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   #if CUDA_ARCH < 350
 356
 357   u32x t;
 358   u32x r;
 359
 360   #if VECT_SIZE == 2
 361
 362   asm ("shr.b32 %4, %2, %6;"
 363        "shr.b32 %5, %3, %6;"
 364        "mad.lo.u32 %0, %2, %7, %4;"
 365        "mad.lo.u32 %1, %3, %7, %5;"
 366       : "=r"(r.s0),
 367         "=r"(r.s1)
 368       : "r"(a.s0),
 369         "r"(a.s1),
 370         "r"(t.s0),
 371         "r"(t.s1),
 372         "r"(n),
 373         "r"(1 << (32 - n)));
 374
 375   #elif VECT_SIZE == 4
 376
 377   asm ("shr.b32 %8,  %4, %12;\n"
 378        "shr.b32 %9,  %5, %12;\n"
 379        "shr.b32 %10, %6, %12;\n"
 380        "shr.b32 %11, %7, %12;\n"
 381        "mad.lo.u32 %0, %4, %13, %8;\n"
 382        "mad.lo.u32 %1, %5, %13, %9;\n"
 383        "mad.lo.u32 %2, %6, %13, %10;\n"
 384        "mad.lo.u32 %3, %7, %13, %11;\n"
 385       : "=r"(r.s0),
 386         "=r"(r.s1),
 387         "=r"(r.s2),
 388         "=r"(r.s3)
 389       : "r"(a.s0),
 390         "r"(a.s1),
 391         "r"(a.s2),
 392         "r"(a.s3),
 393         "r"(t.s0),
 394         "r"(t.s1),
 395         "r"(t.s2),
 396         "r"(t.s3),
 397         "r"(n),
 398         "r"(1 << (32 - n)));
 399
 400   #else
 401
 402   r = rotate (a, n);
 403
 404   #endif
 405
 406   return r;
 407
 408   #else
 409
 410   return rotate (a, n);
 411
 412   #endif
 413 }
 414
 415 inline u32x rotl32 (const u32x a, const u32 n)
 416 {
 417   return rotr32 (a, 32 - n);
 418 }
 419
 420 inline u64x rotr64 (const u64x a, const u32 n)
 421 {
 422   return rotate (a, (u64) 64 - n);
 423 }
 424
 425 inline u64x rotl64 (const u64x a, const u32 n)
 426 {
 427   return rotate (a, (u64) n);
 428 }
 429
 430 inline u32x __byte_perm (const u32x a, const u32x b, const u32x c)
 431 {
 432   u32x r;
 433
 434   #if VECT_SIZE == 1
 435   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r)    : "r"(a),    "r"(b),    "r"(c)   );
 436   #endif
 437
 438   #if VECT_SIZE >= 2
 439   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s0) : "r"(a.s0), "r"(b.s0), "r"(c.s0));
 440   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s1) : "r"(a.s1), "r"(b.s1), "r"(c.s1));
 441   #endif
 442
 443   #if VECT_SIZE >= 4
 444   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s2) : "r"(a.s2), "r"(b.s2), "r"(c.s2));
 445   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s3) : "r"(a.s3), "r"(b.s3), "r"(c.s3));
 446   #endif
 447
 448   #if VECT_SIZE >= 8
 449   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s4) : "r"(a.s4), "r"(b.s4), "r"(c.s4));
 450   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s5) : "r"(a.s5), "r"(b.s5), "r"(c.s5));
 451   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s6) : "r"(a.s6), "r"(b.s6), "r"(c.s6));
 452   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s7) : "r"(a.s7), "r"(b.s7), "r"(c.s7));
 453   #endif
 454
 455   #if VECT_SIZE >= 16
 456   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s8) : "r"(a.s8), "r"(b.s8), "r"(c.s8));
