Upscaling Texture corruption on door transitions

[Triticum0]

Happens when shooting the door during loading transitions.

Recomp.door.Issue.4x.mp4

[MysteriousAeon]

Thanks for the report. Before I build an instrumented version, could you try one small controlled comparison at the same door?

Please restart the game between the two tests and use:

Test A:

• Internal Resolution: 1x
• Texture Upscale: 1x
• Adaptive Widescreen: Off

Test B:

• Internal Resolution: 1x
• Texture Upscale: 4x
• Adaptive Widescreen: Off

Then reproduce the same door transition and shot timing in both.

The important result is whether the corruption happens:

1. only with Texture Upscale 4x
2. with both settings
3. with neither

Adaptive Widescreen should stay off for this test because there is a separate loading-zone / visibility issue being tracked in #18 and #20.

If convenient, a short capture of both runs would be very useful.

[Triticum0]

Test A:

https://github.com/user-attachments/assets/8b47589a-fbca-48f0-8e62-f0c97e42ed10

Test B:

https://github.com/user-attachments/assets/39517c50-13f5-494e-8b21-b8b734acba86

Seems to only occur with Internal Resolution and Texture Upscale are both up-scaled as you can see

Seem like Up-scaling doesn't work with the adaptive widescreen turned off so file new issue.

Also got it briefly on melonds but not when opening the door wasn't recording so no capture as too breif so hard to confirm. Also couldn't easy reproduce it.

[Triticum0]

Might be related melonDS-emu/melonDS#2683

[MysteriousAeon]

Thanks for the melonDS link. I checked that path against ndsrecomp, and both compute-renderer issues described in melonDS #2683 are still present in the vendored renderer.

The strongest current suspect is the per-variant indirect compute dispatch exceeding NVIDIA's 65,535 work-group Z limit at high internal resolutions. The MaxWorkTiles overflow described there is also still present and may be a secondary issue.

Your earlier comparison with Adaptive Widescreen off is unfortunately not a valid HD isolation because of #22, so could you try one smaller test with Adaptive Widescreen kept ON?

Use the same doorway and shot timing for both runs:

Test A:

• Adaptive Widescreen: On
• HD Rendering: On
• Internal Resolution: 1x
• Texture Upscale: 1x

Test B:

• Adaptive Widescreen: On
• HD Rendering: On
• Internal Resolution: 4x
• Texture Upscale: 1x

Please restart the game between runs.

The key result is whether the corruption appears only at 4x internal resolution.

If it does, I'll prepare a controlled Windows A/B comparing:

1. the current renderer
2. the renderer with only the dispatch-limit fix
3. the renderer with both #2683 fixes

[Triticum0]

Here you go

Test A:

1x.mp4

Test B:

4x.mp4

Used macro so actions are identical but small placement differences before starting and random sway made hard to perfect matching up best I could easily do, but in my previous testing (2/3 hour) haven't produced it at 1x.

4x is where macro was copied from.

Also do u want test 2x and 3x as well?

[MysteriousAeon]

I have a controlled three-arm test build for the high-IR corruption.

Test package:
https://github.com/MysteriousAeon/ndsrecomp/releases/download/intel-gl-diagnostic-1/mph-issue19-2683-ab-win64.zip

SHA-256:
b747a014135303621af9160ed59c6a7bdb9aaa836f81cf85d286285453c1ca29

Please place your USA Rev 0 ROM beside README.txt as:

Metroid Prime Hunters.nds

Then run these in order, restarting between each run:

1. current\run-current.bat
2. dispatch-fix\run-dispatch-fix.bat
3. both-fixes\run-both-fixes.bat

Please use the same scene, camera, and door-shot macro timing for all three.

You only need to report whether the corruption occurs in A, B, and C, plus whether C shows any missing geometry or layers.

No need to test 2x or 3x.

[mstan]

v0.4.2-alpha is live now, but it does not include the renderer experiments for this high-internal-resolution corruption:

https://github.com/mstan/MetroidPrimeHuntersRecomp/releases/tag/v0.4.2-alpha

The useful next step here is still the three-arm A/B package from my previous comment:

1. current\run-current.bat
2. dispatch-fix\run-dispatch-fix.bat
3. both-fixes\run-both-fixes.bat

Please use the same doorway / macro timing for all three and just report whether corruption happens in A, B, and C. No need to test 2x or 3x right now.

[Triticum0]

Did a quick test and 2 and 3 both seem to fix issue as no longer able to easily reproduce the issue. Where as current build was reproducible.
Can't tell u it 100% fix as issue still effects melonds. Did have time for video comparisons unfortunately but on another integrated AMD system I did confirm this only effect NVIDIA cards in my testing. 3060 ti and 7730u if your wondering.

Note: Didn't test death effects as I unable to reproduce the issue once during my extensive testing previously.

[mstan]

The fact it can't reproduce readily is a good sign.

https://github.com/mstan/MetroidPrimeHuntersRecomp/releases/tag/v0.4.11-alpha

I asked you in another issue to check this build out. If you continue to not reproduce here, I may consider we close this issue out for now. You can always re-open if you're able to reproduce the behavior.

[Triticum0]

Still an issue only currently effect nivida cards