CoCalc -- blake2b

GitHub Repository: folium-app/Folium
Path: blob/a-new-beginning/SharedDependencies/Sources/cryptopp/blake2b_simd.cpp
² views
1
// blake2_simd.cpp - written and placed in the public domain by
2
//                   Samuel Neves, Jeffrey Walton, Uri Blumenthal
3
//                   and Marcel Raad.
4
//
5
//    This source file uses intrinsics to gain access to ARMv7a/ARMv8a
6
//    NEON, Power8 and SSE4.1 instructions. A separate source file is
7
//    needed because additional CXXFLAGS are required to enable the
8
//    appropriate instructions sets in some build configurations.
9

10
#include "pch.h"
11
#include "config.h"
12
#include "misc.h"
13
#include "blake2.h"
14

15
// Uncomment for benchmarking C++ against SSE2 or NEON.
16
// Do so in both blake2.cpp and blake2_simd.cpp.
17
// #undef CRYPTOPP_SSE41_AVAILABLE
18
// #undef CRYPTOPP_ARM_NEON_AVAILABLE
19
// #undef CRYPTOPP_ALTIVEC_AVAILABLE
20

21
// Disable NEON/ASIMD for Cortex-A53 and A57. The shifts are too slow and C/C++ is about
22
// 3 cpb faster than NEON/ASIMD. Also see http://github.com/weidai11/cryptopp/issues/367.
23
#if (defined(__aarch32__) || defined(__aarch64__)) && defined(CRYPTOPP_SLOW_ARMV8_SHIFT)
24
# undef CRYPTOPP_ARM_NEON_AVAILABLE
25
#endif
26

27
// BLAKE2s bug on AIX 7.1 (POWER7) with XLC 12.01
28
// https://github.com/weidai11/cryptopp/issues/743
29
#if defined(__xlC__) && (__xlC__ < 0x0d01)
30
# define CRYPTOPP_DISABLE_ALTIVEC 1
31
# undef CRYPTOPP_POWER8_AVAILABLE
32
# undef CRYPTOPP_ALTIVEC_AVAILABLE
33
#endif
34

35
#if defined(__XOP__)
36
# if defined(CRYPTOPP_GCC_COMPATIBLE)
37
#  include <x86intrin.h>
38
# endif
39
# include <ammintrin.h>
40
#endif  // XOP
41

42
#if (CRYPTOPP_SSE41_AVAILABLE)
43
# include <emmintrin.h>
44
# include <tmmintrin.h>
45
# include <smmintrin.h>
46
#endif
47

48
#if (CRYPTOPP_ARM_NEON_HEADER)
49
# include <arm_neon.h>
50
#endif
51

52
#if (CRYPTOPP_ARM_ACLE_HEADER)
53
# include <stdint.h>
54
# include <arm_acle.h>
55
#endif
56

57
#if (CRYPTOPP_POWER8_AVAILABLE)
58
# include "ppc_simd.h"
59
#endif
60

61
#if defined(CRYPTOPP_GCC_DIAGNOSTIC_AVAILABLE)
62
/* Ignore "warning: vec_lvsl is deprecated..." */
63
# pragma GCC diagnostic ignored "-Wdeprecated"
64
#endif
65

66
// Squash MS LNK4221 and libtool warnings
67
extern const char BLAKE2B_SIMD_FNAME[] = __FILE__;
68

69
NAMESPACE_BEGIN(CryptoPP)
70

71
// Exported by blake2.cpp
72
extern const word32 BLAKE2S_IV[8];
73
extern const word64 BLAKE2B_IV[8];
74

75
#if CRYPTOPP_SSE41_AVAILABLE
76

77
#define LOADU(p)  _mm_loadu_si128((const __m128i *)(const void*)(p))
78
#define STOREU(p,r) _mm_storeu_si128((__m128i *)(void*)(p), r)
79
#define TOF(reg) _mm_castsi128_ps((reg))
80
#define TOI(reg) _mm_castps_si128((reg))
81

82
void BLAKE2_Compress64_SSE4(const byte* input, BLAKE2b_State& state)
83
{
84
    #define BLAKE2B_LOAD_MSG_0_1(b0, b1) \
85
    do { \
86
    b0 = _mm_unpacklo_epi64(m0, m1); \
87
    b1 = _mm_unpacklo_epi64(m2, m3); \
88
    } while(0)
89

90
    #define BLAKE2B_LOAD_MSG_0_2(b0, b1) \
91
    do { \
92
    b0 = _mm_unpackhi_epi64(m0, m1); \
93
    b1 = _mm_unpackhi_epi64(m2, m3); \
94
    } while(0)
95

96
    #define BLAKE2B_LOAD_MSG_0_3(b0, b1) \
97
    do { \
98
    b0 = _mm_unpacklo_epi64(m4, m5); \
99
    b1 = _mm_unpacklo_epi64(m6, m7); \
100
    } while(0)
101

102
    #define BLAKE2B_LOAD_MSG_0_4(b0, b1) \
103
    do { \
104
    b0 = _mm_unpackhi_epi64(m4, m5); \
105
    b1 = _mm_unpackhi_epi64(m6, m7); \
106
    } while(0)
107

108
    #define BLAKE2B_LOAD_MSG_1_1(b0, b1) \
109
    do { \
110
    b0 = _mm_unpacklo_epi64(m7, m2); \
111
    b1 = _mm_unpackhi_epi64(m4, m6); \
112
    } while(0)
113

114
    #define BLAKE2B_LOAD_MSG_1_2(b0, b1) \
115
    do { \
116
    b0 = _mm_unpacklo_epi64(m5, m4); \
117
    b1 = _mm_alignr_epi8(m3, m7, 8); \
118
    } while(0)
119

120
    #define BLAKE2B_LOAD_MSG_1_3(b0, b1) \
121
    do { \
122
    b0 = _mm_shuffle_epi32(m0, _MM_SHUFFLE(1,0,3,2)); \
123
    b1 = _mm_unpackhi_epi64(m5, m2); \
124
    } while(0)
125

126
    #define BLAKE2B_LOAD_MSG_1_4(b0, b1) \
127
    do { \
128
    b0 = _mm_unpacklo_epi64(m6, m1); \
129
    b1 = _mm_unpackhi_epi64(m3, m1); \
130
    } while(0)
131

132
    #define BLAKE2B_LOAD_MSG_2_1(b0, b1) \
133
    do { \
134
    b0 = _mm_alignr_epi8(m6, m5, 8); \
135
    b1 = _mm_unpackhi_epi64(m2, m7); \
136
    } while(0)
137

138
    #define BLAKE2B_LOAD_MSG_2_2(b0, b1) \
139
    do { \
140
    b0 = _mm_unpacklo_epi64(m4, m0); \
141
    b1 = _mm_blend_epi16(m1, m6, 0xF0); \
142
    } while(0)
143

144
    #define BLAKE2B_LOAD_MSG_2_3(b0, b1) \
145
    do { \
146
    b0 = _mm_blend_epi16(m5, m1, 0xF0); \
147
    b1 = _mm_unpackhi_epi64(m3, m4); \
148
    } while(0)
149

150
    #define BLAKE2B_LOAD_MSG_2_4(b0, b1) \
151
    do { \
152
    b0 = _mm_unpacklo_epi64(m7, m3); \
153
    b1 = _mm_alignr_epi8(m2, m0, 8); \
154
    } while(0)
155

156
    #define BLAKE2B_LOAD_MSG_3_1(b0, b1) \
157
    do { \
158
    b0 = _mm_unpackhi_epi64(m3, m1); \
159
    b1 = _mm_unpackhi_epi64(m6, m5); \
160
    } while(0)
161

162
    #define BLAKE2B_LOAD_MSG_3_2(b0, b1) \
163
    do { \
164
    b0 = _mm_unpackhi_epi64(m4, m0); \
165
    b1 = _mm_unpacklo_epi64(m6, m7); \
166
    } while(0)
167

168
    #define BLAKE2B_LOAD_MSG_3_3(b0, b1) \
169
    do { \
170
    b0 = _mm_blend_epi16(m1, m2, 0xF0); \
171
    b1 = _mm_blend_epi16(m2, m7, 0xF0); \
172
    } while(0)
173

174
    #define BLAKE2B_LOAD_MSG_3_4(b0, b1) \
175
    do { \
176
    b0 = _mm_unpacklo_epi64(m3, m5); \
177
    b1 = _mm_unpacklo_epi64(m0, m4); \
178
    } while(0)
179

180
    #define BLAKE2B_LOAD_MSG_4_1(b0, b1) \
181
    do { \
182
    b0 = _mm_unpackhi_epi64(m4, m2); \
183
    b1 = _mm_unpacklo_epi64(m1, m5); \
184
    } while(0)
185

186
    #define BLAKE2B_LOAD_MSG_4_2(b0, b1) \
187
    do { \
188
    b0 = _mm_blend_epi16(m0, m3, 0xF0); \
189
    b1 = _mm_blend_epi16(m2, m7, 0xF0); \
190
    } while(0)
191

