[ROCFFT] RocFFT fails tests when using ROCm 6.0 or later

[hjabird]

Summary

The current tip of oneMKL interfaces fails unit tests when using ROCm 6.0 or later. Using ROCm 5.4.3 does not fail tests.

The failed tests are real-to-complex multi-dimensional tests.

Version

The issue was introduced with #528 and can be reproduced with oneMKL interfaces developr.

Environment

• AMD MI210
• Reproduceable RocFFT from ROCm 6.0 or 6.1
• Ubuntu 22.04
• ICPX 2024.2 with Codeplay's plugin for AMD (which supports ROCm 6.0.2 for MI210).

Steps to reproduce

Build with RocFFT enabled. Test with

./bin/test_main_dft_ct --gtest_filter=*REAL_SINGLE_in_place_USM*batches_1*

to observe failures.

Observed behavior

Tests give wrong results or memory faults.

Expected behavior

Tests should pass.

[hjabird]

• It seems to be the backward DFT that fails.
• If the tests prior to regression-causing PR are used, the tests still pass. (ie. tests using the INPUT/OUTPUT API with recommit, instead of using FWD/BWD API).
• This problem can also be reproduced on an AMD W6800.

[hjabird]

This turned out to be a rocFFT bug. RocFFT issue ROCm/rocFFT#504

[hjabird]

There is a work-around for the above at https://github.com/hjabird/oneMKL/tree/hjab/fix_rocfft6_issue

Unfortunately there is still are still failing tests - out-of-place complex 4x4x4_fwd_strides_2_4_1_16_bwd_strides_1_4_16_1_batches_2. These tests pass with ROCm 5.4.3, but fail with ROCm 5.7.1 and later.

[hjabird]

This looks like a bug in rocFFT to me. I've described the issue at ROCm/rocFFT#507. I think we oneMKL Interfaces will have to throw unsupported on these tranposing DFTs.

[Rbiessy]

The 2 workarounds have been merged and the issues fixed in more recent versions of rocFFT. Closing.