comparison postproc/swscale_template.c @ 3215:3083616ba4d6

sliced scaleing bugfix

3214:6d3c84c8b2f6

	}
#endif
}


static inline void RENAME(hyscale)(uint16_t *dst, int dstWidth, uint8_t *src, int srcWidth, int xInc)
{
      // *** horizontal scale Y line to temp buffer
#ifdef ARCH_X86
#ifdef HAVE_MMX2
	int i;

			:: "m" (src), "m" (dst), "m" (dstWidth), "m" ((xInc*4)>>16),
			"m" ((xInc*4)&0xFFFF), "m" (xInc&0xFFFF)
			: "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi"
		);
		for(i=dstWidth-1; (i*xInc)>>16 >=srcWidth-1; i--) dst[i] = src[srcWidth-1]*128;
	}
	else
	{
#endif
	//NO MMX just normal asm ...

	}
#endif
}

inline static void RENAME(hcscale)(uint16_t *dst, int dstWidth,
				uint8_t *src1, uint8_t *src2, int srcWidth, int xInc)
{
#ifdef ARCH_X86
#ifdef HAVE_MMX2
	int i;
	if(canMMX2BeUsed)

		:: "m" (src1), "m" (dst), "m" (dstWidth), "m" ((xInc*4)>>16),
		  "m" ((xInc*4)&0xFFFF), "m" (xInc&0xFFFF), "m" (src2)
		: "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi"
	);
		for(i=dstWidth-1; (i*xInc)>>16 >=srcWidth/2-1; i--)
		{
			dst[i] = src1[srcWidth/2-1]*128;
			dst[i+2048] = src2[srcWidth/2-1]*128;
		}
	}
	else
	{
#endif

unsigned int s_xinc2;
//FIXME do we need th +-2 stuff?
unsigned int s_xinc= (srcW << 16) / dstW - 2;
unsigned int s_yinc= (srcH << 16) / dstH + 2;

static int lumDstYInSrc; // points to the dst Pixels center in the source (0 is the center of pixel 0,0 in src)
static int dstY;

// last horzontally interpolated lines, used to avoid unnecessary calculations
static int lastLumSrcY;

// used to detect a horizontal size change
static int old_dstW= -1;
static int old_s_xinc= -1;
#endif

int srcWidth;
int dstUVw;
int i;

if(((dstW + 7)&(~7)) >= dststride) dstW&= ~7;

srcWidth= (dstW*s_xinc + 0x8000)>>16;
dstUVw= fullUVIpol ? dstW : dstW/2;

#ifdef HAVE_MMX2
canMMX2BeUsed= (s_xinc <= 0x10000 && (dstW&31)==0 && (srcWidth&15)==0) ? 1 : 0;
#endif

// match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src to pixel n-2 of dst
// n-2 is the last chrominance sample available
// FIXME this is not perfect, but noone shuld notice the difference, the more correct variant
// would be like the vertical one, but that would require some special code for the
// first and last pixel
if(canMMX2BeUsed) 	s_xinc+= 20;
else			s_xinc = ((srcWidth-2)<<16)/(dstW-2) - 20;

if(fullUVIpol && !(dstbpp==12)) 	s_xinc2= s_xinc>>1;
else					s_xinc2= s_xinc;
  // force calculation of the horizontal interpolation of the first line

  if(srcSliceY ==0){
//	printf("dstW %d, srcw %d, mmx2 %d\n", dstW, srcWidth, canMMX2BeUsed);
	lastLumSrcY=-99;
	lastChrSrcY=-99;
	lumDstYInSrc= s_yinc/2 - 0x8000;
	dstY=0;

    uint16_t *buf0=pix_buf_y[ lumSrcY   &1];	// top line of the interpolated slice
    uint16_t *buf1=pix_buf_y[(lumSrcY+1)&1];	// bottom line of the interpolated slice
    uint16_t *uvbuf0=pix_buf_uv[ chrSrcY   &1];	// top line of the interpolated slice
    uint16_t *uvbuf1=pix_buf_uv[(chrSrcY+1)&1];	// bottom line of the interpolated slice

//    if(lumSrcY>=srcSliceY + srcSliceH) break; // wrong, skips last lines, but they are dupliactes anyway
    if(dstY >= dstH) break;

//	printf("lumSrcY:%d, dstY:%d, yalpha:%d\n", lumSrcY, dstY, yalpha*100/0x1000);

    if((dstY&1) && dstbpp==12) uvalpha=-1;

    dstY++; lumDstYInSrc+=s_yinc;

