[x264-devel] x264 Development Newsletter: Vol 36

[Jason Garrett-Glaser]

I'm now allowed to post full-program AVX2 results, so I'll follow up
on the document included with the newsletter.

Test parameters: 720p, preset medium, ~3.6mbps. Numbers in parenthesis
are standard error.

Overall performance gain from AVX2 with threads: 5.23(11)% (30 samples)
Overall performance gain from AVX2 without threads: 4.11(08)% (10 samples)

An unpaired 2-tailed t-test says that the null hypothesis of "there is
no difference in performance gain with and without threads" is
rejected with acceptance probability p ~= 1.88*10^-9. I suspect the
reason for this difference is that some AVX2 functions, like SAD, are
bottlenecked (e.g. by loads), but reducing the number of arithmetic
instructions frees up execution resources for the other thread sharing
the core.

Jason