swscale.c

从FFMPEG转换而来的H264解码程序,VC下编译..

#endif

//MMX2 versions
#ifdef COMPILE_MMX2
#undef RENAME
#define HAVE_MMX
#define HAVE_MMX2
#undef HAVE_3DNOW
#define RENAME(a) a ## _MMX2
#include "swscale_template.c"
#endif

//3DNOW versions
#ifdef COMPILE_3DNOW
#undef RENAME
#define HAVE_MMX
#undef HAVE_MMX2
#define HAVE_3DNOW
#define RENAME(a) a ## _3DNow
#include "swscale_template.c"
#endif

#endif //ARCH_X86 || ARCH_X86_64

// minor note: the HAVE_xyz is messed up after that line so don't use it

static double getSplineCoeff(double a, double b, double c, double d, double dist)
{
//	printf("%f %f %f %f %f\n", a,b,c,d,dist);
	if(dist<=1.0) 	return ((d*dist + c)*dist + b)*dist +a;
	else		return getSplineCoeff(	0.0,
						 b+ 2.0*c + 3.0*d,
						        c + 3.0*d,
						-b- 3.0*c - 6.0*d,
						dist-1.0);
}

static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outFilterSize, int xInc,
			      int srcW, int dstW, int filterAlign, int one, SwsMethodParams *params,int cpuflags, int debugflags,
			      SwsVector *srcFilter, SwsVector *dstFilter)
{
	int i;
	int filterSize;
	int filter2Size;
	int minFilterSize;
	double *filter=NULL;
	double *filter2=NULL;
#if defined(ARCH_X86) || defined(ARCH_X86_64)
	if(cpuflags & SWS_CPU_CAPS_MMX)
		asm volatile("emms\n\t"::: "memory"); //FIXME this shouldnt be required but it IS (even for non mmx versions)
#endif

	// Note the +1 is for the MMXscaler which reads over the end
	*filterPos = av_malloc((dstW+1)*sizeof(int16_t));

	if(ABS(xInc - 0x10000) <10) // unscaled
	{
		int i;
		filterSize= 1;
		filter= av_malloc(dstW*sizeof(double)*filterSize);
		for(i=0; i<dstW*filterSize; i++) filter[i]=0;

		for(i=0; i<dstW; i++)
		{
			filter[i*filterSize]=1;
			(*filterPos)[i]=i;
		}

	}
	else if(params->method&SWS_POINT) // lame looking point sampling mode
	{
		int i;
		int xDstInSrc;
		filterSize= 1;
		filter= av_malloc(dstW*sizeof(double)*filterSize);

		xDstInSrc= xInc/2 - 0x8000;
		for(i=0; i<dstW; i++)
		{
			int xx= (xDstInSrc - ((filterSize-1)<<15) + (1<<15))>>16;

			(*filterPos)[i]= xx;
			filter[i]= 1.0;
			xDstInSrc+= xInc;
		}
	}
	else if((xInc <= (1<<16) && (params->method&SWS_AREA)) || (params->method&SWS_FAST_BILINEAR)) // bilinear upscale
	{
		int i;
		int xDstInSrc;
		if     (params->method&SWS_BICUBIC) filterSize= 4;
		else if(params->method&SWS_X      ) filterSize= 4;
		else			   filterSize= 2; // SWS_BILINEAR / SWS_AREA
		filter= av_malloc(dstW*sizeof(double)*filterSize);

		xDstInSrc= xInc/2 - 0x8000;
		for(i=0; i<dstW; i++)
		{
			int xx= (xDstInSrc - ((filterSize-1)<<15) + (1<<15))>>16;
			int j;

