~ubuntu-branches/ubuntu/maverick/blender/maverick

« back to all changes in this revision

Viewing changes to extern/fftw/rdft/codelets/hc2r/hb_64.c

  • Committer: Bazaar Package Importer
  • Author(s): Khashayar Naderehvandi, Khashayar Naderehvandi, Alessio Treglia
  • Date: 2009-01-22 16:53:59 UTC
  • mfrom: (14.1.1 experimental)
  • Revision ID: james.westby@ubuntu.com-20090122165359-v0996tn7fbit64ni
Tags: 2.48a+dfsg-1ubuntu1
[ Khashayar Naderehvandi ]
* Merge from debian experimental (LP: #320045), Ubuntu remaining changes:
  - Add patch correcting header file locations.
  - Add libvorbis-dev and libgsm1-dev to Build-Depends.
  - Use avcodec_decode_audio2() in source/blender/src/hddaudio.c

[ Alessio Treglia ]
* Add missing previous changelog entries.

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 Sun Jul  2 16:32:54 EDT 2006 */
 
23
 
 
24
#include "codelet-rdft.h"
 
25
 
 
26
#ifdef HAVE_FMA
 
27
 
 
28
/* Generated by: ../../../genfft/gen_hc2hc -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -dif -name hb_64 -include hb.h */
 
29
 
 
30
/*
 
31
 * This function contains 1038 FP additions, 644 FP multiplications,
 
32
 * (or, 520 additions, 126 multiplications, 518 fused multiply/add),
 
33
 * 234 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_hc2hc.ml,v 1.16 2006-02-12 23:34:12 athena Exp $
 
40
 */
 
41
 
 
42
#include "hb.h"
 
43
 
 
44
static const R *hb_64(R *rio, R *iio, const R *W, stride ios, INT m, INT dist)
 
45
{
 
46
     DK(KP773010453, +0.773010453362736960810906609758469800971041293);
 
47
     DK(KP820678790, +0.820678790828660330972281985331011598767386482);
 
48
     DK(KP956940335, +0.956940335732208864935797886980269969482849206);
 
49
     DK(KP303346683, +0.303346683607342391675883946941299872384187453);
 
50
     DK(KP995184726, +0.995184726672196886244836953109479921575474869);
 
51
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
 
52
     DK(KP098491403, +0.098491403357164253077197521291327432293052451);
 
53
     DK(KP881921264, +0.881921264348355029712756863660388349508442621);
 
54
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
 
55
     DK(KP534511135, +0.534511135950791641089685961295362908582039528);
 
56
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
57
     DK(KP198912367, +0.198912367379658006911597622644676228597850501);
 
58
     DK(KP668178637, +0.668178637919298919997757686523080761552472251);
 
59
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
60
     DK(KP414213562, +0.414213562373095048801688724209698078569671875);
 
61
     INT i;
 
62
     for (i = m - 2; i > 0; i = i - 2, rio = rio + dist, iio = iio - dist, W = W + 126, MAKE_VOLATILE_STRIDE(ios)) {
 
