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 // this one needs to die
  37 #define allx(r) r
  38
  39 static inline u32 l32_from_64_S (u64 a)
  40 {
  41   const u32 r = (u32) (a);
  42
  43   return r;
  44 }
  45
  46 static inline u32 h32_from_64_S (u64 a)
  47 {
  48   a >>= 32;
  49
  50   const u32 r = (u32) (a);
  51
  52   return r;
  53 }
  54
  55 static inline u64 hl32_to_64_S (const u32 a, const u32 b)
  56 {
  57   return as_ulong ((uint2) (b, a));
  58 }
  59
  60 static inline u32x l32_from_64 (u64x a)
  61 {
  62   u32x r;
  63
  64   #if VECT_SIZE == 1
  65   r    = (u32) a;
  66   #endif
  67
  68   #if VECT_SIZE >= 2
  69   r.s0 = (u32) a.s0;
  70   r.s1 = (u32) a.s1;
  71   #endif
  72
  73   #if VECT_SIZE >= 4
  74   r.s2 = (u32) a.s2;
  75   r.s3 = (u32) a.s3;
  76   #endif
  77
  78   #if VECT_SIZE >= 8
  79   r.s4 = (u32) a.s4;
  80   r.s5 = (u32) a.s5;
  81   r.s6 = (u32) a.s6;
  82   r.s7 = (u32) a.s7;
  83   #endif
  84
  85   #if VECT_SIZE >= 16
  86   r.s8 = (u32) a.s8;
  87   r.s9 = (u32) a.s9;
  88   r.sa = (u32) a.sa;
  89   r.sb = (u32) a.sb;
  90   r.sc = (u32) a.sc;
  91   r.sd = (u32) a.sd;
  92   r.se = (u32) a.se;
  93   r.sf = (u32) a.sf;
  94   #endif
  95
  96   return r;
  97 }
  98
  99 static inline u32x h32_from_64 (u64x a)
 100 {
 101   a >>= 32;
 102
 103   u32x r;
 104
 105   #if VECT_SIZE == 1
 106   r    = (u32) a;
 107   #endif
 108
 109   #if VECT_SIZE >= 2
 110   r.s0 = (u32) a.s0;
 111   r.s1 = (u32) a.s1;
 112   #endif
 113
 114   #if VECT_SIZE >= 4
 115   r.s2 = (u32) a.s2;
 116   r.s3 = (u32) a.s3;
 117   #endif
 118
 119   #if VECT_SIZE >= 8
 120   r.s4 = (u32) a.s4;
 121   r.s5 = (u32) a.s5;
 122   r.s6 = (u32) a.s6;
 123   r.s7 = (u32) a.s7;
 124   #endif
 125
 126   #if VECT_SIZE >= 16
 127   r.s8 = (u32) a.s8;
 128   r.s9 = (u32) a.s9;
 129   r.sa = (u32) a.sa;
 130   r.sb = (u32) a.sb;
 131   r.sc = (u32) a.sc;
 132   r.sd = (u32) a.sd;
 133   r.se = (u32) a.se;
 134   r.sf = (u32) a.sf;
 135   #endif
 136
 137   return r;
 138 }
 139
 140 static inline u64x hl32_to_64 (const u32x a, const u32x b)
 141 {
 142   u64x r;
 143
 144   #if VECT_SIZE == 1
 145   r    = as_ulong ((uint2) (b,    a));
 146   #endif
 147
 148   #if VECT_SIZE >= 2
 149   r.s0 = as_ulong ((uint2) (b.s0, a.s0));
 150   r.s1 = as_ulong ((uint2) (b.s1, a.s1));
 151   #endif
 152
 153   #if VECT_SIZE >= 4
 154   r.s2 = as_ulong ((uint2) (b.s2, a.s2));
 155   r.s3 = as_ulong ((uint2) (b.s3, a.s3));
 156   #endif
 157
 158   #if VECT_SIZE >= 8
 159   r.s4 = as_ulong ((uint2) (b.s4, a.s4));
 160   r.s5 = as_ulong ((uint2) (b.s5, a.s5));
 161   r.s6 = as_ulong ((uint2) (b.s6, a.s6));
 162   r.s7 = as_ulong ((uint2) (b.s7, a.s7));
 163   #endif
 164
 165   #if VECT_SIZE >= 16
 166   r.s8 = as_ulong ((uint2) (b.s8, a.s8));
 167   r.s9 = as_ulong ((uint2) (b.s9, a.s9));
 168   r.sa = as_ulong ((uint2) (b.sa, a.sa));
 169   r.sb = as_ulong ((uint2) (b.sb, a.sb));
 170   r.sc = as_ulong ((uint2) (b.sc, a.sc));
 171   r.sd = as_ulong ((uint2) (b.sd, a.sd));
 172   r.se = as_ulong ((uint2) (b.se, a.se));
 173   r.sf = as_ulong ((uint2) (b.sf, a.sf));
 174   #endif
 175
 176   return r;
 177 }
 178
 179 #ifdef IS_AMD
 180 static inline u32 swap32_S (const u32 v)
 181 {
 182   return (as_uint (as_uchar4 (v).s3210));
 183 }
 184
 185 static inline u64 swap64_S (const u64 v)
 186 {
 187   return (as_ulong (as_uchar8 (v).s76543210));
 188 }
 189
 190 static inline u32 rotr32_S (const u32 a, const u32 n)
 191 {
 192   return rotate (a, 32 - n);
 193 }
 194
 195 static inline u32 rotl32_S (const u32 a, const u32 n)
 196 {
 197   return rotate (a, n);
 198 }
 199
 200 static inline u64 rotr64_S (const u64 a, const u32 n)
 201 {
 202   #if DEVICE_TYPE == DEVICE_TYPE_CPU
 203
 204   const u64 r = rotate (a, (u64) 64 - n);
 205
 206   #else
 207
 208   const u32 a0 = h32_from_64_S (a);
 209   const u32 a1 = l32_from_64_S (a);
 210
 211   const u32 t0 = (n >= 32) ? amd_bitalign (a0, a1, n - 32) : amd_bitalign (a1, a0, n);
 212   const u32 t1 = (n >= 32) ? amd_bitalign (a1, a0, n - 32) : amd_bitalign (a0, a1, n);
 213
 214   const u64 r = hl32_to_64_S (t0, t1);
 215
 216   #endif
 217
 218   return r;
 219 }
 220
 221 static inline u64 rotl64_S (const u64 a, const u32 n)
 222 {
 223   return rotr64_S (a, 64 - n);
 224 }
 225
 226 static inline u32x swap32 (const u32x v)
 227 {
 228   return ((v >> 24) & 0x000000ff)
 229        | ((v >>  8) & 0x0000ff00)
 230        | ((v <<  8) & 0x00ff0000)
 231        | ((v << 24) & 0xff000000);
 232 }
 233
 234 static inline u64x swap64 (const u64x v)
 235 {
 236   return ((v >> 56) & 0x00000000000000ff)
 237        | ((v >> 40) & 0x000000000000ff00)
 238        | ((v >> 24) & 0x0000000000ff0000)
 239        | ((v >>  8) & 0x00000000ff000000)
 240        | ((v <<  8) & 0x000000ff00000000)
 241        | ((v << 24) & 0x0000ff0000000000)
 242        | ((v << 40) & 0x00ff000000000000)
 243        | ((v << 56) & 0xff00000000000000);
 244 }
 245
 246 static inline u32x rotr32 (const u32x a, const u32 n)
 247 {
 248   return rotate (a, 32 - n);
 249 }
 250
 251 static inline u32x rotl32 (const u32x a, const u32 n)
 252 {
 253   return rotate (a, n);
 254 }
 255
 256 static inline u64x rotr64 (const u64x a, const u32 n)
 257 {
 258   #if DEVICE_TYPE == DEVICE_TYPE_CPU
 259
 260   const u64x r = rotate (a, (u64) 64 - n);
 261
 262   #else
 263
 264   const u32x a0 = h32_from_64 (a);
 265   const u32x a1 = l32_from_64 (a);
 266
 267   const u32x t0 = (n >= 32) ? amd_bitalign (a0, a1, n - 32) : amd_bitalign (a1, a0, n);
 268   const u32x t1 = (n >= 32) ? amd_bitalign (a1, a0, n - 32) : amd_bitalign (a0, a1, n);
 269
 270   const u64x r = hl32_to_64 (t0, t1);
 271
 272   #endif
 273
 274   return r;
 275 }
 276
 277 static inline u64x rotl64 (const u64x a, const u32 n)
 278 {
 279   return rotr64 (a, 64 - n);
 280 }
 281
 282 static inline u32 __bfe (const u32 a, const u32 b, const u32 c)
 283 {
 284   return amd_bfe (a, b, c);
 285 }
 286
 287 static inline u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
 288 {
 289   return amd_bytealign (a, b, c);
 290 }
 291 #endif
 292
 293 #ifdef IS_NV
 294 static inline u32 swap32_S (const u32 v)
 295 {
 296   u32 r;
 297
 298   asm ("prmt.b32 %0, %1, 0, 0x0123;" : "=r"(r) : "r"(v));
 299
 300   return r;
 301 }
 302
 303 static inline u64 swap64_S (const u64 v)
 304 {
 305   u32 il;
 306   u32 ir;
 307
 308   asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(v));
 309
 310   u32 tl;
 311   u32 tr;
 312
 313   asm ("prmt.b32 %0, %1, 0, 0x0123;" : "=r"(tl) : "r"(il));
 314   asm ("prmt.b32 %0, %1, 0, 0x0123;" : "=r"(tr) : "r"(ir));
 315
 316   u64 r;
 317
 318   asm ("mov.b64 %0, {%1, %2};" : "=l"(r) : "r"(tr), "r"(tl));
 319
 320   return r;
 321 }
 322