			(*filterPos)[i]= xx;
				//Bilinear upscale / linear interpolate / Area averaging
				for(j=0; j<filterSize; j++)
				{
					double d= ABS((xx<<16) - xDstInSrc)/(double)(1<<16);
					double coeff= 1.0 - d;
					if(coeff<0) coeff=0;
					filter[i*filterSize + j]= coeff;
					xx++;
				}
			xDstInSrc+= xInc;
		}
	}
	else
	{
		double xDstInSrc;
		double sizeFactor, filterSizeInSrc;
		const double xInc1= (double)xInc / (double)(1<<16);

		if     (params->method&SWS_BICUBIC)	sizeFactor= 4.0;
		else if(params->method&SWS_X)		sizeFactor= 8.0;
		else if(params->method&SWS_AREA)	sizeFactor= 1.0; //downscale only, for upscale it is bilinear
		else if(params->method&SWS_GAUSS)	sizeFactor= 8.0;   // infinite ;)
		else if(params->method&SWS_LANCZOS)	sizeFactor= params->param ? 2.0*params->param : 6.0;
		else if(params->method&SWS_SINC)	sizeFactor= 20.0; // infinite ;)
		else if(params->method&SWS_SPLINE)	sizeFactor= 20.0;  // infinite ;)
		else if(params->method&SWS_BILINEAR)	sizeFactor= 2.0;
		else {
			sizeFactor= 0.0; //GCC warning killer
			ASSERT(0)
		}

		if(xInc1 <= 1.0)	filterSizeInSrc= sizeFactor; // upscale
		else			filterSizeInSrc= sizeFactor*srcW / (double)dstW;

		filterSize= (int)ceil(1 + filterSizeInSrc); // will be reduced later if possible
		if(filterSize > srcW-2) filterSize=srcW-2;

		filter= av_malloc(dstW*sizeof(double)*filterSize);

		xDstInSrc= xInc1 / 2.0 - 0.5;
		for(i=0; i<dstW; i++)
		{
			int xx= (int)(xDstInSrc - (filterSize-1)*0.5 + 0.5);
			int j;
			(*filterPos)[i]= xx;
			for(j=0; j<filterSize; j++)
			{
				double d= ABS(xx - xDstInSrc)/filterSizeInSrc*sizeFactor;
				double coeff;
				if(params->method & SWS_BICUBIC)
				{
					double A= params->param ? -params->param*0.01 : -0.08; //old default was -0.60

                                        // Equation is from VirtualDub
					if(d<1.0)
						coeff = (1.0 - (A+3.0)*d*d + (A+2.0)*d*d*d);
					else if(d<2.0)
						coeff = (-4.0*A + 8.0*A*d - 5.0*A*d*d + A*d*d*d);
					else
						coeff=0.0;
				}
/*				else if(flags & SWS_X)
				{
					double p= param ? param*0.01 : 0.3;
					coeff = d ? sin(d*PI)/(d*PI) : 1.0;
					coeff*= pow(2.0, - p*d*d);
				}*/
				else if(params->method & SWS_X)
				{
					double A= params->param ? params->param*0.1 : 1.0;

					if(d<1.0)
						coeff = cos(d*PI);
					else
						coeff=-1.0;
					if(coeff<0.0) 	coeff= -pow(-coeff, A);
					else		coeff=  pow( coeff, A);
					coeff= coeff*0.5 + 0.5;
				}
				else if(params->method & SWS_AREA)
				{
					double srcPixelSize= 1.0/xInc1;
					if(d + srcPixelSize/2 < 0.5) coeff= 1.0;
					else if(d - srcPixelSize/2 < 0.5) coeff= (0.5-d)/srcPixelSize + 0.5;
					else coeff=0.0;
				}
				else if(params->method & SWS_GAUSS)
				{
					double p= params->param ? params->param*0.1 : 3.0;
					coeff = pow(2.0, - p*d*d);
				}
				else if(params->method & SWS_SINC)
				{
					coeff = d ? sin(d*PI)/(d*PI) : 1.0;
				}
				else if(params->method & SWS_LANCZOS)
				{
					double p= params->param ? params->param : 3.0;
					coeff = d ? sin(d*PI)*sin(d*PI/p)/(d*d*PI*PI/p) : 1.0;
					if(d>p) coeff=0;
				}
				else if(params->method & SWS_BILINEAR)
				{
					coeff= 1.0 - d;
					if(coeff<0) coeff=0;
				}
				else if(params->method & SWS_SPLINE)
				{
					static const double p=-2.196152422706632;
					coeff = getSplineCoeff(1.0, 0.0, p, -p-1.0, d);
				}
				else {
					coeff= 0.0; //GCC warning killer
					ASSERT(0)
				}

