perf(desktop-gpui): presize flattened RGB buffer

[kyzer1023]

Summary

flatten_rgba_to_rgb_over_white built its output with a per-pixel extend_from_slice. It now allocates vec![0u8; rgba.len() / 4 * 3] once and writes each pixel into dst[i..i + 3], using the same three alpha branches and the same blend. rgba_to_rgb and encode_jpeg_rgba_as_rgb are untouched.

Files: apps/desktop-gpui/kernels/src/codec.rs and a bench case in apps/desktop-gpui/kernels/benches/frame_export.rs. Based on main at 85219015.

Correctness

Output is byte-identical to the old loop for opaque, fully transparent, and mixed-alpha buffers, including a hand-checked alpha == 128 pixel. cap-gpui-kernels lib tests: 42 passed, 1 ignored, on rustc 1.88.0 and 1.99.1. The full desktop-gpui app was not built here because gpui is patched to a local zed-cap checkout.

Benchmarks

Measured under the profile desktop-gpui ships with (its own workspace: opt-level = 3, lto = "thin", codegen-units = 1, strip = "symbols"), rustc 1.99.1, 4 vCPU. Each run is 5 warmups and the median of 25 samples; the table is the median of 5 runs. 3840×2160.

Workload	Before	After	Speedup
Opaque	7.888 ms	6.349 ms	1.24×
Fully transparent	8.593 ms	5.091 ms	1.69×
Mixed alpha	7.937 ms	8.105 ms	0.98×

Opaque frames (the common share case) and transparent frames are faster. Mixed alpha is about 2% slower on stable and flat on 1.88.0.

Overlap

Does not touch rgba_to_rgb / encode_jpeg_rgba_as_rgb (#2672) or PNG filtering (#2674).

Confidence Score: 5/5

This PR appears safe to merge; no actionable issues were found.

Summary

flatten_rgba_to_rgb_over_white now sizes the RGB buffer upfront and writes pixels into it directly.

• Screenshot JPEG exports build flattened RGB frames without per-pixel appends.

No actionable issues found. Tests and benchmarks were inspected, not run.

_{Reviews (1) · Last reviewed commit: "perf(desktop-gpui): presize flattened RG..." · Reviewed by Greptile}

[vercel]

@kyzer1023 is attempting to deploy a commit to the Cap Software Inc Team on Vercel.

A member of the Team first needs to authorize it.