CoCalc -- sph_groestl.c

GitHub Repository: tpruvot/cpuminer-multi
Path: blob/linux/sha3/sph_groestl.c
¹²⁰¹ views
1
/* $Id: groestl.c 260 2011-07-21 01:02:38Z tp $ */
2
/*
3
 * Groestl implementation.
4
 *
5
 * ==========================(LICENSE BEGIN)============================
6
 *
7
 * Copyright (c) 2007-2010  Projet RNRT SAPHIR
8
 * 
9
 * Permission is hereby granted, free of charge, to any person obtaining
10
 * a copy of this software and associated documentation files (the
11
 * "Software"), to deal in the Software without restriction, including
12
 * without limitation the rights to use, copy, modify, merge, publish,
13
 * distribute, sublicense, and/or sell copies of the Software, and to
14
 * permit persons to whom the Software is furnished to do so, subject to
15
 * the following conditions:
16
 * 
17
 * The above copyright notice and this permission notice shall be
18
 * included in all copies or substantial portions of the Software.
19
 * 
20
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
21
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
22
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
23
 * IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
24
 * CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
25
 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
26
 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27
 *
28
 * ===========================(LICENSE END)=============================
29
 *
30
 * @author   Thomas Pornin <[email protected]>
31
 */
32

33
#include <stddef.h>
34
#include <string.h>
35

36
#include "sph_groestl.h"
37

38
#ifdef __cplusplus
39
extern "C"{
40
#endif
41

42
#if SPH_SMALL_FOOTPRINT && !defined SPH_SMALL_FOOTPRINT_GROESTL
43
#define SPH_SMALL_FOOTPRINT_GROESTL   1
44
#endif
45

46
/*
47
 * Apparently, the 32-bit-only version is not faster than the 64-bit
48
 * version unless using the "small footprint" code on a 32-bit machine.
49
 */
50
#if !defined SPH_GROESTL_64
51
#if SPH_SMALL_FOOTPRINT_GROESTL && !SPH_64_TRUE
52
#define SPH_GROESTL_64   0
53
#else
54
#define SPH_GROESTL_64   1
55
#endif
56
#endif
57

58
#if !SPH_64
59
#undef SPH_GROESTL_64
60
#endif
61

62
#ifdef _MSC_VER
63
#pragma warning (disable: 4146)
64
#endif
65

66
/*
67
 * The internal representation may use either big-endian or
68
 * little-endian. Using the platform default representation speeds up
69
 * encoding and decoding between bytes and the matrix columns.
70
 */
71

72
#undef USE_LE
73
#if SPH_GROESTL_LITTLE_ENDIAN
74
#define USE_LE   1
75
#elif SPH_GROESTL_BIG_ENDIAN
76
#define USE_LE   0
77
#elif SPH_LITTLE_ENDIAN
78
#define USE_LE   1
79
#endif
80

81
#if USE_LE
82

83
#define C32e(x)     ((SPH_C32(x) >> 24) \
84
                    | ((SPH_C32(x) >>  8) & SPH_C32(0x0000FF00)) \
85
                    | ((SPH_C32(x) <<  8) & SPH_C32(0x00FF0000)) \
86
                    | ((SPH_C32(x) << 24) & SPH_C32(0xFF000000)))
87
#define dec32e_aligned   sph_dec32le_aligned
88
#define enc32e           sph_enc32le
89
#define B32_0(x)    ((x) & 0xFF)
90
#define B32_1(x)    (((x) >> 8) & 0xFF)
91
#define B32_2(x)    (((x) >> 16) & 0xFF)
92
#define B32_3(x)    ((x) >> 24)
93

94
#define R32u(u, d)   SPH_T32(((u) << 16) | ((d) >> 16))
95
#define R32d(u, d)   SPH_T32(((u) >> 16) | ((d) << 16))
96

97
#define PC32up(j, r)   ((sph_u32)((j) + (r)))
98
#define PC32dn(j, r)   0
99
#define QC32up(j, r)   SPH_C32(0xFFFFFFFF)
100
#define QC32dn(j, r)   (((sph_u32)(r) << 24) ^ SPH_T32(~((sph_u32)(j) << 24)))
101

102
#if SPH_64
103
#define C64e(x)     ((SPH_C64(x) >> 56) \
104
                    | ((SPH_C64(x) >> 40) & SPH_C64(0x000000000000FF00)) \
105
                    | ((SPH_C64(x) >> 24) & SPH_C64(0x0000000000FF0000)) \
106
                    | ((SPH_C64(x) >>  8) & SPH_C64(0x00000000FF000000)) \
107
                    | ((SPH_C64(x) <<  8) & SPH_C64(0x000000FF00000000)) \
108
                    | ((SPH_C64(x) << 24) & SPH_C64(0x0000FF0000000000)) \
109
                    | ((SPH_C64(x) << 40) & SPH_C64(0x00FF000000000000)) \
110
                    | ((SPH_C64(x) << 56) & SPH_C64(0xFF00000000000000)))
111
#define dec64e_aligned   sph_dec64le_aligned
112
#define enc64e           sph_enc64le
113
#define B64_0(x)    ((x) & 0xFF)
114
#define B64_1(x)    (((x) >> 8) & 0xFF)
115
#define B64_2(x)    (((x) >> 16) & 0xFF)
116
#define B64_3(x)    (((x) >> 24) & 0xFF)
117
#define B64_4(x)    (((x) >> 32) & 0xFF)
118
#define B64_5(x)    (((x) >> 40) & 0xFF)
119
#define B64_6(x)    (((x) >> 48) & 0xFF)
120
#define B64_7(x)    ((x) >> 56)
121
#define R64         SPH_ROTL64
122
#define PC64(j, r)  ((sph_u64)((j) + (r)))
123
#define QC64(j, r)  (((sph_u64)(r) << 56) ^ SPH_T64(~((sph_u64)(j) << 56)))
124
#endif
125

126
#else
127

128
#define C32e(x)     SPH_C32(x)
129
#define dec32e_aligned   sph_dec32be_aligned
130
#define enc32e           sph_enc32be
131
#define B32_0(x)    ((x) >> 24)
132
#define B32_1(x)    (((x) >> 16) & 0xFF)
133
#define B32_2(x)    (((x) >> 8) & 0xFF)
134
#define B32_3(x)    ((x) & 0xFF)
135

136
#define R32u(u, d)   SPH_T32(((u) >> 16) | ((d) << 16))
137
#define R32d(u, d)   SPH_T32(((u) << 16) | ((d) >> 16))
138

139
#define PC32up(j, r)   ((sph_u32)((j) + (r)) << 24)
140
#define PC32dn(j, r)   0
141
#define QC32up(j, r)   SPH_C32(0xFFFFFFFF)
142
#define QC32dn(j, r)   ((sph_u32)(r) ^ SPH_T32(~(sph_u32)(j)))
143

144
#if SPH_64
145
#define C64e(x)     SPH_C64(x)
146
#define dec64e_aligned   sph_dec64be_aligned
147
#define enc64e           sph_enc64be
148
#define B64_0(x)    ((x) >> 56)
149
#define B64_1(x)    (((x) >> 48) & 0xFF)
150
#define B64_2(x)    (((x) >> 40) & 0xFF)
151
#define B64_3(x)    (((x) >> 32) & 0xFF)
152
#define B64_4(x)    (((x) >> 24) & 0xFF)
153
#define B64_5(x)    (((x) >> 16) & 0xFF)
154
#define B64_6(x)    (((x) >> 8) & 0xFF)
155
#define B64_7(x)    ((x) & 0xFF)
156
#define R64         SPH_ROTR64
157
#define PC64(j, r)  ((sph_u64)((j) + (r)) << 56)
158
#define QC64(j, r)  ((sph_u64)(r) ^ SPH_T64(~(sph_u64)(j)))
159
#endif
160

161
#endif
162

163
#if SPH_GROESTL_64
164

165
static const sph_u64 T0[] = {
166
	C64e(0xc632f4a5f497a5c6), C64e(0xf86f978497eb84f8),
167
	C64e(0xee5eb099b0c799ee), C64e(0xf67a8c8d8cf78df6),
168
	C64e(0xffe8170d17e50dff), C64e(0xd60adcbddcb7bdd6),
169
	C64e(0xde16c8b1c8a7b1de), C64e(0x916dfc54fc395491),
170
	C64e(0x6090f050f0c05060), C64e(0x0207050305040302),
171
	C64e(0xce2ee0a9e087a9ce), C64e(0x56d1877d87ac7d56),
172
	C64e(0xe7cc2b192bd519e7), C64e(0xb513a662a67162b5),
173
	C64e(0x4d7c31e6319ae64d), C64e(0xec59b59ab5c39aec),
174
	C64e(0x8f40cf45cf05458f), C64e(0x1fa3bc9dbc3e9d1f),
175
	C64e(0x8949c040c0094089), C64e(0xfa68928792ef87fa),
176
	C64e(0xefd03f153fc515ef), C64e(0xb29426eb267febb2),
177
	C64e(0x8ece40c94007c98e), C64e(0xfbe61d0b1ded0bfb),
178
	C64e(0x416e2fec2f82ec41), C64e(0xb31aa967a97d67b3),
179
	C64e(0x5f431cfd1cbefd5f), C64e(0x456025ea258aea45),
180
	C64e(0x23f9dabfda46bf23), C64e(0x535102f702a6f753),
181
	C64e(0xe445a196a1d396e4), C64e(0x9b76ed5bed2d5b9b),
182
	C64e(0x75285dc25deac275), C64e(0xe1c5241c24d91ce1),
183
	C64e(0x3dd4e9aee97aae3d), C64e(0x4cf2be6abe986a4c),
184
	C64e(0x6c82ee5aeed85a6c), C64e(0x7ebdc341c3fc417e),
185
	C64e(0xf5f3060206f102f5), C64e(0x8352d14fd11d4f83),
186
	C64e(0x688ce45ce4d05c68), C64e(0x515607f407a2f451),
187
	C64e(0xd18d5c345cb934d1), C64e(0xf9e1180818e908f9),
188
	C64e(0xe24cae93aedf93e2), C64e(0xab3e9573954d73ab),
189
	C64e(0x6297f553f5c45362), C64e(0x2a6b413f41543f2a),
190
	C64e(0x081c140c14100c08), C64e(0x9563f652f6315295),
191
	C64e(0x46e9af65af8c6546), C64e(0x9d7fe25ee2215e9d),
192
	C64e(0x3048782878602830), C64e(0x37cff8a1f86ea137),
193
	C64e(0x0a1b110f11140f0a), C64e(0x2febc4b5c45eb52f),
194
	C64e(0x0e151b091b1c090e), C64e(0x247e5a365a483624),
195
	C64e(0x1badb69bb6369b1b), C64e(0xdf98473d47a53ddf),
196
	C64e(0xcda76a266a8126cd), C64e(0x4ef5bb69bb9c694e),
197
	C64e(0x7f334ccd4cfecd7f), C64e(0xea50ba9fbacf9fea),
198
	C64e(0x123f2d1b2d241b12), C64e(0x1da4b99eb93a9e1d),
199
	C64e(0x58c49c749cb07458), C64e(0x3446722e72682e34),
200
	C64e(0x3641772d776c2d36), C64e(0xdc11cdb2cda3b2dc),
201
	C64e(0xb49d29ee2973eeb4), C64e(0x5b4d16fb16b6fb5b),
202
	C64e(0xa4a501f60153f6a4), C64e(0x76a1d74dd7ec4d76),
203
	C64e(0xb714a361a37561b7), C64e(0x7d3449ce49face7d),
204
	C64e(0x52df8d7b8da47b52), C64e(0xdd9f423e42a13edd),
205
	C64e(0x5ecd937193bc715e), C64e(0x13b1a297a2269713),
206
	C64e(0xa6a204f50457f5a6), C64e(0xb901b868b86968b9),
207
	C64e(0x0000000000000000), C64e(0xc1b5742c74992cc1),
208
	C64e(0x40e0a060a0806040), C64e(0xe3c2211f21dd1fe3),
209
	C64e(0x793a43c843f2c879), C64e(0xb69a2ced2c77edb6),
210
	C64e(0xd40dd9bed9b3bed4), C64e(0x8d47ca46ca01468d),
211
	C64e(0x671770d970ced967), C64e(0x72afdd4bdde44b72),
212
	C64e(0x94ed79de7933de94), C64e(0x98ff67d4672bd498),
213
	C64e(0xb09323e8237be8b0), C64e(0x855bde4ade114a85),
214
	C64e(0xbb06bd6bbd6d6bbb), C64e(0xc5bb7e2a7e912ac5),
215
	C64e(0x4f7b34e5349ee54f), C64e(0xedd73a163ac116ed),
216
	C64e(0x86d254c55417c586), C64e(0x9af862d7622fd79a),
217
	C64e(0x6699ff55ffcc5566), C64e(0x11b6a794a7229411),
218
	C64e(0x8ac04acf4a0fcf8a), C64e(0xe9d9301030c910e9),
219
	C64e(0x040e0a060a080604), C64e(0xfe66988198e781fe),
220
	C64e(0xa0ab0bf00b5bf0a0), C64e(0x78b4cc44ccf04478),
221
	C64e(0x25f0d5bad54aba25), C64e(0x4b753ee33e96e34b),
222
	C64e(0xa2ac0ef30e5ff3a2), C64e(0x5d4419fe19bafe5d),
223
	C64e(0x80db5bc05b1bc080), C64e(0x0580858a850a8a05),
224
	C64e(0x3fd3ecadec7ead3f), C64e(0x21fedfbcdf42bc21),
225
	C64e(0x70a8d848d8e04870), C64e(0xf1fd0c040cf904f1),
226
	C64e(0x63197adf7ac6df63), C64e(0x772f58c158eec177),
227
	C64e(0xaf309f759f4575af), C64e(0x42e7a563a5846342),
228
	C64e(0x2070503050403020), C64e(0xe5cb2e1a2ed11ae5),
229
	C64e(0xfdef120e12e10efd), C64e(0xbf08b76db7656dbf),
230
	C64e(0x8155d44cd4194c81), C64e(0x18243c143c301418),
231
	C64e(0x26795f355f4c3526), C64e(0xc3b2712f719d2fc3),
232
	C64e(0xbe8638e13867e1be), C64e(0x35c8fda2fd6aa235),
233
	C64e(0x88c74fcc4f0bcc88), C64e(0x2e654b394b5c392e),
234
	C64e(0x936af957f93d5793), C64e(0x55580df20daaf255),
235
	C64e(0xfc619d829de382fc), C64e(0x7ab3c947c9f4477a),
236
	C64e(0xc827efacef8bacc8), C64e(0xba8832e7326fe7ba),
237
	C64e(0x324f7d2b7d642b32), C64e(0xe642a495a4d795e6),
238
	C64e(0xc03bfba0fb9ba0c0), C64e(0x19aab398b3329819),
239
	C64e(0x9ef668d16827d19e), C64e(0xa322817f815d7fa3),
240
	C64e(0x44eeaa66aa886644), C64e(0x54d6827e82a87e54),
241
	C64e(0x3bdde6abe676ab3b), C64e(0x0b959e839e16830b),
242
	C64e(0x8cc945ca4503ca8c), C64e(0xc7bc7b297b9529c7),
243
	C64e(0x6b056ed36ed6d36b), C64e(0x286c443c44503c28),
244
	C64e(0xa72c8b798b5579a7), C64e(0xbc813de23d63e2bc),
245
	C64e(0x1631271d272c1d16), C64e(0xad379a769a4176ad),
246
	C64e(0xdb964d3b4dad3bdb), C64e(0x649efa56fac85664),
247
	C64e(0x74a6d24ed2e84e74), C64e(0x1436221e22281e14),
248
	C64e(0x92e476db763fdb92), C64e(0x0c121e0a1e180a0c),
249
	C64e(0x48fcb46cb4906c48), C64e(0xb88f37e4376be4b8),
250
	C64e(0x9f78e75de7255d9f), C64e(0xbd0fb26eb2616ebd),
251
	C64e(0x43692aef2a86ef43), C64e(0xc435f1a6f193a6c4),
252
	C64e(0x39dae3a8e372a839), C64e(0x31c6f7a4f762a431),
253
	C64e(0xd38a593759bd37d3), C64e(0xf274868b86ff8bf2),
254
	C64e(0xd583563256b132d5), C64e(0x8b4ec543c50d438b),
255
	C64e(0x6e85eb59ebdc596e), C64e(0xda18c2b7c2afb7da),
256
	C64e(0x018e8f8c8f028c01), C64e(0xb11dac64ac7964b1),
257
	C64e(0x9cf16dd26d23d29c), C64e(0x49723be03b92e049),
258
	C64e(0xd81fc7b4c7abb4d8), C64e(0xacb915fa1543faac),
259
	C64e(0xf3fa090709fd07f3), C64e(0xcfa06f256f8525cf),
260
	C64e(0xca20eaafea8fafca), C64e(0xf47d898e89f38ef4),
261
	C64e(0x476720e9208ee947), C64e(0x1038281828201810),
262
	C64e(0x6f0b64d564ded56f), C64e(0xf073838883fb88f0),
263
	C64e(0x4afbb16fb1946f4a), C64e(0x5cca967296b8725c),
264
	C64e(0x38546c246c702438), C64e(0x575f08f108aef157),
265
	C64e(0x732152c752e6c773), C64e(0x9764f351f3355197),
266
	C64e(0xcbae6523658d23cb), C64e(0xa125847c84597ca1),
267
	C64e(0xe857bf9cbfcb9ce8), C64e(0x3e5d6321637c213e),
268
	C64e(0x96ea7cdd7c37dd96), C64e(0x611e7fdc7fc2dc61),
269
	C64e(0x0d9c9186911a860d), C64e(0x0f9b9485941e850f),
270
	C64e(0xe04bab90abdb90e0), C64e(0x7cbac642c6f8427c),
271
	C64e(0x712657c457e2c471), C64e(0xcc29e5aae583aacc),
272
	C64e(0x90e373d8733bd890), C64e(0x06090f050f0c0506),
273
	C64e(0xf7f4030103f501f7), C64e(0x1c2a36123638121c),
274
	C64e(0xc23cfea3fe9fa3c2), C64e(0x6a8be15fe1d45f6a),
275
	C64e(0xaebe10f91047f9ae), C64e(0x69026bd06bd2d069),
276
	C64e(0x17bfa891a82e9117), C64e(0x9971e858e8295899),
277
	C64e(0x3a5369276974273a), C64e(0x27f7d0b9d04eb927),
278
	C64e(0xd991483848a938d9), C64e(0xebde351335cd13eb),
279
	C64e(0x2be5ceb3ce56b32b), C64e(0x2277553355443322),
280
	C64e(0xd204d6bbd6bfbbd2), C64e(0xa9399070904970a9),
281
	C64e(0x07878089800e8907), C64e(0x33c1f2a7f266a733),
282
	C64e(0x2decc1b6c15ab62d), C64e(0x3c5a66226678223c),
283
	C64e(0x15b8ad92ad2a9215), C64e(0xc9a96020608920c9),
284
	C64e(0x875cdb49db154987), C64e(0xaab01aff1a4fffaa),
285
	C64e(0x50d8887888a07850), C64e(0xa52b8e7a8e517aa5),
286
	C64e(0x03898a8f8a068f03), C64e(0x594a13f813b2f859),
287
	C64e(0x09929b809b128009), C64e(0x1a2339173934171a),
288
	C64e(0x651075da75cada65), C64e(0xd784533153b531d7),
289
	C64e(0x84d551c65113c684), C64e(0xd003d3b8d3bbb8d0),
290
	C64e(0x82dc5ec35e1fc382), C64e(0x29e2cbb0cb52b029),
291
	C64e(0x5ac3997799b4775a), C64e(0x1e2d3311333c111e),
292
	C64e(0x7b3d46cb46f6cb7b), C64e(0xa8b71ffc1f4bfca8),
293
	C64e(0x6d0c61d661dad66d), C64e(0x2c624e3a4e583a2c)
294
};
295

296
#if !SPH_SMALL_FOOTPRINT_GROESTL
297

298
static const sph_u64 T1[] = {
299
	C64e(0xc6c632f4a5f497a5), C64e(0xf8f86f978497eb84),
300
	C64e(0xeeee5eb099b0c799), C64e(0xf6f67a8c8d8cf78d),
301
	C64e(0xffffe8170d17e50d), C64e(0xd6d60adcbddcb7bd),
302
	C64e(0xdede16c8b1c8a7b1), C64e(0x91916dfc54fc3954),
303
	C64e(0x606090f050f0c050), C64e(0x0202070503050403),
304
	C64e(0xcece2ee0a9e087a9), C64e(0x5656d1877d87ac7d),
305
	C64e(0xe7e7cc2b192bd519), C64e(0xb5b513a662a67162),
306
	C64e(0x4d4d7c31e6319ae6), C64e(0xecec59b59ab5c39a),
307
	C64e(0x8f8f40cf45cf0545), C64e(0x1f1fa3bc9dbc3e9d),
308
	C64e(0x898949c040c00940), C64e(0xfafa68928792ef87),
309
	C64e(0xefefd03f153fc515), C64e(0xb2b29426eb267feb),
310
	C64e(0x8e8ece40c94007c9), C64e(0xfbfbe61d0b1ded0b),
311
	C64e(0x41416e2fec2f82ec), C64e(0xb3b31aa967a97d67),
312
	C64e(0x5f5f431cfd1cbefd), C64e(0x45456025ea258aea),
313
	C64e(0x2323f9dabfda46bf), C64e(0x53535102f702a6f7),
314
	C64e(0xe4e445a196a1d396), C64e(0x9b9b76ed5bed2d5b),
315
	C64e(0x7575285dc25deac2), C64e(0xe1e1c5241c24d91c),
316
	C64e(0x3d3dd4e9aee97aae), C64e(0x4c4cf2be6abe986a),
317
	C64e(0x6c6c82ee5aeed85a), C64e(0x7e7ebdc341c3fc41),
318
	C64e(0xf5f5f3060206f102), C64e(0x838352d14fd11d4f),
319
	C64e(0x68688ce45ce4d05c), C64e(0x51515607f407a2f4),
320
	C64e(0xd1d18d5c345cb934), C64e(0xf9f9e1180818e908),
321
	C64e(0xe2e24cae93aedf93), C64e(0xabab3e9573954d73),
322
	C64e(0x626297f553f5c453), C64e(0x2a2a6b413f41543f),
323
	C64e(0x08081c140c14100c), C64e(0x959563f652f63152),
324
	C64e(0x4646e9af65af8c65), C64e(0x9d9d7fe25ee2215e),
325
	C64e(0x3030487828786028), C64e(0x3737cff8a1f86ea1),
326
	C64e(0x0a0a1b110f11140f), C64e(0x2f2febc4b5c45eb5),
327
	C64e(0x0e0e151b091b1c09), C64e(0x24247e5a365a4836),
328
	C64e(0x1b1badb69bb6369b), C64e(0xdfdf98473d47a53d),
329
	C64e(0xcdcda76a266a8126), C64e(0x4e4ef5bb69bb9c69),
330
	C64e(0x7f7f334ccd4cfecd), C64e(0xeaea50ba9fbacf9f),
331
	C64e(0x12123f2d1b2d241b), C64e(0x1d1da4b99eb93a9e),
332
	C64e(0x5858c49c749cb074), C64e(0x343446722e72682e),
333
	C64e(0x363641772d776c2d), C64e(0xdcdc11cdb2cda3b2),
334
	C64e(0xb4b49d29ee2973ee), C64e(0x5b5b4d16fb16b6fb),
335
	C64e(0xa4a4a501f60153f6), C64e(0x7676a1d74dd7ec4d),
336
	C64e(0xb7b714a361a37561), C64e(0x7d7d3449ce49face),
337
	C64e(0x5252df8d7b8da47b), C64e(0xdddd9f423e42a13e),
338
	C64e(0x5e5ecd937193bc71), C64e(0x1313b1a297a22697),
339
	C64e(0xa6a6a204f50457f5), C64e(0xb9b901b868b86968),
340
	C64e(0x0000000000000000), C64e(0xc1c1b5742c74992c),
341
	C64e(0x4040e0a060a08060), C64e(0xe3e3c2211f21dd1f),
342
	C64e(0x79793a43c843f2c8), C64e(0xb6b69a2ced2c77ed),
343
	C64e(0xd4d40dd9bed9b3be), C64e(0x8d8d47ca46ca0146),
344
	C64e(0x67671770d970ced9), C64e(0x7272afdd4bdde44b),
345
	C64e(0x9494ed79de7933de), C64e(0x9898ff67d4672bd4),
346
	C64e(0xb0b09323e8237be8), C64e(0x85855bde4ade114a),
347
	C64e(0xbbbb06bd6bbd6d6b), C64e(0xc5c5bb7e2a7e912a),
348
	C64e(0x4f4f7b34e5349ee5), C64e(0xededd73a163ac116),
349
	C64e(0x8686d254c55417c5), C64e(0x9a9af862d7622fd7),
350
	C64e(0x666699ff55ffcc55), C64e(0x1111b6a794a72294),
351
	C64e(0x8a8ac04acf4a0fcf), C64e(0xe9e9d9301030c910),
352
	C64e(0x04040e0a060a0806), C64e(0xfefe66988198e781),
353
	C64e(0xa0a0ab0bf00b5bf0), C64e(0x7878b4cc44ccf044),
354
	C64e(0x2525f0d5bad54aba), C64e(0x4b4b753ee33e96e3),
355
	C64e(0xa2a2ac0ef30e5ff3), C64e(0x5d5d4419fe19bafe),
356
	C64e(0x8080db5bc05b1bc0), C64e(0x050580858a850a8a),
357
	C64e(0x3f3fd3ecadec7ead), C64e(0x2121fedfbcdf42bc),
358
	C64e(0x7070a8d848d8e048), C64e(0xf1f1fd0c040cf904),
359
	C64e(0x6363197adf7ac6df), C64e(0x77772f58c158eec1),
360
	C64e(0xafaf309f759f4575), C64e(0x4242e7a563a58463),
361
	C64e(0x2020705030504030), C64e(0xe5e5cb2e1a2ed11a),
362
	C64e(0xfdfdef120e12e10e), C64e(0xbfbf08b76db7656d),
363
	C64e(0x818155d44cd4194c), C64e(0x1818243c143c3014),
364
	C64e(0x2626795f355f4c35), C64e(0xc3c3b2712f719d2f),
365
	C64e(0xbebe8638e13867e1), C64e(0x3535c8fda2fd6aa2),
366
	C64e(0x8888c74fcc4f0bcc), C64e(0x2e2e654b394b5c39),
367
	C64e(0x93936af957f93d57), C64e(0x5555580df20daaf2),
368
	C64e(0xfcfc619d829de382), C64e(0x7a7ab3c947c9f447),
369
	C64e(0xc8c827efacef8bac), C64e(0xbaba8832e7326fe7),
370
	C64e(0x32324f7d2b7d642b), C64e(0xe6e642a495a4d795),
371
	C64e(0xc0c03bfba0fb9ba0), C64e(0x1919aab398b33298),
372
	C64e(0x9e9ef668d16827d1), C64e(0xa3a322817f815d7f),
373
	C64e(0x4444eeaa66aa8866), C64e(0x5454d6827e82a87e),
374
	C64e(0x3b3bdde6abe676ab), C64e(0x0b0b959e839e1683),
375
	C64e(0x8c8cc945ca4503ca), C64e(0xc7c7bc7b297b9529),
376
	C64e(0x6b6b056ed36ed6d3), C64e(0x28286c443c44503c),
377
	C64e(0xa7a72c8b798b5579), C64e(0xbcbc813de23d63e2),
378
	C64e(0x161631271d272c1d), C64e(0xadad379a769a4176),
379
	C64e(0xdbdb964d3b4dad3b), C64e(0x64649efa56fac856),
380
	C64e(0x7474a6d24ed2e84e), C64e(0x141436221e22281e),
381
	C64e(0x9292e476db763fdb), C64e(0x0c0c121e0a1e180a),
382
	C64e(0x4848fcb46cb4906c), C64e(0xb8b88f37e4376be4),
383
	C64e(0x9f9f78e75de7255d), C64e(0xbdbd0fb26eb2616e),
384
	C64e(0x4343692aef2a86ef), C64e(0xc4c435f1a6f193a6),
385
	C64e(0x3939dae3a8e372a8), C64e(0x3131c6f7a4f762a4),
386
	C64e(0xd3d38a593759bd37), C64e(0xf2f274868b86ff8b),
387
	C64e(0xd5d583563256b132), C64e(0x8b8b4ec543c50d43),
388
	C64e(0x6e6e85eb59ebdc59), C64e(0xdada18c2b7c2afb7),
389
	C64e(0x01018e8f8c8f028c), C64e(0xb1b11dac64ac7964),
390
	C64e(0x9c9cf16dd26d23d2), C64e(0x4949723be03b92e0),
391
	C64e(0xd8d81fc7b4c7abb4), C64e(0xacacb915fa1543fa),
392
	C64e(0xf3f3fa090709fd07), C64e(0xcfcfa06f256f8525),
393
	C64e(0xcaca20eaafea8faf), C64e(0xf4f47d898e89f38e),
394
	C64e(0x47476720e9208ee9), C64e(0x1010382818282018),
395
	C64e(0x6f6f0b64d564ded5), C64e(0xf0f073838883fb88),
396
	C64e(0x4a4afbb16fb1946f), C64e(0x5c5cca967296b872),
397
	C64e(0x3838546c246c7024), C64e(0x57575f08f108aef1),
398
	C64e(0x73732152c752e6c7), C64e(0x979764f351f33551),
399
	C64e(0xcbcbae6523658d23), C64e(0xa1a125847c84597c),
400
	C64e(0xe8e857bf9cbfcb9c), C64e(0x3e3e5d6321637c21),
401
	C64e(0x9696ea7cdd7c37dd), C64e(0x61611e7fdc7fc2dc),
402
	C64e(0x0d0d9c9186911a86), C64e(0x0f0f9b9485941e85),
403
	C64e(0xe0e04bab90abdb90), C64e(0x7c7cbac642c6f842),
404
	C64e(0x71712657c457e2c4), C64e(0xcccc29e5aae583aa),
405
	C64e(0x9090e373d8733bd8), C64e(0x0606090f050f0c05),
406
	C64e(0xf7f7f4030103f501), C64e(0x1c1c2a3612363812),
407
	C64e(0xc2c23cfea3fe9fa3), C64e(0x6a6a8be15fe1d45f),
408
	C64e(0xaeaebe10f91047f9), C64e(0x6969026bd06bd2d0),
409
	C64e(0x1717bfa891a82e91), C64e(0x999971e858e82958),
410
	C64e(0x3a3a536927697427), C64e(0x2727f7d0b9d04eb9),
411
	C64e(0xd9d991483848a938), C64e(0xebebde351335cd13),
412
	C64e(0x2b2be5ceb3ce56b3), C64e(0x2222775533554433),
413
	C64e(0xd2d204d6bbd6bfbb), C64e(0xa9a9399070904970),
414
	C64e(0x0707878089800e89), C64e(0x3333c1f2a7f266a7),
415
	C64e(0x2d2decc1b6c15ab6), C64e(0x3c3c5a6622667822),
416
	C64e(0x1515b8ad92ad2a92), C64e(0xc9c9a96020608920),
417
	C64e(0x87875cdb49db1549), C64e(0xaaaab01aff1a4fff),
418
	C64e(0x5050d8887888a078), C64e(0xa5a52b8e7a8e517a),
419
	C64e(0x0303898a8f8a068f), C64e(0x59594a13f813b2f8),
420
	C64e(0x0909929b809b1280), C64e(0x1a1a233917393417),
421
	C64e(0x65651075da75cada), C64e(0xd7d784533153b531),
422
	C64e(0x8484d551c65113c6), C64e(0xd0d003d3b8d3bbb8),
423
	C64e(0x8282dc5ec35e1fc3), C64e(0x2929e2cbb0cb52b0),
424
	C64e(0x5a5ac3997799b477), C64e(0x1e1e2d3311333c11),
425
	C64e(0x7b7b3d46cb46f6cb), C64e(0xa8a8b71ffc1f4bfc),
426
	C64e(0x6d6d0c61d661dad6), C64e(0x2c2c624e3a4e583a)
427
};
428

