perf(android): normalize variable refresh rate and render resolution #1083

Closed
opened 2026-11-08 00:39:02 +00:00 by SakulFlee · 3 comments
Owner

Summary

Android devices differ in native resolution, DPI, and variable display refresh rate. Orbital currently configures the surface at the native window size and leaves refresh selection to the platform/compositor. That makes phone/tablet performance captures hard to compare: raw FPS can change because of resolution or refresh before any renderer difference is involved.

This is an Android-platform concern independent of the exact phone/tablet settings.

What is missing

  • The active display mode is not consistently exposed to the renderer or capture metadata.
  • There is no engine-level render-scale/internal-resolution control.
  • There is no Android-facing frame cap or refresh-rate preference.
  • Captures do not clearly separate native resolution, render resolution, refresh rate, and present mode.
  • Debug UI access currently also blocks reliable on-device A/B testing (#1082).

Proposed direction

  • Query and record physical resolution, DPI, refresh rate, present mode, and desired frame latency in the metadata sidecar and log.
  • Add an internal render-scale/resolution setting with presets such as native, 0.75x, and 0.5x.
  • Add an optional Android frame cap or refresh-rate preference where the platform APIs allow it, with a documented fallback.
  • Record the active render resolution separately from the surface resolution.
  • Make captures directly comparable by reporting FPS, frame time, render scale, and pixels per frame.
  • Test the same scene at matched resolution and refresh on phone and tablet; compare 60/90/120/144 Hz modes where available.
  • Document that raw FPS across Android devices with different resolution/refresh is not directly comparable.

Acceptance criteria

  • Android captures include resolution, DPI, refresh rate, render scale, present mode, and frame latency.
  • Render scale can be changed on-device without rebuilding.
  • Phone/tablet comparisons can be made at matched resolution and refresh.
  • Desktop behavior is unchanged.
## Summary Android devices differ in native resolution, DPI, and variable display refresh rate. Orbital currently configures the surface at the native window size and leaves refresh selection to the platform/compositor. That makes phone/tablet performance captures hard to compare: raw FPS can change because of resolution or refresh before any renderer difference is involved. This is an Android-platform concern independent of the exact phone/tablet settings. ## What is missing - The active display mode is not consistently exposed to the renderer or capture metadata. - There is no engine-level render-scale/internal-resolution control. - There is no Android-facing frame cap or refresh-rate preference. - Captures do not clearly separate native resolution, render resolution, refresh rate, and present mode. - Debug UI access currently also blocks reliable on-device A/B testing (#1082). ## Proposed direction - Query and record physical resolution, DPI, refresh rate, present mode, and desired frame latency in the metadata sidecar and log. - Add an internal render-scale/resolution setting with presets such as native, 0.75x, and 0.5x. - Add an optional Android frame cap or refresh-rate preference where the platform APIs allow it, with a documented fallback. - Record the active render resolution separately from the surface resolution. - Make captures directly comparable by reporting FPS, frame time, render scale, and pixels per frame. - Test the same scene at matched resolution and refresh on phone and tablet; compare 60/90/120/144 Hz modes where available. - Document that raw FPS across Android devices with different resolution/refresh is not directly comparable. ## Acceptance criteria - Android captures include resolution, DPI, refresh rate, render scale, present mode, and frame latency. - Render scale can be changed on-device without rebuilding. - Phone/tablet comparisons can be made at matched resolution and refresh. - Desktop behavior is unchanged.
Author
Owner

Concrete Android evidence from the phone capture:

  • Surface/window: 1220 x 2582, scale factor 3.25
  • monitor.refresh_rate_mhz: 0 (winit does not provide a usable refresh rate here)
  • Present mode: AutoNoVsync; supported: Mailbox, Fifo; vsync off; desired latency 2
  • The phone was rotated roughly halfway through the run, but the CSV has no per-row resolution/orientation marker. The metadata sidecar is only a startup snapshot and is not refreshed when the surface changes.

This should therefore include per-row capture fields such as window width/height, surface width/height, orientation, render scale, refresh rate and present mode, plus an Android display-mode/refresh fallback when winit reports 0.

