poclbm_noamd.cl 29 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322
  1. // This file is taken and modified from the public-domain poclbm project, and
  2. // we have therefore decided to keep it public-domain in Phoenix.
  3. #define VECTORS
  4. #ifdef VECTORS
  5. typedef uint4 u;
  6. #else
  7. typedef uint u;
  8. #endif
  9. __constant uint K[64] = {
  10. 0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5, 0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
  11. 0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3, 0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
  12. 0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc, 0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
  13. 0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7, 0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967,
  14. 0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13, 0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85,
  15. 0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3, 0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070,
  16. 0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5, 0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3,
  17. 0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208, 0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2
  18. };
  19. #define rotr(x, y) rotate((u)x, (u)(32-y))
  20. #define Ch(x, y, z) (z ^ (x & (y ^ z)))
  21. #define Ma(x, y, z) ((x & z) | (y & (x | z)))
  22. #define Ma2(x, y, z) ((y & z) | (x & (y | z)))
  23. __kernel void search( const uint state0, const uint state1, const uint state2, const uint state3,
  24. const uint state4, const uint state5, const uint state6, const uint state7,
  25. const uint B1, const uint C1, const uint D1,
  26. const uint F1, const uint G1, const uint H1,
  27. const uint base,
  28. const uint fW0, const uint fW1, const uint fW2, const uint fW3, const uint fW15, const uint fW01r, const uint fcty_e, const uint fcty_e2,
  29. __global uint * output)
  30. {
  31. u W0, W1, W2, W3, W4, W5, W6, W7, W8, W9, W10, W11, W12, W13, W14, W15;
  32. u A,B,C,D,E,F,G,H;
  33. u nonce;
  34. uint it;
  35. #ifdef VECTORS
  36. nonce = ((base >> 2) + (get_global_id(0))<<2) + (uint4)(0, 1, 2, 3);
  37. #else
  38. nonce = base + get_global_id(0);
  39. #endif
  40. W3 = nonce + fW3;
  41. E = fcty_e + nonce; A = state0 + E; E = E + fcty_e2;
  42. D = D1 + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B1, C1) + K[ 4] + 0x80000000; H = H1 + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma2(G1, E, F1);
  43. C = C1 + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B1) + K[ 5]; G = G1 + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma2(F1, D, E);
  44. B = B1 + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[ 6]; F = F1 + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  45. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[ 7]; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  46. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[ 8]; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  47. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[ 9]; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  48. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[10]; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  49. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[11]; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  50. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[12]; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  51. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[13]; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  52. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[14]; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  53. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[15] + 0x00000280U; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  54. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[16] + fW0; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  55. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[17] + fW1; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  56. W2 = (rotr(nonce, 7) ^ rotr(nonce, 18) ^ (nonce >> 3U)) + fW2;
  57. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[18] + W2; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  58. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[19] + W3; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  59. W4 = (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U)) + 0x80000000;
  60. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[20] + W4; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  61. W5 = (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  62. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[21] + W5; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  63. W6 = (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U)) + 0x00000280U;
  64. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[22] + W6; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  65. W7 = (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U)) + fW0;
  66. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[23] + W7; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  67. W8 = (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U)) + fW1;
