h264 luma motion compensation in mmx2/3dnow

Originally committed as revision 3437 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/i386/dsputil_mmx_avg.h

@@ -53,6 +53,53 @@ static void DEF(put_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_
 	:"%eax", "memory");
 }
 
+static void DEF(put_pixels4_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
+{
+    __asm __volatile(
+	"testl $1, %0			\n\t"
+	    " jz 1f				\n\t"
+	"movd	(%1), %%mm0		\n\t"
+	"movd	(%2), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"addl	$4, %2			\n\t"
+	PAVGB" %%mm1, %%mm0		\n\t"
+	"movd	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"decl	%0			\n\t"
+	"1:				\n\t"
+	"movd	(%1), %%mm0		\n\t"
+	"addl	%4, %1			\n\t"
+	"movd	(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	PAVGB" (%2), %%mm0		\n\t"
+	PAVGB" 4(%2), %%mm1		\n\t"
+	"movd	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movd	%%mm1, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movd	(%1), %%mm0		\n\t"
+	"addl	%4, %1			\n\t"
+	"movd	(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	PAVGB" 8(%2), %%mm0		\n\t"
+	PAVGB" 12(%2), %%mm1		\n\t"
+	"movd	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movd	%%mm1, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+        "addl	$16, %2			\n\t"
+	"subl	$4, %0			\n\t"
+	"jnz	1b			\n\t"
+#ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used
+	:"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)
+#else
+	:"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)
+#endif
+	:"S"(src1Stride), "D"(dstStride)
+	:"memory"); 
+}
+
+
 static void DEF(put_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
 {
     __asm __volatile(
@@ -173,6 +220,58 @@ static void DEF(put_no_rnd_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src
 	:"memory");*/
 }
 
+static void DEF(avg_pixels4_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
+{
+    __asm __volatile(
+	"testl $1, %0			\n\t"
+	    " jz 1f				\n\t"
+	"movd	(%1), %%mm0		\n\t"
+	"movd	(%2), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"addl	$4, %2			\n\t"
+	PAVGB" %%mm1, %%mm0		\n\t"
+	PAVGB" (%3), %%mm0		\n\t"
+	"movd	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"decl	%0			\n\t"
+	"1:				\n\t"
+	"movd	(%1), %%mm0		\n\t"
+	"addl	%4, %1			\n\t"
+	"movd	(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	PAVGB" (%2), %%mm0		\n\t"
+	PAVGB" 4(%2), %%mm1		\n\t"
+	PAVGB" (%3), %%mm0	 	\n\t"
+	"movd	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	PAVGB" (%3), %%mm1	 	\n\t"
+	"movd	%%mm1, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movd	(%1), %%mm0		\n\t"
+	"addl	%4, %1			\n\t"
+	"movd	(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	PAVGB" 8(%2), %%mm0		\n\t"
+	PAVGB" 12(%2), %%mm1		\n\t"
+	PAVGB" (%3), %%mm0	 	\n\t"
+	"movd	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	PAVGB" (%3), %%mm1	 	\n\t"
+	"movd	%%mm1, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+        "addl	$16, %2			\n\t"
+	"subl	$4, %0			\n\t"
+	"jnz	1b			\n\t"
+#ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used
+	:"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)
+#else
+	:"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)
+#endif
+	:"S"(src1Stride), "D"(dstStride)
+	:"memory"); 
+}
+
+
 static void DEF(avg_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
 {
     __asm __volatile(

libavcodec/i386/dsputil_mmx_rnd.h

@@ -296,6 +296,25 @@ static void DEF(put, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, int lin
 }
 
 // avg_pixels
+static void DEF(avg, pixels4)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    MOVQ_BFE(mm6);
+    JUMPALIGN();
+    do {
+	__asm __volatile(
+	     "movd  %0, %%mm0		\n\t"
+	     "movd  %1, %%mm1		\n\t"
+	     PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
+	     "movd  %%mm2, %0		\n\t"
+	     :"+m"(*block)
+	     :"m"(*pixels)
+	     :"memory");
+	pixels += line_size;
+	block += line_size;
+    }
+    while (--h);
+}
+
 // in case more speed is needed - unroling would certainly help
 static void DEF(avg, pixels8)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
 {