[NFC][SYCL] Stabilize sub_group LIT tests

[vladimirlaz]

• Fixed OCL_ICD_FILENAMES capturing to LIT environment.
• Disable sub_group tests which fail on some ISAs on OpenCL CPU.
• Change local and global workgroup sizes to make them dividend
  of all possible sub_group sizes. That will fix test failures
  in the last sub_group of local group.
• Remove use of deprecated shuffle methods.
• Fixed tests where sub_group local ID exceed sub_group size.

[Pennycook]

Change local and global workgroup sizes to make them dividend of all possible sub_group sizes. That will fix test failures in the last sub_group of local group.

Isn't it important to test that the final sub-group in each work-group still supports these operations? The final sub-group will require some special handling (because in some cases the results of certain shuffles are undefined) but I don't think we should only test evenly divisible cases.

[vladimirlaz]

Change local and global workgroup sizes to make them dividend of all possible sub_group sizes. That will fix test failures in the last sub_group of local group.

Isn't it important to test that the final sub-group in each work-group still supports these operations? The final sub-group will require some special handling (because in some cases the results of certain shuffles are undefined) but I don't think we should only test evenly divisible cases.

In fact it means that we are testing test itself and OpenCL RT because the difference between divisible and non-divisible case is in OpenCL RT and test but not in SYCL RT.

So I removed this case

[Pennycook]

In fact it means that we are testing test itself and OpenCL RT because the difference between divisible and non-divisible case is in OpenCL RT and test but not in SYCL RT.

So I removed this case

I'm not sure I understand. Are you saying that because the SYCL compiler just forwards these functions to __spirv built-ins and doesn't give special handling to uneven divisions, we shouldn't have tests in the SYCL compiler? Instead, testing that the SPIR-V instructions do the correct thing should be handled by OpenCL tests?

[vladimirlaz]

e correct thing should be handled by

Yes, you are right. This is what I was trying to say.

[sergey-semenov]

The changes LGTM

[Pennycook]

This is what I was trying to say.

Ok, got it. Thanks for the explanation.