Release Kaleido 0.6.1 with smoother browsing
All checks were successful
Validate and publish Kaleido / android (push) Successful in 2m24s
All checks were successful
Validate and publish Kaleido / android (push) Successful in 2m24s
Home, Movies and TV stay composed once shown and are revealed instead of rebuilt; hidden screens are unplaced, have no semantics and take no focus. The hero stays composed as a zero-alpha layer. Lazy-list prefetch runs on FramePrefetch, which uses only spare frame time (unlimited only when truly idle) instead of Compose's scheduler, which gave late frames unlimited time and composed whole rows mid-scroll. CacheWindows keeps Compose's cache-window strategy on that scheduler and premeasures nested cards. With any accessibility service enabled, Compose walked and sorted every composed node every 100 ms while browsing; rows and cards outside the viewport now carry no semantics. Cards use a lighter artwork path, lay out both title weights up front, and models are @Immutable. Preview and hero copy compose over the transparent start of their fade. Library sync applies its snapshot after 2 s without input, and moveToRow retries focus so a refreshed card is never left unfocused. profile_transitions.py now counts frames actually dropped on the display. On the onn the 42 measured transitions dropped 24-38 frames, versus 55 for 0.6.0; details are in docs/runtime-verification.md. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
parent
ca5f0e20e5
commit
30e0f30bca
20 changed files with 502 additions and 111 deletions
|
|
@ -4,6 +4,38 @@ For the 2026-10-26 codec/HDR sweep and its native-versus-converted playback find
|
|||
|
||||
Target: the authorized onn at `192.168.50.195:5555`, package `tv.kaleido.app`. Other applications and the launcher were not modified. Screenshots, traces, UI dumps and server-specific logs stay in ignored `evidence/private/`.
|
||||
|
||||
## 2026-10-28/29 browse frame pacing toward 60 fps (0.6.1)
|
||||
|
||||
Measured on the onn with signed release builds, each freshly installed, profile-broadcast and `speed-profile` compiled (the same preparation for 0.6.0 and every candidate; a long-lived `bg-dexopt` install measures faster and is not comparable). `profile_transitions.py` now reports `dropped`: refreshes that repeated a frame, from `DisplayPresentTime`. A gap longer than about six frames counts only when the frame before it ran long enough to cause it. Frame latency alone misleads: the render thread usually waits 8–12 ms in `dequeueBuffer` because the pipeline is full, and buffer stuffing hides many 20–30 ms UI frames. The metric still cannot tell a deliberate pause from a stall, and run-to-run variance is large (the same code measured 24–38 drops), so compare totals across several runs.
|
||||
|
||||
Causes found with atrace, temporary measure sections and `simpleperf` (a temporarily `profileable` release):
|
||||
|
||||
- **Tab switches** disposed one section (~31 ms of `onForgotten`) and composed the next (50–100 ms frames). Home, Movies and TV now stay composed once shown; hidden ones are neither measured, placed nor drawn, carry `clearAndSetSemantics{}`, and take no focus. A hidden screen gives up focus and moves none; shown again it focuses what a new screen would. Movies and TV rows are kept per destination and keep their previous rows while the same library reloads. A retained switch is now one ~25 ms frame (placing and redrawing the screen).
|
||||
- **The hero** was composed again on every return; it stays composed as a zero-alpha layer without semantics.
|
||||
- **Prefetch**: Compose's scheduler treats a frame after a late one as idle and gives its request unlimited time; it composed whole rows (35–45 ms) mid-scroll. `FramePrefetch` uses only what is left of each frame, unlimited only when nothing has drawn for two frames, and rotates past a request that cannot progress (a hidden screen's list) instead of letting it block the queue. `CacheWindows` keeps Compose's cache-window strategy on that scheduler; it must subclass `CacheWindowLogic`, because the list measure policy only keeps scrolled-out items when the strategy is one. Nested prefetch composes and measures each card a row will show (Compose precomposes two). The catalog window is two viewports each way, refilled after row changes.
|
||||
- **Accessibility** was the largest single cost. With any accessibility service enabled (the onn runs Luma's; Button Mapper was removed at the user's request), Compose rebuilds and geometry-sorts its whole semantics tree, placed or not, after every layout change, at most every 100 ms: about 40% of main-thread time while browsing. Rows and cards outside the viewport now carry no semantics (`semanticsOnScreen`, a `Modifier.Node` that observes the list and never recomposes). With accessibility temporarily disabled (restored afterwards), all 42 transitions dropped 16 frames; with Luma's service the best runs dropped 24–33.
