OpenCL/types_ocl.c

   1 /**
   2  * Authors.....: Jens Steube <jens.steube@gmail.com>
   3  *               magnum <john.magnum@hushmail.com>
   4  *
   5  * License.....: MIT
   6  */
   7
   8 #define DEVICE_TYPE_CPU 2
   9 #define DEVICE_TYPE_GPU 4
  10
  11 typedef uchar  u8;
  12 typedef ushort u16;
  13 typedef uint   u32;
  14 typedef ulong  u64;
  15
  16 #ifndef NEW_SIMD_CODE
  17 #undef  VECT_SIZE
  18 #define VECT_SIZE 1
  19 #endif
  20
  21 #define CONCAT(a, b)         a##b
  22 #define VTYPE(type, width) CONCAT(type, width)
  23
  24 #if VECT_SIZE == 1
  25 typedef uchar   u8x;
  26 typedef ushort  u16x;
  27 typedef uint    u32x;
  28 typedef ulong   u64x;
  29 #else
  30 typedef VTYPE(uchar,  VECT_SIZE)  u8x;
  31 typedef VTYPE(ushort, VECT_SIZE) u16x;
  32 typedef VTYPE(uint,   VECT_SIZE) u32x;
  33 typedef VTYPE(ulong,  VECT_SIZE) u64x;
  34 #endif
  35
  36 inline u32 l32_from_64_S (u64 a)
  37 {
  38   const u32 r = (u32) (a);
  39
  40   return r;
  41 }
  42
  43 inline u32 h32_from_64_S (u64 a)
  44 {
  45   a >>= 32;
  46
  47   const u32 r = (u32) (a);
  48
  49   return r;
  50 }
  51
  52 inline u64 hl32_to_64_S (const u32 a, const u32 b)
  53 {
  54   return as_ulong ((uint2) (b, a));
  55 }
  56
  57 inline u32x l32_from_64 (u64x a)
  58 {
  59   u32x r;
  60
  61   #if VECT_SIZE == 1
  62   r    = (u32) a;
  63   #endif
  64
  65   #if VECT_SIZE >= 2
  66   r.s0 = (u32) a.s0;
  67   r.s1 = (u32) a.s1;
  68   #endif
  69
  70   #if VECT_SIZE >= 4
  71   r.s2 = (u32) a.s2;
  72   r.s3 = (u32) a.s3;
  73   #endif
  74
  75   #if VECT_SIZE >= 8
  76   r.s4 = (u32) a.s4;
  77   r.s5 = (u32) a.s5;
  78   r.s6 = (u32) a.s6;
  79   r.s7 = (u32) a.s7;
  80   #endif
  81
  82   #if VECT_SIZE >= 16
  83   r.s8 = (u32) a.s8;
  84   r.s9 = (u32) a.s9;
  85   r.sa = (u32) a.sa;
  86   r.sb = (u32) a.sb;
  87   r.sc = (u32) a.sc;
  88   r.sd = (u32) a.sd;
  89   r.se = (u32) a.se;
  90   r.sf = (u32) a.sf;
  91   #endif
  92
  93   return r;
  94 }
  95
  96 inline u32x h32_from_64 (u64x a)
  97 {
  98   a >>= 32;
  99
 100   u32x r;
 101
 102   #if VECT_SIZE == 1
 103   r    = (u32) a;
 104   #endif
 105
 106   #if VECT_SIZE >= 2
 107   r.s0 = (u32) a.s0;
 108   r.s1 = (u32) a.s1;
 109   #endif
 110
 111   #if VECT_SIZE >= 4
 112   r.s2 = (u32) a.s2;
 113   r.s3 = (u32) a.s3;
 114   #endif
 115
 116   #if VECT_SIZE >= 8
 117   r.s4 = (u32) a.s4;
 118   r.s5 = (u32) a.s5;
 119   r.s6 = (u32) a.s6;
 120   r.s7 = (u32) a.s7;
 121   #endif
 122
 123   #if VECT_SIZE >= 16
 124   r.s8 = (u32) a.s8;
 125   r.s9 = (u32) a.s9;
 126   r.sa = (u32) a.sa;
 127   r.sb = (u32) a.sb;
 128   r.sc = (u32) a.sc;
 129   r.sd = (u32) a.sd;
 130   r.se = (u32) a.se;
 131   r.sf = (u32) a.sf;
 132   #endif
 133
 134   return r;
 135 }
 136
 137 inline u64x hl32_to_64 (const u32x a, const u32x b)
 138 {
 139   u64x r;
 140
 141   #if VECT_SIZE == 1
 142   r    = as_ulong ((uint2) (b,    a));
 143   #endif
 144
 145   #if VECT_SIZE >= 2
 146   r.s0 = as_ulong ((uint2) (b.s0, a.s0));
 147   r.s1 = as_ulong ((uint2) (b.s1, a.s1));
 148   #endif
 149
 150   #if VECT_SIZE >= 4
 151   r.s2 = as_ulong ((uint2) (b.s2, a.s2));
 152   r.s3 = as_ulong ((uint2) (b.s3, a.s3));
 153   #endif
 154
 155   #if VECT_SIZE >= 8
 156   r.s4 = as_ulong ((uint2) (b.s4, a.s4));
 157   r.s5 = as_ulong ((uint2) (b.s5, a.s5));
 158   r.s6 = as_ulong ((uint2) (b.s6, a.s6));
 159   r.s7 = as_ulong ((uint2) (b.s7, a.s7));
 160   #endif
 161
 162   #if VECT_SIZE >= 16
 163   r.s8 = as_ulong ((uint2) (b.s8, a.s8));
 164   r.s9 = as_ulong ((uint2) (b.s9, a.s9));
