~ubuntu-branches/ubuntu/raring/fftw3/raring-proposed

« back to all changes in this revision

Viewing changes to dft/codelets/standard/n1_64.c

  • Committer: Bazaar Package Importer
  • Author(s): Paul Brossier
  • Date: 2006-05-31 13:44:05 UTC
  • mfrom: (1.1.1 upstream)
  • Revision ID: james.westby@ubuntu.com-20060531134405-ol9hrbg6bh81sg0c
Tags: 3.1.1-1
* New upstream release (closes: #350327, #338487, #338501)
* Add --enable-portable-binary to use -mtune instead of -march
* Use --with-gcc-arch=G5 / pentium4 on powerpc / i386
* Updated Standards-Version

Show diffs side-by-side

added added

removed removed

Lines of Context:
 
1
/*
 
2
 * Copyright (c) 2003, 2006 Matteo Frigo
 
3
 * Copyright (c) 2003, 2006 Massachusetts Institute of Technology
 
4
 *
 
5
 * This program is free software; you can redistribute it and/or modify
 
6
 * it under the terms of the GNU General Public License as published by
 
7
 * the Free Software Foundation; either version 2 of the License, or
 
8
 * (at your option) any later version.
 
9
 *
 
10
 * This program is distributed in the hope that it will be useful,
 
11
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 
12
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 
13
 * GNU General Public License for more details.
 
14
 *
 
15
 * You should have received a copy of the GNU General Public License
 
16
 * along with this program; if not, write to the Free Software
 
17
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 
18
 *
 
19
 */
 
20
 
 
21
/* This file was automatically generated --- DO NOT EDIT */
 
22
/* Generated on Tue Mar  7 09:18:20 EST 2006 */
 
23
 
 
24
#include "codelet-dft.h"
 
25
 
 
26
#ifdef HAVE_FMA
 
27
 
 
28
/* Generated by: ../../../genfft/gen_notw -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 64 -name n1_64 -include n.h */
 
29
 
 
30
/*
 
31
 * This function contains 912 FP additions, 392 FP multiplications,
 
32
 * (or, 520 additions, 0 multiplications, 392 fused multiply/add),
 
33
 * 202 stack variables, and 256 memory accesses
 
34
 */
 
35
/*
 
36
 * Generator Id's : 
 
37
 * $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
 
38
 * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
 
39
 * $Id: gen_notw.ml,v 1.30 2006-02-12 23:34:12 athena Exp $
 
40
 */
 
41
 
 
42
#include "n.h"
 
43
 
 
44
static void n1_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
 
45
{
 
46
     DK(KP956940335, +0.956940335732208864935797886980269969482849206);
 
47
     DK(KP881921264, +0.881921264348355029712756863660388349508442621);
 
48
     DK(KP534511135, +0.534511135950791641089685961295362908582039528);
 
49
     DK(KP303346683, +0.303346683607342391675883946941299872384187453);
 
50
     DK(KP995184726, +0.995184726672196886244836953109479921575474869);
 
51
     DK(KP773010453, +0.773010453362736960810906609758469800971041293);
 
52
     DK(KP820678790, +0.820678790828660330972281985331011598767386482);
 
53
     DK(KP098491403, +0.098491403357164253077197521291327432293052451);
 
54
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
 
55
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
 
56
     DK(KP668178637, +0.668178637919298919997757686523080761552472251);
 
57
     DK(KP198912367, +0.198912367379658006911597622644676228597850501);
 
58
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
59
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
60
     DK(KP414213562, +0.414213562373095048801688724209698078569671875);
 