 323 static inline u32 rotr32_S (const u32 a, const u32 n)
 324 {
 325   return rotate (a, 32 - n);
 326 }
 327
 328 static inline u32 rotl32_S (const u32 a, const u32 n)
 329 {
 330   return rotate (a, n);
 331 }
 332
 333 #if CUDA_ARCH >= 350
 334 static inline u64 rotr64_S (const u64 a, const u32 n)
 335 {
 336   u32 il;
 337   u32 ir;
 338
 339   asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a));
 340
 341   u32 tl;
 342   u32 tr;
 343
 344   if (n >= 32)
 345   {
 346     asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 347     asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 348   }
 349   else
 350   {
 351     asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 352     asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 353   }
 354
 355   u64 r;
 356
 357   asm ("mov.b64 %0, {%1, %2};" : "=l"(r) : "r"(tl), "r"(tr));
 358
 359   return r;
 360 }
 361 #else
 362 static inline u64 rotr64_S (const u64 a, const u32 n)
 363 {
 364   return rotate (a, (u64) 64 - n);
 365 }
 366 #endif
 367
 368 static inline u64 rotl64_S (const u64 a, const u32 n)
 369 {
 370   return rotr64_S (a, 64 - n);
 371 }
 372
 373 #if CUDA_ARCH >= 500
 374 static inline u32 lut3_2d_S (const u32 a, const u32 b, const u32 c)
 375 {
 376   u32 r;
 377
 378   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r) : "r" (a), "r" (b), "r" (c));
 379
 380   return r;
 381 }
 382
 383 static inline u32 lut3_39_S (const u32 a, const u32 b, const u32 c)
 384 {
 385   u32 r;
 386
 387   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r) : "r" (a), "r" (b), "r" (c));
 388
 389   return r;
 390 }
 391
 392 static inline u32 lut3_59_S (const u32 a, const u32 b, const u32 c)
 393 {
 394   u32 r;
 395
 396   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r) : "r" (a), "r" (b), "r" (c));
 397
 398   return r;
 399 }
 400
 401 static inline u32 lut3_96_S (const u32 a, const u32 b, const u32 c)
 402 {
 403   u32 r;
 404
 405   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r) : "r" (a), "r" (b), "r" (c));
 406
 407   return r;
 408 }
 409
 410 static inline u32 lut3_e4_S (const u32 a, const u32 b, const u32 c)
 411 {
 412   u32 r;
 413
 414   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r) : "r" (a), "r" (b), "r" (c));
 415
 416   return r;
 417 }
 418
 419 static inline u32 lut3_e8_S (const u32 a, const u32 b, const u32 c)
 420 {
 421   u32 r;
 422
 423   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r) : "r" (a), "r" (b), "r" (c));
 424
 425   return r;
 426 }
 427
 428 static inline u32 lut3_ca_S (const u32 a, const u32 b, const u32 c)
 429 {
 430   u32 r;
 431
 432   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r) : "r" (a), "r" (b), "r" (c));
 433
 434   return r;
 435 }
 436 #endif
 437
 438 static inline u32 __byte_perm_S (const u32 a, const u32 b, const u32 c)
 439 {
 440   u32 r;
 441
 442   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(c));
 443
 444   return r;
 445 }
 446
 447 static inline u32x swap32 (const u32x v)
 448 {
 449   return ((v >> 24) & 0x000000ff)
 450        | ((v >>  8) & 0x0000ff00)
 451        | ((v <<  8) & 0x00ff0000)
 452        | ((v << 24) & 0xff000000);
 453 }
 454
 455 static inline u64x swap64 (const u64x v)
 456 {
 457   return ((v >> 56) & 0x00000000000000ff)
 458        | ((v >> 40) & 0x000000000000ff00)
 459        | ((v >> 24) & 0x0000000000ff0000)
 460        | ((v >>  8) & 0x00000000ff000000)
 461        | ((v <<  8) & 0x000000ff00000000)
 462        | ((v << 24) & 0x0000ff0000000000)
 463        | ((v << 40) & 0x00ff000000000000)
 464        | ((v << 56) & 0xff00000000000000);
 465 }
 466
 467 static inline u32x rotr32 (const u32x a, const u32 n)
 468 {
 469   return rotate (a, 32 - n);
 470 }
 471
 472 static inline u32x rotl32 (const u32x a, const u32 n)
 473 {
 474   return rotate (a, n);
 475 }
 476
 477 #if CUDA_ARCH >= 350
 478 static inline u64x rotr64 (const u64x a, const u32 n)
 479 {
 480   u64x r;
 481
 482   u32 il;
 483   u32 ir;
 484   u32 tl;
 485   u32 tr;
 486
 487   #if VECT_SIZE == 1
 488
 489   asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a));
 490
 491   if (n >= 32)
 492   {
 493     asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 494     asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 495   }
 496   else
 497   {
 498     asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 499     asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 500   }
 501
 502   asm ("mov.b64 %0, {%1, %2};" : "=l"(r) : "r"(tl), "r"(tr));
 503
 504   #endif
 505
 506   #if VECT_SIZE >= 2
 507
 508   {
 509     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.s0));
 510
 511     if (n >= 32)
 512     {
 513       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 514       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 515     }
 516     else
 517     {
 518       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 519       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 520     }
 521
 522     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.s0) : "r"(tl), "r"(tr));
 523   }
 524
 525   {
 526     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.s1));
 527
 528     if (n >= 32)
 529     {
 530       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 531       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 532     }
 533     else
 534     {
 535       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 536       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 537     }
 538
 539     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.s1) : "r"(tl), "r"(tr));
 540   }
 541
 542   #endif
 543
 544   #if VECT_SIZE >= 4
 545
 546   {
 547     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.s2));
 548
 549     if (n >= 32)
 550     {
 551       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 552       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 553     }
 554     else
 555     {
 556       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 557       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 558     }
 559
 560     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.s2) : "r"(tl), "r"(tr));
 561   }
 562
 563   {
 564     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.s3));
 565
 566     if (n >= 32)
 567     {
 568       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 569       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 570     }
 571     else
 572     {
 573       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 574       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 575     }
 576
 577     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.s3) : "r"(tl), "r"(tr));
 578   }
 579
 580   #endif
 581
 582   #if VECT_SIZE >= 8
 583
 584   {
