WIP: Fix stale fixed-resolution scene scale (experimental opt-in) #329

Closed

i900gaming wants to merge 1 commit from fix/fixed-resolution-scene-scale into main

i900gaming commented

2026-11-07 21:04:21 +00:00

(Migrated from github.com)

Fixed Resolution 100% can leave the scene rendering at its previous dynamic scale. This draft adds an off-by-default correction, enabled with HOS_FIX_RES=1, that invalidates the cached scene scale when the fixed request changes. In the tested scene, the main color/depth targets increased from 1536x864 to 2048x1152 at 100%.

This is a draft with unresolved image corruption at 67%/50%. It is not ready to enable by default.

Cause and change

The controller's fixed path can return a resource-refresh status without marking the scene's cached scale dirty. Instrumentation observed a requested scale of 1.5 while the cached scene value remained 0.75 and the dirty flags were 0.

nx2/oot_patches.py now injects a check at the scene update entry. In fixed mode, with the opt-in enabled, it compares the requested scale in the scene's reference units with the cached value and sets the existing scale-dirty bit (8) when they differ. The game's existing update code performs the recalculation. Other dirty bits are preserved; unchanged values do not continuously invalidate the scene. Auto and the existing initialization warm-up from #102 are retained.

The patch includes documentation, a game-free compiled regression test, and a source-regression workflow step. It contains no renderer/filter changes, game assets, generated game code, binaries, saves, or dumps.

Reproducing the opt-in

Re-lift/rebuild the game executable with this source change; replacing only the Vulkan DLL is insufficient.
Add HOS_FIX_RES=1 to hookshot_env.txt beside the executable and restart.
Select Resolution 100% and compare the main 3D target dimensions with the same scene without the opt-in. The presentation/window size alone is not evidence of the internal resolution.
Remove the variable or set it to 0 and restart to restore the previous behavior.

Known problems and limits

67% and 50% produce image noise in the user's test. Auto, 75% and 100% were reported usable. The lower-scale corruption has not been reproduced in a controlled opt-in-off/on comparison, so it is not yet known whether this patch introduces it or exposes an existing problem. The state-transition tests do not establish image correctness at those scales.
A GPU timeout occurred during the investigation (Windows LiveKernelEvent 141). It happened in an earlier diagnostic run, before this correction was executed, with HOS_RES_SCALE=1.5, frame generation and a separate experimental 16x anisotropy override active. The responsible component is unknown. Anisotropy was subsequently disabled; short follow-up tests with the correction, including 2x frame generation, recorded no further timeout. This is not a long-term stability claim, and the timeout is not considered resolved.
This does not provide native 1440p/4K rendering. The validated main 3D targets are 2048x1152, although presentation was 2560x1440. A separate experiment at 125%, including bypassing another engine cap, did not establish larger 3D targets; that experiment is excluded here. Different postprocessing passes may still use smaller targets.
Automated live menu switching was inconclusive because input overlapped with user interaction. Booting at 100% was checked; a complete runtime transition matrix remains unverified.

Validation and coverage

Based on upstream 40c3f851b07a4367a13d7d7bcfb1d5b91cc95dfe. Its nx2/oot_patches.py matches the starting version used for the local fix.
python -I -B tests/test_resolution_scale.py --cxx <clang++> passes on Windows. It compiles the injected logic with memory stubs and checks changed/unchanged scale, preservation of other dirty bits, Auto, opt-in absent/0/1, invalid requested values, reference-size normalization, patch idempotence and layout guards.
Negative control: removing the scale-dirty bit makes the enabled-case regression check fail.
python -I -B tests/test_make_game_cache.py: all six checks pass.
The local executable containing the correction was built and tested with copied saves, in the house and outdoor Kokiri area. Main color/depth target dimensions were checked in renderer logs; 2x frame generation and the FPS overlay still operated. This was an existing Windows build with local renderer fixes, not a fresh end-to-end build of this PR's current-main tree.
Runtime coverage: Windows, Radeon RX 6900 XT, driver 32.0.21045.16010. This identifies the tested configuration; the scene-state correction is not claimed to be AMD-specific. Other GPUs/platforms have not been runtime-tested.

Before this can leave draft

Reproduce and investigate 67%/50% noise against the opt-in-off baseline.
Build and runtime-test the complete current-main tree with this patch.
Check Auto/fixed transitions and scene changes systematically.
Investigate the timeout and run longer stability tests; obtain additional hardware coverage.