61
     INT i;
 
62
     for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
 
63
          E T9b, T9e;
 
64
          {
 
65
               E T7B, T37, T5Z, T8F, Td9, Tf, TcB, TbB, T7C, T62, TdH, T2i, Tcb, Tah, T8G;
 
66
               E T3e, Tu, TdI, Tak, TbC, TbD, Tan, Tda, T2x, T65, T3m, T8I, T7G, T8J, T7J;
 
67
               E T64, T3t, Tdd, TK, Tce, Tas, Tcf, Tav, Tdc, T2N, T6G, T3G, T9k, T7O, T9l;
 
68
               E T7R, T6H, T3N, TdA, T1L, Tct, Tbs, Teo, Tdx, T6Y, T5j, T6V, T5Q, T9z, T8y;
 
69
               E Tcw, Tbb, T9C, T8n, Tdf, TZ, Tch, Taz, Tci, TaC, Tdg, T32, T6J, T3Z, T9n;
 
70
               E T7V, T9o, T7Y, T6K, T46, Tdp, T1g, Tcm, Tb1, Tej, Tdm, T6R, T4q, T6O, T4X;
 
71
               E T9s, T8f, Tcp, TaK, T9v, T84, Tdn, T1v, Tcq, Tb4, Tek, Tds, T6P, T4N, T6S;
 
72
               E T50, T9w, T8i, Tcn, TaV, T9t, T8b, Tdy, T20, Tcx, Tbv, Tep, TdD, T8q, T6W;
 
73
               E T5G, T6Z, T5T, T8t, T9D, T8B, Tcu, Tbm, T8l, T8m;
 
74
               {
 
75
                    E T3s, T3p, T3M, T3J;
 
76
                    {
 
77
                         E Taf, T3d, T3a, Tag;
 
78
                         {
 
79
                              E T35, T3, T5Y, T26, T5X, T6, T36, T29, Tb, T39, Ta, T38, T2d, Tc, T2e;
 
80
                              E T2f;
 
81
                              {
 
82
                                   E T4, T5, T27, T28;
 
83
                                   {
 
84
                                        E T1, T2, T24, T25;
 
85
                                        T1 = ri[0];
 
86
                                        T2 = ri[WS(is, 32)];
 
87
                                        T24 = ii[0];
 
88
                                        T25 = ii[WS(is, 32)];
 
89
                                        T4 = ri[WS(is, 16)];
 
90
                                        T35 = T1 - T2;
 
91
                                        T3 = T1 + T2;
 
92
                                        T5Y = T24 - T25;
 
93
                                        T26 = T24 + T25;
 
94
                                        T5 = ri[WS(is, 48)];
 
95
                                        T27 = ii[WS(is, 16)];
 
96
                                        T28 = ii[WS(is, 48)];
 
97
                                   }
 
98
                                   {
 
99
                                        E T8, T9, T2b, T2c;
 
100
                                        T8 = ri[WS(is, 8)];
 
101
                                        T5X = T4 - T5;
 
102
                                        T6 = T4 + T5;
 
103
                                        T36 = T27 - T28;
 
104
                                        T29 = T27 + T28;
 
105
                                        T9 = ri[WS(is, 40)];
 
106
                                        T2b = ii[WS(is, 8)];
 
107
                                        T2c = ii[WS(is, 40)];
 
108
                                        Tb = ri[WS(is, 56)];
 
109
                                        T39 = T8 - T9;
 
110
                                        Ta = T8 + T9;
 
111
                                        T38 = T2b - T2c;
 
112
                                        T2d = T2b + T2c;
 
113
                                        Tc = ri[WS(is, 24)];
 
114
                                        T2e = ii[WS(is, 56)];
 
115
                                        T2f = ii[WS(is, 24)];
 
116
                                   }
 
117
                              }
 
118
                              {
 
119
                                   E T3b, T3c, T2g, T7, Te, Tbz, Td;
 
120
                                   T7B = T35 + T36;
 
121
                                   T37 = T35 - T36;
 
122
                                   T3b = Tb - Tc;
 
123
                                   Td = Tb + Tc;
 
124
                                   T3c = T2e - T2f;
 
125
                                   T2g = T2e + T2f;
 
126
                                   T5Z = T5X + T5Y;
 
127
                                   T8F = T5Y - T5X;
 
128
                                   Taf = T3 - T6;
 
129
                                   T7 = T3 + T6;
 
130
                                   Te = Ta + Td;
 
131
                                   Tbz = Td - Ta;
 
132
                                   {
 
133
                                        E T2a, T60, T61, TbA, T2h;
 
134
                                        TbA = T26 - T29;
 
135
                                        T2a = T26 + T29;
 
136
                                        T3d = T3b + T3c;
 
137
                                        T60 = T3b - T3c;
 
138
                                        Td9 = T7 - Te;
 
139
                                        Tf = T7 + Te;
 
140
                                        TcB = TbA - Tbz;
 
141
                                        TbB = Tbz + TbA;
 
142
                                        T61 = T39 + T38;
 
143
                                        T3a = T38 - T39;
 
144
                                        T2h = T2d + T2g;
 
145
                                        Tag = T2d - T2g;
 
146
                                        T7C = T61 + T60;
 
147
                                        T62 = T60 - T61;
 
148
                                        TdH = T2a - T2h;
 
149
                                        T2i = T2a + T2h;
 
150
                                   }
 
151
                              }
 
152
                         }
 
153
                         {
 
154
                              E T3j, Ti, T3h, T2l, T3g, Tl, T3k, T2o, Tq, T3q, Tp, T3o, T2s, Tr, T2t;
 
155
                              E T2u;
 
156
                              {
 
157
                                   E Tj, Tk, T2m, T2n;
 
158
                                   {
 
159
                                        E Tg, Th, T2j, T2k;
 
160
                                        Tg = ri[WS(is, 4)];
 
161
                                        Tcb = Taf - Tag;
 
162
                                        Tah = Taf + Tag;
 
163
                                        T8G = T3a + T3d;
 
164
                                        T3e = T3a - T3d;
 
165
                                        Th = ri[WS(is, 36)];
 
166
                                        T2j = ii[WS(is, 4)];
 
167
                                        T2k = ii[WS(is, 36)];
 
168
                                        Tj = ri[WS(is, 20)];
 
169
                                        T3j = Tg - Th;
 
170
                                        Ti = Tg + Th;
 
171
                                        T3h = T2j - T2k;
 
172
                                        T2l = T2j + T2k;
 
173
                                        Tk = ri[WS(is, 52)];
 
174
                                        T2m = ii[WS(is, 20)];
 
175
                                        T2n = ii[WS(is, 52)];
 
176
                                   }
 
177
                                   {
 
178
                                        E Tn, To, T2q, T2r;
 
179
                                        Tn = ri[WS(is, 60)];
 
180
                                        T3g = Tj - Tk;
 
181
                                        Tl = Tj + Tk;
 
182
                                        T3k = T2m - T2n;
 
183
                                        T2o = T2m + T2n;
 
184
                                        To = ri[WS(is, 28)];
 
185
                                        T2q = ii[WS(is, 60)];
 
186
                                        T2r = ii[WS(is, 28)];
 
187
                                        Tq = ri[WS(is, 12)];
 
188
                                        T3q = Tn - To;
 
189
                                        Tp = Tn + To;
 
190
                                        T3o = T2q - T2r;
 
191
                                        T2s = T2q + T2r;
 
192
                                        Tr = ri[WS(is, 44)];
 
193
                                        T2t = ii[WS(is, 12)];
 
194
                                        T2u = ii[WS(is, 44)];
 
195
                                   }
 
196
                              }
 
197
                              {
 
198
                                   E T3n, T3r, T2p, T2w;
 
199
                                   {
 
200
                                        E Tai, Tm, T2v, Tal, Tt, Taj, Ts, Tam;
 
201
                                        Tai = Ti - Tl;
 
202
                                        Tm = Ti + Tl;
 
203
                                        T3n = Tq - Tr;
 
204
                                        Ts = Tq + Tr;
 
205
                                        T3r = T2t - T2u;
 
206
                                        T2v = T2t + T2u;
 
207
                                        Tal = Tp - Ts;
 
208
                                        Tt = Tp + Ts;
 
209
                                        Taj = T2l - T2o;
 
210
                                        T2p = T2l + T2o;
 
211
                                        Tam = T2s - T2v;
 
212
                                        T2w = T2s + T2v;
 
213
                                        Tu = Tm + Tt;
 
214
                                        TdI = Tt - Tm;
 
215
                                        Tak = Tai + Taj;
 
216
                                        TbC = Taj - Tai;
 
217
                                        TbD = Tal + Tam;
 
218
                                        Tan = Tal - Tam;
 
219
                                   }
 
220
                                   {
 
221
                                        E T7F, T7E, T3i, T3l, T7H, T7I;
 
222
                                        T7F = T3h - T3g;
 
223
                                        T3i = T3g + T3h;
 
224
                                        T3l = T3j - T3k;
 
225
                                        T7E = T3j + T3k;
 
226
                                        Tda = T2p - T2w;
 
227
                                        T2x = T2p + T2w;
 
228
                                        T65 = FNMS(KP414213562, T3i, T3l);
 
229
                                        T3m = FMA(KP414213562, T3l, T3i);
 
230
                                        T3s = T3q - T3r;
 
231
                                        T7H = T3q + T3r;
 
232
                                        T7I = T3o - T3n;
 
233
                                        T3p = T3n + T3o;
 
234
                                        T8I = FNMS(KP414213562, T7E, T7F);
 
235
                                        T7G = FMA(KP414213562, T7F, T7E);
 
236
                                        T8J = FMA(KP414213562, T7H, T7I);
 
237
                                        T7J = FNMS(KP414213562, T7I, T7H);
 
238
                                   }
 
239
                              }
 
240
                         }
 
241
                    }
 
242
                    {
 
243
                         E T3H, Ty, T3x, T2B, T3w, TB, T3I, T2E, TI, T2L, T3z, TF, T3E, T3K, T2I;
 
244
                         E T3A;
 
245
                         {
 
246
                              E T2z, T2A, Tz, TA, Tw, Tx, T2C, T2D;
 
247
                              Tw = ri[WS(is, 2)];
 
248
                              Tx = ri[WS(is, 34)];
 
249
                              T2z = ii[WS(is, 2)];
 
250
                              T64 = FMA(KP414213562, T3p, T3s);
 
251
                              T3t = FNMS(KP414213562, T3s, T3p);
 
252
                              T3H = Tw - Tx;
 
253
                              Ty = Tw + Tx;
 
254
                              T2A = ii[WS(is, 34)];
 
255
                              Tz = ri[WS(is, 18)];
 
256
                              TA = ri[WS(is, 50)];
 
257
                              T2C = ii[WS(is, 18)];
 
258
                              T3x = T2z - T2A;
 
259
                              T2B = T2z + T2A;
 
260
                              T3w = Tz - TA;
 
261
                              TB = Tz + TA;
 
262
                              T2D = ii[WS(is, 50)];
 
263
                              {
 
264
                                   E T2J, T3C, T2K, TG, TH;
 
265
                                   TG = ri[WS(is, 58)];
 
266
                                   TH = ri[WS(is, 26)];
 
267
                                   T2J = ii[WS(is, 58)];
 
268
                                   T3I = T2C - T2D;
 
269
                                   T2E = T2C + T2D;
 
270
                                   T3C = TG - TH;
 
271
                                   TI = TG + TH;
 
272
                                   T2K = ii[WS(is, 26)];
 
273
                                   {
 
274
                                        E T2G, T2H, TD, TE, T3D;
 
275
                                        TD = ri[WS(is, 10)];
 
276
                                        TE = ri[WS(is, 42)];
 
277
                                        T3D = T2J - T2K;
 
278
                                        T2L = T2J + T2K;
 
279
                                        T2G = ii[WS(is, 10)];
 
280
                                        T3z = TD - TE;
 
281
                                        TF = TD + TE;
 
282
                                        T2H = ii[WS(is, 42)];
 
283
                                        T3E = T3C - T3D;
 
284
                                        T3K = T3C + T3D;
 
285
                                        T2I = T2G + T2H;
 
286
                                        T3A = T2G - T2H;
 
287
                                   }
 
288
                              }
 
289
                         }
 
290
                         {
 
291
                              E T3L, T3B, T2F, T2M;
 
292
                              {
 
293
                                   E Tat, Taq, Tar, TC, TJ, Tau;
 
294
                                   Tat = Ty - TB;
 
295
                                   TC = Ty + TB;
 
296
                                   TJ = TF + TI;
 
297
                                   Taq = TI - TF;
 
298
                                   T3L = T3A - T3z;
 
299
                                   T3B = T3z + T3A;
 
300
                                   Tdd = TC - TJ;
 
301
                                   TK = TC + TJ;
 
302
                                   Tar = T2B - T2E;
 
303
                                   T2F = T2B + T2E;
 
304
                                   Tau = T2I - T2L;
 
305
                                   T2M = T2I + T2L;
 
306
                                   Tce = Tar - Taq;
 
307
                                   Tas = Taq + Tar;
 
308
                                   Tcf = Tat - Tau;
 
309
                                   Tav = Tat + Tau;
 
310
                              }
 
311
                              {
 
312
                                   E T7M, T7Q, T7N, T3y, T3F, T7P;
 
313
                                   T7M = T3x - T3w;
 
314
                                   T3y = T3w + T3x;
 
315
                                   T3F = T3B - T3E;
 
316
                                   T7Q = T3B + T3E;
 
317
                                   Tdc = T2F - T2M;
 
318
                                   T2N = T2F + T2M;
 
319
                                   T6G = FMA(KP707106781, T3F, T3y);
 
320
                                   T3G = FNMS(KP707106781, T3F, T3y);
 
321
                                   T7N = T3L + T3K;
 
322
                                   T3M = T3K - T3L;
 
323
                                   T3J = T3H - T3I;
 
324
                                   T7P = T3H + T3I;
 
325
                                   T9k = FNMS(KP707106781, T7N, T7M);
 
326
                                   T7O = FMA(KP707106781, T7N, T7M);
 
327
                                   T9l = FNMS(KP707106781, T7Q, T7P);
 
328
                                   T7R = FMA(KP707106781, T7Q, T7P);
 
329
                              }
 
330
                         }
 
331
                    }
 
332
                    {
 
333
                         E T5I, T1z, Tb8, T56, T53, T1C, Tb9, T5L, T1J, Tbq, T58, T1G, T5N, T5h, Tbp;
 
334
                         E T5b;
 
335
                         {
 
336
                              E T54, T55, T1A, T1B, T1x, T1y, T5J, T5K;
 
337
                              T1x = ri[WS(is, 63)];
 
338
                              T1y = ri[WS(is, 31)];
 
339
                              T54 = ii[WS(is, 63)];
 
340
                              T6H = FMA(KP707106781, T3M, T3J);
 
341
                              T3N = FNMS(KP707106781, T3M, T3J);
 
342
                              T5I = T1x - T1y;
 
343
                              T1z = T1x + T1y;
 
344
                              T55 = ii[WS(is, 31)];
 
345
                              T1A = ri[WS(is, 15)];
 
346
                              T1B = ri[WS(is, 47)];
 
347
                              T5J = ii[WS(is, 15)];
 
348
                              Tb8 = T54 + T55;
 
349
                              T56 = T54 - T55;
 
350
                              T53 = T1A - T1B;
 
351
                              T1C = T1A + T1B;
 
352
                              T5K = ii[WS(is, 47)];
 
353
                              {
 
354
                                   E T5e, T5d, T5f, T1H, T1I;
 
355
                                   T1H = ri[WS(is, 55)];
 
356
                                   T1I = ri[WS(is, 23)];
 
357
                                   T5e = ii[WS(is, 55)];
 
358
                                   Tb9 = T5J + T5K;
 
359
                                   T5L = T5J - T5K;
 
360
                                   T5d = T1H - T1I;
 
361
                                   T1J = T1H + T1I;
 
362
                                   T5f = ii[WS(is, 23)];
 
363
                                   {
 
364
                                        E T59, T5a, T1E, T1F, T5g;
 
365
                                        T1E = ri[WS(is, 7)];
 
366
                                        T1F = ri[WS(is, 39)];
 
367
                                        T5g = T5e - T5f;
 
368
                                        Tbq = T5e + T5f;
 
369
                                        T59 = ii[WS(is, 7)];
 
370
                                        T58 = T1E - T1F;
 
371
                                        T1G = T1E + T1F;
 
372
                                        T5a = ii[WS(is, 39)];
 
373
                                        T5N = T5d + T5g;
 
374
                                        T5h = T5d - T5g;
 
375
                                        Tbp = T59 + T5a;
 
376
                                        T5b = T59 - T5a;
 
377
                                   }
 
378
                              }
 
379
                         }
 
380
                         {
 
381
                              E Tb7, T5O, Tba, T57, T5i, T8x, T8w, T5M, T5P;
 
382
                              {
 
383
                                   E Tbo, T5c, Tbr, Tdw, T1D, T1K, Tdv;
 
384
                                   Tbo = T1z - T1C;
 
385
                                   T1D = T1z + T1C;
 
386
                                   T1K = T1G + T1J;
 
387
                                   Tb7 = T1J - T1G;
 
388
                                   T5c = T58 + T5b;
 
389
                                   T5O = T5b - T58;
 
390
                                   TdA = T1D - T1K;
 
391
                                   T1L = T1D + T1K;
 
392
                                   Tbr = Tbp - Tbq;
 
393
                                   Tdw = Tbp + Tbq;
 
394
                                   Tba = Tb8 - Tb9;
 
395
                                   Tdv = Tb8 + Tb9;
 
396
                                   T8l = T56 - T53;
 
397
                                   T57 = T53 + T56;
 
398
                                   Tct = Tbo - Tbr;
 
399
                                   Tbs = Tbo + Tbr;
 
400
                                   Teo = Tdv + Tdw;
 
401
                                   Tdx = Tdv - Tdw;
 
402
                                   T5i = T5c - T5h;
 
403
                                   T8x = T5c + T5h;
 
404
                              }
 
405
                              T8w = T5I + T5L;
 
406
                              T5M = T5I - T5L;
 
407
                              T5P = T5N - T5O;
 
408
                              T8m = T5O + T5N;
 
409
                              T6Y = FMA(KP707106781, T5i, T57);
 
410
                              T5j = FNMS(KP707106781, T5i, T57);
 
411
                              T6V = FMA(KP707106781, T5P, T5M);
 
412
                              T5Q = FNMS(KP707106781, T5P, T5M);
 
413
                              T9z = FNMS(KP707106781, T8x, T8w);
 
414
                              T8y = FMA(KP707106781, T8x, T8w);
 
415
                              Tcw = Tba - Tb7;
 
416
                              Tbb = Tb7 + Tba;
 
417
                         }
 
418
                    }
 
419
               }
 
420
               {
 
421
                    E T82, T83, T45, T42, T87, T8a;
 
422
                    {
 
423
                         E T40, TN, T3Q, T2Q, T3P, TQ, T41, T2T, TX, T30, T3S, TU, T3X, T43, T2X;
 
424
                         E T3T;
 
425
                         {
 
426
                              E T2O, T2P, TO, TP, TL, TM, T2R, T2S;
 
427
                              TL = ri[WS(is, 62)];
 
428
                              TM = ri[WS(is, 30)];
 
429
                              T2O = ii[WS(is, 62)];
 
430
                              T9C = FNMS(KP707106781, T8m, T8l);
 
431
                              T8n = FMA(KP707106781, T8m, T8l);
 
432
                              T40 = TL - TM;
 
433
                              TN = TL + TM;
 
434
                              T2P = ii[WS(is, 30)];
 
435
                              TO = ri[WS(is, 14)];
 
436
                              TP = ri[WS(is, 46)];
 
437
                              T2R = ii[WS(is, 14)];
 
438
                              T3Q = T2O - T2P;
 
439
                              T2Q = T2O + T2P;
 
440
                              T3P = TO - TP;
 
441
                              TQ = TO + TP;
 
442
                              T2S = ii[WS(is, 46)];
 
443
                              {
 
444
                                   E T2Y, T3V, T2Z, TV, TW;
 
445
                                   TV = ri[WS(is, 54)];
 
446
                                   TW = ri[WS(is, 22)];
 
447
                                   T2Y = ii[WS(is, 54)];
 
448
                                   T41 = T2R - T2S;
 
449
                                   T2T = T2R + T2S;
 
450
                                   T3V = TV - TW;
 
451
                                   TX = TV + TW;
 
452
                                   T2Z = ii[WS(is, 22)];
 
453
                                   {
 
454
                                        E T2V, T2W, TS, TT, T3W;
 
455
                                        TS = ri[WS(is, 6)];
 
456
                                        TT = ri[WS(is, 38)];
 
457
                                        T3W = T2Y - T2Z;
 
458
                                        T30 = T2Y + T2Z;
 
459
                                        T2V = ii[WS(is, 6)];
 
460
                                        T3S = TS - TT;
 
461
                                        TU = TS + TT;
 
462
                                        T2W = ii[WS(is, 38)];
 
463
                                        T3X = T3V - T3W;
 
464
                                        T43 = T3V + T3W;
 
465
                                        T2X = T2V + T2W;
 
466
                                        T3T = T2V - T2W;
 
467
                                   }
 
468
                              }
 
469
                         }
 
470
                         {
 
471
                              E T44, T3U, T2U, T31;
 
472
                              {
 
473
                                   E TaA, Tax, Tay, TR, TY, TaB;
 
474
                                   TaA = TN - TQ;
 
475
                                   TR = TN + TQ;
 
476
                                   TY = TU + TX;
 
477
                                   Tax = TX - TU;
 
478
                                   T44 = T3T - T3S;
 
479
                                   T3U = T3S + T3T;
 
480
                                   Tdf = TR - TY;
 
481
                                   TZ = TR + TY;
 
482
                                   Tay = T2Q - T2T;
 
483
                                   T2U = T2Q + T2T;
 
484
                                   TaB = T2X - T30;
 
485
                                   T31 = T2X + T30;
 
486
                                   Tch = Tay - Tax;
 
487
                                   Taz = Tax + Tay;
 
488
                                   Tci = TaA - TaB;
 
489
                                   TaC = TaA + TaB;
 
490
                              }
 
491
                              {
 
492
                                   E T7T, T7X, T7U, T3R, T3Y, T7W;
 
493
                                   T7T = T3Q - T3P;
 
494
                                   T3R = T3P + T3Q;
 
495
                                   T3Y = T3U - T3X;
 
496
                                   T7X = T3U + T3X;
 
497
                                   Tdg = T2U - T31;
 
498
                                   T32 = T2U + T31;
 
499
                                   T6J = FMA(KP707106781, T3Y, T3R);
 
500
                                   T3Z = FNMS(KP707106781, T3Y, T3R);
 
501
                                   T7U = T44 + T43;
 
502
                                   T45 = T43 - T44;
 
503
                                   T42 = T40 - T41;
 
504
                                   T7W = T40 + T41;
 
505
                                   T9n = FNMS(KP707106781, T7U, T7T);
 
506
                                   T7V = FMA(KP707106781, T7U, T7T);
 
507
                                   T9o = FNMS(KP707106781, T7X, T7W);
 
508
                                   T7Y = FMA(KP707106781, T7X, T7W);
 
509
                              }
 
510
                         }
 
511
                    }
 
512
                    {
 
513
                         E T4P, T14, TaH, T4d, T4a, T17, TaI, T4S, T1e, TaZ, T4f, T1b, T4U, T4o, TaY;
 
514
                         E T4i;
 
515
                         {
 
516
                              E T4b, T4c, T15, T16, T12, T13, T4Q, T4R;
 
517
                              T12 = ri[WS(is, 1)];
 
518
                              T13 = ri[WS(is, 33)];
 
519
                              T4b = ii[WS(is, 1)];
 
520
                              T6K = FMA(KP707106781, T45, T42);
 
521
                              T46 = FNMS(KP707106781, T45, T42);
 
522
                              T4P = T12 - T13;
 
523
                              T14 = T12 + T13;
 
524
                              T4c = ii[WS(is, 33)];
 
525
                              T15 = ri[WS(is, 17)];
 
526
                              T16 = ri[WS(is, 49)];
 
527
                              T4Q = ii[WS(is, 17)];
 
528
                              TaH = T4b + T4c;
 
529
                              T4d = T4b - T4c;
 
530
                              T4a = T15 - T16;
 
531
                              T17 = T15 + T16;
 
532
                              T4R = ii[WS(is, 49)];
 
533
                              {
 
534
                                   E T4l, T4k, T4m, T1c, T1d;
 
535
                                   T1c = ri[WS(is, 57)];
 
536
                                   T1d = ri[WS(is, 25)];
 
537
                                   T4l = ii[WS(is, 57)];
 
538
                                   TaI = T4Q + T4R;
 
539
                                   T4S = T4Q - T4R;
 
540
                                   T4k = T1c - T1d;
 
541
                                   T1e = T1c + T1d;
 
542
                                   T4m = ii[WS(is, 25)];
 
543
                                   {
 
544
                                        E T4g, T4h, T19, T1a, T4n;
 
545
                                        T19 = ri[WS(is, 9)];
 
546
                                        T1a = ri[WS(is, 41)];
 
547
                                        T4n = T4l - T4m;
 
548
                                        TaZ = T4l + T4m;
 
549
                                        T4g = ii[WS(is, 9)];
 
550
                                        T4f = T19 - T1a;
 
551
                                        T1b = T19 + T1a;
 
552
                                        T4h = ii[WS(is, 41)];
 
553
                                        T4U = T4k + T4n;
 
554
                                        T4o = T4k - T4n;
 
555
                                        TaY = T4g + T4h;
 
556
                                        T4i = T4g - T4h;
 
557
                                   }
 
558
                              }
 
559
                         }
 
560
                         {
 
561
                              E TaG, T4V, TaJ, T4e, T4p, T8e, T8d, T4T, T4W;
 
562
                              {
 
563
                                   E TaX, T4j, Tb0, Tdl, T18, T1f, Tdk;
 
564
                                   TaX = T14 - T17;
 
565
                                   T18 = T14 + T17;
 
566
                                   T1f = T1b + T1e;
 
567
                                   TaG = T1e - T1b;
 
568
                                   T4j = T4f + T4i;
 
569
                                   T4V = T4i - T4f;
 
570
                                   Tdp = T18 - T1f;
 
571
                                   T1g = T18 + T1f;
 
572
                                   Tb0 = TaY - TaZ;
 
573
                                   Tdl = TaY + TaZ;
 
574
                                   TaJ = TaH - TaI;
 
575
                                   Tdk = TaH + TaI;
 
576
                                   T82 = T4d - T4a;
 
577
                                   T4e = T4a + T4d;
 
578
                                   Tcm = TaX - Tb0;
 
579
                                   Tb1 = TaX + Tb0;
 
580
                                   Tej = Tdk + Tdl;
 
581
                                   Tdm = Tdk - Tdl;
 
582
                                   T4p = T4j - T4o;
 
583
                                   T8e = T4j + T4o;
 
584
                              }
 
585
                              T8d = T4P + T4S;
 
586
                              T4T = T4P - T4S;
 
587
                              T4W = T4U - T4V;
 
588
                              T83 = T4V + T4U;
 
589
                              T6R = FMA(KP707106781, T4p, T4e);
 
590
                              T4q = FNMS(KP707106781, T4p, T4e);
 
591
                              T6O = FMA(KP707106781, T4W, T4T);
 
592
                              T4X = FNMS(KP707106781, T4W, T4T);
 
593
                              T9s = FNMS(KP707106781, T8e, T8d);
 
594
                              T8f = FMA(KP707106781, T8e, T8d);
 
595
                              Tcp = TaJ - TaG;
 
596
                              TaK = TaG + TaJ;
 
597
                         }
 
598
                    }
 
599
                    {
 
600
                         E T85, T4L, TaO, T1n, Tdq, TaN, T86, T4G, T4r, T1q, T4s, TaR, T4z, T4w, T1t;
 
601
                         E T4t;
 
602
                         {
 
603
                              E T4C, T1j, T4D, TaL, T4K, T4H, T1m, T4E;
 
604
                              {
 
605
                                   E T4I, T4J, T1h, T1i, T1k, T1l;
 
606
                                   T1h = ri[WS(is, 5)];
 
607
                                   T1i = ri[WS(is, 37)];
 
608
                                   T4I = ii[WS(is, 5)];
 
609
                                   T9v = FNMS(KP707106781, T83, T82);
 
610
                                   T84 = FMA(KP707106781, T83, T82);
 
611
                                   T4C = T1h - T1i;
 
612
                                   T1j = T1h + T1i;
 
613
                                   T4J = ii[WS(is, 37)];
 
614
                                   T1k = ri[WS(is, 21)];
 
615
                                   T1l = ri[WS(is, 53)];
 
616
                                   T4D = ii[WS(is, 21)];
 
617
                                   TaL = T4I + T4J;
 
618
                                   T4K = T4I - T4J;
 
619
                                   T4H = T1k - T1l;
 
620
                                   T1m = T1k + T1l;
 
621
                                   T4E = ii[WS(is, 53)];
 
622
                              }
 
623
                              {
 
624
                                   E T4x, T4y, T1r, T1s;
 
625
                                   {
 
626
                                        E T1o, T4F, TaM, T1p;
 
627
                                        T1o = ri[WS(is, 61)];
 
628
                                        T85 = T4K - T4H;
 
629
                                        T4L = T4H + T4K;
 
630
                                        TaO = T1j - T1m;
 
631
                                        T1n = T1j + T1m;
 
632
                                        T4F = T4D - T4E;
 
633
                                        TaM = T4D + T4E;
 
634
                                        T1p = ri[WS(is, 29)];
 
635
                                        T4x = ii[WS(is, 61)];
 
636
                                        Tdq = TaL + TaM;
 
637
                                        TaN = TaL - TaM;
 
638
                                        T86 = T4C + T4F;
 
639
                                        T4G = T4C - T4F;
 
640
                                        T4r = T1o - T1p;
 
641
                                        T1q = T1o + T1p;
 
642
                                        T4y = ii[WS(is, 29)];
 
643
                                   }
 
644
                                   T1r = ri[WS(is, 13)];
 
645
                                   T1s = ri[WS(is, 45)];
 
646
                                   T4s = ii[WS(is, 13)];
 
647
                                   TaR = T4x + T4y;
 
648
                                   T4z = T4x - T4y;
 
649
                                   T4w = T1r - T1s;
 
650
                                   T1t = T1r + T1s;
 
651
                                   T4t = ii[WS(is, 45)];
 
652
                              }
 
653
                         }
 
654
                         {
 
655
                              E T88, TaP, T89, TaU, T4Z, T4B, T4M, T4Y, T8g, T8h;
 
656
                              {
 
657
                                   E T4A, Tb2, Tdr, T4v, Tb3;
 
658
                                   {
 
659
                                        E TaQ, T1u, T4u, TaS, TaT;
 
660
                                        T88 = T4z - T4w;
 
661
                                        T4A = T4w + T4z;
 
662
                                        TaQ = T1q - T1t;
 
663
                                        T1u = T1q + T1t;
 
664
                                        T4u = T4s - T4t;
 
665
                                        TaS = T4s + T4t;
 
666
                                        Tb2 = TaO + TaN;
 
667
                                        TaP = TaN - TaO;
 
668
                                        Tdr = TaR + TaS;
 
669
                                        TaT = TaR - TaS;
 
670
                                        T89 = T4r + T4u;
 
671
                                        T4v = T4r - T4u;
 
672
                                        Tdn = T1u - T1n;
 
673
                                        T1v = T1n + T1u;
 
674
                                        Tb3 = TaQ - TaT;
 
675
                                        TaU = TaQ + TaT;
 
676
                                   }
 
677
                                   T4Z = FNMS(KP414213562, T4v, T4A);
 
678
                                   T4B = FMA(KP414213562, T4A, T4v);
 
679
                                   Tcq = Tb2 - Tb3;
 
680
                                   Tb4 = Tb2 + Tb3;
 
681
                                   Tek = Tdq + Tdr;
 
682
                                   Tds = Tdq - Tdr;
 
683
                                   T4M = FNMS(KP414213562, T4L, T4G);
 
684
                                   T4Y = FMA(KP414213562, T4G, T4L);
 
685
                              }
 
686
                              T87 = FNMS(KP414213562, T86, T85);
 
687
                              T8g = FMA(KP414213562, T85, T86);
 
688
                              T6P = T4M + T4B;
 
689
                              T4N = T4B - T4M;
 
690
                              T6S = T4Y + T4Z;
 
691
                              T50 = T4Y - T4Z;
 
692
                              T8h = FNMS(KP414213562, T88, T89);
 
693
                              T8a = FMA(KP414213562, T89, T88);
 
694
                              T9w = T8g - T8h;
 
695
                              T8i = T8g + T8h;
 
696
                              Tcn = TaU - TaP;
 
697
                              TaV = TaP + TaU;
 
698
                         }
 
699
                    }
 
700
                    {
 
701
                         E T8o, T5E, Tbf, T1S, TdB, Tbe, T8p, T5z, T5k, T1V, T5l, Tbi, T5s, T5p, T1Y;
 
702
                         E T5m;
 
703
                         {
 
704
                              E T5v, T1O, T5w, Tbc, T5D, T5A, T1R, T5x;
 
705
                              {
 
706
                                   E T5B, T5C, T1M, T1N, T1P, T1Q;
 
707
                                   T1M = ri[WS(is, 3)];
 
708
                                   T1N = ri[WS(is, 35)];
 
709
                                   T5B = ii[WS(is, 3)];
 
710
                                   T9t = T8a - T87;
 
711
                                   T8b = T87 + T8a;
 
712
                                   T5v = T1M - T1N;
 
713
                                   T1O = T1M + T1N;
 
714
                                   T5C = ii[WS(is, 35)];
 
715
                                   T1P = ri[WS(is, 19)];
 
716
                                   T1Q = ri[WS(is, 51)];
 
717
                                   T5w = ii[WS(is, 19)];
 
718
                                   Tbc = T5B + T5C;
 
719
                                   T5D = T5B - T5C;
 
720
                                   T5A = T1P - T1Q;
 
721
                                   T1R = T1P + T1Q;
 
722
                                   T5x = ii[WS(is, 51)];
 
723
                              }
 
724
                              {
 
725
                                   E T5q, T5r, T1W, T1X;
 
726
                                   {
 
727
                                        E T1T, T5y, Tbd, T1U;
 
728
                                        T1T = ri[WS(is, 59)];
 
729
                                        T8o = T5D - T5A;
 
730
                                        T5E = T5A + T5D;
 
731
                                        Tbf = T1O - T1R;
 
732
                                        T1S = T1O + T1R;
 
733
                                        T5y = T5w - T5x;
 
734
                                        Tbd = T5w + T5x;
 
735
                                        T1U = ri[WS(is, 27)];
 
736
                                        T5q = ii[WS(is, 59)];
 
737
                                        TdB = Tbc + Tbd;
 
738
                                        Tbe = Tbc - Tbd;
 
739
                                        T8p = T5v + T5y;
 
740
                                        T5z = T5v - T5y;
 
741
                                        T5k = T1T - T1U;
 
742
                                        T1V = T1T + T1U;
 
743
                                        T5r = ii[WS(is, 27)];
 
744
                                   }
 
745
                                   T1W = ri[WS(is, 11)];
 
746
                                   T1X = ri[WS(is, 43)];
 
747
                                   T5l = ii[WS(is, 11)];
 
748
                                   Tbi = T5q + T5r;
 
749
                                   T5s = T5q - T5r;
 
750
                                   T5p = T1W - T1X;
 
751
                                   T1Y = T1W + T1X;
 
752
                                   T5m = ii[WS(is, 43)];
 
753
                              }
 
754
                         }
 
755
                         {
 
756
                              E T8r, Tbg, T8s, Tbl, T5S, T5u, T5F, T5R, T8z, T8A;
 
757
                              {
 
758
                                   E T5t, Tbt, TdC, T5o, Tbu;
 
759
                                   {
 
760
                                        E Tbh, T1Z, T5n, Tbj, Tbk;
 
761
                                        T8r = T5s - T5p;
 
762
                                        T5t = T5p + T5s;
 
763
                                        Tbh = T1V - T1Y;
 
764
                                        T1Z = T1V + T1Y;
 
765
                                        T5n = T5l - T5m;
 
766
                                        Tbj = T5l + T5m;
 
767
                                        Tbt = Tbf + Tbe;
 
768
                                        Tbg = Tbe - Tbf;
 
769
                                        TdC = Tbi + Tbj;
 
770
                                        Tbk = Tbi - Tbj;
 
771
                                        T8s = T5k + T5n;
 
772
                                        T5o = T5k - T5n;
 
773
                                        Tdy = T1Z - T1S;
 
774
                                        T20 = T1S + T1Z;
 
775
                                        Tbu = Tbh - Tbk;
 
776
                                        Tbl = Tbh + Tbk;
 
777
                                   }
 
778
                                   T5S = FNMS(KP414213562, T5o, T5t);
 
779
                                   T5u = FMA(KP414213562, T5t, T5o);
 
780
                                   Tcx = Tbt - Tbu;
 
781
                                   Tbv = Tbt + Tbu;
 
782
                                   Tep = TdB + TdC;
 
783
                                   TdD = TdB - TdC;
 
784
                                   T5F = FNMS(KP414213562, T5E, T5z);
 
785
                                   T5R = FMA(KP414213562, T5z, T5E);
 
786
                              }
 
787
                              T8q = FNMS(KP414213562, T8p, T8o);
 
788
                              T8z = FMA(KP414213562, T8o, T8p);
 
789
                              T6W = T5F + T5u;
 
790
                              T5G = T5u - T5F;
 
791
                              T6Z = T5R + T5S;
 
792
                              T5T = T5R - T5S;
 
793
                              T8A = FNMS(KP414213562, T8r, T8s);
 
794
                              T8t = FMA(KP414213562, T8s, T8r);
 
795
                              T9D = T8z - T8A;
 
796
                              T8B = T8z + T8A;
 
797
                              Tcu = Tbl - Tbg;
 
798
                              Tbm = Tbg + Tbl;
 
799
                         }
 
800
                    }
 
801
               }
 
802
               {
 
803
                    E T9A, T8u, TbE, Tao, Td7, Td8;
 
804
                    {
 
805
                         E Teq, Ten, Tex, Teh, TeB, Tev, Tey, Tem, Te9, Tec;
 
806
                         {
 
807
                              E Tef, Teu, Tel, T11, Tei, Tet, T2y, TeI, T23, T22, T33, Teg, TeD, TeG, T34;
 
808
                              E TeH;
 
809
                              {
 
810
                                   E TeE, TeF, Tv, T10, T1w, T21;
 
811
                                   Tef = Tf - Tu;
 
812
                                   Tv = Tf + Tu;
 
813
                                   T10 = TK + TZ;
 
814
                                   Teu = TZ - TK;
 
815
                                   Tel = Tej - Tek;
 
816
                                   TeE = Tej + Tek;
 
817
                                   T9A = T8t - T8q;
 
818
                                   T8u = T8q + T8t;
 
819
                                   TeD = Tv - T10;
 
820
                                   T11 = Tv + T10;
 
821
                                   TeF = Teo + Tep;
 
822
                                   Teq = Teo - Tep;
 
823
                                   Tei = T1g - T1v;
 
824
                                   T1w = T1g + T1v;
 
825
                                   T21 = T1L + T20;
 
826
                                   Ten = T1L - T20;
 
827
                                   Tet = T2i - T2x;
 
828
                                   T2y = T2i + T2x;
 
829
                                   TeI = TeE + TeF;
 
830
                                   TeG = TeE - TeF;
 
831
                                   T23 = T21 - T1w;
 
832
                                   T22 = T1w + T21;
 
833
                                   T33 = T2N + T32;
 
834
                                   Teg = T2N - T32;
 
835
                              }
 
836
                              ro[WS(os, 16)] = TeD + TeG;
 