192
    #define BLAKE2B_LOAD_MSG_4_3(b0, b1) \
193
    do { \
194
    b0 = _mm_blend_epi16(m7, m5, 0xF0); \
195
    b1 = _mm_blend_epi16(m3, m1, 0xF0); \
196
    } while(0)
197

198
    #define BLAKE2B_LOAD_MSG_4_4(b0, b1) \
199
    do { \
200
    b0 = _mm_alignr_epi8(m6, m0, 8); \
201
    b1 = _mm_blend_epi16(m4, m6, 0xF0); \
202
    } while(0)
203

204
    #define BLAKE2B_LOAD_MSG_5_1(b0, b1) \
205
    do { \
206
    b0 = _mm_unpacklo_epi64(m1, m3); \
207
    b1 = _mm_unpacklo_epi64(m0, m4); \
208
    } while(0)
209

210
    #define BLAKE2B_LOAD_MSG_5_2(b0, b1) \
211
    do { \
212
    b0 = _mm_unpacklo_epi64(m6, m5); \
213
    b1 = _mm_unpackhi_epi64(m5, m1); \
214
    } while(0)
215

216
    #define BLAKE2B_LOAD_MSG_5_3(b0, b1) \
217
    do { \
218
    b0 = _mm_blend_epi16(m2, m3, 0xF0); \
219
    b1 = _mm_unpackhi_epi64(m7, m0); \
220
    } while(0)
221

222
    #define BLAKE2B_LOAD_MSG_5_4(b0, b1) \
223
    do { \
224
    b0 = _mm_unpackhi_epi64(m6, m2); \
225
    b1 = _mm_blend_epi16(m7, m4, 0xF0); \
226
    } while(0)
227

228
    #define BLAKE2B_LOAD_MSG_6_1(b0, b1) \
229
    do { \
230
    b0 = _mm_blend_epi16(m6, m0, 0xF0); \
231
    b1 = _mm_unpacklo_epi64(m7, m2); \
232
    } while(0)
233

234
    #define BLAKE2B_LOAD_MSG_6_2(b0, b1) \
235
    do { \
236
    b0 = _mm_unpackhi_epi64(m2, m7); \
237
    b1 = _mm_alignr_epi8(m5, m6, 8); \
238
    } while(0)
239

240
    #define BLAKE2B_LOAD_MSG_6_3(b0, b1) \
241
    do { \
242
    b0 = _mm_unpacklo_epi64(m0, m3); \
243
    b1 = _mm_shuffle_epi32(m4, _MM_SHUFFLE(1,0,3,2)); \
244
    } while(0)
245

246
    #define BLAKE2B_LOAD_MSG_6_4(b0, b1) \
247
    do { \
248
    b0 = _mm_unpackhi_epi64(m3, m1); \
249
    b1 = _mm_blend_epi16(m1, m5, 0xF0); \
250
    } while(0)
251

252
    #define BLAKE2B_LOAD_MSG_7_1(b0, b1) \
253
    do { \
254
    b0 = _mm_unpackhi_epi64(m6, m3); \
255
    b1 = _mm_blend_epi16(m6, m1, 0xF0); \
256
    } while(0)
257

258
    #define BLAKE2B_LOAD_MSG_7_2(b0, b1) \
259
    do { \
260
    b0 = _mm_alignr_epi8(m7, m5, 8); \
261
    b1 = _mm_unpackhi_epi64(m0, m4); \
262
    } while(0)
263

264
    #define BLAKE2B_LOAD_MSG_7_3(b0, b1) \
265
    do { \
266
    b0 = _mm_unpackhi_epi64(m2, m7); \
267
    b1 = _mm_unpacklo_epi64(m4, m1); \
268
    } while(0)
269

270
    #define BLAKE2B_LOAD_MSG_7_4(b0, b1) \
271
    do { \
272
    b0 = _mm_unpacklo_epi64(m0, m2); \
273
    b1 = _mm_unpacklo_epi64(m3, m5); \
274
    } while(0)
275

276
    #define BLAKE2B_LOAD_MSG_8_1(b0, b1) \
277
    do { \
278
    b0 = _mm_unpacklo_epi64(m3, m7); \
279
    b1 = _mm_alignr_epi8(m0, m5, 8); \
280
    } while(0)
281

282
    #define BLAKE2B_LOAD_MSG_8_2(b0, b1) \
283
    do { \
284
    b0 = _mm_unpackhi_epi64(m7, m4); \
285
    b1 = _mm_alignr_epi8(m4, m1, 8); \
286
    } while(0)
287

288
    #define BLAKE2B_LOAD_MSG_8_3(b0, b1) \
289
    do { \
290
    b0 = m6; \
291
    b1 = _mm_alignr_epi8(m5, m0, 8); \
292
    } while(0)
293

294
    #define BLAKE2B_LOAD_MSG_8_4(b0, b1) \
295
    do { \
296
    b0 = _mm_blend_epi16(m1, m3, 0xF0); \
297
    b1 = m2; \
298
    } while(0)
299

300
    #define BLAKE2B_LOAD_MSG_9_1(b0, b1) \
301
    do { \
302
    b0 = _mm_unpacklo_epi64(m5, m4); \
303
    b1 = _mm_unpackhi_epi64(m3, m0); \
304
    } while(0)
305

306
    #define BLAKE2B_LOAD_MSG_9_2(b0, b1) \
307
    do { \
308
    b0 = _mm_unpacklo_epi64(m1, m2); \
309
    b1 = _mm_blend_epi16(m3, m2, 0xF0); \
310
    } while(0)
311

312
    #define BLAKE2B_LOAD_MSG_9_3(b0, b1) \
313
    do { \
314
    b0 = _mm_unpackhi_epi64(m7, m4); \
315
    b1 = _mm_unpackhi_epi64(m1, m6); \
316
    } while(0)
317

318
    #define BLAKE2B_LOAD_MSG_9_4(b0, b1) \
319
    do { \
320
    b0 = _mm_alignr_epi8(m7, m5, 8); \
321
    b1 = _mm_unpacklo_epi64(m6, m0); \
322
    } while(0)
323

324
    #define BLAKE2B_LOAD_MSG_10_1(b0, b1) \
325
    do { \
326
    b0 = _mm_unpacklo_epi64(m0, m1); \
327
    b1 = _mm_unpacklo_epi64(m2, m3); \
328
    } while(0)
329

330
    #define BLAKE2B_LOAD_MSG_10_2(b0, b1) \
331
    do { \
332
    b0 = _mm_unpackhi_epi64(m0, m1); \
333
    b1 = _mm_unpackhi_epi64(m2, m3); \
334
    } while(0)
335

336
    #define BLAKE2B_LOAD_MSG_10_3(b0, b1) \
337
    do { \
338
    b0 = _mm_unpacklo_epi64(m4, m5); \
339
    b1 = _mm_unpacklo_epi64(m6, m7); \
340
    } while(0)
341

342
    #define BLAKE2B_LOAD_MSG_10_4(b0, b1) \
343
    do { \
344
    b0 = _mm_unpackhi_epi64(m4, m5); \
345
    b1 = _mm_unpackhi_epi64(m6, m7); \
346
    } while(0)
347

348
    #define BLAKE2B_LOAD_MSG_11_1(b0, b1) \
349
    do { \
350
    b0 = _mm_unpacklo_epi64(m7, m2); \
351
    b1 = _mm_unpackhi_epi64(m4, m6); \
352
    } while(0)
353

354
    #define BLAKE2B_LOAD_MSG_11_2(b0, b1) \
355
    do { \
356
    b0 = _mm_unpacklo_epi64(m5, m4); \
357
    b1 = _mm_alignr_epi8(m3, m7, 8); \
358
    } while(0)
359

360
    #define BLAKE2B_LOAD_MSG_11_3(b0, b1) \
361
    do { \
362
    b0 = _mm_shuffle_epi32(m0, _MM_SHUFFLE(1,0,3,2)); \
363
    b1 = _mm_unpackhi_epi64(m5, m2); \
364
    } while(0)
365

366
    #define BLAKE2B_LOAD_MSG_11_4(b0, b1) \
367
    do { \
368
    b0 = _mm_unpacklo_epi64(m6, m1); \
369
    b1 = _mm_unpackhi_epi64(m3, m1); \
370
    } while(0)
371

372
#ifdef __XOP__
373
# define MM_ROTI_EPI64(r, c) \
374
    _mm_roti_epi64(r, c)
375
#else
376
# define MM_ROTI_EPI64(x, c) \
377
      (-(c) == 32) ? _mm_shuffle_epi32((x), _MM_SHUFFLE(2,3,0,1))  \
378
    : (-(c) == 24) ? _mm_shuffle_epi8((x), r24) \
379
    : (-(c) == 16) ? _mm_shuffle_epi8((x), r16) \
380
    : (-(c) == 63) ? _mm_xor_si128(_mm_srli_epi64((x), -(c)), _mm_add_epi64((x), (x)))  \
381
    : _mm_xor_si128(_mm_srli_epi64((x), -(c)), _mm_slli_epi64((x), 64-(-(c))))
382
#endif
383

