DX12: Align copies b/w textures and buffers when D3D12_FEATURE_DATA_D3D12_OPTIONS13.UnrestrictedBufferTextureCopyPitchSupported is false

[ErichDonGubler]

Connections

• Likely resolves [d3d12] Panic in end_encoding  #5285.
• A case we need to handle for [d3d12] Support padded rows_per_image for 3D texture writes #4150.

Testing

• @teoxoy and I proved on some custom DLLs copied from an old Windows machine that this works, both on older WARP adapters and when forcefully done in updated DX12 runtimes. So, we're definitely not breaking anything! It's hard to set up an environment to specifically test this, though.

Squash or Rebase?

rebase plz

Checklist

• If this contains user-facing changes, add a CHANGELOG.md entry.

[ErichDonGubler]

Got the CI issue resolved! I was definitely doing something wrong. 👉🏻👈🏻 From community chat:

@ErichDonGubler:

I just paired with Teo, and we actually FIGGERED IT OUT!

@teoxoy and I sideloaded a bunch of DLLs from an ancient Windows 10 computer I had, which happened to have a WARP adapter that did not support low-alignment buffer-to-texture copies.

...turns out I accidentally swapped the alignment requirement branches. 😅

We had to work around:

• Not having the right DLLs (because new versions have lower alignment needs) and stealing them from an old machine I happened to have around.
• Crashing when trying to compile indirect buffer validation shaders (because old versions?)
• Feature unification working against us when trying to use WGPU_VALIDATION_INDIRECT_CALL=0
• Erich being dumb and necessitating debugging in the first place

It was an hours-long adventure, lemmetellyawhat.

@teoxoy:

It was fun :)

@cwfitzgerald:

oh boy!

that sounds like an adventure indeed

[teoxoy]

Looks good overall but we should also enable some CTS tests like the one in #5285.

[ErichDonGubler]

Doing some branch pushes to check that webgpu:api,operation,command_buffer,image_copy:undefined_params:* fails before and after, and hopefully that it catches the bug @teoxoy found with region offsets being swapped on the tex-to-buf path. 🤞🏻

[ErichDonGubler]

Alright, so:

• https://github.com/erichdongubler-mozilla/wgpu/actions/runs/18361970647/job/52307306029 is based on trunk, and has 4 failures, including some in 2d copies and 3d copies.
• https://github.com/erichdongubler-mozilla/wgpu/actions/runs/18361971738/job/52307309175 is based on this PR's contents before the fix Teo identified for offsets in tex-to-buf. There are still 2d failures here, but fewer.
• https://github.com/gfx-rs/wgpu/actions/runs/18362103668/job/52307694538?pr=7721 implements Teo's fix, and confirms that things get greener. Yay! Unfortunately, there are still breakages with 3d copies, so this whole thing isn't passing yet, but I'd say this is a nice confirmation that we're fixing something.

I'll figure out the exact set of CTS tests to use tomorrow. If nothing else, I can use a (verbosely requested) subset of webgpu:api,operation,command_buffer,image_copy:undefined_params:*.

[ErichDonGubler]

I added baseline expectations (plus failures) for webgpu:api,operation,command_buffer,image_copy:undefined_params:* to our CTS list with 932567d. I have amended the current tip of the PR to remove fails-if(dx12) from some of the 2d variants of the added tests, thus proving that we fix something with this PR. Nice!

[ErichDonGubler]

It's odd that webgpu:api,operation,command_buffer,image_copy:undefined_params:initMethod="WriteTexture";checkMethod="PartialCopyT2B";dimension="2d" still fails after this PR's fix. I suspect there's more work to do here to make this correct.

[ErichDonGubler]

Ah, it seems the full path that's failing is a single subtest exercising a 3rd dimension for the "2d" case: copySize=[3,3,2];origin=["_undef_","_undef_","_undef_"];bytesPerRow=256;rowsPerImage=3

Definitely follow-up work for something more edge-casey. Whew!

[ErichDonGubler]

onuts, I almost forgot to move this out of the previous release section.

@cwfitzgerald: I'm strongly considering experimenting with adding some CI that fails a PR that adds things to the CHANGELOG outside of the Unreleased section, unless opted out by a label. Thoughts?

[ErichDonGubler]

I guess we don't have to make it a required check, either. It can simply be a signal that something might be wrong, at first.

[torokati44]

Oh, is a backport of this to the 25/26/27 series, and corresponding patch releases feasible? 👀 It certainly would be appreciated IMHO!

[ErichDonGubler]

@torokati44: Two things:

1. For optional fixes like this, we generally expect the community to handle backporting effort (which should be light here, since you can omit the test coverage). So, if you're willing, then we look forward to a PR from you against v26 and v27. 😄
2. In our maintainers' meeting today, we know we're willing to field backports to v26 (esp. for Bevy) and v27. Was there a reason you wanted to backport to v25, specifically?

[torokati44]

@ErichDonGubler:

1. Alright, I'll see what I can do w.r.t backporting PRs.

2. No, not really, anymore, since chore: Update wgpu to 26, egui to match ruffle-rs/ruffle#21730 got merged (but not yet chore: Update wgpu to 27, egui to 0.33 ruffle-rs/ruffle#21895)