<div style="line-height:1.7;color:#000000;font-size:14px;font-family:arial"><div> </div><pre><br>At 2015-03-11 17:48:16,sumalatha@multicorewareinc.com wrote:
># HG changeset patch
># User Sumalatha Polureddy<sumalatha@multicorewareinc.com>
># Date 1426067286 -19800
># Node ID bc9f8dbe8f655fe0b1a218bb6abb8fa88536c6dd
># Parent 01bfd365bf5f5317874b5c0315736ca76176f3df
>asm: avx2 code for sad[64x64] for 8bpp
>
>SSE3
>sad[64x64] 33.47x 2584.70 86508.58
>
>AVX2
>sad[64x64] 34.42x 2557.98 88038.73
</pre><pre>same speed?</pre><pre> </pre><pre>>diff -r 01bfd365bf5f -r bc9f8dbe8f65 source/common/x86/asm-primitives.cpp
>--- a/source/common/x86/asm-primitives.cpp Wed Mar 11 14:24:10 2015 +0530
>+++ b/source/common/x86/asm-primitives.cpp Wed Mar 11 15:18:06 2015 +0530
>@@ -1443,6 +1443,7 @@
> p.pu[LUMA_8x8].satd = x265_pixel_satd_8x8_avx2;
>
> p.pu[LUMA_32x32].sad = x265_pixel_sad_32x32_avx2;
>+ p.pu[LUMA_64x64].sad = x265_pixel_sad_64x64_avx2;
>
> p.pu[LUMA_8x4].sad_x3 = x265_pixel_sad_x3_8x4_avx2;
> p.pu[LUMA_8x8].sad_x3 = x265_pixel_sad_x3_8x8_avx2;
>diff -r 01bfd365bf5f -r bc9f8dbe8f65 source/common/x86/sad-a.asm
>--- a/source/common/x86/sad-a.asm Wed Mar 11 14:24:10 2015 +0530
>+++ b/source/common/x86/sad-a.asm Wed Mar 11 15:18:06 2015 +0530
>@@ -3924,4 +3924,44 @@
> movd eax, xm0
> RET
>
>+INIT_YMM avx2
>+cglobal pixel_sad_64x64, 4,4,5
>+ xorps m0, m0
>+
>+%assign x 0
>+%rep 32
</pre><pre>could you try to compare loop version performance?</pre><pre> </pre><pre>>+ movu m1, [r0] ; first 32 of row 0 of pix0
>+ movu m2, [r2] ; first 32 of row 0 of pix1
>+ movu m3, [r0 + 32] ; second 32 of row 0 of pix0
>+ movu m4, [r2 + 32] ; second 32 of row 0 of pix1
>+
>+ psadbw m1, m2
>+ psadbw m3, m4
>+ paddd m0, m1
>+ paddd m0, m3
</pre><pre>in here, we have more free register, so sum parallel is better, eg: m0+=m1, m2+=m3</pre><pre>>+
>+ movu m1, [r0 + r1] ; first 32 of row 1 of pix0
>+ movu m2, [r2 + r1] ; first 32 of row 1 of pix1
>+ movu m3, [r0 + 32 + r1] ; second 32 of row 1 of pix0
>+ movu m4, [r2 + 32 + r1] ; second 32 of row 1 of pix1
>+
>+ psadbw m1, m2
>+ psadbw m3, m4
>+ paddd m0, m1
>+ paddd m0, m3
>+
>+%assign x x+1
>+ %if x < 32
>+ lea r2, [r2 + 2 * r3]
>+ lea r0, [r0 + 2 * r1]
>+ %endif
>+%endrep
>+
>+ vextracti128 xm1, m0, 1
>+ paddd xm0, xm1
>+ pshufd xm1, xm0, 2
>+ paddd xm0,xm1
>+ movd eax, xm0
>+ RET
>+
> %endif
>_______________________________________________
>x265-devel mailing list
>x265-devel@videolan.org
>https://mailman.videolan.org/listinfo/x265-devel
</pre></div>