A coworker found a bug in fftw linked to genric simd 256 kernels following avx/avx2 kernels and we were able to narrow down the issue.
Repro code (we observed the issue in Ubuntu 21.10, Ubuntu 22.04 docker but not Ubuntu 20.04 which seems to always generated plans without simd256 kernels):
Hello,
A coworker found a bug in fftw linked to genric simd 256 kernels following avx/avx2 kernels and we were able to narrow down the issue.
Repro code (we observed the issue in Ubuntu 21.10, Ubuntu 22.04 docker but not Ubuntu 20.04 which seems to always generated plans without simd256 kernels):
This C code compiled with
gcc fftw_bug.c -O3 path/to/libfftw3.a -lm
yields the sample output:
when compiling libfftw3.a from the fftw-3.3.10 source directory:
However, when disabling the generic simd256 i.e. :
The 1000 test runs will run without any issue in an Ubuntu 22.04 docker image which would have precision issue with the simd256 plans.
Cheers