 585     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.s4));
 586
 587     if (n >= 32)
 588     {
 589       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 590       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 591     }
 592     else
 593     {
 594       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 595       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 596     }
 597
 598     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.s4) : "r"(tl), "r"(tr));
 599   }
 600
 601   {
 602     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.s5));
 603
 604     if (n >= 32)
 605     {
 606       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 607       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 608     }
 609     else
 610     {
 611       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 612       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 613     }
 614
 615     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.s5) : "r"(tl), "r"(tr));
 616   }
 617
 618   {
 619     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.s6));
 620
 621     if (n >= 32)
 622     {
 623       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 624       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 625     }
 626     else
 627     {
 628       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 629       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 630     }
 631
 632     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.s6) : "r"(tl), "r"(tr));
 633   }
 634
 635   {
 636     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.s7));
 637
 638     if (n >= 32)
 639     {
 640       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 641       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 642     }
 643     else
 644     {
 645       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 646       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 647     }
 648
 649     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.s7) : "r"(tl), "r"(tr));
 650   }
 651
 652   #endif
 653
 654   #if VECT_SIZE >= 16
 655
 656   {
 657     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.s8));
 658
 659     if (n >= 32)
 660     {
 661       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 662       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 663     }
 664     else
 665     {
 666       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 667       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 668     }
 669
 670     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.s8) : "r"(tl), "r"(tr));
 671   }
 672
 673   {
 674     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.s9));
 675
 676     if (n >= 32)
 677     {
 678       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 679       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 680     }
 681     else
 682     {
 683       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 684       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 685     }
 686
 687     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.s9) : "r"(tl), "r"(tr));
 688   }
 689
 690   {
 691     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.sa));
 692
 693     if (n >= 32)
 694     {
 695       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 696       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 697     }
 698     else
 699     {
 700       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 701       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 702     }
 703
 704     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.sa) : "r"(tl), "r"(tr));
 705   }
 706
 707   {
 708     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.sb));
 709
 710     if (n >= 32)
 711     {
 712       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 713       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 714     }
 715     else
 716     {
 717       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 718       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 719     }
 720
 721     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.sb) : "r"(tl), "r"(tr));
 722   }
 723
 724   {
 725     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.sc));
 726
 727     if (n >= 32)
 728     {
 729       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 730       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 731     }
 732     else
 733     {
 734       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 735       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 736     }
 737
 738     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.sc) : "r"(tl), "r"(tr));
 739   }
 740
 741   {
 742     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.sd));
 743
 744     if (n >= 32)
 745     {
 746       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 747       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 748     }
 749     else
 750     {
 751       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 752       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 753     }
 754
 755     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.sd) : "r"(tl), "r"(tr));
 756   }
 757
 758   {
 759     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.se));
 760
 761     if (n >= 32)
 762     {
 763       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 764       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 765     }
 766     else
 767     {
 768       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 769       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 770     }
 771
 772     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.se) : "r"(tl), "r"(tr));
 773   }
 774
 775   {
 776     asm ("mov.b64 {%0, %1}, %2;" : "=r"(il), "=r"(ir) : "l"(a.sf));
 777
 778     if (n >= 32)
 779     {
 780       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(ir), "r"(il), "r"(n - 32));
 781       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(il), "r"(ir), "r"(n - 32));
 782     }
 783     else
 784     {
 785       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tl) : "r"(il), "r"(ir), "r"(n));
 786       asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(tr) : "r"(ir), "r"(il), "r"(n));
 787     }
 788
 789     asm ("mov.b64 %0, {%1, %2};" : "=l"(r.sf) : "r"(tl), "r"(tr));
 790   }
 791
 792   #endif
 793
 794   return r;
 795 }
 796 #else
 797 static inline u64x rotr64 (const u64x a, const u32 n)
 798 {
 799   return rotate (a, (u64) 64 - n);
 800 }
 801 #endif
 802
 803 static inline u64x rotl64 (const u64x a, const u32 n)
 804 {
 805   return rotr64 (a, (u64) 64 - n);
 806 }
 807
 808 static inline u32x __byte_perm (const u32x a, const u32x b, const u32x c)
