3D real forward out-of-place transform with sizes 98, 98, 98 fails with HIPFFT_EXEC_FAILED

[ex-rzr]

What is the expected behavior

It should work, other sizes close to this one, as well as in-place transform seem to work ok.

What actually happens

One of OpenMM-HIP tests fails:
StreamHPC/openmm-hip#1
https://github.com/StreamHPC/openmm-hip/blob/master/platforms/hip/tests/TestHipFFTImplHipFFT.cpp#L144

How to reproduce

./hipfft-rider -o -t 2 --length 98 98 98

Running profile with 1 samples
out-of-place
length: 98 98 98
terminate called after throwing an instance of 'std::runtime_error'
  what():  Execution failed
Aborted (core dumped)

Environment

Hardware	description
GPU	MI100, MI210, RX6900XT, does not depend on the GPU model

Software	version
ROCm	5.2

[evetsso]

The problem here is actually in rocFFT, and has been fixed with ROCm/rocFFT@ced7c1c .