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 #define MD4_STEP(f,a,b,c,d,x,K,s)   \
  37 {                                   \
  38   a += K;                           \
  39   a += x;                           \
  40   a += f (b, c, d);                 \
  41   a  = rotl32 (a, s);               \
  42 }
  43
  44 #define MD4_STEP0(f,a,b,c,d,K,s)    \
  45 {                                   \
  46   a += K;                           \
  47   a += f (b, c, d);                 \
  48   a  = rotl32 (a, s);               \
  49 }
  50
  51 #endif
  52
  53 #if defined _MD5_ || defined _MD5H_ || defined _SAPB_ || defined _OLDOFFICE01_ || defined _WPA_ || defined _MD5_SHA1_ || defined _SHA1_MD5_ || defined _NETNTLMV2_ || defined _KRB5PA_ || defined _PBKDF2_MD5_
  54
  55 #ifdef IS_NV
  56 #if CUDA_ARCH >= 500
  57 #define MD5_F(x,y,z)    lut3_ca ((x), (y), (z))
  58 #define MD5_G(x,y,z)    lut3_e4 ((x), (y), (z))
  59 #define MD5_H(x,y,z)    lut3_96 ((x), (y), (z))
  60 #define MD5_H1(x,y,z)   lut3_96 ((x), (y), (z))
  61 #define MD5_H2(x,y,z)   lut3_96 ((x), (y), (z))
  62 #define MD5_I(x,y,z)    lut3_39 ((x), (y), (z))
  63 #else
  64 #define MD5_F(x,y,z)    ((z) ^ ((x) & ((y) ^ (z))))
  65 #define MD5_G(x,y,z)    ((y) ^ ((z) & ((x) ^ (y))))
  66 #define MD5_H(x,y,z)    ((x) ^ (y) ^ (z))
  67 #define MD5_H1(x,y,z)   ((tmp2 = (x) ^ (y)) ^ (z))
  68 #define MD5_H2(x,y,z)   ((x) ^ tmp2)
  69 #define MD5_I(x,y,z)    ((y) ^ ((x) | ~(z)))
  70 #endif
  71 #define MD5_Fo(x,y,z)   (MD5_F((x), (y), (z)))
  72 #define MD5_Go(x,y,z)   (MD5_G((x), (y), (z)))
  73 #endif
  74
  75 #ifdef IS_AMD
  76 #define MD5_F(x,y,z)    ((z) ^ ((x) & ((y) ^ (z))))
  77 #define MD5_G(x,y,z)    ((y) ^ ((z) & ((x) ^ (y))))
  78 #define MD5_H(x,y,z)    ((x) ^ (y) ^ (z))
  79 #define MD5_H1(x,y,z)   ((tmp2 = (x) ^ (y)) ^ (z))
  80 #define MD5_H2(x,y,z)   ((x) ^ tmp2)
  81 #define MD5_I(x,y,z)    (bitselect (0xffffffffU, (x), (z)) ^ (y))
  82 #define MD5_Fo(x,y,z)   (bitselect ((z), (y), (x)))
  83 #define MD5_Go(x,y,z)   (bitselect ((y), (x), (z)))
  84 #endif
  85
  86 #define MD5_STEP(f,a,b,c,d,x,K,s)   \
  87 {                                   \
  88   a += K;                           \
  89   a += x;                           \
  90   a += f (b, c, d);                 \
  91   a  = rotl32 (a, s);               \
  92   a += b;                           \
  93 }
  94
  95 #define MD5_STEP0(f,a,b,c,d,K,s)    \
  96 {                                   \
  97   a += K;                           \
  98   a += f (b, c, d);                 \
  99   a  = rotl32 (a, s);               \
 100   a += b;                           \
 101 }
 102 #endif
 103
 104 #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_
 105
 106 /**
 107  * SHA1 Functions
 108  */
 109
 110 #ifdef IS_NV
 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 #define SHA1_F0o(x,y,z) (SHA1_F0 ((x), (y), (z)))
 121 #define SHA1_F2o(x,y,z) (SHA1_F2 ((x), (y), (z)))
 122 #endif
 123
 124 #ifdef IS_AMD
 125 #define SHA1_F0(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 126 #define SHA1_F1(x,y,z)  ((x) ^ (y) ^ (z))
 127 #define SHA1_F2(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 128 #define SHA1_F0o(x,y,z) (bitselect ((z), (y), (x)))
 129 #define SHA1_F2o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 130 #endif
 131
 132 #define SHA1_STEP(f,a,b,c,d,e,x)    \
 133 {                                   \
 134   e += K;                           \
 135   e += x;                           \
 136   e += f (b, c, d);                 \
 137   e += rotl32 (a,  5u);             \
 138   b  = rotl32 (b, 30u);             \
 139 }
 140
 141 #define SHA1_STEP0(f,a,b,c,d,e,x)   \
 142 {                                   \
 143   e += K;                           \
 144   e += f (b, c, d);                 \
 145   e += rotl32 (a,  5u);             \
 146   b  = rotl32 (b, 30u);             \
 147 }
 148
 149 #define SHA1_STEPX(f,a,b,c,d,e,x)   \