429
static const sph_u64 T2[] = {
430
	C64e(0xa5c6c632f4a5f497), C64e(0x84f8f86f978497eb),
431
	C64e(0x99eeee5eb099b0c7), C64e(0x8df6f67a8c8d8cf7),
432
	C64e(0x0dffffe8170d17e5), C64e(0xbdd6d60adcbddcb7),
433
	C64e(0xb1dede16c8b1c8a7), C64e(0x5491916dfc54fc39),
434
	C64e(0x50606090f050f0c0), C64e(0x0302020705030504),
435
	C64e(0xa9cece2ee0a9e087), C64e(0x7d5656d1877d87ac),
436
	C64e(0x19e7e7cc2b192bd5), C64e(0x62b5b513a662a671),
437
	C64e(0xe64d4d7c31e6319a), C64e(0x9aecec59b59ab5c3),
438
	C64e(0x458f8f40cf45cf05), C64e(0x9d1f1fa3bc9dbc3e),
439
	C64e(0x40898949c040c009), C64e(0x87fafa68928792ef),
440
	C64e(0x15efefd03f153fc5), C64e(0xebb2b29426eb267f),
441
	C64e(0xc98e8ece40c94007), C64e(0x0bfbfbe61d0b1ded),
442
	C64e(0xec41416e2fec2f82), C64e(0x67b3b31aa967a97d),
443
	C64e(0xfd5f5f431cfd1cbe), C64e(0xea45456025ea258a),
444
	C64e(0xbf2323f9dabfda46), C64e(0xf753535102f702a6),
445
	C64e(0x96e4e445a196a1d3), C64e(0x5b9b9b76ed5bed2d),
446
	C64e(0xc27575285dc25dea), C64e(0x1ce1e1c5241c24d9),
447
	C64e(0xae3d3dd4e9aee97a), C64e(0x6a4c4cf2be6abe98),
448
	C64e(0x5a6c6c82ee5aeed8), C64e(0x417e7ebdc341c3fc),
449
	C64e(0x02f5f5f3060206f1), C64e(0x4f838352d14fd11d),
450
	C64e(0x5c68688ce45ce4d0), C64e(0xf451515607f407a2),
451
	C64e(0x34d1d18d5c345cb9), C64e(0x08f9f9e1180818e9),
452
	C64e(0x93e2e24cae93aedf), C64e(0x73abab3e9573954d),
453
	C64e(0x53626297f553f5c4), C64e(0x3f2a2a6b413f4154),
454
	C64e(0x0c08081c140c1410), C64e(0x52959563f652f631),
455
	C64e(0x654646e9af65af8c), C64e(0x5e9d9d7fe25ee221),
456
	C64e(0x2830304878287860), C64e(0xa13737cff8a1f86e),
457
	C64e(0x0f0a0a1b110f1114), C64e(0xb52f2febc4b5c45e),
458
	C64e(0x090e0e151b091b1c), C64e(0x3624247e5a365a48),
459
	C64e(0x9b1b1badb69bb636), C64e(0x3ddfdf98473d47a5),
460
	C64e(0x26cdcda76a266a81), C64e(0x694e4ef5bb69bb9c),
461
	C64e(0xcd7f7f334ccd4cfe), C64e(0x9feaea50ba9fbacf),
462
	C64e(0x1b12123f2d1b2d24), C64e(0x9e1d1da4b99eb93a),
463
	C64e(0x745858c49c749cb0), C64e(0x2e343446722e7268),
464
	C64e(0x2d363641772d776c), C64e(0xb2dcdc11cdb2cda3),
465
	C64e(0xeeb4b49d29ee2973), C64e(0xfb5b5b4d16fb16b6),
466
	C64e(0xf6a4a4a501f60153), C64e(0x4d7676a1d74dd7ec),
467
	C64e(0x61b7b714a361a375), C64e(0xce7d7d3449ce49fa),
468
	C64e(0x7b5252df8d7b8da4), C64e(0x3edddd9f423e42a1),
469
	C64e(0x715e5ecd937193bc), C64e(0x971313b1a297a226),
470
	C64e(0xf5a6a6a204f50457), C64e(0x68b9b901b868b869),
471
	C64e(0x0000000000000000), C64e(0x2cc1c1b5742c7499),
472
	C64e(0x604040e0a060a080), C64e(0x1fe3e3c2211f21dd),
473
	C64e(0xc879793a43c843f2), C64e(0xedb6b69a2ced2c77),
474
	C64e(0xbed4d40dd9bed9b3), C64e(0x468d8d47ca46ca01),
475
	C64e(0xd967671770d970ce), C64e(0x4b7272afdd4bdde4),
476
	C64e(0xde9494ed79de7933), C64e(0xd49898ff67d4672b),
477
	C64e(0xe8b0b09323e8237b), C64e(0x4a85855bde4ade11),
478
	C64e(0x6bbbbb06bd6bbd6d), C64e(0x2ac5c5bb7e2a7e91),
479
	C64e(0xe54f4f7b34e5349e), C64e(0x16ededd73a163ac1),
480
	C64e(0xc58686d254c55417), C64e(0xd79a9af862d7622f),
481
	C64e(0x55666699ff55ffcc), C64e(0x941111b6a794a722),
482
	C64e(0xcf8a8ac04acf4a0f), C64e(0x10e9e9d9301030c9),
483
	C64e(0x0604040e0a060a08), C64e(0x81fefe66988198e7),
484
	C64e(0xf0a0a0ab0bf00b5b), C64e(0x447878b4cc44ccf0),
485
	C64e(0xba2525f0d5bad54a), C64e(0xe34b4b753ee33e96),
486
	C64e(0xf3a2a2ac0ef30e5f), C64e(0xfe5d5d4419fe19ba),
487
	C64e(0xc08080db5bc05b1b), C64e(0x8a050580858a850a),
488
	C64e(0xad3f3fd3ecadec7e), C64e(0xbc2121fedfbcdf42),
489
	C64e(0x487070a8d848d8e0), C64e(0x04f1f1fd0c040cf9),
490
	C64e(0xdf6363197adf7ac6), C64e(0xc177772f58c158ee),
491
	C64e(0x75afaf309f759f45), C64e(0x634242e7a563a584),
492
	C64e(0x3020207050305040), C64e(0x1ae5e5cb2e1a2ed1),
493
	C64e(0x0efdfdef120e12e1), C64e(0x6dbfbf08b76db765),
494
	C64e(0x4c818155d44cd419), C64e(0x141818243c143c30),
495
	C64e(0x352626795f355f4c), C64e(0x2fc3c3b2712f719d),
496
	C64e(0xe1bebe8638e13867), C64e(0xa23535c8fda2fd6a),
497
	C64e(0xcc8888c74fcc4f0b), C64e(0x392e2e654b394b5c),
498
	C64e(0x5793936af957f93d), C64e(0xf25555580df20daa),
499
	C64e(0x82fcfc619d829de3), C64e(0x477a7ab3c947c9f4),
500
	C64e(0xacc8c827efacef8b), C64e(0xe7baba8832e7326f),
501
	C64e(0x2b32324f7d2b7d64), C64e(0x95e6e642a495a4d7),
502
	C64e(0xa0c0c03bfba0fb9b), C64e(0x981919aab398b332),
503
	C64e(0xd19e9ef668d16827), C64e(0x7fa3a322817f815d),
504
	C64e(0x664444eeaa66aa88), C64e(0x7e5454d6827e82a8),
505
	C64e(0xab3b3bdde6abe676), C64e(0x830b0b959e839e16),
506
	C64e(0xca8c8cc945ca4503), C64e(0x29c7c7bc7b297b95),
507
	C64e(0xd36b6b056ed36ed6), C64e(0x3c28286c443c4450),
508
	C64e(0x79a7a72c8b798b55), C64e(0xe2bcbc813de23d63),
509
	C64e(0x1d161631271d272c), C64e(0x76adad379a769a41),
510
	C64e(0x3bdbdb964d3b4dad), C64e(0x5664649efa56fac8),
511
	C64e(0x4e7474a6d24ed2e8), C64e(0x1e141436221e2228),
512
	C64e(0xdb9292e476db763f), C64e(0x0a0c0c121e0a1e18),
513
	C64e(0x6c4848fcb46cb490), C64e(0xe4b8b88f37e4376b),
514
	C64e(0x5d9f9f78e75de725), C64e(0x6ebdbd0fb26eb261),
515
	C64e(0xef4343692aef2a86), C64e(0xa6c4c435f1a6f193),
516
	C64e(0xa83939dae3a8e372), C64e(0xa43131c6f7a4f762),
517
	C64e(0x37d3d38a593759bd), C64e(0x8bf2f274868b86ff),
518
	C64e(0x32d5d583563256b1), C64e(0x438b8b4ec543c50d),
519
	C64e(0x596e6e85eb59ebdc), C64e(0xb7dada18c2b7c2af),
520
	C64e(0x8c01018e8f8c8f02), C64e(0x64b1b11dac64ac79),
521
	C64e(0xd29c9cf16dd26d23), C64e(0xe04949723be03b92),
522
	C64e(0xb4d8d81fc7b4c7ab), C64e(0xfaacacb915fa1543),
523
	C64e(0x07f3f3fa090709fd), C64e(0x25cfcfa06f256f85),
524
	C64e(0xafcaca20eaafea8f), C64e(0x8ef4f47d898e89f3),
525
	C64e(0xe947476720e9208e), C64e(0x1810103828182820),
526
	C64e(0xd56f6f0b64d564de), C64e(0x88f0f073838883fb),
527
	C64e(0x6f4a4afbb16fb194), C64e(0x725c5cca967296b8),
528
	C64e(0x243838546c246c70), C64e(0xf157575f08f108ae),
529
	C64e(0xc773732152c752e6), C64e(0x51979764f351f335),
530
	C64e(0x23cbcbae6523658d), C64e(0x7ca1a125847c8459),
531
	C64e(0x9ce8e857bf9cbfcb), C64e(0x213e3e5d6321637c),
532
	C64e(0xdd9696ea7cdd7c37), C64e(0xdc61611e7fdc7fc2),
533
	C64e(0x860d0d9c9186911a), C64e(0x850f0f9b9485941e),
534
	C64e(0x90e0e04bab90abdb), C64e(0x427c7cbac642c6f8),
535
	C64e(0xc471712657c457e2), C64e(0xaacccc29e5aae583),
536
	C64e(0xd89090e373d8733b), C64e(0x050606090f050f0c),
537
	C64e(0x01f7f7f4030103f5), C64e(0x121c1c2a36123638),
538
	C64e(0xa3c2c23cfea3fe9f), C64e(0x5f6a6a8be15fe1d4),
539
	C64e(0xf9aeaebe10f91047), C64e(0xd06969026bd06bd2),
540
	C64e(0x911717bfa891a82e), C64e(0x58999971e858e829),
541
	C64e(0x273a3a5369276974), C64e(0xb92727f7d0b9d04e),
542
	C64e(0x38d9d991483848a9), C64e(0x13ebebde351335cd),
543
	C64e(0xb32b2be5ceb3ce56), C64e(0x3322227755335544),
544
	C64e(0xbbd2d204d6bbd6bf), C64e(0x70a9a93990709049),
545
	C64e(0x890707878089800e), C64e(0xa73333c1f2a7f266),
546
	C64e(0xb62d2decc1b6c15a), C64e(0x223c3c5a66226678),
547
	C64e(0x921515b8ad92ad2a), C64e(0x20c9c9a960206089),
548
	C64e(0x4987875cdb49db15), C64e(0xffaaaab01aff1a4f),
549
	C64e(0x785050d8887888a0), C64e(0x7aa5a52b8e7a8e51),
550
	C64e(0x8f0303898a8f8a06), C64e(0xf859594a13f813b2),
551
	C64e(0x800909929b809b12), C64e(0x171a1a2339173934),
552
	C64e(0xda65651075da75ca), C64e(0x31d7d784533153b5),
553
	C64e(0xc68484d551c65113), C64e(0xb8d0d003d3b8d3bb),
554
	C64e(0xc38282dc5ec35e1f), C64e(0xb02929e2cbb0cb52),
555
	C64e(0x775a5ac3997799b4), C64e(0x111e1e2d3311333c),
556
	C64e(0xcb7b7b3d46cb46f6), C64e(0xfca8a8b71ffc1f4b),
557
	C64e(0xd66d6d0c61d661da), C64e(0x3a2c2c624e3a4e58)
558
};
559

560
static const sph_u64 T3[] = {
561
	C64e(0x97a5c6c632f4a5f4), C64e(0xeb84f8f86f978497),
562
	C64e(0xc799eeee5eb099b0), C64e(0xf78df6f67a8c8d8c),
563
	C64e(0xe50dffffe8170d17), C64e(0xb7bdd6d60adcbddc),
564
	C64e(0xa7b1dede16c8b1c8), C64e(0x395491916dfc54fc),
565
	C64e(0xc050606090f050f0), C64e(0x0403020207050305),
566
	C64e(0x87a9cece2ee0a9e0), C64e(0xac7d5656d1877d87),
567
	C64e(0xd519e7e7cc2b192b), C64e(0x7162b5b513a662a6),
568
	C64e(0x9ae64d4d7c31e631), C64e(0xc39aecec59b59ab5),
569
	C64e(0x05458f8f40cf45cf), C64e(0x3e9d1f1fa3bc9dbc),
570
	C64e(0x0940898949c040c0), C64e(0xef87fafa68928792),
571
	C64e(0xc515efefd03f153f), C64e(0x7febb2b29426eb26),
572
	C64e(0x07c98e8ece40c940), C64e(0xed0bfbfbe61d0b1d),
573
	C64e(0x82ec41416e2fec2f), C64e(0x7d67b3b31aa967a9),
574
	C64e(0xbefd5f5f431cfd1c), C64e(0x8aea45456025ea25),
575
	C64e(0x46bf2323f9dabfda), C64e(0xa6f753535102f702),
576
	C64e(0xd396e4e445a196a1), C64e(0x2d5b9b9b76ed5bed),
577
	C64e(0xeac27575285dc25d), C64e(0xd91ce1e1c5241c24),
578
	C64e(0x7aae3d3dd4e9aee9), C64e(0x986a4c4cf2be6abe),
579
	C64e(0xd85a6c6c82ee5aee), C64e(0xfc417e7ebdc341c3),
580
	C64e(0xf102f5f5f3060206), C64e(0x1d4f838352d14fd1),
581
	C64e(0xd05c68688ce45ce4), C64e(0xa2f451515607f407),
582
	C64e(0xb934d1d18d5c345c), C64e(0xe908f9f9e1180818),
583
	C64e(0xdf93e2e24cae93ae), C64e(0x4d73abab3e957395),
584
	C64e(0xc453626297f553f5), C64e(0x543f2a2a6b413f41),
585
	C64e(0x100c08081c140c14), C64e(0x3152959563f652f6),
586
	C64e(0x8c654646e9af65af), C64e(0x215e9d9d7fe25ee2),
587
	C64e(0x6028303048782878), C64e(0x6ea13737cff8a1f8),
588
	C64e(0x140f0a0a1b110f11), C64e(0x5eb52f2febc4b5c4),
589
	C64e(0x1c090e0e151b091b), C64e(0x483624247e5a365a),
590
	C64e(0x369b1b1badb69bb6), C64e(0xa53ddfdf98473d47),
591
	C64e(0x8126cdcda76a266a), C64e(0x9c694e4ef5bb69bb),
592
	C64e(0xfecd7f7f334ccd4c), C64e(0xcf9feaea50ba9fba),
593
	C64e(0x241b12123f2d1b2d), C64e(0x3a9e1d1da4b99eb9),
594
	C64e(0xb0745858c49c749c), C64e(0x682e343446722e72),
595
	C64e(0x6c2d363641772d77), C64e(0xa3b2dcdc11cdb2cd),
596
	C64e(0x73eeb4b49d29ee29), C64e(0xb6fb5b5b4d16fb16),
597
	C64e(0x53f6a4a4a501f601), C64e(0xec4d7676a1d74dd7),
598
	C64e(0x7561b7b714a361a3), C64e(0xface7d7d3449ce49),
599
	C64e(0xa47b5252df8d7b8d), C64e(0xa13edddd9f423e42),
600
	C64e(0xbc715e5ecd937193), C64e(0x26971313b1a297a2),
601
	C64e(0x57f5a6a6a204f504), C64e(0x6968b9b901b868b8),
602
	C64e(0x0000000000000000), C64e(0x992cc1c1b5742c74),
603
	C64e(0x80604040e0a060a0), C64e(0xdd1fe3e3c2211f21),
604
	C64e(0xf2c879793a43c843), C64e(0x77edb6b69a2ced2c),
605
	C64e(0xb3bed4d40dd9bed9), C64e(0x01468d8d47ca46ca),
606
	C64e(0xced967671770d970), C64e(0xe44b7272afdd4bdd),
607
	C64e(0x33de9494ed79de79), C64e(0x2bd49898ff67d467),
608
	C64e(0x7be8b0b09323e823), C64e(0x114a85855bde4ade),
609
	C64e(0x6d6bbbbb06bd6bbd), C64e(0x912ac5c5bb7e2a7e),
610
	C64e(0x9ee54f4f7b34e534), C64e(0xc116ededd73a163a),
611
	C64e(0x17c58686d254c554), C64e(0x2fd79a9af862d762),
612
	C64e(0xcc55666699ff55ff), C64e(0x22941111b6a794a7),
613
	C64e(0x0fcf8a8ac04acf4a), C64e(0xc910e9e9d9301030),
614
	C64e(0x080604040e0a060a), C64e(0xe781fefe66988198),
615
	C64e(0x5bf0a0a0ab0bf00b), C64e(0xf0447878b4cc44cc),
616
	C64e(0x4aba2525f0d5bad5), C64e(0x96e34b4b753ee33e),
617
	C64e(0x5ff3a2a2ac0ef30e), C64e(0xbafe5d5d4419fe19),
618
	C64e(0x1bc08080db5bc05b), C64e(0x0a8a050580858a85),
619
	C64e(0x7ead3f3fd3ecadec), C64e(0x42bc2121fedfbcdf),
620
	C64e(0xe0487070a8d848d8), C64e(0xf904f1f1fd0c040c),
621
	C64e(0xc6df6363197adf7a), C64e(0xeec177772f58c158),
622
	C64e(0x4575afaf309f759f), C64e(0x84634242e7a563a5),
623
	C64e(0x4030202070503050), C64e(0xd11ae5e5cb2e1a2e),
624
	C64e(0xe10efdfdef120e12), C64e(0x656dbfbf08b76db7),
625
	C64e(0x194c818155d44cd4), C64e(0x30141818243c143c),
626
	C64e(0x4c352626795f355f), C64e(0x9d2fc3c3b2712f71),
627
	C64e(0x67e1bebe8638e138), C64e(0x6aa23535c8fda2fd),
628
	C64e(0x0bcc8888c74fcc4f), C64e(0x5c392e2e654b394b),
629
	C64e(0x3d5793936af957f9), C64e(0xaaf25555580df20d),
630
	C64e(0xe382fcfc619d829d), C64e(0xf4477a7ab3c947c9),
631
	C64e(0x8bacc8c827efacef), C64e(0x6fe7baba8832e732),
632
	C64e(0x642b32324f7d2b7d), C64e(0xd795e6e642a495a4),
633
	C64e(0x9ba0c0c03bfba0fb), C64e(0x32981919aab398b3),
634
	C64e(0x27d19e9ef668d168), C64e(0x5d7fa3a322817f81),
635
	C64e(0x88664444eeaa66aa), C64e(0xa87e5454d6827e82),
636
	C64e(0x76ab3b3bdde6abe6), C64e(0x16830b0b959e839e),
637
	C64e(0x03ca8c8cc945ca45), C64e(0x9529c7c7bc7b297b),
638
	C64e(0xd6d36b6b056ed36e), C64e(0x503c28286c443c44),
639
	C64e(0x5579a7a72c8b798b), C64e(0x63e2bcbc813de23d),
640
	C64e(0x2c1d161631271d27), C64e(0x4176adad379a769a),
641
	C64e(0xad3bdbdb964d3b4d), C64e(0xc85664649efa56fa),
642
	C64e(0xe84e7474a6d24ed2), C64e(0x281e141436221e22),
643
	C64e(0x3fdb9292e476db76), C64e(0x180a0c0c121e0a1e),
644
	C64e(0x906c4848fcb46cb4), C64e(0x6be4b8b88f37e437),
645
	C64e(0x255d9f9f78e75de7), C64e(0x616ebdbd0fb26eb2),
646
	C64e(0x86ef4343692aef2a), C64e(0x93a6c4c435f1a6f1),
647
	C64e(0x72a83939dae3a8e3), C64e(0x62a43131c6f7a4f7),
648
	C64e(0xbd37d3d38a593759), C64e(0xff8bf2f274868b86),
649
	C64e(0xb132d5d583563256), C64e(0x0d438b8b4ec543c5),
650
	C64e(0xdc596e6e85eb59eb), C64e(0xafb7dada18c2b7c2),
651
	C64e(0x028c01018e8f8c8f), C64e(0x7964b1b11dac64ac),
652
	C64e(0x23d29c9cf16dd26d), C64e(0x92e04949723be03b),
653
	C64e(0xabb4d8d81fc7b4c7), C64e(0x43faacacb915fa15),
654
	C64e(0xfd07f3f3fa090709), C64e(0x8525cfcfa06f256f),
655
	C64e(0x8fafcaca20eaafea), C64e(0xf38ef4f47d898e89),
656
	C64e(0x8ee947476720e920), C64e(0x2018101038281828),
657
	C64e(0xded56f6f0b64d564), C64e(0xfb88f0f073838883),
658
	C64e(0x946f4a4afbb16fb1), C64e(0xb8725c5cca967296),
659
	C64e(0x70243838546c246c), C64e(0xaef157575f08f108),
660
	C64e(0xe6c773732152c752), C64e(0x3551979764f351f3),
661
	C64e(0x8d23cbcbae652365), C64e(0x597ca1a125847c84),
662
	C64e(0xcb9ce8e857bf9cbf), C64e(0x7c213e3e5d632163),
663
	C64e(0x37dd9696ea7cdd7c), C64e(0xc2dc61611e7fdc7f),
664
	C64e(0x1a860d0d9c918691), C64e(0x1e850f0f9b948594),
665
	C64e(0xdb90e0e04bab90ab), C64e(0xf8427c7cbac642c6),
666
	C64e(0xe2c471712657c457), C64e(0x83aacccc29e5aae5),
667
	C64e(0x3bd89090e373d873), C64e(0x0c050606090f050f),
668
	C64e(0xf501f7f7f4030103), C64e(0x38121c1c2a361236),
669
	C64e(0x9fa3c2c23cfea3fe), C64e(0xd45f6a6a8be15fe1),
670
	C64e(0x47f9aeaebe10f910), C64e(0xd2d06969026bd06b),
671
	C64e(0x2e911717bfa891a8), C64e(0x2958999971e858e8),
672
	C64e(0x74273a3a53692769), C64e(0x4eb92727f7d0b9d0),
673
	C64e(0xa938d9d991483848), C64e(0xcd13ebebde351335),
674
	C64e(0x56b32b2be5ceb3ce), C64e(0x4433222277553355),
675
	C64e(0xbfbbd2d204d6bbd6), C64e(0x4970a9a939907090),
676
	C64e(0x0e89070787808980), C64e(0x66a73333c1f2a7f2),
677
	C64e(0x5ab62d2decc1b6c1), C64e(0x78223c3c5a662266),
678
	C64e(0x2a921515b8ad92ad), C64e(0x8920c9c9a9602060),
679
	C64e(0x154987875cdb49db), C64e(0x4fffaaaab01aff1a),
680
	C64e(0xa0785050d8887888), C64e(0x517aa5a52b8e7a8e),
681
	C64e(0x068f0303898a8f8a), C64e(0xb2f859594a13f813),
682
	C64e(0x12800909929b809b), C64e(0x34171a1a23391739),
683
	C64e(0xcada65651075da75), C64e(0xb531d7d784533153),
684
	C64e(0x13c68484d551c651), C64e(0xbbb8d0d003d3b8d3),
685
	C64e(0x1fc38282dc5ec35e), C64e(0x52b02929e2cbb0cb),
686
	C64e(0xb4775a5ac3997799), C64e(0x3c111e1e2d331133),
687
	C64e(0xf6cb7b7b3d46cb46), C64e(0x4bfca8a8b71ffc1f),
688
	C64e(0xdad66d6d0c61d661), C64e(0x583a2c2c624e3a4e)
689
};
690