63
          E Tgv, Tgs, Tgr;
 
64
          {
 
65
               E T8v, Ta9, Tv, T9O, Tj7, TgN, TjA, ThQ, Tj6, ThN, T6y, T2v, T64, T2c, T6x;
 
66
               E T4L, TeV, Tbv, Tfn, Tbg, Tjz, TgG, T65, T4O, T96, T7S, Ta8, T8y, Tfo, TdE;
 
67
               E TeU, TdB, T10, Ta7, T9R, TaN, T2P, T4Q, T8A, T7X, T4R, T38, T8B, T82, T6A;
 
68
               E T69, T6B, T6c, TeZ, Tfq, TdG, TbP, Tjb, TjC, ThS, TgV, Tf2, Tfr, TdH, Tc8;
 
69
               E Tje, TjD, ThT, Th2, T9T, T1w, T9a, T8e, TaF, T9W, T99, T89, T6k, T7a, T5p;
 
70
               E T3D, T6h, T79, T5q, T3Q, Tf7, Tg0, Teg, TcB, Tfa, TfZ, Tef, TcM, Tjj, Tkc;
 
71
               E Tis, Thf, Tjm, Tkb, Tir, Thm, T21, T9Y, T9d, T8p, TaG, Ta1, T9c, T8k, T6r;
 
72
               E T7d, T5s, T4k, T6o, T7c, T5t, T4x, Tfe, Tg3, Tej, Tde, Thq, Tjr, Tfh, Tg2;
 
73
               E Tei, Tdp, ThC, ThD, Tjo, ThB, Tjp, Thx;
 
74
               {
 
75
                    E Tc3, TgW, TgZ, TgX, Tc0, Tf1, Tc6, Th0, TdA, Tdx;
 
76
                    {
 
77
                         E Tdw, T4G, Tb7, T4J, Tb6, Tdv, T24, T7, T27, T2a, Tbe, Tdy, Tbb, Tdz, Te;
 
78
                         E T4D, Tbt, TgK, T2m, Tt, Tbq, TgL, T7Q, T2t, Tj, Tbh, Ti, Tbl, T2g, Tk;
 
79
                         E T2h, T2i;
 
80
                         {
 
81
                              E T1, T2, T4, T5;
 
82
                              {
 
83
                                   E T4E, T4F, T4H, T4I;
 
84
                                   T4E = iio[0];
 
85
                                   T4F = rio[WS(ios, 32)];
 
86
                                   T4H = iio[-WS(ios, 16)];
 
87
                                   T4I = rio[WS(ios, 48)];
 
88
                                   T1 = rio[0];
 
89
                                   Tdw = T4E + T4F;
 
90
                                   T4G = T4E - T4F;
 
91
                                   Tb7 = T4H + T4I;
 
92
                                   T4J = T4H - T4I;
 
93
                                   T2 = iio[-WS(ios, 32)];
 
94
                                   T4 = rio[WS(ios, 16)];
 
95
                                   T5 = iio[-WS(ios, 48)];
 
96
                              }
 
97
                              {
 
98
                                   E Ta, Tba, Tb9, Td;
 
99
                                   {
 
100
                                        E Tbc, T28, Tbd, T29, Tb, Tc, T8, T9;
 
101
                                        {
 
102
                                             E T25, T3, T6, T26;
 
103
                                             T25 = iio[-WS(ios, 8)];
 
104
                                             Tb6 = T1 - T2;
 
105
                                             T3 = T1 + T2;
 
106
                                             Tdv = T4 - T5;
 
107
                                             T6 = T4 + T5;
 
108
                                             T26 = rio[WS(ios, 40)];
 
109
                                             T8 = rio[WS(ios, 8)];
 
110
                                             T9 = iio[-WS(ios, 40)];
 
111
                                             T24 = T3 - T6;
 
112
                                             T7 = T3 + T6;
 
113
                                             Tbc = T25 + T26;
 
114
                                             T27 = T25 - T26;
 
115
                                        }
 
116
                                        T28 = iio[-WS(ios, 24)];
 
117
                                        Tbd = T8 - T9;
 
118
                                        Ta = T8 + T9;
 
119
                                        T29 = rio[WS(ios, 56)];
 
120
                                        Tb = iio[-WS(ios, 56)];
 
121
                                        Tc = rio[WS(ios, 24)];
 
122
                                        Tba = T29 + T28;
 
123
                                        T2a = T28 - T29;
 
124
                                        Tbe = Tbc - Tbd;
 
125
                                        Tdy = Tbd + Tbc;
 
126
                                        Tb9 = Tb - Tc;
 
127
                                        Td = Tb + Tc;
 
128
                                   }
 
129
                                   Tbb = Tb9 - Tba;
 
130
                                   Tdz = Tb9 + Tba;
 
131
                                   Te = Ta + Td;
 
132
                                   T4D = Td - Ta;
 
133
                                   {
 
134
                                        E Tq, Tbo, Tp, Tbs, T2p, Tr, T2q, T2r;
 
135
                                        {
 
136
                                             E Tn, To, T2n, T2o;
 
137
                                             Tn = iio[-WS(ios, 60)];
 
138
                                             To = rio[WS(ios, 28)];
 
139
                                             T2n = iio[-WS(ios, 28)];
 
140
                                             T2o = rio[WS(ios, 60)];
 
141
                                             Tq = rio[WS(ios, 12)];
 
142
                                             Tbo = Tn - To;
 
143
                                             Tp = Tn + To;
 
144
                                             Tbs = T2o + T2n;
 
145
                                             T2p = T2n - T2o;
 
146
                                             Tr = iio[-WS(ios, 44)];
 
147
                                             T2q = iio[-WS(ios, 12)];
 
148
                                             T2r = rio[WS(ios, 44)];
 
149
                                        }
 
150
                                        {
 
151
                                             E Tg, Th, T2e, T2f;
 
152
                                             Tg = rio[WS(ios, 4)];
 
153
                                             {
 
154
                                                  E Tbr, Ts, Tbp, T2s;
 
155
                                                  Tbr = Tq - Tr;
 
156
                                                  Ts = Tq + Tr;
 
157
                                                  Tbp = T2q + T2r;
 
158
                                                  T2s = T2q - T2r;
 
159
                                                  Tbt = Tbr - Tbs;
 
160
                                                  TgK = Tbr + Tbs;
 
161
                                                  T2m = Tp - Ts;
 
162
                                                  Tt = Tp + Ts;
 
163
                                                  Tbq = Tbo - Tbp;
 
164
                                                  TgL = Tbo + Tbp;
 
165
                                                  T7Q = T2p + T2s;
 
166
                                                  T2t = T2p - T2s;
 
167
                                                  Th = iio[-WS(ios, 36)];
 
168
                                             }
 
169
                                             T2e = iio[-WS(ios, 4)];
 
170
                                             T2f = rio[WS(ios, 36)];
 
171
                                             Tj = rio[WS(ios, 20)];
 
172
                                             Tbh = Tg - Th;
 
173
                                             Ti = Tg + Th;
 
174
                                             Tbl = T2e + T2f;
 
175
                                             T2g = T2e - T2f;
 
176
                                             Tk = iio[-WS(ios, 52)];
 
177
                                             T2h = iio[-WS(ios, 20)];
 
178
                                             T2i = rio[WS(ios, 52)];
 
179
                                        }
 
180
                                   }
 
181
                              }
 
182
                         }
 
183
                         {
 
184
                              E T7O, Tbm, Tbj, T7R, Tb8, T4M, T2l, T2u, T4N, T2b, T8x;
 
185
                              {
 
186
                                   E T2d, T2k, ThO, TgJ, TgM, ThP, ThL, ThM;
 
187
                                   {
 
188
                                        E Tf, TgH, TgI, Tu, T7P;
 
189
                                        T7O = T7 - Te;
 
190
                                        Tf = T7 + Te;
 
191
                                        {
 
192
                                             E Tbk, Tl, Tbi, T2j, Tm;
 
193
                                             Tbk = Tj - Tk;
 
194
                                             Tl = Tj + Tk;
 
195
                                             Tbi = T2h + T2i;
 
196
                                             T2j = T2h - T2i;
 
197
                                             Tbm = Tbk + Tbl;
 
198
                                             TgH = Tbl - Tbk;
 
199
                                             T2d = Ti - Tl;
 
200
                                             Tm = Ti + Tl;
 
201
                                             Tbj = Tbh - Tbi;
 
202
                                             TgI = Tbh + Tbi;
 
203
                                             T7P = T2g + T2j;
 
204
                                             T2k = T2g - T2j;
 
205
                                             T8v = Tt - Tm;
 
206
                                             Tu = Tm + Tt;
 
207
                                        }
 
208
                                        Ta9 = T7P + T7Q;
 
209
                                        T7R = T7P - T7Q;
 
210
                                        ThO = FMA(KP414213562, TgH, TgI);
 
211
                                        TgJ = FNMS(KP414213562, TgI, TgH);
 
212
                                        TgM = FNMS(KP414213562, TgL, TgK);
 
213
                                        ThP = FMA(KP414213562, TgK, TgL);
 
214
                                        Tv = Tf + Tu;
 
215
                                        T9O = Tf - Tu;
 
216
                                   }
 
217
                                   Tj7 = TgJ + TgM;
 
218
                                   TgN = TgJ - TgM;
 
219
                                   Tb8 = Tb6 - Tb7;
 
220
                                   ThL = Tb6 + Tb7;
 
221
                                   ThM = Tdy + Tdz;
 
222
                                   TdA = Tdy - Tdz;
 
223
                                   T4M = T2k - T2d;
 
224
                                   T2l = T2d + T2k;
 
225
                                   TjA = ThO - ThP;
 
226
                                   ThQ = ThO + ThP;
 
227
                                   Tj6 = FNMS(KP707106781, ThM, ThL);
 
228
                                   ThN = FMA(KP707106781, ThM, ThL);
 
229
                                   T2u = T2m - T2t;
 
230
                                   T4N = T2m + T2t;
 
231
                                   T2b = T27 - T2a;
 
232
                                   T8x = T27 + T2a;
 
233
                              }
 
234
                              {
 
235
                                   E T8w, TdC, TdD, Tbf, TgF, TgE, T4K, Tbn, Tbu;
 
236
                                   T6y = T2l - T2u;
 
237
                                   T2v = T2l + T2u;
 
238
                                   T64 = T24 - T2b;
 
239
                                   T2c = T24 + T2b;
 
240
                                   T4K = T4G - T4J;
 
241
                                   T8w = T4G + T4J;
 
242
                                   TdC = FMA(KP414213562, Tbj, Tbm);
 
243
                                   Tbn = FNMS(KP414213562, Tbm, Tbj);
 
244
                                   Tbu = FMA(KP414213562, Tbt, Tbq);
 
245
                                   TdD = FNMS(KP414213562, Tbq, Tbt);
 
246
                                   T6x = T4K - T4D;
 
247
                                   T4L = T4D + T4K;
 
248
                                   TeV = Tbu - Tbn;
 
249
                                   Tbv = Tbn + Tbu;
 
250
                                   Tbf = Tbb - Tbe;
 
251
                                   TgF = Tbe + Tbb;
 
252
                                   Tdx = Tdv + Tdw;
 
253
                                   TgE = Tdw - Tdv;
 
254
                                   Tfn = FNMS(KP707106781, Tbf, Tb8);
 
255
                                   Tbg = FMA(KP707106781, Tbf, Tb8);
 
256
                                   Tjz = FNMS(KP707106781, TgF, TgE);
 
257
                                   TgG = FMA(KP707106781, TgF, TgE);
 
258
                                   T65 = T4N - T4M;
 
259
                                   T4O = T4M + T4N;
 
260
                                   T96 = T7O - T7R;
 
261
                                   T7S = T7O + T7R;
 
262
                                   Ta8 = T8x + T8w;
 
263
                                   T8y = T8w - T8x;
 
264
                                   Tfo = TdC - TdD;
 
265
                                   TdE = TdC + TdD;
 
266
                              }
 
267
                         }
 
268
                    }
 
269
                    {
 
270
                         E TbK, TgP, T2x, TC, Tbz, TgS, T7U, T2N, TbV, Tc4, TY, T2Z, Tc5, TbY, T2X;
 
271
                         E T80, TbF, TbL, TJ, T2G, TbM, TbC, T2E, T7V, TbQ, TN, T33, Tc2, T32, Tc1;
 
272
                         E TQ, T34;
 
273
                         {
 
274
                              E T2T, TbW, T2W, TbX;
 
275
                              {
 
276
                                   E Tbx, Ty, T2K, TbJ, T2J, TbI, TB, T2L;
 
277
                                   {
 
278
                                        E T2H, T2I, Tw, Tx, Tz, TA;
 
279
                                        Tw = rio[WS(ios, 2)];
 
280
                                        Tx = iio[-WS(ios, 34)];
 
281
                                        T2H = iio[-WS(ios, 2)];
 
282
                                        TeU = FNMS(KP707106781, TdA, Tdx);
 
283
                                        TdB = FMA(KP707106781, TdA, Tdx);
 
284
                                        Tbx = Tw - Tx;
 
285
                                        Ty = Tw + Tx;
 
286
                                        T2I = rio[WS(ios, 34)];
 
287
                                        Tz = rio[WS(ios, 18)];
 
288
                                        TA = iio[-WS(ios, 50)];
 
289
                                        T2K = iio[-WS(ios, 18)];
 
290
                                        TbJ = T2H + T2I;
 
291
                                        T2J = T2H - T2I;
 
292
                                        TbI = Tz - TA;
 
293
                                        TB = Tz + TA;
 
294
                                        T2L = rio[WS(ios, 50)];
 
295
                                   }
 
296
                                   {
 
297
                                        E TbT, TU, T2U, TbU, TX, T2V;
 
298
                                        {
 
299
                                             E T2R, T2S, TV, TW;
 
300
                                             {
 
301
                                                  E TS, Tby, T2M, TT;
 
302
                                                  TS = rio[WS(ios, 6)];
 
303
                                                  TbK = TbI + TbJ;
 
304
                                                  TgP = TbJ - TbI;
 
305
                                                  T2x = Ty - TB;
 
306
                                                  TC = Ty + TB;
 
307
                                                  Tby = T2K + T2L;
 
308
                                                  T2M = T2K - T2L;
 
309
                                                  TT = iio[-WS(ios, 38)];
 
310
                                                  T2R = iio[-WS(ios, 6)];
 
311
                                                  Tbz = Tbx - Tby;
 
312
                                                  TgS = Tbx + Tby;
 
313
                                                  T7U = T2J + T2M;
 
314
                                                  T2N = T2J - T2M;
 
315
                                                  TbT = TS - TT;
 
316
                                                  TU = TS + TT;
 
317
                                                  T2S = rio[WS(ios, 38)];
 
318
                                             }
 
319
                                             TV = iio[-WS(ios, 54)];
 
320
                                             TW = rio[WS(ios, 22)];
 
321
                                             T2U = iio[-WS(ios, 22)];
 
322
                                             TbU = T2R + T2S;
 
323
                                             T2T = T2R - T2S;
 
324
                                             TbW = TV - TW;
 
325
                                             TX = TV + TW;
 
326
                                             T2V = rio[WS(ios, 54)];
 
327
                                        }
 
328
                                        TbV = TbT - TbU;
 
329
                                        Tc4 = TbT + TbU;
 
330
                                        TY = TU + TX;
 
331
                                        T2Z = TX - TU;
 
332
                                        T2W = T2U - T2V;
 
333
                                        TbX = T2V + T2U;
 
334
                                   }
 
335
                              }
 
336
                              {
 
337
                                   E T2A, TbA, T2D, TbB;
 
338
                                   {
 
339
                                        E TbE, TF, T2B, TbD, TI, T2C;
 
340
                                        {
 
341
                                             E T2y, T2z, TD, TE, TG, TH;
 
342
                                             TD = rio[WS(ios, 10)];
 
343
                                             TE = iio[-WS(ios, 42)];
 
344
                                             Tc5 = TbW + TbX;
 
345
                                             TbY = TbW - TbX;
 
346
                                             T2X = T2T - T2W;
 
347
                                             T80 = T2T + T2W;
 
348
                                             TbE = TD - TE;
 
349
                                             TF = TD + TE;
 
350
                                             T2y = iio[-WS(ios, 10)];
 
351
                                             T2z = rio[WS(ios, 42)];
 
352
                                             TG = iio[-WS(ios, 58)];
 
353
                                             TH = rio[WS(ios, 26)];
 
354
                                             T2B = iio[-WS(ios, 26)];
 
355
                                             TbD = T2y + T2z;
 
356
                                             T2A = T2y - T2z;
 
357
                                             TbA = TG - TH;
 
358
                                             TI = TG + TH;
 
359
                                             T2C = rio[WS(ios, 58)];
 
360
                                        }
 
361
                                        TbF = TbD - TbE;
 
362
                                        TbL = TbE + TbD;
 
363
                                        TJ = TF + TI;
 
364
                                        T2G = TI - TF;
 
365
                                        T2D = T2B - T2C;
 
366
                                        TbB = T2C + T2B;
 
367
                                   }
 
368
                                   {
 
369
                                        E T30, T31, TL, TM, TO, TP;
 
370
                                        TL = iio[-WS(ios, 62)];
 
371
                                        TM = rio[WS(ios, 30)];
 
372
                                        TbM = TbA + TbB;
 
373
                                        TbC = TbA - TbB;
 
374
                                        T2E = T2A - T2D;
 
375
                                        T7V = T2A + T2D;
 
376
                                        TbQ = TL - TM;
 
377
                                        TN = TL + TM;
 
378
                                        T30 = iio[-WS(ios, 30)];
 
379
                                        T31 = rio[WS(ios, 62)];
 
380
                                        TO = rio[WS(ios, 14)];
 
381
                                        TP = iio[-WS(ios, 46)];
 
382
                                        T33 = iio[-WS(ios, 14)];
 
383
                                        Tc2 = T31 + T30;
 
384
                                        T32 = T30 - T31;
 
385
                                        Tc1 = TO - TP;
 
386
                                        TQ = TO + TP;
 
387
                                        T34 = rio[WS(ios, 46)];
 
388
                                   }
 
389
                              }
 
390
                         }
 
391
                         {
 
392
                              E T7Y, TbS, T67, T81, T68, T6a, T2Y, T37, T6b;
 
393
                              {
 
394
                                   E T7T, T2Q, T9P, T7W, T36, T2F, T9Q, T2O;
 
395
                                   {
 
396
                                        E TK, TR, TbR, T35, T7Z, TZ;
 
397
                                        T7T = TC - TJ;
 
398
                                        TK = TC + TJ;
 
399
                                        Tc3 = Tc1 - Tc2;
 
400
                                        TgW = Tc1 + Tc2;
 
401
                                        T2Q = TN - TQ;
 
402
                                        TR = TN + TQ;
 
403
                                        TbR = T33 + T34;
 
404
                                        T35 = T33 - T34;
 
405
                                        T9P = T7V + T7U;
 
406
                                        T7W = T7U - T7V;
 
407
                                        T7Y = TR - TY;
 
408
                                        TZ = TR + TY;
 
409
                                        TbS = TbQ - TbR;
 
410
                                        TgZ = TbQ + TbR;
 
411
                                        T7Z = T32 + T35;
 
412
                                        T36 = T32 - T35;
 
413
                                        T10 = TK + TZ;
 
414
                                        Ta7 = TZ - TK;
 
415
                                        T67 = T2x - T2E;
 
416
                                        T2F = T2x + T2E;
 
417
                                        T9Q = T80 + T7Z;
 
418
                                        T81 = T7Z - T80;
 
419
                                        T68 = T2N - T2G;
 
420
                                        T2O = T2G + T2N;
 
421
                                   }
 
422
                                   T9R = T9P - T9Q;
 
423
                                   TaN = T9P + T9Q;
 
424
                                   T2P = FMA(KP414213562, T2O, T2F);
 
425
                                   T4Q = FNMS(KP414213562, T2F, T2O);
 
426
                                   T8A = T7W - T7T;
 
427
                                   T7X = T7T + T7W;
 
428
                                   T6a = T2Q - T2X;
 
429
                                   T2Y = T2Q + T2X;
 
430
                                   T37 = T2Z + T36;
 
431
                                   T6b = T36 - T2Z;
 
432
                              }
 
433
                              {
 
434
                                   E TgQ, TbH, TeY, TbN, TgT, TbG, TbZ, TeX, TbO;
 
435
                                   T4R = FMA(KP414213562, T2Y, T37);
 
436
                                   T38 = FNMS(KP414213562, T37, T2Y);
 
437
                                   T8B = T7Y + T81;
 
438
                                   T82 = T7Y - T81;
 
439
                                   TgQ = TbF + TbC;
 
440
                                   TbG = TbC - TbF;
 
441
                                   T6A = FMA(KP414213562, T67, T68);
 
442
                                   T69 = FNMS(KP414213562, T68, T67);
 
443
                                   T6B = FNMS(KP414213562, T6a, T6b);
 
444
                                   T6c = FMA(KP414213562, T6b, T6a);
 
445
                                   TbH = FMA(KP707106781, TbG, Tbz);
 
446
                                   TeY = FNMS(KP707106781, TbG, Tbz);
 
447
                                   TbN = TbL - TbM;
 
448
                                   TgT = TbL + TbM;
 
449
                                   TgX = TbV - TbY;
 
450
                                   TbZ = TbV + TbY;
 
451
                                   TeX = FNMS(KP707106781, TbN, TbK);
 
452
                                   TbO = FMA(KP707106781, TbN, TbK);
 
453
                                   {
 
454
                                        E Tja, TgR, Tj9, TgU;
 
455
                                        Tja = FNMS(KP707106781, TgQ, TgP);
 
456
                                        TgR = FMA(KP707106781, TgQ, TgP);
 
457
                                        Tj9 = FNMS(KP707106781, TgT, TgS);
 
458
                                        TgU = FMA(KP707106781, TgT, TgS);
 
459
                                        TeZ = FNMS(KP668178637, TeY, TeX);
 
460
                                        Tfq = FMA(KP668178637, TeX, TeY);
 
461
                                        TdG = FMA(KP198912367, TbH, TbO);
 
462
                                        TbP = FNMS(KP198912367, TbO, TbH);
 
463
                                        Tjb = FNMS(KP668178637, Tja, Tj9);
 
464
                                        TjC = FMA(KP668178637, Tj9, Tja);
 
465
                                        ThS = FMA(KP198912367, TgR, TgU);
 
466
                                        TgV = FNMS(KP198912367, TgU, TgR);
 
467
                                        Tc0 = FMA(KP707106781, TbZ, TbS);
 
468
                                        Tf1 = FNMS(KP707106781, TbZ, TbS);
 
469
                                   }
 
470
                                   Tc6 = Tc4 - Tc5;
 
471
                                   Th0 = Tc4 + Tc5;
 
472
                              }
 
473
                         }
 
474
                    }
 
475
                    {
 
476
                         E Th7, Tjk, Thj, Thk, Tjh, Thi, Tji, The;
 
477
                         {
 
478
                              E TcE, Th5, T3b, T18, Tcd, Thg, T8b, T3L, Tcy, Thb, T3t, T1u, Tcv, Thc, T87;
 
479
                              E T3A, Tcj, TcF, T1f, T3E, TcG, Tcg, T8c, T3i, T1k, Tcm, T1j, Tcq, T3n, T1l;
 
480
                              E T3o, T3p;
 
481
                              {
 
482
                                   E T1r, Tct, T1q, Tcx, T3w, T1s, T3x, T3y;
 
483
                                   {
 
484
                                        E T15, Tcb, T14, TcD, T3H, T16, T3I, T3J;
 
485
                                        {
 
486
                                             E T12, T13, T3F, T3G, Tf0, Tc7;
 
487
                                             T12 = rio[WS(ios, 1)];
 
488
                                             Tf0 = FNMS(KP707106781, Tc6, Tc3);
 
489
                                             Tc7 = FMA(KP707106781, Tc6, Tc3);
 
490
                                             {
 
491
                                                  E Tjd, TgY, Tjc, Th1;
 
492
                                                  Tjd = FNMS(KP707106781, TgX, TgW);
 
493
                                                  TgY = FMA(KP707106781, TgX, TgW);
 
494
                                                  Tjc = FNMS(KP707106781, Th0, TgZ);
 
495
                                                  Th1 = FMA(KP707106781, Th0, TgZ);
 
496
                                                  Tf2 = FMA(KP668178637, Tf1, Tf0);
 
497
                                                  Tfr = FNMS(KP668178637, Tf0, Tf1);
 
498
                                                  TdH = FNMS(KP198912367, Tc0, Tc7);
 
499
                                                  Tc8 = FMA(KP198912367, Tc7, Tc0);
 
500
                                                  Tje = FNMS(KP668178637, Tjd, Tjc);
 
501
                                                  TjD = FMA(KP668178637, Tjc, Tjd);
 
502
                                                  ThT = FMA(KP198912367, TgY, Th1);
 
503
                                                  Th2 = FNMS(KP198912367, Th1, TgY);
 
504
                                                  T13 = iio[-WS(ios, 33)];
 
505
                                             }
 
506
                                             T3F = iio[-WS(ios, 1)];
 
507
                                             T3G = rio[WS(ios, 33)];
 
508
                                             T15 = rio[WS(ios, 17)];
 
509
                                             Tcb = T12 - T13;
 
510
                                             T14 = T12 + T13;
 
511
                                             TcD = T3F + T3G;
 
512
                                             T3H = T3F - T3G;
 
513
                                             T16 = iio[-WS(ios, 49)];
 
514
                                             T3I = iio[-WS(ios, 17)];
 
515
                                             T3J = rio[WS(ios, 49)];
 
516
                                        }
 
517
                                        {
 
518
                                             E T1o, T1p, T3u, T3v;
 
519
                                             T1o = iio[-WS(ios, 61)];
 
520
                                             {
 
521
                                                  E TcC, T17, Tcc, T3K;
 
522
                                                  TcC = T15 - T16;
 
523
                                                  T17 = T15 + T16;
 
524
                                                  Tcc = T3I + T3J;
 
525
                                                  T3K = T3I - T3J;
 
526
                                                  TcE = TcC + TcD;
 
527
                                                  Th5 = TcD - TcC;
 
528
                                                  T3b = T14 - T17;
 
529
                                                  T18 = T14 + T17;
 
530
                                                  Tcd = Tcb - Tcc;
 
531
                                                  Thg = Tcb + Tcc;
 
532
                                                  T8b = T3H + T3K;
 
533
                                                  T3L = T3H - T3K;
 
534
                                                  T1p = rio[WS(ios, 29)];
 
535
                                             }
 
536
                                             T3u = iio[-WS(ios, 29)];
 
537
                                             T3v = rio[WS(ios, 61)];
 
538
                                             T1r = rio[WS(ios, 13)];
 
539
                                             Tct = T1o - T1p;
 
540
                                             T1q = T1o + T1p;
 
541
                                             Tcx = T3v + T3u;
 
542
                                             T3w = T3u - T3v;
 
543
                                             T1s = iio[-WS(ios, 45)];
 
544
                                             T3x = iio[-WS(ios, 13)];
 
545
                                             T3y = rio[WS(ios, 45)];
 
546
                                        }
 
547
                                   }
 
548
                                   {
 
549
                                        E T1c, Tci, T1b, Tch, T3e, T1d, T3f, T3g;
 
550
                                        {
 
551
                                             E T19, T1a, T3c, T3d;
 
552
                                             T19 = rio[WS(ios, 9)];
 
553
                                             {
 
554
                                                  E Tcw, T1t, Tcu, T3z;
 
555
                                                  Tcw = T1r - T1s;
 
556
                                                  T1t = T1r + T1s;
 
557
                                                  Tcu = T3x + T3y;
 
558
                                                  T3z = T3x - T3y;
 
559
                                                  Tcy = Tcw - Tcx;
 
560
                                                  Thb = Tcw + Tcx;
 
561
                                                  T3t = T1q - T1t;
 
562
                                                  T1u = T1q + T1t;
 
563
                                                  Tcv = Tct - Tcu;
 
564
                                                  Thc = Tct + Tcu;
 
565
                                                  T87 = T3w + T3z;
 
566
                                                  T3A = T3w - T3z;
 
567
                                                  T1a = iio[-WS(ios, 41)];
 
568
                                             }
 
569
                                             T3c = iio[-WS(ios, 9)];
 
570
                                             T3d = rio[WS(ios, 41)];
 
571
                                             T1c = iio[-WS(ios, 57)];
 
572
                                             Tci = T19 - T1a;
 
573
                                             T1b = T19 + T1a;
 
574
                                             Tch = T3c + T3d;
 
575
                                             T3e = T3c - T3d;
 
576
                                             T1d = rio[WS(ios, 25)];
 
577
                                             T3f = iio[-WS(ios, 25)];
 
578
                                             T3g = rio[WS(ios, 57)];
 
579
                                        }
 
580
                                        {
 
581
                                             E Tce, T1e, Tcf, T3h;
 
582
                                             Tcj = Tch - Tci;
 
583
                                             TcF = Tci + Tch;
 
584
                                             Tce = T1c - T1d;
 
585
                                             T1e = T1c + T1d;
 
586
                                             Tcf = T3g + T3f;
 
587
                                             T3h = T3f - T3g;
 
588
                                             T1f = T1b + T1e;
 
589
                                             T3E = T1e - T1b;
 
590
                                             TcG = Tce + Tcf;
 
591
                                             Tcg = Tce - Tcf;
 
592
                                             T8c = T3e + T3h;
 
593
                                             T3i = T3e - T3h;
 
594
                                        }
 
595
                                   }
 
596
                                   {
 
597
                                        E T1h, T1i, T3l, T3m;
 
598
                                        T1h = rio[WS(ios, 5)];
 
599
                                        T1i = iio[-WS(ios, 37)];
 
600
                                        T3l = iio[-WS(ios, 5)];
 
601
                                        T3m = rio[WS(ios, 37)];
 
602
                                        T1k = rio[WS(ios, 21)];
 
603
                                        Tcm = T1h - T1i;
 
604
                                        T1j = T1h + T1i;
 
605
                                        Tcq = T3l + T3m;
 
606
                                        T3n = T3l - T3m;
 
607
                                        T1l = iio[-WS(ios, 53)];
 
608
                                        T3o = iio[-WS(ios, 21)];
 
609
                                        T3p = rio[WS(ios, 53)];
 
610
                                   }
 
611
                              }
 
612
                              {
 
613
                                   E T85, Tcr, Th8, T3k, Tco, Th9, T3r, T9U, T9V, T88;
 
614
                                   {
 
615
                                        E T1g, T86, T1v, T8a, T8d;
 
616
                                        T85 = T18 - T1f;
 
617
                                        T1g = T18 + T1f;
 
618
                                        {
 
619
                                             E Tcp, T1m, Tcn, T3q, T1n;
 
620
                                             Tcp = T1k - T1l;
 
621
                                             T1m = T1k + T1l;
 
622
                                             Tcn = T3o + T3p;
 
623
                                             T3q = T3o - T3p;
 
624
                                             Tcr = Tcp + Tcq;
 
625
                                             Th8 = Tcq - Tcp;
 
626
                                             T3k = T1j - T1m;
 
627
                                             T1n = T1j + T1m;
 
628
                                             Tco = Tcm - Tcn;
 
629
                                             Th9 = Tcm + Tcn;
 
630
                                             T86 = T3n + T3q;
 
631
                                             T3r = T3n - T3q;
 
632
                                             T1v = T1n + T1u;
 
633
                                             T8a = T1u - T1n;
 
634
                                        }
 
635
                                        T8d = T8b - T8c;
 
636
                                        T9U = T8c + T8b;
 
637
                                        T9T = T1g - T1v;
 
638
                                        T1w = T1g + T1v;
 
639
                                        T9a = T8d - T8a;
 
640
                                        T8e = T8a + T8d;
 
641
                                        T9V = T86 + T87;
 
642
                                        T88 = T86 - T87;
 
643
                                   }
 
644
                                   {
 
645
                                        E T6f, T3j, T6i, T3M, T3N, T3O, T6j, T3C, T3s, T3B;
 
646
                                        T6f = T3b - T3i;
 
647
                                        T3j = T3b + T3i;
 
648
                                        TaF = T9V + T9U;
 
649
                                        T9W = T9U - T9V;
 
650
                                        T99 = T85 - T88;
 
651
                                        T89 = T85 + T88;
 
652
                                        T6i = T3L - T3E;
 
653
                                        T3M = T3E + T3L;
 
654
                                        T3N = T3r - T3k;
 
655
                                        T3s = T3k + T3r;
 
656
                                        T3B = T3t - T3A;
 
657
                                        T3O = T3t + T3A;
 
658
                                        T6j = T3s - T3B;
 
659
                                        T3C = T3s + T3B;
 
660
                                        {
 
661
                                             E Th6, Tcl, Tf8, Thh, TcJ, TcK, Tf5, TcI, Tf6, TcA;
 
662
                                             {
 
663
                                                  E TcH, T6g, T3P, Tck, Tcs, Tcz;
 
664
                                                  Th6 = Tcj + Tcg;
 
665
                                                  Tck = Tcg - Tcj;
 
666
                                                  T6k = FMA(KP707106781, T6j, T6i);
 
667
                                                  T7a = FNMS(KP707106781, T6j, T6i);
 
668
                                                  T5p = FNMS(KP707106781, T3C, T3j);
 
669
                                                  T3D = FMA(KP707106781, T3C, T3j);
 
670
                                                  T6g = T3O - T3N;
 
671
                                                  T3P = T3N + T3O;
 
672
                                                  Tcl = FMA(KP707106781, Tck, Tcd);
 
673
                                                  Tf8 = FNMS(KP707106781, Tck, Tcd);
 
674
                                                  T6h = FMA(KP707106781, T6g, T6f);
 
675
                                                  T79 = FNMS(KP707106781, T6g, T6f);
 
676
                                                  T5q = FNMS(KP707106781, T3P, T3M);
 
677
                                                  T3Q = FMA(KP707106781, T3P, T3M);
 
678
                                                  TcH = TcF - TcG;
 
679
                                                  Thh = TcF + TcG;
 
680
                                                  TcJ = FMA(KP414213562, Tco, Tcr);
 
681
                                                  Tcs = FNMS(KP414213562, Tcr, Tco);
 
682
                                                  Tcz = FMA(KP414213562, Tcy, Tcv);
 
683
                                                  TcK = FNMS(KP414213562, Tcv, Tcy);
 
684
                                                  Tf5 = FNMS(KP707106781, TcH, TcE);
 
685
                                                  TcI = FMA(KP707106781, TcH, TcE);
 
686
                                                  Tf6 = Tcz - Tcs;
 
687
                                                  TcA = Tcs + Tcz;
 
688
                                             }
 
689
                                             {
 
690
                                                  E Tf9, TcL, Tha, Thd;
 
691
                                                  Tf7 = FMA(KP923879532, Tf6, Tf5);
 
692
                                                  Tg0 = FNMS(KP923879532, Tf6, Tf5);
 
693
                                                  Teg = FNMS(KP923879532, TcA, Tcl);
 
694
                                                  TcB = FMA(KP923879532, TcA, Tcl);
 
695
                                                  Tf9 = TcJ - TcK;
 
696
                                                  TcL = TcJ + TcK;
 
697
                                                  Th7 = FMA(KP707106781, Th6, Th5);
 
698
                                                  Tjk = FNMS(KP707106781, Th6, Th5);
 
699
                                                  Tfa = FMA(KP923879532, Tf9, Tf8);
 
700
                                                  TfZ = FNMS(KP923879532, Tf9, Tf8);
 
701
                                                  Tef = FNMS(KP923879532, TcL, TcI);
 
702
                                                  TcM = FMA(KP923879532, TcL, TcI);
 
703
                                                  Thj = FMA(KP414213562, Th8, Th9);
 
704
                                                  Tha = FNMS(KP414213562, Th9, Th8);
 
705
                                                  Thd = FNMS(KP414213562, Thc, Thb);
 
706
                                                  Thk = FMA(KP414213562, Thb, Thc);
 
707
                                                  Tjh = FNMS(KP707106781, Thh, Thg);
 
708
                                                  Thi = FMA(KP707106781, Thh, Thg);
 
709
                                                  Tji = Tha + Thd;
 
710
                                                  The = Tha - Thd;
 
711
                                             }
 
712
                                        }
 
713
                                   }
 
714
                              }
 
715
                         }
 
716
                         {
 
717
                              E Tdh, Tho, T3S, T1D, TcQ, Thz, T8m, T4s, Tdb, Thu, T4a, T1Z, Td8, Thv, T8i;
 
718
                              E T4h, TcT, Tdi, T1K, T4l, Tdj, TcW, T3Z, T8n, TcZ, T1O, T45, Td3, T44, Td2;
 
719
                              E T1R, T46;
 
720
                              {
 
721
                                   E Td6, T1V, T4e, Tda, T4d, Td9, T1Y, T4f;
 
722
                                   {
 
723
                                        E TcO, T1z, T4p, Tdg, T4o, Tdf, T1C, T4q;
 
724
                                        {
 
725
                                             E T4m, T4n, T1A, T1B;
 
726
                                             {
 
727
                                                  E T1x, Tjl, Thl, T1y;
 
728
                                                  T1x = iio[-WS(ios, 63)];
 
729
                                                  Tjj = FNMS(KP923879532, Tji, Tjh);
 
730
                                                  Tkc = FMA(KP923879532, Tji, Tjh);
 
731
                                                  Tis = FNMS(KP923879532, The, Th7);
 
732
                                                  Thf = FMA(KP923879532, The, Th7);
 
733
                                                  Tjl = Thj - Thk;
 
734
                                                  Thl = Thj + Thk;
 
735
                                                  T1y = rio[WS(ios, 31)];
 
736
                                                  T4m = iio[-WS(ios, 31)];
 
737
                                                  Tjm = FMA(KP923879532, Tjl, Tjk);
 
738
                                                  Tkb = FNMS(KP923879532, Tjl, Tjk);
 
739
                                                  Tir = FNMS(KP923879532, Thl, Thi);
 
740
                                                  Thm = FMA(KP923879532, Thl, Thi);
 
741
                                                  TcO = T1x - T1y;
 
742
                                                  T1z = T1x + T1y;
 
743
                                                  T4n = rio[WS(ios, 63)];
 
744
                                             }
 
745
                                             T1A = rio[WS(ios, 15)];
 
746
                                             T1B = iio[-WS(ios, 47)];
 
747
                                             T4p = iio[-WS(ios, 15)];
 
748
                                             Tdg = T4n + T4m;
 
749
                                             T4o = T4m - T4n;
 
750
                                             Tdf = T1A - T1B;
 
751
                                             T1C = T1A + T1B;
 
752
                                             T4q = rio[WS(ios, 47)];
 
753
                                        }
 
754
                                        {
 
755
                                             E T4b, T4c, T1W, T1X;
 
756
                                             {
 
757
                                                  E T1T, TcP, T4r, T1U;
 
758
                                                  T1T = iio[-WS(ios, 59)];
 
759
                                                  Tdh = Tdf - Tdg;
 
760
                                                  Tho = Tdf + Tdg;
 
761
                                                  T3S = T1z - T1C;
 
762
                                                  T1D = T1z + T1C;
 
763
                                                  TcP = T4p + T4q;
 
764
                                                  T4r = T4p - T4q;
 
765
                                                  T1U = rio[WS(ios, 27)];
 
766
                                                  T4b = iio[-WS(ios, 27)];
 
767
                                                  TcQ = TcO - TcP;
 
768
                                                  Thz = TcO + TcP;
 
769
                                                  T8m = T4o + T4r;
 
770
                                                  T4s = T4o - T4r;
 
771
                                                  Td6 = T1T - T1U;
 
772
                                                  T1V = T1T + T1U;
 
773
                                                  T4c = rio[WS(ios, 59)];
 
774
                                             }
 
775
                                             T1W = rio[WS(ios, 11)];
 
776
                                             T1X = iio[-WS(ios, 43)];
 
777
                                             T4e = iio[-WS(ios, 11)];
 
778
                                             Tda = T4c + T4b;
 
779
                                             T4d = T4b - T4c;
 
780
                                             Td9 = T1W - T1X;
 
781
                                             T1Y = T1W + T1X;
 
782
                                             T4f = rio[WS(ios, 43)];
 
783
                                        }
 
784
                                   }
 
785
                                   {
 
786
                                        E T3V, TcU, T3Y, TcV;
 
787
                                        {
 
788
                                             E TcR, T1G, T3W, TcS, T1J, T3X;
 
789
                                             {
 
790
                                                  E T3T, T3U, T1H, T1I;
 
791
                                                  {
 
792
                                                       E T1E, Td7, T4g, T1F;
 
793
                                                       T1E = rio[WS(ios, 7)];
 
794
                                                       Tdb = Td9 - Tda;
 
795
                                                       Thu = Td9 + Tda;
 
796
                                                       T4a = T1V - T1Y;
 
797
                                                       T1Z = T1V + T1Y;
 
798
                                                       Td7 = T4e + T4f;
 
799
                                                       T4g = T4e - T4f;
 
800
                                                       T1F = iio[-WS(ios, 39)];
 
801
                                                       T3T = iio[-WS(ios, 7)];
 
802
                                                       Td8 = Td6 - Td7;
 
803
                                                       Thv = Td6 + Td7;
 
804
                                                       T8i = T4d + T4g;
 
805
                                                       T4h = T4d - T4g;
 
806
                                                       TcR = T1E - T1F;
 
807
                                                       T1G = T1E + T1F;
 
808
                                                       T3U = rio[WS(ios, 39)];
 
809
                                                  }
 
810
                                                  T1H = iio[-WS(ios, 55)];
 
811
                                                  T1I = rio[WS(ios, 23)];
 
812
                                                  T3W = iio[-WS(ios, 23)];
 
813
                                                  TcS = T3T + T3U;
 
814
                                                  T3V = T3T - T3U;
 
815
                                                  TcU = T1H - T1I;
 
816
                                                  T1J = T1H + T1I;
 
817
                                                  T3X = rio[WS(ios, 55)];
 
818
                                             }
 
819
                                             TcT = TcR - TcS;
 
820
                                             Tdi = TcR + TcS;
 
821
                                             T1K = T1G + T1J;
 
822
                                             T4l = T1J - T1G;
 
823
                                             T3Y = T3W - T3X;
 
824
                                             TcV = T3X + T3W;
 
825
                                        }
 
826
                                        {
 
827
                                             E T42, T43, T1M, T1N, T1P, T1Q;
 
828
                                             T1M = rio[WS(ios, 3)];
 
829
                                             T1N = iio[-WS(ios, 35)];
 
830
                                             Tdj = TcU + TcV;
 
831
                                             TcW = TcU - TcV;
 
832
                                             T3Z = T3V - T3Y;
 
833
                                             T8n = T3V + T3Y;
 
834
                                             TcZ = T1M - T1N;
 
835
                                             T1O = T1M + T1N;
 
836
                                             T42 = iio[-WS(ios, 3)];
 
837
                                             T43 = rio[WS(ios, 35)];
 
838
                                             T1P = rio[WS(ios, 19)];
 
839
                                             T1Q = iio[-WS(ios, 51)];
 
840
                                             T45 = iio[-WS(ios, 19)];
 
841
                                             Td3 = T42 + T43;
 
842
                                             T44 = T42 - T43;
 
843
                                             Td2 = T1P - T1Q;
 
844
                                             T1R = T1P + T1Q;
 
845
                                             T46 = rio[WS(ios, 51)];
 
846
                                        }
 
847
                                   }
 
848
                              }
 
849
                              {
 
850
                                   E T8g, Td4, Thr, T41, T9Z, Ths, Td1, T48, Ta0, T8j, T8h;
 
851
                                   {
 
852
                                        E T1L, T1S, Td0, T47, T8o, T20, T8l;
 
853
                                        T8g = T1D - T1K;
 
854
                                        T1L = T1D + T1K;
 
855
                                        Td4 = Td2 + Td3;
 
856
                                        Thr = Td2 - Td3;
 
857
                                        T41 = T1O - T1R;
 
858
                                        T1S = T1O + T1R;
 
859
                                        Td0 = T45 + T46;
 
860
                                        T47 = T45 - T46;
 
861
                                        T8o = T8m - T8n;
 
862
                                        T9Z = T8n + T8m;
 
863
                                        T20 = T1S + T1Z;
 
864
                                        T8l = T1Z - T1S;
 
865
                                        Ths = TcZ + Td0;
 
866
                                        Td1 = TcZ - Td0;
 
867
                                        T48 = T44 - T47;
 
868
                                        T8h = T44 + T47;
 
869
                                        T21 = T1L + T20;
 
870
                                        T9Y = T1L - T20;
 
871
                                        T9d = T8o - T8l;
 
872
                                        T8p = T8l + T8o;
 
873
                                   }
 
874
                                   Ta0 = T8h + T8i;
 
875
                                   T8j = T8h - T8i;
 
876
                                   {
 
877
                                        E T6m, T40, T6p, T4t, T4u, T4v, T6q, T4j, T49, T4i;
 
878
                                        T6m = T3S - T3Z;
 
879
                                        T40 = T3S + T3Z;
 
880
                                        TaG = Ta0 + T9Z;
 
881
                                        Ta1 = T9Z - Ta0;
 
882
                                        T9c = T8g - T8j;
 
883
                                        T8k = T8g + T8j;
 
884
                                        T6p = T4s - T4l;
 
885
                                        T4t = T4l + T4s;
 
886
                                        T4u = T48 - T41;
 
887
                                        T49 = T41 + T48;
 
888
                                        T4i = T4a - T4h;
 
889
                                        T4v = T4a + T4h;
 
890
                                        T6q = T49 - T4i;
 
891
                                        T4j = T49 + T4i;
 
892
                                        {
 
893
                                             E Thp, TcY, Tff, ThA, Tdm, Tdn, Tfc, Tdl, Tfd, Tdd;
 
894
                                             {
 
895
                                                  E Tdk, T6n, T4w, TcX, Td5, Tdc;
 
896
                                                  Thp = TcT - TcW;
 
897
                                                  TcX = TcT + TcW;
 
898
                                                  T6r = FMA(KP707106781, T6q, T6p);
 
899
                                                  T7d = FNMS(KP707106781, T6q, T6p);
 
900
                                                  T5s = FNMS(KP707106781, T4j, T40);
 
901
                                                  T4k = FMA(KP707106781, T4j, T40);
 
902
                                                  T6n = T4v - T4u;
 
903
                                                  T4w = T4u + T4v;
 
904
                                                  TcY = FMA(KP707106781, TcX, TcQ);
 
905
                                                  Tff = FNMS(KP707106781, TcX, TcQ);
 
906
                                                  T6o = FMA(KP707106781, T6n, T6m);
 
907
                                                  T7c = FNMS(KP707106781, T6n, T6m);
 
908
                                                  T5t = FNMS(KP707106781, T4w, T4t);
 
909
                                                  T4x = FMA(KP707106781, T4w, T4t);
 
910
                                                  Tdk = Tdi - Tdj;
 
911
                                                  ThA = Tdi + Tdj;
 
912
                                                  Tdm = FMA(KP414213562, Td1, Td4);
 
913
                                                  Td5 = FNMS(KP414213562, Td4, Td1);
 
914
                                                  Tdc = FMA(KP414213562, Tdb, Td8);
 
915
                                                  Tdn = FNMS(KP414213562, Td8, Tdb);
 
916
                                                  Tfc = FNMS(KP707106781, Tdk, Tdh);
 
917
                                                  Tdl = FMA(KP707106781, Tdk, Tdh);
 
918
                                                  Tfd = Tdc - Td5;
 
919
                                                  Tdd = Td5 + Tdc;
 
920
                                             }
 
921
                                             {
 
922
                                                  E Tfg, Tdo, Tht, Thw;
 
923
                                                  Tfe = FMA(KP923879532, Tfd, Tfc);
 
924
                                                  Tg3 = FNMS(KP923879532, Tfd, Tfc);
 
925
                                                  Tej = FNMS(KP923879532, Tdd, TcY);
 
926
                                                  Tde = FMA(KP923879532, Tdd, TcY);
 
927
                                                  Tfg = Tdm - Tdn;
 
928
                                                  Tdo = Tdm + Tdn;
 
929
                                                  Thq = FMA(KP707106781, Thp, Tho);
 
930
                                                  Tjr = FNMS(KP707106781, Thp, Tho);
 
931
                                                  Tfh = FMA(KP923879532, Tfg, Tff);
 
932
                                                  Tg2 = FNMS(KP923879532, Tfg, Tff);
 
933
                                                  Tei = FNMS(KP923879532, Tdo, Tdl);
 
934
                                                  Tdp = FMA(KP923879532, Tdo, Tdl);
 
935
                                                  ThC = FNMS(KP414213562, Thr, Ths);
 
936
                                                  Tht = FMA(KP414213562, Ths, Thr);
 
937
                                                  Thw = FNMS(KP414213562, Thv, Thu);
 
938
                                                  ThD = FMA(KP414213562, Thu, Thv);
 
939
                                                  Tjo = FNMS(KP707106781, ThA, Thz);
 
940
                                                  ThB = FMA(KP707106781, ThA, Thz);
 
941
                                                  Tjp = Tht - Thw;
 
942
                                                  Thx = Tht + Thw;
 
943
                                             }
 
944
                                        }
 
945
                                   }
 
946
                              }
 
947
                         }
 
948
                    }
 
949
               }
 
950
               {
 
951
                    E TiX, TiU, TiT, TeL, TeI, TeH;
 
952
                    {
 
953
                         E Tjq, Tkf, Tiv, Thy, Tjt, Tke, Tiu, ThF, T9F, T9C, T9x, T9G, T9B;
 
954
                         {
 
955
                              E Tav, Tas, Tan, Taw, Tar;
 
956
                              {
 
957
                                   E Taa, TaY, Tb3, Tb0, TaX, Tb4;
 
958
                                   {
 
959
                                        E T22, TaL, Tb2, T11, TaM, TaS, TaI, TaO, Tb1;
 
960
                                        {
 
961
                                             E Tjs, ThE, TaH, TaE;
 
962
                                             T22 = T1w + T21;
 
963
                                             TaL = T21 - T1w;
 
964
                                             Tjq = FMA(KP923879532, Tjp, Tjo);
 
965
                                             Tkf = FNMS(KP923879532, Tjp, Tjo);
 
966
                                             Tiv = FNMS(KP923879532, Thx, Thq);
 
967
                                             Thy = FMA(KP923879532, Thx, Thq);
 
968
                                             Tjs = ThD - ThC;
 
969
                                             ThE = ThC + ThD;
 
970
                                             Tb2 = TaF + TaG;
 
971
                                             TaH = TaF - TaG;
 
972
                                             TaE = Tv - T10;
 
973
                                             T11 = Tv + T10;
 
974
                                             Tjt = FMA(KP923879532, Tjs, Tjr);
 
975
                                             Tke = FNMS(KP923879532, Tjs, Tjr);
 
976
                                             Tiu = FNMS(KP923879532, ThE, ThB);
 
977
                                             ThF = FMA(KP923879532, ThE, ThB);
 
978
                                             Taa = Ta8 - Ta9;
 
979
                                             TaM = Ta9 + Ta8;
 
980
                                             TaY = T11 - T22;
 
981
                                             TaS = TaE - TaH;
 
982
                                             TaI = TaE + TaH;
 
983
                                        }
 
984
                                        rio[0] = T11 + T22;
 
985
                                        TaO = TaM - TaN;
 
986
                                        Tb1 = TaN + TaM;
 
987
                                        {
 
988
                                             E TaK, TaD, TaP, TaV, TaQ, TaJ;
 
989
                                             TaK = W[95];
 
990
                                             TaD = W[94];
 
991
                                             iio[-WS(ios, 63)] = Tb2 + Tb1;
 
992
                                             TaP = TaL + TaO;
 
993
                                             TaV = TaO - TaL;
 
994
                                             TaQ = TaK * TaI;
 
995
                                             TaJ = TaD * TaI;
 
996
                                             {
 
997
                                                  E TaU, TaR, TaW, TaT;
 
998
                                                  TaU = W[31];
 
999
                                                  iio[-WS(ios, 15)] = FMA(TaD, TaP, TaQ);
 
1000
                                                  rio[WS(ios, 48)] = FNMS(TaK, TaP, TaJ);
 
1001
                                                  TaR = W[30];
 
1002
                                                  TaW = TaU * TaS;
 
1003
                                                  Tb3 = Tb1 - Tb2;
 
1004
                                                  Tb0 = W[63];
 
1005
                                                  TaT = TaR * TaS;
 
1006
                                                  iio[-WS(ios, 47)] = FMA(TaR, TaV, TaW);
 
