poclbm.cl 30 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323
  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. #ifdef VECTORS
  4. typedef uint2 u;
  5. #else
  6. typedef uint u;
  7. #endif
  8. __constant uint K[64] = {
  9. 0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5, 0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
  10. 0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3, 0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
  11. 0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc, 0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
  12. 0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7, 0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967,
  13. 0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13, 0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85,
  14. 0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3, 0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070,
  15. 0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5, 0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3,
  16. 0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208, 0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2
  17. };
  18. #define BITALIGN
  19. #ifdef BITALIGN
  20. #pragma OPENCL EXTENSION cl_amd_media_ops : enable
  21. #define rotr(x, y) amd_bitalign((u)x, (u)x, (u)y)
  22. #else
  23. #define rotr(x, y) rotate((u)x, (u)(32-y))
  24. #endif
  25. // This part is not from the stock poclbm kernel. It's part of an optimization
  26. // added in the Phoenix Miner.
  27. // Some AMD devices have a BFI_INT opcode, which behaves exactly like the
  28. // SHA-256 Ch function, but provides it in exactly one instruction. If
  29. // detected, use it for Ch. Otherwise, construct Ch out of simpler logical
  30. // primitives.
  31. #ifdef BFI_INT
  32. // Well, slight problem... It turns out BFI_INT isn't actually exposed to
  33. // OpenCL (or CAL IL for that matter) in any way. However, there is
  34. // a similar instruction, BYTE_ALIGN_INT, which is exposed to OpenCL via
  35. // amd_bytealign, takes the same inputs, and provides the same output.
  36. // We can use that as a placeholder for BFI_INT and have the application
  37. // patch it after compilation.
  38. // This is the BFI_INT function
  39. #define Ch(x, y, z) amd_bytealign(x, y, z)
  40. // Ma can also be implemented in terms of BFI_INT...
  41. #define Ma(x, y, z) amd_bytealign((y), (x | z), (z & x))
  42. #else
  43. #define Ch(x, y, z) (z ^ (x & (y ^ z)))
  44. #define Ma(x, y, z) ((x & z) | (y & (x | z)))
  45. #endif
  46. // AMD's KernelAnalyzer throws errors compiling the kernel if we use
  47. // amd_bytealign on constants with vectors enabled, so we use this to avoid
  48. // problems. (this is used 4 times, and likely optimized out by the compiler.)
  49. #define Ma2(x, y, z) ((y & z) | (x & (y | z)))
  50. __kernel void search( const uint state0, const uint state1, const uint state2, const uint state3,
  51. const uint state4, const uint state5, const uint state6, const uint state7,
  52. const uint B1, const uint C1, const uint D1,
  53. const uint F1, const uint G1, const uint H1,
  54. const uint base,
  55. 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,
  56. __global uint * output)
  57. {
  58. u W0, W1, W2, W3, W4, W5, W6, W7, W8, W9, W10, W11, W12, W13, W14, W15;
  59. u A,B,C,D,E,F,G,H;
  60. u nonce;
  61. uint it;
  62. #ifdef VECTORS
  63. nonce = ((base + get_global_id(0))<<1) + (uint2)(0, 1);
  64. #else
  65. nonce = base + get_global_id(0);
  66. #endif
  67. W3 = nonce + fW3;
  68. E = fcty_e + nonce; A = state0 + E; E = E + fcty_e2;
  69. 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);
  70. 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);
  71. 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);
  72. 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);
  73. 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);
  74. 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);
  75. 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);
  76. 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);
  77. 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);
  78. 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);