 457   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s9) : "r"(a.s9), "r"(b.s9), "r"(c.s9));
 458   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sa) : "r"(a.sa), "r"(b.sa), "r"(c.sa));
 459   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sb) : "r"(a.sb), "r"(b.sb), "r"(c.sb));
 460   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sc) : "r"(a.sc), "r"(b.sc), "r"(c.sc));
 461   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sd) : "r"(a.sd), "r"(b.sd), "r"(c.sd));
 462   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.se) : "r"(a.se), "r"(b.se), "r"(c.se));
 463   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sf) : "r"(a.sf), "r"(b.sf), "r"(c.sf));
 464   #endif
 465
 466   return r;
 467 }
 468
 469 inline u32 __bfe (const u32 a, const u32 b, const u32 c)
 470 {
 471   u32 r;
 472
 473   asm ("bfe.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(c));
 474
 475   return r;
 476 }
 477
 478 #if CUDA_ARCH >= 350
 479 inline u32 amd_bytealign (const u32 a, const u32 b, const u32 c)
 480 {
 481   u32 r;
 482
 483   asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(r) : "r"(b), "r"(a), "r"((c & 3) * 8));
 484
 485   return r;
 486 }
 487 #else
 488 inline u32 amd_bytealign (const u32 a, const u32 b, const u32 c)
 489 {
 490   return __byte_perm_S (b, a, (0x76543210 >> ((c & 3) * 4)) & 0xffff);
 491 }
 492 #endif
 493
 494 #endif
 495
 496 #ifdef IS_GENERIC
 497 inline u32 swap32_S (const u32 v)
 498 {
 499   return (as_uint (as_uchar4 (v).s3210));
 500 }
 501
 502 inline u64 swap64_S (const u64 v)
 503 {
 504   return (as_ulong (as_uchar8 (v).s76543210));
 505 }
 506
 507 inline u32 rotr32_S (const u32 a, const u32 n)
 508 {
 509   return rotate (a, 32 - n);
 510 }
 511
 512 inline u32 rotl32_S (const u32 a, const u32 n)
 513 {
 514   return rotate (a, n);
 515 }
 516
 517 inline u64 rotr64_S (const u64 a, const u32 n)
 518 {
 519   return rotate (a, (u64) 64 - n);
 520 }
 521
 522 inline u64 rotl64_S (const u64 a, const u32 n)
 523 {
 524   return rotate (a, (u64) n);
 525 }
 526
 527 inline u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
 528 {
 529   const u64 tmp = ((((u64) a) << 32) | ((u64) b)) >> ((c & 3) * 8);
 530
 531   return (u32) (tmp);
 532 }
 533
 534 inline u32x swap32 (const u32x v)
 535 {
 536   return ((v >> 24) & 0x000000ff)
 537        | ((v >>  8) & 0x0000ff00)
 538        | ((v <<  8) & 0x00ff0000)
 539        | ((v << 24) & 0xff000000);
 540 }
 541
 542 inline u64x swap64 (const u64x v)
 543 {
 544   return ((v >> 56) & 0x00000000000000ff)
 545        | ((v >> 40) & 0x000000000000ff00)
 546        | ((v >> 24) & 0x0000000000ff0000)
 547        | ((v >>  8) & 0x00000000ff000000)
 548        | ((v <<  8) & 0x000000ff00000000)
 549        | ((v << 24) & 0x0000ff0000000000)
 550        | ((v << 40) & 0x00ff000000000000)
 551        | ((v << 56) & 0xff00000000000000);
 552 }
 553
 554 inline u32x rotr32 (const u32x a, const u32 n)
 555 {
 556   return rotate (a, 32 - n);
 557 }
 558