 165   r.sa = as_ulong ((uint2) (b.sa, a.sa));
 166   r.sb = as_ulong ((uint2) (b.sb, a.sb));
 167   r.sc = as_ulong ((uint2) (b.sc, a.sc));
 168   r.sd = as_ulong ((uint2) (b.sd, a.sd));
 169   r.se = as_ulong ((uint2) (b.se, a.se));
 170   r.sf = as_ulong ((uint2) (b.sf, a.sf));
 171   #endif
 172
 173   return r;
 174 }
 175
 176 #ifdef IS_AMD
 177 inline u32 swap32_S (const u32 v)
 178 {
 179   return (as_uint (as_uchar4 (v).s3210));
 180 }
 181
 182 inline u64 swap64_S (const u64 v)
 183 {
 184   return (as_ulong (as_uchar8 (v).s76543210));
 185 }
 186
 187 inline u32 rotr32_S (const u32 a, const u32 n)
 188 {
 189   return rotate (a, 32 - n);
 190 }
 191
 192 inline u32 rotl32_S (const u32 a, const u32 n)
 193 {
 194   return rotate (a, n);
 195 }
 196
 197 inline u64 rotr64_S (const u64 a, const u32 n)
 198 {
 199   const u32 a0 = h32_from_64_S (a);
 200   const u32 a1 = l32_from_64_S (a);
 201
 202   const u32 t0 = (n >= 32) ? amd_bitalign (a0, a1, n - 32) : amd_bitalign (a1, a0, n);
 203   const u32 t1 = (n >= 32) ? amd_bitalign (a1, a0, n - 32) : amd_bitalign (a0, a1, n);
 204
 205   const u64 r = hl32_to_64_S (t0, t1);
 206
 207   return r;
 208 }
 209
 210 inline u64 rotl64_S (const u64 a, const u32 n)
 211 {
 212   return rotr64_S (a, 64 - n);
 213 }
 214
 215 inline u32x swap32 (const u32x v)
 216 {
 217   return ((v >> 24) & 0x000000ff)
 218        | ((v >>  8) & 0x0000ff00)
 219        | ((v <<  8) & 0x00ff0000)
 220        | ((v << 24) & 0xff000000);
 221 }
 222
 223 inline u64x swap64 (const u64x v)
 224 {
 225   return ((v >> 56) & 0x00000000000000ff)
 226        | ((v >> 40) & 0x000000000000ff00)
 227        | ((v >> 24) & 0x0000000000ff0000)
 228        | ((v >>  8) & 0x00000000ff000000)
 229        | ((v <<  8) & 0x000000ff00000000)
 230        | ((v << 24) & 0x0000ff0000000000)
 231        | ((v << 40) & 0x00ff000000000000)
 232        | ((v << 56) & 0xff00000000000000);
 233 }
 234
 235 inline u32x rotr32 (const u32x a, const u32 n)
 236 {
 237   return rotate (a, 32 - n);
 238 }
 239
 240 inline u32x rotl32 (const u32x a, const u32 n)
 241 {
 242   return rotate (a, n);
 243 }
 244
 245 inline u64x rotr64 (const u64x a, const u32 n)
 246 {
 247   const u32x a0 = h32_from_64 (a);
 248   const u32x a1 = l32_from_64 (a);
 249
 250   const u32x t0 = (n >= 32) ? amd_bitalign (a0, a1, n - 32) : amd_bitalign (a1, a0, n);
 251   const u32x t1 = (n >= 32) ? amd_bitalign (a1, a0, n - 32) : amd_bitalign (a0, a1, n);
 252
 253   const u64x r = hl32_to_64 (t0, t1);
 254
 255   return r;
 256 }
 257
 258 inline u64x rotl64 (const u64x a, const u32 n)
 259 {
 260   return rotr64 (a, 64 - n);
 261 }
 262
 263 inline u32 __bfe (const u32 a, const u32 b, const u32 c)
 264 {
 265   return amd_bfe (a, b, c);
 266 }
 267
 268 inline u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
 269 {
 270   return amd_bytealign (a, b, c);
 271 }
 272 #endif
 273
 274 #ifdef IS_NV
 275 inline u32 swap32_S (const u32 v)
 276 {
 277   u32 r;
 278
 279   asm ("prmt.b32 %0, %1, 0, 0x0123;" : "=r"(r) : "r"(v));
 280
 281   return r;
 282 }
 283
 284 inline u64 swap64_S (const u64 v)
 285 {
 286   u32 il;
 287   u32 ir;
 288
 289   asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(v));
 290
 291   u32 tl;
 292   u32 tr;
 293
 294   asm ("prmt.b32 %0, %1, 0, 0x0123;" : "=r"(tl) : "r"(il));
 295   asm ("prmt.b32 %0, %1, 0, 0x0123;" : "=r"(tr) : "r"(ir));
 296
 297   u64 r;
 298
 299   asm ("mov.b64 %0, {%1, %2};" : "=l"(r) : "r"(tr), "r"(tl));
 300
 301   return r;
 302 }
 303
 304 inline u32 rotr32_S (const u32 a, const u32 n)