691
#endif
692

693
static const sph_u64 T4[] = {
694
	C64e(0xf497a5c6c632f4a5), C64e(0x97eb84f8f86f9784),
695
	C64e(0xb0c799eeee5eb099), C64e(0x8cf78df6f67a8c8d),
696
	C64e(0x17e50dffffe8170d), C64e(0xdcb7bdd6d60adcbd),
697
	C64e(0xc8a7b1dede16c8b1), C64e(0xfc395491916dfc54),
698
	C64e(0xf0c050606090f050), C64e(0x0504030202070503),
699
	C64e(0xe087a9cece2ee0a9), C64e(0x87ac7d5656d1877d),
700
	C64e(0x2bd519e7e7cc2b19), C64e(0xa67162b5b513a662),
701
	C64e(0x319ae64d4d7c31e6), C64e(0xb5c39aecec59b59a),
702
	C64e(0xcf05458f8f40cf45), C64e(0xbc3e9d1f1fa3bc9d),
703
	C64e(0xc00940898949c040), C64e(0x92ef87fafa689287),
704
	C64e(0x3fc515efefd03f15), C64e(0x267febb2b29426eb),
705
	C64e(0x4007c98e8ece40c9), C64e(0x1ded0bfbfbe61d0b),
706
	C64e(0x2f82ec41416e2fec), C64e(0xa97d67b3b31aa967),
707
	C64e(0x1cbefd5f5f431cfd), C64e(0x258aea45456025ea),
708
	C64e(0xda46bf2323f9dabf), C64e(0x02a6f753535102f7),
709
	C64e(0xa1d396e4e445a196), C64e(0xed2d5b9b9b76ed5b),
710
	C64e(0x5deac27575285dc2), C64e(0x24d91ce1e1c5241c),
711
	C64e(0xe97aae3d3dd4e9ae), C64e(0xbe986a4c4cf2be6a),
712
	C64e(0xeed85a6c6c82ee5a), C64e(0xc3fc417e7ebdc341),
713
	C64e(0x06f102f5f5f30602), C64e(0xd11d4f838352d14f),
714
	C64e(0xe4d05c68688ce45c), C64e(0x07a2f451515607f4),
715
	C64e(0x5cb934d1d18d5c34), C64e(0x18e908f9f9e11808),
716
	C64e(0xaedf93e2e24cae93), C64e(0x954d73abab3e9573),
717
	C64e(0xf5c453626297f553), C64e(0x41543f2a2a6b413f),
718
	C64e(0x14100c08081c140c), C64e(0xf63152959563f652),
719
	C64e(0xaf8c654646e9af65), C64e(0xe2215e9d9d7fe25e),
720
	C64e(0x7860283030487828), C64e(0xf86ea13737cff8a1),
721
	C64e(0x11140f0a0a1b110f), C64e(0xc45eb52f2febc4b5),
722
	C64e(0x1b1c090e0e151b09), C64e(0x5a483624247e5a36),
723
	C64e(0xb6369b1b1badb69b), C64e(0x47a53ddfdf98473d),
724
	C64e(0x6a8126cdcda76a26), C64e(0xbb9c694e4ef5bb69),
725
	C64e(0x4cfecd7f7f334ccd), C64e(0xbacf9feaea50ba9f),
726
	C64e(0x2d241b12123f2d1b), C64e(0xb93a9e1d1da4b99e),
727
	C64e(0x9cb0745858c49c74), C64e(0x72682e343446722e),
728
	C64e(0x776c2d363641772d), C64e(0xcda3b2dcdc11cdb2),
729
	C64e(0x2973eeb4b49d29ee), C64e(0x16b6fb5b5b4d16fb),
730
	C64e(0x0153f6a4a4a501f6), C64e(0xd7ec4d7676a1d74d),
731
	C64e(0xa37561b7b714a361), C64e(0x49face7d7d3449ce),
732
	C64e(0x8da47b5252df8d7b), C64e(0x42a13edddd9f423e),
733
	C64e(0x93bc715e5ecd9371), C64e(0xa226971313b1a297),
734
	C64e(0x0457f5a6a6a204f5), C64e(0xb86968b9b901b868),
735
	C64e(0x0000000000000000), C64e(0x74992cc1c1b5742c),
736
	C64e(0xa080604040e0a060), C64e(0x21dd1fe3e3c2211f),
737
	C64e(0x43f2c879793a43c8), C64e(0x2c77edb6b69a2ced),
738
	C64e(0xd9b3bed4d40dd9be), C64e(0xca01468d8d47ca46),
739
	C64e(0x70ced967671770d9), C64e(0xdde44b7272afdd4b),
740
	C64e(0x7933de9494ed79de), C64e(0x672bd49898ff67d4),
741
	C64e(0x237be8b0b09323e8), C64e(0xde114a85855bde4a),
742
	C64e(0xbd6d6bbbbb06bd6b), C64e(0x7e912ac5c5bb7e2a),
743
	C64e(0x349ee54f4f7b34e5), C64e(0x3ac116ededd73a16),
744
	C64e(0x5417c58686d254c5), C64e(0x622fd79a9af862d7),
745
	C64e(0xffcc55666699ff55), C64e(0xa722941111b6a794),
746
	C64e(0x4a0fcf8a8ac04acf), C64e(0x30c910e9e9d93010),
747
	C64e(0x0a080604040e0a06), C64e(0x98e781fefe669881),
748
	C64e(0x0b5bf0a0a0ab0bf0), C64e(0xccf0447878b4cc44),
749
	C64e(0xd54aba2525f0d5ba), C64e(0x3e96e34b4b753ee3),
750
	C64e(0x0e5ff3a2a2ac0ef3), C64e(0x19bafe5d5d4419fe),
751
	C64e(0x5b1bc08080db5bc0), C64e(0x850a8a050580858a),
752
	C64e(0xec7ead3f3fd3ecad), C64e(0xdf42bc2121fedfbc),
753
	C64e(0xd8e0487070a8d848), C64e(0x0cf904f1f1fd0c04),
754
	C64e(0x7ac6df6363197adf), C64e(0x58eec177772f58c1),
755
	C64e(0x9f4575afaf309f75), C64e(0xa584634242e7a563),
756
	C64e(0x5040302020705030), C64e(0x2ed11ae5e5cb2e1a),
757
	C64e(0x12e10efdfdef120e), C64e(0xb7656dbfbf08b76d),
758
	C64e(0xd4194c818155d44c), C64e(0x3c30141818243c14),
759
	C64e(0x5f4c352626795f35), C64e(0x719d2fc3c3b2712f),
760
	C64e(0x3867e1bebe8638e1), C64e(0xfd6aa23535c8fda2),
761
	C64e(0x4f0bcc8888c74fcc), C64e(0x4b5c392e2e654b39),
762
	C64e(0xf93d5793936af957), C64e(0x0daaf25555580df2),
763
	C64e(0x9de382fcfc619d82), C64e(0xc9f4477a7ab3c947),
764
	C64e(0xef8bacc8c827efac), C64e(0x326fe7baba8832e7),
765
	C64e(0x7d642b32324f7d2b), C64e(0xa4d795e6e642a495),
766
	C64e(0xfb9ba0c0c03bfba0), C64e(0xb332981919aab398),
767
	C64e(0x6827d19e9ef668d1), C64e(0x815d7fa3a322817f),
768
	C64e(0xaa88664444eeaa66), C64e(0x82a87e5454d6827e),
769
	C64e(0xe676ab3b3bdde6ab), C64e(0x9e16830b0b959e83),
770
	C64e(0x4503ca8c8cc945ca), C64e(0x7b9529c7c7bc7b29),
771
	C64e(0x6ed6d36b6b056ed3), C64e(0x44503c28286c443c),
772
	C64e(0x8b5579a7a72c8b79), C64e(0x3d63e2bcbc813de2),
773
	C64e(0x272c1d161631271d), C64e(0x9a4176adad379a76),
774
	C64e(0x4dad3bdbdb964d3b), C64e(0xfac85664649efa56),
775
	C64e(0xd2e84e7474a6d24e), C64e(0x22281e141436221e),
776
	C64e(0x763fdb9292e476db), C64e(0x1e180a0c0c121e0a),
777
	C64e(0xb4906c4848fcb46c), C64e(0x376be4b8b88f37e4),
778
	C64e(0xe7255d9f9f78e75d), C64e(0xb2616ebdbd0fb26e),
779
	C64e(0x2a86ef4343692aef), C64e(0xf193a6c4c435f1a6),
780
	C64e(0xe372a83939dae3a8), C64e(0xf762a43131c6f7a4),
781
	C64e(0x59bd37d3d38a5937), C64e(0x86ff8bf2f274868b),
782
	C64e(0x56b132d5d5835632), C64e(0xc50d438b8b4ec543),
783
	C64e(0xebdc596e6e85eb59), C64e(0xc2afb7dada18c2b7),
784
	C64e(0x8f028c01018e8f8c), C64e(0xac7964b1b11dac64),
785
	C64e(0x6d23d29c9cf16dd2), C64e(0x3b92e04949723be0),
786
	C64e(0xc7abb4d8d81fc7b4), C64e(0x1543faacacb915fa),
787
	C64e(0x09fd07f3f3fa0907), C64e(0x6f8525cfcfa06f25),
788
	C64e(0xea8fafcaca20eaaf), C64e(0x89f38ef4f47d898e),
789
	C64e(0x208ee947476720e9), C64e(0x2820181010382818),
790
	C64e(0x64ded56f6f0b64d5), C64e(0x83fb88f0f0738388),
791
	C64e(0xb1946f4a4afbb16f), C64e(0x96b8725c5cca9672),
792
	C64e(0x6c70243838546c24), C64e(0x08aef157575f08f1),
793
	C64e(0x52e6c773732152c7), C64e(0xf33551979764f351),
794
	C64e(0x658d23cbcbae6523), C64e(0x84597ca1a125847c),
795
	C64e(0xbfcb9ce8e857bf9c), C64e(0x637c213e3e5d6321),
796
	C64e(0x7c37dd9696ea7cdd), C64e(0x7fc2dc61611e7fdc),
797
	C64e(0x911a860d0d9c9186), C64e(0x941e850f0f9b9485),
798
	C64e(0xabdb90e0e04bab90), C64e(0xc6f8427c7cbac642),
799
	C64e(0x57e2c471712657c4), C64e(0xe583aacccc29e5aa),
800
	C64e(0x733bd89090e373d8), C64e(0x0f0c050606090f05),
801
	C64e(0x03f501f7f7f40301), C64e(0x3638121c1c2a3612),
802
	C64e(0xfe9fa3c2c23cfea3), C64e(0xe1d45f6a6a8be15f),
803
	C64e(0x1047f9aeaebe10f9), C64e(0x6bd2d06969026bd0),
804
	C64e(0xa82e911717bfa891), C64e(0xe82958999971e858),
805
	C64e(0x6974273a3a536927), C64e(0xd04eb92727f7d0b9),
806
	C64e(0x48a938d9d9914838), C64e(0x35cd13ebebde3513),
807
	C64e(0xce56b32b2be5ceb3), C64e(0x5544332222775533),
808
	C64e(0xd6bfbbd2d204d6bb), C64e(0x904970a9a9399070),
809
	C64e(0x800e890707878089), C64e(0xf266a73333c1f2a7),
810
	C64e(0xc15ab62d2decc1b6), C64e(0x6678223c3c5a6622),
811
	C64e(0xad2a921515b8ad92), C64e(0x608920c9c9a96020),
812
	C64e(0xdb154987875cdb49), C64e(0x1a4fffaaaab01aff),
813
	C64e(0x88a0785050d88878), C64e(0x8e517aa5a52b8e7a),
814
	C64e(0x8a068f0303898a8f), C64e(0x13b2f859594a13f8),
815
	C64e(0x9b12800909929b80), C64e(0x3934171a1a233917),
816
	C64e(0x75cada65651075da), C64e(0x53b531d7d7845331),
817
	C64e(0x5113c68484d551c6), C64e(0xd3bbb8d0d003d3b8),
818
	C64e(0x5e1fc38282dc5ec3), C64e(0xcb52b02929e2cbb0),
819
	C64e(0x99b4775a5ac39977), C64e(0x333c111e1e2d3311),
820
	C64e(0x46f6cb7b7b3d46cb), C64e(0x1f4bfca8a8b71ffc),
821
	C64e(0x61dad66d6d0c61d6), C64e(0x4e583a2c2c624e3a)
822
};
823

824
#if !SPH_SMALL_FOOTPRINT_GROESTL
825

826
static const sph_u64 T5[] = {
827
	C64e(0xa5f497a5c6c632f4), C64e(0x8497eb84f8f86f97),
828
	C64e(0x99b0c799eeee5eb0), C64e(0x8d8cf78df6f67a8c),
829
	C64e(0x0d17e50dffffe817), C64e(0xbddcb7bdd6d60adc),
830
	C64e(0xb1c8a7b1dede16c8), C64e(0x54fc395491916dfc),
831
	C64e(0x50f0c050606090f0), C64e(0x0305040302020705),
832
	C64e(0xa9e087a9cece2ee0), C64e(0x7d87ac7d5656d187),
833
	C64e(0x192bd519e7e7cc2b), C64e(0x62a67162b5b513a6),
834
	C64e(0xe6319ae64d4d7c31), C64e(0x9ab5c39aecec59b5),
835
	C64e(0x45cf05458f8f40cf), C64e(0x9dbc3e9d1f1fa3bc),
836
	C64e(0x40c00940898949c0), C64e(0x8792ef87fafa6892),
837
	C64e(0x153fc515efefd03f), C64e(0xeb267febb2b29426),
838
	C64e(0xc94007c98e8ece40), C64e(0x0b1ded0bfbfbe61d),
839
	C64e(0xec2f82ec41416e2f), C64e(0x67a97d67b3b31aa9),
840
	C64e(0xfd1cbefd5f5f431c), C64e(0xea258aea45456025),
841
	C64e(0xbfda46bf2323f9da), C64e(0xf702a6f753535102),
842
	C64e(0x96a1d396e4e445a1), C64e(0x5bed2d5b9b9b76ed),
843
	C64e(0xc25deac27575285d), C64e(0x1c24d91ce1e1c524),
844
	C64e(0xaee97aae3d3dd4e9), C64e(0x6abe986a4c4cf2be),
845
	C64e(0x5aeed85a6c6c82ee), C64e(0x41c3fc417e7ebdc3),
846
	C64e(0x0206f102f5f5f306), C64e(0x4fd11d4f838352d1),
847
	C64e(0x5ce4d05c68688ce4), C64e(0xf407a2f451515607),
848
	C64e(0x345cb934d1d18d5c), C64e(0x0818e908f9f9e118),
849
	C64e(0x93aedf93e2e24cae), C64e(0x73954d73abab3e95),
850
	C64e(0x53f5c453626297f5), C64e(0x3f41543f2a2a6b41),
851
	C64e(0x0c14100c08081c14), C64e(0x52f63152959563f6),
852
	C64e(0x65af8c654646e9af), C64e(0x5ee2215e9d9d7fe2),
853
	C64e(0x2878602830304878), C64e(0xa1f86ea13737cff8),
854
	C64e(0x0f11140f0a0a1b11), C64e(0xb5c45eb52f2febc4),
855
	C64e(0x091b1c090e0e151b), C64e(0x365a483624247e5a),
856
	C64e(0x9bb6369b1b1badb6), C64e(0x3d47a53ddfdf9847),
857
	C64e(0x266a8126cdcda76a), C64e(0x69bb9c694e4ef5bb),
858
	C64e(0xcd4cfecd7f7f334c), C64e(0x9fbacf9feaea50ba),
859
	C64e(0x1b2d241b12123f2d), C64e(0x9eb93a9e1d1da4b9),
860
	C64e(0x749cb0745858c49c), C64e(0x2e72682e34344672),
861
	C64e(0x2d776c2d36364177), C64e(0xb2cda3b2dcdc11cd),
862
	C64e(0xee2973eeb4b49d29), C64e(0xfb16b6fb5b5b4d16),
863
	C64e(0xf60153f6a4a4a501), C64e(0x4dd7ec4d7676a1d7),
864
	C64e(0x61a37561b7b714a3), C64e(0xce49face7d7d3449),
865
	C64e(0x7b8da47b5252df8d), C64e(0x3e42a13edddd9f42),
866
	C64e(0x7193bc715e5ecd93), C64e(0x97a226971313b1a2),
867
	C64e(0xf50457f5a6a6a204), C64e(0x68b86968b9b901b8),
868
	C64e(0x0000000000000000), C64e(0x2c74992cc1c1b574),
869
	C64e(0x60a080604040e0a0), C64e(0x1f21dd1fe3e3c221),
870
	C64e(0xc843f2c879793a43), C64e(0xed2c77edb6b69a2c),
871
	C64e(0xbed9b3bed4d40dd9), C64e(0x46ca01468d8d47ca),
872
	C64e(0xd970ced967671770), C64e(0x4bdde44b7272afdd),
873
	C64e(0xde7933de9494ed79), C64e(0xd4672bd49898ff67),
874
	C64e(0xe8237be8b0b09323), C64e(0x4ade114a85855bde),
875
	C64e(0x6bbd6d6bbbbb06bd), C64e(0x2a7e912ac5c5bb7e),
876
	C64e(0xe5349ee54f4f7b34), C64e(0x163ac116ededd73a),
877
	C64e(0xc55417c58686d254), C64e(0xd7622fd79a9af862),
878
	C64e(0x55ffcc55666699ff), C64e(0x94a722941111b6a7),
879
	C64e(0xcf4a0fcf8a8ac04a), C64e(0x1030c910e9e9d930),
880
	C64e(0x060a080604040e0a), C64e(0x8198e781fefe6698),
881
	C64e(0xf00b5bf0a0a0ab0b), C64e(0x44ccf0447878b4cc),
882
	C64e(0xbad54aba2525f0d5), C64e(0xe33e96e34b4b753e),
883
	C64e(0xf30e5ff3a2a2ac0e), C64e(0xfe19bafe5d5d4419),
884
	C64e(0xc05b1bc08080db5b), C64e(0x8a850a8a05058085),
885
	C64e(0xadec7ead3f3fd3ec), C64e(0xbcdf42bc2121fedf),
886
	C64e(0x48d8e0487070a8d8), C64e(0x040cf904f1f1fd0c),
887
	C64e(0xdf7ac6df6363197a), C64e(0xc158eec177772f58),
888
	C64e(0x759f4575afaf309f), C64e(0x63a584634242e7a5),
889
	C64e(0x3050403020207050), C64e(0x1a2ed11ae5e5cb2e),
890
	C64e(0x0e12e10efdfdef12), C64e(0x6db7656dbfbf08b7),
891
	C64e(0x4cd4194c818155d4), C64e(0x143c30141818243c),
892
	C64e(0x355f4c352626795f), C64e(0x2f719d2fc3c3b271),
893
	C64e(0xe13867e1bebe8638), C64e(0xa2fd6aa23535c8fd),
894
	C64e(0xcc4f0bcc8888c74f), C64e(0x394b5c392e2e654b),
895
	C64e(0x57f93d5793936af9), C64e(0xf20daaf25555580d),
896
	C64e(0x829de382fcfc619d), C64e(0x47c9f4477a7ab3c9),
897
	C64e(0xacef8bacc8c827ef), C64e(0xe7326fe7baba8832),
898
	C64e(0x2b7d642b32324f7d), C64e(0x95a4d795e6e642a4),
899
	C64e(0xa0fb9ba0c0c03bfb), C64e(0x98b332981919aab3),
900
	C64e(0xd16827d19e9ef668), C64e(0x7f815d7fa3a32281),
901
	C64e(0x66aa88664444eeaa), C64e(0x7e82a87e5454d682),
902
	C64e(0xabe676ab3b3bdde6), C64e(0x839e16830b0b959e),
903
	C64e(0xca4503ca8c8cc945), C64e(0x297b9529c7c7bc7b),
904
	C64e(0xd36ed6d36b6b056e), C64e(0x3c44503c28286c44),
905
	C64e(0x798b5579a7a72c8b), C64e(0xe23d63e2bcbc813d),
906
	C64e(0x1d272c1d16163127), C64e(0x769a4176adad379a),
907
	C64e(0x3b4dad3bdbdb964d), C64e(0x56fac85664649efa),
908
	C64e(0x4ed2e84e7474a6d2), C64e(0x1e22281e14143622),
909
	C64e(0xdb763fdb9292e476), C64e(0x0a1e180a0c0c121e),
910
	C64e(0x6cb4906c4848fcb4), C64e(0xe4376be4b8b88f37),
911
	C64e(0x5de7255d9f9f78e7), C64e(0x6eb2616ebdbd0fb2),
912
	C64e(0xef2a86ef4343692a), C64e(0xa6f193a6c4c435f1),
913
	C64e(0xa8e372a83939dae3), C64e(0xa4f762a43131c6f7),
914
	C64e(0x3759bd37d3d38a59), C64e(0x8b86ff8bf2f27486),
915
	C64e(0x3256b132d5d58356), C64e(0x43c50d438b8b4ec5),
916
	C64e(0x59ebdc596e6e85eb), C64e(0xb7c2afb7dada18c2),
917
	C64e(0x8c8f028c01018e8f), C64e(0x64ac7964b1b11dac),
918
	C64e(0xd26d23d29c9cf16d), C64e(0xe03b92e04949723b),
919
	C64e(0xb4c7abb4d8d81fc7), C64e(0xfa1543faacacb915),
920
	C64e(0x0709fd07f3f3fa09), C64e(0x256f8525cfcfa06f),
921
	C64e(0xafea8fafcaca20ea), C64e(0x8e89f38ef4f47d89),
922
	C64e(0xe9208ee947476720), C64e(0x1828201810103828),
923
	C64e(0xd564ded56f6f0b64), C64e(0x8883fb88f0f07383),
924
	C64e(0x6fb1946f4a4afbb1), C64e(0x7296b8725c5cca96),
925
	C64e(0x246c70243838546c), C64e(0xf108aef157575f08),
926
	C64e(0xc752e6c773732152), C64e(0x51f33551979764f3),
927
	C64e(0x23658d23cbcbae65), C64e(0x7c84597ca1a12584),
928
	C64e(0x9cbfcb9ce8e857bf), C64e(0x21637c213e3e5d63),
929
	C64e(0xdd7c37dd9696ea7c), C64e(0xdc7fc2dc61611e7f),
930
	C64e(0x86911a860d0d9c91), C64e(0x85941e850f0f9b94),
931
	C64e(0x90abdb90e0e04bab), C64e(0x42c6f8427c7cbac6),
932
	C64e(0xc457e2c471712657), C64e(0xaae583aacccc29e5),
933
	C64e(0xd8733bd89090e373), C64e(0x050f0c050606090f),
934
	C64e(0x0103f501f7f7f403), C64e(0x123638121c1c2a36),
935
	C64e(0xa3fe9fa3c2c23cfe), C64e(0x5fe1d45f6a6a8be1),
936
	C64e(0xf91047f9aeaebe10), C64e(0xd06bd2d06969026b),
937
	C64e(0x91a82e911717bfa8), C64e(0x58e82958999971e8),
938
	C64e(0x276974273a3a5369), C64e(0xb9d04eb92727f7d0),
939
	C64e(0x3848a938d9d99148), C64e(0x1335cd13ebebde35),
940
	C64e(0xb3ce56b32b2be5ce), C64e(0x3355443322227755),
941
	C64e(0xbbd6bfbbd2d204d6), C64e(0x70904970a9a93990),
942
	C64e(0x89800e8907078780), C64e(0xa7f266a73333c1f2),
943
	C64e(0xb6c15ab62d2decc1), C64e(0x226678223c3c5a66),
944
	C64e(0x92ad2a921515b8ad), C64e(0x20608920c9c9a960),
945
	C64e(0x49db154987875cdb), C64e(0xff1a4fffaaaab01a),
946
	C64e(0x7888a0785050d888), C64e(0x7a8e517aa5a52b8e),
947
	C64e(0x8f8a068f0303898a), C64e(0xf813b2f859594a13),
948
	C64e(0x809b12800909929b), C64e(0x173934171a1a2339),
949
	C64e(0xda75cada65651075), C64e(0x3153b531d7d78453),
950
	C64e(0xc65113c68484d551), C64e(0xb8d3bbb8d0d003d3),
951
	C64e(0xc35e1fc38282dc5e), C64e(0xb0cb52b02929e2cb),
952
	C64e(0x7799b4775a5ac399), C64e(0x11333c111e1e2d33),
953
	C64e(0xcb46f6cb7b7b3d46), C64e(0xfc1f4bfca8a8b71f),
954
	C64e(0xd661dad66d6d0c61), C64e(0x3a4e583a2c2c624e)
955
};
956

957
static const sph_u64 T6[] = {
958
	C64e(0xf4a5f497a5c6c632), C64e(0x978497eb84f8f86f),
959
	C64e(0xb099b0c799eeee5e), C64e(0x8c8d8cf78df6f67a),
960
	C64e(0x170d17e50dffffe8), C64e(0xdcbddcb7bdd6d60a),
961
	C64e(0xc8b1c8a7b1dede16), C64e(0xfc54fc395491916d),
962
	C64e(0xf050f0c050606090), C64e(0x0503050403020207),
963
	C64e(0xe0a9e087a9cece2e), C64e(0x877d87ac7d5656d1),
964
	C64e(0x2b192bd519e7e7cc), C64e(0xa662a67162b5b513),
965
	C64e(0x31e6319ae64d4d7c), C64e(0xb59ab5c39aecec59),
966
	C64e(0xcf45cf05458f8f40), C64e(0xbc9dbc3e9d1f1fa3),
967
	C64e(0xc040c00940898949), C64e(0x928792ef87fafa68),
968
	C64e(0x3f153fc515efefd0), C64e(0x26eb267febb2b294),
969
	C64e(0x40c94007c98e8ece), C64e(0x1d0b1ded0bfbfbe6),
970
	C64e(0x2fec2f82ec41416e), C64e(0xa967a97d67b3b31a),
971
	C64e(0x1cfd1cbefd5f5f43), C64e(0x25ea258aea454560),
972
	C64e(0xdabfda46bf2323f9), C64e(0x02f702a6f7535351),
973
	C64e(0xa196a1d396e4e445), C64e(0xed5bed2d5b9b9b76),
974
	C64e(0x5dc25deac2757528), C64e(0x241c24d91ce1e1c5),
975
	C64e(0xe9aee97aae3d3dd4), C64e(0xbe6abe986a4c4cf2),
976
	C64e(0xee5aeed85a6c6c82), C64e(0xc341c3fc417e7ebd),
977
	C64e(0x060206f102f5f5f3), C64e(0xd14fd11d4f838352),
978
	C64e(0xe45ce4d05c68688c), C64e(0x07f407a2f4515156),
979
	C64e(0x5c345cb934d1d18d), C64e(0x180818e908f9f9e1),
980
	C64e(0xae93aedf93e2e24c), C64e(0x9573954d73abab3e),
981
	C64e(0xf553f5c453626297), C64e(0x413f41543f2a2a6b),
982
	C64e(0x140c14100c08081c), C64e(0xf652f63152959563),
983
	C64e(0xaf65af8c654646e9), C64e(0xe25ee2215e9d9d7f),
984
	C64e(0x7828786028303048), C64e(0xf8a1f86ea13737cf),
985
	C64e(0x110f11140f0a0a1b), C64e(0xc4b5c45eb52f2feb),
986
	C64e(0x1b091b1c090e0e15), C64e(0x5a365a483624247e),
987
	C64e(0xb69bb6369b1b1bad), C64e(0x473d47a53ddfdf98),
988
	C64e(0x6a266a8126cdcda7), C64e(0xbb69bb9c694e4ef5),
989
	C64e(0x4ccd4cfecd7f7f33), C64e(0xba9fbacf9feaea50),
990
	C64e(0x2d1b2d241b12123f), C64e(0xb99eb93a9e1d1da4),
991
	C64e(0x9c749cb0745858c4), C64e(0x722e72682e343446),
992
	C64e(0x772d776c2d363641), C64e(0xcdb2cda3b2dcdc11),
993
	C64e(0x29ee2973eeb4b49d), C64e(0x16fb16b6fb5b5b4d),
994
	C64e(0x01f60153f6a4a4a5), C64e(0xd74dd7ec4d7676a1),
995
	C64e(0xa361a37561b7b714), C64e(0x49ce49face7d7d34),
996
	C64e(0x8d7b8da47b5252df), C64e(0x423e42a13edddd9f),
997
	C64e(0x937193bc715e5ecd), C64e(0xa297a226971313b1),
998
	C64e(0x04f50457f5a6a6a2), C64e(0xb868b86968b9b901),
999
	C64e(0x0000000000000000), C64e(0x742c74992cc1c1b5),
1000
	C64e(0xa060a080604040e0), C64e(0x211f21dd1fe3e3c2),
1001
	C64e(0x43c843f2c879793a), C64e(0x2ced2c77edb6b69a),
1002
	C64e(0xd9bed9b3bed4d40d), C64e(0xca46ca01468d8d47),
1003
	C64e(0x70d970ced9676717), C64e(0xdd4bdde44b7272af),
1004
	C64e(0x79de7933de9494ed), C64e(0x67d4672bd49898ff),
1005
	C64e(0x23e8237be8b0b093), C64e(0xde4ade114a85855b),
1006
	C64e(0xbd6bbd6d6bbbbb06), C64e(0x7e2a7e912ac5c5bb),
1007
	C64e(0x34e5349ee54f4f7b), C64e(0x3a163ac116ededd7),
1008
	C64e(0x54c55417c58686d2), C64e(0x62d7622fd79a9af8),
1009
	C64e(0xff55ffcc55666699), C64e(0xa794a722941111b6),
1010
	C64e(0x4acf4a0fcf8a8ac0), C64e(0x301030c910e9e9d9),
1011
	C64e(0x0a060a080604040e), C64e(0x988198e781fefe66),
1012
	C64e(0x0bf00b5bf0a0a0ab), C64e(0xcc44ccf0447878b4),
1013
	C64e(0xd5bad54aba2525f0), C64e(0x3ee33e96e34b4b75),
1014
	C64e(0x0ef30e5ff3a2a2ac), C64e(0x19fe19bafe5d5d44),
1015
	C64e(0x5bc05b1bc08080db), C64e(0x858a850a8a050580),
1016
	C64e(0xecadec7ead3f3fd3), C64e(0xdfbcdf42bc2121fe),
1017
	C64e(0xd848d8e0487070a8), C64e(0x0c040cf904f1f1fd),
1018
	C64e(0x7adf7ac6df636319), C64e(0x58c158eec177772f),
1019
	C64e(0x9f759f4575afaf30), C64e(0xa563a584634242e7),
1020
	C64e(0x5030504030202070), C64e(0x2e1a2ed11ae5e5cb),
1021
	C64e(0x120e12e10efdfdef), C64e(0xb76db7656dbfbf08),
1022
	C64e(0xd44cd4194c818155), C64e(0x3c143c3014181824),
1023
	C64e(0x5f355f4c35262679), C64e(0x712f719d2fc3c3b2),
1024
	C64e(0x38e13867e1bebe86), C64e(0xfda2fd6aa23535c8),
1025
	C64e(0x4fcc4f0bcc8888c7), C64e(0x4b394b5c392e2e65),
1026
	C64e(0xf957f93d5793936a), C64e(0x0df20daaf2555558),
1027
	C64e(0x9d829de382fcfc61), C64e(0xc947c9f4477a7ab3),
1028
	C64e(0xefacef8bacc8c827), C64e(0x32e7326fe7baba88),
1029
	C64e(0x7d2b7d642b32324f), C64e(0xa495a4d795e6e642),
1030
	C64e(0xfba0fb9ba0c0c03b), C64e(0xb398b332981919aa),
1031
	C64e(0x68d16827d19e9ef6), C64e(0x817f815d7fa3a322),
1032
	C64e(0xaa66aa88664444ee), C64e(0x827e82a87e5454d6),
1033
	C64e(0xe6abe676ab3b3bdd), C64e(0x9e839e16830b0b95),
1034
	C64e(0x45ca4503ca8c8cc9), C64e(0x7b297b9529c7c7bc),
1035
	C64e(0x6ed36ed6d36b6b05), C64e(0x443c44503c28286c),
1036
	C64e(0x8b798b5579a7a72c), C64e(0x3de23d63e2bcbc81),
1037
	C64e(0x271d272c1d161631), C64e(0x9a769a4176adad37),
1038
	C64e(0x4d3b4dad3bdbdb96), C64e(0xfa56fac85664649e),
1039
	C64e(0xd24ed2e84e7474a6), C64e(0x221e22281e141436),
1040
	C64e(0x76db763fdb9292e4), C64e(0x1e0a1e180a0c0c12),
1041
	C64e(0xb46cb4906c4848fc), C64e(0x37e4376be4b8b88f),
1042
	C64e(0xe75de7255d9f9f78), C64e(0xb26eb2616ebdbd0f),
1043
	C64e(0x2aef2a86ef434369), C64e(0xf1a6f193a6c4c435),
1044
	C64e(0xe3a8e372a83939da), C64e(0xf7a4f762a43131c6),
1045
	C64e(0x593759bd37d3d38a), C64e(0x868b86ff8bf2f274),
1046
	C64e(0x563256b132d5d583), C64e(0xc543c50d438b8b4e),
1047
	C64e(0xeb59ebdc596e6e85), C64e(0xc2b7c2afb7dada18),
1048
	C64e(0x8f8c8f028c01018e), C64e(0xac64ac7964b1b11d),
1049
	C64e(0x6dd26d23d29c9cf1), C64e(0x3be03b92e0494972),
1050
	C64e(0xc7b4c7abb4d8d81f), C64e(0x15fa1543faacacb9),
1051
	C64e(0x090709fd07f3f3fa), C64e(0x6f256f8525cfcfa0),
1052
	C64e(0xeaafea8fafcaca20), C64e(0x898e89f38ef4f47d),
1053
	C64e(0x20e9208ee9474767), C64e(0x2818282018101038),
1054
	C64e(0x64d564ded56f6f0b), C64e(0x838883fb88f0f073),
1055
	C64e(0xb16fb1946f4a4afb), C64e(0x967296b8725c5cca),
1056
	C64e(0x6c246c7024383854), C64e(0x08f108aef157575f),
1057
	C64e(0x52c752e6c7737321), C64e(0xf351f33551979764),
1058
	C64e(0x6523658d23cbcbae), C64e(0x847c84597ca1a125),
1059
	C64e(0xbf9cbfcb9ce8e857), C64e(0x6321637c213e3e5d),
1060
	C64e(0x7cdd7c37dd9696ea), C64e(0x7fdc7fc2dc61611e),
1061
	C64e(0x9186911a860d0d9c), C64e(0x9485941e850f0f9b),
1062
	C64e(0xab90abdb90e0e04b), C64e(0xc642c6f8427c7cba),
1063
	C64e(0x57c457e2c4717126), C64e(0xe5aae583aacccc29),
1064
	C64e(0x73d8733bd89090e3), C64e(0x0f050f0c05060609),
1065
	C64e(0x030103f501f7f7f4), C64e(0x36123638121c1c2a),
1066
	C64e(0xfea3fe9fa3c2c23c), C64e(0xe15fe1d45f6a6a8b),
1067
	C64e(0x10f91047f9aeaebe), C64e(0x6bd06bd2d0696902),
1068
	C64e(0xa891a82e911717bf), C64e(0xe858e82958999971),
1069
	C64e(0x69276974273a3a53), C64e(0xd0b9d04eb92727f7),
1070
	C64e(0x483848a938d9d991), C64e(0x351335cd13ebebde),
1071
	C64e(0xceb3ce56b32b2be5), C64e(0x5533554433222277),
1072
	C64e(0xd6bbd6bfbbd2d204), C64e(0x9070904970a9a939),
1073
	C64e(0x8089800e89070787), C64e(0xf2a7f266a73333c1),
1074
	C64e(0xc1b6c15ab62d2dec), C64e(0x66226678223c3c5a),
1075
	C64e(0xad92ad2a921515b8), C64e(0x6020608920c9c9a9),
1076
	C64e(0xdb49db154987875c), C64e(0x1aff1a4fffaaaab0),
1077
	C64e(0x887888a0785050d8), C64e(0x8e7a8e517aa5a52b),
1078
	C64e(0x8a8f8a068f030389), C64e(0x13f813b2f859594a),
1079
	C64e(0x9b809b1280090992), C64e(0x39173934171a1a23),
1080
	C64e(0x75da75cada656510), C64e(0x533153b531d7d784),
1081
	C64e(0x51c65113c68484d5), C64e(0xd3b8d3bbb8d0d003),
1082
	C64e(0x5ec35e1fc38282dc), C64e(0xcbb0cb52b02929e2),
1083
	C64e(0x997799b4775a5ac3), C64e(0x3311333c111e1e2d),
1084
	C64e(0x46cb46f6cb7b7b3d), C64e(0x1ffc1f4bfca8a8b7),
1085
	C64e(0x61d661dad66d6d0c), C64e(0x4e3a4e583a2c2c62)
1086
};
1087

