put/avg_pixels16

fixing 2 small qpel bugs

Originally committed as revision 915 to svn://svn.ffmpeg.org/ffmpeg/trunk

3 files changed, 325 insertions, 61 deletions

diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index fed1818743..4336e4bde0 100644
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -343,7 +343,7 @@ static void add_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line
     } while (--i);
 }
 
-static void put_pixels_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void put_pixels8_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     __asm __volatile(
 	 "lea (%3, %3), %%eax		\n\t"
@@ -369,6 +369,40 @@ static void put_pixels_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int
 	);
 }
 
+static void put_pixels16_mmx(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+{
+    __asm __volatile(
+	 "lea (%3, %3), %%eax		\n\t"
+	 ".balign 8			\n\t"
+	 "1:				\n\t"
+	 "movq (%1), %%mm0		\n\t"
+	 "movq 8(%1), %%mm4		\n\t"
+	 "movq (%1, %3), %%mm1		\n\t"
+	 "movq 8(%1, %3), %%mm5		\n\t"
+     	 "movq %%mm0, (%2)		\n\t"
+     	 "movq %%mm4, 8(%2)		\n\t"
+	 "movq %%mm1, (%2, %3)		\n\t"
+	 "movq %%mm5, 8(%2, %3)		\n\t"
+	 "addl %%eax, %1		\n\t"
+         "addl %%eax, %2       		\n\t"
+	 "movq (%1), %%mm0		\n\t"
+	 "movq 8(%1), %%mm4		\n\t"
+	 "movq (%1, %3), %%mm1		\n\t"
+	 "movq 8(%1, %3), %%mm5		\n\t"
+	 "movq %%mm0, (%2)		\n\t"
+	 "movq %%mm4, 8(%2)		\n\t"
+	 "movq %%mm1, (%2, %3)		\n\t"
+	 "movq %%mm5, 8(%2, %3)		\n\t"
+	 "addl %%eax, %1		\n\t"
+	 "addl %%eax, %2       		\n\t"
+	 "subl $4, %0			\n\t"
+	 "jnz 1b			\n\t"
+	 : "+g"(h), "+r" (pixels),  "+r" (block)
+	 : "r"(line_size)
+	 : "%eax", "memory"
+	);
+}
+
 static void clear_blocks_mmx(DCTELEM *blocks)
 {
     __asm __volatile(
@@ -424,25 +458,45 @@ void dsputil_init_mmx(void)
         pix_abs8x8_y2 = pix_abs8x8_y2_mmx;
         pix_abs8x8_xy2= pix_abs8x8_xy2_mmx;
 
