mpeg2 qscale flag

a few bugfixes for the C pp code
minor optimization

Originally committed as revision 1786 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/avcodec.h

@@ -15,8 +15,8 @@ extern "C" {
 
 #define LIBAVCODEC_VERSION_INT 0x000406
 #define LIBAVCODEC_VERSION     "0.4.6"
-#define LIBAVCODEC_BUILD       4663
-#define LIBAVCODEC_BUILD_STR   "4663"
+#define LIBAVCODEC_BUILD       4664
+#define LIBAVCODEC_BUILD_STR   "4664"
 
 #define LIBAVCODEC_IDENT	"FFmpeg" LIBAVCODEC_VERSION "b" LIBAVCODEC_BUILD_STR
 
@@ -314,8 +314,15 @@ static const int Motion_Est_QTab[] = { ME_ZERO, ME_PHODS, ME_LOG,
      * - encoding: unused\
      * - decoding: set by lavc\
      */\
-    int repeat_pict;
+    int repeat_pict;\
+    \
+    /**\
+     * \
+     */\
+    int qscale_type;\
 
+#define FF_QSCALE_TYPE_MPEG1	0
+#define FF_QSCALE_TYPE_MPEG2	1
 
 #define FF_BUFFER_TYPE_INTERNAL 1
 #define FF_BUFFER_TYPE_USER     2 ///< Direct rendering buffers

libavcodec/libpostproc/postprocess.c

@@ -199,8 +199,9 @@ static inline int isHorizDC(uint8_t src[], int stride, PPContext *c)
 {
 	int numEq= 0;
 	int y;
-	const int dcOffset= ((c->QP*c->ppMode.baseDcDiff)>>8) + 1;
+	const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
 	const int dcThreshold= dcOffset*2 + 1;
+
 	for(y=0; y<BLOCK_SIZE; y++)
 	{
 		if(((unsigned)(src[0] - src[1] + dcOffset)) < dcThreshold) numEq++;
@@ -221,8 +222,9 @@ static inline int isHorizDC(uint8_t src[], int stride, PPContext *c)
 static inline int isVertDC_C(uint8_t src[], int stride, PPContext *c){
 	int numEq= 0;
 	int y;
-	const int dcOffset= ((c->QP*c->ppMode.baseDcDiff)>>8) + 1;
+	const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
 	const int dcThreshold= dcOffset*2 + 1;
+
 	src+= stride*4; // src points to begin of the 8x8 Block
 	for(y=0; y<BLOCK_SIZE-1; y++)
 	{
@@ -735,12 +737,13 @@ static void reallocAlign(void **p, int alignment, int size){
 	memset(*p, 0, size);
 }
 
-static void reallocBuffers(PPContext *c, int width, int height, int stride){
+static void reallocBuffers(PPContext *c, int width, int height, int stride, int qpStride){
 	int mbWidth = (width+15)>>4;
 	int mbHeight= (height+15)>>4;
 	int i;
 
 	c->stride= stride;
+	c->qpStride= qpStride;
 
 	reallocAlign((void **)&c->tempDst, 8, stride*24);
 	reallocAlign((void **)&c->tempSrc, 8, stride*24);
@@ -757,7 +760,8 @@ static void reallocBuffers(PPContext *c, int width, int height, int stride){
 	}
 
 	reallocAlign((void **)&c->deintTemp, 8, 2*width+32);
-	reallocAlign((void **)&c->nonBQPTable, 8, mbWidth*mbHeight*sizeof(QP_STORE_T));
+	reallocAlign((void **)&c->nonBQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
+	reallocAlign((void **)&c->stdQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
 	reallocAlign((void **)&c->forcedQPTable, 8, mbWidth*sizeof(QP_STORE_T));
 }
 
@@ -772,6 +776,7 @@ static void global_init(){
 pp_context_t *pp_get_context(int width, int height, int cpuCaps){
 	PPContext *c= memalign(32, sizeof(PPContext));
 	int stride= (width+15)&(~15); //assumed / will realloc if needed
+	int qpStride= (width+15)/16 + 2; //assumed / will realloc if needed
         
 	global_init();
 
@@ -785,7 +790,7 @@ pp_context_t *pp_get_context(int width, int height, int cpuCaps){
 		c->vChromaSubSample= 1;
 	}
 
-	reallocBuffers(c, width, height, stride);
+	reallocBuffers(c, width, height, stride, qpStride);
         
 	c->frameNum=-1;
 
@@ -804,6 +809,7 @@ void pp_free_context(void *vc){
 	free(c->tempDst);
 	free(c->tempSrc);
 	free(c->deintTemp);
+	free(c->stdQPTable);
 	free(c->nonBQPTable);
 	free(c->forcedQPTable);
         
