super mega ffmpeg tree sync

CVS patchset: 5615
CVS date: 2003/10/27 15:24:38

2 files changed, 139 insertions, 94 deletions

diff --git a/src/libffmpeg/libavcodec/libpostproc/postprocess.c b/src/libffmpeg/libavcodec/libpostproc/postprocess.c
index 470c889d8..0db5341e3 100644
--- a/src/libffmpeg/libavcodec/libpostproc/postprocess.c
+++ b/src/libffmpeg/libavcodec/libpostproc/postprocess.c
@@ -213,9 +213,77 @@ static inline int isVertDC_C(uint8_t src[], int stride, PPContext *c){
 
 static inline int isHorizMinMaxOk(uint8_t src[], int stride, int QP)
 {
-	if(abs(src[0] - src[7]) > 2*QP) return 0;
+	int i;
+#if 1
+	for(i=0; i<2; i++){
+		if((unsigned)(src[0] - src[5] + 2*QP) > 4*QP) return 0;
+		src += stride;
+		if((unsigned)(src[2] - src[7] + 2*QP) > 4*QP) return 0;
+		src += stride;
+		if((unsigned)(src[4] - src[1] + 2*QP) > 4*QP) return 0;
+		src += stride;
+		if((unsigned)(src[6] - src[3] + 2*QP) > 4*QP) return 0;
+		src += stride;
+	}
+#else        
+	for(i=0; i<8; i++){
+		if((unsigned)(src[0] - src[7] + 2*QP) > 4*QP) return 0;
+		src += stride;
+	}
+#endif
+	return 1;
+}
 
+static inline int isVertMinMaxOk_C(uint8_t src[], int stride, int QP)
+{
+#if 1
+#if 1
+	int x;
+	src+= stride*4;
+	for(x=0; x<BLOCK_SIZE; x+=4)
+	{
+		if((unsigned)(src[  x + 0*stride] - src[  x + 5*stride] + 2*QP) > 4*QP) return 0;
+		if((unsigned)(src[1+x + 2*stride] - src[1+x + 7*stride] + 2*QP) > 4*QP) return 0;
+		if((unsigned)(src[2+x + 4*stride] - src[2+x + 1*stride] + 2*QP) > 4*QP) return 0;
+		if((unsigned)(src[3+x + 6*stride] - src[3+x + 3*stride] + 2*QP) > 4*QP) return 0;
+	}
+#else
+	int x;
+	src+= stride*3;
+	for(x=0; x<BLOCK_SIZE; x++)
+	{
+		if((unsigned)(src[x + stride] - src[x + (stride<<3)] + 2*QP) > 4*QP) return 0;
+	}
+#endif
 	return 1;
+#else
+	int x;
+	src+= stride*4;
+	for(x=0; x<BLOCK_SIZE; x++)
+	{
+		int min=255;
+		int max=0;
+		int y;
+		for(y=0; y<8; y++){
+			int v= src[x + y*stride];
+			if(v>max) max=v;
+			if(v<min) min=v;
+		}
+		if(max-min > 2*QP) return 0;
+	}
+	return 1;
+#endif
+}
+
+static inline int vertClassify_C(uint8_t src[], int stride, PPContext *c){
+	if( isVertDC_C(src, stride, c) ){
+		if( isVertMinMaxOk_C(src, stride, c->QP) )
+			return 1;
+		else
+			return 0;
+	}else{
+		return 2;
+	}
 }
 