837
                              ro[WS(os, 48)] = TeD - TeG;
 
838
                              ro[0] = T11 + T22;
 
839
                              ro[WS(os, 32)] = T11 - T22;
 
840
                              T34 = T2y - T33;
 
841
                              TeH = T2y + T33;
 
842
                              io[0] = TeH + TeI;
 
843
                              io[WS(os, 32)] = TeH - TeI;
 
844
                              io[WS(os, 48)] = T34 - T23;
 
845
                              io[WS(os, 16)] = T23 + T34;
 
846
                              Tex = Tef - Teg;
 
847
                              Teh = Tef + Teg;
 
848
                              TeB = Teu + Tet;
 
849
                              Tev = Tet - Teu;
 
850
                              Tey = Tel - Tei;
 
851
                              Tem = Tei + Tel;
 
852
                         }
 
853
                         {
 
854
                              E TdV, Tdb, TdJ, Te5, TdE, Tdz, Te6, Tdi, Teb, Te3, TdZ, TdY, TdW, TdM, TdR;
 
855
                              E Tdu;
 
856
                              {
 
857
                                   E TdL, Tde, Tdh, TdK, Tez, Ter;
 
858
                                   TdV = Td9 + Tda;
 
859
                                   Tdb = Td9 - Tda;
 
860
                                   TdJ = TdH - TdI;
 
861
                                   Te5 = TdI + TdH;
 
862
                                   Tez = Ten + Teq;
 
863
                                   Ter = Ten - Teq;
 
864
                                   TdL = Tdd + Tdc;
 
865
                                   Tde = Tdc - Tdd;
 
866
                                   {
 
867
                                        E TeA, TeC, Tew, Tes;
 
868
                                        TeA = Tey - Tez;
 
869
                                        TeC = Tey + Tez;
 
870
                                        Tew = Ter - Tem;
 
871
                                        Tes = Tem + Ter;
 
872
                                        ro[WS(os, 24)] = FMA(KP707106781, TeA, Tex);
 
873
                                        ro[WS(os, 56)] = FNMS(KP707106781, TeA, Tex);
 
874
                                        io[WS(os, 8)] = FMA(KP707106781, TeC, TeB);
 
875
                                        io[WS(os, 40)] = FNMS(KP707106781, TeC, TeB);
 
876
                                        io[WS(os, 24)] = FMA(KP707106781, Tew, Tev);
 
877
                                        io[WS(os, 56)] = FNMS(KP707106781, Tew, Tev);
 
878
                                        ro[WS(os, 8)] = FMA(KP707106781, Tes, Teh);
 
879
                                        ro[WS(os, 40)] = FNMS(KP707106781, Tes, Teh);
 
880
                                        Tdh = Tdf + Tdg;
 
881
                                        TdK = Tdf - Tdg;
 
882
                                   }
 
883
                                   {
 
884
                                        E Te1, Te2, Tdo, Tdt;
 
885
                                        TdE = TdA - TdD;
 
886
                                        Te1 = TdA + TdD;
 
887
                                        Te2 = Tdy + Tdx;
 
888
                                        Tdz = Tdx - Tdy;
 
889
                                        Te6 = Tde + Tdh;
 
890
                                        Tdi = Tde - Tdh;
 
891
                                        Teb = FMA(KP414213562, Te1, Te2);
 
892
                                        Te3 = FNMS(KP414213562, Te2, Te1);
 
893
                                        TdZ = Tdn + Tdm;
 
894
                                        Tdo = Tdm - Tdn;
 
895
                                        Tdt = Tdp - Tds;
 
896
                                        TdY = Tdp + Tds;
 
897
                                        TdW = TdL + TdK;
 
898
                                        TdM = TdK - TdL;
 
899
                                        TdR = FNMS(KP414213562, Tdo, Tdt);
 
900
                                        Tdu = FMA(KP414213562, Tdt, Tdo);
 
901
                                   }
 
902
                              }
 
903
                              {
 
904
                                   E TdT, Tea, Te0, TdU;
 
905
                                   {
 
906
                                        E Tdj, TdQ, TdF, TdP, TdN, TdS, TdO, TdG;
 
907
                                        TdT = FNMS(KP707106781, Tdi, Tdb);
 
908
                                        Tdj = FMA(KP707106781, Tdi, Tdb);
 
909
                                        Tea = FNMS(KP414213562, TdY, TdZ);
 
910
                                        Te0 = FMA(KP414213562, TdZ, TdY);
 
911
                                        TdQ = FMA(KP414213562, Tdz, TdE);
 
912
                                        TdF = FNMS(KP414213562, TdE, Tdz);
 
913
                                        TdP = FMA(KP707106781, TdM, TdJ);
 
914
                                        TdN = FNMS(KP707106781, TdM, TdJ);
 
915
                                        TdS = TdQ - TdR;
 
916
                                        TdU = TdR + TdQ;
 
917
                                        TdO = Tdu + TdF;
 
918
                                        TdG = Tdu - TdF;
 
919
                                        io[WS(os, 12)] = FMA(KP923879532, TdS, TdP);
 
920
                                        io[WS(os, 44)] = FNMS(KP923879532, TdS, TdP);
 
921
                                        ro[WS(os, 12)] = FMA(KP923879532, TdG, Tdj);
 
922
                                        ro[WS(os, 44)] = FNMS(KP923879532, TdG, Tdj);
 
923
                                        io[WS(os, 60)] = FMA(KP923879532, TdO, TdN);
 
924
                                        io[WS(os, 28)] = FNMS(KP923879532, TdO, TdN);
 
925
                                   }
 
926
                                   {
 
927
                                        E Te8, Te7, Ted, Tee, TdX, Te4;
 
928
                                        Te9 = FNMS(KP707106781, TdW, TdV);
 
929
                                        TdX = FMA(KP707106781, TdW, TdV);
 
930
                                        Te4 = Te0 + Te3;
 
931
                                        Te8 = Te3 - Te0;
 
932
                                        Te7 = FNMS(KP707106781, Te6, Te5);
 
933
                                        Ted = FMA(KP707106781, Te6, Te5);
 
934
                                        ro[WS(os, 60)] = FMA(KP923879532, TdU, TdT);
 
935
                                        ro[WS(os, 28)] = FNMS(KP923879532, TdU, TdT);
 
936
                                        ro[WS(os, 4)] = FMA(KP923879532, Te4, TdX);
 
937
                                        ro[WS(os, 36)] = FNMS(KP923879532, Te4, TdX);
 
938
                                        Tee = Tea + Teb;
 
939
                                        Tec = Tea - Teb;
 
940
                                        io[WS(os, 4)] = FMA(KP923879532, Tee, Ted);
 
941
                                        io[WS(os, 36)] = FNMS(KP923879532, Tee, Ted);
 
942
                                        io[WS(os, 20)] = FMA(KP923879532, Te8, Te7);
 
943
                                        io[WS(os, 52)] = FNMS(KP923879532, Te8, Te7);
 
944
                                   }
 
945
                              }
 
946
                         }
 
947
                         {
 
948
                              E TcP, Tcd, TcZ, TcD, Tcy, Tcv, TcT, Td0, Tck, Td4, TcX, TcS, TcK, Tcs, TcQ;
 
949
                              E TcG;
 
950
                              {
 
951
                                   E TcF, Tcg, Tcj, TcE, TcV, TcW, Tcc, TcC, Tco, Tcr;
 
952
                                   TbE = TbC + TbD;
 
953
                                   Tcc = TbC - TbD;
 
954
                                   TcC = Tan - Tak;
 
955
                                   Tao = Tak + Tan;
 
956
                                   TcF = FNMS(KP414213562, Tce, Tcf);
 
957
                                   Tcg = FMA(KP414213562, Tcf, Tce);
 
958
                                   ro[WS(os, 20)] = FMA(KP923879532, Tec, Te9);
 
959
                                   ro[WS(os, 52)] = FNMS(KP923879532, Tec, Te9);
 
960
                                   TcP = FNMS(KP707106781, Tcc, Tcb);
 
961
                                   Tcd = FMA(KP707106781, Tcc, Tcb);
 
962
                                   TcZ = FNMS(KP707106781, TcC, TcB);
 
963
                                   TcD = FMA(KP707106781, TcC, TcB);
 
964
                                   Tcj = FNMS(KP414213562, Tci, Tch);
 
965
                                   TcE = FMA(KP414213562, Tch, Tci);
 
966
                                   Tcy = FNMS(KP707106781, Tcx, Tcw);
 
967
                                   TcV = FMA(KP707106781, Tcx, Tcw);
 
968
                                   TcW = FMA(KP707106781, Tcu, Tct);
 
969
                                   Tcv = FNMS(KP707106781, Tcu, Tct);
 
970
                                   TcT = FMA(KP707106781, Tcn, Tcm);
 
971
                                   Tco = FNMS(KP707106781, Tcn, Tcm);
 
972
                                   Td0 = Tcg + Tcj;
 
973
                                   Tck = Tcg - Tcj;
 
974
                                   Td4 = FMA(KP198912367, TcV, TcW);
 
975
                                   TcX = FNMS(KP198912367, TcW, TcV);
 
976
                                   Tcr = FNMS(KP707106781, Tcq, Tcp);
 
977
                                   TcS = FMA(KP707106781, Tcq, Tcp);
 
978
                                   TcK = FNMS(KP668178637, Tco, Tcr);
 
979
                                   Tcs = FMA(KP668178637, Tcr, Tco);
 
980
                                   TcQ = TcF + TcE;
 
981
                                   TcG = TcE - TcF;
 
982
                              }
 
983
                              {
 
984
                                   E TcJ, Td5, TcU, TcM;
 
985
                                   {
 
986
                                        E Tcl, TcL, Tcz, TcN, TcH, TcO, TcI, TcA;
 
987
                                        TcJ = FNMS(KP923879532, Tck, Tcd);
 
988
                                        Tcl = FMA(KP923879532, Tck, Tcd);
 
989
                                        Td5 = FNMS(KP198912367, TcS, TcT);
 
990
                                        TcU = FMA(KP198912367, TcT, TcS);
 
991
                                        TcL = FMA(KP668178637, Tcv, Tcy);
 
992
                                        Tcz = FNMS(KP668178637, Tcy, Tcv);
 
993
                                        TcN = FMA(KP923879532, TcG, TcD);
 
994
                                        TcH = FNMS(KP923879532, TcG, TcD);
 
995
                                        TcO = TcK + TcL;
 
996
                                        TcM = TcK - TcL;
 
997
                                        TcI = Tcz - Tcs;
 
998
                                        TcA = Tcs + Tcz;
 
999
                                        io[WS(os, 6)] = FMA(KP831469612, TcO, TcN);
 
1000
                                        io[WS(os, 38)] = FNMS(KP831469612, TcO, TcN);
 
1001
                                        ro[WS(os, 6)] = FMA(KP831469612, TcA, Tcl);
 
1002
                                        ro[WS(os, 38)] = FNMS(KP831469612, TcA, Tcl);
 
1003
                                        io[WS(os, 22)] = FMA(KP831469612, TcI, TcH);
 
1004
                                        io[WS(os, 54)] = FNMS(KP831469612, TcI, TcH);
 
1005
                                   }
 
1006
                                   {
 
1007
                                        E Td2, Td1, Td3, Td6, TcR, TcY;
 
1008
                                        Td7 = FMA(KP923879532, TcQ, TcP);
 
1009
                                        TcR = FNMS(KP923879532, TcQ, TcP);
 
1010
                                        TcY = TcU - TcX;
 
1011
                                        Td2 = TcU + TcX;
 
1012
                                        Td1 = FMA(KP923879532, Td0, TcZ);
 
1013
                                        Td3 = FNMS(KP923879532, Td0, TcZ);
 
1014
                                        ro[WS(os, 22)] = FMA(KP831469612, TcM, TcJ);
 
1015
                                        ro[WS(os, 54)] = FNMS(KP831469612, TcM, TcJ);
 
1016
                                        ro[WS(os, 14)] = FMA(KP980785280, TcY, TcR);
 
1017
                                        ro[WS(os, 46)] = FNMS(KP980785280, TcY, TcR);
 
1018
                                        Td6 = Td4 - Td5;
 
1019
                                        Td8 = Td5 + Td4;
 
1020
                                        io[WS(os, 14)] = FMA(KP980785280, Td6, Td3);
 
1021
                                        io[WS(os, 46)] = FNMS(KP980785280, Td6, Td3);
 
1022
                                        io[WS(os, 62)] = FMA(KP980785280, Td2, Td1);
 
1023
                                        io[WS(os, 30)] = FNMS(KP980785280, Td2, Td1);
 
1024
                                   }
 
1025
                              }
 
1026
                         }
 
1027
                    }
 
1028
                    {
 
1029
                         E T3f, T66, T63, T3u, T7z, T7A, Tc5, Tc8;
 
1030
                         {
 
1031
                              E TbR, Tap, Tc1, TbF, Tbw, Tbn, TbV, Tc2, TaE, Tc7, TbZ, TbU, TbN, Tb6, TbS;
 
1032
                              E TbI;
 
1033
                              {
 
1034
                                   E TbH, Taw, TaD, TbG, TbX, TbY, TaW, Tb5;
 
1035
                                   TbH = FMA(KP414213562, Tas, Tav);
 
1036
                                   Taw = FNMS(KP414213562, Tav, Tas);
 
1037
                                   ro[WS(os, 62)] = FMA(KP980785280, Td8, Td7);
 
1038
                                   ro[WS(os, 30)] = FNMS(KP980785280, Td8, Td7);
 
1039
                                   TbR = FMA(KP707106781, Tao, Tah);
 
1040
                                   Tap = FNMS(KP707106781, Tao, Tah);
 
1041
                                   Tc1 = FMA(KP707106781, TbE, TbB);
 
1042
                                   TbF = FNMS(KP707106781, TbE, TbB);
 
1043
                                   TaD = FMA(KP414213562, TaC, Taz);
 
1044
                                   TbG = FNMS(KP414213562, Taz, TaC);
 
1045
                                   Tbw = FNMS(KP707106781, Tbv, Tbs);
 
1046
                                   TbX = FMA(KP707106781, Tbv, Tbs);
 
1047
                                   TbY = FMA(KP707106781, Tbm, Tbb);
 
1048
                                   Tbn = FNMS(KP707106781, Tbm, Tbb);
 
1049
                                   TbV = FMA(KP707106781, TaV, TaK);
 
1050
                                   TaW = FNMS(KP707106781, TaV, TaK);
 
1051
                                   Tc2 = Taw + TaD;
 
1052
                                   TaE = Taw - TaD;
 
1053
                                   Tc7 = FMA(KP198912367, TbX, TbY);
 
1054
                                   TbZ = FNMS(KP198912367, TbY, TbX);
 
1055
                                   Tb5 = FNMS(KP707106781, Tb4, Tb1);
 
1056
                                   TbU = FMA(KP707106781, Tb4, Tb1);
 
1057
                                   TbN = FNMS(KP668178637, TaW, Tb5);
 
1058
                                   Tb6 = FMA(KP668178637, Tb5, TaW);
 
1059
                                   TbS = TbH + TbG;
 
1060
                                   TbI = TbG - TbH;
 
1061
                              }
 
1062
                              {
 
1063
                                   E TbP, Tc6, TbW, TbQ;
 
1064
                                   {
 
1065
                                        E TaF, TbM, Tbx, TbL, TbJ, TbO, TbK, Tby;
 
1066
                                        TbP = FNMS(KP923879532, TaE, Tap);
 
1067
                                        TaF = FMA(KP923879532, TaE, Tap);
 
1068
                                        Tc6 = FNMS(KP198912367, TbU, TbV);
 
1069
                                        TbW = FMA(KP198912367, TbV, TbU);
 
1070
                                        TbM = FMA(KP668178637, Tbn, Tbw);
 
1071
                                        Tbx = FNMS(KP668178637, Tbw, Tbn);
 
1072
                                        TbL = FMA(KP923879532, TbI, TbF);
 
1073
                                        TbJ = FNMS(KP923879532, TbI, TbF);
 
1074
                                        TbO = TbM - TbN;
 
1075
                                        TbQ = TbN + TbM;
 
1076
                                        TbK = Tb6 + Tbx;
 
1077
                                        Tby = Tb6 - Tbx;
 
1078
                                        io[WS(os, 10)] = FMA(KP831469612, TbO, TbL);
 
1079
                                        io[WS(os, 42)] = FNMS(KP831469612, TbO, TbL);
 
1080
                                        ro[WS(os, 10)] = FMA(KP831469612, Tby, TaF);
 
1081
                                        ro[WS(os, 42)] = FNMS(KP831469612, Tby, TaF);
 
1082
                                        io[WS(os, 58)] = FMA(KP831469612, TbK, TbJ);
 
1083
                                        io[WS(os, 26)] = FNMS(KP831469612, TbK, TbJ);
 
1084
                                   }
 
1085
                                   {
 
1086
                                        E Tc4, Tc3, Tc9, Tca, TbT, Tc0;
 
1087
                                        Tc5 = FNMS(KP923879532, TbS, TbR);
 
1088
                                        TbT = FMA(KP923879532, TbS, TbR);
 
1089
                                        Tc0 = TbW + TbZ;
 
1090
                                        Tc4 = TbZ - TbW;
 
1091
                                        Tc3 = FNMS(KP923879532, Tc2, Tc1);
 
1092
                                        Tc9 = FMA(KP923879532, Tc2, Tc1);
 
1093
                                        ro[WS(os, 58)] = FMA(KP831469612, TbQ, TbP);
 
1094
                                        ro[WS(os, 26)] = FNMS(KP831469612, TbQ, TbP);
 
1095
                                        ro[WS(os, 2)] = FMA(KP980785280, Tc0, TbT);
 
1096
                                        ro[WS(os, 34)] = FNMS(KP980785280, Tc0, TbT);
 
1097
                                        Tca = Tc6 + Tc7;
 
1098
                                        Tc8 = Tc6 - Tc7;
 
1099
                                        io[WS(os, 2)] = FMA(KP980785280, Tca, Tc9);
 
1100
                                        io[WS(os, 34)] = FNMS(KP980785280, Tca, Tc9);
 
1101
                                        io[WS(os, 18)] = FMA(KP980785280, Tc4, Tc3);
 
1102
                                        io[WS(os, 50)] = FNMS(KP980785280, Tc4, Tc3);
 
1103
                                   }
 
1104
                              }
 
1105
                         }
 
1106
                         {
 
1107
                              E T7h, T6F, T70, T6X, T7x, T7m, T7w, T7p, T7s, T6M, T7c, T6U, T7r, T75, T7i;
 
1108
                              E T78;
 
1109
                              {
 
1110
                                   E T6T, T6Q, T77, T6I, T6L, T76, T73, T74;
 
1111
                                   {
 
1112
                                        E T7k, T7l, T6D, T6E, T7n, T7o;
 
1113
                                        T3f = FMA(KP707106781, T3e, T37);
 
1114
                                        T6D = FNMS(KP707106781, T3e, T37);
 
1115
                                        T6E = T65 + T64;
 
1116
                                        T66 = T64 - T65;
 
1117
                                        T6T = FNMS(KP923879532, T6S, T6R);
 
1118
                                        T7k = FMA(KP923879532, T6S, T6R);
 
1119
                                        ro[WS(os, 18)] = FMA(KP980785280, Tc8, Tc5);
 
