swscale.c

uclinux 下的vlc播放器源代码

				{
					double A= param ? -param*0.01 : -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(flags & SWS_X)
				{
					double A= param ? 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(flags & 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(flags & SWS_GAUSS)
				{
					double p= param ? param*0.1 : 3.0;
					coeff = pow(2.0, - p*d*d);
				}
				else if(flags & SWS_SINC)
				{
					coeff = d ? sin(d*PI)/(d*PI) : 1.0;
				}
				else if(flags & SWS_LANCZOS)
				{
					double p= param ? 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(flags & SWS_BILINEAR)
				{
					coeff= 1.0 - d;
					if(coeff<0) coeff=0;
				}
				else if(flags & SWS_SPLINE)
				{
					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
	   free(filter);
	*/
	ASSERT(filterSize>0)
	filter2Size= filterSize;
	if(srcFilter) filter2Size+= srcFilter->length - 1;
	if(dstFilter) filter2Size+= dstFilter->length - 1;
	ASSERT(filter2Size>0)
	filter2= (double*)memalign(8, 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);
	}
	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 (flags & 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;
        }

	ASSERT(minFilterSize > 0)
	filterSize= (minFilterSize +(filterAlign-1)) & (~(filterAlign-1));
	ASSERT(filterSize > 0)
	filter= (double*)memalign(8, filterSize*dstW*sizeof(double));
	*outFilterSize= filterSize;

	if(flags&SWS_PRINT_INFO)
		MSG_INFO("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];
		}
	}
	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= MAX(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= MIN(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
	*outFilter= (int16_t*)memalign(8, *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)];
	}

	free(filter);
}

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

	int xpos, i;

	// create an optimized horizontal scaling routine

	//code fragment

	asm volatile(
		"jmp 9f				\n\t"
	// Begin
		"0:				\n\t"
		"movq (%%edx, %%eax), %%mm3	\n\t" 
		"movd (%%ecx, %%esi), %%mm0	\n\t" 
		"movd 1(%%ecx, %%esi), %%mm1	\n\t"
		"punpcklbw %%mm7, %%mm1		\n\t"
		"punpcklbw %%mm7, %%mm0		\n\t"
		"pshufw $0xFF, %%mm1, %%mm1	\n\t"
		"1:				\n\t"
		"pshufw $0xFF, %%mm0, %%mm0	\n\t"
		"2:				\n\t"
		"psubw %%mm1, %%mm0		\n\t"
		"movl 8(%%ebx, %%eax), %%esi	\n\t"
		"pmullw %%mm3, %%mm0		\n\t"
		"psllw $7, %%mm1		\n\t"
		"paddw %%mm1, %%mm0		\n\t"

		"movq %%mm0, (%%edi, %%eax)	\n\t"

		"addl $8, %%eax			\n\t"
	// End
		"9:				\n\t"
//		"int $3\n\t"
		"leal 0b, %0			\n\t"
		"leal 1b, %1			\n\t"
		"leal 2b, %2			\n\t"
		"decl %1			\n\t"
		"decl %2			\n\t"
		"subl %0, %1			\n\t"
		"subl %0, %2			\n\t"
		"leal 9b, %3			\n\t"
		"subl %0, %3			\n\t"


		:"=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
		"=r" (fragmentLengthA)
	);

	asm volatile(
		"jmp 9f				\n\t"
	// Begin
		"0:				\n\t"
		"movq (%%edx, %%eax), %%mm3	\n\t" 
		"movd (%%ecx, %%esi), %%mm0	\n\t" 
		"punpcklbw %%mm7, %%mm0		\n\t"
		"pshufw $0xFF, %%mm0, %%mm1	\n\t"
		"1:				\n\t"
		"pshufw $0xFF, %%mm0, %%mm0	\n\t"
		"2:				\n\t"
		"psubw %%mm1, %%mm0		\n\t"
		"movl 8(%%ebx, %%eax), %%esi	\n\t"
		"pmullw %%mm3, %%mm0		\n\t"
		"psllw $7, %%mm1		\n\t"
		"paddw %%mm1, %%mm0		\n\t"