 static inline void doHorizDefFilter(uint8_t dst[], int stride, int QP)
@@ -481,18 +549,18 @@ char *pp_help=
 "tn:64:128:256\n"
 "Filters			Options\n"
 "short	long name	short	long option	Description\n"
-"*	*		a	autoq		cpu power dependant enabler\n"
-"			c	chrom		chrominance filtring enabled\n"
-"			y	nochrom		chrominance filtring disabled\n"
-"hb	hdeblock	(2 Threshold)		horizontal deblocking filter\n"
+"*	*		a	autoq		CPU power dependent enabler\n"
+"			c	chrom		chrominance filtering enabled\n"
+"			y	nochrom		chrominance filtering disabled\n"
+"hb	hdeblock	(2 threshold)		horizontal deblocking filter\n"
 "	1. difference factor: default=32, higher -> more deblocking\n"
 "	2. flatness threshold: default=39, lower -> more deblocking\n"
 "			the h & v deblocking filters share these\n"
-"			so u cant set different thresholds for h / v\n"
-"vb	vdeblock	(2 Threshold)		vertical deblocking filter\n"
-"h1	x1hdeblock				Experimental h deblock filter 1\n"
-"v1	x1vdeblock				Experimental v deblock filter 1\n"
-"dr	dering					Deringing filter\n"
+"			so you can't set different thresholds for h / v\n"
+"vb	vdeblock	(2 threshold)		vertical deblocking filter\n"
+"h1	x1hdeblock				experimental h deblock filter 1\n"
+"v1	x1vdeblock				experimental v deblock filter 1\n"
+"dr	dering					deringing filter\n"
 "al	autolevels				automatic brightness / contrast\n"
 "			f	fullyrange	stretch luminance to (0..255)\n"
 "lb	linblenddeint				linear blend deinterlacer\n"
@@ -502,9 +570,9 @@ char *pp_help=
 "fd	ffmpegdeint				ffmpeg deinterlacer\n"
 "de	default					hb:a,vb:a,dr:a,al\n"
 "fa	fast					h1:a,v1:a,dr:a,al\n"
-"tn	tmpnoise	(3 Thresholds)		Temporal Noise Reducer\n"
+"tn	tmpnoise	(3 threshold)		temporal noise reducer\n"
 "			1. <= 2. <= 3.		larger -> stronger filtering\n"
-"fq	forceQuant	<quantizer>		Force quantizer\n"
+"fq	forceQuant	<quantizer>		force quantizer\n"
 ;
 
 pp_mode_t *pp_get_mode_by_name_and_quality(char *name, int quality)
diff --git a/src/libffmpeg/libavcodec/libpostproc/postprocess_template.c b/src/libffmpeg/libavcodec/libpostproc/postprocess_template.c
index 48027e417..582ad28f2 100644
--- a/src/libffmpeg/libavcodec/libpostproc/postprocess_template.c
+++ b/src/libffmpeg/libavcodec/libpostproc/postprocess_template.c
@@ -49,65 +49,85 @@
 	"paddb " #a ", " #b " \n\t"
 #endif
 
-
 //FIXME? |255-0| = 1 (shouldnt be a problem ...)
 #ifdef HAVE_MMX
 /**
  * Check if the middle 8x8 Block in the given 8x16 block is flat
  */
-static inline int RENAME(isVertDC)(uint8_t src[], int stride, PPContext *c){
-	int numEq= 0;
+static inline int RENAME(vertClassify)(uint8_t src[], int stride, PPContext *c){
+	int numEq= 0, dcOk;
 	src+= stride*4; // src points to begin of the 8x8 Block
 asm volatile(
-		"leal (%1, %2), %%eax				\n\t"
+		"movq %0, %%mm7					\n\t" 
+		"movq %1, %%mm6					\n\t" 
+                : : "m" (c->mmxDcOffset[c->nonBQP]),  "m" (c->mmxDcThreshold[c->nonBQP])
+                );
+                
+asm volatile(
+		"leal (%2, %3), %%eax				\n\t"
 //	0	1	2	3	4	5	6	7	8	9
 //	%1	eax	eax+%2	eax+2%2	%1+4%2	ecx	ecx+%2	ecx+2%2	%1+8%2	ecx+4%2
-		"movq %3, %%mm7					\n\t" 
-		"movq %4, %%mm6					\n\t" 
 
-		"movq (%1), %%mm0				\n\t"
+		"movq (%2), %%mm0				\n\t"
 		"movq (%%eax), %%mm1				\n\t"
+                "movq %%mm0, %%mm3				\n\t"
+                "movq %%mm0, %%mm4				\n\t"
+                PMAXUB(%%mm1, %%mm4)
+                PMINUB(%%mm1, %%mm3, %%mm5)
 		"psubb %%mm1, %%mm0				\n\t" // mm0 = differnece
 		"paddb %%mm7, %%mm0				\n\t"
 		"pcmpgtb %%mm6, %%mm0				\n\t"
 