1120
                                        ro[WS(os, 50)] = FNMS(KP980785280, Tc8, Tc5);
 
1121
                                        T7h = FMA(KP923879532, T6E, T6D);
 
1122
                                        T6F = FNMS(KP923879532, T6E, T6D);
 
1123
                                        T7l = FMA(KP923879532, T6P, T6O);
 
1124
                                        T6Q = FNMS(KP923879532, T6P, T6O);
 
1125
                                        T70 = FNMS(KP923879532, T6Z, T6Y);
 
1126
                                        T7n = FMA(KP923879532, T6Z, T6Y);
 
1127
                                        T7o = FMA(KP923879532, T6W, T6V);
 
1128
                                        T6X = FNMS(KP923879532, T6W, T6V);
 
1129
                                        T77 = FNMS(KP198912367, T6G, T6H);
 
1130
                                        T6I = FMA(KP198912367, T6H, T6G);
 
1131
                                        T7x = FNMS(KP098491403, T7k, T7l);
 
1132
                                        T7m = FMA(KP098491403, T7l, T7k);
 
1133
                                        T7w = FMA(KP098491403, T7n, T7o);
 
1134
                                        T7p = FNMS(KP098491403, T7o, T7n);
 
1135
                                        T6L = FNMS(KP198912367, T6K, T6J);
 
1136
                                        T76 = FMA(KP198912367, T6J, T6K);
 
1137
                                   }
 
1138
                                   T63 = FMA(KP707106781, T62, T5Z);
 
1139
                                   T73 = FNMS(KP707106781, T62, T5Z);
 
1140
                                   T7s = T6I + T6L;
 
1141
                                   T6M = T6I - T6L;
 
1142
                                   T7c = FNMS(KP820678790, T6Q, T6T);
 
1143
                                   T6U = FMA(KP820678790, T6T, T6Q);
 
1144
                                   T74 = T3m + T3t;
 
1145
                                   T3u = T3m - T3t;
 
1146
                                   T7r = FMA(KP923879532, T74, T73);
 
1147
                                   T75 = FNMS(KP923879532, T74, T73);
 
1148
                                   T7i = T77 + T76;
 
1149
                                   T78 = T76 - T77;
 
1150
                              }
 
1151
                              {
 
1152
                                   E T7b, T6N, T7f, T79, T71, T7d;
 
1153
                                   T7b = FNMS(KP980785280, T6M, T6F);
 
1154
                                   T6N = FMA(KP980785280, T6M, T6F);
 
1155
                                   T7f = FMA(KP980785280, T78, T75);
 
1156
                                   T79 = FNMS(KP980785280, T78, T75);
 
1157
                                   T71 = FNMS(KP820678790, T70, T6X);
 
1158
                                   T7d = FMA(KP820678790, T6X, T70);
 
1159
                                   {
 
1160
                                        E T7u, T7t, T7v, T7y, T7j, T7q;
 
1161
                                        T7z = FMA(KP980785280, T7i, T7h);
 
1162
                                        T7j = FNMS(KP980785280, T7i, T7h);
 
1163
                                        T7q = T7m - T7p;
 
1164
                                        T7u = T7m + T7p;
 
1165
                                        {
 
1166
                                             E T7g, T7e, T72, T7a;
 
1167
                                             T7g = T7c + T7d;
 
1168
                                             T7e = T7c - T7d;
 
1169
                                             T72 = T6U + T71;
 
1170
                                             T7a = T71 - T6U;
 
1171
                                             ro[WS(os, 23)] = FMA(KP773010453, T7e, T7b);
 
1172
                                             ro[WS(os, 55)] = FNMS(KP773010453, T7e, T7b);
 
1173
                                             io[WS(os, 7)] = FMA(KP773010453, T7g, T7f);
 
1174
                                             io[WS(os, 39)] = FNMS(KP773010453, T7g, T7f);
 
1175
                                             io[WS(os, 23)] = FMA(KP773010453, T7a, T79);
 
1176
                                             io[WS(os, 55)] = FNMS(KP773010453, T7a, T79);
 
1177
                                             ro[WS(os, 7)] = FMA(KP773010453, T72, T6N);
 
1178
                                             ro[WS(os, 39)] = FNMS(KP773010453, T72, T6N);
 
1179
                                             ro[WS(os, 47)] = FNMS(KP995184726, T7q, T7j);
 
1180
                                             ro[WS(os, 15)] = FMA(KP995184726, T7q, T7j);
 
1181
                                        }
 
1182
                                        T7t = FMA(KP980785280, T7s, T7r);
 
1183
                                        T7v = FNMS(KP980785280, T7s, T7r);
 
1184
                                        T7y = T7w - T7x;
 
1185
                                        T7A = T7x + T7w;
 
1186
                                        io[WS(os, 15)] = FMA(KP995184726, T7y, T7v);
 
1187
                                        io[WS(os, 47)] = FNMS(KP995184726, T7y, T7v);
 
1188
                                        io[WS(os, 63)] = FMA(KP995184726, T7u, T7t);
 
1189
                                        io[WS(os, 31)] = FNMS(KP995184726, T7u, T7t);
 
1190
                                   }
 
1191
                              }
 
1192
                         }
 
1193
                         {
 
1194
                              E T7D, T8K, T8H, T7K, Tad, Tae, T6x, T6A;
 
1195
                              {
 
1196
                                   E T9V, T9j, T9E, T9B, Tab, Ta0, Taa, Ta3, Ta6, T9q, T9Q, T9y, Ta5, T9J, T9W;
 
1197
                                   E T9M;
 
1198
                                   {
 
1199
                                        E T9x, T9u, T9L, T9m, T9p, T9K, T9H, T9I;
 
1200
                                        {
 
1201
                                             E T9Y, T9Z, T9h, T9i, Ta1, Ta2;
 
1202
                                             T7D = FMA(KP707106781, T7C, T7B);
 
1203
                                             T9h = FNMS(KP707106781, T7C, T7B);
 
1204
                                             T9i = T8I - T8J;
 
1205
                                             T8K = T8I + T8J;
 
1206
                                             T9x = FNMS(KP923879532, T9w, T9v);
 
1207
                                             T9Y = FMA(KP923879532, T9w, T9v);
 
1208
                                             ro[WS(os, 63)] = FMA(KP995184726, T7A, T7z);
 
1209
                                             ro[WS(os, 31)] = FNMS(KP995184726, T7A, T7z);
 
1210
                                             T9V = FNMS(KP923879532, T9i, T9h);
 
1211
                                             T9j = FMA(KP923879532, T9i, T9h);
 
1212
                                             T9Z = FMA(KP923879532, T9t, T9s);
 
1213
                                             T9u = FNMS(KP923879532, T9t, T9s);
 
1214
                                             T9E = FNMS(KP923879532, T9D, T9C);
 
1215
                                             Ta1 = FMA(KP923879532, T9D, T9C);
 
1216
                                             Ta2 = FMA(KP923879532, T9A, T9z);
 
1217
                                             T9B = FNMS(KP923879532, T9A, T9z);
 
1218
                                             T9L = FNMS(KP668178637, T9k, T9l);
 
1219
                                             T9m = FMA(KP668178637, T9l, T9k);
 
1220
                                             Tab = FNMS(KP303346683, T9Y, T9Z);
 
1221
                                             Ta0 = FMA(KP303346683, T9Z, T9Y);
 
1222
                                             Taa = FMA(KP303346683, Ta1, Ta2);
 
1223
                                             Ta3 = FNMS(KP303346683, Ta2, Ta1);
 
1224
                                             T9p = FNMS(KP668178637, T9o, T9n);
 
1225
                                             T9K = FMA(KP668178637, T9n, T9o);
 
1226
                                        }
 
1227
                                        T8H = FMA(KP707106781, T8G, T8F);
 
1228
                                        T9H = FNMS(KP707106781, T8G, T8F);
 
1229
                                        Ta6 = T9m + T9p;
 
1230
                                        T9q = T9m - T9p;
 
1231
                                        T9Q = FNMS(KP534511135, T9u, T9x);
 
1232
                                        T9y = FMA(KP534511135, T9x, T9u);
 
1233
                                        T9I = T7J - T7G;
 
1234
                                        T7K = T7G + T7J;
 
1235
                                        Ta5 = FNMS(KP923879532, T9I, T9H);
 
1236
                                        T9J = FMA(KP923879532, T9I, T9H);
 
1237
                                        T9W = T9L + T9K;
 
1238
                                        T9M = T9K - T9L;
 
1239
                                   }
 
1240
                                   {
 
1241
                                        E T9P, T9r, T9T, T9N, T9F, T9R;
 
1242
                                        T9P = FNMS(KP831469612, T9q, T9j);
 
1243
                                        T9r = FMA(KP831469612, T9q, T9j);
 
1244
                                        T9T = FMA(KP831469612, T9M, T9J);
 
1245
                                        T9N = FNMS(KP831469612, T9M, T9J);
 
1246
                                        T9F = FNMS(KP534511135, T9E, T9B);
 
1247
                                        T9R = FMA(KP534511135, T9B, T9E);
 
1248
                                        {
 
1249
                                             E Ta8, Ta7, Ta9, Tac, T9X, Ta4;
 
1250
                                             Tad = FMA(KP831469612, T9W, T9V);
 
1251
                                             T9X = FNMS(KP831469612, T9W, T9V);
 
1252
                                             Ta4 = Ta0 - Ta3;
 
1253
                                             Ta8 = Ta0 + Ta3;
 
1254
                                             {
 
1255
                                                  E T9U, T9S, T9G, T9O;
 
1256
                                                  T9U = T9Q + T9R;
 
1257
                                                  T9S = T9Q - T9R;
 
1258
                                                  T9G = T9y + T9F;
 
1259
                                                  T9O = T9F - T9y;
 
1260
                                                  ro[WS(os, 21)] = FMA(KP881921264, T9S, T9P);
 
1261
                                                  ro[WS(os, 53)] = FNMS(KP881921264, T9S, T9P);
 
1262
                                                  io[WS(os, 5)] = FMA(KP881921264, T9U, T9T);
 
1263
                                                  io[WS(os, 37)] = FNMS(KP881921264, T9U, T9T);
 
1264
                                                  io[WS(os, 21)] = FMA(KP881921264, T9O, T9N);
 
1265
                                                  io[WS(os, 53)] = FNMS(KP881921264, T9O, T9N);
 
1266
                                                  ro[WS(os, 5)] = FMA(KP881921264, T9G, T9r);
 
1267
                                                  ro[WS(os, 37)] = FNMS(KP881921264, T9G, T9r);
 
1268
                                                  ro[WS(os, 45)] = FNMS(KP956940335, Ta4, T9X);
 
1269
                                                  ro[WS(os, 13)] = FMA(KP956940335, Ta4, T9X);
 
1270
                                             }
 
1271
                                             Ta7 = FMA(KP831469612, Ta6, Ta5);
 
1272
                                             Ta9 = FNMS(KP831469612, Ta6, Ta5);
 
1273
                                             Tac = Taa - Tab;
 
1274
                                             Tae = Tab + Taa;
 
1275
                                             io[WS(os, 13)] = FMA(KP956940335, Tac, Ta9);
 
1276
                                             io[WS(os, 45)] = FNMS(KP956940335, Tac, Ta9);
 
1277
                                             io[WS(os, 61)] = FMA(KP956940335, Ta8, Ta7);
 
1278
                                             io[WS(os, 29)] = FNMS(KP956940335, Ta8, Ta7);
 
1279
                                        }
 
1280
                                   }
 
1281
                              }
 
1282
                              {
 
1283
                                   E T6j, T3v, T5U, T5H, T6y, T6o, T6z, T6r, T6u, T48, T6f, T52, T6t, T67, T6k;
 
1284
                                   E T6a;
 
1285
                                   {
 
1286
                                        E T51, T4O, T69, T3O, T47, T68;
 
1287
                                        {
 
1288
                                             E T6m, T6n, T6p, T6q;
 
1289
                                             T51 = FNMS(KP923879532, T50, T4X);
 
1290
                                             T6m = FMA(KP923879532, T50, T4X);
 
1291
                                             ro[WS(os, 61)] = FMA(KP956940335, Tae, Tad);
 
1292
                                             ro[WS(os, 29)] = FNMS(KP956940335, Tae, Tad);
 
1293
                                             T6j = FMA(KP923879532, T3u, T3f);
 
1294
                                             T3v = FNMS(KP923879532, T3u, T3f);
 
1295
                                             T6n = FMA(KP923879532, T4N, T4q);
 
1296
                                             T4O = FNMS(KP923879532, T4N, T4q);
 
1297
                                             T5U = FNMS(KP923879532, T5T, T5Q);
 
1298
                                             T6p = FMA(KP923879532, T5T, T5Q);
 
1299
                                             T6q = FMA(KP923879532, T5G, T5j);
 
1300
                                             T5H = FNMS(KP923879532, T5G, T5j);
 
1301
                                             T69 = FMA(KP668178637, T3G, T3N);
 
1302
                                             T3O = FNMS(KP668178637, T3N, T3G);
 
1303
                                             T6y = FNMS(KP303346683, T6m, T6n);
 
1304
                                             T6o = FMA(KP303346683, T6n, T6m);
 
1305
                                             T6z = FMA(KP303346683, T6p, T6q);
 
1306
                                             T6r = FNMS(KP303346683, T6q, T6p);
 
1307
                                             T47 = FMA(KP668178637, T46, T3Z);
 
1308
                                             T68 = FNMS(KP668178637, T3Z, T46);
 
1309
                                        }
 
1310
                                        T6u = T3O + T47;
 
1311
                                        T48 = T3O - T47;
 
1312
                                        T6f = FNMS(KP534511135, T4O, T51);
 
1313
                                        T52 = FMA(KP534511135, T51, T4O);
 
1314
                                        T6t = FMA(KP923879532, T66, T63);
 
1315
                                        T67 = FNMS(KP923879532, T66, T63);
 
1316
                                        T6k = T69 + T68;
 
1317
                                        T6a = T68 - T69;
 
1318
                                   }
 
1319
                                   {
 
1320
                                        E T6h, T49, T6d, T6b, T5V, T6e;
 
1321
                                        T6h = FNMS(KP831469612, T48, T3v);
 
1322
                                        T49 = FMA(KP831469612, T48, T3v);
 
1323
                                        T6d = FMA(KP831469612, T6a, T67);
 
1324
                                        T6b = FNMS(KP831469612, T6a, T67);
 
1325
                                        T5V = FNMS(KP534511135, T5U, T5H);
 
1326
                                        T6e = FMA(KP534511135, T5H, T5U);
 
1327
                                        {
 
1328
                                             E T6w, T6v, T6B, T6C, T6l, T6s;
 
1329
                                             T6x = FNMS(KP831469612, T6k, T6j);
 
1330
                                             T6l = FMA(KP831469612, T6k, T6j);
 
1331
                                             T6s = T6o + T6r;
 
1332
                                             T6w = T6r - T6o;
 
1333
                                             {
 
1334
                                                  E T6g, T6i, T5W, T6c;
 
1335
                                                  T6g = T6e - T6f;
 
1336
                                                  T6i = T6f + T6e;
 
1337
                                                  T5W = T52 - T5V;
 
1338
                                                  T6c = T52 + T5V;
 
1339
                                                  ro[WS(os, 59)] = FMA(KP881921264, T6i, T6h);
 
1340
                                                  ro[WS(os, 27)] = FNMS(KP881921264, T6i, T6h);
 
1341
                                                  io[WS(os, 11)] = FMA(KP881921264, T6g, T6d);
 
1342
                                                  io[WS(os, 43)] = FNMS(KP881921264, T6g, T6d);
 
1343
                                                  io[WS(os, 59)] = FMA(KP881921264, T6c, T6b);
 
1344
                                                  io[WS(os, 27)] = FNMS(KP881921264, T6c, T6b);
 
1345
                                                  ro[WS(os, 11)] = FMA(KP881921264, T5W, T49);
 
1346
                                                  ro[WS(os, 43)] = FNMS(KP881921264, T5W, T49);
 
1347
                                                  ro[WS(os, 35)] = FNMS(KP956940335, T6s, T6l);
 
1348
                                                  ro[WS(os, 3)] = FMA(KP956940335, T6s, T6l);
 
1349
                                             }
 
1350
                                             T6v = FNMS(KP831469612, T6u, T6t);
 
1351
                                             T6B = FMA(KP831469612, T6u, T6t);
 
1352
                                             T6C = T6y + T6z;
 
1353
                                             T6A = T6y - T6z;
 
1354
                                             io[WS(os, 3)] = FMA(KP956940335, T6C, T6B);
 
1355
                                             io[WS(os, 35)] = FNMS(KP956940335, T6C, T6B);
 
1356
                                             io[WS(os, 19)] = FMA(KP956940335, T6w, T6v);
 
1357
                                             io[WS(os, 51)] = FNMS(KP956940335, T6w, T6v);
 
1358
                                        }
 
1359
                                   }
 
1360
                              }
 
1361
                              {
 
1362
                                   E T8X, T7L, T8C, T8v, T9c, T92, T9d, T95, T98, T80, T8T, T8k, T97, T8L, T8Y;
 
1363
                                   E T8O;
 
1364
                                   {
 
1365
                                        E T8j, T8c, T8N, T7S, T7Z, T8M;
 
1366
                                        {
 
1367
                                             E T90, T91, T93, T94;
 
1368
                                             T8j = FNMS(KP923879532, T8i, T8f);
 
1369
                                             T90 = FMA(KP923879532, T8i, T8f);
 
1370
                                             ro[WS(os, 19)] = FMA(KP956940335, T6A, T6x);
 
1371
                                             ro[WS(os, 51)] = FNMS(KP956940335, T6A, T6x);
 
1372
                                             T8X = FMA(KP923879532, T7K, T7D);
 
1373
                                             T7L = FNMS(KP923879532, T7K, T7D);
 
1374
                                             T91 = FMA(KP923879532, T8b, T84);
 
1375
                                             T8c = FNMS(KP923879532, T8b, T84);
 
1376
                                             T8C = FNMS(KP923879532, T8B, T8y);
 
1377
                                             T93 = FMA(KP923879532, T8B, T8y);
 
1378
                                             T94 = FMA(KP923879532, T8u, T8n);
 
1379
                                             T8v = FNMS(KP923879532, T8u, T8n);
 
1380
                                             T8N = FMA(KP198912367, T7O, T7R);
 
1381
                                             T7S = FNMS(KP198912367, T7R, T7O);
 
1382
                                             T9c = FNMS(KP098491403, T90, T91);
 
1383
                                             T92 = FMA(KP098491403, T91, T90);
 
1384
                                             T9d = FMA(KP098491403, T93, T94);
 
1385
                                             T95 = FNMS(KP098491403, T94, T93);
 
1386
                                             T7Z = FMA(KP198912367, T7Y, T7V);
 
1387
                                             T8M = FNMS(KP198912367, T7V, T7Y);
 
1388
                                        }
 
1389
                                        T98 = T7S + T7Z;
 
1390
                                        T80 = T7S - T7Z;
 
1391
                                        T8T = FNMS(KP820678790, T8c, T8j);
 
1392
                                        T8k = FMA(KP820678790, T8j, T8c);
 
1393
                                        T97 = FMA(KP923879532, T8K, T8H);
 
1394
                                        T8L = FNMS(KP923879532, T8K, T8H);
 
1395
                                        T8Y = T8N + T8M;
 
1396
                                        T8O = T8M - T8N;
 
1397
                                   }
 
1398
                                   {
 
1399
                                        E T8V, T81, T8R, T8P, T8D, T8S;
 
1400
                                        T8V = FNMS(KP980785280, T80, T7L);
 
1401
                                        T81 = FMA(KP980785280, T80, T7L);
 
1402
                                        T8R = FMA(KP980785280, T8O, T8L);
 
1403
                                        T8P = FNMS(KP980785280, T8O, T8L);
 
1404
                                        T8D = FNMS(KP820678790, T8C, T8v);
 
1405
                                        T8S = FMA(KP820678790, T8v, T8C);
 
1406
                                        {
 
1407
                                             E T9a, T99, T9f, T9g, T8Z, T96;
 
1408
                                             T9b = FNMS(KP980785280, T8Y, T8X);
 
1409
                                             T8Z = FMA(KP980785280, T8Y, T8X);
 
1410
                                             T96 = T92 + T95;
 
1411
                                             T9a = T95 - T92;
 
1412
                                             {
 
1413
                                                  E T8U, T8W, T8E, T8Q;
 
1414
                                                  T8U = T8S - T8T;
 
1415
                                                  T8W = T8T + T8S;
 
1416
                                                  T8E = T8k - T8D;
 
1417
                                                  T8Q = T8k + T8D;
 
1418
                                                  ro[WS(os, 57)] = FMA(KP773010453, T8W, T8V);
 
