OpenMP Offload Fortran Tests Pass with Flang-new

Runtimes OpenMP
,
1

We are excited to share, we have now added support for the OpenMP Offload V&V’s Fortran tests in the llvm-test-suite. PR #119

Currently:

AMD GPUs Pass Rate is 5%

NVIDIA GPUs Pass Rate is 15%

Please find the CI Pipeline here: LLVM OpenMP Offload CI

This work is from a continuation of our efforts from the Exascale Computing Project (ECP) SOLLVE activities and moving into DOE’s S4PST project, we started developing a CI/CD to test LLVM Clang’s implementation of OpenMP Offloading onto any to all type of GPUs. Part of the CI is using the OpenMP V&V testsuite to check the implementation. Though, we were only checking C and C++ with Clang. Now we have now included Fortran testing of Flang-new to the CI.

You can see more detailed results here: Results :: OpenMP Validation & Verification

5 Likes
2

Thanks for adding this CI.

Would you know why the different GPUs have different pass rates?

Please report any crashes in github issues Issues · llvm/llvm-project · GitHub.

3

The main problem that we see is:

“PluginInterface” error: Failure to init kernel: Error in hsa_executable_get_symbol_by_name(__omp_offloading_85_abc42cb__QFPtest_target_enter_exit_data_if_l70.kd): HSA_STATUS_ERROR_INVALID_SYMBOL_NAME: There is no symbol with the given name.
omptarget error: Failed to load kernel

But we haven’t had time to create the issue on LLVM. We would do it soon.

Thanks

4

Seems pretty straightforward if it’s looking up a kernel name that doesn’t exist / isn’t exported. It’s possible someone isn’t setting the visibility right (i.e. the kernel is hidden and doesn’t show up in the ELF symbol table) or it’s just never emitted.

I’d recommend getting the binary the OpenMP runtime is trying to load and doing llvm-readelf --notes to see if said kernel exists.

5

@ajarmusch @fel-cab Do you have any updates on the pass rates?

6

An updated pass rate with the V&V for both NVIDIA and AMD as of today is

AMD GPUs Pass Rate ~ 11%

NVIDIA GPUs Pass Rate ~ 11%

The two main changes are the pass rate for AMD increased, bringing the AMD and NVIDIA GPUs pass rate to 11%. As well as the NVIDIA Pass Rate decreased, we will report the failing tests in GitHub issue.

1 Like