-		"movq (%%eax,%2), %%mm2				\n\t"
+		"movq (%%eax,%3), %%mm2				\n\t"
+                PMAXUB(%%mm2, %%mm4)
+                PMINUB(%%mm2, %%mm3, %%mm5)
 		"psubb %%mm2, %%mm1				\n\t"
 		"paddb %%mm7, %%mm1				\n\t"
 		"pcmpgtb %%mm6, %%mm1				\n\t"
 		"paddb %%mm1, %%mm0				\n\t"
 
-		"movq (%%eax, %2, 2), %%mm1			\n\t"
+		"movq (%%eax, %3, 2), %%mm1			\n\t"
+                PMAXUB(%%mm1, %%mm4)
+                PMINUB(%%mm1, %%mm3, %%mm5)
 		"psubb %%mm1, %%mm2				\n\t"
 		"paddb %%mm7, %%mm2				\n\t"
 		"pcmpgtb %%mm6, %%mm2				\n\t"
 		"paddb %%mm2, %%mm0				\n\t"
 		
-		"leal (%%eax, %2, 4), %%eax			\n\t"
+		"leal (%%eax, %3, 4), %%eax			\n\t"
 
-		"movq (%1, %2, 4), %%mm2			\n\t"
+		"movq (%2, %3, 4), %%mm2			\n\t"
+                PMAXUB(%%mm2, %%mm4)
+                PMINUB(%%mm2, %%mm3, %%mm5)
 		"psubb %%mm2, %%mm1				\n\t"
 		"paddb %%mm7, %%mm1				\n\t"
 		"pcmpgtb %%mm6, %%mm1				\n\t"
 		"paddb %%mm1, %%mm0				\n\t"
 
 		"movq (%%eax), %%mm1				\n\t"
+                PMAXUB(%%mm1, %%mm4)
+                PMINUB(%%mm1, %%mm3, %%mm5)
 		"psubb %%mm1, %%mm2				\n\t"
 		"paddb %%mm7, %%mm2				\n\t"
 		"pcmpgtb %%mm6, %%mm2				\n\t"
 		"paddb %%mm2, %%mm0				\n\t"
 
-		"movq (%%eax, %2), %%mm2			\n\t"
+		"movq (%%eax, %3), %%mm2			\n\t"
+                PMAXUB(%%mm2, %%mm4)
+                PMINUB(%%mm2, %%mm3, %%mm5)
 		"psubb %%mm2, %%mm1				\n\t"
 		"paddb %%mm7, %%mm1				\n\t"
 		"pcmpgtb %%mm6, %%mm1				\n\t"
 		"paddb %%mm1, %%mm0				\n\t"
 
-		"movq (%%eax, %2, 2), %%mm1			\n\t"
+		"movq (%%eax, %3, 2), %%mm1			\n\t"
+                PMAXUB(%%mm1, %%mm4)
+                PMINUB(%%mm1, %%mm3, %%mm5)
 		"psubb %%mm1, %%mm2				\n\t"
 		"paddb %%mm7, %%mm2				\n\t"
 		"pcmpgtb %%mm6, %%mm2				\n\t"
 		"paddb %%mm2, %%mm0				\n\t"
+		"psubusb %%mm3, %%mm4				\n\t"
 
