~ubuntu-branches/ubuntu/utopic/fftw3/utopic

« back to all changes in this revision

Viewing changes to dft/simd/common/n1bv_128.c

  • Committer: Package Import Robot
  • Author(s): Matthias Klose
  • Date: 2011-12-14 13:21:22 UTC
  • mfrom: (3.1.5 sid)
  • Revision ID: package-import@ubuntu.com-20111214132122-l4avyl2kkr7vq5aj
Tags: 3.3-1ubuntu1
* Merge with Debian; remaining changes:
  - Revert the ARM workaround.

Show diffs side-by-side

added added

removed removed

Lines of Context:
 
1
/*
 
2
 * Copyright (c) 2003, 2007-11 Matteo Frigo
 
3
 * Copyright (c) 2003, 2007-11 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 Wed Jul 27 06:13:45 EDT 2011 */
 
23
 
 
24
#include "codelet-dft.h"
 
25
 
 
26
#ifdef HAVE_FMA
 
27
 
 
28
/* Generated by: ../../../genfft/gen_notw_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 128 -name n1bv_128 -include n1b.h */
 
29
 
 
30
/*
 
31
 * This function contains 1082 FP additions, 642 FP multiplications,
 
32
 * (or, 440 additions, 0 multiplications, 642 fused multiply/add),
 
33
 * 295 stack variables, 31 constants, and 256 memory accesses
 
34
 */
 
35
#include "n1b.h"
 
36
 
 
37
static void n1bv_128(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
 
38
{
 
39
     DVK(KP903989293, +0.903989293123443331586200297230537048710132025);
 
40
     DVK(KP941544065, +0.941544065183020778412509402599502357185589796);
 
41
     DVK(KP357805721, +0.357805721314524104672487743774474392487532769);
 
42
     DVK(KP472964775, +0.472964775891319928124438237972992463904131113);
 
43
     DVK(KP857728610, +0.857728610000272069902269984284770137042490799);
 
44
     DVK(KP970031253, +0.970031253194543992603984207286100251456865962);
 
45
     DVK(KP250486960, +0.250486960191305461595702160124721208578685568);
 
46
     DVK(KP998795456, +0.998795456205172392714771604759100694443203615);
 
47
     DVK(KP740951125, +0.740951125354959091175616897495162729728955309);
 
48
     DVK(KP599376933, +0.599376933681923766271389869014404232837890546);
 
49
     DVK(KP906347169, +0.906347169019147157946142717268914412664134293);
 
50
     DVK(KP049126849, +0.049126849769467254105343321271313617079695752);
 
51
     DVK(KP989176509, +0.989176509964780973451673738016243063983689533);
 
52
     DVK(KP803207531, +0.803207531480644909806676512963141923879569427);
 
53
     DVK(KP741650546, +0.741650546272035369581266691172079863842265220);
 
54
     DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
 
55
     DVK(KP148335987, +0.148335987538347428753676511486911367000625355);
 
56
     DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
 
57
     DVK(KP303346683, +0.303346683607342391675883946941299872384187453);
 
58
     DVK(KP534511135, +0.534511135950791641089685961295362908582039528);
 
59
     DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
 
60
     DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
 
61
     DVK(KP820678790, +0.820678790828660330972281985331011598767386482);
 
62
     DVK(KP098491403, +0.098491403357164253077197521291327432293052451);
 
63
     DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
 
64
     DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
 
65
     DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
 
66
     DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
 
67
     DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
68
     DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
 
69
     DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
70
     {
 
71
          INT i;
 
72
          const R *xi;
 
73
          R *xo;
 
74
          xi = ii;
 
75
          xo = io;
 
76
          for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
 
77
               V T6a, T5J, T6b, T5K, T6B, T6C, T6J, T6A, T6o, T6j, T6r, T68, T6d, T5O, T5R;
 
78
               V T6e, T6D, T6K;
 
79
               {
 
80
                    V Tad, TcZ, T6Z, T8T, T4U, Tr, Tfq, TgG, Ted, Tgf, Td0, Tcc, T9k, T84, Tb6;
 
81
                    V Tbt, Td8, TdK, TeK, Tgq, TeV, Tgt, T7q, T94, T3p, T5X, T7B, T97, T2G, T5U;
 
82
                    V TbD, Tc0, Tdf, TdN, Tf5, Tgx, Tfg, TgA, T7J, T9b, T4E, T64, T7U, T9e, T3V;
 
83
                    V T61, Td2, Td3, T85, T72, T4V, TI, Tcd, Tas, TgH, Tek, Tgg, Tft, T86, T75;
 
84
                    V T4W, TZ, TaI, Tcf, Tdo, TdG, Tgi, Tet, Tgj, Teq, T8X, T7a, T5M, T1B, T8W;
 
85
                    V T7d, T5N, T1s, TaX, Tcg, Tdr, TdH, Tgl, TeC, Tgm, Tez, T90, T7h, T5P, T2c;
 
86
                    V T8Z, T7k, T5Q, T23, T3Y, T49, TdL, Tdb, Tbu, Tbl, Tgu, TeR, Tgr, TeY, Tf6;
 
87
                    V TbG, T5V, T3s, T5Y, T3f, T95, T7E, T98, T7x, T4g, T4f, T4q, TbH, T41, TbI;
 
88
                    V T44, T4h, T4j, T4k, Tf9, TbN;
 
89
                    {
 
90
                         V Tu, TF, Ty, TL, TW, Tah, Tx, Tag, Tee, Tz, TM, TN, Teh, Tan, TP;
 
91
                         V TQ;
 
92
                         {
 
93
                              V TeG, T2A, Tbq, TeT, Tbp, TeH, T3m, T2x, Td6, T7o, T2q, T3l, T7z, Tbr, T2D;
 
94
                              V T82, T83;
 
95
                              {
 
96
                                   V Ta7, T3, Ta8, T4O, Taa, Tab, Ta, T4P, Te, Tc6, Th, Tc7, Tl, Tc9, Tca;
 
97
                                   V To;
 
98
                                   {
 
99
                                        V T1, T2, T4M, T4N;
 
100
                                        T1 = LD(&(xi[0]), ivs, &(xi[0]));
 
101
                                        T2 = LD(&(xi[WS(is, 64)]), ivs, &(xi[0]));
 
102
                                        T4M = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
 
103
                                        T4N = LD(&(xi[WS(is, 96)]), ivs, &(xi[0]));
 
104
                                        {
 
105
                                             V T4, T5, T7, T8;
 
106
                                             T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
 
107
                                             T5 = LD(&(xi[WS(is, 80)]), ivs, &(xi[0]));
 
108
                                             T7 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0]));
 
109
                                             T8 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
 
110
                                             {
 
111
                                                  V Tc, T6, T9, Td, Tf, Tg;
 
112
                                                  Tc = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
 
113
                                                  Ta7 = VADD(T1, T2);
 
114
                                                  T3 = VSUB(T1, T2);
 
115
                                                  Ta8 = VADD(T4M, T4N);
 
116
                                                  T4O = VSUB(T4M, T4N);
 
117
                                                  Taa = VADD(T4, T5);
 
118
                                                  T6 = VSUB(T4, T5);
 
119
                                                  Tab = VADD(T7, T8);
 
120
                                                  T9 = VSUB(T7, T8);
 
121
                                                  Td = LD(&(xi[WS(is, 72)]), ivs, &(xi[0]));
 
122
                                                  Tf = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
 
123
                                                  Tg = LD(&(xi[WS(is, 104)]), ivs, &(xi[0]));
 
124
                                                  {
 
125
                                                       V Tj, Tk, Tm, Tn;
 
126
                                                       Tj = LD(&(xi[WS(is, 120)]), ivs, &(xi[0]));
 
127
                                                       Tk = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
 
128
                                                       Tm = LD(&(xi[WS(is, 88)]), ivs, &(xi[0]));
 
129
                                                       Tn = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
 
130
                                                       Ta = VADD(T6, T9);
 
131
                                                       T4P = VSUB(T6, T9);
 
132
                                                       Te = VSUB(Tc, Td);
 
133
                                                       Tc6 = VADD(Tc, Td);
 
134
                                                       Th = VSUB(Tf, Tg);
 
135
                                                       Tc7 = VADD(Tf, Tg);
 
136
                                                       Tl = VSUB(Tj, Tk);
 
137
                                                       Tc9 = VADD(Tj, Tk);
 
138
                                                       Tca = VADD(Tn, Tm);
 
139
                                                       To = VSUB(Tm, Tn);
 
140
                                                  }
 
141
                                             }
 
142
                                        }
 
143
                                   }
 
144
                                   {
 
145
                                        V T6X, Tb, Te9, Ta9, Tc8, Tea, T4R, Ti, Tfo, Tac, Tp, T4S, Tcb, Teb, T4Q;
 
146
                                        T6X = VFNMS(LDK(KP707106781), Ta, T3);
 
147
                                        Tb = VFMA(LDK(KP707106781), Ta, T3);
 
148
                                        Te9 = VSUB(Ta7, Ta8);
 
149
                                        Ta9 = VADD(Ta7, Ta8);
 
150
                                        Tc8 = VADD(Tc6, Tc7);
 
151
                                        Tea = VSUB(Tc6, Tc7);
 
152
                                        T4R = VFMA(LDK(KP414213562), Te, Th);
 
153
                                        Ti = VFNMS(LDK(KP414213562), Th, Te);
 
154
                                        Tfo = VSUB(Taa, Tab);
 
155
                                        Tac = VADD(Taa, Tab);
 
156
                                        Tp = VFNMS(LDK(KP414213562), To, Tl);
 
157
                                        T4S = VFMA(LDK(KP414213562), Tl, To);
 
158
                                        Tcb = VADD(Tc9, Tca);
 
159
                                        Teb = VSUB(Tc9, Tca);
 
160
                                        T4Q = VFMA(LDK(KP707106781), T4P, T4O);
 
161
                                        T82 = VFNMS(LDK(KP707106781), T4P, T4O);
 
162
                                        {
 
163
                                             V T4T, T6Y, Tq, Tfp, Tec;
 
164
                                             T4T = VSUB(T4R, T4S);
 
165
                                             T6Y = VADD(T4R, T4S);
 
166
                                             T83 = VSUB(Ti, Tp);
 
167
                                             Tq = VADD(Ti, Tp);
 
168
                                             Tfp = VSUB(Tea, Teb);
 
169
                                             Tec = VADD(Tea, Teb);
 
170
                                             Tad = VSUB(Ta9, Tac);
 
171
                                             TcZ = VADD(Ta9, Tac);
 
172
                                             T6Z = VFMA(LDK(KP923879532), T6Y, T6X);
 
173
                                             T8T = VFNMS(LDK(KP923879532), T6Y, T6X);
 
174
                                             T4U = VFMA(LDK(KP923879532), T4T, T4Q);
 
175
                                             T6a = VFNMS(LDK(KP923879532), T4T, T4Q);
 
176
                                             Tr = VFMA(LDK(KP923879532), Tq, Tb);
 
177
                                             T5J = VFNMS(LDK(KP923879532), Tq, Tb);
 
178
                                             Tfq = VFMA(LDK(KP707106781), Tfp, Tfo);
 
179
                                             TgG = VFNMS(LDK(KP707106781), Tfp, Tfo);
 
180
                                             Ted = VFMA(LDK(KP707106781), Tec, Te9);
 
181
                                             Tgf = VFNMS(LDK(KP707106781), Tec, Te9);
 
182
                                             Td0 = VADD(Tc8, Tcb);
 
183
                                             Tcc = VSUB(Tc8, Tcb);
 
184
                                        }
 
185
                                   }
 
186
                              }
 
187
                              {
 
188
                                   V T2i, T3j, Tb2, T2B, Tb5, T3k, T2p, T2C;
 
189
                                   {
 
190
                                        V T2m, Tb0, Tb1, Tb3, T2l, T2n;
 
191
                                        {
 
192
                                             V T2g, T2h, T3h, T3i, T2j, T2k;
 
193
                                             T2g = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
 
194
                                             T2h = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)]));
 
195
                                             T3h = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
 
196
                                             T3i = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)]));
 
197
                                             T2j = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
 
198
                                             T2k = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)]));
 
199
                                             T2m = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)]));
 
200
                                             T9k = VFMA(LDK(KP923879532), T83, T82);
 
201
                                             T84 = VFNMS(LDK(KP923879532), T83, T82);
 
202
                                             T2i = VSUB(T2g, T2h);
 
203
                                             Tb0 = VADD(T2g, T2h);
 
204
                                             T3j = VSUB(T3h, T3i);
 
205
                                             Tb1 = VADD(T3h, T3i);
 
206
                                             Tb3 = VADD(T2j, T2k);
 
207
                                             T2l = VSUB(T2j, T2k);
 
208
                                             T2n = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
 
209
                                        }
 
210
                                        {
 
211
                                             V T2r, T2s, T2u, T2v;
 
212
                                             T2r = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
 
213
                                             T2s = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)]));
 
214
                                             T2u = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
 
215
                                             T2v = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)]));
 
216
                                             TeG = VSUB(Tb0, Tb1);
 
217
                                             Tb2 = VADD(Tb0, Tb1);
 
218
                                             {
 
219
                                                  V T2y, T2z, Tb4, T2o, Tbn, T2t, Tbo, T2w;
 
220
                                                  T2y = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)]));
 
221
                                                  T2z = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
 
222
                                                  Tb4 = VADD(T2m, T2n);
 
223
                                                  T2o = VSUB(T2m, T2n);
 
224
                                                  Tbn = VADD(T2r, T2s);
 
225
                                                  T2t = VSUB(T2r, T2s);
 
226
                                                  Tbo = VADD(T2u, T2v);
 
227
                                                  T2w = VSUB(T2u, T2v);
 
228
                                                  T2B = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
 
229
                                                  T2A = VSUB(T2y, T2z);
 
230
                                                  Tbq = VADD(T2y, T2z);
 
231
                                                  TeT = VSUB(Tb3, Tb4);
 
232
                                                  Tb5 = VADD(Tb3, Tb4);
 
233
                                                  T3k = VSUB(T2l, T2o);
 
234
                                                  T2p = VADD(T2l, T2o);
 
235
                                                  Tbp = VADD(Tbn, Tbo);
 
236
                                                  TeH = VSUB(Tbn, Tbo);
 
237
                                                  T3m = VFMA(LDK(KP414213562), T2t, T2w);
 
238
                                                  T2x = VFNMS(LDK(KP414213562), T2w, T2t);
 
239
                                                  T2C = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)]));
 
240
                                             }
 
241
                                        }
 
242
                                   }
 
243
                                   Td6 = VADD(Tb2, Tb5);
 
244
                                   Tb6 = VSUB(Tb2, Tb5);
 
245
                                   T7o = VFNMS(LDK(KP707106781), T2p, T2i);
 
246
                                   T2q = VFMA(LDK(KP707106781), T2p, T2i);
 
247
                                   T3l = VFMA(LDK(KP707106781), T3k, T3j);
 
248
                                   T7z = VFNMS(LDK(KP707106781), T3k, T3j);
 
249
                                   Tbr = VADD(T2B, T2C);
 
250
                                   T2D = VSUB(T2B, T2C);
 
251
                              }
 
252
                              {
 
253
                                   V Tf1, Tfe, Tf2, TbZ, T3M, T4B, Tdd, T3F, T7H, T4A, T7S, TbW, Tf3, T4C, T3T;
 
254
                                   {
 
255
                                        V T3x, T4y, Tbz, T3Q, TbC, T4z, T3E, T3R, T3P, TbU, TbV, T3S;
 
256
                                        {
 
257
                                             V T3y, T3z, T3B, T3C;
 
258
                                             {
 
259
                                                  V T3v, T3w, T4w, T4x;
 
260
                                                  T3v = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)]));
 
261
                                                  T3w = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
 
262
                                                  T4w = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)]));
 
263
                                                  T4x = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
 
264
                                                  T3y = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
 
265
                                                  {
 
266
                                                       V Tbs, TeI, T3n, T2E, Tbx;
 
267
                                                       Tbs = VADD(Tbq, Tbr);
 
268
                                                       TeI = VSUB(Tbq, Tbr);
 
269
                                                       T3n = VFNMS(LDK(KP414213562), T2A, T2D);
 
270
                                                       T2E = VFMA(LDK(KP414213562), T2D, T2A);
 
271
                                                       T3x = VSUB(T3v, T3w);
 
272
                                                       Tbx = VADD(T3v, T3w);
 
273
                                                       {
 
274
                                                            V Tby, Td7, TeJ, TeU;
 
275
                                                            T4y = VSUB(T4w, T4x);
 
276
                                                            Tby = VADD(T4x, T4w);
 
277
                                                            Td7 = VADD(Tbp, Tbs);
 
278
                                                            Tbt = VSUB(Tbp, Tbs);
 
279
                                                            TeJ = VADD(TeH, TeI);
 
280
                                                            TeU = VSUB(TeH, TeI);
 
281
                                                            {
 
282
                                                                 V T7p, T3o, T7A, T2F;
 
283
                                                                 T7p = VSUB(T3m, T3n);
 
284
                                                                 T3o = VADD(T3m, T3n);
 
285
                                                                 T7A = VSUB(T2x, T2E);
 
286
                                                                 T2F = VADD(T2x, T2E);
 
287
                                                                 Tbz = VADD(Tbx, Tby);
 
288
                                                                 Tf1 = VSUB(Tbx, Tby);
 
289
                                                                 Td8 = VADD(Td6, Td7);
 
290
                                                                 TdK = VSUB(Td6, Td7);
 
291
                                                                 TeK = VFMA(LDK(KP707106781), TeJ, TeG);
 
292
                                                                 Tgq = VFNMS(LDK(KP707106781), TeJ, TeG);
 
293
                                                                 TeV = VFMA(LDK(KP707106781), TeU, TeT);
 
294
                                                                 Tgt = VFNMS(LDK(KP707106781), TeU, TeT);
 
295
                                                                 T7q = VFMA(LDK(KP923879532), T7p, T7o);
 
296
                                                                 T94 = VFNMS(LDK(KP923879532), T7p, T7o);
 
297
                                                                 T3p = VFMA(LDK(KP923879532), T3o, T3l);
 
298
                                                                 T5X = VFNMS(LDK(KP923879532), T3o, T3l);
 
299
                                                                 T7B = VFNMS(LDK(KP923879532), T7A, T7z);
 
300
                                                                 T97 = VFMA(LDK(KP923879532), T7A, T7z);
 
301
                                                                 T2G = VFMA(LDK(KP923879532), T2F, T2q);
 
302
                                                                 T5U = VFNMS(LDK(KP923879532), T2F, T2q);
 
303
                                                                 T3z = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)]));
 
304
                                                            }
 
305
                                                       }
 
306
                                                  }
 
307
                                                  T3B = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)]));
 
308
                                                  T3C = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
 
309
                                             }
 
310
                                             {
 
311
                                                  V T3G, T3H, T3J, T3K;
 
312
                                                  T3G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
 
313
                                                  T3H = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)]));
 
314
                                                  T3J = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)]));
 
315
                                                  T3K = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
 
316
                                                  {
 
317
                                                       V T3N, T3A, TbA, T3D, TbB, T3I, TbX, T3L, TbY, T3O;
 
318
                                                       T3N = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)]));
 
319
                                                       T3A = VSUB(T3y, T3z);
 
320
                                                       TbA = VADD(T3y, T3z);
 
321
                                                       T3D = VSUB(T3B, T3C);
 
322
                                                       TbB = VADD(T3B, T3C);
 
323
                                                       T3I = VSUB(T3G, T3H);
 
324
                                                       TbX = VADD(T3G, T3H);
 
325
                                                       T3L = VSUB(T3J, T3K);
 
326
                                                       TbY = VADD(T3K, T3J);
 
327
                                                       T3O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
 
328
                                                       T3Q = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)]));
 
