src/external/Random123-1.08/include/Random123/threefry.h

   1 /*
   2 Copyright 2010-2011, D. E. Shaw Research.
   3 All rights reserved.
   4
   5 Redistribution and use in source and binary forms, with or without
   6 modification, are permitted provided that the following conditions are
   7 met:
   8
   9 * Redistributions of source code must retain the above copyright
  10   notice, this list of conditions, and the following disclaimer.
  11
  12 * Redistributions in binary form must reproduce the above copyright
  13   notice, this list of conditions, and the following disclaimer in the
  14   documentation and/or other materials provided with the distribution.
  15
  16 * Neither the name of D. E. Shaw Research nor the names of its
  17   contributors may be used to endorse or promote products derived from
  18   this software without specific prior written permission.
  19
  20 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  23 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  24 OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  25 SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  26 LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  27 DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  28 THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  29 (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  30 OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  31 */
  32 #ifndef _threefry_dot_h_
  33 #define _threefry_dot_h_
  34 #include "features/compilerfeatures.h"
  35 #include "array.h"
  36
  37 /** \cond HIDDEN_FROM_DOXYGEN */
  38 /* Significant parts of this file were copied from
  39    from:
  40       Skein_FinalRnd/ReferenceImplementation/skein.h
  41       Skein_FinalRnd/ReferenceImplementation/skein_block.c
  42
  43    in http://csrc.nist.gov/groups/ST/hash/sha-3/Round3/documents/Skein_FinalRnd.zip
  44
  45    This file has been modified so that it may no longer perform its originally
  46    intended function.  If you're looking for a Skein or Threefish source code,
  47    please consult the original file.
  48
  49    The original file had the following header:
  50 **************************************************************************
  51 **
  52 ** Interface declarations and internal definitions for Skein hashing.
  53 **
  54 ** Source code author: Doug Whiting, 2008.
  55 **
  56 ** This algorithm and source code is released to the public domain.
  57 **
  58 ***************************************************************************
  59
  60 */
  61
  62 /* See comment at the top of philox.h for the macro pre-process
  63    strategy. */
  64
  65 /* Rotation constants: */
  66 enum r123_enum_threefry64x4 {
  67     /* These are the R_256 constants from the Threefish reference sources
  68        with names changed to R_64x4... */
  69     R_64x4_0_0=14, R_64x4_0_1=16,
  70     R_64x4_1_0=52, R_64x4_1_1=57,
  71     R_64x4_2_0=23, R_64x4_2_1=40,
  72     R_64x4_3_0= 5, R_64x4_3_1=37,
  73     R_64x4_4_0=25, R_64x4_4_1=33,
  74     R_64x4_5_0=46, R_64x4_5_1=12,
  75     R_64x4_6_0=58, R_64x4_6_1=22,
  76     R_64x4_7_0=32, R_64x4_7_1=32
  77 };
  78
  79 enum r123_enum_threefry64x2 {
  80     /*
  81     // Output from skein_rot_search: (srs64_B64-X1000)
  82     // Random seed = 1. BlockSize = 128 bits. sampleCnt =  1024. rounds =  8, minHW_or=57
  83     // Start: Tue Mar  1 10:07:48 2011
  84     // rMin = 0.136. #0325[*15] [CRC=455A682F. hw_OR=64. cnt=16384. blkSize= 128].format
  85     */
  86     R_64x2_0_0=16,
  87     R_64x2_1_0=42,
  88     R_64x2_2_0=12,
  89     R_64x2_3_0=31,
  90     R_64x2_4_0=16,
  91     R_64x2_5_0=32,
  92     R_64x2_6_0=24,
  93     R_64x2_7_0=21
  94     /* 4 rounds: minHW =  4  [  4  4  4  4 ]
  95     // 5 rounds: minHW =  8  [  8  8  8  8 ]
  96     // 6 rounds: minHW = 16  [ 16 16 16 16 ]
  97     // 7 rounds: minHW = 32  [ 32 32 32 32 ]
  98     // 8 rounds: minHW = 64  [ 64 64 64 64 ]
  99     // 9 rounds: minHW = 64  [ 64 64 64 64 ]
 100     //10 rounds: minHW = 64  [ 64 64 64 64 ]
 101     //11 rounds: minHW = 64  [ 64 64 64 64 ] */
 102 };
 103
 104 enum r123_enum_threefry32x4 {
 105     /* Output from skein_rot_search: (srs-B128-X5000.out)
 106     // Random seed = 1. BlockSize = 64 bits. sampleCnt =  1024. rounds =  8, minHW_or=28
 107     // Start: Mon Aug 24 22:41:36 2009
 108     // ...
