[x264-devel] x86: AVX2 high bit-depth pixel_sad_x3/pixel_sad_x4

[Henrik Gramner]

x264 | branch: master | Henrik Gramner <henrik at gramner.com> | Tue Apr 16 23:27:46 2013 +0200| [c2852be748c66f1ff25f38133d5efbd6059bed6c] | committer: Jason Garrett-Glaser

x86: AVX2 high bit-depth pixel_sad_x3/pixel_sad_x4

Also reduce the number of xmm registers used by sse2/ssse3 pixel_sad_x3.

> http://git.videolan.org/gitweb.cgi/x264.git/?a=commit;h=c2852be748c66f1ff25f38133d5efbd6059bed6c
---

 common/pixel.c         |    2 ++
 common/x86/sad16-a.asm |   35 ++++++++++++++++++++++-------------
 2 files changed, 24 insertions(+), 13 deletions(-)

diff --git a/common/pixel.c b/common/pixel.c
index bbe43be..13b566c 100644
--- a/common/pixel.c
+++ b/common/pixel.c
@@ -991,6 +991,8 @@ void x264_pixel_init( int cpu, x264_pixel_function_t *pixf )
     {
         INIT2( sad, _avx2 );
         INIT2_NAME( sad_aligned, sad, _avx2 );
+        INIT2( sad_x3, _avx2 );
+        INIT2( sad_x4, _avx2 );
         pixf->vsad = x264_pixel_vsad_avx2;
     }
 #endif // HAVE_MMX
diff --git a/common/x86/sad16-a.asm b/common/x86/sad16-a.asm
index a3a6892..8c48723 100644
--- a/common/x86/sad16-a.asm
+++ b/common/x86/sad16-a.asm
@@ -258,14 +258,14 @@ SAD 16,  8
     HADDW    m2, m5
 %endif
 %if UNIX64
-    movd [r5+0], m0
-    movd [r5+4], m1
-    movd [r5+8], m2
+    movd [r5+0], xm0
+    movd [r5+4], xm1
+    movd [r5+8], xm2
 %else
     mov      r0, r5mp
-    movd [r0+0], m0
-    movd [r0+4], m1
-    movd [r0+8], m2
+    movd [r0+0], xm0
+    movd [r0+4], xm1
+    movd [r0+8], xm2
 %endif
     RET
 %endmacro
@@ -354,10 +354,10 @@ SAD 16,  8
     HADDW     m3, m7
 %endif
     mov       r0, r6mp
-    movd [r0+ 0], m0
-    movd [r0+ 4], m1
-    movd [r0+ 8], m2
-    movd [r0+12], m3
+    movd [r0+ 0], xm0
+    movd [r0+ 4], xm1
+    movd [r0+ 8], xm2
+    movd [r0+12], xm3
     RET
 %endmacro
 
@@ -453,7 +453,7 @@ cglobal pixel_vsad, 3,3
     RET
 
 ;-----------------------------------------------------------------------------
-; void pixel_sad_xK_MxN( uint16_t *fenc, uint16_t *pix0, uint16_t *pix1,
+; void pixel_sad_xN_WxH( uint16_t *fenc, uint16_t *pix0, uint16_t *pix1,
 ;                        uint16_t *pix2, intptr_t i_stride, int scores[3] )
 ;-----------------------------------------------------------------------------
 %macro SAD_X 3
@@ -497,29 +497,38 @@ SAD_X 3,  4,  4
 SAD_X 4,  4,  8
 SAD_X 4,  4,  4
 INIT_XMM ssse3
-%define XMM_REGS 9
+%define XMM_REGS 7
 SAD_X 3, 16, 16
 SAD_X 3, 16,  8
 SAD_X 3,  8, 16
 SAD_X 3,  8,  8
 SAD_X 3,  8,  4
+%define XMM_REGS 9
 SAD_X 4, 16, 16
 SAD_X 4, 16,  8
 SAD_X 4,  8, 16
 SAD_X 4,  8,  8
 SAD_X 4,  8,  4
 INIT_XMM sse2
-%define XMM_REGS 11
+%define XMM_REGS 8
 SAD_X 3, 16, 16
 SAD_X 3, 16,  8
 SAD_X 3,  8, 16
 SAD_X 3,  8,  8
 SAD_X 3,  8,  4
+%define XMM_REGS 11
 SAD_X 4, 16, 16
 SAD_X 4, 16,  8
 SAD_X 4,  8, 16
 SAD_X 4,  8,  8
 SAD_X 4,  8,  4
+INIT_YMM avx2
+%define XMM_REGS 7
+SAD_X 3, 16, 16
+SAD_X 3, 16,  8
+%define XMM_REGS 9
+SAD_X 4, 16, 16
+SAD_X 4, 16,  8
 
 ;-----------------------------------------------------------------------------
 ; void intra_sad_x3_4x4( uint16_t *fenc, uint16_t *fdec, int res[3] );