[FFmpeg-devel] [PATCH v2 3/3] avfilter/vf_colordetect: add x86 SIMD implementation

[Niklas Haas]

On Thu, 17 Jul 2025 11:41:56 +0200 Niklas Haas <ffmpeg at haasn.xyz> wrote:
> On Wed, 16 Jul 2025 17:25:12 -0300 James Almer <jamrial at gmail.com> wrote:
> > On 7/16/2025 1:24 PM, Niklas Haas wrote:
> > > +cglobal detect_alpha%1_%3, 6, 7, 6, color, color_stride, alpha, alpha_stride, width, height, x
> > > +    pxor m0, m0
> > > +    add colorq, widthq
> > > +    add alphaq, widthq
> > > +    neg widthq
> > > +%ifidn %3, limited
> > > +%if ARCH_X86_64
> > > +    movq xm3, r6mp ; p
> > > +    movq xm4, r7mp ; q
> > > +    movq xm5, r8mp ; k
> >
> > They are int arguments, so no point using movq.
> >
> > > +%else
> > > +    movd xm3, r6mp ; p
> > > +    movd xm4, r7mp ; q
> > > +    movd xm5, r8mp ; k
> > > +%endif
> > > +    vpbroadcast%2 m3, xm3
> > > +    vpbroadcast%2 m4, xm4
> > > +    vpbroadcast%2 m5, xm5
> > > +%endif
> > > +.lineloop:
> > > +    mov xq, widthq
> > > +    .loop:
> > > +    %ifidn %3, full
> > > +        movu m1, [colorq + xq]
> > > +        movu m2, [alphaq + xq]
> > > +        pmaxu%1 m1, m2
> > > +    %else
> > > +        pmovzx%1%2 m1, [colorq + xq]
> > > +        pmovzx%1%2 m2, [alphaq + xq]
> > > +        pmull%2 m1, m3
> > > +        pmull%2 m2, m4
> >
> > pmulld is painfully slow, so try to instead use pmaddwd. It should be
> > doable because the input are words.
>
> Wouldn't this add together adjacent pairs of words, which is not what we want?
>
> Or do you mean I should construct two intermediate registers like
> {p, 0, p, 0, p, 0, ...} and {0, p, 0, p, 0, p, ...} and then use two separate
> pmaddwd instructions to process a full register?

I gave that a try, but it fails on the signedness requirement of `pmaddwd`.
I tried instead using `pmullw` + `pmulhuw`, and then repacking the data with
`puncklwd/punckhwd`, but it ended up the same performance as a single `pmulld`,
while also making the code a lot less readable.

I will stick with the `pmulld` for now, it is the cleanest solution.

>
> > _______________________________________________
> > ffmpeg-devel mailing list
> > ffmpeg-devel at ffmpeg.org
> > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> >
> > To unsubscribe, visit link above, or email
> > ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".