 809 {
 810   u32x r;
 811
 812   #if VECT_SIZE == 1
 813   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r)    : "r"(a),    "r"(b),    "r"(c)   );
 814   #endif
 815
 816   #if VECT_SIZE >= 2
 817   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s0) : "r"(a.s0), "r"(b.s0), "r"(c.s0));
 818   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s1) : "r"(a.s1), "r"(b.s1), "r"(c.s1));
 819   #endif
 820
 821   #if VECT_SIZE >= 4
 822   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s2) : "r"(a.s2), "r"(b.s2), "r"(c.s2));
 823   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s3) : "r"(a.s3), "r"(b.s3), "r"(c.s3));
 824   #endif
 825
 826   #if VECT_SIZE >= 8
 827   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s4) : "r"(a.s4), "r"(b.s4), "r"(c.s4));
 828   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s5) : "r"(a.s5), "r"(b.s5), "r"(c.s5));
 829   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s6) : "r"(a.s6), "r"(b.s6), "r"(c.s6));
 830   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s7) : "r"(a.s7), "r"(b.s7), "r"(c.s7));
 831   #endif
 832
 833   #if VECT_SIZE >= 16
 834   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s8) : "r"(a.s8), "r"(b.s8), "r"(c.s8));
 835   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.s9) : "r"(a.s9), "r"(b.s9), "r"(c.s9));
 836   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sa) : "r"(a.sa), "r"(b.sa), "r"(c.sa));
 837   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sb) : "r"(a.sb), "r"(b.sb), "r"(c.sb));
 838   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sc) : "r"(a.sc), "r"(b.sc), "r"(c.sc));
 839   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sd) : "r"(a.sd), "r"(b.sd), "r"(c.sd));
 840   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.se) : "r"(a.se), "r"(b.se), "r"(c.se));
 841   asm ("prmt.b32 %0, %1, %2, %3;" : "=r"(r.sf) : "r"(a.sf), "r"(b.sf), "r"(c.sf));
 842   #endif
 843
 844   return r;
 845 }
 846
 847 static inline u32 __bfe (const u32 a, const u32 b, const u32 c)
 848 {
 849   u32 r;
 850
 851   asm ("bfe.u32 %0, %1, %2, %3;" : "=r"(r) : "r"(a), "r"(b), "r"(c));
 852
 853   return r;
 854 }
 855
 856 #if CUDA_ARCH >= 350
 857 static inline u32 amd_bytealign (const u32 a, const u32 b, const u32 c)
 858 {
 859   u32 r;
 860
 861   asm ("shf.r.wrap.b32 %0, %1, %2, %3;" : "=r"(r) : "r"(b), "r"(a), "r"((c & 3) * 8));
 862
 863   return r;
 864 }
 865 #else
 866 static inline u32 amd_bytealign (const u32 a, const u32 b, const u32 c)
 867 {
 868   return __byte_perm_S (b, a, (0x76543210 >> ((c & 3) * 4)) & 0xffff);
 869 }
 870 #endif
 871
 872 #if CUDA_ARCH >= 500
 873 static inline u32x lut3_2d (const u32x a, const u32x b, const u32x c)
 874 {
 875   u32x r;
 876
 877   #if VECT_SIZE == 1
 878   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r) : "r" (a), "r" (b), "r" (c));
 879   #endif
 880
 881   #if VECT_SIZE >= 2
 882   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.s0) : "r" (a.s0), "r" (b.s0), "r" (c.s0));
 883   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.s1) : "r" (a.s1), "r" (b.s1), "r" (c.s1));
 884   #endif
 885
 886   #if VECT_SIZE >= 4
 887   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.s2) : "r" (a.s2), "r" (b.s2), "r" (c.s2));
 888   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.s3) : "r" (a.s3), "r" (b.s3), "r" (c.s3));
 889   #endif
 890
 891   #if VECT_SIZE >= 8
 892   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.s4) : "r" (a.s4), "r" (b.s4), "r" (c.s4));
 893   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.s5) : "r" (a.s5), "r" (b.s5), "r" (c.s5));
 894   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.s6) : "r" (a.s6), "r" (b.s6), "r" (c.s6));
 895   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.s7) : "r" (a.s7), "r" (b.s7), "r" (c.s7));
 896   #endif
 897
 898   #if VECT_SIZE >= 16
 899   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.s8) : "r" (a.s8), "r" (b.s8), "r" (c.s8));
 900   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.s9) : "r" (a.s9), "r" (b.s9), "r" (c.s9));
 901   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.sa) : "r" (a.sa), "r" (b.sa), "r" (c.sa));
 902   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.sb) : "r" (a.sb), "r" (b.sb), "r" (c.sb));
 903   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.sc) : "r" (a.sc), "r" (b.sc), "r" (c.sc));
 904   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.sd) : "r" (a.sd), "r" (b.sd), "r" (c.sd));
 905   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.se) : "r" (a.se), "r" (b.se), "r" (c.se));
 906   asm ("lop3.b32 %0, %1, %2, %3, 0x2d;" : "=r" (r.sf) : "r" (a.sf), "r" (b.sf), "r" (c.sf));
 907   #endif
 908
 909   return r;
 910 }
 911
 912 static inline u32x lut3_39 (const u32x a, const u32x b, const u32x c)
 913 {
 914   u32x r;
 915
 916   #if VECT_SIZE == 1
 917   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r) : "r" (a), "r" (b), "r" (c));
 918   #endif
 919
 920   #if VECT_SIZE >= 2
 921   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.s0) : "r" (a.s0), "r" (b.s0), "r" (c.s0));
 922   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.s1) : "r" (a.s1), "r" (b.s1), "r" (c.s1));
 923   #endif
 924
 925   #if VECT_SIZE >= 4
 926   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.s2) : "r" (a.s2), "r" (b.s2), "r" (c.s2));
 927   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.s3) : "r" (a.s3), "r" (b.s3), "r" (c.s3));
 928   #endif
 929
 930   #if VECT_SIZE >= 8
 931   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.s4) : "r" (a.s4), "r" (b.s4), "r" (c.s4));
 932   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.s5) : "r" (a.s5), "r" (b.s5), "r" (c.s5));
 933   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.s6) : "r" (a.s6), "r" (b.s6), "r" (c.s6));
 934   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.s7) : "r" (a.s7), "r" (b.s7), "r" (c.s7));
 935   #endif
 936
 937   #if VECT_SIZE >= 16
 938   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.s8) : "r" (a.s8), "r" (b.s8), "r" (c.s8));
 939   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.s9) : "r" (a.s9), "r" (b.s9), "r" (c.s9));
 940   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.sa) : "r" (a.sa), "r" (b.sa), "r" (c.sa));
 941   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.sb) : "r" (a.sb), "r" (b.sb), "r" (c.sb));
 942   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.sc) : "r" (a.sc), "r" (b.sc), "r" (c.sc));
 943   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.sd) : "r" (a.sd), "r" (b.sd), "r" (c.sd));
 944   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.se) : "r" (a.se), "r" (b.se), "r" (c.se));
 945   asm ("lop3.b32 %0, %1, %2, %3, 0x39;" : "=r" (r.sf) : "r" (a.sf), "r" (b.sf), "r" (c.sf));