 150 {                                   \
 151   e += x;                           \
 152   e += f (b, c, d);                 \
 153   e += rotl32 (a,  5u);             \
 154   b  = rotl32 (b, 30u);             \
 155 }
 156
 157 #define SHA1_STEP_PE(f,a,b,c,d,e,x) \
 158 {                                   \
 159   e += x;                           \
 160   e += f (b, c, d);                 \
 161   e += rotl32 (a,  5u);             \
 162 }
 163
 164 #define SHA1_STEP_PB(f,a,b,c,d,e,x) \
 165 {                                   \
 166   e += K;                           \
 167   b  = rotl32 (b, 30u);             \
 168 }
 169 #endif
 170
 171 #if defined _SHA256_ || defined _PDF17L8_ || defined _SEVEN_ZIP_ || defined _ANDROIDFDE_ || defined _CLOUDKEY_ || defined _SCRYPT_ || defined _PBKDF2_SHA256_  || defined _SHA256_SHA1_ || defined _MS_DRSR_
 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 #ifdef IS_NV
 181 #if CUDA_ARCH >= 500
 182 #define SHA256_F0(x,y,z)  lut3_e8 ((x), (y), (z))
 183 #define SHA256_F1(x,y,z)  lut3_ca ((x), (y), (z))
 184 #else
 185 #define SHA256_F0(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 186 #define SHA256_F1(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 187 #endif
 188 #define SHA256_F0o(x,y,z) (SHA256_F0 ((x), (y), (z)))
 189 #define SHA256_F1o(x,y,z) (SHA256_F1 ((x), (y), (z)))
 190 #endif
 191
 192 #ifdef IS_AMD
 193 #define SHA256_F0(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 194 #define SHA256_F1(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 195 #define SHA256_F0o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 196 #define SHA256_F1o(x,y,z) (bitselect ((z), (y), (x)))
 197 #endif
 198
 199 #define SHA256_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 200 {                                               \
 201   h += K;                                       \
 202   h += x;                                       \
 203   h += SHA256_S3 (e);                           \
 204   h += F1 (e,f,g);                              \
 205   d += h;                                       \
 206   h += SHA256_S2 (a);                           \
 207   h += F0 (a,b,c);                              \
 208 }
 209
 210 #define SHA256_EXPAND(x,y,z,w) (SHA256_S1 (x) + y + SHA256_S0 (z) + w)
 211
 212 #endif
 213
 214 #if defined _SHA384_ || defined _PDF17L8_
 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_NV
 227 #define SHA384_F0o(x,y,z) (SHA384_F0 ((x), (y), (z)))
 228 #define SHA384_F1o(x,y,z) (SHA384_F1 ((x), (y), (z)))
 229 #endif
 230
 231 #ifdef IS_AMD
 232 #define SHA384_F0o(x,y,z) (bitselect ((z), (y), (x)))
 233 #define SHA384_F1o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 234 #endif
 235
 236 #define SHA384_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 237 {                                               \
 238   u64   temp0;                                  \
 239   temp0  = K;                                   \
 240   temp0 += x;                                   \
 241   temp0 += h;                                   \
 242   temp0 += SHA384_S1 (e);                       \
 243   temp0 += F0 (e, f, g);                        \
 244   d     += temp0;                               \
 245   h      = SHA384_S0 (a);                       \
 246   h     += F1 (a, b, c);                        \
 247   h     += temp0;                               \
 248 }
 249
 250 #define SHA384_EXPAND(x,y,z,w) (SHA384_S3 (x) + y + SHA384_S2 (z) + w)
 251 #endif
 252
 253 #if defined _SHA512_ || defined _CLOUDKEY_ || defined _OFFICE2013_ || defined _PDF17L8_ || defined _PBKDF2_SHA512_
 254 /**
 255  * SHA512 Functions
 256  */
 257
 258 #define SHIFT_RIGHT_64(x,n) ((x) >> (n))
 259
 260 #define SHA512_S0(x) (rotr64 ((x), 28) ^ rotr64 ((x), 34) ^ rotr64 ((x), 39))
 261 #define SHA512_S1(x) (rotr64 ((x), 14) ^ rotr64 ((x), 18) ^ rotr64 ((x), 41))
