BJX2, Color Cell Encode (ConGfx/Doom/Quake)

/** Color Cell Encoder, Inner Function
  * src: source framebuffer, current frame
  * lsrc: scratch framebuffer, to detect/skip unchanging blocks.
  * rdsta: first part of destination
  * rdstb: second part of destination
  * Note: that rdsta and rdstb need not be adjacent.
  * Note: also that screen width is hard-coded in this version.
  */
void    VID_ConGfx_EncBlock16P(u16 *src, u16 *lsrc, u32 *rdsta, u32 *rdstb)
{
    u16 *cs, *ct;
    int cmin, cmax;
    register int cavg, calo, cahi;
    int clra, clrb;
    u32 px2;
    u32 dxa, dxb;
    u64 lpx0, lpx1, lpx2, lpx3, lpx4;
    register int k0, k1;
    int i, j, k;

    //NOTE: BGBCC Codegen Hint.
    //Hints register allocator to enable use of all 32 GPRs.
    __hint_use_egpr();

    //Note: BGBCC is kinda stupid, so branches result in spilling most of the registers...

    //Detect if pixels haven't changed
    ct=lsrc; cs=src; i=0;
    lpx0=*(u64 *)cs;    lpx4=*(u64 *)ct;
    cs+=BASEWIDTH;      ct+=BASEWIDTH;
    i|=(lpx0!=lpx4);
    lpx1=*(u64 *)cs;    lpx4=*(u64 *)ct;
    cs+=BASEWIDTH;      ct+=BASEWIDTH;
    i|=(lpx1!=lpx4);
    lpx2=*(u64 *)cs;    lpx4=*(u64 *)ct;
    cs+=BASEWIDTH;      ct+=BASEWIDTH;
    i|=(lpx2!=lpx4);
    lpx3=*(u64 *)cs;    lpx4=*(u64 *)ct;
    i|=(lpx3!=lpx4);

    if(!i)  return;     //Hasn't changed, do nothing

    //Save new pixels to scratch buffer
    ct=lsrc;
    *(u64 *)ct=lpx0;    ct+=BASEWIDTH;
    *(u64 *)ct=lpx1;    ct+=BASEWIDTH;
    *(u64 *)ct=lpx2;    ct+=BASEWIDTH;
    *(u64 *)ct=lpx3;

    calo=65536; cahi=-1;

    //NOTE: BGBCC: __int_min() and __int_max() are intrinsics wrapping the CSELT instruction.

    //Find the lightest and darkest pixels:
    k0=(u16)(lpx0    );     k1=(u16)(lpx0>>16);
    calo=__int_min(calo, k0);       cahi=__int_max(cahi, k0);
    k0=(u16)(lpx0>>32);
    calo=__int_min(calo, k1);       cahi=__int_max(cahi, k1);
    k1=(u16)(lpx0>>48);
    calo=__int_min(calo, k0);       cahi=__int_max(cahi, k0);
    calo=__int_min(calo, k1);       cahi=__int_max(cahi, k1);

    k0=(u16)(lpx1    );     k1=(u16)(lpx1>>16);
    calo=__int_min(calo, k0);       cahi=__int_max(cahi, k0);
    k0=(u16)(lpx1>>32);
    calo=__int_min(calo, k1);       cahi=__int_max(cahi, k1);
    k1=(u16)(lpx1>>48);
    calo=__int_min(calo, k0);       cahi=__int_max(cahi, k0);
    calo=__int_min(calo, k1);       cahi=__int_max(cahi, k1);

    k0=(u16)(lpx2    );     k1=(u16)(lpx2>>16);
    calo=__int_min(calo, k0);       cahi=__int_max(cahi, k0);
    k0=(u16)(lpx2>>32);
    calo=__int_min(calo, k1);       cahi=__int_max(cahi, k1);
    k1=(u16)(lpx2>>48);
    calo=__int_min(calo, k0);       cahi=__int_max(cahi, k0);
    calo=__int_min(calo, k1);       cahi=__int_max(cahi, k1);

    k0=(u16)(lpx3    );             k1=(u16)(lpx3>>16);
    calo=__int_min(calo, k0);       cahi=__int_max(cahi, k0);
    k0=(u16)(lpx3>>32);
    calo=__int_min(calo, k1);       cahi=__int_max(cahi, k1);
    k1=(u16)(lpx3>>48);
    calo=__int_min(calo, k0);       cahi=__int_max(cahi, k0);
    calo=__int_min(calo, k1);       cahi=__int_max(cahi, k1);