1088
static const sph_u64 T7[] = {
1089
	C64e(0x32f4a5f497a5c6c6), C64e(0x6f978497eb84f8f8),
1090
	C64e(0x5eb099b0c799eeee), C64e(0x7a8c8d8cf78df6f6),
1091
	C64e(0xe8170d17e50dffff), C64e(0x0adcbddcb7bdd6d6),
1092
	C64e(0x16c8b1c8a7b1dede), C64e(0x6dfc54fc39549191),
1093
	C64e(0x90f050f0c0506060), C64e(0x0705030504030202),
1094
	C64e(0x2ee0a9e087a9cece), C64e(0xd1877d87ac7d5656),
1095
	C64e(0xcc2b192bd519e7e7), C64e(0x13a662a67162b5b5),
1096
	C64e(0x7c31e6319ae64d4d), C64e(0x59b59ab5c39aecec),
1097
	C64e(0x40cf45cf05458f8f), C64e(0xa3bc9dbc3e9d1f1f),
1098
	C64e(0x49c040c009408989), C64e(0x68928792ef87fafa),
1099
	C64e(0xd03f153fc515efef), C64e(0x9426eb267febb2b2),
1100
	C64e(0xce40c94007c98e8e), C64e(0xe61d0b1ded0bfbfb),
1101
	C64e(0x6e2fec2f82ec4141), C64e(0x1aa967a97d67b3b3),
1102
	C64e(0x431cfd1cbefd5f5f), C64e(0x6025ea258aea4545),
1103
	C64e(0xf9dabfda46bf2323), C64e(0x5102f702a6f75353),
1104
	C64e(0x45a196a1d396e4e4), C64e(0x76ed5bed2d5b9b9b),
1105
	C64e(0x285dc25deac27575), C64e(0xc5241c24d91ce1e1),
1106
	C64e(0xd4e9aee97aae3d3d), C64e(0xf2be6abe986a4c4c),
1107
	C64e(0x82ee5aeed85a6c6c), C64e(0xbdc341c3fc417e7e),
1108
	C64e(0xf3060206f102f5f5), C64e(0x52d14fd11d4f8383),
1109
	C64e(0x8ce45ce4d05c6868), C64e(0x5607f407a2f45151),
1110
	C64e(0x8d5c345cb934d1d1), C64e(0xe1180818e908f9f9),
1111
	C64e(0x4cae93aedf93e2e2), C64e(0x3e9573954d73abab),
1112
	C64e(0x97f553f5c4536262), C64e(0x6b413f41543f2a2a),
1113
	C64e(0x1c140c14100c0808), C64e(0x63f652f631529595),
1114
	C64e(0xe9af65af8c654646), C64e(0x7fe25ee2215e9d9d),
1115
	C64e(0x4878287860283030), C64e(0xcff8a1f86ea13737),
1116
	C64e(0x1b110f11140f0a0a), C64e(0xebc4b5c45eb52f2f),
1117
	C64e(0x151b091b1c090e0e), C64e(0x7e5a365a48362424),
1118
	C64e(0xadb69bb6369b1b1b), C64e(0x98473d47a53ddfdf),
1119
	C64e(0xa76a266a8126cdcd), C64e(0xf5bb69bb9c694e4e),
1120
	C64e(0x334ccd4cfecd7f7f), C64e(0x50ba9fbacf9feaea),
1121
	C64e(0x3f2d1b2d241b1212), C64e(0xa4b99eb93a9e1d1d),
1122
	C64e(0xc49c749cb0745858), C64e(0x46722e72682e3434),
1123
	C64e(0x41772d776c2d3636), C64e(0x11cdb2cda3b2dcdc),
1124
	C64e(0x9d29ee2973eeb4b4), C64e(0x4d16fb16b6fb5b5b),
1125
	C64e(0xa501f60153f6a4a4), C64e(0xa1d74dd7ec4d7676),
1126
	C64e(0x14a361a37561b7b7), C64e(0x3449ce49face7d7d),
1127
	C64e(0xdf8d7b8da47b5252), C64e(0x9f423e42a13edddd),
1128
	C64e(0xcd937193bc715e5e), C64e(0xb1a297a226971313),
1129
	C64e(0xa204f50457f5a6a6), C64e(0x01b868b86968b9b9),
1130
	C64e(0x0000000000000000), C64e(0xb5742c74992cc1c1),
1131
	C64e(0xe0a060a080604040), C64e(0xc2211f21dd1fe3e3),
1132
	C64e(0x3a43c843f2c87979), C64e(0x9a2ced2c77edb6b6),
1133
	C64e(0x0dd9bed9b3bed4d4), C64e(0x47ca46ca01468d8d),
1134
	C64e(0x1770d970ced96767), C64e(0xafdd4bdde44b7272),
1135
	C64e(0xed79de7933de9494), C64e(0xff67d4672bd49898),
1136
	C64e(0x9323e8237be8b0b0), C64e(0x5bde4ade114a8585),
1137
	C64e(0x06bd6bbd6d6bbbbb), C64e(0xbb7e2a7e912ac5c5),
1138
	C64e(0x7b34e5349ee54f4f), C64e(0xd73a163ac116eded),
1139
	C64e(0xd254c55417c58686), C64e(0xf862d7622fd79a9a),
1140
	C64e(0x99ff55ffcc556666), C64e(0xb6a794a722941111),
1141
	C64e(0xc04acf4a0fcf8a8a), C64e(0xd9301030c910e9e9),
1142
	C64e(0x0e0a060a08060404), C64e(0x66988198e781fefe),
1143
	C64e(0xab0bf00b5bf0a0a0), C64e(0xb4cc44ccf0447878),
1144
	C64e(0xf0d5bad54aba2525), C64e(0x753ee33e96e34b4b),
1145
	C64e(0xac0ef30e5ff3a2a2), C64e(0x4419fe19bafe5d5d),
1146
	C64e(0xdb5bc05b1bc08080), C64e(0x80858a850a8a0505),
1147
	C64e(0xd3ecadec7ead3f3f), C64e(0xfedfbcdf42bc2121),
1148
	C64e(0xa8d848d8e0487070), C64e(0xfd0c040cf904f1f1),
1149
	C64e(0x197adf7ac6df6363), C64e(0x2f58c158eec17777),
1150
	C64e(0x309f759f4575afaf), C64e(0xe7a563a584634242),
1151
	C64e(0x7050305040302020), C64e(0xcb2e1a2ed11ae5e5),
1152
	C64e(0xef120e12e10efdfd), C64e(0x08b76db7656dbfbf),
1153
	C64e(0x55d44cd4194c8181), C64e(0x243c143c30141818),
1154
	C64e(0x795f355f4c352626), C64e(0xb2712f719d2fc3c3),
1155
	C64e(0x8638e13867e1bebe), C64e(0xc8fda2fd6aa23535),
1156
	C64e(0xc74fcc4f0bcc8888), C64e(0x654b394b5c392e2e),
1157
	C64e(0x6af957f93d579393), C64e(0x580df20daaf25555),
1158
	C64e(0x619d829de382fcfc), C64e(0xb3c947c9f4477a7a),
1159
	C64e(0x27efacef8bacc8c8), C64e(0x8832e7326fe7baba),
1160
	C64e(0x4f7d2b7d642b3232), C64e(0x42a495a4d795e6e6),
1161
	C64e(0x3bfba0fb9ba0c0c0), C64e(0xaab398b332981919),
1162
	C64e(0xf668d16827d19e9e), C64e(0x22817f815d7fa3a3),
1163
	C64e(0xeeaa66aa88664444), C64e(0xd6827e82a87e5454),
1164
	C64e(0xdde6abe676ab3b3b), C64e(0x959e839e16830b0b),
1165
	C64e(0xc945ca4503ca8c8c), C64e(0xbc7b297b9529c7c7),
1166
	C64e(0x056ed36ed6d36b6b), C64e(0x6c443c44503c2828),
1167
	C64e(0x2c8b798b5579a7a7), C64e(0x813de23d63e2bcbc),
1168
	C64e(0x31271d272c1d1616), C64e(0x379a769a4176adad),
1169
	C64e(0x964d3b4dad3bdbdb), C64e(0x9efa56fac8566464),
1170
	C64e(0xa6d24ed2e84e7474), C64e(0x36221e22281e1414),
1171
	C64e(0xe476db763fdb9292), C64e(0x121e0a1e180a0c0c),
1172
	C64e(0xfcb46cb4906c4848), C64e(0x8f37e4376be4b8b8),
1173
	C64e(0x78e75de7255d9f9f), C64e(0x0fb26eb2616ebdbd),
1174
	C64e(0x692aef2a86ef4343), C64e(0x35f1a6f193a6c4c4),
1175
	C64e(0xdae3a8e372a83939), C64e(0xc6f7a4f762a43131),
1176
	C64e(0x8a593759bd37d3d3), C64e(0x74868b86ff8bf2f2),
1177
	C64e(0x83563256b132d5d5), C64e(0x4ec543c50d438b8b),
1178
	C64e(0x85eb59ebdc596e6e), C64e(0x18c2b7c2afb7dada),
1179
	C64e(0x8e8f8c8f028c0101), C64e(0x1dac64ac7964b1b1),
1180
	C64e(0xf16dd26d23d29c9c), C64e(0x723be03b92e04949),
1181
	C64e(0x1fc7b4c7abb4d8d8), C64e(0xb915fa1543faacac),
1182
	C64e(0xfa090709fd07f3f3), C64e(0xa06f256f8525cfcf),
1183
	C64e(0x20eaafea8fafcaca), C64e(0x7d898e89f38ef4f4),
1184
	C64e(0x6720e9208ee94747), C64e(0x3828182820181010),
1185
	C64e(0x0b64d564ded56f6f), C64e(0x73838883fb88f0f0),
1186
	C64e(0xfbb16fb1946f4a4a), C64e(0xca967296b8725c5c),
1187
	C64e(0x546c246c70243838), C64e(0x5f08f108aef15757),
1188
	C64e(0x2152c752e6c77373), C64e(0x64f351f335519797),
1189
	C64e(0xae6523658d23cbcb), C64e(0x25847c84597ca1a1),
1190
	C64e(0x57bf9cbfcb9ce8e8), C64e(0x5d6321637c213e3e),
1191
	C64e(0xea7cdd7c37dd9696), C64e(0x1e7fdc7fc2dc6161),
1192
	C64e(0x9c9186911a860d0d), C64e(0x9b9485941e850f0f),
1193
	C64e(0x4bab90abdb90e0e0), C64e(0xbac642c6f8427c7c),
1194
	C64e(0x2657c457e2c47171), C64e(0x29e5aae583aacccc),
1195
	C64e(0xe373d8733bd89090), C64e(0x090f050f0c050606),
1196
	C64e(0xf4030103f501f7f7), C64e(0x2a36123638121c1c),
1197
	C64e(0x3cfea3fe9fa3c2c2), C64e(0x8be15fe1d45f6a6a),
1198
	C64e(0xbe10f91047f9aeae), C64e(0x026bd06bd2d06969),
1199
	C64e(0xbfa891a82e911717), C64e(0x71e858e829589999),
1200
	C64e(0x5369276974273a3a), C64e(0xf7d0b9d04eb92727),
1201
	C64e(0x91483848a938d9d9), C64e(0xde351335cd13ebeb),
1202
	C64e(0xe5ceb3ce56b32b2b), C64e(0x7755335544332222),
1203
	C64e(0x04d6bbd6bfbbd2d2), C64e(0x399070904970a9a9),
1204
	C64e(0x878089800e890707), C64e(0xc1f2a7f266a73333),
1205
	C64e(0xecc1b6c15ab62d2d), C64e(0x5a66226678223c3c),
1206
	C64e(0xb8ad92ad2a921515), C64e(0xa96020608920c9c9),
1207
	C64e(0x5cdb49db15498787), C64e(0xb01aff1a4fffaaaa),
1208
	C64e(0xd8887888a0785050), C64e(0x2b8e7a8e517aa5a5),
1209
	C64e(0x898a8f8a068f0303), C64e(0x4a13f813b2f85959),
1210
	C64e(0x929b809b12800909), C64e(0x2339173934171a1a),
1211
	C64e(0x1075da75cada6565), C64e(0x84533153b531d7d7),
1212
	C64e(0xd551c65113c68484), C64e(0x03d3b8d3bbb8d0d0),
1213
	C64e(0xdc5ec35e1fc38282), C64e(0xe2cbb0cb52b02929),
1214
	C64e(0xc3997799b4775a5a), C64e(0x2d3311333c111e1e),
1215
	C64e(0x3d46cb46f6cb7b7b), C64e(0xb71ffc1f4bfca8a8),
1216
	C64e(0x0c61d661dad66d6d), C64e(0x624e3a4e583a2c2c)
1217
};
1218

1219
#endif
1220

1221
#define DECL_STATE_SMALL \
1222
	sph_u64 H[8];
1223

1224
#define READ_STATE_SMALL(sc)   do { \
1225
		memcpy(H, (sc)->state.wide, sizeof H); \
1226
	} while (0)
1227

1228
#define WRITE_STATE_SMALL(sc)   do { \
1229
		memcpy((sc)->state.wide, H, sizeof H); \
1230
	} while (0)
1231

1232
#if SPH_SMALL_FOOTPRINT_GROESTL
1233

1234
#define RSTT(d, a, b0, b1, b2, b3, b4, b5, b6, b7)   do { \
1235
		t[d] = T0[B64_0(a[b0])] \
1236
			^ R64(T0[B64_1(a[b1])],  8) \
1237
			^ R64(T0[B64_2(a[b2])], 16) \
1238
			^ R64(T0[B64_3(a[b3])], 24) \
1239
			^ T4[B64_4(a[b4])] \
1240
			^ R64(T4[B64_5(a[b5])],  8) \
1241
			^ R64(T4[B64_6(a[b6])], 16) \
1242
			^ R64(T4[B64_7(a[b7])], 24); \
1243
	} while (0)
1244

1245
#else
1246

1247
#define RSTT(d, a, b0, b1, b2, b3, b4, b5, b6, b7)   do { \
1248
		t[d] = T0[B64_0(a[b0])] \
1249
			^ T1[B64_1(a[b1])] \
1250
			^ T2[B64_2(a[b2])] \
1251
			^ T3[B64_3(a[b3])] \
1252
			^ T4[B64_4(a[b4])] \
1253
			^ T5[B64_5(a[b5])] \
1254
			^ T6[B64_6(a[b6])] \
1255
			^ T7[B64_7(a[b7])]; \
1256
	} while (0)
1257

1258
#endif
1259

1260
#define ROUND_SMALL_P(a, r)   do { \
1261
		sph_u64 t[8]; \
1262
		a[0] ^= PC64(0x00, r); \
1263
		a[1] ^= PC64(0x10, r); \
1264
		a[2] ^= PC64(0x20, r); \
1265
		a[3] ^= PC64(0x30, r); \
1266
		a[4] ^= PC64(0x40, r); \
1267
		a[5] ^= PC64(0x50, r); \
1268
		a[6] ^= PC64(0x60, r); \
1269
		a[7] ^= PC64(0x70, r); \
1270
		RSTT(0, a, 0, 1, 2, 3, 4, 5, 6, 7); \
1271
		RSTT(1, a, 1, 2, 3, 4, 5, 6, 7, 0); \
1272
		RSTT(2, a, 2, 3, 4, 5, 6, 7, 0, 1); \
1273
		RSTT(3, a, 3, 4, 5, 6, 7, 0, 1, 2); \
1274
		RSTT(4, a, 4, 5, 6, 7, 0, 1, 2, 3); \
1275
		RSTT(5, a, 5, 6, 7, 0, 1, 2, 3, 4); \
1276
		RSTT(6, a, 6, 7, 0, 1, 2, 3, 4, 5); \
1277
		RSTT(7, a, 7, 0, 1, 2, 3, 4, 5, 6); \
1278
		a[0] = t[0]; \
1279
		a[1] = t[1]; \
1280
		a[2] = t[2]; \
1281
		a[3] = t[3]; \
1282
		a[4] = t[4]; \
1283
		a[5] = t[5]; \
1284
		a[6] = t[6]; \
1285
		a[7] = t[7]; \
1286
	} while (0)
1287

1288
#define ROUND_SMALL_Q(a, r)   do { \
1289
		sph_u64 t[8]; \
1290
		a[0] ^= QC64(0x00, r); \
1291
		a[1] ^= QC64(0x10, r); \
1292
		a[2] ^= QC64(0x20, r); \
1293
		a[3] ^= QC64(0x30, r); \
1294
		a[4] ^= QC64(0x40, r); \
1295
		a[5] ^= QC64(0x50, r); \
1296
		a[6] ^= QC64(0x60, r); \
1297
		a[7] ^= QC64(0x70, r); \
1298
		RSTT(0, a, 1, 3, 5, 7, 0, 2, 4, 6); \
1299
		RSTT(1, a, 2, 4, 6, 0, 1, 3, 5, 7); \
1300
		RSTT(2, a, 3, 5, 7, 1, 2, 4, 6, 0); \
1301
		RSTT(3, a, 4, 6, 0, 2, 3, 5, 7, 1); \
1302
		RSTT(4, a, 5, 7, 1, 3, 4, 6, 0, 2); \
1303
		RSTT(5, a, 6, 0, 2, 4, 5, 7, 1, 3); \
1304
		RSTT(6, a, 7, 1, 3, 5, 6, 0, 2, 4); \
1305
		RSTT(7, a, 0, 2, 4, 6, 7, 1, 3, 5); \
1306
		a[0] = t[0]; \
1307
		a[1] = t[1]; \
1308
		a[2] = t[2]; \
1309
		a[3] = t[3]; \
1310
		a[4] = t[4]; \
1311
		a[5] = t[5]; \
1312
		a[6] = t[6]; \
1313
		a[7] = t[7]; \
1314
	} while (0)
1315

1316
#if SPH_SMALL_FOOTPRINT_GROESTL
1317

1318
#define PERM_SMALL_P(a)   do { \
1319
		int r; \
1320
		for (r = 0; r < 10; r ++) \
1321
			ROUND_SMALL_P(a, r); \
1322
	} while (0)
1323

1324
#define PERM_SMALL_Q(a)   do { \
1325
		int r; \
1326
		for (r = 0; r < 10; r ++) \
1327
			ROUND_SMALL_Q(a, r); \
1328
	} while (0)
1329

1330
#else
1331

1332
/*
1333
 * Apparently, unrolling more than that confuses GCC, resulting in
1334
 * lower performance, even though L1 cache would be no problem.
1335
 */
1336
#define PERM_SMALL_P(a)   do { \
1337
		int r; \
1338
		for (r = 0; r < 10; r += 2) { \
1339
			ROUND_SMALL_P(a, r + 0); \
1340
			ROUND_SMALL_P(a, r + 1); \
1341
		} \
1342
	} while (0)
1343

1344
#define PERM_SMALL_Q(a)   do { \
1345
		int r; \
1346
		for (r = 0; r < 10; r += 2) { \
1347
			ROUND_SMALL_Q(a, r + 0); \
1348
			ROUND_SMALL_Q(a, r + 1); \
1349
		} \
1350
	} while (0)
1351

1352
#endif
1353

1354
#define COMPRESS_SMALL   do { \
1355
		sph_u64 g[8], m[8]; \
1356
		size_t u; \
1357
		for (u = 0; u < 8; u ++) { \
1358
			m[u] = dec64e_aligned(buf + (u << 3)); \
1359
			g[u] = m[u] ^ H[u]; \
1360
		} \
1361
		PERM_SMALL_P(g); \
1362
		PERM_SMALL_Q(m); \
1363
		for (u = 0; u < 8; u ++) \
1364
			H[u] ^= g[u] ^ m[u]; \
1365
	} while (0)
1366

1367
#define FINAL_SMALL   do { \
1368
		sph_u64 x[8]; \
1369
		size_t u; \
1370
		memcpy(x, H, sizeof x); \
1371
		PERM_SMALL_P(x); \
1372
		for (u = 0; u < 8; u ++) \
1373
			H[u] ^= x[u]; \
1374
	} while (0)
1375

1376
#define DECL_STATE_BIG \
1377
	sph_u64 H[16];
1378

1379
#define READ_STATE_BIG(sc)   do { \
1380
		memcpy(H, (sc)->state.wide, sizeof H); \
1381
	} while (0)
1382

1383
#define WRITE_STATE_BIG(sc)   do { \
1384
		memcpy((sc)->state.wide, H, sizeof H); \
1385
	} while (0)
1386

1387
#if SPH_SMALL_FOOTPRINT_GROESTL
1388

1389
#define RBTT(d, a, b0, b1, b2, b3, b4, b5, b6, b7)   do { \
1390
		t[d] = T0[B64_0(a[b0])] \
1391
			^ R64(T0[B64_1(a[b1])],  8) \
1392
			^ R64(T0[B64_2(a[b2])], 16) \
1393
			^ R64(T0[B64_3(a[b3])], 24) \
1394
			^ T4[B64_4(a[b4])] \
1395
			^ R64(T4[B64_5(a[b5])],  8) \
1396
			^ R64(T4[B64_6(a[b6])], 16) \
1397
			^ R64(T4[B64_7(a[b7])], 24); \
1398
	} while (0)
1399

1400
#else
1401

1402
#define RBTT(d, a, b0, b1, b2, b3, b4, b5, b6, b7)   do { \
1403
		t[d] = T0[B64_0(a[b0])] \
1404
			^ T1[B64_1(a[b1])] \
1405
			^ T2[B64_2(a[b2])] \
1406
			^ T3[B64_3(a[b3])] \
1407
			^ T4[B64_4(a[b4])] \
1408
			^ T5[B64_5(a[b5])] \
1409
			^ T6[B64_6(a[b6])] \
1410
			^ T7[B64_7(a[b7])]; \
1411
	} while (0)
1412

1413
#endif
1414

1415
#if SPH_SMALL_FOOTPRINT_GROESTL
1416

1417
#define ROUND_BIG_P(a, r)   do { \
1418
		sph_u64 t[16]; \
1419
		size_t u; \
1420
		a[0x0] ^= PC64(0x00, r); \
1421
		a[0x1] ^= PC64(0x10, r); \
1422
		a[0x2] ^= PC64(0x20, r); \
1423
		a[0x3] ^= PC64(0x30, r); \
1424
		a[0x4] ^= PC64(0x40, r); \
1425
		a[0x5] ^= PC64(0x50, r); \
1426
		a[0x6] ^= PC64(0x60, r); \
1427
		a[0x7] ^= PC64(0x70, r); \
1428
		a[0x8] ^= PC64(0x80, r); \
1429
		a[0x9] ^= PC64(0x90, r); \
1430
		a[0xA] ^= PC64(0xA0, r); \
1431
		a[0xB] ^= PC64(0xB0, r); \
1432
		a[0xC] ^= PC64(0xC0, r); \
1433
		a[0xD] ^= PC64(0xD0, r); \
1434
		a[0xE] ^= PC64(0xE0, r); \
1435
		a[0xF] ^= PC64(0xF0, r); \
1436
		for (u = 0; u < 16; u += 4) { \
1437
			RBTT(u + 0, a, u + 0, (u + 1) & 0xF, \
1438
				(u + 2) & 0xF, (u + 3) & 0xF, (u + 4) & 0xF, \
1439
				(u + 5) & 0xF, (u + 6) & 0xF, (u + 11) & 0xF); \
1440
			RBTT(u + 1, a, u + 1, (u + 2) & 0xF, \
1441
				(u + 3) & 0xF, (u + 4) & 0xF, (u + 5) & 0xF, \
1442
				(u + 6) & 0xF, (u + 7) & 0xF, (u + 12) & 0xF); \
1443
			RBTT(u + 2, a, u + 2, (u + 3) & 0xF, \
1444
				(u + 4) & 0xF, (u + 5) & 0xF, (u + 6) & 0xF, \
1445
				(u + 7) & 0xF, (u + 8) & 0xF, (u + 13) & 0xF); \
1446
			RBTT(u + 3, a, u + 3, (u + 4) & 0xF, \
1447
				(u + 5) & 0xF, (u + 6) & 0xF, (u + 7) & 0xF, \
1448
				(u + 8) & 0xF, (u + 9) & 0xF, (u + 14) & 0xF); \
1449
		} \
1450
		memcpy(a, t, sizeof t); \
1451
	} while (0)
1452

1453
#define ROUND_BIG_Q(a, r)   do { \
1454
		sph_u64 t[16]; \
1455
		size_t u; \
1456
		a[0x0] ^= QC64(0x00, r); \
1457
		a[0x1] ^= QC64(0x10, r); \
1458
		a[0x2] ^= QC64(0x20, r); \
1459
		a[0x3] ^= QC64(0x30, r); \
1460
		a[0x4] ^= QC64(0x40, r); \
1461
		a[0x5] ^= QC64(0x50, r); \
1462
		a[0x6] ^= QC64(0x60, r); \
1463
		a[0x7] ^= QC64(0x70, r); \
1464
		a[0x8] ^= QC64(0x80, r); \
1465
		a[0x9] ^= QC64(0x90, r); \
1466
		a[0xA] ^= QC64(0xA0, r); \
1467
		a[0xB] ^= QC64(0xB0, r); \
1468
		a[0xC] ^= QC64(0xC0, r); \
1469
		a[0xD] ^= QC64(0xD0, r); \
1470
		a[0xE] ^= QC64(0xE0, r); \
1471
		a[0xF] ^= QC64(0xF0, r); \
1472
		for (u = 0; u < 16; u += 4) { \
1473
			RBTT(u + 0, a, (u + 1) & 0xF, (u + 3) & 0xF, \
1474
				(u + 5) & 0xF, (u + 11) & 0xF, (u + 0) & 0xF, \
1475
				(u + 2) & 0xF, (u + 4) & 0xF, (u + 6) & 0xF); \
1476
			RBTT(u + 1, a, (u + 2) & 0xF, (u + 4) & 0xF, \
1477
				(u + 6) & 0xF, (u + 12) & 0xF, (u + 1) & 0xF, \
1478
				(u + 3) & 0xF, (u + 5) & 0xF, (u + 7) & 0xF); \
1479
			RBTT(u + 2, a, (u + 3) & 0xF, (u + 5) & 0xF, \
1480
				(u + 7) & 0xF, (u + 13) & 0xF, (u + 2) & 0xF, \
1481
				(u + 4) & 0xF, (u + 6) & 0xF, (u + 8) & 0xF); \
1482
			RBTT(u + 3, a, (u + 4) & 0xF, (u + 6) & 0xF, \
1483
				(u + 8) & 0xF, (u + 14) & 0xF, (u + 3) & 0xF, \
1484
				(u + 5) & 0xF, (u + 7) & 0xF, (u + 9) & 0xF); \
1485
		} \
1486
		memcpy(a, t, sizeof t); \
1487
	} while (0)
1488

1489
#else
1490

1491
#define ROUND_BIG_P(a, r)   do { \
1492
		sph_u64 t[16]; \
1493
		a[0x0] ^= PC64(0x00, r); \
1494
		a[0x1] ^= PC64(0x10, r); \
1495
		a[0x2] ^= PC64(0x20, r); \
1496
		a[0x3] ^= PC64(0x30, r); \
1497
		a[0x4] ^= PC64(0x40, r); \
1498
		a[0x5] ^= PC64(0x50, r); \
1499
		a[0x6] ^= PC64(0x60, r); \
1500
		a[0x7] ^= PC64(0x70, r); \
1501
		a[0x8] ^= PC64(0x80, r); \
1502
		a[0x9] ^= PC64(0x90, r); \
1503
		a[0xA] ^= PC64(0xA0, r); \
1504
		a[0xB] ^= PC64(0xB0, r); \
1505
		a[0xC] ^= PC64(0xC0, r); \
1506
		a[0xD] ^= PC64(0xD0, r); \
1507
		a[0xE] ^= PC64(0xE0, r); \
1508
		a[0xF] ^= PC64(0xF0, r); \
1509
		RBTT(0x0, a, 0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0xB); \
1510
		RBTT(0x1, a, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0xC); \
1511
		RBTT(0x2, a, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x8, 0xD); \
1512
		RBTT(0x3, a, 0x3, 0x4, 0x5, 0x6, 0x7, 0x8, 0x9, 0xE); \
1513
		RBTT(0x4, a, 0x4, 0x5, 0x6, 0x7, 0x8, 0x9, 0xA, 0xF); \
1514
		RBTT(0x5, a, 0x5, 0x6, 0x7, 0x8, 0x9, 0xA, 0xB, 0x0); \
1515
		RBTT(0x6, a, 0x6, 0x7, 0x8, 0x9, 0xA, 0xB, 0xC, 0x1); \
1516
		RBTT(0x7, a, 0x7, 0x8, 0x9, 0xA, 0xB, 0xC, 0xD, 0x2); \
1517
		RBTT(0x8, a, 0x8, 0x9, 0xA, 0xB, 0xC, 0xD, 0xE, 0x3); \
1518
		RBTT(0x9, a, 0x9, 0xA, 0xB, 0xC, 0xD, 0xE, 0xF, 0x4); \
1519
		RBTT(0xA, a, 0xA, 0xB, 0xC, 0xD, 0xE, 0xF, 0x0, 0x5); \
1520
		RBTT(0xB, a, 0xB, 0xC, 0xD, 0xE, 0xF, 0x0, 0x1, 0x6); \
1521
		RBTT(0xC, a, 0xC, 0xD, 0xE, 0xF, 0x0, 0x1, 0x2, 0x7); \
1522
		RBTT(0xD, a, 0xD, 0xE, 0xF, 0x0, 0x1, 0x2, 0x3, 0x8); \
1523
		RBTT(0xE, a, 0xE, 0xF, 0x0, 0x1, 0x2, 0x3, 0x4, 0x9); \
1524
		RBTT(0xF, a, 0xF, 0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0xA); \
1525
		a[0x0] = t[0x0]; \
1526
		a[0x1] = t[0x1]; \
1527
		a[0x2] = t[0x2]; \
1528
		a[0x3] = t[0x3]; \
1529
		a[0x4] = t[0x4]; \
1530
		a[0x5] = t[0x5]; \
1531
		a[0x6] = t[0x6]; \
1532
		a[0x7] = t[0x7]; \
1533
		a[0x8] = t[0x8]; \
1534
		a[0x9] = t[0x9]; \
1535
		a[0xA] = t[0xA]; \
1536
		a[0xB] = t[0xB]; \
1537
		a[0xC] = t[0xC]; \
1538
		a[0xD] = t[0xD]; \
1539
		a[0xE] = t[0xE]; \
1540
		a[0xF] = t[0xF]; \
1541
	} while (0)
1542