-        put_pixels_tab[0] = put_pixels_mmx;
-        put_pixels_tab[1] = put_pixels_x2_mmx;
-        put_pixels_tab[2] = put_pixels_y2_mmx;
-        put_pixels_tab[3] = put_pixels_xy2_mmx;
-
-        put_no_rnd_pixels_tab[0] = put_pixels_mmx;
-        put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx;
-        put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx;
-        put_no_rnd_pixels_tab[3] = put_no_rnd_pixels_xy2_mmx;
-
-        avg_pixels_tab[0] = avg_pixels_mmx;
-        avg_pixels_tab[1] = avg_pixels_x2_mmx;
-        avg_pixels_tab[2] = avg_pixels_y2_mmx;
-        avg_pixels_tab[3] = avg_pixels_xy2_mmx;
-
-        avg_no_rnd_pixels_tab[0] = avg_no_rnd_pixels_mmx;
-        avg_no_rnd_pixels_tab[1] = avg_no_rnd_pixels_x2_mmx;
-        avg_no_rnd_pixels_tab[2] = avg_no_rnd_pixels_y2_mmx;
-        avg_no_rnd_pixels_tab[3] = avg_no_rnd_pixels_xy2_mmx;
+        put_pixels_tab[0][0] = put_pixels16_mmx;
+        put_pixels_tab[0][1] = put_pixels16_x2_mmx;
+        put_pixels_tab[0][2] = put_pixels16_y2_mmx;
+        put_pixels_tab[0][3] = put_pixels16_xy2_mmx;
+
+        put_no_rnd_pixels_tab[0][0] = put_pixels16_mmx;
+        put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx;
+        put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx;
+        put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_mmx;
+
+        avg_pixels_tab[0][0] = avg_pixels16_mmx;
+        avg_pixels_tab[0][1] = avg_pixels16_x2_mmx;
+        avg_pixels_tab[0][2] = avg_pixels16_y2_mmx;
+        avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx;
+
+        avg_no_rnd_pixels_tab[0][0] = avg_no_rnd_pixels16_mmx;
+        avg_no_rnd_pixels_tab[0][1] = avg_no_rnd_pixels16_x2_mmx;
+        avg_no_rnd_pixels_tab[0][2] = avg_no_rnd_pixels16_y2_mmx;
+        avg_no_rnd_pixels_tab[0][3] = avg_no_rnd_pixels16_xy2_mmx;
+        
+        put_pixels_tab[1][0] = put_pixels8_mmx;
+        put_pixels_tab[1][1] = put_pixels8_x2_mmx;
+        put_pixels_tab[1][2] = put_pixels8_y2_mmx;
+        put_pixels_tab[1][3] = put_pixels8_xy2_mmx;
+
+        put_no_rnd_pixels_tab[1][0] = put_pixels8_mmx;
+        put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx;
+        put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx;
+        put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_mmx;
+
+        avg_pixels_tab[1][0] = avg_pixels8_mmx;
+        avg_pixels_tab[1][1] = avg_pixels8_x2_mmx;
+        avg_pixels_tab[1][2] = avg_pixels8_y2_mmx;
+        avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx;
+
+        avg_no_rnd_pixels_tab[1][0] = avg_no_rnd_pixels8_mmx;
+        avg_no_rnd_pixels_tab[1][1] = avg_no_rnd_pixels8_x2_mmx;
+        avg_no_rnd_pixels_tab[1][2] = avg_no_rnd_pixels8_y2_mmx;
+        avg_no_rnd_pixels_tab[1][3] = avg_no_rnd_pixels8_xy2_mmx;
 
         if (mm_flags & MM_MMXEXT) {
             pix_abs16x16    = pix_abs16x16_mmx2;
@@ -455,25 +509,45 @@ void dsputil_init_mmx(void)
             pix_abs8x8_y2 = pix_abs8x8_y2_mmx2;
             pix_abs8x8_xy2= pix_abs8x8_xy2_mmx2;
 
-            put_pixels_tab[1] = put_pixels_x2_mmx2;
-            put_pixels_tab[2] = put_pixels_y2_mmx2;
-            put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx2;
-            put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx2;
-
-            avg_pixels_tab[0] = avg_pixels_mmx2;
-            avg_pixels_tab[1] = avg_pixels_x2_mmx2;
-            avg_pixels_tab[2] = avg_pixels_y2_mmx2;
-            avg_pixels_tab[3] = avg_pixels_xy2_mmx2;
+            put_pixels_tab[0][1] = put_pixels16_x2_mmx2;
+            put_pixels_tab[0][2] = put_pixels16_y2_mmx2;
+            put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx2;
+            put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx2;
+
+            avg_pixels_tab[0][0] = avg_pixels16_mmx2;
+            avg_pixels_tab[0][1] = avg_pixels16_x2_mmx2;
+            avg_pixels_tab[0][2] = avg_pixels16_y2_mmx2;
+            avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx2;
+
+            put_pixels_tab[1][1] = put_pixels8_x2_mmx2;
+            put_pixels_tab[1][2] = put_pixels8_y2_mmx2;
+            put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx2;
+            put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx2;
+
+            avg_pixels_tab[1][0] = avg_pixels8_mmx2;
+            avg_pixels_tab[1][1] = avg_pixels8_x2_mmx2;
+            avg_pixels_tab[1][2] = avg_pixels8_y2_mmx2;
+            avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx2;
         } else if (mm_flags & MM_3DNOW) {
-            put_pixels_tab[1] = put_pixels_x2_3dnow;
-            put_pixels_tab[2] = put_pixels_y2_3dnow;
-            put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_3dnow;
-            put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_3dnow;
-
-            avg_pixels_tab[0] = avg_pixels_3dnow;
-            avg_pixels_tab[1] = avg_pixels_x2_3dnow;
-            avg_pixels_tab[2] = avg_pixels_y2_3dnow;
-            avg_pixels_tab[3] = avg_pixels_xy2_3dnow;
+            put_pixels_tab[0][1] = put_pixels16_x2_3dnow;
+            put_pixels_tab[0][2] = put_pixels16_y2_3dnow;
+            put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_3dnow;
+            put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_3dnow;
+
+            avg_pixels_tab[0][0] = avg_pixels16_3dnow;
+            avg_pixels_tab[0][1] = avg_pixels16_x2_3dnow;
+            avg_pixels_tab[0][2] = avg_pixels16_y2_3dnow;
+            avg_pixels_tab[0][3] = avg_pixels16_xy2_3dnow;
+            
+            put_pixels_tab[1][1] = put_pixels8_x2_3dnow;
+            put_pixels_tab[1][2] = put_pixels8_y2_3dnow;
+            put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_3dnow;
+            put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_3dnow;
+
+            avg_pixels_tab[1][0] = avg_pixels8_3dnow;
+            avg_pixels_tab[1][1] = avg_pixels8_x2_3dnow;
+            avg_pixels_tab[1][2] = avg_pixels8_y2_3dnow;
+            avg_pixels_tab[1][3] = avg_pixels8_xy2_3dnow;
         }
 
