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

[Ivan Kalvachev]

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

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

>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.