1419
                                                  ro[WS(os, 25)] = FNMS(KP773010453, T8W, T8V);
 
1420
                                                  io[WS(os, 9)] = FMA(KP773010453, T8U, T8R);
 
1421
                                                  io[WS(os, 41)] = FNMS(KP773010453, T8U, T8R);
 
1422
                                                  io[WS(os, 57)] = FMA(KP773010453, T8Q, T8P);
 
1423
                                                  io[WS(os, 25)] = FNMS(KP773010453, T8Q, T8P);
 
1424
                                                  ro[WS(os, 9)] = FMA(KP773010453, T8E, T81);
 
1425
                                                  ro[WS(os, 41)] = FNMS(KP773010453, T8E, T81);
 
1426
                                                  ro[WS(os, 33)] = FNMS(KP995184726, T96, T8Z);
 
1427
                                                  ro[WS(os, 1)] = FMA(KP995184726, T96, T8Z);
 
1428
                                             }
 
1429
                                             T99 = FNMS(KP980785280, T98, T97);
 
1430
                                             T9f = FMA(KP980785280, T98, T97);
 
1431
                                             T9g = T9c + T9d;
 
1432
                                             T9e = T9c - T9d;
 
1433
                                             io[WS(os, 1)] = FMA(KP995184726, T9g, T9f);
 
1434
                                             io[WS(os, 33)] = FNMS(KP995184726, T9g, T9f);
 
1435
                                             io[WS(os, 17)] = FMA(KP995184726, T9a, T99);
 
1436
                                             io[WS(os, 49)] = FNMS(KP995184726, T9a, T99);
 
1437
                                        }
 
1438
                                   }
 
1439
                              }
 
1440
                         }
 
1441
                    }
 
1442
               }
 
1443
          }
 
1444
          ro[WS(os, 17)] = FMA(KP995184726, T9e, T9b);
 
1445
          ro[WS(os, 49)] = FNMS(KP995184726, T9e, T9b);
 
1446
     }
 
1447
}
 
1448
 
 
1449
static const kdft_desc desc = { 64, "n1_64", {520, 0, 392, 0}, &GENUS, 0, 0, 0, 0 };
 
1450
void X(codelet_n1_64) (planner *p) {
 
1451
     X(kdft_register) (p, n1_64, &desc);
 
1452
}
 
1453
 
 
1454
#else                           /* HAVE_FMA */
 
1455
 
 
1456
/* Generated by: ../../../genfft/gen_notw -compact -variables 4 -pipeline-latency 4 -n 64 -name n1_64 -include n.h */
 
1457
 
 
1458
/*
 
1459
 * This function contains 912 FP additions, 248 FP multiplications,
 
1460
 * (or, 808 additions, 144 multiplications, 104 fused multiply/add),
 
1461
 * 172 stack variables, and 256 memory accesses
 
1462
 */
 
1463
/*
 
1464
 * Generator Id's : 
 
1465
 * $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
 
1466
 * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
 
1467
 * $Id: gen_notw.ml,v 1.30 2006-02-12 23:34:12 athena Exp $
 
1468
 */
 
1469
 
 
1470
#include "n.h"
 
1471
 
 
1472
static void n1_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
 
1473
{
 
1474
     DK(KP773010453, +0.773010453362736960810906609758469800971041293);
 
1475
     DK(KP634393284, +0.634393284163645498215171613225493370675687095);
 
1476
     DK(KP098017140, +0.098017140329560601994195563888641845861136673);
 
1477
     DK(KP995184726, +0.995184726672196886244836953109479921575474869);
 
1478
     DK(KP881921264, +0.881921264348355029712756863660388349508442621);
 
1479
     DK(KP471396736, +0.471396736825997648556387625905254377657460319);
 
1480
     DK(KP290284677, +0.290284677254462367636192375817395274691476278);
 
1481
     DK(KP956940335, +0.956940335732208864935797886980269969482849206);
 
1482
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
 
1483
     DK(KP555570233, +0.555570233019602224742830813948532874374937191);
 
1484
     DK(KP195090322, +0.195090322016128267848284868477022240927691618);
 
1485
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
 
1486
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
1487
     DK(KP382683432, +0.382683432365089771728459984030398866761344562);
 
1488
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
1489
     INT i;
 
