vaguely fast DXT1 encoder.

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

void BGBBTJ_BCn_EncodeBlockDXT1F(byte *block,
    byte *rgba, int xstride, int ystride, int nclr)
{
    byte clr0[4], clr1[4];
    byte pxy[16];
    int p0, p1, p2, p3;
    int mcr, mcg, mcb, mcy;
    int ncr, ncg, ncb, ncy;
    int acr, acg, acb;
    int cr, cg, cb, cy, acy;
    int i, j, k;

    mcr=255; mcg=255; mcb=255; mcy=255;
    ncr=  0; ncg=  0; ncb=  0; ncy=  0;
    acr=  0; acg=  0; acb=  0;

    for(i=0; i<4; i++)
        for(j=0; j<4; j++)
    {
        k=(i*ystride)+(j*xstride);
        cr=rgba[k+0];
        cg=rgba[k+1];
        cb=rgba[k+2];
        cy=(cr+2*cg+cb)>>2;
        pxy[i*4+j]=cy;

        if(cy<mcy) { mcr=cr; mcg=cg; mcb=cb; mcy=cy; }
        if(cy>ncy) { ncr=cr; ncg=cg; ncb=cb; ncy=cy; }
        acr+=cr;
        acg+=cg;
        acb+=cb;
    }

    acr=acr>>4;
    acg=acg>>4;
    acb=acb>>4;

    clr0[0]=(mcr*11+acr*5)>>4;
    clr0[1]=(mcg*11+acg*5)>>4;
    clr0[2]=(mcb*11+acb*5)>>4;
    clr1[0]=(ncr*11+acr*5)>>4;
    clr1[1]=(ncg*11+acg*5)>>4;
    clr1[2]=(ncb*11+acb*5)>>4;

    block[0]=((clr0[1]<<3)&0xE0) | (clr0[2]>>3);
    block[1]=(clr0[0]&0xF8) | (clr0[1]>>5);
    block[2]=((clr1[1]<<3)&0xE0) | (clr1[2]>>3);
    block[3]=(clr1[0]&0xF8) | (clr1[1]>>5);

    clr0[3]=(clr0[0]+2*clr0[1]+clr0[2])>>2;
    clr1[3]=(clr1[0]+2*clr1[1]+clr1[2])>>2;

    acy=(clr0[3]+clr1[3])>>1;

    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[ 4])>>16);
    p1=-((acy-pxy[ 5])>>16);
    p2=-((acy-pxy[ 6])>>16);
    p3=-((acy-pxy[ 7])>>16);
    block[5]=(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[6]=(p3<<6)|(p2<<4)|(p1<<2)|p0;
    p0=-((acy-pxy[12])>>16);
    p1=-((acy-pxy[13])>>16);
    p2=-((acy-pxy[14])>>16);
    p3=-((acy-pxy[15])>>16);
    block[7]=(p3<<6)|(p2<<4)|(p1<<2)|p0;
}

BGBBTJ_API void BGBBTJ_BCn_EncodeImageDXT1F(byte *block,
    byte *rgba, int xs, int ys, int stride)
{
    int xs1, ys1;
    int i, j;

    xs1=xs>>2; ys1=ys>>2;
    for(i=0; i<ys1; i++)
        for(j=0; j<xs1; j++)
    {
        BGBBTJ_BCn_EncodeBlockDXT1F(
            block+(i*xs1+j)*8,
            rgba+(i*4*xs+j*4)*stride,
            stride, xs*stride, 3);
    }
}