329
                                                       Tfe = VSUB(TbB, TbA);
 
330
                                                       TbC = VADD(TbA, TbB);
 
331
                                                       T4z = VSUB(T3D, T3A);
 
332
                                                       T3E = VADD(T3A, T3D);
 
333
                                                       T3R = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
 
334
                                                       Tf2 = VSUB(TbX, TbY);
 
335
                                                       TbZ = VADD(TbX, TbY);
 
336
                                                       T3M = VFMA(LDK(KP414213562), T3L, T3I);
 
337
                                                       T4B = VFNMS(LDK(KP414213562), T3I, T3L);
 
338
                                                       T3P = VSUB(T3N, T3O);
 
339
                                                       TbU = VADD(T3N, T3O);
 
340
                                                  }
 
341
                                             }
 
342
                                        }
 
343
                                        Tdd = VADD(Tbz, TbC);
 
344
                                        TbD = VSUB(Tbz, TbC);
 
345
                                        TbV = VADD(T3R, T3Q);
 
346
                                        T3S = VSUB(T3Q, T3R);
 
347
                                        T3F = VFMA(LDK(KP707106781), T3E, T3x);
 
348
                                        T7H = VFNMS(LDK(KP707106781), T3E, T3x);
 
349
                                        T4A = VFMA(LDK(KP707106781), T4z, T4y);
 
350
                                        T7S = VFNMS(LDK(KP707106781), T4z, T4y);
 
351
                                        TbW = VADD(TbU, TbV);
 
352
                                        Tf3 = VSUB(TbU, TbV);
 
353
                                        T4C = VFMA(LDK(KP414213562), T3P, T3S);
 
354
                                        T3T = VFNMS(LDK(KP414213562), T3S, T3P);
 
355
                                   }
 
356
                                   {
 
357
                                        V TD, Tae, TE, TJ, TK, TU, TV;
 
358
                                        {
 
359
                                             V Ts, Tt, Tde, Tf4, Tff;
 
360
                                             Ts = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
 
361
                                             Tt = LD(&(xi[WS(is, 68)]), ivs, &(xi[0]));
 
362
                                             TD = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
 
363
                                             Tde = VADD(TbZ, TbW);
 
364
                                             Tc0 = VSUB(TbW, TbZ);
 
365
                                             Tf4 = VADD(Tf2, Tf3);
 
366
                                             Tff = VSUB(Tf3, Tf2);
 
367
                                             {
 
368
                                                  V T7I, T4D, T7T, T3U;
 
369
                                                  T7I = VSUB(T4C, T4B);
 
370
                                                  T4D = VADD(T4B, T4C);
 
371
                                                  T7T = VSUB(T3T, T3M);
 
372
                                                  T3U = VADD(T3M, T3T);
 
373
                                                  Tae = VADD(Ts, Tt);
 
374
                                                  Tu = VSUB(Ts, Tt);
 
375
                                                  Tdf = VADD(Tdd, Tde);
 
376
                                                  TdN = VSUB(Tdd, Tde);
 
377
                                                  Tf5 = VFMA(LDK(KP707106781), Tf4, Tf1);
 
378
                                                  Tgx = VFNMS(LDK(KP707106781), Tf4, Tf1);
 
379
                                                  Tfg = VFMA(LDK(KP707106781), Tff, Tfe);
 
380
                                                  TgA = VFNMS(LDK(KP707106781), Tff, Tfe);
 
381
                                                  T7J = VFMA(LDK(KP923879532), T7I, T7H);
 
382
                                                  T9b = VFNMS(LDK(KP923879532), T7I, T7H);
 
383
                                                  T4E = VFMA(LDK(KP923879532), T4D, T4A);
 
384
                                                  T64 = VFNMS(LDK(KP923879532), T4D, T4A);
 
385
                                                  T7U = VFNMS(LDK(KP923879532), T7T, T7S);
 
386
                                                  T9e = VFMA(LDK(KP923879532), T7T, T7S);
 
387
                                                  T3V = VFMA(LDK(KP923879532), T3U, T3F);
 
388
                                                  T61 = VFNMS(LDK(KP923879532), T3U, T3F);
 
389
                                                  TE = LD(&(xi[WS(is, 100)]), ivs, &(xi[0]));
 
390
                                             }
 
391
                                        }
 
392
                                        TJ = LD(&(xi[WS(is, 124)]), ivs, &(xi[0]));
 
393
                                        TK = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
 
394
                                        TU = LD(&(xi[WS(is, 92)]), ivs, &(xi[0]));
 
395
                                        TV = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
 
396
                                        {
 
397
                                             V Tal, Tam, Tv, Tw, Taf;
 
398
                                             Tv = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
 
399
                                             Tw = LD(&(xi[WS(is, 84)]), ivs, &(xi[0]));
 
400
                                             Taf = VADD(TD, TE);
 
401
                                             TF = VSUB(TD, TE);
 
402
                                             Ty = LD(&(xi[WS(is, 116)]), ivs, &(xi[0]));
 
403
                                             TL = VSUB(TJ, TK);
 
404
                                             Tal = VADD(TJ, TK);
 
405
                                             TW = VSUB(TU, TV);
 
406
                                             Tam = VADD(TV, TU);
 
407
                                             Tah = VADD(Tv, Tw);
 
408
                                             Tx = VSUB(Tv, Tw);
 
409
                                             Tag = VADD(Tae, Taf);
 
410
                                             Tee = VSUB(Tae, Taf);
 
411
                                             Tz = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
 
412
                                             TM = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
 
413
                                             TN = LD(&(xi[WS(is, 76)]), ivs, &(xi[0]));
 
414
                                             Teh = VSUB(Tal, Tam);
 
415
                                             Tan = VADD(Tal, Tam);
 
416
                                             TP = LD(&(xi[WS(is, 108)]), ivs, &(xi[0]));
 
417
                                             TQ = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
 
418
                                        }
 
419
                                   }
 
420
                              }
 
421
                         }
 
422
                         {
 
423
                              V Tev, TeA, Tdp, TaP, Tew, TaV, T1U, T29, T7f, T1N, T28, T7i, Tex, TaS, T21;
 
424
                              V T2a;
 
425
                              {
 
426
                                   V Tem, Ter, Ten, TaD, T1j, T1y, TaA, Tdm, T1c, T78, T7b, T1x, TaG, Teo, T1z;
 
427
                                   V T1q;
 
428
                                   {
 
429
                                        V T14, T1v, Taw, Taz, T1b, T1w, T1n, T1o, T1m, TaE, TaF, T1p;
 
430
                                        {
 
431
                                             V Tau, Tav, T15, T16, T18, T19;
 
432
                                             {
 
433
                                                  V T12, Tai, TA, Tao, TO, T13;
 
434
                                                  T12 = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
 
435
                                                  Tai = VADD(Ty, Tz);
 
436
                                                  TA = VSUB(Ty, Tz);
 
437
                                                  Tao = VADD(TM, TN);
 
438
                                                  TO = VSUB(TM, TN);
 
439
                                                  T13 = LD(&(xi[WS(is, 66)]), ivs, &(xi[0]));
 
440
                                                  {
 
441
                                                       V T1t, Tap, TR, Taj, Tef, TG, TB, T1u;
 
442
                                                       T1t = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
 
443
                                                       Tap = VADD(TP, TQ);
 
444
                                                       TR = VSUB(TP, TQ);
 
445
                                                       Taj = VADD(Tah, Tai);
 
446
                                                       Tef = VSUB(Tah, Tai);
 
447
                                                       TG = VSUB(Tx, TA);
 
448
                                                       TB = VADD(Tx, TA);
 
449
                                                       Tau = VADD(T12, T13);
 
450
                                                       T14 = VSUB(T12, T13);
 
451
                                                       T1u = LD(&(xi[WS(is, 98)]), ivs, &(xi[0]));
 
452
                                                       {
 
453
                                                            V Taq, Tei, TX, TS, Tak;
 
454
                                                            Taq = VADD(Tao, Tap);
 
455
                                                            Tei = VSUB(Tap, Tao);
 
456
                                                            TX = VSUB(TR, TO);
 
457
                                                            TS = VADD(TO, TR);
 
458
                                                            Tak = VSUB(Tag, Taj);
 
459
                                                            Td2 = VADD(Tag, Taj);
 
460
                                                            {
 
461
                                                                 V Teg, Tfr, T71, TH;
 
462
                                                                 Teg = VFNMS(LDK(KP414213562), Tef, Tee);
 
463
                                                                 Tfr = VFMA(LDK(KP414213562), Tee, Tef);
 
464
                                                                 T71 = VFNMS(LDK(KP707106781), TG, TF);
 
465
                                                                 TH = VFMA(LDK(KP707106781), TG, TF);
 
466
                                                                 {
 
467
                                                                      V T70, TC, Tar, Tej, Tfs;
 
468
                                                                      T70 = VFNMS(LDK(KP707106781), TB, Tu);
 
469
                                                                      TC = VFMA(LDK(KP707106781), TB, Tu);
 
470
                                                                      Tar = VSUB(Tan, Taq);
 
471
                                                                      Td3 = VADD(Tan, Taq);
 
472
                                                                      Tej = VFNMS(LDK(KP414213562), Tei, Teh);
 
473
                                                                      Tfs = VFMA(LDK(KP414213562), Teh, Tei);
 
474
                                                                      {
 
475
                                                                           V T74, TY, T73, TT;
 
476
                                                                           T74 = VFNMS(LDK(KP707106781), TX, TW);
 
477
                                                                           TY = VFMA(LDK(KP707106781), TX, TW);
 
478
                                                                           T73 = VFNMS(LDK(KP707106781), TS, TL);
 
479
                                                                           TT = VFMA(LDK(KP707106781), TS, TL);
 
480
                                                                           T85 = VFNMS(LDK(KP668178637), T70, T71);
 
481
                                                                           T72 = VFMA(LDK(KP668178637), T71, T70);
 
482
                                                                           T4V = VFMA(LDK(KP198912367), TC, TH);
 
483
                                                                           TI = VFNMS(LDK(KP198912367), TH, TC);
 
484
                                                                           Tcd = VSUB(Tak, Tar);
 
485
                                                                           Tas = VADD(Tak, Tar);
 
486
                                                                           TgH = VSUB(Teg, Tej);
 
487
                                                                           Tek = VADD(Teg, Tej);
 
488
                                                                           Tgg = VADD(Tfr, Tfs);
 
489
                                                                           Tft = VSUB(Tfr, Tfs);
 
490
                                                                           T86 = VFNMS(LDK(KP668178637), T73, T74);
 
491
                                                                           T75 = VFMA(LDK(KP668178637), T74, T73);
 
492
                                                                           T4W = VFMA(LDK(KP198912367), TT, TY);
 
493
                                                                           TZ = VFNMS(LDK(KP198912367), TY, TT);
 
494
                                                                           Tav = VADD(T1t, T1u);
 
495
                                                                           T1v = VSUB(T1t, T1u);
 
496
                                                                      }
 
497
                                                                 }
 
498
                                                            }
 
499
                                                       }
 
500
                                                  }
 
501
                                             }
 
502
                                             T15 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
 
503
                                             T16 = LD(&(xi[WS(is, 82)]), ivs, &(xi[0]));
 
504
                                             T18 = LD(&(xi[WS(is, 114)]), ivs, &(xi[0]));
 
505
                                             T19 = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
 
506
                                             {
 
507
                                                  V T1d, T1e, T1g, T1h, Tax, T17, Tay, T1a;
 
508
                                                  T1d = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
 
509
                                                  Taw = VADD(Tau, Tav);
 
510
                                                  Tem = VSUB(Tau, Tav);
 
511
                                                  T1e = LD(&(xi[WS(is, 74)]), ivs, &(xi[0]));
 
512
                                                  T1g = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
 
513
                                                  T1h = LD(&(xi[WS(is, 106)]), ivs, &(xi[0]));
 
514
                                                  Tax = VADD(T15, T16);
 
515
                                                  T17 = VSUB(T15, T16);
 
516
                                                  Tay = VADD(T18, T19);
 
517
                                                  T1a = VSUB(T18, T19);
 
518
                                                  {
 
519
                                                       V T1k, T1f, TaB, T1i, TaC, T1l;
 
520
                                                       T1k = LD(&(xi[WS(is, 122)]), ivs, &(xi[0]));
 
521
                                                       T1f = VSUB(T1d, T1e);
 
522
                                                       TaB = VADD(T1d, T1e);
 
523
                                                       T1i = VSUB(T1g, T1h);
 
524
                                                       TaC = VADD(T1g, T1h);
 
525
                                                       T1l = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
 
526
                                                       Taz = VADD(Tax, Tay);
 
527
                                                       Ter = VSUB(Tax, Tay);
 
528
                                                       T1b = VADD(T17, T1a);
 
529
                                                       T1w = VSUB(T17, T1a);
 
530
                                                       T1n = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
 
531
                                                       T1o = LD(&(xi[WS(is, 90)]), ivs, &(xi[0]));
 
532
                                                       Ten = VSUB(TaB, TaC);
 
533
                                                       TaD = VADD(TaB, TaC);
 
534
                                                       T1j = VFNMS(LDK(KP414213562), T1i, T1f);
 
535
                                                       T1y = VFMA(LDK(KP414213562), T1f, T1i);
 
536
                                                       T1m = VSUB(T1k, T1l);
 
537
                                                       TaE = VADD(T1k, T1l);
 
538
                                                  }
 
539
                                             }
 
540
                                        }
 
541
                                        TaA = VSUB(Taw, Taz);
 
542
                                        Tdm = VADD(Taw, Taz);
 
543
                                        TaF = VADD(T1n, T1o);
 
544
                                        T1p = VSUB(T1n, T1o);
 
545
                                        T1c = VFMA(LDK(KP707106781), T1b, T14);
 
546
                                        T78 = VFNMS(LDK(KP707106781), T1b, T14);
 
547
                                        T7b = VFNMS(LDK(KP707106781), T1w, T1v);
 
548
                                        T1x = VFMA(LDK(KP707106781), T1w, T1v);
 
549
                                        TaG = VADD(TaE, TaF);
 
550
                                        Teo = VSUB(TaE, TaF);
 
551
                                        T1z = VFNMS(LDK(KP414213562), T1m, T1p);
 
552
                                        T1q = VFMA(LDK(KP414213562), T1p, T1m);
 
553
                                   }
 
554
                                   {
 
555
                                        V T1F, T26, T1Q, TaT, TaL, TaO, T27, T1M, T1Y, T1Z, TaU, T1T, TaQ, T1X, T20;
 
556
                                        V TaR;
 
557
                                        {
 
558
                                             V T24, TaJ, T25, T1G, T1H, T1J, T1K, T1D, T1E;
 
559
                                             T1D = LD(&(xi[WS(is, 126)]), ivs, &(xi[0]));
 
560
                                             T1E = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
 
561
                                             T24 = LD(&(xi[WS(is, 94)]), ivs, &(xi[0]));
 
562
                                             {
 
563
                                                  V TaH, Tdn, Tes, Tep;
 
564
                                                  TaH = VSUB(TaD, TaG);
 
565
                                                  Tdn = VADD(TaD, TaG);
 
566
                                                  Tes = VSUB(Ten, Teo);
 
567
                                                  Tep = VADD(Ten, Teo);
 
568
                                                  {
 
569
                                                       V T79, T1A, T7c, T1r;
 
570
                                                       T79 = VSUB(T1y, T1z);
 
571
                                                       T1A = VADD(T1y, T1z);
 
572
                                                       T7c = VSUB(T1j, T1q);
 
573
                                                       T1r = VADD(T1j, T1q);
 
574
                                                       TaJ = VADD(T1D, T1E);
 
575
                                                       T1F = VSUB(T1D, T1E);
 
576
                                                       TaI = VFNMS(LDK(KP414213562), TaH, TaA);
 
577
                                                       Tcf = VFMA(LDK(KP414213562), TaA, TaH);
 
578
                                                       Tdo = VADD(Tdm, Tdn);
 
579
                                                       TdG = VSUB(Tdm, Tdn);
 
580
                                                       Tgi = VFNMS(LDK(KP707106781), Tes, Ter);
 
581
                                                       Tet = VFMA(LDK(KP707106781), Tes, Ter);
 
582
                                                       Tgj = VFNMS(LDK(KP707106781), Tep, Tem);
 
583
                                                       Teq = VFMA(LDK(KP707106781), Tep, Tem);
 
584
                                                       T8X = VFNMS(LDK(KP923879532), T79, T78);
 
585
                                                       T7a = VFMA(LDK(KP923879532), T79, T78);
 
586
                                                       T5M = VFNMS(LDK(KP923879532), T1A, T1x);
 
587
                                                       T1B = VFMA(LDK(KP923879532), T1A, T1x);
 
588
                                                       T8W = VFMA(LDK(KP923879532), T7c, T7b);
 
589
                                                       T7d = VFNMS(LDK(KP923879532), T7c, T7b);
 
590
                                                       T5N = VFNMS(LDK(KP923879532), T1r, T1c);
 
591
                                                       T1s = VFMA(LDK(KP923879532), T1r, T1c);
 
592
                                                       T25 = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
 
593
                                                  }
 
594
                                             }
 
595
                                             T1G = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
 
596
                                             T1H = LD(&(xi[WS(is, 78)]), ivs, &(xi[0]));
 
597
                                             T1J = LD(&(xi[WS(is, 110)]), ivs, &(xi[0]));
 
598
                                             T1K = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
 
599
                                             {
 
600
                                                  V T1R, T1I, TaM, T1L, TaN, T1S, T1O, T1P, TaK, T1V, T1W;
 
601
                                                  T1O = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
 
602
                                                  T1P = LD(&(xi[WS(is, 70)]), ivs, &(xi[0]));
 
603
                                                  T26 = VSUB(T24, T25);
 
604
                                                  TaK = VADD(T25, T24);
 
605
                                                  T1R = LD(&(xi[WS(is, 102)]), ivs, &(xi[0]));
 
606
                                                  T1I = VSUB(T1G, T1H);
 
607
                                                  TaM = VADD(T1G, T1H);
 
608
                                                  T1L = VSUB(T1J, T1K);
 
609
                                                  TaN = VADD(T1J, T1K);
 
610
                                                  T1Q = VSUB(T1O, T1P);
 
611
                                                  TaT = VADD(T1O, T1P);
 
612
                                                  Tev = VSUB(TaJ, TaK);
 
613
                                                  TaL = VADD(TaJ, TaK);
 
614
                                                  T1S = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
 
615
                                                  T1V = LD(&(xi[WS(is, 118)]), ivs, &(xi[0]));
 
616
                                                  T1W = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
 
617
                                                  TeA = VSUB(TaN, TaM);
 
618
                                                  TaO = VADD(TaM, TaN);
 
619
                                                  T27 = VSUB(T1L, T1I);
 
620
                                                  T1M = VADD(T1I, T1L);
 
621
                                                  T1Y = LD(&(xi[WS(is, 86)]), ivs, &(xi[0]));
 
622
                                                  T1Z = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
 
623
                                                  TaU = VADD(T1S, T1R);
 
624
                                                  T1T = VSUB(T1R, T1S);
 
625
                                                  TaQ = VADD(T1V, T1W);
 
626
                                                  T1X = VSUB(T1V, T1W);
 
627
                                             }
 
628
                                        }
 
629
                                        Tdp = VADD(TaL, TaO);
 
630
                                        TaP = VSUB(TaL, TaO);
 
631
                                        T20 = VSUB(T1Y, T1Z);
 
632
                                        TaR = VADD(T1Z, T1Y);
 
633
                                        Tew = VSUB(TaT, TaU);
 
634
                                        TaV = VADD(TaT, TaU);
 
635
                                        T1U = VFMA(LDK(KP414213562), T1T, T1Q);
 
636
                                        T29 = VFNMS(LDK(KP414213562), T1Q, T1T);
 
637
                                        T7f = VFNMS(LDK(KP707106781), T1M, T1F);
 
638
                                        T1N = VFMA(LDK(KP707106781), T1M, T1F);
 
639
                                        T28 = VFMA(LDK(KP707106781), T27, T26);
 
640
                                        T7i = VFNMS(LDK(KP707106781), T27, T26);
 
641
                                        Tex = VSUB(TaQ, TaR);
 
642
                                        TaS = VADD(TaQ, TaR);
 
643
                                        T21 = VFNMS(LDK(KP414213562), T20, T1X);
 
644
                                        T2a = VFMA(LDK(KP414213562), T1X, T20);
 
645
                                   }
 
646
                              }
 
647
                              {
 
648
                                   V T2J, T2U, T30, T3b, TeL, Tb9, TeO, Tbg, T2M, Tba, T2P, Tbb, T34, Tbh, T33;
 
649
                                   V T35;
 
650
                                   {
 
651
                                        V T2H, T2I, T2S, T2T, T2Y, T2Z, T39, T3a;
 
652
                                        T2H = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
 
653
                                        {
 
654
                                             V Tdq, TaW, Tey, TeB;
 
655
                                             Tdq = VADD(TaV, TaS);
 
656
                                             TaW = VSUB(TaS, TaV);
 
657
                                             Tey = VADD(Tew, Tex);
 
658
                                             TeB = VSUB(Tex, Tew);
 
659
                                             {
 
660
                                                  V T2b, T7g, T22, T7j;
 
661
                                                  T2b = VADD(T29, T2a);
 
662
                                                  T7g = VSUB(T2a, T29);
 
663
                                                  T22 = VADD(T1U, T21);
 
664
                                                  T7j = VSUB(T21, T1U);
 
665
                                                  TaX = VFNMS(LDK(KP414213562), TaW, TaP);
 
666
                                                  Tcg = VFMA(LDK(KP414213562), TaP, TaW);
 
667
                                                  Tdr = VADD(Tdp, Tdq);
 
668
                                                  TdH = VSUB(Tdp, Tdq);
 
669
                                                  Tgl = VFNMS(LDK(KP707106781), TeB, TeA);
 
670
                                                  TeC = VFMA(LDK(KP707106781), TeB, TeA);
 
671
                                                  Tgm = VFNMS(LDK(KP707106781), Tey, Tev);
 
672
                                                  Tez = VFMA(LDK(KP707106781), Tey, Tev);
 
673
                                                  T90 = VFNMS(LDK(KP923879532), T7g, T7f);
 
674
                                                  T7h = VFMA(LDK(KP923879532), T7g, T7f);
 
675
                                                  T5P = VFNMS(LDK(KP923879532), T2b, T28);
 
676
                                                  T2c = VFMA(LDK(KP923879532), T2b, T28);
 
677
                                                  T8Z = VFMA(LDK(KP923879532), T7j, T7i);
 
678
                                                  T7k = VFNMS(LDK(KP923879532), T7j, T7i);
 
679
                                                  T5Q = VFNMS(LDK(KP923879532), T22, T1N);
 
680
                                                  T23 = VFMA(LDK(KP923879532), T22, T1N);
 
681
                                                  T2I = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)]));
 
682
                                             }
 
683
                                        }
 
684
                                        T2S = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
 
685
                                        T2T = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)]));
 
686
                                        T2Y = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)]));
 
687
                                        T2Z = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
 
688
                                        T39 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
 
689
                                        T3a = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)]));
 
690
                                        {
 
691
                                             V T2K, Tbe, Tbf, T2L, T2N, T2O, Tb7, Tb8, T31, T32;
 
692
                                             T2K = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
 
693
                                             T2J = VSUB(T2H, T2I);
 
694
                                             Tb7 = VADD(T2H, T2I);
 
695
                                             T2U = VSUB(T2S, T2T);
 
696
                                             Tb8 = VADD(T2S, T2T);
 
697
                                             T30 = VSUB(T2Y, T2Z);
 
698
                                             Tbe = VADD(T2Y, T2Z);
 
699
                                             T3b = VSUB(T39, T3a);
 
700
                                             Tbf = VADD(T39, T3a);
 
701
                                             T2L = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)]));
 
702
                                             T2N = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)]));
 
703
                                             T2O = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
 
704
                                             TeL = VSUB(Tb7, Tb8);
 
705
                                             Tb9 = VADD(Tb7, Tb8);
 
706
                                             T31 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
 
707
                                             T32 = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)]));
 
708
                                             TeO = VSUB(Tbe, Tbf);
 
709
                                             Tbg = VADD(Tbe, Tbf);
 
710
                                             T2M = VSUB(T2K, T2L);
 
711
                                             Tba = VADD(T2K, T2L);
 
712
                                             T2P = VSUB(T2N, T2O);
 
713
                                             Tbb = VADD(T2N, T2O);
 
714
                                             T34 = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)]));
 
715
                                             Tbh = VADD(T31, T32);
 
716
                                             T33 = VSUB(T31, T32);
 
717
                                             T35 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
 
718
                                        }
 
719
                                   }
 
720
                                   {
 
721
                                        V T4d, T4e, T4o, T4p;
 
722
                                        {
 
723
                                             V T2X, T3q, T7t, T7C, T3r, T3e, T7D, T7w;
 
724
                                             {
 
725
                                                  V T47, TbE, Tbd, Td9, TeW, TeN, T7s, T2W, T7r, T2R, TeP, Tbj, T37, T3c, T48;
 
726
                                                  {
 
727
                                                       V T3W, T3X, TeM, Tbc, T2Q, T2V, Tbi, T36;
 
728
                                                       T3W = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
 
729
                                                       T3X = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)]));
 
730
                                                       TeM = VSUB(Tba, Tbb);
 
731
                                                       Tbc = VADD(Tba, Tbb);
 
732
                                                       T2Q = VADD(T2M, T2P);
 
733
                                                       T2V = VSUB(T2M, T2P);
 
734
                                                       T47 = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)]));
 