 305 {
 306   return rotate (a, 32 - n);
 307 }
 308
 309 inline u32 rotl32_S (const u32 a, const u32 n)
 310 {
 311   return rotate (a, n);
 312 }
 313
 314 inline u64 rotr64_S (const u64 a, const u32 n)
 315 {
 316   return rotate (a, (u64) 64 - n);
 317 }
 318
 319 inline u64 rotl64_S (const u64 a, const u32 n)
 320 {
 321   return rotr64_S (a, 64 - n);
 322 }
 323
 324 inline u32 __byte_perm_S (const u32 a, const u32 b, const u32 c)
 325 {
 326   u32 r;
 327
 328   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(c));
 329
 330   return r;
 331 }
 332
 333 inline u32x swap32 (const u32x v)
 334 {
 335   return ((v >> 24) & 0x000000ff)
 336        | ((v >>  8) & 0x0000ff00)
 337        | ((v <<  8) & 0x00ff0000)
 338        | ((v << 24) & 0xff000000);
 339 }
 340
 341 inline u64x swap64 (const u64x v)
 342 {
 343   return ((v >> 56) & 0x00000000000000ff)
 344        | ((v >> 40) & 0x000000000000ff00)
 345        | ((v >> 24) & 0x0000000000ff0000)
 346        | ((v >>  8) & 0x00000000ff000000)
 347        | ((v <<  8) & 0x000000ff00000000)
 348        | ((v << 24) & 0x0000ff0000000000)
 349        | ((v << 40) & 0x00ff000000000000)
 350        | ((v << 56) & 0xff00000000000000);
 351 }
 352
 353 inline u32x rotr32 (const u32x a, const u32 n)
 354 {
 355   return rotate (a, 32 - n);
 356 }
 357
 358 inline u32x rotl32 (const u32x a, const u32 n)
 359 {
 360   return rotate (a, n);
 361 }
 362
 363 inline u64x rotr64 (const u64x a, const u32 n)
 364 {
 365   return rotate (a, (u64) 64 - n);
 366 }
 367
 368 inline u64x rotl64 (const u64x a, const u32 n)
 369 {
 370   return rotate (a, (u64) n);
 371 }
 372
 373 inline u32x __byte_perm (const u32x a, const u32x b, const u32x c)
 374 {
 375   u32x r;
 376
 377   #if VECT_SIZE == 1
 378   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r)    : "r"(a),    "r"(b),    "r"(c)   );
 379   #endif
 380
 381   #if VECT_SIZE >= 2
 382   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s0) : "r"(a.s0), "r"(b.s0), "r"(c.s0));
 383   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s1) : "r"(a.s1), "r"(b.s1), "r"(c.s1));
 384   #endif
 385
 386   #if VECT_SIZE >= 4
 387   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s2) : "r"(a.s2), "r"(b.s2), "r"(c.s2));
 388   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s3) : "r"(a.s3), "r"(b.s3), "r"(c.s3));
 389   #endif
 390
 391   #if VECT_SIZE >= 8
 392   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s4) : "r"(a.s4), "r"(b.s4), "r"(c.s4));
 393   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s5) : "r"(a.s5), "r"(b.s5), "r"(c.s5));
 394   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s6) : "r"(a.s6), "r"(b.s6), "r"(c.s6));
 395   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s7) : "r"(a.s7), "r"(b.s7), "r"(c.s7));
 396   #endif
 397
 398   #if VECT_SIZE >= 16
 399   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s8) : "r"(a.s8), "r"(b.s8), "r"(c.s8));
 400   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s9) : "r"(a.s9), "r"(b.s9), "r"(c.s9));
 401   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sa) : "r"(a.sa), "r"(b.sa), "r"(c.sa));
 402   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sb) : "r"(a.sb), "r"(b.sb), "r"(c.sb));
 403   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sc) : "r"(a.sc), "r"(b.sc), "r"(c.sc));
 404   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sd) : "r"(a.sd), "r"(b.sd), "r"(c.sd));
 405   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.se) : "r"(a.se), "r"(b.se), "r"(c.se));