 109     // rMin = 0.472. #0A4B[*33] [CRC=DD1ECE0F. hw_OR=31. cnt=16384. blkSize= 128].format    */
 110     R_32x4_0_0=10, R_32x4_0_1=26,
 111     R_32x4_1_0=11, R_32x4_1_1=21,
 112     R_32x4_2_0=13, R_32x4_2_1=27,
 113     R_32x4_3_0=23, R_32x4_3_1= 5,
 114     R_32x4_4_0= 6, R_32x4_4_1=20,
 115     R_32x4_5_0=17, R_32x4_5_1=11,
 116     R_32x4_6_0=25, R_32x4_6_1=10,
 117     R_32x4_7_0=18, R_32x4_7_1=20
 118
 119     /* 4 rounds: minHW =  3  [  3  3  3  3 ]
 120     // 5 rounds: minHW =  7  [  7  7  7  7 ]
 121     // 6 rounds: minHW = 12  [ 13 12 13 12 ]
 122     // 7 rounds: minHW = 22  [ 22 23 22 23 ]
 123     // 8 rounds: minHW = 31  [ 31 31 31 31 ]
 124     // 9 rounds: minHW = 32  [ 32 32 32 32 ]
 125     //10 rounds: minHW = 32  [ 32 32 32 32 ]
 126     //11 rounds: minHW = 32  [ 32 32 32 32 ] */
 127
 128 };
 129
 130 enum r123_enum_threefry32x2 {
 131     /* Output from skein_rot_search (srs32x2-X5000.out)
 132     // Random seed = 1. BlockSize = 64 bits. sampleCnt =  1024. rounds =  8, minHW_or=28
 133     // Start: Tue Jul 12 11:11:33 2011
 134     // rMin = 0.334. #0206[*07] [CRC=1D9765C0. hw_OR=32. cnt=16384. blkSize=  64].format   */
 135     R_32x2_0_0=13,
 136     R_32x2_1_0=15,
 137     R_32x2_2_0=26,
 138     R_32x2_3_0= 6,
 139     R_32x2_4_0=17,
 140     R_32x2_5_0=29,
 141     R_32x2_6_0=16,
 142     R_32x2_7_0=24
 143
 144     /* 4 rounds: minHW =  4  [  4  4  4  4 ]
 145     // 5 rounds: minHW =  6  [  6  8  6  8 ]
 146     // 6 rounds: minHW =  9  [  9 12  9 12 ]
 147     // 7 rounds: minHW = 16  [ 16 24 16 24 ]
 148     // 8 rounds: minHW = 32  [ 32 32 32 32 ]
 149     // 9 rounds: minHW = 32  [ 32 32 32 32 ]
 150     //10 rounds: minHW = 32  [ 32 32 32 32 ]
 151     //11 rounds: minHW = 32  [ 32 32 32 32 ] */
 152     };
 153
 154 enum r123_enum_threefry_wcnt {
 155     WCNT2=2,
 156     WCNT4=4
 157 };
 158 R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(uint64_t RotL_64(uint64_t x, unsigned int N));
 159 R123_CUDA_DEVICE R123_STATIC_INLINE uint64_t RotL_64(uint64_t x, unsigned int N)
 160 {
 161     return (x << (N & 63)) | (x >> ((64-N) & 63));
 162 }
 163
 164 R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(uint32_t RotL_32(uint32_t x, unsigned int N));
 165 R123_CUDA_DEVICE R123_STATIC_INLINE uint32_t RotL_32(uint32_t x, unsigned int N)
 166 {
 167     return (x << (N & 31)) | (x >> ((32-N) & 31));
 168 }
 169
 170 #define SKEIN_MK_64(hi32,lo32)  ((lo32) + (((uint64_t) (hi32)) << 32))
 171 #define SKEIN_KS_PARITY64         SKEIN_MK_64(0x1BD11BDA,0xA9FC1A22)
 172 #define SKEIN_KS_PARITY32         0x1BD11BDA
 173
 174 #ifndef THREEFRY2x32_DEFAULT_ROUNDS
 175 #define THREEFRY2x32_DEFAULT_ROUNDS 20
 176 #endif
 177
 178 #ifndef THREEFRY2x64_DEFAULT_ROUNDS
 179 #define THREEFRY2x64_DEFAULT_ROUNDS 20
 180 #endif
 181
 182 #ifndef THREEFRY4x32_DEFAULT_ROUNDS
 183 #define THREEFRY4x32_DEFAULT_ROUNDS 20
 184 #endif
 185
 186 #ifndef THREEFRY4x64_DEFAULT_ROUNDS
 187 #define THREEFRY4x64_DEFAULT_ROUNDS 20
 188 #endif
 189
 190 #define _threefry2x_tpl(W)                                              \
 191 typedef struct r123array2x##W threefry2x##W##_ctr_t;                          \
 192 typedef struct r123array2x##W threefry2x##W##_key_t;                          \
 193 typedef struct r123array2x##W threefry2x##W##_ukey_t;                          \
 194 R123_CUDA_DEVICE R123_STATIC_INLINE threefry2x##W##_key_t threefry2x##W##keyinit(threefry2x##W##_ukey_t uk) { return uk; } \
 195 R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(threefry2x##W##_ctr_t threefry2x##W##_R(unsigned int Nrounds, threefry2x##W##_ctr_t in, threefry2x##W##_key_t k)); \
 196 R123_CUDA_DEVICE R123_STATIC_INLINE                                          \
 197 threefry2x##W##_ctr_t threefry2x##W##_R(unsigned int Nrounds, threefry2x##W##_ctr_t in, threefry2x##W##_key_t k){ \
 198     threefry2x##W##_ctr_t X;                                              \
 199     uint##W##_t ks[2+1];                                          \
 200     int  i; /* avoid size_t to avoid need for stddef.h */                   \
 201     R123_ASSERT(Nrounds<=32);                                           \
 202     ks[2] =  SKEIN_KS_PARITY##W;                                   \
 203     for (i=0;i < 2; i++)                                        \
 204         {                                                               \
 205             ks[i] = k.v[i];                                             \
 206             X.v[i]  = in.v[i];                                          \
 207             ks[2] ^= k.v[i];                                    \
 208         }                                                               \
 209                                                                         \
 210     /* Insert initial key before round 0 */                             \
 211     X.v[0] += ks[0]; X.v[1] += ks[1];                                   \
 212                                                                         \
 213     if(Nrounds>0){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_0_0); X.v[1] ^= X.v[0]; } \
 214     if(Nrounds>1){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_1_0); X.v[1] ^= X.v[0]; } \
 215     if(Nrounds>2){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_2_0); X.v[1] ^= X.v[0]; } \
 216     if(Nrounds>3){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_3_0); X.v[1] ^= X.v[0]; } \
 217     if(Nrounds>3){                                                      \
 218         /* InjectKey(r=1) */                                            \
 219         X.v[0] += ks[1]; X.v[1] += ks[2];                               \
 220         X.v[1] += 1;     /* X.v[2-1] += r  */                   \
 221     }                                                                   \
 222     if(Nrounds>4){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_4_0); X.v[1] ^= X.v[0]; } \
 223     if(Nrounds>5){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_5_0); X.v[1] ^= X.v[0]; } \
 224     if(Nrounds>6){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_6_0); X.v[1] ^= X.v[0]; } \
 225     if(Nrounds>7){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_7_0); X.v[1] ^= X.v[0]; } \
 226     if(Nrounds>7){                                                      \
 227         /* InjectKey(r=2) */                                            \
 228         X.v[0] += ks[2]; X.v[1] += ks[0];                               \
 229         X.v[1] += 2;                                                    \
 230     }                                                                   \
 231     if(Nrounds>8){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_0_0); X.v[1] ^= X.v[0]; } \
 232     if(Nrounds>9){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_1_0); X.v[1] ^= X.v[0]; } \
 233     if(Nrounds>10){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_2_0); X.v[1] ^= X.v[0]; } \
 234     if(Nrounds>11){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_3_0); X.v[1] ^= X.v[0]; } \
 235     if(Nrounds>11){                                                     \
 236         /* InjectKey(r=3) */                                            \
 237         X.v[0] += ks[0]; X.v[1] += ks[1];                               \
 238         X.v[1] += 3;                                                    \
 239     }                                                                   \
 240     if(Nrounds>12){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_4_0); X.v[1] ^= X.v[0]; } \
 241     if(Nrounds>13){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_5_0); X.v[1] ^= X.v[0]; } \
 242     if(Nrounds>14){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_6_0); X.v[1] ^= X.v[0]; } \
 243     if(Nrounds>15){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_7_0); X.v[1] ^= X.v[0]; } \
