comparison libpostproc/postprocess.c @ 791:4f61ca80b6c1 libavcodec

better deblocking filter

790:b9156f8e6747

	(the if/else stuff per block is slowing things down)
compare the quality & speed of all filters
split this huge file
optimize c versions
try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks
put fastmemcpy back
...
*/

//Changelog: use the CVS log

	uint8_t *tempSrcBlock;
	uint8_t *deintTemp;

	uint64_t __attribute__((aligned(8))) pQPb;
	uint64_t __attribute__((aligned(8))) pQPb2;
	
	uint64_t __attribute__((aligned(8))) mmxDcOffset;
	uint64_t __attribute__((aligned(8))) mmxDcThreshold;

	int QP;
	int dcOffset;
	int dcThreshold;

	int frameNum;

	PPMode ppMode;
} PPContext;

 */
static inline int isHorizDC(uint8_t src[], int stride, PPContext *c)
{
	int numEq= 0;
	int y;
	const int dcOffset= c->dcOffset;
	const int dcThreshold= c->dcThreshold;
	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->dcOffset;
	const int dcThreshold= c->dcThreshold;
	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++;

					}
				}
				else if(filters[i].mask == V_DEBLOCK || filters[i].mask == H_DEBLOCK)
				{
					int o;
					ppMode.maxDcDiff=1;
//					hFlatnessThreshold= 40;
//					vFlatnessThreshold= 40;

					for(o=0; options[o]!=NULL && o<2; o++)
					{
						char *tail;
						int val= strtol(options[o], &tail, 0);
						if(tail==options[o]) break;

						numOfUnknownOptions--;
						if(o==0) ppMode.maxDcDiff= val;
						else ppMode.flatnessThreshold= val;
					}
				}
				else if(filters[i].mask == FORCE_QUANT)
				{

}

void *getPPContext(int width, int height){
	PPContext *c= memalign(32, sizeof(PPContext));
	int i;

	c->tempBlocks= (uint8_t*)memalign(8, 2*16*8);
	c->yHistogram= (uint64_t*)memalign(8, 256*sizeof(uint64_t));
	for(i=0; i<256; i++)
		c->yHistogram[i]= width*height/64*15/256;

	c->tempDst= (uint8_t*)memalign(8, 1024*24);
	c->tempSrc= (uint8_t*)memalign(8, 1024*24);
	c->tempDstBlock= (uint8_t*)memalign(8, 1024*24);
	c->tempSrcBlock= (uint8_t*)memalign(8, 1024*24);
	c->deintTemp= (uint8_t*)memalign(8, width+16);

	c->frameNum=-1;

	return c;
}

	free(c->tempDst);
	free(c->tempSrc);
	free(c->tempDstBlock);
	free(c->tempSrcBlock);
	free(c->deintTemp);
	
	free(c);
}

//FIXME move this shit away from here

}


void  postprocess(uint8_t * src[3], int srcStride[3],
                 uint8_t * dst[3], int dstStride[3],
                 int horizontalSize, int verticalSize,
                 QP_STORE_T *QP_store,  int QPStride,
		 PPMode *mode,  void *c)
{

	QP_STORE_T quantArray[2048/8];

	if(QP_store==NULL || (mode->lumMode & FORCE_QUANT)) 
	{
		int i;
		QP_store= quantArray;

		if(mode->lumMode & FORCE_QUANT)
			for(i=0; i<2048/8; i++) quantArray[i]= mode->forcedQuant;
		else
			for(i=0; i<2048/8; i++) quantArray[i]= 1;
	}

	if(firstTime2 && verbose)
	{
		printf("using npp filters 0x%X/0x%X\n", mode->lumMode, mode->chromMode);
		firstTime2=0;
	}

	postProcess(src[0], srcStride[0], dst[0], dstStride[0],
		horizontalSize, verticalSize, QP_store, QPStride, 0, mode, c);

	horizontalSize = (horizontalSize+1)>> 1;
	verticalSize   = (verticalSize+1)>>1;

	if(mode->chromMode)
	{
		postProcess(src[1], srcStride[1], dst[1], dstStride[1],
			horizontalSize, verticalSize, QP_store, QPStride, 1, mode, c);
		postProcess(src[2], srcStride[2], dst[2], dstStride[2],
			horizontalSize, verticalSize, QP_store, QPStride, 2, mode, c);
	}
	else if(srcStride[1] == dstStride[1] && srcStride[2] == dstStride[2])
	{
		memcpy(dst[1], src[1], srcStride[1]*verticalSize);
		memcpy(dst[2], src[2], srcStride[2]*verticalSize);
	}
	else
	{
		int y;
		for(y=0; y<verticalSize; y++)
		{
			memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), horizontalSize);
			memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), horizontalSize);
		}
	}
}

