amd/gpu_serpent256_amd.c

   1 /* This is an independent implementation of the encryption algorithm:   */
   2 /*                                                                      */
   3 /*         Serpent by Ross Anderson, Eli Biham and Lars Knudsen         */
   4 /*                                                                      */
   5 /* which is a candidate algorithm in the Advanced Encryption Standard   */
   6 /* programme of the US National Institute of Standards and Technology.  */
   7 /*                                                                      */
   8 /* Copyright in this implementation is held by Dr B R Gladman but I     */
   9 /* hereby give permission for its free direct or derivative use subject */
  10 /* to acknowledgment of its origin and compliance with any conditions   */
  11 /* that the originators of the algorithm place on its exploitation.     */
  12 /*                                                                      */
  13 /* Dr Brian Gladman (gladman@seven77.demon.co.uk) 14th January 1999     */
  14 /*                                                                      */
  15 /* -------------------------------------------------------------------- */
  16 /*                                                                      */
  17 /* Cleaned and optimized for GPU use with oclHashcat by Jens Steube     */
  18
  19 /* 15 terms */
  20
  21 #define sb0(a,b,c,d,e,f,g,h)    \
  22     t1 = a ^ d;     \
  23     t2 = a & d;     \
  24     t3 = c ^ t1;    \
  25     t6 = b & t1;    \
  26     t4 = b ^ t3;    \
  27     t10 = ~t3;      \
  28     h = t2 ^ t4;    \
  29     t7 = a ^ t6;    \
  30     t14 = ~t7;      \
  31     t8 = c | t7;    \
  32     t11 = t3 ^ t7;  \
  33     g = t4 ^ t8;    \
  34     t12 = h & t11;  \
  35     f = t10 ^ t12;  \
  36     e = t12 ^ t14
  37
  38 /* 15 terms */
  39
  40 #define ib0(a,b,c,d,e,f,g,h)    \
  41     t1 = ~a;        \
  42     t2 = a ^ b;     \
  43     t3 = t1 | t2;   \
  44     t4 = d ^ t3;    \
  45     t7 = d & t2;    \
  46     t5 = c ^ t4;    \
  47     t8 = t1 ^ t7;   \
  48     g = t2 ^ t5;    \
  49     t11 = a & t4;   \
  50     t9 = g & t8;    \
  51     t14 = t5 ^ t8;  \
  52     f = t4 ^ t9;    \
  53     t12 = t5 | f;   \
  54     h = t11 ^ t12;  \
  55     e = h ^ t14
  56
  57 /* 14 terms!  */
  58
  59 #define sb1(a,b,c,d,e,f,g,h)    \
  60     t1 = ~a;        \
  61     t2 = b ^ t1;    \
  62     t3 = a | t2;    \
  63     t4 = d | t2;    \
  64     t5 = c ^ t3;    \
  65     g = d ^ t5;     \
  66     t7 = b ^ t4;    \
  67     t8 = t2 ^ g;    \
  68     t9 = t5 & t7;   \
  69     h = t8 ^ t9;    \
  70     t11 = t5 ^ t7;  \
  71     f = h ^ t11;    \
  72     t13 = t8 & t11; \
  73     e = t5 ^ t13
  74
  75 /* 17 terms */
  76
  77 #define ib1(a,b,c,d,e,f,g,h)    \
  78     t1 = a ^ d;     \
  79     t2 = a & b;     \
  80     t3 = b ^ c;     \
  81     t4 = a ^ t3;    \
  82     t5 = b | d;     \
  83     t7 = c | t1;    \
  84     h = t4 ^ t5;    \
  85     t8 = b ^ t7;    \
  86     t11 = ~t2;      \
  87     t9 = t4 & t8;   \
  88     f = t1 ^ t9;    \
  89     t13 = t9 ^ t11; \
  90     t12 = h & f;    \
  91     g = t12 ^ t13;  \
  92     t15 = a & d;    \
  93     t16 = c ^ t13;  \
  94     e = t15 ^ t16
  95
  96 /* 16 terms */
  97
  98 #define sb2(a,b,c,d,e,f,g,h)    \
  99     t1 = ~a;        \