 244     if(Nrounds>15){                                                     \
 245         /* InjectKey(r=4) */                                            \
 246         X.v[0] += ks[1]; X.v[1] += ks[2];                               \
 247         X.v[1] += 4;                                                    \
 248     }                                                                   \
 249     if(Nrounds>16){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_0_0); X.v[1] ^= X.v[0]; } \
 250     if(Nrounds>17){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_1_0); X.v[1] ^= X.v[0]; } \
 251     if(Nrounds>18){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_2_0); X.v[1] ^= X.v[0]; } \
 252     if(Nrounds>19){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_3_0); X.v[1] ^= X.v[0]; } \
 253     if(Nrounds>19){                                                     \
 254         /* InjectKey(r=5) */                                            \
 255         X.v[0] += ks[2]; X.v[1] += ks[0];                               \
 256         X.v[1] += 5;                                                    \
 257     }                                                                   \
 258     if(Nrounds>20){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_4_0); X.v[1] ^= X.v[0]; } \
 259     if(Nrounds>21){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_5_0); X.v[1] ^= X.v[0]; } \
 260     if(Nrounds>22){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_6_0); X.v[1] ^= X.v[0]; } \
 261     if(Nrounds>23){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_7_0); X.v[1] ^= X.v[0]; } \
 262     if(Nrounds>23){                                                     \
 263         /* InjectKey(r=6) */                                            \
 264         X.v[0] += ks[0]; X.v[1] += ks[1];                               \
 265         X.v[1] += 6;                                                    \
 266     }                                                                   \
 267     if(Nrounds>24){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_0_0); X.v[1] ^= X.v[0]; } \
 268     if(Nrounds>25){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_1_0); X.v[1] ^= X.v[0]; } \
 269     if(Nrounds>26){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_2_0); X.v[1] ^= X.v[0]; } \
 270     if(Nrounds>27){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_3_0); X.v[1] ^= X.v[0]; } \
 271     if(Nrounds>27){                                                     \
 272         /* InjectKey(r=7) */                                            \
 273         X.v[0] += ks[1]; X.v[1] += ks[2];                               \
 274         X.v[1] += 7;                                                    \
 275     }                                                                   \
 276     if(Nrounds>28){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_4_0); X.v[1] ^= X.v[0]; } \
 277     if(Nrounds>29){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_5_0); X.v[1] ^= X.v[0]; } \
 278     if(Nrounds>30){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_6_0); X.v[1] ^= X.v[0]; } \
 279     if(Nrounds>31){  X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x2_7_0); X.v[1] ^= X.v[0]; } \
 280     if(Nrounds>31){                                                     \
 281         /* InjectKey(r=8) */                                            \
 282         X.v[0] += ks[2]; X.v[1] += ks[0];                               \
 283         X.v[1] += 8;                                                    \
 284     }                                                                   \
 285     return X;                                                           \
 286 }                                                                       \
 287  /** @ingroup ThreefryNxW */                                            \
 288 enum r123_enum_threefry2x##W { threefry2x##W##_rounds = THREEFRY2x##W##_DEFAULT_ROUNDS };       \
 289 R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(threefry2x##W##_ctr_t threefry2x##W(threefry2x##W##_ctr_t in, threefry2x##W##_key_t k)); \
 290 R123_CUDA_DEVICE R123_STATIC_INLINE                                     \
 291 threefry2x##W##_ctr_t threefry2x##W(threefry2x##W##_ctr_t in, threefry2x##W##_key_t k){ \
 292     return threefry2x##W##_R(threefry2x##W##_rounds, in, k);            \
 293 }
 294
 295
 296 #define _threefry4x_tpl(W)                                              \
 297 typedef struct r123array4x##W threefry4x##W##_ctr_t;                        \
 298 typedef struct r123array4x##W threefry4x##W##_key_t;                        \
 299 typedef struct r123array4x##W threefry4x##W##_ukey_t;                        \
 300 R123_CUDA_DEVICE R123_STATIC_INLINE threefry4x##W##_key_t threefry4x##W##keyinit(threefry4x##W##_ukey_t uk) { return uk; } \
 301 R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(threefry4x##W##_ctr_t threefry4x##W##_R(unsigned int Nrounds, threefry4x##W##_ctr_t in, threefry4x##W##_key_t k)); \
 302 R123_CUDA_DEVICE R123_STATIC_INLINE                                          \
 303 threefry4x##W##_ctr_t threefry4x##W##_R(unsigned int Nrounds, threefry4x##W##_ctr_t in, threefry4x##W##_key_t k){ \
 304     threefry4x##W##_ctr_t X;                                            \
 305     uint##W##_t ks[4+1];                                            \
 306     int  i; /* avoid size_t to avoid need for stddef.h */                   \
 307     R123_ASSERT(Nrounds<=72);                                           \
 308     ks[4] =  SKEIN_KS_PARITY##W;                                    \
 309     for (i=0;i < 4; i++)                                            \
 310         {                                                               \
 311             ks[i] = k.v[i];                                             \
 312             X.v[i]  = in.v[i];                                          \
 313             ks[4] ^= k.v[i];                                        \
 314         }                                                               \
 315                                                                         \
 316     /* Insert initial key before round 0 */                             \
 317     X.v[0] += ks[0]; X.v[1] += ks[1]; X.v[2] += ks[2]; X.v[3] += ks[3]; \
 318                                                                         \
 319     if(Nrounds>0){                                                      \
 320         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
 321         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
 322     }                                                                   \
 323     if(Nrounds>1){                                                      \
 324         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
 325         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
 326     }                                                                   \
 327     if(Nrounds>2){                                                      \
 328         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
 329         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
 330     }                                                                   \
 331     if(Nrounds>3){                                                      \
 332         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
 333         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
 334     }                                                                   \
 335     if(Nrounds>3){                                                      \
 336         /* InjectKey(r=1) */                                            \
