perf(android): Adreno 810 tablet is volumetrics-bound at ~162 ms #1089

Closed
opened 2026-11-10 15:02:11 +00:00 by SakulFlee · 4 comments
Owner

Summary

After the light-range fix (b7086621), the Adreno 810 tablet is no longer deferred-bound. It is volumetrics-bound:

Metric Tablet (2560x1600)
FPS 4.47
frame 219.5 ms
gpu_scene_total 38.11 ms
gpu_deferred 16.01 ms
gpu_light_culling 12.84 ms
gpu_volumetrics 161.78 ms
gpu_tonemap 13.24 ms
gpu_aa 4.34 ms

gpu_volumetrics is ~74% of the frame time. The phone’s volumetric cost at 1220x2582 is only ~12 ms, so the tablet’s ~162 ms is ~13× higher despite only ~1.3× the pixel count.

The tablet A/B against 5be0937d confirms the range fix helped there too (deferred 76.43 → 16.02 ms, scene 100.13 → 38.40 ms, FPS 3.0 → 4.5), but the volumetric pass remains the new bottleneck.

Likely contributors

  • The all-in-one scene uses VolumetricQuality::Low with max_distance: 30. Adaptive depth slicing raises the froxel grid to 192 slices, giving roughly 213×133×192 ≈ 5.4M froxels on the tablet.
  • The inject shader loops over all lights per froxel and evaluates shadows. The light ranges are now local, but the froxel shader still has no froxel-level light list.
  • gpu_volumetrics currently aggregates inject + temporal accumulate + composite; the exact split is unknown on Adreno.
  • gpu_light_culling is also 12.84 ms on Adreno vs ~1.0 ms on Mali, so the tiled-light path needs a net-benefit check there.

Next measurements

  1. Tablet with volumetrics disabled (keep everything else): establishes the non-volumetric frame floor.
  2. temporal_history: false: isolates the accumulate/reprojection cost.
  3. max_distance: 10/15/20/30 at VolumetricQuality::Low: shows how much of the cost is froxel depth resolution.
  4. A custom grid, e.g. Custom { grid_x: 128, grid_y: 80, grid_z: 64, composite_steps: 32 }, to separate grid size from light/shadow evaluation.
  5. 1280×800 diagnostic run to measure the pixel/froxel scaling (not a proposed fix).
  6. Tablet forced-full-loop vs tiled light culling to check whether the 12.84 ms culling pass is net-positive on Adreno.

If the results are still ambiguous, add sub-timestamps for volumetric inject, accumulate, and composite and expose them in the CSV/debug UI.

