-decoding into a UNORM one. The legacy hatch never builds the _SRGB pipelines and keeps the old path exactly; so does the weave-on-scanout split, whose bridge planes are typed-UNORM D3D11 textures. Measured (hardware A/B pending): cube_handle_vk_win, DXR_LOCAL2D_PANEL=1, opaque checker authored 40/235 -- expect _SRGB panel 40/235; UNORM panel 110/246 (one encode); UNORM under DXR_COLOR_LEGACY_UNORM_ENCODED=1 40/235. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…SRGB (#1795) A pure-zones / Local2D commit never reached the compose decision (it lives on the projection write path), so every one was a passthrough: a UNORM zone or Local2D source was sampled raw into the encoded atlas and its declared format ignored (a true-linear UNORM panel came out too dark). service_composite_zones_frame() now decides for a commit with no projection layer: when some sampled source is not _SRGB and the legacy hatch is off, it opens the same private _SRGB-view compose target as a composing projection commit (client_ensure_compose_target, the same !compose_was_active || !compose_mirrors_atlas seed, under the immediate_ctx_mutex the pass already holds), so zones_resolve_src_srv() takes its format-honest branch, the blends run in linear, and the existing end-of-commit whole-atlas publish copies the result. An all-_SRGB commit stays the byte-identical passthrough (no compose target, no extra copy). compose_was_active is now recorded on every painting pure-zones commit, so a passthrough one forces the next re-seed, and a composing Local2D-only commit stamps atlas_holds_srgb_bytes (encoded). ADR-044 §7. Mixed projection + zones / Local2D commits already composed (has_local_2d counts as an extra layer); unchanged. Measured: (hardware A/B pending) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…Local2D strip UNORM Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…1795) Windows VK always runs the weave-on-scanout split on the hybrid box (ADR-039 same-adapter engage), where the over-flatten and the 2D-under backdrop render straight into the bridge planes. Those are typed-UNORM D3D11 textures with no _SRGB view, so 8b00d0283 left that path byte-passthrough and a UNORM panel still read 40/235. The split now flattens into the private MUTABLE scratch through its _SRGB view, exactly as the non-split path does, and vkCmdCopyImage's the region into the cleared plane: a raw texel copy between two B8G8R8A8 images (never a blit), so the plane carries the same ENCODED bytes. The copy is recorded in the plane pass's own command buffer, so it sits inside the existing plane-write window (timeline signal/wait, keyed-mutex timing-only ordering #1274) and, like the direct flatten, runs on app frames only (#868/#875). The plane's VkFormat is now exposed by comp_vk_deposit_plane_get; anything but B8G8R8A8, the legacy hatch, or a missing scratch falls back to the direct plane flatten. Measured (hardware A/B pending): cube_handle_vk_win, DXR_LOCAL2D_PANEL=1, opaque checker authored 40/235 -- expect _SRGB panel 40/235; UNORM panel 110/246; UNORM under DXR_COLOR_LEGACY_UNORM_ENCODED=1 40/235. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
dfattal
marked this pull request as ready for review
October 3, 2026 17:47
This was referenced Oct 3, 2026
Collaborator
Author
|
Numeric follow-up (with the #1810 capture fix): the D3D11 service pure-zones path is now measured, not just eyeballed. In |
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.
The deferred half of #1795 (ADR-044 §7): Local2D layers honour their declared format on every in-process backend and on the D3D11 service's pure-zones commits.
Blocked on apps declaring Local2D(cleared, see Measured):_SRGBEach of these would wash out on this runtime until it is rebuilt.
What changes
_SRGBRTV for the flatten. The source view follows the target._SRGBRTV,_SRGBflatten PSOs. The masked composite views the TYPELESS scratch typed.GL_SRGB8_ALPHA8flatten target withGL_FRAMEBUFFER_SRGBon for the flatten only, thenglCopyImageSubDatainto the unchanged RGBA8 scratches._SRGB} with opt-in_SRGBflatten pipelines. Under the output-device split (always engaged on hybrid Windows boxes, even on one adapter), the flatten goes to the scratch and is thenvkCmdCopyImaged into the B8G8R8A8 bridge plane._SRGB-view compose target only when some zone/Local2D source is not_SRGB. All-_SRGBcommits stay byte-identical (passthrough).The legacy hatch
DXR_COLOR_LEGACY_UNORM_ENCODED=1keeps every old path.Measured (win box)
Test app:
cube_handle_*_win,DXR_LOCAL2D_PANEL=1, opaque checker authored 40/235. Captured with a screen grab /PrintWindow, because Local2D is flattened after the atlas._SRGBpanelcube_zones_d3d11_winunderDISPLAYXR_ZONES=1:_SRGBstrip looked correct;capture_framereturnedok:false.The test apps gain a
DXR_SWAPCHAIN_ENCODING=unormoverride for the panel/strip, the same switch displayxr-common uses.🤖 Generated with Claude Code