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 #ifdef IS_NV
 107 #if CUDA_ARCH >= 500
 108 #define SHA1_F0(x,y,z)  lut3_ca ((x), (y), (z))
 109 #define SHA1_F1(x,y,z)  lut3_96 ((x), (y), (z))
 110 #define SHA1_F2(x,y,z)  lut3_e8 ((x), (y), (z))
 111 #else
 112 #define SHA1_F0(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 113 #define SHA1_F1(x,y,z)  ((x) ^ (y) ^ (z))
 114 #define SHA1_F2(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 115 #endif
 116 #define SHA1_F0o(x,y,z) (SHA1_F0 ((x), (y), (z)))
 117 #define SHA1_F2o(x,y,z) (SHA1_F2 ((x), (y), (z)))
 118 #endif
 119
 120 #ifdef IS_AMD
 121 #define SHA1_F0(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 122 #define SHA1_F1(x,y,z)  ((x) ^ (y) ^ (z))
 123 #define SHA1_F2(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 124 #define SHA1_F0o(x,y,z) (bitselect ((z), (y), (x)))
 125 #define SHA1_F2o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 126 #endif
 127
 128 #define SHA1_STEP(f,a,b,c,d,e,x)    \
 129 {                                   \
 130   e += K;                           \
 131   e += x;                           \
 132   e += f (b, c, d);                 \
 133   e += rotl32 (a,  5u);             \
 134   b  = rotl32 (b, 30u);             \
 135 }
 136
 137 #define SHA1_STEP0(f,a,b,c,d,e,x)   \
 138 {                                   \
 139   e += K;                           \
 140   e += f (b, c, d);                 \
 141   e += rotl32 (a,  5u);             \
 142   b  = rotl32 (b, 30u);             \
 143 }
 144
 145 #define SHA1_STEPX(f,a,b,c,d,e,x)   \
 146 {                                   \
 147   e += x;                           \
 148   e += f (b, c, d);                 \
 149   e += rotl32 (a,  5u);             \
 150   b  = rotl32 (b, 30u);             \
 151 }
 152
 153 #define SHA1_STEP_PE(f,a,b,c,d,e,x) \
 154 {                                   \
 155   e += x;                           \
 156   e += f (b, c, d);                 \
 157   e += rotl32 (a,  5u);             \
 158 }
 159
 160 #define SHA1_STEP_PB(f,a,b,c,d,e,x) \
 161 {                                   \
 162   e += K;                           \
 163   b  = rotl32 (b, 30u);             \
 164 }
 165 #endif
 166
 167 #if defined _SHA256_ || defined _PDF17L8_ || defined _SEVEN_ZIP_ || defined _ANDROIDFDE_ || defined _CLOUDKEY_ || defined _SCRYPT_ || defined _PBKDF2_SHA256_  || defined _SHA256_SHA1_ || defined _MS_DRSR_
 168
 169 #define SHIFT_RIGHT_32(x,n) ((x) >> (n))
 170
 171 #define SHA256_S0(x) (rotl32 ((x), 25u) ^ rotl32 ((x), 14u) ^ SHIFT_RIGHT_32 ((x),  3u))
 172 #define SHA256_S1(x) (rotl32 ((x), 15u) ^ rotl32 ((x), 13u) ^ SHIFT_RIGHT_32 ((x), 10u))
 173 #define SHA256_S2(x) (rotl32 ((x), 30u) ^ rotl32 ((x), 19u) ^ rotl32 ((x), 10u))
 174 #define SHA256_S3(x) (rotl32 ((x), 26u) ^ rotl32 ((x), 21u) ^ rotl32 ((x),  7u))
 175
 176 #ifdef IS_NV
 177 #if CUDA_ARCH >= 500
 178 #define SHA256_F0(x,y,z)  lut3_e8 ((x), (y), (z))
 179 #define SHA256_F1(x,y,z)  lut3_ca ((x), (y), (z))
 180 #else