384
#define BLAKE2B_G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \
385
    row1l = _mm_add_epi64(_mm_add_epi64(row1l, b0), row2l); \
386
    row1h = _mm_add_epi64(_mm_add_epi64(row1h, b1), row2h); \
387
    \
388
    row4l = _mm_xor_si128(row4l, row1l); \
389
    row4h = _mm_xor_si128(row4h, row1h); \
390
    \
391
    row4l = MM_ROTI_EPI64(row4l, -32); \
392
    row4h = MM_ROTI_EPI64(row4h, -32); \
393
    \
394
    row3l = _mm_add_epi64(row3l, row4l); \
395
    row3h = _mm_add_epi64(row3h, row4h); \
396
    \
397
    row2l = _mm_xor_si128(row2l, row3l); \
398
    row2h = _mm_xor_si128(row2h, row3h); \
399
    \
400
    row2l = MM_ROTI_EPI64(row2l, -24); \
401
    row2h = MM_ROTI_EPI64(row2h, -24);
402

403
#define BLAKE2B_G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \
404
    row1l = _mm_add_epi64(_mm_add_epi64(row1l, b0), row2l); \
405
    row1h = _mm_add_epi64(_mm_add_epi64(row1h, b1), row2h); \
406
    \
407
    row4l = _mm_xor_si128(row4l, row1l); \
408
    row4h = _mm_xor_si128(row4h, row1h); \
409
    \
410
    row4l = MM_ROTI_EPI64(row4l, -16); \
411
    row4h = MM_ROTI_EPI64(row4h, -16); \
412
    \
413
    row3l = _mm_add_epi64(row3l, row4l); \
414
    row3h = _mm_add_epi64(row3h, row4h); \
415
    \
416
    row2l = _mm_xor_si128(row2l, row3l); \
417
    row2h = _mm_xor_si128(row2h, row3h); \
418
    \
419
    row2l = MM_ROTI_EPI64(row2l, -63); \
420
    row2h = MM_ROTI_EPI64(row2h, -63); \
421

422
#define BLAKE2B_DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \
423
    t0 = row4l;\
424
    t1 = row2l;\
425
    row4l = row3l;\
426
    row3l = row3h;\
427
    row3h = row4l;\
428
    row4l = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t0, t0)); \
429
    row4h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row4h, row4h)); \
430
    row2l = _mm_unpackhi_epi64(row2l, _mm_unpacklo_epi64(row2h, row2h)); \
431
    row2h = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(t1, t1))
432

433
#define BLAKE2B_UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \
434
    t0 = row3l;\
435
    row3l = row3h;\
436
    row3h = t0;\
437
    t0 = row2l;\
438
    t1 = row4l;\
439
    row2l = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(row2l, row2l)); \
440
    row2h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row2h, row2h)); \
441
    row4l = _mm_unpackhi_epi64(row4l, _mm_unpacklo_epi64(row4h, row4h)); \
442
    row4h = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t1, t1))
443

444
#define BLAKE2B_ROUND(r) \
445
    BLAKE2B_LOAD_MSG_ ##r ##_1(b0, b1); \
446
    BLAKE2B_G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
447
    BLAKE2B_LOAD_MSG_ ##r ##_2(b0, b1); \
448
    BLAKE2B_G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
449
    BLAKE2B_DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h); \
450
    BLAKE2B_LOAD_MSG_ ##r ##_3(b0, b1); \
451
    BLAKE2B_G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
452
    BLAKE2B_LOAD_MSG_ ##r ##_4(b0, b1); \
453
    BLAKE2B_G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
454
    BLAKE2B_UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h);
455

456
    __m128i row1l, row1h;
457
    __m128i row2l, row2h;
458
    __m128i row3l, row3h;
459
    __m128i row4l, row4h;
460
    __m128i b0, b1;
461
    __m128i t0, t1;
462

463
    const __m128i r16 = _mm_setr_epi8(2, 3, 4, 5, 6, 7, 0, 1, 10, 11, 12, 13, 14, 15, 8, 9);
464
    const __m128i r24 = _mm_setr_epi8(3, 4, 5, 6, 7, 0, 1, 2, 11, 12, 13, 14, 15, 8, 9, 10);
465

466
    const __m128i m0 = LOADU(input + 00);
467
    const __m128i m1 = LOADU(input + 16);
468
    const __m128i m2 = LOADU(input + 32);
469
    const __m128i m3 = LOADU(input + 48);
470
    const __m128i m4 = LOADU(input + 64);
471
    const __m128i m5 = LOADU(input + 80);
472
    const __m128i m6 = LOADU(input + 96);
473
    const __m128i m7 = LOADU(input + 112);
474

475
    row1l = LOADU(state.h()+0);
476
    row1h = LOADU(state.h()+2);
477
    row2l = LOADU(state.h()+4);
478
    row2h = LOADU(state.h()+6);
479
    row3l = LOADU(BLAKE2B_IV+0);
480
    row3h = LOADU(BLAKE2B_IV+2);
481
    row4l = _mm_xor_si128(LOADU(BLAKE2B_IV+4), LOADU(state.t()+0));
482
    row4h = _mm_xor_si128(LOADU(BLAKE2B_IV+6), LOADU(state.f()+0));
483

484
    BLAKE2B_ROUND(0);
485
    BLAKE2B_ROUND(1);
486
    BLAKE2B_ROUND(2);
487
    BLAKE2B_ROUND(3);
488
    BLAKE2B_ROUND(4);
489
    BLAKE2B_ROUND(5);
490
    BLAKE2B_ROUND(6);
491
    BLAKE2B_ROUND(7);
492
    BLAKE2B_ROUND(8);
493
    BLAKE2B_ROUND(9);
494
    BLAKE2B_ROUND(10);
495
    BLAKE2B_ROUND(11);
496

497
    row1l = _mm_xor_si128(row3l, row1l);
498
    row1h = _mm_xor_si128(row3h, row1h);
499
    STOREU(state.h()+0, _mm_xor_si128(LOADU(state.h()+0), row1l));
500
    STOREU(state.h()+2, _mm_xor_si128(LOADU(state.h()+2), row1h));
501
    row2l = _mm_xor_si128(row4l, row2l);
502
    row2h = _mm_xor_si128(row4h, row2h);
503
    STOREU(state.h()+4, _mm_xor_si128(LOADU(state.h()+4), row2l));
504
    STOREU(state.h()+6, _mm_xor_si128(LOADU(state.h()+6), row2h));
505
}
506
#endif  // CRYPTOPP_SSE41_AVAILABLE
507

508
#if CRYPTOPP_ARM_NEON_AVAILABLE
509
void BLAKE2_Compress64_NEON(const byte* input, BLAKE2b_State& state)
510
{
511
    #define BLAKE2B_LOAD_MSG_0_1(b0, b1) \
512
    do { b0 = vcombine_u64(vget_low_u64(m0), vget_low_u64(m1)); b1 = vcombine_u64(vget_low_u64(m2), vget_low_u64(m3)); } while(0)
513

514
    #define BLAKE2B_LOAD_MSG_0_2(b0, b1) \
515
    do { b0 = vcombine_u64(vget_high_u64(m0), vget_high_u64(m1)); b1 = vcombine_u64(vget_high_u64(m2), vget_high_u64(m3)); } while(0)
516

517
    #define BLAKE2B_LOAD_MSG_0_3(b0, b1) \
518
    do { b0 = vcombine_u64(vget_low_u64(m4), vget_low_u64(m5)); b1 = vcombine_u64(vget_low_u64(m6), vget_low_u64(m7)); } while(0)
519

520
    #define BLAKE2B_LOAD_MSG_0_4(b0, b1) \
521
    do { b0 = vcombine_u64(vget_high_u64(m4), vget_high_u64(m5)); b1 = vcombine_u64(vget_high_u64(m6), vget_high_u64(m7)); } while(0)
522

523
    #define BLAKE2B_LOAD_MSG_1_1(b0, b1) \
524
    do { b0 = vcombine_u64(vget_low_u64(m7), vget_low_u64(m2)); b1 = vcombine_u64(vget_high_u64(m4), vget_high_u64(m6)); } while(0)
525

526
    #define BLAKE2B_LOAD_MSG_1_2(b0, b1) \
527
    do { b0 = vcombine_u64(vget_low_u64(m5), vget_low_u64(m4)); b1 = vextq_u64(m7, m3, 1); } while(0)
528

529
    #define BLAKE2B_LOAD_MSG_1_3(b0, b1) \
530
    do { b0 = vextq_u64(m0, m0, 1); b1 = vcombine_u64(vget_high_u64(m5), vget_high_u64(m2)); } while(0)
531