1543
#define ROUND_BIG_Q(a, r)   do { \
1544
		sph_u64 t[16]; \
1545
		a[0x0] ^= QC64(0x00, r); \
1546
		a[0x1] ^= QC64(0x10, r); \
1547
		a[0x2] ^= QC64(0x20, r); \
1548
		a[0x3] ^= QC64(0x30, r); \
1549
		a[0x4] ^= QC64(0x40, r); \
1550
		a[0x5] ^= QC64(0x50, r); \
1551
		a[0x6] ^= QC64(0x60, r); \
1552
		a[0x7] ^= QC64(0x70, r); \
1553
		a[0x8] ^= QC64(0x80, r); \
1554
		a[0x9] ^= QC64(0x90, r); \
1555
		a[0xA] ^= QC64(0xA0, r); \
1556
		a[0xB] ^= QC64(0xB0, r); \
1557
		a[0xC] ^= QC64(0xC0, r); \
1558
		a[0xD] ^= QC64(0xD0, r); \
1559
		a[0xE] ^= QC64(0xE0, r); \
1560
		a[0xF] ^= QC64(0xF0, r); \
1561
		RBTT(0x0, a, 0x1, 0x3, 0x5, 0xB, 0x0, 0x2, 0x4, 0x6); \
1562
		RBTT(0x1, a, 0x2, 0x4, 0x6, 0xC, 0x1, 0x3, 0x5, 0x7); \
1563
		RBTT(0x2, a, 0x3, 0x5, 0x7, 0xD, 0x2, 0x4, 0x6, 0x8); \
1564
		RBTT(0x3, a, 0x4, 0x6, 0x8, 0xE, 0x3, 0x5, 0x7, 0x9); \
1565
		RBTT(0x4, a, 0x5, 0x7, 0x9, 0xF, 0x4, 0x6, 0x8, 0xA); \
1566
		RBTT(0x5, a, 0x6, 0x8, 0xA, 0x0, 0x5, 0x7, 0x9, 0xB); \
1567
		RBTT(0x6, a, 0x7, 0x9, 0xB, 0x1, 0x6, 0x8, 0xA, 0xC); \
1568
		RBTT(0x7, a, 0x8, 0xA, 0xC, 0x2, 0x7, 0x9, 0xB, 0xD); \
1569
		RBTT(0x8, a, 0x9, 0xB, 0xD, 0x3, 0x8, 0xA, 0xC, 0xE); \
1570
		RBTT(0x9, a, 0xA, 0xC, 0xE, 0x4, 0x9, 0xB, 0xD, 0xF); \
1571
		RBTT(0xA, a, 0xB, 0xD, 0xF, 0x5, 0xA, 0xC, 0xE, 0x0); \
1572
		RBTT(0xB, a, 0xC, 0xE, 0x0, 0x6, 0xB, 0xD, 0xF, 0x1); \
1573
		RBTT(0xC, a, 0xD, 0xF, 0x1, 0x7, 0xC, 0xE, 0x0, 0x2); \
1574
		RBTT(0xD, a, 0xE, 0x0, 0x2, 0x8, 0xD, 0xF, 0x1, 0x3); \
1575
		RBTT(0xE, a, 0xF, 0x1, 0x3, 0x9, 0xE, 0x0, 0x2, 0x4); \
1576
		RBTT(0xF, a, 0x0, 0x2, 0x4, 0xA, 0xF, 0x1, 0x3, 0x5); \
1577
		a[0x0] = t[0x0]; \
1578
		a[0x1] = t[0x1]; \
1579
		a[0x2] = t[0x2]; \
1580
		a[0x3] = t[0x3]; \
1581
		a[0x4] = t[0x4]; \
1582
		a[0x5] = t[0x5]; \
1583
		a[0x6] = t[0x6]; \
1584
		a[0x7] = t[0x7]; \
1585
		a[0x8] = t[0x8]; \
1586
		a[0x9] = t[0x9]; \
1587
		a[0xA] = t[0xA]; \
1588
		a[0xB] = t[0xB]; \
1589
		a[0xC] = t[0xC]; \
1590
		a[0xD] = t[0xD]; \
1591
		a[0xE] = t[0xE]; \
1592
		a[0xF] = t[0xF]; \
1593
	} while (0)
1594

1595
#endif
1596

1597
#define PERM_BIG_P(a)   do { \
1598
		int r; \
1599
		for (r = 0; r < 14; r += 2) { \
1600
			ROUND_BIG_P(a, r + 0); \
1601
			ROUND_BIG_P(a, r + 1); \
1602
		} \
1603
	} while (0)
1604

1605
#define PERM_BIG_Q(a)   do { \
1606
		int r; \
1607
		for (r = 0; r < 14; r += 2) { \
1608
			ROUND_BIG_Q(a, r + 0); \
1609
			ROUND_BIG_Q(a, r + 1); \
1610
		} \
1611
	} while (0)
1612

1613
/* obsolete
1614
#if SPH_SMALL_FOOTPRINT_GROESTL
1615

1616
#define COMPRESS_BIG   do { \
1617
		sph_u64 g[16], m[16], *ya; \
1618
		const sph_u64 *yc; \
1619
		size_t u; \
1620
		int i; \
1621
		for (u = 0; u < 16; u ++) { \
1622
			m[u] = dec64e_aligned(buf + (u << 3)); \
1623
			g[u] = m[u] ^ H[u]; \
1624
		} \
1625
		ya = g; \
1626
		yc = CP; \
1627
		for (i = 0; i < 2; i ++) { \
1628
			PERM_BIG(ya, yc); \
1629
			ya = m; \
1630
			yc = CQ; \
1631
		} \
1632
		for (u = 0; u < 16; u ++) { \
1633
			H[u] ^= g[u] ^ m[u]; \
1634
		} \
1635
	} while (0)
1636

1637
#else
1638
*/
1639

1640
#define COMPRESS_BIG   do { \
1641
		sph_u64 g[16], m[16]; \
1642
		size_t u; \
1643
		for (u = 0; u < 16; u ++) { \
1644
			m[u] = dec64e_aligned(buf + (u << 3)); \
1645
			g[u] = m[u] ^ H[u]; \
1646
		} \
1647
		PERM_BIG_P(g); \
1648
		PERM_BIG_Q(m); \
1649
		for (u = 0; u < 16; u ++) { \
1650
			H[u] ^= g[u] ^ m[u]; \
1651
		} \
1652
	} while (0)
1653

1654
/* obsolete
1655
#endif
1656
*/
1657

1658
#define FINAL_BIG   do { \
1659
		sph_u64 x[16]; \
1660
		size_t u; \
1661
		memcpy(x, H, sizeof x); \
1662
		PERM_BIG_P(x); \
1663
		for (u = 0; u < 16; u ++) \
1664
			H[u] ^= x[u]; \
1665
	} while (0)
1666

1667
#else
1668

1669
static const sph_u32 T0up[] = {
1670
	C32e(0xc632f4a5), C32e(0xf86f9784), C32e(0xee5eb099), C32e(0xf67a8c8d),
1671
	C32e(0xffe8170d), C32e(0xd60adcbd), C32e(0xde16c8b1), C32e(0x916dfc54),
1672
	C32e(0x6090f050), C32e(0x02070503), C32e(0xce2ee0a9), C32e(0x56d1877d),
1673
	C32e(0xe7cc2b19), C32e(0xb513a662), C32e(0x4d7c31e6), C32e(0xec59b59a),
1674
	C32e(0x8f40cf45), C32e(0x1fa3bc9d), C32e(0x8949c040), C32e(0xfa689287),
1675
	C32e(0xefd03f15), C32e(0xb29426eb), C32e(0x8ece40c9), C32e(0xfbe61d0b),
1676
	C32e(0x416e2fec), C32e(0xb31aa967), C32e(0x5f431cfd), C32e(0x456025ea),
1677
	C32e(0x23f9dabf), C32e(0x535102f7), C32e(0xe445a196), C32e(0x9b76ed5b),
1678
	C32e(0x75285dc2), C32e(0xe1c5241c), C32e(0x3dd4e9ae), C32e(0x4cf2be6a),
1679
	C32e(0x6c82ee5a), C32e(0x7ebdc341), C32e(0xf5f30602), C32e(0x8352d14f),
1680
	C32e(0x688ce45c), C32e(0x515607f4), C32e(0xd18d5c34), C32e(0xf9e11808),
1681
	C32e(0xe24cae93), C32e(0xab3e9573), C32e(0x6297f553), C32e(0x2a6b413f),
1682
	C32e(0x081c140c), C32e(0x9563f652), C32e(0x46e9af65), C32e(0x9d7fe25e),
1683
	C32e(0x30487828), C32e(0x37cff8a1), C32e(0x0a1b110f), C32e(0x2febc4b5),
1684
	C32e(0x0e151b09), C32e(0x247e5a36), C32e(0x1badb69b), C32e(0xdf98473d),
1685
	C32e(0xcda76a26), C32e(0x4ef5bb69), C32e(0x7f334ccd), C32e(0xea50ba9f),
1686
	C32e(0x123f2d1b), C32e(0x1da4b99e), C32e(0x58c49c74), C32e(0x3446722e),
1687
	C32e(0x3641772d), C32e(0xdc11cdb2), C32e(0xb49d29ee), C32e(0x5b4d16fb),
1688
	C32e(0xa4a501f6), C32e(0x76a1d74d), C32e(0xb714a361), C32e(0x7d3449ce),
1689
	C32e(0x52df8d7b), C32e(0xdd9f423e), C32e(0x5ecd9371), C32e(0x13b1a297),
1690
	C32e(0xa6a204f5), C32e(0xb901b868), C32e(0x00000000), C32e(0xc1b5742c),
1691
	C32e(0x40e0a060), C32e(0xe3c2211f), C32e(0x793a43c8), C32e(0xb69a2ced),
1692
	C32e(0xd40dd9be), C32e(0x8d47ca46), C32e(0x671770d9), C32e(0x72afdd4b),
1693
	C32e(0x94ed79de), C32e(0x98ff67d4), C32e(0xb09323e8), C32e(0x855bde4a),
1694
	C32e(0xbb06bd6b), C32e(0xc5bb7e2a), C32e(0x4f7b34e5), C32e(0xedd73a16),
1695
	C32e(0x86d254c5), C32e(0x9af862d7), C32e(0x6699ff55), C32e(0x11b6a794),
1696
	C32e(0x8ac04acf), C32e(0xe9d93010), C32e(0x040e0a06), C32e(0xfe669881),
1697
	C32e(0xa0ab0bf0), C32e(0x78b4cc44), C32e(0x25f0d5ba), C32e(0x4b753ee3),
1698
	C32e(0xa2ac0ef3), C32e(0x5d4419fe), C32e(0x80db5bc0), C32e(0x0580858a),
1699
	C32e(0x3fd3ecad), C32e(0x21fedfbc), C32e(0x70a8d848), C32e(0xf1fd0c04),
1700
	C32e(0x63197adf), C32e(0x772f58c1), C32e(0xaf309f75), C32e(0x42e7a563),
1701
	C32e(0x20705030), C32e(0xe5cb2e1a), C32e(0xfdef120e), C32e(0xbf08b76d),
1702
	C32e(0x8155d44c), C32e(0x18243c14), C32e(0x26795f35), C32e(0xc3b2712f),
1703
	C32e(0xbe8638e1), C32e(0x35c8fda2), C32e(0x88c74fcc), C32e(0x2e654b39),
1704
	C32e(0x936af957), C32e(0x55580df2), C32e(0xfc619d82), C32e(0x7ab3c947),
1705
	C32e(0xc827efac), C32e(0xba8832e7), C32e(0x324f7d2b), C32e(0xe642a495),
1706
	C32e(0xc03bfba0), C32e(0x19aab398), C32e(0x9ef668d1), C32e(0xa322817f),
1707
	C32e(0x44eeaa66), C32e(0x54d6827e), C32e(0x3bdde6ab), C32e(0x0b959e83),
1708
	C32e(0x8cc945ca), C32e(0xc7bc7b29), C32e(0x6b056ed3), C32e(0x286c443c),
1709
	C32e(0xa72c8b79), C32e(0xbc813de2), C32e(0x1631271d), C32e(0xad379a76),
1710
	C32e(0xdb964d3b), C32e(0x649efa56), C32e(0x74a6d24e), C32e(0x1436221e),
1711
	C32e(0x92e476db), C32e(0x0c121e0a), C32e(0x48fcb46c), C32e(0xb88f37e4),
1712
	C32e(0x9f78e75d), C32e(0xbd0fb26e), C32e(0x43692aef), C32e(0xc435f1a6),
1713
	C32e(0x39dae3a8), C32e(0x31c6f7a4), C32e(0xd38a5937), C32e(0xf274868b),
1714
	C32e(0xd5835632), C32e(0x8b4ec543), C32e(0x6e85eb59), C32e(0xda18c2b7),
1715
	C32e(0x018e8f8c), C32e(0xb11dac64), C32e(0x9cf16dd2), C32e(0x49723be0),
1716
	C32e(0xd81fc7b4), C32e(0xacb915fa), C32e(0xf3fa0907), C32e(0xcfa06f25),
1717
	C32e(0xca20eaaf), C32e(0xf47d898e), C32e(0x476720e9), C32e(0x10382818),
1718
	C32e(0x6f0b64d5), C32e(0xf0738388), C32e(0x4afbb16f), C32e(0x5cca9672),
1719
	C32e(0x38546c24), C32e(0x575f08f1), C32e(0x732152c7), C32e(0x9764f351),
1720
	C32e(0xcbae6523), C32e(0xa125847c), C32e(0xe857bf9c), C32e(0x3e5d6321),
1721
	C32e(0x96ea7cdd), C32e(0x611e7fdc), C32e(0x0d9c9186), C32e(0x0f9b9485),
1722
	C32e(0xe04bab90), C32e(0x7cbac642), C32e(0x712657c4), C32e(0xcc29e5aa),
1723
	C32e(0x90e373d8), C32e(0x06090f05), C32e(0xf7f40301), C32e(0x1c2a3612),
1724
	C32e(0xc23cfea3), C32e(0x6a8be15f), C32e(0xaebe10f9), C32e(0x69026bd0),
1725
	C32e(0x17bfa891), C32e(0x9971e858), C32e(0x3a536927), C32e(0x27f7d0b9),
1726
	C32e(0xd9914838), C32e(0xebde3513), C32e(0x2be5ceb3), C32e(0x22775533),
1727
	C32e(0xd204d6bb), C32e(0xa9399070), C32e(0x07878089), C32e(0x33c1f2a7),
1728
	C32e(0x2decc1b6), C32e(0x3c5a6622), C32e(0x15b8ad92), C32e(0xc9a96020),
1729
	C32e(0x875cdb49), C32e(0xaab01aff), C32e(0x50d88878), C32e(0xa52b8e7a),
1730
	C32e(0x03898a8f), C32e(0x594a13f8), C32e(0x09929b80), C32e(0x1a233917),
1731
	C32e(0x651075da), C32e(0xd7845331), C32e(0x84d551c6), C32e(0xd003d3b8),
1732
	C32e(0x82dc5ec3), C32e(0x29e2cbb0), C32e(0x5ac39977), C32e(0x1e2d3311),
1733
	C32e(0x7b3d46cb), C32e(0xa8b71ffc), C32e(0x6d0c61d6), C32e(0x2c624e3a)
1734
};
1735

1736
static const sph_u32 T0dn[] = {
1737
	C32e(0xf497a5c6), C32e(0x97eb84f8), C32e(0xb0c799ee), C32e(0x8cf78df6),
1738
	C32e(0x17e50dff), C32e(0xdcb7bdd6), C32e(0xc8a7b1de), C32e(0xfc395491),
1739
	C32e(0xf0c05060), C32e(0x05040302), C32e(0xe087a9ce), C32e(0x87ac7d56),
1740
	C32e(0x2bd519e7), C32e(0xa67162b5), C32e(0x319ae64d), C32e(0xb5c39aec),
1741
	C32e(0xcf05458f), C32e(0xbc3e9d1f), C32e(0xc0094089), C32e(0x92ef87fa),
1742
	C32e(0x3fc515ef), C32e(0x267febb2), C32e(0x4007c98e), C32e(0x1ded0bfb),
1743
	C32e(0x2f82ec41), C32e(0xa97d67b3), C32e(0x1cbefd5f), C32e(0x258aea45),
1744
	C32e(0xda46bf23), C32e(0x02a6f753), C32e(0xa1d396e4), C32e(0xed2d5b9b),
1745
	C32e(0x5deac275), C32e(0x24d91ce1), C32e(0xe97aae3d), C32e(0xbe986a4c),
1746
	C32e(0xeed85a6c), C32e(0xc3fc417e), C32e(0x06f102f5), C32e(0xd11d4f83),
1747
	C32e(0xe4d05c68), C32e(0x07a2f451), C32e(0x5cb934d1), C32e(0x18e908f9),
1748
	C32e(0xaedf93e2), C32e(0x954d73ab), C32e(0xf5c45362), C32e(0x41543f2a),
1749
	C32e(0x14100c08), C32e(0xf6315295), C32e(0xaf8c6546), C32e(0xe2215e9d),
1750
	C32e(0x78602830), C32e(0xf86ea137), C32e(0x11140f0a), C32e(0xc45eb52f),
1751
	C32e(0x1b1c090e), C32e(0x5a483624), C32e(0xb6369b1b), C32e(0x47a53ddf),
1752
	C32e(0x6a8126cd), C32e(0xbb9c694e), C32e(0x4cfecd7f), C32e(0xbacf9fea),
1753
	C32e(0x2d241b12), C32e(0xb93a9e1d), C32e(0x9cb07458), C32e(0x72682e34),
1754
	C32e(0x776c2d36), C32e(0xcda3b2dc), C32e(0x2973eeb4), C32e(0x16b6fb5b),
1755
	C32e(0x0153f6a4), C32e(0xd7ec4d76), C32e(0xa37561b7), C32e(0x49face7d),
1756
	C32e(0x8da47b52), C32e(0x42a13edd), C32e(0x93bc715e), C32e(0xa2269713),
1757
	C32e(0x0457f5a6), C32e(0xb86968b9), C32e(0x00000000), C32e(0x74992cc1),
1758
	C32e(0xa0806040), C32e(0x21dd1fe3), C32e(0x43f2c879), C32e(0x2c77edb6),
1759
	C32e(0xd9b3bed4), C32e(0xca01468d), C32e(0x70ced967), C32e(0xdde44b72),
1760
	C32e(0x7933de94), C32e(0x672bd498), C32e(0x237be8b0), C32e(0xde114a85),
1761
	C32e(0xbd6d6bbb), C32e(0x7e912ac5), C32e(0x349ee54f), C32e(0x3ac116ed),
1762
	C32e(0x5417c586), C32e(0x622fd79a), C32e(0xffcc5566), C32e(0xa7229411),
1763
	C32e(0x4a0fcf8a), C32e(0x30c910e9), C32e(0x0a080604), C32e(0x98e781fe),
1764
	C32e(0x0b5bf0a0), C32e(0xccf04478), C32e(0xd54aba25), C32e(0x3e96e34b),
1765
	C32e(0x0e5ff3a2), C32e(0x19bafe5d), C32e(0x5b1bc080), C32e(0x850a8a05),
1766
	C32e(0xec7ead3f), C32e(0xdf42bc21), C32e(0xd8e04870), C32e(0x0cf904f1),
1767
	C32e(0x7ac6df63), C32e(0x58eec177), C32e(0x9f4575af), C32e(0xa5846342),
1768
	C32e(0x50403020), C32e(0x2ed11ae5), C32e(0x12e10efd), C32e(0xb7656dbf),
1769
	C32e(0xd4194c81), C32e(0x3c301418), C32e(0x5f4c3526), C32e(0x719d2fc3),
1770
	C32e(0x3867e1be), C32e(0xfd6aa235), C32e(0x4f0bcc88), C32e(0x4b5c392e),
1771
	C32e(0xf93d5793), C32e(0x0daaf255), C32e(0x9de382fc), C32e(0xc9f4477a),
1772
	C32e(0xef8bacc8), C32e(0x326fe7ba), C32e(0x7d642b32), C32e(0xa4d795e6),
1773
	C32e(0xfb9ba0c0), C32e(0xb3329819), C32e(0x6827d19e), C32e(0x815d7fa3),
1774
	C32e(0xaa886644), C32e(0x82a87e54), C32e(0xe676ab3b), C32e(0x9e16830b),
1775
	C32e(0x4503ca8c), C32e(0x7b9529c7), C32e(0x6ed6d36b), C32e(0x44503c28),
1776
	C32e(0x8b5579a7), C32e(0x3d63e2bc), C32e(0x272c1d16), C32e(0x9a4176ad),
1777
	C32e(0x4dad3bdb), C32e(0xfac85664), C32e(0xd2e84e74), C32e(0x22281e14),
1778
	C32e(0x763fdb92), C32e(0x1e180a0c), C32e(0xb4906c48), C32e(0x376be4b8),
1779
	C32e(0xe7255d9f), C32e(0xb2616ebd), C32e(0x2a86ef43), C32e(0xf193a6c4),
1780
	C32e(0xe372a839), C32e(0xf762a431), C32e(0x59bd37d3), C32e(0x86ff8bf2),
1781
	C32e(0x56b132d5), C32e(0xc50d438b), C32e(0xebdc596e), C32e(0xc2afb7da),
1782
	C32e(0x8f028c01), C32e(0xac7964b1), C32e(0x6d23d29c), C32e(0x3b92e049),
1783
	C32e(0xc7abb4d8), C32e(0x1543faac), C32e(0x09fd07f3), C32e(0x6f8525cf),
1784
	C32e(0xea8fafca), C32e(0x89f38ef4), C32e(0x208ee947), C32e(0x28201810),
1785
	C32e(0x64ded56f), C32e(0x83fb88f0), C32e(0xb1946f4a), C32e(0x96b8725c),
1786
	C32e(0x6c702438), C32e(0x08aef157), C32e(0x52e6c773), C32e(0xf3355197),
1787
	C32e(0x658d23cb), C32e(0x84597ca1), C32e(0xbfcb9ce8), C32e(0x637c213e),
1788
	C32e(0x7c37dd96), C32e(0x7fc2dc61), C32e(0x911a860d), C32e(0x941e850f),
1789
	C32e(0xabdb90e0), C32e(0xc6f8427c), C32e(0x57e2c471), C32e(0xe583aacc),
1790
	C32e(0x733bd890), C32e(0x0f0c0506), C32e(0x03f501f7), C32e(0x3638121c),
1791
	C32e(0xfe9fa3c2), C32e(0xe1d45f6a), C32e(0x1047f9ae), C32e(0x6bd2d069),
1792
	C32e(0xa82e9117), C32e(0xe8295899), C32e(0x6974273a), C32e(0xd04eb927),
1793
	C32e(0x48a938d9), C32e(0x35cd13eb), C32e(0xce56b32b), C32e(0x55443322),
1794
	C32e(0xd6bfbbd2), C32e(0x904970a9), C32e(0x800e8907), C32e(0xf266a733),
1795
	C32e(0xc15ab62d), C32e(0x6678223c), C32e(0xad2a9215), C32e(0x608920c9),
1796
	C32e(0xdb154987), C32e(0x1a4fffaa), C32e(0x88a07850), C32e(0x8e517aa5),
1797
	C32e(0x8a068f03), C32e(0x13b2f859), C32e(0x9b128009), C32e(0x3934171a),
1798
	C32e(0x75cada65), C32e(0x53b531d7), C32e(0x5113c684), C32e(0xd3bbb8d0),
1799
	C32e(0x5e1fc382), C32e(0xcb52b029), C32e(0x99b4775a), C32e(0x333c111e),
1800
	C32e(0x46f6cb7b), C32e(0x1f4bfca8), C32e(0x61dad66d), C32e(0x4e583a2c)
1801
};
1802

1803
static const sph_u32 T1up[] = {
1804
	C32e(0xc6c632f4), C32e(0xf8f86f97), C32e(0xeeee5eb0), C32e(0xf6f67a8c),
1805
	C32e(0xffffe817), C32e(0xd6d60adc), C32e(0xdede16c8), C32e(0x91916dfc),
1806
	C32e(0x606090f0), C32e(0x02020705), C32e(0xcece2ee0), C32e(0x5656d187),
1807
	C32e(0xe7e7cc2b), C32e(0xb5b513a6), C32e(0x4d4d7c31), C32e(0xecec59b5),
1808
	C32e(0x8f8f40cf), C32e(0x1f1fa3bc), C32e(0x898949c0), C32e(0xfafa6892),
1809
	C32e(0xefefd03f), C32e(0xb2b29426), C32e(0x8e8ece40), C32e(0xfbfbe61d),
1810
	C32e(0x41416e2f), C32e(0xb3b31aa9), C32e(0x5f5f431c), C32e(0x45456025),
1811
	C32e(0x2323f9da), C32e(0x53535102), C32e(0xe4e445a1), C32e(0x9b9b76ed),
1812
	C32e(0x7575285d), C32e(0xe1e1c524), C32e(0x3d3dd4e9), C32e(0x4c4cf2be),
1813
	C32e(0x6c6c82ee), C32e(0x7e7ebdc3), C32e(0xf5f5f306), C32e(0x838352d1),
1814
	C32e(0x68688ce4), C32e(0x51515607), C32e(0xd1d18d5c), C32e(0xf9f9e118),
1815
	C32e(0xe2e24cae), C32e(0xabab3e95), C32e(0x626297f5), C32e(0x2a2a6b41),
1816
	C32e(0x08081c14), C32e(0x959563f6), C32e(0x4646e9af), C32e(0x9d9d7fe2),
1817
	C32e(0x30304878), C32e(0x3737cff8), C32e(0x0a0a1b11), C32e(0x2f2febc4),
1818
	C32e(0x0e0e151b), C32e(0x24247e5a), C32e(0x1b1badb6), C32e(0xdfdf9847),
1819
	C32e(0xcdcda76a), C32e(0x4e4ef5bb), C32e(0x7f7f334c), C32e(0xeaea50ba),
1820
	C32e(0x12123f2d), C32e(0x1d1da4b9), C32e(0x5858c49c), C32e(0x34344672),
1821
	C32e(0x36364177), C32e(0xdcdc11cd), C32e(0xb4b49d29), C32e(0x5b5b4d16),
1822
	C32e(0xa4a4a501), C32e(0x7676a1d7), C32e(0xb7b714a3), C32e(0x7d7d3449),
1823
	C32e(0x5252df8d), C32e(0xdddd9f42), C32e(0x5e5ecd93), C32e(0x1313b1a2),
1824
	C32e(0xa6a6a204), C32e(0xb9b901b8), C32e(0x00000000), C32e(0xc1c1b574),
1825
	C32e(0x4040e0a0), C32e(0xe3e3c221), C32e(0x79793a43), C32e(0xb6b69a2c),
1826
	C32e(0xd4d40dd9), C32e(0x8d8d47ca), C32e(0x67671770), C32e(0x7272afdd),
1827
	C32e(0x9494ed79), C32e(0x9898ff67), C32e(0xb0b09323), C32e(0x85855bde),
1828
	C32e(0xbbbb06bd), C32e(0xc5c5bb7e), C32e(0x4f4f7b34), C32e(0xededd73a),
1829
	C32e(0x8686d254), C32e(0x9a9af862), C32e(0x666699ff), C32e(0x1111b6a7),
1830
	C32e(0x8a8ac04a), C32e(0xe9e9d930), C32e(0x04040e0a), C32e(0xfefe6698),
1831
	C32e(0xa0a0ab0b), C32e(0x7878b4cc), C32e(0x2525f0d5), C32e(0x4b4b753e),
1832
	C32e(0xa2a2ac0e), C32e(0x5d5d4419), C32e(0x8080db5b), C32e(0x05058085),
1833
	C32e(0x3f3fd3ec), C32e(0x2121fedf), C32e(0x7070a8d8), C32e(0xf1f1fd0c),
1834
	C32e(0x6363197a), C32e(0x77772f58), C32e(0xafaf309f), C32e(0x4242e7a5),
1835
	C32e(0x20207050), C32e(0xe5e5cb2e), C32e(0xfdfdef12), C32e(0xbfbf08b7),
1836
	C32e(0x818155d4), C32e(0x1818243c), C32e(0x2626795f), C32e(0xc3c3b271),
1837
	C32e(0xbebe8638), C32e(0x3535c8fd), C32e(0x8888c74f), C32e(0x2e2e654b),
1838
	C32e(0x93936af9), C32e(0x5555580d), C32e(0xfcfc619d), C32e(0x7a7ab3c9),
1839
	C32e(0xc8c827ef), C32e(0xbaba8832), C32e(0x32324f7d), C32e(0xe6e642a4),
1840
	C32e(0xc0c03bfb), C32e(0x1919aab3), C32e(0x9e9ef668), C32e(0xa3a32281),
1841
	C32e(0x4444eeaa), C32e(0x5454d682), C32e(0x3b3bdde6), C32e(0x0b0b959e),
1842
	C32e(0x8c8cc945), C32e(0xc7c7bc7b), C32e(0x6b6b056e), C32e(0x28286c44),
1843
	C32e(0xa7a72c8b), C32e(0xbcbc813d), C32e(0x16163127), C32e(0xadad379a),
1844
	C32e(0xdbdb964d), C32e(0x64649efa), C32e(0x7474a6d2), C32e(0x14143622),
1845
	C32e(0x9292e476), C32e(0x0c0c121e), C32e(0x4848fcb4), C32e(0xb8b88f37),
1846
	C32e(0x9f9f78e7), C32e(0xbdbd0fb2), C32e(0x4343692a), C32e(0xc4c435f1),
1847
	C32e(0x3939dae3), C32e(0x3131c6f7), C32e(0xd3d38a59), C32e(0xf2f27486),
1848
	C32e(0xd5d58356), C32e(0x8b8b4ec5), C32e(0x6e6e85eb), C32e(0xdada18c2),
1849
	C32e(0x01018e8f), C32e(0xb1b11dac), C32e(0x9c9cf16d), C32e(0x4949723b),
1850
	C32e(0xd8d81fc7), C32e(0xacacb915), C32e(0xf3f3fa09), C32e(0xcfcfa06f),
1851
	C32e(0xcaca20ea), C32e(0xf4f47d89), C32e(0x47476720), C32e(0x10103828),
1852
	C32e(0x6f6f0b64), C32e(0xf0f07383), C32e(0x4a4afbb1), C32e(0x5c5cca96),
1853
	C32e(0x3838546c), C32e(0x57575f08), C32e(0x73732152), C32e(0x979764f3),
1854
	C32e(0xcbcbae65), C32e(0xa1a12584), C32e(0xe8e857bf), C32e(0x3e3e5d63),
1855
	C32e(0x9696ea7c), C32e(0x61611e7f), C32e(0x0d0d9c91), C32e(0x0f0f9b94),
1856
	C32e(0xe0e04bab), C32e(0x7c7cbac6), C32e(0x71712657), C32e(0xcccc29e5),
1857
	C32e(0x9090e373), C32e(0x0606090f), C32e(0xf7f7f403), C32e(0x1c1c2a36),
1858
	C32e(0xc2c23cfe), C32e(0x6a6a8be1), C32e(0xaeaebe10), C32e(0x6969026b),
1859
	C32e(0x1717bfa8), C32e(0x999971e8), C32e(0x3a3a5369), C32e(0x2727f7d0),
1860
	C32e(0xd9d99148), C32e(0xebebde35), C32e(0x2b2be5ce), C32e(0x22227755),
1861
	C32e(0xd2d204d6), C32e(0xa9a93990), C32e(0x07078780), C32e(0x3333c1f2),
1862
	C32e(0x2d2decc1), C32e(0x3c3c5a66), C32e(0x1515b8ad), C32e(0xc9c9a960),
1863
	C32e(0x87875cdb), C32e(0xaaaab01a), C32e(0x5050d888), C32e(0xa5a52b8e),
1864
	C32e(0x0303898a), C32e(0x59594a13), C32e(0x0909929b), C32e(0x1a1a2339),
1865
	C32e(0x65651075), C32e(0xd7d78453), C32e(0x8484d551), C32e(0xd0d003d3),
1866
	C32e(0x8282dc5e), C32e(0x2929e2cb), C32e(0x5a5ac399), C32e(0x1e1e2d33),
1867
	C32e(0x7b7b3d46), C32e(0xa8a8b71f), C32e(0x6d6d0c61), C32e(0x2c2c624e)
1868
};
1869