         /* idct */
@@ -552,21 +626,22 @@ void bit_exact_idct_put(UINT8 *dest, int line_size, INT16 *block){
 void dsputil_set_bit_exact_mmx(void)
 {
     if (mm_flags & MM_MMX) {
-        if (mm_flags & MM_MMXEXT) {
-            put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx;
-            put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx;
-            avg_pixels_tab[3] = avg_pixels_xy2_mmx;
+    
+        /* MMX2 & 3DNOW */
+        put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_mmx;
+        put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_mmx;
+        avg_pixels_tab[0][3] = avg_pixels16_xy2_mmx;
+        put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels8_x2_mmx;
+        put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels8_y2_mmx;
+        avg_pixels_tab[1][3] = avg_pixels8_xy2_mmx;
 
+        if (mm_flags & MM_MMXEXT) {
             pix_abs16x16_x2  = pix_abs16x16_x2_mmx;
             pix_abs16x16_y2  = pix_abs16x16_y2_mmx;
             pix_abs16x16_xy2 = pix_abs16x16_xy2_mmx;
             pix_abs8x8_x2 = pix_abs8x8_x2_mmx;
             pix_abs8x8_y2 = pix_abs8x8_y2_mmx;
             pix_abs8x8_xy2= pix_abs8x8_xy2_mmx;
-        } else if (mm_flags & MM_3DNOW) {
-            put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx;
-            put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx;
-            avg_pixels_tab[3] = avg_pixels_xy2_mmx;
         }
 #ifdef SIMPLE_IDCT
         if(ff_idct_put==gen_idct_put && ff_idct == simple_idct_mmx)
diff --git a/libavcodec/i386/dsputil_mmx_avg.h b/libavcodec/i386/dsputil_mmx_avg.h
index a16ccc88b0..6873432ce8 100644
--- a/libavcodec/i386/dsputil_mmx_avg.h
+++ b/libavcodec/i386/dsputil_mmx_avg.h
@@ -25,7 +25,7 @@
 /* XXX: we use explicit registers to avoid a gcc 2.95.2 register asm
    clobber bug - now it will work with 2.95.2 and also with -fPIC
  */
-static void DEF(put_pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(put_pixels8_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     __asm __volatile(
 	"lea (%3, %3), %%eax		\n\t"
@@ -52,9 +52,49 @@ static void DEF(put_pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size,
 	:"r" (line_size)
 	:"%eax", "memory");
 }
+
+static void DEF(put_pixels16_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+{
+    __asm __volatile(
+	"lea (%3, %3), %%eax		\n\t"
+	"1:				\n\t"
+	"movq (%1), %%mm0		\n\t"
+	"movq (%1, %3), %%mm1		\n\t"
+	"movq 8(%1), %%mm2		\n\t"
+	"movq 8(%1, %3), %%mm3		\n\t"
+	PAVGB" 1(%1), %%mm0		\n\t"
+	PAVGB" 1(%1, %3), %%mm1		\n\t"
+	PAVGB" 9(%1), %%mm2		\n\t"
+	PAVGB" 9(%1, %3), %%mm3		\n\t"
+	"movq %%mm0, (%2)		\n\t"
+	"movq %%mm1, (%2, %3)		\n\t"
+	"movq %%mm2, 8(%2)		\n\t"
+	"movq %%mm3, 8(%2, %3)		\n\t"
+	"addl %%eax, %1			\n\t"
+	"addl %%eax, %2			\n\t"
+	"movq (%1), %%mm0		\n\t"
+	"movq (%1, %3), %%mm1		\n\t"
+	"movq 8(%1), %%mm2		\n\t"
+	"movq 8(%1, %3), %%mm3		\n\t"
+	PAVGB" 1(%1), %%mm0		\n\t"
+	PAVGB" 1(%1, %3), %%mm1		\n\t"
+	PAVGB" 9(%1), %%mm2		\n\t"
+	PAVGB" 9(%1, %3), %%mm3		\n\t"
+	"addl %%eax, %1			\n\t"
+	"movq %%mm0, (%2)		\n\t"
+	"movq %%mm1, (%2, %3)		\n\t"
+	"movq %%mm2, 8(%2)		\n\t"
+	"movq %%mm3, 8(%2, %3)		\n\t"
+	"addl %%eax, %2			\n\t"
+	"subl $4, %0			\n\t"
+	"jnz 1b				\n\t"
+	:"+g"(h), "+S"(pixels), "+D"(block)
+	:"r" (line_size)
+	:"%eax", "memory");
+}
  