 181 #define SHA256_F0(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 182 #define SHA256_F1(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 183 #endif
 184 #define SHA256_F0o(x,y,z) (SHA256_F0 ((x), (y), (z)))
 185 #define SHA256_F1o(x,y,z) (SHA256_F1 ((x), (y), (z)))
 186 #endif
 187
 188 #ifdef IS_AMD
 189 #define SHA256_F0(x,y,z)  (((x) & (y)) | ((z) & ((x) ^ (y))))
 190 #define SHA256_F1(x,y,z)  ((z) ^ ((x) & ((y) ^ (z))))
 191 #define SHA256_F0o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 192 #define SHA256_F1o(x,y,z) (bitselect ((z), (y), (x)))
 193 #endif
 194
 195 #define SHA256_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 196 {                                               \
 197   h += K;                                       \
 198   h += x;                                       \
 199   h += SHA256_S3 (e);                           \
 200   h += F1 (e,f,g);                              \
 201   d += h;                                       \
 202   h += SHA256_S2 (a);                           \
 203   h += F0 (a,b,c);                              \
 204 }
 205
 206 #define SHA256_EXPAND(x,y,z,w) (SHA256_S1 (x) + y + SHA256_S0 (z) + w)
 207
 208 #endif
 209
 210 #if defined _SHA384_ || defined _PDF17L8_
 211
 212 #define SHIFT_RIGHT_64(x,n) ((x) >> (n))
 213
 214 #define SHA384_S0(x) (rotr64 ((x), 28) ^ rotr64 ((x), 34) ^ rotr64 ((x), 39))
 215 #define SHA384_S1(x) (rotr64 ((x), 14) ^ rotr64 ((x), 18) ^ rotr64 ((x), 41))
 216 #define SHA384_S2(x) (rotr64 ((x),  1) ^ rotr64 ((x),  8) ^ SHIFT_RIGHT_64 ((x), 7))
 217 #define SHA384_S3(x) (rotr64 ((x), 19) ^ rotr64 ((x), 61) ^ SHIFT_RIGHT_64 ((x), 6))
 218
 219 #define SHA384_F0(x,y,z) ((z) ^ ((x) & ((y) ^ (z))))
 220 #define SHA384_F1(x,y,z) (((x) & (y)) | ((z) & ((x) ^ (y))))
 221
 222 #ifdef IS_NV
 223 #define SHA384_F0o(x,y,z) (SHA384_F0 ((x), (y), (z)))
 224 #define SHA384_F1o(x,y,z) (SHA384_F1 ((x), (y), (z)))
 225 #endif
 226
 227 #ifdef IS_AMD
 228 #define SHA384_F0o(x,y,z) (bitselect ((z), (y), (x)))
 229 #define SHA384_F1o(x,y,z) (bitselect ((x), (y), ((x) ^ (z))))
 230 #endif
 231
 232 #define SHA384_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 233 {                                               \
 234   u64   temp0;                                  \
 235   temp0  = K;                                   \
 236   temp0 += x;                                   \
 237   temp0 += h;                                   \
 238   temp0 += SHA384_S1 (e);                       \
 239   temp0 += F0 (e, f, g);                        \
 240   d     += temp0;                               \
 241   h      = SHA384_S0 (a);                       \
 242   h     += F1 (a, b, c);                        \
 243   h     += temp0;                               \
 244 }
 245
 246 #define SHA384_EXPAND(x,y,z,w) (SHA384_S3 (x) + y + SHA384_S2 (z) + w)
 247 #endif
 248
 249 #if defined _SHA512_ || defined _CLOUDKEY_ || defined _OFFICE2013_ || defined _PDF17L8_ || defined _PBKDF2_SHA512_
 250
 251 #define SHIFT_RIGHT_64(x,n) ((x) >> (n))
 252
 253 #define SHA512_S0(x) (rotr64 ((x), 28) ^ rotr64 ((x), 34) ^ rotr64 ((x), 39))