 406   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sf) : "r"(a.sf), "r"(b.sf), "r"(c.sf));
 407   #endif
 408
 409   return r;
 410 }
 411
 412 inline u32 __bfe (const u32 a, const u32 b, const u32 c)
 413 {
 414   u32 r;
 415
 416   asm ("bfe.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(c));
 417
 418   return r;
 419 }
 420
 421 #if CUDA_ARCH >= 350
 422 inline u32 amd_bytealign (const u32 a, const u32 b, const u32 c)
 423 {
 424   u32 r;
 425
 426   asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(r) : "r"(b), "r"(a), "r"((c & 3) * 8));
 427
 428   return r;
 429 }
 430 #else
 431 inline u32 amd_bytealign (const u32 a, const u32 b, const u32 c)
 432 {
 433   return __byte_perm_S (b, a, (0x76543210 >> ((c & 3) * 4)) & 0xffff);
 434 }
 435 #endif
 436
 437 #endif
 438
 439 #ifdef IS_GENERIC
 440 inline u32 swap32_S (const u32 v)
 441 {
 442   return (as_uint (as_uchar4 (v).s3210));
 443 }
 444
 445 inline u64 swap64_S (const u64 v)
 446 {
 447   return (as_ulong (as_uchar8 (v).s76543210));
 448 }
 449
 450 inline u32 rotr32_S (const u32 a, const u32 n)
 451 {
 452   return rotate (a, 32 - n);
 453 }
 454
 455 inline u32 rotl32_S (const u32 a, const u32 n)
 456 {
 457   return rotate (a, n);
 458 }
 459
 460 inline u64 rotr64_S (const u64 a, const u32 n)
 461 {
 462   return rotate (a, (u64) 64 - n);
 463 }
 464
 465 inline u64 rotl64_S (const u64 a, const u32 n)
 466 {
 467   return rotate (a, (u64) n);
 468 }
 469
 470 inline u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
 471 {
 472   const u64 tmp = ((((u64) a) << 32) | ((u64) b)) >> ((c & 3) * 8);
 473
 474   return (u32) (tmp);
 475 }
 476
 477 inline u32x swap32 (const u32x v)
 478 {
 479   return ((v >> 24) & 0x000000ff)
 480        | ((v >>  8) & 0x0000ff00)
 481        | ((v <<  8) & 0x00ff0000)
 482        | ((v << 24) & 0xff000000);
 483 }
 484
 485 inline u64x swap64 (const u64x v)
 486 {
 487   return ((v >> 56) & 0x00000000000000ff)
 488        | ((v >> 40) & 0x000000000000ff00)
 489        | ((v >> 24) & 0x0000000000ff0000)
 490        | ((v >>  8) & 0x00000000ff000000)
 491        | ((v <<  8) & 0x000000ff00000000)
 492        | ((v << 24) & 0x0000ff0000000000)
 493        | ((v << 40) & 0x00ff000000000000)
 494        | ((v << 56) & 0xff00000000000000);
 495 }
 496
 497 inline u32x rotr32 (const u32x a, const u32 n)
 498 {
 499   return rotate (a, 32 - n);
 500 }
 501
 502 inline u32x rotl32 (const u32x a, const u32 n)
 503 {
 504   return rotate (a, n);
 505 }
 506
 507 inline u64x rotr64 (const u64x a, const u32 n)
 508 {
 509   return rotate (a, (u64) 64 - n);
 510 }
 511
 512 inline u64x rotl64 (const u64x a, const u32 n)
 513 {
 514   return rotate (a, (u64) n);
 515 }
 516
 517 inline u32 __bfe (const u32 a, const u32 b, const u32 c)
 518 {
 519   #define BIT(x)      (1 << (x))
 520   #define BIT_MASK(x) (BIT (x) - 1)
 521   #define BFE(x,y,z)  (((x) >> (y)) & BIT_MASK (z))
 522
 523   return BFE (a, b, c);
 524 }
 525
 526 inline u32x amd_bytealign (const u32x a, const u32x b, const u32 c)
 527 {
 528   #if VECT_SIZE == 1
 529   const u64x tmp = ((((u64x) (a)) << 32) | ((u64x) (b))) >> ((c & 3) * 8);
 530
 531   return (u32x) (tmp);
 532   #endif
 533
 534   #if VECT_SIZE == 2
 535   const u64x tmp = ((((u64x) (a.s0, a.s1)) << 32) | ((u64x) (b.s0, b.s1))) >> ((c & 3) * 8);
 536
 537   return (u32x) (tmp.s0, tmp.s1);
 538   #endif
 539
 540   #if VECT_SIZE == 4
 541   const u64x tmp = ((((u64x) (a.s0, a.s1, a.s2, a.s3)) << 32) | ((u64x) (b.s0, b.s1, b.s2, b.s3))) >> ((c & 3) * 8);
 542
 543   return (u32x) (tmp.s0, tmp.s1, tmp.s2, tmp.s3);
