Browse Source

Merge pull request #136 from Diapolo/master

force constants in diakgcn to be uint
nfactor-troky
Con Kolivas 13 years ago
parent
commit
66a19f8294
  1. 504
      diakgcn120222.cl

504
diakgcn120222.cl

@ -1,4 +1,4 @@ @@ -1,4 +1,4 @@
// DiaKGCN 22-02-2012 - OpenCL kernel by Diapolo
// DiaKGCN 23-02-2012 - OpenCL kernel by Diapolo
//
// Parts and / or ideas for this kernel are based upon the public-domain poclbm project, the phatk kernel by Phateus and the DiabloMiner kernel by DiabloD3.
// The kernel was rewritten by me (Diapolo) and is still public-domain!
@ -76,29 +76,29 @@ __kernel @@ -76,29 +76,29 @@ __kernel
V[4] += PreVal0addK7 + nonce + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = PreVal0addK7 + nonce + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0xd807aa98 + V[7] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0xd807aa98 + V[7] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0xd807aa98U + V[7] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0xd807aa98U + V[7] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0x12835b01 + V[6] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x12835b01 + V[6] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0x12835b01U + V[6] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x12835b01U + V[6] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[1] += 0x243185be + V[5] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x243185be + V[5] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0x243185beU + V[5] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x243185beU + V[5] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0x550c7dc3 + V[4] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x550c7dc3 + V[4] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0x550c7dc3U + V[4] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x550c7dc3U + V[4] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0x72be5d74 + V[3] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x72be5d74 + V[3] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0x72be5d74U + V[3] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x72be5d74U + V[3] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0x80deb1fe + V[2] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x80deb1fe + V[2] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0x80deb1feU + V[2] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x80deb1feU + V[2] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0x9bdc06a7 + V[1] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x9bdc06a7 + V[1] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0x9bdc06a7U + V[1] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x9bdc06a7U + V[1] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0xc19bf3f4 + V[0] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0xc19bf3f4 + V[0] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0xc19bf3f4U + V[0] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0xc19bf3f4U + V[0] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += W16addK16 + V[7] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = W16addK16 + V[7] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
@ -110,9 +110,9 @@ __kernel @@ -110,9 +110,9 @@ __kernel
W[0] = PreW18 + rotr25(nonce);
W[1] = PreW19 + nonce;
W[2] = 0x80000000 + rotr15(W[0]);
W[2] = 0x80000000U + rotr15(W[0]);
W[3] = rotr15(W[1]);
W[4] = 0x00000280 + rotr15(W[2]);
W[4] = 0x00000280U + rotr15(W[2]);
W[5] = W16 + rotr15(W[3]);
W[6] = W17 + rotr15(W[4]);
W[7] = W[0] + rotr15(W[5]);
@ -120,58 +120,58 @@ __kernel @@ -120,58 +120,58 @@ __kernel
W[9] = W[2] + rotr15(W[7]);
W[10] = W[3] + rotr15(W[8]);
W[11] = W[4] + rotr15(W[9]);
W[12] = W[5] + 0x00a00055 + rotr15(W[10]);
W[12] = W[5] + 0x00a00055U + rotr15(W[10]);
W[13] = W[6] + PreW31 + rotr15(W[11]);
W[14] = W[7] + PreW32 + rotr15(W[12]);
W[15] = W[8] + W17 + rotr15(W[13]) + rotr25(W[0]);
V[1] += 0x0fc19dc6 + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x0fc19dc6 + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0x0fc19dc6U + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x0fc19dc6U + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0x240ca1cc + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x240ca1cc + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0x240ca1ccU + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x240ca1ccU + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0x2de92c6f + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x2de92c6f + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0x2de92c6fU + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x2de92c6fU + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0x4a7484aa + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x4a7484aa + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0x4a7484aaU + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x4a7484aaU + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0x5cb0a9dc + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x5cb0a9dc + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0x5cb0a9dcU + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x5cb0a9dcU + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0x76f988da + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x76f988da + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0x76f988daU + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x76f988daU + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0x983e5152 + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x983e5152 + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0x983e5152U + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x983e5152U + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0xa831c66d + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0xa831c66d + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0xa831c66dU + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0xa831c66dU + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[1] += 0xb00327c8 + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0xb00327c8 + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0xb00327c8U + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0xb00327c8U + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0xbf597fc7 + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0xbf597fc7 + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0xbf597fc7U + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0xbf597fc7U + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0xc6e00bf3 + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0xc6e00bf3 + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0xc6e00bf3U + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0xc6e00bf3U + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0xd5a79147 + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0xd5a79147 + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0xd5a79147U + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0xd5a79147U + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0x06ca6351 + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x06ca6351 + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0x06ca6351U + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x06ca6351U + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0x14292967 + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x14292967 + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0x14292967U + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x14292967U + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0x27b70a85 + V[7] + W[14] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x27b70a85 + V[7] + W[14] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0x27b70a85U + V[7] + W[14] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x27b70a85U + V[7] + W[14] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0x2e1b2138 + V[6] + W[15] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x2e1b2138 + V[6] + W[15] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0x2e1b2138U + V[6] + W[15] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x2e1b2138U + V[6] + W[15] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
//----------------------------------------------------------------------------------
@ -192,53 +192,53 @@ __kernel @@ -192,53 +192,53 @@ __kernel
W[14] = W[14] + W[7] + rotr15(W[12]) + rotr25(W[15]);
W[15] = W[15] + W[8] + rotr15(W[13]) + rotr25( W[0]);
V[1] += 0x4d2c6dfc + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x4d2c6dfc + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0x4d2c6dfcU + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x4d2c6dfcU + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0x53380d13 + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x53380d13 + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0x53380d13U + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x53380d13U + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0x650a7354 + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x650a7354 + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0x650a7354U + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x650a7354U + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0x766a0abb + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x766a0abb + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0x766a0abbU + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x766a0abbU + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0x81c2c92e + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x81c2c92e + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0x81c2c92eU + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x81c2c92eU + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0x92722c85 + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x92722c85 + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0x92722c85U + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x92722c85U + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0xa2bfe8a1 + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0xa2bfe8a1 + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0xa2bfe8a1U + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0xa2bfe8a1U + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0xa81a664b + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0xa81a664b + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0xa81a664bU + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0xa81a664bU + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[1] += 0xc24b8b70 + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0xc24b8b70 + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0xc24b8b70U + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0xc24b8b70U + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0xc76c51a3 + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0xc76c51a3 + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0xc76c51a3U + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0xc76c51a3U + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0xd192e819 + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0xd192e819 + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0xd192e819U + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0xd192e819U + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0xd6990624 + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0xd6990624 + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0xd6990624U + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0xd6990624U + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0xf40e3585 + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0xf40e3585 + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0xf40e3585U + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0xf40e3585U + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0x106aa070 + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x106aa070 + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0x106aa070U + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x106aa070U + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0x19a4c116 + V[7] + W[14] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x19a4c116 + V[7] + W[14] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0x19a4c116U + V[7] + W[14] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x19a4c116U + V[7] + W[14] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0x1e376c08 + V[6] + W[15] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x1e376c08 + V[6] + W[15] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0x1e376c08U + V[6] + W[15] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x1e376c08U + V[6] + W[15] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