1007
                                                  TaX = W[62];
 
1008
                                                  Tb4 = Tb0 * TaY;
 
1009
                                                  rio[WS(ios, 16)] = FNMS(TaU, TaV, TaT);
 
1010
                                             }
 
1011
                                        }
 
1012
                                   }
 
1013
                                   {
 
1014
                                        E Tao, T9S, T9X, Tab, Tat, Ta2, Tap, Tae, TaZ, Tac, Tad;
 
1015
                                        TaZ = TaX * TaY;
 
1016
                                        iio[-WS(ios, 31)] = FMA(TaX, Tb3, Tb4);
 
1017
                                        Tao = T9O - T9R;
 
1018
                                        T9S = T9O + T9R;
 
1019
                                        rio[WS(ios, 32)] = FNMS(Tb0, Tb3, TaZ);
 
1020
                                        Tac = T9W - T9T;
 
1021
                                        T9X = T9T + T9W;
 
1022
                                        Tab = Ta7 + Taa;
 
1023
                                        Tat = Taa - Ta7;
 
1024
                                        Ta2 = T9Y - Ta1;
 
1025
                                        Tad = T9Y + Ta1;
 
1026
                                        Tap = Tad - Tac;
 
1027
                                        Tae = Tac + Tad;
 
1028
                                        {
 
1029
                                             E Tal, Taq, Tai, TaB, TaA, Taz;
 
1030
                                             {
 
1031
                                                  E Ta6, Tau, Taf, Tay, Ta4, T9N, Ta3;
 
1032
                                                  Ta6 = W[111];
 
1033
                                                  Tau = T9X - Ta2;
 
1034
                                                  Ta3 = T9X + Ta2;
 
1035
                                                  Taf = FMA(KP707106781, Tae, Tab);
 
1036
                                                  Tal = FNMS(KP707106781, Tae, Tab);
 
1037
                                                  Tay = FNMS(KP707106781, Tap, Tao);
 
1038
                                                  Taq = FMA(KP707106781, Tap, Tao);
 
1039
                                                  Ta4 = FMA(KP707106781, Ta3, T9S);
 
1040
                                                  Tai = FNMS(KP707106781, Ta3, T9S);
 
1041
                                                  T9N = W[110];
 
1042
                                                  {
 
1043
                                                       E Tax, TaC, Tag, Ta5;
 
1044
                                                       TaB = FNMS(KP707106781, Tau, Tat);
 
1045
                                                       Tav = FMA(KP707106781, Tau, Tat);
 
1046
                                                       TaA = W[79];
 
1047
                                                       Tag = Ta6 * Ta4;
 
1048
                                                       Ta5 = T9N * Ta4;
 
1049
                                                       Tax = W[78];
 
1050
                                                       TaC = TaA * Tay;
 
1051
                                                       iio[-WS(ios, 7)] = FMA(T9N, Taf, Tag);
 
1052
                                                       rio[WS(ios, 56)] = FNMS(Ta6, Taf, Ta5);
 
1053
                                                       Taz = Tax * Tay;
 
1054
                                                       iio[-WS(ios, 23)] = FMA(Tax, TaB, TaC);
 
1055
                                                  }
 
1056
                                             }
 
1057
                                             {
 
1058
                                                  E Tak, Tah, Tam, Taj;
 
1059
                                                  Tak = W[47];
 
1060
                                                  rio[WS(ios, 40)] = FNMS(TaA, TaB, Taz);
 
1061
                                                  Tah = W[46];
 
1062
                                                  Tam = Tak * Tai;
 
1063
                                                  Tas = W[15];
 
1064
                                                  Taj = Tah * Tai;
 
1065
                                                  Tan = W[14];
 
1066
                                                  iio[-WS(ios, 39)] = FMA(Tah, Tal, Tam);
 
1067
                                                  Taw = Tas * Taq;
 
1068
                                                  rio[WS(ios, 24)] = FNMS(Tak, Tal, Taj);
 
1069
                                                  Tar = Tan * Taq;
 
1070
                                             }
 
1071
                                        }
 
1072
                                   }
 
1073
                              }
 
1074
                              {
 
1075
                                   E T9j, T97, T9k, T8X, T8U, T8P, T8Y, T8T;
 
1076
                                   {
 
1077
                                        E T8E, T8f, T8D, T8V, T84, T8Q, T8q, T8F, T8z, T8C, T83;
 
1078
                                        T9j = T8y - T8v;
 
1079
                                        T8z = T8v + T8y;
 
1080
                                        iio[-WS(ios, 55)] = FMA(Tan, Tav, Taw);
 
1081
                                        T97 = T8B - T8A;
 
1082
                                        T8C = T8A + T8B;
 
1083
                                        rio[WS(ios, 8)] = FNMS(Tas, Tav, Tar);
 
1084
                                        T9k = T7X - T82;
 
1085
                                        T83 = T7X + T82;
 
1086
                                        T8E = FNMS(KP414213562, T89, T8e);
 
1087
                                        T8f = FMA(KP414213562, T8e, T89);
 
1088
                                        T8D = FMA(KP707106781, T8C, T8z);
 
1089
                                        T8V = FNMS(KP707106781, T8C, T8z);
 
1090
                                        T84 = FMA(KP707106781, T83, T7S);
 
1091
                                        T8Q = FNMS(KP707106781, T83, T7S);
 
1092
                                        T8q = FNMS(KP414213562, T8p, T8k);
 
1093
                                        T8F = FMA(KP414213562, T8k, T8p);
 
1094
                                        {
 
1095
                                             E T8N, T8S, T8K, T93, T92, T91;
 
1096
                                             {
 
1097
                                                  E T8u, T8W, T8H, T90, T8s, T7N, T8R, T8G, T8r;
 
1098
                                                  T8u = W[119];
 
1099
                                                  T8R = T8F - T8E;
 
1100
                                                  T8G = T8E + T8F;
 
1101
                                                  T8r = T8f + T8q;
 
1102
                                                  T8W = T8f - T8q;
 
1103
                                                  T8N = FNMS(KP923879532, T8G, T8D);
 
1104
                                                  T8H = FMA(KP923879532, T8G, T8D);
 
1105
                                                  T8S = FMA(KP923879532, T8R, T8Q);
 
1106
                                                  T90 = FNMS(KP923879532, T8R, T8Q);
 
1107
                                                  T8K = FNMS(KP923879532, T8r, T84);
 
1108
                                                  T8s = FMA(KP923879532, T8r, T84);
 
1109
                                                  T7N = W[118];
 
1110
                                                  {
 
1111
                                                       E T8Z, T94, T8I, T8t;
 
1112
                                                       T8X = FMA(KP923879532, T8W, T8V);
 
1113
                                                       T93 = FNMS(KP923879532, T8W, T8V);
 
1114
                                                       T92 = W[87];
 
1115
                                                       T8I = T8u * T8s;
 
1116
                                                       T8t = T7N * T8s;
 
1117
                                                       T8Z = W[86];
 
1118
                                                       T94 = T92 * T90;
 
1119
                                                       iio[-WS(ios, 3)] = FMA(T7N, T8H, T8I);
 
1120
                                                       rio[WS(ios, 60)] = FNMS(T8u, T8H, T8t);
 
1121
                                                       T91 = T8Z * T90;
 
1122
                                                       iio[-WS(ios, 19)] = FMA(T8Z, T93, T94);
 
1123
                                                  }
 
1124
                                             }
 
1125
                                             {
 
1126
                                                  E T8M, T8J, T8O, T8L;
 
1127
                                                  T8M = W[55];
 
1128
                                                  rio[WS(ios, 44)] = FNMS(T92, T93, T91);
 
1129
                                                  T8J = W[54];
 
1130
                                                  T8O = T8M * T8K;
 
1131
                                                  T8U = W[23];
 
1132
                                                  T8L = T8J * T8K;
 
1133
                                                  T8P = W[22];
 
1134
                                                  iio[-WS(ios, 35)] = FMA(T8J, T8N, T8O);
 
1135
                                                  T8Y = T8U * T8S;
 
1136
                                                  rio[WS(ios, 28)] = FNMS(T8M, T8N, T8L);
 
1137
                                                  T8T = T8P * T8S;
 
1138
                                             }
 
1139
                                        }
 
1140
                                   }
 
1141
                                   {
 
1142
                                        E T9b, T9m, T9D, T9l, T9y, T98, T9n, T9e;
 
1143
                                        iio[-WS(ios, 51)] = FMA(T8P, T8X, T8Y);
 
1144
                                        rio[WS(ios, 12)] = FNMS(T8U, T8X, T8T);
 
1145
                                        T9b = FNMS(KP414213562, T9a, T99);
 
1146
                                        T9m = FMA(KP414213562, T99, T9a);
 
1147
                                        T9D = FNMS(KP707106781, T9k, T9j);
 
1148
                                        T9l = FMA(KP707106781, T9k, T9j);
 
1149
                                        T9y = FNMS(KP707106781, T97, T96);
 
1150
                                        T98 = FMA(KP707106781, T97, T96);
 
1151
                                        T9n = FNMS(KP414213562, T9c, T9d);
 
1152
                                        T9e = FMA(KP414213562, T9d, T9c);
 
1153
                                        {
 
1154
                                             E T9v, T9A, T9s, T9L, T9K, T9J;
 
1155
                                             {
 
1156
                                                  E T9i, T9E, T9p, T9I, T9g, T95, T9o, T9z, T9f;
 
1157
                                                  T9i = W[7];
 
1158
                                                  T9o = T9m + T9n;
 
1159
                                                  T9z = T9m - T9n;
 
1160
                                                  T9E = T9e - T9b;
 
1161
                                                  T9f = T9b + T9e;
 
1162
                                                  T9p = FMA(KP923879532, T9o, T9l);
 
1163
                                                  T9v = FNMS(KP923879532, T9o, T9l);
 
1164
                                                  T9I = FNMS(KP923879532, T9z, T9y);
 
1165
                                                  T9A = FMA(KP923879532, T9z, T9y);
 
1166
                                                  T9g = FMA(KP923879532, T9f, T98);
 
1167
                                                  T9s = FNMS(KP923879532, T9f, T98);
 
1168
                                                  T95 = W[6];
 
1169
                                                  {
 
1170
                                                       E T9H, T9M, T9q, T9h;
 
1171
                                                       T9L = FNMS(KP923879532, T9E, T9D);
 
1172
                                                       T9F = FMA(KP923879532, T9E, T9D);
 
1173
                                                       T9K = W[39];
 
1174
                                                       T9q = T9i * T9g;
 
1175
                                                       T9h = T95 * T9g;
 
1176
                                                       T9H = W[38];
 
1177
                                                       T9M = T9K * T9I;
 
1178
                                                       iio[-WS(ios, 59)] = FMA(T95, T9p, T9q);
 
1179
                                                       rio[WS(ios, 4)] = FNMS(T9i, T9p, T9h);
 
1180
                                                       T9J = T9H * T9I;
 
1181
                                                       iio[-WS(ios, 43)] = FMA(T9H, T9L, T9M);
 
1182
                                                  }
 
1183
                                             }
 
1184
                                             {
 
1185
                                                  E T9u, T9r, T9w, T9t;
 
1186
                                                  T9u = W[71];
 
1187
                                                  rio[WS(ios, 20)] = FNMS(T9K, T9L, T9J);
 
1188
                                                  T9r = W[70];
 
1189
                                                  T9w = T9u * T9s;
 
1190
                                                  T9C = W[103];
 
1191
                                                  T9t = T9r * T9s;
 
1192
                                                  T9x = W[102];
 
1193
                                                  iio[-WS(ios, 27)] = FMA(T9r, T9v, T9w);
 
1194
                                                  T9G = T9C * T9A;
 
1195
                                                  rio[WS(ios, 36)] = FNMS(T9u, T9v, T9t);
 
1196
                                                  T9B = T9x * T9A;
 
1197
                                             }
 
1198
                                        }
 
1199
                                   }
 
1200
                              }
 
1201
                         }
 
1202
                         {
 
1203
                              E TjB, Tjf, Tj8, TjE, TiB, Tip, Tio, TiC, Tif, Tic, Tib, TkH, TkE, TkD, TjZ;
 
1204
                              E TjW, TjV;
 
1205
                              {
 
1206
                                   E TkK, Tkx, Tki, TkC, TkN, Tku, Tkr;
 
1207
                                   {
 
1208
                                        E Tkd, Tkg, Tka, TkF, TkG, Tkq, Tkl, Tkm;
 
1209
                                        {
 
1210
                                             E Tko, Tkp, Tk8, Tk9;
 
1211
                                             TjB = FMA(KP923879532, TjA, Tjz);
 
1212
                                             Tk8 = FNMS(KP923879532, TjA, Tjz);
 
1213
                                             iio[-WS(ios, 11)] = FMA(T9x, T9F, T9G);
 
1214
                                             Tjf = Tjb + Tje;
 
1215
                                             Tk9 = Tje - Tjb;
 
1216
                                             rio[WS(ios, 52)] = FNMS(T9C, T9F, T9B);
 
1217
                                             Tkd = FNMS(KP534511135, Tkc, Tkb);
 
1218
                                             Tko = FMA(KP534511135, Tkb, Tkc);
 
1219
                                             Tkp = FMA(KP534511135, Tke, Tkf);
 
1220
                                             Tkg = FNMS(KP534511135, Tkf, Tke);
 
1221
                                             Tka = FMA(KP831469612, Tk9, Tk8);
 
1222
                                             TkF = FNMS(KP831469612, Tk9, Tk8);
 
1223
                                             TkG = Tko - Tkp;
 
1224
                                             Tkq = Tko + Tkp;
 
1225
                                             Tj8 = FNMS(KP923879532, Tj7, Tj6);
 
1226
                                             Tkl = FMA(KP923879532, Tj7, Tj6);
 
1227
                                             Tkm = TjC + TjD;
 
1228
                                             TjE = TjC - TjD;
 
1229
                                        }
 
1230
                                        {
 
1231
                                             E TkB, Tkh, TkA, Tkn;
 
1232
                                             TkH = FMA(KP881921264, TkG, TkF);
 
1233
                                             TkK = FNMS(KP881921264, TkG, TkF);
 
1234
                                             TkB = Tkd + Tkg;
 
1235
                                             Tkh = Tkd - Tkg;
 
1236
                                             TkA = FNMS(KP831469612, Tkm, Tkl);
 
1237
                                             Tkn = FMA(KP831469612, Tkm, Tkl);
 
1238
                                             Tkx = FNMS(KP881921264, Tkh, Tka);
 
1239
                                             Tki = FMA(KP881921264, Tkh, Tka);
 
1240
                                             TkC = FNMS(KP881921264, TkB, TkA);
 
1241
                                             TkN = FMA(KP881921264, TkB, TkA);
 
1242
                                             Tku = FNMS(KP881921264, Tkq, Tkn);
 
1243
                                             Tkr = FMA(KP881921264, Tkq, Tkn);
 
1244
                                        }
 
1245
                                   }
 
1246
                                   {
 
1247
                                        E TkM, TkL, Tk7, Tkk, Tkt, Tkw;
 
1248
                                        Tk7 = W[116];
 
1249
                                        Tkk = W[117];
 
1250
                                        {
 
1251
                                             E TkJ, Tks, Tkj, TkO;
 
1252
                                             TkJ = W[84];
 
1253
                                             TkM = W[85];
 
1254
                                             Tks = Tk7 * Tkr;
 
1255
                                             Tkj = Tk7 * Tki;
 
1256
                                             TkO = TkJ * TkN;
 
1257
                                             TkL = TkJ * TkK;
 
1258
                                             rio[WS(ios, 59)] = FNMS(Tkk, Tki, Tks);
 
1259
                                             iio[-WS(ios, 4)] = FMA(Tkk, Tkr, Tkj);
 
1260
                                             rio[WS(ios, 43)] = FNMS(TkM, TkK, TkO);
 
1261
                                        }
 
1262
                                        iio[-WS(ios, 20)] = FMA(TkM, TkN, TkL);
 
1263
                                        Tkt = W[52];
 
1264
                                        Tkw = W[53];
 
1265
                                        {
 
1266
                                             E Tkz, Tky, Tkv, TkI;
 
1267
                                             Tkz = W[20];
 
1268
                                             TkE = W[21];
 
1269
                                             Tky = Tkt * Tkx;
 
1270
                                             Tkv = Tkt * Tku;
 
1271
                                             TkI = Tkz * TkH;
 
1272
                                             TkD = Tkz * TkC;
 
1273
                                             iio[-WS(ios, 36)] = FMA(Tkw, Tku, Tky);
 
1274
                                             rio[WS(ios, 27)] = FNMS(Tkw, Tkx, Tkv);
 
1275
                                             iio[-WS(ios, 52)] = FMA(TkE, TkC, TkI);
 
1276
                                        }
 
1277
                                   }
 
1278
                              }
 
1279
                              rio[WS(ios, 11)] = FNMS(TkE, TkH, TkD);
 
1280
                              {
 
1281
                                   E Tii, Ti5, ThI, Tia, Til, Ti2, ThZ;
 
1282
                                   {
 
1283
                                        E Thn, ThG, Tid, Th4, Tie, ThY, ThR, ThU;
 
1284
                                        {
 
1285
                                             E TgO, Th3, ThW, ThX;
 
1286
                                             TiB = FNMS(KP923879532, TgN, TgG);
 
1287
                                             TgO = FMA(KP923879532, TgN, TgG);
 
1288
                                             Th3 = TgV - Th2;
 
1289
                                             Tip = TgV + Th2;
 
1290
                                             Thn = FNMS(KP098491403, Thm, Thf);
 
1291
                                             ThW = FMA(KP098491403, Thf, Thm);
 
1292
                                             ThX = FMA(KP098491403, Thy, ThF);
 
1293
                                             ThG = FNMS(KP098491403, ThF, Thy);
 
1294
                                             Tid = FNMS(KP980785280, Th3, TgO);
 
1295
                                             Th4 = FMA(KP980785280, Th3, TgO);
 
1296
                                             Tie = ThW - ThX;
 
1297
                                             ThY = ThW + ThX;
 
1298
                                             Tio = FNMS(KP923879532, ThQ, ThN);
 
1299
                                             ThR = FMA(KP923879532, ThQ, ThN);
 
1300
                                             ThU = ThS + ThT;
 
1301
                                             TiC = ThS - ThT;
 
1302
                                        }
 
1303
                                        {
 
1304
                                             E Ti9, ThH, Ti8, ThV;
 
1305
                                             Tif = FMA(KP995184726, Tie, Tid);
 
1306
                                             Tii = FNMS(KP995184726, Tie, Tid);
 
1307
                                             Ti9 = Thn + ThG;
 
1308
                                             ThH = Thn - ThG;
 
1309
                                             Ti8 = FNMS(KP980785280, ThU, ThR);
 
1310
                                             ThV = FMA(KP980785280, ThU, ThR);
 
1311
                                             Ti5 = FNMS(KP995184726, ThH, Th4);
 
1312
                                             ThI = FMA(KP995184726, ThH, Th4);
 
1313
                                             Tia = FNMS(KP995184726, Ti9, Ti8);
 
1314
                                             Til = FMA(KP995184726, Ti9, Ti8);
 
1315
                                             Ti2 = FNMS(KP995184726, ThY, ThV);
 
1316
                                             ThZ = FMA(KP995184726, ThY, ThV);
 
1317
                                        }
 
1318
                                   }
 
1319
                                   {
 
1320
                                        E Tik, Tij, TgD, ThK, Ti1, Ti4;
 
1321
                                        TgD = W[124];
 
1322
                                        ThK = W[125];
 
1323
                                        {
 
1324
                                             E Tih, Ti0, ThJ, Tim;
 
1325
                                             Tih = W[92];
 
1326
                                             Tik = W[93];
 
1327
                                             Ti0 = TgD * ThZ;
 
1328
                                             ThJ = TgD * ThI;
 
1329
                                             Tim = Tih * Til;
 
1330
                                             Tij = Tih * Tii;
 
1331
                                             rio[WS(ios, 63)] = FNMS(ThK, ThI, Ti0);
 
1332
                                             iio[0] = FMA(ThK, ThZ, ThJ);
 
1333
                                             rio[WS(ios, 47)] = FNMS(Tik, Tii, Tim);
 
1334
                                        }
 
1335
                                        iio[-WS(ios, 16)] = FMA(Tik, Til, Tij);
 
1336
                                        Ti1 = W[60];
 
1337
                                        Ti4 = W[61];
 
1338
                                        {
 
1339
                                             E Ti7, Ti6, Ti3, Tig;
 
1340
                                             Ti7 = W[28];
 
1341
                                             Tic = W[29];
 
1342
                                             Ti6 = Ti1 * Ti5;
 
1343
                                             Ti3 = Ti1 * Ti2;
 
1344
                                             Tig = Ti7 * Tif;
 
1345
                                             Tib = Ti7 * Tia;
 
1346
                                             iio[-WS(ios, 32)] = FMA(Ti4, Ti2, Ti6);
 
1347
                                             rio[WS(ios, 31)] = FNMS(Ti4, Ti5, Ti3);
 
1348
                                             iio[-WS(ios, 48)] = FMA(Tic, Tia, Tig);
 
1349
                                        }
 
1350
                                   }
 
1351
                              }
 
1352
                              rio[WS(ios, 15)] = FNMS(Tic, Tif, Tib);
 
1353
                              {
 
1354
                                   E Tk2, TjP, Tjw, TjU, Tk5, TjM, TjJ;
 
1355
                                   {
 
1356
                                        E Tjn, Tju, TjX, Tjg, TjY, TjI, TjG, TjH;
 
1357
                                        Tjn = FNMS(KP303346683, Tjm, Tjj);
 
1358
                                        TjG = FMA(KP303346683, Tjj, Tjm);
 
1359
                                        TjH = FMA(KP303346683, Tjq, Tjt);
 
1360
                                        Tju = FNMS(KP303346683, Tjt, Tjq);
 
1361
                                        TjX = FNMS(KP831469612, Tjf, Tj8);
 
1362
                                        Tjg = FMA(KP831469612, Tjf, Tj8);
 
1363
                                        TjY = TjG + TjH;
 
1364
                                        TjI = TjG - TjH;
 
1365
                                        {
 
1366
                                             E TjT, Tjv, TjS, TjF;
 
1367
                                             TjZ = FMA(KP956940335, TjY, TjX);
 
1368
                                             Tk2 = FNMS(KP956940335, TjY, TjX);
 
1369
                                             TjT = Tju - Tjn;
 
1370
                                             Tjv = Tjn + Tju;
 
1371
                                             TjS = FNMS(KP831469612, TjE, TjB);
 
1372
                                             TjF = FMA(KP831469612, TjE, TjB);
 
1373
                                             TjP = FNMS(KP956940335, Tjv, Tjg);
 
1374
                                             Tjw = FMA(KP956940335, Tjv, Tjg);
 
1375
                                             TjU = FMA(KP956940335, TjT, TjS);
 
1376
                                             Tk5 = FNMS(KP956940335, TjT, TjS);
 
1377
                                             TjM = FNMS(KP956940335, TjI, TjF);
 
1378
                                             TjJ = FMA(KP956940335, TjI, TjF);
 
1379
                                        }
 
1380
                                   }
 
1381
                                   {
 
1382
                                        E Tk4, Tk3, Tj5, Tjy, TjL, TjO;
 
1383
                                        Tj5 = W[4];
 
1384
                                        Tjy = W[5];
 
1385
                                        {
 
1386
                                             E Tk1, TjK, Tjx, Tk6;
 
1387
                                             Tk1 = W[36];
 
1388
                                             Tk4 = W[37];
 
1389
                                             TjK = Tj5 * TjJ;
 
1390
                                             Tjx = Tj5 * Tjw;
 
1391
                                             Tk6 = Tk1 * Tk5;
 
1392
                                             Tk3 = Tk1 * Tk2;
 
1393
                                             iio[-WS(ios, 60)] = FMA(Tjy, Tjw, TjK);
 
1394
                                             rio[WS(ios, 3)] = FNMS(Tjy, TjJ, Tjx);
 
1395
                                             iio[-WS(ios, 44)] = FMA(Tk4, Tk2, Tk6);
 
1396
                                        }
 
1397
                                        rio[WS(ios, 19)] = FNMS(Tk4, Tk5, Tk3);
 
1398
                                        TjL = W[68];
 
1399
                                        TjO = W[69];
 
1400
                                        {
 
1401
                                             E TjR, TjQ, TjN, Tk0;
 
1402
                                             TjR = W[100];
 
1403
                                             TjW = W[101];
 
1404
                                             TjQ = TjL * TjP;
 
1405
                                             TjN = TjL * TjM;
 
1406
                                             Tk0 = TjR * TjZ;
 
1407
                                             TjV = TjR * TjU;
 
1408
                                             rio[WS(ios, 35)] = FNMS(TjO, TjM, TjQ);
 
1409
                                             iio[-WS(ios, 28)] = FMA(TjO, TjP, TjN);
 
1410
                                             rio[WS(ios, 51)] = FNMS(TjW, TjU, Tk0);
 
1411
                                        }
 
1412
                                   }
 
1413
                              }
 
1414
                              iio[-WS(ios, 12)] = FMA(TjW, TjZ, TjV);
 
1415
                              {
 
1416
                                   E Tj0, TiN, Tiy, TiS, Tj3, TiK, TiH;
 
1417
                                   {
 
1418
                                        E Tit, Tiw, TiV, Tiq, TiW, TiG, TiE, TiF;
 
1419
                                        Tit = FNMS(KP820678790, Tis, Tir);
 
1420
                                        TiE = FMA(KP820678790, Tir, Tis);
 
1421
                                        TiF = FMA(KP820678790, Tiu, Tiv);
 
1422
                                        Tiw = FNMS(KP820678790, Tiv, Tiu);
 
1423
                                        TiV = FMA(KP980785280, Tip, Tio);
 
1424
                                        Tiq = FNMS(KP980785280, Tip, Tio);
 
1425
                                        TiW = TiE + TiF;
 
1426
                                        TiG = TiE - TiF;
 
1427
                                        {
 
1428
                                             E TiR, Tix, TiQ, TiD;
 
1429
                                             TiX = FMA(KP773010453, TiW, TiV);
 
1430
                                             Tj0 = FNMS(KP773010453, TiW, TiV);
 
1431
                                             TiR = Tiw - Tit;
 
1432
                                             Tix = Tit + Tiw;
 
1433
                                             TiQ = FNMS(KP980785280, TiC, TiB);
 
1434
                                             TiD = FMA(KP980785280, TiC, TiB);
 
1435
                                             TiN = FNMS(KP773010453, Tix, Tiq);
 
1436
                                             Tiy = FMA(KP773010453, Tix, Tiq);
 
1437
                                             TiS = FMA(KP773010453, TiR, TiQ);
 
1438
                                             Tj3 = FNMS(KP773010453, TiR, TiQ);
 
1439
                                             TiK = FNMS(KP773010453, TiG, TiD);
 
1440
                                             TiH = FMA(KP773010453, TiG, TiD);
 
1441
                                        }
 
1442
                                   }
 
1443
                                   {
 
1444
                                        E Tj2, Tj1, Tin, TiA, TiJ, TiM;
 
1445
                                        Tin = W[12];
 
1446
                                        TiA = W[13];
 
1447
                                        {
 
1448
                                             E TiZ, TiI, Tiz, Tj4;
 
1449
                                             TiZ = W[44];
 
1450
                                             Tj2 = W[45];
 
1451
                                             TiI = Tin * TiH;
 
1452
                                             Tiz = Tin * Tiy;
 
1453
                                             Tj4 = TiZ * Tj3;
 
1454
                                             Tj1 = TiZ * Tj0;
 
1455
                                             iio[-WS(ios, 56)] = FMA(TiA, Tiy, TiI);
 
