perf(desktop-gpui): accelerate screenshot-export and frame kernels

[mihir0209]

Rebased onto main (2b2bf84). Speeds up the GPUI screenshot-export path. All outputs are bit-identical to the current code.

Changes

• kernels/src/screenshot.rs — flatten_onto_white fast-paths fully-opaque pixels (skip) and fully-transparent pixels (fill white) instead of running three integer divisions per pixel, and hoists the per-pixel 255 * (255 - alpha) recompute. This has four production callers, including the unannotated-composite fast path added in perf: speed up unannotated screenshot exports #2493.
• kernels/src/codec.rs — new flatten_rgba_to_rgb_over_white fuses the Share-JPEG flatten plus the alpha strip into a single pass with one allocation, replacing two full-frame passes and a 33 MB transient at 4K. rgba_to_rgb pre-sizes its output instead of growing it through flat_map + collect.
• src/screenshot_export.rs — encode_jpeg uses the fused kernel.
• kernels/benches/frame_export.rs — before/after bench in the style of screenshot_blit.rs (old code preserved as mod baseline, equality asserted, alternating execution order, CSV output). Run with cargo bench -p cap-gpui-kernels --bench frame_export.

Benchmarks

Median of 25 samples after 5 warmup, alternating before/after order on the same buffers, opt-level = 3:

workload	before	after	speedup
flatten opaque 1080p	4.56 ms	1.48 ms	3.09x
flatten opaque 4K	18.41 ms	6.01 ms	3.06x
flatten mixed-alpha 4K	18.30 ms	7.70 ms	2.38x
export flatten + strip 4K (fused)	49.19 ms	15.20 ms	3.24x
rgba_to_rgb 4K	31.00 ms	19.44 ms	1.59x

The 4K Share-JPEG pre-encode stage drops ~34 ms and one 33 MB transient allocation.

Note on the last row: after this PR encode_jpeg uses the fused kernel, so rgba_to_rgb no longer has a production caller and is kept as a public-kernels improvement, covered by the bench.

Correctness

cargo test -p cap-gpui-kernels — 34 passed, 0 failed. The pre-existing exhaustive all-channel/all-alpha flatten test, the JPEG round-trip test, and the new fused_flatten_to_rgb_matches_two_step test (all 65,536 channel/alpha pairs, empty input, trailing bytes) all pass, so outputs are bit-identical.

A previous revision of this PR also changed frame::swap_red_blue to a u32 swizzle. The in-crate benchmark showed that as a 5.5x regression when called across a crate boundary (the copy_from_slice lowers to a per-pixel memcpy there), and a direct-index rewrite measured as a wash, so that change has been dropped rather than shipped.

Confidence Score: 5/5

The PR appears safe to merge; no actionable issues remain.

Summary

Speeds up screenshot export without changing the pixel formulas.

• Screenshot exports flatten onto white faster by skipping work for solid and clear pixels.
• Share JPEG exports turn RGBA into white-backed RGB in one pass.

All three earlier findings are addressed: benchmark columns stay correctly labeled, the fused function has unit tests, and the unnecessary comment is gone. The supplied threads were manually resolved by mihir0209 without explanatory replies.

The review used the full PR diff because the current and previously reviewed commits have different merge bases. Tests and benchmarks were inspected, not executed.

_{Reviews (6) · Last reviewed commit: "perf(desktop-gpui): speed up the screens..." · Reviewed by Greptile}

[vercel]

Someone is attempting to deploy a commit to the Cap Software Inc Team on Vercel.

A member of the Team first needs to authorize it.

[richiemcilroy]

Thanks for the PR. Greptile currently has this at 4/5, and we only merge PRs once Greptile scores the current head 5/5. Please address the remaining Greptile feedback, then comment @greptileai to request a re-review, and repeat until it reaches 5/5.

[mihir0209]

@greptileai please re-review. Addressed both P2 findings: (1) bench now keeps (baseline, candidate) column order on alternating samples in both emit and run_swap; (2) added exhaustive fused_flatten_to_rgb_matches_two_step unit test in kernels/src/codec.rs covering all channel/alpha pairs, empty input, and trailing bytes.

[mihir0209]

@greptileai both P2 threads are now resolved (fixes in 4e38f37: bench column order kept as (baseline, candidate) on alternating samples, plus exhaustive fused unit test). Please re-review the current head for the 5/5.

[mihir0209]

@greptileai run-command comment removed (2719e65), thread resolved. Please re-review the current head for the 5/5.

[mihir0209]

@richiemcilroy all Greptile feedback is addressed and verified on the current head (2719e65): bench column order fixed, exhaustive fused unit test added, run-command comment removed. All three review threads resolved and the Greptile Review run is green with no new findings. Ready for your 5/5 check whenever suits.

[mihir0209]

@greptileai fixed the CI compile error in the new unit test (pinned Vec:: for the empty-input assert, 69a7035). Please re-review the current head.

[richiemcilroy]

Thanks for this! It now conflicts with main after today's merges. Could you merge or rebase onto the latest main? Once it's up to date I'll run the full engine verify and real-capture suites on it and merge if they pass.

[mihir0209]

@greptileai rebased onto main (2b2bf84) with a single clean commit, and the benchmark caught a regression in my own change: the rame::swap_red_blue u32 swizzle is a 5.5x slowdown across the crate boundary, so it has been dropped. The PR now only carries the flatten fast-path, the fused flatten-to-RGB kernel and the pre-sized strip, with corrected before/after numbers from the in-crate bench. Please re-review the current head.

[richiemcilroy]

Verified: the alpha 0/255 fast paths are exact against the original formula, the fused flatten is tested exhaustively against the two-step path, CI green, Greptile 5/5. Thanks!