 100     t2 = b ^ d;     \
 101     t3 = c & t1;    \
 102     t13 = d | t1;   \
 103     e = t2 ^ t3;    \
 104     t5 = c ^ t1;    \
 105     t6 = c ^ e;     \
 106     t7 = b & t6;    \
 107     t10 = e | t5;   \
 108     h = t5 ^ t7;    \
 109     t9 = d | t7;    \
 110     t11 = t9 & t10; \
 111     t14 = t2 ^ h;   \
 112     g = a ^ t11;    \
 113     t15 = g ^ t13;  \
 114     f = t14 ^ t15
 115
 116 /* 16 terms */
 117
 118 #define ib2(a,b,c,d,e,f,g,h)    \
 119     t1 = b ^ d;     \
 120     t2 = ~t1;       \
 121     t3 = a ^ c;     \
 122     t4 = c ^ t1;    \
 123     t7 = a | t2;    \
 124     t5 = b & t4;    \
 125     t8 = d ^ t7;    \
 126     t11 = ~t4;      \
 127     e = t3 ^ t5;    \
 128     t9 = t3 | t8;   \
 129     t14 = d & t11;  \
 130     h = t1 ^ t9;    \
 131     t12 = e | h;    \
 132     f = t11 ^ t12;  \
 133     t15 = t3 ^ t12; \
 134     g = t14 ^ t15
 135
 136 /* 17 terms */
 137
 138 #define sb3(a,b,c,d,e,f,g,h)    \
 139     t1 = a ^ c;     \
 140     t2 = d ^ t1;    \
 141     t3 = a & t2;    \
 142     t4 = d ^ t3;    \
 143     t5 = b & t4;    \
 144     g = t2 ^ t5;    \
 145     t7 = a | g;     \
 146     t8 = b | d;     \
 147     t11 = a | d;    \
 148     t9 = t4 & t7;   \
 149     f = t8 ^ t9;    \
 150     t12 = b ^ t11;  \
 151     t13 = g ^ t9;   \
 152     t15 = t3 ^ t8;  \
 153     h = t12 ^ t13;  \
 154     t16 = c & t15;  \
 155     e = t12 ^ t16
 156
 157 /* 16 term solution that performs less well than 17 term one
 158    in my environment (PPro/PII)
 159
 160 #define sb3(a,b,c,d,e,f,g,h)    \
 161     t1 = a ^ b;     \
 162     t2 = a & c;     \
 163     t3 = a | d;     \
 164     t4 = c ^ d;     \
 165     t5 = t1 & t3;   \
 166     t6 = t2 | t5;   \
 167     g = t4 ^ t6;    \
 168     t8 = b ^ t3;    \
 169     t9 = t6 ^ t8;   \
 170     t10 = t4 & t9;  \
 171     e = t1 ^ t10;   \
 172     t12 = g & e;    \
 173     f = t9 ^ t12;   \
 174     t14 = b | d;    \
 175     t15 = t4 ^ t12; \
 176     h = t14 ^ t15
 177 */
 178
 179 /* 17 terms */
 180
 181 #define ib3(a,b,c,d,e,f,g,h)    \
 182     t1 = b ^ c;     \
 183     t2 = b | c;     \
 184     t3 = a ^ c;     \
 185     t7 = a ^ d;     \
 186     t4 = t2 ^ t3;   \
 187     t5 = d | t4;    \
 188     t9 = t2 ^ t7;   \
 189     e = t1 ^ t5;    \
 190     t8 = t1 | t5;   \
 191     t11 = a & t4;   \
 192     g = t8 ^ t9;    \
 193     t12 = e | t9;   \
 194     f = t11 ^ t12;  \
 195     t14 = a & g;    \
 196     t15 = t2 ^ t14; \
 197     t16 = e & t15;  \
 198     h = t4 ^ t16
 199
 200 /* 15 terms */
 201
 202 #define sb4(a,b,c,d,e,f,g,h)    \
 203     t1 = a ^ d;     \
 204     t2 = d & t1;    \
 205     t3 = c ^ t2;    \
 206     t4 = b | t3;    \
 207     h = t1 ^ t4;    \
 208     t6 = ~b;        \
 209     t7 = t1 | t6;   \
 210     e = t3 ^ t7;    \
 211     t9 = a & e;     \
 212     t10 = t1 ^ t6;  \
 213     t11 = t4 & t10; \
 214     g = t9 ^ t11;   \
 215     t13 = a ^ t3;   \
 216     t14 = t10 & g;  \
 217     f = t13 ^ t14
 218
 219 /* 17 terms */
 220
 221 #define ib4(a,b,c,d,e,f,g,h)    \
 222     t1 = c ^ d;     \
 223     t2 = c | d;     \
 224     t3 = b ^ t2;    \
 225     t4 = a & t3;    \
 226     f = t1 ^ t4;    \
 227     t6 = a ^ d;     \
 228     t7 = b | d;     \
 229     t8 = t6 & t7;   \
 230     h = t3 ^ t8;    \
 231     t10 = ~a;       \
 232     t11 = c ^ h;    \
 233     t12 = t10 | t11;\
