[x264-devel] commit: SSE4 and SSSE3 versions of some intra_sad functions ( Jason Garrett-Glaser )

git at videolan.org git at videolan.org
Thu Jul 15 04:08:45 CEST 2010


x264 | branch: stable | Jason Garrett-Glaser <darkshikari at gmail.com> | Fri Jun 18 13:58:11 2010 -0700| [ef0c268a2b1a42ff234ef27a535c16b825452ae0] | committer: Jason Garrett-Glaser 

SSE4 and SSSE3 versions of some intra_sad functions
Primarily Nehalem-optimized.

> http://git.videolan.org/gitweb.cgi/x264.git/?a=commit;h=ef0c268a2b1a42ff234ef27a535c16b825452ae0
---

 common/pixel.c       |    3 +
 common/x86/pixel.h   |    2 +
 common/x86/sad-a.asm |  107 ++++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 112 insertions(+), 0 deletions(-)

diff --git a/common/pixel.c b/common/pixel.c
index a8cb1df..8441c7a 100644
--- a/common/pixel.c
+++ b/common/pixel.c
@@ -856,6 +856,9 @@ void x264_pixel_init( int cpu, x264_pixel_function_t *pixf )
         }
         pixf->sa8d[PIXEL_16x16]= x264_pixel_sa8d_16x16_sse4;
         pixf->sa8d[PIXEL_8x8]  = x264_pixel_sa8d_8x8_sse4;
+        pixf->intra_sad_x3_4x4 = x264_intra_sad_x3_4x4_sse4;
+        /* Slower on Conroe, so only enable under SSE4 */
+        pixf->intra_sad_x3_8x8  = x264_intra_sad_x3_8x8_ssse3;
     }
 #endif //HAVE_MMX
 
diff --git a/common/x86/pixel.h b/common/x86/pixel.h
index 9bba683..b1b916d 100644
--- a/common/x86/pixel.h
+++ b/common/x86/pixel.h
@@ -80,6 +80,7 @@ DECL_PIXELS( uint64_t, hadamard_ac, sse4,   ( uint8_t *pix, int i_stride ))
 void x264_intra_satd_x3_4x4_mmxext  ( uint8_t *, uint8_t *, int * );
 void x264_intra_satd_x3_4x4_ssse3   ( uint8_t *, uint8_t *, int * );
 void x264_intra_sad_x3_4x4_mmxext   ( uint8_t *, uint8_t *, int * );
+void x264_intra_sad_x3_4x4_sse4     ( uint8_t *, uint8_t *, int * );
 void x264_intra_satd_x3_8x8c_mmxext ( uint8_t *, uint8_t *, int * );
 void x264_intra_satd_x3_8x8c_ssse3  ( uint8_t *, uint8_t *, int * );
 void x264_intra_sad_x3_8x8c_mmxext  ( uint8_t *, uint8_t *, int * );
@@ -93,6 +94,7 @@ void x264_intra_sa8d_x3_8x8_mmxext  ( uint8_t *, uint8_t *, int * );
 void x264_intra_sa8d_x3_8x8_sse2    ( uint8_t *, uint8_t *, int * );
 void x264_intra_sa8d_x3_8x8_ssse3   ( uint8_t *, uint8_t *, int * );
 void x264_intra_sad_x3_8x8_mmxext   ( uint8_t *, uint8_t *, int * );
+void x264_intra_sad_x3_8x8_ssse3    ( uint8_t *, uint8_t *, int * );
 void x264_intra_sa8d_x3_8x8_core_mmxext( uint8_t *, int16_t [2][8], int * );
 void x264_intra_sa8d_x3_8x8_core_sse2  ( uint8_t *, int16_t [2][8], int * );
 void x264_intra_sa8d_x3_8x8_core_ssse3 ( uint8_t *, int16_t [2][8], int * );
diff --git a/common/x86/sad-a.asm b/common/x86/sad-a.asm
index 72c1789..10a365c 100644
--- a/common/x86/sad-a.asm
+++ b/common/x86/sad-a.asm
@@ -26,6 +26,19 @@
 %include "x86inc.asm"
 %include "x86util.asm"
 
+SECTION_RODATA
+
+h4x4_pred_shuf: db 3,3,3,3,7,7,7,7,11,11,11,11,15,15,15,15
+h4x4_pred_shuf2: db 3,7,11,15,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1
+h8x8_pred_shuf: times 8 db 1
+                times 8 db 0
+                times 8 db 3
+                times 8 db 2
+                times 8 db 5
+                times 8 db 4
+                times 8 db 7
+                times 8 db 6
+
 SECTION .text
 
 cextern pb_3
@@ -303,6 +316,40 @@ cglobal intra_sad_x3_4x4_mmxext, 3,3
     movd   [r2+4], mm1 ;H prediction cost
     RET
 
