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 #ifdef IS_NV
   9 #if CUDA_ARCH >= 500
  10 #define MD4_F(x,y,z)    lut3_ca ((x), (y), (z))
  11 #define MD4_G(x,y,z)    lut3_e8 ((x), (y), (z))
  12 #define MD4_H(x,y,z)    lut3_96 ((x), (y), (z))
  13 #define MD4_H1(x,y,z)   lut3_96 ((x), (y), (z))
  14 #define MD4_H2(x,y,z)   lut3_96 ((x), (y), (z))
  15 #else
  16 #define MD4_F(x,y,z)    (((x) & (y)) | ((~(x)) & (z)))
  17 #define MD4_G(x,y,z)    (((x) & (y)) | ((x) & (z)) | ((y) & (z)))
  18 #define MD4_H(x,y,z)    ((x) ^ (y) ^ (z))
  19 #define MD4_H1(x,y,z)   ((tmp2 = (x) ^ (y)) ^ (z))
  20 #define MD4_H2(x,y,z)   ((x) ^ tmp2)
  21 #endif
  22 #define MD4_Fo(x,y,z)   (MD4_F((x), (y), (z)))
  23 #define MD4_Go(x,y,z)   (MD4_G((x), (y), (z)))
  24 #endif
  25
  26 #ifdef IS_AMD
  27 #define MD4_F(x,y,z)    (((x) & (y)) | ((~(x)) & (z)))
  28 #define MD4_G(x,y,z)    (((x) & (y)) | ((x) & (z)) | ((y) & (z)))
  29 #define MD4_H(x,y,z)    ((x) ^ (y) ^ (z))
  30 #define MD4_H1(x,y,z)   ((tmp2 = (x) ^ (y)) ^ (z))
  31 #define MD4_H2(x,y,z)   ((x) ^ tmp2)
  32 #define MD4_Fo(x,y,z)   (bitselect ((z), (y), (x)))
  33 #define MD4_Go(x,y,z)   (bitselect ((x), (y), ((x) ^ (z))))
  34 #endif
  35
  36 #ifdef IS_GENERIC
  37 #define MD4_F(x,y,z)    (((x) & (y)) | ((~(x)) & (z)))
  38 #define MD4_G(x,y,z)    (((x) & (y)) | ((x) & (z)) | ((y) & (z)))
  39 #define MD4_H(x,y,z)    ((x) ^ (y) ^ (z))
  40 #define MD4_H1(x,y,z)   (MD4_H((x), (y), (z)))
  41 #define MD4_H2(x,y,z)   (MD4_H((x), (y), (z)))
  42 #define MD4_Fo(x,y,z)   (MD4_F((x), (y), (z)))
  43 #define MD4_Go(x,y,z)   (MD4_G((x), (y), (z)))
  44 #endif
  45
  46 #define MD4_STEP(f,a,b,c,d,x,K,s)   \
  47 {                                   \
  48   a += K;                           \
  49   a += x;                           \
  50   a += f (b, c, d);                 \
  51   a  = rotl32 (a, s);               \
  52 }
  53
  54 #define MD4_STEP0(f,a,b,c,d,K,s)    \
  55 {                                   \
  56   a += K;                           \
  57   a += f (b, c, d);                 \
  58   a  = rotl32 (a, s);               \
  59 }
  60
  61 #endif
  62
  63 #if defined _MD5_ || defined _MD5H_ || defined _SAPB_ || defined _OLDOFFICE01_ || defined _WPA_ || defined _MD5_SHA1_ || defined _SHA1_MD5_ || defined _NETNTLMV2_ || defined _KRB5PA_ || defined _PBKDF2_MD5_
  64
  65 #ifdef IS_NV
  66 #if CUDA_ARCH >= 500
  67 #define MD5_F(x,y,z)    lut3_ca ((x), (y), (z))
  68 #define MD5_G(x,y,z)    lut3_e4 ((x), (y), (z))
  69 #define MD5_H(x,y,z)    lut3_96 ((x), (y), (z))
  70 #define MD5_H1(x,y,z)   lut3_96 ((x), (y), (z))
  71 #define MD5_H2(x,y,z)   lut3_96 ((x), (y), (z))
  72 #define MD5_I(x,y,z)    lut3_39 ((x), (y), (z))
  73 #else
  74 #define MD5_F(x,y,z)    ((z) ^ ((x) & ((y) ^ (z))))
  75 #define MD5_G(x,y,z)    ((y) ^ ((z) & ((x) ^ (y))))
  76 #define MD5_H(x,y,z)    ((x) ^ (y) ^ (z))
  77 #define MD5_H1(x,y,z)   ((tmp2 = (x) ^ (y)) ^ (z))
  78 #define MD5_H2(x,y,z)   ((x) ^ tmp2)
  79 #define MD5_I(x,y,z)    ((y) ^ ((x) | ~(z)))
  80 #endif
  81 #define MD5_Fo(x,y,z)   (MD5_F((x), (y), (z)))
  82 #define MD5_Go(x,y,z)   (MD5_G((x), (y), (z)))
  83 #endif
  84
  85 #ifdef IS_AMD
  86 #define MD5_F(x,y,z)    ((z) ^ ((x) & ((y) ^ (z))))
  87 #define MD5_G(x,y,z)    ((y) ^ ((z) & ((x) ^ (y))))