 /* GL: this function does incorrect rounding if overflow */
-static void DEF(put_no_rnd_pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(put_no_rnd_pixels8_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     MOVQ_BONE(mm6);
     __asm __volatile(
@@ -91,7 +131,7 @@ static void DEF(put_no_rnd_pixels_x2)(UINT8 *block, const UINT8 *pixels, int lin
 	:"%eax", "memory");
 }
 
-static void DEF(put_pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(put_pixels8_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     __asm __volatile(
 	"lea (%3, %3), %%eax		\n\t"
@@ -122,7 +162,7 @@ static void DEF(put_pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size,
 }
 
 /* GL: this function does incorrect rounding if overflow */
-static void DEF(put_no_rnd_pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(put_no_rnd_pixels8_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     MOVQ_BONE(mm6);
     __asm __volatile(
@@ -155,7 +195,7 @@ static void DEF(put_no_rnd_pixels_y2)(UINT8 *block, const UINT8 *pixels, int lin
 	:"%eax", "memory");
 }
 
-static void DEF(avg_pixels)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(avg_pixels8)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     __asm __volatile(
 	"lea (%3, %3), %%eax		\n\t"
@@ -183,7 +223,7 @@ static void DEF(avg_pixels)(UINT8 *block, const UINT8 *pixels, int line_size, in
 	:"%eax", "memory");
 }
 
-static void DEF(avg_pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(avg_pixels8_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     __asm __volatile(
 	"lea (%3, %3), %%eax		\n\t"
@@ -215,7 +255,7 @@ static void DEF(avg_pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size,
 	:"%eax", "memory");
 }
 
-static void DEF(avg_pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(avg_pixels8_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     __asm __volatile(
 	"lea (%3, %3), %%eax		\n\t"
@@ -254,7 +294,7 @@ static void DEF(avg_pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size,
 }
 