1456
                                             rio[WS(ios, 7)] = FNMS(TiA, TiH, Tiz);
 
1457
                                             iio[-WS(ios, 40)] = FMA(Tj2, Tj0, Tj4);
 
1458
                                        }
 
1459
                                        rio[WS(ios, 23)] = FNMS(Tj2, Tj3, Tj1);
 
1460
                                        TiJ = W[76];
 
1461
                                        TiM = W[77];
 
1462
                                        {
 
1463
                                             E TiP, TiO, TiL, TiY;
 
1464
                                             TiP = W[108];
 
1465
                                             TiU = W[109];
 
1466
                                             TiO = TiJ * TiN;
 
1467
                                             TiL = TiJ * TiK;
 
1468
                                             TiY = TiP * TiX;
 
1469
                                             TiT = TiP * TiS;
 
1470
                                             rio[WS(ios, 39)] = FNMS(TiM, TiK, TiO);
 
1471
                                             iio[-WS(ios, 24)] = FMA(TiM, TiN, TiL);
 
1472
                                             rio[WS(ios, 55)] = FNMS(TiU, TiS, TiY);
 
1473
                                        }
 
1474
                                   }
 
1475
                              }
 
1476
                         }
 
1477
                    }
 
1478
                    iio[-WS(ios, 8)] = FMA(TiU, TiX, TiT);
 
1479
                    {
 
1480
                         E T5z, T5n, T5m, T5A, Tep, Ted, Tec, Teq, Te3, Te0, TdZ;
 
1481
                         {
 
1482
                              E T5d, T5a, T55, T5e, T59;
 
1483
                              {
 
1484
                                   E T5b, T4T, T4z, T5c, T57, T4W, T3a, T56, T4P, T4S;
 
1485
                                   T5z = FNMS(KP707106781, T4O, T4L);
 
1486
                                   T4P = FMA(KP707106781, T4O, T4L);
 
1487
                                   T4S = T4Q + T4R;
 
1488
                                   T5n = T4R - T4Q;
 
1489
                                   {
 
1490
                                        E T4U, T4V, T2w, T39, T3R, T4y;
 
1491
                                        T4U = FNMS(KP198912367, T3D, T3Q);
 
1492
                                        T3R = FMA(KP198912367, T3Q, T3D);
 
1493
                                        T4y = FNMS(KP198912367, T4x, T4k);
 
1494
                                        T4V = FMA(KP198912367, T4k, T4x);
 
1495
                                        T5b = FNMS(KP923879532, T4S, T4P);
 
1496
                                        T4T = FMA(KP923879532, T4S, T4P);
 
1497
                                        T4z = T3R + T4y;
 
1498
                                        T5c = T3R - T4y;
 
1499
                                        T2w = FMA(KP707106781, T2v, T2c);
 
1500
                                        T5m = FNMS(KP707106781, T2v, T2c);
 
1501
                                        T5A = T2P - T38;
 
1502
                                        T39 = T2P + T38;
 
1503
                                        T57 = T4V - T4U;
 
1504
                                        T4W = T4U + T4V;
 
1505
                                        T3a = FMA(KP923879532, T39, T2w);
 
1506
                                        T56 = FNMS(KP923879532, T39, T2w);
 
1507
                                   }
 
1508
                                   {
 
1509
                                        E T53, T58, T50, T5j, T5i, T5h;
 
1510
                                        {
 
1511
                                             E T4X, T4C, T5g, T4A, T23;
 
1512
                                             T53 = FNMS(KP980785280, T4W, T4T);
 
1513
                                             T4X = FMA(KP980785280, T4W, T4T);
 
1514
                                             T4C = W[123];
 
1515
                                             T5g = FNMS(KP980785280, T57, T56);
 
1516
                                             T58 = FMA(KP980785280, T57, T56);
 
1517
                                             T4A = FMA(KP980785280, T4z, T3a);
 
1518
                                             T50 = FNMS(KP980785280, T4z, T3a);
 
1519
                                             T23 = W[122];
 
1520
                                             {
 
1521
                                                  E T5f, T5k, T4Y, T4B;
 
1522
                                                  T5j = FNMS(KP980785280, T5c, T5b);
 
1523
                                                  T5d = FMA(KP980785280, T5c, T5b);
 
1524
                                                  T5i = W[91];
 
1525
                                                  T4Y = T4C * T4A;
 
1526
                                                  T4B = T23 * T4A;
 
1527
                                                  T5f = W[90];
 
1528
                                                  T5k = T5i * T5g;
 
1529
                                                  iio[-WS(ios, 1)] = FMA(T23, T4X, T4Y);
 
1530
                                                  rio[WS(ios, 62)] = FNMS(T4C, T4X, T4B);
 
1531
                                                  T5h = T5f * T5g;
 
1532
                                                  iio[-WS(ios, 17)] = FMA(T5f, T5j, T5k);
 
1533
                                             }
 
1534
                                        }
 
1535
                                        {
 
1536
                                             E T52, T4Z, T54, T51;
 
1537
                                             T52 = W[59];
 
1538
                                             rio[WS(ios, 46)] = FNMS(T5i, T5j, T5h);
 
1539
                                             T4Z = W[58];
 
1540
                                             T54 = T52 * T50;
 
1541
                                             T5a = W[27];
 
1542
                                             T51 = T4Z * T50;
 
1543
                                             T55 = W[26];
 
1544
                                             iio[-WS(ios, 33)] = FMA(T4Z, T53, T54);
 
1545
                                             T5e = T5a * T58;
 
1546
                                             rio[WS(ios, 30)] = FNMS(T52, T53, T51);
 
1547
                                             T59 = T55 * T58;
 
1548
                                        }
 
1549
                                   }
 
1550
                              }
 
1551
                              {
 
1552
                                   E Te6, TdT, Tds, TdY, Te9, TdQ, TdN;
 
1553
                                   {
 
1554
                                        E TcN, Tdq, Tca, Te1, Te2, TdM, TdF, TdI;
 
1555
                                        {
 
1556
                                             E TdK, TdL, Tbw, Tc9;
 
1557
                                             Tep = FNMS(KP923879532, Tbv, Tbg);
 
1558
                                             Tbw = FMA(KP923879532, Tbv, Tbg);
 
1559
                                             iio[-WS(ios, 49)] = FMA(T55, T5d, T5e);
 
1560
                                             Ted = Tc8 - TbP;
 
1561
                                             Tc9 = TbP + Tc8;
 
1562
                                             rio[WS(ios, 14)] = FNMS(T5a, T5d, T59);
 
1563
                                             TcN = FNMS(KP098491403, TcM, TcB);
 
1564
                                             TdK = FMA(KP098491403, TcB, TcM);
 
1565
                                             TdL = FNMS(KP098491403, Tde, Tdp);
 
1566
                                             Tdq = FMA(KP098491403, Tdp, Tde);
 
1567
                                             Tca = FMA(KP980785280, Tc9, Tbw);
 
1568
                                             Te1 = FNMS(KP980785280, Tc9, Tbw);
 
1569
                                             Te2 = TdK - TdL;
 
1570
                                             TdM = TdK + TdL;
 
1571
                                             Tec = FNMS(KP923879532, TdE, TdB);
 
1572
                                             TdF = FMA(KP923879532, TdE, TdB);
 
1573
                                             TdI = TdG + TdH;
 
1574
                                             Teq = TdG - TdH;
 
1575
                                        }
 
1576
                                        {
 
1577
                                             E TdX, Tdr, TdW, TdJ;
 
1578
                                             Te3 = FMA(KP995184726, Te2, Te1);
 
1579
                                             Te6 = FNMS(KP995184726, Te2, Te1);
 
1580
                                             TdX = Tdq - TcN;
 
1581
                                             Tdr = TcN + Tdq;
 
1582
                                             TdW = FNMS(KP980785280, TdI, TdF);
 
1583
                                             TdJ = FMA(KP980785280, TdI, TdF);
 
1584
                                             TdT = FNMS(KP995184726, Tdr, Tca);
 
1585
                                             Tds = FMA(KP995184726, Tdr, Tca);
 
1586
                                             TdY = FMA(KP995184726, TdX, TdW);
 
1587
                                             Te9 = FNMS(KP995184726, TdX, TdW);
 
1588
                                             TdQ = FNMS(KP995184726, TdM, TdJ);
 
1589
                                             TdN = FMA(KP995184726, TdM, TdJ);
 
1590
                                        }
 
1591
                                   }
 
1592
                                   {
 
1593
                                        E Te8, Te7, Tb5, Tdu, TdP, TdS;
 
1594
                                        Tb5 = W[0];
 
1595
                                        Tdu = W[1];
 
1596
                                        {
 
1597
                                             E Te5, TdO, Tdt, Tea;
 
1598
                                             Te5 = W[32];
 
1599
                                             Te8 = W[33];
 
1600
                                             TdO = Tb5 * TdN;
 
1601
                                             Tdt = Tb5 * Tds;
 
1602
                                             Tea = Te5 * Te9;
 
1603
                                             Te7 = Te5 * Te6;
 
1604
                                             iio[-WS(ios, 62)] = FMA(Tdu, Tds, TdO);
 
1605
                                             rio[WS(ios, 1)] = FNMS(Tdu, TdN, Tdt);
 
1606
                                             iio[-WS(ios, 46)] = FMA(Te8, Te6, Tea);
 
1607
                                        }
 
1608
                                        rio[WS(ios, 17)] = FNMS(Te8, Te9, Te7);
 
1609
                                        TdP = W[64];
 
1610
                                        TdS = W[65];
 
1611
                                        {
 
1612
                                             E TdV, TdU, TdR, Te4;
 
1613
                                             TdV = W[96];
 
1614
                                             Te0 = W[97];
 
1615
                                             TdU = TdP * TdT;
 
1616
                                             TdR = TdP * TdQ;
 
1617
                                             Te4 = TdV * Te3;
 
1618
                                             TdZ = TdV * TdY;
 
1619
                                             rio[WS(ios, 33)] = FNMS(TdS, TdQ, TdU);
 
1620
                                             iio[-WS(ios, 30)] = FMA(TdS, TdT, TdR);
 
1621
                                             rio[WS(ios, 49)] = FNMS(Te0, TdY, Te4);
 
1622
                                        }
 
1623
                                   }
 
1624
                              }
 
1625
                         }
 
1626
                         iio[-WS(ios, 14)] = FMA(Te0, Te3, TdZ);
 
1627
                         {
 
1628
                              E T5V, T5S, T5N, T5W, T5R;
 
1629
                              {
 
1630
                                   E T5T, T5B, T5v, T5U, T5P, T5E, T5o, T5O;
 
1631
                                   {
 
1632
                                        E T5C, T5D, T5r, T5u;
 
1633
                                        T5C = FMA(KP668178637, T5p, T5q);
 
1634
                                        T5r = FNMS(KP668178637, T5q, T5p);
 
1635
                                        T5u = FMA(KP668178637, T5t, T5s);
 
1636
                                        T5D = FNMS(KP668178637, T5s, T5t);
 
1637
                                        T5T = FNMS(KP923879532, T5A, T5z);
 
1638
                                        T5B = FMA(KP923879532, T5A, T5z);
 
1639
                                        T5v = T5r + T5u;
 
1640
                                        T5U = T5u - T5r;
 
1641
                                        T5P = T5C - T5D;
 
1642
                                        T5E = T5C + T5D;
 
1643
                                        T5o = FMA(KP923879532, T5n, T5m);
 
1644
                                        T5O = FNMS(KP923879532, T5n, T5m);
 
1645
                                   }
 
1646
                                   {
 
1647
                                        E T5L, T5Q, T5I, T61, T60, T5Z;
 
1648
                                        {
 
1649
                                             E T5F, T5y, T5Y, T5w, T5l;
 
1650
                                             T5L = FNMS(KP831469612, T5E, T5B);
 
1651
                                             T5F = FMA(KP831469612, T5E, T5B);
 
1652
                                             T5y = W[11];
 
1653
                                             T5Y = FNMS(KP831469612, T5P, T5O);
 
1654
                                             T5Q = FMA(KP831469612, T5P, T5O);
 
1655
                                             T5w = FMA(KP831469612, T5v, T5o);
 
1656
                                             T5I = FNMS(KP831469612, T5v, T5o);
 
1657
                                             T5l = W[10];
 
1658
                                             {
 
1659
                                                  E T5X, T62, T5G, T5x;
 
1660
                                                  T61 = FNMS(KP831469612, T5U, T5T);
 
1661
                                                  T5V = FMA(KP831469612, T5U, T5T);
 
1662
                                                  T60 = W[43];
 
1663
                                                  T5G = T5y * T5w;
 
1664
                                                  T5x = T5l * T5w;
 
1665
                                                  T5X = W[42];
 
1666
                                                  T62 = T60 * T5Y;
 
1667
                                                  iio[-WS(ios, 57)] = FMA(T5l, T5F, T5G);
 
1668
                                                  rio[WS(ios, 6)] = FNMS(T5y, T5F, T5x);
 
1669
                                                  T5Z = T5X * T5Y;
 
1670
                                                  iio[-WS(ios, 41)] = FMA(T5X, T61, T62);
 
1671
                                             }
 
1672
                                        }
 
1673
                                        {
 
1674
                                             E T5K, T5H, T5M, T5J;
 
1675
                                             T5K = W[75];
 
1676
                                             rio[WS(ios, 22)] = FNMS(T60, T61, T5Z);
 
1677
                                             T5H = W[74];
 
1678
                                             T5M = T5K * T5I;
 
1679
                                             T5S = W[107];
 
1680
                                             T5J = T5H * T5I;
 
1681
                                             T5N = W[106];
 
1682
                                             iio[-WS(ios, 25)] = FMA(T5H, T5L, T5M);
 
1683
                                             T5W = T5S * T5Q;
 
1684
                                             rio[WS(ios, 38)] = FNMS(T5K, T5L, T5J);
 
1685
                                             T5R = T5N * T5Q;
 
1686
                                        }
 
1687
                                   }
 
1688
                              }
 
1689
                              {
 
1690
                                   E TeO, TeB, Tem, TeG, TeR, Tey, Tev;
 
1691
                                   {
 
1692
                                        E Teh, Tek, Tee, TeJ, TeK, Teu, Tes, Tet;
 
1693
                                        iio[-WS(ios, 9)] = FMA(T5N, T5V, T5W);
 
1694
                                        rio[WS(ios, 54)] = FNMS(T5S, T5V, T5R);
 
1695
                                        Teh = FNMS(KP820678790, Teg, Tef);
 
1696
                                        Tes = FMA(KP820678790, Tef, Teg);
 
1697
                                        Tet = FNMS(KP820678790, Tei, Tej);
 
1698
                                        Tek = FMA(KP820678790, Tej, Tei);
 
1699
                                        Tee = FMA(KP980785280, Ted, Tec);
 
1700
                                        TeJ = FNMS(KP980785280, Ted, Tec);
 
1701
                                        TeK = Tes - Tet;
 
1702
                                        Teu = Tes + Tet;
 
1703
                                        {
 
1704
                                             E TeF, Tel, TeE, Ter;
 
1705
                                             TeL = FMA(KP773010453, TeK, TeJ);
 
1706
                                             TeO = FNMS(KP773010453, TeK, TeJ);
 
1707
                                             TeF = Tek - Teh;
 
1708
                                             Tel = Teh + Tek;
 
1709
                                             TeE = FNMS(KP980785280, Teq, Tep);
 
1710
                                             Ter = FMA(KP980785280, Teq, Tep);
 
1711
                                             TeB = FNMS(KP773010453, Tel, Tee);
 
1712
                                             Tem = FMA(KP773010453, Tel, Tee);
 
1713
                                             TeG = FMA(KP773010453, TeF, TeE);
 
1714
                                             TeR = FNMS(KP773010453, TeF, TeE);
 
1715
                                             Tey = FNMS(KP773010453, Teu, Ter);
 
1716
                                             Tev = FMA(KP773010453, Teu, Ter);
 
1717
                                        }
 
1718
                                   }
 
1719
                                   {
 
1720
                                        E TeQ, TeP, Teb, Teo, Tex, TeA;
 
1721
                                        Teb = W[112];
 
1722
                                        Teo = W[113];
 
1723
                                        {
 
1724
                                             E TeN, Tew, Ten, TeS;
 
1725
                                             TeN = W[80];
 
1726
                                             TeQ = W[81];
 
1727
                                             Tew = Teb * Tev;
 
1728
                                             Ten = Teb * Tem;
 
1729
                                             TeS = TeN * TeR;
 
1730
                                             TeP = TeN * TeO;
 
1731
                                             rio[WS(ios, 57)] = FNMS(Teo, Tem, Tew);
 
1732
                                             iio[-WS(ios, 6)] = FMA(Teo, Tev, Ten);
 
1733
                                             rio[WS(ios, 41)] = FNMS(TeQ, TeO, TeS);
 
1734
                                        }
 
1735
                                        iio[-WS(ios, 22)] = FMA(TeQ, TeR, TeP);
 
1736
                                        Tex = W[48];
 
1737
                                        TeA = W[49];
 
1738
                                        {
 
1739
                                             E TeD, TeC, Tez, TeM;
 
1740
                                             TeD = W[16];
 
1741
                                             TeI = W[17];
 
1742
                                             TeC = Tex * TeB;
 
1743
                                             Tez = Tex * Tey;
 
1744
                                             TeM = TeD * TeL;
 
1745
                                             TeH = TeD * TeG;
 
1746
                                             iio[-WS(ios, 38)] = FMA(TeA, Tey, TeC);
 
1747
                                             rio[WS(ios, 25)] = FNMS(TeA, TeB, Tez);
 
1748
                                             iio[-WS(ios, 54)] = FMA(TeI, TeG, TeM);
 
1749
                                        }
 
1750
                                   }
 
1751
                              }
 
1752
                         }
 
1753
                    }
 
1754
                    rio[WS(ios, 9)] = FNMS(TeI, TeL, TeH);
 
1755
                    {
 
1756
                         E T7j, T77, T76, T7k, Tg9, TfX, TfW, Tga, TfN, TfK, TfJ;
 
1757
                         {
 
1758
                              E T6X, T6U, T6P, T6Y, T6T;
 
1759
                              {
 
1760
                                   E T6V, T6D, T6t, T6W, T6R, T6G, T6e, T6Q, T6z, T6C;
 
1761
                                   T7j = FNMS(KP707106781, T6y, T6x);
 
1762
                                   T6z = FMA(KP707106781, T6y, T6x);
 
1763
                                   T6C = T6A + T6B;
 
1764
                                   T77 = T6A - T6B;
 
1765
                                   {
 
1766
                                        E T6E, T6F, T66, T6d, T6l, T6s;
 
1767
                                        T6E = FMA(KP198912367, T6h, T6k);
 
1768
                                        T6l = FNMS(KP198912367, T6k, T6h);
 
1769
                                        T6s = FMA(KP198912367, T6r, T6o);
 
1770
                                        T6F = FNMS(KP198912367, T6o, T6r);
 
1771
                                        T6V = FNMS(KP923879532, T6C, T6z);
 
1772
                                        T6D = FMA(KP923879532, T6C, T6z);
 
1773
                                        T6t = T6l + T6s;
 
1774
                                        T6W = T6s - T6l;
 
1775
                                        T66 = FMA(KP707106781, T65, T64);
 
1776
                                        T76 = FNMS(KP707106781, T65, T64);
 
1777
                                        T7k = T6c - T69;
 
1778
                                        T6d = T69 + T6c;
 
1779
                                        T6R = T6E - T6F;
 
1780
                                        T6G = T6E + T6F;
 
1781
                                        T6e = FMA(KP923879532, T6d, T66);
 
1782
                                        T6Q = FNMS(KP923879532, T6d, T66);
 
1783
                                   }
 
1784
                                   {
 
1785
                                        E T6N, T6S, T6K, T73, T72, T71;
 
1786
                                        {
 
1787
                                             E T6H, T6w, T70, T6u, T63;
 
1788
                                             T6N = FNMS(KP980785280, T6G, T6D);
 
1789
                                             T6H = FMA(KP980785280, T6G, T6D);
 
1790
                                             T6w = W[3];
 
1791
                                             T70 = FNMS(KP980785280, T6R, T6Q);
 
1792
                                             T6S = FMA(KP980785280, T6R, T6Q);
 
1793
                                             T6u = FMA(KP980785280, T6t, T6e);
 
1794
                                             T6K = FNMS(KP980785280, T6t, T6e);
 
1795
                                             T63 = W[2];
 
1796
                                             {
 
1797
                                                  E T6Z, T74, T6I, T6v;
 
1798
                                                  T73 = FNMS(KP980785280, T6W, T6V);
 
1799
                                                  T6X = FMA(KP980785280, T6W, T6V);
 
1800
                                                  T72 = W[35];
 
1801
                                                  T6I = T6w * T6u;
 
1802
                                                  T6v = T63 * T6u;
 
1803
                                                  T6Z = W[34];
 
1804
                                                  T74 = T72 * T70;
 
1805
                                                  iio[-WS(ios, 61)] = FMA(T63, T6H, T6I);
 
1806
                                                  rio[WS(ios, 2)] = FNMS(T6w, T6H, T6v);
 
1807
                                                  T71 = T6Z * T70;
 
1808
                                                  iio[-WS(ios, 45)] = FMA(T6Z, T73, T74);
 
1809
                                             }
 
1810
                                        }
 
1811
                                        {
 
1812
                                             E T6M, T6J, T6O, T6L;
 
1813
                                             T6M = W[67];
 
1814
                                             rio[WS(ios, 18)] = FNMS(T72, T73, T71);
 
1815
                                             T6J = W[66];
 
1816
                                             T6O = T6M * T6K;
 
1817
                                             T6U = W[99];
 
1818
                                             T6L = T6J * T6K;
 
1819
                                             T6P = W[98];
 
1820
                                             iio[-WS(ios, 29)] = FMA(T6J, T6N, T6O);
 
1821
                                             T6Y = T6U * T6S;
 
1822
                                             rio[WS(ios, 34)] = FNMS(T6M, T6N, T6L);
 
1823
                                             T6T = T6P * T6S;
 
1824
                                        }
 
1825
                                   }
 
1826
                              }
 
1827
                              {
 
1828
                                   E TfQ, TfD, Tfk, TfI, TfT, TfA, Tfx;
 
1829
                                   {
 
1830
                                        E Tfb, Tfi, Tf4, TfL, TfM, Tfw, Tfp, Tfs;
 
1831
                                        {
 
1832
                                             E Tfu, Tfv, TeW, Tf3;
 
1833
                                             Tg9 = FNMS(KP923879532, TeV, TeU);
 
1834
                                             TeW = FMA(KP923879532, TeV, TeU);
 
1835
                                             iio[-WS(ios, 13)] = FMA(T6P, T6X, T6Y);
 
1836
                                             TfX = Tf2 - TeZ;
 
1837
                                             Tf3 = TeZ + Tf2;
 
1838
                                             rio[WS(ios, 50)] = FNMS(T6U, T6X, T6T);
 
1839
                                             Tfb = FNMS(KP303346683, Tfa, Tf7);
 
1840
                                             Tfu = FMA(KP303346683, Tf7, Tfa);
 
1841
                                             Tfv = FNMS(KP303346683, Tfe, Tfh);
 
1842
                                             Tfi = FMA(KP303346683, Tfh, Tfe);
 
1843
                                             Tf4 = FMA(KP831469612, Tf3, TeW);
 
1844
                                             TfL = FNMS(KP831469612, Tf3, TeW);
 
1845
                                             TfM = Tfu - Tfv;
 
1846
                                             Tfw = Tfu + Tfv;
 
1847
                                             TfW = FNMS(KP923879532, Tfo, Tfn);
 
1848
                                             Tfp = FMA(KP923879532, Tfo, Tfn);
 
1849
                                             Tfs = Tfq + Tfr;
 
1850
                                             Tga = Tfq - Tfr;
 
1851
                                        }
 
1852
                                        {
 
1853
                                             E TfH, Tfj, TfG, Tft;
 
1854
                                             TfN = FMA(KP956940335, TfM, TfL);
 
1855
                                             TfQ = FNMS(KP956940335, TfM, TfL);
 
1856
                                             TfH = Tfi - Tfb;
 
1857
                                             Tfj = Tfb + Tfi;
 
1858
                                             TfG = FNMS(KP831469612, Tfs, Tfp);
 
1859
                                             Tft = FMA(KP831469612, Tfs, Tfp);
 
1860
                                             TfD = FNMS(KP956940335, Tfj, Tf4);
 
1861
                                             Tfk = FMA(KP956940335, Tfj, Tf4);
 
1862
                                             TfI = FMA(KP956940335, TfH, TfG);
 
1863
                                             TfT = FNMS(KP956940335, TfH, TfG);
 
1864
                                             TfA = FNMS(KP956940335, Tfw, Tft);
 
1865
                                             Tfx = FMA(KP956940335, Tfw, Tft);
 
1866
                                        }
 
1867
                                   }
 
1868
                                   {
 
1869
                                        E TfS, TfR, TeT, Tfm, Tfz, TfC;
 
1870
                                        TeT = W[120];
 
1871
                                        Tfm = W[121];
 
1872
                                        {
 
1873
                                             E TfP, Tfy, Tfl, TfU;
 
1874
                                             TfP = W[88];
 
1875
                                             TfS = W[89];
 
1876
                                             Tfy = TeT * Tfx;
 
1877
                                             Tfl = TeT * Tfk;
 
1878
                                             TfU = TfP * TfT;
 
1879
                                             TfR = TfP * TfQ;
 
1880
                                             rio[WS(ios, 61)] = FNMS(Tfm, Tfk, Tfy);
 
1881
                                             iio[-WS(ios, 2)] = FMA(Tfm, Tfx, Tfl);
 
1882
                                             rio[WS(ios, 45)] = FNMS(TfS, TfQ, TfU);
 
1883
                                        }
 
1884
                                        iio[-WS(ios, 18)] = FMA(TfS, TfT, TfR);
 
1885
                                        Tfz = W[56];
 
1886
                                        TfC = W[57];
 
1887
                                        {
 
1888
                                             E TfF, TfE, TfB, TfO;
 
1889
                                             TfF = W[24];
 
1890
                                             TfK = W[25];
 
1891
                                             TfE = Tfz * TfD;
 
1892
                                             TfB = Tfz * TfA;
 
1893
                                             TfO = TfF * TfN;
 
1894
                                             TfJ = TfF * TfI;
 
1895
                                             iio[-WS(ios, 34)] = FMA(TfC, TfA, TfE);
 
1896
                                             rio[WS(ios, 29)] = FNMS(TfC, TfD, TfB);
 
1897
                                             iio[-WS(ios, 50)] = FMA(TfK, TfI, TfO);
 
1898
                                        }
 
1899
                                   }
 
1900
                              }
 
1901
                         }
 
1902
                         rio[WS(ios, 13)] = FNMS(TfK, TfN, TfJ);
 
