xform

#define rotl(x,r) (x << r) | (x >> (64-r))
#define r(x,s,t,u,v,r1,r2,r3,r4)\
    x^=s;x=rotl(x,r1);x+=t;x=rotl(x,r2);x^=u;x=rotl(x,r3);x+=v;x=rotl(x,r4)

#define rotr(x,r) (x >> r) | (x << (64-r))
#define rinv(x,s,t,u,v,r1,r2,r3,r4)\
    x=rotr(x,r4);x-=v;x=rotr(x,r3);x^=u;x=rotr(x,r2);x-=t;x=rotr(x,r1);x^=s

void f(uint64_t x[8])
{
    uint64_t a, b, c, d, e, f, g, h;
    a = 0x243F6A8885A308D3 + x[0]; a=rotl(a, 3);
    b = 0x13198A2E03707344 + x[1]; b=rotl(b, 5);
    c = 0xA4093822299F31D0 + x[2]; c=rotl(c, 7);
    d = 0x082EFA98EC4E6C89 + x[3]; d=rotl(d,11);
    e = 0x452821E638D01377 + x[4]; e=rotl(e,13);
    f = 0xBE5466CF34E90C6C + x[5]; f=rotl(f,17);
    g = 0xC0AC29B7C97C50DD + x[6]; g=rotl(g,19);
    h = 0x3F84D5B5B5470917 + x[7]; h=rotl(h,23);

    r(e,a,b,c,d,41,21,17,29);
    r(f,b,c,d,a,33,49, 1,17);
    r(g,c,d,a,b,41,11,17, 3);
    r(h,d,a,b,c,17, 7,33,55);

    r(a,e,f,g,h, 9,27,17,51);
    r(b,f,g,h,e,25,61,49,53);
    r(c,g,h,e,f,49,23,33,39);
    r(d,h,e,f,g,57,51,49,27);

    e += 0x9216D5D98979FB1B; e=rotl(e,47);
    f += 0xD1310BA698DFB5AC; f=rotl(f,53);
    g += 0x2FFD72DBD01ADFB7; g=rotl(g,59);
    h += 0xB8E1AFED6A267E96; h=rotl(h,61);

    r(e,a,b,c,d,33,15, 1,47);
    r(f,b,c,d,a,57,13,49,37);
    r(g,c,d,a,b,25, 3,49,11);
    r(h,d,a,b,c, 9,37,17,13);

    a += 0xBA7C9045F12C7F99; a=rotl(a,29);
    b += 0x24A19947B3916CF7; b=rotl(b,37);
    c += 0x0801F2E2858EFC16; c=rotl(c,41);
    d += 0x636920D871574E69; d=rotl(d,43);

    r(a,e,f,g,h, 9, 5,17,45);
    r(b,f,g,h,e,25,29,49,21);
    r(c,g,h,e,f,17,57,33, 9);
    r(d,h,e,f,g,57,19,49,59);

    e += 0xA458FEA3F4933D7E; e=rotl(e, 3);
    f += 0x0D95748F728EB658; f=rotl(f, 5);
    g += 0x718BCD5882154AEE; g=rotl(g, 7);
    h += 0x7B54A41DC25A59B5; h=rotl(h,11);

    r(e,a,b,c,d, 9,27,17,51);
    r(f,b,c,d,a,25,61,49,53);
    r(g,c,d,a,b,49,23,33,39);
    r(h,d,a,b,c,57,51,49,28);

    a += 0x9C30D5392AF26013; a=rotl(a,13);
    b += 0xC5D1B023286085F0; b=rotl(b,17);
    c += 0xCA417918B8DB38EF; c=rotl(c,19);
    d += 0x8E79DCB0603A180E; d=rotl(d,23);

    r(a,e,f,g,h,41,21,17,29);
    r(b,f,g,h,e,33,49, 1,17);
    r(c,g,h,e,f,41,11,17, 3);
    r(d,h,e,f,g,17, 7,33,55);