735
                                                       Tbi = VADD(T34, T35);
 
736
                                                       T36 = VSUB(T34, T35);
 
737
                                                       TbE = VADD(T3W, T3X);
 
738
                                                       T3Y = VSUB(T3W, T3X);
 
739
                                                       Tbd = VSUB(Tb9, Tbc);
 
740
                                                       Td9 = VADD(Tb9, Tbc);
 
741
                                                       TeW = VFMA(LDK(KP414213562), TeL, TeM);
 
742
                                                       TeN = VFNMS(LDK(KP414213562), TeM, TeL);
 
743
                                                       T7s = VFNMS(LDK(KP707106781), T2V, T2U);
 
744
                                                       T2W = VFMA(LDK(KP707106781), T2V, T2U);
 
745
                                                       T7r = VFNMS(LDK(KP707106781), T2Q, T2J);
 
746
                                                       T2R = VFMA(LDK(KP707106781), T2Q, T2J);
 
747
                                                       TeP = VSUB(Tbh, Tbi);
 
748
                                                       Tbj = VADD(Tbh, Tbi);
 
749
                                                       T37 = VADD(T33, T36);
 
750
                                                       T3c = VSUB(T33, T36);
 
751
                                                       T48 = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
 
752
                                                  }
 
753
                                                  T2X = VFNMS(LDK(KP198912367), T2W, T2R);
 
754
                                                  T3q = VFMA(LDK(KP198912367), T2R, T2W);
 
755
                                                  T7t = VFMA(LDK(KP668178637), T7s, T7r);
 
756
                                                  T7C = VFNMS(LDK(KP668178637), T7r, T7s);
 
757
                                                  {
 
758
                                                       V Tbk, Tda, TeX, TeQ;
 
759
                                                       Tbk = VSUB(Tbg, Tbj);
 
760
                                                       Tda = VADD(Tbg, Tbj);
 
761
                                                       TeX = VFNMS(LDK(KP414213562), TeO, TeP);
 
762
                                                       TeQ = VFMA(LDK(KP414213562), TeP, TeO);
 
763
                                                       {
 
764
                                                            V T7v, T3d, T7u, T38, TbF;
 
765
                                                            T7v = VFNMS(LDK(KP707106781), T3c, T3b);
 
766
                                                            T3d = VFMA(LDK(KP707106781), T3c, T3b);
 
767
                                                            T7u = VFNMS(LDK(KP707106781), T37, T30);
 
768
                                                            T38 = VFMA(LDK(KP707106781), T37, T30);
 
769
                                                            T49 = VSUB(T47, T48);
 
770
                                                            TbF = VADD(T48, T47);
 
771
                                                            TdL = VSUB(Td9, Tda);
 
772
                                                            Tdb = VADD(Td9, Tda);
 
773
                                                            Tbu = VSUB(Tbd, Tbk);
 
774
                                                            Tbl = VADD(Tbd, Tbk);
 
775
                                                            Tgu = VSUB(TeN, TeQ);
 
776
                                                            TeR = VADD(TeN, TeQ);
 
777
                                                            Tgr = VSUB(TeW, TeX);
 
778
                                                            TeY = VADD(TeW, TeX);
 
779
                                                            T3r = VFNMS(LDK(KP198912367), T38, T3d);
 
780
                                                            T3e = VFMA(LDK(KP198912367), T3d, T38);
 
781
                                                            T7D = VFMA(LDK(KP668178637), T7u, T7v);
 
782
                                                            T7w = VFNMS(LDK(KP668178637), T7v, T7u);
 
783
                                                            Tf6 = VSUB(TbE, TbF);
 
784
                                                            TbG = VADD(TbE, TbF);
 
785
                                                       }
 
786
                                                  }
 
787
                                             }
 
788
                                             T4d = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)]));
 
789
                                             T5V = VSUB(T3q, T3r);
 
790
                                             T3s = VADD(T3q, T3r);
 
791
                                             T5Y = VSUB(T2X, T3e);
 
792
                                             T3f = VADD(T2X, T3e);
 
793
                                             T95 = VSUB(T7D, T7C);
 
794
                                             T7E = VADD(T7C, T7D);
 
795
                                             T98 = VSUB(T7t, T7w);
 
796
                                             T7x = VADD(T7t, T7w);
 
797
                                             T4e = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
 
798
                                             T4o = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)]));
 
799
                                             T4p = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
 
800
                                        }
 
801
                                        {
 
802
                                             V T3Z, T40, T42, T43, TbL, TbM;
 
803
                                             T3Z = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
 
804
                                             T40 = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)]));
 
805
                                             T42 = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)]));
 
806
                                             T43 = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
 
807
                                             T4g = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
 
808
                                             T4f = VSUB(T4d, T4e);
 
809
                                             TbL = VADD(T4d, T4e);
 
810
                                             T4q = VSUB(T4o, T4p);
 
811
                                             TbM = VADD(T4p, T4o);
 
812
                                             TbH = VADD(T3Z, T40);
 
813
                                             T41 = VSUB(T3Z, T40);
 
814
                                             TbI = VADD(T42, T43);
 
815
                                             T44 = VSUB(T42, T43);
 
816
                                             T4h = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)]));
 
817
                                             T4j = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)]));
 
818
                                             T4k = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
 
819
                                             Tf9 = VSUB(TbL, TbM);
 
820
                                             TbN = VADD(TbL, TbM);
 
821
                                        }
 
822
                                   }
 
823
                              }
 
824
                         }
 
825
                    }
 
826
                    {
 
827
                         V TgB, Tgy, T62, T4H, T65, T4u, T9c, T7X, T9f, T7Q, Tg0, Tga, TfF, TeF, TfT;
 
828
                         V TfU, TfP, Tg7, TfI, Tfy, Tfz, Tf0, TfA, Tfl, Tg1, TfS;
 
829
                         {
 
830
                              V Tc1, TbS, Tfc, Tfj, TdX, Te5, TdZ, TdR, Te7, Te3, TdU, Te4;
 
831
                              {
 
832
                                   V TdF, TdS, Tdx, Td5, TdO, TdE, TdC, Tdt, Tdk;
 
833
                                   {
 
834
                                        V Tdc, TdA, T4F, T4c, T7V, T7M, T4G, T4t, T7W, T7P, TdB, Tdj;
 
835
                                        {
 
836
                                             V Td1, Tdg, TbK, Tf8, Tfh, T4b, T7L, T46, T7K, TbQ, Tfa, T4r, T4m, Td4;
 
837
                                             TdF = VSUB(TcZ, Td0);
 
838
                                             Td1 = VADD(TcZ, Td0);
 
839
                                             {
 
840
                                                  V TbJ, Tf7, T4a, T45;
 
841
                                                  TbJ = VADD(TbH, TbI);
 
842
                                                  Tf7 = VSUB(TbI, TbH);
 
843
                                                  T4a = VSUB(T44, T41);
 
844
                                                  T45 = VADD(T41, T44);
 
845
                                                  {
 
846
                                                       V TbO, T4i, TbP, T4l;
 
847
                                                       TbO = VADD(T4g, T4h);
 
848
                                                       T4i = VSUB(T4g, T4h);
 
849
                                                       TbP = VADD(T4j, T4k);
 
850
                                                       T4l = VSUB(T4j, T4k);
 
851
                                                       Tdg = VADD(TbG, TbJ);
 
852
                                                       TbK = VSUB(TbG, TbJ);
 
853
                                                       Tf8 = VFMA(LDK(KP414213562), Tf7, Tf6);
 
854
                                                       Tfh = VFNMS(LDK(KP414213562), Tf6, Tf7);
 
855
                                                       T4b = VFMA(LDK(KP707106781), T4a, T49);
 
856
                                                       T7L = VFNMS(LDK(KP707106781), T4a, T49);
 
857
                                                       T46 = VFMA(LDK(KP707106781), T45, T3Y);
 
858
                                                       T7K = VFNMS(LDK(KP707106781), T45, T3Y);
 
859
                                                       TbQ = VADD(TbO, TbP);
 
860
                                                       Tfa = VSUB(TbP, TbO);
 
861
                                                       T4r = VSUB(T4l, T4i);
 
862
                                                       T4m = VADD(T4i, T4l);
 
863
                                                       Td4 = VADD(Td2, Td3);
 
864
                                                       TdS = VSUB(Td2, Td3);
 
865
                                                  }
 
866
                                             }
 
867
                                             Tdc = VSUB(Td8, Tdb);
 
868
                                             TdA = VADD(Td8, Tdb);
 
869
                                             T4F = VFNMS(LDK(KP198912367), T46, T4b);
 
870
                                             T4c = VFMA(LDK(KP198912367), T4b, T46);
 
871
                                             T7V = VFMA(LDK(KP668178637), T7K, T7L);
 
872
                                             T7M = VFNMS(LDK(KP668178637), T7L, T7K);
 
873
                                             {
 
874
                                                  V Tdh, TbR, Tfb, Tfi;
 
875
                                                  Tdh = VADD(TbN, TbQ);
 
876
                                                  TbR = VSUB(TbN, TbQ);
 
877
                                                  Tfb = VFNMS(LDK(KP414213562), Tfa, Tf9);
 
878
                                                  Tfi = VFMA(LDK(KP414213562), Tf9, Tfa);
 
879
                                                  {
 
880
                                                       V T4s, T7O, T4n, T7N, Tdi;
 
881
                                                       T4s = VFMA(LDK(KP707106781), T4r, T4q);
 
882
                                                       T7O = VFNMS(LDK(KP707106781), T4r, T4q);
 
883
                                                       T4n = VFMA(LDK(KP707106781), T4m, T4f);
 
884
                                                       T7N = VFNMS(LDK(KP707106781), T4m, T4f);
 
885
                                                       Tdx = VADD(Td1, Td4);
 
886
                                                       Td5 = VSUB(Td1, Td4);
 
887
                                                       TdO = VSUB(Tdh, Tdg);
 
888
                                                       Tdi = VADD(Tdg, Tdh);
 
889
                                                       Tc1 = VSUB(TbR, TbK);
 
890
                                                       TbS = VADD(TbK, TbR);
 
891
                                                       TgB = VSUB(Tfb, Tf8);
 
892
                                                       Tfc = VADD(Tf8, Tfb);
 
893
                                                       Tgy = VSUB(Tfi, Tfh);
 
894
                                                       Tfj = VADD(Tfh, Tfi);
 
895
                                                       T4G = VFMA(LDK(KP198912367), T4n, T4s);
 
896
                                                       T4t = VFNMS(LDK(KP198912367), T4s, T4n);
 
897
                                                       T7W = VFNMS(LDK(KP668178637), T7N, T7O);
 
898
                                                       T7P = VFMA(LDK(KP668178637), T7O, T7N);
 
899
                                                       TdB = VADD(Tdf, Tdi);
 
900
                                                       Tdj = VSUB(Tdf, Tdi);
 
901
                                                  }
 
902
                                             }
 
903
                                        }
 
904
                                        T62 = VSUB(T4G, T4F);
 
905
                                        T4H = VADD(T4F, T4G);
 
906
                                        T65 = VSUB(T4t, T4c);
 
907
                                        T4u = VADD(T4c, T4t);
 
908
                                        T9c = VSUB(T7V, T7W);
 
909
                                        T7X = VADD(T7V, T7W);
 
910
                                        T9f = VSUB(T7P, T7M);
 
911
                                        T7Q = VADD(T7M, T7P);
 
912
                                        TdE = VADD(TdA, TdB);
 
913
                                        TdC = VSUB(TdA, TdB);
 
914
                                        Tdt = VSUB(Tdc, Tdj);
 
915
                                        Tdk = VADD(Tdc, Tdj);
 
916
                                   }
 
917
                                   {
 
918
                                        V TdT, Tdl, Tdv, TdJ, Te1, Te2, TdQ, Tdz, TdD, Tdu, Tdw;
 
919
                                        {
 
920
                                             V TdI, TdM, TdV, TdW, TdP, Tds, Tdy;
 
921
                                             TdI = VADD(TdG, TdH);
 
922
                                             TdT = VSUB(TdG, TdH);
 
923
                                             TdM = VFNMS(LDK(KP414213562), TdL, TdK);
 
924
                                             TdV = VFMA(LDK(KP414213562), TdK, TdL);
 
925
                                             TdW = VFMA(LDK(KP414213562), TdN, TdO);
 
926
                                             TdP = VFNMS(LDK(KP414213562), TdO, TdN);
 
927
                                             Tdl = VFNMS(LDK(KP707106781), Tdk, Td5);
 
928
                                             Tdv = VFMA(LDK(KP707106781), Tdk, Td5);
 
929
                                             Tds = VSUB(Tdo, Tdr);
 
930
                                             Tdy = VADD(Tdo, Tdr);
 
931
                                             TdJ = VFMA(LDK(KP707106781), TdI, TdF);
 
932
                                             Te1 = VFNMS(LDK(KP707106781), TdI, TdF);
 
933
                                             TdX = VSUB(TdV, TdW);
 
934
                                             Te2 = VADD(TdV, TdW);
 
935
                                             Te5 = VSUB(TdM, TdP);
 
936
                                             TdQ = VADD(TdM, TdP);
 
937
                                             Tdz = VSUB(Tdx, Tdy);
 
938
                                             TdD = VADD(Tdx, Tdy);
 
939
                                             Tdu = VFNMS(LDK(KP707106781), Tdt, Tds);
 
940
                                             Tdw = VFMA(LDK(KP707106781), Tdt, Tds);
 
941
                                        }
 
942
                                        TdZ = VFMA(LDK(KP923879532), TdQ, TdJ);
 
943
                                        TdR = VFNMS(LDK(KP923879532), TdQ, TdJ);
 
944
                                        Te7 = VFMA(LDK(KP923879532), Te2, Te1);
 
945
                                        Te3 = VFNMS(LDK(KP923879532), Te2, Te1);
 
946
                                        ST(&(xo[0]), VADD(TdD, TdE), ovs, &(xo[0]));
 
947
                                        ST(&(xo[WS(os, 64)]), VSUB(TdD, TdE), ovs, &(xo[0]));
 
948
                                        ST(&(xo[WS(os, 32)]), VFMAI(TdC, Tdz), ovs, &(xo[0]));
 
949
                                        ST(&(xo[WS(os, 96)]), VFNMSI(TdC, Tdz), ovs, &(xo[0]));
 
950
                                        ST(&(xo[WS(os, 112)]), VFNMSI(Tdw, Tdv), ovs, &(xo[0]));
 
951
                                        ST(&(xo[WS(os, 16)]), VFMAI(Tdw, Tdv), ovs, &(xo[0]));
 
952
                                        ST(&(xo[WS(os, 80)]), VFMAI(Tdu, Tdl), ovs, &(xo[0]));
 
953
                                        ST(&(xo[WS(os, 48)]), VFNMSI(Tdu, Tdl), ovs, &(xo[0]));
 
954
                                        TdU = VFMA(LDK(KP707106781), TdT, TdS);
 
955
                                        Te4 = VFNMS(LDK(KP707106781), TdT, TdS);
 
956
                                   }
 
957
                              }
 
958
                              {
 
959
                                   V Tcx, TcJ, TcI, Tcy, TcA, Tbm, Tcp, TaZ, Tcs, Tci, Tbv, TcB, TcD, TbT, Tc2;
 
960
                                   V TcE, Tat, TaY;
 
961
                                   Tcx = VFNMS(LDK(KP707106781), Tas, Tad);
 
962
                                   Tat = VFMA(LDK(KP707106781), Tas, Tad);
 
963
                                   TaY = VADD(TaI, TaX);
 
964
                                   TcJ = VSUB(TaI, TaX);
 
965
                                   {
 
966
                                        V Tce, Tch, Te8, Te6, TdY, Te0;
 
967
                                        TcI = VFNMS(LDK(KP707106781), Tcd, Tcc);
 
968
                                        Tce = VFMA(LDK(KP707106781), Tcd, Tcc);
 
969
                                        Tch = VSUB(Tcf, Tcg);
 
970
                                        Tcy = VADD(Tcf, Tcg);
 
971
                                        Te8 = VFNMS(LDK(KP923879532), Te5, Te4);
 
972
                                        Te6 = VFMA(LDK(KP923879532), Te5, Te4);
 
973
                                        TdY = VFNMS(LDK(KP923879532), TdX, TdU);
 
974
                                        Te0 = VFMA(LDK(KP923879532), TdX, TdU);
 
975
                                        TcA = VFNMS(LDK(KP707106781), Tbl, Tb6);
 
976
                                        Tbm = VFMA(LDK(KP707106781), Tbl, Tb6);
 
977
                                        Tcp = VFNMS(LDK(KP923879532), TaY, Tat);
 
978
                                        TaZ = VFMA(LDK(KP923879532), TaY, Tat);
 
979
                                        Tcs = VFNMS(LDK(KP923879532), Tch, Tce);
 
980
                                        Tci = VFMA(LDK(KP923879532), Tch, Tce);
 
981
                                        ST(&(xo[WS(os, 88)]), VFNMSI(Te6, Te3), ovs, &(xo[0]));
 
982
                                        ST(&(xo[WS(os, 40)]), VFMAI(Te6, Te3), ovs, &(xo[0]));
 
983
                                        ST(&(xo[WS(os, 104)]), VFMAI(Te8, Te7), ovs, &(xo[0]));
 
984
                                        ST(&(xo[WS(os, 24)]), VFNMSI(Te8, Te7), ovs, &(xo[0]));
 
985
                                        ST(&(xo[WS(os, 8)]), VFMAI(Te0, TdZ), ovs, &(xo[0]));
 
986
                                        ST(&(xo[WS(os, 120)]), VFNMSI(Te0, TdZ), ovs, &(xo[0]));
 
987
                                        ST(&(xo[WS(os, 72)]), VFMAI(TdY, TdR), ovs, &(xo[0]));
 
988
                                        ST(&(xo[WS(os, 56)]), VFNMSI(TdY, TdR), ovs, &(xo[0]));
 
989
                                        Tbv = VFMA(LDK(KP707106781), Tbu, Tbt);
 
990
                                        TcB = VFNMS(LDK(KP707106781), Tbu, Tbt);
 
991
                                        TcD = VFNMS(LDK(KP707106781), TbS, TbD);
 
992
                                        TbT = VFMA(LDK(KP707106781), TbS, TbD);
 
993
                                        Tc2 = VFMA(LDK(KP707106781), Tc1, Tc0);
 
994
                                        TcE = VFNMS(LDK(KP707106781), Tc1, Tc0);
 
995
                                   }
 
996
                                   {
 
997
                                        V TcR, Tcz, TcU, TcK, Tcq, Tcl, Tct, Tc4;
 
998
                                        {
 
999
                                             V Tcj, Tbw, Tck, Tc3;
 
1000
                                             Tcj = VFMA(LDK(KP198912367), Tbm, Tbv);
 
1001
                                             Tbw = VFNMS(LDK(KP198912367), Tbv, Tbm);
 
1002
                                             Tck = VFMA(LDK(KP198912367), TbT, Tc2);
 
1003
                                             Tc3 = VFNMS(LDK(KP198912367), Tc2, TbT);
 
1004
                                             TcR = VFNMS(LDK(KP923879532), Tcy, Tcx);
 
1005
                                             Tcz = VFMA(LDK(KP923879532), Tcy, Tcx);
 
1006
                                             TcU = VFMA(LDK(KP923879532), TcJ, TcI);
 
1007
                                             TcK = VFNMS(LDK(KP923879532), TcJ, TcI);
 
1008
                                             Tcq = VADD(Tcj, Tck);
 
1009
                                             Tcl = VSUB(Tcj, Tck);
 
1010
                                             Tct = VSUB(Tbw, Tc3);
 
1011
                                             Tc4 = VADD(Tbw, Tc3);
 
1012
                                        }
 
1013
                                        {
 
1014
                                             V TfN, Tel, TfY, Tfu, Tfv, Tfw, TcT, TcX, TcQ, TcO, TcW, TcY, TcP, TcH, TfZ;
 
1015
                                             V TeE;
 
1016
                                             {
 
1017
                                                  V Teu, TcS, TcN, TcV, TcG, TeD;
 
1018
                                                  TfN = VFNMS(LDK(KP923879532), Tek, Ted);
 
1019
                                                  Tel = VFMA(LDK(KP923879532), Tek, Ted);
 
1020
                                                  {
 
1021
                                                       V TcL, TcC, Tcr, Tcv;
 
1022
                                                       TcL = VFNMS(LDK(KP668178637), TcA, TcB);
 
1023
                                                       TcC = VFMA(LDK(KP668178637), TcB, TcA);
 
1024
                                                       Tcr = VFNMS(LDK(KP980785280), Tcq, Tcp);
 
1025
                                                       Tcv = VFMA(LDK(KP980785280), Tcq, Tcp);
 
1026
                                                       {
 
1027
                                                            V Tco, Tcm, Tcu, Tcw;
 
1028
                                                            Tco = VFMA(LDK(KP980785280), Tcl, Tci);
 
1029
                                                            Tcm = VFNMS(LDK(KP980785280), Tcl, Tci);
 
1030
                                                            Tcu = VFMA(LDK(KP980785280), Tct, Tcs);
 
1031
                                                            Tcw = VFNMS(LDK(KP980785280), Tct, Tcs);
 
1032
                                                            {
 
1033
                                                                 V Tcn, Tc5, TcM, TcF;
 
1034
                                                                 Tcn = VFMA(LDK(KP980785280), Tc4, TaZ);
 
1035
                                                                 Tc5 = VFNMS(LDK(KP980785280), Tc4, TaZ);
 
1036
                                                                 TcM = VFNMS(LDK(KP668178637), TcD, TcE);
 
1037
                                                                 TcF = VFMA(LDK(KP668178637), TcE, TcD);
 
1038
                                                                 TfY = VFNMS(LDK(KP923879532), Tft, Tfq);
 
1039
                                                                 Tfu = VFMA(LDK(KP923879532), Tft, Tfq);
 
1040
                                                                 Tfv = VFMA(LDK(KP198912367), Teq, Tet);
 
1041
                                                                 Teu = VFNMS(LDK(KP198912367), Tet, Teq);
 
1042
                                                                 ST(&(xo[WS(os, 92)]), VFNMSI(Tcu, Tcr), ovs, &(xo[0]));
 
1043
                                                                 ST(&(xo[WS(os, 36)]), VFMAI(Tcu, Tcr), ovs, &(xo[0]));
 
1044
                                                                 ST(&(xo[WS(os, 100)]), VFMAI(Tcw, Tcv), ovs, &(xo[0]));
 
1045
                                                                 ST(&(xo[WS(os, 28)]), VFNMSI(Tcw, Tcv), ovs, &(xo[0]));
 
1046
                                                                 ST(&(xo[WS(os, 4)]), VFMAI(Tco, Tcn), ovs, &(xo[0]));
 
1047
                                                                 ST(&(xo[WS(os, 124)]), VFNMSI(Tco, Tcn), ovs, &(xo[0]));
 
1048
                                                                 ST(&(xo[WS(os, 68)]), VFMAI(Tcm, Tc5), ovs, &(xo[0]));
 
1049
                                                                 ST(&(xo[WS(os, 60)]), VFNMSI(Tcm, Tc5), ovs, &(xo[0]));
 
1050
                                                                 TcS = VADD(TcL, TcM);
 
1051
                                                                 TcN = VSUB(TcL, TcM);
 
1052
                                                                 TcV = VSUB(TcC, TcF);
 
1053
                                                                 TcG = VADD(TcC, TcF);
 
1054
                                                                 TeD = VFNMS(LDK(KP198912367), TeC, Tez);
 
1055
                                                                 Tfw = VFMA(LDK(KP198912367), Tez, TeC);
 
1056
                                                            }
 
1057
                                                       }
 
1058
                                                  }
 
1059
                                                  TcT = VFMA(LDK(KP831469612), TcS, TcR);
 
1060
                                                  TcX = VFNMS(LDK(KP831469612), TcS, TcR);
 
1061
                                                  TcQ = VFMA(LDK(KP831469612), TcN, TcK);
 
1062
                                                  TcO = VFNMS(LDK(KP831469612), TcN, TcK);
 
1063
                                                  TcW = VFNMS(LDK(KP831469612), TcV, TcU);
 
1064
                                                  TcY = VFMA(LDK(KP831469612), TcV, TcU);
 
1065
                                                  TcP = VFMA(LDK(KP831469612), TcG, Tcz);
 
1066
                                                  TcH = VFNMS(LDK(KP831469612), TcG, Tcz);
 
1067
                                                  TfZ = VSUB(Teu, TeD);
 
1068
                                                  TeE = VADD(Teu, TeD);
 
1069
                                             }
 
1070
                                             {
 
1071
                                                  V TfQ, TeS, TfO, Tfx, TeZ, TfR, Tfd, Tfk;
 
1072
                                                  TfQ = VFNMS(LDK(KP923879532), TeR, TeK);
 
1073
                                                  TeS = VFMA(LDK(KP923879532), TeR, TeK);
 
1074
                                                  ST(&(xo[WS(os, 84)]), VFMAI(TcW, TcT), ovs, &(xo[0]));
 
1075
                                                  ST(&(xo[WS(os, 44)]), VFNMSI(TcW, TcT), ovs, &(xo[0]));
 
1076
                                                  ST(&(xo[WS(os, 108)]), VFNMSI(TcY, TcX), ovs, &(xo[0]));
 
1077
                                                  ST(&(xo[WS(os, 20)]), VFMAI(TcY, TcX), ovs, &(xo[0]));
 
1078
                                                  ST(&(xo[WS(os, 116)]), VFMAI(TcQ, TcP), ovs, &(xo[0]));
 
1079
                                                  ST(&(xo[WS(os, 12)]), VFNMSI(TcQ, TcP), ovs, &(xo[0]));
 
1080
                                                  ST(&(xo[WS(os, 52)]), VFMAI(TcO, TcH), ovs, &(xo[0]));
 
1081
                                                  ST(&(xo[WS(os, 76)]), VFNMSI(TcO, TcH), ovs, &(xo[0]));
 
1082
                                                  Tg0 = VFNMS(LDK(KP980785280), TfZ, TfY);
 
1083
                                                  Tga = VFMA(LDK(KP980785280), TfZ, TfY);
 
1084
                                                  TfF = VFNMS(LDK(KP980785280), TeE, Tel);
 
1085
                                                  TeF = VFMA(LDK(KP980785280), TeE, Tel);
 
1086
                                                  TfO = VADD(Tfv, Tfw);
 
1087
                                                  Tfx = VSUB(Tfv, Tfw);
 
1088
                                                  TeZ = VFMA(LDK(KP923879532), TeY, TeV);
 
1089
                                                  TfR = VFNMS(LDK(KP923879532), TeY, TeV);
 
1090
                                                  TfT = VFNMS(LDK(KP923879532), Tfc, Tf5);
 
1091
                                                  Tfd = VFMA(LDK(KP923879532), Tfc, Tf5);
 
1092
                                                  Tfk = VFMA(LDK(KP923879532), Tfj, Tfg);
 
1093
                                                  TfU = VFNMS(LDK(KP923879532), Tfj, Tfg);
 
1094
                                                  TfP = VFMA(LDK(KP980785280), TfO, TfN);
 
1095
                                                  Tg7 = VFNMS(LDK(KP980785280), TfO, TfN);
 
1096
                                                  TfI = VFNMS(LDK(KP980785280), Tfx, Tfu);
 
1097
                                                  Tfy = VFMA(LDK(KP980785280), Tfx, Tfu);
 
1098
                                                  Tfz = VFMA(LDK(KP098491403), TeS, TeZ);
 
1099
                                                  Tf0 = VFNMS(LDK(KP098491403), TeZ, TeS);
 
1100
                                                  TfA = VFMA(LDK(KP098491403), Tfd, Tfk);
 
1101
                                                  Tfl = VFNMS(LDK(KP098491403), Tfk, Tfd);
 
1102
                                                  Tg1 = VFNMS(LDK(KP820678790), TfQ, TfR);
 
1103
                                                  TfS = VFMA(LDK(KP820678790), TfR, TfQ);
 
1104
                                             }
 
1105
                                        }
 
1106
                                   }
 
1107
                              }
 
1108
                         }
 
1109
                         {
 
1110
                              V T8x, T8y, T8F, T8w, T8k, T8f, T8n, T80, T9l, T76, T87, T8U, T89, T7e, T7l;
 
1111
                              V T8a;
 
1112
                              {
 
1113
                                   V The, Tho, TgT, Tgp, Th7, Th8, Thf, Th6, Th3, Thl, TgW, TgM, TgU, TgP, TgX;
 
1114
                                   V TgE;
 
1115
                                   {
 
1116
                                        V Th1, TgI, TgJ, TgK;
 
1117
                                        {
 
1118
                                             V Tgh, Thc, Tgk, TfG, TfB, TfJ, Tfm, Tg2, TfV, Tgn, TfL, TfH;
 
1119
                                             Th1 = VFMA(LDK(KP923879532), Tgg, Tgf);
 
1120
                                             Tgh = VFNMS(LDK(KP923879532), Tgg, Tgf);
 
1121
                                             Thc = VFNMS(LDK(KP923879532), TgH, TgG);
 
1122
                                             TgI = VFMA(LDK(KP923879532), TgH, TgG);
 
1123
                                             TgJ = VFMA(LDK(KP668178637), Tgi, Tgj);
 
1124
                                             Tgk = VFNMS(LDK(KP668178637), Tgj, Tgi);
 
1125
                                             TfG = VADD(Tfz, TfA);
 
1126
                                             TfB = VSUB(Tfz, TfA);
 
1127
                                             TfJ = VSUB(Tf0, Tfl);
 
1128
                                             Tfm = VADD(Tf0, Tfl);
 
1129
                                             Tg2 = VFNMS(LDK(KP820678790), TfT, TfU);
 
1130
                                             TfV = VFMA(LDK(KP820678790), TfU, TfT);
 
1131
                                             Tgn = VFNMS(LDK(KP668178637), Tgm, Tgl);
 
1132
                                             TgK = VFMA(LDK(KP668178637), Tgl, Tgm);
 
1133
                                             TfL = VFMA(LDK(KP995184726), TfG, TfF);
 
1134
                                             TfH = VFNMS(LDK(KP995184726), TfG, TfF);
 
1135
                                             {
 
1136
                                                  V TfE, TfC, TfM, TfK;
 
1137
                                                  TfE = VFMA(LDK(KP995184726), TfB, Tfy);
 
1138
                                                  TfC = VFNMS(LDK(KP995184726), TfB, Tfy);
 
1139
                                                  TfM = VFNMS(LDK(KP995184726), TfJ, TfI);
 
1140
                                                  TfK = VFMA(LDK(KP995184726), TfJ, TfI);
 
1141
                                                  {
 
1142
                                                       V TfD, Tfn, Tg8, Tg3;
 
1143
                                                       TfD = VFMA(LDK(KP995184726), Tfm, TeF);
 
1144
                                                       Tfn = VFNMS(LDK(KP995184726), Tfm, TeF);
 
1145
                                                       Tg8 = VADD(Tg1, Tg2);
 
1146
                                                       Tg3 = VSUB(Tg1, Tg2);
 
1147
                                                       {
 
1148
                                                            V Tgb, TfW, Thd, Tgo;
 
1149
                                                            Tgb = VSUB(TfS, TfV);
 
1150
                                                            TfW = VADD(TfS, TfV);
 
1151
                                                            Thd = VSUB(Tgk, Tgn);
 
1152
                                                            Tgo = VADD(Tgk, Tgn);
 
1153
                                                            ST(&(xo[WS(os, 98)]), VFMAI(TfM, TfL), ovs, &(xo[0]));
 
1154
                                                            ST(&(xo[WS(os, 30)]), VFNMSI(TfM, TfL), ovs, &(xo[0]));
 
1155
                                                            ST(&(xo[WS(os, 94)]), VFNMSI(TfK, TfH), ovs, &(xo[0]));
 
1156
                                                            ST(&(xo[WS(os, 34)]), VFMAI(TfK, TfH), ovs, &(xo[0]));
 
1157
                                                            ST(&(xo[WS(os, 2)]), VFMAI(TfE, TfD), ovs, &(xo[0]));
 
1158
                                                            ST(&(xo[WS(os, 126)]), VFNMSI(TfE, TfD), ovs, &(xo[0]));
 
1159
                                                            ST(&(xo[WS(os, 66)]), VFMAI(TfC, Tfn), ovs, &(xo[0]));
 
1160
                                                            ST(&(xo[WS(os, 62)]), VFNMSI(TfC, Tfn), ovs, &(xo[0]));
 
1161
                                                            {
 
1162
                                                                 V Tgd, Tg9, Tg6, Tg4;
 
1163
                                                                 Tgd = VFNMS(LDK(KP773010453), Tg8, Tg7);
 
1164
                                                                 Tg9 = VFMA(LDK(KP773010453), Tg8, Tg7);
 
1165
                                                                 Tg6 = VFMA(LDK(KP773010453), Tg3, Tg0);
 
1166
                                                                 Tg4 = VFNMS(LDK(KP773010453), Tg3, Tg0);
 
1167
                                                                 {
 
1168
                                                                      V Tge, Tgc, Tg5, TfX;
 
1169
                                                                      Tge = VFMA(LDK(KP773010453), Tgb, Tga);
 
1170
                                                                      Tgc = VFNMS(LDK(KP773010453), Tgb, Tga);
 
1171
                                                                      Tg5 = VFMA(LDK(KP773010453), TfW, TfP);
 
1172
                                                                      TfX = VFNMS(LDK(KP773010453), TfW, TfP);
 
1173
                                                                      The = VFMA(LDK(KP831469612), Thd, Thc);
 
1174
                                                                      Tho = VFNMS(LDK(KP831469612), Thd, Thc);
 
1175
                                                                      TgT = VFMA(LDK(KP831469612), Tgo, Tgh);
 
1176
                                                                      Tgp = VFNMS(LDK(KP831469612), Tgo, Tgh);
 
1177
                                                                      ST(&(xo[WS(os, 110)]), VFNMSI(Tge, Tgd), ovs, &(xo[0]));
 
1178
                                                                      ST(&(xo[WS(os, 18)]), VFMAI(Tge, Tgd), ovs, &(xo[0]));
 
1179
                                                                      ST(&(xo[WS(os, 82)]), VFMAI(Tgc, Tg9), ovs, &(xo[0]));
 
1180
                                                                      ST(&(xo[WS(os, 46)]), VFNMSI(Tgc, Tg9), ovs, &(xo[0]));
 
1181
                                                                      ST(&(xo[WS(os, 114)]), VFMAI(Tg6, Tg5), ovs, &(xo[0]));
 
1182
                                                                      ST(&(xo[WS(os, 14)]), VFNMSI(Tg6, Tg5), ovs, &(xo[0]));
 
1183
                                                                      ST(&(xo[WS(os, 50)]), VFMAI(Tg4, TfX), ovs, &(xo[0]));
 
1184
                                                                      ST(&(xo[WS(os, 78)]), VFNMSI(Tg4, TfX), ovs, &(xo[0]));
 
1185
                                                                 }
 
1186
                                                            }
 
1187
                                                       }
 
1188
                                                  }
 
1189
                                             }
 
1190
                                        }
 
1191
                                        {
 
1192
                                             V Th4, Tgs, Tgv, Th5, Tgz, TgC, Th2, TgL;
 
1193
                                             Th4 = VFMA(LDK(KP923879532), Tgr, Tgq);
 
1194
                                             Tgs = VFNMS(LDK(KP923879532), Tgr, Tgq);
 
1195
                                             Tgv = VFMA(LDK(KP923879532), Tgu, Tgt);
 
1196
                                             Th5 = VFNMS(LDK(KP923879532), Tgu, Tgt);
 
1197
                                             Th7 = VFMA(LDK(KP923879532), Tgy, Tgx);
 
1198
                                             Tgz = VFNMS(LDK(KP923879532), Tgy, Tgx);
 
1199
                                             TgC = VFMA(LDK(KP923879532), TgB, TgA);
 
1200
                                             Th8 = VFNMS(LDK(KP923879532), TgB, TgA);
 
1201
                                             Th2 = VADD(TgJ, TgK);
 
1202
                                             TgL = VSUB(TgJ, TgK);
 
1203
                                             {
 
1204
                                                  V TgN, Tgw, TgO, TgD;
 
1205
                                                  TgN = VFMA(LDK(KP534511135), Tgs, Tgv);
 
1206
                                                  Tgw = VFNMS(LDK(KP534511135), Tgv, Tgs);
 
1207
                                                  TgO = VFMA(LDK(KP534511135), Tgz, TgC);
 
1208
                                                  TgD = VFNMS(LDK(KP534511135), TgC, Tgz);
 
1209
                                                  Thf = VFNMS(LDK(KP303346683), Th4, Th5);
 
1210
                                                  Th6 = VFMA(LDK(KP303346683), Th5, Th4);
 
1211
                                                  Th3 = VFMA(LDK(KP831469612), Th2, Th1);
 
1212
                                                  Thl = VFNMS(LDK(KP831469612), Th2, Th1);
 
1213
                                                  TgW = VFNMS(LDK(KP831469612), TgL, TgI);
 
1214
                                                  TgM = VFMA(LDK(KP831469612), TgL, TgI);
 
1215
                                                  TgU = VADD(TgN, TgO);
 
1216
                                                  TgP = VSUB(TgN, TgO);
 
1217
                                                  TgX = VSUB(Tgw, TgD);
 
1218
                                                  TgE = VADD(Tgw, TgD);
 
1219
                                             }
 
1220
                                        }
 
1221
                                   }
 
1222
                                   {
 
1223
                                        V T8u, T8v, T7R, T8d, T7G, Thm, Thh, Thp, Tha, T7Y, Thr, Thn;
 
1224
                                        {
 
1225
                                             V T7y, T7F, TgZ, TgV;
 
1226
                                             T8u = VFNMS(LDK(KP831469612), T7x, T7q);
 
1227
                                             T7y = VFMA(LDK(KP831469612), T7x, T7q);
 
1228
                                             T7F = VFMA(LDK(KP831469612), T7E, T7B);
 
1229
                                             T8v = VFNMS(LDK(KP831469612), T7E, T7B);
 
1230
                                             T8x = VFNMS(LDK(KP831469612), T7Q, T7J);
 
1231
                                             T7R = VFMA(LDK(KP831469612), T7Q, T7J);
 
1232
                                             TgZ = VFMA(LDK(KP881921264), TgU, TgT);
 
1233
                                             TgV = VFNMS(LDK(KP881921264), TgU, TgT);
 
1234
                                             {
 
1235
                                                  V TgS, TgQ, Th0, TgY;
 
1236
                                                  TgS = VFMA(LDK(KP881921264), TgP, TgM);
 
1237
                                                  TgQ = VFNMS(LDK(KP881921264), TgP, TgM);
 
1238
                                                  Th0 = VFNMS(LDK(KP881921264), TgX, TgW);
 
1239
                                                  TgY = VFMA(LDK(KP881921264), TgX, TgW);
 
1240
                                                  {
 
1241
                                                       V TgR, TgF, Thg, Th9;
 
1242
                                                       TgR = VFMA(LDK(KP881921264), TgE, Tgp);
 
1243
                                                       TgF = VFNMS(LDK(KP881921264), TgE, Tgp);
 
1244
                                                       Thg = VFNMS(LDK(KP303346683), Th7, Th8);
 
1245
                                                       Th9 = VFMA(LDK(KP303346683), Th8, Th7);
 
1246
                                                       T8d = VFNMS(LDK(KP148335987), T7y, T7F);
 
1247
                                                       T7G = VFMA(LDK(KP148335987), T7F, T7y);
 
1248
                                                       ST(&(xo[WS(os, 106)]), VFMAI(Th0, TgZ), ovs, &(xo[0]));
 
1249
                                                       ST(&(xo[WS(os, 22)]), VFNMSI(Th0, TgZ), ovs, &(xo[0]));
 
1250
                                                       ST(&(xo[WS(os, 86)]), VFNMSI(TgY, TgV), ovs, &(xo[0]));
 
1251
                                                       ST(&(xo[WS(os, 42)]), VFMAI(TgY, TgV), ovs, &(xo[0]));
 
1252
                                                       ST(&(xo[WS(os, 10)]), VFMAI(TgS, TgR), ovs, &(xo[0]));
 
1253
                                                       ST(&(xo[WS(os, 118)]), VFNMSI(TgS, TgR), ovs, &(xo[0]));
 
1254
                                                       ST(&(xo[WS(os, 74)]), VFMAI(TgQ, TgF), ovs, &(xo[0]));
 
1255
                                                       ST(&(xo[WS(os, 54)]), VFNMSI(TgQ, TgF), ovs, &(xo[0]));
 
1256
                                                       Thm = VADD(Thf, Thg);
 
1257
                                                       Thh = VSUB(Thf, Thg);
 
1258
                                                       Thp = VSUB(Th6, Th9);
 
1259
                                                       Tha = VADD(Th6, Th9);
 
1260
                                                       T7Y = VFMA(LDK(KP831469612), T7X, T7U);
 
1261
                                                       T8y = VFNMS(LDK(KP831469612), T7X, T7U);
 
1262
                                                  }
 
1263
                                             }
 
1264
                                        }
 
1265
                                        Thr = VFNMS(LDK(KP956940335), Thm, Thl);
 
1266
                                        Thn = VFMA(LDK(KP956940335), Thm, Thl);
 
1267
                                        {
 
1268
                                             V Thk, Thi, Ths, Thq;
 
1269
                                             Thk = VFMA(LDK(KP956940335), Thh, The);
 
1270
                                             Thi = VFNMS(LDK(KP956940335), Thh, The);
 
1271
                                             Ths = VFMA(LDK(KP956940335), Thp, Tho);
 
1272
                                             Thq = VFNMS(LDK(KP956940335), Thp, Tho);
 
1273
                                             {
 
1274
                                                  V Thj, Thb, T8e, T7Z;
 
1275
                                                  Thj = VFMA(LDK(KP956940335), Tha, Th3);
 
1276
                                                  Thb = VFNMS(LDK(KP956940335), Tha, Th3);
 
1277
                                                  T8e = VFNMS(LDK(KP148335987), T7R, T7Y);
 
1278
                                                  T7Z = VFMA(LDK(KP148335987), T7Y, T7R);
 
1279
                                                  T8F = VFMA(LDK(KP741650546), T8u, T8v);
 
1280
                                                  T8w = VFNMS(LDK(KP741650546), T8v, T8u);
 
1281
                                                  ST(&(xo[WS(os, 102)]), VFNMSI(Ths, Thr), ovs, &(xo[0]));
 
1282
                                                  ST(&(xo[WS(os, 26)]), VFMAI(Ths, Thr), ovs, &(xo[0]));
 
1283
                                                  ST(&(xo[WS(os, 90)]), VFMAI(Thq, Thn), ovs, &(xo[0]));
 
1284
                                                  ST(&(xo[WS(os, 38)]), VFNMSI(Thq, Thn), ovs, &(xo[0]));
 
1285
                                                  ST(&(xo[WS(os, 122)]), VFMAI(Thk, Thj), ovs, &(xo[0]));
 
1286
                                                  ST(&(xo[WS(os, 6)]), VFNMSI(Thk, Thj), ovs, &(xo[0]));
 
1287
                                                  ST(&(xo[WS(os, 58)]), VFMAI(Thi, Thb), ovs, &(xo[0]));
 
1288
                                                  ST(&(xo[WS(os, 70)]), VFNMSI(Thi, Thb), ovs, &(xo[0]));
 
1289
                                                  T8k = VADD(T8d, T8e);
 
1290
                                                  T8f = VSUB(T8d, T8e);
 
1291
                                                  T8n = VSUB(T7G, T7Z);
 
1292
                                                  T80 = VADD(T7G, T7Z);
 
1293
                                             }
 
1294
                                        }
 
1295
                                        T9l = VSUB(T72, T75);
 
1296
                                        T76 = VADD(T72, T75);
 
1297
                                        T87 = VSUB(T85, T86);
 
1298
                                        T8U = VADD(T85, T86);
 
1299
                                        T89 = VFNMS(LDK(KP303346683), T7a, T7d);
 
1300
                                        T7e = VFMA(LDK(KP303346683), T7d, T7a);
 
1301
                                        T7l = VFMA(LDK(KP303346683), T7k, T7h);
 
1302
                                        T8a = VFNMS(LDK(KP303346683), T7h, T7k);
 
1303
                                   }
 
1304
                              }
 
1305
                              {
 
1306
                                   V T11, T5h, T5a, T55, T5d, T4K, T5C, T5x, T5F, T5q, T4X, T4Z, T1C, T2d, T50;
 
1307
                                   {
 
1308
                                        V T5k, T3g, T3t, T5l, T5n, T4v, T4I, T5o, T8G, T8z;
 
1309
                                        T5k = VFNMS(LDK(KP980785280), T3f, T2G);
 
1310
                                        T3g = VFMA(LDK(KP980785280), T3f, T2G);
 
1311
                                        T8G = VFMA(LDK(KP741650546), T8x, T8y);
 
1312
                                        T8z = VFNMS(LDK(KP741650546), T8y, T8x);
 
1313
                                        {
 
1314
                                             V T8r, T77, T8C, T88;
 
1315
                                             T8r = VFNMS(LDK(KP831469612), T76, T6Z);
 
1316
                                             T77 = VFMA(LDK(KP831469612), T76, T6Z);
 
1317
                                             T8C = VFNMS(LDK(KP831469612), T87, T84);
 
1318
                                             T88 = VFMA(LDK(KP831469612), T87, T84);
 
1319
                                             {
 
1320
                                                  V T8D, T7m, T8s, T8b;
 
1321
                                                  T8D = VSUB(T7e, T7l);
 
1322
                                                  T7m = VADD(T7e, T7l);
 
1323
                                                  T8s = VADD(T89, T8a);
 
1324
                                                  T8b = VSUB(T89, T8a);
 
1325
                                                  {
 
1326
                                                       V T8M, T8H, T8P, T8A;
 
1327
                                                       T8M = VADD(T8F, T8G);
 
1328
                                                       T8H = VSUB(T8F, T8G);
 
1329
                                                       T8P = VSUB(T8w, T8z);
 
1330
                                                       T8A = VADD(T8w, T8z);
 
1331
                                                       {
 
1332
                                                            V T8E, T8O, T8j, T7n;
 
1333
                                                            T8E = VFMA(LDK(KP956940335), T8D, T8C);
 
1334
                                                            T8O = VFNMS(LDK(KP956940335), T8D, T8C);
 
1335
                                                            T8j = VFNMS(LDK(KP956940335), T7m, T77);
 
1336
                                                            T7n = VFMA(LDK(KP956940335), T7m, T77);
 
1337
                                                            {
 
1338
                                                                 V T8t, T8L, T8m, T8c;
 
1339
                                                                 T8t = VFNMS(LDK(KP956940335), T8s, T8r);
 
1340
                                                                 T8L = VFMA(LDK(KP956940335), T8s, T8r);
 
1341
                                                                 T8m = VFNMS(LDK(KP956940335), T8b, T88);
 
1342
                                                                 T8c = VFMA(LDK(KP956940335), T8b, T88);
 
1343
                                                                 {
 
1344
                                                                      V T8K, T8I, T8S, T8Q;
 
1345
                                                                      T8K = VFMA(LDK(KP803207531), T8H, T8E);
 
1346
                                                                      T8I = VFNMS(LDK(KP803207531), T8H, T8E);
 
1347
                                                                      T8S = VFNMS(LDK(KP803207531), T8P, T8O);
 
1348
                                                                      T8Q = VFMA(LDK(KP803207531), T8P, T8O);
 
1349
                                                                      {
 
1350
                                                                           V T8p, T8l, T8h, T81;
 
1351
                                                                           T8p = VFNMS(LDK(KP989176509), T8k, T8j);
 
1352
                                                                           T8l = VFMA(LDK(KP989176509), T8k, T8j);
 
1353
                                                                           T8h = VFMA(LDK(KP989176509), T80, T7n);
 
1354
                                                                           T81 = VFNMS(LDK(KP989176509), T80, T7n);
 
1355
                                                                           {
 
1356
                                                                                V T8J, T8B, T8R, T8N;
 
1357
                                                                                T8J = VFMA(LDK(KP803207531), T8A, T8t);
 
1358
                                                                                T8B = VFNMS(LDK(KP803207531), T8A, T8t);
 
1359
                                                                                T8R = VFMA(LDK(KP803207531), T8M, T8L);
 
1360
                                                                                T8N = VFNMS(LDK(KP803207531), T8M, T8L);
 
1361
                                                                                {
 
1362
                                                                                     V T8q, T8o, T8i, T8g;
 
1363
                                                                                     T8q = VFMA(LDK(KP989176509), T8n, T8m);
 
1364
                                                                                     T8o = VFNMS(LDK(KP989176509), T8n, T8m);
 
1365
                                                                                     T8i = VFMA(LDK(KP989176509), T8f, T8c);
 
1366
                                                                                     T8g = VFNMS(LDK(KP989176509), T8f, T8c);
 
1367
                                                                                     ST(&(xo[WS(os, 13)]), VFMAI(T8K, T8J), ovs, &(xo[WS(os, 1)]));
 
1368
                                                                                     ST(&(xo[WS(os, 115)]), VFNMSI(T8K, T8J), ovs, &(xo[WS(os, 1)]));
 
1369
                                                                                     ST(&(xo[WS(os, 77)]), VFMAI(T8I, T8B), ovs, &(xo[WS(os, 1)]));
 
1370
                                                                                     ST(&(xo[WS(os, 51)]), VFNMSI(T8I, T8B), ovs, &(xo[WS(os, 1)]));
 
1371
                                                                                     ST(&(xo[WS(os, 109)]), VFMAI(T8S, T8R), ovs, &(xo[WS(os, 1)]));
 
1372
                                                                                     ST(&(xo[WS(os, 19)]), VFNMSI(T8S, T8R), ovs, &(xo[WS(os, 1)]));
 
1373
                                                                                     ST(&(xo[WS(os, 83)]), VFNMSI(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
 
1374
                                                                                     ST(&(xo[WS(os, 45)]), VFMAI(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
 
1375
                                                                                     ST(&(xo[WS(os, 99)]), VFNMSI(T8q, T8p), ovs, &(xo[WS(os, 1)]));
 
1376
                                                                                     ST(&(xo[WS(os, 29)]), VFMAI(T8q, T8p), ovs, &(xo[WS(os, 1)]));
 
1377
                                                                                     ST(&(xo[WS(os, 93)]), VFMAI(T8o, T8l), ovs, &(xo[WS(os, 1)]));
 
1378
                                                                                     ST(&(xo[WS(os, 35)]), VFNMSI(T8o, T8l), ovs, &(xo[WS(os, 1)]));
 
1379
                                                                                     ST(&(xo[WS(os, 125)]), VFMAI(T8i, T8h), ovs, &(xo[WS(os, 1)]));
 
1380
                                                                                     ST(&(xo[WS(os, 3)]), VFNMSI(T8i, T8h), ovs, &(xo[WS(os, 1)]));
 
1381
                                                                                     ST(&(xo[WS(os, 61)]), VFMAI(T8g, T81), ovs, &(xo[WS(os, 1)]));
 
1382
                                                                                     ST(&(xo[WS(os, 67)]), VFNMSI(T8g, T81), ovs, &(xo[WS(os, 1)]));
 
1383
                                                                                     T3t = VFMA(LDK(KP980785280), T3s, T3p);
 
1384
                                                                                     T5l = VFNMS(LDK(KP980785280), T3s, T3p);
 
1385
                                                                                }
 
1386
                                                                           }
 
1387
                                                                      }
 
1388
                                                                 }
 
1389
                                                            }
 
1390
                                                       }
 
1391
                                                  }
 
1392
                                             }
 
1393
                                        }
 
1394
                                        T5n = VFNMS(LDK(KP980785280), T4u, T3V);
 
1395
                                        T4v = VFMA(LDK(KP980785280), T4u, T3V);
 
1396
                                        T4I = VFMA(LDK(KP980785280), T4H, T4E);
 
1397
                                        T5o = VFNMS(LDK(KP980785280), T4H, T4E);
 
1398
                                        {
 
1399
                                             V T53, T3u, T54, T4J, T5v, T5m, T5w, T5p, T10;
 
1400
                                             T6b = VSUB(TI, TZ);
 
1401
                                             T10 = VADD(TI, TZ);
 
1402
                                             T53 = VFMA(LDK(KP049126849), T3g, T3t);
 
1403
                                             T3u = VFNMS(LDK(KP049126849), T3t, T3g);
 
1404
                                             T54 = VFMA(LDK(KP049126849), T4v, T4I);
 
1405
                                             T4J = VFNMS(LDK(KP049126849), T4I, T4v);
 
1406
                                             T5v = VFNMS(LDK(KP906347169), T5k, T5l);
 
1407
                                             T5m = VFMA(LDK(KP906347169), T5l, T5k);
 
1408
                                             T5w = VFNMS(LDK(KP906347169), T5n, T5o);
 
1409
                                             T5p = VFMA(LDK(KP906347169), T5o, T5n);
 
1410
                                             T11 = VFMA(LDK(KP980785280), T10, Tr);
 
1411
                                             T5h = VFNMS(LDK(KP980785280), T10, Tr);
 
1412
                                             T5a = VADD(T53, T54);
 
1413
                                             T55 = VSUB(T53, T54);
 
1414
                                             T5d = VSUB(T3u, T4J);
 
1415
                                             T4K = VADD(T3u, T4J);
 
1416
                                             T5C = VADD(T5v, T5w);
 
1417
                                             T5x = VSUB(T5v, T5w);
 
1418
                                             T5F = VSUB(T5m, T5p);
 
1419
                                             T5q = VADD(T5m, T5p);
 
1420
                                             T4X = VSUB(T4V, T4W);
 
1421
                                             T5K = VADD(T4V, T4W);
 
1422
                                        }
 
1423
                                        T4Z = VFMA(LDK(KP098491403), T1s, T1B);
 
1424
                                        T1C = VFNMS(LDK(KP098491403), T1B, T1s);
 
1425
                                        T2d = VFNMS(LDK(KP098491403), T2c, T23);
 
1426
                                        T50 = VFMA(LDK(KP098491403), T23, T2c);
 
1427
                                   }
 
1428
                                   {
 
1429
                                        V T9y, T9t, T9B, T9i, T9n, T9o, T9F, T8V, T9Q, T9m, T9R, T92, Ta0, T9V, Ta3;
 
1430
                                        V T9O;
 
1431
                                        {
 
1432
                                             V T9I, T9J, T9L, T9d, T5s, T4Y, T5t, T2e, T5i, T51, T9r, T9a, T9g, T9M, T96;
 
1433
                                             V T99;
 
1434
                                             T9I = VFMA(LDK(KP831469612), T95, T94);
 
1435
                                             T96 = VFNMS(LDK(KP831469612), T95, T94);
 
1436
                                             T99 = VFNMS(LDK(KP831469612), T98, T97);
 
1437
                                             T9J = VFMA(LDK(KP831469612), T98, T97);
 
1438
                                             T9L = VFMA(LDK(KP831469612), T9c, T9b);
 
1439
                                             T9d = VFNMS(LDK(KP831469612), T9c, T9b);
 
1440
                                             T5s = VFNMS(LDK(KP980785280), T4X, T4U);
 
1441
                                             T4Y = VFMA(LDK(KP980785280), T4X, T4U);
 
1442
                                             T5t = VSUB(T1C, T2d);
 
1443
                                             T2e = VADD(T1C, T2d);
 
1444
                                             T5i = VADD(T4Z, T50);
 
1445
                                             T51 = VSUB(T4Z, T50);
 
1446
                                             T9r = VFNMS(LDK(KP599376933), T96, T99);
 
1447
                                             T9a = VFMA(LDK(KP599376933), T99, T96);
 
1448
                                             T9g = VFNMS(LDK(KP831469612), T9f, T9e);
 
1449
                                             T9M = VFMA(LDK(KP831469612), T9f, T9e);
 
1450
                                             {
 
1451
                                                  V T5u, T5E, T8Y, T91;
 
1452
                                                  T5u = VFNMS(LDK(KP995184726), T5t, T5s);
 
1453
                                                  T5E = VFMA(LDK(KP995184726), T5t, T5s);
 
1454
                                                  {
 
1455
                                                       V T59, T2f, T5j, T5B;
 
1456
                                                       T59 = VFNMS(LDK(KP995184726), T2e, T11);
 
1457
                                                       T2f = VFMA(LDK(KP995184726), T2e, T11);
 
1458
                                                       T5j = VFMA(LDK(KP995184726), T5i, T5h);
 
1459
                                                       T5B = VFNMS(LDK(KP995184726), T5i, T5h);
 
1460
                                                       {
 
1461
                                                            V T5c, T52, T9s, T9h;
 
1462
                                                            T5c = VFNMS(LDK(KP995184726), T51, T4Y);
 
1463
                                                            T52 = VFMA(LDK(KP995184726), T51, T4Y);
 
1464
                                                            T9s = VFNMS(LDK(KP599376933), T9d, T9g);
 
1465
                                                            T9h = VFMA(LDK(KP599376933), T9g, T9d);
 
1466
                                                            {
 
1467
                                                                 V T5A, T5y, T5I, T5G;
 
1468
                                                                 T5A = VFMA(LDK(KP740951125), T5x, T5u);
 
1469
                                                                 T5y = VFNMS(LDK(KP740951125), T5x, T5u);
 
1470
                                                                 T5I = VFMA(LDK(KP740951125), T5F, T5E);
 
1471
                                                                 T5G = VFNMS(LDK(KP740951125), T5F, T5E);
 
1472
                                                                 {
 
1473
                                                                      V T5f, T5b, T57, T4L;
 
1474
                                                                      T5f = VFMA(LDK(KP998795456), T5a, T59);
 
1475
                                                                      T5b = VFNMS(LDK(KP998795456), T5a, T59);
 
1476
                                                                      T57 = VFMA(LDK(KP998795456), T4K, T2f);
 
1477
                                                                      T4L = VFNMS(LDK(KP998795456), T4K, T2f);
 
1478
                                                                      {
 
1479
                                                                           V T5z, T5r, T5H, T5D;
 
1480
                                                                           T5z = VFMA(LDK(KP740951125), T5q, T5j);
 
1481
                                                                           T5r = VFNMS(LDK(KP740951125), T5q, T5j);
 
1482
                                                                           T5H = VFNMS(LDK(KP740951125), T5C, T5B);
 
1483
                                                                           T5D = VFMA(LDK(KP740951125), T5C, T5B);
 
1484
                                                                           {
 
1485
                                                                                V T5g, T5e, T58, T56;
 
1486
                                                                                T5g = VFNMS(LDK(KP998795456), T5d, T5c);
 
1487
                                                                                T5e = VFMA(LDK(KP998795456), T5d, T5c);
 
1488
                                                                                T58 = VFMA(LDK(KP998795456), T55, T52);
 
1489
                                                                                T56 = VFNMS(LDK(KP998795456), T55, T52);
 
1490
                                                                                T9y = VADD(T9r, T9s);
 
1491
                                                                                T9t = VSUB(T9r, T9s);
 
1492
                                                                                T9B = VSUB(T9a, T9h);
 
1493
                                                                                T9i = VADD(T9a, T9h);
 
1494
                                                                                ST(&(xo[WS(os, 113)]), VFMAI(T5A, T5z), ovs, &(xo[WS(os, 1)]));
 
1495
                                                                                ST(&(xo[WS(os, 15)]), VFNMSI(T5A, T5z), ovs, &(xo[WS(os, 1)]));
 
1496
                                                                                ST(&(xo[WS(os, 49)]), VFMAI(T5y, T5r), ovs, &(xo[WS(os, 1)]));
 
1497
                                                                                ST(&(xo[WS(os, 79)]), VFNMSI(T5y, T5r), ovs, &(xo[WS(os, 1)]));
 
1498
                                                                                ST(&(xo[WS(os, 111)]), VFNMSI(T5I, T5H), ovs, &(xo[WS(os, 1)]));
 
1499
                                                                                ST(&(xo[WS(os, 17)]), VFMAI(T5I, T5H), ovs, &(xo[WS(os, 1)]));
 
1500
                                                                                ST(&(xo[WS(os, 81)]), VFMAI(T5G, T5D), ovs, &(xo[WS(os, 1)]));
 
1501
                                                                                ST(&(xo[WS(os, 47)]), VFNMSI(T5G, T5D), ovs, &(xo[WS(os, 1)]));
 
1502
                                                                                ST(&(xo[WS(os, 97)]), VFMAI(T5g, T5f), ovs, &(xo[WS(os, 1)]));
 
1503
                                                                                ST(&(xo[WS(os, 31)]), VFNMSI(T5g, T5f), ovs, &(xo[WS(os, 1)]));
 
1504
                                                                                ST(&(xo[WS(os, 95)]), VFNMSI(T5e, T5b), ovs, &(xo[WS(os, 1)]));
 
1505
                                                                                ST(&(xo[WS(os, 33)]), VFMAI(T5e, T5b), ovs, &(xo[WS(os, 1)]));
 
1506
                                                                                ST(&(xo[WS(os, 1)]), VFMAI(T58, T57), ovs, &(xo[WS(os, 1)]));
 
1507
                                                                                ST(&(xo[WS(os, 127)]), VFNMSI(T58, T57), ovs, &(xo[WS(os, 1)]));
 
1508
                                                                                ST(&(xo[WS(os, 65)]), VFMAI(T56, T4L), ovs, &(xo[WS(os, 1)]));
 
1509
                                                                                ST(&(xo[WS(os, 63)]), VFNMSI(T56, T4L), ovs, &(xo[WS(os, 1)]));
 
1510
                                                                           }
 
1511
                                                                      }
 
1512
                                                                 }
 
1513
                                                            }
 
1514
                                                       }
 
1515
                                                  }
 
1516
                                                  T9n = VFNMS(LDK(KP534511135), T8W, T8X);
 
1517
                                                  T8Y = VFMA(LDK(KP534511135), T8X, T8W);
 
1518
                                                  T91 = VFMA(LDK(KP534511135), T90, T8Z);
 
1519
                                                  T9o = VFNMS(LDK(KP534511135), T8Z, T90);
 
1520
                                                  {
 
1521
                                                       V T9T, T9K, T9U, T9N;
 
1522
                                                       T9T = VFMA(LDK(KP250486960), T9I, T9J);
 
1523
                                                       T9K = VFNMS(LDK(KP250486960), T9J, T9I);
 
1524
                                                       T9U = VFMA(LDK(KP250486960), T9L, T9M);
 
1525
                                                       T9N = VFNMS(LDK(KP250486960), T9M, T9L);
 
1526
                                                       T9F = VFNMS(LDK(KP831469612), T8U, T8T);
 
1527
                                                       T8V = VFMA(LDK(KP831469612), T8U, T8T);
 
1528
                                                       T9Q = VFMA(LDK(KP831469612), T9l, T9k);
 
1529
                                                       T9m = VFNMS(LDK(KP831469612), T9l, T9k);
 
1530
                                                       T9R = VSUB(T8Y, T91);
 
1531
                                                       T92 = VADD(T8Y, T91);
 
1532
                                                       Ta0 = VADD(T9T, T9U);
 
1533
                                                       T9V = VSUB(T9T, T9U);
 
1534
                                                       Ta3 = VSUB(T9K, T9N);
 
1535
                                                       T9O = VADD(T9K, T9N);
 
1536
                                                  }
 
1537
                                             }
 
1538
                                        }
 
1539
                                        {
 
1540
                                             V T6y, T6z, T63, T9Y, T9W, Ta6, Ta4, T9D, T9z, T9v, T9j, T6h, T60, T9H, T9Z;
 
1541
                                             V T9A, T9q, T66, T9X, T9P;
 
1542
                                             {
 
1543
                                                  V T5W, T9S, Ta2, T9x, T93, T5Z, T9G, T9p;
 
1544
                                                  T6y = VFMA(LDK(KP980785280), T5V, T5U);
 
1545
                                                  T5W = VFNMS(LDK(KP980785280), T5V, T5U);
 
1546
                                                  T9S = VFMA(LDK(KP881921264), T9R, T9Q);
 
1547
                                                  Ta2 = VFNMS(LDK(KP881921264), T9R, T9Q);
 
1548
                                                  T9x = VFNMS(LDK(KP881921264), T92, T8V);
 
1549
                                                  T93 = VFMA(LDK(KP881921264), T92, T8V);
 
1550
                                                  T5Z = VFMA(LDK(KP980785280), T5Y, T5X);
 
1551
                                                  T6z = VFNMS(LDK(KP980785280), T5Y, T5X);
 
1552
                                                  T6B = VFMA(LDK(KP980785280), T62, T61);
 
1553
                                                  T63 = VFNMS(LDK(KP980785280), T62, T61);
 
1554
                                                  T9G = VADD(T9n, T9o);
 
1555
                                                  T9p = VSUB(T9n, T9o);
 
1556
                                                  T9Y = VFMA(LDK(KP970031253), T9V, T9S);
 
1557
                                                  T9W = VFNMS(LDK(KP970031253), T9V, T9S);
 
1558
                                                  Ta6 = VFNMS(LDK(KP970031253), Ta3, Ta2);
 
1559
                                                  Ta4 = VFMA(LDK(KP970031253), Ta3, Ta2);
 
1560
                                                  T9D = VFNMS(LDK(KP857728610), T9y, T9x);
 
1561
                                                  T9z = VFMA(LDK(KP857728610), T9y, T9x);
 
1562
                                                  T9v = VFMA(LDK(KP857728610), T9i, T93);
 
1563
                                                  T9j = VFNMS(LDK(KP857728610), T9i, T93);
 
1564
                                                  T6h = VFMA(LDK(KP472964775), T5W, T5Z);
 
1565
                                                  T60 = VFNMS(LDK(KP472964775), T5Z, T5W);
 
1566
                                                  T9H = VFMA(LDK(KP881921264), T9G, T9F);
 
1567
                                                  T9Z = VFNMS(LDK(KP881921264), T9G, T9F);
 
1568
                                                  T9A = VFMA(LDK(KP881921264), T9p, T9m);
 
1569
                                                  T9q = VFNMS(LDK(KP881921264), T9p, T9m);
 
1570
                                                  T66 = VFMA(LDK(KP980785280), T65, T64);
 
1571
                                                  T6C = VFNMS(LDK(KP980785280), T65, T64);
 
1572
                                             }
 
1573
                                             T9X = VFMA(LDK(KP970031253), T9O, T9H);
 
1574
                                             T9P = VFNMS(LDK(KP970031253), T9O, T9H);
 
1575
                                             {
 
1576
                                                  V Ta5, Ta1, T9E, T9C;
 
1577
                                                  Ta5 = VFMA(LDK(KP970031253), Ta0, T9Z);
 
1578
                                                  Ta1 = VFNMS(LDK(KP970031253), Ta0, T9Z);
 
1579
                                                  T9E = VFMA(LDK(KP857728610), T9B, T9A);
 
1580
                                                  T9C = VFNMS(LDK(KP857728610), T9B, T9A);
 
1581
                                                  {
 
1582
                                                       V T9w, T9u, T6i, T67;
 
1583
                                                       T9w = VFMA(LDK(KP857728610), T9t, T9q);
 
1584
                                                       T9u = VFNMS(LDK(KP857728610), T9t, T9q);
 
1585
                                                       T6i = VFMA(LDK(KP472964775), T63, T66);
 
1586
                                                       T67 = VFNMS(LDK(KP472964775), T66, T63);
 
1587
                                                       T6J = VFNMS(LDK(KP357805721), T6y, T6z);
 
1588
                                                       T6A = VFMA(LDK(KP357805721), T6z, T6y);
 
1589
                                                       ST(&(xo[WS(os, 5)]), VFMAI(T9Y, T9X), ovs, &(xo[WS(os, 1)]));
 
1590
                                                       ST(&(xo[WS(os, 123)]), VFNMSI(T9Y, T9X), ovs, &(xo[WS(os, 1)]));
 
1591
                                                       ST(&(xo[WS(os, 69)]), VFMAI(T9W, T9P), ovs, &(xo[WS(os, 1)]));
 
1592
                                                       ST(&(xo[WS(os, 59)]), VFNMSI(T9W, T9P), ovs, &(xo[WS(os, 1)]));
 
1593
                                                       ST(&(xo[WS(os, 101)]), VFMAI(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
 
1594
                                                       ST(&(xo[WS(os, 27)]), VFNMSI(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
 
1595
                                                       ST(&(xo[WS(os, 91)]), VFNMSI(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
 
1596
                                                       ST(&(xo[WS(os, 37)]), VFMAI(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
 
1597
                                                       ST(&(xo[WS(os, 107)]), VFNMSI(T9E, T9D), ovs, &(xo[WS(os, 1)]));
 
1598
                                                       ST(&(xo[WS(os, 21)]), VFMAI(T9E, T9D), ovs, &(xo[WS(os, 1)]));
 
1599
                                                       ST(&(xo[WS(os, 85)]), VFMAI(T9C, T9z), ovs, &(xo[WS(os, 1)]));
 
1600
                                                       ST(&(xo[WS(os, 43)]), VFNMSI(T9C, T9z), ovs, &(xo[WS(os, 1)]));
 
1601
                                                       ST(&(xo[WS(os, 117)]), VFMAI(T9w, T9v), ovs, &(xo[WS(os, 1)]));
 
1602
                                                       ST(&(xo[WS(os, 11)]), VFNMSI(T9w, T9v), ovs, &(xo[WS(os, 1)]));
 
1603
                                                       ST(&(xo[WS(os, 53)]), VFMAI(T9u, T9j), ovs, &(xo[WS(os, 1)]));
 
1604
                                                       ST(&(xo[WS(os, 75)]), VFNMSI(T9u, T9j), ovs, &(xo[WS(os, 1)]));
 
1605
                                                       T6o = VADD(T6h, T6i);
 
1606
                                                       T6j = VSUB(T6h, T6i);
 
1607
                                                       T6r = VSUB(T60, T67);
 
1608
                                                       T68 = VADD(T60, T67);
 
1609
                                                  }
 
1610
                                             }
 
1611
                                             T6d = VFMA(LDK(KP820678790), T5M, T5N);
 
1612
                                             T5O = VFNMS(LDK(KP820678790), T5N, T5M);
 
1613
                                             T5R = VFNMS(LDK(KP820678790), T5Q, T5P);
 
1614
                                             T6e = VFMA(LDK(KP820678790), T5P, T5Q);
 
1615
                                        }
 
1616
                                   }
 
1617
                              }
 
1618
                         }
 
1619
                    }
 
1620
               }
 
1621
               T6D = VFMA(LDK(KP357805721), T6C, T6B);
 
1622
               T6K = VFNMS(LDK(KP357805721), T6B, T6C);
 
1623
               {
 
1624
                    V T5L, T6v, T6c, T6G;
 
1625
                    T5L = VFNMS(LDK(KP980785280), T5K, T5J);
 
1626
                    T6v = VFMA(LDK(KP980785280), T5K, T5J);
 
1627
                    T6c = VFMA(LDK(KP980785280), T6b, T6a);
 
1628
                    T6G = VFNMS(LDK(KP980785280), T6b, T6a);
 
1629
                    {
 
1630
                         V T5S, T6H, T6f, T6w;
 
1631
                         T5S = VADD(T5O, T5R);
 
1632
                         T6H = VSUB(T5O, T5R);
 
1633
                         T6f = VSUB(T6d, T6e);
 
1634
                         T6w = VADD(T6d, T6e);
 
1635
                         {
 
1636
                              V T6L, T6Q, T6E, T6T;
 
1637
                              T6L = VSUB(T6J, T6K);
 
1638
                              T6Q = VADD(T6J, T6K);
 
1639
                              T6E = VADD(T6A, T6D);
 
1640
                              T6T = VSUB(T6A, T6D);
 
1641
                              {
 
1642
                                   V T6S, T6I, T5T, T6n;
 
1643
                                   T6S = VFNMS(LDK(KP773010453), T6H, T6G);
 
1644
                                   T6I = VFMA(LDK(KP773010453), T6H, T6G);
 
1645
                                   T5T = VFNMS(LDK(KP773010453), T5S, T5L);
 
1646
                                   T6n = VFMA(LDK(KP773010453), T5S, T5L);
 
1647
                                   {
 
1648
                                        V T6P, T6x, T6g, T6q;
 
1649
                                        T6P = VFNMS(LDK(KP773010453), T6w, T6v);
 
1650
                                        T6x = VFMA(LDK(KP773010453), T6w, T6v);
 
1651
                                        T6g = VFMA(LDK(KP773010453), T6f, T6c);
 
1652
                                        T6q = VFNMS(LDK(KP773010453), T6f, T6c);
 
1653
                                        {
 
1654
                                             V T6M, T6O, T6U, T6W;
 
1655
                                             T6M = VFNMS(LDK(KP941544065), T6L, T6I);
 
1656
                                             T6O = VFMA(LDK(KP941544065), T6L, T6I);
 
1657
                                             T6U = VFNMS(LDK(KP941544065), T6T, T6S);
 
1658
                                             T6W = VFMA(LDK(KP941544065), T6T, T6S);
 
1659
                                             {
 
1660
                                                  V T6p, T6t, T69, T6l;
 
1661
                                                  T6p = VFNMS(LDK(KP903989293), T6o, T6n);
 
1662
                                                  T6t = VFMA(LDK(KP903989293), T6o, T6n);
 
1663
                                                  T69 = VFNMS(LDK(KP903989293), T68, T5T);
 
1664
                                                  T6l = VFMA(LDK(KP903989293), T68, T5T);
 
1665
                                                  {
 
1666
                                                       V T6F, T6N, T6R, T6V;
 
1667
                                                       T6F = VFNMS(LDK(KP941544065), T6E, T6x);
 
1668
                                                       T6N = VFMA(LDK(KP941544065), T6E, T6x);
 
1669
                                                       T6R = VFMA(LDK(KP941544065), T6Q, T6P);
 
1670
                                                       T6V = VFNMS(LDK(KP941544065), T6Q, T6P);
 
1671
                                                       {
 
1672
                                                            V T6s, T6u, T6k, T6m;
 
1673
                                                            T6s = VFMA(LDK(KP903989293), T6r, T6q);
 
1674
                                                            T6u = VFNMS(LDK(KP903989293), T6r, T6q);
 
1675
                                                            T6k = VFNMS(LDK(KP903989293), T6j, T6g);
 
1676
                                                            T6m = VFMA(LDK(KP903989293), T6j, T6g);
 
1677
                                                            ST(&(xo[WS(os, 121)]), VFMAI(T6O, T6N), ovs, &(xo[WS(os, 1)]));
 
1678
                                                            ST(&(xo[WS(os, 7)]), VFNMSI(T6O, T6N), ovs, &(xo[WS(os, 1)]));
 
1679
                                                            ST(&(xo[WS(os, 57)]), VFMAI(T6M, T6F), ovs, &(xo[WS(os, 1)]));
 
1680
                                                            ST(&(xo[WS(os, 71)]), VFNMSI(T6M, T6F), ovs, &(xo[WS(os, 1)]));
 
1681
                                                            ST(&(xo[WS(os, 103)]), VFNMSI(T6W, T6V), ovs, &(xo[WS(os, 1)]));
 
1682
                                                            ST(&(xo[WS(os, 25)]), VFMAI(T6W, T6V), ovs, &(xo[WS(os, 1)]));
 
1683
                                                            ST(&(xo[WS(os, 89)]), VFMAI(T6U, T6R), ovs, &(xo[WS(os, 1)]));
 
1684
                                                            ST(&(xo[WS(os, 39)]), VFNMSI(T6U, T6R), ovs, &(xo[WS(os, 1)]));
 
1685
                                                            ST(&(xo[WS(os, 105)]), VFMAI(T6u, T6t), ovs, &(xo[WS(os, 1)]));
 
1686
                                                            ST(&(xo[WS(os, 23)]), VFNMSI(T6u, T6t), ovs, &(xo[WS(os, 1)]));
 
1687
                                                            ST(&(xo[WS(os, 87)]), VFNMSI(T6s, T6p), ovs, &(xo[WS(os, 1)]));
 
1688
                                                            ST(&(xo[WS(os, 41)]), VFMAI(T6s, T6p), ovs, &(xo[WS(os, 1)]));
 
1689
                                                            ST(&(xo[WS(os, 9)]), VFMAI(T6m, T6l), ovs, &(xo[WS(os, 1)]));
 
1690
                                                            ST(&(xo[WS(os, 119)]), VFNMSI(T6m, T6l), ovs, &(xo[WS(os, 1)]));
 
1691
                                                            ST(&(xo[WS(os, 73)]), VFMAI(T6k, T69), ovs, &(xo[WS(os, 1)]));
 
1692
                                                            ST(&(xo[WS(os, 55)]), VFNMSI(T6k, T69), ovs, &(xo[WS(os, 1)]));
 
1693
                                                       }
 
1694
                                                  }
 
1695
                                             }
 
1696
                                        }
 
1697
                                   }
 
1698
                              }
 
1699
                         }
 
1700
                    }
 
1701
               }
 
1702
          }
 
1703
     }
 
1704
     VLEAVE();
 
1705
}
 
1706
 
 
1707
static const kdft_desc desc = { 128, XSIMD_STRING("n1bv_128"), {440, 0, 642, 0}, &GENUS, 0, 0, 0, 0 };
 
1708
 
 
1709
void XSIMD(codelet_n1bv_128) (planner *p) {
 
1710
     X(kdft_register) (p, n1bv_128, &desc);
 
1711
}
 
1712
 
 
1713
#else                           /* HAVE_FMA */
 
1714
 
 
1715
/* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 128 -name n1bv_128 -include n1b.h */
 
1716
 
 
1717
/*
 
1718
 * This function contains 1082 FP additions, 330 FP multiplications,
 
1719
 * (or, 938 additions, 186 multiplications, 144 fused multiply/add),
 
1720
 * 194 stack variables, 31 constants, and 256 memory accesses
 
1721
 */
 
1722
#include "n1b.h"
 
1723
 
 
1724
static void n1bv_128(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
 
1725
{
 
1726
     DVK(KP146730474, +0.146730474455361751658850129646717819706215317);
 
1727
     DVK(KP989176509, +0.989176509964780973451673738016243063983689533);
 
1728
     DVK(KP595699304, +0.595699304492433343467036528829969889511926338);
 
1729
     DVK(KP803207531, +0.803207531480644909806676512963141923879569427);
 
1730
     DVK(KP049067674, +0.049067674327418014254954976942682658314745363);
 
1731
     DVK(KP998795456, +0.998795456205172392714771604759100694443203615);
 
1732
     DVK(KP671558954, +0.671558954847018400625376850427421803228750632);
 
1733
     DVK(KP740951125, +0.740951125354959091175616897495162729728955309);
 
1734
     DVK(KP514102744, +0.514102744193221726593693838968815772608049120);
 
1735
     DVK(KP857728610, +0.857728610000272069902269984284770137042490799);
 
1736
     DVK(KP242980179, +0.242980179903263889948274162077471118320990783);
 
1737
     DVK(KP970031253, +0.970031253194543992603984207286100251456865962);
 
1738
     DVK(KP427555093, +0.427555093430282094320966856888798534304578629);
 
1739
     DVK(KP903989293, +0.903989293123443331586200297230537048710132025);
 
1740
     DVK(KP336889853, +0.336889853392220050689253212619147570477766780);
 
1741
     DVK(KP941544065, +0.941544065183020778412509402599502357185589796);
 
1742
     DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
 
1743
     DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
 
1744
     DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
 
1745
     DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
 
1746
     DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
 
1747
     DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
 
1748
     DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
 
1749
     DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
 
1750
     DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
 
1751
     DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
 
1752
     DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
 
1753
     DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
 
1754
     DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
 
1755
     DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
1756
     DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
1757
     {
 
1758
          INT i;
 
1759
          const R *xi;
 
1760
          R *xo;
 
1761
          xi = ii;
 
1762
          xo = io;
 
