-
Notifications
You must be signed in to change notification settings - Fork 34
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
ARM iGPU Mali-G52 Failures #736
Comments
@franz Is there a single reason why most of these fail? |
IIRC there were multiple reasons. The major one was that the Mali does not support FP64 and there is no emulation (AFAIK). Some of the tests have "int" or "unsigned" in the name, but they have shared kernel code for all variants including "double", and because of that, they driver failed to compile the SPIRV. Another reason was that some tests required more memory than the driver could allocate, this is perhaps fixable with some kernel parameter tuning, i haven't tried. Then there are a bunch of memset/memfill tests that fail, i have no idea why, possibly bugs in Mali drivers. |
This we can filter out now with IIRC only a couple of tests failed due to this. |
Updated list of failing tests:
Will work on root-causing these in for 1.3 |
For non-timeout failures:
|
Conclusions:
|
Does it support the KHR subgroup shuffles? Did you notice in devicelib.cl:
If it supports the KHR one, we could detect at build time that we are building on non-Intel and use that by the default. Ideally, this would be selected at .spv linkage time in chipStar. |
Seems like it is supported. |
So seems like this is one of the issues but there are more - I haven't investigated every single SPIR-V. |
The text was updated successfully, but these errors were encountered: