[FFmpeg-devel] [PATCH] avcodec/vp9: ipred_dr_16x16_16 avx2 implementation

[gh0st]

>I know unaligned loads are not as slow as they used to be,
>but could m1 be produced by m2 and palignr?

I am not sure, can you clarify your question?

>From the comment I assume you don't use the extra two bytes
>that you get from the load, as you mark them as "*"
>generic undefined values

No, those two extra bytes are actually used, that's the above/left corner
pixel.
If you look in the vp9dsp_template.c file, there is a macro defined
diag_downright_
that's top[-1] in the body. Sorry for this ambiguous marking, but it's used
in
other ipred_dr functions so I decided to follow it.

>"cnt" doesn't seem to be used.

Yes indeed, I mislooked that, thanks.

On Fri, Jun 9, 2017 at 6:03 PM, Ivan Kalvachev <ikalvachev at gmail.com> wrote:

> On 6/8/17, Ilia Valiakhmetov <zakne0ne at gmail.com> wrote:
> > vp9_diag_downright_16x16_12bpp_c: 149.0
> > vp9_diag_downright_16x16_12bpp_sse2: 67.8
> > vp9_diag_downright_16x16_12bpp_ssse3: 45.6
> > vp9_diag_downright_16x16_12bpp_avx: 36.6
> > vp9_diag_downright_16x16_12bpp_avx2: 25.5
> >
> > ~30% faster than avx
> >
> > Signed-off-by: Ilia Valiakhmetov <zakne0ne at gmail.com>
> > ---
> >  libavcodec/x86/vp9dsp_init_16bpp.c    |  2 ++
> >  libavcodec/x86/vp9intrapred_16bpp.asm | 56
> > +++++++++++++++++++++++++++++++++++
> >  2 files changed, 58 insertions(+)
> >
> > diff --git a/libavcodec/x86/vp9dsp_init_16bpp.c
> > b/libavcodec/x86/vp9dsp_init_16bpp.c
> > index d1b8fcd..8d1aa13 100644
> > --- a/libavcodec/x86/vp9dsp_init_16bpp.c
> > +++ b/libavcodec/x86/vp9dsp_init_16bpp.c
> > @@ -52,6 +52,7 @@ decl_ipred_fns(dc,      16, mmxext, sse2);
> >  decl_ipred_fns(dc_top,  16, mmxext, sse2);
> >  decl_ipred_fns(dc_left, 16, mmxext, sse2);
> >  decl_ipred_fn(dl,       16,     16, avx2);
> > +decl_ipred_fn(dr,       16,     16, avx2);
> >  decl_ipred_fn(dl,       32,     16, avx2);
> >
> >  #define decl_ipred_dir_funcs(type) \
> > @@ -136,6 +137,7 @@ av_cold void ff_vp9dsp_init_16bpp_x86(VP9DSPContext
> > *dsp)
> >          init_fpel_func(1, 1,  64, avg, _16, avx2);
> >          init_fpel_func(0, 1, 128, avg, _16, avx2);
> >          init_ipred_func(dl, DIAG_DOWN_LEFT, 16, 16, avx2);
> > +        init_ipred_func(dr, DIAG_DOWN_RIGHT, 16, 16, avx2);
> >          init_ipred_func(dl, DIAG_DOWN_LEFT, 32, 16, avx2);
> >      }
> >
> > diff --git a/libavcodec/x86/vp9intrapred_16bpp.asm
> > b/libavcodec/x86/vp9intrapred_16bpp.asm
> > index 92333bc..67b98b1 100644
> > --- a/libavcodec/x86/vp9intrapred_16bpp.asm
> > +++ b/libavcodec/x86/vp9intrapred_16bpp.asm
> > @@ -1170,6 +1170,62 @@ DR_FUNCS 2
> >  INIT_XMM avx
> >  DR_FUNCS 2
> >
> > +%if HAVE_AVX2_EXTERNAL
> > +INIT_YMM avx2
> > +cglobal vp9_ipred_dr_16x16_16, 4, 6, 7, dst, stride, l, a
> > +    mova                    m0, [lq]                   ;
> klmnopqrstuvwxyz
> > +    movu                    m1, [aq-2]                 ;
> *abcdefghijklmno
> > +    mova                    m2, [aq]                   ;
> abcdefghijklmnop
>


> From the comment I assume you don't use the extra two bytes
> that you get from the load, as you mark them as "*"
> generic undefined values
>
> > +    vperm2i128              m4, m2, m2, q2001          ;
> ijklmnop........
> > +    vpalignr                m5, m4, m2, 2              ;
> bcdefghijklmnop.
> > +    vperm2i128              m3, m0, m1, q0201          ;
> stuvwxyz*abcdefg
> > +    LOWPASS                  1,  2,  5                 ;
> ABCDEFGHIJKLMNO.
> > +    vpalignr                m4, m3, m0, 2              ;
> lmnopqrstuvwxyz*
> > +    vpalignr                m5, m3, m0, 4              ;
> mnopqrstuvwxyz*a
> > +    LOWPASS                  0,  4,  5                 ;
> LMNOPQRSTUVWXYZ#
> > +    vperm2i128              m5, m0, m1, q0201          ;
> TUVWXYZ#ABCDEFGH
> > +    DEFINE_ARGS dst, stride, stride3, stride5, dst3, cnt
>
> "cnt" doesn't seem to be used.
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>