				filter[i*filterSize + j]= coeff;
				xx++;
			}
			xDstInSrc+= xInc1;
		}
	}

	/* apply src & dst Filter to filter -> filter2
	   av_free(filter);
	*/
	ASSERT(filterSize>0)
	filter2Size= filterSize;
	if(srcFilter) filter2Size+= srcFilter->length - 1;
	if(dstFilter) filter2Size+= dstFilter->length - 1;
	ASSERT(filter2Size>0)
	filter2= av_malloc(filter2Size*dstW*sizeof(double));

	for(i=0; i<dstW; i++)
	{
		int j;
		SwsVector scaleFilter;
		SwsVector *outVec;

		scaleFilter.coeff= filter + i*filterSize;
		scaleFilter.length= filterSize;

		if(srcFilter) outVec= sws_getConvVec(srcFilter, &scaleFilter);
		else	      outVec= &scaleFilter;

		ASSERT(outVec->length == filter2Size)
		//FIXME dstFilter

		for(j=0; j<outVec->length; j++)
		{
			filter2[i*filter2Size + j]= outVec->coeff[j];
		}

		(*filterPos)[i]+= (filterSize-1)/2 - (filter2Size-1)/2;

		if(outVec != &scaleFilter) sws_freeVec(outVec);
	}
	av_free(filter); filter=NULL;

	/* try to reduce the filter-size (step1 find size and shift left) */
	// Assume its near normalized (*0.5 or *2.0 is ok but * 0.001 is not)
	minFilterSize= 0;
	for(i=dstW-1; i>=0; i--)
	{
		int min= filter2Size;
		int j;
		double cutOff=0.0;

		/* get rid off near zero elements on the left by shifting left */
		for(j=0; j<filter2Size; j++)
		{
			int k;
			cutOff += ABS(filter2[i*filter2Size]);

			if(cutOff > SWS_MAX_REDUCE_CUTOFF) break;

			/* preserve Monotonicity because the core can't handle the filter otherwise */
			if(i<dstW-1 && (*filterPos)[i] >= (*filterPos)[i+1]) break;

			// Move filter coeffs left
			for(k=1; k<filter2Size; k++)
				filter2[i*filter2Size + k - 1]= filter2[i*filter2Size + k];
			filter2[i*filter2Size + k - 1]= 0.0;
			(*filterPos)[i]++;
		}

		cutOff=0.0;
		/* count near zeros on the right */
		for(j=filter2Size-1; j>0; j--)
		{
			cutOff += ABS(filter2[i*filter2Size + j]);

			if(cutOff > SWS_MAX_REDUCE_CUTOFF) break;
			min--;
		}

		if(min>minFilterSize) minFilterSize= min;
	}

        if (cpuflags & SWS_CPU_CAPS_ALTIVEC) {
          // we can handle the special case 4,
          // so we don't want to go to the full 8
          if (minFilterSize < 5)
            filterAlign = 4;

          // we really don't want to waste our time
          // doing useless computation, so fall-back on
          // the scalar C code for very small filter.
          // vectorizing is worth it only if you have
          // decent-sized vector.
          if (minFilterSize < 3)
            filterAlign = 1;
        }

        if (cpuflags & SWS_CPU_CAPS_MMX) {
                // special case for unscaled vertical filtering
                if(minFilterSize == 1 && filterAlign == 2)
                        filterAlign= 1;
        }

