[x265] [PATCH] asm: avx2 code for sad_x3[16xN] for 10 bpp

[Steve Borho]

On 05/18, sumalatha at multicorewareinc.com wrote:
> # HG changeset patch
> # User Sumalatha Polureddy
> # Date 1431930810 -19800
> #      Mon May 18 12:03:30 2015 +0530
> # Node ID e9283ef43b7c4eb32a435a835c4ffac2f1163d70
> # Parent  8592bf81d0848279fa79cd1487406cb516dffe99
> asm: avx2 code for sad_x3[16xN] for 10 bpp
> 
> sse2:
> sad_x3[ 16x4]  2.93x    680.82          1996.91
> sad_x3[ 16x8]  3.03x    1266.26         3834.18
> sad_x3[16x12]  3.07x    1834.17         5631.97
> sad_x3[16x16]  3.06x    2413.24         7380.88
> sad_x3[16x32]  2.82x    5554.36         15654.50
> sad_x3[16x64]  2.80x    10161.18        28493.52

sse2 was the highest SIMD arch we had for these sad primitives? ouch

> avx2:
> sad_x3[ 16x4]  4.82x    404.45          1948.78
> sad_x3[ 16x8]  5.85x    634.65          3714.40
> sad_x3[16x12]  6.17x    885.30          5465.97
> sad_x3[16x16]  6.28x    1170.04         7350.87
> sad_x3[16x32]  5.34x    2909.76         15547.79
> sad_x3[16x64]  6.12x    5071.22         31043.80

this series will go a long way in catching up Main10 perf with Main on
AVX2 capable CPUs

> diff -r 8592bf81d084 -r e9283ef43b7c source/common/x86/asm-primitives.cpp
> --- a/source/common/x86/asm-primitives.cpp	Thu May 14 17:12:14 2015 +0530
> +++ b/source/common/x86/asm-primitives.cpp	Mon May 18 12:03:30 2015 +0530
> @@ -1329,6 +1329,13 @@
>          p.pu[LUMA_64x48].sad = x265_pixel_sad_64x48_avx2;
>          p.pu[LUMA_64x64].sad = x265_pixel_sad_64x64_avx2;
>  
> +        p.pu[LUMA_16x4].sad_x3 = x265_pixel_sad_x3_16x4_avx2;
> +        p.pu[LUMA_16x8].sad_x3 = x265_pixel_sad_x3_16x8_avx2;
> +        p.pu[LUMA_16x12].sad_x3 = x265_pixel_sad_x3_16x12_avx2;
> +        p.pu[LUMA_16x16].sad_x3 = x265_pixel_sad_x3_16x16_avx2;
> +        p.pu[LUMA_16x32].sad_x3 = x265_pixel_sad_x3_16x32_avx2;
> +        p.pu[LUMA_16x64].sad_x3 = x265_pixel_sad_x3_16x64_avx2;
> +
>          p.pu[LUMA_16x4].convert_p2s = x265_filterPixelToShort_16x4_avx2;
>          p.pu[LUMA_16x8].convert_p2s = x265_filterPixelToShort_16x8_avx2;
>          p.pu[LUMA_16x12].convert_p2s = x265_filterPixelToShort_16x12_avx2;
> diff -r 8592bf81d084 -r e9283ef43b7c source/common/x86/sad16-a.asm
> --- a/source/common/x86/sad16-a.asm	Thu May 14 17:12:14 2015 +0530
> +++ b/source/common/x86/sad16-a.asm	Mon May 18 12:03:30 2015 +0530
> @@ -1485,8 +1485,12 @@
>  SAD_X 4,  8,  4
>  INIT_YMM avx2
>  %define XMM_REGS 7
> -SAD_X 3, 16, 16
> +SAD_X 3, 16,  4
>  SAD_X 3, 16,  8
> +SAD_X 3, 16,  12
> +SAD_X 3, 16,  16
> +SAD_X 3, 16,  32
> +SAD_X 3, 16,  64
>  %define XMM_REGS 9
>  SAD_X 4, 16, 16
>  SAD_X 4, 16,  8
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel

-- 
Steve Borho