 337         X.v[0] += ks[1]; X.v[1] += ks[2]; X.v[2] += ks[3]; X.v[3] += ks[4]; \
 338         X.v[4-1] += 1;     /* X.v[WCNT4-1] += r  */                 \
 339     }                                                                   \
 340                                                                         \
 341     if(Nrounds>4){                                                      \
 342         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
 343         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
 344     }                                                                   \
 345     if(Nrounds>5){                                                      \
 346         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
 347         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
 348     }                                                                   \
 349     if(Nrounds>6){                                                      \
 350         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
 351         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
 352     }                                                                   \
 353     if(Nrounds>7){                                                      \
 354         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
 355         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
 356     }                                                                   \
 357     if(Nrounds>7){                                                      \
 358         /* InjectKey(r=2) */                                            \
 359         X.v[0] += ks[2]; X.v[1] += ks[3]; X.v[2] += ks[4]; X.v[3] += ks[0]; \
 360         X.v[4-1] += 2;     /* X.v[WCNT4-1] += r  */                 \
 361     }                                                                   \
 362                                                                         \
 363     if(Nrounds>8){                                                      \
 364         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
 365         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
 366     }                                                                   \
 367     if(Nrounds>9){                                                      \
 368         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
 369         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
 370     }                                                                   \
 371     if(Nrounds>10){                                                     \
 372         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
 373         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
 374     }                                                                   \
 375     if(Nrounds>11){                                                     \
 376         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
 377         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
 378     }                                                                   \
 379     if(Nrounds>11){                                                     \
 380         /* InjectKey(r=3) */                                            \
 381         X.v[0] += ks[3]; X.v[1] += ks[4]; X.v[2] += ks[0]; X.v[3] += ks[1]; \
 382         X.v[4-1] += 3;     /* X.v[WCNT4-1] += r  */                 \
 383     }                                                                   \
 384                                                                         \
 385     if(Nrounds>12){                                                     \
 386         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
 387         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
 388     }                                                                   \
 389     if(Nrounds>13){                                                     \
 390         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
 391         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
 392     }                                                                   \
 393     if(Nrounds>14){                                                     \
 394         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
 395         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
 396     }                                                                   \
 397     if(Nrounds>15){                                                     \
 398         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
 399         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
 400     }                                                                   \
 401     if(Nrounds>15){                                                     \
 402         /* InjectKey(r=1) */                                            \
 403         X.v[0] += ks[4]; X.v[1] += ks[0]; X.v[2] += ks[1]; X.v[3] += ks[2]; \
 404         X.v[4-1] += 4;     /* X.v[WCNT4-1] += r  */                 \
 405     }                                                                   \
 406                                                                         \
 407     if(Nrounds>16){                                                     \
 408         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
 409         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
 410     }                                                                   \
 411     if(Nrounds>17){                                                     \
 412         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
 413         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
 414     }                                                                   \
 415     if(Nrounds>18){                                                     \
 416         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
 417         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
 418     }                                                                   \
 419     if(Nrounds>19){                                                     \
 420         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
 421         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
 422     }                                                                   \
 423     if(Nrounds>19){                                                     \
 424         /* InjectKey(r=1) */                                            \
 425         X.v[0] += ks[0]; X.v[1] += ks[1]; X.v[2] += ks[2]; X.v[3] += ks[3]; \
 426         X.v[4-1] += 5;     /* X.v[WCNT4-1] += r  */                 \
 427     }                                                                   \
 428                                                                         \
 429     if(Nrounds>20){                                                     \
 430         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
 431         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
 432     }                                                                   \
 433     if(Nrounds>21){                                                     \
 434         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
 435         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
 436     }                                                                   \
 437     if(Nrounds>22){                                                     \
 438         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
 439         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
 440     }                                                                   \
 441     if(Nrounds>23){                                                     \
 442         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