  68. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[24] + W8; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  69. W9 = W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  70. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[25] + W9; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  71. W10 = W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  72. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[26] + W10; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  73. W11 = W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  74. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[27] + W11; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  75. W12 = W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  76. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[28] + W12; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  77. W13 = W6 + (rotr(W11, 17) ^ rotr(W11, 19) ^ (W11 >> 10U));
  78. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[29] + W13; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  79. W14 = 0x00a00055U + W7 + (rotr(W12, 17) ^ rotr(W12, 19) ^ (W12 >> 10U));
  80. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[30] + W14; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  81. W15 = fW15 + W8 + (rotr(W13, 17) ^ rotr(W13, 19) ^ (W13 >> 10U));
  82. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[31] + W15; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  83. W0 = fW01r + W9 + (rotr(W14, 17) ^ rotr(W14, 19) ^ (W14 >> 10U));
  84. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[32] + W0; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  85. W1 = fW1 + (rotr(W2, 7) ^ rotr(W2, 18) ^ (W2 >> 3U)) + W10 + (rotr(W15, 17) ^ rotr(W15, 19) ^ (W15 >> 10U));
  86. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[33] + W1; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  87. W2 = W2 + (rotr(W3, 7) ^ rotr(W3, 18) ^ (W3 >> 3U)) + W11 + (rotr(W0, 17) ^ rotr(W0, 19) ^ (W0 >> 10U));
  88. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[34] + W2; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  89. W3 = W3 + (rotr(W4, 7) ^ rotr(W4, 18) ^ (W4 >> 3U)) + W12 + (rotr(W1, 17) ^ rotr(W1, 19) ^ (W1 >> 10U));
  90. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[35] + W3; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  91. W4 = W4 + (rotr(W5, 7) ^ rotr(W5, 18) ^ (W5 >> 3U)) + W13 + (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U));
  92. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[36] + W4; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  93. W5 = W5 + (rotr(W6, 7) ^ rotr(W6, 18) ^ (W6 >> 3U)) + W14 + (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  94. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[37] + W5; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  95. W6 = W6 + (rotr(W7, 7) ^ rotr(W7, 18) ^ (W7 >> 3U)) + W15 + (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U));
  96. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[38] + W6; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  97. W7 = W7 + (rotr(W8, 7) ^ rotr(W8, 18) ^ (W8 >> 3U)) + W0 + (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U));
  98. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[39] + W7; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  99. W8 = W8 + (rotr(W9, 7) ^ rotr(W9, 18) ^ (W9 >> 3U)) + W1 + (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U));
  100. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[40] + W8; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  101. W9 = W9 + (rotr(W10, 7) ^ rotr(W10, 18) ^ (W10 >> 3U)) + W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  102. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[41] + W9; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  103. W10 = W10 + (rotr(W11, 7) ^ rotr(W11, 18) ^ (W11 >> 3U)) + W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  104. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[42] + W10; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  105. W11 = W11 + (rotr(W12, 7) ^ rotr(W12, 18) ^ (W12 >> 3U)) + W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  106. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[43] + W11; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  107. W12 = W12 + (rotr(W13, 7) ^ rotr(W13, 18) ^ (W13 >> 3U)) + W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  108. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[44] + W12; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  109. W13 = W13 + (rotr(W14, 7) ^ rotr(W14, 18) ^ (W14 >> 3U)) + W6 + (rotr(W11, 17) ^ rotr(W11, 19) ^ (W11 >> 10U));
  110. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[45] + W13; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  111. W14 = W14 + (rotr(W15, 7) ^ rotr(W15, 18) ^ (W15 >> 3U)) + W7 + (rotr(W12, 17) ^ rotr(W12, 19) ^ (W12 >> 10U));
  112. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[46] + W14; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  113. W15 = W15 + (rotr(W0, 7) ^ rotr(W0, 18) ^ (W0 >> 3U)) + W8 + (rotr(W13, 17) ^ rotr(W13, 19) ^ (W13 >> 10U));
  114. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[47] + W15; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  115. W0 = W0 + (rotr(W1, 7) ^ rotr(W1, 18) ^ (W1 >> 3U)) + W9 + (rotr(W14, 17) ^ rotr(W14, 19) ^ (W14 >> 10U));
  116. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[48] + W0; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  117. W1 = W1 + (rotr(W2, 7) ^ rotr(W2, 18) ^ (W2 >> 3U)) + W10 + (rotr(W15, 17) ^ rotr(W15, 19) ^ (W15 >> 10U));
  118. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[49] + W1; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  119. W2 = W2 + (rotr(W3, 7) ^ rotr(W3, 18) ^ (W3 >> 3U)) + W11 + (rotr(W0, 17) ^ rotr(W0, 19) ^ (W0 >> 10U));
