include/kernel_functions.c

   1 /**
   2  * Author......: Jens Steube <jens.steube@gmail.com>
   3  * License.....: MIT
   4  */
   5
   6 #if defined _MD4_ || defined _DCC2_ || defined _NETNTLMV2_ || defined _KRB5PA_ || defined _MS_DRSR_
   7 /**
   8  * MD4 Functions
   9  */
  10
  11 #if __CUDA_ARCH__ >= 500
  12 #define MD4_F(x,y,z)    lut3_ca ((x), (y), (z))
  13 #define MD4_G(x,y,z)    lut3_e8 ((x), (y), (z))
  14 #define MD4_H(x,y,z)    lut3_96 ((x), (y), (z))
  15 #define MD4_H1(x,y,z)   lut3_96 ((x), (y), (z))
  16 #define MD4_H2(x,y,z)   lut3_96 ((x), (y), (z))
  17 #else
  18 #define MD4_F(x,y,z)    (((x) & (y)) | ((~(x)) & (z)))
  19 #define MD4_G(x,y,z)    (((x) & (y)) | ((x) & (z)) | ((y) & (z)))
  20 #define MD4_H(x,y,z)    ((x) ^ (y) ^ (z))
  21 #define MD4_H1(x,y,z)   ((tmp2 = (x) ^ (y)) ^ (z))
  22 #define MD4_H2(x,y,z)   ((x) ^ tmp2)
  23 #endif
  24
  25 #ifdef IS_AMD
  26 #define MD4_Fo(x,y,z)   (bitselect ((z), (y), (x)))
  27 #define MD4_Go(x,y,z)   (bitselect ((x), (y), ((x) ^ (z))))
  28 #else
  29 #define MD4_Fo(x,y,z)   (MD4_F((x), (y), (z)))
  30 #define MD4_Go(x,y,z)   (MD4_G((x), (y), (z)))
  31 #endif
  32
  33 #define MD4_STEP(f,a,b,c,d,x,K,s)   \
  34 {                                   \
  35   a += K;                           \
  36   a += x;                           \
  37   a += f (b, c, d);                 \
  38   a  = rotl32 (a, s);               \
  39 }
  40
  41 #define MD4_STEP0(f,a,b,c,d,K,s)    \
  42 {                                   \
  43   a += K;                           \
  44   a += f (b, c, d);                 \
  45   a  = rotl32 (a, s);               \
  46 }
  47
  48 #endif
  49
  50 #if defined _MD5_ || defined _MD5H_ || defined _SAPB_ || defined _OLDOFFICE01_ || defined _WPA_ || defined _MD5_SHA1_ || defined _SHA1_MD5_ || defined _NETNTLMV2_ || defined _KRB5PA_ || defined _PBKDF2_MD5_
  51 /**
  52  * MD5 Functions
  53  */
  54
  55 #if __CUDA_ARCH__ >= 500
  56 #define MD5_F(x,y,z)    lut3_ca ((x), (y), (z))
  57 #define MD5_G(x,y,z)    lut3_e4 ((x), (y), (z))
  58 #define MD5_H(x,y,z)    lut3_96 ((x), (y), (z))
  59 #define MD5_H1(x,y,z)   lut3_96 ((x), (y), (z))
  60 #define MD5_H2(x,y,z)   lut3_96 ((x), (y), (z))
  61 #else
  62 #define MD5_F(x,y,z)    ((z) ^ ((x) & ((y) ^ (z))))
  63 #define MD5_G(x,y,z)    ((y) ^ ((z) & ((x) ^ (y))))
  64 #define MD5_H(x,y,z)    ((x) ^ (y) ^ (z))
  65 #define MD5_H1(x,y,z)   ((tmp2 = (x) ^ (y)) ^ (z))
  66 #define MD5_H2(x,y,z)   ((x) ^ tmp2)
  67 #endif
  68
  69 #ifdef IS_AMD
  70 #define MD5_I(x,y,z)    (bitselect (0xffffffffU, (x), (z)) ^ (y))
  71 #else
  72 #if __CUDA_ARCH__ >= 500
  73 #define MD5_I(x,y,z)    lut3_39 ((x), (y), (z))
  74 #else
  75 #define MD5_I(x,y,z)    ((y) ^ ((x) | ~(z)))
  76 #endif
  77 #endif
  78
  79 #ifdef IS_AMD
  80 #define MD5_Fo(x,y,z)   (bitselect ((z), (y), (x)))
  81 #define MD5_Go(x,y,z)   (bitselect ((y), (x), (z)))
  82 #else
  83 #define MD5_Fo(x,y,z)   (MD5_F((x), (y), (z)))
  84 #define MD5_Go(x,y,z)   (MD5_G((x), (y), (z)))
  85 #endif
  86
  87 #define MD5_STEP(f,a,b,c,d,x,K,s)   \
  88 {                                   \
  89   a += K;                           \
  90   a += x;                           \
  91   a += f (b, c, d);                 \
  92   a  = rotl32 (a, s);               \
  93   a += b;                           \
  94 }
  95
  96 #define MD5_STEP0(f,a,b,c,d,K,s)    \