Acceptance criteria

  • The dominant volumetric component is identified on Adreno with measured A/B data.
  • The fix reduces the tablet’s volumetric cost without changing the phone baseline materially.
  • No automatic resolution/refresh changes; #1083 remains separate.
## Summary After the light-range fix (`b7086621`), the Adreno 810 tablet is no longer deferred-bound. It is volumetrics-bound: | Metric | Tablet (2560x1600) | |---|---:| | FPS | 4.47 | | frame | 219.5 ms | | `gpu_scene_total` | 38.11 ms | | `gpu_deferred` | 16.01 ms | | `gpu_light_culling` | 12.84 ms | | `gpu_volumetrics` | **161.78 ms** | | `gpu_tonemap` | 13.24 ms | | `gpu_aa` | 4.34 ms | `gpu_volumetrics` is ~74% of the frame time. The phone’s volumetric cost at 1220x2582 is only ~12 ms, so the tablet’s ~162 ms is ~13× higher despite only ~1.3× the pixel count. The tablet A/B against `5be0937d` confirms the range fix helped there too (deferred 76.43 → 16.02 ms, scene 100.13 → 38.40 ms, FPS 3.0 → 4.5), but the volumetric pass remains the new bottleneck. ## Likely contributors - The all-in-one scene uses `VolumetricQuality::Low` with `max_distance: 30`. Adaptive depth slicing raises the froxel grid to 192 slices, giving roughly 213×133×192 ≈ 5.4M froxels on the tablet. - The inject shader loops over all lights per froxel and evaluates shadows. The light ranges are now local, but the froxel shader still has no froxel-level light list. - `gpu_volumetrics` currently aggregates inject + temporal accumulate + composite; the exact split is unknown on Adreno. - `gpu_light_culling` is also 12.84 ms on Adreno vs ~1.0 ms on Mali, so the tiled-light path needs a net-benefit check there. ## Next measurements 1. Tablet with volumetrics disabled (keep everything else): establishes the non-volumetric frame floor. 2. `temporal_history: false`: isolates the accumulate/reprojection cost. 3. `max_distance: 10/15/20/30` at `VolumetricQuality::Low`: shows how much of the cost is froxel depth resolution. 4. A custom grid, e.g. `Custom { grid_x: 128, grid_y: 80, grid_z: 64, composite_steps: 32 }`, to separate grid size from light/shadow evaluation. 5. 1280×800 diagnostic run to measure the pixel/froxel scaling (not a proposed fix). 6. Tablet forced-full-loop vs tiled light culling to check whether the 12.84 ms culling pass is net-positive on Adreno. If the results are still ambiguous, add sub-timestamps for volumetric inject, accumulate, and composite and expose them in the CSV/debug UI. ## Acceptance criteria - The dominant volumetric component is identified on Adreno with measured A/B data. - The fix reduces the tablet’s volumetric cost without changing the phone baseline materially. - No automatic resolution/refresh changes; #1083 remains separate.
Author
Owner

Tablet volumetric/culling batch completed on b7086621 (Adreno 810, 2560x1600, 17 lights):

Variant gpu_volumetrics gpu_light_culling Scene FPS
baseline, tile 16 159.81 13.08 38.40 4.50
V0 volumetrics off 0.00 13.06 38.34 16.82
V1 history off 118.37 12.95 38.17 5.60
V2 max distance 10 62.79 13.07 38.34 8.19
V3 custom 128×80×64 34.45 13.08 38.36 10.62
V4 shadows off 147.96 12.78 33.09 4.94
V5 directional only 83.91 12.72 27.47 7.88
V6 800×1280 39.93 2.69 9.79 17.50
C1 full light loop 163.37 0.02 53.68 4.23
C2 tile 8 158.59 7.06 32.53 4.67
C2 tile 32 158.97 19.44 44.68 4.36

Order of levers: froxel grid/depth ≫ per-light inject loop > temporal history > shadowing. The 192-slice low-quality grid is the largest single cost; custom 128×80×64 saves 125 ms. Tiled culling is net-positive on Adreno, and tile 8 is better than 16/32 there.

Added volumetric sub-timestamps on feat/perf-timings in 7060c5bf:

gpu_volumetric_inject
gpu_volumetric_accumulate
gpu_volumetric_composite
gpu_volumetrics (aggregate)

Next: capture those three columns on the tablet, then choose a mobile volumetric grid/depth tier and evaluate TILE_SIZE = 8 as a default. No defaults were changed in this batch.