@@ -823,9 +829,11 @@ void  pp_postprocess(uint8_t * src[3], int srcStride[3],
 	PPMode *mode = (PPMode*)vm;
 	PPContext *c = (PPContext*)vc;
         int minStride= MAX(srcStride[0], dstStride[0]);
-	
-	if(c->stride < minStride)
-		reallocBuffers(c, width, height, minStride);
+
+	if(c->stride < minStride || c->qpStride < QPStride)
+		reallocBuffers(c, width, height, 
+				MAX(minStride, c->stride), 
+				MAX(c->qpStride, QPStride));
 
 	if(QP_store==NULL || (mode->lumMode & FORCE_QUANT)) 
 	{
@@ -837,6 +845,20 @@ void  pp_postprocess(uint8_t * src[3], int srcStride[3],
 		else
 			for(i=0; i<mbWidth; i++) QP_store[i]= 1;
 	}
+//printf("pict_type:%d\n", pict_type);
+
+	if(pict_type & PP_PICT_TYPE_QP2){
+		int i;
+		const int count= mbHeight * QPStride;
+		for(i=0; i<(count>>2); i++){
+			((uint32_t*)c->stdQPTable)[i] = (((uint32_t*)QP_store)[i]>>1) & 0x7F7F7F7F;
+		}
+		for(i<<=2; i<count; i++){
+			c->stdQPTable[i] = QP_store[i]>>1;
+		}
+                QP_store= c->stdQPTable;
+	}
+
 if(0){
 int x,y;
 for(y=0; y<mbHeight; y++){
@@ -847,18 +869,16 @@ for(y=0; y<mbHeight; y++){
 }
 	printf("\n");
 }
-//printf("pict_type:%d\n", pict_type);
 
-	if(pict_type!=3)
+	if((pict_type&7)!=3)
 	{
-		int x,y;
-		for(y=0; y<mbHeight; y++){
-			for(x=0; x<mbWidth; x++){
-				int qscale= QP_store[x + y*QPStride];
-				if(qscale&~31)
-				    qscale=31;
-				c->nonBQPTable[y*mbWidth + x]= qscale;
-			}
+		int i;
+		const int count= mbHeight * QPStride;
+		for(i=0; i<(count>>2); i++){
+			((uint32_t*)c->nonBQPTable)[i] = ((uint32_t*)QP_store)[i] & 0x1F1F1F1F;
+		}
+		for(i<<=2; i<count; i++){
+			c->nonBQPTable[i] = QP_store[i] & 0x1F;
 		}
 	}
 

libavcodec/libpostproc/postprocess.h

@@ -66,6 +66,8 @@ void pp_free_context(pp_context_t *ppContext);
 #define PP_FORMAT_411    (0x00000002|PP_FORMAT)
 #define PP_FORMAT_444    (0x00000000|PP_FORMAT)
 
+#define PP_PICT_TYPE_QP2  0x00000010 ///< MPEG2 style QScale
+
 #ifdef __cplusplus
 }
 #endif

libavcodec/libpostproc/postprocess_internal.h

@@ -137,6 +137,7 @@ typedef struct PPContext{
 	uint64_t __attribute__((aligned(8))) mmxDcOffset[32];
 	uint64_t __attribute__((aligned(8))) mmxDcThreshold[32];
 
+	QP_STORE_T *stdQPTable;       ///< used to fix MPEG2 style qscale
 	QP_STORE_T *nonBQPTable;
 	QP_STORE_T *forcedQPTable;
 
@@ -147,7 +148,8 @@ typedef struct PPContext{
 	
 	int cpuCaps;
         
-	int stride; ///<size of some buffers (needed to realloc them if needed)
+	int qpStride; ///<size of qp buffers (needed to realloc them if needed)
+	int stride;   ///<size of some buffers (needed to realloc them if needed)
         
 	int hChromaSubSample;
 	int vChromaSubSample;
@@ -155,3 +157,4 @@ typedef struct PPContext{
 	PPMode ppMode;
 } PPContext;
 
+

libavcodec/libpostproc/postprocess_template.c

@@ -2993,7 +2993,7 @@ static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int
 		uint8_t *tempBlock2= c.tempBlocks + 8;
 #endif
 		int8_t *QPptr= &QPs[(y>>qpVShift)*QPStride];
-		int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*mbWidth];
+		int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*QPStride];
 		int QP=0;
 		/* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
 		   if not than use a temporary buffer */

libavcodec/mpeg12.c

@@ -2001,8 +2001,11 @@ eos: //end of slice
     if (s->mb_y<<field_pic == s->mb_height && !s->first_field) {
         /* end of image */
 
-        if(s->mpeg2)
-            s->qscale >>=1;
+        if(s->mpeg2){
+            s->current_picture_ptr->qscale_type= FF_QSCALE_TYPE_MPEG2;
+        }else
+            s->current_picture_ptr->qscale_type= FF_QSCALE_TYPE_MPEG1;
+
         ff_er_frame_end(s);
 
         MPV_frame_end(s);