  97 {                                   \
  98   a += K;                           \
  99   a += f (b, c, d);                 \
 100   a  = rotl32 (a, s);               \
 101   a += b;                           \
 102 }
 103 #endif
 104
 105 #if defined _SHA1_ || defined _SAPG_ || defined _OFFICE2007_ || defined _OFFICE2010_ || defined _OLDOFFICE34_ || defined _ANDROIDFDE_ || defined _DCC2_ || defined _WPA_ || defined _MD5_SHA1_ || defined _SHA1_MD5_ || defined _PSAFE2_ || defined _LOTUS8_ || defined _PBKDF2_SHA1_ || defined _RAR3_ || defined _SHA256_SHA1_
 106
 107 /**
 108  * SHA1 Functions
 109  */
 110
 111 #if __CUDA_ARCH__ >= 500
 112 #define SHA1_F0(x,y,z)  lut3_ca ((x), (y), (z))
 113 #define SHA1_F1(x,y,z)  lut3_96 ((x), (y), (z))
 114 #define SHA1_F2(x,y,z)  lut3_e8 ((x), (y), (z))
 115 #else
 116 #define SHA1_F0(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 117 #define SHA1_F1(x,y,z)  ((x) ^ (y) ^ (z))
 118 #define SHA1_F2(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 119 #endif
 120
 121 #ifdef IS_AMD
 122 #define SHA1_F0o(x,y,z) (bitselect ((z), (y), (x)))
 123 #define SHA1_F2o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 124 #else
 125 #define SHA1_F0o(x,y,z) (SHA1_F0 ((x), (y), (z)))
 126 #define SHA1_F2o(x,y,z) (SHA1_F2 ((x), (y), (z)))
 127 #endif
 128
 129 #define SHA1_STEP(f,a,b,c,d,e,x)    \
 130 {                                   \
 131   e += K;                           \
 132   e += x;                           \
 133   e += f (b, c, d);                 \
 134   e += rotl32 (a,  5u);             \
 135   b  = rotl32 (b, 30u);             \
 136 }
 137
 138 #define SHA1_STEP0(f,a,b,c,d,e,x)   \
 139 {                                   \
 140   e += K;                           \
 141   e += f (b, c, d);                 \
 142   e += rotl32 (a,  5u);             \
 143   b  = rotl32 (b, 30u);             \
 144 }
 145
 146 #define SHA1_STEPX(f,a,b,c,d,e,x)   \
 147 {                                   \
 148   e += x;                           \
 149   e += f (b, c, d);                 \
 150   e += rotl32 (a,  5u);             \
 151   b  = rotl32 (b, 30u);             \
 152 }
 153
 154 #define SHA1_STEP_PE(f,a,b,c,d,e,x) \
 155 {                                   \
 156   e += x;                           \
 157   e += f (b, c, d);                 \
 158   e += rotl32 (a,  5u);             \
 159 }
 160
 161 #define SHA1_STEP_PB(f,a,b,c,d,e,x) \
 162 {                                   \
 163   e += K;                           \
 164   b  = rotl32 (b, 30u);             \
 165 }
 166 #endif
 167
 168 #if defined _SHA256_ || defined _PDF17L8_ || defined _SEVEN_ZIP_ || defined _ANDROIDFDE_ || defined _CLOUDKEY_ || defined _SCRYPT_ || defined _PBKDF2_SHA256_  || defined _SHA256_SHA1_ || defined _MS_DRSR_
 169 /**
 170  * SHA256 Functions
 171  */
 172
 173 #define SHIFT_RIGHT_32(x,n) ((x) >> (n))
 174
 175 #define SHA256_S0(x) (rotl32 ((x), 25u) ^ rotl32 ((x), 14u) ^ SHIFT_RIGHT_32 ((x),  3u))
 176 #define SHA256_S1(x) (rotl32 ((x), 15u) ^ rotl32 ((x), 13u) ^ SHIFT_RIGHT_32 ((x), 10u))
 177 #define SHA256_S2(x) (rotl32 ((x), 30u) ^ rotl32 ((x), 19u) ^ rotl32 ((x), 10u))
 178 #define SHA256_S3(x) (rotl32 ((x), 26u) ^ rotl32 ((x), 21u) ^ rotl32 ((x),  7u))
 179
 180 #if __CUDA_ARCH__ >= 500
 181 #define SHA256_F0(x,y,z)  lut3_e8 ((x), (y), (z))
 182 #define SHA256_F1(x,y,z)  lut3_ca ((x), (y), (z))
 183 #else
