[x265] [PATCH] asm: 16bpp support for satd_64xN
yuvaraj at multicorewareinc.com
yuvaraj at multicorewareinc.com
Mon Dec 2 14:21:45 CET 2013
# HG changeset patch
# User Yuvaraj Venkatesh <yuvaraj at multicorewareinc.com>
# Date 1385990487 -19800
# Mon Dec 02 18:51:27 2013 +0530
# Node ID 55597521ae176f6628c4bf3951b0e454be2f33eb
# Parent fa3a3eced7228599400f9403dba159d433d05222
asm: 16bpp support for satd_64xN
diff -r fa3a3eced722 -r 55597521ae17 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Mon Dec 02 17:50:12 2013 +0530
+++ b/source/common/x86/asm-primitives.cpp Mon Dec 02 18:51:27 2013 +0530
@@ -489,6 +489,8 @@
if (cpuMask & X265_CPU_SSE2)
{
INIT6(satd, _sse2);
+ HEVC_SATD(sse2);
+ p.satd[LUMA_4x4] = x265_pixel_satd_4x4_mmx2;
p.satd[LUMA_4x16] = x265_pixel_satd_4x16_sse2;
p.satd[LUMA_8x32] = x265_pixel_satd_8x32_sse2;
p.satd[LUMA_16x4] = x265_pixel_satd_16x4_sse2;
@@ -496,14 +498,9 @@
p.satd[LUMA_16x32] = x265_pixel_satd_16x32_sse2;
p.satd[LUMA_16x64] = x265_pixel_satd_16x64_sse2;
p.satd[LUMA_12x16] = x265_pixel_satd_12x16_sse2;
- p.satd[LUMA_24x32] = x265_pixel_satd_24x32_sse2;
- p.satd[LUMA_48x64] = x265_pixel_satd_48x64_sse2;
p.satd[LUMA_32x8] = x265_pixel_satd_32x8_sse2;
p.satd[LUMA_32x16] = x265_pixel_satd_32x16_sse2;
p.satd[LUMA_32x24] = x265_pixel_satd_32x24_sse2;
- p.satd[LUMA_32x32] = x265_pixel_satd_32x32_sse2;
- p.satd[LUMA_32x64] = x265_pixel_satd_32x64_sse2;
-
p.sa8d_inter[LUMA_8x8] = x265_pixel_sa8d_8x8_sse2;
p.sa8d_inter[LUMA_16x16] = x265_pixel_sa8d_16x16_sse2;
diff -r fa3a3eced722 -r 55597521ae17 source/common/x86/pixel-a.asm
--- a/source/common/x86/pixel-a.asm Mon Dec 02 17:50:12 2013 +0530
+++ b/source/common/x86/pixel-a.asm Mon Dec 02 18:51:27 2013 +0530
@@ -4255,35 +4255,42 @@
mov r7, r2
call pixel_satd_8x8_internal
call pixel_satd_8x8_internal
- lea r0, [r6 + 8]
- lea r2, [r7 + 8]
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 16]
- lea r2, [r7 + 16]
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 24]
- lea r2, [r7 + 24]
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 32]
- lea r2, [r7 + 32]
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 40]
- lea r2, [r7 + 40]
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 48]
- lea r2, [r7 + 48]
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 56]
- lea r2, [r7 + 56]
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- SATD_END_SSE2 m6
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 8*SIZEOF_PIXEL]
+ lea r2, [r7 + 8*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 16*SIZEOF_PIXEL]
+ lea r2, [r7 + 16*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 24*SIZEOF_PIXEL]
+ lea r2, [r7 + 24*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 32*SIZEOF_PIXEL]
+ lea r2, [r7 + 32*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 40*SIZEOF_PIXEL]
+ lea r2, [r7 + 40*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 48*SIZEOF_PIXEL]
+ lea r2, [r7 + 48*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 56*SIZEOF_PIXEL]
+ lea r2, [r7 + 56*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_END_SSE2 m6, m7
%else
cglobal pixel_satd_64x16, 4,7,8,0-4 ;if !WIN64
SATD_START_SSE2 m6, m7
@@ -4291,42 +4298,50 @@
mov [rsp], r2
call pixel_satd_8x8_internal
call pixel_satd_8x8_internal
- lea r0, [r6 + 8]
+ pxor m7, m7
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 8*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2,8
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 16]
+ add r2,8*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 16*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2,16
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 24]
+ add r2,16*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 24*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2,24
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 32]
+ add r2,24*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 32*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2,32
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 40]
+ add r2,32*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 40*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2,40
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 48]
+ add r2,40*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 48*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2,48
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- lea r0, [r6 + 56]
+ add r2,48*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_ACCUM m6, m0, m7
+ lea r0, [r6 + 56*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2,56
- call pixel_satd_8x8_internal
- call pixel_satd_8x8_internal
- SATD_END_SSE2 m6
+ add r2,56*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_END_SSE2 m6, m7
%endif
%if WIN64
@@ -4338,44 +4353,44 @@
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
- lea r0, [r6 + 8]
- lea r2, [r7 + 8]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 16]
- lea r2, [r7 + 16]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 24]
- lea r2, [r7 + 24]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 32]
- lea r2, [r7 + 32]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 40]
- lea r2, [r7 + 40]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 48]
- lea r2, [r7 + 48]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 56]
- lea r2, [r7 + 56]
+ lea r0, [r6 + 8*SIZEOF_PIXEL]
+ lea r2, [r7 + 8*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 16*SIZEOF_PIXEL]
+ lea r2, [r7 + 16*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 24*SIZEOF_PIXEL]
+ lea r2, [r7 + 24*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 32*SIZEOF_PIXEL]
+ lea r2, [r7 + 32*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 40*SIZEOF_PIXEL]
+ lea r2, [r7 + 40*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 48*SIZEOF_PIXEL]
+ lea r2, [r7 + 48*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 56*SIZEOF_PIXEL]
+ lea r2, [r7 + 56*SIZEOF_PIXEL]
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
@@ -4396,51 +4411,51 @@
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
- lea r0, [r6 + 8]
+ lea r0, [r6 + 8*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 8
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 16]
+ add r2, 8*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 16*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 16
