/*
Copyright (C) 2013 by Brendan G Bohannon
Email: cr88192@gmail.com
Copying: http://pastebin.com/iJxtZHm6
 */

void BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(byte *block,
	short *pxy, short *pxu, short *pxv)
{
	int p0, p1, p2, p3;
	int mcy, ncy, acy;
	int acr, acg, acb;
	int cy, cu, cv;
	int cra, cga, cba;
	int crb, cgb, cbb;
	int u1, v1, l0, l1, l2;
	int i, j, k;

	cu=(pxu[0]+pxu[1]+pxu[8]+pxu[9])>>2;
	cv=(pxv[0]+pxv[1]+pxv[8]+pxv[9])>>2;

	mcy=pxy[0];	ncy=mcy;
	i=pxy[ 1];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[ 2];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[ 3];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[ 8];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[ 9];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[10];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[11];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[16];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[17];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[18];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[19];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[24];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[25];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[26];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[27];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);

	acy=(mcy+ncy)>>1;

	u1=cu-128; v1=cv-128;
	l0=91881*v1;
	l1=-22554*u1-46802*v1;
	l2=116130*u1;

	//EDIT: fixed math
	cra=(65536*mcy+l0)>>16;	cga=(65536*mcy+l1)>>16;	cba=(65536*mcy+l2)>>16;
	crb=(65536*ncy+l0)>>16;	cgb=(65536*ncy+l1)>>16;	cbb=(65536*ncy+l2)>>16;

	if((cra|cga|cba|crb|cgb|cbb)&(~255))
	{
		cra=(cra<0)?0:((cra>255)?255:cra);
		cga=(cga<0)?0:((cga>255)?255:cga);
		cba=(cba<0)?0:((cba>255)?255:cba);
		crb=(crb<0)?0:((crb>255)?255:crb);
		cgb=(cgb<0)?0:((cgb>255)?255:cgb);
		cbb=(cbb<0)?0:((cbb>255)?255:cbb);
	}

	i=((cra<<8)&0xF800) | ((cga<<3)&0x07E0) | (cba>>3);
	j=((crb<<8)&0xF800) | ((cgb<<3)&0x07E0) | (cbb>>3);

	block[0]=i;		block[1]=i>>8;
	block[2]=j;		block[3]=j>>8;

	p0=-((acy-pxy[ 0])>>16);
	p1=-((acy-pxy[ 1])>>16);
	p2=-((acy-pxy[ 2])>>16);
	p3=-((acy-pxy[ 3])>>16);
	block[4]=(p3<<6)|(p2<<4)|(p1<<2)|p0;
	p0=-((acy-pxy[ 8])>>16);
	p1=-((acy-pxy[ 9])>>16);
	p2=-((acy-pxy[10])>>16);
	p3=-((acy-pxy[11])>>16);
	block[5]=(p3<<6)|(p2<<4)|(p1<<2)|p0;
	p0=-((acy-pxy[16])>>16);
	p1=-((acy-pxy[17])>>16);
	p2=-((acy-pxy[18])>>16);
	p3=-((acy-pxy[19])>>16);
	block[6]=(p3<<6)|(p2<<4)|(p1<<2)|p0;
	p0=-((acy-pxy[24])>>16);
	p1=-((acy-pxy[25])>>16);
	p2=-((acy-pxy[26])>>16);
	p3=-((acy-pxy[27])>>16);
	block[7]=(p3<<6)|(p2<<4)|(p1<<2)|p0;
}


/*
Alternate Version
Flipped vertically, uses DXT5 interpolation (via index table).
true DXT1 is color1/color2/average/transparent
	{ 0,0,0,0, 0,0,2,2, 2,2,1,1, 1,1,1,1 };
The use of interpolation is mostly for slightly higher image quality.
*/

void BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(byte *block,
	short *pxy, short *pxu, short *pxv)
{
	static const char idxtab[16]=
		{ 0,0,0,0, 0,0,2,2, 3,3,1,1, 1,1,1,1 };

	int p0, p1, p2, p3;
	int mcy, ncy, acy;
	int acr, acg, acb;
	int cy, cu, cv;
	int cra, cga, cba;
	int crb, cgb, cbb;
	int u1, v1, l0, l1, l2;
	int i, j, k;

	cu=(pxu[0]+pxu[1]+pxu[8]+pxu[9])>>2;
	cv=(pxv[0]+pxv[1]+pxv[8]+pxv[9])>>2;