532
    #define BLAKE2B_LOAD_MSG_1_4(b0, b1) \
533
    do { b0 = vcombine_u64(vget_low_u64(m6), vget_low_u64(m1)); b1 = vcombine_u64(vget_high_u64(m3), vget_high_u64(m1)); } while(0)
534

535
    #define BLAKE2B_LOAD_MSG_2_1(b0, b1) \
536
    do { b0 = vextq_u64(m5, m6, 1); b1 = vcombine_u64(vget_high_u64(m2), vget_high_u64(m7)); } while(0)
537

538
    #define BLAKE2B_LOAD_MSG_2_2(b0, b1) \
539
    do { b0 = vcombine_u64(vget_low_u64(m4), vget_low_u64(m0)); b1 = vcombine_u64(vget_low_u64(m1), vget_high_u64(m6)); } while(0)
540

541
    #define BLAKE2B_LOAD_MSG_2_3(b0, b1) \
542
    do { b0 = vcombine_u64(vget_low_u64(m5), vget_high_u64(m1)); b1 = vcombine_u64(vget_high_u64(m3), vget_high_u64(m4)); } while(0)
543

544
    #define BLAKE2B_LOAD_MSG_2_4(b0, b1) \
545
    do { b0 = vcombine_u64(vget_low_u64(m7), vget_low_u64(m3)); b1 = vextq_u64(m0, m2, 1); } while(0)
546

547
    #define BLAKE2B_LOAD_MSG_3_1(b0, b1) \
548
    do { b0 = vcombine_u64(vget_high_u64(m3), vget_high_u64(m1)); b1 = vcombine_u64(vget_high_u64(m6), vget_high_u64(m5)); } while(0)
549

550
    #define BLAKE2B_LOAD_MSG_3_2(b0, b1) \
551
    do { b0 = vcombine_u64(vget_high_u64(m4), vget_high_u64(m0)); b1 = vcombine_u64(vget_low_u64(m6), vget_low_u64(m7)); } while(0)
552

553
    #define BLAKE2B_LOAD_MSG_3_3(b0, b1) \
554
    do { b0 = vcombine_u64(vget_low_u64(m1), vget_high_u64(m2)); b1 = vcombine_u64(vget_low_u64(m2), vget_high_u64(m7)); } while(0)
555

556
    #define BLAKE2B_LOAD_MSG_3_4(b0, b1) \
557
    do { b0 = vcombine_u64(vget_low_u64(m3), vget_low_u64(m5)); b1 = vcombine_u64(vget_low_u64(m0), vget_low_u64(m4)); } while(0)
558

559
    #define BLAKE2B_LOAD_MSG_4_1(b0, b1) \
560
    do { b0 = vcombine_u64(vget_high_u64(m4), vget_high_u64(m2)); b1 = vcombine_u64(vget_low_u64(m1), vget_low_u64(m5)); } while(0)
561

562
    #define BLAKE2B_LOAD_MSG_4_2(b0, b1) \
563
    do { b0 = vcombine_u64(vget_low_u64(m0), vget_high_u64(m3)); b1 = vcombine_u64(vget_low_u64(m2), vget_high_u64(m7)); } while(0)
564

565
    #define BLAKE2B_LOAD_MSG_4_3(b0, b1) \
566
    do { b0 = vcombine_u64(vget_low_u64(m7), vget_high_u64(m5)); b1 = vcombine_u64(vget_low_u64(m3), vget_high_u64(m1)); } while(0)
567

568
    #define BLAKE2B_LOAD_MSG_4_4(b0, b1) \
569
    do { b0 = vextq_u64(m0, m6, 1); b1 = vcombine_u64(vget_low_u64(m4), vget_high_u64(m6)); } while(0)
570

571
    #define BLAKE2B_LOAD_MSG_5_1(b0, b1) \
572
    do { b0 = vcombine_u64(vget_low_u64(m1), vget_low_u64(m3)); b1 = vcombine_u64(vget_low_u64(m0), vget_low_u64(m4)); } while(0)
573

574
    #define BLAKE2B_LOAD_MSG_5_2(b0, b1) \
575
    do { b0 = vcombine_u64(vget_low_u64(m6), vget_low_u64(m5)); b1 = vcombine_u64(vget_high_u64(m5), vget_high_u64(m1)); } while(0)
576

577
    #define BLAKE2B_LOAD_MSG_5_3(b0, b1) \
578
    do { b0 = vcombine_u64(vget_low_u64(m2), vget_high_u64(m3)); b1 = vcombine_u64(vget_high_u64(m7), vget_high_u64(m0)); } while(0)
579

580
    #define BLAKE2B_LOAD_MSG_5_4(b0, b1) \
581
    do { b0 = vcombine_u64(vget_high_u64(m6), vget_high_u64(m2)); b1 = vcombine_u64(vget_low_u64(m7), vget_high_u64(m4)); } while(0)
582

583
    #define BLAKE2B_LOAD_MSG_6_1(b0, b1) \
584
    do { b0 = vcombine_u64(vget_low_u64(m6), vget_high_u64(m0)); b1 = vcombine_u64(vget_low_u64(m7), vget_low_u64(m2)); } while(0)
585

586
    #define BLAKE2B_LOAD_MSG_6_2(b0, b1) \
587
    do { b0 = vcombine_u64(vget_high_u64(m2), vget_high_u64(m7)); b1 = vextq_u64(m6, m5, 1); } while(0)
588

589
    #define BLAKE2B_LOAD_MSG_6_3(b0, b1) \
590
    do { b0 = vcombine_u64(vget_low_u64(m0), vget_low_u64(m3)); b1 = vextq_u64(m4, m4, 1); } while(0)
591

592
    #define BLAKE2B_LOAD_MSG_6_4(b0, b1) \
593
    do { b0 = vcombine_u64(vget_high_u64(m3), vget_high_u64(m1)); b1 = vcombine_u64(vget_low_u64(m1), vget_high_u64(m5)); } while(0)
594

595
    #define BLAKE2B_LOAD_MSG_7_1(b0, b1) \
596
    do { b0 = vcombine_u64(vget_high_u64(m6), vget_high_u64(m3)); b1 = vcombine_u64(vget_low_u64(m6), vget_high_u64(m1)); } while(0)
597

598
    #define BLAKE2B_LOAD_MSG_7_2(b0, b1) \
599
    do { b0 = vextq_u64(m5, m7, 1); b1 = vcombine_u64(vget_high_u64(m0), vget_high_u64(m4)); } while(0)
600

601
    #define BLAKE2B_LOAD_MSG_7_3(b0, b1) \
602
    do { b0 = vcombine_u64(vget_high_u64(m2), vget_high_u64(m7)); b1 = vcombine_u64(vget_low_u64(m4), vget_low_u64(m1)); } while(0)
603

604
    #define BLAKE2B_LOAD_MSG_7_4(b0, b1) \
605
    do { b0 = vcombine_u64(vget_low_u64(m0), vget_low_u64(m2)); b1 = vcombine_u64(vget_low_u64(m3), vget_low_u64(m5)); } while(0)
606

607
    #define BLAKE2B_LOAD_MSG_8_1(b0, b1) \
608
    do { b0 = vcombine_u64(vget_low_u64(m3), vget_low_u64(m7)); b1 = vextq_u64(m5, m0, 1); } while(0)
609

610
    #define BLAKE2B_LOAD_MSG_8_2(b0, b1) \
611
    do { b0 = vcombine_u64(vget_high_u64(m7), vget_high_u64(m4)); b1 = vextq_u64(m1, m4, 1); } while(0)
612

613
    #define BLAKE2B_LOAD_MSG_8_3(b0, b1) \
614
    do { b0 = m6; b1 = vextq_u64(m0, m5, 1); } while(0)
615

616
    #define BLAKE2B_LOAD_MSG_8_4(b0, b1) \
617
    do { b0 = vcombine_u64(vget_low_u64(m1), vget_high_u64(m3)); b1 = m2; } while(0)
618

619
    #define BLAKE2B_LOAD_MSG_9_1(b0, b1) \
620
    do { b0 = vcombine_u64(vget_low_u64(m5), vget_low_u64(m4)); b1 = vcombine_u64(vget_high_u64(m3), vget_high_u64(m0)); } while(0)
621

622
    #define BLAKE2B_LOAD_MSG_9_2(b0, b1) \
623
    do { b0 = vcombine_u64(vget_low_u64(m1), vget_low_u64(m2)); b1 = vcombine_u64(vget_low_u64(m3), vget_high_u64(m2)); } while(0)
624

625
    #define BLAKE2B_LOAD_MSG_9_3(b0, b1) \
626
    do { b0 = vcombine_u64(vget_high_u64(m7), vget_high_u64(m4)); b1 = vcombine_u64(vget_high_u64(m1), vget_high_u64(m6)); } while(0)
627

