[Markgatcha]

On Windows and Linux, capture_target_thumbnail grabs a full-resolution screenshot and hands it to fit_centered_capture to build the picker card. That fits by box-averaging the source down by a whole factor first and letting Lanczos finish the fit. For a 1440p display the box pass was about 25ms of the 66ms the thumbnail took, and almost all of it was per-pixel overhead rather than arithmetic.

The loop made two decisions that only depend on the column, but made them for every pixel: which BlockDivisor to divide by (if index + 1 == reduced_width), and where the source block starts (columns.chunks(factor * CHANNELS)). Only the last column averages a short block, so it can take its own divisor outside the loop, and every other block is just index * factor * CHANNELS. Splitting the last column out and indexing directly is 2.8x on that phase in isolation, with byte-identical output.

Whole fit, medians of three interleaved builds per side:

capture	before	after	speedup
2560x1440	63.7 ms	34.4 ms	1.69x
3440x1440	80.4 ms	42.7 ms	1.66x
3840x2160	68.5 ms	46.0 ms	1.49x
5120x2880	77.4 ms	56.8 ms	1.39x
3840x2160 window	72.0 ms	57.0 ms	1.39x
1920x1080	45.7 ms	44.6 ms	unchanged

1080p can't move: prescale_factor comes out below 2 at that size, so it never enters box_downscale.

cargo test -p cap-gpui-kernels passes. box_downscale_averages_blocks_and_keeps_partial_edges recomputes the expected block averages by hand, including the short edge blocks that this loop now handles separately, and capture_fit_stays_within_tight_bounds_of_the_exact_fit covers the fitted output.

One note on method, since it decides whether the table is worth anything: this machine drifts between builds. The same unchanged 1440p capture measured 61.6ms in one pair and 63.7ms in the next, and the bench's own control column moved by up to 20% between rebuilds. So each row is a rebuild of main followed by a rebuild of this branch, three times over, with the speedup taken as the median of the three and each side normalised against the image crate column that benches/thumbnail_capture.rs measures in the same run.

Confidence Score: 5/5

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

Summary

Speeds up box_downscale by handling the final column separately and moving the shared sum loop into block_sums.

• Large capture thumbnails skip edge checks inside the pixel loop.

_{Reviews (1) · Last reviewed commit: "perf(desktop-gpui): keep the edge-column..." · Reviewed by Greptile}

[vercel]

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

A member of the Team first needs to authorize it.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-25T04:11:58.025316Z We couldn’t complete this review. Comment @codex review to try again.	dd73afc	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.