Hey there, I'm currently facing an issue testing my code on a RadeonVII. My Adrenaline version is 22.5.2
Basically, there are certain shader computations that are wrongly memoized/cached at workgroup level (8x8 pixel blocks). (The application uses a Vulkan 1.1 context.) The following shader snippet is used during runtime to check whether or not a certain combined 2D_ARRAY sampler has more than 1 layers. If so, it is considered a 'terrain texture'.
For greater context, the combined sampler is fetched out of a variable count descriptor set (depending on an instance ID associated with a fragment):
Please note that InstID is evaluated per fragment (this is all in a fragment shader -- attached at the bottom of the post):
However, this value is seemingly only evaluated and cached per 8x8 block resulting in artifacts at runtime:
This is the expected output (using a temporary workaround):
I created a short-circuit in shader code to try and isolate the issue using the following snippet (outputs white for terrain, black for all else):
The issue is visible more clearly. Here is the wrong output:
And here is the expected output (with a workaround):
Please note that I hope to not rely on the workaround (which is to check if alpha on layer 1 is zero) as something very similar to this issue pops up in other passes that I have not covered in this post. It oddly seems pervasive.(EDIT: Fixed this separate issue via using nonuniformEXT() in all necessary places). Additionally, I do not see issues like this when testing the build on any of my nVidia cards (2080 Ti, 2060 OC or the 1050Ti).
I've uploaded the above test builds isolating the issue:
(Release build): http://www.toomuchvoltage.com/highomega_boolean_wgp_issue.zip
(Debug build): http://www.toomuchvoltage.com/highomega_boolean_wgp_issue_debugbin.zip
I'm also attaching the shader code creating the above for closer examination as well:
http://www.toomuchvoltage.com/gatherresolvesimple.frag
Your examination of this is highly appreciated.
With much thanks,
Baktash.
In addition, I would like to post the other instance where such a caching issue seems to appear. This is basically from another pass that is supposed to combine irradiance caching with other values. I've once again short-circuited the shader to highlight the issue:
Incorrect output (RadeonVII) (note the distant staircase artifacts highlighted in red):
Expected output (rendered on a GTX 1050Ti):
I have not been able to fully isolate the origin of this artifact. However, the full shader code can be found here:http://www.toomuchvoltage.com/modulate.fragI can provide a separate build for this upon request.Thanks again,Baktash.
EDIT: This issue has been resolved. Explicitly sized sampler arrays were accessed without nonuniformEXT(). The original issue still persists however.
Hey everyone, I believe we now have a greater degree of certainty regarding this being a driver issue.
The Radeon GPU Analyzer produced this compiled shader code using a simplified version of the fragment shader provided.
https://shader-playground.timjones.io/ddb98c8f6a51b16e3cd89ea3fa8b6725
It is apparent that nonuniformEXT() is not taking effect on the textureSize() call thus resulting in undefined behavior on Radeon drivers.
More specifically:
As can be witnessed from above, the image_get_resinfo instruction reads out of register v0 to determine image dimensions. However, v0 itself is loaded only from the first lane (which is possibly shared among the entire workgroup). Consulting the following article, it appears that nonuniformEXT() is simply not taking effect:
https://gpuopen.com/learn/porting-detroit-2/