1490
     for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
 
1491
          E T37, T7B, T8F, T5Z, Tf, Td9, TbB, TcB, T62, T7C, T2i, TdH, Tah, Tcb, T3e;
 
1492
          E T8G, Tu, TdI, Tak, TbD, Tan, TbC, T2x, Tda, T3m, T65, T7G, T8J, T7J, T8I;
 
1493
          E T3t, T64, TK, Tdd, Tas, Tce, Tav, Tcf, T2N, Tdc, T3G, T6G, T7O, T9k, T7R;
 
1494
          E T9l, T3N, T6H, T1L, Tdv, Tbs, Tcw, TdC, Teo, T5j, T6V, T5Q, T6Y, T8y, T9C;
 
1495
          E Tbb, Tct, T8n, T9z, TZ, Tdf, Taz, Tch, TaC, Tci, T32, Tdg, T3Z, T6J, T7V;
 
1496
          E T9n, T7Y, T9o, T46, T6K, T1g, Tdp, Tb1, Tcm, Tdm, Tej, T4q, T6R, T4X, T6O;
 
1497
          E T8f, T9s, TaK, Tcp, T84, T9v, T1v, Tdn, Tb4, Tcq, Tds, Tek, T4N, T6P, T50;
 
1498
          E T6S, T8i, T9w, TaV, Tcn, T8b, T9t, T20, TdD, Tbv, Tcu, Tdy, Tep, T5G, T6Z;
 
1499
          E T5T, T6W, T8B, T9A, Tbm, Tcx, T8u, T9D;
 
1500
          {
 
1501
               E T3, T35, T26, T5Y, T6, T5X, T29, T36, Ta, T39, T2d, T38, Td, T3b, T2g;
 
1502
               E T3c;
 
1503
               {
 
1504
                    E T1, T2, T24, T25;
 
1505
                    T1 = ri[0];
 
1506
                    T2 = ri[WS(is, 32)];
 
1507
                    T3 = T1 + T2;
 
1508
                    T35 = T1 - T2;
 
1509
                    T24 = ii[0];
 
1510
                    T25 = ii[WS(is, 32)];
 
1511
                    T26 = T24 + T25;
 
1512
                    T5Y = T24 - T25;
 
1513
               }
 
1514
               {
 
1515
                    E T4, T5, T27, T28;
 
1516
                    T4 = ri[WS(is, 16)];
 
1517
                    T5 = ri[WS(is, 48)];
 
1518
                    T6 = T4 + T5;
 
1519
                    T5X = T4 - T5;
 
1520
                    T27 = ii[WS(is, 16)];
 
1521
                    T28 = ii[WS(is, 48)];
 
1522
                    T29 = T27 + T28;
 
1523
                    T36 = T27 - T28;
 
1524
               }
 
1525
               {
 
1526
                    E T8, T9, T2b, T2c;
 
1527
                    T8 = ri[WS(is, 8)];
 
1528
                    T9 = ri[WS(is, 40)];
 
1529
                    Ta = T8 + T9;
 
1530
                    T39 = T8 - T9;
 
1531
                    T2b = ii[WS(is, 8)];
 
1532
                    T2c = ii[WS(is, 40)];
 
1533
                    T2d = T2b + T2c;
 
1534
                    T38 = T2b - T2c;
 
1535
               }
 
1536
               {
 
1537
                    E Tb, Tc, T2e, T2f;
 
1538
                    Tb = ri[WS(is, 56)];
 
1539
                    Tc = ri[WS(is, 24)];
 
1540
                    Td = Tb + Tc;
 
1541
                    T3b = Tb - Tc;
 
1542
                    T2e = ii[WS(is, 56)];
 
1543
                    T2f = ii[WS(is, 24)];
 
1544
                    T2g = T2e + T2f;
 
1545
                    T3c = T2e - T2f;
 
1546
               }
 
1547
               {
 
1548
                    E T7, Te, T2a, T2h;
 
1549
                    T37 = T35 - T36;
 
1550
                    T7B = T35 + T36;
 
1551
                    T8F = T5Y - T5X;
 
1552
                    T5Z = T5X + T5Y;
 
1553
                    T7 = T3 + T6;
 
1554
                    Te = Ta + Td;
 
1555
                    Tf = T7 + Te;
 
1556
                    Td9 = T7 - Te;
 
1557
                    {
 
1558
                         E Tbz, TbA, T60, T61;
 
1559
                         Tbz = T26 - T29;
 
1560
                         TbA = Td - Ta;
 
1561
                         TbB = Tbz - TbA;
 
1562
                         TcB = TbA + Tbz;
 
1563
                         T60 = T3b - T3c;
 
1564
                         T61 = T39 + T38;
 
1565
                         T62 = KP707106781 * (T60 - T61);
 
1566
                         T7C = KP707106781 * (T61 + T60);
 
1567
                    }
 
1568
                    T2a = T26 + T29;
 
1569
                    T2h = T2d + T2g;
 
1570
                    T2i = T2a + T2h;
 
1571
                    TdH = T2a - T2h;
 
1572
                    {
 
1573
                         E Taf, Tag, T3a, T3d;
 
1574
                         Taf = T3 - T6;
 
1575
                         Tag = T2d - T2g;
 
1576
                         Tah = Taf - Tag;
 
1577
                         Tcb = Taf + Tag;
 
1578
                         T3a = T38 - T39;
 
1579
                         T3d = T3b + T3c;
 
1580
                         T3e = KP707106781 * (T3a - T3d);
 
1581
                         T8G = KP707106781 * (T3a + T3d);
 
1582
                    }
 
1583
               }
 
1584
          }
 
1585
          {
 
1586
               E Ti, T3j, T2l, T3h, Tl, T3g, T2o, T3k, Tp, T3q, T2s, T3o, Ts, T3n, T2v;
 
1587
               E T3r;
 
1588
               {
 
1589
                    E Tg, Th, T2j, T2k;
 
1590
                    Tg = ri[WS(is, 4)];
 
1591
                    Th = ri[WS(is, 36)];
 
1592
                    Ti = Tg + Th;
 
1593
                    T3j = Tg - Th;
 
1594
                    T2j = ii[WS(is, 4)];
 
1595
                    T2k = ii[WS(is, 36)];
 
1596
                    T2l = T2j + T2k;
 
1597
                    T3h = T2j - T2k;
 
1598
               }
 
1599
               {
 
1600
                    E Tj, Tk, T2m, T2n;
 
1601
                    Tj = ri[WS(is, 20)];
 
1602
                    Tk = ri[WS(is, 52)];
 
1603
                    Tl = Tj + Tk;
 
1604
                    T3g = Tj - Tk;
 
1605
                    T2m = ii[WS(is, 20)];
 
1606
                    T2n = ii[WS(is, 52)];
 
1607
                    T2o = T2m + T2n;
 
1608
                    T3k = T2m - T2n;
 
1609
               }
 
1610
               {
 
1611
                    E Tn, To, T2q, T2r;
 
1612
                    Tn = ri[WS(is, 60)];
 
1613
                    To = ri[WS(is, 28)];
 
1614
                    Tp = Tn + To;
 
1615
                    T3q = Tn - To;
 
1616
                    T2q = ii[WS(is, 60)];
 
1617
                    T2r = ii[WS(is, 28)];
 
1618
                    T2s = T2q + T2r;
 
1619
                    T3o = T2q - T2r;
 
1620
               }
 
1621
               {
 
1622
                    E Tq, Tr, T2t, T2u;
 
1623
                    Tq = ri[WS(is, 12)];
 
1624
                    Tr = ri[WS(is, 44)];
 
1625
                    Ts = Tq + Tr;
 
1626
                    T3n = Tq - Tr;
 
1627
                    T2t = ii[WS(is, 12)];
 
1628
                    T2u = ii[WS(is, 44)];
 
1629
                    T2v = T2t + T2u;
 
1630
                    T3r = T2t - T2u;
 
1631
               }
 
1632
               {
 
1633
                    E Tm, Tt, Tai, Taj;
 
1634
                    Tm = Ti + Tl;
 
1635
                    Tt = Tp + Ts;
 
1636
                    Tu = Tm + Tt;
 
1637
                    TdI = Tt - Tm;
 
1638
                    Tai = T2l - T2o;
 
1639
                    Taj = Ti - Tl;
 
1640
                    Tak = Tai - Taj;
 
1641
                    TbD = Taj + Tai;
 
1642
               }
 
1643
               {
 
1644
                    E Tal, Tam, T2p, T2w;
 
1645
                    Tal = Tp - Ts;
 
1646
                    Tam = T2s - T2v;
 
1647
                    Tan = Tal + Tam;
 
1648
                    TbC = Tal - Tam;
 
1649
                    T2p = T2l + T2o;
 
1650
                    T2w = T2s + T2v;
 
1651
                    T2x = T2p + T2w;
 
1652
                    Tda = T2p - T2w;
 
1653
               }
 
1654
               {
 
1655
                    E T3i, T3l, T7E, T7F;
 
1656
                    T3i = T3g + T3h;
 
1657
                    T3l = T3j - T3k;
 
1658
                    T3m = FNMS(KP923879532, T3l, KP382683432 * T3i);
 
1659
                    T65 = FMA(KP923879532, T3i, KP382683432 * T3l);
 
1660
                    T7E = T3h - T3g;
 
1661
                    T7F = T3j + T3k;
 
1662
                    T7G = FNMS(KP382683432, T7F, KP923879532 * T7E);
 
1663
                    T8J = FMA(KP382683432, T7E, KP923879532 * T7F);
 
1664
               }
 
1665
               {
 
1666
                    E T7H, T7I, T3p, T3s;
 
1667
                    T7H = T3o - T3n;
 
1668
                    T7I = T3q + T3r;
 
1669
                    T7J = FMA(KP923879532, T7H, KP382683432 * T7I);
 
1670
                    T8I = FNMS(KP382683432, T7H, KP923879532 * T7I);
 
1671
                    T3p = T3n + T3o;
 
1672
                    T3s = T3q - T3r;
 
1673
                    T3t = FMA(KP382683432, T3p, KP923879532 * T3s);
 
1674
                    T64 = FNMS(KP923879532, T3p, KP382683432 * T3s);
 
1675
               }
 
1676
          }
 
1677
          {
 
1678
               E Ty, T3H, T2B, T3x, TB, T3w, T2E, T3I, TI, T3L, T2L, T3B, TF, T3K, T2I;
 
1679
               E T3E;
 
1680
               {
 
1681
                    E Tw, Tx, T2C, T2D;
 
1682
                    Tw = ri[WS(is, 2)];
 
1683
                    Tx = ri[WS(is, 34)];
 
1684
                    Ty = Tw + Tx;
 
1685
                    T3H = Tw - Tx;
 
1686
                    {
 
1687
                         E T2z, T2A, Tz, TA;
 
1688
                         T2z = ii[WS(is, 2)];
 
1689
                         T2A = ii[WS(is, 34)];
 
1690
                         T2B = T2z + T2A;
 
1691
                         T3x = T2z - T2A;
 
1692
                         Tz = ri[WS(is, 18)];
 
1693
                         TA = ri[WS(is, 50)];
 
1694
                         TB = Tz + TA;
 
1695
                         T3w = Tz - TA;
 
1696
                    }
 
1697
                    T2C = ii[WS(is, 18)];
 
1698
                    T2D = ii[WS(is, 50)];
 
1699
                    T2E = T2C + T2D;
 
1700
                    T3I = T2C - T2D;
 
1701
                    {
 
1702
                         E TG, TH, T3z, T2J, T2K, T3A;
 
1703
                         TG = ri[WS(is, 58)];
 
1704
                         TH = ri[WS(is, 26)];
 
1705
                         T3z = TG - TH;
 
1706
                         T2J = ii[WS(is, 58)];
 
1707
                         T2K = ii[WS(is, 26)];
 
1708
                         T3A = T2J - T2K;
 
1709
                         TI = TG + TH;
 
1710
                         T3L = T3z + T3A;
 
1711
                         T2L = T2J + T2K;
 
1712
                         T3B = T3z - T3A;
 
1713
                    }
 
1714
                    {
 
1715
                         E TD, TE, T3C, T2G, T2H, T3D;
 
1716
                         TD = ri[WS(is, 10)];
 
1717
                         TE = ri[WS(is, 42)];
 
1718
                         T3C = TD - TE;
 
1719
                         T2G = ii[WS(is, 10)];
 
1720
                         T2H = ii[WS(is, 42)];
 
1721
                         T3D = T2G - T2H;
 
1722
                         TF = TD + TE;
 
1723
                         T3K = T3D - T3C;
 
1724
                         T2I = T2G + T2H;
 
1725
                         T3E = T3C + T3D;
 
1726
                    }
 
1727
               }
 
1728
               {
 
1729
                    E TC, TJ, Taq, Tar;
 
1730
                    TC = Ty + TB;
 
1731
                    TJ = TF + TI;
 
1732
                    TK = TC + TJ;
 
1733
                    Tdd = TC - TJ;
 
1734
                    Taq = T2B - T2E;
 
1735
                    Tar = TI - TF;
 
1736
                    Tas = Taq - Tar;
 
1737
                    Tce = Tar + Taq;
 
1738
               }
 
1739
               {
 
1740
                    E Tat, Tau, T2F, T2M;
 
1741
                    Tat = Ty - TB;
 
1742
                    Tau = T2I - T2L;
 
1743
                    Tav = Tat - Tau;
 
1744
                    Tcf = Tat + Tau;
 
1745
                    T2F = T2B + T2E;
 
1746
                    T2M = T2I + T2L;
 
1747
                    T2N = T2F + T2M;
 
1748
                    Tdc = T2F - T2M;
 
1749
               }
 
1750
               {
 
1751
                    E T3y, T3F, T7M, T7N;
 
1752
                    T3y = T3w + T3x;
 
1753
                    T3F = KP707106781 * (T3B - T3E);
 
1754
                    T3G = T3y - T3F;
 
1755
                    T6G = T3y + T3F;
 
1756
                    T7M = T3x - T3w;
 
1757
                    T7N = KP707106781 * (T3K + T3L);
 
1758
                    T7O = T7M - T7N;
 
1759
                    T9k = T7M + T7N;
 
1760
               }
 
1761
               {
 
1762
                    E T7P, T7Q, T3J, T3M;
 
1763
                    T7P = T3H + T3I;
 
1764
                    T7Q = KP707106781 * (T3E + T3B);
 
1765
                    T7R = T7P - T7Q;
 
1766
                    T9l = T7P + T7Q;
 
1767
                    T3J = T3H - T3I;
 
1768
                    T3M = KP707106781 * (T3K - T3L);
 
1769
                    T3N = T3J - T3M;
 
1770
                    T6H = T3J + T3M;
 
1771
               }
 
1772
          }
 
1773
          {
 
1774
               E T1z, T53, T5L, Tbo, T1C, T5I, T56, Tbp, T1J, Tb9, T5h, T5N, T1G, Tb8, T5c;
 
1775
               E T5O;
 
1776
               {
 
1777
                    E T1x, T1y, T54, T55;
 
1778
                    T1x = ri[WS(is, 63)];
 
1779
                    T1y = ri[WS(is, 31)];
 
1780
                    T1z = T1x + T1y;
 
1781
                    T53 = T1x - T1y;
 
1782
                    {
 
1783
                         E T5J, T5K, T1A, T1B;
 
1784
                         T5J = ii[WS(is, 63)];
 
1785
                         T5K = ii[WS(is, 31)];
 
1786
                         T5L = T5J - T5K;
 
1787
                         Tbo = T5J + T5K;
 
1788
                         T1A = ri[WS(is, 15)];
 
1789
                         T1B = ri[WS(is, 47)];
 
1790
                         T1C = T1A + T1B;
 
1791
                         T5I = T1A - T1B;
 
1792
                    }
 
1793
                    T54 = ii[WS(is, 15)];
 
1794
                    T55 = ii[WS(is, 47)];
 
1795
                    T56 = T54 - T55;
 
1796
                    Tbp = T54 + T55;
 
1797
                    {
 
1798
                         E T1H, T1I, T5d, T5e, T5f, T5g;
 
1799
                         T1H = ri[WS(is, 55)];
 
1800
                         T1I = ri[WS(is, 23)];
 
1801
                         T5d = T1H - T1I;
 
1802
                         T5e = ii[WS(is, 55)];
 
1803
                         T5f = ii[WS(is, 23)];
 
1804
                         T5g = T5e - T5f;
 
1805
                         T1J = T1H + T1I;
 
1806
                         Tb9 = T5e + T5f;
 
1807
                         T5h = T5d + T5g;
 
1808
                         T5N = T5d - T5g;
 
1809
                    }
 
1810
                    {
 
1811
                         E T1E, T1F, T5b, T58, T59, T5a;
 
1812
                         T1E = ri[WS(is, 7)];
 
1813
                         T1F = ri[WS(is, 39)];
 
1814
                         T5b = T1E - T1F;
 
1815
                         T58 = ii[WS(is, 7)];
 
1816
                         T59 = ii[WS(is, 39)];
 
1817
                         T5a = T58 - T59;
 
1818
                         T1G = T1E + T1F;
 
1819
                         Tb8 = T58 + T59;
 
1820
                         T5c = T5a - T5b;
 
1821
                         T5O = T5b + T5a;
 
1822
                    }
 
1823
               }
 
1824
               {
 
1825
                    E T1D, T1K, Tbq, Tbr;
 
1826
                    T1D = T1z + T1C;
 
1827
                    T1K = T1G + T1J;
 
1828
                    T1L = T1D + T1K;
 
1829
                    Tdv = T1D - T1K;
 
1830
                    Tbq = Tbo - Tbp;
 
1831
                    Tbr = T1J - T1G;
 
1832
                    Tbs = Tbq - Tbr;
 
1833
                    Tcw = Tbr + Tbq;
 
1834
               }
 
1835
               {
 
1836
                    E TdA, TdB, T57, T5i;
 
1837
                    TdA = Tbo + Tbp;
 
1838
                    TdB = Tb8 + Tb9;
 
1839
                    TdC = TdA - TdB;
 
1840
                    Teo = TdA + TdB;
 
1841
                    T57 = T53 - T56;
 
1842
                    T5i = KP707106781 * (T5c - T5h);
 
1843
                    T5j = T57 - T5i;
 
1844
                    T6V = T57 + T5i;
 
1845
               }
 
1846
               {
 
1847
                    E T5M, T5P, T8w, T8x;
 
1848
                    T5M = T5I + T5L;
 
1849
                    T5P = KP707106781 * (T5N - T5O);
 
1850
                    T5Q = T5M - T5P;
 
1851
                    T6Y = T5M + T5P;
 
1852
                    T8w = T5L - T5I;
 
1853
                    T8x = KP707106781 * (T5c + T5h);
 
1854
                    T8y = T8w - T8x;
 
1855
                    T9C = T8w + T8x;
 
1856
               }
 
1857
               {
 
1858
                    E Tb7, Tba, T8l, T8m;
 
1859
                    Tb7 = T1z - T1C;
 
1860
                    Tba = Tb8 - Tb9;
 
1861
                    Tbb = Tb7 - Tba;
 
1862
                    Tct = Tb7 + Tba;
 
1863
                    T8l = T53 + T56;
 
1864
                    T8m = KP707106781 * (T5O + T5N);
 
1865
                    T8n = T8l - T8m;
 
1866
                    T9z = T8l + T8m;
 
1867
               }
 
1868
          }
 
1869
          {
 
1870
               E TN, T40, T2Q, T3Q, TQ, T3P, T2T, T41, TX, T44, T30, T3U, TU, T43, T2X;
 
1871
               E T3X;
 
1872
               {
 
1873
                    E TL, TM, T2R, T2S;
 
1874
                    TL = ri[WS(is, 62)];
 
1875
                    TM = ri[WS(is, 30)];
 
1876
                    TN = TL + TM;
 
1877
                    T40 = TL - TM;
 
1878
                    {
 
1879
                         E T2O, T2P, TO, TP;
 
1880
                         T2O = ii[WS(is, 62)];
 
1881
                         T2P = ii[WS(is, 30)];
 
1882
                         T2Q = T2O + T2P;
 
1883
                         T3Q = T2O - T2P;
 
1884
                         TO = ri[WS(is, 14)];
 
1885
                         TP = ri[WS(is, 46)];
 
1886
                         TQ = TO + TP;
 
1887
                         T3P = TO - TP;
 
1888
                    }
 
1889
                    T2R = ii[WS(is, 14)];
 
1890
                    T2S = ii[WS(is, 46)];
 
1891
                    T2T = T2R + T2S;
 
1892
                    T41 = T2R - T2S;
 
1893
                    {
 
1894
                         E TV, TW, T3S, T2Y, T2Z, T3T;
 
1895
                         TV = ri[WS(is, 54)];
 
1896
                         TW = ri[WS(is, 22)];
 
1897
                         T3S = TV - TW;
 
1898
                         T2Y = ii[WS(is, 54)];
 
1899
                         T2Z = ii[WS(is, 22)];
 
1900
                         T3T = T2Y - T2Z;
 
1901
                         TX = TV + TW;
 
1902
                         T44 = T3S + T3T;
 
1903
                         T30 = T2Y + T2Z;
 
1904
                         T3U = T3S - T3T;
 
1905
                    }
 
1906
                    {
 
1907
                         E TS, TT, T3V, T2V, T2W, T3W;
 
1908
                         TS = ri[WS(is, 6)];
 
1909
                         TT = ri[WS(is, 38)];
 
1910
                         T3V = TS - TT;
 
1911
                         T2V = ii[WS(is, 6)];
 
1912
                         T2W = ii[WS(is, 38)];
 
1913
                         T3W = T2V - T2W;
 
1914
                         TU = TS + TT;
 
1915
                         T43 = T3W - T3V;
 
1916
                         T2X = T2V + T2W;
 
1917
                         T3X = T3V + T3W;
 
1918
                    }
 
1919
               }
 
1920
               {
 
1921
                    E TR, TY, Tax, Tay;
 
1922
                    TR = TN + TQ;
 
1923
                    TY = TU + TX;
 
1924
                    TZ = TR + TY;
 
1925
                    Tdf = TR - TY;
 
1926
                    Tax = T2Q - T2T;
 
1927
                    Tay = TX - TU;
 
1928
                    Taz = Tax - Tay;
 
1929
                    Tch = Tay + Tax;
 
1930
               }
 
1931
               {
 
1932
                    E TaA, TaB, T2U, T31;
 
1933
                    TaA = TN - TQ;
 
1934
                    TaB = T2X - T30;
 
1935
                    TaC = TaA - TaB;
 
1936
                    Tci = TaA + TaB;
 
1937
                    T2U = T2Q + T2T;
 
1938
                    T31 = T2X + T30;
 
1939
                    T32 = T2U + T31;
 
1940
                    Tdg = T2U - T31;
 
1941
               }
 
1942
               {
 
1943
                    E T3R, T3Y, T7T, T7U;
 
1944
                    T3R = T3P + T3Q;
 
1945
                    T3Y = KP707106781 * (T3U - T3X);
 
1946
                    T3Z = T3R - T3Y;
 
1947
                    T6J = T3R + T3Y;
 
1948
                    T7T = T40 + T41;
 
1949
                    T7U = KP707106781 * (T3X + T3U);
 
1950
                    T7V = T7T - T7U;
 
1951
                    T9n = T7T + T7U;
 
1952
               }
 
1953
               {
 
1954
                    E T7W, T7X, T42, T45;
 
1955
                    T7W = T3Q - T3P;
 
1956
                    T7X = KP707106781 * (T43 + T44);
 
1957
                    T7Y = T7W - T7X;
 
1958
                    T9o = T7W + T7X;
 
1959
                    T42 = T40 - T41;
 
1960
                    T45 = KP707106781 * (T43 - T44);
 
1961
                    T46 = T42 - T45;
 
1962
                    T6K = T42 + T45;
 
1963
               }
 
1964
          }
 
1965
          {
 
1966
               E T14, T4P, T4d, TaG, T17, T4a, T4S, TaH, T1e, TaZ, T4j, T4V, T1b, TaY, T4o;
 
1967
               E T4U;
 
1968
               {
 
1969
                    E T12, T13, T4Q, T4R;
 
1970
                    T12 = ri[WS(is, 1)];
 
1971
                    T13 = ri[WS(is, 33)];
 
1972
                    T14 = T12 + T13;
 
1973
                    T4P = T12 - T13;
 
1974
                    {
 
1975
                         E T4b, T4c, T15, T16;
 
1976
                         T4b = ii[WS(is, 1)];
 
1977
                         T4c = ii[WS(is, 33)];
 
1978
                         T4d = T4b - T4c;
 
1979
                         TaG = T4b + T4c;
 
1980
                         T15 = ri[WS(is, 17)];
 
1981
                         T16 = ri[WS(is, 49)];
 
1982
                         T17 = T15 + T16;
 
1983
                         T4a = T15 - T16;
 
1984
                    }
 
1985
                    T4Q = ii[WS(is, 17)];
 
1986
                    T4R = ii[WS(is, 49)];
 
1987
                    T4S = T4Q - T4R;
 
1988
                    TaH = T4Q + T4R;
 
1989
                    {
 
1990
                         E T1c, T1d, T4f, T4g, T4h, T4i;
 
1991
                         T1c = ri[WS(is, 57)];
 
1992
                         T1d = ri[WS(is, 25)];
 
1993
                         T4f = T1c - T1d;
 
1994
                         T4g = ii[WS(is, 57)];
 
1995
                         T4h = ii[WS(is, 25)];
 
1996
                         T4i = T4g - T4h;
 
1997
                         T1e = T1c + T1d;
 
1998
                         TaZ = T4g + T4h;
 
1999
                         T4j = T4f - T4i;
 
2000
                         T4V = T4f + T4i;
 
2001
                    }
 
2002
                    {
 
2003
                         E T19, T1a, T4k, T4l, T4m, T4n;
 
2004
                         T19 = ri[WS(is, 9)];
 
2005
                         T1a = ri[WS(is, 41)];
 
2006
                         T4k = T19 - T1a;
 
2007
                         T4l = ii[WS(is, 9)];
 
2008
                         T4m = ii[WS(is, 41)];
 
2009
                         T4n = T4l - T4m;
 
2010
                         T1b = T19 + T1a;
 
2011
                         TaY = T4l + T4m;
 
2012
                         T4o = T4k + T4n;
 
2013
                         T4U = T4n - T4k;
 
2014
                    }
 
2015
               }
 
2016
               {
 
2017
                    E T18, T1f, TaX, Tb0;
 
2018
                    T18 = T14 + T17;
 
2019
                    T1f = T1b + T1e;
 
2020
                    T1g = T18 + T1f;
 
2021
                    Tdp = T18 - T1f;
 
2022
                    TaX = T14 - T17;
 
2023
                    Tb0 = TaY - TaZ;
 
2024
                    Tb1 = TaX - Tb0;
 
2025
                    Tcm = TaX + Tb0;
 
2026
               }
 
2027
               {
 
2028
                    E Tdk, Tdl, T4e, T4p;
 
2029
                    Tdk = TaG + TaH;
 
2030
                    Tdl = TaY + TaZ;
 
2031
                    Tdm = Tdk - Tdl;
 
2032
                    Tej = Tdk + Tdl;
 
2033
                    T4e = T4a + T4d;
 
2034
                    T4p = KP707106781 * (T4j - T4o);
 
2035
                    T4q = T4e - T4p;
 
2036
                    T6R = T4e + T4p;
 
2037
               }
 
2038
               {
 
2039
                    E T4T, T4W, T8d, T8e;
 
2040
                    T4T = T4P - T4S;
 
2041
                    T4W = KP707106781 * (T4U - T4V);
 
2042
                    T4X = T4T - T4W;
 
2043
                    T6O = T4T + T4W;
 
2044
                    T8d = T4P + T4S;
 
2045
                    T8e = KP707106781 * (T4o + T4j);
 
2046
                    T8f = T8d - T8e;
 
2047
                    T9s = T8d + T8e;
 
2048
               }
 
2049
               {
 
2050
                    E TaI, TaJ, T82, T83;
 
2051
                    TaI = TaG - TaH;
 
2052
                    TaJ = T1e - T1b;
 
2053
                    TaK = TaI - TaJ;
 
2054
                    Tcp = TaJ + TaI;
 
2055
                    T82 = T4d - T4a;
 
2056
                    T83 = KP707106781 * (T4U + T4V);
 
2057
                    T84 = T82 - T83;
 
2058
