X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;ds=sidebyside;f=libavcodec%2Fi386%2Fdsputil_mmx_rnd.h;h=a56374b63ecbbeb35d09a1d7b29b0829af2df03b;hb=79396ac68573628f3b59142601db86a0f4ba7ad5;hp=dc70c9c8ed4f0536ba205acd93e2a53dd4f3bff8;hpb=d3cd07425a3781709f53effae6d21d3b3504cdaa;p=ffmpeg

diff --git a/libavcodec/i386/dsputil_mmx_rnd.h b/libavcodec/i386/dsputil_mmx_rnd.h
index dc70c9c8ed4..a56374b63ec 100644
--- a/libavcodec/i386/dsputil_mmx_rnd.h
+++ b/libavcodec/i386/dsputil_mmx_rnd.h
@@ -1,6 +1,7 @@
 /*
  * DSP utils mmx functions are compiled twice for rnd/no_rnd
  * Copyright (c) 2000, 2001 Fabrice Bellard.
+ * Copyright (c) 2003-2004 Michael Niedermayer <michaelni@gmx.at>
  *
  * This library is free software; you can redistribute it and/or
  * modify it under the terms of the GNU Lesser General Public
@@ -22,11 +23,11 @@
  */
 
 // 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_t *block, const uint8_t *pixels, int line_size, int h)
 {
     MOVQ_BFE(mm6);
     __asm __volatile(
-	"lea	(%3, %3), %%eax		\n\t"
+	"lea	(%3, %3), %%"REG_a"	\n\t"
 	".balign 8			\n\t"
 	"1:				\n\t"
 	"movq	(%1), %%mm0		\n\t"
@@ -36,8 +37,8 @@ static void DEF(put, pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size
 	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
 	"movq	%%mm4, (%2)		\n\t"
 	"movq	%%mm5, (%2, %3)		\n\t"
-	"addl	%%eax, %1		\n\t"
-	"addl	%%eax, %2		\n\t"
+	"add	%%"REG_a", %1		\n\t"
+	"add	%%"REG_a", %2		\n\t"
 	"movq	(%1), %%mm0		\n\t"
 	"movq	1(%1), %%mm1		\n\t"
 	"movq	(%1, %3), %%mm2		\n\t"
@@ -45,45 +46,190 @@ static void DEF(put, pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size
 	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
 	"movq	%%mm4, (%2)		\n\t"
 	"movq	%%mm5, (%2, %3)		\n\t"
-	"addl	%%eax, %1		\n\t"
-	"addl	%%eax, %2		\n\t"
+	"add	%%"REG_a", %1		\n\t"
+	"add	%%"REG_a", %2		\n\t"
 	"subl	$4, %0			\n\t"
 	"jnz	1b			\n\t"
 	:"+g"(h), "+S"(pixels), "+D"(block)
-	:"r"(line_size)
-	:"eax", "memory");
+	:"r"((long)line_size)
+	:REG_a, "memory");
 }
 
