Repository navigation
fix(compositor): restore the 2.0 export speed on release/v2.0.0 - #984
Merged
Merged
Conversation
Every layer is drawn by one WGSL pixel shader that branches on its mode, and ACO allocates registers for the heaviest branch. With the ray-traced cursor model, its click impact and the device frame (modes 15-17) in it, the shader needed 168 VGPRs and ran 6 waves per SIMD instead of 32: the wallpaper, the shadow and the screen paid that occupancy on every pixel. On a Radeon 610M the benchmark export went from 31 s (1.11.0-rc.1) to 57 s (2.0.0-rc.12). Compile layer.wgsl twice, with and without LAYER_MODELS, and draw only the three modelled layers through the second pipeline. The flat one needs 48 VGPRs (20 waves). Export 57 s -> 36.5 s, same pixels. (cherry picked from commit 1bd3e4d)
`select` evaluates both of its operands, so the screen's corner mask computed three continuous-corner SDFs per pixel (sd_round_rect, and sd_screen_under_bar which calls it twice) where one was used: ~0.6 ms per 1080p frame on the two compute units of a Radeon 610M. Branch on the uniform instead, in the screen tail and in the tilted mode 8. sd_round_rect also returns the plain box distance away from the corners, where the exponent and the extent cancel out, so exp2 and the divisions run only near them. Screen draw 1.53 -> 0.81 ms; the export is unchanged to the byte. (cherry picked from commit 9c7f661)
The cursor trail accumulates in its own target and was then copied back over the whole output: 0.5 ms per 1080p frame on a Radeon 610M, on every frame where the cursor moves, for a sprite a few dozen pixels wide. Only the trail's quads are ever painted into `accum`, so the copy is scissored to their union. (cherry picked from commit 1204aca)
… Linux Pass 1 sampled the 3840 px wallpaper without mips onto the whole output every frame: 1.65 ms on a Radeon 610M, more than the screen itself, for a picture that does not change. Keep the background as pass 1 and its blur left it, keyed on everything that decides it (clear colour, layer, blur amount, render size), and copy it into the render target while the key holds. An animated background is never kept, nor one whose image failed to load. Benchmark export on that machine: 34.1 s -> 30.9 s, against 31.3 s for 1.11.0-rc.1 in the same session and 57 s for 2.0.0-rc.12. The MP4 is byte-identical to rc.12's. (cherry picked from commit c00e5b7)
The models pipeline was chosen at the three call sites known to draw a 3D model: the cursor model, its click impact and the device frame. The device's shadow is a mode-17 model too (device_shadow_cb builds on device_frame_cb), and it went through the flat pipeline, which compiles no model and drew nothing. LayerCB::needs_models now says which layers need the models variant, shared by the three backends, and every Linux layer is drawn by draw_layer, which reads it from the LayerBind that make_bind returns. A call site can no longer get it wrong: forcing the flag off fails the device shadow, device frame, cursor trail and click impact tests. (cherry picked from commit 294b0eb)
…shaders Same split as on Linux: ps_main is compiled once without LAYER_MODELS for every layer, and once with it for the cursor model, its click impact, the device and its shadow, chosen per draw from LayerCB::needs_models. D3D11 gets ps_main_models from build.rs (a D3D_SHADER_MACRO on the same source); Metal compiles a second library with the macro prefixed, and draw_layer switches to the models variant of the pass's pipeline for the draw. The flat ps_main compiles to 40 KB of DXIL against 137 KB with the models (dxc 1.9, ps_6_0). Neither backend can be built from Linux: compile_hlsl was type-checked against x86_64-pc-windows-msvc, the HLSL compiled with dxc, and every_shader_entry_point_compiles now builds the Metal variant too. (cherry picked from commit 48c2778)
…nd macOS The HLSL and Metal ports of 9c7f661. `?:` with two SDF calls computed three continuous-corner SDFs per screen pixel where one is used (HLSL evaluates both operands), and sd_round_rect paid for its exponent away from the corners, where it cancels out. A uniform branch, and the box distance off the corners. (cherry picked from commit 527f934)
…ground The Linux background cache built its key every frame from the image path's bytes, so a wallpaper stored as a data URI was copied in full, several megabytes, on every frame. The key is now a BackgroundKey kept with the cached copy (the scene's background, blur, fallback colour, render size) and compared field by field against the requested one; it is only built, and the background cloned, when a new background is kept. Shared in frame_geometry so the other backends can use the same rule. (cherry picked from commit 73de470)
Contributor
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Advanced Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
This was referenced Oct 3, 2026
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.
Summary
Cherry-picks the code of #982 onto the frozen
release/v2.0.0: the Linux export of 2.0.0-rc.12 runs at 2.43× the ffmpeg floor against 1.29× for 1.11.0-rc.1 on the same machine and session, and this restores it (1.29×, byte-identical MP4). Full write-up in #982.The eight code commits,
git cherry-pick -xin order; the two docs commits of #982 are left onmain, as the release contract asks:crates/here is identical to #982's tip (git diff 73de470e -- crates/is empty), so #982's CI results carry over. #983 (the same optimizations ported to Windows and macOS) is not included: it is not a regression fix.After merge: rerun
prerelease.ymlwithrc_number13 to tag the branch tip.Type of change
Release impact
Desktop impact
Testing
See #982: Linux GPU tests and benchmark on a Radeon 610M; Windows and macOS compiled and tested in CI only (the macOS runner renders the 3D-model tests on Metal), not measured on hardware.
🤖 Generated with Claude Code