                    T9v = T82 + T83;
 
2059
               }
 
2060
          }
 
2061
          {
 
2062
               E T1j, TaR, T1m, TaS, T4G, T4L, TaT, TaQ, T89, T88, T1q, TaM, T1t, TaN, T4v;
 
2063
               E T4A, TaO, TaL, T86, T85;
 
2064
               {
 
2065
                    E T4H, T4F, T4C, T4K;
 
2066
                    {
 
2067
                         E T1h, T1i, T4D, T4E;
 
2068
                         T1h = ri[WS(is, 5)];
 
2069
                         T1i = ri[WS(is, 37)];
 
2070
                         T1j = T1h + T1i;
 
2071
                         T4H = T1h - T1i;
 
2072
                         T4D = ii[WS(is, 5)];
 
2073
                         T4E = ii[WS(is, 37)];
 
2074
                         T4F = T4D - T4E;
 
2075
                         TaR = T4D + T4E;
 
2076
                    }
 
2077
                    {
 
2078
                         E T1k, T1l, T4I, T4J;
 
2079
                         T1k = ri[WS(is, 21)];
 
2080
                         T1l = ri[WS(is, 53)];
 
2081
                         T1m = T1k + T1l;
 
2082
                         T4C = T1k - T1l;
 
2083
                         T4I = ii[WS(is, 21)];
 
2084
                         T4J = ii[WS(is, 53)];
 
2085
                         T4K = T4I - T4J;
 
2086
                         TaS = T4I + T4J;
 
2087
                    }
 
2088
                    T4G = T4C + T4F;
 
2089
                    T4L = T4H - T4K;
 
2090
                    TaT = TaR - TaS;
 
2091
                    TaQ = T1j - T1m;
 
2092
                    T89 = T4H + T4K;
 
2093
                    T88 = T4F - T4C;
 
2094
               }
 
2095
               {
 
2096
                    E T4r, T4z, T4w, T4u;
 
2097
                    {
 
2098
                         E T1o, T1p, T4x, T4y;
 
2099
                         T1o = ri[WS(is, 61)];
 
2100
                         T1p = ri[WS(is, 29)];
 
2101
                         T1q = T1o + T1p;
 
2102
                         T4r = T1o - T1p;
 
2103
                         T4x = ii[WS(is, 61)];
 
2104
                         T4y = ii[WS(is, 29)];
 
2105
                         T4z = T4x - T4y;
 
2106
                         TaM = T4x + T4y;
 
2107
                    }
 
2108
                    {
 
2109
                         E T1r, T1s, T4s, T4t;
 
2110
                         T1r = ri[WS(is, 13)];
 
2111
                         T1s = ri[WS(is, 45)];
 
2112
                         T1t = T1r + T1s;
 
2113
                         T4w = T1r - T1s;
 
2114
                         T4s = ii[WS(is, 13)];
 
2115
                         T4t = ii[WS(is, 45)];
 
2116
                         T4u = T4s - T4t;
 
2117
                         TaN = T4s + T4t;
 
2118
                    }
 
2119
                    T4v = T4r - T4u;
 
2120
                    T4A = T4w + T4z;
 
2121
                    TaO = TaM - TaN;
 
2122
                    TaL = T1q - T1t;
 
2123
                    T86 = T4z - T4w;
 
2124
                    T85 = T4r + T4u;
 
2125
               }
 
2126
               {
 
2127
                    E T1n, T1u, Tb2, Tb3;
 
2128
                    T1n = T1j + T1m;
 
2129
                    T1u = T1q + T1t;
 
2130
                    T1v = T1n + T1u;
 
2131
                    Tdn = T1u - T1n;
 
2132
                    Tb2 = TaT - TaQ;
 
2133
                    Tb3 = TaL + TaO;
 
2134
                    Tb4 = KP707106781 * (Tb2 - Tb3);
 
2135
                    Tcq = KP707106781 * (Tb2 + Tb3);
 
2136
               }
 
2137
               {
 
2138
                    E Tdq, Tdr, T4B, T4M;
 
2139
                    Tdq = TaR + TaS;
 
2140
                    Tdr = TaM + TaN;
 
2141
                    Tds = Tdq - Tdr;
 
2142
                    Tek = Tdq + Tdr;
 
2143
                    T4B = FNMS(KP923879532, T4A, KP382683432 * T4v);
 
2144
                    T4M = FMA(KP923879532, T4G, KP382683432 * T4L);
 
2145
                    T4N = T4B - T4M;
 
2146
                    T6P = T4M + T4B;
 
2147
               }
 
2148
               {
 
2149
                    E T4Y, T4Z, T8g, T8h;
 
2150
                    T4Y = FNMS(KP923879532, T4L, KP382683432 * T4G);
 
2151
                    T4Z = FMA(KP382683432, T4A, KP923879532 * T4v);
 
2152
                    T50 = T4Y - T4Z;
 
2153
                    T6S = T4Y + T4Z;
 
2154
                    T8g = FNMS(KP382683432, T89, KP923879532 * T88);
 
2155
                    T8h = FMA(KP923879532, T86, KP382683432 * T85);
 
2156
                    T8i = T8g - T8h;
 
2157
                    T9w = T8g + T8h;
 
2158
               }
 
2159
               {
 
2160
                    E TaP, TaU, T87, T8a;
 
2161
                    TaP = TaL - TaO;
 
2162
                    TaU = TaQ + TaT;
 
2163
                    TaV = KP707106781 * (TaP - TaU);
 
2164
                    Tcn = KP707106781 * (TaU + TaP);
 
2165
                    T87 = FNMS(KP382683432, T86, KP923879532 * T85);
 
2166
                    T8a = FMA(KP382683432, T88, KP923879532 * T89);
 
2167
                    T8b = T87 - T8a;
 
2168
                    T9t = T8a + T87;
 
2169
               }
 
2170
          }
 
2171
          {
 
2172
               E T1O, Tbc, T1R, Tbd, T5o, T5t, Tbf, Tbe, T8p, T8o, T1V, Tbi, T1Y, Tbj, T5z;
 
2173
               E T5E, Tbk, Tbh, T8s, T8r;
 
2174
               {
 
2175
                    E T5p, T5n, T5k, T5s;
 
2176
                    {
 
2177
                         E T1M, T1N, T5l, T5m;
 
2178
                         T1M = ri[WS(is, 3)];
 
2179
                         T1N = ri[WS(is, 35)];
 
2180
                         T1O = T1M + T1N;
 
2181
                         T5p = T1M - T1N;
 
2182
                         T5l = ii[WS(is, 3)];
 
2183
                         T5m = ii[WS(is, 35)];
 
2184
                         T5n = T5l - T5m;
 
2185
                         Tbc = T5l + T5m;
 
2186
                    }
 
2187
                    {
 
2188
                         E T1P, T1Q, T5q, T5r;
 
2189
                         T1P = ri[WS(is, 19)];
 
2190
                         T1Q = ri[WS(is, 51)];
 
2191
                         T1R = T1P + T1Q;
 
2192
                         T5k = T1P - T1Q;
 
2193
                         T5q = ii[WS(is, 19)];
 
2194
                         T5r = ii[WS(is, 51)];
 
2195
                         T5s = T5q - T5r;
 
2196
                         Tbd = T5q + T5r;
 
2197
                    }
 
2198
                    T5o = T5k + T5n;
 
2199
                    T5t = T5p - T5s;
 
2200
                    Tbf = T1O - T1R;
 
2201
                    Tbe = Tbc - Tbd;
 
2202
                    T8p = T5p + T5s;
 
2203
                    T8o = T5n - T5k;
 
2204
               }
 
2205
               {
 
2206
                    E T5A, T5y, T5v, T5D;
 
2207
                    {
 
2208
                         E T1T, T1U, T5w, T5x;
 
2209
                         T1T = ri[WS(is, 59)];
 
2210
                         T1U = ri[WS(is, 27)];
 
2211
                         T1V = T1T + T1U;
 
2212
                         T5A = T1T - T1U;
 
2213
                         T5w = ii[WS(is, 59)];
 
2214
                         T5x = ii[WS(is, 27)];
 
2215
                         T5y = T5w - T5x;
 
2216
                         Tbi = T5w + T5x;
 
2217
                    }
 
2218
                    {
 
2219
                         E T1W, T1X, T5B, T5C;
 
2220
                         T1W = ri[WS(is, 11)];
 
2221
                         T1X = ri[WS(is, 43)];
 
2222
                         T1Y = T1W + T1X;
 
2223
                         T5v = T1W - T1X;
 
2224
                         T5B = ii[WS(is, 11)];
 
2225
                         T5C = ii[WS(is, 43)];
 
2226
                         T5D = T5B - T5C;
 
2227
                         Tbj = T5B + T5C;
 
2228
                    }
 
2229
                    T5z = T5v + T5y;
 
2230
                    T5E = T5A - T5D;
 
2231
                    Tbk = Tbi - Tbj;
 
2232
                    Tbh = T1V - T1Y;
 
2233
                    T8s = T5A + T5D;
 
2234
                    T8r = T5y - T5v;
 
2235
               }
 
2236
               {
 
2237
                    E T1S, T1Z, Tbt, Tbu;
 
2238
                    T1S = T1O + T1R;
 
2239
                    T1Z = T1V + T1Y;
 
2240
                    T20 = T1S + T1Z;
 
2241
                    TdD = T1Z - T1S;
 
2242
                    Tbt = Tbh - Tbk;
 
2243
                    Tbu = Tbf + Tbe;
 
2244
                    Tbv = KP707106781 * (Tbt - Tbu);
 
2245
                    Tcu = KP707106781 * (Tbu + Tbt);
 
2246
               }
 
2247
               {
 
2248
                    E Tdw, Tdx, T5u, T5F;
 
2249
                    Tdw = Tbc + Tbd;
 
2250
                    Tdx = Tbi + Tbj;
 
2251
                    Tdy = Tdw - Tdx;
 
2252
                    Tep = Tdw + Tdx;
 
2253
                    T5u = FNMS(KP923879532, T5t, KP382683432 * T5o);
 
2254
                    T5F = FMA(KP382683432, T5z, KP923879532 * T5E);
 
2255
                    T5G = T5u - T5F;
 
2256
                    T6Z = T5u + T5F;
 
2257
               }
 
2258
               {
 
2259
                    E T5R, T5S, T8z, T8A;
 
2260
                    T5R = FNMS(KP923879532, T5z, KP382683432 * T5E);
 
2261
                    T5S = FMA(KP923879532, T5o, KP382683432 * T5t);
 
2262
                    T5T = T5R - T5S;
 
2263
                    T6W = T5S + T5R;
 
2264
                    T8z = FNMS(KP382683432, T8r, KP923879532 * T8s);
 
2265
                    T8A = FMA(KP382683432, T8o, KP923879532 * T8p);
 
2266
                    T8B = T8z - T8A;
 
2267
                    T9A = T8A + T8z;
 
2268
               }
 
2269
               {
 
2270
                    E Tbg, Tbl, T8q, T8t;
 
2271
                    Tbg = Tbe - Tbf;
 
2272
                    Tbl = Tbh + Tbk;
 
2273
                    Tbm = KP707106781 * (Tbg - Tbl);
 
2274
                    Tcx = KP707106781 * (Tbg + Tbl);
 
2275
                    T8q = FNMS(KP382683432, T8p, KP923879532 * T8o);
 
2276
                    T8t = FMA(KP923879532, T8r, KP382683432 * T8s);
 
2277
                    T8u = T8q - T8t;
 
2278
                    T9D = T8q + T8t;
 
2279
               }
 
2280
          }
 
2281
          {
 
2282
               E T11, TeD, TeG, TeI, T22, T23, T34, TeH;
 
2283
               {
 
2284
                    E Tv, T10, TeE, TeF;
 
2285
                    Tv = Tf + Tu;
 
2286
                    T10 = TK + TZ;
 
2287
                    T11 = Tv + T10;
 
2288
                    TeD = Tv - T10;
 
2289
                    TeE = Tej + Tek;
 
2290
                    TeF = Teo + Tep;
 
2291
                    TeG = TeE - TeF;
 
2292
                    TeI = TeE + TeF;
 
2293
               }
 
2294
               {
 
2295
                    E T1w, T21, T2y, T33;
 
2296
                    T1w = T1g + T1v;
 
2297
                    T21 = T1L + T20;
 
2298
                    T22 = T1w + T21;
 
2299
                    T23 = T21 - T1w;
 
2300
                    T2y = T2i + T2x;
 
2301
                    T33 = T2N + T32;
 
2302
                    T34 = T2y - T33;
 
2303
                    TeH = T2y + T33;
 
2304
               }
 
2305
               ro[WS(os, 32)] = T11 - T22;
 
2306
               io[WS(os, 32)] = TeH - TeI;
 
2307
               ro[0] = T11 + T22;
 
2308
               io[0] = TeH + TeI;
 
2309
               io[WS(os, 16)] = T23 + T34;
 
2310
               ro[WS(os, 16)] = TeD + TeG;
 
2311
               io[WS(os, 48)] = T34 - T23;
 
2312
               ro[WS(os, 48)] = TeD - TeG;
 
2313
          }
 
2314
          {
 
2315
               E Teh, Tex, Tev, TeB, Tem, Tey, Ter, Tez;
 
2316
               {
 
2317
                    E Tef, Teg, Tet, Teu;
 
2318
                    Tef = Tf - Tu;
 
2319
                    Teg = T2N - T32;
 
2320
                    Teh = Tef + Teg;
 
2321
                    Tex = Tef - Teg;
 
2322
                    Tet = T2i - T2x;
 
2323
                    Teu = TZ - TK;
 
2324
                    Tev = Tet - Teu;
 
2325
                    TeB = Teu + Tet;
 
2326
               }
 
2327
               {
 
2328
                    E Tei, Tel, Ten, Teq;
 
2329
                    Tei = T1g - T1v;
 
2330
                    Tel = Tej - Tek;
 
2331
                    Tem = Tei + Tel;
 
2332
                    Tey = Tel - Tei;
 
2333
                    Ten = T1L - T20;
 
2334
                    Teq = Teo - Tep;
 
2335
                    Ter = Ten - Teq;
 
2336
                    Tez = Ten + Teq;
 
2337
               }
 
2338
               {
 
2339
                    E Tes, TeC, Tew, TeA;
 
2340
                    Tes = KP707106781 * (Tem + Ter);
 
2341
                    ro[WS(os, 40)] = Teh - Tes;
 
2342
                    ro[WS(os, 8)] = Teh + Tes;
 
2343
                    TeC = KP707106781 * (Tey + Tez);
 
2344
                    io[WS(os, 40)] = TeB - TeC;
 
2345
                    io[WS(os, 8)] = TeB + TeC;
 
2346
                    Tew = KP707106781 * (Ter - Tem);
 
2347
                    io[WS(os, 56)] = Tev - Tew;
 
2348
                    io[WS(os, 24)] = Tev + Tew;
 
2349
                    TeA = KP707106781 * (Tey - Tez);
 
2350
                    ro[WS(os, 56)] = Tex - TeA;
 
2351
                    ro[WS(os, 24)] = Tex + TeA;
 
2352
               }
 
2353
          }
 
2354
          {
 
2355
               E Tdb, TdV, Te5, TdJ, Tdi, Te6, Te3, Teb, TdM, TdW, Tdu, TdQ, Te0, Tea, TdF;
 
2356
               E TdR;
 
2357
               {
 
2358
                    E Tde, Tdh, Tdo, Tdt;
 
2359
                    Tdb = Td9 - Tda;
 
2360
                    TdV = Td9 + Tda;
 
2361
                    Te5 = TdI + TdH;
 
2362
                    TdJ = TdH - TdI;
 
2363
                    Tde = Tdc - Tdd;
 
2364
                    Tdh = Tdf + Tdg;
 
2365
                    Tdi = KP707106781 * (Tde - Tdh);
 
2366
                    Te6 = KP707106781 * (Tde + Tdh);
 
2367
                    {
 
2368
                         E Te1, Te2, TdK, TdL;
 
2369
                         Te1 = Tdv + Tdy;
 
2370
                         Te2 = TdD + TdC;
 
2371
                         Te3 = FNMS(KP382683432, Te2, KP923879532 * Te1);
 
2372
                         Teb = FMA(KP923879532, Te2, KP382683432 * Te1);
 
2373
                         TdK = Tdf - Tdg;
 
2374
                         TdL = Tdd + Tdc;
 
2375
                         TdM = KP707106781 * (TdK - TdL);
 
2376
                         TdW = KP707106781 * (TdL + TdK);
 
2377
                    }
 
2378
                    Tdo = Tdm - Tdn;
 
2379
                    Tdt = Tdp - Tds;
 
2380
                    Tdu = FMA(KP923879532, Tdo, KP382683432 * Tdt);
 
2381
                    TdQ = FNMS(KP923879532, Tdt, KP382683432 * Tdo);
 
2382
                    {
 
2383
                         E TdY, TdZ, Tdz, TdE;
 
2384
                         TdY = Tdn + Tdm;
 
2385
                         TdZ = Tdp + Tds;
 
2386
                         Te0 = FMA(KP382683432, TdY, KP923879532 * TdZ);
 
2387
                         Tea = FNMS(KP382683432, TdZ, KP923879532 * TdY);
 
2388
                         Tdz = Tdv - Tdy;
 
2389
                         TdE = TdC - TdD;
 
2390
                         TdF = FNMS(KP923879532, TdE, KP382683432 * Tdz);
 
2391
                         TdR = FMA(KP382683432, TdE, KP923879532 * Tdz);
 
2392
                    }
 
2393
               }
 
2394
               {
 
2395
                    E Tdj, TdG, TdT, TdU;
 
2396
                    Tdj = Tdb + Tdi;
 
2397
                    TdG = Tdu + TdF;
 
2398
                    ro[WS(os, 44)] = Tdj - TdG;
 
2399
                    ro[WS(os, 12)] = Tdj + TdG;
 
2400
                    TdT = TdJ + TdM;
 
2401
                    TdU = TdQ + TdR;
 
2402
                    io[WS(os, 44)] = TdT - TdU;
 
2403
                    io[WS(os, 12)] = TdT + TdU;
 
2404
               }
 
2405
               {
 
2406
                    E TdN, TdO, TdP, TdS;
 
2407
                    TdN = TdJ - TdM;
 
2408
                    TdO = TdF - Tdu;
 
2409
                    io[WS(os, 60)] = TdN - TdO;
 
2410
                    io[WS(os, 28)] = TdN + TdO;
 
2411
                    TdP = Tdb - Tdi;
 
2412
                    TdS = TdQ - TdR;
 
2413
                    ro[WS(os, 60)] = TdP - TdS;
 
2414
                    ro[WS(os, 28)] = TdP + TdS;
 
2415
               }
 
2416
               {
 
2417
                    E TdX, Te4, Ted, Tee;
 
2418
                    TdX = TdV + TdW;
 
2419
                    Te4 = Te0 + Te3;
 
2420
                    ro[WS(os, 36)] = TdX - Te4;
 
2421
                    ro[WS(os, 4)] = TdX + Te4;
 
2422
                    Ted = Te5 + Te6;
 
2423
                    Tee = Tea + Teb;
 
2424
                    io[WS(os, 36)] = Ted - Tee;
 
2425
                    io[WS(os, 4)] = Ted + Tee;
 
2426
               }
 
2427
               {
 
2428
                    E Te7, Te8, Te9, Tec;
 
2429
                    Te7 = Te5 - Te6;
 
2430
                    Te8 = Te3 - Te0;
 
2431
                    io[WS(os, 52)] = Te7 - Te8;
 
2432
                    io[WS(os, 20)] = Te7 + Te8;
 
2433
                    Te9 = TdV - TdW;
 
2434
                    Tec = Tea - Teb;
 
2435
                    ro[WS(os, 52)] = Te9 - Tec;
 
2436
                    ro[WS(os, 20)] = Te9 + Tec;
 
2437
               }
 
2438
          }
 
2439
          {
 
2440
               E Tcd, TcP, TcD, TcZ, Tck, Td0, TcX, Td5, Tcs, TcK, TcG, TcQ, TcU, Td4, Tcz;
 
2441
               E TcL, Tcc, TcC;
 
2442
               Tcc = KP707106781 * (TbD + TbC);
 
2443
               Tcd = Tcb - Tcc;
 
2444
               TcP = Tcb + Tcc;
 
2445
               TcC = KP707106781 * (Tak + Tan);
 
2446
               TcD = TcB - TcC;
 
2447
               TcZ = TcB + TcC;
 
2448
               {
 
2449
                    E Tcg, Tcj, TcV, TcW;
 
2450
                    Tcg = FNMS(KP382683432, Tcf, KP923879532 * Tce);
 
2451
                    Tcj = FMA(KP923879532, Tch, KP382683432 * Tci);
 
2452
                    Tck = Tcg - Tcj;
 
2453
                    Td0 = Tcg + Tcj;
 
2454
                    TcV = Tct + Tcu;
 
2455
                    TcW = Tcw + Tcx;
 
2456
                    TcX = FNMS(KP195090322, TcW, KP980785280 * TcV);
 
2457
                    Td5 = FMA(KP195090322, TcV, KP980785280 * TcW);
 
2458
               }
 
2459
               {
 
2460
                    E Tco, Tcr, TcE, TcF;
 
2461
                    Tco = Tcm - Tcn;
 
2462
                    Tcr = Tcp - Tcq;
 
2463
                    Tcs = FMA(KP555570233, Tco, KP831469612 * Tcr);
 
2464
                    TcK = FNMS(KP831469612, Tco, KP555570233 * Tcr);
 
2465
                    TcE = FNMS(KP382683432, Tch, KP923879532 * Tci);
 
2466
                    TcF = FMA(KP382683432, Tce, KP923879532 * Tcf);
 
2467
                    TcG = TcE - TcF;
 
2468
                    TcQ = TcF + TcE;
 
2469
               }
 
2470
               {
 
2471
                    E TcS, TcT, Tcv, Tcy;
 
2472
                    TcS = Tcm + Tcn;
 
2473
                    TcT = Tcp + Tcq;
 
2474
                    TcU = FMA(KP980785280, TcS, KP195090322 * TcT);
 
2475
                    Td4 = FNMS(KP195090322, TcS, KP980785280 * TcT);
 
2476
                    Tcv = Tct - Tcu;
 
2477
                    Tcy = Tcw - Tcx;
 
2478
                    Tcz = FNMS(KP831469612, Tcy, KP555570233 * Tcv);
 
2479
                    TcL = FMA(KP831469612, Tcv, KP555570233 * Tcy);
 
2480
               }
 
2481
               {
 
2482
                    E Tcl, TcA, TcN, TcO;
 
2483
                    Tcl = Tcd + Tck;
 
2484
                    TcA = Tcs + Tcz;
 
2485
                    ro[WS(os, 42)] = Tcl - TcA;
 
2486
                    ro[WS(os, 10)] = Tcl + TcA;
 
2487
                    TcN = TcD + TcG;
 
2488
                    TcO = TcK + TcL;
 
2489
                    io[WS(os, 42)] = TcN - TcO;
 
2490
                    io[WS(os, 10)] = TcN + TcO;
 
2491
               }
 
2492
               {
 
2493
                    E TcH, TcI, TcJ, TcM;
 
2494
                    TcH = TcD - TcG;
 
2495
                    TcI = Tcz - Tcs;
 
2496
                    io[WS(os, 58)] = TcH - TcI;
 
2497
                    io[WS(os, 26)] = TcH + TcI;
 
2498
                    TcJ = Tcd - Tck;
 
2499
                    TcM = TcK - TcL;
 
2500
                    ro[WS(os, 58)] = TcJ - TcM;
 
2501
                    ro[WS(os, 26)] = TcJ + TcM;
 
2502
               }
 
2503
               {
 
2504
                    E TcR, TcY, Td7, Td8;
 
2505
                    TcR = TcP + TcQ;
 
2506
                    TcY = TcU + TcX;
 
2507
                    ro[WS(os, 34)] = TcR - TcY;
 
2508
                    ro[WS(os, 2)] = TcR + TcY;
 
2509
                    Td7 = TcZ + Td0;
 
2510
                    Td8 = Td4 + Td5;
 
2511
                    io[WS(os, 34)] = Td7 - Td8;
 
2512
                    io[WS(os, 2)] = Td7 + Td8;
 
2513
               }
 
2514
               {
 
2515
                    E Td1, Td2, Td3, Td6;
 
2516
                    Td1 = TcZ - Td0;
 
2517
                    Td2 = TcX - TcU;
 
2518
                    io[WS(os, 50)] = Td1 - Td2;
 
2519
                    io[WS(os, 18)] = Td1 + Td2;
 
2520
                    Td3 = TcP - TcQ;
 
2521
                    Td6 = Td4 - Td5;
 
2522
                    ro[WS(os, 50)] = Td3 - Td6;
 
2523
                    ro[WS(os, 18)] = Td3 + Td6;
 
2524
               }
 
2525
          }
 
2526
          {
 
2527
               E Tap, TbR, TbF, Tc1, TaE, Tc2, TbZ, Tc7, Tb6, TbM, TbI, TbS, TbW, Tc6, Tbx;
 
2528
               E TbN, Tao, TbE;
 
2529
               Tao = KP707106781 * (Tak - Tan);
 
2530
               Tap = Tah - Tao;
 
2531
               TbR = Tah + Tao;
 
2532
               TbE = KP707106781 * (TbC - TbD);
 
2533
               TbF = TbB - TbE;
 
2534
               Tc1 = TbB + TbE;
 
2535
               {
 
2536
                    E Taw, TaD, TbX, TbY;
 
2537
                    Taw = FNMS(KP923879532, Tav, KP382683432 * Tas);
 
2538
                    TaD = FMA(KP382683432, Taz, KP923879532 * TaC);
 
2539
                    TaE = Taw - TaD;
 
2540
                    Tc2 = Taw + TaD;
 
2541
                    TbX = Tbb + Tbm;
 
2542
                    TbY = Tbs + Tbv;
 
2543
                    TbZ = FNMS(KP555570233, TbY, KP831469612 * TbX);
 
2544
                    Tc7 = FMA(KP831469612, TbY, KP555570233 * TbX);
 
2545
               }
 
2546
               {
 
2547
                    E TaW, Tb5, TbG, TbH;
 
2548
                    TaW = TaK - TaV;
 
2549
                    Tb5 = Tb1 - Tb4;
 
2550
                    Tb6 = FMA(KP980785280, TaW, KP195090322 * Tb5);
 
2551
                    TbM = FNMS(KP980785280, Tb5, KP195090322 * TaW);
 
2552
                    TbG = FNMS(KP923879532, Taz, KP382683432 * TaC);
 
2553
                    TbH = FMA(KP923879532, Tas, KP382683432 * Tav);
 
2554
                    TbI = TbG - TbH;
 
2555
                    TbS = TbH + TbG;
 
2556
               }
 
2557
               {
 
2558
                    E TbU, TbV, Tbn, Tbw;
 
2559
                    TbU = TaK + TaV;
 
2560
                    TbV = Tb1 + Tb4;
 
2561
                    TbW = FMA(KP555570233, TbU, KP831469612 * TbV);
 
2562
                    Tc6 = FNMS(KP555570233, TbV, KP831469612 * TbU);
 
2563
                    Tbn = Tbb - Tbm;
 
2564
                    Tbw = Tbs - Tbv;
 
2565
                    Tbx = FNMS(KP980785280, Tbw, KP195090322 * Tbn);
 
2566
                    TbN = FMA(KP195090322, Tbw, KP980785280 * Tbn);
 
2567
               }
 
2568
               {
 
2569
                    E TaF, Tby, TbP, TbQ;
 
2570
                    TaF = Tap + TaE;
 
2571
                    Tby = Tb6 + Tbx;
 
2572
                    ro[WS(os, 46)] = TaF - Tby;
 
2573
                    ro[WS(os, 14)] = TaF + Tby;
 
2574
                    TbP = TbF + TbI;
 
2575
                    TbQ = TbM + TbN;
 
2576
                    io[WS(os, 46)] = TbP - TbQ;
 
2577
                    io[WS(os, 14)] = TbP + TbQ;
 
2578
               }
 
2579
               {
 
2580
                    E TbJ, TbK, TbL, TbO;
 
2581
                    TbJ = TbF - TbI;
 
2582
                    TbK = Tbx - Tb6;
 
2583
                    io[WS(os, 62)] = TbJ - TbK;
 
2584
                    io[WS(os, 30)] = TbJ + TbK;
 
2585
                    TbL = Tap - TaE;
 
2586
                    TbO = TbM - TbN;
 
2587
                    ro[WS(os, 62)] = TbL - TbO;
 
2588
                    ro[WS(os, 30)] = TbL + TbO;
 
2589
               }
 
2590
               {
 
2591
                    E TbT, Tc0, Tc9, Tca;
 
2592
                    TbT = TbR + TbS;
 
2593
                    Tc0 = TbW + TbZ;
 
2594
                    ro[WS(os, 38)] = TbT - Tc0;
 
2595
                    ro[WS(os, 6)] = TbT + Tc0;
 
2596
                    Tc9 = Tc1 + Tc2;
 
2597
                    Tca = Tc6 + Tc7;
 
2598
                    io[WS(os, 38)] = Tc9 - Tca;
 
2599
                    io[WS(os, 6)] = Tc9 + Tca;
 
2600
               }
 
2601
               {
 
2602
                    E Tc3, Tc4, Tc5, Tc8;
 
2603
                    Tc3 = Tc1 - Tc2;
 
2604
                    Tc4 = TbZ - TbW;
 
2605
                    io[WS(os, 54)] = Tc3 - Tc4;
 