 946   #endif
 947
 948   return r;
 949 }
 950
 951 static inline u32x lut3_59 (const u32x a, const u32x b, const u32x c)
 952 {
 953   u32x r;
 954
 955   #if VECT_SIZE == 1
 956   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r) : "r" (a), "r" (b), "r" (c));
 957   #endif
 958
 959   #if VECT_SIZE >= 2
 960   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.s0) : "r" (a.s0), "r" (b.s0), "r" (c.s0));
 961   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.s1) : "r" (a.s1), "r" (b.s1), "r" (c.s1));
 962   #endif
 963
 964   #if VECT_SIZE >= 4
 965   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.s2) : "r" (a.s2), "r" (b.s2), "r" (c.s2));
 966   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.s3) : "r" (a.s3), "r" (b.s3), "r" (c.s3));
 967   #endif
 968
 969   #if VECT_SIZE >= 8
 970   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.s4) : "r" (a.s4), "r" (b.s4), "r" (c.s4));
 971   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.s5) : "r" (a.s5), "r" (b.s5), "r" (c.s5));
 972   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.s6) : "r" (a.s6), "r" (b.s6), "r" (c.s6));
 973   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.s7) : "r" (a.s7), "r" (b.s7), "r" (c.s7));
 974   #endif
 975
 976   #if VECT_SIZE >= 16
 977   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.s8) : "r" (a.s8), "r" (b.s8), "r" (c.s8));
 978   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.s9) : "r" (a.s9), "r" (b.s9), "r" (c.s9));
 979   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.sa) : "r" (a.sa), "r" (b.sa), "r" (c.sa));
 980   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.sb) : "r" (a.sb), "r" (b.sb), "r" (c.sb));
 981   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.sc) : "r" (a.sc), "r" (b.sc), "r" (c.sc));
 982   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.sd) : "r" (a.sd), "r" (b.sd), "r" (c.sd));
 983   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.se) : "r" (a.se), "r" (b.se), "r" (c.se));
 984   asm ("lop3.b32 %0, %1, %2, %3, 0x59;" : "=r" (r.sf) : "r" (a.sf), "r" (b.sf), "r" (c.sf));
 985   #endif
 986
 987   return r;
 988 }
 989
 990 static inline u32x lut3_96 (const u32x a, const u32x b, const u32x c)
 991 {
 992   u32x r;
 993
 994   #if VECT_SIZE == 1
 995   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r)    : "r" (a),    "r" (b),    "r" (c));
 996   #endif
 997
 998   #if VECT_SIZE >= 2
 999   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.s0) : "r" (a.s0), "r" (b.s0), "r" (c.s0));
1000   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.s1) : "r" (a.s1), "r" (b.s1), "r" (c.s1));
1001   #endif
1002
1003   #if VECT_SIZE >= 4
1004   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.s2) : "r" (a.s2), "r" (b.s2), "r" (c.s2));
1005   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.s3) : "r" (a.s3), "r" (b.s3), "r" (c.s3));
1006   #endif
1007
1008   #if VECT_SIZE >= 8
1009   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.s4) : "r" (a.s4), "r" (b.s4), "r" (c.s4));
1010   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.s5) : "r" (a.s5), "r" (b.s5), "r" (c.s5));
1011   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.s6) : "r" (a.s6), "r" (b.s6), "r" (c.s6));
1012   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.s7) : "r" (a.s7), "r" (b.s7), "r" (c.s7));
1013   #endif
1014
1015   #if VECT_SIZE >= 16
1016   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.s8) : "r" (a.s8), "r" (b.s8), "r" (c.s8));
1017   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.s9) : "r" (a.s9), "r" (b.s9), "r" (c.s9));
1018   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.sa) : "r" (a.sa), "r" (b.sa), "r" (c.sa));
1019   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.sb) : "r" (a.sb), "r" (b.sb), "r" (c.sb));
1020   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.sc) : "r" (a.sc), "r" (b.sc), "r" (c.sc));
1021   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.sd) : "r" (a.sd), "r" (b.sd), "r" (c.sd));
1022   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.se) : "r" (a.se), "r" (b.se), "r" (c.se));
1023   asm ("lop3.b32 %0, %1, %2, %3, 0x96;" : "=r" (r.sf) : "r" (a.sf), "r" (b.sf), "r" (c.sf));
1024   #endif
1025
1026   return r;
1027 }
1028
1029 static inline u32x lut3_e4 (const u32x a, const u32x b, const u32x c)
1030 {
1031   u32x r;
1032
1033   #if VECT_SIZE == 1
1034   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r)    : "r" (a),    "r" (b),    "r" (c));
1035   #endif
1036
1037   #if VECT_SIZE >= 2
1038   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.s0) : "r" (a.s0), "r" (b.s0), "r" (c.s0));
1039   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.s1) : "r" (a.s1), "r" (b.s1), "r" (c.s1));
1040   #endif
1041
1042   #if VECT_SIZE >= 4
1043   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.s2) : "r" (a.s2), "r" (b.s2), "r" (c.s2));
1044   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.s3) : "r" (a.s3), "r" (b.s3), "r" (c.s3));
1045   #endif
1046
1047   #if VECT_SIZE >= 8
1048   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.s4) : "r" (a.s4), "r" (b.s4), "r" (c.s4));
1049   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.s5) : "r" (a.s5), "r" (b.s5), "r" (c.s5));
1050   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.s6) : "r" (a.s6), "r" (b.s6), "r" (c.s6));
1051   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.s7) : "r" (a.s7), "r" (b.s7), "r" (c.s7));
1052   #endif
1053
1054   #if VECT_SIZE >= 16
1055   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.s8) : "r" (a.s8), "r" (b.s8), "r" (c.s8));
1056   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.s9) : "r" (a.s9), "r" (b.s9), "r" (c.s9));
1057   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.sa) : "r" (a.sa), "r" (b.sa), "r" (c.sa));
1058   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.sb) : "r" (a.sb), "r" (b.sb), "r" (c.sb));
1059   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.sc) : "r" (a.sc), "r" (b.sc), "r" (c.sc));
1060   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.sd) : "r" (a.sd), "r" (b.sd), "r" (c.sd));
1061   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.se) : "r" (a.se), "r" (b.se), "r" (c.se));
1062   asm ("lop3.b32 %0, %1, %2, %3, 0xe4;" : "=r" (r.sf) : "r" (a.sf), "r" (b.sf), "r" (c.sf));
1063   #endif
1064
1065   return r;
1066 }
1067
1068 static inline u32x lut3_e8 (const u32x a, const u32x b, const u32x c)
1069 {
1070   u32x r;
1071
1072   #if VECT_SIZE == 1
1073   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r)    : "r" (a),    "r" (b),    "r" (c));
1074   #endif
1075
1076   #if VECT_SIZE >= 2
1077   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.s0) : "r" (a.s0), "r" (b.s0), "r" (c.s0));
