Repository navigation
Conversation
Galaxy S23 — ChromePhysical SM-S918B, Chrome 154; native touch input, with the regional dust dataset enabled or disabled before mounting. Chrome trace frame-delivery rates:
Single recordings per scenario; three-repeat variability is not measured here. Both conditions contain dropped/partial reporter records, which are not unique missed frames. The dust adds a drag deficit while coast is approximately at parity. Recorded GPU command-buffer flushes take 22–50 ms during input; a 150.087 ms presentation gap includes a 130.854 ms flush and 44.36 ms of concurrent main-thread work. The responsible primitive or driver wait is unresolved. iPad — SafariPhysical iPad12,1, iPadOS 27.0.1; viewport 810 × 994 CSS px, DPR 2. Three repeats for each dust-on/off condition at the opening view, supplied saved view and inside. The same paired sequence is off/on, on/off, off/on. Eighteen complete recordings contain both drag and coast, yielding six scenarios. Jankmonster’s recorder targets Chrome/CDP. Its report accepts the iPad’s exported WebKit CPU trace, but displayed-frame and dropped/partial counters require Chromium PipelineReporter events: the fresh device trace imports as usable with
Means ± sample SD across three runs; differences calculated before rounding. These are RAF callback intervals, not display/compositor frame delivery. Compare dust on/off within each device; Safari RAF fps and the S23’s presentation fps are different metrics. Page-dispatched pointer input, not native touch: a 60-frame horizontal fling, true application coast, then a 60-frame held return. Application motion events define the disjoint drag, coast, still and boundary bins. The fixed endpoint is 180 CSS px from the start. Driven time includes a 120 ms finger rest. The Inspector detaches before input; no profiler or native screen sampler is active during timing. All 18 runs complete 660 intervals with one coast onset and no skipped steps. Raw timestamps replay exactly through the phase-summary function; all frame and late counts are conserved. Dust off removes the regional dataset before mount, including foreground boards, cirrus and stellar transmission. The start receipts show 88 foreground elements and six cirrus faces on, zero off. The application code and prepared bank match this PR; capture-helper changes are confined to performance tooling. The dust-on transports are restored. Thermal and Low Power state were not sampled; normal browser cache is enabled. Opening and saved-view dust-on coast vary substantially. Repeat counts and interval durations
Boundary-crossing intervals, including release frames, remain in the mixed bin. The whole-program column includes every phase; still and unknown intervals remain in the raw records. Estimated missed 60 Hz opportunities sum nominal refresh slots from intervals over 25 ms; they are not actual dropped frames. iPad viewsNative device screens were captured separately from timing and inspected with
Prepared rendering and matched GPU imagesNative Metal Chrome, 1372 × 871 CSS px, DPR 1. Opening, saved, 90°, behind and inside captures each mount 88 foreground elements and six cirrus faces, with no page errors or console warnings. The source recipe reproduces geometry, decoded foreground RGBA, cirrus bytes and stellar attenuation exactly. Device timings above use these same drawing payloads. Pixelmatch 7.2.0, threshold 0.1, includeAA=false: 0 / 779,240 differing pixels in each of the five views. Each comparison uses the same unscaled crop (x=360, y=70, width=1012, height=770), selected to exclude the shell, sidebar and release badge. These are separate real-site reference and rendered captures; this checks appearance, not scientific accuracy or displayed-frame delivery. Reference:
Rendered:
Difference:
|
apresmoi
changed the title
880ecdb to
ddd7ebf
Compare
ddd7ebf to
3a688b0
Compare
The Pleiades dataset combines Gaia and Hipparcos stars with the surrounding Milky Way dust cloud. Foreground dust scatters the bright stars' light; wider cirrus is faintly lit by Galactic starlight, and the full field dims and reddens background stars. The page opens from Earth's direction. Positions, images and lighting stay fixed while the camera turns.
Member reddening and Gaia constrain placement near 136 pc. Edenhofer and Leike provide broad depth information; WISE, Spitzer and IRIS constrain structure on the sky. Published B-star parameters and Draine grains supply single-scattering light, with the modified-MMP Galactic field at U = 1. The NOIRLab optical image checks the result; it does not supply dust texture. The depth input retains 32,657 native-extracted samples. Sources, processing and assumptions: Pleiades, dust model.
Limits: cloud thickness and infrared-derived mass are uncertain; bright-star depths are adopted. Fixed Earth-reference lighting and finite billboards approximate other viewpoints; a large element can clip visible support on the opposite phone orbit. Galaxy S23 drag is 10–14 fps below dust off, with coast near parity; both conditions contain dropped/partial-frame reporter records. The slowdown is GPU-side; its cause and a recorded 131 ms command-buffer flush remain unresolved. On iPad, mean RAF drag pacing is about 4 fps lower; opening coast is about 46 versus 57 fps, with substantial variation. RAF gaps are not hardware dropped-frame counts. Native iPad screens show faint straight-edged dark boundaries in the Earth view; that visual check remains open.
The source recipe reproduces the retained geometry, foreground RGBA, cirrus bytes and stellar attenuation exactly. An empty destination restores all 14 runtime files and 25 numerical inputs from R2.
Clean: full typecheck, color spelling, hash placement, literal styles, prepared delivery, architecture, affected renderer/objects/site tests, shipped-form checks and production site build.