628
    #define BLAKE2B_LOAD_MSG_9_4(b0, b1) \
629
    do { b0 = vextq_u64(m5, m7, 1); b1 = vcombine_u64(vget_low_u64(m6), vget_low_u64(m0)); } while(0)
630

631
    #define BLAKE2B_LOAD_MSG_10_1(b0, b1) \
632
    do { b0 = vcombine_u64(vget_low_u64(m0), vget_low_u64(m1)); b1 = vcombine_u64(vget_low_u64(m2), vget_low_u64(m3)); } while(0)
633

634
    #define BLAKE2B_LOAD_MSG_10_2(b0, b1) \
635
    do { b0 = vcombine_u64(vget_high_u64(m0), vget_high_u64(m1)); b1 = vcombine_u64(vget_high_u64(m2), vget_high_u64(m3)); } while(0)
636

637
    #define BLAKE2B_LOAD_MSG_10_3(b0, b1) \
638
    do { b0 = vcombine_u64(vget_low_u64(m4), vget_low_u64(m5)); b1 = vcombine_u64(vget_low_u64(m6), vget_low_u64(m7)); } while(0)
639

640
    #define BLAKE2B_LOAD_MSG_10_4(b0, b1) \
641
    do { b0 = vcombine_u64(vget_high_u64(m4), vget_high_u64(m5)); b1 = vcombine_u64(vget_high_u64(m6), vget_high_u64(m7)); } while(0)
642

643
    #define BLAKE2B_LOAD_MSG_11_1(b0, b1) \
644
    do { b0 = vcombine_u64(vget_low_u64(m7), vget_low_u64(m2)); b1 = vcombine_u64(vget_high_u64(m4), vget_high_u64(m6)); } while(0)
645

646
    #define BLAKE2B_LOAD_MSG_11_2(b0, b1) \
647
    do { b0 = vcombine_u64(vget_low_u64(m5), vget_low_u64(m4)); b1 = vextq_u64(m7, m3, 1); } while(0)
648

649
    #define BLAKE2B_LOAD_MSG_11_3(b0, b1) \
650
    do { b0 = vextq_u64(m0, m0, 1); b1 = vcombine_u64(vget_high_u64(m5), vget_high_u64(m2)); } while(0)
651

652
    #define BLAKE2B_LOAD_MSG_11_4(b0, b1) \
653
    do { b0 = vcombine_u64(vget_low_u64(m6), vget_low_u64(m1)); b1 = vcombine_u64(vget_high_u64(m3), vget_high_u64(m1)); } while(0)
654

655
    #define vrorq_n_u64_32(x) vreinterpretq_u64_u32(vrev64q_u32(vreinterpretq_u32_u64((x))))
656

657
    #define vrorq_n_u64_24(x) vcombine_u64( \
658
        vreinterpret_u64_u8(vext_u8(vreinterpret_u8_u64(vget_low_u64(x)), vreinterpret_u8_u64(vget_low_u64(x)), 3)), \
659
        vreinterpret_u64_u8(vext_u8(vreinterpret_u8_u64(vget_high_u64(x)), vreinterpret_u8_u64(vget_high_u64(x)), 3)))
660

661
    #define vrorq_n_u64_16(x) vcombine_u64( \
662
        vreinterpret_u64_u8(vext_u8(vreinterpret_u8_u64(vget_low_u64(x)), vreinterpret_u8_u64(vget_low_u64(x)), 2)), \
663
        vreinterpret_u64_u8(vext_u8(vreinterpret_u8_u64(vget_high_u64(x)), vreinterpret_u8_u64(vget_high_u64(x)), 2)))
664

665
    #define vrorq_n_u64_63(x) veorq_u64(vaddq_u64(x, x), vshrq_n_u64(x, 63))
666

667
    #define BLAKE2B_G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \
668
    do { \
669
      row1l = vaddq_u64(vaddq_u64(row1l, b0), row2l); \
670
      row1h = vaddq_u64(vaddq_u64(row1h, b1), row2h); \
671
      row4l = veorq_u64(row4l, row1l); row4h = veorq_u64(row4h, row1h); \
672
      row4l = vrorq_n_u64_32(row4l); row4h = vrorq_n_u64_32(row4h); \
673
      row3l = vaddq_u64(row3l, row4l); row3h = vaddq_u64(row3h, row4h); \
674
      row2l = veorq_u64(row2l, row3l); row2h = veorq_u64(row2h, row3h); \
675
      row2l = vrorq_n_u64_24(row2l); row2h = vrorq_n_u64_24(row2h); \
676
    } while(0)
677

678
    #define BLAKE2B_G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \
679
    do { \
680
      row1l = vaddq_u64(vaddq_u64(row1l, b0), row2l); \
681
      row1h = vaddq_u64(vaddq_u64(row1h, b1), row2h); \
682
      row4l = veorq_u64(row4l, row1l); row4h = veorq_u64(row4h, row1h); \
683
      row4l = vrorq_n_u64_16(row4l); row4h = vrorq_n_u64_16(row4h); \
684
      row3l = vaddq_u64(row3l, row4l); row3h = vaddq_u64(row3h, row4h); \
685
      row2l = veorq_u64(row2l, row3l); row2h = veorq_u64(row2h, row3h); \
686
      row2l = vrorq_n_u64_63(row2l); row2h = vrorq_n_u64_63(row2h); \
687
    } while(0)
688

689
    #define BLAKE2B_DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \
690
    do { \
691
      uint64x2_t t0 = vextq_u64(row2l, row2h, 1); \
692
      uint64x2_t t1 = vextq_u64(row2h, row2l, 1); \
693
      row2l = t0; row2h = t1; t0 = row3l;  row3l = row3h; row3h = t0; \
694
      t0 = vextq_u64(row4h, row4l, 1); t1 = vextq_u64(row4l, row4h, 1); \
695
      row4l = t0; row4h = t1; \
696
    } while(0)
697

698
    #define BLAKE2B_UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \
699
    do { \
700
      uint64x2_t t0 = vextq_u64(row2h, row2l, 1); \
701
      uint64x2_t t1 = vextq_u64(row2l, row2h, 1); \
702
      row2l = t0; row2h = t1; t0 = row3l; row3l = row3h; row3h = t0; \
703
      t0 = vextq_u64(row4l, row4h, 1); t1 = vextq_u64(row4h, row4l, 1); \
704
      row4l = t0; row4h = t1; \
705
    } while(0)
706

707
    #define BLAKE2B_ROUND(r) \
708
    do { \
709
      uint64x2_t b0, b1; \
710
      BLAKE2B_LOAD_MSG_ ##r ##_1(b0, b1); \
711
      BLAKE2B_G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
712
      BLAKE2B_LOAD_MSG_ ##r ##_2(b0, b1); \
713
      BLAKE2B_G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
714
      BLAKE2B_DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h); \
715
      BLAKE2B_LOAD_MSG_ ##r ##_3(b0, b1); \
716
      BLAKE2B_G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
717
      BLAKE2B_LOAD_MSG_ ##r ##_4(b0, b1); \
718
      BLAKE2B_G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
719
      BLAKE2B_UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h); \
720
    } while(0)
721

722
    const uint64x2_t m0 = vreinterpretq_u64_u8(vld1q_u8(input +  00));
723
    const uint64x2_t m1 = vreinterpretq_u64_u8(vld1q_u8(input +  16));
724
    const uint64x2_t m2 = vreinterpretq_u64_u8(vld1q_u8(input +  32));
725
    const uint64x2_t m3 = vreinterpretq_u64_u8(vld1q_u8(input +  48));
726
    const uint64x2_t m4 = vreinterpretq_u64_u8(vld1q_u8(input +  64));
727
    const uint64x2_t m5 = vreinterpretq_u64_u8(vld1q_u8(input +  80));
728
    const uint64x2_t m6 = vreinterpretq_u64_u8(vld1q_u8(input +  96));
729
    const uint64x2_t m7 = vreinterpretq_u64_u8(vld1q_u8(input + 112));
730

731
    uint64x2_t row1l, row1h, row2l, row2h;
732
    uint64x2_t row3l, row3h, row4l, row4h;
733

734
    const uint64x2_t h0 = row1l = vld1q_u64(state.h()+0);
735
    const uint64x2_t h1 = row1h = vld1q_u64(state.h()+2);
736
    const uint64x2_t h2 = row2l = vld1q_u64(state.h()+4);
737
    const uint64x2_t h3 = row2h = vld1q_u64(state.h()+6);
738

739
    row3l = vld1q_u64(BLAKE2B_IV+0);
740
    row3h = vld1q_u64(BLAKE2B_IV+2);
741
    row4l = veorq_u64(vld1q_u64(BLAKE2B_IV+4), vld1q_u64(state.t()+0));
742
    row4h = veorq_u64(vld1q_u64(BLAKE2B_IV+6), vld1q_u64(state.f()+0));
743

