Re-enable CUDA tests #1542
Comments
|
A lot of work of the community and us went into fixing LIT testing to work as expected for CUDA - without marking tests as XFAIL it would have taken longer with more breakage slipping through. Without a lot of visible and invisible work of Alexander there would to this time be no reliable and deterministic LIT testing. Sometimes I mark tests as XFAIL that we expect to pass in the future. This way the tests will automatically show up as unexpected passes when running tests after adding features. We have a roadmap for many of the CUDA related LIT XFAILs (and UNSUPPORTEDs) and our whole team takes responsibility for these and is actively working on them. Some XFAILS will be fixed soon, some later, some are not part of what we plan to achieve and we will not work on them. Tests we are aware off and are tracking will have a TODO or comment added to them about the problem. Tests without any comment typicall fall into the following categories:
|
|
That is fine, we just need a bug/series of bugs to track these. So, this is here to track them. At least one is disabled because of design implementation details of __builtin_unique_stable_name, which is how this came up. It is disappointing that it was never brought up via bug/etc, since we could have fixed this trivially immediately (I've since fixed it in the upstream, so we're awaiting the pulldown). We need to analyze the changes that will need to happen to make sure they do not have ABI implications, otherwise they might never be able to be fixed. |
|
@erichkeane, it looks like @vladimirlaz has some problems with pulldown, so if |
|
I don't know if it necessarily IS required soon, just before we hit a release that locks us into ABI. Presumably, we could have these problems with some of the things that CUDA disabled as well, so I just want to make sure we have a plan for HOW to fix these before a release locks us into not being able to make said change. |
|
ABI, ABI... Ahem... Is it a proof of success when we end up into ABI stability discussion? :-) |
Any language worth using has to end up in an ABI discussion eventually. As a person shipping a compiler frontend, ABI compatibility between versions and compilers is a constant headache. |
|
Fix of tests for |
|
PR including fix for |
|
The following (merged) PR triaged most LIT tests to mark missing features or features that cannot be supported by CUDA as
|
|
|
|
That leaves xfails due to missing builtins - and we are actively working on them. |
|
|
|
|
|
|
Unnamed Lambdas should now be supported for CUDA I changed the text format so that it should now work, if you're still having a problem, please let me know! |
|
I see, thanks! So nothing to do with the mangling anymore. Not sure how the USM matters here (other than perhaps it is used in the test?). |
|
Both tests ( |
|
Great, thanks for clarifying. |
|
PRs adding builtins:
|
|
usm/queue_wait.cpp - PR #1577 |
|
SYCL :: basic_tests/image/image_write.cpp fails on CUDA device. |
|
I don't see any more |
|
+@vladimirlaz, I think we should check https://github.com/intel/llvm-test-suite/tree/intel/SYCL as most of these tests moved from |
|
I think there are a handful of xfails left in there, but would it make sense to open a new ticket on that repo for tracking those? |
|
Simple grepping showed that there are some: |
|
To my knowledge, some of them have both |
You are right. |
I'm okay with that. |
I think it may be cleaner to just restart it and link. That way people who read it won't have to go through all the previous discussions on fixed issues to get to the relevant information. Thank you though! If you are okay with that solution, I will go find the relevant XFAILs and list them in a new issue on llvm-test-suite. I'll leave out the ones with |
|
Sounds good to me. Thanks! |
|
I have opened the continuation issue here: intel/llvm-test-suite#249 |
|
Closing this one, because we have a more up-to-date tracker for those issues here: intel/llvm-test-suite#249 |
In 7a9a425 Alexander disabled quite a few SYCL tests for CUDA with XFAIL:CUDA in order to get the commit in. These all need to be re-enabled. This likely can be handled by a few different people, but Alexander should own re-enabling that which he disabled.
Note that most have a TODO associated with them.
The following test all have XFAIL:cuda in them, so they need to be analyzed and fixed:
usm/queue_wait.cpp
usm/allocator_vector_fail.cpp
usm/math.cpp
usm/mixed2template.cpp
usm/smemll.cpp
usm/depends_on.cpp
usm/memset.cpp
usm/mixed_queue.cpp
usm/dmemll.cpp
usm/memcpy.cpp
usm/allocator_vector.cpp
usm/allocatorll.cpp
usm/hmemll.cpp
usm/memadvise.cpp
usm/badmalloc.cpp
usm/mixed2.cpp
usm/mixed.cpp
device-code-split/aot-gpu.cpp
device-code-split/split-per-kernel.cpp
device-code-split/split-per-source-main.cpp
regression/kernel_name_class.cpp
regression/kernel_unnamed.cpp
separate-compile/sycl-external.cpp
ordered_queue/oq_kernels.cpp
ordered_queue/ordered_buffs.cpp
basic_tests/buffer/subbuffer.cpp
basic_tests/buffer/buffer.cpp
basic_tests/parallel_for_range.cpp
basic_tests/scalar_vec_access.cpp
basic_tests/vec_op.cpp
basic_tests/image.cpp
basic_tests/compare_exchange_strong.cpp
basic_tests/image_array.cpp
basic_tests/stream/stream.cpp
basic_tests/stream/auto_flush.cpp
basic_tests/parallel_for_indexers.cpp
built-ins/vector_common.cpp
built-ins/scalar_geometric.cpp
built-ins/vector_integer.cpp
built-ins/vector_geometric.cpp
built-ins/scalar_relational.cpp
built-ins/vector_math.cpp
built-ins/vector_relational.cpp
built-ins/nan.cpp
built-ins/scalar_common.cpp
built-ins/scalar_integer.cpp
built-ins/scalar_math.cpp
built-ins/printf.cpp
scheduler/ReleaseResourcesTest.cpp
scheduler/DataMovement.cpp
The text was updated successfully, but these errors were encountered: