Conversation
This reverts commit a6503e5.
|
Instead of rolling back the whole MR, can we just fix the failure? It's just some validation that's acting up |
|
Or more specifically, I don't quite understand why this is failing now but didn't fail inside the PR |
|
Most likely due to updated validation layers. Interesting timing though. |
|
Validation layers haven't updated since mid october |
|
Indeed, CI logs show the same versions being installed. Idk why the error didn't show up previously then, odd. But I think it's correct nonetheless. |
|
Heh, it was a mesa update. We went from 23.2.1 to 23.3.0 and lavapipe learned how to make NV12 textures 😆 |
|
Ah, that explains it :) |
|
I'm going to close this, I don't want to revert the whole PR. We should either:
|
|
@teoxoy I can successfully create nv12 texture and r8/rg8 views on windows with vulkan backend. I think this is an implementation issue with llvmpipe. In addition, we must add the plane view formats to the format list, otherwise the corresponding view cannot be created. I consider disabling the nv12 feature for llvmpipe. |
I don't think this is the case, we don't have to add the individual aspect formats for combined depth-stencil formats either. I added references to the Vulkan spec in #4573 (comment). If things work on one Vulkan implementation despite the Vulkan spec saying otherwise that doesn't mean it's correct. |
|
We have to set References:
from https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VkImageViewCreateInfo.html
from https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VkImageCreateFlagBits.html
from https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VkImageCreateInfo.html |
|
We can't add image format list at all, including the original format: |
|
Yeah, that shouldn't be necessary for the NV12 format; |
I don't think this is reasonable, because |
|
That's not consistent with combined depth-stencil formats, users of the API don't have to pass the aspect formats to
We can enforce this without users having to pass the plane formats to |
|
Okay, it seems better to have plane as an aspect. |
Reverts #4573 due to CI failures that make it difficult to merge other PRs.
See also #4573 (comment).