Tablet volumetric/culling batch completed on `b7086621` (Adreno 810, 2560x1600, 17 lights): | Variant | `gpu_volumetrics` | `gpu_light_culling` | Scene | FPS | |---|---:|---:|---:|---:| | baseline, tile 16 | 159.81 | 13.08 | 38.40 | 4.50 | | V0 volumetrics off | 0.00 | 13.06 | 38.34 | 16.82 | | V1 history off | 118.37 | 12.95 | 38.17 | 5.60 | | V2 max distance 10 | 62.79 | 13.07 | 38.34 | 8.19 | | V3 custom 128×80×64 | 34.45 | 13.08 | 38.36 | 10.62 | | V4 shadows off | 147.96 | 12.78 | 33.09 | 4.94 | | V5 directional only | 83.91 | 12.72 | 27.47 | 7.88 | | V6 800×1280 | 39.93 | 2.69 | 9.79 | 17.50 | | C1 full light loop | 163.37 | 0.02 | 53.68 | 4.23 | | C2 tile 8 | 158.59 | 7.06 | 32.53 | 4.67 | | C2 tile 32 | 158.97 | 19.44 | 44.68 | 4.36 | Order of levers: froxel grid/depth ≫ per-light inject loop > temporal history > shadowing. The 192-slice low-quality grid is the largest single cost; custom 128×80×64 saves 125 ms. Tiled culling is net-positive on Adreno, and tile 8 is better than 16/32 there. Added volumetric sub-timestamps on `feat/perf-timings` in `7060c5bf`: ```text gpu_volumetric_inject gpu_volumetric_accumulate gpu_volumetric_composite gpu_volumetrics (aggregate) ``` Next: capture those three columns on the tablet, then choose a mobile volumetric grid/depth tier and evaluate `TILE_SIZE = 8` as a default. No defaults were changed in this batch.
Author
Owner

Implemented the first mobile tier and Adreno-friendly tile size on feat/perf-timings:

3233c0b5 perf(renderer): add mobile volumetric tier and 8px light tiles

  • VolumetricQuality::Mobile: X/Y grid resolution / 20, fixed 64 depth slices, 32 composite steps. At 2560x1600 this is 128x80x64, matching the best measured custom variant.
  • Adaptive depth slicing is skipped for Mobile.
  • All-in-one example selects Mobile on Android and keeps Low on desktop.
  • TILE_SIZE is now 8. Tablet measurement was 7.06 ms culling / 32.53 ms scene versus 13.08 / 38.40 at 16; Mali was neutral-to-positive.

The test project’s copied AllInOne/src/lib.rs must be synced to the updated Examples/all-in-one/src/lib.rs for the Android Mobile tier to take effect. Otherwise it will still use Low.

Expected next capture:

  • Tablet: volumetrics near the V3 custom 128x80x64 result (~34 ms), scene near ~32.5 ms with tile 8, FPS materially above 4.5.
  • Phone: Mobile grid is much smaller than Low; verify visual quality and that the scene does not become too coarse.

After that capture, remaining tablet costs should be re-split with the new volumetric sub-timestamps. Composite (~26 ms) and per-light inject are likely the next candidates; history can follow.

Implemented the first mobile tier and Adreno-friendly tile size on `feat/perf-timings`: `3233c0b5 perf(renderer): add mobile volumetric tier and 8px light tiles` - `VolumetricQuality::Mobile`: X/Y grid `resolution / 20`, fixed 64 depth slices, 32 composite steps. At 2560x1600 this is 128x80x64, matching the best measured custom variant. - Adaptive depth slicing is skipped for Mobile. - All-in-one example selects Mobile on Android and keeps Low on desktop. - `TILE_SIZE` is now 8. Tablet measurement was 7.06 ms culling / 32.53 ms scene versus 13.08 / 38.40 at 16; Mali was neutral-to-positive. The test project’s copied `AllInOne/src/lib.rs` must be synced to the updated `Examples/all-in-one/src/lib.rs` for the Android Mobile tier to take effect. Otherwise it will still use Low. Expected next capture: - Tablet: volumetrics near the V3 custom 128x80x64 result (~34 ms), scene near ~32.5 ms with tile 8, FPS materially above 4.5. - Phone: Mobile grid is much smaller than Low; verify visual quality and that the scene does not become too coarse. After that capture, remaining tablet costs should be re-split with the new volumetric sub-timestamps. Composite (~26 ms) and per-light inject are likely the next candidates; history can follow.
Author
Owner

Mobile tier 3233c0b5 verified on both devices:

Metric Tablet Mobile Tablet Low (previous) Phone Mobile
FPS 11.27 4.47 30.41
frame 88.0 ms 224.0 ms —
gpu_volumetrics 34.71 ms 164.08 ms 9.84 ms
inject 11.10 92.67 sensor artifact*
accumulate 2.51 45.08 0.66
composite 21.10 26.33 9.16
gpu_light_culling 7.12 12.84 — (8 px tiles)
gpu_scene_total 32.67 38.11 —

*Mali inject timestamps remain ~0.02 ms; the aggregate and other stages are reliable.