 544   #endif
 545
 546   #if VECT_SIZE == 8
 547   const u64x tmp = ((((u64x) (a.s0, a.s1, a.s2, a.s3, a.s4, a.s5, a.s6, a.s7)) << 32) | ((u64x) (b.s0, b.s1, b.s2, b.s3, b.s4, b.s5, b.s6, b.s7))) >> ((c & 3) * 8);
 548
 549   return (u32x) (tmp.s0, tmp.s1, tmp.s2, tmp.s3, tmp.s4, tmp.s5, tmp.s6, tmp.s7);
 550   #endif
 551
 552   #if VECT_SIZE == 16
 553   const u64x tmp = ((((u64x) (a.s0, a.s1, a.s2, a.s3, a.s4, a.s5, a.s6, a.s7, a.s8, a.s9, a.sa, a.sb, a.sc, a.sd, a.se, a.sf)) << 32) | ((u64x) (b.s0, b.s1, b.s2, b.s3, b.s4, b.s5, b.s6, b.s7, b.s8, b.s9, b.sa, b.sb, b.sc, b.sd, b.se, b.sf))) >> ((c & 3) * 8);
 554
 555   return (u32x) (tmp.s0, tmp.s1, tmp.s2, tmp.s3, tmp.s4, tmp.s5, tmp.s6, tmp.s7, tmp.s8, tmp.s9, tmp.sa, tmp.sb, tmp.sc, tmp.sd, tmp.se, tmp.sf);
 556   #endif
 557 }
 558 #endif
 559
 560 typedef struct
 561 {
 562   #if   defined _DES_
 563   u32  digest_buf[4];
 564   #elif defined _MD4_
 565   u32  digest_buf[4];
 566   #elif defined _MD5_
 567   u32  digest_buf[4];
 568   #elif defined _MD5H_
 569   u32  digest_buf[4];
 570   #elif defined _SHA1_
 571   u32  digest_buf[5];
 572   #elif defined _BCRYPT_
 573   u32  digest_buf[6];
 574   #elif defined _SHA256_
 575   u32  digest_buf[8];
 576   #elif defined _SHA384_
 577   u32  digest_buf[16];
 578   #elif defined _SHA512_
 579   u32  digest_buf[16];
 580   #elif defined _KECCAK_
 581   u32  digest_buf[50];
 582   #elif defined _RIPEMD160_
 583   u32  digest_buf[5];
 584   #elif defined _WHIRLPOOL_
 585   u32  digest_buf[16];
 586   #elif defined _GOST_
 587   u32  digest_buf[8];
 588   #elif defined _GOST2012_256_
 589   u32  digest_buf[8];
 590   #elif defined _GOST2012_512_
 591   u32  digest_buf[16];
 592   #elif defined _SAPB_
 593   u32  digest_buf[4];
 594   #elif defined _SAPG_
 595   u32  digest_buf[5];
 596   #elif defined _MYSQL323_
 597   u32  digest_buf[4];
 598   #elif defined _LOTUS5_
 599   u32  digest_buf[4];
 600   #elif defined _LOTUS6_
 601   u32  digest_buf[4];
 602   #elif defined _SCRYPT_
 603   u32  digest_buf[8];
 604   #elif defined _LOTUS8_
 605   u32  digest_buf[4];
 606   #elif defined _OFFICE2007_
 607   u32  digest_buf[4];
 608   #elif defined _OFFICE2010_
 609   u32  digest_buf[4];
 610   #elif defined _OFFICE2013_
 611   u32  digest_buf[4];
 612   #elif defined _OLDOFFICE01_
 613   u32  digest_buf[4];
 614   #elif defined _OLDOFFICE34_
 615   u32  digest_buf[4];
 616   #elif defined _SIPHASH_
 617   u32  digest_buf[4];
 618   #elif defined _PBKDF2_MD5_
 619   u32  digest_buf[32];
 620   #elif defined _PBKDF2_SHA1_
 621   u32  digest_buf[32];
 622   #elif defined _PBKDF2_SHA256_
 623   u32  digest_buf[32];
 624   #elif defined _PBKDF2_SHA512_
 625   u32  digest_buf[32];
 626   #elif defined _PDF17L8_
 627   u32  digest_buf[8];
 628   #elif defined _CRC32_
 629   u32  digest_buf[4];
 630   #elif defined _SEVEN_ZIP_
 631   u32  digest_buf[4];
 632   #elif defined _ANDROIDFDE_
 633   u32  digest_buf[4];
 634   #elif defined _DCC2_
 635   u32  digest_buf[4];
 636   #elif defined _WPA_
 637   u32  digest_buf[4];
 638   #elif defined _MD5_SHA1_
 639   u32  digest_buf[4];
 640   #elif defined _SHA1_MD5_
 641   u32  digest_buf[5];
 642   #elif defined _NETNTLMV2_
 643   u32  digest_buf[4];
 644   #elif defined _KRB5PA_
 645   u32  digest_buf[4];
 646   #elif defined _CLOUDKEY_
 647   u32  digest_buf[8];
 648   #elif defined _SCRYPT_
 649   u32  digest_buf[4];
 650   #elif defined _PSAFE2_
 651   u32  digest_buf[5];
 652   #elif defined _LOTUS8_
 653   u32  digest_buf[4];
 654   #elif defined _RAR3_
 655   u32  digest_buf[4];
