Image x/y extend, alpha, nearest neighbor sampling

[dfrg]

This adds support for image extend modes, alpha and nearest neighbor sampling to fine and plumbs support for all through the pipeline.

Extend modes with bilinear filtering (bottom right is x repeat and y reflect):

Extend modes with nearest neighbor filtering:

Image alpha:

[waywardmonkeys]

LGTM. I don't know if we need to quantize the alpha down to 8 bits and store it with the other stuff, but okay for now.

[DJMcNab]

I think this should have a changelog entry.

I'd also like to see some new tests here; these should be very similar to the existing tests, and so should be straightforward. Indeed, one of the tests needs to be updated.

Otherwise, only relatively peripheral concerns.

[dfrg]

Thanks for the feedback, Daniel. I'm on vacation and out of town for the next week but I may have some time over the weekend to address all concerns and get this landed.

[waywardmonkeys]

@dfrg I took the liberty of updating this branch to fix the merge conflicts and a couple of other things. Feel free to force push and remove my changes if you have others locally that you prefer.

[waywardmonkeys]

Odd that the nearest neighbor snapshot that I generated on my machine is slightly different than the one generated in CI.

[dfrg]

@dfrg I took the liberty of updating this branch to fix the merge conflicts and a couple of other things. Feel free to force push and remove my changes if you have others locally that you prefer.

Not a problem. Your changes look good. Thanks for pushing this forward.

[dfrg]

Odd that the nearest neighbor snapshot that I generated on my machine is slightly different than the one generated in CI.

I’ve also noticed inconsistent failures between local tests and CI. I’m not sure how to resolve this properly but I’ll play with this one and see if I can get it to pass as this PR would be nice to land for 0.4.

[dfrg]

The border on the left side of the two right images looks off to me. Which platform/GPU did you use to generate these?

[waywardmonkeys]

M3 Pro Max laptop, macOS 15.2.

[dfrg]

Thanks, I’ll test on my M3 in a bit and see if I can reproduce.

[dfrg]

So there’s actually a fairly thick border around the entire image and this appears on the previous version as well. Note that this only seems to occur when generating snapshots on macOS. The winit demo is fine on macOS and the snapshots appear as they should on Windows.

I don’t have a lot of time to continue investigating this right now but I’ll come back to it later this week.

[dfrg]

Additional data point: on my win11/rtx3070 machine, the COLR snapshot tests fail locally on main even though they pass in CI.

[DJMcNab]

Discussion in office hours yesterday suggests that we should just skip the testing for now.

I don't have a Mac to dig into this further myself. I understand that this does occur consistently on macOS (and so that with_winit doesn't in itself fix it) - Bruce has done some testing which suggests this.
It's only seems to be a one-pixel failure, which won't be a massive issue at realistic screen sizes.
We certainly should track this in an issue, though (and ideally even fix it).

[nicoburns]

Given that this is new functionality (and thus not a regression), I agree that shipping this with a known bug is reasonable. It's the sort of the thing that could likely be fixed in a patch release (if we can work out what's wrong (and it doesn't require a wgpu upgrade I suppose)).

[dfrg]

I suspect this might be a very slight difference in path rendering leading to over-extension of the image. I’ll do some tests to confirm this but in the meantime, I’m fine with disabling these failing tests and merging.