  120. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[50] + W2; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  121. W3 = W3 + (rotr(W4, 7) ^ rotr(W4, 18) ^ (W4 >> 3U)) + W12 + (rotr(W1, 17) ^ rotr(W1, 19) ^ (W1 >> 10U));
  122. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[51] + W3; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  123. W4 = W4 + (rotr(W5, 7) ^ rotr(W5, 18) ^ (W5 >> 3U)) + W13 + (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U));
  124. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[52] + W4; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  125. W5 = W5 + (rotr(W6, 7) ^ rotr(W6, 18) ^ (W6 >> 3U)) + W14 + (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  126. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[53] + W5; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  127. W6 = W6 + (rotr(W7, 7) ^ rotr(W7, 18) ^ (W7 >> 3U)) + W15 + (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U));
  128. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[54] + W6; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  129. W7 = W7 + (rotr(W8, 7) ^ rotr(W8, 18) ^ (W8 >> 3U)) + W0 + (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U));
  130. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[55] + W7; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  131. W8 = W8 + (rotr(W9, 7) ^ rotr(W9, 18) ^ (W9 >> 3U)) + W1 + (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U));
  132. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[56] + W8; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  133. W9 = W9 + (rotr(W10, 7) ^ rotr(W10, 18) ^ (W10 >> 3U)) + W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  134. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[57] + W9; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  135. W10 = W10 + (rotr(W11, 7) ^ rotr(W11, 18) ^ (W11 >> 3U)) + W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  136. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[58] + W10; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  137. W11 = W11 + (rotr(W12, 7) ^ rotr(W12, 18) ^ (W12 >> 3U)) + W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  138. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[59] + W11; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  139. W12 = W12 + (rotr(W13, 7) ^ rotr(W13, 18) ^ (W13 >> 3U)) + W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  140. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[60] + W12; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  141. W13 = W13 + (rotr(W14, 7) ^ rotr(W14, 18) ^ (W14 >> 3U)) + W6 + (rotr(W11, 17) ^ rotr(W11, 19) ^ (W11 >> 10U));
  142. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[61] + W13; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  143. W14 = W14 + (rotr(W15, 7) ^ rotr(W15, 18) ^ (W15 >> 3U)) + W7 + (rotr(W12, 17) ^ rotr(W12, 19) ^ (W12 >> 10U));
  144. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[62] + W14; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  145. W15 = W15 + (rotr(W0, 7) ^ rotr(W0, 18) ^ (W0 >> 3U)) + W8 + (rotr(W13, 17) ^ rotr(W13, 19) ^ (W13 >> 10U));
  146. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[63] + W15; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  147. W0 = A + state0; W1 = B + state1;
  148. W2 = C + state2; W3 = D + state3;
  149. W4 = E + state4; W5 = F + state5;
  150. W6 = G + state6; W7 = H + state7;
  151. H = 0xb0edbdd0 + K[ 0] + W0; D = 0xa54ff53a + H; H = H + 0x08909ae5U;
  152. G = 0x1f83d9abU + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + (0x9b05688cU ^ (D & 0xca0b3af3U)) + K[ 1] + W1; C = 0x3c6ef372U + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma2(0xbb67ae85U, H, 0x6a09e667U);
  153. F = 0x9b05688cU + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, 0x510e527fU) + K[ 2] + W2; B = 0xbb67ae85U + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma2(0x6a09e667U, G, H);
  154. E = 0x510e527fU + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[ 3] + W3; A = 0x6a09e667U + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  155. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[ 4] + W4; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  156. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[ 5] + W5; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  157. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[ 6] + W6; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  158. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[ 7] + W7; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  159. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[ 8] + 0x80000000; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  160. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[ 9]; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  161. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[10]; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  162. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[11]; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  163. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[12]; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  164. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[13]; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  165. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[14]; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  166. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[15] + 0x00000100U; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  167. W0 = W0 + (rotr(W1, 7) ^ rotr(W1, 18) ^ (W1 >> 3U));
  168. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[16] + W0; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  169. W1 = W1 + (rotr(W2, 7) ^ rotr(W2, 18) ^ (W2 >> 3U)) + 0x00a00000U;