 656   #elif defined _SHA256_SHA1_
 657   u32  digest_buf[8];
 658   #elif defined _MS_DRSR_
 659   u32  digest_buf[8];
 660   #elif defined _ANDROIDFDE_SAMSUNG_
 661   u32  digest_buf[8];
 662   #elif defined _RAR5_
 663   u32  digest_buf[4];
 664   #elif defined _KRB5TGS_
 665   u32  digest_buf[4];
 666   #elif defined _AXCRYPT_
 667   u32  digest_buf[4];
 668   #elif defined _KEEPASS_
 669   u32  digest_buf[4];
 670   #endif
 671
 672 } digest_t;
 673
 674 typedef struct
 675 {
 676   u32 salt_buf[16];
 677   u32 salt_buf_pc[8];
 678
 679   u32 salt_len;
 680   u32 salt_iter;
 681   u32 salt_sign[2];
 682
 683   u32 keccak_mdlen;
 684   u32 truecrypt_mdlen;
 685
 686   u32 digests_cnt;
 687   u32 digests_done;
 688
 689   u32 digests_offset;
 690
 691   u32 scrypt_N;
 692   u32 scrypt_r;
 693   u32 scrypt_p;
 694   u32 scrypt_tmto;
 695   u32 scrypt_phy;
 696
 697 } salt_t;
 698
 699 typedef struct
 700 {
 701   int V;
 702   int R;
 703   int P;
 704
 705   int enc_md;
 706
 707   u32 id_buf[8];
 708   u32 u_buf[32];
 709   u32 o_buf[32];
 710
 711   int id_len;
 712   int o_len;
 713   int u_len;
 714
 715   u32 rc4key[2];
 716   u32 rc4data[2];
 717
 718 } pdf_t;
 719
 720 typedef struct
 721 {
 722   u32 pke[25];
 723   u32 eapol[64];
 724   int eapol_size;
 725   int keyver;
 726   u8  orig_mac1[6];
 727   u8  orig_mac2[6];
 728   u8  orig_nonce1[32];
 729   u8  orig_nonce2[32];
 730
 731 } wpa_t;
 732
 733 typedef struct
 734 {
 735   u32 cry_master_buf[64];
 736   u32 ckey_buf[64];
 737   u32 public_key_buf[64];
 738
 739   u32 cry_master_len;
 740   u32 ckey_len;
 741   u32 public_key_len;
 742
 743 } bitcoin_wallet_t;
 744
 745 typedef struct
 746 {
 747   u32 salt_buf[30];
 748   u32 salt_len;
 749
 750   u32 esalt_buf[38];
 751   u32 esalt_len;
 752
 753 } sip_t;
 754
 755 typedef struct
 756 {
 757   u32 data[384];
 758
 759 } androidfde_t;
 760
 761 typedef struct
 762 {
 763   u32 nr_buf[16];
 764   u32 nr_len;
 765
 766   u32 msg_buf[128];
 767   u32 msg_len;
 768
 769 } ikepsk_t;
 770
 771 typedef struct
 772 {
 773   u32 user_len;
 774   u32 domain_len;
 775   u32 srvchall_len;
 776   u32 clichall_len;
 777
 778   u32 userdomain_buf[64];
 779   u32 chall_buf[256];
 780
 781 } netntlm_t;
 782
 783 typedef struct
 784 {
 785   u32 user[16];
 786   u32 realm[16];
 787   u32 salt[32];
 788   u32 timestamp[16];
 789   u32 checksum[4];
 790
 791 } krb5pa_t;
 792
 793 typedef struct
 794 {
 795   u32 account_info[512];
 796   u32 checksum[4];
 797   u32 edata2[2560];
 798   u32 edata2_len;
 799
 800 } krb5tgs_t;
 801
 802 typedef struct
 803 {
 804   u32 salt_buf[16];
 805   u32 data_buf[112];
 806   u32 keyfile_buf[16];
 807
 808 } tc_t;
 809
 810 typedef struct
 811 {
 812   u32 salt_buf[16];
 813
 814 } pbkdf2_md5_t;
 815
 816 typedef struct
 817 {
 818   u32 salt_buf[16];
 819
 820 } pbkdf2_sha1_t;
 821
 822 typedef struct
 823 {
 824   u32 salt_buf[16];
 825
 826 } pbkdf2_sha256_t;
 827
 828 typedef struct
 829 {
 830   u32 salt_buf[32];
 831
 832 } pbkdf2_sha512_t;
 833
 834 typedef struct
 835 {
 836   u32 salt_buf[128];
 837   u32 salt_len;
 838
 839 } rakp_t;
 840
 841 typedef struct
 842 {
 843   u32 data_len;
 844   u32 data_buf[512];
 845
 846 } cloudkey_t;
 847
 848 typedef struct
 849 {
 850   u32 encryptedVerifier[4];
 851   u32 encryptedVerifierHash[5];
 852
 853   u32 keySize;
 854
 855 } office2007_t;
 856
 857 typedef struct
 858 {
 859   u32 encryptedVerifier[4];
 860   u32 encryptedVerifierHash[8];
 861
 862 } office2010_t;
 863
 864 typedef struct
 865 {
 866   u32 encryptedVerifier[4];
