comparison libpostproc/postprocess.c @ 1196:2e06398e4647 libavcodec

mpeg2 qscale flag a few bugfixes for the C pp code minor optimization

1195:f7522f310c7e

 */
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 dcThreshold= dcOffset*2 + 1;
	for(y=0; y<BLOCK_SIZE; y++)
	{
		if(((unsigned)(src[0] - src[1] + dcOffset)) < dcThreshold) numEq++;
		if(((unsigned)(src[1] - src[2] + dcOffset)) < dcThreshold) numEq++;
		if(((unsigned)(src[2] - src[3] + dcOffset)) < dcThreshold) numEq++;

 * Check if the middle 8x8 Block in the given 8x16 block is flat
 */
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 dcThreshold= dcOffset*2 + 1;
	src+= stride*4; // src points to begin of the 8x8 Block
	for(y=0; y<BLOCK_SIZE-1; y++)
	{
		if(((unsigned)(src[0] - src[0+stride] + dcOffset)) < dcThreshold) numEq++;
		if(((unsigned)(src[1] - src[1+stride] + dcOffset)) < dcThreshold) numEq++;

	if(*p) free(*p);
	*p= memalign(alignment, size);
	memset(*p, 0, size);
}

static void reallocBuffers(PPContext *c, int width, int height, int stride){
	int mbWidth = (width+15)>>4;
	int mbHeight= (height+15)>>4;
	int i;

	c->stride= stride;

	reallocAlign((void **)&c->tempDst, 8, stride*24);
	reallocAlign((void **)&c->tempSrc, 8, stride*24);
	reallocAlign((void **)&c->tempBlocks, 8, 2*16*8);
	reallocAlign((void **)&c->yHistogram, 8, 256*sizeof(uint64_t));

		reallocAlign((void **)&c->tempBlured[i], 8, stride*mbHeight*16 + 17*1024);
		reallocAlign((void **)&c->tempBluredPast[i], 8, 256*((height+7)&(~7))/2 + 17*1024);//FIXME size
	}

	reallocAlign((void **)&c->deintTemp, 8, 2*width+32);
	reallocAlign((void **)&c->nonBQPTable, 8, mbWidth*mbHeight*sizeof(QP_STORE_T));
	reallocAlign((void **)&c->forcedQPTable, 8, mbWidth*sizeof(QP_STORE_T));
}

static void global_init(){
	int i;

}

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
        
	global_init();

	memset(c, 0, sizeof(PPContext));
	c->cpuCaps= cpuCaps;

	}else{
		c->hChromaSubSample= 1;
		c->vChromaSubSample= 1;
	}

	reallocBuffers(c, width, height, stride);
        
	c->frameNum=-1;

	return c;
}

	free(c->tempBlocks);
	free(c->yHistogram);
	free(c->tempDst);
	free(c->tempSrc);
	free(c->deintTemp);
	free(c->nonBQPTable);
	free(c->forcedQPTable);
        
	memset(c, 0, sizeof(PPContext));

	int mbWidth = (width+15)>>4;
	int mbHeight= (height+15)>>4;
	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(QP_store==NULL || (mode->lumMode & FORCE_QUANT)) 
	{
		int i;
		QP_store= c->forcedQPTable;

		if(mode->lumMode & FORCE_QUANT)
			for(i=0; i<mbWidth; i++) QP_store[i]= mode->forcedQuant;
		else
			for(i=0; i<mbWidth; i++) QP_store[i]= 1;
	}
if(0){
int x,y;
for(y=0; y<mbHeight; y++){
	for(x=0; x<mbWidth; x++){
		printf("%2d ", QP_store[x + y*QPStride]);
	}
	printf("\n");
}
	printf("\n");
}
//printf("pict_type:%d\n", pict_type);

	if(pict_type!=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;
			}
		}
	}

	if(verbose>2)
	{

1196:2e06398e4647

 */
static inline int isHorizDC(uint8_t src[], int stride, PPContext *c)
{
	int numEq= 0;
	int y;
	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++;
		if(((unsigned)(src[1] - src[2] + dcOffset)) < dcThreshold) numEq++;
		if(((unsigned)(src[2] - src[3] + dcOffset)) < dcThreshold) numEq++;

 * Check if the middle 8x8 Block in the given 8x16 block is flat
 */
static inline int isVertDC_C(uint8_t src[], int stride, PPContext *c){
	int numEq= 0;
	int y;
	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++)
	{
		if(((unsigned)(src[0] - src[0+stride] + dcOffset)) < dcThreshold) numEq++;
		if(((unsigned)(src[1] - src[1+stride] + dcOffset)) < dcThreshold) numEq++;

	if(*p) free(*p);
	*p= memalign(alignment, size);
	memset(*p, 0, size);
}

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);
	reallocAlign((void **)&c->tempBlocks, 8, 2*16*8);
	reallocAlign((void **)&c->yHistogram, 8, 256*sizeof(uint64_t));

		reallocAlign((void **)&c->tempBlured[i], 8, stride*mbHeight*16 + 17*1024);
		reallocAlign((void **)&c->tempBluredPast[i], 8, 256*((height+7)&(~7))/2 + 17*1024);//FIXME size
	}

	reallocAlign((void **)&c->deintTemp, 8, 2*width+32);
	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));
}

static void global_init(){
	int i;

}

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();

	memset(c, 0, sizeof(PPContext));
	c->cpuCaps= cpuCaps;

	}else{
		c->hChromaSubSample= 1;
		c->vChromaSubSample= 1;
	}

	reallocBuffers(c, width, height, stride, qpStride);
        
	c->frameNum=-1;

	return c;
}

	free(c->tempBlocks);
	free(c->yHistogram);
	free(c->tempDst);
	free(c->tempSrc);
	free(c->deintTemp);
	free(c->stdQPTable);
	free(c->nonBQPTable);
	free(c->forcedQPTable);
        
	memset(c, 0, sizeof(PPContext));

	int mbWidth = (width+15)>>4;
	int mbHeight= (height+15)>>4;
	PPMode *mode = (PPMode*)vm;
	PPContext *c = (PPContext*)vc;
        int minStride= MAX(srcStride[0], dstStride[0]);

	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)) 
	{
		int i;
		QP_store= c->forcedQPTable;

		if(mode->lumMode & FORCE_QUANT)
			for(i=0; i<mbWidth; i++) QP_store[i]= mode->forcedQuant;
		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++){
	for(x=0; x<mbWidth; x++){
		printf("%2d ", QP_store[x + y*QPStride]);
	}
	printf("\n");
}
	printf("\n");
}

	if((pict_type&7)!=3)
	{
		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;
		}
	}

	if(verbose>2)
	{