  79. 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);
  80. 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);
  81. 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);
  82. 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);
  83. W2 = (rotr(nonce, 7) ^ rotr(nonce, 18) ^ (nonce >> 3U)) + fW2;
  84. 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);
  85. 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);
  86. W4 = (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U)) + 0x80000000;
  87. 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);
  88. W5 = (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  89. 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);
  90. W6 = (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U)) + 0x00000280U;
  91. 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);
  92. W7 = (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U)) + fW0;
  93. 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);
  94. W8 = (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U)) + fW1;
  95. 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);
  96. W9 = W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  97. 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);
  98. W10 = W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  99. 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);
  100. W11 = W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  101. 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);
  102. W12 = W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  103. 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);
  104. W13 = W6 + (rotr(W11, 17) ^ rotr(W11, 19) ^ (W11 >> 10U));
  105. 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);
  106. W14 = 0x00a00055U + W7 + (rotr(W12, 17) ^ rotr(W12, 19) ^ (W12 >> 10U));
  107. 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);
  108. W15 = fW15 + W8 + (rotr(W13, 17) ^ rotr(W13, 19) ^ (W13 >> 10U));
  109. 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);
  110. W0 = fW01r + W9 + (rotr(W14, 17) ^ rotr(W14, 19) ^ (W14 >> 10U));
  111. 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);
  112. W1 = fW1 + (rotr(W2, 7) ^ rotr(W2, 18) ^ (W2 >> 3U)) + W10 + (rotr(W15, 17) ^ rotr(W15, 19) ^ (W15 >> 10U));
  113. 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);
  114. W2 = W2 + (rotr(W3, 7) ^ rotr(W3, 18) ^ (W3 >> 3U)) + W11 + (rotr(W0, 17) ^ rotr(W0, 19) ^ (W0 >> 10U));
  115. 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);
  116. W3 = W3 + (rotr(W4, 7) ^ rotr(W4, 18) ^ (W4 >> 3U)) + W12 + (rotr(W1, 17) ^ rotr(W1, 19) ^ (W1 >> 10U));
  117. 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);
  118. W4 = W4 + (rotr(W5, 7) ^ rotr(W5, 18) ^ (W5 >> 3U)) + W13 + (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U));
  119. 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);
  120. W5 = W5 + (rotr(W6, 7) ^ rotr(W6, 18) ^ (W6 >> 3U)) + W14 + (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  121. 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);
  122. W6 = W6 + (rotr(W7, 7) ^ rotr(W7, 18) ^ (W7 >> 3U)) + W15 + (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U));
  123. 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);
  124. W7 = W7 + (rotr(W8, 7) ^ rotr(W8, 18) ^ (W8 >> 3U)) + W0 + (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U));
  125. 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);
  126. W8 = W8 + (rotr(W9, 7) ^ rotr(W9, 18) ^ (W9 >> 3U)) + W1 + (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U));
  127. 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);
  128. W9 = W9 + (rotr(W10, 7) ^ rotr(W10, 18) ^ (W10 >> 3U)) + W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  129. 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);
  130. W10 = W10 + (rotr(W11, 7) ^ rotr(W11, 18) ^ (W11 >> 3U)) + W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  131. 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);
  132. W11 = W11 + (rotr(W12, 7) ^ rotr(W12, 18) ^ (W12 >> 3U)) + W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  133. 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);
  134. W12 = W12 + (rotr(W13, 7) ^ rotr(W13, 18) ^ (W13 >> 3U)) + W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  135. 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);