	mcy=pxy[0];	ncy=mcy;
	i=pxy[ 1];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[ 2];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[ 3];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[ 8];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[ 9];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[10];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[11];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[16];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[17];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[18];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[19];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[24];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[25];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[26];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);
	i=pxy[27];	j=(i-mcy)>>24; k=(ncy-i)>>24;
	mcy=(j&i)|((~j)&mcy); ncy=(k&i)|((~k)&ncy);

	acy=(mcy+ncy)>>1;

	u1=cu-128; v1=cv-128;
	l0=359*v1;
	l1=-88*u1-183*v1;
	l2=454*u1;

	cra=(256*mcy+l0)>>8;	cga=(256*mcy+l1)>>8;	cba=(256*mcy+l2)>>8;
	crb=(256*ncy+l0)>>8;	cgb=(256*ncy+l1)>>8;	cbb=(256*ncy+l2)>>8;
	if((cra|cga|cba|crb|cgb|cbb)&(~255))
	{
		cra=(cra<0)?0:((cra>255)?255:cra);
		cga=(cga<0)?0:((cga>255)?255:cga);
		cba=(cba<0)?0:((cba>255)?255:cba);
		crb=(crb<0)?0:((crb>255)?255:crb);
		cgb=(cgb<0)?0:((cgb>255)?255:cgb);
		cbb=(cbb<0)?0:((cbb>255)?255:cbb);
	}

	i=((cra<<8)&0xF800) | ((cga<<3)&0x07E0) | (cba>>3);
	j=((crb<<8)&0xF800) | ((cgb<<3)&0x07E0) | (cbb>>3);
	block[0]=i;		block[1]=i>>8;
	block[2]=j;		block[3]=j>>8;

	if(i==j)
	{
		block[4]=0;	block[5]=0;
		block[6]=0;	block[7]=0;
		return;
	}

//	l0=(256+64)/(ncy-acy+1);
	l0=32768/(ncy-acy+1);		//Quality FIX

	p0=idxtab[(((pxy[24]-acy)*l0)>>13)+8];
	p1=idxtab[(((pxy[25]-acy)*l0)>>13)+8];
	p2=idxtab[(((pxy[26]-acy)*l0)>>13)+8];
	p3=idxtab[(((pxy[27]-acy)*l0)>>13)+8];
	block[4]=(p3<<6)|(p2<<4)|(p1<<2)|p0;
	p0=idxtab[(((pxy[16]-acy)*l0)>>13)+8];
	p1=idxtab[(((pxy[17]-acy)*l0)>>13)+8];
	p2=idxtab[(((pxy[18]-acy)*l0)>>13)+8];
	p3=idxtab[(((pxy[19]-acy)*l0)>>13)+8];
	block[5]=(p3<<6)|(p2<<4)|(p1<<2)|p0;
	p0=idxtab[(((pxy[ 8]-acy)*l0)>>13)+8];
	p1=idxtab[(((pxy[ 9]-acy)*l0)>>13)+8];
	p2=idxtab[(((pxy[10]-acy)*l0)>>13)+8];
	p3=idxtab[(((pxy[11]-acy)*l0)>>13)+8];
	block[6]=(p3<<6)|(p2<<4)|(p1<<2)|p0;
	p0=idxtab[(((pxy[ 0]-acy)*l0)>>13)+8];
	p1=idxtab[(((pxy[ 1]-acy)*l0)>>13)+8];
	p2=idxtab[(((pxy[ 2]-acy)*l0)>>13)+8];
	p3=idxtab[(((pxy[ 3]-acy)*l0)>>13)+8];
	block[7]=(p3<<6)|(p2<<4)|(p1<<2)|p0;
}


//Logic for YUV -> DXT

int BGBBTJ_JPG_DecodeColorTransform_YCbCr420_DXT1(
	BGBBTJ_JPG_Context *ctx, byte *obuf, int xs, int ys, int stride)
{
	short *yblk, *ublk, *vblk;
	byte *oblk;
	int xs1, ys1, xs2, ys2, xs3, ys3, xs4, ys4;
	int i, j, k, i1;

	xs1=xs>>1; ys1=ys>>1;
	xs2=xs>>2; ys2=ys>>2;
	xs3=xs>>3; ys3=ys>>3;
	xs4=xs>>4; ys4=ys>>4;

	for(i=0; i<ys4; i++)
		for(j=0; j<xs4; j++)
	{
		k=i*xs4+j;
		ublk=ctx->jpg_sibuf[1]+k*64;
		vblk=ctx->jpg_sibuf[2]+k*64;