-static void DEF(put, pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void attribute_unused DEF(put, pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
 {
     MOVQ_BFE(mm6);
     __asm __volatile(
-	"lea (%3, %3), %%eax		\n\t"
+	"testl $1, %0			\n\t"
+        " jz 1f				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	(%2), %%mm1		\n\t"
+	"add	%4, %1			\n\t"
+	"add	$8, %2			\n\t"
+	PAVGB(%%mm0, %%mm1, %%mm4, %%mm6)
+	"movq	%%mm4, (%3)		\n\t"
+	"add	%5, %3			\n\t"
+        "decl	%0			\n\t"
+	".balign 8			\n\t"
+	"1:				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	(%2), %%mm1		\n\t"
+	"add	%4, %1			\n\t"
+	"movq	(%1), %%mm2		\n\t"
+	"movq	8(%2), %%mm3		\n\t"
+	"add	%4, %1			\n\t"
+	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
+	"movq	%%mm4, (%3)		\n\t"
+	"add	%5, %3			\n\t"
+	"movq	%%mm5, (%3)		\n\t"
+	"add	%5, %3			\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	16(%2), %%mm1		\n\t"
+	"add	%4, %1			\n\t"
+	"movq	(%1), %%mm2		\n\t"
+	"movq	24(%2), %%mm3		\n\t"
+	"add	%4, %1			\n\t"
+	"add	$32, %2			\n\t"
+	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
+	"movq	%%mm4, (%3)		\n\t"
+	"add	%5, %3			\n\t"
+	"movq	%%mm5, (%3)		\n\t"
+	"add	%5, %3			\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"((long)src1Stride), "D"((long)dstStride)
+	:"memory");
+}
+
+static void DEF(put, pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    MOVQ_BFE(mm6);
+    __asm __volatile(
+	"lea	(%3, %3), %%"REG_a"	\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"
+	"add	%%"REG_a", %1		\n\t"
+	"add	%%"REG_a", %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"
+	"add	%%"REG_a", %1		\n\t"
+	"add	%%"REG_a", %2		\n\t"
+	"subl	$4, %0			\n\t"
+	"jnz	1b			\n\t"
+	:"+g"(h), "+S"(pixels), "+D"(block)
+	:"r"((long)line_size)
+	:REG_a, "memory");
+}
+
+static void attribute_unused DEF(put, pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
+{
+    MOVQ_BFE(mm6);
+    __asm __volatile(
+	"testl $1, %0			\n\t"
+        " jz 1f				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	(%2), %%mm1		\n\t"
+	"movq	8(%1), %%mm2		\n\t"
+	"movq	8(%2), %%mm3		\n\t"
+	"add	%4, %1			\n\t"
+	"add	$16, %2			\n\t"
+	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
+	"movq	%%mm4, (%3)		\n\t"
+	"movq	%%mm5, 8(%3)		\n\t"
+	"add	%5, %3			\n\t"
+	"decl	%0			\n\t"
+	".balign 8			\n\t"
+	"1:				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	(%2), %%mm1		\n\t"
+	"movq	8(%1), %%mm2		\n\t"
+	"movq	8(%2), %%mm3		\n\t"
+	"add	%4, %1			\n\t"
+	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
+	"movq	%%mm4, (%3)		\n\t"
+	"movq	%%mm5, 8(%3)		\n\t"
+	"add	%5, %3			\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	16(%2), %%mm1		\n\t"
+	"movq	8(%1), %%mm2		\n\t"
+	"movq	24(%2), %%mm3		\n\t"
+	"add	%4, %1			\n\t"
+	PAVGBP(%%mm0, %%mm1, %%mm4,   %%mm2, %%mm3, %%mm5)
+	"movq	%%mm4, (%3)		\n\t"
+	"movq	%%mm5, 8(%3)		\n\t"
+	"add	%5, %3			\n\t"
+	"add	$32, %2			\n\t"
+	"subl	$2, %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"((long)src1Stride), "D"((long)dstStride)
+	:"memory"); 
+}
+
+static void DEF(put, pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    MOVQ_BFE(mm6);
+    __asm __volatile(
+	"lea (%3, %3), %%"REG_a"	\n\t"
 	"movq (%1), %%mm0		\n\t"
 	".balign 8			\n\t"
 	"1:				\n\t"
 	"movq	(%1, %3), %%mm1		\n\t"
-	"movq	(%1, %%eax),%%mm2	\n\t"
+	"movq	(%1, %%"REG_a"),%%mm2	\n\t"
 	PAVGBP(%%mm1, %%mm0, %%mm4,   %%mm2, %%mm1, %%mm5)
 	"movq	%%mm4, (%2)		\n\t"
 	"movq	%%mm5, (%2, %3)		\n\t"
-	"addl	%%eax, %1		\n\t"
-	"addl	%%eax, %2		\n\t"
+	"add	%%"REG_a", %1		\n\t"
+	"add	%%"REG_a", %2		\n\t"
 	"movq	(%1, %3), %%mm1		\n\t"
-	"movq	(%1, %%eax),%%mm0	\n\t"
+	"movq	(%1, %%"REG_a"),%%mm0	\n\t"
 	PAVGBP(%%mm1, %%mm2, %%mm4,   %%mm0, %%mm1, %%mm5)
 	"movq	%%mm4, (%2)		\n\t"
 	"movq	%%mm5, (%2, %3)		\n\t"
-	"addl	%%eax, %1		\n\t"
-	"addl	%%eax, %2		\n\t"
+	"add	%%"REG_a", %1		\n\t"
+	"add	%%"REG_a", %2		\n\t"
 	"subl	$4, %0			\n\t"
 	"jnz	1b			\n\t"
 	:"+g"(h), "+S"(pixels), "+D"(block)
-	:"r"(line_size)
-	:"eax", "memory");
+	:"r"((long)line_size)
+	:REG_a, "memory");
 }
 