1903
                         {
 
1904
                              E T7F, T7C, T7x, T7G, T7B;
 
1905
                              {
 
1906
                                   E T7D, T7l, T7f, T7E, T7z, T7o, T78, T7y;
 
1907
                                   {
 
1908
                                        E T7m, T7n, T7b, T7e;
 
1909
                                        T7m = FNMS(KP668178637, T79, T7a);
 
1910
                                        T7b = FMA(KP668178637, T7a, T79);
 
1911
                                        T7e = FNMS(KP668178637, T7d, T7c);
 
1912
                                        T7n = FMA(KP668178637, T7c, T7d);
 
1913
                                        T7D = FNMS(KP923879532, T7k, T7j);
 
1914
                                        T7l = FMA(KP923879532, T7k, T7j);
 
1915
                                        T7f = T7b + T7e;
 
1916
                                        T7E = T7b - T7e;
 
1917
                                        T7z = T7n - T7m;
 
1918
                                        T7o = T7m + T7n;
 
1919
                                        T78 = FMA(KP923879532, T77, T76);
 
1920
                                        T7y = FNMS(KP923879532, T77, T76);
 
1921
                                   }
 
1922
                                   {
 
1923
                                        E T7v, T7A, T7s, T7L, T7K, T7J;
 
1924
                                        {
 
1925
                                             E T7p, T7i, T7I, T7g, T75;
 
1926
                                             T7v = FNMS(KP831469612, T7o, T7l);
 
1927
                                             T7p = FMA(KP831469612, T7o, T7l);
 
1928
                                             T7i = W[115];
 
1929
                                             T7I = FNMS(KP831469612, T7z, T7y);
 
1930
                                             T7A = FMA(KP831469612, T7z, T7y);
 
1931
                                             T7g = FMA(KP831469612, T7f, T78);
 
1932
                                             T7s = FNMS(KP831469612, T7f, T78);
 
1933
                                             T75 = W[114];
 
1934
                                             {
 
1935
                                                  E T7H, T7M, T7q, T7h;
 
1936
                                                  T7L = FNMS(KP831469612, T7E, T7D);
 
1937
                                                  T7F = FMA(KP831469612, T7E, T7D);
 
1938
                                                  T7K = W[83];
 
1939
                                                  T7q = T7i * T7g;
 
1940
                                                  T7h = T75 * T7g;
 
1941
                                                  T7H = W[82];
 
1942
                                                  T7M = T7K * T7I;
 
1943
                                                  iio[-WS(ios, 5)] = FMA(T75, T7p, T7q);
 
1944
                                                  rio[WS(ios, 58)] = FNMS(T7i, T7p, T7h);
 
1945
                                                  T7J = T7H * T7I;
 
1946
                                                  iio[-WS(ios, 21)] = FMA(T7H, T7L, T7M);
 
1947
                                             }
 
1948
                                        }
 
1949
                                        {
 
1950
                                             E T7u, T7r, T7w, T7t;
 
1951
                                             T7u = W[51];
 
1952
                                             rio[WS(ios, 42)] = FNMS(T7K, T7L, T7J);
 
1953
                                             T7r = W[50];
 
1954
                                             T7w = T7u * T7s;
 
1955
                                             T7C = W[19];
 
1956
                                             T7t = T7r * T7s;
 
1957
                                             T7x = W[18];
 
1958
                                             iio[-WS(ios, 37)] = FMA(T7r, T7v, T7w);
 
1959
                                             T7G = T7C * T7A;
 
1960
                                             rio[WS(ios, 26)] = FNMS(T7u, T7v, T7t);
 
1961
                                             T7B = T7x * T7A;
 
1962
                                        }
 
1963
                                   }
 
1964
                              }
 
1965
                              {
 
1966
                                   E Tgy, Tgl, Tg6, Tgq, TgB, Tgi, Tgf;
 
1967
                                   {
 
1968
                                        E Tg1, Tg4, TfY, Tgt, Tgu, Tge, Tgc, Tgd;
 
1969
                                        iio[-WS(ios, 53)] = FMA(T7x, T7F, T7G);
 
1970
                                        rio[WS(ios, 10)] = FNMS(T7C, T7F, T7B);
 
1971
                                        Tg1 = FNMS(KP534511135, Tg0, TfZ);
 
1972
                                        Tgc = FMA(KP534511135, TfZ, Tg0);
 
1973
                                        Tgd = FNMS(KP534511135, Tg2, Tg3);
 
1974
                                        Tg4 = FMA(KP534511135, Tg3, Tg2);
 
1975
                                        TfY = FMA(KP831469612, TfX, TfW);
 
1976
                                        Tgt = FNMS(KP831469612, TfX, TfW);
 
1977
                                        Tgu = Tgc - Tgd;
 
1978
                                        Tge = Tgc + Tgd;
 
1979
                                        {
 
1980
                                             E Tgp, Tg5, Tgo, Tgb;
 
1981
                                             Tgv = FMA(KP881921264, Tgu, Tgt);
 
1982
                                             Tgy = FNMS(KP881921264, Tgu, Tgt);
 
1983
                                             Tgp = Tg4 - Tg1;
 
1984
                                             Tg5 = Tg1 + Tg4;
 
1985
                                             Tgo = FNMS(KP831469612, Tga, Tg9);
 
1986
                                             Tgb = FMA(KP831469612, Tga, Tg9);
 
1987
                                             Tgl = FNMS(KP881921264, Tg5, TfY);
 
1988
                                             Tg6 = FMA(KP881921264, Tg5, TfY);
 
1989
                                             Tgq = FMA(KP881921264, Tgp, Tgo);
 
1990
                                             TgB = FNMS(KP881921264, Tgp, Tgo);
 
1991
                                             Tgi = FNMS(KP881921264, Tge, Tgb);
 
1992
                                             Tgf = FMA(KP881921264, Tge, Tgb);
 
1993
                                        }
 
1994
                                   }
 
1995
                                   {
 
1996
                                        E TgA, Tgz, TfV, Tg8, Tgh, Tgk;
 
1997
                                        TfV = W[8];
 
1998
                                        Tg8 = W[9];
 
1999
                                        {
 
2000
                                             E Tgx, Tgg, Tg7, TgC;
 
2001
                                             Tgx = W[40];
 
2002
                                             TgA = W[41];
 
2003
                                             Tgg = TfV * Tgf;
 
2004
                                             Tg7 = TfV * Tg6;
 
2005
                                             TgC = Tgx * TgB;
 
2006
                                             Tgz = Tgx * Tgy;
 
2007
                                             iio[-WS(ios, 58)] = FMA(Tg8, Tg6, Tgg);
 
2008
                                             rio[WS(ios, 5)] = FNMS(Tg8, Tgf, Tg7);
 
2009
                                             iio[-WS(ios, 42)] = FMA(TgA, Tgy, TgC);
 
2010
                                        }
 
2011
                                        rio[WS(ios, 21)] = FNMS(TgA, TgB, Tgz);
 
2012
                                        Tgh = W[72];
 
2013
                                        Tgk = W[73];
 
2014
                                        {
 
2015
                                             E Tgn, Tgm, Tgj, Tgw;
 
2016
                                             Tgn = W[104];
 
2017
                                             Tgs = W[105];
 
2018
                                             Tgm = Tgh * Tgl;
 
2019
                                             Tgj = Tgh * Tgi;
 
2020
                                             Tgw = Tgn * Tgv;
 
2021
                                             Tgr = Tgn * Tgq;
 
2022
                                             rio[WS(ios, 37)] = FNMS(Tgk, Tgi, Tgm);
 
2023
                                             iio[-WS(ios, 26)] = FMA(Tgk, Tgl, Tgj);
 
2024
                                             rio[WS(ios, 53)] = FNMS(Tgs, Tgq, Tgw);
 
2025
                                        }
 
2026
                                   }
 
2027
                              }
 
2028
                         }
 
2029
                    }
 
2030
               }
 
2031
          }
 
2032
          iio[-WS(ios, 10)] = FMA(Tgs, Tgv, Tgr);
 
2033
     }
 
2034
     return W;
 
2035
}
 
2036
 
 
2037
static const tw_instr twinstr[] = {
 
2038
     {TW_FULL, 0, 64},
 
2039
     {TW_NEXT, 1, 0}
 
2040
};
 
2041
 
 
2042
static const hc2hc_desc desc = { 64, "hb_64", twinstr, &GENUS, {520, 126, 518, 0}, 0, 0, 0 };
 
2043
 
 
2044
void X(codelet_hb_64) (planner *p) {
 
2045
     X(khc2hc_register) (p, hb_64, &desc);
 
2046
}
 
2047
#else                           /* HAVE_FMA */
 
2048
 
 
2049
/* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -dif -name hb_64 -include hb.h */
 
2050
 
 
2051
/*
 
2052
 * This function contains 1038 FP additions, 500 FP multiplications,
 
2053
 * (or, 808 additions, 270 multiplications, 230 fused multiply/add),
 
2054
 * 196 stack variables, and 256 memory accesses
 
2055
 */
 
2056
/*
 
2057
 * Generator Id's : 
 
2058
 * $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
 
2059
 * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
 
2060
 * $Id: gen_hc2hc.ml,v 1.16 2006-02-12 23:34:12 athena Exp $
 
2061
 */
 
2062
 
 
2063
#include "hb.h"
 
2064
 
 
2065
static const R *hb_64(R *rio, R *iio, const R *W, stride ios, INT m, INT dist)
 
2066
{
 
2067
     DK(KP634393284, +0.634393284163645498215171613225493370675687095);
 
2068
     DK(KP773010453, +0.773010453362736960810906609758469800971041293);
 
2069
     DK(KP098017140, +0.098017140329560601994195563888641845861136673);
 
2070
     DK(KP995184726, +0.995184726672196886244836953109479921575474869);
 
2071
     DK(KP471396736, +0.471396736825997648556387625905254377657460319);
 
2072
     DK(KP881921264, +0.881921264348355029712756863660388349508442621);
 
2073
     DK(KP290284677, +0.290284677254462367636192375817395274691476278);
 
2074
     DK(KP956940335, +0.956940335732208864935797886980269969482849206);
 
2075
     DK(KP195090322, +0.195090322016128267848284868477022240927691618);
 
2076
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
 
2077
     DK(KP555570233, +0.555570233019602224742830813948532874374937191);
 
2078
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
 
2079
     DK(KP382683432, +0.382683432365089771728459984030398866761344562);
 
2080
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
2081
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
2082
     INT i;
 
2083
     for (i = m - 2; i > 0; i = i - 2, rio = rio + dist, iio = iio - dist, W = W + 126, MAKE_VOLATILE_STRIDE(ios)) {
 