The tablet files submitted with this run are actually phone duplicates (same Mali-G720/1220x2582 metadata; CSV byte-identical to an earlier phone capture), so they are not useful for the tablet comparison.

Concrete Android evidence from the phone capture: - Surface/window: 1220 x 2582, scale factor 3.25 - `monitor.refresh_rate_mhz`: **0** (winit does not provide a usable refresh rate here) - Present mode: `AutoNoVsync`; supported: `Mailbox`, `Fifo`; vsync off; desired latency 2 - The phone was rotated roughly halfway through the run, but the CSV has no per-row resolution/orientation marker. The metadata sidecar is only a startup snapshot and is not refreshed when the surface changes. This should therefore include per-row capture fields such as window width/height, surface width/height, orientation, render scale, refresh rate and present mode, plus an Android display-mode/refresh fallback when winit reports 0. The tablet files submitted with this run are actually phone duplicates (same Mali-G720/1220x2582 metadata; CSV byte-identical to an earlier phone capture), so they are not useful for the tablet comparison.
Author
Owner

Tablet evidence added:

  • Adreno 810, 2560 x 1600, scale factor 2.0
  • monitor.refresh_rate_mhz: 0 (same missing-refresh problem as the phone)
  • present mode AutoNoVsync; supported [Mailbox, Fifo]; vsync off; desired latency 2
  • no iced, so the tablet is a clean GPU test
  • 17 usable samples, ~2.7 FPS; frame ~358 ms; gpu_complete ~567 ms

The phone and tablet have different pixel counts and GPU classes, so their raw FPS cannot be compared directly. Matching render resolution and refresh (or at least recording them per row) is required.

Tablet evidence added: - Adreno 810, 2560 x 1600, scale factor 2.0 - `monitor.refresh_rate_mhz`: 0 (same missing-refresh problem as the phone) - present mode `AutoNoVsync`; supported `[Mailbox, Fifo]`; vsync off; desired latency 2 - no iced, so the tablet is a clean GPU test - 17 usable samples, ~2.7 FPS; frame ~358 ms; `gpu_complete` ~567 ms The phone and tablet have different pixel counts and GPU classes, so their raw FPS cannot be compared directly. Matching render resolution and refresh (or at least recording them per row) is required.
Author
Owner

Resolution A/B confirms the issue is real and high-impact:

  • Tablet native 2560x1600: ~2.63 FPS, frame ~372 ms, gpu_complete ~680 ms
  • Tablet 1280x800: ~8.25 FPS, frame ~111 ms, gpu_complete ~204 ms
  • That is 4x fewer pixels and ~3.3-3.4x faster frame/GPU completion.

The tablet display reports an active 48 Hz mode (supported 120/90/60/50/48/30), while winit metadata says refresh_rate_mhz = 0. present_mode is AutoNoVsync; supported modes are Mailbox/Fifo. Thermal status is 0 and GPU is ~40 C, so this is not throttling.

This strengthens the request for:

  • a first-class render scale/internal resolution setting;
  • explicit refresh-rate / present-mode selection;
  • per-row window, surface, orientation, render scale and refresh fields;
  • matched-resolution phone/tablet comparisons.
Resolution A/B confirms the issue is real and high-impact: - Tablet native 2560x1600: ~2.63 FPS, frame ~372 ms, `gpu_complete` ~680 ms - Tablet 1280x800: ~8.25 FPS, frame ~111 ms, `gpu_complete` ~204 ms - That is 4x fewer pixels and ~3.3-3.4x faster frame/GPU completion. The tablet display reports an active 48 Hz mode (supported 120/90/60/50/48/30), while winit metadata says `refresh_rate_mhz = 0`. `present_mode` is `AutoNoVsync`; supported modes are Mailbox/Fifo. Thermal status is 0 and GPU is ~40 C, so this is not throttling. This strengthens the request for: - a first-class render scale/internal resolution setting; - explicit refresh-rate / present-mode selection; - per-row window, surface, orientation, render scale and refresh fields; - matched-resolution phone/tablet comparisons.
Sign in to join this conversation.
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
SakulFlee/Orbital#1083
No description provided.