 234     e = t3 ^ t12;   \
 235     t14 = c | t4;   \
 236     t15 = t7 ^ t14; \
 237     t16 = h | t10;  \
 238     g = t15 ^ t16
 239
 240 /* 16 terms */
 241
 242 #define sb5(a,b,c,d,e,f,g,h)    \
 243     t1 = ~a;        \
 244     t2 = a ^ b;     \
 245     t3 = a ^ d;     \
 246     t4 = c ^ t1;    \
 247     t5 = t2 | t3;   \
 248     e = t4 ^ t5;    \
 249     t7 = d & e;     \
 250     t8 = t2 ^ e;    \
 251     t10 = t1 | e;   \
 252     f = t7 ^ t8;    \
 253     t11 = t2 | t7;  \
 254     t12 = t3 ^ t10; \
 255     t14 = b ^ t7;   \
 256     g = t11 ^ t12;  \
 257     t15 = f & t12;  \
 258     h = t14 ^ t15
 259
 260 /* 16 terms */
 261
 262 #define ib5(a,b,c,d,e,f,g,h)    \
 263     t1 = ~c;        \
 264     t2 = b & t1;    \
 265     t3 = d ^ t2;    \
 266     t4 = a & t3;    \
 267     t5 = b ^ t1;    \
 268     h = t4 ^ t5;    \
 269     t7 = b | h;     \
 270     t8 = a & t7;    \
 271     f = t3 ^ t8;    \
 272     t10 = a | d;    \
 273     t11 = t1 ^ t7;  \
 274     e = t10 ^ t11;  \
 275     t13 = a ^ c;    \
 276     t14 = b & t10;  \
 277     t15 = t4 | t13; \
 278     g = t14 ^ t15
 279
 280 /* 15 terms */
 281
 282 #define sb6(a,b,c,d,e,f,g,h)    \
 283     t1 = ~a;        \
 284     t2 = a ^ d;     \
 285     t3 = b ^ t2;    \
 286     t4 = t1 | t2;   \
 287     t5 = c ^ t4;    \
 288     f = b ^ t5;     \
 289     t13 = ~t5;      \
 290     t7 = t2 | f;    \
 291     t8 = d ^ t7;    \
 292     t9 = t5 & t8;   \
 293     g = t3 ^ t9;    \
 294     t11 = t5 ^ t8;  \
 295     e = g ^ t11;    \
 296     t14 = t3 & t11; \
 297     h = t13 ^ t14
 298
 299 /* 15 terms */
 300
 301 #define ib6(a,b,c,d,e,f,g,h)    \
 302     t1 = ~a;        \
 303     t2 = a ^ b;     \
 304     t3 = c ^ t2;    \
 305     t4 = c | t1;    \
 306     t5 = d ^ t4;    \
 307     t13 = d & t1;   \
 308     f = t3 ^ t5;    \
 309     t7 = t3 & t5;   \
 310     t8 = t2 ^ t7;   \
 311     t9 = b | t8;    \
 312     h = t5 ^ t9;    \
 313     t11 = b | h;    \
 314     e = t8 ^ t11;   \
 315     t14 = t3 ^ t11; \
 316     g = t13 ^ t14
 317
 318 /* 17 terms */
 319
 320 #define sb7(a,b,c,d,e,f,g,h)    \
 321     t1 = ~c;        \
 322     t2 = b ^ c;     \
 323     t3 = b | t1;    \
 324     t4 = d ^ t3;    \
 325     t5 = a & t4;    \
 326     t7 = a ^ d;     \
 327     h = t2 ^ t5;    \
 328     t8 = b ^ t5;    \
 329     t9 = t2 | t8;   \
 330     t11 = d & t3;   \
 331     f = t7 ^ t9;    \
 332     t12 = t5 ^ f;   \
 333     t15 = t1 | t4;  \
 334     t13 = h & t12;  \
 335     g = t11 ^ t13;  \
 336     t16 = t12 ^ g;  \
 337     e = t15 ^ t16
 338
 339 /* 17 terms */
 340
 341 #define ib7(a,b,c,d,e,f,g,h)    \
 342     t1 = a & b;     \
 343     t2 = a | b;     \
 344     t3 = c | t1;    \
 345     t4 = d & t2;    \
 346     h = t3 ^ t4;    \
 347     t6 = ~d;        \
 348     t7 = b ^ t4;    \
 349     t8 = h ^ t6;    \
 350     t11 = c ^ t7;   \
 351     t9 = t7 | t8;   \
 352     f = a ^ t9;     \
 353     t12 = d | f;    \
 354     e = t11 ^ t12;  \
 355     t14 = a & h;    \
 356     t15 = t3 ^ f;   \
 357     t16 = e ^ t14;  \
 358     g = t15 ^ t16
 359
 360 #define k_xor(r,a,b,c,d)    \
 361     a ^= ks[4 * r +  8]; \
 362     b ^= ks[4 * r +  9]; \
 363     c ^= ks[4 * r + 10]; \
 364     d ^= ks[4 * r + 11]
 365
 366 #define k_set(r,a,b,c,d)    \
 367     a = ks[4 * r +  8];  \
 368     b = ks[4 * r +  9];  \
 369     c = ks[4 * r + 10];  \