-static void DEF(put, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(put, pixels8_xy2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
 {
     MOVQ_ZERO(mm7);
     SET_RND(mm6); // =2 for rnd  and  =1 for no_rnd version
@@ -98,12 +244,12 @@ static void DEF(put, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_siz
 	"punpckhbw %%mm7, %%mm5		\n\t"
 	"paddusw %%mm0, %%mm4		\n\t"
 	"paddusw %%mm1, %%mm5		\n\t"
-	"xorl	%%eax, %%eax		\n\t"
-	"addl	%3, %1			\n\t"
+	"xor	%%"REG_a", %%"REG_a"	\n\t"
+	"add	%3, %1			\n\t"
 	".balign 8      		\n\t"
 	"1:				\n\t"
-	"movq	(%1, %%eax), %%mm0	\n\t"
-	"movq	1(%1, %%eax), %%mm2	\n\t"
+	"movq	(%1, %%"REG_a"), %%mm0	\n\t"
+	"movq	1(%1, %%"REG_a"), %%mm2	\n\t"
 	"movq	%%mm0, %%mm1		\n\t"
 	"movq	%%mm2, %%mm3		\n\t"
 	"punpcklbw %%mm7, %%mm0		\n\t"
@@ -119,11 +265,11 @@ static void DEF(put, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_siz
 	"psrlw	$2, %%mm4		\n\t"
 	"psrlw	$2, %%mm5		\n\t"
 	"packuswb  %%mm5, %%mm4		\n\t"
-	"movq	%%mm4, (%2, %%eax)	\n\t"
-	"addl	%3, %%eax		\n\t"
+	"movq	%%mm4, (%2, %%"REG_a")	\n\t"
+	"add	%3, %%"REG_a"		\n\t"
 
-	"movq	(%1, %%eax), %%mm2	\n\t" // 0 <-> 2   1 <-> 3
-	"movq	1(%1, %%eax), %%mm4	\n\t"
+	"movq	(%1, %%"REG_a"), %%mm2	\n\t" // 0 <-> 2   1 <-> 3
+	"movq	1(%1, %%"REG_a"), %%mm4	\n\t"
 	"movq	%%mm2, %%mm3		\n\t"
 	"movq	%%mm4, %%mm5		\n\t"
 	"punpcklbw %%mm7, %%mm2		\n\t"
@@ -139,19 +285,57 @@ static void DEF(put, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_siz
 	"psrlw	$2, %%mm0		\n\t"
 	"psrlw	$2, %%mm1		\n\t"
 	"packuswb  %%mm1, %%mm0		\n\t"
-	"movq	%%mm0, (%2, %%eax)	\n\t"
-	"addl	%3, %%eax		\n\t"
+	"movq	%%mm0, (%2, %%"REG_a")	\n\t"
+	"add	%3, %%"REG_a"		\n\t"
 
 	"subl	$2, %0			\n\t"
 	"jnz	1b			\n\t"
 	:"+g"(h), "+S"(pixels)
-	:"D"(block), "r"(line_size)
-	:"eax", "memory");
+	:"D"(block), "r"((long)line_size)
+	:REG_a, "memory");
 }
 
