AVX2+FMA?

Issue #22 resolved
Former user created an issue

Hi,

I've heard that shtns can utilize AVX2+FMA to double performance over just using AVX, however there's no mention of this in your overview.

Is this something that is currently supported, and if so does it require an additional compilation option when installing shtns?

Comments (2)

  1. Nathanaël Schaeffer repo owner

    Hello,

    there is nothing special to do to enable that. By default SHtns will be compiled with -march=native which will pick up the details for you. If you want to compare with and without, try replacing -march=native with -march=avx in the Makefile generated by ./configure

    AVX2+FMA gave me a x1.7 to x1.8 speedup compared to AVX.

  2. Log in to comment