  136. W13 = W13 + (rotr(W14, 7) ^ rotr(W14, 18) ^ (W14 >> 3U)) + W6 + (rotr(W11, 17) ^ rotr(W11, 19) ^ (W11 >> 10U));
  137. 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);
  138. W14 = W14 + (rotr(W15, 7) ^ rotr(W15, 18) ^ (W15 >> 3U)) + W7 + (rotr(W12, 17) ^ rotr(W12, 19) ^ (W12 >> 10U));
  139. 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);
  140. W15 = W15 + (rotr(W0, 7) ^ rotr(W0, 18) ^ (W0 >> 3U)) + W8 + (rotr(W13, 17) ^ rotr(W13, 19) ^ (W13 >> 10U));
  141. 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);
  142. W0 = W0 + (rotr(W1, 7) ^ rotr(W1, 18) ^ (W1 >> 3U)) + W9 + (rotr(W14, 17) ^ rotr(W14, 19) ^ (W14 >> 10U));
  143. 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);
  144. W1 = W1 + (rotr(W2, 7) ^ rotr(W2, 18) ^ (W2 >> 3U)) + W10 + (rotr(W15, 17) ^ rotr(W15, 19) ^ (W15 >> 10U));
  145. 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);
  146. W2 = W2 + (rotr(W3, 7) ^ rotr(W3, 18) ^ (W3 >> 3U)) + W11 + (rotr(W0, 17) ^ rotr(W0, 19) ^ (W0 >> 10U));
  147. 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);
  148. W3 = W3 + (rotr(W4, 7) ^ rotr(W4, 18) ^ (W4 >> 3U)) + W12 + (rotr(W1, 17) ^ rotr(W1, 19) ^ (W1 >> 10U));
  149. 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);
  150. W4 = W4 + (rotr(W5, 7) ^ rotr(W5, 18) ^ (W5 >> 3U)) + W13 + (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U));
  151. 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);
  152. W5 = W5 + (rotr(W6, 7) ^ rotr(W6, 18) ^ (W6 >> 3U)) + W14 + (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  153. 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);
  154. W6 = W6 + (rotr(W7, 7) ^ rotr(W7, 18) ^ (W7 >> 3U)) + W15 + (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U));
  155. 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);
  156. W7 = W7 + (rotr(W8, 7) ^ rotr(W8, 18) ^ (W8 >> 3U)) + W0 + (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U));
  157. 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);
  158. W8 = W8 + (rotr(W9, 7) ^ rotr(W9, 18) ^ (W9 >> 3U)) + W1 + (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U));
  159. 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);
  160. W9 = W9 + (rotr(W10, 7) ^ rotr(W10, 18) ^ (W10 >> 3U)) + W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  161. 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);
  162. W10 = W10 + (rotr(W11, 7) ^ rotr(W11, 18) ^ (W11 >> 3U)) + W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  163. 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);
  164. W11 = W11 + (rotr(W12, 7) ^ rotr(W12, 18) ^ (W12 >> 3U)) + W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  165. 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);
  166. W12 = W12 + (rotr(W13, 7) ^ rotr(W13, 18) ^ (W13 >> 3U)) + W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  167. 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);
  168. W13 = W13 + (rotr(W14, 7) ^ rotr(W14, 18) ^ (W14 >> 3U)) + W6 + (rotr(W11, 17) ^ rotr(W11, 19) ^ (W11 >> 10U));
  169. 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);
  170. W14 = W14 + (rotr(W15, 7) ^ rotr(W15, 18) ^ (W15 >> 3U)) + W7 + (rotr(W12, 17) ^ rotr(W12, 19) ^ (W12 >> 10U));
  171. 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);
  172. W15 = W15 + (rotr(W0, 7) ^ rotr(W0, 18) ^ (W0 >> 3U)) + W8 + (rotr(W13, 17) ^ rotr(W13, 19) ^ (W13 >> 10U));
  173. 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);
  174. W0 = A + state0; W1 = B + state1;
  175. W2 = C + state2; W3 = D + state3;
  176. W4 = E + state4; W5 = F + state5;
  177. W6 = G + state6; W7 = H + state7;
  178. H = 0xb0edbdd0 + K[ 0] + W0; D = 0xa54ff53a + H; H = H + 0x08909ae5U;
  179. 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);
  180. 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);
  181. 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);
  182. 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);
  183. 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);
  184. 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);
  185. 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);
  186. 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);
  187. 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);