1763
          for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
 
1764
               V T49, T6e, Tev, TgK, TfA, TgL, T4U, T5J, T7R, T9o, Tah, TdG, Tcw, TdB, T84;
 
1765
               V T8T, Tfk, Tfo, T1G, T64, Tgs, Th6, T2p, T62, T7t, T9c, Tce, Tdm, T7i, T9e;
 
1766
               V Tc8, Tdp, TgF, TgG, T4q, T4V, TeC, Tfx, T4H, T4W, T7X, T86, Tcr, TdH, T7U;
 
1767
               V T85, Taw, TdC, Tf3, Tf7, Tr, T5X, Tgl, Th3, T1a, T5V, T7a, T95, TbD, Tdf;
 
1768
               V T6Z, T97, Tbx, Tdi, Tgy, Tgz, TgA, TaN, Tdv, TeK, Tfu, T2W, T5M, T35, T5N;
 
1769
               V T7F, T8X, TaI, Tdu, T7C, T8W, TgB, TgC, TgD, Tb4, Tdy, TeR, Tfv, T3x, T5P;
 
1770
               V T3G, T5Q, T7M, T90, TaZ, Tdx, T7J, T8Z, Tbm, Tdg, TbG, Tdj, Tgo, Th4, Tf0;
 
1771
               V Tf8, T76, T98, T7d, T94, T10, T5Y, T1d, T5U, TbX, Tdn, Tch, Tdq, Tgv, Th7;
 
1772
               V Tfh, Tfp, T7p, T9f, T7w, T9b, T2f, T65, T2s, T61;
 
1773
               {
 
1774
                    V T47, Ta8, T4O, Ta7, T44, Tcu, T4P, Tct, Taa, Tab, T3P, Tac, T4R, Tad, Tae;
 
1775
                    V T3W, Taf, T4S;
 
1776
                    {
 
1777
                         V T45, T46, T4M, T4N;
 
1778
                         T45 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
 
1779
                         T46 = LD(&(xi[WS(is, 96)]), ivs, &(xi[0]));
 
1780
                         T47 = VSUB(T45, T46);
 
1781
                         Ta8 = VADD(T45, T46);
 
1782
                         T4M = LD(&(xi[0]), ivs, &(xi[0]));
 
1783
                         T4N = LD(&(xi[WS(is, 64)]), ivs, &(xi[0]));
 
1784
                         T4O = VSUB(T4M, T4N);
 
1785
                         Ta7 = VADD(T4M, T4N);
 
1786
                    }
 
1787
                    {
 
1788
                         V T3Y, T3Z, T40, T41, T42, T43;
 
1789
                         T3Y = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
 
1790
                         T3Z = LD(&(xi[WS(is, 80)]), ivs, &(xi[0]));
 
1791
                         T40 = VSUB(T3Y, T3Z);
 
1792
                         T41 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0]));
 
1793
                         T42 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
 
1794
                         T43 = VSUB(T41, T42);
 
1795
                         T44 = VMUL(LDK(KP707106781), VSUB(T40, T43));
 
1796
                         Tcu = VADD(T41, T42);
 
1797
                         T4P = VMUL(LDK(KP707106781), VADD(T40, T43));
 
1798
                         Tct = VADD(T3Y, T3Z);
 
1799
                    }
 
1800
                    {
 
1801
                         V T3L, T3O, T3S, T3V;
 
1802
                         {
 
1803
                              V T3J, T3K, T3M, T3N;
 
1804
                              T3J = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
 
1805
                              T3K = LD(&(xi[WS(is, 72)]), ivs, &(xi[0]));
 
1806
                              T3L = VSUB(T3J, T3K);
 
1807
                              Taa = VADD(T3J, T3K);
 
1808
                              T3M = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
 
1809
                              T3N = LD(&(xi[WS(is, 104)]), ivs, &(xi[0]));
 
1810
                              T3O = VSUB(T3M, T3N);
 
1811
                              Tab = VADD(T3M, T3N);
 
1812
                         }
 
1813
                         T3P = VFNMS(LDK(KP382683432), T3O, VMUL(LDK(KP923879532), T3L));
 
1814
                         Tac = VSUB(Taa, Tab);
 
1815
                         T4R = VFMA(LDK(KP382683432), T3L, VMUL(LDK(KP923879532), T3O));
 
1816
                         {
 
1817
                              V T3Q, T3R, T3T, T3U;
 
1818
                              T3Q = LD(&(xi[WS(is, 120)]), ivs, &(xi[0]));
 
1819
                              T3R = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
 
1820
                              T3S = VSUB(T3Q, T3R);
 
1821
                              Tad = VADD(T3Q, T3R);
 
1822
                              T3T = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
 
1823
                              T3U = LD(&(xi[WS(is, 88)]), ivs, &(xi[0]));
 
1824
                              T3V = VSUB(T3T, T3U);
 
1825
                              Tae = VADD(T3T, T3U);
 
1826
                         }
 
1827
                         T3W = VFMA(LDK(KP923879532), T3S, VMUL(LDK(KP382683432), T3V));
 
1828
                         Taf = VSUB(Tad, Tae);
 
1829
                         T4S = VFNMS(LDK(KP382683432), T3S, VMUL(LDK(KP923879532), T3V));
 
1830
                    }
 
1831
                    {
 
1832
                         V T3X, T48, Tet, Teu;
 
1833
                         T3X = VSUB(T3P, T3W);
 
1834
                         T48 = VSUB(T44, T47);
 
1835
                         T49 = VSUB(T3X, T48);
 
1836
                         T6e = VADD(T48, T3X);
 
1837
                         Tet = VADD(Ta7, Ta8);
 
1838
                         Teu = VADD(Tct, Tcu);
 
1839
                         Tev = VSUB(Tet, Teu);
 
1840
                         TgK = VADD(Tet, Teu);
 
1841
                    }
 
1842
                    {
 
1843
                         V Tfy, Tfz, T4Q, T4T;
 
1844
                         Tfy = VADD(Taa, Tab);
 
1845
                         Tfz = VADD(Tad, Tae);
 
1846
                         TfA = VSUB(Tfy, Tfz);
 
1847
                         TgL = VADD(Tfy, Tfz);
 
1848
                         T4Q = VSUB(T4O, T4P);
 
1849
                         T4T = VSUB(T4R, T4S);
 
1850
                         T4U = VSUB(T4Q, T4T);
 
1851
                         T5J = VADD(T4Q, T4T);
 
1852
                    }
 
1853
                    {
 
1854
                         V T7P, T7Q, Ta9, Tag;
 
1855
                         T7P = VADD(T4R, T4S);
 
1856
                         T7Q = VADD(T47, T44);
 
1857
                         T7R = VSUB(T7P, T7Q);
 
1858
                         T9o = VADD(T7Q, T7P);
 
1859
                         Ta9 = VSUB(Ta7, Ta8);
 
1860
                         Tag = VMUL(LDK(KP707106781), VADD(Tac, Taf));
 
1861
                         Tah = VSUB(Ta9, Tag);
 
1862
                         TdG = VADD(Ta9, Tag);
 
1863
                    }
 
1864
                    {
 
1865
                         V Tcs, Tcv, T82, T83;
 
1866
                         Tcs = VMUL(LDK(KP707106781), VSUB(Tac, Taf));
 
1867
                         Tcv = VSUB(Tct, Tcu);
 
1868
                         Tcw = VSUB(Tcs, Tcv);
 
1869
                         TdB = VADD(Tcv, Tcs);
 
1870
                         T82 = VADD(T4O, T4P);
 
1871
                         T83 = VADD(T3P, T3W);
 
1872
                         T84 = VSUB(T82, T83);
 
1873
                         T8T = VADD(T82, T83);
 
1874
                    }
 
1875
               }
 
1876
               {
 
1877
                    V Tca, Tcb, T1i, Tfm, T2n, Tc5, Tc6, T1p, Tfn, T2k, T1x, Tfi, T2h, Tc0, T1E;
 
1878
                    V Tfj, T2i, Tc3, T1l, T1o, Tcc, Tcd;
 
1879
                    {
 
1880
                         V T1g, T1h, T2l, T2m;
 
1881
                         T1g = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)]));
 
1882
                         T1h = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
 
1883
                         Tca = VADD(T1g, T1h);
 
1884
                         T2l = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
 
1885
                         T2m = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)]));
 
1886
                         Tcb = VADD(T2l, T2m);
 
1887
                         T1i = VSUB(T1g, T1h);
 
1888
                         Tfm = VADD(Tca, Tcb);
 
1889
                         T2n = VSUB(T2l, T2m);
 
1890
                    }
 
1891
                    {
 
1892
                         V T1j, T1k, T1m, T1n;
 
1893
                         T1j = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
 
1894
                         T1k = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)]));
 
1895
                         T1l = VSUB(T1j, T1k);
 
1896
                         Tc5 = VADD(T1j, T1k);
 
1897
                         T1m = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)]));
 
1898
                         T1n = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
 
1899
                         T1o = VSUB(T1m, T1n);
 
1900
                         Tc6 = VADD(T1m, T1n);
 
1901
                    }
 
1902
                    T1p = VMUL(LDK(KP707106781), VADD(T1l, T1o));
 
1903
                    Tfn = VADD(Tc5, Tc6);
 
1904
                    T2k = VMUL(LDK(KP707106781), VSUB(T1l, T1o));
 
1905
                    {
 
1906
                         V T1t, TbY, T1w, TbZ;
 
1907
                         {
 
1908
                              V T1r, T1s, T1u, T1v;
 
1909
                              T1r = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
 
1910
                              T1s = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)]));
 
1911
                              T1t = VSUB(T1r, T1s);
 
1912
                              TbY = VADD(T1r, T1s);
 
1913
                              T1u = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
 
1914
                              T1v = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)]));
 
1915
                              T1w = VSUB(T1u, T1v);
 
1916
                              TbZ = VADD(T1u, T1v);
 
1917
                         }
 
1918
                         T1x = VFMA(LDK(KP382683432), T1t, VMUL(LDK(KP923879532), T1w));
 
1919
                         Tfi = VADD(TbY, TbZ);
 
1920
                         T2h = VFNMS(LDK(KP382683432), T1w, VMUL(LDK(KP923879532), T1t));
 
1921
                         Tc0 = VSUB(TbY, TbZ);
 
1922
                    }
 
1923
                    {
 
1924
                         V T1A, Tc2, T1D, Tc1;
 
1925
                         {
 
1926
                              V T1y, T1z, T1B, T1C;
 
1927
                              T1y = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
 
1928
                              T1z = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)]));
 
1929
                              T1A = VSUB(T1y, T1z);
 
1930
                              Tc2 = VADD(T1y, T1z);
 
1931
                              T1B = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)]));
 
1932
                              T1C = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
 
1933
                              T1D = VSUB(T1B, T1C);
 
1934
                              Tc1 = VADD(T1B, T1C);
 
1935
                         }
 
1936
                         T1E = VFNMS(LDK(KP382683432), T1D, VMUL(LDK(KP923879532), T1A));
 
1937
                         Tfj = VADD(Tc1, Tc2);
 
1938
                         T2i = VFMA(LDK(KP923879532), T1D, VMUL(LDK(KP382683432), T1A));
 
1939
                         Tc3 = VSUB(Tc1, Tc2);
 
1940
                    }
 
1941
                    Tfk = VSUB(Tfi, Tfj);
 
1942
                    Tfo = VSUB(Tfm, Tfn);
 
1943
                    {
 
1944
                         V T1q, T1F, Tgq, Tgr;
 
1945
                         T1q = VSUB(T1i, T1p);
 
1946
                         T1F = VSUB(T1x, T1E);
 
1947
                         T1G = VSUB(T1q, T1F);
 
1948
                         T64 = VADD(T1q, T1F);
 
1949
                         Tgq = VADD(Tfm, Tfn);
 
1950
                         Tgr = VADD(Tfi, Tfj);
 
1951
                         Tgs = VSUB(Tgq, Tgr);
 
1952
                         Th6 = VADD(Tgq, Tgr);
 
1953
                    }
 
1954
                    {
 
1955
                         V T2j, T2o, T7r, T7s;
 
1956
                         T2j = VSUB(T2h, T2i);
 
1957
                         T2o = VSUB(T2k, T2n);
 
1958
                         T2p = VSUB(T2j, T2o);
 
1959
                         T62 = VADD(T2o, T2j);
 
1960
                         T7r = VADD(T1x, T1E);
 
1961
                         T7s = VADD(T2n, T2k);
 
1962
                         T7t = VSUB(T7r, T7s);
 
1963
                         T9c = VADD(T7s, T7r);
 
1964
                    }
 
1965
                    Tcc = VSUB(Tca, Tcb);
 
1966
                    Tcd = VMUL(LDK(KP707106781), VADD(Tc0, Tc3));
 
1967
                    Tce = VSUB(Tcc, Tcd);
 
1968
                    Tdm = VADD(Tcc, Tcd);
 
1969
                    {
 
1970
                         V T7g, T7h, Tc4, Tc7;
 
1971
                         T7g = VADD(T1i, T1p);
 
1972
                         T7h = VADD(T2h, T2i);
 
1973
                         T7i = VSUB(T7g, T7h);
 
1974
                         T9e = VADD(T7g, T7h);
 
1975
                         Tc4 = VMUL(LDK(KP707106781), VSUB(Tc0, Tc3));
 
1976
                         Tc7 = VSUB(Tc5, Tc6);
 
1977
                         Tc8 = VSUB(Tc4, Tc7);
 
1978
                         Tdp = VADD(Tc7, Tc4);
 
1979
                    }
 
1980
               }
 
1981
               {
 
1982
                    V T4c, Tew, T4o, Tak, T4A, Tez, T4E, Tau, T4j, Tex, T4l, Tan, T4x, TeA, T4F;
 
1983
                    V Tar, Tcp, Tcq;
 
1984
                    {
 
1985
                         V T4a, T4b, Tai, T4m, T4n, Taj;
 
1986
                         T4a = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
 
1987
                         T4b = LD(&(xi[WS(is, 68)]), ivs, &(xi[0]));
 
1988
                         Tai = VADD(T4a, T4b);
 
1989
                         T4m = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
 
1990
                         T4n = LD(&(xi[WS(is, 100)]), ivs, &(xi[0]));
 
1991
                         Taj = VADD(T4m, T4n);
 
1992
                         T4c = VSUB(T4a, T4b);
 
1993
                         Tew = VADD(Tai, Taj);
 
1994
                         T4o = VSUB(T4m, T4n);
 
1995
                         Tak = VSUB(Tai, Taj);
 
1996
                    }
 
1997
                    {
 
1998
                         V T4y, T4z, Tat, T4C, T4D, Tas;
 
1999
                         T4y = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
 
2000
                         T4z = LD(&(xi[WS(is, 92)]), ivs, &(xi[0]));
 
2001
                         Tat = VADD(T4y, T4z);
 
2002
                         T4C = LD(&(xi[WS(is, 124)]), ivs, &(xi[0]));
 
2003
                         T4D = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
 
2004
                         Tas = VADD(T4C, T4D);
 
2005
                         T4A = VSUB(T4y, T4z);
 
2006
                         Tez = VADD(Tas, Tat);
 
2007
                         T4E = VSUB(T4C, T4D);
 
2008
                         Tau = VSUB(Tas, Tat);
 
2009
                    }
 
2010
                    {
 
2011
                         V T4f, Tal, T4i, Tam;
 
2012
                         {
 
2013
                              V T4d, T4e, T4g, T4h;
 
2014
                              T4d = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
 
2015
                              T4e = LD(&(xi[WS(is, 84)]), ivs, &(xi[0]));
 
2016
                              T4f = VSUB(T4d, T4e);
 
2017
                              Tal = VADD(T4d, T4e);
 
2018
                              T4g = LD(&(xi[WS(is, 116)]), ivs, &(xi[0]));
 
2019
                              T4h = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
 
2020
                              T4i = VSUB(T4g, T4h);
 
2021
                              Tam = VADD(T4g, T4h);
 
2022
                         }
 
2023
                         T4j = VMUL(LDK(KP707106781), VADD(T4f, T4i));
 
2024
                         Tex = VADD(Tal, Tam);
 
2025
                         T4l = VMUL(LDK(KP707106781), VSUB(T4f, T4i));
 
2026
                         Tan = VSUB(Tal, Tam);
 
2027
                    }
 
2028
                    {
 
2029
                         V T4t, Tap, T4w, Taq;
 
2030
                         {
 
2031
                              V T4r, T4s, T4u, T4v;
 
2032
                              T4r = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
 
2033
                              T4s = LD(&(xi[WS(is, 76)]), ivs, &(xi[0]));
 
2034
                              T4t = VSUB(T4r, T4s);
 
2035
                              Tap = VADD(T4r, T4s);
 
2036
                              T4u = LD(&(xi[WS(is, 108)]), ivs, &(xi[0]));
 
2037
                              T4v = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
 
2038
                              T4w = VSUB(T4u, T4v);
 
2039
                              Taq = VADD(T4u, T4v);
 
2040
                         }
 
2041
                         T4x = VMUL(LDK(KP707106781), VSUB(T4t, T4w));
 
2042
                         TeA = VADD(Tap, Taq);
 
2043
                         T4F = VMUL(LDK(KP707106781), VADD(T4t, T4w));
 
2044
                         Tar = VSUB(Tap, Taq);
 
2045
                    }
 
2046
                    TgF = VADD(Tew, Tex);
 
2047
                    TgG = VADD(Tez, TeA);
 
2048
                    {
 
2049
                         V T4k, T4p, Tey, TeB;
 
2050
                         T4k = VSUB(T4c, T4j);
 
2051
                         T4p = VSUB(T4l, T4o);
 
2052
                         T4q = VFNMS(LDK(KP555570233), T4p, VMUL(LDK(KP831469612), T4k));
 
2053
                         T4V = VFMA(LDK(KP831469612), T4p, VMUL(LDK(KP555570233), T4k));
 
2054
                         Tey = VSUB(Tew, Tex);
 
2055
                         TeB = VSUB(Tez, TeA);
 
2056
                         TeC = VMUL(LDK(KP707106781), VADD(Tey, TeB));
 
2057
                         Tfx = VMUL(LDK(KP707106781), VSUB(Tey, TeB));
 
2058
                    }
 
2059
                    {
 
2060
                         V T4B, T4G, T7V, T7W;
 
2061
                         T4B = VSUB(T4x, T4A);
 
2062
                         T4G = VSUB(T4E, T4F);
 
2063
                         T4H = VFMA(LDK(KP555570233), T4B, VMUL(LDK(KP831469612), T4G));
 
2064
                         T4W = VFNMS(LDK(KP555570233), T4G, VMUL(LDK(KP831469612), T4B));
 
2065
                         T7V = VADD(T4A, T4x);
 
2066
                         T7W = VADD(T4E, T4F);
 
2067
                         T7X = VFMA(LDK(KP195090322), T7V, VMUL(LDK(KP980785280), T7W));
 
2068
                         T86 = VFNMS(LDK(KP195090322), T7W, VMUL(LDK(KP980785280), T7V));
 
2069
                    }
 
2070
                    Tcp = VFNMS(LDK(KP382683432), Tan, VMUL(LDK(KP923879532), Tak));
 
2071
                    Tcq = VFMA(LDK(KP923879532), Tau, VMUL(LDK(KP382683432), Tar));
 
2072
                    Tcr = VSUB(Tcp, Tcq);
 
2073
                    TdH = VADD(Tcp, Tcq);
 
2074
                    {
 
2075
                         V T7S, T7T, Tao, Tav;
 
2076
                         T7S = VADD(T4c, T4j);
 
2077
                         T7T = VADD(T4o, T4l);
 
2078
                         T7U = VFNMS(LDK(KP195090322), T7T, VMUL(LDK(KP980785280), T7S));
 
2079
                         T85 = VFMA(LDK(KP980785280), T7T, VMUL(LDK(KP195090322), T7S));
 
2080
                         Tao = VFMA(LDK(KP382683432), Tak, VMUL(LDK(KP923879532), Tan));
 
2081
                         Tav = VFNMS(LDK(KP382683432), Tau, VMUL(LDK(KP923879532), Tar));
 
2082
                         Taw = VSUB(Tao, Tav);
 
2083
                         TdC = VADD(Tao, Tav);
 
2084
                    }
 
2085
               }
 
2086
               {
 
2087
                    V Tbz, TbA, T3, Tf5, T18, Tbu, Tbv, Ta, Tf6, T15, Ti, Tf1, T12, Tbp, Tp;
 
2088
                    V Tf2, T13, Tbs, T6, T9, TbB, TbC;
 
2089
                    {
 
2090
                         V T1, T2, T16, T17;
 
2091
                         T1 = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
 
2092
                         T2 = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)]));
 
2093
                         Tbz = VADD(T1, T2);
 
2094
                         T16 = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
 
2095
                         T17 = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)]));
 
2096
                         TbA = VADD(T16, T17);
 
2097
                         T3 = VSUB(T1, T2);
 
2098
                         Tf5 = VADD(Tbz, TbA);
 
2099
                         T18 = VSUB(T16, T17);
 
2100
                    }
 
2101
                    {
 
2102
                         V T4, T5, T7, T8;
 
2103
                         T4 = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
 
2104
                         T5 = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)]));
 
2105
                         T6 = VSUB(T4, T5);
 
2106
                         Tbu = VADD(T4, T5);
 
2107
                         T7 = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)]));
 
2108
                         T8 = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
 
2109
                         T9 = VSUB(T7, T8);
 
2110
                         Tbv = VADD(T7, T8);
 
2111
                    }
 
2112
                    Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
 
2113
                    Tf6 = VADD(Tbu, Tbv);
 
2114
                    T15 = VMUL(LDK(KP707106781), VSUB(T6, T9));
 
2115
                    {
 
2116
                         V Te, Tbn, Th, Tbo;
 
2117
                         {
 
2118
                              V Tc, Td, Tf, Tg;
 
2119
                              Tc = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
 
2120
                              Td = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)]));
 
2121
                              Te = VSUB(Tc, Td);
 
2122
                              Tbn = VADD(Tc, Td);
 
2123
                              Tf = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
 
2124
                              Tg = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)]));
 
2125
                              Th = VSUB(Tf, Tg);
 
2126
                              Tbo = VADD(Tf, Tg);
 
2127
                         }
 
2128
                         Ti = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
 
2129
                         Tf1 = VADD(Tbn, Tbo);
 
2130
                         T12 = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
 
2131
                         Tbp = VSUB(Tbn, Tbo);
 
2132
                    }
 
2133
                    {
 
2134
                         V Tl, Tbr, To, Tbq;
 
2135
                         {
 
2136
                              V Tj, Tk, Tm, Tn;
 
2137
                              Tj = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
 
2138
                              Tk = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)]));
 
2139
                              Tl = VSUB(Tj, Tk);
 
2140
                              Tbr = VADD(Tj, Tk);
 
2141
                              Tm = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)]));
 
2142
                              Tn = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
 
2143
                              To = VSUB(Tm, Tn);
 
2144
                              Tbq = VADD(Tm, Tn);
 
2145
                         }
 
2146
                         Tp = VFNMS(LDK(KP382683432), To, VMUL(LDK(KP923879532), Tl));
 
2147
                         Tf2 = VADD(Tbq, Tbr);
 
2148
                         T13 = VFMA(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
 
2149
                         Tbs = VSUB(Tbq, Tbr);
 
2150
                    }
 
2151
                    Tf3 = VSUB(Tf1, Tf2);
 
2152
                    Tf7 = VSUB(Tf5, Tf6);
 
2153
                    {
 
2154
                         V Tb, Tq, Tgj, Tgk;
 
2155
                         Tb = VSUB(T3, Ta);
 
2156
                         Tq = VSUB(Ti, Tp);
 
2157
                         Tr = VSUB(Tb, Tq);
 
2158
                         T5X = VADD(Tb, Tq);
 
2159
                         Tgj = VADD(Tf5, Tf6);
 
2160
                         Tgk = VADD(Tf1, Tf2);
 
2161
                         Tgl = VSUB(Tgj, Tgk);
 
2162
                         Th3 = VADD(Tgj, Tgk);
 
2163
                    }
 
2164
                    {
 
2165
                         V T14, T19, T78, T79;
 
2166
                         T14 = VSUB(T12, T13);
 
2167
                         T19 = VSUB(T15, T18);
 
2168
                         T1a = VSUB(T14, T19);
 
2169
                         T5V = VADD(T19, T14);
 
2170
                         T78 = VADD(Ti, Tp);
 
2171
                         T79 = VADD(T18, T15);
 
2172
                         T7a = VSUB(T78, T79);
 
2173
                         T95 = VADD(T79, T78);
 
2174
                    }
 
2175
                    TbB = VSUB(Tbz, TbA);
 
2176
                    TbC = VMUL(LDK(KP707106781), VADD(Tbp, Tbs));
 
2177