 // avg_pixels
+static void attribute_unused 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, pixels)(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+static void DEF(avg, pixels8)(uint8_t *block, const uint8_t *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"
+	     :"+m"(*block)
+	     :"m"(*pixels)
+	     :"memory");
+	pixels += line_size;
+	block += line_size;
+    }
+    while (--h);
+}
+
+static void DEF(avg, pixels16)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
 {
     MOVQ_BFE(mm6);
     JUMPALIGN();
@@ -161,6 +345,10 @@ static void DEF(avg, pixels)(UINT8 *block, const UINT8 *pixels, int line_size, i
 	     "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");
@@ -170,7 +358,48 @@ 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, pixels8_x2)(uint8_t *block, const uint8_t *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"
+	    :"+m"(*block)
+	    :"m"(*pixels)
+	    :"memory");
+	pixels += line_size;
+	block += line_size;
+    } while (--h);
+}
+
+static __attribute__((unused)) void DEF(avg, pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
+{
+    MOVQ_BFE(mm6);
+    JUMPALIGN();
+    do {
+	__asm __volatile(
+	    "movq  %1, %%mm0		\n\t"
+	    "movq  %2, %%mm1		\n\t"
+	    "movq  %0, %%mm3		\n\t"
+	    PAVGB(%%mm0, %%mm1, %%mm2, %%mm6)
+	    PAVGB(%%mm3, %%mm2, %%mm0, %%mm6)
+	    "movq  %%mm0, %0		\n\t"
+	    :"+m"(*dst)
+	    :"m"(*src1), "m"(*src2)
+	    :"memory");
+	dst += dstStride;
+        src1 += src1Stride;
+        src2 += 8;
+    } while (--h);
+}
+
+static void DEF(avg, pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
 {
     MOVQ_BFE(mm6);
     JUMPALIGN();
@@ -182,6 +411,12 @@ static void DEF(avg, pixels_x2)(UINT8 *block, const UINT8 *pixels, int line_size
 	    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");
@@ -190,16 +425,43 @@ 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 __attribute__((unused)) void DEF(avg, pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
+{
+    MOVQ_BFE(mm6);
+    JUMPALIGN();
+    do {
+	__asm __volatile(
+	    "movq  %1, %%mm0		\n\t"
+	    "movq  %2, %%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  8%2, %%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"(*dst)
+	    :"m"(*src1), "m"(*src2)
+	    :"memory");
+	dst += dstStride;
+        src1 += src1Stride;
+        src2 += 16;
+    } while (--h);
+}
+
+static void DEF(avg, pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
 {
     MOVQ_BFE(mm6);
     __asm __volatile(
-	"lea	(%3, %3), %%eax		\n\t"
+	"lea	(%3, %3), %%"REG_a"	\n\t"
 	"movq	(%1), %%mm0		\n\t"
 	".balign 8			\n\t"
 	"1:				\n\t"
 	"movq	(%1, %3), %%mm1		\n\t"
-	"movq	(%1, %%eax), %%mm2	\n\t"
+	"movq	(%1, %%"REG_a"), %%mm2	\n\t"
 	PAVGBP(%%mm1, %%mm0, %%mm4,   %%mm2, %%mm1, %%mm5)
 	"movq	(%2), %%mm3		\n\t"
 	PAVGB(%%mm3, %%mm4, %%mm0, %%mm6)
@@ -207,30 +469,30 @@ static void DEF(avg, pixels_y2)(UINT8 *block, const UINT8 *pixels, int line_size
 	PAVGB(%%mm3, %%mm5, %%mm1, %%mm6)
 	"movq	%%mm0, (%2)		\n\t"
 	"movq	%%mm1, (%2, %3)		\n\t"
-	"addl	%%eax, %1		\n\t"
-	"addl	%%eax, %2		\n\t"
+	"add	%%"REG_a", %1		\n\t"
+	"add	%%"REG_a", %2		\n\t"
 
 	"movq	(%1, %3), %%mm1		\n\t"
-	"movq	(%1, %%eax), %%mm0	\n\t"
+	"movq	(%1, %%"REG_a"), %%mm0	\n\t"
 	PAVGBP(%%mm1, %%mm2, %%mm4,   %%mm0, %%mm1, %%mm5)
 	"movq	(%2), %%mm3		\n\t"
-	PAVGB(%%mm3, %%mm4, %%mm0, %%mm6)
+	PAVGB(%%mm3, %%mm4, %%mm2, %%mm6)
 	"movq	(%2, %3), %%mm3		\n\t"
 	PAVGB(%%mm3, %%mm5, %%mm1, %%mm6)
-	"movq	%%mm0, (%2)		\n\t"
+	"movq	%%mm2, (%2)		\n\t"
 	"movq	%%mm1, (%2, %3)		\n\t"
-	"addl	%%eax, %1		\n\t"
-	"addl	%%eax, %2		\n\t"
+	"add	%%"REG_a", %1		\n\t"
+	"add	%%"REG_a", %2		\n\t"
 
 	"subl	$4, %0			\n\t"
 	"jnz	1b			\n\t"
 	:"+g"(h), "+S"(pixels), "+D"(block)
-	:"r"(line_size)
-	:"eax", "memory");
+	:"r"((long)line_size)
+	:REG_a, "memory");
 }
 