1078   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.s1) : "r" (a.s1), "r" (b.s1), "r" (c.s1));
1079   #endif
1080
1081   #if VECT_SIZE >= 4
1082   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.s2) : "r" (a.s2), "r" (b.s2), "r" (c.s2));
1083   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.s3) : "r" (a.s3), "r" (b.s3), "r" (c.s3));
1084   #endif
1085
1086   #if VECT_SIZE >= 8
1087   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.s4) : "r" (a.s4), "r" (b.s4), "r" (c.s4));
1088   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.s5) : "r" (a.s5), "r" (b.s5), "r" (c.s5));
1089   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.s6) : "r" (a.s6), "r" (b.s6), "r" (c.s6));
1090   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.s7) : "r" (a.s7), "r" (b.s7), "r" (c.s7));
1091   #endif
1092
1093   #if VECT_SIZE >= 16
1094   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.s8) : "r" (a.s8), "r" (b.s8), "r" (c.s8));
1095   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.s9) : "r" (a.s9), "r" (b.s9), "r" (c.s9));
1096   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.sa) : "r" (a.sa), "r" (b.sa), "r" (c.sa));
1097   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.sb) : "r" (a.sb), "r" (b.sb), "r" (c.sb));
1098   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.sc) : "r" (a.sc), "r" (b.sc), "r" (c.sc));
1099   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.sd) : "r" (a.sd), "r" (b.sd), "r" (c.sd));
1100   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.se) : "r" (a.se), "r" (b.se), "r" (c.se));
1101   asm ("lop3.b32 %0, %1, %2, %3, 0xe8;" : "=r" (r.sf) : "r" (a.sf), "r" (b.sf), "r" (c.sf));
1102   #endif
1103
1104   return r;
1105 }
1106
1107 static inline u32x lut3_ca (const u32x a, const u32x b, const u32x c)
1108 {
1109   u32x r;
1110
1111   #if VECT_SIZE == 1
1112   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r)    : "r" (a),    "r" (b),    "r" (c));
1113   #endif
1114
1115   #if VECT_SIZE >= 2
1116   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.s0) : "r" (a.s0), "r" (b.s0), "r" (c.s0));
1117   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.s1) : "r" (a.s1), "r" (b.s1), "r" (c.s1));
1118   #endif
1119
1120   #if VECT_SIZE >= 4
1121   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.s2) : "r" (a.s2), "r" (b.s2), "r" (c.s2));
1122   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.s3) : "r" (a.s3), "r" (b.s3), "r" (c.s3));
1123   #endif
1124
1125   #if VECT_SIZE >= 8
1126   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.s4) : "r" (a.s4), "r" (b.s4), "r" (c.s4));
1127   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.s5) : "r" (a.s5), "r" (b.s5), "r" (c.s5));
1128   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.s6) : "r" (a.s6), "r" (b.s6), "r" (c.s6));
1129   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.s7) : "r" (a.s7), "r" (b.s7), "r" (c.s7));
1130   #endif
1131
1132   #if VECT_SIZE >= 16
1133   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.s8) : "r" (a.s8), "r" (b.s8), "r" (c.s8));
1134   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.s9) : "r" (a.s9), "r" (b.s9), "r" (c.s9));
1135   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.sa) : "r" (a.sa), "r" (b.sa), "r" (c.sa));
1136   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.sb) : "r" (a.sb), "r" (b.sb), "r" (c.sb));
1137   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.sc) : "r" (a.sc), "r" (b.sc), "r" (c.sc));
1138   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.sd) : "r" (a.sd), "r" (b.sd), "r" (c.sd));
1139   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.se) : "r" (a.se), "r" (b.se), "r" (c.se));
1140   asm ("lop3.b32 %0, %1, %2, %3, 0xca;" : "=r" (r.sf) : "r" (a.sf), "r" (b.sf), "r" (c.sf));
1141   #endif
1142
1143   return r;
1144 }
1145
1146 #endif
1147 #endif
1148
1149 #ifdef IS_GENERIC
1150 static inline u32 swap32_S (const u32 v)
1151 {
1152   return (as_uint (as_uchar4 (v).s3210));
1153 }
1154
1155 static inline u64 swap64_S (const u64 v)
1156 {
1157   return (as_ulong (as_uchar8 (v).s76543210));
1158 }
1159
1160 static inline u32 rotr32_S (const u32 a, const u32 n)
1161 {
1162   return rotate (a, 32 - n);
1163 }
1164
1165 static inline u32 rotl32_S (const u32 a, const u32 n)
1166 {
1167   return rotate (a, n);
1168 }
1169
1170 static inline u64 rotr64_S (const u64 a, const u32 n)
1171 {
1172   return rotate (a, (u64) 64 - n);
1173 }
1174
1175 static inline u64 rotl64_S (const u64 a, const u32 n)
1176 {
1177   return rotate (a, (u64) n);
1178 }
1179
1180 static inline u32 amd_bytealign_S (const u32 a, const u32 b, const u32 c)
1181 {
1182   const u64 tmp = ((((u64) a) << 32) | ((u64) b)) >> ((c & 3) * 8);
1183
1184   return (u32) (tmp);
1185 }
1186
1187 static inline u32x swap32 (const u32x v)
1188 {
1189   return ((v >> 24) & 0x000000ff)
1190        | ((v >>  8) & 0x0000ff00)
1191        | ((v <<  8) & 0x00ff0000)
1192        | ((v << 24) & 0xff000000);
1193 }
1194
1195 static inline u64x swap64 (const u64x v)
1196 {
1197   return ((v >> 56) & 0x00000000000000ff)
1198        | ((v >> 40) & 0x000000000000ff00)
1199        | ((v >> 24) & 0x0000000000ff0000)
1200        | ((v >>  8) & 0x00000000ff000000)
1201        | ((v <<  8) & 0x000000ff00000000)
1202        | ((v << 24) & 0x0000ff0000000000)
1203        | ((v << 40) & 0x00ff000000000000)
1204        | ((v << 56) & 0xff00000000000000);
1205 }
1206
1207 static inline u32x rotr32 (const u32x a, const u32 n)
1208 {
1209   return rotate (a, 32 - n);
1210 }
1211
1212 static inline u32x rotl32 (const u32x a, const u32 n)
1213 {
1214   return rotate (a, n);
1215 }
1216
1217 static inline u64x rotr64 (const u64x a, const u32 n)
1218 {
1219   return rotate (a, (u64) 64 - n);
1220 }
1221
1222 static inline u64x rotl64 (const u64x a, const u32 n)
1223 {
1224   return rotate (a, (u64) n);
1225 }
1226
1227 static inline u32 __bfe (const u32 a, const u32 b, const u32 c)
1228 {
1229   #define BIT(x)      (1 << (x))
1230   #define BIT_MASK(x) (BIT (x) - 1)
1231   #define BFE(x,y,z)  (((x) >> (y)) & BIT_MASK (z))
1232
1233   return BFE (a, b, c);
1234 }
1235
1236 static inline u32x amd_bytealign (const u32x a, const u32x b, const u32 c)
1237 {
1238   #if VECT_SIZE == 1
1239   const u64x tmp = ((((u64x) (a)) << 32) | ((u64x) (b))) >> ((c & 3) * 8);
1240
1241   return (u32x) (tmp);
1242   #endif
1243
1244   #if VECT_SIZE == 2
1245   const u64x tmp = ((((u64x) (a.s0, a.s1)) << 32) | ((u64x) (b.s0, b.s1))) >> ((c & 3) * 8);
1246
1247   return (u32x) (tmp.s0, tmp.s1);
1248   #endif
1249
1250   #if VECT_SIZE == 4
1251   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);
1252
1253   return (u32x) (tmp.s0, tmp.s1, tmp.s2, tmp.s3);
1254   #endif
1255
1256   #if VECT_SIZE == 8