1870
static const sph_u32 T1dn[] = {
1871
	C32e(0xa5f497a5), C32e(0x8497eb84), C32e(0x99b0c799), C32e(0x8d8cf78d),
1872
	C32e(0x0d17e50d), C32e(0xbddcb7bd), C32e(0xb1c8a7b1), C32e(0x54fc3954),
1873
	C32e(0x50f0c050), C32e(0x03050403), C32e(0xa9e087a9), C32e(0x7d87ac7d),
1874
	C32e(0x192bd519), C32e(0x62a67162), C32e(0xe6319ae6), C32e(0x9ab5c39a),
1875
	C32e(0x45cf0545), C32e(0x9dbc3e9d), C32e(0x40c00940), C32e(0x8792ef87),
1876
	C32e(0x153fc515), C32e(0xeb267feb), C32e(0xc94007c9), C32e(0x0b1ded0b),
1877
	C32e(0xec2f82ec), C32e(0x67a97d67), C32e(0xfd1cbefd), C32e(0xea258aea),
1878
	C32e(0xbfda46bf), C32e(0xf702a6f7), C32e(0x96a1d396), C32e(0x5bed2d5b),
1879
	C32e(0xc25deac2), C32e(0x1c24d91c), C32e(0xaee97aae), C32e(0x6abe986a),
1880
	C32e(0x5aeed85a), C32e(0x41c3fc41), C32e(0x0206f102), C32e(0x4fd11d4f),
1881
	C32e(0x5ce4d05c), C32e(0xf407a2f4), C32e(0x345cb934), C32e(0x0818e908),
1882
	C32e(0x93aedf93), C32e(0x73954d73), C32e(0x53f5c453), C32e(0x3f41543f),
1883
	C32e(0x0c14100c), C32e(0x52f63152), C32e(0x65af8c65), C32e(0x5ee2215e),
1884
	C32e(0x28786028), C32e(0xa1f86ea1), C32e(0x0f11140f), C32e(0xb5c45eb5),
1885
	C32e(0x091b1c09), C32e(0x365a4836), C32e(0x9bb6369b), C32e(0x3d47a53d),
1886
	C32e(0x266a8126), C32e(0x69bb9c69), C32e(0xcd4cfecd), C32e(0x9fbacf9f),
1887
	C32e(0x1b2d241b), C32e(0x9eb93a9e), C32e(0x749cb074), C32e(0x2e72682e),
1888
	C32e(0x2d776c2d), C32e(0xb2cda3b2), C32e(0xee2973ee), C32e(0xfb16b6fb),
1889
	C32e(0xf60153f6), C32e(0x4dd7ec4d), C32e(0x61a37561), C32e(0xce49face),
1890
	C32e(0x7b8da47b), C32e(0x3e42a13e), C32e(0x7193bc71), C32e(0x97a22697),
1891
	C32e(0xf50457f5), C32e(0x68b86968), C32e(0x00000000), C32e(0x2c74992c),
1892
	C32e(0x60a08060), C32e(0x1f21dd1f), C32e(0xc843f2c8), C32e(0xed2c77ed),
1893
	C32e(0xbed9b3be), C32e(0x46ca0146), C32e(0xd970ced9), C32e(0x4bdde44b),
1894
	C32e(0xde7933de), C32e(0xd4672bd4), C32e(0xe8237be8), C32e(0x4ade114a),
1895
	C32e(0x6bbd6d6b), C32e(0x2a7e912a), C32e(0xe5349ee5), C32e(0x163ac116),
1896
	C32e(0xc55417c5), C32e(0xd7622fd7), C32e(0x55ffcc55), C32e(0x94a72294),
1897
	C32e(0xcf4a0fcf), C32e(0x1030c910), C32e(0x060a0806), C32e(0x8198e781),
1898
	C32e(0xf00b5bf0), C32e(0x44ccf044), C32e(0xbad54aba), C32e(0xe33e96e3),
1899
	C32e(0xf30e5ff3), C32e(0xfe19bafe), C32e(0xc05b1bc0), C32e(0x8a850a8a),
1900
	C32e(0xadec7ead), C32e(0xbcdf42bc), C32e(0x48d8e048), C32e(0x040cf904),
1901
	C32e(0xdf7ac6df), C32e(0xc158eec1), C32e(0x759f4575), C32e(0x63a58463),
1902
	C32e(0x30504030), C32e(0x1a2ed11a), C32e(0x0e12e10e), C32e(0x6db7656d),
1903
	C32e(0x4cd4194c), C32e(0x143c3014), C32e(0x355f4c35), C32e(0x2f719d2f),
1904
	C32e(0xe13867e1), C32e(0xa2fd6aa2), C32e(0xcc4f0bcc), C32e(0x394b5c39),
1905
	C32e(0x57f93d57), C32e(0xf20daaf2), C32e(0x829de382), C32e(0x47c9f447),
1906
	C32e(0xacef8bac), C32e(0xe7326fe7), C32e(0x2b7d642b), C32e(0x95a4d795),
1907
	C32e(0xa0fb9ba0), C32e(0x98b33298), C32e(0xd16827d1), C32e(0x7f815d7f),
1908
	C32e(0x66aa8866), C32e(0x7e82a87e), C32e(0xabe676ab), C32e(0x839e1683),
1909
	C32e(0xca4503ca), C32e(0x297b9529), C32e(0xd36ed6d3), C32e(0x3c44503c),
1910
	C32e(0x798b5579), C32e(0xe23d63e2), C32e(0x1d272c1d), C32e(0x769a4176),
1911
	C32e(0x3b4dad3b), C32e(0x56fac856), C32e(0x4ed2e84e), C32e(0x1e22281e),
1912
	C32e(0xdb763fdb), C32e(0x0a1e180a), C32e(0x6cb4906c), C32e(0xe4376be4),
1913
	C32e(0x5de7255d), C32e(0x6eb2616e), C32e(0xef2a86ef), C32e(0xa6f193a6),
1914
	C32e(0xa8e372a8), C32e(0xa4f762a4), C32e(0x3759bd37), C32e(0x8b86ff8b),
1915
	C32e(0x3256b132), C32e(0x43c50d43), C32e(0x59ebdc59), C32e(0xb7c2afb7),
1916
	C32e(0x8c8f028c), C32e(0x64ac7964), C32e(0xd26d23d2), C32e(0xe03b92e0),
1917
	C32e(0xb4c7abb4), C32e(0xfa1543fa), C32e(0x0709fd07), C32e(0x256f8525),
1918
	C32e(0xafea8faf), C32e(0x8e89f38e), C32e(0xe9208ee9), C32e(0x18282018),
1919
	C32e(0xd564ded5), C32e(0x8883fb88), C32e(0x6fb1946f), C32e(0x7296b872),
1920
	C32e(0x246c7024), C32e(0xf108aef1), C32e(0xc752e6c7), C32e(0x51f33551),
1921
	C32e(0x23658d23), C32e(0x7c84597c), C32e(0x9cbfcb9c), C32e(0x21637c21),
1922
	C32e(0xdd7c37dd), C32e(0xdc7fc2dc), C32e(0x86911a86), C32e(0x85941e85),
1923
	C32e(0x90abdb90), C32e(0x42c6f842), C32e(0xc457e2c4), C32e(0xaae583aa),
1924
	C32e(0xd8733bd8), C32e(0x050f0c05), C32e(0x0103f501), C32e(0x12363812),
1925
	C32e(0xa3fe9fa3), C32e(0x5fe1d45f), C32e(0xf91047f9), C32e(0xd06bd2d0),
1926
	C32e(0x91a82e91), C32e(0x58e82958), C32e(0x27697427), C32e(0xb9d04eb9),
1927
	C32e(0x3848a938), C32e(0x1335cd13), C32e(0xb3ce56b3), C32e(0x33554433),
1928
	C32e(0xbbd6bfbb), C32e(0x70904970), C32e(0x89800e89), C32e(0xa7f266a7),
1929
	C32e(0xb6c15ab6), C32e(0x22667822), C32e(0x92ad2a92), C32e(0x20608920),
1930
	C32e(0x49db1549), C32e(0xff1a4fff), C32e(0x7888a078), C32e(0x7a8e517a),
1931
	C32e(0x8f8a068f), C32e(0xf813b2f8), C32e(0x809b1280), C32e(0x17393417),
1932
	C32e(0xda75cada), C32e(0x3153b531), C32e(0xc65113c6), C32e(0xb8d3bbb8),
1933
	C32e(0xc35e1fc3), C32e(0xb0cb52b0), C32e(0x7799b477), C32e(0x11333c11),
1934
	C32e(0xcb46f6cb), C32e(0xfc1f4bfc), C32e(0xd661dad6), C32e(0x3a4e583a)
1935
};
1936

1937
static const sph_u32 T2up[] = {
1938
	C32e(0xa5c6c632), C32e(0x84f8f86f), C32e(0x99eeee5e), C32e(0x8df6f67a),
1939
	C32e(0x0dffffe8), C32e(0xbdd6d60a), C32e(0xb1dede16), C32e(0x5491916d),
1940
	C32e(0x50606090), C32e(0x03020207), C32e(0xa9cece2e), C32e(0x7d5656d1),
1941
	C32e(0x19e7e7cc), C32e(0x62b5b513), C32e(0xe64d4d7c), C32e(0x9aecec59),
1942
	C32e(0x458f8f40), C32e(0x9d1f1fa3), C32e(0x40898949), C32e(0x87fafa68),
1943
	C32e(0x15efefd0), C32e(0xebb2b294), C32e(0xc98e8ece), C32e(0x0bfbfbe6),
1944
	C32e(0xec41416e), C32e(0x67b3b31a), C32e(0xfd5f5f43), C32e(0xea454560),
1945
	C32e(0xbf2323f9), C32e(0xf7535351), C32e(0x96e4e445), C32e(0x5b9b9b76),
1946
	C32e(0xc2757528), C32e(0x1ce1e1c5), C32e(0xae3d3dd4), C32e(0x6a4c4cf2),
1947
	C32e(0x5a6c6c82), C32e(0x417e7ebd), C32e(0x02f5f5f3), C32e(0x4f838352),
1948
	C32e(0x5c68688c), C32e(0xf4515156), C32e(0x34d1d18d), C32e(0x08f9f9e1),
1949
	C32e(0x93e2e24c), C32e(0x73abab3e), C32e(0x53626297), C32e(0x3f2a2a6b),
1950
	C32e(0x0c08081c), C32e(0x52959563), C32e(0x654646e9), C32e(0x5e9d9d7f),
1951
	C32e(0x28303048), C32e(0xa13737cf), C32e(0x0f0a0a1b), C32e(0xb52f2feb),
1952
	C32e(0x090e0e15), C32e(0x3624247e), C32e(0x9b1b1bad), C32e(0x3ddfdf98),
1953
	C32e(0x26cdcda7), C32e(0x694e4ef5), C32e(0xcd7f7f33), C32e(0x9feaea50),
1954
	C32e(0x1b12123f), C32e(0x9e1d1da4), C32e(0x745858c4), C32e(0x2e343446),
1955
	C32e(0x2d363641), C32e(0xb2dcdc11), C32e(0xeeb4b49d), C32e(0xfb5b5b4d),
1956
	C32e(0xf6a4a4a5), C32e(0x4d7676a1), C32e(0x61b7b714), C32e(0xce7d7d34),
1957
	C32e(0x7b5252df), C32e(0x3edddd9f), C32e(0x715e5ecd), C32e(0x971313b1),
1958
	C32e(0xf5a6a6a2), C32e(0x68b9b901), C32e(0x00000000), C32e(0x2cc1c1b5),
1959
	C32e(0x604040e0), C32e(0x1fe3e3c2), C32e(0xc879793a), C32e(0xedb6b69a),
1960
	C32e(0xbed4d40d), C32e(0x468d8d47), C32e(0xd9676717), C32e(0x4b7272af),
1961
	C32e(0xde9494ed), C32e(0xd49898ff), C32e(0xe8b0b093), C32e(0x4a85855b),
1962
	C32e(0x6bbbbb06), C32e(0x2ac5c5bb), C32e(0xe54f4f7b), C32e(0x16ededd7),
1963
	C32e(0xc58686d2), C32e(0xd79a9af8), C32e(0x55666699), C32e(0x941111b6),
1964
	C32e(0xcf8a8ac0), C32e(0x10e9e9d9), C32e(0x0604040e), C32e(0x81fefe66),
1965
	C32e(0xf0a0a0ab), C32e(0x447878b4), C32e(0xba2525f0), C32e(0xe34b4b75),
1966
	C32e(0xf3a2a2ac), C32e(0xfe5d5d44), C32e(0xc08080db), C32e(0x8a050580),
1967
	C32e(0xad3f3fd3), C32e(0xbc2121fe), C32e(0x487070a8), C32e(0x04f1f1fd),
1968
	C32e(0xdf636319), C32e(0xc177772f), C32e(0x75afaf30), C32e(0x634242e7),
1969
	C32e(0x30202070), C32e(0x1ae5e5cb), C32e(0x0efdfdef), C32e(0x6dbfbf08),
1970
	C32e(0x4c818155), C32e(0x14181824), C32e(0x35262679), C32e(0x2fc3c3b2),
1971
	C32e(0xe1bebe86), C32e(0xa23535c8), C32e(0xcc8888c7), C32e(0x392e2e65),
1972
	C32e(0x5793936a), C32e(0xf2555558), C32e(0x82fcfc61), C32e(0x477a7ab3),
1973
	C32e(0xacc8c827), C32e(0xe7baba88), C32e(0x2b32324f), C32e(0x95e6e642),
1974
	C32e(0xa0c0c03b), C32e(0x981919aa), C32e(0xd19e9ef6), C32e(0x7fa3a322),
1975
	C32e(0x664444ee), C32e(0x7e5454d6), C32e(0xab3b3bdd), C32e(0x830b0b95),
1976
	C32e(0xca8c8cc9), C32e(0x29c7c7bc), C32e(0xd36b6b05), C32e(0x3c28286c),
1977
	C32e(0x79a7a72c), C32e(0xe2bcbc81), C32e(0x1d161631), C32e(0x76adad37),
1978
	C32e(0x3bdbdb96), C32e(0x5664649e), C32e(0x4e7474a6), C32e(0x1e141436),
1979
	C32e(0xdb9292e4), C32e(0x0a0c0c12), C32e(0x6c4848fc), C32e(0xe4b8b88f),
1980
	C32e(0x5d9f9f78), C32e(0x6ebdbd0f), C32e(0xef434369), C32e(0xa6c4c435),
1981
	C32e(0xa83939da), C32e(0xa43131c6), C32e(0x37d3d38a), C32e(0x8bf2f274),
1982
	C32e(0x32d5d583), C32e(0x438b8b4e), C32e(0x596e6e85), C32e(0xb7dada18),
1983
	C32e(0x8c01018e), C32e(0x64b1b11d), C32e(0xd29c9cf1), C32e(0xe0494972),
1984
	C32e(0xb4d8d81f), C32e(0xfaacacb9), C32e(0x07f3f3fa), C32e(0x25cfcfa0),
1985
	C32e(0xafcaca20), C32e(0x8ef4f47d), C32e(0xe9474767), C32e(0x18101038),
1986
	C32e(0xd56f6f0b), C32e(0x88f0f073), C32e(0x6f4a4afb), C32e(0x725c5cca),
1987
	C32e(0x24383854), C32e(0xf157575f), C32e(0xc7737321), C32e(0x51979764),
1988
	C32e(0x23cbcbae), C32e(0x7ca1a125), C32e(0x9ce8e857), C32e(0x213e3e5d),
1989
	C32e(0xdd9696ea), C32e(0xdc61611e), C32e(0x860d0d9c), C32e(0x850f0f9b),
1990
	C32e(0x90e0e04b), C32e(0x427c7cba), C32e(0xc4717126), C32e(0xaacccc29),
1991
	C32e(0xd89090e3), C32e(0x05060609), C32e(0x01f7f7f4), C32e(0x121c1c2a),
1992
	C32e(0xa3c2c23c), C32e(0x5f6a6a8b), C32e(0xf9aeaebe), C32e(0xd0696902),
1993
	C32e(0x911717bf), C32e(0x58999971), C32e(0x273a3a53), C32e(0xb92727f7),
1994
	C32e(0x38d9d991), C32e(0x13ebebde), C32e(0xb32b2be5), C32e(0x33222277),
1995
	C32e(0xbbd2d204), C32e(0x70a9a939), C32e(0x89070787), C32e(0xa73333c1),
1996
	C32e(0xb62d2dec), C32e(0x223c3c5a), C32e(0x921515b8), C32e(0x20c9c9a9),
1997
	C32e(0x4987875c), C32e(0xffaaaab0), C32e(0x785050d8), C32e(0x7aa5a52b),
1998
	C32e(0x8f030389), C32e(0xf859594a), C32e(0x80090992), C32e(0x171a1a23),
1999
	C32e(0xda656510), C32e(0x31d7d784), C32e(0xc68484d5), C32e(0xb8d0d003),
2000
	C32e(0xc38282dc), C32e(0xb02929e2), C32e(0x775a5ac3), C32e(0x111e1e2d),
2001
	C32e(0xcb7b7b3d), C32e(0xfca8a8b7), C32e(0xd66d6d0c), C32e(0x3a2c2c62)
2002
};
2003

2004
static const sph_u32 T2dn[] = {
2005
	C32e(0xf4a5f497), C32e(0x978497eb), C32e(0xb099b0c7), C32e(0x8c8d8cf7),
2006
	C32e(0x170d17e5), C32e(0xdcbddcb7), C32e(0xc8b1c8a7), C32e(0xfc54fc39),
2007
	C32e(0xf050f0c0), C32e(0x05030504), C32e(0xe0a9e087), C32e(0x877d87ac),
2008
	C32e(0x2b192bd5), C32e(0xa662a671), C32e(0x31e6319a), C32e(0xb59ab5c3),
2009
	C32e(0xcf45cf05), C32e(0xbc9dbc3e), C32e(0xc040c009), C32e(0x928792ef),
2010
	C32e(0x3f153fc5), C32e(0x26eb267f), C32e(0x40c94007), C32e(0x1d0b1ded),
2011
	C32e(0x2fec2f82), C32e(0xa967a97d), C32e(0x1cfd1cbe), C32e(0x25ea258a),
2012
	C32e(0xdabfda46), C32e(0x02f702a6), C32e(0xa196a1d3), C32e(0xed5bed2d),
2013
	C32e(0x5dc25dea), C32e(0x241c24d9), C32e(0xe9aee97a), C32e(0xbe6abe98),
2014
	C32e(0xee5aeed8), C32e(0xc341c3fc), C32e(0x060206f1), C32e(0xd14fd11d),
2015
	C32e(0xe45ce4d0), C32e(0x07f407a2), C32e(0x5c345cb9), C32e(0x180818e9),
2016
	C32e(0xae93aedf), C32e(0x9573954d), C32e(0xf553f5c4), C32e(0x413f4154),
2017
	C32e(0x140c1410), C32e(0xf652f631), C32e(0xaf65af8c), C32e(0xe25ee221),
2018
	C32e(0x78287860), C32e(0xf8a1f86e), C32e(0x110f1114), C32e(0xc4b5c45e),
2019
	C32e(0x1b091b1c), C32e(0x5a365a48), C32e(0xb69bb636), C32e(0x473d47a5),
2020
	C32e(0x6a266a81), C32e(0xbb69bb9c), C32e(0x4ccd4cfe), C32e(0xba9fbacf),
2021
	C32e(0x2d1b2d24), C32e(0xb99eb93a), C32e(0x9c749cb0), C32e(0x722e7268),
2022
	C32e(0x772d776c), C32e(0xcdb2cda3), C32e(0x29ee2973), C32e(0x16fb16b6),
2023
	C32e(0x01f60153), C32e(0xd74dd7ec), C32e(0xa361a375), C32e(0x49ce49fa),
2024
	C32e(0x8d7b8da4), C32e(0x423e42a1), C32e(0x937193bc), C32e(0xa297a226),
2025
	C32e(0x04f50457), C32e(0xb868b869), C32e(0x00000000), C32e(0x742c7499),
2026
	C32e(0xa060a080), C32e(0x211f21dd), C32e(0x43c843f2), C32e(0x2ced2c77),
2027
	C32e(0xd9bed9b3), C32e(0xca46ca01), C32e(0x70d970ce), C32e(0xdd4bdde4),
2028
	C32e(0x79de7933), C32e(0x67d4672b), C32e(0x23e8237b), C32e(0xde4ade11),
2029
	C32e(0xbd6bbd6d), C32e(0x7e2a7e91), C32e(0x34e5349e), C32e(0x3a163ac1),
2030
	C32e(0x54c55417), C32e(0x62d7622f), C32e(0xff55ffcc), C32e(0xa794a722),
2031
	C32e(0x4acf4a0f), C32e(0x301030c9), C32e(0x0a060a08), C32e(0x988198e7),
2032
	C32e(0x0bf00b5b), C32e(0xcc44ccf0), C32e(0xd5bad54a), C32e(0x3ee33e96),
2033
	C32e(0x0ef30e5f), C32e(0x19fe19ba), C32e(0x5bc05b1b), C32e(0x858a850a),
2034
	C32e(0xecadec7e), C32e(0xdfbcdf42), C32e(0xd848d8e0), C32e(0x0c040cf9),
2035
	C32e(0x7adf7ac6), C32e(0x58c158ee), C32e(0x9f759f45), C32e(0xa563a584),
2036
	C32e(0x50305040), C32e(0x2e1a2ed1), C32e(0x120e12e1), C32e(0xb76db765),
2037
	C32e(0xd44cd419), C32e(0x3c143c30), C32e(0x5f355f4c), C32e(0x712f719d),
2038
	C32e(0x38e13867), C32e(0xfda2fd6a), C32e(0x4fcc4f0b), C32e(0x4b394b5c),
2039
	C32e(0xf957f93d), C32e(0x0df20daa), C32e(0x9d829de3), C32e(0xc947c9f4),
2040
	C32e(0xefacef8b), C32e(0x32e7326f), C32e(0x7d2b7d64), C32e(0xa495a4d7),
2041
	C32e(0xfba0fb9b), C32e(0xb398b332), C32e(0x68d16827), C32e(0x817f815d),
2042
	C32e(0xaa66aa88), C32e(0x827e82a8), C32e(0xe6abe676), C32e(0x9e839e16),
2043
	C32e(0x45ca4503), C32e(0x7b297b95), C32e(0x6ed36ed6), C32e(0x443c4450),
2044
	C32e(0x8b798b55), C32e(0x3de23d63), C32e(0x271d272c), C32e(0x9a769a41),
2045
	C32e(0x4d3b4dad), C32e(0xfa56fac8), C32e(0xd24ed2e8), C32e(0x221e2228),
2046
	C32e(0x76db763f), C32e(0x1e0a1e18), C32e(0xb46cb490), C32e(0x37e4376b),
2047
	C32e(0xe75de725), C32e(0xb26eb261), C32e(0x2aef2a86), C32e(0xf1a6f193),
2048
	C32e(0xe3a8e372), C32e(0xf7a4f762), C32e(0x593759bd), C32e(0x868b86ff),
2049
	C32e(0x563256b1), C32e(0xc543c50d), C32e(0xeb59ebdc), C32e(0xc2b7c2af),
2050
	C32e(0x8f8c8f02), C32e(0xac64ac79), C32e(0x6dd26d23), C32e(0x3be03b92),
2051
	C32e(0xc7b4c7ab), C32e(0x15fa1543), C32e(0x090709fd), C32e(0x6f256f85),
2052
	C32e(0xeaafea8f), C32e(0x898e89f3), C32e(0x20e9208e), C32e(0x28182820),
2053
	C32e(0x64d564de), C32e(0x838883fb), C32e(0xb16fb194), C32e(0x967296b8),
2054
	C32e(0x6c246c70), C32e(0x08f108ae), C32e(0x52c752e6), C32e(0xf351f335),
2055
	C32e(0x6523658d), C32e(0x847c8459), C32e(0xbf9cbfcb), C32e(0x6321637c),
2056
	C32e(0x7cdd7c37), C32e(0x7fdc7fc2), C32e(0x9186911a), C32e(0x9485941e),
2057
	C32e(0xab90abdb), C32e(0xc642c6f8), C32e(0x57c457e2), C32e(0xe5aae583),
2058
	C32e(0x73d8733b), C32e(0x0f050f0c), C32e(0x030103f5), C32e(0x36123638),
2059
	C32e(0xfea3fe9f), C32e(0xe15fe1d4), C32e(0x10f91047), C32e(0x6bd06bd2),
2060
	C32e(0xa891a82e), C32e(0xe858e829), C32e(0x69276974), C32e(0xd0b9d04e),
2061
	C32e(0x483848a9), C32e(0x351335cd), C32e(0xceb3ce56), C32e(0x55335544),
2062
	C32e(0xd6bbd6bf), C32e(0x90709049), C32e(0x8089800e), C32e(0xf2a7f266),
2063
	C32e(0xc1b6c15a), C32e(0x66226678), C32e(0xad92ad2a), C32e(0x60206089),
2064
	C32e(0xdb49db15), C32e(0x1aff1a4f), C32e(0x887888a0), C32e(0x8e7a8e51),
2065
	C32e(0x8a8f8a06), C32e(0x13f813b2), C32e(0x9b809b12), C32e(0x39173934),
2066
	C32e(0x75da75ca), C32e(0x533153b5), C32e(0x51c65113), C32e(0xd3b8d3bb),
2067
	C32e(0x5ec35e1f), C32e(0xcbb0cb52), C32e(0x997799b4), C32e(0x3311333c),
2068
	C32e(0x46cb46f6), C32e(0x1ffc1f4b), C32e(0x61d661da), C32e(0x4e3a4e58)
2069
};
2070