 559 inline u32x rotl32 (const u32x a, const u32 n)
 560 {
 561   return rotate (a, n);
 562 }
 563
 564 inline u64x rotr64 (const u64x a, const u32 n)
 565 {
 566   return rotate (a, (u64) 64 - n);
 567 }
 568
 569 inline u64x rotl64 (const u64x a, const u32 n)
 570 {
 571   return rotate (a, (u64) n);
 572 }
 573
 574 inline u32 __bfe (const u32 a, const u32 b, const u32 c)
 575 {
 576   #define BIT(x)      (1 << (x))
 577   #define BIT_MASK(x) (BIT (x) - 1)
 578   #define BFE(x,y,z)  (((x) >> (y)) & BIT_MASK (z))
 579
 580   return BFE (a, b, c);
 581 }
 582
 583 inline u32x amd_bytealign (const u32x a, const u32x b, const u32 c)
 584 {
 585   #if VECT_SIZE == 1
 586   const u64x tmp = ((((u64x) (a)) << 32) | ((u64x) (b))) >> ((c & 3) * 8);
 587
 588   return (u32x) (tmp);
 589   #endif
 590
 591   #if VECT_SIZE == 2
 592   const u64x tmp = ((((u64x) (a.s0, a.s1)) << 32) | ((u64x) (b.s0, b.s1))) >> ((c & 3) * 8);
 593
 594   return (u32x) (tmp.s0, tmp.s1);
 595   #endif
 596
 597   #if VECT_SIZE == 4
 598   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);
 599
 600   return (u32x) (tmp.s0, tmp.s1, tmp.s2, tmp.s3);
 601   #endif
 602
 603   #if VECT_SIZE == 8
 604   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);
 605
 606   return (u32x) (tmp.s0, tmp.s1, tmp.s2, tmp.s3, tmp.s4, tmp.s5, tmp.s6, tmp.s7);
 607   #endif
 608
 609   #if VECT_SIZE == 16
 610   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);
 611
 612   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);
 613   #endif
 614 }
 615 #endif
 616
 617 typedef struct
 618 {
 619   #if   defined _DES_
 620   u32  digest_buf[4];
 621   #elif defined _MD4_
 622   u32  digest_buf[4];
 623   #elif defined _MD5_
 624   u32  digest_buf[4];
 625   #elif defined _MD5H_
 626   u32  digest_buf[4];
 627   #elif defined _SHA1_
 628   u32  digest_buf[5];
 629   #elif defined _BCRYPT_
 630   u32  digest_buf[6];
 631   #elif defined _SHA256_
 632   u32  digest_buf[8];
 633   #elif defined _SHA384_
 634   u32  digest_buf[16];
 635   #elif defined _SHA512_
 636   u32  digest_buf[16];
 637   #elif defined _KECCAK_
 638   u32  digest_buf[50];
 639   #elif defined _RIPEMD160_
 640   u32  digest_buf[5];
 641   #elif defined _WHIRLPOOL_
 642   u32  digest_buf[16];
 643   #elif defined _GOST_
 644   u32  digest_buf[8];
 645   #elif defined _GOST2012_256_
 646   u32  digest_buf[8];
 647   #elif defined _GOST2012_512_
 648   u32  digest_buf[16];
 649   #elif defined _SAPB_
 650   u32  digest_buf[4];
 651   #elif defined _SAPG_
 652   u32  digest_buf[5];
 653   #elif defined _MYSQL323_
 654   u32  digest_buf[4];
 655   #elif defined _LOTUS5_
 656   u32  digest_buf[4];
 657   #elif defined _LOTUS6_
 658   u32  digest_buf[4];
 659   #elif defined _SCRYPT_
 660   u32  digest_buf[8];
 661   #elif defined _LOTUS8_
 662   u32  digest_buf[4];
 663   #elif defined _OFFICE2007_
 664   u32  digest_buf[4];
 665   #elif defined _OFFICE2010_
 666   u32  digest_buf[4];
 667   #elif defined _OFFICE2013_
 668   u32  digest_buf[4];