 867   u32 encryptedVerifierHash[8];
 868
 869 } office2013_t;
 870
 871 typedef struct
 872 {
 873   u32 version;
 874   u32 encryptedVerifier[4];
 875   u32 encryptedVerifierHash[4];
 876   u32 rc4key[2];
 877
 878 } oldoffice01_t;
 879
 880 typedef struct
 881 {
 882   u32 version;
 883   u32 encryptedVerifier[4];
 884   u32 encryptedVerifierHash[5];
 885   u32 rc4key[2];
 886
 887 } oldoffice34_t;
 888
 889 typedef struct
 890 {
 891   u32 salt_buf[128];
 892   u32 salt_len;
 893
 894   u32 pc_digest[5];
 895   u32 pc_offset;
 896
 897 } pstoken_t;
 898
 899 typedef struct
 900 {
 901   u32 version;
 902   u32 algorithm;
 903
 904   /* key-file handling */
 905   u32 keyfile_len;
 906   u32 keyfile[8];
 907
 908   u32 final_random_seed[8];
 909   u32 transf_random_seed[8];
 910   u32 enc_iv[4];
 911   u32 contents_hash[8];
 912
 913   /* specific to version 1 */
 914   u32 contents_len;
 915   u32 contents[75000];
 916
 917   /* specific to version 2 */
 918   u32 expected_bytes[8];
 919
 920 } keepass_t;
 921
 922 typedef struct
 923 {
 924   u32 digest[4];
 925   u32 out[4];
 926
 927 } pdf14_tmp_t;
 928
 929 typedef struct
 930 {
 931   union
 932   {
 933     u32 dgst32[16];
 934     u64 dgst64[8];
 935   };
 936
 937   u32 dgst_len;
 938   u32 W_len;
 939
 940 } pdf17l8_tmp_t;
 941
 942 typedef struct
 943 {
 944   u32 digest_buf[4];
 945
 946 } phpass_tmp_t;
 947
 948 typedef struct
 949 {
 950   u32 digest_buf[4];
 951
 952 } md5crypt_tmp_t;
 953
 954 typedef struct
 955 {
 956   u32 alt_result[8];
 957
 958   u32 p_bytes[4];
 959   u32 s_bytes[4];
 960
 961 } sha256crypt_tmp_t;
 962
 963 typedef struct
 964 {
 965   u64 l_alt_result[8];
 966
 967   u64 l_p_bytes[2];
 968   u64 l_s_bytes[2];
 969
 970 } sha512crypt_tmp_t;
 971
 972 typedef struct
 973 {
 974   u32 ipad[5];
 975   u32 opad[5];
 976
 977   u32 dgst[10];
 978   u32 out[10];
 979
 980 } wpa_tmp_t;
 981
 982 typedef struct
 983 {
 984   u64 dgst[8];
 985
 986 } bitcoin_wallet_tmp_t;
 987
 988 typedef struct
 989 {
 990   u32 ipad[5];
 991   u32 opad[5];
 992
 993   u32 dgst[5];
 994   u32 out[4];
 995
 996 } dcc2_tmp_t;
 997
 998 typedef struct
 999 {
1000   u32 E[18];
1001
1002   u32 P[18];
1003
1004   u32 S0[256];
1005   u32 S1[256];
1006   u32 S2[256];
1007   u32 S3[256];
1008
1009 } bcrypt_tmp_t;
1010
1011 typedef struct
1012 {
1013   u32 digest[2];
1014
1015   u32 P[18];
1016
1017   u32 S0[256];
1018   u32 S1[256];
1019   u32 S2[256];
1020   u32 S3[256];
1021
1022 } pwsafe2_tmp_t;
1023
1024 typedef struct
1025 {
1026   u32 digest_buf[8];
1027
1028 } pwsafe3_tmp_t;
1029
1030 typedef struct
1031 {
1032   u32 digest_buf[5];
1033
1034 } androidpin_tmp_t;
1035
1036 typedef struct
1037 {
1038   u32 ipad[5];
1039   u32 opad[5];
1040
1041   u32 dgst[10];
1042   u32 out[10];
1043
1044 } androidfde_tmp_t;
1045
1046 typedef struct
1047 {
1048   u32 ipad[16];
1049   u32 opad[16];
1050
1051   u32 dgst[64];
1052   u32 out[64];
1053
1054 } tc_tmp_t;
1055
1056 typedef struct
1057 {
1058   u64 ipad[8];
1059   u64 opad[8];
1060
1061   u64 dgst[32];
1062   u64 out[32];
1063
1064 } tc64_tmp_t;
1065
1066 typedef struct
1067 {
1068   u32 ipad[4];
1069   u32 opad[4];
1070
1071   u32 dgst[32];
1072   u32 out[32];
1073
1074 } pbkdf2_md5_tmp_t;
1075
1076 typedef struct
1077 {
1078   u32 ipad[5];
1079   u32 opad[5];
1080
1081   u32 dgst[32];
1082   u32 out[32];
1083
1084 } pbkdf2_sha1_tmp_t;
1085
1086 typedef struct
1087 {
1088   u32 ipad[8];
1089   u32 opad[8];
1090
1091   u32 dgst[32];
1092   u32 out[32];
1093
1094 } pbkdf2_sha256_tmp_t;
1095
1096 typedef struct
