Add three missing llvm-spirv commits from Khronos

[sarnex]

The below three commits are missing from intel/llvm llvm-spirv and it's causing differences in internal testing:
KhronosGroup/SPIRV-LLVM-Translator@352ea14
KhronosGroup/SPIRV-LLVM-Translator@85815e7
KhronosGroup/SPIRV-LLVM-Translator@aded5af

I also had to add a customization to skip interface global variables.

Signed-off-by: Sarnie, Nick [email protected]

[MrSidims]

Thanks for doing this!
Unfortunately I don't think we ready for them. Here are the reasons:
KhronosGroup/SPIRV-LLVM-Translator@352ea14 this patch lists GVs in the Interfaces of EntryPoint. Unfortunately there was a bug in IGC, which doesn't handle Interfaces at all. AFAIK the bug is now fixed, but these drivers are not yet in CI.

UPD: ah, missed this from the description: I also had to add a customization to skip interface global variables.

KhronosGroup/SPIRV-LLVM-Translator@85815e7 this patch is addressing SPIR-V 1.4 requirement and we haven't yet switched to it in this repo. There is PR for that #7493 but fresh-new pre-commit runs are showing some tests with very fishy error message failing. I haven't yet investigated them, but it's on my list. Though, probably we still can merge this patch.

There are actually more 'customizations' in intel/llvm, they are being tracked here: #7592 feel free to update the list if you something that was not mentioned.

[sarnex]

@MrSidims You are realy good at sniping my draft PRs

Yeah after looking at these CI failures I came to the same conclusion, I'll close this for now

[MrSidims]

If you don't mind, I'd re-open it :)

The CI failures I see are not related to GPU driver lacking SPIR-V 1.4 support. Not sure about this very PR though. So it's worth a stand-alone investigation

[sarnex]

@MrSidims Sure, I still need to investigate the remaining CI issues, I think there were only ~5 test failures

[sarnex]

@bader Any idea what's going on with the OCL x64 failure? I merged sycl HEAD into my branch before running CI.

Starting download for sycl_linux_default
Error: Unable to find an artifact with the name: sycl_linux_default
Run actions/upload-artifact@v1
Error: Path does not exist /localdisk2/github/_work/llvm/llvm/build/results_ocl_x64_default.json
Error: Exit code 1 returned from process: file name '/localdisk2/github/bin.2.301.1/Runner.PluginHost', arguments 'action "GitHub.Runner.Plugins.Artifact.PublishArtifact, Runner.Plugins"'.

[bader]

@bader Any idea what's going on with the OCL x64 failure?

No.

I merged sycl HEAD into my branch before running CI.

Please, try it again.

[sarnex]

@MrSidims Okay it seems like the 5 failures on OCL CPU are all caused by an old driver. If I use a newer driver locally it works fine.

I have three options here:

1. Disable these 5 tests on OCL CPU until new driver
2. Try to upgrade the driver in CI here.
3. Hold off on this change until new driver

What do you recommend? 2) could turn into a nightmare :)

[MrSidims]

@cdai2 hi, since CPU runtime is now public, may I ask you whether we expect to have it used in intel/llvm for testing?

[sarnex]

Note I didn't test if the open source version works, I only tested an internal version.

[MrSidims]

@sarnex

Disable these 5 tests on OCL CPU until new driver

I see
llvm-foreach: Segmentation fault (core dumped)
so I would vote for 'no'. Do you know, what exactly is missing in the currently used CPU driver?

[sarnex]

@MrSidims

Do you know, what exactly is missing in the currently used CPU driver?

No idea, but I can investigate. Sorry what do you mean by 'vote for no'? Do you support one of the 1-3 options?

[MrSidims]

@MrSidims

Do you know, what exactly is missing in the currently used CPU driver?

No idea, but I can investigate. Sorry what do you mean by 'vote for no'? Do you support one of the 1-3 options?

Sorry :) I would vote for not disabling the tests, until we know, what causes segfault.

[asudarsa]

Hi @sarnex

Thanks for looking at this. Just a quick comment. Will it make it easier to test/debug if we submit the three PRs separately?

[sarnex]

@asudarsa Unfortunately we can't, they are interdependent. We can't merge this anyway because of the OCL CPU issue.

[sarnex]

finally new ocl cpu driver, is there hope? let's find out on tonight's episode of i doubt it

[github-actions]

This pull request is stale because it has been open 180 days with no activity. Remove stale label or comment or this will be automatically closed in 30 days.

[github-actions]

This pull request was closed because it has been stalled for 30 days with no activity.