 669   #elif defined _OLDOFFICE01_
 670   u32  digest_buf[4];
 671   #elif defined _OLDOFFICE34_
 672   u32  digest_buf[4];
 673   #elif defined _SIPHASH_
 674   u32  digest_buf[4];
 675   #elif defined _PBKDF2_MD5_
 676   u32  digest_buf[32];
 677   #elif defined _PBKDF2_SHA1_
 678   u32  digest_buf[32];
 679   #elif defined _PBKDF2_SHA256_
 680   u32  digest_buf[32];
 681   #elif defined _PBKDF2_SHA512_
 682   u32  digest_buf[32];
 683   #elif defined _PDF17L8_
 684   u32  digest_buf[8];
 685   #elif defined _CRC32_
 686   u32  digest_buf[4];
 687   #elif defined _SEVEN_ZIP_
 688   u32  digest_buf[4];
 689   #elif defined _ANDROIDFDE_
 690   u32  digest_buf[4];
 691   #elif defined _DCC2_
 692   u32  digest_buf[4];
 693   #elif defined _WPA_
 694   u32  digest_buf[4];
 695   #elif defined _MD5_SHA1_
 696   u32  digest_buf[4];
 697   #elif defined _SHA1_MD5_
 698   u32  digest_buf[5];
 699   #elif defined _NETNTLMV2_
 700   u32  digest_buf[4];
 701   #elif defined _KRB5PA_
 702   u32  digest_buf[4];
 703   #elif defined _CLOUDKEY_
 704   u32  digest_buf[8];
 705   #elif defined _SCRYPT_
 706   u32  digest_buf[4];
 707   #elif defined _PSAFE2_
 708   u32  digest_buf[5];
 709   #elif defined _LOTUS8_
 710   u32  digest_buf[4];
 711   #elif defined _RAR3_
 712   u32  digest_buf[4];
 713   #elif defined _SHA256_SHA1_
 714   u32  digest_buf[8];
 715   #elif defined _MS_DRSR_
 716   u32  digest_buf[8];
 717   #elif defined _ANDROIDFDE_SAMSUNG_
 718   u32  digest_buf[8];
 719   #elif defined _RAR5_
 720   u32  digest_buf[4];
 721   #elif defined _KRB5TGS_
 722   u32  digest_buf[4];
 723   #elif defined _AXCRYPT_
 724   u32  digest_buf[4];
 725   #elif defined _KEEPASS_
 726   u32  digest_buf[4];
 727   #endif
 728
 729 } digest_t;
 730
 731 typedef struct
 732 {
 733   u32 salt_buf[16];
 734   u32 salt_buf_pc[8];
 735
 736   u32 salt_len;
 737   u32 salt_iter;
 738   u32 salt_sign[2];
 739
 740   u32 keccak_mdlen;
 741   u32 truecrypt_mdlen;
 742
 743   u32 digests_cnt;
 744   u32 digests_done;
 745
 746   u32 digests_offset;
 747
 748   u32 scrypt_N;
 749   u32 scrypt_r;
 750   u32 scrypt_p;
 751   u32 scrypt_tmto;
 752   u32 scrypt_phy;
 753
 754 } salt_t;
 755
 756 typedef struct
 757 {
 758   int V;
 759   int R;
 760   int P;
 761
 762   int enc_md;
 763
 764   u32 id_buf[8];
 765   u32 u_buf[32];
 766   u32 o_buf[32];
 767
 768   int id_len;
 769   int o_len;
 770   int u_len;
 771
 772   u32 rc4key[2];
 773   u32 rc4data[2];
 774
 775 } pdf_t;
 776
 777 typedef struct
 778 {
 779   u32 pke[25];
 780   u32 eapol[64];
 781   int eapol_size;
 782   int keyver;
 783   u8  orig_mac1[6];
 784   u8  orig_mac2[6];
 785   u8  orig_nonce1[32];
 786   u8  orig_nonce2[32];
 787
 788 } wpa_t;
 789
 790 typedef struct
 791 {
 792   u32 cry_master_buf[64];
 793   u32 ckey_buf[64];
 794   u32 public_key_buf[64];
 795
 796   u32 cry_master_len;
 797   u32 ckey_len;
 798   u32 public_key_len;
 799
 800 } bitcoin_wallet_t;
 801
 802 typedef struct
 803 {
 804   u32 salt_buf[30];