  88 #define MD5_H(x,y,z)    ((x) ^ (y) ^ (z))
  89 #define MD5_H1(x,y,z)   ((tmp2 = (x) ^ (y)) ^ (z))
  90 #define MD5_H2(x,y,z)   ((x) ^ tmp2)
  91 #define MD5_I(x,y,z)    (bitselect (0xffffffffU, (x), (z)) ^ (y))
  92 #define MD5_Fo(x,y,z)   (bitselect ((z), (y), (x)))
  93 #define MD5_Go(x,y,z)   (bitselect ((y), (x), (z)))
  94 #endif
  95
  96 #ifdef IS_GENERIC
  97 #define MD5_F(x,y,z)    ((z) ^ ((x) & ((y) ^ (z))))
  98 #define MD5_G(x,y,z)    ((y) ^ ((z) & ((x) ^ (y))))
  99 #define MD5_H(x,y,z)    ((x) ^ (y) ^ (z))
 100 #define MD5_H1(x,y,z)   (MD5_H((x), (y), (z)))
 101 #define MD5_H2(x,y,z)   (MD5_H((x), (y), (z)))
 102 #define MD5_I(x,y,z)    ((y) ^ ((x) | ~(z)))
 103 #define MD5_Fo(x,y,z)   (MD5_F((x), (y), (z)))
 104 #define MD5_Go(x,y,z)   (MD5_G((x), (y), (z)))
 105 #endif
 106
 107 #define MD5_STEP(f,a,b,c,d,x,K,s)   \
 108 {                                   \
 109   a += K;                           \
 110   a += x;                           \
 111   a += f (b, c, d);                 \
 112   a  = rotl32 (a, s);               \
 113   a += b;                           \
 114 }
 115
 116 #define MD5_STEP0(f,a,b,c,d,K,s)    \
 117 {                                   \
 118   a += K;                           \
 119   a += f (b, c, d);                 \
 120   a  = rotl32 (a, s);               \
 121   a += b;                           \
 122 }
 123 #endif
 124
 125 #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_
 126
 127 #ifdef IS_NV
 128 #if CUDA_ARCH >= 500
 129 #define SHA1_F0(x,y,z)  lut3_ca ((x), (y), (z))
 130 #define SHA1_F1(x,y,z)  lut3_96 ((x), (y), (z))
 131 #define SHA1_F2(x,y,z)  lut3_e8 ((x), (y), (z))
 132 #else
 133 #define SHA1_F0(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 134 #define SHA1_F1(x,y,z)  ((x) ^ (y) ^ (z))
 135 #define SHA1_F2(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 136 #endif
 137 #define SHA1_F0o(x,y,z) (SHA1_F0 ((x), (y), (z)))
 138 #define SHA1_F2o(x,y,z) (SHA1_F2 ((x), (y), (z)))
 139 #endif
 140
 141 #ifdef IS_AMD
 142 #define SHA1_F0(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 143 #define SHA1_F1(x,y,z)  ((x) ^ (y) ^ (z))
 144 #define SHA1_F2(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 145 #define SHA1_F0o(x,y,z) (bitselect ((z), (y), (x)))
 146 #define SHA1_F2o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 147 #endif
 148
 149 #ifdef IS_GENERIC
 150 #define SHA1_F0(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 151 #define SHA1_F1(x,y,z)  ((x) ^ (y) ^ (z))
 152 #define SHA1_F2(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 153 #define SHA1_F0o(x,y,z) (SHA1_F0 ((x), (y), (z)))
 154 #define SHA1_F2o(x,y,z) (SHA1_F2 ((x), (y), (z)))
 155 #endif
 156
 157 #define SHA1_STEP(f,a,b,c,d,e,x)    \
 158 {                                   \
 159   e += K;                           \
 160   e += x;                           \
 161   e += f (b, c, d);                 \
 162   e += rotl32 (a,  5u);             \
 163   b  = rotl32 (b, 30u);             \
 164 }
 165
 166 #define SHA1_STEP0(f,a,b,c,d,e,x)   \
 167 {                                   \
 168   e += K;                           \
 169   e += f (b, c, d);                 \
 170   e += rotl32 (a,  5u);             \
 171   b  = rotl32 (b, 30u);             \