 254 #define SHA512_S1(x) (rotr64 ((x), 14) ^ rotr64 ((x), 18) ^ rotr64 ((x), 41))
 255 #define SHA512_S2(x) (rotr64 ((x),  1) ^ rotr64 ((x),  8) ^ SHIFT_RIGHT_64 ((x), 7))
 256 #define SHA512_S3(x) (rotr64 ((x), 19) ^ rotr64 ((x), 61) ^ SHIFT_RIGHT_64 ((x), 6))
 257
 258 #define SHA512_F0(x,y,z) ((z) ^ ((x) & ((y) ^ (z))))
 259 #define SHA512_F1(x,y,z) (((x) & (y)) | ((z) & ((x) ^ (y))))
 260
 261 #ifdef IS_NV
 262 #define SHA512_F0o(x,y,z) (SHA512_F0 ((x), (y), (z)))
 263 #define SHA512_F1o(x,y,z) (SHA512_F1 ((x), (y), (z)))
 264 #endif
 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 #endif
 270
 271 #define SHA512_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K)  \
 272 {                                               \
 273   u64   temp0;                                  \
 274   temp0  = K;                                   \
 275   temp0 += x;                                   \
 276   temp0 += h;                                   \
 277   temp0 += SHA512_S1 (e);                       \
 278   temp0 += F0 (e, f, g);                        \
 279   d     += temp0;                               \
 280   h      = SHA512_S0 (a);                       \
 281   h     += F1 (a, b, c);                        \
 282   h     += temp0;                               \
 283 }
 284
 285 #define SHA512_EXPAND(x,y,z,w) (SHA512_S3 (x) + y + SHA512_S2 (z) + w)
 286
 287 #define SHA512_S2_WO(x) (rotr64 ((x),  1) ^ rotr64 ((x),  8) ^ SHIFT_RIGHT_64 ((x), 7))
 288 #define SHA512_S3_WO(x) (rotr64 ((x), 19) ^ rotr64 ((x), 61) ^ SHIFT_RIGHT_64 ((x), 6))
 289
 290 #define SHA512_EXPAND_WO(x,y,z,w) (SHA512_S3_WO (x) + y + SHA512_S2_WO (z) + w)
 291 #endif
 292
 293 #ifdef _RIPEMD160_
 294
 295 #ifdef IS_NV
 296 #if CUDA_ARCH >= 500
 297 #define RIPEMD160_F(x,y,z)    lut3_96 ((x), (y), (z))
 298 #define RIPEMD160_G(x,y,z)    lut3_ca ((x), (y), (z))
 299 #define RIPEMD160_H(x,y,z)    lut3_59 ((x), (y), (z))
 300 #define RIPEMD160_I(x,y,z)    lut3_e4 ((x), (y), (z))
 301 #define RIPEMD160_J(x,y,z)    lut3_2d ((x), (y), (z))
 302 #else
 303 #define RIPEMD160_F(x,y,z)    ((x) ^ (y) ^ (z))
 304 #define RIPEMD160_G(x,y,z)    ((z) ^ ((x) & ((y) ^ (z)))) /* x ? y : z */
 305 #define RIPEMD160_H(x,y,z)    (((x) | ~(y)) ^ (z))
 306 #define RIPEMD160_I(x,y,z)    ((y) ^ ((z) & ((x) ^ (y)))) /* z ? x : y */
 307 #define RIPEMD160_J(x,y,z)    ((x) ^ ((y) | ~(z)))
 308 #endif
 309 #define RIPEMD160_Go(x,y,z)   (RIPEMD160_G ((x), (y), (z)))
 310 #define RIPEMD160_Io(x,y,z)   (RIPEMD160_I ((x), (y), (z)))
 311 #endif
 312
 313 #ifdef IS_AMD
 314 #define RIPEMD160_F(x,y,z)    ((x) ^ (y) ^ (z))
 315 #define RIPEMD160_G(x,y,z)    ((z) ^ ((x) & ((y) ^ (z)))) /* x ? y : z */
 316 #define RIPEMD160_H(x,y,z)    (((x) | ~(y)) ^ (z))
 317 #define RIPEMD160_I(x,y,z)    ((y) ^ ((z) & ((x) ^ (y)))) /* z ? x : y */
 318 #define RIPEMD160_J(x,y,z)    ((x) ^ ((y) | ~(z)))
 319 #define RIPEMD160_Go(x,y,z)   (bitselect ((z), (y), (x)))
 320 #define RIPEMD160_Io(x,y,z)   (bitselect ((y), (x), (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