  170. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[17] + W1; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  171. W2 = W2 + (rotr(W3, 7) ^ rotr(W3, 18) ^ (W3 >> 3U)) + (rotr(W0, 17) ^ rotr(W0, 19) ^ (W0 >> 10U));
  172. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[18] + W2; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  173. W3 = W3 + (rotr(W4, 7) ^ rotr(W4, 18) ^ (W4 >> 3U)) + (rotr(W1, 17) ^ rotr(W1, 19) ^ (W1 >> 10U));
  174. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[19] + W3; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  175. W4 = W4 + (rotr(W5, 7) ^ rotr(W5, 18) ^ (W5 >> 3U)) + (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U));
  176. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[20] + W4; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  177. W5 = W5 + (rotr(W6, 7) ^ rotr(W6, 18) ^ (W6 >> 3U)) + (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  178. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[21] + W5; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  179. W6 = W6 + (rotr(W7, 7) ^ rotr(W7, 18) ^ (W7 >> 3U)) + 0x00000100U + (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U));
  180. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[22] + W6; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  181. W7 = W7 + 0x11002000U + W0 + (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U));
  182. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[23] + W7; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  183. W8 = 0x80000000 + W1 + (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U));
  184. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[24] + W8; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  185. W9 = W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  186. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[25] + W9; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  187. W10 = W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  188. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[26] + W10; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  189. W11 = W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  190. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[27] + W11; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  191. W12 = W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  192. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[28] + W12; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  193. W13 = W6 + (rotr(W11, 17) ^ rotr(W11, 19) ^ (W11 >> 10U));
  194. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[29] + W13; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  195. W14 = 0x00400022U + W7 + (rotr(W12, 17) ^ rotr(W12, 19) ^ (W12 >> 10U));
  196. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[30] + W14; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  197. W15 = 0x00000100U + (rotr(W0, 7) ^ rotr(W0, 18) ^ (W0 >> 3U)) + W8 + (rotr(W13, 17) ^ rotr(W13, 19) ^ (W13 >> 10U));
  198. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[31] + W15; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  199. W0 = W0 + (rotr(W1, 7) ^ rotr(W1, 18) ^ (W1 >> 3U)) + W9 + (rotr(W14, 17) ^ rotr(W14, 19) ^ (W14 >> 10U));
  200. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[32] + W0; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  201. W1 = W1 + (rotr(W2, 7) ^ rotr(W2, 18) ^ (W2 >> 3U)) + W10 + (rotr(W15, 17) ^ rotr(W15, 19) ^ (W15 >> 10U));
  202. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[33] + W1; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  203. W2 = W2 + (rotr(W3, 7) ^ rotr(W3, 18) ^ (W3 >> 3U)) + W11 + (rotr(W0, 17) ^ rotr(W0, 19) ^ (W0 >> 10U));
  204. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[34] + W2; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  205. W3 = W3 + (rotr(W4, 7) ^ rotr(W4, 18) ^ (W4 >> 3U)) + W12 + (rotr(W1, 17) ^ rotr(W1, 19) ^ (W1 >> 10U));
  206. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[35] + W3; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  207. W4 = W4 + (rotr(W5, 7) ^ rotr(W5, 18) ^ (W5 >> 3U)) + W13 + (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U));
  208. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[36] + W4; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  209. W5 = W5 + (rotr(W6, 7) ^ rotr(W6, 18) ^ (W6 >> 3U)) + W14 + (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  210. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[37] + W5; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  211. W6 = W6 + (rotr(W7, 7) ^ rotr(W7, 18) ^ (W7 >> 3U)) + W15 + (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U));
  212. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[38] + W6; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  213. W7 = W7 + (rotr(W8, 7) ^ rotr(W8, 18) ^ (W8 >> 3U)) + W0 + (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U));
  214. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[39] + W7; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  215. W8 = W8 + (rotr(W9, 7) ^ rotr(W9, 18) ^ (W9 >> 3U)) + W1 + (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U));