+cglobal intra_sad_x3_4x4_sse4, 3,3
+    movd       xmm4, [r1+FDEC_STRIDE*0-4]
+    pinsrd     xmm4, [r1+FDEC_STRIDE*1-4], 1
+    pinsrd     xmm4, [r1+FDEC_STRIDE*2-4], 2
+    pinsrd     xmm4, [r1+FDEC_STRIDE*3-4], 3
+    movd       xmm2, [r1-FDEC_STRIDE]
+    pxor       xmm3, xmm3
+    movdqa     xmm5, xmm4
+    pshufb     xmm4, [h4x4_pred_shuf2] ; EFGH
+    pshufb     xmm5, [h4x4_pred_shuf]  ; EEEEFFFFGGGGHHHH
+    pshufd     xmm0, xmm2, 0           ; ABCDABCDABCDABCD
+    punpckldq  xmm2, xmm4              ; ABCDEFGH
+    psadbw     xmm2, xmm3
+    movd       xmm1, [r0+FENC_STRIDE*0]
+    pinsrd     xmm1, [r0+FENC_STRIDE*1], 1
+    pinsrd     xmm1, [r0+FENC_STRIDE*2], 2
+    pinsrd     xmm1, [r0+FENC_STRIDE*3], 3
+    psadbw     xmm0, xmm1
+    psadbw     xmm5, xmm1
+    psraw      xmm2, 2
+    pavgw      xmm2, xmm3
+    pshufb     xmm2, xmm3              ; DC prediction
+    movdqa     xmm3, xmm0
+    punpcklqdq xmm0, xmm5
+    punpckhqdq xmm3, xmm5
+    psadbw     xmm2, xmm1
+    paddw      xmm0, xmm3
+    movhlps    xmm4, xmm2
+    packusdw   xmm0, xmm0
+    paddw      xmm2, xmm4
+    movq       [r2], xmm0              ; V/H prediction costs
+    movd     [r2+8], xmm2              ; DC prediction cost
+    RET
+
 ;-----------------------------------------------------------------------------
 ; void intra_sad_x3_8x8( uint8_t *fenc, uint8_t edge[33], int res[3]);
 ;-----------------------------------------------------------------------------
@@ -370,6 +417,66 @@ cglobal intra_sad_x3_8x8_mmxext, 3,3
     movd  [r2+8], m1
     RET
 
+INIT_XMM
+cglobal intra_sad_x3_8x8_ssse3, 3,4,9
+%ifdef PIC
+    lea        r11, [h8x8_pred_shuf]
+%define shuf r11
+%else
+%define shuf h8x8_pred_shuf
+%endif
+    movq       m0, [r1+7]   ; left pixels
+    movq       m1, [r1+16]  ; top pixels
+    pxor       m2, m2
+    pxor       m3, m3
+    psadbw     m2, m0
+    psadbw     m3, m1
+    paddw      m2, m3
+    pxor       m3, m3       ; V score accumulator
+    psraw      m2, 3
+    pavgw      m2, m3
+    punpcklqdq m1, m1       ; V prediction
+    pshufb     m2, m3       ; DC prediction
+    pxor       m4, m4       ; H score accumulator
+    pxor       m5, m5       ; DC score accumulator
+    mov       r3d, 6
+.loop:
+    movq        m6, [r0+FENC_STRIDE*0]
+    movhps      m6, [r0+FENC_STRIDE*1]
+    movdqa      m7, m0
+    pshufb      m7, [shuf+r3*8] ; H prediction
+%ifdef ARCH_X86_64
+    movdqa      m8, m1
+    psadbw      m7, m6
+    psadbw      m8, m6
+    psadbw      m6, m2
+    paddw       m4, m7
+    paddw       m3, m8
+    paddw       m5, m6
+%else
+    psadbw      m7, m6
+    paddw       m4, m7
+    movdqa      m7, m1
+    psadbw      m7, m6
+    psadbw      m6, m2
+    paddw       m3, m7
+    paddw       m5, m6
+%endif
+    add         r0, FENC_STRIDE*2
+    sub        r3d, 2
+    jge .loop
+
+    movhlps     m0, m3
+    movhlps     m1, m4
+    movhlps     m2, m5
+    paddw       m3, m0
+    paddw       m4, m1
+    paddw       m5, m2
+    movd    [r2+0], m3
+    movd    [r2+4], m4
+    movd    [r2+8], m5
+    RET
+
 ;-----------------------------------------------------------------------------
 ; void intra_sad_x3_8x8c( uint8_t *fenc, uint8_t *fdec, int res[3] );
 ;-----------------------------------------------------------------------------



More information about the x264-devel mailing list