 443         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
 444     }                                                                   \
 445     if(Nrounds>23){                                                     \
 446         /* InjectKey(r=1) */                                            \
 447         X.v[0] += ks[1]; X.v[1] += ks[2]; X.v[2] += ks[3]; X.v[3] += ks[4]; \
 448         X.v[4-1] += 6;     /* X.v[WCNT4-1] += r  */                 \
 449     }                                                                   \
 450                                                                         \
 451     if(Nrounds>24){                                                     \
 452         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
 453         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
 454     }                                                                   \
 455     if(Nrounds>25){                                                     \
 456         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
 457         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
 458     }                                                                   \
 459     if(Nrounds>26){                                                     \
 460         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
 461         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
 462     }                                                                   \
 463     if(Nrounds>27){                                                     \
 464         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
 465         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
 466     }                                                                   \
 467     if(Nrounds>27){                                                     \
 468         /* InjectKey(r=1) */                                            \
 469         X.v[0] += ks[2]; X.v[1] += ks[3]; X.v[2] += ks[4]; X.v[3] += ks[0]; \
 470         X.v[4-1] += 7;     /* X.v[WCNT4-1] += r  */                 \
 471     }                                                                   \
 472                                                                         \
 473     if(Nrounds>28){                                                     \
 474         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
 475         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
 476     }                                                                   \
 477     if(Nrounds>29){                                                     \
 478         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
 479         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
 480     }                                                                   \
 481     if(Nrounds>30){                                                     \
 482         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
 483         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
 484     }                                                                   \
 485     if(Nrounds>31){                                                     \
 486         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
 487         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
 488     }                                                                   \
 489     if(Nrounds>31){                                                     \
 490         /* InjectKey(r=1) */                                            \
 491         X.v[0] += ks[3]; X.v[1] += ks[4]; X.v[2] += ks[0]; X.v[3] += ks[1]; \
 492         X.v[4-1] += 8;     /* X.v[WCNT4-1] += r  */                 \
 493     }                                                                   \
 494                                                                         \
 495     if(Nrounds>32){                                                     \
 496         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
 497         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
 498     }                                                                   \
 499     if(Nrounds>33){                                                     \
 500         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
 501         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
 502     }                                                                   \
 503     if(Nrounds>34){                                                     \
 504         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
 505         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
 506     }                                                                   \
 507     if(Nrounds>35){                                                     \
 508         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
 509         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
 510     }                                                                   \
 511     if(Nrounds>35){                                                     \
 512         /* InjectKey(r=1) */                                            \
 513         X.v[0] += ks[4]; X.v[1] += ks[0]; X.v[2] += ks[1]; X.v[3] += ks[2]; \
 514         X.v[4-1] += 9;     /* X.v[WCNT4-1] += r  */                 \
 515     }                                                                   \
 516                                                                         \
 517     if(Nrounds>36){                                                     \
 518         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
 519         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
 520     }                                                                   \
 521     if(Nrounds>37){                                                     \
 522         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
 523         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
 524     }                                                                   \
 525     if(Nrounds>38){                                                     \
 526         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
 527         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
 528     }                                                                   \
 529     if(Nrounds>39){                                                     \
 530         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
 531         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
 532     }                                                                   \
 533     if(Nrounds>39){                                                     \
 534         /* InjectKey(r=1) */                                            \
 535         X.v[0] += ks[0]; X.v[1] += ks[1]; X.v[2] += ks[2]; X.v[3] += ks[3]; \
 536         X.v[4-1] += 10;     /* X.v[WCNT4-1] += r  */                 \
 537     }                                                                   \
 538                                                                         \
 539     if(Nrounds>40){                                                     \
 540         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
 541         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
 542     }                                                                   \
 543     if(Nrounds>41){                                                     \
 544         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
 545         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
 546     }                                                                   \
 547     if(Nrounds>42){                                                     \
