10000l fix and use more mmx2/3dnow code for mpeg4 qpel which has been written and commited long time ago but appearently never used, qpel motion compensation is 5% faster

now

Originally committed as revision 3435 to svn://svn.ffmpeg.org/ffmpeg/trunk

2 files changed, 329 insertions, 39 deletions

diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index 51280a256c..3053a4e9d3 100644
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -1334,7 +1334,7 @@ static void diff_bytes_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
 static void sub_hfyu_median_prediction_mmx2(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top){
     int i=0;
     uint8_t l, lt;
-
+    
     asm volatile(
         "1:				\n\t"
         "movq  -1(%1, %0), %%mm0	\n\t" // LT
@@ -2046,7 +2046,7 @@ static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride)
     uint64_t temp[8];\
     uint8_t * const half= (uint8_t*)temp;\
     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\
-    OPNAME ## pixels8_l2_mmx(dst, src, half, stride, stride, 8);\
+    OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\
 }\
 \
 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
@@ -2057,14 +2057,14 @@ static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride)
     uint64_t temp[8];\
     uint8_t * const half= (uint8_t*)temp;\
     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, stride, 8);\
-    OPNAME ## pixels8_l2_mmx(dst, src+1, half, stride, stride, 8);\
+    OPNAME ## pixels8_l2_ ## MMX(dst, src+1, half, stride, stride, 8);\
 }\
 \
 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t temp[8];\
     uint8_t * const half= (uint8_t*)temp;\
     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\
-    OPNAME ## pixels8_l2_mmx(dst, src, half, stride, stride, 8);\
+    OPNAME ## pixels8_l2_ ## MMX(dst, src, half, stride, stride, 8);\
 }\
 \
 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
@@ -2075,43 +2075,43 @@ static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride)
     uint64_t temp[8];\
     uint8_t * const half= (uint8_t*)temp;\
     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, 8, stride);\
-    OPNAME ## pixels8_l2_mmx(dst, src+stride, half, stride, stride, 8);\
+    OPNAME ## pixels8_l2_ ## MMX(dst, src+stride, half, stride, stride, 8);\
 }\
 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[8 + 9];\
     uint8_t * const halfH= ((uint8_t*)half) + 64;\
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
-    put ## RND ## pixels8_l2_mmx(halfH, src, halfH, 8, stride, 9);\
+    put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
-    OPNAME ## pixels8_l2_mmx(dst, halfH, halfHV, stride, 8, 8);\
+    OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
 }\
 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[8 + 9];\
     uint8_t * const halfH= ((uint8_t*)half) + 64;\
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
-    put ## RND ## pixels8_l2_mmx(halfH, src+1, halfH, 8, stride, 9);\
+    put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
-    OPNAME ## pixels8_l2_mmx(dst, halfH, halfHV, stride, 8, 8);\
+    OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
 }\
 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[8 + 9];\
     uint8_t * const halfH= ((uint8_t*)half) + 64;\
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
-    put ## RND ## pixels8_l2_mmx(halfH, src, halfH, 8, stride, 9);\
+    put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
-    OPNAME ## pixels8_l2_mmx(dst, halfH+8, halfHV, stride, 8, 8);\
+    OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
 }\
 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[8 + 9];\
     uint8_t * const halfH= ((uint8_t*)half) + 64;\
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
-    put ## RND ## pixels8_l2_mmx(halfH, src+1, halfH, 8, stride, 9);\
+    put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
-    OPNAME ## pixels8_l2_mmx(dst, halfH+8, halfHV, stride, 8, 8);\
+    OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
 }\
 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[8 + 9];\
@@ -2119,7 +2119,7 @@ static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride)
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
-    OPNAME ## pixels8_l2_mmx(dst, halfH, halfHV, stride, 8, 8);\
+    OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, stride, 8, 8);\
 }\
 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[8 + 9];\
@@ -2127,20 +2127,20 @@ static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride)
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
     put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
-    OPNAME ## pixels8_l2_mmx(dst, halfH+8, halfHV, stride, 8, 8);\
+    OPNAME ## pixels8_l2_ ## MMX(dst, halfH+8, halfHV, stride, 8, 8);\
 }\
 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[8 + 9];\
     uint8_t * const halfH= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
-    put ## RND ## pixels8_l2_mmx(halfH, src, halfH, 8, stride, 9);\
+    put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, stride, 9);\
     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
 }\
 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[8 + 9];\
     uint8_t * const halfH= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, stride, 9);\
-    put ## RND ## pixels8_l2_mmx(halfH, src+1, halfH, 8, stride, 9);\
+    put ## RND ## pixels8_l2_ ## MMX(halfH, src+1, halfH, 8, stride, 9);\
     OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, stride, 8);\
 }\
 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
@@ -2157,7 +2157,7 @@ static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride
     uint64_t temp[32];\
     uint8_t * const half= (uint8_t*)temp;\
     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\
-    OPNAME ## pixels16_l2_mmx(dst, src, half, stride, stride, 16);\
+    OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\
 }\
 \
 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
@@ -2168,14 +2168,14 @@ static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride
     uint64_t temp[32];\
     uint8_t * const half= (uint8_t*)temp;\
     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, stride, 16);\
-    OPNAME ## pixels16_l2_mmx(dst, src+1, half, stride, stride, 16);\
+    OPNAME ## pixels16_l2_ ## MMX(dst, src+1, half, stride, stride, 16);\
 }\
 \
 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t temp[32];\
     uint8_t * const half= (uint8_t*)temp;\
     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\
-    OPNAME ## pixels16_l2_mmx(dst, src, half, stride, stride, 16);\
+    OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, stride, 16);\
 }\
 \
 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
@@ -2186,43 +2186,43 @@ static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride
     uint64_t temp[32];\
     uint8_t * const half= (uint8_t*)temp;\
     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, stride);\
-    OPNAME ## pixels16_l2_mmx(dst, src+stride, half, stride, stride, 16);\
+    OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, stride, stride, 16);\
 }\
 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[16*2 + 17*2];\
     uint8_t * const halfH= ((uint8_t*)half) + 256;\
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
-    put ## RND ## pixels16_l2_mmx(halfH, src, halfH, 16, stride, 17);\
+    put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
-    OPNAME ## pixels16_l2_mmx(dst, halfH, halfHV, stride, 16, 16);\
+    OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
 }\
 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[16*2 + 17*2];\
     uint8_t * const halfH= ((uint8_t*)half) + 256;\
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
-    put ## RND ## pixels16_l2_mmx(halfH, src+1, halfH, 16, stride, 17);\
+    put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
-    OPNAME ## pixels16_l2_mmx(dst, halfH, halfHV, stride, 16, 16);\
+    OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
 }\
 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[16*2 + 17*2];\
     uint8_t * const halfH= ((uint8_t*)half) + 256;\
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
-    put ## RND ## pixels16_l2_mmx(halfH, src, halfH, 16, stride, 17);\
+    put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
-    OPNAME ## pixels16_l2_mmx(dst, halfH+16, halfHV, stride, 16, 16);\
+    OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
 }\
 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[16*2 + 17*2];\
     uint8_t * const halfH= ((uint8_t*)half) + 256;\
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
-    put ## RND ## pixels16_l2_mmx(halfH, src+1, halfH, 16, stride, 17);\
+    put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
-    OPNAME ## pixels16_l2_mmx(dst, halfH+16, halfHV, stride, 16, 16);\
+    OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
 }\
 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[16*2 + 17*2];\
@@ -2230,7 +2230,7 @@ static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
-    OPNAME ## pixels16_l2_mmx(dst, halfH, halfHV, stride, 16, 16);\
+    OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, stride, 16, 16);\
 }\
 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[16*2 + 17*2];\
@@ -2238,20 +2238,20 @@ static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride
     uint8_t * const halfHV= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
     put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, 16, 16);\
-    OPNAME ## pixels16_l2_mmx(dst, halfH+16, halfHV, stride, 16, 16);\
+    OPNAME ## pixels16_l2_ ## MMX(dst, halfH+16, halfHV, stride, 16, 16);\
 }\
 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[17*2];\
     uint8_t * const halfH= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
-    put ## RND ## pixels16_l2_mmx(halfH, src, halfH, 16, stride, 17);\
+    put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, stride, 17);\
     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
 }\
 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
     uint64_t half[17*2];\
     uint8_t * const halfH= ((uint8_t*)half);\
     put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, stride, 17);\
-    put ## RND ## pixels16_l2_mmx(halfH, src+1, halfH, 16, stride, 17);\
+    put ## RND ## pixels16_l2_ ## MMX(halfH, src+1, halfH, 16, stride, 17);\
     OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, stride, 16);\
 }\
 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
diff --git a/libavcodec/i386/dsputil_mmx_avg.h b/libavcodec/i386/dsputil_mmx_avg.h
index 052aad75c0..030bd37d02 100644
--- a/libavcodec/i386/dsputil_mmx_avg.h
+++ b/libavcodec/i386/dsputil_mmx_avg.h
@@ -53,9 +53,140 @@ static void DEF(put_pixels8_x2)(uint8_t *block, const uint8_t *pixels, int line_
 	:"%eax", "memory");
 }
 