 184 #define SHA256_F0(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 185 #define SHA256_F1(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 186 #endif
 187
 188 #ifdef IS_AMD
 189 #define SHA256_F0o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 190 #define SHA256_F1o(x,y,z) (bitselect ((z), (y), (x)))
 191 #else
 192 #define SHA256_F0o(x,y,z) (SHA256_F0 ((x), (y), (z)))
 193 #define SHA256_F1o(x,y,z) (SHA256_F1 ((x), (y), (z)))
 194 #endif
 195
 196 #define SHA256_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 197 {                                               \
 198   h += K;                                       \
 199   h += x;                                       \
 200   h += SHA256_S3 (e);                           \
 201   h += F1 (e,f,g);                              \
 202   d += h;                                       \
 203   h += SHA256_S2 (a);                           \
 204   h += F0 (a,b,c);                              \
 205 }
 206
 207 #define SHA256_EXPAND(x,y,z,w) (SHA256_S1 (x) + y + SHA256_S0 (z) + w)
 208 #endif
 209
 210
 211 #if defined _SHA384_ || defined _PDF17L8_
 212 /**
 213  * SHA384 Functions
 214  */
 215
 216 #define SHIFT_RIGHT_64(x,n) ((x) >> (n))
 217
 218 #define SHA384_S0(x) (rotr64 ((x), 28) ^ rotr64 ((x), 34) ^ rotr64 ((x), 39))
 219 #define SHA384_S1(x) (rotr64 ((x), 14) ^ rotr64 ((x), 18) ^ rotr64 ((x), 41))
 220 #define SHA384_S2(x) (rotr64 ((x),  1) ^ rotr64 ((x),  8) ^ SHIFT_RIGHT_64 ((x), 7))
 221 #define SHA384_S3(x) (rotr64 ((x), 19) ^ rotr64 ((x), 61) ^ SHIFT_RIGHT_64 ((x), 6))
 222
 223 #define SHA384_F0(x,y,z) ((z) ^ ((x) & ((y) ^ (z))))
 224 #define SHA384_F1(x,y,z) (((x) & (y)) | ((z) & ((x) ^ (y))))
 225
 226 #ifdef IS_AMD
 227 #define SHA384_F0o(x,y,z) (bitselect ((z), (y), (x)))
 228 #define SHA384_F1o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 229 #else
 230 #define SHA384_F0o(x,y,z) (SHA384_F0 ((x), (y), (z)))
 231 #define SHA384_F1o(x,y,z) (SHA384_F1 ((x), (y), (z)))
 232 #endif
 233
 234 #define SHA384_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 235 {                                               \
 236   u64x   temp0;                                 \
 237   temp0  = K;                                   \
 238   temp0 += x;                                   \
 239   temp0 += h;                                   \
 240   temp0 += SHA384_S1 (e);                       \
 241   temp0 += F0 (e, f, g);                        \
 242   d     += temp0;                               \
 243   h      = SHA384_S0 (a);                       \
 244   h     += F1 (a, b, c);                        \
 245   h     += temp0;                               \
 246 }
 247
 248 #define SHA384_EXPAND(x,y,z,w) (SHA384_S3 (x) + y + SHA384_S2 (z) + w)
 249 #endif
 250
 251 #if defined _SHA512_ || defined _CLOUDKEY_ || defined _OFFICE2013_ || defined _PDF17L8_ || defined _PBKDF2_SHA512_
 252 /**
 253  * SHA512 Functions
 254  */
 255
 256 #define SHIFT_RIGHT_64(x,n) ((x) >> (n))
 257
 258 #define SHA512_S0(x) (rotr64 ((x), 28) ^ rotr64 ((x), 34) ^ rotr64 ((x), 39))
 259 #define SHA512_S1(x) (rotr64 ((x), 14) ^ rotr64 ((x), 18) ^ rotr64 ((x), 41))
 260 #define SHA512_S2(x) (rotr64 ((x),  1) ^ rotr64 ((x),  8) ^ SHIFT_RIGHT_64 ((x), 7))
 261 #define SHA512_S3(x) (rotr64 ((x), 19) ^ rotr64 ((x), 61) ^ SHIFT_RIGHT_64 ((x), 6))
 262
 263 #define SHA512_F0(x,y,z) ((z) ^ ((x) & ((y) ^ (z))))