2606
                    io[WS(os, 22)] = Tc3 + Tc4;
 
2607
                    Tc5 = TbR - TbS;
 
2608
                    Tc8 = Tc6 - Tc7;
 
2609
                    ro[WS(os, 54)] = Tc5 - Tc8;
 
2610
                    ro[WS(os, 22)] = Tc5 + Tc8;
 
2611
               }
 
2612
          }
 
2613
          {
 
2614
               E T6F, T7h, T7m, T7w, T7p, T7x, T6M, T7s, T6U, T7c, T75, T7r, T78, T7i, T71;
 
2615
               E T7d;
 
2616
               {
 
2617
                    E T6D, T6E, T7k, T7l;
 
2618
                    T6D = T37 + T3e;
 
2619
                    T6E = T65 + T64;
 
2620
                    T6F = T6D - T6E;
 
2621
                    T7h = T6D + T6E;
 
2622
                    T7k = T6O + T6P;
 
2623
                    T7l = T6R + T6S;
 
2624
                    T7m = FMA(KP956940335, T7k, KP290284677 * T7l);
 
2625
                    T7w = FNMS(KP290284677, T7k, KP956940335 * T7l);
 
2626
               }
 
2627
               {
 
2628
                    E T7n, T7o, T6I, T6L;
 
2629
                    T7n = T6V + T6W;
 
2630
                    T7o = T6Y + T6Z;
 
2631
                    T7p = FNMS(KP290284677, T7o, KP956940335 * T7n);
 
2632
                    T7x = FMA(KP290284677, T7n, KP956940335 * T7o);
 
2633
                    T6I = FNMS(KP555570233, T6H, KP831469612 * T6G);
 
2634
                    T6L = FMA(KP831469612, T6J, KP555570233 * T6K);
 
2635
                    T6M = T6I - T6L;
 
2636
                    T7s = T6I + T6L;
 
2637
               }
 
2638
               {
 
2639
                    E T6Q, T6T, T73, T74;
 
2640
                    T6Q = T6O - T6P;
 
2641
                    T6T = T6R - T6S;
 
2642
                    T6U = FMA(KP471396736, T6Q, KP881921264 * T6T);
 
2643
                    T7c = FNMS(KP881921264, T6Q, KP471396736 * T6T);
 
2644
                    T73 = T5Z + T62;
 
2645
                    T74 = T3m + T3t;
 
2646
                    T75 = T73 - T74;
 
2647
                    T7r = T73 + T74;
 
2648
               }
 
2649
               {
 
2650
                    E T76, T77, T6X, T70;
 
2651
                    T76 = FNMS(KP555570233, T6J, KP831469612 * T6K);
 
2652
                    T77 = FMA(KP555570233, T6G, KP831469612 * T6H);
 
2653
                    T78 = T76 - T77;
 
2654
                    T7i = T77 + T76;
 
2655
                    T6X = T6V - T6W;
 
2656
                    T70 = T6Y - T6Z;
 
2657
                    T71 = FNMS(KP881921264, T70, KP471396736 * T6X);
 
2658
                    T7d = FMA(KP881921264, T6X, KP471396736 * T70);
 
2659
               }
 
2660
               {
 
2661
                    E T6N, T72, T7f, T7g;
 
2662
                    T6N = T6F + T6M;
 
2663
                    T72 = T6U + T71;
 
2664
                    ro[WS(os, 43)] = T6N - T72;
 
2665
                    ro[WS(os, 11)] = T6N + T72;
 
2666
                    T7f = T75 + T78;
 
2667
                    T7g = T7c + T7d;
 
2668
                    io[WS(os, 43)] = T7f - T7g;
 
2669
                    io[WS(os, 11)] = T7f + T7g;
 
2670
               }
 
2671
               {
 
2672
                    E T79, T7a, T7b, T7e;
 
2673
                    T79 = T75 - T78;
 
2674
                    T7a = T71 - T6U;
 
2675
                    io[WS(os, 59)] = T79 - T7a;
 
2676
                    io[WS(os, 27)] = T79 + T7a;
 
2677
                    T7b = T6F - T6M;
 
2678
                    T7e = T7c - T7d;
 
2679
                    ro[WS(os, 59)] = T7b - T7e;
 
2680
                    ro[WS(os, 27)] = T7b + T7e;
 
2681
               }
 
2682
               {
 
2683
                    E T7j, T7q, T7z, T7A;
 
2684
                    T7j = T7h + T7i;
 
2685
                    T7q = T7m + T7p;
 
2686
                    ro[WS(os, 35)] = T7j - T7q;
 
2687
                    ro[WS(os, 3)] = T7j + T7q;
 
2688
                    T7z = T7r + T7s;
 
2689
                    T7A = T7w + T7x;
 
2690
                    io[WS(os, 35)] = T7z - T7A;
 
2691
                    io[WS(os, 3)] = T7z + T7A;
 
2692
               }
 
2693
               {
 
2694
                    E T7t, T7u, T7v, T7y;
 
2695
                    T7t = T7r - T7s;
 
2696
                    T7u = T7p - T7m;
 
2697
                    io[WS(os, 51)] = T7t - T7u;
 
2698
                    io[WS(os, 19)] = T7t + T7u;
 
2699
                    T7v = T7h - T7i;
 
2700
                    T7y = T7w - T7x;
 
2701
                    ro[WS(os, 51)] = T7v - T7y;
 
2702
                    ro[WS(os, 19)] = T7v + T7y;
 
2703
               }
 
2704
          }
 
2705
          {
 
2706
               E T9j, T9V, Ta0, Taa, Ta3, Tab, T9q, Ta6, T9y, T9Q, T9J, Ta5, T9M, T9W, T9F;
 
2707
               E T9R;
 
2708
               {
 
2709
                    E T9h, T9i, T9Y, T9Z;
 
2710
                    T9h = T7B + T7C;
 
2711
                    T9i = T8J + T8I;
 
2712
                    T9j = T9h - T9i;
 
2713
                    T9V = T9h + T9i;
 
2714
                    T9Y = T9s + T9t;
 
2715
                    T9Z = T9v + T9w;
 
2716
                    Ta0 = FMA(KP995184726, T9Y, KP098017140 * T9Z);
 
2717
                    Taa = FNMS(KP098017140, T9Y, KP995184726 * T9Z);
 
2718
               }
 
2719
               {
 
2720
                    E Ta1, Ta2, T9m, T9p;
 
2721
                    Ta1 = T9z + T9A;
 
2722
                    Ta2 = T9C + T9D;
 
2723
                    Ta3 = FNMS(KP098017140, Ta2, KP995184726 * Ta1);
 
2724
                    Tab = FMA(KP098017140, Ta1, KP995184726 * Ta2);
 
2725
                    T9m = FNMS(KP195090322, T9l, KP980785280 * T9k);
 
2726
                    T9p = FMA(KP195090322, T9n, KP980785280 * T9o);
 
2727
                    T9q = T9m - T9p;
 
2728
                    Ta6 = T9m + T9p;
 
2729
               }
 
2730
               {
 
2731
                    E T9u, T9x, T9H, T9I;
 
2732
                    T9u = T9s - T9t;
 
2733
                    T9x = T9v - T9w;
 
2734
                    T9y = FMA(KP634393284, T9u, KP773010453 * T9x);
 
2735
                    T9Q = FNMS(KP773010453, T9u, KP634393284 * T9x);
 
2736
                    T9H = T8F + T8G;
 
2737
                    T9I = T7G + T7J;
 
2738
                    T9J = T9H - T9I;
 
2739
                    Ta5 = T9H + T9I;
 
2740
               }
 
2741
               {
 
2742
                    E T9K, T9L, T9B, T9E;
 
2743
                    T9K = FNMS(KP195090322, T9o, KP980785280 * T9n);
 
2744
                    T9L = FMA(KP980785280, T9l, KP195090322 * T9k);
 
2745
                    T9M = T9K - T9L;
 
2746
                    T9W = T9L + T9K;
 
2747
                    T9B = T9z - T9A;
 
2748
                    T9E = T9C - T9D;
 
2749
                    T9F = FNMS(KP773010453, T9E, KP634393284 * T9B);
 
2750
                    T9R = FMA(KP773010453, T9B, KP634393284 * T9E);
 
2751
               }
 
2752
               {
 
2753
                    E T9r, T9G, T9T, T9U;
 
2754
                    T9r = T9j + T9q;
 
2755
                    T9G = T9y + T9F;
 
2756
                    ro[WS(os, 41)] = T9r - T9G;
 
2757
                    ro[WS(os, 9)] = T9r + T9G;
 
2758
                    T9T = T9J + T9M;
 
2759
                    T9U = T9Q + T9R;
 
2760
                    io[WS(os, 41)] = T9T - T9U;
 
2761
                    io[WS(os, 9)] = T9T + T9U;
 
2762
               }
 
2763
               {
 
2764
                    E T9N, T9O, T9P, T9S;
 
2765
                    T9N = T9J - T9M;
 
2766
                    T9O = T9F - T9y;
 
2767
                    io[WS(os, 57)] = T9N - T9O;
 
2768
                    io[WS(os, 25)] = T9N + T9O;
 
2769
                    T9P = T9j - T9q;
 
2770
                    T9S = T9Q - T9R;
 
2771
                    ro[WS(os, 57)] = T9P - T9S;
 
2772
                    ro[WS(os, 25)] = T9P + T9S;
 
2773
               }
 
2774
               {
 
2775
                    E T9X, Ta4, Tad, Tae;
 
2776
                    T9X = T9V + T9W;
 
2777
                    Ta4 = Ta0 + Ta3;
 
2778
                    ro[WS(os, 33)] = T9X - Ta4;
 
2779
                    ro[WS(os, 1)] = T9X + Ta4;
 
2780
                    Tad = Ta5 + Ta6;
 
2781
                    Tae = Taa + Tab;
 
2782
                    io[WS(os, 33)] = Tad - Tae;
 
2783
                    io[WS(os, 1)] = Tad + Tae;
 
2784
               }
 
2785
               {
 
2786
                    E Ta7, Ta8, Ta9, Tac;
 
2787
                    Ta7 = Ta5 - Ta6;
 
2788
                    Ta8 = Ta3 - Ta0;
 
2789
                    io[WS(os, 49)] = Ta7 - Ta8;
 
2790
                    io[WS(os, 17)] = Ta7 + Ta8;
 
2791
                    Ta9 = T9V - T9W;
 
2792
                    Tac = Taa - Tab;
 
2793
                    ro[WS(os, 49)] = Ta9 - Tac;
 
2794
                    ro[WS(os, 17)] = Ta9 + Tac;
 
2795
               }
 
2796
          }
 
2797
          {
 
2798
               E T3v, T6j, T6o, T6y, T6r, T6z, T48, T6u, T52, T6e, T67, T6t, T6a, T6k, T5V;
 
2799
               E T6f;
 
2800
               {
 
2801
                    E T3f, T3u, T6m, T6n;
 
2802
                    T3f = T37 - T3e;
 
2803
                    T3u = T3m - T3t;
 
2804
                    T3v = T3f - T3u;
 
2805
                    T6j = T3f + T3u;
 
2806
                    T6m = T4q + T4N;
 
2807
                    T6n = T4X + T50;
 
2808
                    T6o = FMA(KP634393284, T6m, KP773010453 * T6n);
 
2809
                    T6y = FNMS(KP634393284, T6n, KP773010453 * T6m);
 
2810
               }
 
2811
               {
 
2812
                    E T6p, T6q, T3O, T47;
 
2813
                    T6p = T5j + T5G;
 
2814
                    T6q = T5Q + T5T;
 
2815
                    T6r = FNMS(KP634393284, T6q, KP773010453 * T6p);
 
2816
                    T6z = FMA(KP773010453, T6q, KP634393284 * T6p);
 
2817
                    T3O = FNMS(KP980785280, T3N, KP195090322 * T3G);
 
2818
                    T47 = FMA(KP195090322, T3Z, KP980785280 * T46);
 
2819
                    T48 = T3O - T47;
 
2820
                    T6u = T3O + T47;
 
2821
               }
 
2822
               {
 
2823
                    E T4O, T51, T63, T66;
 
2824
                    T4O = T4q - T4N;
 
2825
                    T51 = T4X - T50;
 
2826
                    T52 = FMA(KP995184726, T4O, KP098017140 * T51);
 
2827
                    T6e = FNMS(KP995184726, T51, KP098017140 * T4O);
 
2828
                    T63 = T5Z - T62;
 
2829
                    T66 = T64 - T65;
 
2830
                    T67 = T63 - T66;
 
2831
                    T6t = T63 + T66;
 
2832
               }
 
2833
               {
 
2834
                    E T68, T69, T5H, T5U;
 
2835
                    T68 = FNMS(KP980785280, T3Z, KP195090322 * T46);
 
2836
                    T69 = FMA(KP980785280, T3G, KP195090322 * T3N);
 
2837
                    T6a = T68 - T69;
 
2838
                    T6k = T69 + T68;
 
2839
                    T5H = T5j - T5G;
 
2840
                    T5U = T5Q - T5T;
 
2841
                    T5V = FNMS(KP995184726, T5U, KP098017140 * T5H);
 
2842
                    T6f = FMA(KP098017140, T5U, KP995184726 * T5H);
 
2843
               }
 
2844
               {
 
2845
                    E T49, T5W, T6h, T6i;
 
2846
                    T49 = T3v + T48;
 
2847
                    T5W = T52 + T5V;
 
2848
                    ro[WS(os, 47)] = T49 - T5W;
 
2849
                    ro[WS(os, 15)] = T49 + T5W;
 
2850
                    T6h = T67 + T6a;
 
2851
                    T6i = T6e + T6f;
 
2852
                    io[WS(os, 47)] = T6h - T6i;
 
2853
                    io[WS(os, 15)] = T6h + T6i;
 
2854
               }
 
2855
               {
 
2856
                    E T6b, T6c, T6d, T6g;
 
2857
                    T6b = T67 - T6a;
 
2858
                    T6c = T5V - T52;
 
2859
                    io[WS(os, 63)] = T6b - T6c;
 
2860
                    io[WS(os, 31)] = T6b + T6c;
 
2861
                    T6d = T3v - T48;
 
2862
                    T6g = T6e - T6f;
 
2863
                    ro[WS(os, 63)] = T6d - T6g;
 
2864
                    ro[WS(os, 31)] = T6d + T6g;
 
2865
               }
 
2866
               {
 
2867
                    E T6l, T6s, T6B, T6C;
 
2868
                    T6l = T6j + T6k;
 
2869
                    T6s = T6o + T6r;
 
2870
                    ro[WS(os, 39)] = T6l - T6s;
 
2871
                    ro[WS(os, 7)] = T6l + T6s;
 
2872
                    T6B = T6t + T6u;
 
2873
                    T6C = T6y + T6z;
 
2874
                    io[WS(os, 39)] = T6B - T6C;
 
2875
                    io[WS(os, 7)] = T6B + T6C;
 
2876
               }
 
2877
               {
 
2878
                    E T6v, T6w, T6x, T6A;
 
2879
                    T6v = T6t - T6u;
 
2880
                    T6w = T6r - T6o;
 
2881
                    io[WS(os, 55)] = T6v - T6w;
 
2882
                    io[WS(os, 23)] = T6v + T6w;
 
2883
                    T6x = T6j - T6k;
 
2884
                    T6A = T6y - T6z;
 
2885
                    ro[WS(os, 55)] = T6x - T6A;
 
2886
                    ro[WS(os, 23)] = T6x + T6A;
 
2887
               }
 
2888
          }
 
2889
          {
 
2890
               E T7L, T8X, T92, T9c, T95, T9d, T80, T98, T8k, T8S, T8L, T97, T8O, T8Y, T8D;
 
2891
               E T8T;
 
2892
               {
 
2893
                    E T7D, T7K, T90, T91;
 
2894
                    T7D = T7B - T7C;
 
2895
                    T7K = T7G - T7J;
 
2896
                    T7L = T7D - T7K;
 
2897
                    T8X = T7D + T7K;
 
2898
                    T90 = T84 + T8b;
 
2899
                    T91 = T8f + T8i;
 
2900
                    T92 = FMA(KP471396736, T90, KP881921264 * T91);
 
2901
                    T9c = FNMS(KP471396736, T91, KP881921264 * T90);
 
2902
               }
 
2903
               {
 
2904
                    E T93, T94, T7S, T7Z;
 
2905
                    T93 = T8n + T8u;
 
2906
                    T94 = T8y + T8B;
 
2907
                    T95 = FNMS(KP471396736, T94, KP881921264 * T93);
 
2908
                    T9d = FMA(KP881921264, T94, KP471396736 * T93);
 
2909
                    T7S = FNMS(KP831469612, T7R, KP555570233 * T7O);
 
2910
                    T7Z = FMA(KP831469612, T7V, KP555570233 * T7Y);
 
2911
                    T80 = T7S - T7Z;
 
2912
                    T98 = T7S + T7Z;
 
2913
               }
 
2914
               {
 
2915
                    E T8c, T8j, T8H, T8K;
 
2916
                    T8c = T84 - T8b;
 
2917
                    T8j = T8f - T8i;
 
2918
                    T8k = FMA(KP956940335, T8c, KP290284677 * T8j);
 
2919
                    T8S = FNMS(KP956940335, T8j, KP290284677 * T8c);
 
2920
                    T8H = T8F - T8G;
 
2921
                    T8K = T8I - T8J;
 
2922
                    T8L = T8H - T8K;
 
2923
                    T97 = T8H + T8K;
 
2924
               }
 
2925
               {
 
2926
                    E T8M, T8N, T8v, T8C;
 
2927
                    T8M = FNMS(KP831469612, T7Y, KP555570233 * T7V);
 
2928
                    T8N = FMA(KP555570233, T7R, KP831469612 * T7O);
 
2929
                    T8O = T8M - T8N;
 
2930
                    T8Y = T8N + T8M;
 
2931
                    T8v = T8n - T8u;
 
2932
                    T8C = T8y - T8B;
 
2933
                    T8D = FNMS(KP956940335, T8C, KP290284677 * T8v);
 
2934
                    T8T = FMA(KP290284677, T8C, KP956940335 * T8v);
 
2935
               }
 
2936
               {
 
2937
                    E T81, T8E, T8V, T8W;
 
2938
                    T81 = T7L + T80;
 
2939
                    T8E = T8k + T8D;
 
2940
                    ro[WS(os, 45)] = T81 - T8E;
 
2941
                    ro[WS(os, 13)] = T81 + T8E;
 
2942
                    T8V = T8L + T8O;
 
2943
                    T8W = T8S + T8T;
 
2944
                    io[WS(os, 45)] = T8V - T8W;
 
2945
                    io[WS(os, 13)] = T8V + T8W;
 
2946
               }
 
2947
               {
 
2948
                    E T8P, T8Q, T8R, T8U;
 
2949
                    T8P = T8L - T8O;
 
2950
                    T8Q = T8D - T8k;
 
2951
                    io[WS(os, 61)] = T8P - T8Q;
 
2952
                    io[WS(os, 29)] = T8P + T8Q;
 
2953
                    T8R = T7L - T80;
 
2954
                    T8U = T8S - T8T;
 
2955
                    ro[WS(os, 61)] = T8R - T8U;
 
2956
                    ro[WS(os, 29)] = T8R + T8U;
 
2957
               }
 
2958
               {
 
2959
                    E T8Z, T96, T9f, T9g;
 
2960
                    T8Z = T8X + T8Y;
 
2961
                    T96 = T92 + T95;
 
2962
                    ro[WS(os, 37)] = T8Z - T96;
 
2963
                    ro[WS(os, 5)] = T8Z + T96;
 
2964
                    T9f = T97 + T98;
 
2965
                    T9g = T9c + T9d;
 
2966
                    io[WS(os, 37)] = T9f - T9g;
 
2967
                    io[WS(os, 5)] = T9f + T9g;
 
2968
               }
 
2969
               {
 
2970
                    E T99, T9a, T9b, T9e;
 
2971
                    T99 = T97 - T98;
 
2972
                    T9a = T95 - T92;
 
2973
                    io[WS(os, 53)] = T99 - T9a;
 
2974
                    io[WS(os, 21)] = T99 + T9a;
 
2975
                    T9b = T8X - T8Y;
 
2976
                    T9e = T9c - T9d;
 
2977
                    ro[WS(os, 53)] = T9b - T9e;
 
2978
                    ro[WS(os, 21)] = T9b + T9e;
 
2979
               }
 
2980
          }
 
2981
     }
 
2982
}
 
2983
 
 
2984
static const kdft_desc desc = { 64, "n1_64", {808, 144, 104, 0}, &GENUS, 0, 0, 0, 0 };
 
2985
void X(codelet_n1_64) (planner *p) {
 
2986
     X(kdft_register) (p, n1_64, &desc);
 
2987
}
 
2988
 
 
2989
#endif                          /* HAVE_FMA */