744
    BLAKE2B_ROUND(0);
745
    BLAKE2B_ROUND(1);
746
    BLAKE2B_ROUND(2);
747
    BLAKE2B_ROUND(3);
748
    BLAKE2B_ROUND(4);
749
    BLAKE2B_ROUND(5);
750
    BLAKE2B_ROUND(6);
751
    BLAKE2B_ROUND(7);
752
    BLAKE2B_ROUND(8);
753
    BLAKE2B_ROUND(9);
754
    BLAKE2B_ROUND(10);
755
    BLAKE2B_ROUND(11);
756

757
    vst1q_u64(state.h()+0, veorq_u64(h0, veorq_u64(row1l, row3l)));
758
    vst1q_u64(state.h()+2, veorq_u64(h1, veorq_u64(row1h, row3h)));
759
    vst1q_u64(state.h()+4, veorq_u64(h2, veorq_u64(row2l, row4l)));
760
    vst1q_u64(state.h()+6, veorq_u64(h3, veorq_u64(row2h, row4h)));
761
}
762
#endif  // CRYPTOPP_ARM_NEON_AVAILABLE
763

764
#if (CRYPTOPP_POWER8_AVAILABLE)
765

766
inline uint64x2_p VecLoad64(const void* p)
767
{
768
    return (uint64x2_p)vec_xl(0, CONST_V32_CAST(p));
769
}
770

771
inline uint64x2_p VecLoad64LE(const void* p, const uint8x16_p le_mask)
772
{
773
#if defined(CRYPTOPP_BIG_ENDIAN)
774
    const uint32x4_p v = vec_xl(0, CONST_V32_CAST(p));
775
    return (uint64x2_p)VecPermute(v, v, le_mask);
776
#else
777
    CRYPTOPP_UNUSED(le_mask);
778
    return (uint64x2_p)vec_xl(0, CONST_V32_CAST(p));
779
#endif
780
}
781

782
inline void VecStore64(void* p, const uint64x2_p x)
783
{
784
    vec_xst((uint32x4_p)x, 0, NCONST_V32_CAST(p));
785
}
786

787
inline void VecStore64LE(void* p, const uint64x2_p x, const uint8x16_p le_mask)
788
{
789
#if defined(CRYPTOPP_BIG_ENDIAN)
790
    const uint64x2_p v = VecPermute(x, x, le_mask);
791
    vec_xst((uint32x4_p)v, 0, NCONST_V32_CAST(p));
792
#else
793
    CRYPTOPP_UNUSED(le_mask);
794
    vec_xst((uint32x4_p)x, 0, NCONST_V32_CAST(p));
795
#endif
796
}
797

798
#if defined(CRYPTOPP_BIG_ENDIAN)
799
#define vec_shl_8(a,b) (uint64x2_p)vec_sld((uint8x16_p)a,(uint8x16_p)b,8)
800
#else
801
#define vec_shl_8(a,b) (uint64x2_p)vec_sld((uint8x16_p)b,(uint8x16_p)a,8)
802
#endif
803

804
#define vec_merge_hi(a, b) vec_mergeh(a,b)
805
#define vec_merge_hi_lo(a, b) vec_mergeh(a,(uint64x2_p)vec_sld((uint8x16_p)b,(uint8x16_p)b,8))
806
#define vec_merge_lo(a, b) vec_mergel(a,b)
807

808
void BLAKE2_Compress64_POWER8(const byte* input, BLAKE2b_State& state)
809
{
810
    #define BLAKE2B_LOAD_MSG_0_1(b0, b1) \
811
    do { \
812
         b0 = vec_merge_hi(m0, m1); \
813
         b1 = vec_merge_hi(m2, m3); \
814
    } while(0)
815

816
    #define BLAKE2B_LOAD_MSG_0_2(b0, b1) \
817
    do { \
818
         b0 = vec_merge_lo(m0, m1); \
819
         b1 = vec_merge_lo(m2, m3); \
820
    } while(0)
821

822
    #define BLAKE2B_LOAD_MSG_0_3(b0, b1) \
823
    do { \
824
         b0 = vec_merge_hi(m4, m5); \
825
         b1 = vec_merge_hi(m6, m7); \
826
    } while(0)
827

828
    #define BLAKE2B_LOAD_MSG_0_4(b0, b1) \
829
    do { \
830
         b0 = vec_merge_lo(m4, m5); \
831
         b1 = vec_merge_lo(m6, m7); \
832
    } while(0)
833

834
    #define BLAKE2B_LOAD_MSG_1_1(b0, b1) \
835
    do { \
836
         b0 = vec_merge_hi(m7, m2); \
837
         b1 = vec_merge_lo(m4, m6); \
838
    } while(0)
839

840
    #define BLAKE2B_LOAD_MSG_1_2(b0, b1) \
841
    do { \
842
         b0 = vec_merge_hi(m5, m4); \
843
         b1 = vec_shl_8(m7, m3); \
844
    } while(0)
845

846
    #define BLAKE2B_LOAD_MSG_1_3(b0, b1) \
847
    do { \
848
         b0 = vec_shl_8(m0, m0); \
849
         b1 = vec_merge_lo(m5, m2); \
850
    } while(0)
851

852
    #define BLAKE2B_LOAD_MSG_1_4(b0, b1) \
853
    do { \
854
         b0 = vec_merge_hi(m6, m1); \
855
         b1 = vec_merge_lo(m3, m1); \
856
    } while(0)
857

858
    #define BLAKE2B_LOAD_MSG_2_1(b0, b1) \
859
    do { \
860
         b0 = vec_shl_8(m5, m6); \
861
         b1 = vec_merge_lo(m2, m7); \
862
    } while(0)
863

864
    #define BLAKE2B_LOAD_MSG_2_2(b0, b1) \
865
    do { \
866
         b0 = vec_merge_hi(m4, m0); \
867
         b1 = vec_merge_hi_lo(m1, m6); \
868
    } while(0)
869

870
    #define BLAKE2B_LOAD_MSG_2_3(b0, b1) \
871
       do { \
872
         b0 = vec_merge_hi_lo(m5, m1); \
873
         b1 = vec_merge_lo(m3, m4); \
874
    } while(0)
875

876
    #define BLAKE2B_LOAD_MSG_2_4(b0, b1) \
877
       do { \
878
         b0 = vec_merge_hi(m7, m3); \
879
         b1 = vec_shl_8(m0, m2); \
880
    } while(0)
881

882
    #define BLAKE2B_LOAD_MSG_3_1(b0, b1) \
883
       do { \
884
         b0 = vec_merge_lo(m3, m1); \
885
         b1 = vec_merge_lo(m6, m5); \
886
    } while(0)
887

888
    #define BLAKE2B_LOAD_MSG_3_2(b0, b1) \
889
       do { \
890
         b0 = vec_merge_lo(m4, m0); \
891
         b1 = vec_merge_hi(m6, m7); \
892
    } while(0)
893

894
    #define BLAKE2B_LOAD_MSG_3_3(b0, b1) \
895
       do { \
896
         b0 = vec_merge_hi_lo(m1, m2); \
897
         b1 = vec_merge_hi_lo(m2, m7); \
898
    } while(0)
899

900
    #define BLAKE2B_LOAD_MSG_3_4(b0, b1) \
901
       do { \
902
         b0 = vec_merge_hi(m3, m5); \
903
         b1 = vec_merge_hi(m0, m4); \
904
    } while(0)
905

906
    #define BLAKE2B_LOAD_MSG_4_1(b0, b1) \
907
       do { \
908
         b0 = vec_merge_lo(m4, m2); \
909
         b1 = vec_merge_hi(m1, m5); \
910
    } while(0)
911

912
    #define BLAKE2B_LOAD_MSG_4_2(b0, b1) \
913
       do { \
914
         b0 = vec_merge_hi_lo(m0, m3); \
915
         b1 = vec_merge_hi_lo(m2, m7); \
916
    } while(0)
917

918
    #define BLAKE2B_LOAD_MSG_4_3(b0, b1) \
919
       do { \
920
         b0 = vec_merge_hi_lo(m7, m5); \
921
         b1 = vec_merge_hi_lo(m3, m1); \
922
    } while(0)
923

924
    #define BLAKE2B_LOAD_MSG_4_4(b0, b1) \
925
       do { \
926
         b0 = vec_shl_8(m0, m6); \
927
         b1 = vec_merge_hi_lo(m4, m6); \
928
    } while(0)
929

930
    #define BLAKE2B_LOAD_MSG_5_1(b0, b1) \
931
       do { \
932
         b0 = vec_merge_hi(m1, m3); \
933
         b1 = vec_merge_hi(m0, m4); \
934
    } while(0)
935

936
    #define BLAKE2B_LOAD_MSG_5_2(b0, b1) \
937
       do { \
938
         b0 = vec_merge_hi(m6, m5); \
939
         b1 = vec_merge_lo(m5, m1); \
940
    } while(0)
941