 805   u32 salt_len;
 806
 807   u32 esalt_buf[38];
 808   u32 esalt_len;
 809
 810 } sip_t;
 811
 812 typedef struct
 813 {
 814   u32 data[384];
 815
 816 } androidfde_t;
 817
 818 typedef struct
 819 {
 820   u32 nr_buf[16];
 821   u32 nr_len;
 822
 823   u32 msg_buf[128];
 824   u32 msg_len;
 825
 826 } ikepsk_t;
 827
 828 typedef struct
 829 {
 830   u32 user_len;
 831   u32 domain_len;
 832   u32 srvchall_len;
 833   u32 clichall_len;
 834
 835   u32 userdomain_buf[64];
 836   u32 chall_buf[256];
 837
 838 } netntlm_t;
 839
 840 typedef struct
 841 {
 842   u32 user[16];
 843   u32 realm[16];
 844   u32 salt[32];
 845   u32 timestamp[16];
 846   u32 checksum[4];
 847
 848 } krb5pa_t;
 849
 850 typedef struct
 851 {
 852   u32 account_info[512];
 853   u32 checksum[4];
 854   u32 edata2[2560];
 855   u32 edata2_len;
 856
 857 } krb5tgs_t;
 858
 859 typedef struct
 860 {
 861   u32 salt_buf[16];
 862   u32 data_buf[112];
 863   u32 keyfile_buf[16];
 864
 865 } tc_t;
 866
 867 typedef struct
 868 {
 869   u32 salt_buf[16];
 870
 871 } pbkdf2_md5_t;
 872
 873 typedef struct
 874 {
 875   u32 salt_buf[16];
 876
 877 } pbkdf2_sha1_t;
 878
 879 typedef struct
 880 {
 881   u32 salt_buf[16];
 882
 883 } pbkdf2_sha256_t;
 884
 885 typedef struct
 886 {
 887   u32 salt_buf[32];
 888
 889 } pbkdf2_sha512_t;
 890
 891 typedef struct
 892 {
 893   u32 salt_buf[128];
 894   u32 salt_len;
 895
 896 } rakp_t;
 897
 898 typedef struct
 899 {
 900   u32 data_len;
 901   u32 data_buf[512];
 902
 903 } cloudkey_t;
 904
 905 typedef struct
 906 {
 907   u32 encryptedVerifier[4];
 908   u32 encryptedVerifierHash[5];
 909
 910   u32 keySize;
 911
 912 } office2007_t;
 913
 914 typedef struct
 915 {
 916   u32 encryptedVerifier[4];
 917   u32 encryptedVerifierHash[8];
 918
 919 } office2010_t;
 920
 921 typedef struct
 922 {
 923   u32 encryptedVerifier[4];
 924   u32 encryptedVerifierHash[8];
 925
 926 } office2013_t;
 927
 928 typedef struct
 929 {
 930   u32 version;
 931   u32 encryptedVerifier[4];
 932   u32 encryptedVerifierHash[4];
 933   u32 rc4key[2];
 934
 935 } oldoffice01_t;
 936
 937 typedef struct
 938 {
 939   u32 version;
 940   u32 encryptedVerifier[4];
 941   u32 encryptedVerifierHash[5];
 942   u32 rc4key[2];
 943
 944 } oldoffice34_t;
 945
 946 typedef struct
 947 {
 948   u32 salt_buf[128];
 949   u32 salt_len;
 950
 951   u32 pc_digest[5];
 952   u32 pc_offset;
 953
 954 } pstoken_t;
 955
 956 typedef struct
 957 {
 958   u32 version;
 959   u32 algorithm;
 960
 961   /* key-file handling */
 962   u32 keyfile_len;
 963   u32 keyfile[8];
 964
 965   u32 final_random_seed[8];
 966   u32 transf_random_seed[8];
 967   u32 enc_iv[4];
 968   u32 contents_hash[8];
 969
 970   /* specific to version 1 */
 971   u32 contents_len;
 972   u32 contents[75000];
 973
 974   /* specific to version 2 */
 975   u32 expected_bytes[8];
 976
 977 } keepass_t;
 978
 979 typedef struct
 980 {
 981   u32 digest[4];
 982   u32 out[4];
 983
 984 } pdf14_tmp_t;
 985
