comparison postproc/postprocess.c @ 2285:4840e356d0d3

fixed a bug in the tmp buffer fixed the color range for yuv fixed the width %8!=0 bug (another 1% speed loss)

2284:0c5fea3a0b91

static uint64_t temp2=0;
static uint64_t temp3=0;
static uint64_t temp4=0;
static uint64_t temp5=0;
static uint64_t pQPb=0;
static uint8_t tempBlock[16*16];

int hFlatnessThreshold= 56 - 16;
int vFlatnessThreshold= 56 - 16;

//amount of "black" u r willing to loose to get a brightness corrected picture
double maxClippedThreshold= 0.01;

int maxAllowedY=255;
//FIXME can never make a movieŽs black brighter (anyone needs that?)
int minAllowedY=0;

#ifdef TIMING
static inline long long rdtsc()
{
	long long l;

	/* Temporary buffers for handling the last row(s) */
	static uint8_t *tempDst= NULL;
	static uint8_t *tempSrc= NULL;

#ifdef TIMING
	long long T0, T1, memcpyTime=0, vertTime=0, horizTime=0, sumTime, diffTime=0;
	sumTime= rdtsc();
#endif

	if(tempDst==NULL)
	{
		tempDst= (uint8_t*)memalign(8, 1024*24);
		tempSrc= (uint8_t*)memalign(8, 1024*24);
	}

	if(!yHistogram)
	{
		int i;
		yHistogram= (uint64_t*)malloc(8*256);
		for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256;
	}

	if(!isColor)
	{
		uint64_t sum= 0;

			memcpy(tempDst, dstBlock, dstStride*MIN(height-y, 5) );
			dstBlock= tempDst;
			srcBlock= tempSrc;
		}

		// finish 1 block before the next otherwise weŽll might have a problem
		// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
		for(x=0; x<width; x+=BLOCK_SIZE)
		{
			const int stride= dstStride;

			prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
*/
#endif

			if(!isColor) yHistogram[ srcBlock[srcStride*5] ]++;

			blockCopy(dstBlock + dstStride*5, dstStride,
				srcBlock + srcStride*5, srcStride, 8, mode & LEVEL_FIX);

			if(mode & LINEAR_IPOL_DEINT_FILTER)

				T0=T1;
#endif
			}

			/* check if we have a previous block to deblock it with dstBlock */
			if(x - 8 >= 0 && x<width)
			{
#ifdef MORE_TIMING
				T0= rdtsc();
#endif
				if(mode & H_DEBLOCK)

			}
			else if(y!=0)
				dering(dstBlock - stride*9 + width-9, stride, QP);
			//FIXME dering filter will not be applied to last block (bottom right)

			dstBlock+=8;
			srcBlock+=8;
		}

		/* did we use a tmp buffer */
		if(y+15 > height)
		{
			uint8_t *dstBlock= &(dst[y*dstStride]);
			memcpy(dstBlock, tempDst, dstStride*(height-y) );
		}
	}

2285:4840e356d0d3

static uint64_t temp2=0;
static uint64_t temp3=0;
static uint64_t temp4=0;
static uint64_t temp5=0;
static uint64_t pQPb=0;
static uint8_t tempBlock[16*16]; //used so the horizontal code gets aligned data

int hFlatnessThreshold= 56 - 16;
int vFlatnessThreshold= 56 - 16;

//amount of "black" u r willing to loose to get a brightness corrected picture
double maxClippedThreshold= 0.01;

int maxAllowedY=255;
//FIXME can never make a movieŽs black brighter (anyone needs that?)
int minAllowedY=16;

#ifdef TIMING
static inline long long rdtsc()
{
	long long l;

	/* Temporary buffers for handling the last row(s) */
	static uint8_t *tempDst= NULL;
	static uint8_t *tempSrc= NULL;

	/* Temporary buffers for handling the last block */
	static uint8_t *tempDstBlock= NULL;
	static uint8_t *tempSrcBlock= NULL;

	uint8_t *dstBlockPtrBackup;
	uint8_t *srcBlockPtrBackup;

#ifdef TIMING
	long long T0, T1, memcpyTime=0, vertTime=0, horizTime=0, sumTime, diffTime=0;
	sumTime= rdtsc();
#endif

	if(tempDst==NULL)
	{
		tempDst= (uint8_t*)memalign(8, 1024*24);
		tempSrc= (uint8_t*)memalign(8, 1024*24);
		tempDstBlock= (uint8_t*)memalign(8, 1024*24);
		tempSrcBlock= (uint8_t*)memalign(8, 1024*24);
	}

	if(!yHistogram)
	{
		int i;
		yHistogram= (uint64_t*)malloc(8*256);
		for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256;

		if(mode & FULL_Y_RANGE)
		{
			maxAllowedY=255;
			minAllowedY=0;
		}
	}

	if(!isColor)
	{
		uint64_t sum= 0;

			memcpy(tempDst, dstBlock, dstStride*MIN(height-y, 5) );
			dstBlock= tempDst;
			srcBlock= tempSrc;
		}

		// From this point on it is guranteed that we can read and write 16 lines downward
		// finish 1 block before the next otherwise weŽll might have a problem
		// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
		for(x=0; x<width; x+=BLOCK_SIZE)
		{
			const int stride= dstStride;

			prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
*/
#endif

			if(!isColor) yHistogram[ srcBlock[srcStride*5] ]++;

			//can we mess with a 8x16 block, if not use a temp buffer, yes again
			if(x+7 >= width)
			{
				int i;
				dstBlockPtrBackup= dstBlock;
				srcBlockPtrBackup= srcBlock;

				for(i=0;i<BLOCK_SIZE*2; i++)
				{
					memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x);
					memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x);
				}

				dstBlock= tempDstBlock;
				srcBlock= tempSrcBlock;
			}

			blockCopy(dstBlock + dstStride*5, dstStride,
				srcBlock + srcStride*5, srcStride, 8, mode & LEVEL_FIX);

			if(mode & LINEAR_IPOL_DEINT_FILTER)

				T0=T1;
#endif
			}

			/* check if we have a previous block to deblock it with dstBlock */
			if(x - 8 >= 0)
			{
#ifdef MORE_TIMING
				T0= rdtsc();
#endif
				if(mode & H_DEBLOCK)

			}
			else if(y!=0)
				dering(dstBlock - stride*9 + width-9, stride, QP);
			//FIXME dering filter will not be applied to last block (bottom right)

			/* did we use a tmp-block buffer */
			if(x+7 >= width)
			{
				int i;
				dstBlock= dstBlockPtrBackup;
				srcBlock= srcBlockPtrBackup;

				for(i=0;i<BLOCK_SIZE*2; i++)
				{
					memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x);
				}
			}

			dstBlock+=8;
			srcBlock+=8;
		}

		/* did we use a tmp buffer */
		if(y+15 >= height)
		{
			uint8_t *dstBlock= &(dst[y*dstStride]);
			memcpy(dstBlock, tempDst, dstStride*(height-y) );
		}
	}