В Sun, 7 Apr 2019 13:57:38 +0500
Nikita Zlobin <cook60020tmp@xxxxxxx> пишет:
Adding to discussion, initiated by me long time ago, about using -O3
in yoshimi, then audio software at all - someone notes, someone noted
that -O3 is where vectorization is enabled (they compare
vectorization ops with scalar SSE ops, that are not vectorized).
https://stackoverflow.com/a/8080094
I'm currently working at first audio app (spectroscope, which will
later evolve into spectral editing helpers), and decided to try
overlook, how good code is simd-ed. While searching (without success)
about why integers ops are not simd-ed, despite SIMD has instructions
for both FP and int operations, i casually discovered this post.
More exactly, it happened while experimenting with graphical
post-processing for instant spectrum view. Since app is gtk3 based and
spectrum is created by cairo (probably not best approach), due to
cairo color formats (best avilable are argb32, rgb24 and rgb30),
there are close to not FP ops in my post-proc code.