2084
          E Tf, T7i, Tfa, ThM, Tgp, ThH, T2c, T5O, T4T, T6n, Tcp, Ted, TcA, TdE, T87;
 
2085
          E T9o, TK, T93, T2P, T4F, Tfo, Thz, T5T, T6j, Tbx, TdI, Tfl, ThA, T7r, T81;
 
2086
          E TbE, TdH, TZ, T94, T38, T4G, Tfv, ThC, T5W, T6k, TbQ, TdK, Tfs, ThD, T7w;
 
2087
          E T82, TbX, TdL, Tu, T84, Tfh, ThG, Tgm, ThN, T2v, T6m, T4K, T5P, Tce, TdF;
 
2088
          E TcD, Tec, T7l, T9p, T1L, T20, T9c, T9d, T9e, T9f, T40, T66, Tg1, Thu, Tg8;
 
2089
          E Thv, Tg5, Thr, T4n, T67, T4j, T69, T4w, T6a, TaT, TdW, Tb8, TdZ, TfU, Ths;
 
2090
          E T7O, T8y, T7T, T8z, Tbc, TdX, Tbj, Te0, T1g, T1v, T97, T98, T99, T9a, T3j;
 
2091
          E T5Z, TfI, Thk, TfP, Thl, TfM, Tho, T3G, T60, T3C, T62, T3P, T63, Tak, TdQ;
 
2092
          E Tav, TdT, TfB, Thn, T7D, T8v, T7I, T8w, TaD, TdP, TaG, TdS;
 
2093
          {
 
2094
               E T3, Tcm, T4O, Tcv, T6, Tcu, T4R, Tcn, Td, Tcy, T2a, Tch, Ta, Tcx, T27;
 
2095
               E Tck;
 
2096
               {
 
2097
                    E T1, T2, T4P, T4Q;
 
2098
                    T1 = rio[0];
 
2099
                    T2 = iio[-WS(ios, 32)];
 
2100
                    T3 = T1 + T2;
 
2101
                    Tcm = T1 - T2;
 
2102
                    {
 
2103
                         E T4M, T4N, T4, T5;
 
2104
                         T4M = iio[0];
 
2105
                         T4N = rio[WS(ios, 32)];
 
2106
                         T4O = T4M - T4N;
 
2107
                         Tcv = T4M + T4N;
 
2108
                         T4 = rio[WS(ios, 16)];
 
2109
                         T5 = iio[-WS(ios, 48)];
 
2110
                         T6 = T4 + T5;
 
2111
                         Tcu = T4 - T5;
 
2112
                    }
 
2113
                    T4P = iio[-WS(ios, 16)];
 
2114
                    T4Q = rio[WS(ios, 48)];
 
2115
                    T4R = T4P - T4Q;
 
2116
                    Tcn = T4P + T4Q;
 
2117
                    {
 
2118
                         E Tb, Tc, Tcf, T28, T29, Tcg;
 
2119
                         Tb = iio[-WS(ios, 56)];
 
2120
                         Tc = rio[WS(ios, 24)];
 
2121
                         Tcf = Tb - Tc;
 
2122
                         T28 = iio[-WS(ios, 24)];
 
2123
                         T29 = rio[WS(ios, 56)];
 
2124
                         Tcg = T29 + T28;
 
2125
                         Td = Tb + Tc;
 
2126
                         Tcy = Tcf + Tcg;
 
2127
                         T2a = T28 - T29;
 
2128
                         Tch = Tcf - Tcg;
 
2129
                    }
 
2130
                    {
 
2131
                         E T8, T9, Tcj, T25, T26, Tci;
 
2132
                         T8 = rio[WS(ios, 8)];
 
2133
                         T9 = iio[-WS(ios, 40)];
 
2134
                         Tcj = T8 - T9;
 
2135
                         T25 = iio[-WS(ios, 8)];
 
2136
                         T26 = rio[WS(ios, 40)];
 
2137
                         Tci = T25 + T26;
 
2138
                         Ta = T8 + T9;
 
2139
                         Tcx = Tcj + Tci;
 
2140
                         T27 = T25 - T26;
 
2141
                         Tck = Tci - Tcj;
 
2142
                    }
 
2143
               }
 
2144
               {
 
2145
                    E T7, Te, Tf8, Tf9;
 
2146
                    T7 = T3 + T6;
 
2147
                    Te = Ta + Td;
 
2148
                    Tf = T7 + Te;
 
2149
                    T7i = T7 - Te;
 
2150
                    Tf8 = Tcv - Tcu;
 
2151
                    Tf9 = KP707106781 * (Tck + Tch);
 
2152
                    Tfa = Tf8 + Tf9;
 
2153
                    ThM = Tf8 - Tf9;
 
2154
               }
 
2155
               {
 
2156
                    E Tgn, Tgo, T24, T2b;
 
2157
                    Tgn = KP707106781 * (Tcx + Tcy);
 
2158
                    Tgo = Tcm + Tcn;
 
2159
                    Tgp = Tgn + Tgo;
 
2160
                    ThH = Tgo - Tgn;
 
2161
                    T24 = T3 - T6;
 
2162
                    T2b = T27 - T2a;
 
2163
                    T2c = T24 + T2b;
 
2164
                    T5O = T24 - T2b;
 
2165
               }
 
2166
               {
 
2167
                    E T4L, T4S, Tcl, Tco;
 
2168
                    T4L = Td - Ta;
 
2169
                    T4S = T4O - T4R;
 
2170
                    T4T = T4L + T4S;
 
2171
                    T6n = T4S - T4L;
 
2172
                    Tcl = KP707106781 * (Tch - Tck);
 
2173
                    Tco = Tcm - Tcn;
 
2174
                    Tcp = Tcl + Tco;
 
2175
                    Ted = Tco - Tcl;
 
2176
               }
 
2177
               {
 
2178
                    E Tcw, Tcz, T85, T86;
 
2179
                    Tcw = Tcu + Tcv;
 
2180
                    Tcz = KP707106781 * (Tcx - Tcy);
 
2181
                    TcA = Tcw + Tcz;
 
2182
                    TdE = Tcw - Tcz;
 
2183
                    T85 = T4O + T4R;
 
2184
                    T86 = T27 + T2a;
 
2185
                    T87 = T85 - T86;
 
2186
                    T9o = T86 + T85;
 
2187
               }
 
2188
          }
 
2189
          {
 
2190
               E TC, Tby, T2x, Tbu, T2N, Tbz, T7o, Tbv, TJ, TbB, TbC, T2E, T2G, Tbp, Tbs;
 
2191
               E T7p, Tfj, Tfk;
 
2192
               {
 
2193
                    E Tw, Tx, Ty, Tz, TA, TB;
 
2194
                    Tw = rio[WS(ios, 2)];
 
2195
                    Tx = iio[-WS(ios, 34)];
 
2196
                    Ty = Tw + Tx;
 
2197
                    Tz = rio[WS(ios, 18)];
 
2198
                    TA = iio[-WS(ios, 50)];
 
2199
                    TB = Tz + TA;
 
2200
                    TC = Ty + TB;
 
2201
                    Tby = Tz - TA;
 
2202
                    T2x = Ty - TB;
 
2203
                    Tbu = Tw - Tx;
 
2204
               }
 
2205
               {
 
2206
                    E T2H, T2I, T2J, T2K, T2L, T2M;
 
2207
                    T2H = iio[-WS(ios, 2)];
 
2208
                    T2I = rio[WS(ios, 34)];
 
2209
                    T2J = T2H - T2I;
 
2210
                    T2K = iio[-WS(ios, 18)];
 
2211
                    T2L = rio[WS(ios, 50)];
 
2212
                    T2M = T2K - T2L;
 
2213
                    T2N = T2J - T2M;
 
2214
                    Tbz = T2H + T2I;
 
2215
                    T7o = T2J + T2M;
 
2216
                    Tbv = T2K + T2L;
 
2217
               }
 
2218
               {
 
2219
                    E TF, Tbr, T2A, Tbq, TI, Tbn, T2D, Tbo;
 
2220
                    {
 
2221
                         E TD, TE, T2y, T2z;
 
2222
                         TD = rio[WS(ios, 10)];
 
2223
                         TE = iio[-WS(ios, 42)];
 
2224
                         TF = TD + TE;
 
2225
                         Tbr = TD - TE;
 
2226
                         T2y = iio[-WS(ios, 10)];
 
2227
                         T2z = rio[WS(ios, 42)];
 
2228
                         T2A = T2y - T2z;
 
2229
                         Tbq = T2y + T2z;
 
2230
                    }
 
2231
                    {
 
2232
                         E TG, TH, T2B, T2C;
 
2233
                         TG = iio[-WS(ios, 58)];
 
2234
                         TH = rio[WS(ios, 26)];
 
2235
                         TI = TG + TH;
 
2236
                         Tbn = TG - TH;
 
2237
                         T2B = iio[-WS(ios, 26)];
 
2238
                         T2C = rio[WS(ios, 58)];
 
2239
                         T2D = T2B - T2C;
 
2240
                         Tbo = T2C + T2B;
 
2241
                    }
 
2242
                    TJ = TF + TI;
 
2243
                    TbB = Tbr + Tbq;
 
2244
                    TbC = Tbn + Tbo;
 
2245
                    T2E = T2A - T2D;
 
2246
                    T2G = TI - TF;
 
2247
                    Tbp = Tbn - Tbo;
 
2248
                    Tbs = Tbq - Tbr;
 
2249
                    T7p = T2A + T2D;
 
2250
               }
 
2251
               TK = TC + TJ;
 
2252
               T93 = T7p + T7o;
 
2253
               {
 
2254
                    E T2F, T2O, Tfm, Tfn;
 
2255
                    T2F = T2x + T2E;
 
2256
                    T2O = T2G + T2N;
 
2257
                    T2P = FMA(KP923879532, T2F, KP382683432 * T2O);
 
2258
                    T4F = FNMS(KP382683432, T2F, KP923879532 * T2O);
 
2259
                    Tfm = KP707106781 * (TbB + TbC);
 
2260
                    Tfn = Tbu + Tbv;
 
2261
                    Tfo = Tfm + Tfn;
 
2262
                    Thz = Tfn - Tfm;
 
2263
               }
 
2264
               {
 
2265
                    E T5R, T5S, Tbt, Tbw;
 
2266
                    T5R = T2x - T2E;
 
2267
                    T5S = T2N - T2G;
 
2268
                    T5T = FNMS(KP382683432, T5S, KP923879532 * T5R);
 
2269
                    T6j = FMA(KP382683432, T5R, KP923879532 * T5S);
 
2270
                    Tbt = KP707106781 * (Tbp - Tbs);
 
2271
                    Tbw = Tbu - Tbv;
 
2272
                    Tbx = Tbt + Tbw;
 
2273
                    TdI = Tbw - Tbt;
 
2274
               }
 
2275
               Tfj = Tbz - Tby;
 
2276
               Tfk = KP707106781 * (Tbs + Tbp);
 
2277
               Tfl = Tfj + Tfk;
 
2278
               ThA = Tfj - Tfk;
 
2279
               {
 
2280
                    E T7n, T7q, TbA, TbD;
 
2281
                    T7n = TC - TJ;
 
2282
                    T7q = T7o - T7p;
 
2283
                    T7r = T7n + T7q;
 
2284
                    T81 = T7q - T7n;
 
2285
                    TbA = Tby + Tbz;
 
2286
                    TbD = KP707106781 * (TbB - TbC);
 
2287
                    TbE = TbA + TbD;
 
2288
                    TdH = TbA - TbD;
 
2289
               }
 
2290
          }
 
2291
          {
 
2292
               E TR, TbU, T2Q, TbN, T36, TbV, T7t, TbO, TY, TbR, TbS, T2X, T2Z, TbI, TbL;
 
2293
               E T7u, Tfq, Tfr;
 
2294
               {
 
2295
                    E TL, TM, TN, TO, TP, TQ;
 
2296
                    TL = iio[-WS(ios, 62)];
 
2297
                    TM = rio[WS(ios, 30)];
 
2298
                    TN = TL + TM;
 
2299
                    TO = rio[WS(ios, 14)];
 
2300
                    TP = iio[-WS(ios, 46)];
 
2301
                    TQ = TO + TP;
 
2302
                    TR = TN + TQ;
 
2303
                    TbU = TL - TM;
 
2304
                    T2Q = TN - TQ;
 
2305
                    TbN = TO - TP;
 
2306
               }
 
2307
               {
 
2308
                    E T30, T31, T32, T33, T34, T35;
 
2309
                    T30 = iio[-WS(ios, 30)];
 
2310
                    T31 = rio[WS(ios, 62)];
 
2311
                    T32 = T30 - T31;
 
2312
                    T33 = iio[-WS(ios, 14)];
 
2313
                    T34 = rio[WS(ios, 46)];
 
2314
                    T35 = T33 - T34;
 
2315
                    T36 = T32 - T35;
 
2316
                    TbV = T33 + T34;
 
2317
                    T7t = T32 + T35;
 
2318
                    TbO = T31 + T30;
 
2319
               }
 
2320
               {
 
2321
                    E TU, TbG, T2T, TbH, TX, TbJ, T2W, TbK;
 
2322
                    {
 
2323
                         E TS, TT, T2R, T2S;
 
2324
                         TS = rio[WS(ios, 6)];
 
2325
                         TT = iio[-WS(ios, 38)];
 
2326
                         TU = TS + TT;
 
2327
                         TbG = TS - TT;
 
2328
                         T2R = iio[-WS(ios, 6)];
 
2329
                         T2S = rio[WS(ios, 38)];
 
2330
                         T2T = T2R - T2S;
 
2331
                         TbH = T2R + T2S;
 
2332
                    }
 
2333
                    {
 
2334
                         E TV, TW, T2U, T2V;
 
2335
                         TV = iio[-WS(ios, 54)];
 
2336
                         TW = rio[WS(ios, 22)];
 
2337
                         TX = TV + TW;
 
2338
                         TbJ = TV - TW;
 
2339
                         T2U = iio[-WS(ios, 22)];
 
2340
                         T2V = rio[WS(ios, 54)];
 
2341
                         T2W = T2U - T2V;
 
2342
                         TbK = T2V + T2U;
 
2343
                    }
 
2344
                    TY = TU + TX;
 
2345
                    TbR = TbJ - TbK;
 
2346
                    TbS = TbH - TbG;
 
2347
                    T2X = T2T - T2W;
 
2348
                    T2Z = TX - TU;
 
2349
                    TbI = TbG + TbH;
 
2350
                    TbL = TbJ + TbK;
 
2351
                    T7u = T2T + T2W;
 
2352
               }
 
2353
               TZ = TR + TY;
 
2354
               T94 = T7u + T7t;
 
2355
               {
 
2356
                    E T2Y, T37, Tft, Tfu;
 
2357
                    T2Y = T2Q + T2X;
 
2358
                    T37 = T2Z + T36;
 
2359
                    T38 = FNMS(KP382683432, T37, KP923879532 * T2Y);
 
2360
                    T4G = FMA(KP382683432, T2Y, KP923879532 * T37);
 
2361
                    Tft = KP707106781 * (TbI + TbL);
 
2362
                    Tfu = TbU + TbV;
 
2363
                    Tfv = Tft + Tfu;
 
2364
                    ThC = Tfu - Tft;
 
2365
               }
 
2366
               {
 
2367
                    E T5U, T5V, TbM, TbP;
 
2368
                    T5U = T2Q - T2X;
 
2369
                    T5V = T36 - T2Z;
 
2370
                    T5W = FMA(KP923879532, T5U, KP382683432 * T5V);
 
2371
                    T6k = FNMS(KP382683432, T5U, KP923879532 * T5V);
 
2372
                    TbM = KP707106781 * (TbI - TbL);
 
2373
                    TbP = TbN - TbO;
 
2374
                    TbQ = TbM + TbP;
 
2375
                    TdK = TbP - TbM;
 
2376
               }
 
2377
               Tfq = KP707106781 * (TbS + TbR);
 
2378
               Tfr = TbN + TbO;
 
2379
               Tfs = Tfq - Tfr;
 
2380
               ThD = Tfq + Tfr;
 
2381
               {
 
2382
                    E T7s, T7v, TbT, TbW;
 
2383
                    T7s = TR - TY;
 
2384
                    T7v = T7t - T7u;
 
2385
                    T7w = T7s - T7v;
 
2386
                    T82 = T7s + T7v;
 
2387
                    TbT = KP707106781 * (TbR - TbS);
 
2388
                    TbW = TbU - TbV;
 
2389
                    TbX = TbT + TbW;
 
2390
                    TdL = TbW - TbT;
 
2391
               }
 
2392
          }
 
2393
          {
 
2394
               E Ti, T2g, Tl, T2j, T2d, T2k, Tfc, Tfb, Tc5, Tc2, Tp, T2p, Ts, T2s, T2m;
 
2395
               E T2t, Tff, Tfe, Tcc, Tc9;
 
2396
               {
 
2397
                    E Tc0, Tc4, Tc3, Tc1;
 
2398
                    {
 
2399
                         E Tg, Th, T2e, T2f;
 
2400
                         Tg = rio[WS(ios, 4)];
 
2401
                         Th = iio[-WS(ios, 36)];
 
2402
                         Ti = Tg + Th;
 
2403
                         Tc0 = Tg - Th;
 
2404
                         T2e = iio[-WS(ios, 4)];
 
2405
                         T2f = rio[WS(ios, 36)];
 
2406
                         T2g = T2e - T2f;
 
2407
                         Tc4 = T2e + T2f;
 
2408
                    }
 
2409
                    {
 
2410
                         E Tj, Tk, T2h, T2i;
 
2411
                         Tj = rio[WS(ios, 20)];
 
2412
                         Tk = iio[-WS(ios, 52)];
 
2413
                         Tl = Tj + Tk;
 
2414
                         Tc3 = Tj - Tk;
 
2415
                         T2h = iio[-WS(ios, 20)];
 
2416
                         T2i = rio[WS(ios, 52)];
 
2417
                         T2j = T2h - T2i;
 
2418
                         Tc1 = T2h + T2i;
 
2419
                    }
 
2420
                    T2d = Ti - Tl;
 
2421
                    T2k = T2g - T2j;
 
2422
                    Tfc = Tc0 + Tc1;
 
2423
                    Tfb = Tc4 - Tc3;
 
2424
                    Tc5 = Tc3 + Tc4;
 
2425
                    Tc2 = Tc0 - Tc1;
 
2426
               }
 
2427
               {
 
2428
                    E Tc7, Tcb, Tca, Tc8;
 
2429
                    {
 
2430
                         E Tn, To, T2n, T2o;
 
2431
                         Tn = iio[-WS(ios, 60)];
 
2432
                         To = rio[WS(ios, 28)];
 
2433
                         Tp = Tn + To;
 
2434
                         Tc7 = Tn - To;
 
2435
                         T2n = iio[-WS(ios, 28)];
 
2436
                         T2o = rio[WS(ios, 60)];
 
2437
                         T2p = T2n - T2o;
 
2438
                         Tcb = T2o + T2n;
 
2439
                    }
 
2440
                    {
 
2441
                         E Tq, Tr, T2q, T2r;
 
2442
                         Tq = rio[WS(ios, 12)];
 
2443
                         Tr = iio[-WS(ios, 44)];
 
2444
                         Ts = Tq + Tr;
 
2445
                         Tca = Tq - Tr;
 
2446
                         T2q = iio[-WS(ios, 12)];
 
2447
                         T2r = rio[WS(ios, 44)];
 
2448
                         T2s = T2q - T2r;
 
2449
                         Tc8 = T2q + T2r;
 
2450
                    }
 
2451
                    T2m = Tp - Ts;
 
2452
                    T2t = T2p - T2s;
 
2453
                    Tff = Tca + Tcb;
 
2454
                    Tfe = Tc7 + Tc8;
 
2455
                    Tcc = Tca - Tcb;
 
2456
                    Tc9 = Tc7 - Tc8;
 
2457
               }
 
2458
               {
 
2459
                    E Tm, Tt, Tfd, Tfg;
 
2460
                    Tm = Ti + Tl;
 
2461
                    Tt = Tp + Ts;
 
2462
                    Tu = Tm + Tt;
 
2463
                    T84 = Tt - Tm;
 
2464
                    Tfd = FNMS(KP382683432, Tfc, KP923879532 * Tfb);
 
2465
                    Tfg = FNMS(KP923879532, Tff, KP382683432 * Tfe);
 
2466
                    Tfh = Tfd + Tfg;
 
2467
                    ThG = Tfg - Tfd;
 
2468
               }
 
2469
               {
 
2470
                    E Tgk, Tgl, T2l, T2u;
 
2471
                    Tgk = FMA(KP382683432, Tfb, KP923879532 * Tfc);
 
2472
                    Tgl = FMA(KP923879532, Tfe, KP382683432 * Tff);
 
2473
                    Tgm = Tgk + Tgl;
 
2474
                    ThN = Tgk - Tgl;
 
2475
                    T2l = T2d + T2k;
 
2476
                    T2u = T2m - T2t;
 
2477
                    T2v = KP707106781 * (T2l + T2u);
 
2478
                    T6m = KP707106781 * (T2l - T2u);
 
2479
               }
 
2480
               {
 
2481
                    E T4I, T4J, Tc6, Tcd;
 
2482
                    T4I = T2k - T2d;
 
2483
                    T4J = T2m + T2t;
 
2484
                    T4K = KP707106781 * (T4I + T4J);
 
2485
                    T5P = KP707106781 * (T4J - T4I);
 
2486
                    Tc6 = FNMS(KP382683432, Tc5, KP923879532 * Tc2);
 
2487
                    Tcd = FMA(KP923879532, Tc9, KP382683432 * Tcc);
 
2488
                    Tce = Tc6 + Tcd;
 
2489
                    TdF = Tcd - Tc6;
 
2490
               }
 
2491
               {
 
2492
                    E TcB, TcC, T7j, T7k;
 
2493
                    TcB = FMA(KP923879532, Tc5, KP382683432 * Tc2);
 
2494
                    TcC = FNMS(KP382683432, Tc9, KP923879532 * Tcc);
 
2495
                    TcD = TcB + TcC;
 
2496
                    Tec = TcB - TcC;
 
2497
                    T7j = T2g + T2j;
 
2498
                    T7k = T2s + T2p;
 
2499
                    T7l = T7j - T7k;
 
2500
                    T9p = T7j + T7k;
 
2501
               }
 
2502
          }
 
2503
          {
 
2504
               E T1z, T1C, T1D, Tbg, TaQ, T4r, T4u, T7Q, Tbh, TaR, T1G, T3V, T1J, T3Y, T1K;
 
2505
               E T7R, Tbe, Tbd, TaO, TaL, T1S, TfV, TfW, T41, T48, TaW, TaZ, T7L, T1Z, TfY;
 
2506
               E TfZ, T4a, T4h, Tb3, Tb6, T7M;
 
2507
               {
 
2508
                    E T1x, T1y, T1A, T1B;
 
2509
                    T1x = iio[-WS(ios, 63)];
 
2510
                    T1y = rio[WS(ios, 31)];
 
2511
                    T1z = T1x + T1y;
 
2512
                    T1A = rio[WS(ios, 15)];
 
2513
                    T1B = iio[-WS(ios, 47)];
 
2514
                    T1C = T1A + T1B;
 
2515
                    T1D = T1z + T1C;
 
2516
                    Tbg = T1x - T1y;
 
2517
                    TaQ = T1A - T1B;
 
2518
               }
 
2519
               {
 
2520
                    E T4p, T4q, T4s, T4t;
 
2521
                    T4p = iio[-WS(ios, 31)];
 
2522
                    T4q = rio[WS(ios, 63)];
 
2523
                    T4r = T4p - T4q;
 
2524
                    T4s = iio[-WS(ios, 15)];
 
2525
                    T4t = rio[WS(ios, 47)];
 
2526
                    T4u = T4s - T4t;
 
2527
                    T7Q = T4r + T4u;
 
2528
                    Tbh = T4s + T4t;
 
2529
                    TaR = T4q + T4p;
 
2530
               }
 
2531
               {
 
2532
                    E TaJ, TaK, TaM, TaN;
 
2533
                    {
 
2534
                         E T1E, T1F, T3T, T3U;
 
2535
                         T1E = rio[WS(ios, 7)];
 
2536
                         T1F = iio[-WS(ios, 39)];
 
2537
                         T1G = T1E + T1F;
 
2538
                         TaJ = T1E - T1F;
 
2539
                         T3T = iio[-WS(ios, 7)];
 
2540
                         T3U = rio[WS(ios, 39)];
 
2541
                         T3V = T3T - T3U;
 
2542
                         TaK = T3T + T3U;
 
2543
                    }
 
2544
                    {
 
2545
                         E T1H, T1I, T3W, T3X;
 
2546
                         T1H = iio[-WS(ios, 55)];
 
2547
                         T1I = rio[WS(ios, 23)];
 
2548
                         T1J = T1H + T1I;
 
2549
                         TaM = T1H - T1I;
 
2550
                         T3W = iio[-WS(ios, 23)];
 
2551
                         T3X = rio[WS(ios, 55)];
 
2552
                         T3Y = T3W - T3X;
 
2553
                         TaN = T3X + T3W;
 
2554
                    }
 
2555
                    T1K = T1G + T1J;
 
2556
                    T7R = T3V + T3Y;
 
2557
                    Tbe = TaK - TaJ;
 
2558
                    Tbd = TaM - TaN;
 
2559
                    TaO = TaM + TaN;
 
2560
                    TaL = TaJ + TaK;
 
2561
               }
 
2562
               {
 
2563
                    E T1O, TaX, T44, TaV, T1R, TaU, T47, TaY;
 
2564
                    {
 
2565
                         E T1M, T1N, T42, T43;
 
2566
                         T1M = rio[WS(ios, 3)];
 
2567
                         T1N = iio[-WS(ios, 35)];
 
2568
                         T1O = T1M + T1N;
 
2569
                         TaX = T1M - T1N;
 
2570
                         T42 = iio[-WS(ios, 3)];
 
2571
                         T43 = rio[WS(ios, 35)];
 
2572
                         T44 = T42 - T43;
 
2573
                         TaV = T42 + T43;
 
2574
                    }
 
2575
                    {
 
2576
                         E T1P, T1Q, T45, T46;
 
2577
                         T1P = rio[WS(ios, 19)];
 
2578
                         T1Q = iio[-WS(ios, 51)];
 
2579
                         T1R = T1P + T1Q;
 
2580
                         TaU = T1P - T1Q;
 
2581
                         T45 = iio[-WS(ios, 19)];
 
2582
                         T46 = rio[WS(ios, 51)];
 
2583
                         T47 = T45 - T46;
 
2584
                         TaY = T45 + T46;
 
2585
                    }
 
2586
                    T1S = T1O + T1R;
 
2587
                    TfV = TaV - TaU;
 
2588
                    TfW = TaX + TaY;
 
2589
                    T41 = T1O - T1R;
 
2590
                    T48 = T44 - T47;
 
2591
                    TaW = TaU + TaV;
 
2592
                    TaZ = TaX - TaY;
 
2593
                    T7L = T44 + T47;
 
2594
               }
 
2595
               {
 
2596
                    E T1V, Tb4, T4d, Tb2, T1Y, Tb1, T4g, Tb5;
 
2597
                    {
 
2598
                         E T1T, T1U, T4b, T4c;
 
2599
                         T1T = iio[-WS(ios, 59)];
 
2600
                         T1U = rio[WS(ios, 27)];
 
2601
                         T1V = T1T + T1U;
 
2602
                         Tb4 = T1T - T1U;
 
2603
                         T4b = iio[-WS(ios, 27)];
 
2604
                         T4c = rio[WS(ios, 59)];
 
2605
                         T4d = T4b - T4c;
 
2606
                         Tb2 = T4c + T4b;
 
2607
                    }
 
2608
                    {
 
2609
                         E T1W, T1X, T4e, T4f;
 
2610
                         T1W = rio[WS(ios, 11)];
 
2611
                         T1X = iio[-WS(ios, 43)];
 
2612
                         T1Y = T1W + T1X;
 
2613
                         Tb1 = T1W - T1X;
 
2614
                         T4e = iio[-WS(ios, 11)];
 
2615
                         T4f = rio[WS(ios, 43)];
 
2616
                         T4g = T4e - T4f;
 
2617
                         Tb5 = T4e + T4f;
 
2618
                    }
 
2619
                    T1Z = T1V + T1Y;
 
2620
                    TfY = Tb4 + Tb5;
 
2621
                    TfZ = Tb1 + Tb2;
 
2622
                    T4a = T1V - T1Y;
 
2623
                    T4h = T4d - T4g;
 
2624
                    Tb3 = Tb1 - Tb2;
 
2625
                    Tb6 = Tb4 - Tb5;
 
2626
                    T7M = T4g + T4d;
 
2627
               }
 
2628
               T1L = T1D + T1K;
 
2629
               T20 = T1S + T1Z;
 
2630
               T9c = T1L - T20;
 
2631
               T9d = T7R + T7Q;
 
2632
               T9e = T7L + T7M;
 
2633
               T9f = T9d - T9e;
 
2634
               {
 
2635
                    E T3S, T3Z, TfX, Tg0;
 
2636
                    T3S = T1z - T1C;
 
2637
                    T3Z = T3V - T3Y;
 
2638
                    T40 = T3S + T3Z;
 
2639
                    T66 = T3S - T3Z;
 
2640
                    TfX = FNMS(KP382683432, TfW, KP923879532 * TfV);
 
2641
                    Tg0 = FNMS(KP923879532, TfZ, KP382683432 * TfY);
 
2642
                    Tg1 = TfX + Tg0;
 
2643
                    Thu = Tg0 - TfX;
 
2644
               }
 
2645
               {
 
2646
                    E Tg6, Tg7, Tg3, Tg4;
 
2647
                    Tg6 = KP707106781 * (TaL + TaO);
 
2648
                    Tg7 = Tbg + Tbh;
 
2649
                    Tg8 = Tg6 + Tg7;
 
2650
                    Thv = Tg7 - Tg6;
 
2651
                    Tg3 = FMA(KP382683432, TfV, KP923879532 * TfW);
 
2652
                    Tg4 = FMA(KP923879532, TfY, KP382683432 * TfZ);
 
2653
                    Tg5 = Tg3 + Tg4;
 
2654
                    Thr = Tg3 - Tg4;
 
2655
               }
 
2656
               {
 
2657
                    E T4l, T4m, T49, T4i;
 
2658
                    T4l = T48 - T41;
 
2659
                    T4m = T4a + T4h;
 
2660
                    T4n = KP707106781 * (T4l + T4m);
 
2661
                    T67 = KP707106781 * (T4m - T4l);
 
2662
                    T49 = T41 + T48;
 
2663
                    T4i = T4a - T4h;
 
2664
                    T4j = KP707106781 * (T49 + T4i);
 
2665
                    T69 = KP707106781 * (T49 - T4i);
 
2666
               }
 
2667
               {
 
2668
                    E T4o, T4v, TaP, TaS;
 
2669
                    T4o = T1J - T1G;
 
2670
                    T4v = T4r - T4u;
 
2671
                    T4w = T4o + T4v;
 
2672
                    T6a = T4v - T4o;
 
2673
                    TaP = KP707106781 * (TaL - TaO);
 
2674
                    TaS = TaQ - TaR;
 
2675
                    TaT = TaP + TaS;
 
2676
                    TdW = TaS - TaP;
 
2677
               }
 
2678
               {
 
2679
                    E Tb0, Tb7, TfS, TfT;
 
2680
                    Tb0 = FMA(KP923879532, TaW, KP382683432 * TaZ);
 
2681
                    Tb7 = FNMS(KP382683432, Tb6, KP923879532 * Tb3);
 
2682
                    Tb8 = Tb0 + Tb7;
 
2683
                    TdZ = Tb0 - Tb7;
 
2684
                    TfS = KP707106781 * (Tbe + Tbd);
 
2685
                    TfT = TaQ + TaR;
 
2686
                    TfU = TfS - TfT;
 
2687
                    Ths = TfS + TfT;
 
2688
               }
 
2689
               {
 
2690
                    E T7K, T7N, T7P, T7S;
 
2691
                    T7K = T1D - T1K;
 
2692
                    T7N = T7L - T7M;
 
2693
                    T7O = T7K + T7N;
 
2694
                    T8y = T7K - T7N;
 
2695
                    T7P = T1Z - T1S;
 
2696
                    T7S = T7Q - T7R;
 
2697
                    T7T = T7P + T7S;
 
2698
                    T8z = T7S - T7P;
 
2699
               }
 
2700
               {
 
2701
                    E Tba, Tbb, Tbf, Tbi;
 
2702
                    Tba = FNMS(KP382683432, TaW, KP923879532 * TaZ);
 
2703
                    Tbb = FMA(KP923879532, Tb6, KP382683432 * Tb3);
 
2704
                    Tbc = Tba + Tbb;
 
2705
                    TdX = Tbb - Tba;
 
2706
                    Tbf = KP707106781 * (Tbd - Tbe);
 
2707
                    Tbi = Tbg - Tbh;
 
2708
                    Tbj = Tbf + Tbi;
 
2709
                    Te0 = Tbi - Tbf;
 
2710
               }
 
2711
          }
 
2712
          {
 
2713
               E T14, T17, T18, Tax, Tas, T3K, T3N, T7F, Tay, Tat, T1b, T3e, T1e, T3h, T1f;
 
2714
               E T7G, TaB, TaA, Taq, Tan, T1n, TfC, TfD, T3k, T3r, Ta8, Tab, T7A, T1u, TfF;
 
2715
               E TfG, T3t, T3A, Taf, Tai, T7B;
 
2716
               {
 
2717
                    E T12, T13, T15, T16;
 
2718
                    T12 = rio[WS(ios, 1)];
 
2719
                    T13 = iio[-WS(ios, 33)];
 
2720
                    T14 = T12 + T13;
 
2721
                    T15 = rio[WS(ios, 17)];
 
2722
                    T16 = iio[-WS(ios, 49)];
 
2723
                    T17 = T15 + T16;
 
2724
                    T18 = T14 + T17;
 
2725
                    Tax = T15 - T16;
 
2726
                    Tas = T12 - T13;
 
2727
               }
 
2728
               {
 
2729
                    E T3I, T3J, T3L, T3M;
 
2730
                    T3I = iio[-WS(ios, 1)];
 
2731
                    T3J = rio[WS(ios, 33)];
 
2732
                    T3K = T3I - T3J;
 
2733
                    T3L = iio[-WS(ios, 17)];
 
2734
                    T3M = rio[WS(ios, 49)];
 
2735
                    T3N = T3L - T3M;
 
2736
                    T7F = T3K + T3N;
 
2737
                    Tay = T3I + T3J;
 
2738
                    Tat = T3L + T3M;
 
2739
               }
 
2740
               {
 
2741
                    E Tap, Tao, Tal, Tam;
 
2742
                    {
 
2743
                         E T19, T1a, T3c, T3d;
 
2744
                         T19 = rio[WS(ios, 9)];
 
2745
                         T1a = iio[-WS(ios, 41)];
 
2746
                         T1b = T19 + T1a;
 
2747
                         Tap = T19 - T1a;
 
2748
                         T3c = iio[-WS(ios, 9)];
 
2749
                         T3d = rio[WS(ios, 41)];
 
2750
                         T3e = T3c - T3d;
 
2751
                         Tao = T3c + T3d;
 
2752
                    }
 
2753
                    {
 
2754
                         E T1c, T1d, T3f, T3g;
 
2755
                         T1c = iio[-WS(ios, 57)];
 
2756
                         T1d = rio[WS(ios, 25)];
 
2757
                         T1e = T1c + T1d;
 
2758
                         Tal = T1c - T1d;
 
2759
                         T3f = iio[-WS(ios, 25)];
 
2760
                         T3g = rio[WS(ios, 57)];
 
2761
                         T3h = T3f - T3g;
 
2762
                         Tam = T3g + T3f;
 
2763
                    }
 
2764
                    T1f = T1b + T1e;
 
2765
                    T7G = T3e + T3h;
 
2766
                    TaB = Tal + Tam;
 
2767
                    TaA = Tap + Tao;
 
2768
                    Taq = Tao - Tap;
 
2769
                    Tan = Tal - Tam;
 
2770
               }
 
2771
               {
 
2772
                    E T1j, Ta6, T3n, Taa, T1m, Ta9, T3q, Ta7;
 
2773
                    {
 
2774
                         E T1h, T1i, T3l, T3m;
 
2775
                         T1h = rio[WS(ios, 5)];
 
2776
                         T1i = iio[-WS(ios, 37)];
 
2777
                         T1j = T1h + T1i;
 
2778
                         Ta6 = T1h - T1i;
 
2779
                         T3l = iio[-WS(ios, 5)];
 
2780
                         T3m = rio[WS(ios, 37)];
 
2781
                         T3n = T3l - T3m;
 
2782
                         Taa = T3l + T3m;
 
2783
                    }
 
2784
                    {
 
2785
                         E T1k, T1l, T3o, T3p;
 
2786
                         T1k = rio[WS(ios, 21)];
 
2787
                         T1l = iio[-WS(ios, 53)];
 
2788
                         T1m = T1k + T1l;
 
2789
                         Ta9 = T1k - T1l;
 
2790
                         T3o = iio[-WS(ios, 21)];
 
2791
                         T3p = rio[WS(ios, 53)];
 
2792
                         T3q = T3o - T3p;
 
2793
                         Ta7 = T3o + T3p;
 
2794
                    }
 
2795
                    T1n = T1j + T1m;
 
2796
                    TfC = Taa - Ta9;
 
2797
                    TfD = Ta6 + Ta7;
 
2798
                    T3k = T1j - T1m;
 
2799
                    T3r = T3n - T3q;
 
2800
                    Ta8 = Ta6 - Ta7;
 
2801
                    Tab = Ta9 + Taa;
 
2802
                    T7A = T3n + T3q;
 
2803
               }
 
2804
               {
 
2805
                    E T1q, Tad, T3w, Tah, T1t, Tag, T3z, Tae;
 
2806
                    {
 
2807
                         E T1o, T1p, T3u, T3v;
 
2808
                         T1o = iio[-WS(ios, 61)];
 
2809
                         T1p = rio[WS(ios, 29)];
 
2810
                         T1q = T1o + T1p;
 
2811
                         Tad = T1o - T1p;
 
2812
                         T3u = iio[-WS(ios, 29)];
 
2813
                         T3v = rio[WS(ios, 61)];
 
2814
                         T3w = T3u - T3v;
 
2815
                         Tah = T3v + T3u;
 
2816
                    }
 
2817
                    {
 
2818
                         E T1r, T1s, T3x, T3y;
 
2819
                         T1r = rio[WS(ios, 13)];
 
2820
                         T1s = iio[-WS(ios, 45)];
 
2821
                         T1t = T1r + T1s;
 
2822
                         Tag = T1r - T1s;
 
2823
                         T3x = iio[-WS(ios, 13)];
 
2824
                         T3y = rio[WS(ios, 45)];
 
2825
                         T3z = T3x - T3y;
 
2826
                         Tae = T3x + T3y;
 
2827
                    }
 
2828
                    T1u = T1q + T1t;
 
2829
                    TfF = Tad + Tae;
 
2830
                    TfG = Tag + Tah;
 
2831
                    T3t = T1q - T1t;
 
2832
                    T3A = T3w - T3z;
 
2833
                    Taf = Tad - Tae;
 
2834
                    Tai = Tag - Tah;
 
2835
                    T7B = T3z + T3w;
 
2836
               }
 
2837
               T1g = T18 + T1f;
 
2838
               T1v = T1n + T1u;
 
2839
               T97 = T1g - T1v;
 
2840
               T98 = T7G + T7F;
 
2841
               T99 = T7A + T7B;
 
2842
               T9a = T98 - T99;
 
2843
               {
 
2844
                    E T3b, T3i, TfE, TfH;
 
2845
                    T3b = T14 - T17;
 
2846
                    T3i = T3e - T3h;
 
2847
                    T3j = T3b + T3i;
 
2848
                    T5Z = T3b - T3i;
 
2849
                    TfE = FNMS(KP382683432, TfD, KP923879532 * TfC);
 
2850
                    TfH = FNMS(KP923879532, TfG, KP382683432 * TfF);
 
2851
                    TfI = TfE + TfH;
 
2852
                    Thk = TfH - TfE;
 
2853
               }
 
2854
               {
 
2855
                    E TfN, TfO, TfK, TfL;
 
2856
                    TfN = KP707106781 * (TaA + TaB);
 
2857
                    TfO = Tas + Tat;
 
2858
                    TfP = TfN + TfO;
 
2859
                    Thl = TfO - TfN;
 
2860
                    TfK = FMA(KP382683432, TfC, KP923879532 * TfD);
 
2861
                    TfL = FMA(KP923879532, TfF, KP382683432 * TfG);
 
2862
                    TfM = TfK + TfL;
 
2863
                    Tho = TfK - TfL;
 
2864
               }
 
2865
               {
 
2866
                    E T3E, T3F, T3s, T3B;
 
2867
                    T3E = T3r - T3k;
 
2868
                    T3F = T3t + T3A;
 
2869
                    T3G = KP707106781 * (T3E + T3F);
 
2870
                    T60 = KP707106781 * (T3F - T3E);
 
2871
                    T3s = T3k + T3r;
 
2872
                    T3B = T3t - T3A;
 
2873
                    T3C = KP707106781 * (T3s + T3B);
 
2874
                    T62 = KP707106781 * (T3s - T3B);
 
2875
               }
 
2876
               {
 
2877
                    E T3H, T3O, Tac, Taj;
 
2878
                    T3H = T1e - T1b;
 
2879
                    T3O = T3K - T3N;
 
2880
                    T3P = T3H + T3O;
 
2881
                    T63 = T3O - T3H;
 
2882
                    Tac = FNMS(KP382683432, Tab, KP923879532 * Ta8);
 
2883
                    Taj = FMA(KP923879532, Taf, KP382683432 * Tai);
 
2884
                    Tak = Tac + Taj;
 
2885
                    TdQ = Taj - Tac;
 
2886
               }
 
2887
               {
 
2888
                    E Tar, Tau, Tfz, TfA;
 
2889
                    Tar = KP707106781 * (Tan - Taq);
 
2890
                    Tau = Tas - Tat;
 
2891
                    Tav = Tar + Tau;
 
2892
                    TdT = Tau - Tar;
 
2893
                    Tfz = Tay - Tax;
 
2894
                    TfA = KP707106781 * (Taq + Tan);
 
2895
                    TfB = Tfz + TfA;
 
2896
                    Thn = Tfz - TfA;
 
2897
               }
 
2898
               {
 
2899
                    E T7z, T7C, T7E, T7H;
 
2900
                    T7z = T18 - T1f;
 
2901
                    T7C = T7A - T7B;
 
2902
                    T7D = T7z + T7C;
 
2903
                    T8v = T7z - T7C;
 
2904
                    T7E = T1u - T1n;
 
2905
                    T7H = T7F - T7G;
 
2906
                    T7I = T7E + T7H;
 
2907
                    T8w = T7H - T7E;
 
2908
               }
 
2909
               {
 
2910
                    E Taz, TaC, TaE, TaF;
 
2911
                    Taz = Tax + Tay;
 
2912
                    TaC = KP707106781 * (TaA - TaB);
 
2913
                    TaD = Taz + TaC;
 
2914
                    TdP = Taz - TaC;
 
2915
                    TaE = FMA(KP923879532, Tab, KP382683432 * Ta8);
 
2916
                    TaF = FNMS(KP382683432, Taf, KP923879532 * Tai);
 
2917
                    TaG = TaE + TaF;
 
2918
                    TdS = TaE - TaF;
 
2919
               }
 
2920
          }
 
2921
          {
 
2922
               E T11, T9K, T9T, Ta2, T22, T9Q, T9N, Ta3;
 
2923
               {
 
2924
                    E Tv, T10, T9R, T9S;
 
2925
                    Tv = Tf + Tu;
 
2926
                    T10 = TK + TZ;
 
2927
                    T11 = Tv + T10;
 
2928
                    T9K = Tv - T10;
 
2929
                    T9R = T9p + T9o;
 
2930
                    T9S = T93 + T94;
 
2931
                    T9T = T9R - T9S;
 
2932
                    Ta2 = T9S + T9R;
 
2933
               }
 
2934
               {
 
2935
                    E T1w, T21, T9L, T9M;
 
2936
                    T1w = T1g + T1v;
 
2937
                    T21 = T1L + T20;
 
2938
                    T22 = T1w + T21;
 
2939
                    T9Q = T21 - T1w;
 
2940
                    T9L = T99 + T98;
 
2941
                    T9M = T9e + T9d;
 
2942
                    T9N = T9L - T9M;
 
2943
                    Ta3 = T9L + T9M;
 
2944
               }
 
2945
               rio[0] = T11 + T22;
 
2946
               iio[-WS(ios, 63)] = Ta3 + Ta2;
 
2947
               {
 
2948
                    E T9O, T9U, T9J, T9P;
 
2949
                    T9O = T9K + T9N;
 
2950
                    T9U = T9Q + T9T;
 
2951
                    T9J = W[94];
 
2952
                    T9P = W[95];
 
2953
                    rio[WS(ios, 48)] = FNMS(T9P, T9U, T9J * T9O);
 
2954
                    iio[-WS(ios, 15)] = FMA(T9P, T9O, T9J * T9U);
 
2955
               }
 
2956
               {
 
2957
                    E T9W, T9Y, T9V, T9X;
 
2958
                    T9W = T9K - T9N;
 
2959
                    T9Y = T9T - T9Q;
 
2960
                    T9V = W[30];
 
2961
                    T9X = W[31];
 
2962
                    rio[WS(ios, 16)] = FNMS(T9X, T9Y, T9V * T9W);
 
2963
                    iio[-WS(ios, 47)] = FMA(T9X, T9W, T9V * T9Y);
 