1257   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);
1258
1259   return (u32x) (tmp.s0, tmp.s1, tmp.s2, tmp.s3, tmp.s4, tmp.s5, tmp.s6, tmp.s7);
1260   #endif
1261
1262   #if VECT_SIZE == 16
1263   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);
1264
1265   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);
1266   #endif
1267 }
1268 #endif
1269
1270 typedef struct
1271 {
1272   #if   defined _DES_
1273   u32  digest_buf[4];
1274   #elif defined _MD4_
1275   u32  digest_buf[4];
1276   #elif defined _MD5_
1277   u32  digest_buf[4];
1278   #elif defined _MD5H_
1279   u32  digest_buf[4];
1280   #elif defined _SHA1_
1281   u32  digest_buf[5];
1282   #elif defined _BCRYPT_
1283   u32  digest_buf[6];
1284   #elif defined _SHA256_
1285   u32  digest_buf[8];
1286   #elif defined _SHA384_
1287   u32  digest_buf[16];
1288   #elif defined _SHA512_
1289   u32  digest_buf[16];
1290   #elif defined _KECCAK_
1291   u32  digest_buf[50];
1292   #elif defined _RIPEMD160_
1293   u32  digest_buf[5];
1294   #elif defined _WHIRLPOOL_
1295   u32  digest_buf[16];
1296   #elif defined _GOST_
1297   u32  digest_buf[8];
1298   #elif defined _GOST2012_256_
1299   u32  digest_buf[8];
1300   #elif defined _GOST2012_512_
1301   u32  digest_buf[16];
1302   #elif defined _SAPB_
1303   u32  digest_buf[4];
1304   #elif defined _SAPG_
1305   u32  digest_buf[5];
1306   #elif defined _MYSQL323_
1307   u32  digest_buf[4];
1308   #elif defined _LOTUS5_
1309   u32  digest_buf[4];
1310   #elif defined _LOTUS6_
1311   u32  digest_buf[4];
1312   #elif defined _SCRYPT_
1313   u32  digest_buf[8];
1314   #elif defined _LOTUS8_
1315   u32  digest_buf[4];
1316   #elif defined _OFFICE2007_
1317   u32  digest_buf[4];
1318   #elif defined _OFFICE2010_
1319   u32  digest_buf[4];
1320   #elif defined _OFFICE2013_
1321   u32  digest_buf[4];
1322   #elif defined _OLDOFFICE01_
1323   u32  digest_buf[4];
1324   #elif defined _OLDOFFICE34_
1325   u32  digest_buf[4];
1326   #elif defined _SIPHASH_
1327   u32  digest_buf[4];
1328   #elif defined _PBKDF2_MD5_
1329   u32  digest_buf[32];
1330   #elif defined _PBKDF2_SHA1_
1331   u32  digest_buf[32];
1332   #elif defined _PBKDF2_SHA256_
1333   u32  digest_buf[32];
1334   #elif defined _PBKDF2_SHA512_
1335   u32  digest_buf[32];
1336   #elif defined _PDF17L8_
1337   u32  digest_buf[8];
1338   #elif defined _CRC32_
1339   u32  digest_buf[4];
1340   #elif defined _SEVEN_ZIP_
1341   u32  digest_buf[4];
1342   #elif defined _ANDROIDFDE_
1343   u32  digest_buf[4];
1344   #elif defined _DCC2_
1345   u32  digest_buf[4];
1346   #elif defined _WPA_
1347   u32  digest_buf[4];
1348   #elif defined _MD5_SHA1_
1349   u32  digest_buf[4];
1350   #elif defined _SHA1_MD5_
1351   u32  digest_buf[5];
1352   #elif defined _NETNTLMV2_
1353   u32  digest_buf[4];
1354   #elif defined _KRB5PA_
1355   u32  digest_buf[4];
1356   #elif defined _CLOUDKEY_
1357   u32  digest_buf[8];
1358   #elif defined _SCRYPT_
1359   u32  digest_buf[4];
1360   #elif defined _PSAFE2_
1361   u32  digest_buf[5];
1362   #elif defined _LOTUS8_
1363   u32  digest_buf[4];
1364   #elif defined _RAR3_
1365   u32  digest_buf[4];
1366   #elif defined _SHA256_SHA1_
1367   u32  digest_buf[8];
1368   #elif defined _MS_DRSR_
1369   u32  digest_buf[8];
1370   #elif defined _ANDROIDFDE_SAMSUNG_
1371   u32  digest_buf[8];
1372   #elif defined _RAR5_
1373   u32  digest_buf[4];
1374   #elif defined _KRB5TGS_
1375   u32  digest_buf[4];
1376   #elif defined _AXCRYPT_
1377   u32  digest_buf[4];
1378   #elif defined _KEEPASS_
1379   u32  digest_buf[4];
1380   #endif
1381
1382 } digest_t;
1383
1384 typedef struct
1385 {
1386   u32 salt_buf[16];
1387   u32 salt_buf_pc[8];
1388
1389   u32 salt_len;
1390   u32 salt_iter;
1391   u32 salt_sign[2];
1392
1393   u32 keccak_mdlen;
1394   u32 truecrypt_mdlen;
1395
1396   u32 digests_cnt;
1397   u32 digests_done;
1398
1399   u32 digests_offset;
1400
1401   u32 scrypt_N;
1402   u32 scrypt_r;
1403   u32 scrypt_p;
1404   u32 scrypt_tmto;
1405   u32 scrypt_phy;
1406
1407 } salt_t;
1408
1409 typedef struct
1410 {
1411   int V;
1412   int R;
1413   int P;
1414
1415   int enc_md;
1416
1417   u32 id_buf[8];
1418   u32 u_buf[32];
1419   u32 o_buf[32];
1420
1421   int id_len;
1422   int o_len;
1423   int u_len;
1424
1425   u32 rc4key[2];
1426   u32 rc4data[2];
1427
1428 } pdf_t;
1429
1430 typedef struct
1431 {
1432   u32 pke[25];
1433   u32 eapol[64];
1434   int eapol_size;
1435   int keyver;
1436
1437 } wpa_t;
1438
1439 typedef struct
1440 {
1441   u32 cry_master_buf[64];
1442   u32 ckey_buf[64];
1443   u32 public_key_buf[64];
1444
1445   u32 cry_master_len;
1446   u32 ckey_len;
1447   u32 public_key_len;
1448
1449 } bitcoin_wallet_t;
1450
1451 typedef struct
1452 {
1453   u32 salt_buf[30];
1454   u32 salt_len;
1455
1456   u32 esalt_buf[38];
1457   u32 esalt_len;
1458
1459 } sip_t;
1460
1461 typedef struct
1462 {
1463   u32 data[384];
1464
1465 } androidfde_t;
1466
1467 typedef struct
1468 {
1469   u32 nr_buf[16];
1470   u32 nr_len;
1471
1472   u32 msg_buf[128];
1473   u32 msg_len;
1474
1475 } ikepsk_t;
1476
1477 typedef struct
1478 {
1479   u32 user_len;
1480   u32 domain_len;
1481   u32 srvchall_len;
1482   u32 clichall_len;
1483
1484   u32 userdomain_buf[64];
1485   u32 chall_buf[256];
1486
1487 } netntlm_t;
1488
1489 typedef struct
1490 {
1491   u32 user[16];
1492   u32 realm[16];
1493   u32 salt[32];
1494   u32 timestamp[16];
1495   u32 checksum[4];
1496
1497 } krb5pa_t;
1498
1499 typedef struct
1500 {
1501   u32 account_info[512];
1502   u32 checksum[4];
1503   u32 edata2[2560];
1504   u32 edata2_len;
1505
1506 } krb5tgs_t;
1507
1508 typedef struct
1509 {
1510   u32 salt_buf[16];
1511   u32 data_buf[112];
1512   u32 keyfile_buf[16];
1513
1514 } tc_t;
1515
1516 typedef struct
1517 {
1518   u32 salt_buf[16];
1519
1520 } pbkdf2_md5_t;
1521
1522 typedef struct
1523 {
1524   u32 salt_buf[16];
1525
1526 } pbkdf2_sha1_t;
1527
1528 typedef struct
1529 {
1530   u32 salt_buf[16];
1531
1532 } pbkdf2_sha256_t;
1533
1534 typedef struct
1535 {
1536   u32 salt_buf[32];
1537
1538 } pbkdf2_sha512_t;
1539
1540 typedef struct
1541 {
1542   u32 salt_buf[128];
1543   u32 salt_len;
1544
1545 } rakp_t;
1546
1547 typedef struct
1548 {
1549   u32 data_len;
1550   u32 data_buf[512];
1551
1552 } cloudkey_t;
1553
1554 typedef struct
1555 {
1556   u32 encryptedVerifier[4];
1557   u32 encryptedVerifierHash[5];
1558
1559   u32 keySize;
1560
1561 } office2007_t;
1562
1563 typedef struct
1564 {
1565   u32 encryptedVerifier[4];
1566   u32 encryptedVerifierHash[8];
1567
1568 } office2010_t;
1569
1570 typedef struct
1571 {
1572   u32 encryptedVerifier[4];
1573   u32 encryptedVerifierHash[8];
1574
1575 } office2013_t;
1576
1577 typedef struct
1578 {
1579   u32 version;
1580   u32 encryptedVerifier[4];
1581   u32 encryptedVerifierHash[4];
1582   u32 rc4key[2];
