Use sa8d instead of satd for i8x8 search.

+.01 dB, -.5% speed

git-svn-id: svn://svn.videolan.org/x264/trunk@512 df754926-b1dd-0310-bc7b-ec298dee348c

diff --git a/common/i386/pixel-a.asm b/common/i386/pixel-a.asm
index 834841e860e7bd9bea51488b2db793970bd2937a..e7278fd68d629239d58b04247d1486ebea731602 100644 (file)
--- a/common/i386/pixel-a.asm
+++ b/common/i386/pixel-a.asm
@@ -847,14 +847,14 @@ x264_pixel_satd_16x16_mmxext:
 %endmacro
 
 %macro SUM4x8_MM 0
-    movq [spill], mm7
-    MMX_ABS  mm0, mm7
-    MMX_ABS  mm1, mm7
-    MMX_ABS  mm2, mm7
-    MMX_ABS  mm3, mm7
+    movq [spill],   mm6
+    movq [spill+8], mm7
+    MMX_ABS_TWO  mm0, mm1, mm6, mm7
+    MMX_ABS_TWO  mm2, mm3, mm6, mm7
     paddw    mm0, mm2
     paddw    mm1, mm3
-    movq     mm7, [spill]
+    movq     mm6, [spill]
+    movq     mm7, [spill+8]
     MMX_ABS_TWO  mm4, mm5, mm2, mm3
     MMX_ABS_TWO  mm6, mm7, mm2, mm3
     paddw    mm4, mm6
@@ -870,14 +870,14 @@ ALIGN 16
 ;-----------------------------------------------------------------------------
 x264_pixel_sa8d_8x8_mmxext:
     SATD_START
-    sub    esp, 0x68
-%define args  esp+0x6c
+    sub    esp, 0x70
+%define args  esp+0x74
 %define spill esp+0x60
     LOAD_DIFF_4x8P 0
     HADAMARD1x8 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7
 
     movq   [spill], mm0
-    TRANSPOSE4x4 mm4, mm5, mm6, mm7, mm0 ; abcd-t -> adtc
+    TRANSPOSE4x4 mm4, mm5, mm6, mm7, mm0
     movq   [esp+0x00], mm4
     movq   [esp+0x08], mm7
     movq   [esp+0x10], mm0
@@ -894,13 +894,13 @@ x264_pixel_sa8d_8x8_mmxext:
     LOAD_DIFF_4x8P 4
     HADAMARD1x8 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7
 
-    movq   [spill], mm4
-    TRANSPOSE4x4 mm0, mm1, mm2, mm3, mm4
+    movq   [spill], mm7
+    TRANSPOSE4x4 mm0, mm1, mm2, mm3, mm7
     movq   [esp+0x40], mm0
     movq   [esp+0x48], mm3
-    movq   [esp+0x50], mm4
+    movq   [esp+0x50], mm7
     movq   [esp+0x58], mm2
-    movq   mm4, [spill]
+    movq   mm7, [spill]
     TRANSPOSE4x4 mm4, mm5, mm6, mm7, mm0
     movq   mm5, [esp+0x00]
     movq   mm1, [esp+0x08]
@@ -933,7 +933,7 @@ x264_pixel_sa8d_8x8_mmxext:
     mov    ecx, eax ; preserve rounding for 16x16
     add    eax, 1
     shr    eax, 1
-    add    esp, 0x68
+    add    esp, 0x70
     pop    ebx
     ret
 %undef args

diff --git a/encoder/analyse.c b/encoder/analyse.c
index 819c51ff41ba2d4111808451404b83d76f39a944..95f111610454071b25a024a9e3de4ae131e2da12 100644 (file)
--- a/encoder/analyse.c
+++ b/encoder/analyse.c
@@ -496,6 +496,7 @@ static void x264_mb_analyse_intra( x264_t *h, x264_mb_analysis_t *a, int i_satd_
     /* 8x8 prediction selection */
     if( flags & X264_ANALYSE_I8x8 )
     {
+        x264_pixel_cmp_t sa8d = (*h->pixf.mbcmp == *h->pixf.sad) ? h->pixf.sad[PIXEL_8x8] : h->pixf.sa8d[PIXEL_8x8];
         int i_satd_thresh = a->b_mbrd ? COST_MAX : X264_MIN( i_satd_inter, a->i_satd_i16x16 );
         int i_cost = 0;
 
@@ -520,10 +521,8 @@ static void x264_mb_analyse_intra( x264_t *h, x264_mb_analysis_t *a, int i_satd_
 
                 h->predict_8x8[i_mode]( p_dst_by, h->mb.i_neighbour8[idx] );
 
-                /* could use sa8d, but it doesn't seem worth the speed cost (without mmx at least) */
-                i_satd = h->pixf.mbcmp[PIXEL_8x8]( p_dst_by, FDEC_STRIDE,
-                                                  p_src_by, FENC_STRIDE )
-                      + a->i_lambda * (i_pred_mode == x264_mb_pred_mode4x4_fix(i_mode) ? 1 : 4);
+                i_satd = sa8d( p_dst_by, FDEC_STRIDE, p_src_by, FENC_STRIDE )
+                       + a->i_lambda * (i_pred_mode == x264_mb_pred_mode4x4_fix(i_mode) ? 1 : 4);
 
                 COPY2_IF_LT( i_best, i_satd, a->i_predict8x8[idx], i_mode );
                 a->i_satd_i8x8_dir[i_mode][idx] = i_satd;
@@ -588,7 +587,7 @@ static void x264_mb_analyse_intra( x264_t *h, x264_mb_analysis_t *a, int i_satd_
 
                 i_satd = h->pixf.mbcmp[PIXEL_4x4]( p_dst_by, FDEC_STRIDE,
                                                   p_src_by, FENC_STRIDE )
-                      + a->i_lambda * (i_pred_mode == x264_mb_pred_mode4x4_fix(i_mode) ? 1 : 4);
+                       + a->i_lambda * (i_pred_mode == x264_mb_pred_mode4x4_fix(i_mode) ? 1 : 4);
 
                 COPY2_IF_LT( i_best, i_satd, a->i_predict4x4[idx], i_mode );
             }