 370     d = ks[4 * r + 11]
 371
 372 #define k_get(r,a,b,c,d)    \
 373     ks[4 * r +  8] = a;  \
 374     ks[4 * r +  9] = b;  \
 375     ks[4 * r + 10] = c;  \
 376     ks[4 * r + 11] = d
 377
 378 /* the linear transformation and its inverse    */
 379
 380 #define rot(a,b,c,d)    \
 381     a = rotl32(a, 13);    \
 382     c = rotl32(c, 3);     \
 383     d ^= c ^ (a << 3);  \
 384     b ^= a ^ c;         \
 385     d = rotl32(d, 7);     \
 386     b = rotl32(b, 1);     \
 387     a ^= b ^ d;         \
 388     c ^= d ^ (b << 7);  \
 389     a = rotl32(a, 5);     \
 390     c = rotl32(c, 22)
 391
 392 #define irot(a,b,c,d)   \
 393     c = rotr32(c, 22);    \
 394     a = rotr32(a, 5);     \
 395     c ^= d ^ (b << 7);  \
 396     a ^= b ^ d;         \
 397     d = rotr32(d, 7);     \
 398     b = rotr32(b, 1);     \
 399     d ^= c ^ (a << 3);  \
 400     b ^= a ^ c;         \
 401     c = rotr32(c, 3);     \
 402     a = rotr32(a, 13)
 403
 404 static void serpent256_set_key (u32 *ks, const u32 *ukey)
 405 {
 406   #pragma unroll
 407   for (int i = 0; i < 8; i++)
 408   {
 409     ks[i] = ukey[i];
 410   }
 411
 412   #pragma unroll
 413   for (int i = 0; i < 132; i++)
 414   {
 415     ks[i + 8] = rotl32 (ks[i + 7] ^ ks[i + 5] ^ ks[i + 3] ^ ks[i + 0] ^ 0x9e3779b9 ^ i, 11);
 416   }
 417
 418   u32  a,b,c,d,e,f,g,h;
 419   u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
 420
 421   k_set( 0,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get( 0,e,f,g,h);
 422   k_set( 1,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get( 1,e,f,g,h);
 423   k_set( 2,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get( 2,e,f,g,h);
 424   k_set( 3,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get( 3,e,f,g,h);
 425   k_set( 4,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get( 4,e,f,g,h);
 426   k_set( 5,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get( 5,e,f,g,h);
 427   k_set( 6,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get( 6,e,f,g,h);
 428   k_set( 7,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get( 7,e,f,g,h);
 429   k_set( 8,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get( 8,e,f,g,h);
 430   k_set( 9,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get( 9,e,f,g,h);
 431   k_set(10,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get(10,e,f,g,h);
 432   k_set(11,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get(11,e,f,g,h);
 433   k_set(12,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get(12,e,f,g,h);
 434   k_set(13,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get(13,e,f,g,h);
 435   k_set(14,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get(14,e,f,g,h);
 436   k_set(15,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get(15,e,f,g,h);
 437   k_set(16,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(16,e,f,g,h);
 438   k_set(17,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get(17,e,f,g,h);
 439   k_set(18,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get(18,e,f,g,h);
 440   k_set(19,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get(19,e,f,g,h);
 441   k_set(20,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get(20,e,f,g,h);
 442   k_set(21,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get(21,e,f,g,h);
 443   k_set(22,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get(22,e,f,g,h);