1097 {
1098   u64 ipad[8];
1099   u64 opad[8];
1100
1101   u64 dgst[16];
1102   u64 out[16];
1103
1104 } pbkdf2_sha512_tmp_t;
1105
1106 typedef struct
1107 {
1108   u64 out[8];
1109
1110 } ecryptfs_tmp_t;
1111
1112 typedef struct
1113 {
1114   u64 ipad[8];
1115   u64 opad[8];
1116
1117   u64 dgst[16];
1118   u64 out[16];
1119
1120 } oraclet_tmp_t;
1121
1122 typedef struct
1123 {
1124   u32 ipad[5];
1125   u32 opad[5];
1126
1127   u32 dgst[5];
1128   u32 out[5];
1129
1130 } agilekey_tmp_t;
1131
1132 typedef struct
1133 {
1134   u32 ipad[5];
1135   u32 opad[5];
1136
1137   u32 dgst1[5];
1138   u32 out1[5];
1139
1140   u32 dgst2[5];
1141   u32 out2[5];
1142
1143 } mywallet_tmp_t;
1144
1145 typedef struct
1146 {
1147   u32 ipad[5];
1148   u32 opad[5];
1149
1150   u32 dgst[5];
1151   u32 out[5];
1152
1153 } sha1aix_tmp_t;
1154
1155 typedef struct
1156 {
1157   u32 ipad[8];
1158   u32 opad[8];
1159
1160   u32 dgst[8];
1161   u32 out[8];
1162
1163 } sha256aix_tmp_t;
1164
1165 typedef struct
1166 {
1167   u64 ipad[8];
1168   u64 opad[8];
1169
1170   u64 dgst[8];
1171   u64 out[8];
1172
1173 } sha512aix_tmp_t;
1174
1175 typedef struct
1176 {
1177   u32 ipad[8];
1178   u32 opad[8];
1179
1180   u32 dgst[8];
1181   u32 out[8];
1182
1183 } lastpass_tmp_t;
1184
1185 typedef struct
1186 {
1187   u64 digest_buf[8];
1188
1189 } drupal7_tmp_t;
1190
1191 typedef struct
1192 {
1193   u32 ipad[5];
1194   u32 opad[5];
1195
1196   u32 dgst[5];
1197   u32 out[5];
1198
1199 } lotus8_tmp_t;
1200
1201 typedef struct
1202 {
1203   u32 out[5];
1204
1205 } office2007_tmp_t;
1206
1207 typedef struct
1208 {
1209   u32 out[5];
1210
1211 } office2010_tmp_t;
1212
1213 typedef struct
1214 {
1215   u64 out[8];
1216
1217 } office2013_tmp_t;
1218
1219 typedef struct
1220 {
1221   u32 digest_buf[5];
1222
1223 } saph_sha1_tmp_t;
1224
1225 typedef struct
1226 {
1227   u32 block[16];
1228
1229   u32 dgst[8];
1230
1231   u32 block_len;
1232   u32 final_len;
1233
1234 } seven_zip_tmp_t;
1235
1236 typedef struct
1237 {
1238   u32 KEK[5];
1239
1240   u32 lsb[4];
1241   u32 cipher[4];
1242
1243 } axcrypt_tmp_t;
1244
1245 typedef struct
1246 {
1247   u32 tmp_digest[8];
1248
1249 } keepass_tmp_t;
1250
1251 typedef struct
1252 {
1253   u32 Kc[16];
1254   u32 Kd[16];
1255
1256   u32 iv[2];
1257
1258 } bsdicrypt_tmp_t;
1259
1260 typedef struct
1261 {
1262   u32 dgst[17][5];
1263
1264 } rar3_tmp_t;
1265
1266 typedef struct
1267 {
1268   u32 user[16];
1269
1270 } cram_md5_t;
1271
1272 typedef struct
1273 {
1274   u32 iv_buf[4];
1275   u32 iv_len;
1276
1277   u32 salt_buf[4];
1278   u32 salt_len;
1279
1280   u32 crc;
1281
1282   u32 data_buf[96];
1283   u32 data_len;
1284
1285   u32 unpack_size;
1286
1287 } seven_zip_t;
1288
1289 typedef struct
1290 {
1291   u32 key;
1292   u64 val;
1293
1294 } hcstat_table_t;
1295
1296 typedef struct
1297 {
1298   u32 cs_buf[0x100];
1299   u32 cs_len;
1300
1301 } cs_t;
1302
1303 typedef struct
1304 {
1305   u32 cmds[0x100];
1306
1307 } kernel_rule_t;
1308
1309 typedef struct
1310 {
1311   u32 gidvid;
1312   u32 il_pos;
1313
1314 } plain_t;
1315
1316 typedef struct
1317 {
1318   u32 i[16];
1319
1320   u32 pw_len;
1321
1322   u32 alignment_placeholder_1;
1323   u32 alignment_placeholder_2;
1324   u32 alignment_placeholder_3;
1325
1326 } pw_t;
1327
1328 typedef struct
1329 {
1330   u32 i;
1331
1332 } bf_t;
1333
1334 typedef struct
1335 {
1336   u32 i[8];
1337
1338   u32 pw_len;
1339
1340 } comb_t;
1341
1342 typedef struct
1343 {
1344   u32 b[32];
1345
1346 } bs_word_t;
1347
1348 typedef struct
1349 {
1350   uint4 P[64];
1351
1352 } scrypt_tmp_t;