2964
               }
 
2965
               {
 
2966
                    E Ta0, Ta4, T9Z, Ta1;
 
2967
                    Ta0 = T11 - T22;
 
2968
                    Ta4 = Ta2 - Ta3;
 
2969
                    T9Z = W[62];
 
2970
                    Ta1 = W[63];
 
2971
                    rio[WS(ios, 32)] = FNMS(Ta1, Ta4, T9Z * Ta0);
 
2972
                    iio[-WS(ios, 31)] = FMA(Ta1, Ta0, T9Z * Ta4);
 
2973
               }
 
2974
          }
 
2975
          {
 
2976
               E T96, T9y, T9r, T9D, T9h, T9C, T9m, T9z;
 
2977
               {
 
2978
                    E T92, T95, T9n, T9q;
 
2979
                    T92 = Tf - Tu;
 
2980
                    T95 = T93 - T94;
 
2981
                    T96 = T92 + T95;
 
2982
                    T9y = T92 - T95;
 
2983
                    T9n = TZ - TK;
 
2984
                    T9q = T9o - T9p;
 
2985
                    T9r = T9n + T9q;
 
2986
                    T9D = T9q - T9n;
 
2987
               }
 
2988
               {
 
2989
                    E T9b, T9g, T9k, T9l;
 
2990
                    T9b = T97 + T9a;
 
2991
                    T9g = T9c - T9f;
 
2992
                    T9h = KP707106781 * (T9b + T9g);
 
2993
                    T9C = KP707106781 * (T9b - T9g);
 
2994
                    T9k = T9a - T97;
 
2995
                    T9l = T9c + T9f;
 
2996
                    T9m = KP707106781 * (T9k + T9l);
 
2997
                    T9z = KP707106781 * (T9l - T9k);
 
2998
               }
 
2999
               {
 
3000
                    E T9i, T9s, T91, T9j;
 
3001
                    T9i = T96 + T9h;
 
3002
                    T9s = T9m + T9r;
 
3003
                    T91 = W[110];
 
3004
                    T9j = W[111];
 
3005
                    rio[WS(ios, 56)] = FNMS(T9j, T9s, T91 * T9i);
 
3006
                    iio[-WS(ios, 7)] = FMA(T9j, T9i, T91 * T9s);
 
3007
               }
 
3008
               {
 
3009
                    E T9G, T9I, T9F, T9H;
 
3010
                    T9G = T9y - T9z;
 
3011
                    T9I = T9D - T9C;
 
3012
                    T9F = W[78];
 
3013
                    T9H = W[79];
 
3014
                    rio[WS(ios, 40)] = FNMS(T9H, T9I, T9F * T9G);
 
3015
                    iio[-WS(ios, 23)] = FMA(T9H, T9G, T9F * T9I);
 
3016
               }
 
3017
               {
 
3018
                    E T9u, T9w, T9t, T9v;
 
3019
                    T9u = T96 - T9h;
 
3020
                    T9w = T9r - T9m;
 
3021
                    T9t = W[46];
 
3022
                    T9v = W[47];
 
3023
                    rio[WS(ios, 24)] = FNMS(T9v, T9w, T9t * T9u);
 
3024
                    iio[-WS(ios, 39)] = FMA(T9v, T9u, T9t * T9w);
 
3025
               }
 
3026
               {
 
3027
                    E T9A, T9E, T9x, T9B;
 
3028
                    T9A = T9y + T9z;
 
3029
                    T9E = T9C + T9D;
 
3030
                    T9x = W[14];
 
3031
                    T9B = W[15];
 
3032
                    rio[WS(ios, 8)] = FNMS(T9B, T9E, T9x * T9A);
 
3033
                    iio[-WS(ios, 55)] = FMA(T9B, T9A, T9x * T9E);
 
3034
               }
 
3035
          }
 
3036
          {
 
3037
               E T8u, T8Q, T8J, T8V, T8B, T8U, T8G, T8R;
 
3038
               {
 
3039
                    E T8s, T8t, T8H, T8I;
 
3040
                    T8s = T7i - T7l;
 
3041
                    T8t = KP707106781 * (T82 - T81);
 
3042
                    T8u = T8s + T8t;
 
3043
                    T8Q = T8s - T8t;
 
3044
                    T8H = KP707106781 * (T7r - T7w);
 
3045
                    T8I = T87 - T84;
 
3046
                    T8J = T8H + T8I;
 
3047
                    T8V = T8I - T8H;
 
3048
               }
 
3049
               {
 
3050
                    E T8x, T8A, T8E, T8F;
 
3051
                    T8x = FNMS(KP382683432, T8w, KP923879532 * T8v);
 
3052
                    T8A = FMA(KP923879532, T8y, KP382683432 * T8z);
 
3053
                    T8B = T8x + T8A;
 
3054
                    T8U = T8A - T8x;
 
3055
                    T8E = FMA(KP382683432, T8v, KP923879532 * T8w);
 
3056
                    T8F = FNMS(KP382683432, T8y, KP923879532 * T8z);
 
3057
                    T8G = T8E + T8F;
 
3058
                    T8R = T8E - T8F;
 
3059
               }
 
3060
               {
 
3061
                    E T8C, T8K, T8r, T8D;
 
3062
                    T8C = T8u + T8B;
 
3063
                    T8K = T8G + T8J;
 
3064
                    T8r = W[6];
 
3065
                    T8D = W[7];
 
3066
                    rio[WS(ios, 4)] = FNMS(T8D, T8K, T8r * T8C);
 
3067
                    iio[-WS(ios, 59)] = FMA(T8D, T8C, T8r * T8K);
 
3068
               }
 
3069
               {
 
3070
                    E T8Y, T90, T8X, T8Z;
 
3071
                    T8Y = T8Q - T8R;
 
3072
                    T90 = T8V - T8U;
 
3073
                    T8X = W[38];
 
3074
                    T8Z = W[39];
 
3075
                    rio[WS(ios, 20)] = FNMS(T8Z, T90, T8X * T8Y);
 
3076
                    iio[-WS(ios, 43)] = FMA(T8Z, T8Y, T8X * T90);
 
3077
               }
 
3078
               {
 
3079
                    E T8M, T8O, T8L, T8N;
 
3080
                    T8M = T8u - T8B;
 
3081
                    T8O = T8J - T8G;
 
3082
                    T8L = W[70];
 
3083
                    T8N = W[71];
 
3084
                    rio[WS(ios, 36)] = FNMS(T8N, T8O, T8L * T8M);
 
3085
                    iio[-WS(ios, 27)] = FMA(T8N, T8M, T8L * T8O);
 
3086
               }
 
3087
               {
 
3088
                    E T8S, T8W, T8P, T8T;
 
3089
                    T8S = T8Q + T8R;
 
3090
                    T8W = T8U + T8V;
 
3091
                    T8P = W[102];
 
3092
                    T8T = W[103];
 
3093
                    rio[WS(ios, 52)] = FNMS(T8T, T8W, T8P * T8S);
 
3094
                    iio[-WS(ios, 11)] = FMA(T8T, T8S, T8P * T8W);
 
3095
               }
 
3096
          }
 
3097
          {
 
3098
               E T7y, T8g, T89, T8l, T7V, T8k, T80, T8h;
 
3099
               {
 
3100
                    E T7m, T7x, T83, T88;
 
3101
                    T7m = T7i + T7l;
 
3102
                    T7x = KP707106781 * (T7r + T7w);
 
3103
                    T7y = T7m + T7x;
 
3104
                    T8g = T7m - T7x;
 
3105
                    T83 = KP707106781 * (T81 + T82);
 
3106
                    T88 = T84 + T87;
 
3107
                    T89 = T83 + T88;
 
3108
                    T8l = T88 - T83;
 
3109
               }
 
3110
               {
 
3111
                    E T7J, T7U, T7Y, T7Z;
 
3112
                    T7J = FMA(KP923879532, T7D, KP382683432 * T7I);
 
3113
                    T7U = FNMS(KP382683432, T7T, KP923879532 * T7O);
 
3114
                    T7V = T7J + T7U;
 
3115
                    T8k = T7J - T7U;
 
3116
                    T7Y = FNMS(KP382683432, T7D, KP923879532 * T7I);
 
3117
                    T7Z = FMA(KP382683432, T7O, KP923879532 * T7T);
 
3118
                    T80 = T7Y + T7Z;
 
3119
                    T8h = T7Z - T7Y;
 
3120
               }
 
3121
               {
 
3122
                    E T7W, T8a, T7h, T7X;
 
3123
                    T7W = T7y + T7V;
 
3124
                    T8a = T80 + T89;
 
3125
                    T7h = W[118];
 
3126
                    T7X = W[119];
 
3127
                    rio[WS(ios, 60)] = FNMS(T7X, T8a, T7h * T7W);
 
3128
                    iio[-WS(ios, 3)] = FMA(T7X, T7W, T7h * T8a);
 
3129
               }
 
3130
               {
 
3131
                    E T8o, T8q, T8n, T8p;
 
3132
                    T8o = T8g - T8h;
 
3133
                    T8q = T8l - T8k;
 
3134
                    T8n = W[86];
 
3135
                    T8p = W[87];
 
3136
                    rio[WS(ios, 44)] = FNMS(T8p, T8q, T8n * T8o);
 
3137
                    iio[-WS(ios, 19)] = FMA(T8p, T8o, T8n * T8q);
 
3138
               }
 
3139
               {
 
3140
                    E T8c, T8e, T8b, T8d;
 
3141
                    T8c = T7y - T7V;
 
3142
                    T8e = T89 - T80;
 
3143
                    T8b = W[54];
 
3144
                    T8d = W[55];
 
3145
                    rio[WS(ios, 28)] = FNMS(T8d, T8e, T8b * T8c);
 
3146
                    iio[-WS(ios, 35)] = FMA(T8d, T8c, T8b * T8e);
 
3147
               }
 
3148
               {
 
3149
                    E T8i, T8m, T8f, T8j;
 
3150
                    T8i = T8g + T8h;
 
3151
                    T8m = T8k + T8l;
 
3152
                    T8f = W[22];
 
3153
                    T8j = W[23];
 
3154
                    rio[WS(ios, 12)] = FNMS(T8j, T8m, T8f * T8i);
 
3155
                    iio[-WS(ios, 51)] = FMA(T8j, T8i, T8f * T8m);
 
3156
               }
 
3157
          }
 
3158
          {
 
3159
               E T6K, T76, T6Z, T7b, T6R, T7a, T6W, T77;
 
3160
               {
 
3161
                    E T6I, T6J, T6X, T6Y;
 
3162
                    T6I = T5O - T5P;
 
3163
                    T6J = T6j - T6k;
 
3164
                    T6K = T6I + T6J;
 
3165
                    T76 = T6I - T6J;
 
3166
                    T6X = T5W - T5T;
 
3167
                    T6Y = T6n - T6m;
 
3168
                    T6Z = T6X + T6Y;
 
3169
                    T7b = T6Y - T6X;
 
3170
                    {
 
3171
                         E T6N, T6U, T6Q, T6V;
 
3172
                         {
 
3173
                              E T6L, T6M, T6O, T6P;
 
3174
                              T6L = T5Z - T60;
 
3175
                              T6M = T63 - T62;
 
3176
                              T6N = FMA(KP831469612, T6L, KP555570233 * T6M);
 
3177
                              T6U = FNMS(KP555570233, T6L, KP831469612 * T6M);
 
3178
                              T6O = T66 - T67;
 
3179
                              T6P = T6a - T69;
 
3180
                              T6Q = FNMS(KP555570233, T6P, KP831469612 * T6O);
 
3181
                              T6V = FMA(KP555570233, T6O, KP831469612 * T6P);
 
3182
                         }
 
3183
                         T6R = T6N + T6Q;
 
3184
                         T7a = T6N - T6Q;
 
3185
                         T6W = T6U + T6V;
 
3186
                         T77 = T6V - T6U;
 
3187
                    }
 
3188
               }
 
3189
               {
 
3190
                    E T6S, T70, T6H, T6T;
 
3191
                    T6S = T6K + T6R;
 
3192
                    T70 = T6W + T6Z;
 
3193
                    T6H = W[114];
 
3194
                    T6T = W[115];
 
3195
                    rio[WS(ios, 58)] = FNMS(T6T, T70, T6H * T6S);
 
3196
                    iio[-WS(ios, 5)] = FMA(T6T, T6S, T6H * T70);
 
3197
               }
 
3198
               {
 
3199
                    E T7e, T7g, T7d, T7f;
 
3200
                    T7e = T76 - T77;
 
3201
                    T7g = T7b - T7a;
 
3202
                    T7d = W[82];
 
3203
                    T7f = W[83];
 
3204
                    rio[WS(ios, 42)] = FNMS(T7f, T7g, T7d * T7e);
 
3205
                    iio[-WS(ios, 21)] = FMA(T7f, T7e, T7d * T7g);
 
3206
               }
 
3207
               {
 
3208
                    E T72, T74, T71, T73;
 
3209
                    T72 = T6K - T6R;
 
3210
                    T74 = T6Z - T6W;
 
3211
                    T71 = W[50];
 
3212
                    T73 = W[51];
 
3213
                    rio[WS(ios, 26)] = FNMS(T73, T74, T71 * T72);
 
3214
                    iio[-WS(ios, 37)] = FMA(T73, T72, T71 * T74);
 
3215
               }
 
3216
               {
 
3217
                    E T78, T7c, T75, T79;
 
3218
                    T78 = T76 + T77;
 
3219
                    T7c = T7a + T7b;
 
3220
                    T75 = W[18];
 
3221
                    T79 = W[19];
 
3222
                    rio[WS(ios, 10)] = FNMS(T79, T7c, T75 * T78);
 
3223
                    iio[-WS(ios, 53)] = FMA(T79, T78, T75 * T7c);
 
3224
               }
 
3225
          }
 
3226
          {
 
3227
               E T3a, T52, T4V, T57, T4z, T56, T4E, T53;
 
3228
               {
 
3229
                    E T2w, T39, T4H, T4U;
 
3230
                    T2w = T2c + T2v;
 
3231
                    T39 = T2P + T38;
 
3232
                    T3a = T2w + T39;
 
3233
                    T52 = T2w - T39;
 
3234
                    T4H = T4F + T4G;
 
3235
                    T4U = T4K + T4T;
 
3236
                    T4V = T4H + T4U;
 
3237
                    T57 = T4U - T4H;
 
3238
                    {
 
3239
                         E T3R, T4C, T4y, T4D;
 
3240
                         {
 
3241
                              E T3D, T3Q, T4k, T4x;
 
3242
                              T3D = T3j + T3C;
 
3243
                              T3Q = T3G + T3P;
 
3244
                              T3R = FMA(KP980785280, T3D, KP195090322 * T3Q);
 
3245
                              T4C = FNMS(KP195090322, T3D, KP980785280 * T3Q);
 
3246
                              T4k = T40 + T4j;
 
3247
                              T4x = T4n + T4w;
 
3248
                              T4y = FNMS(KP195090322, T4x, KP980785280 * T4k);
 
3249
                              T4D = FMA(KP195090322, T4k, KP980785280 * T4x);
 
3250
                         }
 
3251
                         T4z = T3R + T4y;
 
3252
                         T56 = T3R - T4y;
 
3253
                         T4E = T4C + T4D;
 
3254
                         T53 = T4D - T4C;
 
3255
                    }
 
3256
               }
 
3257
               {
 
3258
                    E T4A, T4W, T23, T4B;
 
3259
                    T4A = T3a + T4z;
 
3260
                    T4W = T4E + T4V;
 
3261
                    T23 = W[122];
 
3262
                    T4B = W[123];
 
3263
                    rio[WS(ios, 62)] = FNMS(T4B, T4W, T23 * T4A);
 
3264
                    iio[-WS(ios, 1)] = FMA(T4B, T4A, T23 * T4W);
 
3265
               }
 
3266
               {
 
3267
                    E T5a, T5c, T59, T5b;
 
3268
                    T5a = T52 - T53;
 
3269
                    T5c = T57 - T56;
 
3270
                    T59 = W[90];
 
3271
                    T5b = W[91];
 
3272
                    rio[WS(ios, 46)] = FNMS(T5b, T5c, T59 * T5a);
 
3273
                    iio[-WS(ios, 17)] = FMA(T5b, T5a, T59 * T5c);
 
3274
               }
 
3275
               {
 
3276
                    E T4Y, T50, T4X, T4Z;
 
3277
                    T4Y = T3a - T4z;
 
3278
                    T50 = T4V - T4E;
 
3279
                    T4X = W[58];
 
3280
                    T4Z = W[59];
 
3281
                    rio[WS(ios, 30)] = FNMS(T4Z, T50, T4X * T4Y);
 
3282
                    iio[-WS(ios, 33)] = FMA(T4Z, T4Y, T4X * T50);
 
3283
               }
 
3284
               {
 
3285
                    E T54, T58, T51, T55;
 
3286
                    T54 = T52 + T53;
 
3287
                    T58 = T56 + T57;
 
3288
                    T51 = W[26];
 
3289
                    T55 = W[27];
 
3290
                    rio[WS(ios, 14)] = FNMS(T55, T58, T51 * T54);
 
3291
                    iio[-WS(ios, 49)] = FMA(T55, T54, T51 * T58);
 
3292
               }
 
3293
          }
 
3294
          {
 
3295
               E T5g, T5C, T5v, T5H, T5n, T5G, T5s, T5D;
 
3296
               {
 
3297
                    E T5e, T5f, T5t, T5u;
 
3298
                    T5e = T2c - T2v;
 
3299
                    T5f = T4G - T4F;
 
3300
                    T5g = T5e + T5f;
 
3301
                    T5C = T5e - T5f;
 
3302
                    T5t = T2P - T38;
 
3303
                    T5u = T4T - T4K;
 
3304
                    T5v = T5t + T5u;
 
3305
                    T5H = T5u - T5t;
 
3306
                    {
 
3307
                         E T5j, T5q, T5m, T5r;
 
3308
                         {
 
3309
                              E T5h, T5i, T5k, T5l;
 
3310
                              T5h = T3j - T3C;
 
3311
                              T5i = T3P - T3G;
 
3312
                              T5j = FNMS(KP555570233, T5i, KP831469612 * T5h);
 
3313
                              T5q = FMA(KP555570233, T5h, KP831469612 * T5i);
 
3314
                              T5k = T40 - T4j;
 
3315
                              T5l = T4w - T4n;
 
3316
                              T5m = FMA(KP831469612, T5k, KP555570233 * T5l);
 
3317
                              T5r = FNMS(KP555570233, T5k, KP831469612 * T5l);
 
3318
                         }
 
3319
                         T5n = T5j + T5m;
 
3320
                         T5G = T5m - T5j;
 
3321
                         T5s = T5q + T5r;
 
3322
                         T5D = T5q - T5r;
 
3323
                    }
 
3324
               }
 
3325
               {
 
3326
                    E T5o, T5w, T5d, T5p;
 
3327
                    T5o = T5g + T5n;
 
3328
                    T5w = T5s + T5v;
 
3329
                    T5d = W[10];
 
3330
                    T5p = W[11];
 
3331
                    rio[WS(ios, 6)] = FNMS(T5p, T5w, T5d * T5o);
 
3332
                    iio[-WS(ios, 57)] = FMA(T5p, T5o, T5d * T5w);
 
3333
               }
 
3334
               {
 
3335
                    E T5K, T5M, T5J, T5L;
 
3336
                    T5K = T5C - T5D;
 
3337
                    T5M = T5H - T5G;
 
3338
                    T5J = W[42];
 
3339
                    T5L = W[43];
 
3340
                    rio[WS(ios, 22)] = FNMS(T5L, T5M, T5J * T5K);
 
3341
                    iio[-WS(ios, 41)] = FMA(T5L, T5K, T5J * T5M);
 
3342
               }
 
3343
               {
 
3344
                    E T5y, T5A, T5x, T5z;
 
3345
                    T5y = T5g - T5n;
 
3346
                    T5A = T5v - T5s;
 
3347
                    T5x = W[74];
 
3348
                    T5z = W[75];
 
3349
                    rio[WS(ios, 38)] = FNMS(T5z, T5A, T5x * T5y);
 
3350
                    iio[-WS(ios, 25)] = FMA(T5z, T5y, T5x * T5A);
 
3351
               }
 
3352
               {
 
3353
                    E T5E, T5I, T5B, T5F;
 
3354
                    T5E = T5C + T5D;
 
3355
                    T5I = T5G + T5H;
 
3356
                    T5B = W[106];
 
3357
                    T5F = W[107];
 
3358
                    rio[WS(ios, 54)] = FNMS(T5F, T5I, T5B * T5E);
 
3359
                    iio[-WS(ios, 9)] = FMA(T5F, T5E, T5B * T5I);
 
3360
               }
 
3361
          }
 
3362
          {
 
3363
               E T5Y, T6w, T6p, T6B, T6d, T6A, T6i, T6x;
 
3364
               {
 
3365
                    E T5Q, T5X, T6l, T6o;
 
3366
                    T5Q = T5O + T5P;
 
3367
                    T5X = T5T + T5W;
 
3368
                    T5Y = T5Q + T5X;
 
3369
                    T6w = T5Q - T5X;
 
3370
                    T6l = T6j + T6k;
 
3371
                    T6o = T6m + T6n;
 
3372
                    T6p = T6l + T6o;
 
3373
                    T6B = T6o - T6l;
 
3374
                    {
 
3375
                         E T65, T6g, T6c, T6h;
 
3376
                         {
 
3377
                              E T61, T64, T68, T6b;
 
3378
                              T61 = T5Z + T60;
 
3379
                              T64 = T62 + T63;
 
3380
                              T65 = FNMS(KP195090322, T64, KP980785280 * T61);
 
3381
                              T6g = FMA(KP195090322, T61, KP980785280 * T64);
 
3382
                              T68 = T66 + T67;
 
3383
                              T6b = T69 + T6a;
 
3384
                              T6c = FMA(KP980785280, T68, KP195090322 * T6b);
 
3385
                              T6h = FNMS(KP195090322, T68, KP980785280 * T6b);
 
3386
                         }
 
3387
                         T6d = T65 + T6c;
 
3388
                         T6A = T6c - T65;
 
3389
                         T6i = T6g + T6h;
 
3390
                         T6x = T6g - T6h;
 
3391
                    }
 
3392
               }
 
3393
               {
 
3394
                    E T6e, T6q, T5N, T6f;
 
3395
                    T6e = T5Y + T6d;
 
3396
                    T6q = T6i + T6p;
 
3397
                    T5N = W[2];
 
3398
                    T6f = W[3];
 
3399
                    rio[WS(ios, 2)] = FNMS(T6f, T6q, T5N * T6e);
 
3400
                    iio[-WS(ios, 61)] = FMA(T6f, T6e, T5N * T6q);
 
3401
               }
 
3402
               {
 
3403
                    E T6E, T6G, T6D, T6F;
 
3404
                    T6E = T6w - T6x;
 
3405
                    T6G = T6B - T6A;
 
3406
                    T6D = W[34];
 
3407
                    T6F = W[35];
 
3408
                    rio[WS(ios, 18)] = FNMS(T6F, T6G, T6D * T6E);
 
3409
                    iio[-WS(ios, 45)] = FMA(T6F, T6E, T6D * T6G);
 
3410
               }
 
3411
               {
 
3412
                    E T6s, T6u, T6r, T6t;
 
3413
                    T6s = T5Y - T6d;
 
3414
                    T6u = T6p - T6i;
 
3415
                    T6r = W[66];
 
3416
                    T6t = W[67];
 
3417
                    rio[WS(ios, 34)] = FNMS(T6t, T6u, T6r * T6s);
 
3418
                    iio[-WS(ios, 29)] = FMA(T6t, T6s, T6r * T6u);
 
3419
               }
 
3420
               {
 
3421
                    E T6y, T6C, T6v, T6z;
 
3422
                    T6y = T6w + T6x;
 
3423
                    T6C = T6A + T6B;
 
3424
                    T6v = W[98];
 
3425
                    T6z = W[99];
 
3426
                    rio[WS(ios, 50)] = FNMS(T6z, T6C, T6v * T6y);
 
3427
                    iio[-WS(ios, 13)] = FMA(T6z, T6y, T6v * T6C);
 
3428
               }
 
3429
          }
 
3430
          {
 
3431
               E TdO, Tf1, Teq, TeH, Tef, TeW, Ten, TeM, Te3, Ter, Te8, Tem, TeE, Tf0, TeP;
 
3432
               E TeX;
 
3433
               {
 
3434
                    E TdG, TeG, TdN, TeF, TdJ, TdM;
 
3435
                    TdG = TdE + TdF;
 
3436
                    TeG = Ted - Tec;
 
3437
                    TdJ = FNMS(KP555570233, TdI, KP831469612 * TdH);
 
3438
                    TdM = FMA(KP831469612, TdK, KP555570233 * TdL);
 
3439
                    TdN = TdJ + TdM;
 
3440
                    TeF = TdM - TdJ;
 
3441
                    TdO = TdG + TdN;
 
3442
                    Tf1 = TeG - TeF;
 
3443
                    Teq = TdG - TdN;
 
3444
                    TeH = TeF + TeG;
 
3445
               }
 
3446
               {
 
3447
                    E Tee, TeK, Teb, TeL, Te9, Tea;
 
3448
                    Tee = Tec + Ted;
 
3449
                    TeK = TdE - TdF;
 
3450
                    Te9 = FMA(KP555570233, TdH, KP831469612 * TdI);
 
3451
                    Tea = FNMS(KP555570233, TdK, KP831469612 * TdL);
 
3452
                    Teb = Te9 + Tea;
 
3453
                    TeL = Te9 - Tea;
 
3454
                    Tef = Teb + Tee;
 
3455
                    TeW = TeK - TeL;
 
3456
                    Ten = Tee - Teb;
 
3457
                    TeM = TeK + TeL;
 
3458
               }
 
3459
               {
 
3460
                    E TdV, Te6, Te2, Te7;
 
3461
                    {
 
3462
                         E TdR, TdU, TdY, Te1;
 
3463
                         TdR = TdP + TdQ;
 
3464
                         TdU = TdS + TdT;
 
3465
                         TdV = FNMS(KP290284677, TdU, KP956940335 * TdR);
 
3466
                         Te6 = FMA(KP290284677, TdR, KP956940335 * TdU);
 
3467
                         TdY = TdW + TdX;
 
3468
                         Te1 = TdZ + Te0;
 
3469
                         Te2 = FMA(KP956940335, TdY, KP290284677 * Te1);
 
3470
                         Te7 = FNMS(KP290284677, TdY, KP956940335 * Te1);
 
3471
                    }
 
3472
                    Te3 = TdV + Te2;
 
3473
                    Ter = Te6 - Te7;
 
3474
                    Te8 = Te6 + Te7;
 
3475
                    Tem = Te2 - TdV;
 
3476
               }
 
3477
               {
 
3478
                    E TeA, TeN, TeD, TeO;
 
3479
                    {
 
3480
                         E Tey, Tez, TeB, TeC;
 
3481
                         Tey = TdT - TdS;
 
3482
                         Tez = TdP - TdQ;
 
3483
                         TeA = FNMS(KP471396736, Tez, KP881921264 * Tey);
 
3484
                         TeN = FMA(KP881921264, Tez, KP471396736 * Tey);
 
3485
                         TeB = TdW - TdX;
 
3486
                         TeC = Te0 - TdZ;
 
3487
                         TeD = FMA(KP471396736, TeB, KP881921264 * TeC);
 
3488
                         TeO = FNMS(KP471396736, TeC, KP881921264 * TeB);
 
3489
                    }
 
3490
                    TeE = TeA + TeD;
 
3491
                    Tf0 = TeN - TeO;
 
3492
                    TeP = TeN + TeO;
 
3493
                    TeX = TeD - TeA;
 
3494
               }
 
3495
               {
 
3496
                    E Te4, Teg, TdD, Te5;
 
3497
                    Te4 = TdO + Te3;
 
3498
                    Teg = Te8 + Tef;
 
3499
                    TdD = W[120];
 
3500
                    Te5 = W[121];
 
3501
                    iio[-WS(ios, 2)] = FMA(TdD, Te4, Te5 * Teg);
 