 // 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_t *block, const uint8_t *pixels, int line_size, int h)
 {
     MOVQ_ZERO(mm7);
     SET_RND(mm6); // =2 for rnd  and  =1 for no_rnd version
@@ -245,12 +507,12 @@ static void DEF(avg, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_siz
 	"punpckhbw %%mm7, %%mm5		\n\t"
 	"paddusw %%mm0, %%mm4		\n\t"
 	"paddusw %%mm1, %%mm5		\n\t"
-	"xorl	%%eax, %%eax		\n\t"
-	"addl	%3, %1			\n\t"
+	"xor	%%"REG_a", %%"REG_a"	\n\t"
+	"add	%3, %1			\n\t"
 	".balign 8			\n\t"
 	"1:				\n\t"
-	"movq	(%1, %%eax), %%mm0	\n\t"
-	"movq	1(%1, %%eax), %%mm2	\n\t"
+	"movq	(%1, %%"REG_a"), %%mm0	\n\t"
+	"movq	1(%1, %%"REG_a"), %%mm2	\n\t"
 	"movq	%%mm0, %%mm1		\n\t"
 	"movq	%%mm2, %%mm3		\n\t"
 	"punpcklbw %%mm7, %%mm0		\n\t"
@@ -265,16 +527,16 @@ static void DEF(avg, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_siz
 	"paddusw %%mm1, %%mm5		\n\t"
 	"psrlw	$2, %%mm4		\n\t"
 	"psrlw	$2, %%mm5		\n\t"
-		"movq	(%2, %%eax), %%mm3	\n\t"
+		"movq	(%2, %%"REG_a"), %%mm3	\n\t"
 	"packuswb  %%mm5, %%mm4		\n\t"
 		"pcmpeqd %%mm2, %%mm2	\n\t"
 		"paddb %%mm2, %%mm2	\n\t"
 		PAVGB(%%mm3, %%mm4, %%mm5, %%mm2)
-		"movq	%%mm5, (%2, %%eax)	\n\t"
-	"addl	%3, %%eax		\n\t"
+		"movq	%%mm5, (%2, %%"REG_a")	\n\t"
+	"add	%3, %%"REG_a"		\n\t"
 
-	"movq	(%1, %%eax), %%mm2	\n\t" // 0 <-> 2   1 <-> 3
-	"movq	1(%1, %%eax), %%mm4	\n\t"
+	"movq	(%1, %%"REG_a"), %%mm2	\n\t" // 0 <-> 2   1 <-> 3
+	"movq	1(%1, %%"REG_a"), %%mm4	\n\t"
 	"movq	%%mm2, %%mm3		\n\t"
 	"movq	%%mm4, %%mm5		\n\t"
 	"punpcklbw %%mm7, %%mm2		\n\t"
@@ -289,17 +551,40 @@ static void DEF(avg, pixels_xy2)(UINT8 *block, const UINT8 *pixels, int line_siz
 	"paddusw %%mm5, %%mm1		\n\t"
 	"psrlw	$2, %%mm0		\n\t"
 	"psrlw	$2, %%mm1		\n\t"
-		"movq	(%2, %%eax), %%mm3	\n\t"
+		"movq	(%2, %%"REG_a"), %%mm3	\n\t"
 	"packuswb  %%mm1, %%mm0		\n\t"
 		"pcmpeqd %%mm2, %%mm2	\n\t"
 		"paddb %%mm2, %%mm2	\n\t"
 		PAVGB(%%mm3, %%mm0, %%mm1, %%mm2)
-		"movq	%%mm1, (%2, %%eax)	\n\t"
-	"addl	%3, %%eax		\n\t"
+		"movq	%%mm1, (%2, %%"REG_a")	\n\t"
+	"add	%3, %%"REG_a"		\n\t"
 
 	"subl	$2, %0			\n\t"
 	"jnz	1b			\n\t"
 	:"+g"(h), "+S"(pixels)
-	:"D"(block), "r"(line_size)
-	:"eax", "memory");
+	:"D"(block), "r"((long)line_size)
+	:REG_a, "memory");
 }
+
+//FIXME optimize
+static void DEF(put, pixels16_y2)(uint8_t *block, const uint8_t *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_t *block, const uint8_t *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_t *block, const uint8_t *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_t *block, const uint8_t *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);
+}
+
+