#define rotl(x,y) rotate(x,y)
#define Ch(x,y,z) bitselect(z,y,x)
#define Maj(x,y,z) Ch((x^z),y,z)

#define EndianSwap(n) (rotl(n & ES[0], 24U)|rotl(n & ES[1], 8U))

#define Tr2(x)      (rotl(x, 30U) ^ rotl(x, 19U) ^ rotl(x, 10U))
#define Tr1(x)      (rotl(x, 26U) ^ rotl(x, 21U) ^ rotl(x, 7U))
#define Wr2(x)      (rotl(x, 25U) ^ rotl(x, 14U) ^ (x>>3U))
#define Wr1(x)      (rotl(x, 15U) ^ rotl(x, 13U) ^ (x>>10U))


#define RND(a, b, c, d, e, f, g, h, k)  \
    h += Tr1(e);            \
    h += Ch(e, f, g);       \
    h += k;             \
    d += h;             \
    h += Tr2(a);            \
    h += Maj(a, b, c);

x = W[ 3].z
//x2 = W[ 0].y
// + W[ 2].y + Wr2(W[ 0].y
W[ 0].x += (rotl(x, 15U) ^ rotl(x, 13U) ^ (x>>10U)) + W[ 2].y + (rotl(W[ 0].y, 25U) ^ rotl(W[ 0].y, 14U) ^ (W[ 0].y>>3U))
h += (rotl(e, 26U) ^ rotl(e, 21U) ^ rotl(e, 7U))
h += bitselect(g,f,e)
h += W[0].x+ K[15];
d += h;
h += (rotl(a, 30U) ^ rotl(a, 19U) ^ rotl(a, 10U))
h += bitselect(c,b,(a^c))


(rotl(x, 15U) ^ rotl(x, 13U) ^ (x>>10U)) + (rotl(y, 25U) ^ rotl(y, 14U) ^ (y>>3U))


    W[ 0].x += Wr1(W[ 3].z) + W[ 2].y + Wr2(W[ 0].y);
    RND(A,B,C,D,E,F,G,H, W[0].x+ K[15]);
    W[ 0].y += Wr1(W[ 3].w) + W[ 2].z + Wr2(W[ 0].z);
    RND(H,A,B,C,D,E,F,G, W[0].y+ K[16]);
    W[ 0].z += Wr1(W[ 0].x) + W[ 2].w + Wr2(W[ 0].w);
    RND(G,H,A,B,C,D,E,F, W[0].z+ K[17]);
    W[ 0].w += Wr1(W[ 0].y) + W[ 3].x + Wr2(W[ 1].x);
    RND(F,G,H,A,B,C,D,E, W[0].w+ K[18]);

    W[ 1].x += Wr1(W[ 0].z) + W[ 3].y + Wr2(W[ 1].y);
    RND(E,F,G,H,A,B,C,D, W[1].x+ K[19]);
    W[ 1].y += Wr1(W[ 0].w) + W[ 3].z + Wr2(W[ 1].z);
    RND(D,E,F,G,H,A,B,C, W[1].y+ K[20]);
    W[ 1].z += Wr1(W[ 1].x) + W[ 3].w + Wr2(W[ 1].w);
    RND(C,D,E,F,G,H,A,B, W[1].z+ K[21]);
    W[ 1].w += Wr1(W[ 1].y) + W[ 0].x + Wr2(W[ 2].x);
    RND(B,C,D,E,F,G,H,A, W[1].w+ K[22]);