  216. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[40] + W8; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  217. W9 = W9 + (rotr(W10, 7) ^ rotr(W10, 18) ^ (W10 >> 3U)) + W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  218. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[41] + W9; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  219. W10 = W10 + (rotr(W11, 7) ^ rotr(W11, 18) ^ (W11 >> 3U)) + W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  220. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[42] + W10; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  221. W11 = W11 + (rotr(W12, 7) ^ rotr(W12, 18) ^ (W12 >> 3U)) + W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  222. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[43] + W11; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  223. W12 = W12 + (rotr(W13, 7) ^ rotr(W13, 18) ^ (W13 >> 3U)) + W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  224. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[44] + W12; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  225. W13 = W13 + (rotr(W14, 7) ^ rotr(W14, 18) ^ (W14 >> 3U)) + W6 + (rotr(W11, 17) ^ rotr(W11, 19) ^ (W11 >> 10U));
  226. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[45] + W13; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  227. W14 = W14 + (rotr(W15, 7) ^ rotr(W15, 18) ^ (W15 >> 3U)) + W7 + (rotr(W12, 17) ^ rotr(W12, 19) ^ (W12 >> 10U));
  228. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[46] + W14; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  229. W15 = W15 + (rotr(W0, 7) ^ rotr(W0, 18) ^ (W0 >> 3U)) + W8 + (rotr(W13, 17) ^ rotr(W13, 19) ^ (W13 >> 10U));
  230. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[47] + W15; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  231. W0 = W0 + (rotr(W1, 7) ^ rotr(W1, 18) ^ (W1 >> 3U)) + W9 + (rotr(W14, 17) ^ rotr(W14, 19) ^ (W14 >> 10U));
  232. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[48] + W0; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  233. W1 = W1 + (rotr(W2, 7) ^ rotr(W2, 18) ^ (W2 >> 3U)) + W10 + (rotr(W15, 17) ^ rotr(W15, 19) ^ (W15 >> 10U));
  234. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[49] + W1; C = C + G; G = G + (rotr(H, 2) ^ rotr(H, 13) ^ rotr(H, 22)) + Ma(B, H, A);
  235. W2 = W2 + (rotr(W3, 7) ^ rotr(W3, 18) ^ (W3 >> 3U)) + W11 + (rotr(W0, 17) ^ rotr(W0, 19) ^ (W0 >> 10U));
  236. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[50] + W2; B = B + F; F = F + (rotr(G, 2) ^ rotr(G, 13) ^ rotr(G, 22)) + Ma(A, G, H);
  237. W3 = W3 + (rotr(W4, 7) ^ rotr(W4, 18) ^ (W4 >> 3U)) + W12 + (rotr(W1, 17) ^ rotr(W1, 19) ^ (W1 >> 10U));
  238. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[51] + W3; A = A + E; E = E + (rotr(F, 2) ^ rotr(F, 13) ^ rotr(F, 22)) + Ma(H, F, G);
  239. W4 = W4 + (rotr(W5, 7) ^ rotr(W5, 18) ^ (W5 >> 3U)) + W13 + (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U));
  240. D = D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[52] + W4; H = H + D; D = D + (rotr(E, 2) ^ rotr(E, 13) ^ rotr(E, 22)) + Ma(G, E, F);
  241. W5 = W5 + (rotr(W6, 7) ^ rotr(W6, 18) ^ (W6 >> 3U)) + W14 + (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  242. C = C + (rotr(H, 6) ^ rotr(H, 11) ^ rotr(H, 25)) + Ch(H, A, B) + K[53] + W5; G = G + C; C = C + (rotr(D, 2) ^ rotr(D, 13) ^ rotr(D, 22)) + Ma(F, D, E);
  243. W6 = W6 + (rotr(W7, 7) ^ rotr(W7, 18) ^ (W7 >> 3U)) + W15 + (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U));
  244. B = B + (rotr(G, 6) ^ rotr(G, 11) ^ rotr(G, 25)) + Ch(G, H, A) + K[54] + W6; F = F + B; B = B + (rotr(C, 2) ^ rotr(C, 13) ^ rotr(C, 22)) + Ma(E, C, D);
  245. W7 = W7 + (rotr(W8, 7) ^ rotr(W8, 18) ^ (W8 >> 3U)) + W0 + (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U));
  246. A = A + (rotr(F, 6) ^ rotr(F, 11) ^ rotr(F, 25)) + Ch(F, G, H) + K[55] + W7; E = E + A; A = A + (rotr(B, 2) ^ rotr(B, 13) ^ rotr(B, 22)) + Ma(D, B, C);
  247. W8 = W8 + (rotr(W9, 7) ^ rotr(W9, 18) ^ (W9 >> 3U)) + W1 + (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U));
  248. H = H + (rotr(E, 6) ^ rotr(E, 11) ^ rotr(E, 25)) + Ch(E, F, G) + K[56] + W8; D = D + H; H = H + (rotr(A, 2) ^ rotr(A, 13) ^ rotr(A, 22)) + Ma(C, A, B);
  249. W9 = W9 + (rotr(W10, 7) ^ rotr(W10, 18) ^ (W10 >> 3U)) + W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  250. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[57] + W9; C = C + G;
  251. W10 = W10 + (rotr(W11, 7) ^ rotr(W11, 18) ^ (W11 >> 3U)) + W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  252. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[58] + W10; B = B + F;
  253. W11 = W11 + (rotr(W12, 7) ^ rotr(W12, 18) ^ (W12 >> 3U)) + W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  254. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[59] + W11; A = A + E;
  255. W12 = W12 + (rotr(W13, 7) ^ rotr(W13, 18) ^ (W13 >> 3U)) + W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  256. H = H + D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[60] + W12;
  257. H+=0x5be0cd19U;
  258. #ifdef VECTORS
  259. if (H.x == 0)
  260. {
  261. for (it = 0; it != 127; it++) {
  262. if (!output[it]) {
  263. output[it] = nonce.x;
  264. output[127] = 1;
  265. break;
  266. }
  267. }
  268. }
  269. if (H.y == 0)
  270. {
  271. for (it = 0; it != 127; it++) {
  272. if (!output[it]) {
  273. output[it] = nonce.y;
  274. output[127] = 1;
  275. break;
  276. }
  277. }
  278. }
  279. if (H.z == 0)
  280. {
  281. for (it = 0; it != 127; it++) {
  282. if (!output[it]) {
  283. output[it] = nonce.z;
  284. output[127] = 1;
  285. break;
  286. }
  287. }
  288. }
  289. if (H.w == 0)
  290. {
  291. for (it = 0; it != 127; it++) {
  292. if (!output[it]) {
  293. output[it] = nonce.w;
  294. output[127] = 1;
  295. break;
  296. }
  297. }
  298. }
  299. #else
  300. if (H == 0)
  301. {
  302. for (it = 0; it != 127; it++) {
  303. if (!output[it]) {
  304. output[it] = nonce;
  305. output[127] = 1;
  306. break;
  307. }
  308. }
  309. }
  310. #endif
  311. }