//----------------------------------------------------------------------------------
@ -257,92 +257,92 @@ __kernel @@ -257,92 +257,92 @@ __kernel
W[12] = W[12] + W[5] + rotr15(W[10]) + rotr25(W[13]);
W[13] = W[13] + W[6] + rotr15(W[11]) + rotr25(W[14]);
V[1] += 0x2748774c + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x2748774c + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0x2748774cU + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x2748774cU + V[5] + W[0] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0x34b0bcb5 + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x34b0bcb5 + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0x34b0bcb5U + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x34b0bcb5U + V[4] + W[1] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0x391c0cb3 + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x391c0cb3 + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0x391c0cb3U + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x391c0cb3U + V[3] + W[2] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0x4ed8aa4a + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x4ed8aa4a + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0x4ed8aa4aU + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x4ed8aa4aU + V[2] + W[3] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0x5b9cca4f + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x5b9cca4f + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0x5b9cca4fU + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x5b9cca4fU + V[1] + W[4] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0x682e6ff3 + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x682e6ff3 + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0x682e6ff3U + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x682e6ff3U + V[0] + W[5] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0x748f82ee + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x748f82ee + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0x748f82eeU + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x748f82eeU + V[7] + W[6] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0x78a5636f + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x78a5636f + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0x78a5636fU + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x78a5636fU + V[6] + W[7] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[1] += 0x84c87814 + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x84c87814 + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0x84c87814U + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x84c87814U + V[5] + W[8] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0x8cc70208 + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x8cc70208 + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0x8cc70208U + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x8cc70208U + V[4] + W[9] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0x90befffa + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x90befffa + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0x90befffaU + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x90befffaU + V[3] + W[10] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0xa4506ceb + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0xa4506ceb + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0xa4506cebU + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0xa4506cebU + V[2] + W[11] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0xbef9a3f7 + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0xbef9a3f7 + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0xbef9a3f7U + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0xbef9a3f7U + V[1] + W[12] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0xc67178f2 + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0xc67178f2 + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0xc67178f2U + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0xc67178f2U + V[0] + W[13] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
//----------------------------------------------------------------------------------
W[0] = state0 + V[0] + rotr25(state1 + V[1]);
W[1] = state1 + V[1] + 0x00a00000 + rotr25(state2 + V[2]);
W[1] = state1 + V[1] + 0x00a00000U + rotr25(state2 + V[2]);
W[2] = state2 + V[2] + rotr15(W[0]) + rotr25(state3 + V[3]);
W[3] = state3 + V[3] + rotr15(W[1]) + rotr25(state4 + V[4]);
W[4] = state4 + V[4] + rotr15(W[2]) + rotr25(state5 + V[5]);
W[5] = state5 + V[5] + rotr15(W[3]) + rotr25(state6 + V[6]);
W[6] = state6 + V[6] + 0x00000100 + rotr15(W[4]) + rotr25(state7 + V[7]);
W[7] = state7 + V[7] + W[0] + 0x11002000 + rotr15(W[5]);
W[8] = W[1] + 0x80000000 + rotr15(W[6]);
W[6] = state6 + V[6] + 0x00000100U + rotr15(W[4]) + rotr25(state7 + V[7]);
W[7] = state7 + V[7] + W[0] + 0x11002000U + rotr15(W[5]);
W[8] = W[1] + 0x80000000U + rotr15(W[6]);
W[9] = W[2] + rotr15(W[7]);
W[10] = W[3] + rotr15(W[8]);
W[11] = W[4] + rotr15(W[9]);
W[12] = W[5] + rotr15(W[10]);
W[13] = W[6] + rotr15(W[11]);
W[14] = W[7] + 0x00400022 + rotr15(W[12]);
W[15] = W[8] + 0x00000100 + rotr15(W[13]) + rotr25(W[0]);
W[14] = W[7] + 0x00400022U + rotr15(W[12]);
W[15] = W[8] + 0x00000100U + rotr15(W[13]) + rotr25(W[0]);
// 0x71374491 + 0x1f83d9ab + state1
// 0x71374491U + 0x1f83d9abU + state1
const u state1AaddV1 = state1A + V[1];
// 0xb5c0fbcf + 0x9b05688c + state2
// 0xb5c0fbcfU + 0x9b05688cU + state2
const u state2AaddV2 = state2A + V[2];
// 0x510e527f + 0xe9b5dba5 + state3
