[Performance] Remove redundant image scheduler hop#80
Merged
Conversation
NachoSoto
added a commit
that referenced
this pull request
Jul 12, 2026
## Summary This branch combines the independently reviewed changes from: - #77 — skip unchanged UIKit image sizes - #78 — streamline SwiftUI geometry observation - #79 — preserve `UIImage.scale` in the disk cache - #80 — remove the redundant image-scheduler hop - #81 — lazily initialize SwiftUI image state - #82 — reduce multicast contention while preserving subscription semantics ## Component benchmarks | Change | WatchChess workload | Result | |---|---|---:| | UIKit unchanged-size guard (#77) | 128,000 origin-only mutations | **-32.24% mean**, scheduler calls 128,000 → 0 | | SwiftUI geometry (#78) | 100 real cells × 80 relayouts | **-2.86% mean**, 95% CI for saving 9.053–45.532 ms | | Scheduler hop (#80) | 96 cold + 96 repeated requests | scheduler operations **-50%**; wall time neutral | | Lazy SwiftUI state (#81) | 20,000 body reconstructions | **-4.44% mean**, 12/12 paired wins | | Multicast completed hits (#82) | 50,000 real `PieceView` hits | **-89.3% / 9.33× faster** | | Multicast distinct keys (#82) | 24 concurrent keys | neutral: -1.4% mean, 95% CI spans zero | ## Combined WatchChess benchmark Release WatchChess, iPhone 17 Pro Max / iOS 26.5 simulators, 10 samples per variant and flow. Negative CPU/hitch changes are improvements. ### CPU | Flow | CPU time, baseline → combined | Change (paired-bootstrap 95% CI) | Mean one-core-equivalent CPU | |---|---:|---:|---:| | Tournament-list scrolling | 2.230 → 2.264 s | +1.54% (-2.79% to +6.09%): neutral | 12.01% → 12.07% | | Rounds scrolling | 1.059 → 1.024 s | **-3.25% (-5.55% to -1.27%)** | 13.03% → 12.79% | | Open game | 0.452 → 0.439 s | **-2.89% (-5.36% to -0.12%)** | **13.47% → 13.11%** (-2.63%; CI -4.37% to -0.87%) | ### FPS and hitches `FPS` below is a main-run-loop `CADisplayLink` callback proxy over the exact UI-action window, not compositor presentation FPS. | Flow | FPS, baseline → combined | FPS change (95% CI) | Hitches/min, baseline → combined | Hitch change (95% CI) | |---|---:|---:|---:|---:| | Tournament-list scrolling | 58.93 → 59.00 | +0.12% (-0.20% to +0.48%): neutral | 40.67 → 37.39 | -8.06% (-29.75% to +14.24%): neutral | | Rounds scrolling | 58.53 → 58.98 | **+0.77% (+0.13% to +1.39%)** | 63.24 → 51.68 | -18.28% (-35.36% to +5.33%): trend, CI spans zero | | Open game | 54.14 → 54.07 | -0.14% (-1.20% to +0.91%): neutral | 131.04 → 145.44 | +10.99% (-5.26% to +29.68%): neutral | Missed-refresh rates were also neutral for tournaments (-3.62%, CI -21.79% to +14.17%) and game (+2.99%, CI -5.93% to +13.06%). Rounds trended lower (-26.69%, CI -47.25% to +4.29%). ## Methodology and controls - two-pass, pass-outer crossover: pass 1 baseline on simulator A then candidate on B; pass 2 candidate on A then baseline on B - five measurements per pass, variant, and flow; 10 total per variant/flow - engine analysis forced off (`currentEngineStrength = 1`) and validated inside every app launch/report - exact 24-entry real API fixture in cache-only mode; any cache miss failed the run - identical 147-path warm AsyncImageView cache coverage per variant - exact AsyncImageView path/hash/mtime validation before sample 1 and after every five-sample block - API path/size/hash validation before and after every block - arithmetic-mean CPU headline; duration-weighted ratio-of-sums FPS/hitch rates - deterministic 100,000-resample, pass-stratified paired bootstrap confidence intervals - symmetric full cache hashing before each measured block; these are deliberately warm-cache measurements - hitches: clipped display-link interval at least 1.5× its target refresh period All earlier measurements with uncontrolled engine state or unequal cache warmness were quarantined and excluded. ## Rendering and cache regression checks - cold game and rounds screenshots were pixel-identical; cold tournament app content was pixel-identical outside dynamic system chrome - all four pass-1 seed/verify tournament screenshots were pixel-identical across the 1320×2590 app region; the only difference was 5,708 pixels in the translucent system search-field blur - six gated 8 FPS recordings covered tournament scrolling, rounds scrolling, and opening a game for both variants - manual contact-sheet review found no blank flashes, stale substitutions, missing portraits, or persistent placeholders; both game recordings showed the same normal transition placeholders and matching fully rendered end states - all six recordings passed pre/post AsyncImageView and API cache audits - baseline and candidate kept identical 147 relative cache paths; each variant's hashes and mtimes remained stable The different cache bytes between variants are intentional: #79 stores image scale in a versioned envelope. Legacy raw-PNG entries become one normal miss and regenerate in the corrected format. Existing remote images may therefore re-download once; offline users can temporarily miss legacy cached images until connectivity returns. ## #79 and #82 semantics #79 and the [#82 synchronous-subscriber review finding](#82 (comment)) are different fixes, and both are included: - #79 fixes PNG disk-cache scale metadata. - #82 attaches the first subscriber before starting synchronous upstream work, preserving first-request miss / repeat-request hit semantics. #82's `lifetime.observeEnded` intentionally holds the wrapper and in-flight entry only until producer completion or cancellation. Weak-reference tests verify release after both paths, while cancellation still disposes upstream work; no retain cycle or leak was observed. ## Validation - `swiftlint --config .swiftlint.yml --strict`: 0 violations - full iOS simulator suite: 69 XCTest/Quick + 13 Swift Testing tests = **82 passed** - focused multicast concurrency, synchronous, retry, memory-warning, cancellation, and weak-lifetime coverage - WatchChess Release cache preflights, 60 CPU samples, 60 frame reports, and six visual recordings passed their evidence checks - CI updated to macOS 26 / Xcode 26.5
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
Performance
Isolated AsyncImageView benchmark
20,000 unique size requests across 10 runs:
WatchChess integration benchmark
Release-hosted WatchChess benchmark with 96 real
PieceViewrequests and 96 repeated hits per sample, 10 strictly alternating samples per variant:The paired cold delta was +0.831 ms (95% CI -1.515 to +3.177 ms); the paired warm delta was -0.097 ms (95% CI -0.631 to +0.436 ms). The app workload confirms the operation reduction, but not a measurable wall-time change.
Rendering and cache parity
Baseline and candidate matched UIKit and SwiftUI image/snapshot hashes, point and pixel sizes, data/size invalidation sequences, cold/repeated cache semantics, and persistent
miss → hit → hit after renderer recreationbehavior. Every renderer started on the requested image scheduler and every UI callback ran on the main thread.The benchmark also reproduced a pre-existing disk-reload scale-metadata loss in both variants: raw pixels are preserved while
36 pt @2xreloads as72 pt @1x. This PR does not change it.Validation
swiftlint --config .swiftlint.yml --strict