|
||||
- **Cards**: `CardArtwork` draws a memory-cache hit directly and otherwise loads then draws without recomposing, instead of Coil's painter (~2.5 ms per card). Staging applies only while a section first enters. Both title weights are laid out with the card and focus swaps their layers, so a move no longer re-measures two titles. Models are `@Immutable`, so equal rows and cards skip recomposition after a refresh.
|
||||
- **The preview and hero copy** compose their parts over the first frames of their fade while still transparent (12 ms in one frame before).
|
||||
- **Library sync** refreshes every 60 s, and Plex varies some Movies hubs on each request, so each refresh rebuilt rows; landing mid-move it caused a recurring 80–150 ms frame. The staged snapshot now waits for 2 s without a key before it is validated and applied. That exposed a latent bug: after a refresh replaced the focused card, the replacement could compose a frame after focus was requested, leaving nothing focused. `moveToRow` now retries for a few frames.
|
||||
|
||||
Dropped frames across three cycles of all 14 transitions (cycle 1 is the first pass after launch), 0.6.0 versus this build, Luma's service enabled:
|
||||
|
||||
| Transition | 0.6.0 | this build |
|
||||
| --- | --- | --- |
|
||||
| home: tab -> hero | 5 / 0 / 0 | 5 / 0 / 0 |
|
||||
| home: row -> row | 4 / 0 / 2 | 4 / 1 / 1 |
|
||||
| home: hero <- row | 1 / 1 / 1 | 0 / 0 / 0 |
|
||||
| tabs: home -> movies | 7 / 4 / 5 | 6 / 2 / 1 (the first composes Movies) |
|
||||
| movies: row -> row | 1 / 2 / 0 | 1 / 0 / 1 |
|
||||
| tabs: home <- movies | 2 / 7 / 6 | 2 / 1 / 2 |
|
||||
| all 42 transitions | 55 | 33 (24–38 across runs) |
|
||||
|
||||
On the TCL (two cycles, 28 transitions) 0.6.0 dropped 39 and this build 31; Movies → Home frames fell from 118 ms to 30 ms of UI time. The TCL is GPU-bound: the preview crossfade takes 17–18 ms of GPU time per frame (two full-screen backdrop layers), against 7 ms at rest. Its ADB reset sometimes failed and its later frames arrived every 25 ms, so treat those numbers as indicative.
|
||||
|
||||
Tried and reverted: delaying tab activation until the highlight settles (not measurably better, adds latency); a graphics layer per retained screen (the redraw was not avoided); preloading Movies/TV after startup and composing them unseen (first visit 7 → 2 drops, but it added post-startup load and coincided with an accessibility-dump anomaly); capping idle prefetch to one frame (rows were unprepared for quick moves). Silo's Android source was reviewed: standard Compose for TV with default prefetch, per-card spring focus animations, Coil 3 and a Macrobenchmark baseline profile; its settled-focus prefetch matches Kaleido's `anticipate`. JIT was ruled out here (about 16 compilations in a 6 s trace). On Google TV the default-Home preference is ignored (`launcherx` keeps resolving), and disabling `launcherx` would hand Home to Google's setup recovery activity, so Luma still needs its service.
|
||||
|
||||
Checks: `Build.ps1 -Release -Test` passed; `repro_runtime.py browse` passed (stable geometry, GPU p95 6 ms); `verify_ui.py browse` (8), `repro_runtime.py sections`, `verify_movie_details.py` and `verify_library_jump.py` passed. `verify_library_jump.py` now reads the first of a rail entry's labels (`# #`); it failed identically on 0.6.0. Since Button Mapper's removal, `verify_details_layout.py` fails at the synopsis dialog on 0.6.0 and this build alike: uiautomator dumps the window behind the dialog, while the dialog is shown with Close focused. The goal of no dropped frames is not met: tab switches still have one long frame, row moves sometimes drop one or two, the first Movies visit composes the section, and the Fire TV was not measured. Playback suites were not run; playback code did not change.
|
||||
|
||||
## 2026-10-28 playback quality, audio selection and show shuffle (0.5.0)
|
||||
|
||||
- Original quality now uses the selected media version's source bitrate; Infinity War shows `19.1 Mbps`. No source bitrate is invented when Plex omits it. The quality menu and audio-track popup were checked in actual device screenshots.
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue