Bit Reproducibility with plan_many_dft with variable howmany about fftw3 HOT 2 CLOSED

This behavior is expected.

You have a mental model where FFTW computes a plan for HOWMANY=1 and then performs a loop over HOWMANY. In reality, HOWMANY is part of the plan specification and FFTW is free to use different code paths depending on HOWMANY. For example, HOWMANY=1 may use scalar code or code that uses AVX over the "N" dimension, whereas HOWMANY=4 may use AVX over the HOWMANY dimension. In general, the roundoff properties will differ for different values of HOWMANY.

If you want/need invariance w.r.t. HOWMANY, plan for HOWMANY=1 and write the loop yourself.

from fftw3.

Thank you very much for the explanation!

On Sat, Jun 20, 2015, 6:13 AM Matteo Frigo [email protected] wrote:

Closed #44 #44.

—
Reply to this email directly or view it on GitHub
#44 (comment).

from fftw3.