[FFmpeg-devel] [PATCH] PPC64: Add versions of functions in libswscale/input.c optimized for POWER8 VSX SIMD.

Ronald S. Bultje rsbultje at gmail.com
Sun Jul 10 21:30:20 EEST 2016


On Sun, Jul 10, 2016 at 1:25 PM, Michael Niedermayer <michael at niedermayer.cc
> wrote:

> It could be reverted, it could be left for the next developer to use
> as starting point, i have no oppinion on this.

So, I want to comment on this specifically: I don't think it can.

I don't specifically mean that it's too slow, although that's a reason by
itself. The _specific_ reason I don't think it should be used as a basis
for really anything is that we _do not know what it does_. We don't know if
it uses simd instructions for the scalar operants (+=, *=, etc.) and there
is no way for us to verify other than actually understanding the complete
problem space.

At that point, we might just as well claim the bounty ourselves, which
defeats the whole purpose.

We should revert it, and we should confirm that future commits are fully
simd with no vector-scalar conversion operations. We should, conversely,
also confirm that C did not use auto-vectorization. Without either of these
questions answered, we cannot possibly know whether these optimizations
make sense or not.

And it's up to whoever wants to resolve #5570 to answer these questions.
Not up to us. Certainly not up to me.


More information about the ffmpeg-devel mailing list