 548         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
 549         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
 550     }                                                                   \
 551     if(Nrounds>43){                                                     \
 552         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
 553         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
 554     }                                                                   \
 555     if(Nrounds>43){                                                     \
 556         /* InjectKey(r=1) */                                            \
 557         X.v[0] += ks[1]; X.v[1] += ks[2]; X.v[2] += ks[3]; X.v[3] += ks[4]; \
 558         X.v[4-1] += 11;     /* X.v[WCNT4-1] += r  */                \
 559     }                                                                   \
 560                                                                         \
 561     if(Nrounds>44){                                                     \
 562         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
 563         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
 564     }                                                                   \
 565     if(Nrounds>45){                                                     \
 566         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
 567         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
 568     }                                                                   \
 569     if(Nrounds>46){                                                     \
 570         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
 571         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
 572     }                                                                   \
 573     if(Nrounds>47){                                                     \
 574         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
 575         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
 576     }                                                                   \
 577     if(Nrounds>47){                                                     \
 578         /* InjectKey(r=1) */                                            \
 579         X.v[0] += ks[2]; X.v[1] += ks[3]; X.v[2] += ks[4]; X.v[3] += ks[0]; \
 580         X.v[4-1] += 12;     /* X.v[WCNT4-1] += r  */                 \
 581     }                                                                   \
 582                                                                         \
 583     if(Nrounds>48){                                                     \
 584         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
 585         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
 586     }                                                                   \
 587     if(Nrounds>49){                                                     \
 588         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
 589         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
 590     }                                                                   \
 591     if(Nrounds>50){                                                     \
 592         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
 593         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
 594     }                                                                   \
 595     if(Nrounds>51){                                                     \
 596         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
 597         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
 598     }                                                                   \
 599     if(Nrounds>51){                                                     \
 600         /* InjectKey(r=1) */                                            \
 601         X.v[0] += ks[3]; X.v[1] += ks[4]; X.v[2] += ks[0]; X.v[3] += ks[1]; \
 602         X.v[4-1] += 13;     /* X.v[WCNT4-1] += r  */                 \
 603     }                                                                   \
 604                                                                         \
 605     if(Nrounds>52){                                                     \
 606         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
 607         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
 608     }                                                                   \
 609     if(Nrounds>53){                                                     \
 610         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
 611         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
 612     }                                                                   \
 613     if(Nrounds>54){                                                     \
 614         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
 615         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
 616     }                                                                   \
 617     if(Nrounds>55){                                                     \
 618         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
 619         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
 620     }                                                                   \
 621     if(Nrounds>55){                                                     \
 622         /* InjectKey(r=1) */                                            \
 623         X.v[0] += ks[4]; X.v[1] += ks[0]; X.v[2] += ks[1]; X.v[3] += ks[2]; \
 624         X.v[4-1] += 14;     /* X.v[WCNT4-1] += r  */                 \
 625     }                                                                   \
 626                                                                         \
 627     if(Nrounds>56){                                                     \
 628         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
 629         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
 630     }                                                                   \
 631     if(Nrounds>57){                                                     \
 632         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
 633         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
 634     }                                                                   \
 635     if(Nrounds>58){                                                     \
 636         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
 637         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
 638     }                                                                   \
 639     if(Nrounds>59){                                                     \
 640         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
 641         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
 642     }                                                                   \
 643     if(Nrounds>59){                                                     \
 644         /* InjectKey(r=1) */                                            \
 645         X.v[0] += ks[0]; X.v[1] += ks[1]; X.v[2] += ks[2]; X.v[3] += ks[3]; \
 646         X.v[4-1] += 15;     /* X.v[WCNT4-1] += r  */                 \
 647     }                                                                   \
 648                                                                         \
 649     if(Nrounds>60){                                                     \
 650         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
 651         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
 652     }                                                                   \