 262 #define SHA512_S2(x) (rotr64 ((x),  1) ^ rotr64 ((x),  8) ^ SHIFT_RIGHT_64 ((x), 7))
 263 #define SHA512_S3(x) (rotr64 ((x), 19) ^ rotr64 ((x), 61) ^ SHIFT_RIGHT_64 ((x), 6))
 264
 265 #define SHA512_F0(x,y,z) ((z) ^ ((x) & ((y) ^ (z))))
 266 #define SHA512_F1(x,y,z) (((x) & (y)) | ((z) & ((x) ^ (y))))
 267
 268 #ifdef IS_NV
 269 #define SHA512_F0o(x,y,z) (SHA512_F0 ((x), (y), (z)))
 270 #define SHA512_F1o(x,y,z) (SHA512_F1 ((x), (y), (z)))
 271 #endif
 272
 273 #ifdef IS_AMD
 274 #define SHA512_F0o(x,y,z) (bitselect ((z), (y), (x)))
 275 #define SHA512_F1o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 276 #endif
 277
 278 #define SHA512_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 279 {                                               \
 280   u64   temp0;                                  \
 281   temp0  = K;                                   \
 282   temp0 += x;                                   \
 283   temp0 += h;                                   \
 284   temp0 += SHA512_S1 (e);                       \
 285   temp0 += F0 (e, f, g);                        \
 286   d     += temp0;                               \
 287   h      = SHA512_S0 (a);                       \
 288   h     += F1 (a, b, c);                        \
 289   h     += temp0;                               \
 290 }
 291
 292 #define SHA512_EXPAND(x,y,z,w) (SHA512_S3 (x) + y + SHA512_S2 (z) + w)
 293
 294 #define SHA512_S2_WO(x) (rotate ((x), 64- 1ull) ^ rotate ((x), 64- 8ull) ^ SHIFT_RIGHT_64 ((x), 7))
 295 #define SHA512_S3_WO(x) (rotate ((x), 64-19ull) ^ rotate ((x), 64-61ull) ^ SHIFT_RIGHT_64 ((x), 6))
 296
 297 #define SHA512_EXPAND_WO(x,y,z,w) (SHA512_S3_WO (x) + y + SHA512_S2_WO (z) + w)
 298 #endif
 299
 300 #ifdef _RIPEMD160_
 301
 302 #ifdef IS_NV
 303 #if CUDA_ARCH >= 500
 304 #define RIPEMD160_F(x,y,z)    lut3_96 ((x), (y), (z))
 305 #define RIPEMD160_G(x,y,z)    lut3_ca ((x), (y), (z))
 306 #define RIPEMD160_H(x,y,z)    lut3_59 ((x), (y), (z))
 307 #define RIPEMD160_I(x,y,z)    lut3_e4 ((x), (y), (z))
 308 #define RIPEMD160_J(x,y,z)    lut3_2d ((x), (y), (z))
 309 #else
 310 #define RIPEMD160_F(x,y,z)    ((x) ^ (y) ^ (z))
 311 #define RIPEMD160_G(x,y,z)    ((z) ^ ((x) & ((y) ^ (z)))) /* x ? y : z */
 312 #define RIPEMD160_H(x,y,z)    (((x) | ~(y)) ^ (z))
 313 #define RIPEMD160_I(x,y,z)    ((y) ^ ((z) & ((x) ^ (y)))) /* z ? x : y */
 314 #define RIPEMD160_J(x,y,z)    ((x) ^ ((y) | ~(z)))
 315 #endif
 316 #define RIPEMD160_Go(x,y,z)   (RIPEMD160_G ((x), (y), (z)))
 317 #define RIPEMD160_Io(x,y,z)   (RIPEMD160_I ((x), (y), (z)))
 318 #endif
 319
 320 #ifdef IS_AMD
 321 #define RIPEMD160_F(x,y,z)    ((x) ^ (y) ^ (z))
 322 #define RIPEMD160_G(x,y,z)    ((z) ^ ((x) & ((y) ^ (z)))) /* x ? y : z */
 323 #define RIPEMD160_H(x,y,z)    (((x) | ~(y)) ^ (z))
 324 #define RIPEMD160_I(x,y,z)    ((y) ^ ((z) & ((x) ^ (y)))) /* z ? x : y */
 325 #define RIPEMD160_J(x,y,z)    ((x) ^ ((y) | ~(z)))
 326 #define RIPEMD160_Go(x,y,z)   (bitselect ((z), (y), (x)))
 327 #define RIPEMD160_Io(x,y,z)   (bitselect ((y), (x), (z)))
 328 #endif
 329
 330 #define RIPEMD160_STEP(f,a,b,c,d,e,x,K,s) \
 331 {                                         \
 332   a += K;                                 \
 333   a += x;                                 \
 334   a += f (b, c, d);                       \
 335   a  = rotl32 (a, s);                     \
 336   a += e;                                 \
 337   c  = rotl32 (c, 10u);                   \
 338 }
 339
 340 #define ROTATE_LEFT_WORKAROUND_BUG(a,n) ((a << n) | (a >> (32 - n)))
 341
 342 #define RIPEMD160_STEP_WORKAROUND_BUG(f,a,b,c,d,e,x,K,s)  \
 343 {                                         \
 344   a += K;                                 \
 345   a += x;                                 \
 346   a += f (b, c, d);                       \
 347   a  = ROTATE_LEFT_WORKAROUND_BUG (a, s); \
 348   a += e;                                 \
 349   c  = rotl32 (c, 10u);                   \
 350 }
 351
 352 #endif