3502
                    rio[WS(ios, 61)] = FNMS(Te5, Te4, TdD * Teg);
 
3503
               }
 
3504
               {
 
3505
                    E TeY, Tf2, TeV, TeZ;
 
3506
                    TeY = TeW + TeX;
 
3507
                    Tf2 = Tf0 + Tf1;
 
3508
                    TeV = W[104];
 
3509
                    TeZ = W[105];
 
3510
                    iio[-WS(ios, 10)] = FMA(TeV, TeY, TeZ * Tf2);
 
3511
                    rio[WS(ios, 53)] = FNMS(TeZ, TeY, TeV * Tf2);
 
3512
               }
 
3513
               {
 
3514
                    E Tf4, Tf6, Tf3, Tf5;
 
3515
                    Tf4 = Tf1 - Tf0;
 
3516
                    Tf6 = TeW - TeX;
 
3517
                    Tf3 = W[40];
 
3518
                    Tf5 = W[41];
 
3519
                    rio[WS(ios, 21)] = FNMS(Tf5, Tf6, Tf3 * Tf4);
 
3520
                    iio[-WS(ios, 42)] = FMA(Tf3, Tf6, Tf5 * Tf4);
 
3521
               }
 
3522
               {
 
3523
                    E Tei, Tek, Teh, Tej;
 
3524
                    Tei = Tef - Te8;
 
3525
                    Tek = TdO - Te3;
 
3526
                    Teh = W[56];
 
3527
                    Tej = W[57];
 
3528
                    rio[WS(ios, 29)] = FNMS(Tej, Tek, Teh * Tei);
 
3529
                    iio[-WS(ios, 34)] = FMA(Teh, Tek, Tej * Tei);
 
3530
               }
 
3531
               {
 
3532
                    E Teo, Tes, Tel, Tep;
 
3533
                    Teo = Tem + Ten;
 
3534
                    Tes = Teq + Ter;
 
3535
                    Tel = W[24];
 
3536
                    Tep = W[25];
 
3537
                    rio[WS(ios, 13)] = FNMS(Tep, Tes, Tel * Teo);
 
3538
                    iio[-WS(ios, 50)] = FMA(Tel, Tes, Tep * Teo);
 
3539
               }
 
3540
               {
 
3541
                    E TeI, TeQ, Tex, TeJ;
 
3542
                    TeI = TeE + TeH;
 
3543
                    TeQ = TeM + TeP;
 
3544
                    Tex = W[8];
 
3545
                    TeJ = W[9];
 
3546
                    rio[WS(ios, 5)] = FNMS(TeJ, TeQ, Tex * TeI);
 
3547
                    iio[-WS(ios, 58)] = FMA(Tex, TeQ, TeJ * TeI);
 
3548
               }
 
3549
               {
 
3550
                    E TeS, TeU, TeR, TeT;
 
3551
                    TeS = TeM - TeP;
 
3552
                    TeU = TeH - TeE;
 
3553
                    TeR = W[72];
 
3554
                    TeT = W[73];
 
3555
                    iio[-WS(ios, 26)] = FMA(TeR, TeS, TeT * TeU);
 
3556
                    rio[WS(ios, 37)] = FNMS(TeT, TeS, TeR * TeU);
 
3557
               }
 
3558
               {
 
3559
                    E Teu, Tew, Tet, Tev;
 
3560
                    Teu = Teq - Ter;
 
3561
                    Tew = Ten - Tem;
 
3562
                    Tet = W[88];
 
3563
                    Tev = W[89];
 
3564
                    iio[-WS(ios, 18)] = FMA(Tet, Teu, Tev * Tew);
 
3565
                    rio[WS(ios, 45)] = FNMS(Tev, Teu, Tet * Tew);
 
3566
               }
 
3567
          }
 
3568
          {
 
3569
               E Tcr, Tdw, TcX, Td6, TcI, Tdt, TcS, Tdl, Tbm, TcW, TcL, TcT, Tdd, Tdx, Tdi;
 
3570
               E Tds;
 
3571
               {
 
3572
                    E Tcq, Td4, TbZ, Td5, TbF, TbY;
 
3573
                    Tcq = Tce + Tcp;
 
3574
                    Td4 = TcA - TcD;
 
3575
                    TbF = FNMS(KP195090322, TbE, KP980785280 * Tbx);
 
3576
                    TbY = FMA(KP195090322, TbQ, KP980785280 * TbX);
 
3577
                    TbZ = TbF + TbY;
 
3578
                    Td5 = TbY - TbF;
 
3579
                    Tcr = TbZ + Tcq;
 
3580
                    Tdw = Td4 - Td5;
 
3581
                    TcX = Tcq - TbZ;
 
3582
                    Td6 = Td4 + Td5;
 
3583
               }
 
3584
               {
 
3585
                    E TcE, Tdk, TcH, Tdj, TcF, TcG;
 
3586
                    TcE = TcA + TcD;
 
3587
                    Tdk = Tcp - Tce;
 
3588
                    TcF = FMA(KP980785280, TbE, KP195090322 * Tbx);
 
3589
                    TcG = FNMS(KP195090322, TbX, KP980785280 * TbQ);
 
3590
                    TcH = TcF + TcG;
 
3591
                    Tdj = TcF - TcG;
 
3592
                    TcI = TcE + TcH;
 
3593
                    Tdt = Tdk - Tdj;
 
3594
                    TcS = TcE - TcH;
 
3595
                    Tdl = Tdj + Tdk;
 
3596
               }
 
3597
               {
 
3598
                    E TaI, TcJ, Tbl, TcK;
 
3599
                    {
 
3600
                         E Taw, TaH, Tb9, Tbk;
 
3601
                         Taw = Tak + Tav;
 
3602
                         TaH = TaD + TaG;
 
3603
                         TaI = FNMS(KP098017140, TaH, KP995184726 * Taw);
 
3604
                         TcJ = FMA(KP995184726, TaH, KP098017140 * Taw);
 
3605
                         Tb9 = TaT + Tb8;
 
3606
                         Tbk = Tbc + Tbj;
 
3607
                         Tbl = FMA(KP098017140, Tb9, KP995184726 * Tbk);
 
3608
                         TcK = FNMS(KP098017140, Tbk, KP995184726 * Tb9);
 
3609
                    }
 
3610
                    Tbm = TaI + Tbl;
 
3611
                    TcW = TcJ - TcK;
 
3612
                    TcL = TcJ + TcK;
 
3613
                    TcT = Tbl - TaI;
 
3614
               }
 
3615
               {
 
3616
                    E Td9, Tdg, Tdc, Tdh;
 
3617
                    {
 
3618
                         E Td7, Td8, Tda, Tdb;
 
3619
                         Td7 = TaD - TaG;
 
3620
                         Td8 = Tav - Tak;
 
3621
                         Td9 = FNMS(KP634393284, Td8, KP773010453 * Td7);
 
3622
                         Tdg = FMA(KP634393284, Td7, KP773010453 * Td8);
 
3623
                         Tda = TaT - Tb8;
 
3624
                         Tdb = Tbj - Tbc;
 
3625
                         Tdc = FMA(KP773010453, Tda, KP634393284 * Tdb);
 
3626
                         Tdh = FNMS(KP634393284, Tda, KP773010453 * Tdb);
 
3627
                    }
 
3628
                    Tdd = Td9 + Tdc;
 
3629
                    Tdx = Tdg - Tdh;
 
3630
                    Tdi = Tdg + Tdh;
 
3631
                    Tds = Tdc - Td9;
 
3632
               }
 
3633
               {
 
3634
                    E Tcs, TcM, Ta5, Tct;
 
3635
                    Tcs = Tbm + Tcr;
 
3636
                    TcM = TcI + TcL;
 
3637
                    Ta5 = W[0];
 
3638
                    Tct = W[1];
 
3639
                    rio[WS(ios, 1)] = FNMS(Tct, TcM, Ta5 * Tcs);
 
3640
                    iio[-WS(ios, 62)] = FMA(Ta5, TcM, Tct * Tcs);
 
3641
               }
 
3642
               {
 
3643
                    E Tdu, Tdy, Tdr, Tdv;
 
3644
                    Tdu = Tds + Tdt;
 
3645
                    Tdy = Tdw + Tdx;
 
3646
                    Tdr = W[16];
 
3647
                    Tdv = W[17];
 
3648
                    rio[WS(ios, 9)] = FNMS(Tdv, Tdy, Tdr * Tdu);
 
3649
                    iio[-WS(ios, 54)] = FMA(Tdr, Tdy, Tdv * Tdu);
 
3650
               }
 
3651
               {
 
3652
                    E TdA, TdC, Tdz, TdB;
 
3653
                    TdA = Tdw - Tdx;
 
3654
                    TdC = Tdt - Tds;
 
3655
                    Tdz = W[80];
 
3656
                    TdB = W[81];
 
3657
                    iio[-WS(ios, 22)] = FMA(Tdz, TdA, TdB * TdC);
 
3658
                    rio[WS(ios, 41)] = FNMS(TdB, TdA, Tdz * TdC);
 
3659
               }
 
3660
               {
 
3661
                    E TcO, TcQ, TcN, TcP;
 
3662
                    TcO = TcI - TcL;
 
3663
                    TcQ = Tcr - Tbm;
 
3664
                    TcN = W[64];
 
3665
                    TcP = W[65];
 
3666
                    iio[-WS(ios, 30)] = FMA(TcN, TcO, TcP * TcQ);
 
3667
                    rio[WS(ios, 33)] = FNMS(TcP, TcO, TcN * TcQ);
 
3668
               }
 
3669
               {
 
3670
                    E TcU, TcY, TcR, TcV;
 
3671
                    TcU = TcS + TcT;
 
3672
                    TcY = TcW + TcX;
 
3673
                    TcR = W[96];
 
3674
                    TcV = W[97];
 
3675
                    iio[-WS(ios, 14)] = FMA(TcR, TcU, TcV * TcY);
 
3676
                    rio[WS(ios, 49)] = FNMS(TcV, TcU, TcR * TcY);
 
3677
               }
 
3678
               {
 
3679
                    E Tde, Tdm, Td3, Tdf;
 
3680
                    Tde = Td6 + Tdd;
 
3681
                    Tdm = Tdi + Tdl;
 
3682
                    Td3 = W[112];
 
3683
                    Tdf = W[113];
 
3684
                    iio[-WS(ios, 6)] = FMA(Td3, Tde, Tdf * Tdm);
 
3685
                    rio[WS(ios, 57)] = FNMS(Tdf, Tde, Td3 * Tdm);
 
3686
               }
 
3687
               {
 
3688
                    E Tdo, Tdq, Tdn, Tdp;
 
3689
                    Tdo = Tdl - Tdi;
 
3690
                    Tdq = Td6 - Tdd;
 
3691
                    Tdn = W[48];
 
3692
                    Tdp = W[49];
 
3693
                    rio[WS(ios, 25)] = FNMS(Tdp, Tdq, Tdn * Tdo);
 
3694
                    iio[-WS(ios, 38)] = FMA(Tdn, Tdq, Tdp * Tdo);
 
3695
               }
 
3696
               {
 
3697
                    E Td0, Td2, TcZ, Td1;
 
3698
                    Td0 = TcX - TcW;
 
3699
                    Td2 = TcS - TcT;
 
3700
                    TcZ = W[32];
 
3701
                    Td1 = W[33];
 
3702
                    rio[WS(ios, 17)] = FNMS(Td1, Td2, TcZ * Td0);
 
3703
                    iio[-WS(ios, 46)] = FMA(TcZ, Td2, Td1 * Td0);
 
3704
               }
 
3705
          }
 
3706
          {
 
3707
               E Tfy, Thd, TgC, TgT, Tgr, Th8, Tgz, TgY, Tgb, TgD, Tgg, Tgy, TgQ, Thc, Th1;
 
3708
               E Th9;
 
3709
               {
 
3710
                    E Tfi, TgS, Tfx, TgR, Tfp, Tfw;
 
3711
                    Tfi = Tfa + Tfh;
 
3712
                    TgS = Tgp - Tgm;
 
3713
                    Tfp = FNMS(KP195090322, Tfo, KP980785280 * Tfl);
 
3714
                    Tfw = FMA(KP980785280, Tfs, KP195090322 * Tfv);
 
3715
                    Tfx = Tfp + Tfw;
 
3716
                    TgR = Tfw - Tfp;
 
3717
                    Tfy = Tfi + Tfx;
 
3718
                    Thd = TgS - TgR;
 
3719
                    TgC = Tfi - Tfx;
 
3720
                    TgT = TgR + TgS;
 
3721
               }
 
3722
               {
 
3723
                    E Tgq, TgW, Tgj, TgX, Tgh, Tgi;
 
3724
                    Tgq = Tgm + Tgp;
 
3725
                    TgW = Tfa - Tfh;
 
3726
                    Tgh = FMA(KP195090322, Tfl, KP980785280 * Tfo);
 
3727
                    Tgi = FNMS(KP195090322, Tfs, KP980785280 * Tfv);
 
3728
                    Tgj = Tgh + Tgi;
 
3729
                    TgX = Tgh - Tgi;
 
3730
                    Tgr = Tgj + Tgq;
 
3731
                    Th8 = TgW - TgX;
 
3732
                    Tgz = Tgq - Tgj;
 
3733
                    TgY = TgW + TgX;
 
3734
               }
 
3735
               {
 
3736
                    E TfR, Tge, Tga, Tgf;
 
3737
                    {
 
3738
                         E TfJ, TfQ, Tg2, Tg9;
 
3739
                         TfJ = TfB + TfI;
 
3740
                         TfQ = TfM + TfP;
 
3741
                         TfR = FNMS(KP098017140, TfQ, KP995184726 * TfJ);
 
3742
                         Tge = FMA(KP098017140, TfJ, KP995184726 * TfQ);
 
3743
                         Tg2 = TfU + Tg1;
 
3744
                         Tg9 = Tg5 + Tg8;
 
3745
                         Tga = FMA(KP995184726, Tg2, KP098017140 * Tg9);
 
3746
                         Tgf = FNMS(KP098017140, Tg2, KP995184726 * Tg9);
 
3747
                    }
 
3748
                    Tgb = TfR + Tga;
 
3749
                    TgD = Tge - Tgf;
 
3750
                    Tgg = Tge + Tgf;
 
3751
                    Tgy = Tga - TfR;
 
3752
               }
 
3753
               {
 
3754
                    E TgM, TgZ, TgP, Th0;
 
3755
                    {
 
3756
                         E TgK, TgL, TgN, TgO;
 
3757
                         TgK = TfP - TfM;
 
3758
                         TgL = TfB - TfI;
 
3759
                         TgM = FNMS(KP634393284, TgL, KP773010453 * TgK);
 
3760
                         TgZ = FMA(KP773010453, TgL, KP634393284 * TgK);
 
3761
                         TgN = TfU - Tg1;
 
3762
                         TgO = Tg8 - Tg5;
 
3763
                         TgP = FMA(KP634393284, TgN, KP773010453 * TgO);
 
3764
                         Th0 = FNMS(KP634393284, TgO, KP773010453 * TgN);
 
3765
                    }
 
3766
                    TgQ = TgM + TgP;
 
3767
                    Thc = TgZ - Th0;
 
3768
                    Th1 = TgZ + Th0;
 
3769
                    Th9 = TgP - TgM;
 
3770
               }
 
3771
               {
 
3772
                    E Tgc, Tgs, Tf7, Tgd;
 
3773
                    Tgc = Tfy + Tgb;
 
3774
                    Tgs = Tgg + Tgr;
 
3775
                    Tf7 = W[124];
 
3776
                    Tgd = W[125];
 
3777
                    iio[0] = FMA(Tf7, Tgc, Tgd * Tgs);
 
3778
                    rio[WS(ios, 63)] = FNMS(Tgd, Tgc, Tf7 * Tgs);
 
3779
               }
 
3780
               {
 
3781
                    E Tha, The, Th7, Thb;
 
3782
                    Tha = Th8 + Th9;
 
3783
                    The = Thc + Thd;
 
3784
                    Th7 = W[108];
 
3785
                    Thb = W[109];
 
3786
                    iio[-WS(ios, 8)] = FMA(Th7, Tha, Thb * The);
 
3787
                    rio[WS(ios, 55)] = FNMS(Thb, Tha, Th7 * The);
 
3788
               }
 
3789
               {
 
3790
                    E Thg, Thi, Thf, Thh;
 
3791
                    Thg = Thd - Thc;
 
3792
                    Thi = Th8 - Th9;
 
3793
                    Thf = W[44];
 
3794
                    Thh = W[45];
 
3795
                    rio[WS(ios, 23)] = FNMS(Thh, Thi, Thf * Thg);
 
3796
                    iio[-WS(ios, 40)] = FMA(Thf, Thi, Thh * Thg);
 
3797
               }
 
3798
               {
 
3799
                    E Tgu, Tgw, Tgt, Tgv;
 
3800
                    Tgu = Tgr - Tgg;
 
3801
                    Tgw = Tfy - Tgb;
 
3802
                    Tgt = W[60];
 
3803
                    Tgv = W[61];
 
3804
                    rio[WS(ios, 31)] = FNMS(Tgv, Tgw, Tgt * Tgu);
 
3805
                    iio[-WS(ios, 32)] = FMA(Tgt, Tgw, Tgv * Tgu);
 
3806
               }
 
3807
               {
 
3808
                    E TgA, TgE, Tgx, TgB;
 
3809
                    TgA = Tgy + Tgz;
 
3810
                    TgE = TgC + TgD;
 
3811
                    Tgx = W[28];
 
3812
                    TgB = W[29];
 
3813
                    rio[WS(ios, 15)] = FNMS(TgB, TgE, Tgx * TgA);
 
3814
                    iio[-WS(ios, 48)] = FMA(Tgx, TgE, TgB * TgA);
 
3815
               }
 
3816
               {
 
3817
                    E TgU, Th2, TgJ, TgV;
 
3818
                    TgU = TgQ + TgT;
 
3819
                    Th2 = TgY + Th1;
 
3820
                    TgJ = W[12];
 
3821
                    TgV = W[13];
 
3822
                    rio[WS(ios, 7)] = FNMS(TgV, Th2, TgJ * TgU);
 
3823
                    iio[-WS(ios, 56)] = FMA(TgJ, Th2, TgV * TgU);
 
3824
               }
 
3825
               {
 
3826
                    E Th4, Th6, Th3, Th5;
 
3827
                    Th4 = TgY - Th1;
 
3828
                    Th6 = TgT - TgQ;
 
3829
                    Th3 = W[76];
 
3830
                    Th5 = W[77];
 
3831
                    iio[-WS(ios, 24)] = FMA(Th3, Th4, Th5 * Th6);
 
3832
                    rio[WS(ios, 39)] = FNMS(Th5, Th4, Th3 * Th6);
 
3833
               }
 
3834
               {
 
3835
                    E TgG, TgI, TgF, TgH;
 
3836
                    TgG = TgC - TgD;
 
3837
                    TgI = Tgz - Tgy;
 
3838
                    TgF = W[92];
 
3839
                    TgH = W[93];
 
3840
                    iio[-WS(ios, 16)] = FMA(TgF, TgG, TgH * TgI);
 
3841
                    rio[WS(ios, 47)] = FNMS(TgH, TgG, TgF * TgI);
 
3842
               }
 
3843
          }
 
3844
          {
 
3845
               E ThJ, TiG, Ti7, Tig, ThS, TiD, Ti2, Tiv, Thy, Ti6, ThV, Ti3, Tin, TiH, Tis;
 
3846
               E TiC;
 
3847
               {
 
3848
                    E ThI, Tie, ThF, Tif, ThB, ThE;
 
3849
                    ThI = ThG + ThH;
 
3850
                    Tie = ThM - ThN;
 
3851
                    ThB = FNMS(KP555570233, ThA, KP831469612 * Thz);
 
3852
                    ThE = FNMS(KP555570233, ThD, KP831469612 * ThC);
 
3853
                    ThF = ThB + ThE;
 
3854
                    Tif = ThE - ThB;
 
3855
                    ThJ = ThF + ThI;
 
3856
                    TiG = Tie - Tif;
 
3857
                    Ti7 = ThI - ThF;
 
3858
                    Tig = Tie + Tif;
 
3859
               }
 
3860
               {
 
3861
                    E ThO, Tiu, ThR, Tit, ThP, ThQ;
 
3862
                    ThO = ThM + ThN;
 
3863
                    Tiu = ThH - ThG;
 
3864
                    ThP = FMA(KP831469612, ThA, KP555570233 * Thz);
 
3865
                    ThQ = FMA(KP831469612, ThD, KP555570233 * ThC);
 
3866
                    ThR = ThP - ThQ;
 
3867
                    Tit = ThP + ThQ;
 
3868
                    ThS = ThO + ThR;
 
3869
                    TiD = Tiu - Tit;
 
3870
                    Ti2 = ThO - ThR;
 
3871
                    Tiv = Tit + Tiu;
 
3872
               }
 
3873
               {
 
3874
                    E Thq, ThT, Thx, ThU;
 
3875
                    {
 
3876
                         E Thm, Thp, Tht, Thw;
 
3877
                         Thm = Thk + Thl;
 
3878
                         Thp = Thn + Tho;
 
3879
                         Thq = FNMS(KP290284677, Thp, KP956940335 * Thm);
 
3880
                         ThT = FMA(KP956940335, Thp, KP290284677 * Thm);
 
3881
                         Tht = Thr - Ths;
 
3882
                         Thw = Thu + Thv;
 
3883
                         Thx = FMA(KP290284677, Tht, KP956940335 * Thw);
 
3884
                         ThU = FNMS(KP290284677, Thw, KP956940335 * Tht);
 
3885
                    }
 
3886
                    Thy = Thq + Thx;
 
3887
                    Ti6 = ThT - ThU;
 
3888
                    ThV = ThT + ThU;
 
3889
                    Ti3 = Thx - Thq;
 
3890
               }
 
3891
               {
 
3892
                    E Tij, Tiq, Tim, Tir;
 
3893
                    {
 
3894
                         E Tih, Tii, Tik, Til;
 
3895
                         Tih = Thn - Tho;
 
3896
                         Tii = Thl - Thk;
 
3897
                         Tij = FNMS(KP471396736, Tii, KP881921264 * Tih);
 
3898
                         Tiq = FMA(KP471396736, Tih, KP881921264 * Tii);
 
3899
                         Tik = Thv - Thu;
 
3900
                         Til = Ths + Thr;
 
3901
                         Tim = FNMS(KP881921264, Til, KP471396736 * Tik);
 
3902
                         Tir = FMA(KP471396736, Til, KP881921264 * Tik);
 
3903
                    }
 
3904
                    Tin = Tij + Tim;
 
3905
                    TiH = Tiq - Tir;
 
3906
                    Tis = Tiq + Tir;
 
3907
                    TiC = Tim - Tij;
 
3908
               }
 
3909
               {
 
3910
                    E ThK, ThW, Thj, ThL;
 
3911
                    ThK = Thy + ThJ;
 
3912
                    ThW = ThS + ThV;
 
3913
                    Thj = W[4];
 
3914
                    ThL = W[5];
 
3915
                    rio[WS(ios, 3)] = FNMS(ThL, ThW, Thj * ThK);
 
3916
                    iio[-WS(ios, 60)] = FMA(Thj, ThW, ThL * ThK);
 
3917
               }
 
3918
               {
 
3919
                    E TiE, TiI, TiB, TiF;
 
3920
                    TiE = TiC + TiD;
 
3921
                    TiI = TiG + TiH;
 
3922
                    TiB = W[20];
 
3923
                    TiF = W[21];
 
3924
                    rio[WS(ios, 11)] = FNMS(TiF, TiI, TiB * TiE);
 
3925
                    iio[-WS(ios, 52)] = FMA(TiB, TiI, TiF * TiE);
 
3926
               }
 
3927
               {
 
3928
                    E TiK, TiM, TiJ, TiL;
 
3929
                    TiK = TiG - TiH;
 
3930
                    TiM = TiD - TiC;
 
3931
                    TiJ = W[84];
 
3932
                    TiL = W[85];
 
3933
                    iio[-WS(ios, 20)] = FMA(TiJ, TiK, TiL * TiM);
 
3934
                    rio[WS(ios, 43)] = FNMS(TiL, TiK, TiJ * TiM);
 
3935
               }
 
3936
               {
 
3937
                    E ThY, Ti0, ThX, ThZ;
 
3938
                    ThY = ThS - ThV;
 
3939
                    Ti0 = ThJ - Thy;
 
3940
                    ThX = W[68];
 
3941
                    ThZ = W[69];
 
3942
                    iio[-WS(ios, 28)] = FMA(ThX, ThY, ThZ * Ti0);
 
3943
                    rio[WS(ios, 35)] = FNMS(ThZ, ThY, ThX * Ti0);
 
3944
               }
 
3945
               {
 
3946
                    E Ti4, Ti8, Ti1, Ti5;
 
3947
                    Ti4 = Ti2 + Ti3;
 
3948
                    Ti8 = Ti6 + Ti7;
 
3949
                    Ti1 = W[100];
 
3950
                    Ti5 = W[101];
 
3951
                    iio[-WS(ios, 12)] = FMA(Ti1, Ti4, Ti5 * Ti8);
 
3952
                    rio[WS(ios, 51)] = FNMS(Ti5, Ti4, Ti1 * Ti8);
 
3953
               }
 
3954
               {
 
3955
                    E Tio, Tiw, Tid, Tip;
 
3956
                    Tio = Tig + Tin;
 
3957
                    Tiw = Tis + Tiv;
 
3958
                    Tid = W[116];
 
3959
                    Tip = W[117];
 
3960
                    iio[-WS(ios, 4)] = FMA(Tid, Tio, Tip * Tiw);
 
3961
                    rio[WS(ios, 59)] = FNMS(Tip, Tio, Tid * Tiw);
 
3962
               }
 
3963
               {
 
3964
                    E Tiy, TiA, Tix, Tiz;
 
3965
                    Tiy = Tiv - Tis;
 
3966
                    TiA = Tig - Tin;
 
3967
                    Tix = W[52];
 
3968
                    Tiz = W[53];
 
3969
                    rio[WS(ios, 27)] = FNMS(Tiz, TiA, Tix * Tiy);
 
3970
                    iio[-WS(ios, 36)] = FMA(Tix, TiA, Tiz * Tiy);
 
3971
               }
 
3972
               {
 
3973
                    E Tia, Tic, Ti9, Tib;
 
3974
                    Tia = Ti7 - Ti6;
 
3975
                    Tic = Ti2 - Ti3;
 
3976
                    Ti9 = W[36];
 
3977
                    Tib = W[37];
 
3978
                    rio[WS(ios, 19)] = FNMS(Tib, Tic, Ti9 * Tia);
 
3979
                    iio[-WS(ios, 44)] = FMA(Ti9, Tic, Tib * Tia);
 
3980
               }
 
3981
          }
 
3982
     }
 
3983
     return W;
 
3984
}
 
3985
 
 
3986
static const tw_instr twinstr[] = {
 
3987
     {TW_FULL, 0, 64},
 
3988
     {TW_NEXT, 1, 0}
 
3989
};
 
3990
 
 
3991
static const hc2hc_desc desc = { 64, "hb_64", twinstr, &GENUS, {808, 270, 230, 0}, 0, 0, 0 };
 
3992
 
 
3993
void X(codelet_hb_64) (planner *p) {
 
3994
     X(khc2hc_register) (p, hb_64, &desc);
 
3995
}
 
3996
#endif                          /* HAVE_FMA */