1583
1584 } oldoffice01_t;
1585
1586 typedef struct
1587 {
1588   u32 version;
1589   u32 encryptedVerifier[4];
1590   u32 encryptedVerifierHash[5];
1591   u32 rc4key[2];
1592
1593 } oldoffice34_t;
1594
1595 typedef struct
1596 {
1597   u32 version;
1598   u32 algorithm;
1599
1600   /* key-file handling */
1601   u32 keyfile_len;
1602   u32 keyfile[8];
1603
1604   u32 final_random_seed[8];
1605   u32 transf_random_seed[8];
1606   u32 enc_iv[4];
1607   u32 contents_hash[8];
1608
1609   /* specific to version 1 */
1610   u32 contents_len;
1611   u32 contents[12500];
1612
1613   /* specific to version 2 */
1614   u32 expected_bytes[8];
1615
1616 } keepass_t;
1617
1618 typedef struct
1619 {
1620   u32 digest[4];
1621   u32 out[4];
1622
1623 } pdf14_tmp_t;
1624
1625 typedef struct
1626 {
1627   union
1628   {
1629     u32 dgst32[16];
1630     u64 dgst64[8];
1631   };
1632
1633   u32 dgst_len;
1634   u32 W_len;
1635
1636 } pdf17l8_tmp_t;
1637
1638 typedef struct
1639 {
1640   u32 digest_buf[4];
1641
1642 } phpass_tmp_t;
1643
1644 typedef struct
1645 {
1646   u32 digest_buf[4];
1647
1648 } md5crypt_tmp_t;
1649
1650 typedef struct
1651 {
1652   u32 alt_result[8];
1653
1654   u32 p_bytes[4];
1655   u32 s_bytes[4];
1656
1657 } sha256crypt_tmp_t;
1658
1659 typedef struct
1660 {
1661   u64 l_alt_result[8];
1662
1663   u64 l_p_bytes[2];
1664   u64 l_s_bytes[2];
1665
1666 } sha512crypt_tmp_t;
1667
1668 typedef struct
1669 {
1670   u32 ipad[5];
1671   u32 opad[5];
1672
1673   u32 dgst[10];
1674   u32 out[10];
1675
1676 } wpa_tmp_t;
1677
1678 typedef struct
1679 {
1680   u64 dgst[8];
1681
1682 } bitcoin_wallet_tmp_t;
1683
1684 typedef struct
1685 {
1686   u32 ipad[5];
1687   u32 opad[5];
1688
1689   u32 dgst[5];
1690   u32 out[4];
1691
1692 } dcc2_tmp_t;
1693
1694 typedef struct
1695 {
1696   u32 E[18];
1697
1698   u32 P[18];
1699
1700   u32 S0[256];
1701   u32 S1[256];
1702   u32 S2[256];
1703   u32 S3[256];
1704
1705 } bcrypt_tmp_t;
1706
1707 typedef struct
1708 {
1709   u32 digest[2];
1710
1711   u32 P[18];
1712
1713   u32 S0[256];
1714   u32 S1[256];
1715   u32 S2[256];
1716   u32 S3[256];
1717
1718 } pwsafe2_tmp_t;
1719
1720 typedef struct
1721 {
1722   u32 digest_buf[8];
1723
1724 } pwsafe3_tmp_t;
1725
1726 typedef struct
1727 {
1728   u32 digest_buf[5];
1729
1730 } androidpin_tmp_t;
1731
1732 typedef struct
1733 {
1734   u32 ipad[5];
1735   u32 opad[5];
1736
1737   u32 dgst[10];
1738   u32 out[10];
1739
1740 } androidfde_tmp_t;
1741
1742 typedef struct
1743 {
1744   u32 ipad[16];
1745   u32 opad[16];
1746
1747   u32 dgst[64];
1748   u32 out[64];
1749
1750 } tc_tmp_t;
1751
1752 typedef struct
1753 {
1754   u64 ipad[8];
1755   u64 opad[8];
1756
1757   u64 dgst[32];
1758   u64 out[32];
1759
1760 } tc64_tmp_t;
1761
1762 typedef struct
1763 {
1764   u32 ipad[4];
1765   u32 opad[4];
1766
1767   u32 dgst[32];
1768   u32 out[32];
1769
1770 } pbkdf2_md5_tmp_t;
1771
1772 typedef struct
1773 {
1774   u32 ipad[5];
1775   u32 opad[5];
1776
1777   u32 dgst[32];
1778   u32 out[32];
1779
1780 } pbkdf2_sha1_tmp_t;
1781
1782 typedef struct
1783 {
1784   u32 ipad[8];
1785   u32 opad[8];
1786
1787   u32 dgst[32];
1788   u32 out[32];
1789
1790 } pbkdf2_sha256_tmp_t;
1791
1792 typedef struct
1793 {
1794   u64 ipad[8];
1795   u64 opad[8];
1796
1797   u64 dgst[16];
1798   u64 out[16];
1799
1800 } pbkdf2_sha512_tmp_t;
1801
1802 typedef struct
1803 {
1804   u64 out[8];
1805
1806 } ecryptfs_tmp_t;
1807
1808 typedef struct
1809 {
1810   u64 ipad[8];
1811   u64 opad[8];
1812
1813   u64 dgst[16];
1814   u64 out[16];
1815
1816 } oraclet_tmp_t;
1817
1818 typedef struct
1819 {
1820   u32 ipad[5];
1821   u32 opad[5];
1822
1823   u32 dgst[5];
1824   u32 out[5];
1825
1826 } agilekey_tmp_t;
1827
1828 typedef struct
1829 {
1830   u32 ipad[5];
1831   u32 opad[5];
1832
1833   u32 dgst1[5];
1834   u32 out1[5];
1835
1836   u32 dgst2[5];
1837   u32 out2[5];
1838
1839 } mywallet_tmp_t;
1840
1841 typedef struct
1842 {
1843   u32 ipad[5];
1844   u32 opad[5];
1845
1846   u32 dgst[5];
1847   u32 out[5];
1848
1849 } sha1aix_tmp_t;
1850
1851 typedef struct
1852 {
1853   u32 ipad[8];
1854   u32 opad[8];
1855
1856   u32 dgst[8];
1857   u32 out[8];
1858
1859 } sha256aix_tmp_t;
1860
1861 typedef struct
1862 {
1863   u64 ipad[8];
1864   u64 opad[8];
1865
1866   u64 dgst[8];
1867   u64 out[8];
1868
1869 } sha512aix_tmp_t;
1870
1871 typedef struct
1872 {
1873   u32 ipad[8];
1874   u32 opad[8];
1875
1876   u32 dgst[8];
1877   u32 out[8];
1878
1879 } lastpass_tmp_t;
1880
1881 typedef struct
1882 {
1883   u64 digest_buf[8];
1884
1885 } drupal7_tmp_t;
1886
1887 typedef struct
1888 {
1889   u32 ipad[5];
1890   u32 opad[5];
1891
1892   u32 dgst[5];
1893   u32 out[5];
1894
1895 } lotus8_tmp_t;
1896
1897 typedef struct
1898 {
1899   u32 out[5];
1900
1901 } office2007_tmp_t;
1902
1903 typedef struct
1904 {
1905   u32 out[5];
1906
1907 } office2010_tmp_t;
1908
1909 typedef struct
1910 {
1911   u64 out[8];
1912
1913 } office2013_tmp_t;
1914
1915 typedef struct
1916 {
1917   u32 digest_buf[5];
1918
1919 } saph_sha1_tmp_t;
1920
1921 typedef struct
1922 {
1923   u32 block[16];
1924
1925   u32 dgst[8];
1926
1927   u32 block_len;
1928   u32 final_len;
1929
1930 } seven_zip_tmp_t;
1931
1932 typedef struct
1933 {
1934   u32 KEK[5];
1935
1936   u32 lsb[4];
1937   u32 cipher[4];
1938
1939 } axcrypt_tmp_t;
1940
1941 typedef struct
1942 {
1943   u32 tmp_digest[8];
1944
1945 } keepass_tmp_t;
1946
1947 typedef struct
1948 {
1949   u32 Kc[16];
1950   u32 Kd[16];
1951
1952   u32 iv[2];
1953
1954 } bsdicrypt_tmp_t;
1955
1956 typedef struct
1957 {
1958   u32 dgst[17][5];
1959
1960 } rar3_tmp_t;
1961
1962 typedef struct
1963 {
1964   u32 user[16];
1965
1966 } cram_md5_t;
1967
1968 typedef struct
1969 {
1970   u32 iv_buf[4];
1971   u32 iv_len;
1972
1973   u32 salt_buf[4];
1974   u32 salt_len;
1975
1976   u32 crc;
1977
1978   u32 data_buf[96];
1979   u32 data_len;
1980
1981   u32 unpack_size;
1982
1983 } seven_zip_t;
1984
1985 typedef struct
1986 {
1987   u32 key;
1988   u64 val;
1989
1990 } hcstat_table_t;
1991
1992 typedef struct
1993 {
1994   u32 cs_buf[0x100];
1995   u32 cs_len;
1996
1997 } cs_t;
1998
1999 typedef struct
2000 {
2001   u32 cmds[0x100];
2002
2003 } kernel_rule_t;
2004
2005 typedef struct
2006 {
2007   u32 gidvid;
2008   u32 il_pos;
2009
2010 } plain_t;
2011
2012 typedef struct
2013 {
2014   u32 i[16];
2015
2016   u32 pw_len;
2017
2018   u32 alignment_placeholder_1;
2019   u32 alignment_placeholder_2;
2020   u32 alignment_placeholder_3;
2021
2022 } pw_t;
2023
2024 typedef struct
2025 {
2026   u32 i;
2027
2028 } bf_t;
2029
2030 typedef struct
2031 {
2032   u32 i[8];
2033
2034   u32 pw_len;
2035
2036 } comb_t;
2037
2038 typedef struct
2039 {
2040   u32 b[32];
2041
2042 } bs_word_t;
2043
2044 typedef struct
2045 {
2046   uint4 P[64];
2047
2048 } scrypt_tmp_t;