    //only interpolate the src line horizontally if we didnt do it allready
	if(lastLumSrcY!=lumSrcY)
	{
		unsigned char *src;
		// skip if first line has been horiz scaled alleady
		if(lastLumSrcY != lumSrcY-1)
		{
			// check if first line is before any available src lines
			if(lumSrcY-1 < srcSliceY ) src=srcptr[0]+(0                   )*stride[0];
			else			   src=srcptr[0]+(lumSrcY-srcSliceY -1)*stride[0];

			RENAME(hyscale)(buf0, dstW, src, srcWidth, s_xinc);
		}
		// check if second line is after any available src lines
		if(lumSrcY-srcSliceY  >= srcSliceH) src=srcptr[0]+(srcSliceH-1       )*stride[0];
		else				    src=srcptr[0]+(lumSrcY-srcSliceY )*stride[0];

		// the min() is required to avoid reuseing lines which where not available
		lastLumSrcY= MIN(lumSrcY, srcSliceY +srcSliceH-1);
		RENAME(hyscale)(buf1, dstW, src, srcWidth, s_xinc);
	}
//	printf("%d %d %d %d\n", y, chrSrcY, lastChrSrcY, h);
      // *** horizontal scale U and V lines to temp buffer
	if(lastChrSrcY!=chrSrcY)
	{

				src2= srcptr[2]+(0)*stride[2];
			}else{
				src1= srcptr[1]+(chrSrcY-srcSliceY /2-1)*stride[1];
				src2= srcptr[2]+(chrSrcY-srcSliceY /2-1)*stride[2];
			}
			RENAME(hcscale)(uvbuf0, dstUVw, src1, src2, srcWidth, s_xinc2);
		}

		// check if second line is after any available src lines
		if(chrSrcY - srcSliceY /2 >= srcSliceH/2)
		{

			src2= srcptr[2]+(srcSliceH/2-1)*stride[2];
		}else{
			src1= srcptr[1]+(chrSrcY-srcSliceY /2)*stride[1];
			src2= srcptr[2]+(chrSrcY-srcSliceY /2)*stride[2];
		}
		RENAME(hcscale)(uvbuf1, dstUVw, src1, src2, srcWidth, s_xinc2);

		// the min() is required to avoid reuseing lines which where not available
		lastChrSrcY= MIN(chrSrcY, srcSliceY /2+srcSliceH/2-1);
	}
#ifdef HAVE_MMX

3215:3083616ba4d6

	}
#endif
}


static inline void RENAME(hyscale)(uint16_t *dst, int dstWidth, uint8_t *src, int srcW, int xInc)
{
      // *** horizontal scale Y line to temp buffer
#ifdef ARCH_X86
#ifdef HAVE_MMX2
	int i;

			:: "m" (src), "m" (dst), "m" (dstWidth), "m" ((xInc*4)>>16),
			"m" ((xInc*4)&0xFFFF), "m" (xInc&0xFFFF)
			: "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi"
		);
		for(i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--) dst[i] = src[srcW-1]*128;
	}
	else
	{
#endif
	//NO MMX just normal asm ...

	}
#endif
}

inline static void RENAME(hcscale)(uint16_t *dst, int dstWidth,
				uint8_t *src1, uint8_t *src2, int srcW, int xInc)
{
#ifdef ARCH_X86
#ifdef HAVE_MMX2
	int i;
	if(canMMX2BeUsed)

		:: "m" (src1), "m" (dst), "m" (dstWidth), "m" ((xInc*4)>>16),
		  "m" ((xInc*4)&0xFFFF), "m" (xInc&0xFFFF), "m" (src2)
		: "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi"
	);
		for(i=dstWidth-1; (i*xInc)>>16 >=srcW/2-1; i--)
		{
			dst[i] = src1[srcW/2-1]*128;
			dst[i+2048] = src2[srcW/2-1]*128;
		}
	}
	else
	{
#endif

unsigned int s_xinc2;
//FIXME do we need th +-2 stuff?
unsigned int s_xinc= (srcW << 16) / dstW - 2;
unsigned int s_yinc= (srcH << 16) / dstH + 2;
			  
static int lumDstYInSrc; // points to the dst Pixels center in the source (0 is the center of pixel 0,0 in src)
static int dstY;

// last horzontally interpolated lines, used to avoid unnecessary calculations
static int lastLumSrcY;

// used to detect a horizontal size change
static int old_dstW= -1;
static int old_s_xinc= -1;
#endif

int dstUVw;
int i;

if(((dstW + 7)&(~7)) >= dststride) dstW&= ~7;

dstUVw= fullUVIpol ? dstW : dstW/2;

