[Bug 105496] Using a single-plane imageview from a multi-plane image is broken

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=105496

--- Comment #14 from Lionel Landwerlin <lionel.g.landwerlin at linux.intel.com> ---
(In reply to atomnuker from comment #13)
> (In reply to Lionel Landwerlin from comment #12)
> > (In reply to atomnuker from comment #11)
> > > (In reply to Lionel Landwerlin from comment #10)
> > > > (In reply to atomnuker from comment #9)
> > > > > (In reply to Lionel Landwerlin from comment #8)
> > > > > > 
> > > > > > We don't support storage on multiplanar formats indeed.
> > > > > > 
> > > > > > But R8_UNORM should be reported as supporting storage.
> > > > > > 
> > > > > 
> > > > > The latter (and RGBA) does indeed but since you're attaching the original
> > > > > multiplane image to the descriptor (albeit indirectly though a derived
> > > > > imageview) its that VkFormat that needs to have STORAGE bit supported. It
> > > > > already supports the SAMPLE bit (needed for YUV->RGB converting samplers) so
> > > > > I guess its just a matter of flagging STORAGE as well.
> > > > 
> > > > It's more or less luck that it works for this particular couple of formats.
> > > > For something like G8B8G8R8_422_UNORM the hardware really can't do it.
> > > > So I don't think we want to enable storage on anything ycbcr.
> > > Yes you can. Its listed right there in the compatible formats. Check out the
> > > earlier link I posted. 422 images support VkImageView of R8 for the luma
> > > plane and R8G8 for the chroma planes (so you get vec2() for the chroma plane
> > > of such images). In my use case I need to be able to do this for all YUV
> > > images, since otherwise they'd be pretty much useless. And you can do that
> > > with all YUV images like it says in the spec.
> 
> I should correct myself, this is for NV12. For 422 you use the same
> procedure as 420 but with the image resolution being halved horizontally.
> 
> > > 
> > > 
> > > > 
> > > > Rereading the shader from your repo : 
> > > > 
> > > > #version 460
> > > > 
> > > > layout (local_size_x = 16, local_size_y = 16) in;
> > > > layout (set = 0, binding = 0, rgba8) uniform readonly image2D input_img;
> > > > layout (set = 0, binding = 1, rgba8) uniform writeonly image2D output_img;
> > > > 
> > > > void main()
> > > > {
> > > >     vec4 result = imageLoad(input_img, ivec2(gl_GlobalInvocationID.xy));
> > > >     imageStore(output_img, ivec2(gl_GlobalInvocationID.xy), result);
> > 
> > If you swizzle the result here, I'm pretty sure you won't get what you
> > expect.
> > Hopefully that convinces you that this isn't supported.
> 
> Well of course it won't work, this copies 1 plane only. For multiplane
> images I make both the output images and input images arrays and I iterate
> over them. The result vec4 only contains 1 pixel value in the .r member.
> Like the spec says, creating a single plane imageview from a multiplane
> format exposes 1 plane and 1 plane only.

The ffmpeg repo I built locally had the image below bound to a G8_B8_R8 image.

layout (set = 0, binding = 0, rgba8) uniform readonly image2D input_img;

But somehow it should contain all the planes?

> 
> I have pushed a new version to the same repository which works for 420, 422
> and 444 images. Tested. It also works for NV12 images with minimal
> modifications (iterating over 2 planes instead of 3 and with different
> format on the second plane).
> 
> If you have nothing more to argue about then I propose we end this
> conversation and instead focus on fixing the flags.

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-3d-bugs/attachments/20180316/17b8c7e5/attachment-0001.html>