 // Note this is not correctly rounded, but this function is only used for b frames so it doesnt matter 
-static void DEF(avg_pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(avg_pixels8_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     MOVQ_BONE(mm6);
     __asm __volatile(
@@ -294,3 +334,34 @@ static void DEF(avg_pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_size
 	:"r" (line_size)
 	:"%eax",  "memory");
 }
+
+//FIXME the following could be optimized too ...
+static void DEF(put_no_rnd_pixels16_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(put_no_rnd_pixels8_x2)(block  , pixels  , line_size, h);
+    DEF(put_no_rnd_pixels8_x2)(block+8, pixels+8, line_size, h);
+}
+static void DEF(put_pixels16_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(put_pixels8_y2)(block  , pixels  , line_size, h);
+    DEF(put_pixels8_y2)(block+8, pixels+8, line_size, h);
+}
+static void DEF(put_no_rnd_pixels16_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(put_no_rnd_pixels8_y2)(block  , pixels  , line_size, h);
+    DEF(put_no_rnd_pixels8_y2)(block+8, pixels+8, line_size, h);
+}
+static void DEF(avg_pixels16)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(avg_pixels8)(block  , pixels  , line_size, h);
+    DEF(avg_pixels8)(block+8, pixels+8, line_size, h);
+}
+static void DEF(avg_pixels16_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(avg_pixels8_x2)(block  , pixels  , line_size, h);
+    DEF(avg_pixels8_x2)(block+8, pixels+8, line_size, h);
+}
+static void DEF(avg_pixels16_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(avg_pixels8_y2)(block  , pixels  , line_size, h);
+    DEF(avg_pixels8_y2)(block+8, pixels+8, line_size, h);
+}
+static void DEF(avg_pixels16_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(avg_pixels8_xy2)(block  , pixels  , line_size, h);
+    DEF(avg_pixels8_xy2)(block+8, pixels+8, line_size, h);
+}
+
diff --git a/libavcodec/i386/dsputil_mmx_rnd.h b/libavcodec/i386/dsputil_mmx_rnd.h
index 873f4b3e1e..3605e03f9c 100644
--- a/libavcodec/i386/dsputil_mmx_rnd.h
+++ b/libavcodec/i386/dsputil_mmx_rnd.h
@@ -22,7 +22,7 @@
  */
 
 // put_pixels
-static void DEF(put, pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(put, pixels8_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     MOVQ_BFE(mm6);
     __asm __volatile(
@@ -54,7 +54,53 @@ static void DEF(put, pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size
 	:"eax", "memory");
 }
 
-static void DEF(put, pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(put, pixels16_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+{
+    MOVQ_BFE(mm6);
+    __asm __volatile(
+	"lea	(%3, %3), %%eax		\n\t"
+	".balign 8			\n\t"
+	"1:				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	1(%1), %%mm1		\n\t"
+	"movq	(%1, %3), %%mm2		\n\t"
+	"movq	1(%1, %3), %%mm3	\n\t"
+	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
+	"movq	%%mm4, (%2)		\n\t"
+	"movq	%%mm5, (%2, %3)		\n\t"
+	"movq	8(%1), %%mm0		\n\t"
+	"movq	9(%1), %%mm1		\n\t"
+	"movq	8(%1, %3), %%mm2	\n\t"
+	"movq	9(%1, %3), %%mm3	\n\t"
+	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
+	"movq	%%mm4, 8(%2)		\n\t"
+	"movq	%%mm5, 8(%2, %3)	\n\t"
+	"addl	%%eax, %1		\n\t"
+	"addl	%%eax, %2		\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	1(%1), %%mm1		\n\t"
+	"movq	(%1, %3), %%mm2		\n\t"
+	"movq	1(%1, %3), %%mm3	\n\t"
+	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
+	"movq	%%mm4, (%2)		\n\t"
+	"movq	%%mm5, (%2, %3)		\n\t"
+	"movq	8(%1), %%mm0		\n\t"
+	"movq	9(%1), %%mm1		\n\t"
+	"movq	8(%1, %3), %%mm2	\n\t"
+	"movq	9(%1, %3), %%mm3	\n\t"
+	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
+	"movq	%%mm4, 8(%2)		\n\t"
+	"movq	%%mm5, 8(%2, %3)		\n\t"
+	"addl	%%eax, %1		\n\t"
+	"addl	%%eax, %2		\n\t"
+	"subl	$4, %0			\n\t"
+	"jnz	1b			\n\t"
+	:"+g"(h), "+S"(pixels), "+D"(block)
+	:"r"(line_size)
+	:"eax", "memory");
+}
+
+static void DEF(put, pixels8_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     MOVQ_BFE(mm6);
     __asm __volatile(
@@ -83,7 +129,7 @@ static void DEF(put, pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size
 	:"eax", "memory");
 }
 