// 0x510e527fU + 0xe9b5dba5U + state3
const u state3AaddV3 = state3A + V[3];
// 0x3956c25b + state4
// 0x3956c25bU + state4
const u state4AaddV4 = state4A + V[4];
// 0x59f111f1 + state5
// 0x59f111f1U + state5
const u state5AaddV5 = state5A + V[5];
// 0x923f82a4 + state6
// 0x923f82a4U + state6
const u state6AaddV6 = state6A + V[6];
// 0xab1c5ed5 + state7
// 0xab1c5ed5U + state7
const u state7AaddV7 = state7A + V[7];
// 0x98c7e2a2 + state0
// 0x98c7e2a2U + state0
V[3] = state0A + V[0];
// 0xfc08884d + state0
// 0xfc08884dU + state0
V[7] = state0B + V[0];
V[0] = 0x6a09e667;
V[1] = 0xbb67ae85;
V[2] = 0x3c6ef372;
V[4] = 0x510e527f;
V[5] = 0x9b05688c;
V[6] = 0x1f83d9ab;
V[0] = 0x6a09e667U;
V[1] = 0xbb67ae85U;
V[2] = 0x3c6ef372U;
V[4] = 0x510e527fU;
V[5] = 0x9b05688cU;
V[6] = 0x1f83d9abU;
V[2] += state1AaddV1 + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = state1AaddV1 + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
@ -365,77 +365,77 @@ __kernel @@ -365,77 +365,77 @@ __kernel
V[4] += state7AaddV7 + V[0] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = state7AaddV7 + V[0] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0x5807aa98 + V[7] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x5807aa98 + V[7] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0x5807aa98U + V[7] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x5807aa98U + V[7] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0x12835b01 + V[6] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x12835b01 + V[6] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0x12835b01U + V[6] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x12835b01U + V[6] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[1] += 0x243185be + V[5] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x243185be + V[5] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0x243185beU + V[5] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x243185beU + V[5] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0x550c7dc3 + V[4] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x550c7dc3 + V[4] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0x550c7dc3U + V[4] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x550c7dc3U + V[4] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0x72be5d74 + V[3] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x72be5d74 + V[3] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0x72be5d74U + V[3] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x72be5d74U + V[3] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0x80deb1fe + V[2] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x80deb1fe + V[2] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0x80deb1feU + V[2] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x80deb1feU + V[2] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0x9bdc06a7 + V[1] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x9bdc06a7 + V[1] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0x9bdc06a7U + V[1] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x9bdc06a7U + V[1] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0xc19bf274 + V[0] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0xc19bf274 + V[0] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0xc19bf274U + V[0] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0xc19bf274U + V[0] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0xe49b69c1 + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0xe49b69c1 + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0xe49b69c1U + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0xe49b69c1U + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0xefbe4786 + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0xefbe4786 + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0xefbe4786U + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0xefbe4786U + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[1] += 0x0fc19dc6 + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x0fc19dc6 + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0x0fc19dc6U + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x0fc19dc6U + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0x240ca1cc + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x240ca1cc + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0x240ca1ccU + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x240ca1ccU + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0x2de92c6f + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x2de92c6f + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0x2de92c6fU + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x2de92c6fU + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0x4a7484aa + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x4a7484aa + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0x4a7484aaU + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x4a7484aaU + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0x5cb0a9dc + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x5cb0a9dc + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0x5cb0a9dcU + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x5cb0a9dcU + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0x76f988da + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x76f988da + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0x76f988daU + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x76f988daU + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0x983e5152 + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x983e5152 + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0x983e5152U + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x983e5152U + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0xa831c66d + V[6] + W[9] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0xa831c66d + V[6] + W[9] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0xa831c66dU + V[6] + W[9] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0xa831c66dU + V[6] + W[9] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[1] += 0xb00327c8 + V[5] + W[10] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0xb00327c8 + V[5] + W[10] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0xb00327c8U + V[5] + W[10] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0xb00327c8U + V[5] + W[10] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0xbf597fc7 + V[4] + W[11] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0xbf597fc7 + V[4] + W[11] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0xbf597fc7U + V[4] + W[11] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0xbf597fc7U + V[4] + W[11] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0xc6e00bf3 + V[3] + W[12] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0xc6e00bf3 + V[3] + W[12] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0xc6e00bf3U + V[3] + W[12] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0xc6e00bf3U + V[3] + W[12] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0xd5a79147 + V[2] + W[13] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0xd5a79147 + V[2] + W[13] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0xd5a79147U + V[2] + W[13] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0xd5a79147U + V[2] + W[13] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0x06ca6351 + V[1] + W[14] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x06ca6351 + V[1] + W[14] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0x06ca6351U + V[1] + W[14] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x06ca6351U + V[1] + W[14] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0x14292967 + V[0] + W[15] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x14292967 + V[0] + W[15] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0x14292967U + V[0] + W[15] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x14292967U + V[0] + W[15] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
//----------------------------------------------------------------------------------
@ -456,53 +456,53 @@ __kernel @@ -456,53 +456,53 @@ __kernel
W[14] = W[14] + W[7] + rotr15(W[12]) + rotr25(W[15]);
W[15] = W[15] + W[8] + rotr15(W[13]) + rotr25( W[0]);
V[3] += 0x27b70a85 + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x27b70a85 + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0x27b70a85U + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x27b70a85U + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0x2e1b2138 + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x2e1b2138 + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0x2e1b2138U + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x2e1b2138U + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[1] += 0x4d2c6dfc + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x4d2c6dfc + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0x4d2c6dfcU + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x4d2c6dfcU + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0x53380d13 + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x53380d13 + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0x53380d13U + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x53380d13U + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0x650a7354 + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x650a7354 + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0x650a7354U + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x650a7354U + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0x766a0abb + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x766a0abb + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0x766a0abbU + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x766a0abbU + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0x81c2c92e + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x81c2c92e + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0x81c2c92eU + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x81c2c92eU + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0x92722c85 + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x92722c85 + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0x92722c85U + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x92722c85U + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0xa2bfe8a1 + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0xa2bfe8a1 + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0xa2bfe8a1U + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0xa2bfe8a1U + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0xa81a664b + V[6] + W[9] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0xa81a664b + V[6] + W[9] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0xa81a664bU + V[6] + W[9] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0xa81a664bU + V[6] + W[9] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[1] += 0xc24b8b70 + V[5] + W[10] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0xc24b8b70 + V[5] + W[10] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0xc24b8b70U + V[5] + W[10] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0xc24b8b70U + V[5] + W[10] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0xc76c51a3 + V[4] + W[11] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0xc76c51a3 + V[4] + W[11] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0xc76c51a3U + V[4] + W[11] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0xc76c51a3U + V[4] + W[11] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0xd192e819 + V[3] + W[12] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0xd192e819 + V[3] + W[12] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0xd192e819U + V[3] + W[12] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0xd192e819U + V[3] + W[12] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0xd6990624 + V[2] + W[13] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0xd6990624 + V[2] + W[13] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0xd6990624U + V[2] + W[13] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0xd6990624U + V[2] + W[13] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0xf40e3585 + V[1] + W[14] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0xf40e3585 + V[1] + W[14] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0xf40e3585U + V[1] + W[14] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0xf40e3585U + V[1] + W[14] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0x106aa070 + V[0] + W[15] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x106aa070 + V[0] + W[15] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0x106aa070U + V[0] + W[15] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x106aa070U + V[0] + W[15] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
