Enable architectures for libhipcxx (fixes #2504)

[obersteiner]

Motivation

So far libhipcxx only enables a very limited number of architectures. With this PR we want to widen this support.

Test Plan

Ideally we want to test all added architectures.

Test Logs

Libhipcxx is passing all Linux architectures that we could test via TheRock:

https://github.com/ROCm/TheRock/actions/runs/22229769520?pr=2946

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[obersteiner]

Todo: We need to disable the changes in amdgpu_family_matrix.py before merging.

The failing gfx90a tests seem to be a node issue as other packages have similar issues (and we test with gfx90a on our dev CI which is passing all of our tests)

[marbre]

These changes need to be reverted before the merge but are okay for testing.

[marbre]

https://github.com/ROCm/TheRock/actions/runs/21752139280/job/63742765914?pr=2946#step:10:3072

[20](https://github.com/ROCm/TheRock/actions/runs/21752139280/job/63742765914?pr=2946#step:10:3025)
  5: Failed Tests (25):
  5:   libhip++ :: cuda/float/bfloat_cos_comparison.pass.cpp
  5:   libhip++ :: cuda/float/bfloat_exp_comparison.pass.cpp
  5:   libhip++ :: cuda/float/bfloat_log_comparison.pass.cpp
  5:   libhip++ :: cuda/float/bfloat_sin_comparison.pass.cpp
  5:   libhip++ :: cuda/float/bfloat_sqrt_comparison.pass.cpp
  5:   libhip++ :: cuda/float/half_cos_comparison.pass.cpp
  5:   libhip++ :: cuda/float/half_exp_comparison.pass.cpp
  5:   libhip++ :: cuda/float/half_log_comparison.pass.cpp
  5:   libhip++ :: cuda/float/half_sin_comparison.pass.cpp
  5:   libhip++ :: cuda/float/half_sqrt_comparison.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/atomic_cuda_float.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/atomic_cuda_generic.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/atomic_cuda_signed.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/atomic_cuda_unsigned.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/atomic_std_float.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/atomic_std_generic.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/atomic_std_signed.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/atomic_std_unsigned.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/flag.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/reference_cuda.pass.cpp
  5:   libhip++ :: heterogeneous/atomic/reference_std.pass.cpp
  5:   libhip++ :: heterogeneous/optional.pass.cpp
  5:   libhip++ :: heterogeneous/pair.pass.cpp
  5:   libhip++ :: heterogeneous/tuple.pass.cpp
  5:   libhip++ :: heterogeneous/variant.pass.cpp

[marbre]

@geomin12 @amd-justchen can you help looking into the potential node-issue? I think we should wait for a green signal and not fully rely on the local testing done by @obersteiner and team. We should make sure tests passed at least once before merging :)

[obersteiner]

@geomin, @amd-justchen @marbre: In the logs I see errors like out of memory exceptions that I also see for example for thrust. We do not use a lot of memory in our tests (we do not have any large numeric tests but instead have only very small tests). An example for the rocThrust error https://github.com/ROCm/TheRock/actions/runs/21752139280/job/63742765864?pr=2946#step:10:19181

[obersteiner]

Those tests are passing now.

[marbre]

There are test failures for the so far tested gfx90a, see below. I assume this should be fixed before allowing to run on other gfx90X architectures?

[geomin12]

the gfx90X failures are being resolved, noted in #3506

[geomin12]

i'm glad the tests work :)

but let's please revert this back to original state, we lack the machines to run all tests

i would post the test logs in the PR description as proof of working, then potentially add skip-ci label as this is already proven working and no need to take CI resources

[obersteiner]

I have posted the link to the old test into the description + dropped the commit that changed the file + added the skip-ci label.

[geomin12]

lgtm, thanks for this update

[geomin12]

I'll need to fix skip-ci label, looks like it isn't working properly with other labels. will fix this in a following PR

I removed the labels, so skip-ci will work

[marbre]

All currently linked checks seem to have been canceled (or CI runs might be still queued due do the force push)? Code-change wise this looks good but I don't have any signal from CI for your changes. Therefore leaving it to @geomin12 and you to judge.

Dismissing as outlined in the last review comment.