                    TbD = VSUB(TbB, TbC);
 
2178
                    Tdf = VADD(TbB, TbC);
 
2179
                    {
 
2180
                         V T6X, T6Y, Tbt, Tbw;
 
2181
                         T6X = VADD(T3, Ta);
 
2182
                         T6Y = VADD(T12, T13);
 
2183
                         T6Z = VSUB(T6X, T6Y);
 
2184
                         T97 = VADD(T6X, T6Y);
 
2185
                         Tbt = VMUL(LDK(KP707106781), VSUB(Tbp, Tbs));
 
2186
                         Tbw = VSUB(Tbu, Tbv);
 
2187
                         Tbx = VSUB(Tbt, Tbw);
 
2188
                         Tdi = VADD(Tbw, Tbt);
 
2189
                    }
 
2190
               }
 
2191
               {
 
2192
                    V TaK, TaJ, T2U, TeE, T2Z, TaF, TaG, T2R, TeF, T30, T2C, TeH, T32, TaA, T2J;
 
2193
                    V TeI, T33, TaD, T2N, T2Q, TaL, TaM;
 
2194
                    {
 
2195
                         V T2S, T2T, T2X, T2Y;
 
2196
                         T2S = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
 
2197
                         T2T = LD(&(xi[WS(is, 98)]), ivs, &(xi[0]));
 
2198
                         TaK = VADD(T2S, T2T);
 
2199
                         T2X = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
 
2200
                         T2Y = LD(&(xi[WS(is, 66)]), ivs, &(xi[0]));
 
2201
                         TaJ = VADD(T2X, T2Y);
 
2202
                         T2U = VSUB(T2S, T2T);
 
2203
                         TeE = VADD(TaJ, TaK);
 
2204
                         T2Z = VSUB(T2X, T2Y);
 
2205
                    }
 
2206
                    {
 
2207
                         V T2L, T2M, T2O, T2P;
 
2208
                         T2L = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
 
2209
                         T2M = LD(&(xi[WS(is, 82)]), ivs, &(xi[0]));
 
2210
                         T2N = VSUB(T2L, T2M);
 
2211
                         TaF = VADD(T2L, T2M);
 
2212
                         T2O = LD(&(xi[WS(is, 114)]), ivs, &(xi[0]));
 
2213
                         T2P = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
 
2214
                         T2Q = VSUB(T2O, T2P);
 
2215
                         TaG = VADD(T2O, T2P);
 
2216
                    }
 
2217
                    T2R = VMUL(LDK(KP707106781), VSUB(T2N, T2Q));
 
2218
                    TeF = VADD(TaF, TaG);
 
2219
                    T30 = VMUL(LDK(KP707106781), VADD(T2N, T2Q));
 
2220
                    {
 
2221
                         V T2y, Tay, T2B, Taz;
 
2222
                         {
 
2223
                              V T2w, T2x, T2z, T2A;
 
2224
                              T2w = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
 
2225
                              T2x = LD(&(xi[WS(is, 74)]), ivs, &(xi[0]));
 
2226
                              T2y = VSUB(T2w, T2x);
 
2227
                              Tay = VADD(T2w, T2x);
 
2228
                              T2z = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
 
2229
                              T2A = LD(&(xi[WS(is, 106)]), ivs, &(xi[0]));
 
2230
                              T2B = VSUB(T2z, T2A);
 
2231
                              Taz = VADD(T2z, T2A);
 
2232
                         }
 
2233
                         T2C = VFNMS(LDK(KP382683432), T2B, VMUL(LDK(KP923879532), T2y));
 
2234
                         TeH = VADD(Tay, Taz);
 
2235
                         T32 = VFMA(LDK(KP382683432), T2y, VMUL(LDK(KP923879532), T2B));
 
2236
                         TaA = VSUB(Tay, Taz);
 
2237
                    }
 
2238
                    {
 
2239
                         V T2F, TaB, T2I, TaC;
 
2240
                         {
 
2241
                              V T2D, T2E, T2G, T2H;
 
2242
                              T2D = LD(&(xi[WS(is, 122)]), ivs, &(xi[0]));
 
2243
                              T2E = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
 
2244
                              T2F = VSUB(T2D, T2E);
 
2245
                              TaB = VADD(T2D, T2E);
 
2246
                              T2G = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
 
2247
                              T2H = LD(&(xi[WS(is, 90)]), ivs, &(xi[0]));
 
2248
                              T2I = VSUB(T2G, T2H);
 
2249
                              TaC = VADD(T2G, T2H);
 
2250
                         }
 
2251
                         T2J = VFMA(LDK(KP923879532), T2F, VMUL(LDK(KP382683432), T2I));
 
2252
                         TeI = VADD(TaB, TaC);
 
2253
                         T33 = VFNMS(LDK(KP382683432), T2F, VMUL(LDK(KP923879532), T2I));
 
2254
                         TaD = VSUB(TaB, TaC);
 
2255
                    }
 
2256
                    Tgy = VADD(TeE, TeF);
 
2257
                    Tgz = VADD(TeH, TeI);
 
2258
                    TgA = VSUB(Tgy, Tgz);
 
2259
                    TaL = VSUB(TaJ, TaK);
 
2260
                    TaM = VMUL(LDK(KP707106781), VADD(TaA, TaD));
 
2261
                    TaN = VSUB(TaL, TaM);
 
2262
                    Tdv = VADD(TaL, TaM);
 
2263
                    {
 
2264
                         V TeG, TeJ, T2K, T2V;
 
2265
                         TeG = VSUB(TeE, TeF);
 
2266
                         TeJ = VSUB(TeH, TeI);
 
2267
                         TeK = VFMA(LDK(KP382683432), TeG, VMUL(LDK(KP923879532), TeJ));
 
2268
                         Tfu = VFNMS(LDK(KP382683432), TeJ, VMUL(LDK(KP923879532), TeG));
 
2269
                         T2K = VSUB(T2C, T2J);
 
2270
                         T2V = VSUB(T2R, T2U);
 
2271
                         T2W = VSUB(T2K, T2V);
 
2272
                         T5M = VADD(T2V, T2K);
 
2273
                    }
 
2274
                    {
 
2275
                         V T31, T34, T7D, T7E;
 
2276
                         T31 = VSUB(T2Z, T30);
 
2277
                         T34 = VSUB(T32, T33);
 
2278
                         T35 = VSUB(T31, T34);
 
2279
                         T5N = VADD(T31, T34);
 
2280
                         T7D = VADD(T32, T33);
 
2281
                         T7E = VADD(T2U, T2R);
 
2282
                         T7F = VSUB(T7D, T7E);
 
2283
                         T8X = VADD(T7E, T7D);
 
2284
                    }
 
2285
                    {
 
2286
                         V TaE, TaH, T7A, T7B;
 
2287
                         TaE = VMUL(LDK(KP707106781), VSUB(TaA, TaD));
 
2288
                         TaH = VSUB(TaF, TaG);
 
2289
                         TaI = VSUB(TaE, TaH);
 
2290
                         Tdu = VADD(TaH, TaE);
 
2291
                         T7A = VADD(T2Z, T30);
 
2292
                         T7B = VADD(T2C, T2J);
 
2293
                         T7C = VSUB(T7A, T7B);
 
2294
                         T8W = VADD(T7A, T7B);
 
2295
                    }
 
2296
               }
 
2297
               {
 
2298
                    V Tb1, Tb0, T3v, TeO, T3A, TaW, TaX, T3s, TeP, T3B, T3d, TeL, T3D, TaR, T3k;
 
2299
                    V TeM, T3E, TaU, T3o, T3r, Tb2, Tb3;
 
2300
                    {
 
2301
                         V T3t, T3u, T3y, T3z;
 
2302
                         T3t = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
 
2303
                         T3u = LD(&(xi[WS(is, 94)]), ivs, &(xi[0]));
 
2304
                         Tb1 = VADD(T3t, T3u);
 
2305
                         T3y = LD(&(xi[WS(is, 126)]), ivs, &(xi[0]));
 
2306
                         T3z = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
 
2307
                         Tb0 = VADD(T3y, T3z);
 
2308
                         T3v = VSUB(T3t, T3u);
 
2309
                         TeO = VADD(Tb0, Tb1);
 
2310
                         T3A = VSUB(T3y, T3z);
 
2311
                    }
 
2312
                    {
 
2313
                         V T3m, T3n, T3p, T3q;
 
2314
                         T3m = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
 
2315
                         T3n = LD(&(xi[WS(is, 78)]), ivs, &(xi[0]));
 
2316
                         T3o = VSUB(T3m, T3n);
 
2317
                         TaW = VADD(T3m, T3n);
 
2318
                         T3p = LD(&(xi[WS(is, 110)]), ivs, &(xi[0]));
 
2319
                         T3q = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
 
2320
                         T3r = VSUB(T3p, T3q);
 
2321
                         TaX = VADD(T3p, T3q);
 
2322
                    }
 
2323
                    T3s = VMUL(LDK(KP707106781), VSUB(T3o, T3r));
 
2324
                    TeP = VADD(TaW, TaX);
 
2325
                    T3B = VMUL(LDK(KP707106781), VADD(T3o, T3r));
 
2326
                    {
 
2327
                         V T39, TaP, T3c, TaQ;
 
2328
                         {
 
2329
                              V T37, T38, T3a, T3b;
 
2330
                              T37 = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
 
2331
                              T38 = LD(&(xi[WS(is, 70)]), ivs, &(xi[0]));
 
2332
                              T39 = VSUB(T37, T38);
 
2333
                              TaP = VADD(T37, T38);
 
2334
                              T3a = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
 
2335
                              T3b = LD(&(xi[WS(is, 102)]), ivs, &(xi[0]));
 
2336
                              T3c = VSUB(T3a, T3b);
 
2337
                              TaQ = VADD(T3a, T3b);
 
2338
                         }
 
2339
                         T3d = VFNMS(LDK(KP382683432), T3c, VMUL(LDK(KP923879532), T39));
 
2340
                         TeL = VADD(TaP, TaQ);
 
2341
                         T3D = VFMA(LDK(KP382683432), T39, VMUL(LDK(KP923879532), T3c));
 
2342
                         TaR = VSUB(TaP, TaQ);
 
2343
                    }
 
2344
                    {
 
2345
                         V T3g, TaS, T3j, TaT;
 
2346
                         {
 
2347
                              V T3e, T3f, T3h, T3i;
 
2348
                              T3e = LD(&(xi[WS(is, 118)]), ivs, &(xi[0]));
 
2349
                              T3f = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
 
2350
                              T3g = VSUB(T3e, T3f);
 
2351
                              TaS = VADD(T3e, T3f);
 
2352
                              T3h = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
 
2353
                              T3i = LD(&(xi[WS(is, 86)]), ivs, &(xi[0]));
 
2354
                              T3j = VSUB(T3h, T3i);
 
2355
                              TaT = VADD(T3h, T3i);
 
2356
                         }
 
2357
                         T3k = VFMA(LDK(KP923879532), T3g, VMUL(LDK(KP382683432), T3j));
 
2358
                         TeM = VADD(TaS, TaT);
 
2359
                         T3E = VFNMS(LDK(KP382683432), T3g, VMUL(LDK(KP923879532), T3j));
 
2360
                         TaU = VSUB(TaS, TaT);
 
2361
                    }
 
2362
                    TgB = VADD(TeO, TeP);
 
2363
                    TgC = VADD(TeL, TeM);
 
2364
                    TgD = VSUB(TgB, TgC);
 
2365
                    Tb2 = VSUB(Tb0, Tb1);
 
2366
                    Tb3 = VMUL(LDK(KP707106781), VADD(TaR, TaU));
 
2367
                    Tb4 = VSUB(Tb2, Tb3);
 
2368
                    Tdy = VADD(Tb2, Tb3);
 
2369
                    {
 
2370
                         V TeN, TeQ, T3l, T3w;
 
2371
                         TeN = VSUB(TeL, TeM);
 
2372
                         TeQ = VSUB(TeO, TeP);
 
2373
                         TeR = VFNMS(LDK(KP382683432), TeQ, VMUL(LDK(KP923879532), TeN));
 
2374
                         Tfv = VFMA(LDK(KP923879532), TeQ, VMUL(LDK(KP382683432), TeN));
 
2375
                         T3l = VSUB(T3d, T3k);
 
2376
                         T3w = VSUB(T3s, T3v);
 
2377
                         T3x = VSUB(T3l, T3w);
 
2378
                         T5P = VADD(T3w, T3l);
 
2379
                    }
 
2380
                    {
 
2381
                         V T3C, T3F, T7K, T7L;
 
2382
                         T3C = VSUB(T3A, T3B);
 
2383
                         T3F = VSUB(T3D, T3E);
 
2384
                         T3G = VSUB(T3C, T3F);
 
2385
                         T5Q = VADD(T3C, T3F);
 
2386
                         T7K = VADD(T3A, T3B);
 
2387
                         T7L = VADD(T3d, T3k);
 
2388
                         T7M = VSUB(T7K, T7L);
 
2389
                         T90 = VADD(T7K, T7L);
 
2390
                    }
 
2391
                    {
 
2392
                         V TaV, TaY, T7H, T7I;
 
2393
                         TaV = VMUL(LDK(KP707106781), VSUB(TaR, TaU));
 
2394
                         TaY = VSUB(TaW, TaX);
 
2395
                         TaZ = VSUB(TaV, TaY);
 
2396
                         Tdx = VADD(TaY, TaV);
 
2397
                         T7H = VADD(T3D, T3E);
 
2398
                         T7I = VADD(T3v, T3s);
 
2399
                         T7J = VSUB(T7H, T7I);
 
2400
                         T8Z = VADD(T7I, T7H);
 
2401
                    }
 
2402
               }
 
2403
               {
 
2404
                    V TB, TeU, TF, Tba, TS, TeX, TW, Tbh, Ty, TeV, TG, Tbd, TP, TeY, TX;
 
2405
                    V Tbk;
 
2406
                    {
 
2407
                         V Tz, TA, Tb9, TD, TE, Tb8;
 
2408
                         Tz = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
 
2409
                         TA = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)]));
 
2410
                         Tb9 = VADD(Tz, TA);
 
2411
                         TD = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
 
2412
                         TE = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)]));
 
2413
                         Tb8 = VADD(TD, TE);
 
2414
                         TB = VSUB(Tz, TA);
 
2415
                         TeU = VADD(Tb8, Tb9);
 
2416
                         TF = VSUB(TD, TE);
 
2417
                         Tba = VSUB(Tb8, Tb9);
 
2418
                    }
 
2419
                    {
 
2420
                         V TQ, TR, Tbg, TU, TV, Tbf;
 
2421
                         TQ = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
 
2422
                         TR = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)]));
 
2423
                         Tbg = VADD(TQ, TR);
 
2424
                         TU = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)]));
 
2425
                         TV = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
 
2426
                         Tbf = VADD(TU, TV);
 
2427
                         TS = VSUB(TQ, TR);
 
2428
                         TeX = VADD(Tbf, Tbg);
 
2429
                         TW = VSUB(TU, TV);
 
2430
                         Tbh = VSUB(Tbf, Tbg);
 
2431
                    }
 
2432
                    {
 
2433
                         V Tu, Tbb, Tx, Tbc;
 
2434
                         {
 
2435
                              V Ts, Tt, Tv, Tw;
 
2436
                              Ts = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
 
2437
                              Tt = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)]));
 
2438
                              Tu = VSUB(Ts, Tt);
 
2439
                              Tbb = VADD(Ts, Tt);
 
2440
                              Tv = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)]));
 
2441
                              Tw = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
 
2442
                              Tx = VSUB(Tv, Tw);
 
2443
                              Tbc = VADD(Tv, Tw);
 
2444
                         }
 
2445
                         Ty = VMUL(LDK(KP707106781), VSUB(Tu, Tx));
 
2446
                         TeV = VADD(Tbb, Tbc);
 
2447
                         TG = VMUL(LDK(KP707106781), VADD(Tu, Tx));
 
2448
                         Tbd = VSUB(Tbb, Tbc);
 
2449
                    }
 
2450
                    {
 
2451
                         V TL, Tbi, TO, Tbj;
 
2452
                         {
 
2453
                              V TJ, TK, TM, TN;
 
2454
                              TJ = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
 
2455
                              TK = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)]));
 
2456
                              TL = VSUB(TJ, TK);
 
2457
                              Tbi = VADD(TJ, TK);
 
2458
                              TM = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)]));
 
2459
                              TN = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
 
2460
                              TO = VSUB(TM, TN);
 
2461
                              Tbj = VADD(TM, TN);
 
2462
                         }
 
2463
                         TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
 
2464
                         TeY = VADD(Tbi, Tbj);
 
2465
                         TX = VMUL(LDK(KP707106781), VADD(TL, TO));
 
2466
                         Tbk = VSUB(Tbi, Tbj);
 
2467
                    }
 
2468
                    {
 
2469
                         V Tbe, Tbl, TeW, TeZ;
 
2470
                         Tbe = VFNMS(LDK(KP382683432), Tbd, VMUL(LDK(KP923879532), Tba));
 
2471
                         Tbl = VFMA(LDK(KP923879532), Tbh, VMUL(LDK(KP382683432), Tbk));
 
2472
                         Tbm = VSUB(Tbe, Tbl);
 
2473
                         Tdg = VADD(Tbe, Tbl);
 
2474
                         {
 
2475
                              V TbE, TbF, Tgm, Tgn;
 
2476
                              TbE = VFMA(LDK(KP382683432), Tba, VMUL(LDK(KP923879532), Tbd));
 
2477
                              TbF = VFNMS(LDK(KP382683432), Tbh, VMUL(LDK(KP923879532), Tbk));
 
2478
                              TbG = VSUB(TbE, TbF);
 
2479
                              Tdj = VADD(TbE, TbF);
 
2480
                              Tgm = VADD(TeU, TeV);
 
2481
                              Tgn = VADD(TeX, TeY);
 
2482
                              Tgo = VSUB(Tgm, Tgn);
 
2483
                              Th4 = VADD(Tgm, Tgn);
 
2484
                         }
 
2485
                         TeW = VSUB(TeU, TeV);
 
2486
                         TeZ = VSUB(TeX, TeY);
 
2487
                         Tf0 = VMUL(LDK(KP707106781), VSUB(TeW, TeZ));
 
2488
                         Tf8 = VMUL(LDK(KP707106781), VADD(TeW, TeZ));
 
2489
                         {
 
2490
                              V T72, T7b, T75, T7c;
 
2491
                              {
 
2492
                                   V T70, T71, T73, T74;
 
2493
                                   T70 = VADD(TB, Ty);
 
2494
                                   T71 = VADD(TF, TG);
 
2495
                                   T72 = VFMA(LDK(KP980785280), T70, VMUL(LDK(KP195090322), T71));
 
2496
                                   T7b = VFNMS(LDK(KP195090322), T70, VMUL(LDK(KP980785280), T71));
 
2497
                                   T73 = VADD(TS, TP);
 
2498
                                   T74 = VADD(TW, TX);
 
2499
                                   T75 = VFNMS(LDK(KP195090322), T74, VMUL(LDK(KP980785280), T73));
 
2500
                                   T7c = VFMA(LDK(KP195090322), T73, VMUL(LDK(KP980785280), T74));
 
2501
                              }
 
2502
                              T76 = VSUB(T72, T75);
 
2503
                              T98 = VADD(T7b, T7c);
 
2504
                              T7d = VSUB(T7b, T7c);
 
2505
                              T94 = VADD(T72, T75);
 
2506
                         }
 
2507
                         {
 
2508
                              V TI, T1b, TZ, T1c;
 
2509
                              {
 
2510
                                   V TC, TH, TT, TY;
 
2511
                                   TC = VSUB(Ty, TB);
 
2512
                                   TH = VSUB(TF, TG);
 
2513
                                   TI = VFMA(LDK(KP831469612), TC, VMUL(LDK(KP555570233), TH));
 
2514
                                   T1b = VFNMS(LDK(KP555570233), TC, VMUL(LDK(KP831469612), TH));
 
2515
                                   TT = VSUB(TP, TS);
 
2516
                                   TY = VSUB(TW, TX);
 
2517
                                   TZ = VFNMS(LDK(KP555570233), TY, VMUL(LDK(KP831469612), TT));
 
2518
                                   T1c = VFMA(LDK(KP555570233), TT, VMUL(LDK(KP831469612), TY));
 
2519
                              }
 
2520
                              T10 = VSUB(TI, TZ);
 
2521
                              T5Y = VADD(T1b, T1c);
 
2522
                              T1d = VSUB(T1b, T1c);
 
2523
                              T5U = VADD(TI, TZ);
 
2524
                         }
 
2525
                    }
 
2526
               }
 
2527
               {
 
2528
                    V T1Q, Tfb, T1U, TbL, T27, Tfe, T2b, TbS, T1N, Tfc, T1V, TbO, T24, Tff, T2c;
 
2529
                    V TbV;
 
2530
                    {
 
2531
                         V T1O, T1P, TbK, T1S, T1T, TbJ;
 
2532
                         T1O = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
 
2533
                         T1P = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)]));
 
2534
                         TbK = VADD(T1O, T1P);
 
2535
                         T1S = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
 
2536
                         T1T = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)]));
 
2537
                         TbJ = VADD(T1S, T1T);
 
2538
                         T1Q = VSUB(T1O, T1P);
 
2539
                         Tfb = VADD(TbJ, TbK);
 
2540
                         T1U = VSUB(T1S, T1T);
 
2541
                         TbL = VSUB(TbJ, TbK);
 
2542
                    }
 
2543
                    {
 
2544
                         V T25, T26, TbR, T29, T2a, TbQ;
 
2545
                         T25 = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
 
2546
                         T26 = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)]));
 
2547
                         TbR = VADD(T25, T26);
 
2548
                         T29 = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)]));
 
2549
                         T2a = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
 
2550
                         TbQ = VADD(T29, T2a);
 
2551
                         T27 = VSUB(T25, T26);
 
2552
                         Tfe = VADD(TbQ, TbR);
 
2553
                         T2b = VSUB(T29, T2a);
 
2554
                         TbS = VSUB(TbQ, TbR);
 
2555
                    }
 
2556
                    {
 
2557
                         V T1J, TbM, T1M, TbN;
 
2558
                         {
 
2559
                              V T1H, T1I, T1K, T1L;
 
2560
                              T1H = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
 
2561
                              T1I = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)]));
 
2562
                              T1J = VSUB(T1H, T1I);
 
2563
                              TbM = VADD(T1H, T1I);
 
2564
                              T1K = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)]));
 
2565
                              T1L = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
 
2566
                              T1M = VSUB(T1K, T1L);
 
2567
                              TbN = VADD(T1K, T1L);
 
2568
                         }
 
2569
                         T1N = VMUL(LDK(KP707106781), VSUB(T1J, T1M));
 
2570
                         Tfc = VADD(TbM, TbN);
 
2571
                         T1V = VMUL(LDK(KP707106781), VADD(T1J, T1M));
 
2572
                         TbO = VSUB(TbM, TbN);
 
2573
                    }
 
2574
                    {
 
2575
                         V T20, TbT, T23, TbU;
 
2576
                         {
 
2577
                              V T1Y, T1Z, T21, T22;
 
2578
                              T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
 
2579
                              T1Z = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)]));
 
2580
                              T20 = VSUB(T1Y, T1Z);
 
2581
                              TbT = VADD(T1Y, T1Z);
 
2582
                              T21 = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)]));
 
2583
                              T22 = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
 
2584
                              T23 = VSUB(T21, T22);
 
2585
                              TbU = VADD(T21, T22);
 
2586
                         }
 
2587
                         T24 = VMUL(LDK(KP707106781), VSUB(T20, T23));
 
2588
                         Tff = VADD(TbT, TbU);
 
2589
                         T2c = VMUL(LDK(KP707106781), VADD(T20, T23));
 
2590
                         TbV = VSUB(TbT, TbU);
 
2591
                    }
 
2592
                    {
 
2593
                         V TbP, TbW, Tfd, Tfg;
 
2594
                         TbP = VFNMS(LDK(KP382683432), TbO, VMUL(LDK(KP923879532), TbL));
 
