Additional validation of texture copies #7935
Conversation
* Copies must not overlap. * Copies of multisampled or depth/stencil formats must span the entire texture. * Move no-op for zero-size copies after parameter validation. Closes gfx-rs#2951 Fixes gfx-rs#7844
andyleiserson
force-pushed
the
texture-copy-validation
branch
from
0b8f788 to
fa7961b
Compare
Sounds like we need to add more CTS tests. |
wgpu-core/src/command/transfer.rs
Outdated
| src_texture.desc.format, | ||
| destination, | ||
| dst_texture.desc.format, |
There was a problem hiding this comment.
Nit: The format of the textures is the same since src = dst.
| /// `validate_texture_copy_range`. | ||
| /// | ||
| /// [srtc]: https://gpuweb.github.io/gpuweb/#abstract-opdef-set-of-subresources-for-texture-copy | ||
| pub(crate) fn validate_copy_within_same_texture<T>( |
There was a problem hiding this comment.
This fn is not allowing disjoint mip levels but should, per spec.
There was a problem hiding this comment.
Thanks, fixed. I'll also open a PR to add this to the CTS (it was noted as a TODO, but not implemented).
wgpu-core/src/command/transfer.rs
Outdated
| buffer_layout: source.layout, | ||
| size: hal_copy_size, | ||
| }, | ||
| ] |
There was a problem hiding this comment.
This case shouldn't be allowed by validation (same for copyTextureToBuffer and writeTexture).
If texture.format is a depth-or-stencil format format:
texelCopyTextureInfo.aspect must refer to a single aspect of texture.format.
There was a problem hiding this comment.
Good point, we only need this case for copyTextureToTexture.
andyleiserson
force-pushed
the
texture-copy-validation
branch
from
fa7961b to
04cd49f
Compare
In studying whether #2951 is still an issue, I identified some other texture validation that is missing. This adds it, and also fixes the issue where we skip parameter validation for zero-size copies.
I suspect that this may fix #7844 by rejecting an illegal operation before dx12 panics.I've also included a fix for #7844 that splits depth+stencil copies into two operations, although I'm not sure if this is the right fix -- it appears that metal does support a combined copy in copyTextureToTexture (but not in copyBufferToTexture), so maybe this should be pushed to the backends.Testing
Enables all the
copyTextureToTextureCTS tests. However, the CTS isn't complete coverage of the validation logic I added. In particular, it doesn't cover the logic that allows copies among different aspects of the same texture. I'm not sure if that even makes sense (copying between depth and stencil?), or if it's just the spec being thorough, but I would welcome a suggestion of an interesting test to add here.Squash or Rebase? Squash
Checklist
cargo fmt.taplo format.cargo clippy --tests. If applicable, add:--target wasm32-unknown-unknowncargo xtask testto run tests.CHANGELOG.mdentry. Maybe it's worth one overall changelog entry mentioning this, the vertex/index buffer offset change, and any other added validation that gets added.