Composite is now 61% of tablet volumetrics and ~93% of phone volumetrics. de1d9883 reduces the Mobile composite from 32 to 16 integration steps; the 128x80x64 grid and 64 depth slices are unchanged. Next capture should show composite near half its current cost; check for banding at the 16-step setting.

The 8 px tiles also produced a larger Mali win than expected: deferred 21.13 → 7.44 ms and phone FPS 19.94 → 30.41.

Mobile tier `3233c0b5` verified on both devices: | Metric | Tablet Mobile | Tablet Low (previous) | Phone Mobile | |---|---:|---:|---:| | FPS | 11.27 | 4.47 | 30.41 | | frame | 88.0 ms | 224.0 ms | — | | `gpu_volumetrics` | 34.71 ms | 164.08 ms | 9.84 ms | | inject | 11.10 | 92.67 | sensor artifact* | | accumulate | 2.51 | 45.08 | 0.66 | | composite | 21.10 | 26.33 | 9.16 | | `gpu_light_culling` | 7.12 | 12.84 | — (8 px tiles) | | `gpu_scene_total` | 32.67 | 38.11 | — | *Mali inject timestamps remain ~0.02 ms; the aggregate and other stages are reliable. Composite is now 61% of tablet volumetrics and ~93% of phone volumetrics. `de1d9883` reduces the Mobile composite from 32 to 16 integration steps; the 128x80x64 grid and 64 depth slices are unchanged. Next capture should show composite near half its current cost; check for banding at the 16-step setting. The 8 px tiles also produced a larger Mali win than expected: deferred 21.13 → 7.44 ms and phone FPS 19.94 → 30.41.
Author
Owner

16-step Mobile composite verified on both devices at de1d9883:

Metric Tablet 32 steps Tablet 16 steps Phone 32 steps Phone 16 steps
FPS 11.27 12.88 30.41 35.76
gpu_volumetrics 34.71 24.01 9.84 5.42
gpu_volumetric_composite 21.10 10.39 9.16 4.68
inject 11.10 11.08 timestamp artifact artifact
accumulate 2.51 2.54 0.66 0.72

No visible fog banding in the captured views; 16 steps is acceptable for the current scene. Tablet is now 12.88 FPS / 77.65 ms against a 16.8 FPS / 59.3 ms non-volumetric floor.

The remaining tablet volumetric cost is inject ~11.1 ms, composite ~10.4 ms, accumulate ~2.5 ms. Composite is no longer the 21 ms bottleneck. This closes the Adreno volumetric tier issue; further inject work is optional and has diminishing headroom against the non-volumetric floor.

16-step Mobile composite verified on both devices at `de1d9883`: | Metric | Tablet 32 steps | Tablet 16 steps | Phone 32 steps | Phone 16 steps | |---|---:|---:|---:|---:| | FPS | 11.27 | **12.88** | 30.41 | **35.76** | | `gpu_volumetrics` | 34.71 | **24.01** | 9.84 | **5.42** | | `gpu_volumetric_composite` | 21.10 | **10.39** | 9.16 | **4.68** | | inject | 11.10 | 11.08 | timestamp artifact | artifact | | accumulate | 2.51 | 2.54 | 0.66 | 0.72 | No visible fog banding in the captured views; 16 steps is acceptable for the current scene. Tablet is now 12.88 FPS / 77.65 ms against a 16.8 FPS / 59.3 ms non-volumetric floor. The remaining tablet volumetric cost is inject ~11.1 ms, composite ~10.4 ms, accumulate ~2.5 ms. Composite is no longer the 21 ms bottleneck. This closes the Adreno volumetric tier issue; further inject work is optional and has diminishing headroom against the non-volumetric floor.
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#1089
No description provided.