 		"						\n\t"
 #ifdef HAVE_MMX2
@@ -124,69 +144,27 @@ asm volatile(
 		"psrlq $32, %%mm0				\n\t"
 		"paddb %%mm1, %%mm0				\n\t"
 #endif
-		"movd %%mm0, %0					\n\t"
-		: "=r" (numEq)
-		: "r" (src), "r" (stride), "m" (c->mmxDcOffset[c->nonBQP]),  "m" (c->mmxDcThreshold[c->nonBQP])
-		: "%eax"
-		);
-	numEq= (-numEq) &0xFF;
-	return numEq > c->ppMode.flatnessThreshold;
-}
-#endif
-
-static inline int RENAME(isVertMinMaxOk)(uint8_t src[], int stride, PPContext *c)
-{
-#ifdef HAVE_MMX
-	int isOk;
-	src+= stride*3;
-	asm volatile(
-		"movq (%1, %2), %%mm0				\n\t"
-		"movq (%1, %2, 8), %%mm1			\n\t"
-		"movq %%mm0, %%mm2				\n\t"
-		"psubusb %%mm1, %%mm0				\n\t"
-		"psubusb %%mm2, %%mm1				\n\t"
-		"por %%mm1, %%mm0				\n\t" // ABS Diff
-
-		"movq %3, %%mm7					\n\t" // QP,..., QP
+                "movq %4, %%mm7					\n\t" // QP,..., QP
 		"paddusb %%mm7, %%mm7				\n\t" // 2QP ... 2QP
-		"psubusb %%mm7, %%mm0				\n\t" // Diff <= 2QP -> 0
-		"packssdw %%mm0, %%mm0				\n\t"
+		"psubusb %%mm7, %%mm4				\n\t" // Diff <= 2QP -> 0
+		"packssdw %%mm4, %%mm4				\n\t"
 		"movd %%mm0, %0					\n\t"
-		: "=r" (isOk)
+		"movd %%mm4, %1					\n\t"
+
+		: "=r" (numEq), "=r" (dcOk)
 		: "r" (src), "r" (stride), "m" (c->pQPb)
+		: "%eax"
 		);
-	return isOk==0;
-#else
-#if 1
-	int x;
-	const int QP= c->QP;
-	src+= stride*3;
-	for(x=0; x<BLOCK_SIZE; x++)
-	{
-		if((unsigned)(src[x + stride] - src[x + (stride<<3)] + 2*QP) > 4*QP) return 0;
-	}
 
-	return 1;
-#else
-	int x;
-	const int QP= c->QP;
-	src+= stride*4;
-	for(x=0; x<BLOCK_SIZE; x++)
-	{
-		int min=255;
-		int max=0;
-		int y;
-		for(y=0; y<8; y++){
-			int v= src[x + y*stride];
-			if(v>max) max=v;
-			if(v<min) min=v;
-		}
-		if(max-min > 2*QP) return 0;
-	}
-	return 1;
-#endif
-#endif
+	numEq= (-numEq) &0xFF;
+	if(numEq > c->ppMode.flatnessThreshold){
+            if(dcOk) return 0;
+            else     return 1;
+        }else{
+            return 2;
+        }
 }
+#endif
 
 /**
  * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
@@ -3119,12 +3097,11 @@ static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int
 					RENAME(vertX1Filter)(dstBlock, stride, &c);
 				else if(mode & V_DEBLOCK)
 				{
-					if( RENAME(isVertDC)(dstBlock, stride, &c))
-					{
-						if(RENAME(isVertMinMaxOk)(dstBlock, stride, &c))
-							RENAME(doVertLowPass)(dstBlock, stride, &c);
-					}
-					else
+					const int t= RENAME(vertClassify)(dstBlock, stride, &c);
+
+					if(t==1)
+						RENAME(doVertLowPass)(dstBlock, stride, &c);
+					else if(t==2)
 						RENAME(doVertDefFilter)(dstBlock, stride, &c);
 				}
 			}
@@ -3140,12 +3117,12 @@ static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int
 					RENAME(vertX1Filter)(tempBlock1, 16, &c);
 				else if(mode & H_DEBLOCK)
 				{
-					if( RENAME(isVertDC)(tempBlock1, 16, &c))
-					{
-						if(RENAME(isVertMinMaxOk)(tempBlock1, 16, &c))
-							RENAME(doVertLowPass)(tempBlock1, 16, &c);
-					}
-					else
+//START_TIMER
+					const int t= RENAME(vertClassify)(tempBlock1, 16, &c);
+//STOP_TIMER("dc & minmax")
+                                        if(t==1)
+						RENAME(doVertLowPass)(tempBlock1, 16, &c);
+					else if(t==2)
 						RENAME(doVertDefFilter)(tempBlock1, 16, &c);
 				}