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 24]
+ add r2, 16*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 24*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 24
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 32]
+ add r2, 24*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 32*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 32
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 40]
+ add r2, 32*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 40*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 40
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 48]
+ add r2, 40*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 48*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 48
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 56]
+ add r2, 48*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 56*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 56
+ add r2, 56*SIZEOF_PIXEL
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
@@ -4465,56 +4480,56 @@
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
- lea r0, [r6 + 8]
- lea r2, [r7 + 8]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 16]
- lea r2, [r7 + 16]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 24]
- lea r2, [r7 + 24]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 32]
- lea r2, [r7 + 32]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 40]
- lea r2, [r7 + 40]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 48]
- lea r2, [r7 + 48]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 56]
- lea r2, [r7 + 56]
+ lea r0, [r6 + 8*SIZEOF_PIXEL]
+ lea r2, [r7 + 8*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 16*SIZEOF_PIXEL]
+ lea r2, [r7 + 16*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 24*SIZEOF_PIXEL]
+ lea r2, [r7 + 24*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 32*SIZEOF_PIXEL]
+ lea r2, [r7 + 32*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 40*SIZEOF_PIXEL]
+ lea r2, [r7 + 40*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 48*SIZEOF_PIXEL]
+ lea r2, [r7 + 48*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 56*SIZEOF_PIXEL]
+ lea r2, [r7 + 56*SIZEOF_PIXEL]
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
@@ -4539,63 +4554,63 @@
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
- lea r0, [r6 + 8]
+ lea r0, [r6 + 8*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 8
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 16]
+ add r2, 8*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 16*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 16
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 24]
+ add r2, 16*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 24*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 24
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 32]
+ add r2, 24*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 32*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 32
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 40]
+ add r2, 32*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 40*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 40
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 48]
+ add r2, 40*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 48*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 48
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 56]
+ add r2, 48*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 56*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 56
+ add r2, 56*SIZEOF_PIXEL
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
@@ -4624,68 +4639,68 @@
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
- lea r0, [r6 + 8]
- lea r2, [r7 + 8]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 16]
- lea r2, [r7 + 16]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 24]
- lea r2, [r7 + 24]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 32]
- lea r2, [r7 + 32]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 40]
- lea r2, [r7 + 40]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 48]
- lea r2, [r7 + 48]
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 56]
- lea r2, [r7 + 56]
+ lea r0, [r6 + 8*SIZEOF_PIXEL]
+ lea r2, [r7 + 8*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 16*SIZEOF_PIXEL]
+ lea r2, [r7 + 16*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 24*SIZEOF_PIXEL]
+ lea r2, [r7 + 24*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 32*SIZEOF_PIXEL]
+ lea r2, [r7 + 32*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 40*SIZEOF_PIXEL]
+ lea r2, [r7 + 40*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 48*SIZEOF_PIXEL]
+ lea r2, [r7 + 48*SIZEOF_PIXEL]
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 56*SIZEOF_PIXEL]
+ lea r2, [r7 + 56*SIZEOF_PIXEL]
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
@@ -4714,75 +4729,75 @@
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
- lea r0, [r6 + 8]
+ lea r0, [r6 + 8*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 8
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 16]
+ add r2, 8*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 16*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 16
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 24]
+ add r2, 16*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 24*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 24
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 32]
+ add r2, 24*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 32*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 32
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 40]
+ add r2, 32*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 40*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 40
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 48]
+ add r2, 40*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 48*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 48
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- call pixel_satd_8x8_internal2
- lea r0, [r6 + 56]
+ add r2, 48*SIZEOF_PIXEL
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ call pixel_satd_8x8_internal2
+ lea r0, [r6 + 56*SIZEOF_PIXEL]
mov r2, [rsp]
- add r2, 56
+ add r2, 56*SIZEOF_PIXEL
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
call pixel_satd_8x8_internal2
More information about the x265-devel
mailing list