		"movq %%mm0, (%%edi, %%eax)	\n\t"

		"addl $8, %%eax			\n\t"
	// End
		"9:				\n\t"
//		"int $3\n\t"
		"leal 0b, %0			\n\t"
		"leal 1b, %1			\n\t"
		"leal 2b, %2			\n\t"
		"decl %1			\n\t"
		"decl %2			\n\t"
		"subl %0, %1			\n\t"
		"subl %0, %2			\n\t"
		"leal 9b, %3			\n\t"
		"subl %0, %3			\n\t"


		:"=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
		"=r" (fragmentLengthB)
	);

	xpos= 0; //lumXInc/2 - 0x8000; // difference between pixel centers
	fragmentPos=0;
	
	for(i=0; i<dstW/numSplits; i++)
	{
		int xx=xpos>>16;

		if((i&3) == 0)
		{
			int a=0;
			int b=((xpos+xInc)>>16) - xx;
			int c=((xpos+xInc*2)>>16) - xx;
			int d=((xpos+xInc*3)>>16) - xx;

			filter[i  ] = (( xpos         & 0xFFFF) ^ 0xFFFF)>>9;
			filter[i+1] = (((xpos+xInc  ) & 0xFFFF) ^ 0xFFFF)>>9;
			filter[i+2] = (((xpos+xInc*2) & 0xFFFF) ^ 0xFFFF)>>9;
			filter[i+3] = (((xpos+xInc*3) & 0xFFFF) ^ 0xFFFF)>>9;
			filterPos[i/2]= xx;

			if(d+1<4)
			{
				int maxShift= 3-(d+1);
				int shift=0;

				memcpy(funnyCode + fragmentPos, fragmentB, fragmentLengthB);

				funnyCode[fragmentPos + imm8OfPShufW1B]=
					(a+1) | ((b+1)<<2) | ((c+1)<<4) | ((d+1)<<6);
				funnyCode[fragmentPos + imm8OfPShufW2B]=
					a | (b<<2) | (c<<4) | (d<<6);

				if(i+3>=dstW) shift=maxShift; //avoid overread
				else if((filterPos[i/2]&3) <= maxShift) shift=filterPos[i/2]&3; //Align

				if(shift && i>=shift)
				{
					funnyCode[fragmentPos + imm8OfPShufW1B]+= 0x55*shift;
					funnyCode[fragmentPos + imm8OfPShufW2B]+= 0x55*shift;
					filterPos[i/2]-=shift;
				}

				fragmentPos+= fragmentLengthB;
			}
			else
			{
				int maxShift= 3-d;
				int shift=0;

				memcpy(funnyCode + fragmentPos, fragmentA, fragmentLengthA);

				funnyCode[fragmentPos + imm8OfPShufW1A]=
				funnyCode[fragmentPos + imm8OfPShufW2A]=
					a | (b<<2) | (c<<4) | (d<<6);

				if(i+4>=dstW) shift=maxShift; //avoid overread
				else if((filterPos[i/2]&3) <= maxShift) shift=filterPos[i/2]&3; //partial align

				if(shift && i>=shift)
				{
					funnyCode[fragmentPos + imm8OfPShufW1A]+= 0x55*shift;
					funnyCode[fragmentPos + imm8OfPShufW2A]+= 0x55*shift;
					filterPos[i/2]-=shift;
				}

				fragmentPos+= fragmentLengthA;
			}

			funnyCode[fragmentPos]= RET;
		}
		xpos+=xInc;
	}
	filterPos[i/2]= xpos>>16; // needed to jump to the next part
}
#endif // ARCH_X86

static void globalInit(){
    // generating tables:
    int i;
    for(i=0; i<768; i++){
	int c= MIN(MAX(i-256, 0), 255);
	clip_table[i]=c;
    }
}

static SwsFunc getSwsFunc(int flags){
    
#ifdef RUNTIME_CPUDETECT
#ifdef ARCH_X86
	// ordered per speed fasterst first
	if(flags & SWS_CPU_CAPS_MMX2)
		return swScale_MMX2;
	else if(flags & SWS_CPU_CAPS_3DNOW)