 444   k_set(23,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get(23,e,f,g,h);
 445   k_set(24,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(24,e,f,g,h);
 446   k_set(25,a,b,c,d); sb2(a,b,c,d,e,f,g,h); k_get(25,e,f,g,h);
 447   k_set(26,a,b,c,d); sb1(a,b,c,d,e,f,g,h); k_get(26,e,f,g,h);
 448   k_set(27,a,b,c,d); sb0(a,b,c,d,e,f,g,h); k_get(27,e,f,g,h);
 449   k_set(28,a,b,c,d); sb7(a,b,c,d,e,f,g,h); k_get(28,e,f,g,h);
 450   k_set(29,a,b,c,d); sb6(a,b,c,d,e,f,g,h); k_get(29,e,f,g,h);
 451   k_set(30,a,b,c,d); sb5(a,b,c,d,e,f,g,h); k_get(30,e,f,g,h);
 452   k_set(31,a,b,c,d); sb4(a,b,c,d,e,f,g,h); k_get(31,e,f,g,h);
 453   k_set(32,a,b,c,d); sb3(a,b,c,d,e,f,g,h); k_get(32,e,f,g,h);
 454 }
 455
 456 static void serpent256_encrypt (const u32 *ks, const u32 *in, u32 *out)
 457 {
 458   u32  a,b,c,d,e,f,g,h;
 459   u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
 460
 461   a = in[0];
 462   b = in[1];
 463   c = in[2];
 464   d = in[3];
 465
 466   k_xor( 0,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 467   k_xor( 1,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 468   k_xor( 2,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 469   k_xor( 3,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 470   k_xor( 4,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 471   k_xor( 5,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 472   k_xor( 6,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 473   k_xor( 7,e,f,g,h); sb7(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 474   k_xor( 8,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 475   k_xor( 9,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 476   k_xor(10,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 477   k_xor(11,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 478   k_xor(12,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 479   k_xor(13,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 480   k_xor(14,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 481   k_xor(15,e,f,g,h); sb7(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 482   k_xor(16,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 483   k_xor(17,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 484   k_xor(18,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 485   k_xor(19,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 486   k_xor(20,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 487   k_xor(21,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 488   k_xor(22,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 489   k_xor(23,e,f,g,h); sb7(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 490   k_xor(24,a,b,c,d); sb0(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 491   k_xor(25,e,f,g,h); sb1(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 492   k_xor(26,a,b,c,d); sb2(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 493   k_xor(27,e,f,g,h); sb3(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 494   k_xor(28,a,b,c,d); sb4(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 495   k_xor(29,e,f,g,h); sb5(e,f,g,h,a,b,c,d); rot(a,b,c,d);
 496   k_xor(30,a,b,c,d); sb6(a,b,c,d,e,f,g,h); rot(e,f,g,h);
 497   k_xor(31,e,f,g,h); sb7(e,f,g,h,a,b,c,d);
 498   k_xor(32,a,b,c,d);
 499
 500   out[0] = a;
 501   out[1] = b;
 502   out[2] = c;
 503   out[3] = d;
 504 }
 505