-static __attribute__((unused)) void DEF(put_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
+static void DEF(put_pixels8_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"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	(%2), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"addl	$8, %2			\n\t"
+	PAVGB" %%mm1, %%mm0		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"decl	%0			\n\t"
+	"1:				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"addl	%4, %1			\n\t"
+	"movq	(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	PAVGB" (%2), %%mm0		\n\t"
+	PAVGB" 8(%2), %%mm1		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movq	%%mm1, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"addl	%4, %1			\n\t"
+	"movq	(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	PAVGB" 16(%2), %%mm0		\n\t"
+	PAVGB" 24(%2), %%mm1		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movq	%%mm1, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+        "addl	$32, %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"); 
+//the following should be used, though better not with gcc ...
+/*	:"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst)
+	:"r"(src1Stride), "r"(dstStride)
+	:"memory");*/
+}
+
+static void DEF(put_no_rnd_pixels8_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
+{
+    __asm __volatile(
+	"pcmpeqb %%mm6, %%mm6	\n\t"
+	"testl $1, %0			\n\t"
+	    " jz 1f				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	(%2), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"addl	$8, %2			\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
+	PAVGB" %%mm1, %%mm0		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"decl	%0			\n\t"
+	"1:				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"addl	%4, %1			\n\t"
+	"movq	(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"movq	(%2), %%mm2		\n\t"
+	"movq	8(%2), %%mm3		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
+	"pxor %%mm6, %%mm2		\n\t"
+	"pxor %%mm6, %%mm3		\n\t"
+	PAVGB" %%mm2, %%mm0		\n\t"
+	PAVGB" %%mm3, %%mm1		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movq	%%mm1, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"addl	%4, %1			\n\t"
+	"movq	(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"movq	16(%2), %%mm2		\n\t"
+	"movq	24(%2), %%mm3		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
+	"pxor %%mm6, %%mm2		\n\t"
+	"pxor %%mm6, %%mm3		\n\t"
+	PAVGB" %%mm2, %%mm0		\n\t"
+	PAVGB" %%mm3, %%mm1		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movq	%%mm1, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+        "addl	$32, %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"); 
+//the following should be used, though better not with gcc ...
+/*	:"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst)
+	:"r"(src1Stride), "r"(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(
+	"testl $1, %0			\n\t"
+	    " jz 1f				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	(%2), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"addl	$8, %2			\n\t"
+	PAVGB" %%mm1, %%mm0		\n\t"
+	PAVGB" (%3), %%mm0		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"decl	%0			\n\t"
 	"1:				\n\t"
 	"movq	(%1), %%mm0		\n\t"
 	"addl	%4, %1			\n\t"
@@ -63,8 +194,10 @@ static __attribute__((unused)) void DEF(put_pixels8_l2)(uint8_t *dst, uint8_t *s
 	"addl	%4, %1			\n\t"
 	PAVGB" (%2), %%mm0		\n\t"
 	PAVGB" 8(%2), %%mm1		\n\t"
+	PAVGB" (%3), %%mm0	 	\n\t"
 	"movq	%%mm0, (%3)		\n\t"
 	"addl	%5, %3			\n\t"
+	PAVGB" (%3), %%mm1	 	\n\t"
 	"movq	%%mm1, (%3)		\n\t"
 	"addl	%5, %3			\n\t"
 	"movq	(%1), %%mm0		\n\t"
@@ -73,16 +206,26 @@ static __attribute__((unused)) void DEF(put_pixels8_l2)(uint8_t *dst, uint8_t *s
 	"addl	%4, %1			\n\t"
 	PAVGB" 16(%2), %%mm0		\n\t"
 	PAVGB" 24(%2), %%mm1		\n\t"
+	PAVGB" (%3), %%mm0	 	\n\t"
 	"movq	%%mm0, (%3)		\n\t"
 	"addl	%5, %3			\n\t"
+	PAVGB" (%3), %%mm1	 	\n\t"
 	"movq	%%mm1, (%3)		\n\t"
 	"addl	%5, %3			\n\t"
         "addl	$32, %2			\n\t"
 	"subl	$4, %0			\n\t"
 	"jnz	1b			\n\t"
-	:"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst)
+#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"); 
+//the following should be used, though better not with gcc ...
+/*	:"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst)
 	:"r"(src1Stride), "r"(dstStride)
-	:"memory");
+	:"memory");*/
 }
 
 static void DEF(put_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
@@ -125,15 +268,79 @@ static void DEF(put_pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line
 	:"%eax", "memory");
 }
 