 653     if(Nrounds>61){                                                     \
 654         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
 655         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
 656     }                                                                   \
 657     if(Nrounds>62){                                                     \
 658         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
 659         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
 660     }                                                                   \
 661     if(Nrounds>63){                                                     \
 662         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
 663         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
 664     }                                                                   \
 665     if(Nrounds>63){                                                     \
 666         /* InjectKey(r=1) */                                            \
 667         X.v[0] += ks[1]; X.v[1] += ks[2]; X.v[2] += ks[3]; X.v[3] += ks[4]; \
 668         X.v[4-1] += 16;     /* X.v[WCNT4-1] += r  */                 \
 669     }                                                                   \
 670                                                                         \
 671     if(Nrounds>64){                                                     \
 672         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_0_0); X.v[1] ^= X.v[0]; \
 673         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_0_1); X.v[3] ^= X.v[2]; \
 674     }                                                                   \
 675     if(Nrounds>65){                                                     \
 676         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_1_0); X.v[3] ^= X.v[0]; \
 677         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_1_1); X.v[1] ^= X.v[2]; \
 678     }                                                                   \
 679     if(Nrounds>66){                                                     \
 680         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_2_0); X.v[1] ^= X.v[0]; \
 681         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_2_1); X.v[3] ^= X.v[2]; \
 682     }                                                                   \
 683     if(Nrounds>67){                                                     \
 684         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_3_0); X.v[3] ^= X.v[0]; \
 685         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_3_1); X.v[1] ^= X.v[2]; \
 686     }                                                                   \
 687     if(Nrounds>67){                                                     \
 688         /* InjectKey(r=1) */                                            \
 689         X.v[0] += ks[2]; X.v[1] += ks[3]; X.v[2] += ks[4]; X.v[3] += ks[0]; \
 690         X.v[4-1] += 17;     /* X.v[WCNT4-1] += r  */                 \
 691     }                                                                   \
 692                                                                         \
 693     if(Nrounds>68){                                                     \
 694         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_4_0); X.v[1] ^= X.v[0]; \
 695         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_4_1); X.v[3] ^= X.v[2]; \
 696     }                                                                   \
 697     if(Nrounds>69){                                                     \
 698         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_5_0); X.v[3] ^= X.v[0]; \
 699         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_5_1); X.v[1] ^= X.v[2]; \
 700     }                                                                   \
 701     if(Nrounds>70){                                                     \
 702         X.v[0] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_6_0); X.v[1] ^= X.v[0]; \
 703         X.v[2] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_6_1); X.v[3] ^= X.v[2]; \
 704     }                                                                   \
 705     if(Nrounds>71){                                                     \
 706         X.v[0] += X.v[3]; X.v[3] = RotL_##W(X.v[3],R_##W##x4_7_0); X.v[3] ^= X.v[0]; \
 707         X.v[2] += X.v[1]; X.v[1] = RotL_##W(X.v[1],R_##W##x4_7_1); X.v[1] ^= X.v[2]; \
 708     }                                                                   \
 709     if(Nrounds>71){                                                     \
 710         /* InjectKey(r=1) */                                            \
 711         X.v[0] += ks[3]; X.v[1] += ks[4]; X.v[2] += ks[0]; X.v[3] += ks[1]; \
 712         X.v[4-1] += 18;     /* X.v[WCNT4-1] += r  */                 \
 713     }                                                                   \
 714                                                                         \
 715     return X;                                                           \
 716 }                                                                       \
 717  /** @ingroup ThreefryNxW */                                            \
 718 enum r123_enum_threefry4x##W { threefry4x##W##_rounds = THREEFRY4x##W##_DEFAULT_ROUNDS };       \
 719 R123_CUDA_DEVICE R123_STATIC_INLINE R123_FORCE_INLINE(threefry4x##W##_ctr_t threefry4x##W(threefry4x##W##_ctr_t in, threefry4x##W##_key_t k)); \
 720 R123_CUDA_DEVICE R123_STATIC_INLINE                                     \
 721 threefry4x##W##_ctr_t threefry4x##W(threefry4x##W##_ctr_t in, threefry4x##W##_key_t k){ \
 722     return threefry4x##W##_R(threefry4x##W##_rounds, in, k);            \
 723 }
 724 /** \endcond */
 725
 726 _threefry2x_tpl(64)
 727 _threefry2x_tpl(32)
 728 _threefry4x_tpl(64)
 729 _threefry4x_tpl(32)
 730
 731 /* gcc4.5 and 4.6 seem to optimize a macro-ized threefryNxW better
 732    than a static inline function.  Why?  */
 733 #define threefry2x32(c,k) threefry2x32_R(threefry2x32_rounds, c, k)
 734 #define threefry4x32(c,k) threefry4x32_R(threefry4x32_rounds, c, k)
 735 #define threefry2x64(c,k) threefry2x64_R(threefry2x64_rounds, c, k)
 736 #define threefry4x64(c,k) threefry4x64_R(threefry4x64_rounds, c, k)
 737
 738 #ifdef __cplusplus
 739 /** \cond HIDDEN_FROM_DOXYGEN */
 740 #define _threefryNxWclass_tpl(NxW)                                      \
 741 namespace r123{                                                     \
 742 template<unsigned int R>                                                  \
 743  struct Threefry##NxW##_R{                                              \
 744     typedef threefry##NxW##_ctr_t ctr_type;                             \
 745     typedef threefry##NxW##_key_t key_type;                             \
 746     typedef threefry##NxW##_key_t ukey_type;                            \
 747     static const unsigned int rounds=R;                                 \
 748    inline R123_CUDA_DEVICE R123_FORCE_INLINE(ctr_type operator()(ctr_type ctr, key_type key)){ \
 749         R123_STATIC_ASSERT(R<=72, "threefry is only unrolled up to 72 rounds\n"); \
 750         return threefry##NxW##_R(R, ctr, key);                              \
 751     }                                                                   \
 752 };                                                                      \
 753  typedef Threefry##NxW##_R<threefry##NxW##_rounds> Threefry##NxW;       \
 754 } // namespace r123
 755
 756 /** \endcond */
 757
 758 _threefryNxWclass_tpl(2x32)
 759 _threefryNxWclass_tpl(4x32)
 760 _threefryNxWclass_tpl(2x64)
 761 _threefryNxWclass_tpl(4x64)
 762
 763 /* The _tpl macros don't quite work to do string-pasting inside comments.