942
    #define BLAKE2B_LOAD_MSG_5_3(b0, b1) \
943
       do { \
944
         b0 = vec_merge_hi_lo(m2, m3); \
945
         b1 = vec_merge_lo(m7, m0); \
946
    } while(0)
947

948
    #define BLAKE2B_LOAD_MSG_5_4(b0, b1) \
949
       do { \
950
         b0 = vec_merge_lo(m6, m2); \
951
         b1 = vec_merge_hi_lo(m7, m4); \
952
    } while(0)
953

954
    #define BLAKE2B_LOAD_MSG_6_1(b0, b1) \
955
       do { \
956
         b0 = vec_merge_hi_lo(m6, m0); \
957
         b1 = vec_merge_hi(m7, m2); \
958
    } while(0)
959

960
    #define BLAKE2B_LOAD_MSG_6_2(b0, b1) \
961
       do { \
962
         b0 = vec_merge_lo(m2, m7); \
963
         b1 = vec_shl_8(m6, m5); \
964
    } while(0)
965

966
    #define BLAKE2B_LOAD_MSG_6_3(b0, b1) \
967
       do { \
968
         b0 = vec_merge_hi(m0, m3); \
969
         b1 = vec_shl_8(m4, m4); \
970
    } while(0)
971

972
    #define BLAKE2B_LOAD_MSG_6_4(b0, b1) \
973
       do { \
974
         b0 = vec_merge_lo(m3, m1); \
975
         b1 = vec_merge_hi_lo(m1, m5); \
976
    } while(0)
977

978
    #define BLAKE2B_LOAD_MSG_7_1(b0, b1) \
979
       do { \
980
         b0 = vec_merge_lo(m6, m3); \
981
         b1 = vec_merge_hi_lo(m6, m1); \
982
    } while(0)
983

984
    #define BLAKE2B_LOAD_MSG_7_2(b0, b1) \
985
       do { \
986
         b0 = vec_shl_8(m5, m7); \
987
         b1 = vec_merge_lo(m0, m4); \
988
    } while(0)
989

990
    #define BLAKE2B_LOAD_MSG_7_3(b0, b1) \
991
       do { \
992
         b0 = vec_merge_lo(m2, m7); \
993
         b1 = vec_merge_hi(m4, m1); \
994
    } while(0)
995

996
    #define BLAKE2B_LOAD_MSG_7_4(b0, b1) \
997
       do { \
998
         b0 = vec_merge_hi(m0, m2); \
999
         b1 = vec_merge_hi(m3, m5); \
1000
    } while(0)
1001

1002
    #define BLAKE2B_LOAD_MSG_8_1(b0, b1) \
1003
       do { \
1004
         b0 = vec_merge_hi(m3, m7); \
1005
         b1 = vec_shl_8(m5, m0); \
1006
    } while(0)
1007

1008
    #define BLAKE2B_LOAD_MSG_8_2(b0, b1) \
1009
       do { \
1010
         b0 = vec_merge_lo(m7, m4); \
1011
         b1 = vec_shl_8(m1, m4); \
1012
    } while(0)
1013

1014
    #define BLAKE2B_LOAD_MSG_8_3(b0, b1) \
1015
       do { \
1016
         b0 = m6; \
1017
         b1 = vec_shl_8(m0, m5); \
1018
    } while(0)
1019

1020
    #define BLAKE2B_LOAD_MSG_8_4(b0, b1) \
1021
       do { \
1022
         b0 = vec_merge_hi_lo(m1, m3); \
1023
         b1 = m2; \
1024
    } while(0)
1025

1026
    #define BLAKE2B_LOAD_MSG_9_1(b0, b1) \
1027
       do { \
1028
         b0 = vec_merge_hi(m5, m4); \
1029
         b1 = vec_merge_lo(m3, m0); \
1030
    } while(0)
1031

1032
    #define BLAKE2B_LOAD_MSG_9_2(b0, b1) \
1033
       do { \
1034
         b0 = vec_merge_hi(m1, m2); \
1035
         b1 = vec_merge_hi_lo(m3, m2); \
1036
    } while(0)
1037

1038
    #define BLAKE2B_LOAD_MSG_9_3(b0, b1) \
1039
       do { \
1040
         b0 = vec_merge_lo(m7, m4); \
1041
         b1 = vec_merge_lo(m1, m6); \
1042
    } while(0)
1043

1044
    #define BLAKE2B_LOAD_MSG_9_4(b0, b1) \
1045
       do { \
1046
         b0 = vec_shl_8(m5, m7); \
1047
         b1 = vec_merge_hi(m6, m0); \
1048
    } while(0)
1049

1050
    #define BLAKE2B_LOAD_MSG_10_1(b0, b1) \
1051
       do { \
1052
         b0 = vec_merge_hi(m0, m1); \
1053
         b1 = vec_merge_hi(m2, m3); \
1054
    } while(0)
1055

1056
    #define BLAKE2B_LOAD_MSG_10_2(b0, b1) \
1057
       do { \
1058
         b0 = vec_merge_lo(m0, m1); \
1059
         b1 = vec_merge_lo(m2, m3); \
1060
    } while(0)
1061

1062
    #define BLAKE2B_LOAD_MSG_10_3(b0, b1) \
1063
       do { \
1064
         b0 = vec_merge_hi(m4, m5); \
1065
         b1 = vec_merge_hi(m6, m7); \
1066
    } while(0)
1067

1068
    #define BLAKE2B_LOAD_MSG_10_4(b0, b1) \
1069
       do { \
1070
         b0 = vec_merge_lo(m4, m5); \
1071
         b1 = vec_merge_lo(m6, m7); \
1072
    } while(0)
1073

1074
    #define BLAKE2B_LOAD_MSG_11_1(b0, b1) \
1075
       do { \
1076
         b0 = vec_merge_hi(m7, m2); \
1077
         b1 = vec_merge_lo(m4, m6); \
1078
    } while(0)
1079

1080
    #define BLAKE2B_LOAD_MSG_11_2(b0, b1) \
1081
       do { \
1082
         b0 = vec_merge_hi(m5, m4); \
1083
         b1 = vec_shl_8(m7, m3); \
1084
    } while(0)
1085

1086
    #define BLAKE2B_LOAD_MSG_11_3(b0, b1) \
1087
       do { \
1088
         b0 = vec_shl_8(m0, m0); \
1089
         b1 = vec_merge_lo(m5, m2); \
1090
    } while(0)
1091

1092
    #define BLAKE2B_LOAD_MSG_11_4(b0, b1) \
1093
       do { \
1094
         b0 = vec_merge_hi(m6, m1); \
1095
         b1 = vec_merge_lo(m3, m1); \
1096
    } while(0)
1097

1098
    // Power8 has packed 64-bit rotate, but in terms of left rotate
1099
    const uint64x2_p ROR16_MASK = { 64-16, 64-16 };
1100
    const uint64x2_p ROR24_MASK = { 64-24, 64-24 };
1101
    const uint64x2_p ROR32_MASK = { 64-32, 64-32 };
1102
    const uint64x2_p ROR63_MASK = { 64-63, 64-63 };
1103

1104
    #define vec_ror_32(x) vec_rl(x, ROR32_MASK)
1105
    #define vec_ror_24(x) vec_rl(x, ROR24_MASK)
1106
    #define vec_ror_16(x) vec_rl(x, ROR16_MASK)
1107
    #define vec_ror_63(x) vec_rl(x, ROR63_MASK)
1108

1109
    #define BLAKE2B_G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \
1110
    do { \
1111
      row1l = VecAdd(VecAdd(row1l, b0), row2l); \
1112
      row1h = VecAdd(VecAdd(row1h, b1), row2h); \
1113
      row4l = VecXor(row4l, row1l); row4h = VecXor(row4h, row1h); \
1114
      row4l = vec_ror_32(row4l); row4h = vec_ror_32(row4h); \
1115
      row3l = VecAdd(row3l, row4l); row3h = VecAdd(row3h, row4h); \
1116
      row2l = VecXor(row2l, row3l); row2h = VecXor(row2h, row3h); \
1117
      row2l = vec_ror_24(row2l); row2h = vec_ror_24(row2h); \
1118
    } while(0)
1119

1120
    #define BLAKE2B_G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \
1121
    do { \
1122
      row1l = VecAdd(VecAdd(row1l, b0), row2l); \
1123
      row1h = VecAdd(VecAdd(row1h, b1), row2h); \
1124
      row4l = VecXor(row4l, row1l); row4h = VecXor(row4h, row1h); \
1125
      row4l = vec_ror_16(row4l); row4h = vec_ror_16(row4h); \
1126
      row3l = VecAdd(row3l, row4l); row3h = VecAdd(row3h, row4h); \
1127
      row2l = VecXor(row2l, row3l); row2h = VecXor(row2h, row3h); \
1128
      row2l = vec_ror_63(row2l); row2h = vec_ror_63(row2h); \
1129
    } while(0)
1130