Fixed Resolution 100% can leave the scene rendering at its previous dynamic scale. This draft adds an **off-by-default correction**, enabled with `HOS_FIX_RES=1`, that invalidates the cached scene scale when the fixed request changes. In the tested scene, the main color/depth targets increased from **1536x864 to 2048x1152** at 100%. **This is a draft with unresolved image corruption at 67%/50%. It is not ready to enable by default.** ## Cause and change The controller's fixed path can return a resource-refresh status without marking the scene's cached scale dirty. Instrumentation observed a requested scale of `1.5` while the cached scene value remained `0.75` and the dirty flags were `0`. `nx2/oot_patches.py` now injects a check at the scene update entry. In fixed mode, with the opt-in enabled, it compares the requested scale in the scene's reference units with the cached value and sets the existing scale-dirty bit (`8`) when they differ. The game's existing update code performs the recalculation. Other dirty bits are preserved; unchanged values do not continuously invalidate the scene. Auto and the existing initialization warm-up from #102 are retained. The patch includes documentation, a game-free compiled regression test, and a source-regression workflow step. It contains no renderer/filter changes, game assets, generated game code, binaries, saves, or dumps. ## Reproducing the opt-in 1. Re-lift/rebuild the game executable with this source change; replacing only the Vulkan DLL is insufficient. 2. Add `HOS_FIX_RES=1` to `hookshot_env.txt` beside the executable and restart. 3. Select Resolution 100% and compare the main 3D target dimensions with the same scene without the opt-in. The presentation/window size alone is not evidence of the internal resolution. 4. Remove the variable or set it to `0` and restart to restore the previous behavior. ## Known problems and limits - **67% and 50% produce image noise in the user's test.** Auto, 75% and 100% were reported usable. The lower-scale corruption has not been reproduced in a controlled opt-in-off/on comparison, so it is not yet known whether this patch introduces it or exposes an existing problem. The state-transition tests do not establish image correctness at those scales. - **A GPU timeout occurred during the investigation** (Windows LiveKernelEvent 141). It happened in an earlier diagnostic run, before this correction was executed, with `HOS_RES_SCALE=1.5`, frame generation and a separate experimental 16x anisotropy override active. The responsible component is unknown. Anisotropy was subsequently disabled; short follow-up tests with the correction, including 2x frame generation, recorded no further timeout. This is not a long-term stability claim, and the timeout is not considered resolved. - **This does not provide native 1440p/4K rendering.** The validated main 3D targets are 2048x1152, although presentation was 2560x1440. A separate experiment at 125%, including bypassing another engine cap, did not establish larger 3D targets; that experiment is excluded here. Different postprocessing passes may still use smaller targets. - Automated live menu switching was inconclusive because input overlapped with user interaction. Booting at 100% was checked; a complete runtime transition matrix remains unverified. ## Validation and coverage - Based on upstream `40c3f851b07a4367a13d7d7bcfb1d5b91cc95dfe`. Its `nx2/oot_patches.py` matches the starting version used for the local fix. - `python -I -B tests/test_resolution_scale.py --cxx <clang++>` passes on Windows. It compiles the injected logic with memory stubs and checks changed/unchanged scale, preservation of other dirty bits, Auto, opt-in absent/0/1, invalid requested values, reference-size normalization, patch idempotence and layout guards. - Negative control: removing the scale-dirty bit makes the enabled-case regression check fail. - `python -I -B tests/test_make_game_cache.py`: all six checks pass. - The local executable containing the correction was built and tested with copied saves, in the house and outdoor Kokiri area. Main color/depth target dimensions were checked in renderer logs; 2x frame generation and the FPS overlay still operated. This was an existing Windows build with local renderer fixes, **not a fresh end-to-end build of this PR's current-main tree**. - Runtime coverage: Windows, Radeon RX 6900 XT, driver `32.0.21045.16010`. This identifies the tested configuration; the scene-state correction is not claimed to be AMD-specific. Other GPUs/platforms have not been runtime-tested. ## Before this can leave draft - [ ] Reproduce and investigate 67%/50% noise against the opt-in-off baseline. - [ ] Build and runtime-test the complete current-main tree with this patch. - [ ] Check Auto/fixed transitions and scene changes systematically. - [ ] Investigate the timeout and run longer stability tests; obtain additional hardware coverage.

NotRexed commented

2026-11-09 05:29:06 +00:00

(Migrated from github.com)

Thank you! #849 (merged) now makes a fixed Resolution reach the 3D, the same stale fixed-scale problem this fixed, so I'm closing this as superseded. Closing to keep the queue clear; if something is still wrong at 67%/50% on current main, a new PR on top of it is welcome.

Commenting is not possible because the repository is archived.

No reviewers