2071
static const sph_u32 T3up[] = {
2072
	C32e(0x97a5c6c6), C32e(0xeb84f8f8), C32e(0xc799eeee), C32e(0xf78df6f6),
2073
	C32e(0xe50dffff), C32e(0xb7bdd6d6), C32e(0xa7b1dede), C32e(0x39549191),
2074
	C32e(0xc0506060), C32e(0x04030202), C32e(0x87a9cece), C32e(0xac7d5656),
2075
	C32e(0xd519e7e7), C32e(0x7162b5b5), C32e(0x9ae64d4d), C32e(0xc39aecec),
2076
	C32e(0x05458f8f), C32e(0x3e9d1f1f), C32e(0x09408989), C32e(0xef87fafa),
2077
	C32e(0xc515efef), C32e(0x7febb2b2), C32e(0x07c98e8e), C32e(0xed0bfbfb),
2078
	C32e(0x82ec4141), C32e(0x7d67b3b3), C32e(0xbefd5f5f), C32e(0x8aea4545),
2079
	C32e(0x46bf2323), C32e(0xa6f75353), C32e(0xd396e4e4), C32e(0x2d5b9b9b),
2080
	C32e(0xeac27575), C32e(0xd91ce1e1), C32e(0x7aae3d3d), C32e(0x986a4c4c),
2081
	C32e(0xd85a6c6c), C32e(0xfc417e7e), C32e(0xf102f5f5), C32e(0x1d4f8383),
2082
	C32e(0xd05c6868), C32e(0xa2f45151), C32e(0xb934d1d1), C32e(0xe908f9f9),
2083
	C32e(0xdf93e2e2), C32e(0x4d73abab), C32e(0xc4536262), C32e(0x543f2a2a),
2084
	C32e(0x100c0808), C32e(0x31529595), C32e(0x8c654646), C32e(0x215e9d9d),
2085
	C32e(0x60283030), C32e(0x6ea13737), C32e(0x140f0a0a), C32e(0x5eb52f2f),
2086
	C32e(0x1c090e0e), C32e(0x48362424), C32e(0x369b1b1b), C32e(0xa53ddfdf),
2087
	C32e(0x8126cdcd), C32e(0x9c694e4e), C32e(0xfecd7f7f), C32e(0xcf9feaea),
2088
	C32e(0x241b1212), C32e(0x3a9e1d1d), C32e(0xb0745858), C32e(0x682e3434),
2089
	C32e(0x6c2d3636), C32e(0xa3b2dcdc), C32e(0x73eeb4b4), C32e(0xb6fb5b5b),
2090
	C32e(0x53f6a4a4), C32e(0xec4d7676), C32e(0x7561b7b7), C32e(0xface7d7d),
2091
	C32e(0xa47b5252), C32e(0xa13edddd), C32e(0xbc715e5e), C32e(0x26971313),
2092
	C32e(0x57f5a6a6), C32e(0x6968b9b9), C32e(0x00000000), C32e(0x992cc1c1),
2093
	C32e(0x80604040), C32e(0xdd1fe3e3), C32e(0xf2c87979), C32e(0x77edb6b6),
2094
	C32e(0xb3bed4d4), C32e(0x01468d8d), C32e(0xced96767), C32e(0xe44b7272),
2095
	C32e(0x33de9494), C32e(0x2bd49898), C32e(0x7be8b0b0), C32e(0x114a8585),
2096
	C32e(0x6d6bbbbb), C32e(0x912ac5c5), C32e(0x9ee54f4f), C32e(0xc116eded),
2097
	C32e(0x17c58686), C32e(0x2fd79a9a), C32e(0xcc556666), C32e(0x22941111),
2098
	C32e(0x0fcf8a8a), C32e(0xc910e9e9), C32e(0x08060404), C32e(0xe781fefe),
2099
	C32e(0x5bf0a0a0), C32e(0xf0447878), C32e(0x4aba2525), C32e(0x96e34b4b),
2100
	C32e(0x5ff3a2a2), C32e(0xbafe5d5d), C32e(0x1bc08080), C32e(0x0a8a0505),
2101
	C32e(0x7ead3f3f), C32e(0x42bc2121), C32e(0xe0487070), C32e(0xf904f1f1),
2102
	C32e(0xc6df6363), C32e(0xeec17777), C32e(0x4575afaf), C32e(0x84634242),
2103
	C32e(0x40302020), C32e(0xd11ae5e5), C32e(0xe10efdfd), C32e(0x656dbfbf),
2104
	C32e(0x194c8181), C32e(0x30141818), C32e(0x4c352626), C32e(0x9d2fc3c3),
2105
	C32e(0x67e1bebe), C32e(0x6aa23535), C32e(0x0bcc8888), C32e(0x5c392e2e),
2106
	C32e(0x3d579393), C32e(0xaaf25555), C32e(0xe382fcfc), C32e(0xf4477a7a),
2107
	C32e(0x8bacc8c8), C32e(0x6fe7baba), C32e(0x642b3232), C32e(0xd795e6e6),
2108
	C32e(0x9ba0c0c0), C32e(0x32981919), C32e(0x27d19e9e), C32e(0x5d7fa3a3),
2109
	C32e(0x88664444), C32e(0xa87e5454), C32e(0x76ab3b3b), C32e(0x16830b0b),
2110
	C32e(0x03ca8c8c), C32e(0x9529c7c7), C32e(0xd6d36b6b), C32e(0x503c2828),
2111
	C32e(0x5579a7a7), C32e(0x63e2bcbc), C32e(0x2c1d1616), C32e(0x4176adad),
2112
	C32e(0xad3bdbdb), C32e(0xc8566464), C32e(0xe84e7474), C32e(0x281e1414),
2113
	C32e(0x3fdb9292), C32e(0x180a0c0c), C32e(0x906c4848), C32e(0x6be4b8b8),
2114
	C32e(0x255d9f9f), C32e(0x616ebdbd), C32e(0x86ef4343), C32e(0x93a6c4c4),
2115
	C32e(0x72a83939), C32e(0x62a43131), C32e(0xbd37d3d3), C32e(0xff8bf2f2),
2116
	C32e(0xb132d5d5), C32e(0x0d438b8b), C32e(0xdc596e6e), C32e(0xafb7dada),
2117
	C32e(0x028c0101), C32e(0x7964b1b1), C32e(0x23d29c9c), C32e(0x92e04949),
2118
	C32e(0xabb4d8d8), C32e(0x43faacac), C32e(0xfd07f3f3), C32e(0x8525cfcf),
2119
	C32e(0x8fafcaca), C32e(0xf38ef4f4), C32e(0x8ee94747), C32e(0x20181010),
2120
	C32e(0xded56f6f), C32e(0xfb88f0f0), C32e(0x946f4a4a), C32e(0xb8725c5c),
2121
	C32e(0x70243838), C32e(0xaef15757), C32e(0xe6c77373), C32e(0x35519797),
2122
	C32e(0x8d23cbcb), C32e(0x597ca1a1), C32e(0xcb9ce8e8), C32e(0x7c213e3e),
2123
	C32e(0x37dd9696), C32e(0xc2dc6161), C32e(0x1a860d0d), C32e(0x1e850f0f),
2124
	C32e(0xdb90e0e0), C32e(0xf8427c7c), C32e(0xe2c47171), C32e(0x83aacccc),
2125
	C32e(0x3bd89090), C32e(0x0c050606), C32e(0xf501f7f7), C32e(0x38121c1c),
2126
	C32e(0x9fa3c2c2), C32e(0xd45f6a6a), C32e(0x47f9aeae), C32e(0xd2d06969),
2127
	C32e(0x2e911717), C32e(0x29589999), C32e(0x74273a3a), C32e(0x4eb92727),
2128
	C32e(0xa938d9d9), C32e(0xcd13ebeb), C32e(0x56b32b2b), C32e(0x44332222),
2129
	C32e(0xbfbbd2d2), C32e(0x4970a9a9), C32e(0x0e890707), C32e(0x66a73333),
2130
	C32e(0x5ab62d2d), C32e(0x78223c3c), C32e(0x2a921515), C32e(0x8920c9c9),
2131
	C32e(0x15498787), C32e(0x4fffaaaa), C32e(0xa0785050), C32e(0x517aa5a5),
2132
	C32e(0x068f0303), C32e(0xb2f85959), C32e(0x12800909), C32e(0x34171a1a),
2133
	C32e(0xcada6565), C32e(0xb531d7d7), C32e(0x13c68484), C32e(0xbbb8d0d0),
2134
	C32e(0x1fc38282), C32e(0x52b02929), C32e(0xb4775a5a), C32e(0x3c111e1e),
2135
	C32e(0xf6cb7b7b), C32e(0x4bfca8a8), C32e(0xdad66d6d), C32e(0x583a2c2c)
2136
};
2137

2138
static const sph_u32 T3dn[] = {
2139
	C32e(0x32f4a5f4), C32e(0x6f978497), C32e(0x5eb099b0), C32e(0x7a8c8d8c),
2140
	C32e(0xe8170d17), C32e(0x0adcbddc), C32e(0x16c8b1c8), C32e(0x6dfc54fc),
2141
	C32e(0x90f050f0), C32e(0x07050305), C32e(0x2ee0a9e0), C32e(0xd1877d87),
2142
	C32e(0xcc2b192b), C32e(0x13a662a6), C32e(0x7c31e631), C32e(0x59b59ab5),
2143
	C32e(0x40cf45cf), C32e(0xa3bc9dbc), C32e(0x49c040c0), C32e(0x68928792),
2144
	C32e(0xd03f153f), C32e(0x9426eb26), C32e(0xce40c940), C32e(0xe61d0b1d),
2145
	C32e(0x6e2fec2f), C32e(0x1aa967a9), C32e(0x431cfd1c), C32e(0x6025ea25),
2146
	C32e(0xf9dabfda), C32e(0x5102f702), C32e(0x45a196a1), C32e(0x76ed5bed),
2147
	C32e(0x285dc25d), C32e(0xc5241c24), C32e(0xd4e9aee9), C32e(0xf2be6abe),
2148
	C32e(0x82ee5aee), C32e(0xbdc341c3), C32e(0xf3060206), C32e(0x52d14fd1),
2149
	C32e(0x8ce45ce4), C32e(0x5607f407), C32e(0x8d5c345c), C32e(0xe1180818),
2150
	C32e(0x4cae93ae), C32e(0x3e957395), C32e(0x97f553f5), C32e(0x6b413f41),
2151
	C32e(0x1c140c14), C32e(0x63f652f6), C32e(0xe9af65af), C32e(0x7fe25ee2),
2152
	C32e(0x48782878), C32e(0xcff8a1f8), C32e(0x1b110f11), C32e(0xebc4b5c4),
2153
	C32e(0x151b091b), C32e(0x7e5a365a), C32e(0xadb69bb6), C32e(0x98473d47),
2154
	C32e(0xa76a266a), C32e(0xf5bb69bb), C32e(0x334ccd4c), C32e(0x50ba9fba),
2155
	C32e(0x3f2d1b2d), C32e(0xa4b99eb9), C32e(0xc49c749c), C32e(0x46722e72),
2156
	C32e(0x41772d77), C32e(0x11cdb2cd), C32e(0x9d29ee29), C32e(0x4d16fb16),
2157
	C32e(0xa501f601), C32e(0xa1d74dd7), C32e(0x14a361a3), C32e(0x3449ce49),
2158
	C32e(0xdf8d7b8d), C32e(0x9f423e42), C32e(0xcd937193), C32e(0xb1a297a2),
2159
	C32e(0xa204f504), C32e(0x01b868b8), C32e(0x00000000), C32e(0xb5742c74),
2160
	C32e(0xe0a060a0), C32e(0xc2211f21), C32e(0x3a43c843), C32e(0x9a2ced2c),
2161
	C32e(0x0dd9bed9), C32e(0x47ca46ca), C32e(0x1770d970), C32e(0xafdd4bdd),
2162
	C32e(0xed79de79), C32e(0xff67d467), C32e(0x9323e823), C32e(0x5bde4ade),
2163
	C32e(0x06bd6bbd), C32e(0xbb7e2a7e), C32e(0x7b34e534), C32e(0xd73a163a),
2164
	C32e(0xd254c554), C32e(0xf862d762), C32e(0x99ff55ff), C32e(0xb6a794a7),
2165
	C32e(0xc04acf4a), C32e(0xd9301030), C32e(0x0e0a060a), C32e(0x66988198),
2166
	C32e(0xab0bf00b), C32e(0xb4cc44cc), C32e(0xf0d5bad5), C32e(0x753ee33e),
2167
	C32e(0xac0ef30e), C32e(0x4419fe19), C32e(0xdb5bc05b), C32e(0x80858a85),
2168
	C32e(0xd3ecadec), C32e(0xfedfbcdf), C32e(0xa8d848d8), C32e(0xfd0c040c),
2169
	C32e(0x197adf7a), C32e(0x2f58c158), C32e(0x309f759f), C32e(0xe7a563a5),
2170
	C32e(0x70503050), C32e(0xcb2e1a2e), C32e(0xef120e12), C32e(0x08b76db7),
2171
	C32e(0x55d44cd4), C32e(0x243c143c), C32e(0x795f355f), C32e(0xb2712f71),
2172
	C32e(0x8638e138), C32e(0xc8fda2fd), C32e(0xc74fcc4f), C32e(0x654b394b),
2173
	C32e(0x6af957f9), C32e(0x580df20d), C32e(0x619d829d), C32e(0xb3c947c9),
2174
	C32e(0x27efacef), C32e(0x8832e732), C32e(0x4f7d2b7d), C32e(0x42a495a4),
2175
	C32e(0x3bfba0fb), C32e(0xaab398b3), C32e(0xf668d168), C32e(0x22817f81),
2176
	C32e(0xeeaa66aa), C32e(0xd6827e82), C32e(0xdde6abe6), C32e(0x959e839e),
2177
	C32e(0xc945ca45), C32e(0xbc7b297b), C32e(0x056ed36e), C32e(0x6c443c44),
2178
	C32e(0x2c8b798b), C32e(0x813de23d), C32e(0x31271d27), C32e(0x379a769a),
2179
	C32e(0x964d3b4d), C32e(0x9efa56fa), C32e(0xa6d24ed2), C32e(0x36221e22),
2180
	C32e(0xe476db76), C32e(0x121e0a1e), C32e(0xfcb46cb4), C32e(0x8f37e437),
2181
	C32e(0x78e75de7), C32e(0x0fb26eb2), C32e(0x692aef2a), C32e(0x35f1a6f1),
2182
	C32e(0xdae3a8e3), C32e(0xc6f7a4f7), C32e(0x8a593759), C32e(0x74868b86),
2183
	C32e(0x83563256), C32e(0x4ec543c5), C32e(0x85eb59eb), C32e(0x18c2b7c2),
2184
	C32e(0x8e8f8c8f), C32e(0x1dac64ac), C32e(0xf16dd26d), C32e(0x723be03b),
2185
	C32e(0x1fc7b4c7), C32e(0xb915fa15), C32e(0xfa090709), C32e(0xa06f256f),
2186
	C32e(0x20eaafea), C32e(0x7d898e89), C32e(0x6720e920), C32e(0x38281828),
2187
	C32e(0x0b64d564), C32e(0x73838883), C32e(0xfbb16fb1), C32e(0xca967296),
2188
	C32e(0x546c246c), C32e(0x5f08f108), C32e(0x2152c752), C32e(0x64f351f3),
2189
	C32e(0xae652365), C32e(0x25847c84), C32e(0x57bf9cbf), C32e(0x5d632163),
2190
	C32e(0xea7cdd7c), C32e(0x1e7fdc7f), C32e(0x9c918691), C32e(0x9b948594),
2191
	C32e(0x4bab90ab), C32e(0xbac642c6), C32e(0x2657c457), C32e(0x29e5aae5),
2192
	C32e(0xe373d873), C32e(0x090f050f), C32e(0xf4030103), C32e(0x2a361236),
2193
	C32e(0x3cfea3fe), C32e(0x8be15fe1), C32e(0xbe10f910), C32e(0x026bd06b),
2194
	C32e(0xbfa891a8), C32e(0x71e858e8), C32e(0x53692769), C32e(0xf7d0b9d0),
2195
	C32e(0x91483848), C32e(0xde351335), C32e(0xe5ceb3ce), C32e(0x77553355),
2196
	C32e(0x04d6bbd6), C32e(0x39907090), C32e(0x87808980), C32e(0xc1f2a7f2),
2197
	C32e(0xecc1b6c1), C32e(0x5a662266), C32e(0xb8ad92ad), C32e(0xa9602060),
2198
	C32e(0x5cdb49db), C32e(0xb01aff1a), C32e(0xd8887888), C32e(0x2b8e7a8e),
2199
	C32e(0x898a8f8a), C32e(0x4a13f813), C32e(0x929b809b), C32e(0x23391739),
2200
	C32e(0x1075da75), C32e(0x84533153), C32e(0xd551c651), C32e(0x03d3b8d3),
2201
	C32e(0xdc5ec35e), C32e(0xe2cbb0cb), C32e(0xc3997799), C32e(0x2d331133),
2202
	C32e(0x3d46cb46), C32e(0xb71ffc1f), C32e(0x0c61d661), C32e(0x624e3a4e)
2203
};
2204

2205
#define DECL_STATE_SMALL \
2206
	sph_u32 H[16];
2207

2208
#define READ_STATE_SMALL(sc)   do { \
2209
		memcpy(H, (sc)->state.narrow, sizeof H); \
2210
	} while (0)
2211

2212
#define WRITE_STATE_SMALL(sc)   do { \
2213
		memcpy((sc)->state.narrow, H, sizeof H); \
2214
	} while (0)
2215

2216
#define XCAT(x, y)    XCAT_(x, y)
2217
#define XCAT_(x, y)   x ## y
2218

2219
#define RSTT(d0, d1, a, b0, b1, b2, b3, b4, b5, b6, b7)   do { \
2220
		t[d0] = T0up[B32_0(a[b0])] \
2221
			^ T1up[B32_1(a[b1])] \
2222
			^ T2up[B32_2(a[b2])] \
2223
			^ T3up[B32_3(a[b3])] \
2224
			^ T0dn[B32_0(a[b4])] \
2225
			^ T1dn[B32_1(a[b5])] \
2226
			^ T2dn[B32_2(a[b6])] \
2227
			^ T3dn[B32_3(a[b7])]; \
2228
		t[d1] = T0dn[B32_0(a[b0])] \
2229
			^ T1dn[B32_1(a[b1])] \
2230
			^ T2dn[B32_2(a[b2])] \
2231
			^ T3dn[B32_3(a[b3])] \
2232
			^ T0up[B32_0(a[b4])] \
2233
			^ T1up[B32_1(a[b5])] \
2234
			^ T2up[B32_2(a[b6])] \
2235
			^ T3up[B32_3(a[b7])]; \
2236
	} while (0)
2237

2238
#define ROUND_SMALL_P(a, r)   do { \
2239
		sph_u32 t[16]; \
2240
		a[0x0] ^= PC32up(0x00, r); \
2241
		a[0x1] ^= PC32dn(0x00, r); \
2242
		a[0x2] ^= PC32up(0x10, r); \
2243
		a[0x3] ^= PC32dn(0x10, r); \
2244
		a[0x4] ^= PC32up(0x20, r); \
2245
		a[0x5] ^= PC32dn(0x20, r); \
2246
		a[0x6] ^= PC32up(0x30, r); \
2247
		a[0x7] ^= PC32dn(0x30, r); \
2248
		a[0x8] ^= PC32up(0x40, r); \
2249
		a[0x9] ^= PC32dn(0x40, r); \
2250
		a[0xA] ^= PC32up(0x50, r); \
2251
		a[0xB] ^= PC32dn(0x50, r); \
2252
		a[0xC] ^= PC32up(0x60, r); \
2253
		a[0xD] ^= PC32dn(0x60, r); \
2254
		a[0xE] ^= PC32up(0x70, r); \
2255
		a[0xF] ^= PC32dn(0x70, r); \
2256
		RSTT(0x0, 0x1, a, 0x0, 0x2, 0x4, 0x6, 0x9, 0xB, 0xD, 0xF); \
2257
		RSTT(0x2, 0x3, a, 0x2, 0x4, 0x6, 0x8, 0xB, 0xD, 0xF, 0x1); \
2258
		RSTT(0x4, 0x5, a, 0x4, 0x6, 0x8, 0xA, 0xD, 0xF, 0x1, 0x3); \
2259
		RSTT(0x6, 0x7, a, 0x6, 0x8, 0xA, 0xC, 0xF, 0x1, 0x3, 0x5); \
2260
		RSTT(0x8, 0x9, a, 0x8, 0xA, 0xC, 0xE, 0x1, 0x3, 0x5, 0x7); \
2261
		RSTT(0xA, 0xB, a, 0xA, 0xC, 0xE, 0x0, 0x3, 0x5, 0x7, 0x9); \
2262
		RSTT(0xC, 0xD, a, 0xC, 0xE, 0x0, 0x2, 0x5, 0x7, 0x9, 0xB); \
2263
		RSTT(0xE, 0xF, a, 0xE, 0x0, 0x2, 0x4, 0x7, 0x9, 0xB, 0xD); \
2264
		memcpy(a, t, sizeof t); \
2265
	} while (0)
2266

2267
#define ROUND_SMALL_Q(a, r)   do { \
2268
		sph_u32 t[16]; \
2269
		a[0x0] ^= QC32up(0x00, r); \
2270
		a[0x1] ^= QC32dn(0x00, r); \
2271
		a[0x2] ^= QC32up(0x10, r); \
2272
		a[0x3] ^= QC32dn(0x10, r); \
2273
		a[0x4] ^= QC32up(0x20, r); \
2274
		a[0x5] ^= QC32dn(0x20, r); \
2275
		a[0x6] ^= QC32up(0x30, r); \
2276
		a[0x7] ^= QC32dn(0x30, r); \
2277
		a[0x8] ^= QC32up(0x40, r); \
2278
		a[0x9] ^= QC32dn(0x40, r); \
2279
		a[0xA] ^= QC32up(0x50, r); \
2280
		a[0xB] ^= QC32dn(0x50, r); \
2281
		a[0xC] ^= QC32up(0x60, r); \
2282
		a[0xD] ^= QC32dn(0x60, r); \
2283
		a[0xE] ^= QC32up(0x70, r); \
2284
		a[0xF] ^= QC32dn(0x70, r); \
2285
		RSTT(0x0, 0x1, a, 0x2, 0x6, 0xA, 0xE, 0x1, 0x5, 0x9, 0xD); \
2286
		RSTT(0x2, 0x3, a, 0x4, 0x8, 0xC, 0x0, 0x3, 0x7, 0xB, 0xF); \
2287
		RSTT(0x4, 0x5, a, 0x6, 0xA, 0xE, 0x2, 0x5, 0x9, 0xD, 0x1); \
2288
		RSTT(0x6, 0x7, a, 0x8, 0xC, 0x0, 0x4, 0x7, 0xB, 0xF, 0x3); \
2289
		RSTT(0x8, 0x9, a, 0xA, 0xE, 0x2, 0x6, 0x9, 0xD, 0x1, 0x5); \
2290
		RSTT(0xA, 0xB, a, 0xC, 0x0, 0x4, 0x8, 0xB, 0xF, 0x3, 0x7); \
2291
		RSTT(0xC, 0xD, a, 0xE, 0x2, 0x6, 0xA, 0xD, 0x1, 0x5, 0x9); \
2292
		RSTT(0xE, 0xF, a, 0x0, 0x4, 0x8, 0xC, 0xF, 0x3, 0x7, 0xB); \
2293
		memcpy(a, t, sizeof t); \
2294
	} while (0)
2295

2296
#if SPH_SMALL_FOOTPRINT_GROESTL
2297

2298
#define PERM_SMALL_P(a)   do { \
2299
		int r; \
2300
		for (r = 0; r < 10; r ++) \
2301
			ROUND_SMALL_P(a, r); \
2302
	} while (0)
2303

2304
#define PERM_SMALL_Q(a)   do { \
2305
		int r; \
2306
		for (r = 0; r < 10; r ++) \
2307
			ROUND_SMALL_Q(a, r); \
2308
	} while (0)
2309

2310
#else
2311

2312
#define PERM_SMALL_P(a)   do { \
2313
		int r; \
2314
		for (r = 0; r < 10; r += 2) { \
2315
			ROUND_SMALL_P(a, r + 0); \
2316
			ROUND_SMALL_P(a, r + 1); \
2317
		} \
2318
	} while (0)
2319

2320
#define PERM_SMALL_Q(a)   do { \
2321
		int r; \
2322
		for (r = 0; r < 10; r += 2) { \
2323
			ROUND_SMALL_Q(a, r + 0); \
2324
			ROUND_SMALL_Q(a, r + 1); \
2325
		} \
2326
	} while (0)
2327

2328
#endif
2329

2330
#define COMPRESS_SMALL   do { \
2331
		sph_u32 g[16], m[16]; \
2332
		size_t u; \
2333
		for (u = 0; u < 16; u ++) { \
2334
			m[u] = dec32e_aligned(buf + (u << 2)); \
2335
			g[u] = m[u] ^ H[u]; \
2336
		} \
2337
		PERM_SMALL_P(g); \
2338
		PERM_SMALL_Q(m); \
2339
		for (u = 0; u < 16; u ++) \
2340
			H[u] ^= g[u] ^ m[u]; \
2341
	} while (0)
2342

2343
#define FINAL_SMALL   do { \
2344
		sph_u32 x[16]; \
2345
		size_t u; \
2346
		memcpy(x, H, sizeof x); \
2347
		PERM_SMALL_P(x); \
2348
		for (u = 0; u < 16; u ++) \
2349
			H[u] ^= x[u]; \
2350
	} while (0)
2351

2352
#define DECL_STATE_BIG \
2353
	sph_u32 H[32];
2354

2355
#define READ_STATE_BIG(sc)   do { \
2356
		memcpy(H, (sc)->state.narrow, sizeof H); \
2357
	} while (0)
2358

2359
#define WRITE_STATE_BIG(sc)   do { \
2360
		memcpy((sc)->state.narrow, H, sizeof H); \
2361
	} while (0)
2362

2363
#if SPH_SMALL_FOOTPRINT_GROESTL
2364

2365
#define RBTT(d0, d1, a, b0, b1, b2, b3, b4, b5, b6, b7)   do { \
2366
		sph_u32 fu2 = T0up[B32_2(a[b2])]; \
2367
		sph_u32 fd2 = T0dn[B32_2(a[b2])]; \
2368
		sph_u32 fu3 = T1up[B32_3(a[b3])]; \
2369
		sph_u32 fd3 = T1dn[B32_3(a[b3])]; \
2370
		sph_u32 fu6 = T0up[B32_2(a[b6])]; \
2371
		sph_u32 fd6 = T0dn[B32_2(a[b6])]; \
2372
		sph_u32 fu7 = T1up[B32_3(a[b7])]; \
2373
		sph_u32 fd7 = T1dn[B32_3(a[b7])]; \
2374
		t[d0] = T0up[B32_0(a[b0])] \
2375
			^ T1up[B32_1(a[b1])] \
2376
			^ R32u(fu2, fd2) \
2377
			^ R32u(fu3, fd3) \
2378
			^ T0dn[B32_0(a[b4])] \
2379
			^ T1dn[B32_1(a[b5])] \
2380
			^ R32d(fu6, fd6) \
2381
			^ R32d(fu7, fd7); \
2382
		t[d1] = T0dn[B32_0(a[b0])] \
2383
			^ T1dn[B32_1(a[b1])] \
2384
			^ R32d(fu2, fd2) \
2385
			^ R32d(fu3, fd3) \
2386
			^ T0up[B32_0(a[b4])] \
2387
			^ T1up[B32_1(a[b5])] \
2388
			^ R32u(fu6, fd6) \
2389
			^ R32u(fu7, fd7); \
2390
	} while (0)
2391

2392
#else
2393

2394
#define RBTT(d0, d1, a, b0, b1, b2, b3, b4, b5, b6, b7)   do { \
2395
		t[d0] = T0up[B32_0(a[b0])] \
2396
			^ T1up[B32_1(a[b1])] \
2397
			^ T2up[B32_2(a[b2])] \
2398
			^ T3up[B32_3(a[b3])] \
2399
			^ T0dn[B32_0(a[b4])] \
2400
			^ T1dn[B32_1(a[b5])] \
2401
			^ T2dn[B32_2(a[b6])] \
2402
			^ T3dn[B32_3(a[b7])]; \
2403
		t[d1] = T0dn[B32_0(a[b0])] \
2404
			^ T1dn[B32_1(a[b1])] \
2405
			^ T2dn[B32_2(a[b2])] \
2406
			^ T3dn[B32_3(a[b3])] \
2407
			^ T0up[B32_0(a[b4])] \
2408
			^ T1up[B32_1(a[b5])] \
2409
			^ T2up[B32_2(a[b6])] \
2410
			^ T3up[B32_3(a[b7])]; \
2411
	} while (0)
2412

2413
#endif
2414

2415
#if SPH_SMALL_FOOTPRINT_GROESTL
2416

2417
#define ROUND_BIG_P(a, r)   do { \
2418
		sph_u32 t[32]; \
2419
		size_t u; \
2420
		a[0x00] ^= PC32up(0x00, r); \
2421
		a[0x01] ^= PC32dn(0x00, r); \
2422
		a[0x02] ^= PC32up(0x10, r); \
2423
		a[0x03] ^= PC32dn(0x10, r); \
2424
		a[0x04] ^= PC32up(0x20, r); \
2425
		a[0x05] ^= PC32dn(0x20, r); \
2426
		a[0x06] ^= PC32up(0x30, r); \
2427
		a[0x07] ^= PC32dn(0x30, r); \
2428
		a[0x08] ^= PC32up(0x40, r); \
2429
		a[0x09] ^= PC32dn(0x40, r); \
2430
		a[0x0A] ^= PC32up(0x50, r); \
2431
		a[0x0B] ^= PC32dn(0x50, r); \
2432
		a[0x0C] ^= PC32up(0x60, r); \
2433
		a[0x0D] ^= PC32dn(0x60, r); \
2434
		a[0x0E] ^= PC32up(0x70, r); \
2435
		a[0x0F] ^= PC32dn(0x70, r); \
2436
		a[0x10] ^= PC32up(0x80, r); \
2437
		a[0x11] ^= PC32dn(0x80, r); \
2438
		a[0x12] ^= PC32up(0x90, r); \
2439
		a[0x13] ^= PC32dn(0x90, r); \
2440
		a[0x14] ^= PC32up(0xA0, r); \
2441
		a[0x15] ^= PC32dn(0xA0, r); \
2442
		a[0x16] ^= PC32up(0xB0, r); \
2443
		a[0x17] ^= PC32dn(0xB0, r); \
2444
		a[0x18] ^= PC32up(0xC0, r); \
2445
		a[0x19] ^= PC32dn(0xC0, r); \
2446
		a[0x1A] ^= PC32up(0xD0, r); \
2447
		a[0x1B] ^= PC32dn(0xD0, r); \
2448
		a[0x1C] ^= PC32up(0xE0, r); \
2449
		a[0x1D] ^= PC32dn(0xE0, r); \
2450
		a[0x1E] ^= PC32up(0xF0, r); \
2451
		a[0x1F] ^= PC32dn(0xF0, r); \
2452
		for (u = 0; u < 32; u += 8) { \
2453
			RBTT(u + 0x00, (u + 0x01) & 0x1F, a, \
2454
				u + 0x00, (u + 0x02) & 0x1F, \
2455
				(u + 0x04) & 0x1F, (u + 0x06) & 0x1F, \
2456
				(u + 0x09) & 0x1F, (u + 0x0B) & 0x1F, \
2457
				(u + 0x0D) & 0x1F, (u + 0x17) & 0x1F); \
2458
			RBTT(u + 0x02, (u + 0x03) & 0x1F, a, \
2459
				u + 0x02, (u + 0x04) & 0x1F, \
2460
				(u + 0x06) & 0x1F, (u + 0x08) & 0x1F, \
2461
				(u + 0x0B) & 0x1F, (u + 0x0D) & 0x1F, \
2462
				(u + 0x0F) & 0x1F, (u + 0x19) & 0x1F); \
2463
			RBTT(u + 0x04, (u + 0x05) & 0x1F, a, \
2464
				u + 0x04, (u + 0x06) & 0x1F, \
2465
				(u + 0x08) & 0x1F, (u + 0x0A) & 0x1F, \
2466
				(u + 0x0D) & 0x1F, (u + 0x0F) & 0x1F, \
2467
				(u + 0x11) & 0x1F, (u + 0x1B) & 0x1F); \
2468
			RBTT(u + 0x06, (u + 0x07) & 0x1F, a, \
2469
				u + 0x06, (u + 0x08) & 0x1F, \
2470
				(u + 0x0A) & 0x1F, (u + 0x0C) & 0x1F, \
2471
				(u + 0x0F) & 0x1F, (u + 0x11) & 0x1F, \
2472
				(u + 0x13) & 0x1F, (u + 0x1D) & 0x1F); \
2473
		} \
2474
		memcpy(a, t, sizeof t); \
2475
	} while (0)
2476