  188. 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);
  189. 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);
  190. 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);
  191. 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);
  192. 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);
  193. 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);
  194. W0 = W0 + (rotr(W1, 7) ^ rotr(W1, 18) ^ (W1 >> 3U));
  195. 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);
  196. W1 = W1 + (rotr(W2, 7) ^ rotr(W2, 18) ^ (W2 >> 3U)) + 0x00a00000U;
  197. 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);
  198. W2 = W2 + (rotr(W3, 7) ^ rotr(W3, 18) ^ (W3 >> 3U)) + (rotr(W0, 17) ^ rotr(W0, 19) ^ (W0 >> 10U));
  199. 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);
  200. W3 = W3 + (rotr(W4, 7) ^ rotr(W4, 18) ^ (W4 >> 3U)) + (rotr(W1, 17) ^ rotr(W1, 19) ^ (W1 >> 10U));
  201. 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);
  202. W4 = W4 + (rotr(W5, 7) ^ rotr(W5, 18) ^ (W5 >> 3U)) + (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U));
  203. 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);
  204. W5 = W5 + (rotr(W6, 7) ^ rotr(W6, 18) ^ (W6 >> 3U)) + (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  205. 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);
  206. W6 = W6 + (rotr(W7, 7) ^ rotr(W7, 18) ^ (W7 >> 3U)) + 0x00000100U + (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U));
  207. 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);
  208. W7 = W7 + 0x11002000U + W0 + (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U));
  209. 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);
  210. W8 = 0x80000000 + W1 + (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U));
  211. 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);
  212. W9 = W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  213. 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);
  214. W10 = W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  215. 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);
  216. W11 = W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  217. 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);
  218. W12 = W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  219. 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);
  220. W13 = W6 + (rotr(W11, 17) ^ rotr(W11, 19) ^ (W11 >> 10U));
  221. 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);
  222. W14 = 0x00400022U + W7 + (rotr(W12, 17) ^ rotr(W12, 19) ^ (W12 >> 10U));
  223. 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);
  224. W15 = 0x00000100U + (rotr(W0, 7) ^ rotr(W0, 18) ^ (W0 >> 3U)) + W8 + (rotr(W13, 17) ^ rotr(W13, 19) ^ (W13 >> 10U));
  225. 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);
  226. W0 = W0 + (rotr(W1, 7) ^ rotr(W1, 18) ^ (W1 >> 3U)) + W9 + (rotr(W14, 17) ^ rotr(W14, 19) ^ (W14 >> 10U));
  227. 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);
  228. W1 = W1 + (rotr(W2, 7) ^ rotr(W2, 18) ^ (W2 >> 3U)) + W10 + (rotr(W15, 17) ^ rotr(W15, 19) ^ (W15 >> 10U));
  229. 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);
  230. W2 = W2 + (rotr(W3, 7) ^ rotr(W3, 18) ^ (W3 >> 3U)) + W11 + (rotr(W0, 17) ^ rotr(W0, 19) ^ (W0 >> 10U));
  231. 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);
  232. W3 = W3 + (rotr(W4, 7) ^ rotr(W4, 18) ^ (W4 >> 3U)) + W12 + (rotr(W1, 17) ^ rotr(W1, 19) ^ (W1 >> 10U));
  233. 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);
  234. W4 = W4 + (rotr(W5, 7) ^ rotr(W5, 18) ^ (W5 >> 3U)) + W13 + (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U));
  235. 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);
  236. W5 = W5 + (rotr(W6, 7) ^ rotr(W6, 18) ^ (W6 >> 3U)) + W14 + (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  237. 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);
  238. W6 = W6 + (rotr(W7, 7) ^ rotr(W7, 18) ^ (W7 >> 3U)) + W15 + (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U));
  239. 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);
  240. W7 = W7 + (rotr(W8, 7) ^ rotr(W8, 18) ^ (W8 >> 3U)) + W0 + (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U));
  241. 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);