 172 }
 173
 174 #define SHA1_STEPX(f,a,b,c,d,e,x)   \
 175 {                                   \
 176   e += x;                           \
 177   e += f (b, c, d);                 \
 178   e += rotl32 (a,  5u);             \
 179   b  = rotl32 (b, 30u);             \
 180 }
 181
 182 #define SHA1_STEP_PE(f,a,b,c,d,e,x) \
 183 {                                   \
 184   e += x;                           \
 185   e += f (b, c, d);                 \
 186   e += rotl32 (a,  5u);             \
 187 }
 188
 189 #define SHA1_STEP_PB(f,a,b,c,d,e,x) \
 190 {                                   \
 191   e += K;                           \
 192   b  = rotl32 (b, 30u);             \
 193 }
 194 #endif
 195
 196 #if defined _SHA256_ || defined _PDF17L8_ || defined _SEVEN_ZIP_ || defined _ANDROIDFDE_ || defined _CLOUDKEY_ || defined _SCRYPT_ || defined _PBKDF2_SHA256_  || defined _SHA256_SHA1_ || defined _MS_DRSR_ || defined _ANDROIDFDE_SAMSUNG_
 197
 198 #define SHIFT_RIGHT_32(x,n) ((x) >> (n))
 199
 200 #define SHA256_S0(x) (rotl32 ((x), 25u) ^ rotl32 ((x), 14u) ^ SHIFT_RIGHT_32 ((x),  3u))
 201 #define SHA256_S1(x) (rotl32 ((x), 15u) ^ rotl32 ((x), 13u) ^ SHIFT_RIGHT_32 ((x), 10u))
 202 #define SHA256_S2(x) (rotl32 ((x), 30u) ^ rotl32 ((x), 19u) ^ rotl32 ((x), 10u))
 203 #define SHA256_S3(x) (rotl32 ((x), 26u) ^ rotl32 ((x), 21u) ^ rotl32 ((x),  7u))
 204
 205 #ifdef IS_NV
 206 #if CUDA_ARCH >= 500
 207 #define SHA256_F0(x,y,z)  lut3_e8 ((x), (y), (z))
 208 #define SHA256_F1(x,y,z)  lut3_ca ((x), (y), (z))
 209 #else
 210 #define SHA256_F0(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 211 #define SHA256_F1(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 212 #endif
 213 #define SHA256_F0o(x,y,z) (SHA256_F0 ((x), (y), (z)))
 214 #define SHA256_F1o(x,y,z) (SHA256_F1 ((x), (y), (z)))
 215 #endif
 216
 217 #ifdef IS_AMD
 218 #define SHA256_F0(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 219 #define SHA256_F1(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 220 #define SHA256_F0o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 221 #define SHA256_F1o(x,y,z) (bitselect ((z), (y), (x)))
 222 #endif
 223
 224 #ifdef IS_GENERIC
 225 #define SHA256_F0(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 226 #define SHA256_F1(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 227 #define SHA256_F0o(x,y,z) (SHA256_F0 ((x), (y), (z)))
 228 #define SHA256_F1o(x,y,z) (SHA256_F1 ((x), (y), (z)))
 229 #endif
 230
 231 #define SHA256_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 232 {                                               \
 233   h += K;                                       \
 234   h += x;                                       \
 235   h += SHA256_S3 (e);                           \
 236   h += F1 (e,f,g);                              \
 237   d += h;                                       \
 238   h += SHA256_S2 (a);                           \
 239   h += F0 (a,b,c);                              \
 240 }
 241
 242 #define SHA256_EXPAND(x,y,z,w) (SHA256_S1 (x) + y + SHA256_S0 (z) + w)
 243
 244 #endif
 245
 246 #if defined _SHA384_ || defined _PDF17L8_
 247
 248 #define SHIFT_RIGHT_64(x,n) ((x) >> (n))
 249
 250 #define SHA384_S0(x) (rotr64 ((x), 28) ^ rotr64 ((x), 34) ^ rotr64 ((x), 39))