-static __attribute__((unused)) void DEF(put_pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
+static void DEF(put_pixels16_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"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	8(%1), %%mm1		\n\t"
+	PAVGB" (%2), %%mm0		\n\t"
+	PAVGB" 8(%2), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"addl	$16, %2			\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"movq	%%mm1, 8(%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"decl	%0			\n\t"
+	"1:				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	8(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	PAVGB" (%2), %%mm0		\n\t"
+	PAVGB" 8(%2), %%mm1		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"movq	%%mm1, 8(%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	8(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	PAVGB" 16(%2), %%mm0		\n\t"
+	PAVGB" 24(%2), %%mm1		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"movq	%%mm1, 8(%3)		\n\t"
+	"addl	%5, %3			\n\t"
+        "addl	$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"(src1Stride), "D"(dstStride)
+	:"memory"); 
+//the following should be used, though better not with gcc ...
+/*	:"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst)
+	:"r"(src1Stride), "r"(dstStride)
+	:"memory");*/
+}
+
+static void DEF(avg_pixels16_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"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	8(%1), %%mm1		\n\t"
+	PAVGB" (%2), %%mm0		\n\t"
+	PAVGB" 8(%2), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"addl	$16, %2			\n\t"
+	PAVGB" (%3), %%mm0		\n\t"
+	PAVGB" 8(%3), %%mm1		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"movq	%%mm1, 8(%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"decl	%0			\n\t"
 	"1:				\n\t"
 	"movq	(%1), %%mm0		\n\t"
 	"movq	8(%1), %%mm1		\n\t"
 	"addl	%4, %1			\n\t"
 	PAVGB" (%2), %%mm0		\n\t"
 	PAVGB" 8(%2), %%mm1		\n\t"
+	PAVGB" (%3), %%mm0		\n\t"
+	PAVGB" 8(%3), %%mm1		\n\t"
 	"movq	%%mm0, (%3)		\n\t"
 	"movq	%%mm1, 8(%3)		\n\t"
 	"addl	%5, %3			\n\t"
@@ -142,15 +349,98 @@ static __attribute__((unused)) void DEF(put_pixels16_l2)(uint8_t *dst, uint8_t *
 	"addl	%4, %1			\n\t"
 	PAVGB" 16(%2), %%mm0		\n\t"
 	PAVGB" 24(%2), %%mm1		\n\t"
+	PAVGB" (%3), %%mm0		\n\t"
+	PAVGB" 8(%3), %%mm1		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"movq	%%mm1, 8(%3)		\n\t"
+	"addl	%5, %3			\n\t"
+        "addl	$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"(src1Stride), "D"(dstStride)
+	:"memory"); 
+//the following should be used, though better not with gcc ...
+/*	:"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst)
+	:"r"(src1Stride), "r"(dstStride)
+	:"memory");*/
+}
+
+static void DEF(put_no_rnd_pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
+{
+    __asm __volatile(
+	"pcmpeqb %%mm6, %%mm6\n\t"
+	"testl $1, %0			\n\t"
+	    " jz 1f				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	8(%1), %%mm1		\n\t"
+	"movq	(%2), %%mm2		\n\t"
+	"movq	8(%2), %%mm3		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
+	"pxor %%mm6, %%mm2		\n\t"
+	"pxor %%mm6, %%mm3		\n\t"
+	PAVGB" %%mm2, %%mm0		\n\t"
+	PAVGB" %%mm3, %%mm1		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"addl	$16, %2			\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"movq	%%mm1, 8(%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"decl	%0			\n\t"
+	"1:				\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	8(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"movq	(%2), %%mm2		\n\t"
+	"movq	8(%2), %%mm3		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
+	"pxor %%mm6, %%mm2		\n\t"
+	"pxor %%mm6, %%mm3		\n\t"
+	PAVGB" %%mm2, %%mm0		\n\t"
+	PAVGB" %%mm3, %%mm1		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
+	"movq	%%mm0, (%3)		\n\t"
+	"movq	%%mm1, 8(%3)		\n\t"
+	"addl	%5, %3			\n\t"
+	"movq	(%1), %%mm0		\n\t"
+	"movq	8(%1), %%mm1		\n\t"
+	"addl	%4, %1			\n\t"
+	"movq	16(%2), %%mm2		\n\t"
+	"movq	24(%2), %%mm3		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
+	"pxor %%mm6, %%mm2		\n\t"
+	"pxor %%mm6, %%mm3		\n\t"
+	PAVGB" %%mm2, %%mm0		\n\t"
+	PAVGB" %%mm3, %%mm1		\n\t"
+	"pxor %%mm6, %%mm0		\n\t"
+	"pxor %%mm6, %%mm1		\n\t"
 	"movq	%%mm0, (%3)		\n\t"
 	"movq	%%mm1, 8(%3)		\n\t"
 	"addl	%5, %3			\n\t"
         "addl	$32, %2			\n\t"
 	"subl	$2, %0			\n\t"
 	"jnz	1b			\n\t"
-	:"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst)
+#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"); 
+//the following should be used, though better not with gcc ...
+/*	:"+g"(h), "+r"(src1), "+r"(src2), "+r"(dst)
 	:"r"(src1Stride), "r"(dstStride)
-	:"memory");
+	:"memory");*/
 }
  
 /* GL: this function does incorrect rounding if overflow */