1131
    #define BLAKE2B_DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \
1132
    do { \
1133
      uint64x2_p t0 = vec_shl_8(row2l, row2h); \
1134
      uint64x2_p t1 = vec_shl_8(row2h, row2l); \
1135
      row2l = t0; row2h = t1; t0 = row3l;  row3l = row3h; row3h = t0; \
1136
      t0 = vec_shl_8(row4h, row4l); t1 = vec_shl_8(row4l, row4h); \
1137
      row4l = t0; row4h = t1; \
1138
    } while(0)
1139

1140
    #define BLAKE2B_UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \
1141
    do { \
1142
      uint64x2_p t0 = vec_shl_8(row2h, row2l); \
1143
      uint64x2_p t1 = vec_shl_8(row2l, row2h); \
1144
      row2l = t0; row2h = t1; t0 = row3l; row3l = row3h; row3h = t0; \
1145
      t0 = vec_shl_8(row4l, row4h); t1 = vec_shl_8(row4h, row4l); \
1146
      row4l = t0; row4h = t1; \
1147
    } while(0)
1148

1149
    #define BLAKE2B_ROUND(r) \
1150
    do { \
1151
      uint64x2_p b0, b1; \
1152
      BLAKE2B_LOAD_MSG_ ##r ##_1(b0, b1); \
1153
      BLAKE2B_G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
1154
      BLAKE2B_LOAD_MSG_ ##r ##_2(b0, b1); \
1155
      BLAKE2B_G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
1156
      BLAKE2B_DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h); \
1157
      BLAKE2B_LOAD_MSG_ ##r ##_3(b0, b1); \
1158
      BLAKE2B_G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
1159
      BLAKE2B_LOAD_MSG_ ##r ##_4(b0, b1); \
1160
      BLAKE2B_G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
1161
      BLAKE2B_UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h); \
1162
    } while(0)
1163

1164
    // Possibly unaligned user messages
1165
    uint64x2_p m0, m1, m2, m3, m4, m5, m6, m7;
1166
    // Endian conversion mask
1167
    const uint8x16_p le_mask = {7,6,5,4, 3,2,1,0, 15,14,13,12, 11,10,9,8};
1168

1169
#if defined(_ARCH_PWR9)
1170
    // POWER9 provides loads for char's and short's
1171
    m0 = (uint64x2_p) vec_xl(  0, CONST_V8_CAST( input ));
1172
    m1 = (uint64x2_p) vec_xl( 16, CONST_V8_CAST( input ));
1173
    m2 = (uint64x2_p) vec_xl( 32, CONST_V8_CAST( input ));
1174
    m3 = (uint64x2_p) vec_xl( 48, CONST_V8_CAST( input ));
1175
    m4 = (uint64x2_p) vec_xl( 64, CONST_V8_CAST( input ));
1176
    m5 = (uint64x2_p) vec_xl( 80, CONST_V8_CAST( input ));
1177
    m6 = (uint64x2_p) vec_xl( 96, CONST_V8_CAST( input ));
1178
    m7 = (uint64x2_p) vec_xl(112, CONST_V8_CAST( input ));
1179

1180
# if defined(CRYPTOPP_BIG_ENDIAN)
1181
    m0 = vec_perm(m0, m0, le_mask);
1182
    m1 = vec_perm(m1, m1, le_mask);
1183
    m2 = vec_perm(m2, m2, le_mask);
1184
    m3 = vec_perm(m3, m3, le_mask);
1185
    m4 = vec_perm(m4, m4, le_mask);
1186
    m5 = vec_perm(m5, m5, le_mask);
1187
    m6 = vec_perm(m6, m6, le_mask);
1188
    m7 = vec_perm(m7, m7, le_mask);
1189
# endif
1190
#else
1191
    // Altivec only provides 16-byte aligned loads
1192
    // http://www.nxp.com/docs/en/reference-manual/ALTIVECPEM.pdf
1193
    m0 = (uint64x2_p) vec_ld(  0, CONST_V8_CAST( input ));
1194
    m1 = (uint64x2_p) vec_ld( 16, CONST_V8_CAST( input ));
1195
    m2 = (uint64x2_p) vec_ld( 32, CONST_V8_CAST( input ));
1196
    m3 = (uint64x2_p) vec_ld( 48, CONST_V8_CAST( input ));
1197
    m4 = (uint64x2_p) vec_ld( 64, CONST_V8_CAST( input ));
1198
    m5 = (uint64x2_p) vec_ld( 80, CONST_V8_CAST( input ));
1199
    m6 = (uint64x2_p) vec_ld( 96, CONST_V8_CAST( input ));
1200
    m7 = (uint64x2_p) vec_ld(112, CONST_V8_CAST( input ));
1201

1202
    // Alignment check for load of the message buffer
1203
    const uintptr_t addr = (uintptr_t)input;
1204
    if (addr%16 == 0)
1205
    {
1206
        // Already aligned. Perform a little-endian swap as required
1207
# if defined(CRYPTOPP_BIG_ENDIAN)
1208
        m0 = vec_perm(m0, m0, le_mask);
1209
        m1 = vec_perm(m1, m1, le_mask);
1210
        m2 = vec_perm(m2, m2, le_mask);
1211
        m3 = vec_perm(m3, m3, le_mask);
1212
        m4 = vec_perm(m4, m4, le_mask);
1213
        m5 = vec_perm(m5, m5, le_mask);
1214
        m6 = vec_perm(m6, m6, le_mask);
1215
        m7 = vec_perm(m7, m7, le_mask);
1216
# endif
1217
    }
1218
    else
1219
    {
1220
        // Not aligned. Fix vectors and perform a little-endian swap as required
1221
        // http://mirror.informatimago.com/next/developer.apple.com/
1222
        //        hardwaredrivers/ve/code_optimization.html
1223
        uint64x2_p ex; uint8x16_p perm;
1224
        ex = (uint64x2_p) vec_ld(112+15, CONST_V8_CAST( input ));
1225
        perm = vec_lvsl(0, CONST_V8_CAST( addr ));
1226

1227
# if defined(CRYPTOPP_BIG_ENDIAN)
1228
        // Combine the vector permute with the little-endian swap
1229
        perm = vec_perm(perm, perm, le_mask);
1230
# endif
1231

1232
        m0 = vec_perm(m0, m1, perm);
1233
        m1 = vec_perm(m1, m2, perm);
1234
        m2 = vec_perm(m2, m3, perm);
1235
        m3 = vec_perm(m3, m4, perm);
1236
        m4 = vec_perm(m4, m5, perm);
1237
        m5 = vec_perm(m5, m6, perm);
1238
        m6 = vec_perm(m6, m7, perm);
1239
        m7 = vec_perm(m7, ex, perm);
1240
    }
1241
#endif
1242

1243
    uint64x2_p row1l, row1h, row2l, row2h;
1244
    uint64x2_p row3l, row3h, row4l, row4h;
1245

1246
    const uint64x2_p h0 = row1l = VecLoad64LE(state.h()+0, le_mask);
1247
    const uint64x2_p h1 = row1h = VecLoad64LE(state.h()+2, le_mask);
1248
    const uint64x2_p h2 = row2l = VecLoad64LE(state.h()+4, le_mask);
1249
    const uint64x2_p h3 = row2h = VecLoad64LE(state.h()+6, le_mask);
1250

1251
    row3l = VecLoad64(BLAKE2B_IV+0);
1252
    row3h = VecLoad64(BLAKE2B_IV+2);
1253
    row4l = VecXor(VecLoad64(BLAKE2B_IV+4), VecLoad64(state.t()+0));
1254
    row4h = VecXor(VecLoad64(BLAKE2B_IV+6), VecLoad64(state.f()+0));
1255

1256
    BLAKE2B_ROUND(0);
1257
    BLAKE2B_ROUND(1);
1258
    BLAKE2B_ROUND(2);
1259
    BLAKE2B_ROUND(3);
1260
    BLAKE2B_ROUND(4);
1261
    BLAKE2B_ROUND(5);
1262
    BLAKE2B_ROUND(6);
1263
    BLAKE2B_ROUND(7);
1264
    BLAKE2B_ROUND(8);
1265
    BLAKE2B_ROUND(9);
1266
    BLAKE2B_ROUND(10);
1267
    BLAKE2B_ROUND(11);
1268

1269
    VecStore64LE(state.h()+0, VecXor(h0, VecXor(row1l, row3l)), le_mask);
1270
    VecStore64LE(state.h()+2, VecXor(h1, VecXor(row1h, row3h)), le_mask);
1271
    VecStore64LE(state.h()+4, VecXor(h2, VecXor(row2l, row4l)), le_mask);
1272
    VecStore64LE(state.h()+6, VecXor(h3, VecXor(row2h, row4h)), le_mask);
1273
}
1274
#endif  // CRYPTOPP_POWER8_AVAILABLE
1275

1276
NAMESPACE_END
1277

1278
Product

Resources

Company