 251 #define SHA384_S1(x) (rotr64 ((x), 14) ^ rotr64 ((x), 18) ^ rotr64 ((x), 41))
 252 #define SHA384_S2(x) (rotr64 ((x),  1) ^ rotr64 ((x),  8) ^ SHIFT_RIGHT_64 ((x), 7))
 253 #define SHA384_S3(x) (rotr64 ((x), 19) ^ rotr64 ((x), 61) ^ SHIFT_RIGHT_64 ((x), 6))
 254
 255 #define SHA384_F0(x,y,z) ((z) ^ ((x) & ((y) ^ (z))))
 256 #define SHA384_F1(x,y,z) (((x) & (y)) | ((z) & ((x) ^ (y))))
 257
 258 #ifdef IS_NV
 259 #define SHA384_F0o(x,y,z) (SHA384_F0 ((x), (y), (z)))
 260 #define SHA384_F1o(x,y,z) (SHA384_F1 ((x), (y), (z)))
 261 #endif
 262
 263 #ifdef IS_AMD
 264 #define SHA384_F0o(x,y,z) (bitselect ((z), (y), (x)))
 265 #define SHA384_F1o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 266 #endif
 267
 268 #ifdef IS_GENERIC
 269 #define SHA384_F0o(x,y,z) (SHA384_F0 ((x), (y), (z)))
 270 #define SHA384_F1o(x,y,z) (SHA384_F1 ((x), (y), (z)))
 271 #endif
 272
 273 #define SHA384_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 274 {                                               \
 275   u64   temp0;                                  \
 276   temp0  = K;                                   \
 277   temp0 += x;                                   \
 278   temp0 += h;                                   \
 279   temp0 += SHA384_S1 (e);                       \
 280   temp0 += F0 (e, f, g);                        \
 281   d     += temp0;                               \
 282   h      = SHA384_S0 (a);                       \
 283   h     += F1 (a, b, c);                        \
 284   h     += temp0;                               \
 285 }
 286
 287 #define SHA384_EXPAND(x,y,z,w) (SHA384_S3 (x) + y + SHA384_S2 (z) + w)
 288 #endif
 289
 290 #if defined _SHA512_ || defined _CLOUDKEY_ || defined _OFFICE2013_ || defined _PDF17L8_ || defined _PBKDF2_SHA512_
 291
 292 #define SHIFT_RIGHT_64(x,n) ((x) >> (n))
 293
 294 #define SHA512_S0(x) (rotr64 ((x), 28) ^ rotr64 ((x), 34) ^ rotr64 ((x), 39))
 295 #define SHA512_S1(x) (rotr64 ((x), 14) ^ rotr64 ((x), 18) ^ rotr64 ((x), 41))
 296 #define SHA512_S2(x) (rotr64 ((x),  1) ^ rotr64 ((x),  8) ^ SHIFT_RIGHT_64 ((x), 7))
 297 #define SHA512_S3(x) (rotr64 ((x), 19) ^ rotr64 ((x), 61) ^ SHIFT_RIGHT_64 ((x), 6))
 298
 299 #define SHA512_F0(x,y,z) ((z) ^ ((x) & ((y) ^ (z))))
 300 #define SHA512_F1(x,y,z) (((x) & (y)) | ((z) & ((x) ^ (y))))
 301
 302 #ifdef IS_NV
 303 #define SHA512_F0o(x,y,z) (SHA512_F0 ((x), (y), (z)))
 304 #define SHA512_F1o(x,y,z) (SHA512_F1 ((x), (y), (z)))
 305 #endif
 306
 307 #ifdef IS_AMD
 308 #define SHA512_F0o(x,y,z) (bitselect ((z), (y), (x)))
 309 #define SHA512_F1o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 310 #endif
 311
 312 #ifdef IS_GENERIC
 313 #define SHA512_F0o(x,y,z) (SHA512_F0 ((x), (y), (z)))
 314 #define SHA512_F1o(x,y,z) (SHA512_F1 ((x), (y), (z)))
 315 #endif
 316
 317 #define SHA512_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 318 {                                               \
 319   u64   temp0;                                  \
 320   temp0  = K;                                   \
 321   temp0 += x;                                   \
 322   temp0 += h;                                   \
 323   temp0 += SHA512_S1 (e);                       \
 324   temp0 += F0 (e, f, g);                        \
 325   d     += temp0;                               \