//printf("%d %d %d %d\n", srcW, srcH, dstW, dstH);
//printf("%d %d %d %d\n", s_xinc, s_yinc, srcSliceY, srcSliceH);

#ifdef HAVE_MMX2
canMMX2BeUsed= (s_xinc <= 0x10000 && (dstW&31)==0 && (srcW&15)==0) ? 1 : 0;
#endif

// match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src to pixel n-2 of dst
// n-2 is the last chrominance sample available
// FIXME this is not perfect, but noone shuld notice the difference, the more correct variant
// would be like the vertical one, but that would require some special code for the
// first and last pixel
if(canMMX2BeUsed) 	s_xinc+= 20;
else			s_xinc = ((srcW-2)<<16)/(dstW-2) - 20;

if(fullUVIpol && !(dstbpp==12)) 	s_xinc2= s_xinc>>1;
else					s_xinc2= s_xinc;
  // force calculation of the horizontal interpolation of the first line

  if(srcSliceY ==0){
//	printf("dstW %d, srcw %d, mmx2 %d\n", dstW, srcW, canMMX2BeUsed);
	lastLumSrcY=-99;
	lastChrSrcY=-99;
	lumDstYInSrc= s_yinc/2 - 0x8000;
	dstY=0;

    uint16_t *buf0=pix_buf_y[ lumSrcY   &1];	// top line of the interpolated slice
    uint16_t *buf1=pix_buf_y[(lumSrcY+1)&1];	// bottom line of the interpolated slice
    uint16_t *uvbuf0=pix_buf_uv[ chrSrcY   &1];	// top line of the interpolated slice
    uint16_t *uvbuf1=pix_buf_uv[(chrSrcY+1)&1];	// bottom line of the interpolated slice

    if(lumSrcY>=srcSliceY + srcSliceH && srcSliceY + srcSliceH < srcH) break;
    if(dstY >= dstH) break;
    
//	printf("lumSrcY:%d, dstY:%d, yalpha:%d\n", lumSrcY, dstY, yalpha*100/0x1000);

    if((dstY&1) && dstbpp==12) uvalpha=-1;

    dstY++; lumDstYInSrc+=s_yinc;

    //only interpolate the src line horizontally if we didnt do it allready
	if(lastLumSrcY!=lumSrcY)
	{
		unsigned char *src;
		
		// skip if first line has been horiz scaled alleady
		if(lastLumSrcY != lumSrcY-1)
		{
			// check if first line is before any available src lines
			if(lumSrcY-1 < srcSliceY ) src=srcptr[0]+(0                   )*stride[0];
			else			   src=srcptr[0]+(lumSrcY-srcSliceY -1)*stride[0];

			RENAME(hyscale)(buf0, dstW, src, srcW, s_xinc);
		}
		// check if second line is after any available src lines
		if(lumSrcY-srcSliceY  >= srcSliceH) src=srcptr[0]+(srcSliceH-1       )*stride[0];
		else				    src=srcptr[0]+(lumSrcY-srcSliceY )*stride[0];

		// the min() is required to avoid reuseing lines which where not available
		lastLumSrcY= MIN(lumSrcY, srcSliceY +srcSliceH-1);
		RENAME(hyscale)(buf1, dstW, src, srcW, s_xinc);
	}
//	printf("%d %d %d %d\n", y, chrSrcY, lastChrSrcY, h);
      // *** horizontal scale U and V lines to temp buffer
	if(lastChrSrcY!=chrSrcY)
	{

				src2= srcptr[2]+(0)*stride[2];
			}else{
				src1= srcptr[1]+(chrSrcY-srcSliceY /2-1)*stride[1];
				src2= srcptr[2]+(chrSrcY-srcSliceY /2-1)*stride[2];
			}
			RENAME(hcscale)(uvbuf0, dstUVw, src1, src2, srcW, s_xinc2);
		}

		// check if second line is after any available src lines
		if(chrSrcY - srcSliceY /2 >= srcSliceH/2)
		{

			src2= srcptr[2]+(srcSliceH/2-1)*stride[2];
		}else{
			src1= srcptr[1]+(chrSrcY-srcSliceY /2)*stride[1];
			src2= srcptr[2]+(chrSrcY-srcSliceY /2)*stride[2];
		}
		RENAME(hcscale)(uvbuf1, dstUVw, src1, src2, srcW, s_xinc2);

		// the min() is required to avoid reuseing lines which where not available
		lastChrSrcY= MIN(chrSrcY, srcSliceY /2+srcSliceH/2-1);
	}
#ifdef HAVE_MMX