clvk icon indicating copy to clipboard operation
clvk copied to clipboard

Update deps and CTS

Open kpet opened this issue 1 year ago • 12 comments

Here are a few regressions I'm observing with this update:

AMD/Intel/llvmpipe (logical and physical addressing)

        Half Ops
                Expected the return code to be 0 but got 1
                Expected the status of 'vloada_half' to be 'pass' but got 'fail'
                Expected the status of 'vstorea_half' to be 'pass' but got 'fail'
                Expected the status of 'vstorea_half_rte' to be 'pass' but got 'fail'
                Expected the status of 'vstorea_half_rtz' to be 'pass' but got 'fail'
                Expected the status of 'vstorea_half_rtp' to be 'pass' but got 'fail'
                Expected the status of 'vstorea_half_rtn' to be 'pass' but got 'fail'

Intel/llvmpipe (logical and physical addressing)

        Images (clFillImage pitch)
                Expected the return code to be 0 but got -11
                Got reference for test '1D' but no result
                Got reference for test '1Darray' but no result
                Got reference for test '2D' but no result
                Got reference for test '2Darray' but no result
                Got reference for test '3D' but no result

Intel (physical addressing)

        Printf
                Expected the return code to be 0 but got 1
                Expected the status of 'address_space_0' to be 'pass' but got 'fail'
                Expected the status of 'address_space_1' to be 'pass' but got 'fail'
                Expected the status of 'address_space_4' to be 'pass' but got 'fail'
        Vectors
                Expected the return code to be 0 but got 1
                Expected the status of 'vec_align_struct' to be 'pass' but got 'fail'

Change-Id: Ie2b11d3f0b0297121a085e76d05da5f74af46aca

kpet avatar May 28 '24 18:05 kpet

FYI @rjodinchr in case the above list of regressions is of interest.

kpet avatar May 28 '24 18:05 kpet

@rjodinchr It looks like the libclc changes aren't passing the Android CI.

kpet avatar May 28 '24 20:05 kpet

@rjodinchr It looks like the libclc changes aren't passing the Android CI.

Is there a way to know/update the cmake version used in this Android CI?

rjodinchr avatar May 29 '24 07:05 rjodinchr

I couldn't find anything obvious in the logs. We should be able to update it and/or start using docker images. Do you know what version is needed?

kpet avatar May 29 '24 08:05 kpet

I couldn't find anything obvious in the logs. We should be able to update it and/or start using docker images. Do you know what version is needed?

I don't know, but maybe the minimum required version in llvm is not well set, that's why the failure is hard to understand with just the logs. I don't have lots of ideas looking at the failure, and I'm not sure how I could try to reproduce it to debug.

rjodinchr avatar May 29 '24 08:05 rjodinchr

I am able to reproduce on ChromeOS. The issue does not seem to be about cmake version.

rjodinchr avatar May 29 '24 16:05 rjodinchr

The issue is with cross compiling. I guess the workaround at the moment, would be to build manually libclc for android.... ☹ That's what I am going to do for ChromeOS.

rjodinchr avatar May 30 '24 13:05 rjodinchr

I have a PR in clspv to allow building with external libclc libraries: https://github.com/google/clspv/pull/1366

rjodinchr avatar May 30 '24 14:05 rjodinchr

I am not able to reproduce the clFillImage issues with llvmpipe.

rjodinchr avatar May 31 '24 09:05 rjodinchr

In fact I'm not able to reproduce anything. I have un clFillImage and Printf with and without physical addressing on llvmpipe and Intel, but everything seems to work well on my side.

rjodinchr avatar May 31 '24 10:05 rjodinchr

How do you want to handle android?

rjodinchr avatar May 31 '24 12:05 rjodinchr

https://chromium-review.googlesource.com/c/chromiumos/overlays/chromiumos-overlay/+/5581355

rjodinchr avatar Jun 10 '24 16:06 rjodinchr

Superseded by @rjodinchr's PR.

kpet avatar Aug 28 '24 17:08 kpet