[x265] [PATCH] asm: Add sse_ss for [16x16], [32x32] & [64x64] for 8bpp avx2

chen chenm003 at 163.com
Mon Oct 12 06:07:21 CEST 2015


At 2015-10-12 11:58:46,ramya at multicorewareinc.com wrote:
># HG changeset patch
># User Ramya Sriraman <ramya at multicorewareinc.com>
># Date 1443592336 -19800
>#      Wed Sep 30 11:22:16 2015 +0530
># Node ID f0d43eb655f0048fe5dd491ee4ad1c0d304b76f8
># Parent  b6156a08b1def3584647f26096866c1a0c11e54a
>asm: Add sse_ss for [16x16],[32x32] & [64x64] for 8bpp avx2
>
>diff -r b6156a08b1de -r f0d43eb655f0 source/common/x86/asm-primitives.cpp
>--- a/source/common/x86/asm-primitives.cpp	Fri Oct 09 20:45:59 2015 +0530
>+++ b/source/common/x86/asm-primitives.cpp	Wed Sep 30 11:22:16 2015 +0530
>@@ -2667,6 +2667,10 @@
> #if X86_64
>     if (cpuMask & X265_CPU_AVX2)
>     {
>+        p.cu[BLOCK_16x16].sse_ss = (pixel_sse_ss_t)PFX(pixel_ssd_ss_16x16_avx2);
>+        p.cu[BLOCK_32x32].sse_ss = (pixel_sse_ss_t)PFX(pixel_ssd_ss_32x32_avx2);
>+        p.cu[BLOCK_64x64].sse_ss = (pixel_sse_ss_t)PFX(pixel_ssd_ss_64x64_avx2);
>+
>         p.cu[BLOCK_16x16].var = PFX(pixel_var_16x16_avx2);
>         p.cu[BLOCK_32x32].var = PFX(pixel_var_32x32_avx2);
>         p.cu[BLOCK_64x64].var = PFX(pixel_var_64x64_avx2);
>diff -r b6156a08b1de -r f0d43eb655f0 source/common/x86/ssd-a.asm
>--- a/source/common/x86/ssd-a.asm	Fri Oct 09 20:45:59 2015 +0530
>+++ b/source/common/x86/ssd-a.asm	Wed Sep 30 11:22:16 2015 +0530
>@@ -1016,8 +1016,171 @@
> SSD_SS_32xN
> SSD_SS_48
> SSD_SS_64xN
>+
>+INIT_YMM avx2
>+cglobal pixel_ssd_ss_16x16, 4,4,3
>+    add     r1d, r1d
>+    add     r3d, r3d
>+    pxor    m2, m2
>+
>+    movu        m0, [r0]
>+    movu        m1, [r0 + r1]
>+    psubw       m0, [r2]
>+    psubw       m1, [r2 + r3]
>+    lea         r0, [r0 + 2 * r1]
>+    lea         r2, [r2 + 2 * r3]
we have more register to buffer r1*3, it will reduce number of LEA

>+    pmaddwd     m0, m0
>+    pmaddwd     m1, m1
>+    paddd       m2, m1
>+    paddd       m2, m0
use m2 and m3 for partial sum, it may broken long dependency link
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20151012/81bd98bf/attachment-0001.html>


More information about the x265-devel mailing list