 986 typedef struct
 987 {
 988   union
 989   {
 990     u32 dgst32[16];
 991     u64 dgst64[8];
 992   };
 993
 994   u32 dgst_len;
 995   u32 W_len;
 996
 997 } pdf17l8_tmp_t;
 998
 999 typedef struct
1000 {
1001   u32 digest_buf[4];
1002
1003 } phpass_tmp_t;
1004
1005 typedef struct
1006 {
1007   u32 digest_buf[4];
1008
1009 } md5crypt_tmp_t;
1010
1011 typedef struct
1012 {
1013   u32 alt_result[8];
1014
1015   u32 p_bytes[4];
1016   u32 s_bytes[4];
1017
1018 } sha256crypt_tmp_t;
1019
1020 typedef struct
1021 {
1022   u64 l_alt_result[8];
1023
1024   u64 l_p_bytes[2];
1025   u64 l_s_bytes[2];
1026
1027 } sha512crypt_tmp_t;
1028
1029 typedef struct
1030 {
1031   u32 ipad[5];
1032   u32 opad[5];
1033
1034   u32 dgst[10];
1035   u32 out[10];
1036
1037 } wpa_tmp_t;
1038
1039 typedef struct
1040 {
1041   u64 dgst[8];
1042
1043 } bitcoin_wallet_tmp_t;
1044
1045 typedef struct
1046 {
1047   u32 ipad[5];
1048   u32 opad[5];
1049
1050   u32 dgst[5];
1051   u32 out[4];
1052
1053 } dcc2_tmp_t;
1054
1055 typedef struct
1056 {
1057   u32 E[18];
1058
1059   u32 P[18];
1060
1061   u32 S0[256];
1062   u32 S1[256];
1063   u32 S2[256];
1064   u32 S3[256];
1065
1066 } bcrypt_tmp_t;
1067
1068 typedef struct
1069 {
1070   u32 digest[2];
1071
1072   u32 P[18];
1073
1074   u32 S0[256];
1075   u32 S1[256];
1076   u32 S2[256];
1077   u32 S3[256];
1078
1079 } pwsafe2_tmp_t;
1080
1081 typedef struct
1082 {
1083   u32 digest_buf[8];
1084
1085 } pwsafe3_tmp_t;
1086
1087 typedef struct
1088 {
1089   u32 digest_buf[5];
1090
1091 } androidpin_tmp_t;
1092
1093 typedef struct
1094 {
1095   u32 ipad[5];
1096   u32 opad[5];
1097
1098   u32 dgst[10];
1099   u32 out[10];
1100
1101 } androidfde_tmp_t;
1102
1103 typedef struct
1104 {
1105   u32 ipad[16];
1106   u32 opad[16];
1107
1108   u32 dgst[64];
1109   u32 out[64];
1110
1111 } tc_tmp_t;
1112
1113 typedef struct
1114 {
1115   u64 ipad[8];
1116   u64 opad[8];
1117
1118   u64 dgst[32];
1119   u64 out[32];
1120
1121 } tc64_tmp_t;
1122
1123 typedef struct
1124 {
1125   u32 ipad[4];
1126   u32 opad[4];
1127
1128   u32 dgst[32];
1129   u32 out[32];
1130
1131 } pbkdf2_md5_tmp_t;
1132
1133 typedef struct
1134 {
1135   u32 ipad[5];
1136   u32 opad[5];
1137
1138   u32 dgst[32];
1139   u32 out[32];
1140
1141 } pbkdf2_sha1_tmp_t;
1142
1143 typedef struct
1144 {
1145   u32 ipad[8];
1146   u32 opad[8];
1147
1148   u32 dgst[32];
1149   u32 out[32];
1150
1151 } pbkdf2_sha256_tmp_t;
1152
1153 typedef struct
1154 {
1155   u64 ipad[8];
1156   u64 opad[8];
1157
1158   u64 dgst[16];
1159   u64 out[16];
1160
1161 } pbkdf2_sha512_tmp_t;
1162
1163 typedef struct
1164 {
1165   u64 out[8];
1166
1167 } ecryptfs_tmp_t;
1168
1169 typedef struct
1170 {
1171   u64 ipad[8];
1172   u64 opad[8];
1173
1174   u64 dgst[16];
1175   u64 out[16];
1176
1177 } oraclet_tmp_t;
1178
1179 typedef struct
1180 {
1181   u32 ipad[5];
1182   u32 opad[5];
1183
1184   u32 dgst[5];
1185   u32 out[5];
1186
1187 } agilekey_tmp_t;
1188
1189 typedef struct
1190 {
1191   u32 ipad[5];
1192   u32 opad[5];
1193