    x[0] = a;
    x[1] = b;
    x[2] = c;
    x[3] = d;
    x[4] = e;
    x[5] = f;
    x[6] = g;
    x[7] = h;
}


void finv(uint64_t x[8])
{
    uint64_t a, b, c, d, e, f, g, h;

    a = x[0];
    b = x[1];
    c = x[2];
    d = x[3];
    e = x[4];
    f = x[5];
    g = x[6];
    h = x[7];

    rinv(a,e,f,g,h,41,21,17,29);
    rinv(b,f,g,h,e,33,49, 1,17);
    rinv(c,g,h,e,f,41,11,17, 3);
    rinv(d,h,e,f,g,17, 7,33,55);

    a=rotr(a,13);a -= 0x9C30D5392AF26013;
    b=rotr(b,17);b -= 0xC5D1B023286085F0;
    c=rotr(c,19);c -= 0xCA417918B8DB38EF;
    d=rotr(d,23);d -= 0x8E79DCB0603A180E;

    rinv(e,a,b,c,d, 9,27,17,51);
    rinv(f,b,c,d,a,25,61,49,53);
    rinv(g,c,d,a,b,49,23,33,39);
    rinv(h,d,a,b,c,57,51,49,28);

    e=rotr(e, 3); e -= 0xA458FEA3F4933D7E;
    f=rotr(f, 5); f -= 0x0D95748F728EB658;
    g=rotr(g, 7); g -= 0x718BCD5882154AEE;
    h=rotr(h,11); h -= 0x7B54A41DC25A59B5;

    rinv(a,e,f,g,h, 9, 5,17,45);
    rinv(b,f,g,h,e,25,29,49,21);
    rinv(c,g,h,e,f,17,57,33, 9);
    rinv(d,h,e,f,g,57,19,49,59);

    a=rotr(a,29); a -= 0xBA7C9045F12C7F99;
    b=rotr(b,37); b -= 0x24A19947B3916CF7;
    c=rotr(c,41); c -= 0x0801F2E2858EFC16;
    d=rotr(d,43); d -= 0x636920D871574E69;

    rinv(e,a,b,c,d,33,15, 1,47);
    rinv(f,b,c,d,a,57,13,49,37);
    rinv(g,c,d,a,b,25, 3,49,11);
    rinv(h,d,a,b,c, 9,37,17,13);

    e=rotr(e,47); e -= 0x9216D5D98979FB1B;
    f=rotr(f,53); f -= 0xD1310BA698DFB5AC;
    g=rotr(g,59); g -= 0x2FFD72DBD01ADFB7;
    h=rotr(h,61); h -= 0xB8E1AFED6A267E96;

    rinv(a,e,f,g,h, 9,27,17,51);
    rinv(b,f,g,h,e,25,61,49,53);
    rinv(c,g,h,e,f,49,23,33,39);
    rinv(d,h,e,f,g,57,51,49,27);

    rinv(e,a,b,c,d,41,21,17,29);
    rinv(f,b,c,d,a,33,49, 1,17);
    rinv(g,c,d,a,b,41,11,17, 3);
    rinv(h,d,a,b,c,17, 7,33,55);

    a=rotr(a, 3); x[0] = a - 0x243F6A8885A308D3;
    b=rotr(b, 5); x[1] = b - 0x13198A2E03707344;
    c=rotr(c, 7); x[2] = c - 0xA4093822299F31D0;
    d=rotr(d,11); x[3] = d - 0x082EFA98EC4E6C89;
    e=rotr(e,13); x[4] = e - 0x452821E638D01377;
    f=rotr(f,17); x[5] = f - 0xBE5466CF34E90C6C;
    g=rotr(g,19); x[6] = g - 0xC0AC29B7C97C50DD;
    h=rotr(h,23); x[7] = h - 0x3F84D5B5B5470917;
}