	ASSERT(minFilterSize > 0)
	filterSize= (minFilterSize +(filterAlign-1)) & (~(filterAlign-1));
	ASSERT(filterSize > 0)
	filter= av_malloc(filterSize*dstW*sizeof(double));
        if(filterSize >= MAX_FILTER_SIZE)
                return -1;
	*outFilterSize= filterSize;

	if(debugflags&SWS_PRINT_INFO)
		MSG_V("SwScaler: reducing / aligning filtersize %d -> %d\n", filter2Size, filterSize);
	/* try to reduce the filter-size (step2 reduce it) */
        for(i=0; i<dstW; i++)
        {
                int j;

		for(j=0; j<filterSize; j++)
		{
			if(j>=filter2Size) filter[i*filterSize + j]= 0.0;
			else		   filter[i*filterSize + j]= filter2[i*filter2Size + j];
		}
	}
	av_free(filter2); filter2=NULL;


	//FIXME try to align filterpos if possible

	//fix borders
	for(i=0; i<dstW; i++)
	{
		int j;
		if((*filterPos)[i] < 0)
		{
			// Move filter coeffs left to compensate for filterPos
			for(j=1; j<filterSize; j++)
			{
				int left= FFMAX(j + (*filterPos)[i], 0);
				filter[i*filterSize + left] += filter[i*filterSize + j];
				filter[i*filterSize + j]=0;
			}
			(*filterPos)[i]= 0;
		}

		if((*filterPos)[i] + filterSize > srcW)
		{
			int shift= (*filterPos)[i] + filterSize - srcW;
			// Move filter coeffs right to compensate for filterPos
			for(j=filterSize-2; j>=0; j--)
			{
				int right= FFMIN(j + shift, filterSize-1);
				filter[i*filterSize +right] += filter[i*filterSize +j];
				filter[i*filterSize +j]=0;
			}
			(*filterPos)[i]= srcW - filterSize;
		}
	}

	// Note the +1 is for the MMXscaler which reads over the end
	/* align at 16 for AltiVec (needed by hScale_altivec_real) */
	*outFilter= av_malloc(*outFilterSize*(dstW+1)*sizeof(int16_t));
	memset(*outFilter, 0, *outFilterSize*(dstW+1)*sizeof(int16_t));

	/* Normalize & Store in outFilter */
	for(i=0; i<dstW; i++)
	{
		int j;
		double error=0;
		double sum=0;
		double scale= one;

		for(j=0; j<filterSize; j++)
		{
			sum+= filter[i*filterSize + j];
		}
		scale/= sum;
		for(j=0; j<*outFilterSize; j++)
		{
			double v= filter[i*filterSize + j]*scale + error;
			int intV= floor(v + 0.5);
			(*outFilter)[i*(*outFilterSize) + j]= intV;
			error = v - intV;
		}
	}

	(*filterPos)[dstW]= (*filterPos)[dstW-1]; // the MMX scaler will read over the end
	for(i=0; i<*outFilterSize; i++)
	{
		int j= dstW*(*outFilterSize);
		(*outFilter)[j + i]= (*outFilter)[j + i - (*outFilterSize)];
	}

	av_free(filter);
        return 0;
}

#if defined(ARCH_X86) || defined(ARCH_X86_64)
static void initMMX2HScaler(int dstW, int xInc, uint8_t *funnyCode, int16_t *filter, int32_t *filterPos, int numSplits)
{
	uint8_t *fragmentA;
	long imm8OfPShufW1A;
	long imm8OfPShufW2A;
	long fragmentLengthA;
	uint8_t *fragmentB;
	long imm8OfPShufW1B;
	long imm8OfPShufW2B;
	long fragmentLengthB;
	int fragmentPos;

	int xpos, i;

	// create an optimized horizontal scaling routine

	//code fragment

	asm volatile(
		"jmp 9f				\n\t"
	// Begin