2477
#define ROUND_BIG_Q(a, r)   do { \
2478
		sph_u32 t[32]; \
2479
		size_t u; \
2480
		a[0x00] ^= QC32up(0x00, r); \
2481
		a[0x01] ^= QC32dn(0x00, r); \
2482
		a[0x02] ^= QC32up(0x10, r); \
2483
		a[0x03] ^= QC32dn(0x10, r); \
2484
		a[0x04] ^= QC32up(0x20, r); \
2485
		a[0x05] ^= QC32dn(0x20, r); \
2486
		a[0x06] ^= QC32up(0x30, r); \
2487
		a[0x07] ^= QC32dn(0x30, r); \
2488
		a[0x08] ^= QC32up(0x40, r); \
2489
		a[0x09] ^= QC32dn(0x40, r); \
2490
		a[0x0A] ^= QC32up(0x50, r); \
2491
		a[0x0B] ^= QC32dn(0x50, r); \
2492
		a[0x0C] ^= QC32up(0x60, r); \
2493
		a[0x0D] ^= QC32dn(0x60, r); \
2494
		a[0x0E] ^= QC32up(0x70, r); \
2495
		a[0x0F] ^= QC32dn(0x70, r); \
2496
		a[0x10] ^= QC32up(0x80, r); \
2497
		a[0x11] ^= QC32dn(0x80, r); \
2498
		a[0x12] ^= QC32up(0x90, r); \
2499
		a[0x13] ^= QC32dn(0x90, r); \
2500
		a[0x14] ^= QC32up(0xA0, r); \
2501
		a[0x15] ^= QC32dn(0xA0, r); \
2502
		a[0x16] ^= QC32up(0xB0, r); \
2503
		a[0x17] ^= QC32dn(0xB0, r); \
2504
		a[0x18] ^= QC32up(0xC0, r); \
2505
		a[0x19] ^= QC32dn(0xC0, r); \
2506
		a[0x1A] ^= QC32up(0xD0, r); \
2507
		a[0x1B] ^= QC32dn(0xD0, r); \
2508
		a[0x1C] ^= QC32up(0xE0, r); \
2509
		a[0x1D] ^= QC32dn(0xE0, r); \
2510
		a[0x1E] ^= QC32up(0xF0, r); \
2511
		a[0x1F] ^= QC32dn(0xF0, r); \
2512
		for (u = 0; u < 32; u += 8) { \
2513
			RBTT(u + 0x00, (u + 0x01) & 0x1F, a, \
2514
				(u + 0x02) & 0x1F, (u + 0x06) & 0x1F, \
2515
				(u + 0x0A) & 0x1F, (u + 0x16) & 0x1F, \
2516
				(u + 0x01) & 0x1F, (u + 0x05) & 0x1F, \
2517
				(u + 0x09) & 0x1F, (u + 0x0D) & 0x1F); \
2518
			RBTT(u + 0x02, (u + 0x03) & 0x1F, a, \
2519
				(u + 0x04) & 0x1F, (u + 0x08) & 0x1F, \
2520
				(u + 0x0C) & 0x1F, (u + 0x18) & 0x1F, \
2521
				(u + 0x03) & 0x1F, (u + 0x07) & 0x1F, \
2522
				(u + 0x0B) & 0x1F, (u + 0x0F) & 0x1F); \
2523
			RBTT(u + 0x04, (u + 0x05) & 0x1F, a, \
2524
				(u + 0x06) & 0x1F, (u + 0x0A) & 0x1F, \
2525
				(u + 0x0E) & 0x1F, (u + 0x1A) & 0x1F, \
2526
				(u + 0x05) & 0x1F, (u + 0x09) & 0x1F, \
2527
				(u + 0x0D) & 0x1F, (u + 0x11) & 0x1F); \
2528
			RBTT(u + 0x06, (u + 0x07) & 0x1F, a, \
2529
				(u + 0x08) & 0x1F, (u + 0x0C) & 0x1F, \
2530
				(u + 0x10) & 0x1F, (u + 0x1C) & 0x1F, \
2531
				(u + 0x07) & 0x1F, (u + 0x0B) & 0x1F, \
2532
				(u + 0x0F) & 0x1F, (u + 0x13) & 0x1F); \
2533
		} \
2534
		memcpy(a, t, sizeof t); \
2535
	} while (0)
2536

2537
#else
2538

2539
#define ROUND_BIG_P(a, r)   do { \
2540
		sph_u32 t[32]; \
2541
		a[0x00] ^= PC32up(0x00, r); \
2542
		a[0x01] ^= PC32dn(0x00, r); \
2543
		a[0x02] ^= PC32up(0x10, r); \
2544
		a[0x03] ^= PC32dn(0x10, r); \
2545
		a[0x04] ^= PC32up(0x20, r); \
2546
		a[0x05] ^= PC32dn(0x20, r); \
2547
		a[0x06] ^= PC32up(0x30, r); \
2548
		a[0x07] ^= PC32dn(0x30, r); \
2549
		a[0x08] ^= PC32up(0x40, r); \
2550
		a[0x09] ^= PC32dn(0x40, r); \
2551
		a[0x0A] ^= PC32up(0x50, r); \
2552
		a[0x0B] ^= PC32dn(0x50, r); \
2553
		a[0x0C] ^= PC32up(0x60, r); \
2554
		a[0x0D] ^= PC32dn(0x60, r); \
2555
		a[0x0E] ^= PC32up(0x70, r); \
2556
		a[0x0F] ^= PC32dn(0x70, r); \
2557
		a[0x10] ^= PC32up(0x80, r); \
2558
		a[0x11] ^= PC32dn(0x80, r); \
2559
		a[0x12] ^= PC32up(0x90, r); \
2560
		a[0x13] ^= PC32dn(0x90, r); \
2561
		a[0x14] ^= PC32up(0xA0, r); \
2562
		a[0x15] ^= PC32dn(0xA0, r); \
2563
		a[0x16] ^= PC32up(0xB0, r); \
2564
		a[0x17] ^= PC32dn(0xB0, r); \
2565
		a[0x18] ^= PC32up(0xC0, r); \
2566
		a[0x19] ^= PC32dn(0xC0, r); \
2567
		a[0x1A] ^= PC32up(0xD0, r); \
2568
		a[0x1B] ^= PC32dn(0xD0, r); \
2569
		a[0x1C] ^= PC32up(0xE0, r); \
2570
		a[0x1D] ^= PC32dn(0xE0, r); \
2571
		a[0x1E] ^= PC32up(0xF0, r); \
2572
		a[0x1F] ^= PC32dn(0xF0, r); \
2573
		RBTT(0x00, 0x01, a, \
2574
			0x00, 0x02, 0x04, 0x06, 0x09, 0x0B, 0x0D, 0x17); \
2575
		RBTT(0x02, 0x03, a, \
2576
			0x02, 0x04, 0x06, 0x08, 0x0B, 0x0D, 0x0F, 0x19); \
2577
		RBTT(0x04, 0x05, a, \
2578
			0x04, 0x06, 0x08, 0x0A, 0x0D, 0x0F, 0x11, 0x1B); \
2579
		RBTT(0x06, 0x07, a, \
2580
			0x06, 0x08, 0x0A, 0x0C, 0x0F, 0x11, 0x13, 0x1D); \
2581
		RBTT(0x08, 0x09, a, \
2582
			0x08, 0x0A, 0x0C, 0x0E, 0x11, 0x13, 0x15, 0x1F); \
2583
		RBTT(0x0A, 0x0B, a, \
2584
			0x0A, 0x0C, 0x0E, 0x10, 0x13, 0x15, 0x17, 0x01); \
2585
		RBTT(0x0C, 0x0D, a, \
2586
			0x0C, 0x0E, 0x10, 0x12, 0x15, 0x17, 0x19, 0x03); \
2587
		RBTT(0x0E, 0x0F, a, \
2588
			0x0E, 0x10, 0x12, 0x14, 0x17, 0x19, 0x1B, 0x05); \
2589
		RBTT(0x10, 0x11, a, \
2590
			0x10, 0x12, 0x14, 0x16, 0x19, 0x1B, 0x1D, 0x07); \
2591
		RBTT(0x12, 0x13, a, \
2592
			0x12, 0x14, 0x16, 0x18, 0x1B, 0x1D, 0x1F, 0x09); \
2593
		RBTT(0x14, 0x15, a, \
2594
			0x14, 0x16, 0x18, 0x1A, 0x1D, 0x1F, 0x01, 0x0B); \
2595
		RBTT(0x16, 0x17, a, \
2596
			0x16, 0x18, 0x1A, 0x1C, 0x1F, 0x01, 0x03, 0x0D); \
2597
		RBTT(0x18, 0x19, a, \
2598
			0x18, 0x1A, 0x1C, 0x1E, 0x01, 0x03, 0x05, 0x0F); \
2599
		RBTT(0x1A, 0x1B, a, \
2600
			0x1A, 0x1C, 0x1E, 0x00, 0x03, 0x05, 0x07, 0x11); \
2601
		RBTT(0x1C, 0x1D, a, \
2602
			0x1C, 0x1E, 0x00, 0x02, 0x05, 0x07, 0x09, 0x13); \
2603
		RBTT(0x1E, 0x1F, a, \
2604
			0x1E, 0x00, 0x02, 0x04, 0x07, 0x09, 0x0B, 0x15); \
2605
		memcpy(a, t, sizeof t); \
2606
	} while (0)
2607

2608
#define ROUND_BIG_Q(a, r)   do { \
2609
		sph_u32 t[32]; \
2610
		a[0x00] ^= QC32up(0x00, r); \
2611
		a[0x01] ^= QC32dn(0x00, r); \
2612
		a[0x02] ^= QC32up(0x10, r); \
2613
		a[0x03] ^= QC32dn(0x10, r); \
2614
		a[0x04] ^= QC32up(0x20, r); \
2615
		a[0x05] ^= QC32dn(0x20, r); \
2616
		a[0x06] ^= QC32up(0x30, r); \
2617
		a[0x07] ^= QC32dn(0x30, r); \
2618
		a[0x08] ^= QC32up(0x40, r); \
2619
		a[0x09] ^= QC32dn(0x40, r); \
2620
		a[0x0A] ^= QC32up(0x50, r); \
2621
		a[0x0B] ^= QC32dn(0x50, r); \
2622
		a[0x0C] ^= QC32up(0x60, r); \
2623
		a[0x0D] ^= QC32dn(0x60, r); \
2624
		a[0x0E] ^= QC32up(0x70, r); \
2625
		a[0x0F] ^= QC32dn(0x70, r); \
2626
		a[0x10] ^= QC32up(0x80, r); \
2627
		a[0x11] ^= QC32dn(0x80, r); \
2628
		a[0x12] ^= QC32up(0x90, r); \
2629
		a[0x13] ^= QC32dn(0x90, r); \
2630
		a[0x14] ^= QC32up(0xA0, r); \
2631
		a[0x15] ^= QC32dn(0xA0, r); \
2632
		a[0x16] ^= QC32up(0xB0, r); \
2633
		a[0x17] ^= QC32dn(0xB0, r); \
2634
		a[0x18] ^= QC32up(0xC0, r); \
2635
		a[0x19] ^= QC32dn(0xC0, r); \
2636
		a[0x1A] ^= QC32up(0xD0, r); \
2637
		a[0x1B] ^= QC32dn(0xD0, r); \
2638
		a[0x1C] ^= QC32up(0xE0, r); \
2639
		a[0x1D] ^= QC32dn(0xE0, r); \
2640
		a[0x1E] ^= QC32up(0xF0, r); \
2641
		a[0x1F] ^= QC32dn(0xF0, r); \
2642
		RBTT(0x00, 0x01, a, \
2643
			0x02, 0x06, 0x0A, 0x16, 0x01, 0x05, 0x09, 0x0D); \
2644
		RBTT(0x02, 0x03, a, \
2645
			0x04, 0x08, 0x0C, 0x18, 0x03, 0x07, 0x0B, 0x0F); \
2646
		RBTT(0x04, 0x05, a, \
2647
			0x06, 0x0A, 0x0E, 0x1A, 0x05, 0x09, 0x0D, 0x11); \
2648
		RBTT(0x06, 0x07, a, \
2649
			0x08, 0x0C, 0x10, 0x1C, 0x07, 0x0B, 0x0F, 0x13); \
2650
		RBTT(0x08, 0x09, a, \
2651
			0x0A, 0x0E, 0x12, 0x1E, 0x09, 0x0D, 0x11, 0x15); \
2652
		RBTT(0x0A, 0x0B, a, \
2653
			0x0C, 0x10, 0x14, 0x00, 0x0B, 0x0F, 0x13, 0x17); \
2654
		RBTT(0x0C, 0x0D, a, \
2655
			0x0E, 0x12, 0x16, 0x02, 0x0D, 0x11, 0x15, 0x19); \
2656
		RBTT(0x0E, 0x0F, a, \
2657
			0x10, 0x14, 0x18, 0x04, 0x0F, 0x13, 0x17, 0x1B); \
2658
		RBTT(0x10, 0x11, a, \
2659
			0x12, 0x16, 0x1A, 0x06, 0x11, 0x15, 0x19, 0x1D); \
2660
		RBTT(0x12, 0x13, a, \
2661
			0x14, 0x18, 0x1C, 0x08, 0x13, 0x17, 0x1B, 0x1F); \
2662
		RBTT(0x14, 0x15, a, \
2663
			0x16, 0x1A, 0x1E, 0x0A, 0x15, 0x19, 0x1D, 0x01); \
2664
		RBTT(0x16, 0x17, a, \
2665
			0x18, 0x1C, 0x00, 0x0C, 0x17, 0x1B, 0x1F, 0x03); \
2666
		RBTT(0x18, 0x19, a, \
2667
			0x1A, 0x1E, 0x02, 0x0E, 0x19, 0x1D, 0x01, 0x05); \
2668
		RBTT(0x1A, 0x1B, a, \
2669
			0x1C, 0x00, 0x04, 0x10, 0x1B, 0x1F, 0x03, 0x07); \
2670
		RBTT(0x1C, 0x1D, a, \
2671
			0x1E, 0x02, 0x06, 0x12, 0x1D, 0x01, 0x05, 0x09); \
2672
		RBTT(0x1E, 0x1F, a, \
2673
			0x00, 0x04, 0x08, 0x14, 0x1F, 0x03, 0x07, 0x0B); \
2674
		memcpy(a, t, sizeof t); \
2675
	} while (0)
2676

2677
#endif
2678

2679
#if SPH_SMALL_FOOTPRINT_GROESTL
2680

2681
#define PERM_BIG_P(a)   do { \
2682
		int r; \
2683
		for (r = 0; r < 14; r ++) \
2684
			ROUND_BIG_P(a, r); \
2685
	} while (0)
2686

2687
#define PERM_BIG_Q(a)   do { \
2688
		int r; \
2689
		for (r = 0; r < 14; r ++) \
2690
			ROUND_BIG_Q(a, r); \
2691
	} while (0)
2692

2693
#else
2694

2695
#define PERM_BIG_P(a)   do { \
2696
		int r; \
2697
		for (r = 0; r < 14; r += 2) { \
2698
			ROUND_BIG_P(a, r + 0); \
2699
			ROUND_BIG_P(a, r + 1); \
2700
		} \
2701
	} while (0)
2702

2703
#define PERM_BIG_Q(a)   do { \
2704
		int r; \
2705
		for (r = 0; r < 14; r += 2) { \
2706
			ROUND_BIG_Q(a, r + 0); \
2707
			ROUND_BIG_Q(a, r + 1); \
2708
		} \
2709
	} while (0)
2710

2711
#endif
2712

2713
#define COMPRESS_BIG   do { \
2714
		sph_u32 g[32], m[32]; \
2715
		size_t u; \
2716
		for (u = 0; u < 32; u ++) { \
2717
			m[u] = dec32e_aligned(buf + (u << 2)); \
2718
			g[u] = m[u] ^ H[u]; \
2719
		} \
2720
		PERM_BIG_P(g); \
2721
		PERM_BIG_Q(m); \
2722
		for (u = 0; u < 32; u ++) \
2723
			H[u] ^= g[u] ^ m[u]; \
2724
	} while (0)
2725

2726
#define FINAL_BIG   do { \
2727
		sph_u32 x[32]; \
2728
		size_t u; \
2729
		memcpy(x, H, sizeof x); \
2730
		PERM_BIG_P(x); \
2731
		for (u = 0; u < 32; u ++) \
2732
			H[u] ^= x[u]; \
2733
	} while (0)
2734

2735
#endif
2736

2737
static void
2738
groestl_small_init(sph_groestl_small_context *sc, unsigned out_size)
2739
{
2740
	size_t u;
2741

2742
	sc->ptr = 0;
2743
#if SPH_GROESTL_64
2744
	for (u = 0; u < 7; u ++)
2745
		sc->state.wide[u] = 0;
2746
#if USE_LE
2747
	sc->state.wide[7] = ((sph_u64)(out_size & 0xFF) << 56)
2748
		| ((sph_u64)(out_size & 0xFF00) << 40);
2749
#else
2750
	sc->state.wide[7] = (sph_u64)out_size;
2751
#endif
2752
#else
2753
	for (u = 0; u < 15; u ++)
2754
		sc->state.narrow[u] = 0;
2755
#if USE_LE
2756
	sc->state.narrow[15] = ((sph_u32)(out_size & 0xFF) << 24)
2757
		| ((sph_u32)(out_size & 0xFF00) << 8);
2758
#else
2759
	sc->state.narrow[15] = (sph_u32)out_size;
2760
#endif
2761
#endif
2762
#if SPH_64
2763
	sc->count = 0;
2764
#else
2765
	sc->count_high = 0;
2766
	sc->count_low = 0;
2767
#endif
2768
}
2769

2770
static void
2771
groestl_small_core(sph_groestl_small_context *sc, const void *data, size_t len)
2772
{
2773
	unsigned char *buf;
2774
	size_t ptr;
2775
	DECL_STATE_SMALL
2776

2777
	buf = sc->buf;
2778
	ptr = sc->ptr;
2779
	if (len < (sizeof sc->buf) - ptr) {
2780
		memcpy(buf + ptr, data, len);
2781
		ptr += len;
2782
		sc->ptr = ptr;
2783
		return;
2784
	}
2785

2786
	READ_STATE_SMALL(sc);
2787
	while (len > 0) {
2788
		size_t clen;
2789

2790
		clen = (sizeof sc->buf) - ptr;
2791
		if (clen > len)
2792
			clen = len;
2793
		memcpy(buf + ptr, data, clen);
2794
		ptr += clen;
2795
		data = (const unsigned char *)data + clen;
2796
		len -= clen;
2797
		if (ptr == sizeof sc->buf) {
2798
			COMPRESS_SMALL;
2799
#if SPH_64
2800
			sc->count ++;
2801
#else
2802
			if ((sc->count_low = SPH_T32(sc->count_low + 1)) == 0)
2803
				sc->count_high = SPH_T32(sc->count_high + 1);
2804
#endif
2805
			ptr = 0;
2806
		}
2807
	}
2808
	WRITE_STATE_SMALL(sc);
2809
	sc->ptr = ptr;
2810
}
2811

2812
static void
2813
groestl_small_close(sph_groestl_small_context *sc,
2814
	unsigned ub, unsigned n, void *dst, size_t out_len)
2815
{
2816
	unsigned char pad[72];
2817
	size_t u, ptr, pad_len;
2818
#if SPH_64
2819
	sph_u64 count;
2820
#else
2821
	sph_u32 count_high, count_low;
2822
#endif
2823
	unsigned z;
2824
	DECL_STATE_SMALL
2825

2826
	ptr = sc->ptr;
2827
	z = 0x80 >> n;
2828
	pad[0] = ((ub & -z) | z) & 0xFF;
2829
	if (ptr < 56) {
2830
		pad_len = 64 - ptr;
2831
#if SPH_64
2832
		count = SPH_T64(sc->count + 1);
2833
#else
2834
		count_low = SPH_T32(sc->count_low + 1);
2835
		count_high = SPH_T32(sc->count_high);
2836
		if (count_low == 0)
2837
			count_high = SPH_T32(count_high + 1);
2838
#endif
2839
	} else {
2840
		pad_len = 128 - ptr;
2841
#if SPH_64
2842
		count = SPH_T64(sc->count + 2);
2843
#else
2844
		count_low = SPH_T32(sc->count_low + 2);
2845
		count_high = SPH_T32(sc->count_high);
2846
		if (count_low <= 1)
2847
			count_high = SPH_T32(count_high + 1);
2848
#endif
2849
	}
2850
	memset(pad + 1, 0, pad_len - 9);
2851
#if SPH_64
2852
	sph_enc64be(pad + pad_len - 8, count);
2853
#else
2854
	sph_enc64be(pad + pad_len - 8, count_high);
2855
	sph_enc64be(pad + pad_len - 4, count_low);
2856
#endif
2857
	groestl_small_core(sc, pad, pad_len);
2858
	READ_STATE_SMALL(sc);
2859
	FINAL_SMALL;
2860
#if SPH_GROESTL_64
2861
	for (u = 0; u < 4; u ++)
2862
		enc64e(pad + (u << 3), H[u + 4]);
2863
#else
2864
	for (u = 0; u < 8; u ++)
2865
		enc32e(pad + (u << 2), H[u + 8]);
2866
#endif
2867
	memcpy(dst, pad + 32 - out_len, out_len);
2868
	groestl_small_init(sc, (unsigned)out_len << 3);
2869
}
2870

2871
static void
2872
groestl_big_init(sph_groestl_big_context *sc, unsigned out_size)
2873
{
2874
	size_t u;
2875

2876
	sc->ptr = 0;
2877
#if SPH_GROESTL_64
2878
	for (u = 0; u < 15; u ++)
2879
		sc->state.wide[u] = 0;
2880
#if USE_LE
2881
	sc->state.wide[15] = ((sph_u64)(out_size & 0xFF) << 56)
2882
		| ((sph_u64)(out_size & 0xFF00) << 40);
2883
#else
2884
	sc->state.wide[15] = (sph_u64)out_size;
2885
#endif
2886
#else
2887
	for (u = 0; u < 31; u ++)
2888
		sc->state.narrow[u] = 0;
2889
#if USE_LE
2890
	sc->state.narrow[31] = ((sph_u32)(out_size & 0xFF) << 24)
2891
		| ((sph_u32)(out_size & 0xFF00) << 8);
2892
#else
2893
	sc->state.narrow[31] = (sph_u32)out_size;
2894
#endif
2895
#endif
2896
#if SPH_64
2897
	sc->count = 0;
2898
#else
2899
	sc->count_high = 0;
2900
	sc->count_low = 0;
2901
#endif
2902
}
2903

2904
static void
2905
groestl_big_core(sph_groestl_big_context *sc, const void *data, size_t len)
2906
{
2907
	unsigned char *buf;
2908
	size_t ptr;
2909
	DECL_STATE_BIG
2910

2911
	buf = sc->buf;
2912
	ptr = sc->ptr;
2913
	if (len < (sizeof sc->buf) - ptr) {
2914
		memcpy(buf + ptr, data, len);
2915
		ptr += len;
2916
		sc->ptr = ptr;
2917
		return;
2918
	}
2919

2920
	READ_STATE_BIG(sc);
2921
	while (len > 0) {
2922
		size_t clen;
2923

2924
		clen = (sizeof sc->buf) - ptr;
2925
		if (clen > len)
2926
			clen = len;
2927
		memcpy(buf + ptr, data, clen);
2928
		ptr += clen;
2929
		data = (const unsigned char *)data + clen;
2930
		len -= clen;
2931
		if (ptr == sizeof sc->buf) {
2932
			COMPRESS_BIG;
2933
#if SPH_64
2934
			sc->count ++;
2935
#else
2936
			if ((sc->count_low = SPH_T32(sc->count_low + 1)) == 0)
2937
				sc->count_high = SPH_T32(sc->count_high + 1);
2938
#endif
2939
			ptr = 0;
2940
		}
2941
	}
2942
	WRITE_STATE_BIG(sc);
2943
	sc->ptr = ptr;
2944
}
2945

2946
static void
2947
groestl_big_close(sph_groestl_big_context *sc,
2948
	unsigned ub, unsigned n, void *dst, size_t out_len)
2949
{
2950
	unsigned char pad[136];
2951
	size_t ptr, pad_len, u;
2952
#if SPH_64
2953
	sph_u64 count;
2954
#else
2955
	sph_u32 count_high, count_low;
2956
#endif
2957
	unsigned z;
2958
	DECL_STATE_BIG
2959

2960
	ptr = sc->ptr;
2961
	z = 0x80 >> n;
2962
	pad[0] = ((ub & -z) | z) & 0xFF;
2963
	if (ptr < 120) {
2964
		pad_len = 128 - ptr;
2965
#if SPH_64
2966
		count = SPH_T64(sc->count + 1);
2967
#else
2968
		count_low = SPH_T32(sc->count_low + 1);
2969
		count_high = SPH_T32(sc->count_high);
2970
		if (count_low == 0)
2971
			count_high = SPH_T32(count_high + 1);
2972
#endif
2973
	} else {
2974
		pad_len = 256 - ptr;
2975
#if SPH_64
2976
		count = SPH_T64(sc->count + 2);
2977
#else
2978
		count_low = SPH_T32(sc->count_low + 2);
2979
		count_high = SPH_T32(sc->count_high);
2980
		if (count_low <= 1)
2981
			count_high = SPH_T32(count_high + 1);
2982
#endif
2983
	}
2984
	memset(pad + 1, 0, pad_len - 9);
2985
#if SPH_64
2986
	sph_enc64be(pad + pad_len - 8, count);
2987
#else
2988
	sph_enc64be(pad + pad_len - 8, count_high);
2989
	sph_enc64be(pad + pad_len - 4, count_low);
2990
#endif
2991
	groestl_big_core(sc, pad, pad_len);
2992
	READ_STATE_BIG(sc);
2993
	FINAL_BIG;
2994
#if SPH_GROESTL_64
2995
	for (u = 0; u < 8; u ++)
2996
		enc64e(pad + (u << 3), H[u + 8]);
2997
#else
2998
	for (u = 0; u < 16; u ++)
2999
		enc32e(pad + (u << 2), H[u + 16]);
3000
#endif
3001
	memcpy(dst, pad + 64 - out_len, out_len);
3002
	groestl_big_init(sc, (unsigned)out_len << 3);
3003
}
3004

3005
/* see sph_groestl.h */
3006
void
3007
sph_groestl224_init(void *cc)
3008
{
3009
	groestl_small_init(cc, 224);
3010
}
3011

3012
/* see sph_groestl.h */
3013
void
3014
sph_groestl224(void *cc, const void *data, size_t len)
3015
{
3016
	groestl_small_core(cc, data, len);
3017
}
3018

3019
/* see sph_groestl.h */
3020
void
3021
sph_groestl224_close(void *cc, void *dst)
3022
{
3023
	groestl_small_close(cc, 0, 0, dst, 28);
3024
}
3025

3026
/* see sph_groestl.h */
3027
void
3028
sph_groestl224_addbits_and_close(void *cc, unsigned ub, unsigned n, void *dst)
3029
{
3030
	groestl_small_close(cc, ub, n, dst, 28);
3031
}
3032

3033
/* see sph_groestl.h */
3034
void
3035
sph_groestl256_init(void *cc)
3036
{
3037
	groestl_small_init(cc, 256);
3038
}
3039

3040
/* see sph_groestl.h */
3041
void
3042
sph_groestl256(void *cc, const void *data, size_t len)
3043
{
3044
	groestl_small_core(cc, data, len);
3045
}
3046

3047
/* see sph_groestl.h */
3048
void
3049
sph_groestl256_close(void *cc, void *dst)
3050
{
3051
	groestl_small_close(cc, 0, 0, dst, 32);
3052
}
3053

3054
/* see sph_groestl.h */
3055
void
3056
sph_groestl256_addbits_and_close(void *cc, unsigned ub, unsigned n, void *dst)
3057
{
3058
	groestl_small_close(cc, ub, n, dst, 32);
3059
}
3060

3061
/* see sph_groestl.h */
3062
void
3063
sph_groestl384_init(void *cc)
3064
{
3065
	groestl_big_init(cc, 384);
3066
}
3067

3068
/* see sph_groestl.h */
3069
void
3070
sph_groestl384(void *cc, const void *data, size_t len)
3071
{
3072
	groestl_big_core(cc, data, len);
3073
}
3074

3075
/* see sph_groestl.h */
3076
void
3077
sph_groestl384_close(void *cc, void *dst)
3078
{
3079
	groestl_big_close(cc, 0, 0, dst, 48);
3080
}
3081

3082
/* see sph_groestl.h */
3083
void
3084
sph_groestl384_addbits_and_close(void *cc, unsigned ub, unsigned n, void *dst)
3085
{
3086
	groestl_big_close(cc, ub, n, dst, 48);
3087
}
3088

3089
/* see sph_groestl.h */
3090
void
3091
sph_groestl512_init(void *cc)
3092
{
3093
	groestl_big_init(cc, 512);
3094
}
3095

3096
/* see sph_groestl.h */
3097
void
3098
sph_groestl512(void *cc, const void *data, size_t len)
3099
{
3100
	groestl_big_core(cc, data, len);
3101
}
3102

3103
/* see sph_groestl.h */
3104
void
3105
sph_groestl512_close(void *cc, void *dst)
3106
{
3107
	groestl_big_close(cc, 0, 0, dst, 64);
3108
}
3109

3110
/* see sph_groestl.h */
3111
void
3112
sph_groestl512_addbits_and_close(void *cc, unsigned ub, unsigned n, void *dst)
3113
{
3114
	groestl_big_close(cc, ub, n, dst, 64);
3115
}
3116

3117
#ifdef __cplusplus
3118
}
3119
#endif
3120

3121
Product

Resources

Company