    //OK, we have range, calc average, lower, and upper threshold:
    cmin=calo;  cmax=cahi;
    cavg=(cmin+cmax)>>1;
    calo=(cmin+cavg)>>1;
    cahi=(cmax+cavg)>>1;

    //Find pixel bits; Note that MUL, DIV, and mem-access are expensive in BJX2
    //Especially don't want DIV, because DIV is implemented in software (hidden C runtime call; 100s of cycles).
    //But, what we do have, is 32 GPRs and a fairly cheap 'a>b' operator...

    px2=0;

    k0=(u16)(lpx0    );             k1=(u16)(lpx0>>16);
    px2=(px2<<2)|((k0>cavg)<<1);    px2|=(k0>cahi)|(calo>k0);
    k0=(u16)(lpx0>>32);
    px2=(px2<<2)|((k1>cavg)<<1);    px2|=(k1>cahi)|(calo>k1);
    k1=(u16)(lpx0>>48);
    px2=(px2<<2)|((k0>cavg)<<1);    px2|=(k0>cahi)|(calo>k0);
    px2=(px2<<2)|((k1>cavg)<<1);    px2|=(k1>cahi)|(calo>k1);

    k0=(u16)(lpx1    );             k1=(u16)(lpx1>>16);
    px2=(px2<<2)|((k0>cavg)<<1);    px2|=(k0>cahi)|(calo>k0);
    k0=(u16)(lpx1>>32);
    px2=(px2<<2)|((k1>cavg)<<1);    px2|=(k1>cahi)|(calo>k1);
    k1=(u16)(lpx1>>48);
    px2=(px2<<2)|((k0>cavg)<<1);    px2|=(k0>cahi)|(calo>k0);
    px2=(px2<<2)|((k1>cavg)<<1);    px2|=(k1>cahi)|(calo>k1);

    k0=(u16)(lpx2    );             k1=(u16)(lpx2>>16);
    px2=(px2<<2)|((k0>cavg)<<1);    px2|=(k0>cahi)|(calo>k0);
    k0=(u16)(lpx2>>32);
    px2=(px2<<2)|((k1>cavg)<<1);    px2|=(k1>cahi)|(calo>k1);
    k1=(u16)(lpx2>>48);
    px2=(px2<<2)|((k0>cavg)<<1);    px2|=(k0>cahi)|(calo>k0);
    px2=(px2<<2)|((k1>cavg)<<1);    px2|=(k1>cahi)|(calo>k1);

    k0=(u16)(lpx3    );             k1=(u16)(lpx3>>16);
    px2=(px2<<2)|((k0>cavg)<<1);    px2|=(k0>cahi)|(calo>k0);
    k0=(u16)(lpx3>>32);
    px2=(px2<<2)|((k1>cavg)<<1);    px2|=(k1>cahi)|(calo>k1);
    k1=(u16)(lpx3>>48);
    px2=(px2<<2)|((k0>cavg)<<1);    px2|=(k0>cahi)|(calo>k0);
    px2=(px2<<2)|((k1>cavg)<<1);    px2|=(k1>cahi)|(calo>k1);

    px2=px2^((~(px2>>1))&0x55555555);

    clrb=cmin>>1;                       clra=cmax>>1;
    dxa=0xC0000000|(clrb<<15)|clra;     dxb=px2;
    *rdsta=dxa;                         *rdstb=dxb;
}