-static void DEF(put, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(put, pixels8_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     MOVQ_ZERO(mm7);
     SET_RND(mm6); // =2 for rnd  and  =1 for no_rnd version
@@ -151,7 +197,7 @@ static void DEF(put, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_siz
 
 // avg_pixels
 // in case more speed is needed - unroling would certainly help
-static void DEF(avg, pixels)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(avg, pixels8)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     MOVQ_BFE(mm6);
     JUMPALIGN();
@@ -170,7 +216,30 @@ static void DEF(avg, pixels)(UINT8 *block, const UINT8 *pixels, int line_size, i
     while (--h);
 }
 
-static void DEF(avg, pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(avg, pixels16)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+{
+    MOVQ_BFE(mm6);
+    JUMPALIGN();
+    do {
+	__asm __volatile(
+	     "movq  %0, %%mm0		\n\t"
+	     "movq  %1, %%mm1		\n\t"
+	     PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
+	     "movq  %%mm2, %0		\n\t"
+	     "movq  8%0, %%mm0		\n\t"
+	     "movq  8%1, %%mm1		\n\t"
+	     PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
+	     "movq  %%mm2, 8%0		\n\t"
+	     :"+m"(*block)
+	     :"m"(*pixels)
+	     :"memory");
+	pixels += line_size;
+	block += line_size;
+    }
+    while (--h);
+}
+
+static void DEF(avg, pixels8_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     MOVQ_BFE(mm6);
     JUMPALIGN();
@@ -190,7 +259,33 @@ static void DEF(avg, pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size
     } while (--h);
 }
 
-static void DEF(avg, pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(avg, pixels16_x2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+{
+    MOVQ_BFE(mm6);
+    JUMPALIGN();
+    do {
+	__asm __volatile(
+	    "movq  %1, %%mm0		\n\t"
+	    "movq  1%1, %%mm1		\n\t"
+	    "movq  %0, %%mm3		\n\t"
+	    PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
+	    PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
+	    "movq  %%mm0, %0		\n\t"
+	    "movq  8%1, %%mm0		\n\t"
+	    "movq  9%1, %%mm1		\n\t"
+	    "movq  8%0, %%mm3		\n\t"
+	    PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
+	    PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
+	    "movq  %%mm0, 8%0		\n\t"
+	    :"+m"(*block)
+	    :"m"(*pixels)
+	    :"memory");
+	pixels += line_size;
+	block += line_size;
+    } while (--h);
+}
+
+static void DEF(avg, pixels8_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     MOVQ_BFE(mm6);
     __asm __volatile(
@@ -230,7 +325,7 @@ static void DEF(avg, pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size
 }
 
 // this routine is 'slightly' suboptimal but mostly unused
-static void DEF(avg, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(avg, pixels8_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
 {
     MOVQ_ZERO(mm7);
     SET_RND(mm6); // =2 for rnd  and  =1 for no_rnd version
@@ -303,3 +398,26 @@ static void DEF(avg, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_siz
 	:"D"(block), "r"(line_size)
 	:"eax", "memory");
 }
+
+//FIXME optimize
+static void DEF(put, pixels16_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(put, pixels8_y2)(block  , pixels  , line_size, h);
+    DEF(put, pixels8_y2)(block+8, pixels+8, line_size, h);
+}
+
+static void DEF(put, pixels16_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(put, pixels8_xy2)(block  , pixels  , line_size, h);
+    DEF(put, pixels8_xy2)(block+8, pixels+8, line_size, h);
+}
+
+static void DEF(avg, pixels16_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(avg, pixels8_y2)(block  , pixels  , line_size, h);
+    DEF(avg, pixels8_y2)(block+8, pixels+8, line_size, h);
+}
+
+static void DEF(avg, pixels16_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h){
+    DEF(avg, pixels8_xy2)(block  , pixels  , line_size, h);
+    DEF(avg, pixels8_xy2)(block+8, pixels+8, line_size, h);
+}
+
+