//----------------------------------------------------------------------------------
@ -520,38 +520,38 @@ __kernel @@ -520,38 +520,38 @@ __kernel
W[11] = W[11] + W[4] + rotr15( W[9]) + rotr25(W[12]);
W[12] = W[12] + W[5] + rotr15(W[10]) + rotr25(W[13]);
V[3] += 0x19a4c116 + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x19a4c116 + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0x19a4c116U + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x19a4c116U + V[7] + W[0] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0x1e376c08 + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x1e376c08 + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[2] += 0x1e376c08U + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[6] = 0x1e376c08U + V[6] + W[1] + ch(V[3], V[4], V[5]) + rotr26(V[3]) + rotr30(V[7]) + ma(V[0], V[1], V[7]);
V[1] += 0x2748774c + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x2748774c + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[1] += 0x2748774cU + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[5] = 0x2748774cU + V[5] + W[2] + ch(V[2], V[3], V[4]) + rotr26(V[2]) + rotr30(V[6]) + ma(V[7], V[0], V[6]);
V[0] += 0x34b0bcb5 + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x34b0bcb5 + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[0] += 0x34b0bcb5U + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[4] = 0x34b0bcb5U + V[4] + W[3] + ch(V[1], V[2], V[3]) + rotr26(V[1]) + rotr30(V[5]) + ma(V[6], V[7], V[5]);
V[7] += 0x391c0cb3 + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x391c0cb3 + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[7] += 0x391c0cb3U + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
V[3] = 0x391c0cb3U + V[3] + W[4] + ch(V[0], V[1], V[2]) + rotr26(V[0]) + rotr30(V[4]) + ma(V[5], V[6], V[4]);
V[6] += 0x4ed8aa4a + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x4ed8aa4a + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[6] += 0x4ed8aa4aU + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]);
V[2] = 0x4ed8aa4aU + V[2] + W[5] + ch(V[7], V[0], V[1]) + rotr26(V[7]) + rotr30(V[3]) + ma(V[4], V[5], V[3]);
V[5] += 0x5b9cca4f + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x5b9cca4f + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[5] += 0x5b9cca4fU + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]);
V[1] = 0x5b9cca4fU + V[1] + W[6] + ch(V[6], V[7], V[0]) + rotr26(V[6]) + rotr30(V[2]) + ma(V[3], V[4], V[2]);
V[4] += 0x682e6ff3 + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x682e6ff3 + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[4] += 0x682e6ff3U + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]);
V[0] = 0x682e6ff3U + V[0] + W[7] + ch(V[5], V[6], V[7]) + rotr26(V[5]) + rotr30(V[1]) + ma(V[2], V[3], V[1]);
V[3] += 0x748f82ee + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x748f82ee + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[3] += 0x748f82eeU + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]);
V[7] = 0x748f82eeU + V[7] + W[8] + ch(V[4], V[5], V[6]) + rotr26(V[4]) + rotr30(V[0]) + ma(V[1], V[2], V[0]);
V[2] += 0x78a5636f + V[6] + W[9] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[2] += 0x78a5636fU + V[6] + W[9] + ch(V[3], V[4], V[5]) + rotr26(V[3]);
V[1] += 0x84c87814 + V[5] + W[10] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[1] += 0x84c87814U + V[5] + W[10] + ch(V[2], V[3], V[4]) + rotr26(V[2]);
V[0] += 0x8cc70208 + V[4] + W[11] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[0] += 0x8cc70208U + V[4] + W[11] + ch(V[1], V[2], V[3]) + rotr26(V[1]);
V[7] += V[3] + W[12] + ch(V[0], V[1], V[2]) + rotr26(V[0]);
@ -560,53 +560,53 @@ __kernel @@ -560,53 +560,53 @@ __kernel
#define NFLAG (0x7F)
#ifdef VECTORS8
bool result = any(V[7] == 0x136032ed);
bool result = any(V[7] == 0x136032edU);
if (result) {
output[FOUND] = FOUND;
if (V[7].s0 == 0x136032ed)
if (V[7].s0 == 0x136032edU)
output[NFLAG & nonce.s0] = nonce.s0;
if (V[7].s1 == 0x136032ed)
if (V[7].s1 == 0x136032edU)
output[NFLAG & nonce.s1] = nonce.s1;
if (V[7].s2 == 0x136032ed)
if (V[7].s2 == 0x136032edU)
output[NFLAG & nonce.s2] = nonce.s2;
if (V[7].s3 == 0x136032ed)
if (V[7].s3 == 0x136032edU)
output[NFLAG & nonce.s3] = nonce.s3;
if (V[7].s4 == 0x136032ed)
if (V[7].s4 == 0x136032edU)
output[NFLAG & nonce.s4] = nonce.s4;
if (V[7].s5 == 0x136032ed)
if (V[7].s5 == 0x136032edU)
output[NFLAG & nonce.s5] = nonce.s5;
if (V[7].s6 == 0x136032ed)
if (V[7].s6 == 0x136032edU)
output[NFLAG & nonce.s6] = nonce.s6;
if (V[7].s7 == 0x136032ed)
if (V[7].s7 == 0x136032edU)
output[NFLAG & nonce.s7] = nonce.s7;
}
#elif defined VECTORS4
bool result = any(V[7] == 0x136032ed);
bool result = any(V[7] == 0x136032edU);
if (result) {
output[FOUND] = FOUND;
if (V[7].x == 0x136032ed)
if (V[7].x == 0x136032edU)
output[NFLAG & nonce.x] = nonce.x;
if (V[7].y == 0x136032ed)
if (V[7].y == 0x136032edU)
output[NFLAG & nonce.y] = nonce.y;
if (V[7].z == 0x136032ed)
if (V[7].z == 0x136032edU)
output[NFLAG & nonce.z] = nonce.z;
if (V[7].w == 0x136032ed)
if (V[7].w == 0x136032edU)
output[NFLAG & nonce.w] = nonce.w;
}
#elif defined VECTORS2
bool result = any(V[7] == 0x136032ed);
bool result = any(V[7] == 0x136032edU);
if (result) {
output[FOUND] = FOUND;
if (V[7].x == 0x136032ed)
if (V[7].x == 0x136032edU)
output[NFLAG & nonce.x] = nonce.x;
if (V[7].y == 0x136032ed)
if (V[7].y == 0x136032edU)
output[NFLAG & nonce.y] = nonce.y;
}
#else
if (V[7] == 0x136032ed)
if (V[7] == 0x136032edU)
output[FOUND] = output[NFLAG & nonce] = nonce;
#endif
}

Loading…
Cancel
Save