 264 #define SHA512_F1(x,y,z) (((x) & (y)) | ((z) & ((x) ^ (y))))
 265
 266 #ifdef IS_AMD
 267 #define SHA512_F0o(x,y,z) (bitselect ((z), (y), (x)))
 268 #define SHA512_F1o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 269 #else
 270 #define SHA512_F0o(x,y,z) (SHA512_F0 ((x), (y), (z)))
 271 #define SHA512_F1o(x,y,z) (SHA512_F1 ((x), (y), (z)))
 272 #endif
 273
 274 #define SHA512_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 275 {                                               \
 276   u64x   temp0;                                 \
 277   temp0  = K;                                   \
 278   temp0 += x;                                   \
 279   temp0 += h;                                   \
 280   temp0 += SHA512_S1 (e);                       \
 281   temp0 += F0 (e, f, g);                        \
 282   d     += temp0;                               \
 283   h      = SHA512_S0 (a);                       \
 284   h     += F1 (a, b, c);                        \
 285   h     += temp0;                               \
 286 }
 287
 288 #define SHA512_EXPAND(x,y,z,w) (SHA512_S3 (x) + y + SHA512_S2 (z) + w)
 289
 290 #define SHA512_S2_WO(x) (rotate ((x), 64- 1ull) ^ rotate ((x), 64- 8ull) ^ SHIFT_RIGHT_64 ((x), 7))
 291 #define SHA512_S3_WO(x) (rotate ((x), 64-19ull) ^ rotate ((x), 64-61ull) ^ SHIFT_RIGHT_64 ((x), 6))
 292
 293 #define SHA512_EXPAND_WO(x,y,z,w) (SHA512_S3_WO (x) + y + SHA512_S2_WO (z) + w)
 294 #endif
 295
 296 #ifdef _RIPEMD160_
 297 /**
 298  * RIPEMD160 Functions
 299  */
 300
 301 #if __CUDA_ARCH__ >= 500
 302 #define RIPEMD160_F(x,y,z)    lut3_96 ((x), (y), (z))
 303 #define RIPEMD160_G(x,y,z)    lut3_ca ((x), (y), (z))
 304 #define RIPEMD160_H(x,y,z)    lut3_59 ((x), (y), (z))
 305 #define RIPEMD160_I(x,y,z)    lut3_e4 ((x), (y), (z))
 306 #define RIPEMD160_J(x,y,z)    lut3_2d ((x), (y), (z))
 307 #else
 308 #define RIPEMD160_F(x,y,z)    ((x) ^ (y) ^ (z))
 309 #define RIPEMD160_G(x,y,z)    ((z) ^ ((x) & ((y) ^ (z)))) /* x ? y : z */
 310 #define RIPEMD160_H(x,y,z)    (((x) | ~(y)) ^ (z))
 311 #define RIPEMD160_I(x,y,z)    ((y) ^ ((z) & ((x) ^ (y)))) /* z ? x : y */
 312 #define RIPEMD160_J(x,y,z)    ((x) ^ ((y) | ~(z)))
 313 #endif
 314
 315 #ifdef IS_AMD
 316 #define RIPEMD160_Go(x,y,z)   (bitselect ((z), (y), (x)))
 317 #define RIPEMD160_Io(x,y,z)   (bitselect ((y), (x), (z)))
 318 #else
 319 #define RIPEMD160_Go(x,y,z)   (RIPEMD160_G ((x), (y), (z)))
 320 #define RIPEMD160_Io(x,y,z)   (RIPEMD160_I ((x), (y), (z)))
 321 #endif
 322
 323 #define RIPEMD160_STEP(f,a,b,c,d,e,x,K,s) \
 324 {                                         \
 325   a += K;                                 \
 326   a += x;                                 \
 327   a += f (b, c, d);                       \
 328   a  = rotl32 (a, s);                     \
 329   a += e;                                 \
 330   c  = rotl32 (c, 10u);                   \
 331 }
 332
 333 #define ROTATE_LEFT_WORKAROUND_BUG(a,n) ((a << n) | (a >> (32 - n)))
 334
 335 #define RIPEMD160_STEP_WORKAROUND_BUG(f,a,b,c,d,e,x,K,s)  \
 336 {                                         \
 337   a += K;                                 \
 338   a += x;                                 \
 339   a += f (b, c, d);                       \
 340   a  = ROTATE_LEFT_WORKAROUND_BUG (a, s); \
 341   a += e;                                 \
 342   c  = rotl32 (c, 10u);                   \
 343 }
 344
 345 #endif
 346