791:4f61ca80b6c1

	(the if/else stuff per block is slowing things down)
compare the quality & speed of all filters
split this huge file
optimize c versions
try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks
...
*/

//Changelog: use the CVS log

	uint8_t *tempSrcBlock;
	uint8_t *deintTemp;

	uint64_t __attribute__((aligned(8))) pQPb;
	uint64_t __attribute__((aligned(8))) pQPb2;

	uint64_t __attribute__((aligned(8))) mmxDcOffset[32];
	uint64_t __attribute__((aligned(8))) mmxDcThreshold[32];
	
	QP_STORE_T *nonBQPTable;
	
	int QP;
	int nonBQP;

	int frameNum;

	PPMode ppMode;
} PPContext;

 */
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++;

					}
				}
				else if(filters[i].mask == V_DEBLOCK || filters[i].mask == H_DEBLOCK)
				{
					int o;
					ppMode.baseDcDiff=256/4;
//					hFlatnessThreshold= 40;
//					vFlatnessThreshold= 40;

					for(o=0; options[o]!=NULL && o<2; o++)
					{
						char *tail;
						int val= strtol(options[o], &tail, 0);
						if(tail==options[o]) break;

						numOfUnknownOptions--;
						if(o==0) ppMode.baseDcDiff= val;
						else ppMode.flatnessThreshold= val;
					}
				}
				else if(filters[i].mask == FORCE_QUANT)
				{

}

void *getPPContext(int width, int height){
	PPContext *c= memalign(32, sizeof(PPContext));
	int i;
	int mbWidth = (width+15)>>4;
	int mbHeight= (height+15)>>4;

	c->tempBlocks= (uint8_t*)memalign(8, 2*16*8);
	c->yHistogram= (uint64_t*)memalign(8, 256*sizeof(uint64_t));
	for(i=0; i<256; i++)
		c->yHistogram[i]= width*height/64*15/256;

	c->tempDst= (uint8_t*)memalign(8, 1024*24);
	c->tempSrc= (uint8_t*)memalign(8, 1024*24);
	c->tempDstBlock= (uint8_t*)memalign(8, 1024*24);
	c->tempSrcBlock= (uint8_t*)memalign(8, 1024*24);
	c->deintTemp= (uint8_t*)memalign(8, width+16);
	c->nonBQPTable= (QP_STORE_T*)memalign(8, mbWidth*mbHeight*sizeof(QP_STORE_T));
	memset(c->nonBQPTable, 0, mbWidth*mbHeight*sizeof(QP_STORE_T));

	c->frameNum=-1;

	return c;
}

	free(c->tempDst);
	free(c->tempSrc);
	free(c->tempDstBlock);
	free(c->tempSrcBlock);
	free(c->deintTemp);
	free(c->nonBQPTable);
	
	free(c);
}

//FIXME move this shit away from here

}


void  postprocess(uint8_t * src[3], int srcStride[3],
                 uint8_t * dst[3], int dstStride[3],
                 int width, int height,
                 QP_STORE_T *QP_store,  int QPStride,
		 PPMode *mode,  void *vc, int pict_type)
{
	int mbWidth = (width+15)>>4;
	int mbHeight= (height+15)>>4;
	QP_STORE_T quantArray[2048/8];
	PPContext *c = (PPContext*)vc;

	if(QP_store==NULL || (mode->lumMode & FORCE_QUANT)) 
	{
		int i;
		QP_store= quantArray;

		if(mode->lumMode & FORCE_QUANT)
			for(i=0; i<2048/8; i++) quantArray[i]= mode->forcedQuant;
		else
			for(i=0; i<2048/8; i++) quantArray[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(firstTime2 && verbose)
	{
		printf("using npp filters 0x%X/0x%X\n", mode->lumMode, mode->chromMode);
		firstTime2=0;
	}

	postProcess(src[0], srcStride[0], dst[0], dstStride[0],
		width, height, QP_store, QPStride, 0, mode, c);

	width  = (width +1)>>1;
	height = (height+1)>>1;

	if(mode->chromMode)
	{
		postProcess(src[1], srcStride[1], dst[1], dstStride[1],
			width, height, QP_store, QPStride, 1, mode, c);
		postProcess(src[2], srcStride[2], dst[2], dstStride[2],
			width, height, QP_store, QPStride, 2, mode, c);
	}
	else if(srcStride[1] == dstStride[1] && srcStride[2] == dstStride[2])
	{
		memcpy(dst[1], src[1], srcStride[1]*height);
		memcpy(dst[2], src[2], srcStride[2]*height);
	}
	else
	{
		int y;
		for(y=0; y<height; y++)
		{
			memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), width);
			memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), width);
		}
	}
}