[FFmpeg-devel] [PATCH] vp9/x86: 16px MC functions (64bit only).

[Ronald S. Bultje]

Hi,

On Wed, Jan 15, 2014 at 8:33 AM, Clément Bœsch <u at pkh.me> wrote:

> On Thu, Dec 26, 2013 at 09:05:37PM -0500, Ronald S. Bultje wrote:
> > Cycle counts for large MCs (old -> new on ped1080p.webm, mx!=0&&my!=0):
>
> decicyle?


No, cycle (yes they take forever!).


> > 16x8:    876 ->   870  (0.7%)
> > 16x16:  1444 ->  1435  (0.7%)
> > 16x32:  2784 ->  2748  (1.3%)
> > 32x16:  2455 ->  2349  (4.5%)
> > 32x32:  4641 ->  4084 (13.6%)
> > 32x64:  9200 ->  7834 (17.4%)
> > 64x32:  8980 ->  7197 (24.8%)
> > 64x64: 17330 -> 13796 (25.6%)
> > Total decoding time goes from 9.326sec to 9.182sec.
> > ---
> >  libavcodec/x86/vp9dsp_init.c |   5 ++
> >  libavcodec/x86/vp9mc.asm     | 122
> +++++++++++++++++++++++++++++++++++++++++++
> >  2 files changed, 127 insertions(+)
> >
> [...]
> > +%if ARCH_X86_64
> > +
> > +%macro filter_vx2_fn 1
> > +%assign %%px mmsize
> > +cglobal %1_8tap_1d_v_ %+ %%px, 6, 8, 14, dst, dstride, src, sstride, h,
> filtery, src4, sstride3
>
> > +    sub       srcq, sstrideq
> > +    lea  sstride3q, [sstrideq*3]
> > +    sub       srcq, sstrideq
> > +    mova       m13, [pw_256]
> > +    sub       srcq, sstrideq
> > +    mova        m8, [filteryq+ 0]
> > +    lea      src4q, [srcq+sstrideq*4]
> > +    mova        m9, [filteryq+16]
> > +    mova       m10, [filteryq+32]
> > +    mova       m11, [filteryq+48]
>
> Untested, but wouldn't it be simpler to have:
>
>     lea  sstride3q, [sstrideq*3]
>     lea      src4q, [srcq+sstrideq]
>     sub       srcq, sstride3q
>     mova       m13, [pw_256]
>     mova        m8, [filteryq+ 0]
>     mova        m9, [filteryq+16]
>     mova       m10, [filteryq+32]
>     mova       m11, [filteryq+48]
>
> ?


I think you're almost right, but we'd have to swap the second lea and the
sub. Feel free to test and if that works and it's faster (or not slower),
feel free to commit. I can maybe try it this weekend.

Ronald