#if 1	//Vertical Flip
		i1=ys4-i-1;

		k=((i*2+0)*xs3)+(j*2+0);
		yblk=ctx->jpg_sibuf[0]+k*64;

		k=((i1*4+3)*xs2)+(j*4+0); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+0, ublk+0, vblk+0);
		k=((i1*4+3)*xs2)+(j*4+1); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+4, ublk+2, vblk+2);
		k=((i1*4+2)*xs2)+(j*4+0); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+32, ublk+16, vblk+16);
		k=((i1*4+2)*xs2)+(j*4+1); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+36, ublk+18, vblk+18);

		k=((i*2+0)*xs3)+(j*2+1);
		yblk=ctx->jpg_sibuf[0]+k*64;
	
		k=((i1*4+3)*xs2)+(j*4+2); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+0, ublk+4, vblk+4);
		k=((i1*4+3)*xs2)+(j*4+3); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+4, ublk+6, vblk+6);
		k=((i1*4+2)*xs2)+(j*4+2); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+32, ublk+20, vblk+20);
		k=((i1*4+2)*xs2)+(j*4+3); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+36, ublk+22, vblk+22);

		k=((i*2+1)*xs3)+(j*2+0);
		yblk=ctx->jpg_sibuf[0]+k*64;
	
		k=((i1*4+1)*xs2)+(j*4+0); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+0, ublk+32, vblk+32);
		k=((i1*4+1)*xs2)+(j*4+1); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+4, ublk+34, vblk+34);
		k=((i1*4+0)*xs2)+(j*4+0); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+32, ublk+48, vblk+48);
		k=((i1*4+0)*xs2)+(j*4+1); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+36, ublk+50, vblk+50);

		k=((i*2+1)*xs3)+(j*2+1);
		yblk=ctx->jpg_sibuf[0]+k*64;
	
		k=((i1*4+1)*xs2)+(j*4+2); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+0, ublk+36, vblk+36);
		k=((i1*4+1)*xs2)+(j*4+3); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+4, ublk+38, vblk+38);
		k=((i1*4+0)*xs2)+(j*4+2); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+32, ublk+52, vblk+52);
		k=((i1*4+0)*xs2)+(j*4+3); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+36, ublk+54, vblk+54);
#endif

#if 0	//No Vertical Flip
		k=((i*2+0)*xs3)+(j*2+0);
		yblk=ctx->jpg_sibuf[0]+k*64;

		k=((i*4+0)*xs2)+(j*4+0); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+0, ublk+0, vblk+0);
		k=((i*4+0)*xs2)+(j*4+1); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+4, ublk+2, vblk+2);
		k=((i*4+1)*xs2)+(j*4+0); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+32, ublk+16, vblk+16);
		k=((i*4+1)*xs2)+(j*4+1); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+36, ublk+18, vblk+18);

		k=((i*2+0)*xs3)+(j*2+1);
		yblk=ctx->jpg_sibuf[0]+k*64;
	
		k=((i*4+0)*xs2)+(j*4+2); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+0, ublk+4, vblk+4);
		k=((i*4+0)*xs2)+(j*4+3); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+4, ublk+6, vblk+6);
		k=((i*4+1)*xs2)+(j*4+2); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+32, ublk+20, vblk+20);
		k=((i*4+1)*xs2)+(j*4+3); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+36, ublk+22, vblk+22);

		k=((i*2+1)*xs3)+(j*2+0);
		yblk=ctx->jpg_sibuf[0]+k*64;
	
		k=((i*4+2)*xs2)+(j*4+0); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+0, ublk+32, vblk+32);
		k=((i*4+2)*xs2)+(j*4+1); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+4, ublk+34, vblk+34);
		k=((i*4+3)*xs2)+(j*4+0); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+32, ublk+48, vblk+48);
		k=((i*4+3)*xs2)+(j*4+1); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+36, ublk+50, vblk+50);

		k=((i*2+1)*xs3)+(j*2+1);
		yblk=ctx->jpg_sibuf[0]+k*64;
	
		k=((i*4+2)*xs2)+(j*4+2); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+0, ublk+36, vblk+36);
		k=((i*4+2)*xs2)+(j*4+3); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+4, ublk+38, vblk+38);
		k=((i*4+3)*xs2)+(j*4+2); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+32, ublk+52, vblk+52);
		k=((i*4+3)*xs2)+(j*4+3); oblk=obuf+k*stride;
		BGBBTJ_BCn_EncodeBlockDXT1F_YUV420(oblk, yblk+36, ublk+54, vblk+54);
#endif

	}
}