 326   h      = SHA512_S0 (a);                       \
 327   h     += F1 (a, b, c);                        \
 328   h     += temp0;                               \
 329 }
 330
 331 #define SHA512_EXPAND(x,y,z,w) (SHA512_S3 (x) + y + SHA512_S2 (z) + w)
 332
 333 #define SHA512_S2_WO(x) (rotate ((x), 64- 1ull) ^ rotate ((x), 64- 8ull) ^ SHIFT_RIGHT_64 ((x), 7))
 334 #define SHA512_S3_WO(x) (rotate ((x), 64-19ull) ^ rotate ((x), 64-61ull) ^ SHIFT_RIGHT_64 ((x), 6))
 335
 336 #define SHA512_EXPAND_WO(x,y,z,w) (SHA512_S3_WO (x) + y + SHA512_S2_WO (z) + w)
 337 #endif
 338
 339 #ifdef _RIPEMD160_
 340
 341 #ifdef IS_NV
 342 #if CUDA_ARCH >= 500
 343 #define RIPEMD160_F(x,y,z)    lut3_96 ((x), (y), (z))
 344 #define RIPEMD160_G(x,y,z)    lut3_ca ((x), (y), (z))
 345 #define RIPEMD160_H(x,y,z)    lut3_59 ((x), (y), (z))
 346 #define RIPEMD160_I(x,y,z)    lut3_e4 ((x), (y), (z))
 347 #define RIPEMD160_J(x,y,z)    lut3_2d ((x), (y), (z))
 348 #else
 349 #define RIPEMD160_F(x,y,z)    ((x) ^ (y) ^ (z))
 350 #define RIPEMD160_G(x,y,z)    ((z) ^ ((x) & ((y) ^ (z)))) /* x ? y : z */
 351 #define RIPEMD160_H(x,y,z)    (((x) | ~(y)) ^ (z))
 352 #define RIPEMD160_I(x,y,z)    ((y) ^ ((z) & ((x) ^ (y)))) /* z ? x : y */
 353 #define RIPEMD160_J(x,y,z)    ((x) ^ ((y) | ~(z)))
 354 #endif
 355 #define RIPEMD160_Go(x,y,z)   (RIPEMD160_G ((x), (y), (z)))
 356 #define RIPEMD160_Io(x,y,z)   (RIPEMD160_I ((x), (y), (z)))
 357 #endif
 358
 359 #ifdef IS_AMD
 360 #define RIPEMD160_F(x,y,z)    ((x) ^ (y) ^ (z))
 361 #define RIPEMD160_G(x,y,z)    ((z) ^ ((x) & ((y) ^ (z)))) /* x ? y : z */
 362 #define RIPEMD160_H(x,y,z)    (((x) | ~(y)) ^ (z))
 363 #define RIPEMD160_I(x,y,z)    ((y) ^ ((z) & ((x) ^ (y)))) /* z ? x : y */
 364 #define RIPEMD160_J(x,y,z)    ((x) ^ ((y) | ~(z)))
 365 #define RIPEMD160_Go(x,y,z)   (bitselect ((z), (y), (x)))
 366 #define RIPEMD160_Io(x,y,z)   (bitselect ((y), (x), (z)))
 367 #endif
 368
 369 #ifdef IS_GENERIC
 370 #define RIPEMD160_F(x,y,z)    ((x) ^ (y) ^ (z))
 371 #define RIPEMD160_G(x,y,z)    ((z) ^ ((x) & ((y) ^ (z)))) /* x ? y : z */
 372 #define RIPEMD160_H(x,y,z)    (((x) | ~(y)) ^ (z))
 373 #define RIPEMD160_I(x,y,z)    ((y) ^ ((z) & ((x) ^ (y)))) /* z ? x : y */
 374 #define RIPEMD160_J(x,y,z)    ((x) ^ ((y) | ~(z)))
 375 #define RIPEMD160_Go(x,y,z)   (RIPEMD160_G ((x), (y), (z)))
 376 #define RIPEMD160_Io(x,y,z)   (RIPEMD160_I ((x), (y), (z)))
 377 #endif
 378
 379 #define RIPEMD160_STEP(f,a,b,c,d,e,x,K,s) \
 380 {                                         \
 381   a += K;                                 \
 382   a += x;                                 \
 383   a += f (b, c, d);                       \
 384   a  = rotl32 (a, s);                     \
 385   a += e;                                 \
 386   c  = rotl32 (c, 10u);                   \
 387 }
 388
 389 #define ROTATE_LEFT_WORKAROUND_BUG(a,n) ((a << n) | (a >> (32 - n)))
 390
 391 #define RIPEMD160_STEP_WORKAROUND_BUG(f,a,b,c,d,e,x,K,s)  \
 392 {                                         \
 393   a += K;                                 \
 394   a += x;                                 \
 395   a += f (b, c, d);                       \
 396   a  = ROTATE_LEFT_WORKAROUND_BUG (a, s); \
 397   a += e;                                 \
 398   c  = rotl32 (c, 10u);                   \
 399 }
 400
 401 #endif