[FFmpeg-devel] [PATCH] avcodec/cfhd: add x86 SIMD

[Alexander Strasser]

Hi Derek,
hi all!

On 2020-08-14 20:22 +0100, Derek Buitenhuis wrote:
> On 14/08/2020 20:13, Paul B Mahol wrote:
> >> Resending because I accidentally replied to James instead of the list.
> >> Woops.
> >>
> >> I guess it was not clear to me this is not the initial thread, since it is
> >> not
> >> a v2 patch, and no other thread is titled this, or seems to include SIMD?
> >> Perhaps
> >> I missed it.
> >>
> >> I still cannot actually locate any benchmarks in the various CFHD threads.
> >> If any were
> >> done, as would be needed, to, like... test ones own SIMD code, they should
> >> be included
> >> in the commit messages.
> > Results differs between various CPUs and environments and also depends on
> > encoded file resolution and quality.
>
> Well, yes. That never stopped anyone from providing information.
>
> > With my local patch I get overall several percent speed increase with
> > only horiz_filter SSE2 applied.
> > I also work on vert_filter SSE2 code, which currently give big speedup
> > with lowest quality encodings and higher resolutions.
> >
> > For example:
> >
> > best quality 1080 60fps progressive yuv422p10 with additonal WIP
> > vertical filter:
> > cpuflags 0 speed : 0x243x realtime
> > cpuflags sse2 speed: 0x353x realtime
> >
> > worst quality 1080 60fps progressive yuv422p10 with additional WIP
> > vertical filter:
> > cpuflags 0 speed: 0x348x realtime
> > cpuflags sse2 speed: 0x811x realtime
>
> Thanks for finally providing some numbers. They should be in the commit
> messgae.
>
> > Also I want reviews to be technical as possible, i have not sent this patch
> > to listen to bad remarks but to get more improvements in assembly code
> > if possible.
>
> Reviews of commit messages and methodolody or lack there of are valid reviews.
>
> > If you are not assembly developer and are not willing to test patches better to
> > stay away from this thread.
>
> See below
>
>
> >
> >> (Nice that we still silently ignore various insults from Paul thrown around
> >> by the way.)
> > What specific insults in this thread?
>
> You wrote one just in this email, calling an ask for a better commit message and
> actual benchmarks "bad remarks", and previous wrote "You are not being helpful at all.".
>
> See also above: "If you are not assembly developer and are not willing to test patches better to
> stay away from this thread."
>
> Frankly, it's disgusting that this community prefers to silently ignore your (and others,
> but lately, mostly your) abusive and unfriendly conduct.

I have just read this thread.

Derek's initial comment was clearly valid I would say:

    I would expect any SIMD patch to include benchmarks showing it
    is actually faster.

IMHO what followed was way too much back and forth and not really
nice at all. IMHO this kind of conversations are not for the good
of the community and probably neither for the individuals having
the conversation.

So a better reply to Derek's concern would have been:

    Here are my (preliminary) benchmarks on my not so
    representative CPU.

or

    I will add benchmarks later and publish them before pushing.


Not really related to this in general is the reproducibility of
the benchmarks. That could be improved I think, but having some
benchmarks included is rather a must have and better than having
none.


[...]

Best regards,
  Alexander