[SYCL] Split 'double type' test into two tests, 1) <test>.cpp; 2) <test>_aspect-fp64.cpp for llvm-test-suite

[myler]

'double' type are not native supported in Intel Gen11/12 and next Gens GPUs in the future, to disable particularly big number of important tests is not reasonable, so split 'double' part into two separate test file.

• split double tests in AtomicRef directory
• split double tests in Basic directory
• split double tests in DeprecatedFeatures directory
• split double tests in DeviceLib directory
• split double tests in ESIMD directory
• split double tests in GroupAlgorithm directory
• split double tests in InlineAsm directory
• split double tests in KernelParams directory
• split double tests in Regression directory
• split double tests in SpecConstants directory
• split double tests in SubGroup directory
• split double tests in USM directory
• add REQUIRES:aspect-fp64 to restrict some 'double' type tests "_double.cpp", "_fp64.cpp", "_fp_extra.cpp", etc.

[myler]

Failures are unrelated:

[2022-08-17T07:35:47.081Z] Failed Tests (1):
[2022-08-17T07:35:47.081Z] SYCL :: InvokeSimd/invoke_simd_conv.cpp

[2022-08-17T07:41:39.229Z] Unexpectedly Passed Tests (2):
[2022-08-17T07:41:39.229Z] SYCL :: SubGroup/reduce_spirv13.cpp
[2022-08-17T07:41:39.229Z] SYCL :: SubGroup/reduce_spirv13_fp64.cpp

[v-klochkov]

Ok for ESIMD. Thank you.

I also have a non-blocking comment regarding the naming.
IMO, it is better to stick to underscores and not mix them with '-'.
E.g. "functions_select_2d_core_aspect_fp64.cpp" looks much better to me than "functions_select_2d_core_aspect-fp64.cpp"

[AlexeySachkov]

LGTM for spec constants

[myler]

Ok for ESIMD. Thank you.

I also have a non-blocking comment regarding the naming. IMO, it is better to stick to underscores and not mix them with '-'. E.g. "functions_select_2d_core_aspect_fp64.cpp" looks much better to me than "functions_select_2d_core_aspect-fp64.cpp"

Renamed these new added file with only underscores.

[myler]

Hi @Pennycook,
Seems other subdirectories are already approved, but may I ask you to review changes in SubGroup and GroupAlgorithm, seems you are the code owner of them.

[myler]

Related failure:

[2022-08-31T02:01:02.817Z] Failed Tests (1):
[2022-08-31T02:01:02.817Z]   SYCL :: ESIMD/aot_mixed.cpp

This error LLVM ERROR: vc::translateBuild: unsupported platform cannot be reproduced.

Running on Intel(R) Iris(R) Xe Graphics
Running ESIMD kernel...
TEST Passed

Unrelated failures:

[2022-08-31T02:34:16.124Z] Failed Tests (4):
[2022-08-31T02:34:16.124Z]   SYCL :: ESIMD/api/replicate_smoke.cpp
[2022-08-31T02:34:16.124Z]   SYCL :: ESIMD/api/simd_view_select_2d_int.cpp
[2022-08-31T02:34:16.124Z]   SYCL :: ESIMD/ext_math.cpp
[2022-08-31T02:34:16.124Z]   SYCL :: ESIMD/matrix_transpose_glb.cpp

[romanovvlad]

A possible alternative solution is here: https://github.com/intel/llvm-test-suite/pull/1190/files

[myler]

Failures are known and unrelated.
Hi @v-klochkov
According to the https://github.com/intel/llvm-test-suite/blob/intel/CONTRIBUTING.md#merge, need maintainer to merge the PR. Could you help me on this?

[v-klochkov]

Failures are known and unrelated. Hi @v-klochkov According to the https://github.com/intel/llvm-test-suite/blob/intel/CONTRIBUTING.md#merge, need maintainer to merge the PR. Could you help me on this?

I did not merge this yet because of the comment from @romanovvlad (#1150 (comment)) and did not have clear understanding if the discussion here #1190 has got some solution/agreement.

That alternative approach used in #1190 has sympathy/votes from at least 3 people.
And I would add my vote to there as well because device.has is specially designed and convenient way to run-or-skip hw specific features.

I gave my approval for this PR, but I'd rather let @romanovvlad to merge this patch if Vlad's comment was non-blocking

[AlexeySachkov]

Personally, I would vote for device code split approach, because it allows us to avoid duplicating coverage for non-optional types. I can help with creating PRs similar to one that @romanovvlad published (the first one to go is #1241)

[pvchupin]

@v-klochkov, @myler , @romanovvlad , @AlexeySachkov
I assume we decided to proceed with alternative approach partially done at #1241 , #1190?

@aelovikov-intel, can you please take a look at remaining ones on Windows? We currently have ~30 fails and it's really unsafe to merge anything else until these fixed.

[pvchupin]

@myler, should we close this PR in favor of alternatives in progress?

[myler]

@myler, should we close this PR in favor of alternatives in progress?

Hi @pvchupin

This solution tends to change the testing framework REQURIES: aspect-fp64 and macros #ifdef ENABLE_FP64, not depends on SYCL standard aspect::fp64 and feature -fsycl-device-code-split=per_kernel

#1190 tends to use SYCL style solution to handle double type tests.

If the final decision is to use if (device.has(aspect::fp64)) + -fsycl-device-code-split=per_kernel to handle double type tests, I agree to close this one. But #1190 did not fix all tests, we may have at least 26 tests contain double type code.

[gmlueck]

If the final decision is to use if (device.has(aspect::fp64)) + -fsycl-device-code-split=per_kernel to handle double type tests, I agree to close this one. But #1190 did not fix all tests, we may have at least 26 tests contain double type code.

That seems like a better solution to me. Once we implement the OptionalDeviceFeatures design, we can remove the "-fsycl-device-code-split=per_kernel", which should be relatively easy.

Are you saying there are 26 test failures that will not be fixed with this technique? Do we know what the cause of those failures is?

[myler]

Close this PR as we have another alternative.