 764    so we just write out the boilerplate documentation four times... */
 765
 766 /**
 767 @defgroup ThreefryNxW Threefry Classes and Typedefs
 768
 769 The ThreefryNxW classes export the member functions, typedefs and
 770 operator overloads required by a @ref CBRNG "CBRNG" class.
 771
 772 As described in
 773 <a href="http://dl.acm.org/citation.cfm?doid=2063405"><i>Parallel Random Numbers:  As Easy as 1, 2, 3</i> </a>,
 774 the Threefry family is closely related to the Threefish block cipher from
 775 <a href="http://www.skein-hash.info/"> Skein Hash Function</a>.
 776 Threefry is \b not suitable for cryptographic use.
 777
 778 Threefry uses integer addition, bitwise rotation, xor and permutation of words to randomize its output.
 779
 780 @class r123::Threefry2x32_R
 781 @ingroup ThreefryNxW
 782
 783 exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
 784
 785 The template argument, ROUNDS, is the number of times the Threefry round
 786 function will be applied.
 787
 788 As of September 2011, the authors know of no statistical flaws with
 789 ROUNDS=13 or more for Threefry2x32.
 790
 791 @typedef r123::Threefry2x32
 792 @ingroup ThreefryNxW
 793   Threefry2x32 is equivalent to Threefry2x32_R<20>.    With 20 rounds,
 794   Threefry2x32 has a considerable safety margin over the minimum number
 795   of rounds with no known statistical flaws, but still has excellent
 796    performance.
 797
 798 @class r123::Threefry2x64_R
 799 @ingroup ThreefryNxW
 800
 801 exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
 802
 803 The template argument, ROUNDS, is the number of times the Threefry round
 804 function will be applied.
 805
 806 In November 2011, the authors discovered that 13 rounds of
 807 Threefry2x64 sequenced by strided, interleaved key and counter
 808 increments failed a very long (longer than the default BigCrush
 809 length) WeightDistrub test.  At the same time, it was confirmed that
 810 14 rounds passes much longer tests (up to 5x10^12 samples) of a
 811 similar nature.  The authors know of no statistical flaws with
 812 ROUNDS=14 or more for Threefry2x64.
 813
 814 @typedef r123::Threefry2x64
 815 @ingroup ThreefryNxW
 816   Threefry2x64 is equivalent to Threefry2x64_R<20>.    With 20 rounds,
 817   Threefry2x64 has a considerable safety margin over the minimum number
 818   of rounds with no known statistical flaws, but still has excellent
 819    performance.
 820
 821
 822
 823 @class r123::Threefry4x32_R
 824 @ingroup ThreefryNxW
 825
 826 exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
 827
 828 The template argument, ROUNDS, is the number of times the Threefry round
 829 function will be applied.
 830
 831 As of September 2011, the authors know of no statistical flaws with
 832 ROUNDS=12 or more for Threefry4x32.
 833
 834 @typedef r123::Threefry4x32
 835 @ingroup ThreefryNxW
 836   Threefry4x32 is equivalent to Threefry4x32_R<20>.    With 20 rounds,
 837   Threefry4x32 has a considerable safety margin over the minimum number
 838   of rounds with no known statistical flaws, but still has excellent
 839    performance.
 840
 841
 842
 843 @class r123::Threefry4x64_R
 844 @ingroup ThreefryNxW
 845
 846 exports the member functions, typedefs and operator overloads required by a @ref CBRNG "CBRNG" class.
 847
 848 The template argument, ROUNDS, is the number of times the Threefry round
 849 function will be applied.
 850
 851 As of September 2011, the authors know of no statistical flaws with
 852 ROUNDS=12 or more for Threefry4x64.
 853
 854 @typedef r123::Threefry4x64
 855 @ingroup ThreefryNxW
 856   Threefry4x64 is equivalent to Threefry4x64_R<20>.    With 20 rounds,
 857   Threefry4x64 has a considerable safety margin over the minimum number
 858   of rounds with no known statistical flaws, but still has excellent
 859    performance.
 860 */
 861
 862 #endif
 863
 864 #endif