 506 static void serpent256_decrypt (const u32 *ks, const u32 *in, u32 *out)
 507 {
 508   u32  a,b,c,d,e,f,g,h;
 509   u32  t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15,t16;
 510
 511   a = in[0];
 512   b = in[1];
 513   c = in[2];
 514   d = in[3];
 515
 516                                        k_xor(32,a,b,c,d);
 517                  ib7(a,b,c,d,e,f,g,h); k_xor(31,e,f,g,h);
 518   irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor(30,a,b,c,d);
 519   irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor(29,e,f,g,h);
 520   irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor(28,a,b,c,d);
 521   irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor(27,e,f,g,h);
 522   irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor(26,a,b,c,d);
 523   irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor(25,e,f,g,h);
 524   irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor(24,a,b,c,d);
 525   irot(a,b,c,d); ib7(a,b,c,d,e,f,g,h); k_xor(23,e,f,g,h);
 526   irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor(22,a,b,c,d);
 527   irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor(21,e,f,g,h);
 528   irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor(20,a,b,c,d);
 529   irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor(19,e,f,g,h);
 530   irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor(18,a,b,c,d);
 531   irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor(17,e,f,g,h);
 532   irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor(16,a,b,c,d);
 533   irot(a,b,c,d); ib7(a,b,c,d,e,f,g,h); k_xor(15,e,f,g,h);
 534   irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor(14,a,b,c,d);
 535   irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor(13,e,f,g,h);
 536   irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor(12,a,b,c,d);
 537   irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor(11,e,f,g,h);
 538   irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor(10,a,b,c,d);
 539   irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor( 9,e,f,g,h);
 540   irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor( 8,a,b,c,d);
 541   irot(a,b,c,d); ib7(a,b,c,d,e,f,g,h); k_xor( 7,e,f,g,h);
 542   irot(e,f,g,h); ib6(e,f,g,h,a,b,c,d); k_xor( 6,a,b,c,d);
 543   irot(a,b,c,d); ib5(a,b,c,d,e,f,g,h); k_xor( 5,e,f,g,h);
 544   irot(e,f,g,h); ib4(e,f,g,h,a,b,c,d); k_xor( 4,a,b,c,d);
 545   irot(a,b,c,d); ib3(a,b,c,d,e,f,g,h); k_xor( 3,e,f,g,h);
 546   irot(e,f,g,h); ib2(e,f,g,h,a,b,c,d); k_xor( 2,a,b,c,d);
 547   irot(a,b,c,d); ib1(a,b,c,d,e,f,g,h); k_xor( 1,e,f,g,h);
 548   irot(e,f,g,h); ib0(e,f,g,h,a,b,c,d); k_xor( 0,a,b,c,d);
 549
 550   out[0] = a;
 551   out[1] = b;
 552   out[2] = c;
 553   out[3] = d;
 554 }
 555
 556 static void serpent256_decrypt_xts (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out)
 557 {
 558   u32 T[4] = { 0 };
 559   u32 Z[4] = { 0 };
 560
 561   out[0] = in[0];
 562   out[1] = in[1];
 563   out[2] = in[2];
 564   out[3] = in[3];
 565
 566   u32 ks[140];
 567
 568   serpent256_set_key (ks, ukey2);
 569   serpent256_encrypt (ks, Z, T);
 570
 571   out[0] ^= T[0];
 572   out[1] ^= T[1];
 573   out[2] ^= T[2];
 574   out[3] ^= T[3];
 575
 576   serpent256_set_key (ks, ukey1);
 577   serpent256_decrypt (ks, out, out);
 578
 579   out[0] ^= T[0];
 580   out[1] ^= T[1];
 581   out[2] ^= T[2];
 582   out[3] ^= T[3];
 583 }