1194   u32 dgst1[5];
1195   u32 out1[5];
1196
1197   u32 dgst2[5];
1198   u32 out2[5];
1199
1200 } mywallet_tmp_t;
1201
1202 typedef struct
1203 {
1204   u32 ipad[5];
1205   u32 opad[5];
1206
1207   u32 dgst[5];
1208   u32 out[5];
1209
1210 } sha1aix_tmp_t;
1211
1212 typedef struct
1213 {
1214   u32 ipad[8];
1215   u32 opad[8];
1216
1217   u32 dgst[8];
1218   u32 out[8];
1219
1220 } sha256aix_tmp_t;
1221
1222 typedef struct
1223 {
1224   u64 ipad[8];
1225   u64 opad[8];
1226
1227   u64 dgst[8];
1228   u64 out[8];
1229
1230 } sha512aix_tmp_t;
1231
1232 typedef struct
1233 {
1234   u32 ipad[8];
1235   u32 opad[8];
1236
1237   u32 dgst[8];
1238   u32 out[8];
1239
1240 } lastpass_tmp_t;
1241
1242 typedef struct
1243 {
1244   u64 digest_buf[8];
1245
1246 } drupal7_tmp_t;
1247
1248 typedef struct
1249 {
1250   u32 ipad[5];
1251   u32 opad[5];
1252
1253   u32 dgst[5];
1254   u32 out[5];
1255
1256 } lotus8_tmp_t;
1257
1258 typedef struct
1259 {
1260   u32 out[5];
1261
1262 } office2007_tmp_t;
1263
1264 typedef struct
1265 {
1266   u32 out[5];
1267
1268 } office2010_tmp_t;
1269
1270 typedef struct
1271 {
1272   u64 out[8];
1273
1274 } office2013_tmp_t;
1275
1276 typedef struct
1277 {
1278   u32 digest_buf[5];
1279
1280 } saph_sha1_tmp_t;
1281
1282 typedef struct
1283 {
1284   u32 block[16];
1285
1286   u32 dgst[8];
1287
1288   u32 block_len;
1289   u32 final_len;
1290
1291 } seven_zip_tmp_t;
1292
1293 typedef struct
1294 {
1295   u32 KEK[5];
1296
1297   u32 lsb[4];
1298   u32 cipher[4];
1299
1300 } axcrypt_tmp_t;
1301
1302 typedef struct
1303 {
1304   u32 tmp_digest[8];
1305
1306 } keepass_tmp_t;
1307
1308 typedef struct
1309 {
1310   u32 Kc[16];
1311   u32 Kd[16];
1312
1313   u32 iv[2];
1314
1315 } bsdicrypt_tmp_t;
1316
1317 typedef struct
1318 {
1319   u32 dgst[17][5];
1320
1321 } rar3_tmp_t;
1322
1323 typedef struct
1324 {
1325   u32 user[16];
1326
1327 } cram_md5_t;
1328
1329 typedef struct
1330 {
1331   u32 iv_buf[4];
1332   u32 iv_len;
1333
1334   u32 salt_buf[4];
1335   u32 salt_len;
1336
1337   u32 crc;
1338
1339   u32 data_buf[96];
1340   u32 data_len;
1341
1342   u32 unpack_size;
1343
1344 } seven_zip_t;
1345
1346 typedef struct
1347 {
1348   u32 key;
1349   u64 val;
1350
1351 } hcstat_table_t;
1352
1353 typedef struct
1354 {
1355   u32 cs_buf[0x100];
1356   u32 cs_len;
1357
1358 } cs_t;
1359
1360 typedef struct
1361 {
1362   u32 cmds[0x100];
1363
1364 } kernel_rule_t;
1365
1366 typedef struct
1367 {
1368   u32 gidvid;
1369   u32 il_pos;
1370
1371 } plain_t;
1372
1373 typedef struct
1374 {
1375   u32 i[16];
1376
1377   u32 pw_len;
1378
1379   u32 alignment_placeholder_1;
1380   u32 alignment_placeholder_2;
1381   u32 alignment_placeholder_3;
1382
1383 } pw_t;
1384
1385 typedef struct
1386 {
1387   u32 i;
1388
1389 } bf_t;
1390
1391 typedef struct
1392 {
1393   u32 i[8];
1394
1395   u32 pw_len;
1396
1397 } comb_t;
1398
1399 typedef struct
1400 {
1401   u32 b[32];
1402
1403 } bs_word_t;
1404
1405 typedef struct
1406 {
1407   uint4 P[64];
1408
1409 } scrypt_tmp_t;