  242. W8 = W8 + (rotr(W9, 7) ^ rotr(W9, 18) ^ (W9 >> 3U)) + W1 + (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U));
  243. 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);
  244. W9 = W9 + (rotr(W10, 7) ^ rotr(W10, 18) ^ (W10 >> 3U)) + W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  245. 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);
  246. W10 = W10 + (rotr(W11, 7) ^ rotr(W11, 18) ^ (W11 >> 3U)) + W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  247. 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);
  248. W11 = W11 + (rotr(W12, 7) ^ rotr(W12, 18) ^ (W12 >> 3U)) + W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  249. 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);
  250. W12 = W12 + (rotr(W13, 7) ^ rotr(W13, 18) ^ (W13 >> 3U)) + W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  251. 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);
  252. W13 = W13 + (rotr(W14, 7) ^ rotr(W14, 18) ^ (W14 >> 3U)) + W6 + (rotr(W11, 17) ^ rotr(W11, 19) ^ (W11 >> 10U));
  253. 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);
  254. W14 = W14 + (rotr(W15, 7) ^ rotr(W15, 18) ^ (W15 >> 3U)) + W7 + (rotr(W12, 17) ^ rotr(W12, 19) ^ (W12 >> 10U));
  255. 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);
  256. W15 = W15 + (rotr(W0, 7) ^ rotr(W0, 18) ^ (W0 >> 3U)) + W8 + (rotr(W13, 17) ^ rotr(W13, 19) ^ (W13 >> 10U));
  257. 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);
  258. W0 = W0 + (rotr(W1, 7) ^ rotr(W1, 18) ^ (W1 >> 3U)) + W9 + (rotr(W14, 17) ^ rotr(W14, 19) ^ (W14 >> 10U));
  259. 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);
  260. W1 = W1 + (rotr(W2, 7) ^ rotr(W2, 18) ^ (W2 >> 3U)) + W10 + (rotr(W15, 17) ^ rotr(W15, 19) ^ (W15 >> 10U));
  261. 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);
  262. W2 = W2 + (rotr(W3, 7) ^ rotr(W3, 18) ^ (W3 >> 3U)) + W11 + (rotr(W0, 17) ^ rotr(W0, 19) ^ (W0 >> 10U));
  263. 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);
  264. W3 = W3 + (rotr(W4, 7) ^ rotr(W4, 18) ^ (W4 >> 3U)) + W12 + (rotr(W1, 17) ^ rotr(W1, 19) ^ (W1 >> 10U));
  265. 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);
  266. W4 = W4 + (rotr(W5, 7) ^ rotr(W5, 18) ^ (W5 >> 3U)) + W13 + (rotr(W2, 17) ^ rotr(W2, 19) ^ (W2 >> 10U));
  267. 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);
  268. W5 = W5 + (rotr(W6, 7) ^ rotr(W6, 18) ^ (W6 >> 3U)) + W14 + (rotr(W3, 17) ^ rotr(W3, 19) ^ (W3 >> 10U));
  269. 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);
  270. W6 = W6 + (rotr(W7, 7) ^ rotr(W7, 18) ^ (W7 >> 3U)) + W15 + (rotr(W4, 17) ^ rotr(W4, 19) ^ (W4 >> 10U));
  271. 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);
  272. W7 = W7 + (rotr(W8, 7) ^ rotr(W8, 18) ^ (W8 >> 3U)) + W0 + (rotr(W5, 17) ^ rotr(W5, 19) ^ (W5 >> 10U));
  273. 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);
  274. W8 = W8 + (rotr(W9, 7) ^ rotr(W9, 18) ^ (W9 >> 3U)) + W1 + (rotr(W6, 17) ^ rotr(W6, 19) ^ (W6 >> 10U));
  275. 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);
  276. W9 = W9 + (rotr(W10, 7) ^ rotr(W10, 18) ^ (W10 >> 3U)) + W2 + (rotr(W7, 17) ^ rotr(W7, 19) ^ (W7 >> 10U));
  277. G = G + (rotr(D, 6) ^ rotr(D, 11) ^ rotr(D, 25)) + Ch(D, E, F) + K[57] + W9; C = C + G;
  278. W10 = W10 + (rotr(W11, 7) ^ rotr(W11, 18) ^ (W11 >> 3U)) + W3 + (rotr(W8, 17) ^ rotr(W8, 19) ^ (W8 >> 10U));
  279. F = F + (rotr(C, 6) ^ rotr(C, 11) ^ rotr(C, 25)) + Ch(C, D, E) + K[58] + W10; B = B + F;
  280. W11 = W11 + (rotr(W12, 7) ^ rotr(W12, 18) ^ (W12 >> 3U)) + W4 + (rotr(W9, 17) ^ rotr(W9, 19) ^ (W9 >> 10U));
  281. E = E + (rotr(B, 6) ^ rotr(B, 11) ^ rotr(B, 25)) + Ch(B, C, D) + K[59] + W11; A = A + E;
  282. W12 = W12 + (rotr(W13, 7) ^ rotr(W13, 18) ^ (W13 >> 3U)) + W5 + (rotr(W10, 17) ^ rotr(W10, 19) ^ (W10 >> 10U));
  283. H = H + D + (rotr(A, 6) ^ rotr(A, 11) ^ rotr(A, 25)) + Ch(A, B, C) + K[60] + W12;
  284. H+=0x5be0cd19U;
  285. #ifdef VECTORS
  286. if (H.x == 0)
  287. {
  288. output[OUTPUT_SIZE] = output[nonce.x & OUTPUT_MASK] = nonce.x;
  289. }
  290. else if (H.y == 0)
  291. {
  292. output[OUTPUT_SIZE] = output[nonce.y & OUTPUT_MASK] = nonce.y;
  293. }
  294. #else
  295. if (H == 0)
  296. {
  297. for (it = 0; it != 127; it++) {
  298. if (!output[it]) {
  299. output[it] = nonce;
  300. output[127] = 1;
  301. break;
  302. }
  303. }
  304. }
  305. #endif
  306. }