2595
                         TbW = VFMA(LDK(KP923879532), TbS, VMUL(LDK(KP382683432), TbV));
 
2596
                         TbX = VSUB(TbP, TbW);
 
2597
                         Tdn = VADD(TbP, TbW);
 
2598
                         {
 
2599
                              V Tcf, Tcg, Tgt, Tgu;
 
2600
                              Tcf = VFMA(LDK(KP382683432), TbL, VMUL(LDK(KP923879532), TbO));
 
2601
                              Tcg = VFNMS(LDK(KP382683432), TbS, VMUL(LDK(KP923879532), TbV));
 
2602
                              Tch = VSUB(Tcf, Tcg);
 
2603
                              Tdq = VADD(Tcf, Tcg);
 
2604
                              Tgt = VADD(Tfb, Tfc);
 
2605
                              Tgu = VADD(Tfe, Tff);
 
2606
                              Tgv = VSUB(Tgt, Tgu);
 
2607
                              Th7 = VADD(Tgt, Tgu);
 
2608
                         }
 
2609
                         Tfd = VSUB(Tfb, Tfc);
 
2610
                         Tfg = VSUB(Tfe, Tff);
 
2611
                         Tfh = VMUL(LDK(KP707106781), VSUB(Tfd, Tfg));
 
2612
                         Tfp = VMUL(LDK(KP707106781), VADD(Tfd, Tfg));
 
2613
                         {
 
2614
                              V T7l, T7u, T7o, T7v;
 
2615
                              {
 
2616
                                   V T7j, T7k, T7m, T7n;
 
2617
                                   T7j = VADD(T1Q, T1N);
 
2618
                                   T7k = VADD(T1U, T1V);
 
2619
                                   T7l = VFMA(LDK(KP980785280), T7j, VMUL(LDK(KP195090322), T7k));
 
2620
                                   T7u = VFNMS(LDK(KP195090322), T7j, VMUL(LDK(KP980785280), T7k));
 
2621
                                   T7m = VADD(T27, T24);
 
2622
                                   T7n = VADD(T2b, T2c);
 
2623
                                   T7o = VFNMS(LDK(KP195090322), T7n, VMUL(LDK(KP980785280), T7m));
 
2624
                                   T7v = VFMA(LDK(KP195090322), T7m, VMUL(LDK(KP980785280), T7n));
 
2625
                              }
 
2626
                              T7p = VSUB(T7l, T7o);
 
2627
                              T9f = VADD(T7u, T7v);
 
2628
                              T7w = VSUB(T7u, T7v);
 
2629
                              T9b = VADD(T7l, T7o);
 
2630
                         }
 
2631
                         {
 
2632
                              V T1X, T2q, T2e, T2r;
 
2633
                              {
 
2634
                                   V T1R, T1W, T28, T2d;
 
2635
                                   T1R = VSUB(T1N, T1Q);
 
2636
                                   T1W = VSUB(T1U, T1V);
 
2637
                                   T1X = VFMA(LDK(KP831469612), T1R, VMUL(LDK(KP555570233), T1W));
 
2638
                                   T2q = VFNMS(LDK(KP555570233), T1R, VMUL(LDK(KP831469612), T1W));
 
2639
                                   T28 = VSUB(T24, T27);
 
2640
                                   T2d = VSUB(T2b, T2c);
 
2641
                                   T2e = VFNMS(LDK(KP555570233), T2d, VMUL(LDK(KP831469612), T28));
 
2642
                                   T2r = VFMA(LDK(KP555570233), T28, VMUL(LDK(KP831469612), T2d));
 
2643
                              }
 
2644
                              T2f = VSUB(T1X, T2e);
 
2645
                              T65 = VADD(T2q, T2r);
 
2646
                              T2s = VSUB(T2q, T2r);
 
2647
                              T61 = VADD(T1X, T2e);
 
2648
                         }
 
2649
                    }
 
2650
               }
 
2651
               {
 
2652
                    V Tgx, TgW, TgR, TgZ, TgI, TgY, TgO, TgV;
 
2653
                    {
 
2654
                         V Tgp, Tgw, TgP, TgQ;
 
2655
                         Tgp = VFNMS(LDK(KP382683432), Tgo, VMUL(LDK(KP923879532), Tgl));
 
2656
                         Tgw = VFMA(LDK(KP923879532), Tgs, VMUL(LDK(KP382683432), Tgv));
 
2657
                         Tgx = VSUB(Tgp, Tgw);
 
2658
                         TgW = VADD(Tgp, Tgw);
 
2659
                         TgP = VFMA(LDK(KP382683432), Tgl, VMUL(LDK(KP923879532), Tgo));
 
2660
                         TgQ = VFNMS(LDK(KP382683432), Tgs, VMUL(LDK(KP923879532), Tgv));
 
2661
                         TgR = VSUB(TgP, TgQ);
 
2662
                         TgZ = VADD(TgP, TgQ);
 
2663
                    }
 
2664
                    {
 
2665
                         V TgE, TgH, TgM, TgN;
 
2666
                         TgE = VMUL(LDK(KP707106781), VSUB(TgA, TgD));
 
2667
                         TgH = VSUB(TgF, TgG);
 
2668
                         TgI = VSUB(TgE, TgH);
 
2669
                         TgY = VADD(TgH, TgE);
 
2670
                         TgM = VSUB(TgK, TgL);
 
2671
                         TgN = VMUL(LDK(KP707106781), VADD(TgA, TgD));
 
2672
                         TgO = VSUB(TgM, TgN);
 
2673
                         TgV = VADD(TgM, TgN);
 
2674
                    }
 
2675
                    {
 
2676
                         V TgJ, TgS, Th1, Th2;
 
2677
                         TgJ = VBYI(VSUB(Tgx, TgI));
 
2678
                         TgS = VSUB(TgO, TgR);
 
2679
                         ST(&(xo[WS(os, 40)]), VADD(TgJ, TgS), ovs, &(xo[0]));
 
2680
                         ST(&(xo[WS(os, 88)]), VSUB(TgS, TgJ), ovs, &(xo[0]));
 
2681
                         Th1 = VSUB(TgV, TgW);
 
2682
                         Th2 = VBYI(VSUB(TgZ, TgY));
 
2683
                         ST(&(xo[WS(os, 72)]), VSUB(Th1, Th2), ovs, &(xo[0]));
 
2684
                         ST(&(xo[WS(os, 56)]), VADD(Th1, Th2), ovs, &(xo[0]));
 
2685
                    }
 
2686
                    {
 
2687
                         V TgT, TgU, TgX, Th0;
 
2688
                         TgT = VBYI(VADD(TgI, Tgx));
 
2689
                         TgU = VADD(TgO, TgR);
 
2690
                         ST(&(xo[WS(os, 24)]), VADD(TgT, TgU), ovs, &(xo[0]));
 
2691
                         ST(&(xo[WS(os, 104)]), VSUB(TgU, TgT), ovs, &(xo[0]));
 
2692
                         TgX = VADD(TgV, TgW);
 
2693
                         Th0 = VBYI(VADD(TgY, TgZ));
 
2694
                         ST(&(xo[WS(os, 120)]), VSUB(TgX, Th0), ovs, &(xo[0]));
 
2695
                         ST(&(xo[WS(os, 8)]), VADD(TgX, Th0), ovs, &(xo[0]));
 
2696
                    }
 
2697
               }
 
2698
               {
 
2699
                    V Th9, Thh, Thq, Ths, Thc, Thm, Thg, Thl, Thn, Thr;
 
2700
                    {
 
2701
                         V Th5, Th8, Tho, Thp;
 
2702
                         Th5 = VSUB(Th3, Th4);
 
2703
                         Th8 = VSUB(Th6, Th7);
 
2704
                         Th9 = VMUL(LDK(KP707106781), VSUB(Th5, Th8));
 
2705
                         Thh = VMUL(LDK(KP707106781), VADD(Th5, Th8));
 
2706
                         Tho = VADD(Th3, Th4);
 
2707
                         Thp = VADD(Th6, Th7);
 
2708
                         Thq = VBYI(VSUB(Tho, Thp));
 
2709
                         Ths = VADD(Tho, Thp);
 
2710
                    }
 
2711
                    {
 
2712
                         V Tha, Thb, The, Thf;
 
2713
                         Tha = VADD(Tgy, Tgz);
 
2714
                         Thb = VADD(TgB, TgC);
 
2715
                         Thc = VSUB(Tha, Thb);
 
2716
                         Thm = VADD(Tha, Thb);
 
2717
                         The = VADD(TgK, TgL);
 
2718
                         Thf = VADD(TgF, TgG);
 
2719
                         Thg = VSUB(The, Thf);
 
2720
                         Thl = VADD(The, Thf);
 
2721
                    }
 
2722
                    Thn = VSUB(Thl, Thm);
 
2723
                    ST(&(xo[WS(os, 96)]), VSUB(Thn, Thq), ovs, &(xo[0]));
 
2724
                    ST(&(xo[WS(os, 32)]), VADD(Thn, Thq), ovs, &(xo[0]));
 
2725
                    Thr = VADD(Thl, Thm);
 
2726
                    ST(&(xo[WS(os, 64)]), VSUB(Thr, Ths), ovs, &(xo[0]));
 
2727
                    ST(&(xo[0]), VADD(Thr, Ths), ovs, &(xo[0]));
 
2728
                    {
 
2729
                         V Thd, Thi, Thj, Thk;
 
2730
                         Thd = VBYI(VSUB(Th9, Thc));
 
2731
                         Thi = VSUB(Thg, Thh);
 
2732
                         ST(&(xo[WS(os, 48)]), VADD(Thd, Thi), ovs, &(xo[0]));
 
2733
                         ST(&(xo[WS(os, 80)]), VSUB(Thi, Thd), ovs, &(xo[0]));
 
2734
                         Thj = VBYI(VADD(Thc, Th9));
 
2735
                         Thk = VADD(Thg, Thh);
 
2736
                         ST(&(xo[WS(os, 16)]), VADD(Thj, Thk), ovs, &(xo[0]));
 
2737
                         ST(&(xo[WS(os, 112)]), VSUB(Thk, Thj), ovs, &(xo[0]));
 
2738
                    }
 
2739
               }
 
2740
               {
 
2741
                    V TeT, TfM, TfC, TfK, Tfs, TfN, TfF, TfJ;
 
2742
                    {
 
2743
                         V TeD, TeS, Tfw, TfB;
 
2744
                         TeD = VSUB(Tev, TeC);
 
2745
                         TeS = VSUB(TeK, TeR);
 
2746
                         TeT = VSUB(TeD, TeS);
 
2747
                         TfM = VADD(TeD, TeS);
 
2748
                         Tfw = VSUB(Tfu, Tfv);
 
2749
                         TfB = VSUB(Tfx, TfA);
 
2750
                         TfC = VSUB(Tfw, TfB);
 
2751
                         TfK = VADD(TfB, Tfw);
 
2752
                         {
 
2753
                              V Tfa, TfD, Tfr, TfE;
 
2754
                              {
 
2755
                                   V Tf4, Tf9, Tfl, Tfq;
 
2756
                                   Tf4 = VSUB(Tf0, Tf3);
 
2757
                                   Tf9 = VSUB(Tf7, Tf8);
 
2758
                                   Tfa = VFMA(LDK(KP831469612), Tf4, VMUL(LDK(KP555570233), Tf9));
 
2759
                                   TfD = VFNMS(LDK(KP555570233), Tf4, VMUL(LDK(KP831469612), Tf9));
 
2760
                                   Tfl = VSUB(Tfh, Tfk);
 
2761
                                   Tfq = VSUB(Tfo, Tfp);
 
2762
                                   Tfr = VFNMS(LDK(KP555570233), Tfq, VMUL(LDK(KP831469612), Tfl));
 
2763
                                   TfE = VFMA(LDK(KP555570233), Tfl, VMUL(LDK(KP831469612), Tfq));
 
2764
                              }
 
2765
                              Tfs = VSUB(Tfa, Tfr);
 
2766
                              TfN = VADD(TfD, TfE);
 
2767
                              TfF = VSUB(TfD, TfE);
 
2768
                              TfJ = VADD(Tfa, Tfr);
 
2769
                         }
 
2770
                    }
 
2771
                    {
 
2772
                         V Tft, TfG, TfP, TfQ;
 
2773
                         Tft = VADD(TeT, Tfs);
 
2774
                         TfG = VBYI(VADD(TfC, TfF));
 
2775
                         ST(&(xo[WS(os, 108)]), VSUB(Tft, TfG), ovs, &(xo[0]));
 
2776
                         ST(&(xo[WS(os, 20)]), VADD(Tft, TfG), ovs, &(xo[0]));
 
2777
                         TfP = VBYI(VADD(TfK, TfJ));
 
2778
                         TfQ = VADD(TfM, TfN);
 
2779
                         ST(&(xo[WS(os, 12)]), VADD(TfP, TfQ), ovs, &(xo[0]));
 
2780
                         ST(&(xo[WS(os, 116)]), VSUB(TfQ, TfP), ovs, &(xo[0]));
 
2781
                    }
 
2782
                    {
 
2783
                         V TfH, TfI, TfL, TfO;
 
2784
                         TfH = VSUB(TeT, Tfs);
 
2785
                         TfI = VBYI(VSUB(TfF, TfC));
 
2786
                         ST(&(xo[WS(os, 84)]), VSUB(TfH, TfI), ovs, &(xo[0]));
 
2787
                         ST(&(xo[WS(os, 44)]), VADD(TfH, TfI), ovs, &(xo[0]));
 
2788
                         TfL = VBYI(VSUB(TfJ, TfK));
 
2789
                         TfO = VSUB(TfM, TfN);
 
2790
                         ST(&(xo[WS(os, 52)]), VADD(TfL, TfO), ovs, &(xo[0]));
 
2791
                         ST(&(xo[WS(os, 76)]), VSUB(TfO, TfL), ovs, &(xo[0]));
 
2792
                    }
 
2793
               }
 
2794
               {
 
2795
                    V TfT, Tge, Tg4, Tgc, Tg0, Tgf, Tg7, Tgb;
 
2796
                    {
 
2797
                         V TfR, TfS, Tg2, Tg3;
 
2798
                         TfR = VADD(Tev, TeC);
 
2799
                         TfS = VADD(Tfu, Tfv);
 
2800
                         TfT = VSUB(TfR, TfS);
 
2801
                         Tge = VADD(TfR, TfS);
 
2802
                         Tg2 = VADD(TeK, TeR);
 
2803
                         Tg3 = VADD(TfA, Tfx);
 
2804
                         Tg4 = VSUB(Tg2, Tg3);
 
2805
                         Tgc = VADD(Tg3, Tg2);
 
2806
                         {
 
2807
                              V TfW, Tg5, TfZ, Tg6;
 
2808
                              {
 
2809
                                   V TfU, TfV, TfX, TfY;
 
2810
                                   TfU = VADD(Tf3, Tf0);
 
2811
                                   TfV = VADD(Tf7, Tf8);
 
2812
                                   TfW = VFMA(LDK(KP980785280), TfU, VMUL(LDK(KP195090322), TfV));
 
2813
                                   Tg5 = VFNMS(LDK(KP195090322), TfU, VMUL(LDK(KP980785280), TfV));
 
2814
                                   TfX = VADD(Tfk, Tfh);
 
2815
                                   TfY = VADD(Tfo, Tfp);
 
2816
                                   TfZ = VFNMS(LDK(KP195090322), TfY, VMUL(LDK(KP980785280), TfX));
 
2817
                                   Tg6 = VFMA(LDK(KP195090322), TfX, VMUL(LDK(KP980785280), TfY));
 
2818
                              }
 
2819
                              Tg0 = VSUB(TfW, TfZ);
 
2820
                              Tgf = VADD(Tg5, Tg6);
 
2821
                              Tg7 = VSUB(Tg5, Tg6);
 
2822
                              Tgb = VADD(TfW, TfZ);
 
2823
                         }
 
2824
                    }
 
2825
                    {
 
2826
                         V Tg1, Tg8, Tgh, Tgi;
 
2827
                         Tg1 = VADD(TfT, Tg0);
 
2828
                         Tg8 = VBYI(VADD(Tg4, Tg7));
 
2829
                         ST(&(xo[WS(os, 100)]), VSUB(Tg1, Tg8), ovs, &(xo[0]));
 
2830
                         ST(&(xo[WS(os, 28)]), VADD(Tg1, Tg8), ovs, &(xo[0]));
 
2831
                         Tgh = VBYI(VADD(Tgc, Tgb));
 
2832
                         Tgi = VADD(Tge, Tgf);
 
2833
                         ST(&(xo[WS(os, 4)]), VADD(Tgh, Tgi), ovs, &(xo[0]));
 
2834
                         ST(&(xo[WS(os, 124)]), VSUB(Tgi, Tgh), ovs, &(xo[0]));
 
2835
                    }
 
2836
                    {
 
2837
                         V Tg9, Tga, Tgd, Tgg;
 
2838
                         Tg9 = VSUB(TfT, Tg0);
 
2839
                         Tga = VBYI(VSUB(Tg7, Tg4));
 
2840
                         ST(&(xo[WS(os, 92)]), VSUB(Tg9, Tga), ovs, &(xo[0]));
 
2841
                         ST(&(xo[WS(os, 36)]), VADD(Tg9, Tga), ovs, &(xo[0]));
 
2842
                         Tgd = VBYI(VSUB(Tgb, Tgc));
 
2843
                         Tgg = VSUB(Tge, Tgf);
 
2844
                         ST(&(xo[WS(os, 60)]), VADD(Tgd, Tgg), ovs, &(xo[0]));
 
2845
                         ST(&(xo[WS(os, 68)]), VSUB(Tgg, Tgd), ovs, &(xo[0]));
 
2846
                    }
 
2847
               }
 
2848
               {
 
2849
                    V Tb7, Td8, TcI, Td0, Tcy, Tda, TcG, TcP, Tck, TcJ, TcB, TcF, TcW, Tdb, Td3;
 
2850
                    V Td7;
 
2851
                    {
 
2852
                         V Tax, TcZ, Tb6, TcY, TaO, Tb5;
 
2853
                         Tax = VSUB(Tah, Taw);
 
2854
                         TcZ = VADD(Tcw, Tcr);
 
2855
                         TaO = VFMA(LDK(KP831469612), TaI, VMUL(LDK(KP555570233), TaN));
 
2856
                         Tb5 = VFNMS(LDK(KP555570233), Tb4, VMUL(LDK(KP831469612), TaZ));
 
2857
                         Tb6 = VSUB(TaO, Tb5);
 
2858
                         TcY = VADD(TaO, Tb5);
 
2859
                         Tb7 = VSUB(Tax, Tb6);
 
2860
                         Td8 = VADD(TcZ, TcY);
 
2861
                         TcI = VADD(Tax, Tb6);
 
2862
                         Td0 = VSUB(TcY, TcZ);
 
2863
                    }
 
2864
                    {
 
2865
                         V Tcx, TcN, Tco, TcO, Tcm, Tcn;
 
2866
                         Tcx = VSUB(Tcr, Tcw);
 
2867
                         TcN = VADD(Tah, Taw);
 
2868
                         Tcm = VFNMS(LDK(KP555570233), TaI, VMUL(LDK(KP831469612), TaN));
 
2869
                         Tcn = VFMA(LDK(KP555570233), TaZ, VMUL(LDK(KP831469612), Tb4));
 
2870
                         Tco = VSUB(Tcm, Tcn);
 
2871
                         TcO = VADD(Tcm, Tcn);
 
2872
                         Tcy = VSUB(Tco, Tcx);
 
2873
                         Tda = VADD(TcN, TcO);
 
2874
                         TcG = VADD(Tcx, Tco);
 
2875
                         TcP = VSUB(TcN, TcO);
 
2876
                    }
 
2877
                    {
 
2878
                         V TbI, Tcz, Tcj, TcA;
 
2879
                         {
 
2880
                              V Tby, TbH, Tc9, Tci;
 
2881
                              Tby = VSUB(Tbm, Tbx);
 
2882
                              TbH = VSUB(TbD, TbG);
 
2883
                              TbI = VFMA(LDK(KP881921264), Tby, VMUL(LDK(KP471396736), TbH));
 
2884
                              Tcz = VFNMS(LDK(KP471396736), Tby, VMUL(LDK(KP881921264), TbH));
 
2885
                              Tc9 = VSUB(TbX, Tc8);
 
2886
                              Tci = VSUB(Tce, Tch);
 
2887
                              Tcj = VFNMS(LDK(KP471396736), Tci, VMUL(LDK(KP881921264), Tc9));
 
2888
                              TcA = VFMA(LDK(KP471396736), Tc9, VMUL(LDK(KP881921264), Tci));
 
2889
                         }
 
2890
                         Tck = VSUB(TbI, Tcj);
 
2891
                         TcJ = VADD(Tcz, TcA);
 
2892
                         TcB = VSUB(Tcz, TcA);
 
2893
                         TcF = VADD(TbI, Tcj);
 
2894
                    }
 
2895
                    {
 
2896
                         V TcS, Td1, TcV, Td2;
 
2897
                         {
 
2898
                              V TcQ, TcR, TcT, TcU;
 
2899
                              TcQ = VADD(Tbx, Tbm);
 
2900
                              TcR = VADD(TbD, TbG);
 
2901
                              TcS = VFMA(LDK(KP956940335), TcQ, VMUL(LDK(KP290284677), TcR));
 
2902
                              Td1 = VFNMS(LDK(KP290284677), TcQ, VMUL(LDK(KP956940335), TcR));
 
2903
                              TcT = VADD(Tc8, TbX);
 
2904
                              TcU = VADD(Tce, Tch);
 
2905
                              TcV = VFNMS(LDK(KP290284677), TcU, VMUL(LDK(KP956940335), TcT));
 
2906
                              Td2 = VFMA(LDK(KP290284677), TcT, VMUL(LDK(KP956940335), TcU));
 
2907
                         }
 
2908
                         TcW = VSUB(TcS, TcV);
 
2909
                         Tdb = VADD(Td1, Td2);
 
2910
                         Td3 = VSUB(Td1, Td2);
 
2911
                         Td7 = VADD(TcS, TcV);
 
2912
                    }
 
2913
                    {
 
2914
                         V Tcl, TcC, Td9, Tdc;
 
2915
                         Tcl = VADD(Tb7, Tck);
 
2916
                         TcC = VBYI(VADD(Tcy, TcB));
 
2917
                         ST(&(xo[WS(os, 106)]), VSUB(Tcl, TcC), ovs, &(xo[0]));
 
2918
                         ST(&(xo[WS(os, 22)]), VADD(Tcl, TcC), ovs, &(xo[0]));
 
2919
                         Td9 = VBYI(VSUB(Td7, Td8));
 
2920
                         Tdc = VSUB(Tda, Tdb);
 
2921
                         ST(&(xo[WS(os, 58)]), VADD(Td9, Tdc), ovs, &(xo[0]));
 
2922
                         ST(&(xo[WS(os, 70)]), VSUB(Tdc, Td9), ovs, &(xo[0]));
 
2923
                    }
 
2924
                    {
 
2925
                         V Tdd, Tde, TcD, TcE;
 
2926
                         Tdd = VBYI(VADD(Td8, Td7));
 
2927
                         Tde = VADD(Tda, Tdb);
 
2928
                         ST(&(xo[WS(os, 6)]), VADD(Tdd, Tde), ovs, &(xo[0]));
 
2929
                         ST(&(xo[WS(os, 122)]), VSUB(Tde, Tdd), ovs, &(xo[0]));
 
2930
                         TcD = VSUB(Tb7, Tck);
 
2931
                         TcE = VBYI(VSUB(TcB, Tcy));
 
2932
                         ST(&(xo[WS(os, 86)]), VSUB(TcD, TcE), ovs, &(xo[0]));
 
2933
                         ST(&(xo[WS(os, 42)]), VADD(TcD, TcE), ovs, &(xo[0]));
 
2934
                    }
 
2935
                    {
 
2936
                         V TcH, TcK, TcX, Td4;
 
2937
                         TcH = VBYI(VSUB(TcF, TcG));
 
2938
                         TcK = VSUB(TcI, TcJ);
 
2939
                         ST(&(xo[WS(os, 54)]), VADD(TcH, TcK), ovs, &(xo[0]));
 
2940
                         ST(&(xo[WS(os, 74)]), VSUB(TcK, TcH), ovs, &(xo[0]));
 
2941
                         TcX = VADD(TcP, TcW);
 
2942
                         Td4 = VBYI(VADD(Td0, Td3));
 
2943
                         ST(&(xo[WS(os, 102)]), VSUB(TcX, Td4), ovs, &(xo[0]));
 
2944
                         ST(&(xo[WS(os, 26)]), VADD(TcX, Td4), ovs, &(xo[0]));
 
2945
                    }
 
2946
                    {
 
2947
                         V Td5, Td6, TcL, TcM;
 
2948
                         Td5 = VSUB(TcP, TcW);
 
2949
                         Td6 = VBYI(VSUB(Td3, Td0));
 
2950
                         ST(&(xo[WS(os, 90)]), VSUB(Td5, Td6), ovs, &(xo[0]));
 
2951
                         ST(&(xo[WS(os, 38)]), VADD(Td5, Td6), ovs, &(xo[0]));
 
2952
                         TcL = VBYI(VADD(TcG, TcF));
 
2953
                         TcM = VADD(TcI, TcJ);
 
2954
                         ST(&(xo[WS(os, 10)]), VADD(TcL, TcM), ovs, &(xo[0]));
 
2955
                         ST(&(xo[WS(os, 118)]), VSUB(TcM, TcL), ovs, &(xo[0]));
 
2956
                    }
 
2957
               }
 
2958
               {
 
2959
                    V TdE, Tel, TdW, Tee, TdM, Teo, TdT, Tea, Tdt, TdX, TdP, TdU, Te7, Tep, Teh;
 
2960
                    V Tem;
 
2961
                    {
 
2962
                         V TdD, Tec, TdA, Ted, Tdw, Tdz;
 
2963
                         TdD = VADD(TdB, TdC);
 
2964
                         Tec = VSUB(TdG, TdH);
 
2965
                         Tdw = VFMA(LDK(KP980785280), Tdu, VMUL(LDK(KP195090322), Tdv));
 
2966
                         Tdz = VFNMS(LDK(KP195090322), Tdy, VMUL(LDK(KP980785280), Tdx));
 
2967
                         TdA = VADD(Tdw, Tdz);
 
2968
                         Ted = VSUB(Tdw, Tdz);
 
2969
                         TdE = VSUB(TdA, TdD);
 
2970
                         Tel = VADD(Tec, Ted);
 
2971
                         TdW = VADD(TdD, TdA);
 
2972
                         Tee = VSUB(Tec, Ted);
 
2973
                    }
 
2974
                    {
 
2975
                         V TdI, Te9, TdL, Te8, TdJ, TdK;
 
2976
                         TdI = VADD(TdG, TdH);
 
2977
                         Te9 = VSUB(TdC, TdB);
 
2978
                         TdJ = VFNMS(LDK(KP195090322), Tdu, VMUL(LDK(KP980785280), Tdv));
 
2979
                         TdK = VFMA(LDK(KP195090322), Tdx, VMUL(LDK(KP980785280), Tdy));
 
2980
                         TdL = VADD(TdJ, TdK);
 
2981
                         Te8 = VSUB(TdJ, TdK);
 
2982
                         TdM = VSUB(TdI, TdL);
 
2983
                         Teo = VADD(Te9, Te8);
 
2984
                         TdT = VADD(TdI, TdL);
 
2985
                         Tea = VSUB(Te8, Te9);
 
2986
                    }
 
2987
                    {
 
2988
                         V Tdl, TdN, Tds, TdO;
 
2989
                         {
 
2990
                              V Tdh, Tdk, Tdo, Tdr;
 
2991
                              Tdh = VADD(Tdf, Tdg);
 
2992
                              Tdk = VADD(Tdi, Tdj);
 
2993
                              Tdl = VFNMS(LDK(KP098017140), Tdk, VMUL(LDK(KP995184726), Tdh));
 
2994
                              TdN = VFMA(LDK(KP098017140), Tdh, VMUL(LDK(KP995184726), Tdk));
 
2995
                              Tdo = VADD(Tdm, Tdn);
 
2996
                              Tdr = VADD(Tdp, Tdq);
 
2997
                              Tds = VFMA(LDK(KP995184726), Tdo, VMUL(LDK(KP098017140), Tdr));
 
2998
                              TdO = VFNMS(LDK(KP098017140), Tdo, VMUL(LDK(KP995184726), Tdr));
 
2999
                         }
 
3000
                         Tdt = VSUB(Tdl, Tds);
 
3001
                         TdX = VADD(TdN, TdO);
 
3002
                         TdP = VSUB(TdN, TdO);
 
3003
                         TdU = VADD(Tdl, Tds);
 
3004
                    }
 
3005
                    {
 
3006
                         V Te3, Tef, Te6, Teg;
 
3007
                         {
 
3008
                              V Te1, Te2, Te4, Te5;
 
3009
                              Te1 = VSUB(Tdf, Tdg);
 
3010
                              Te2 = VSUB(Tdj, Tdi);
 
3011
                              Te3 = VFNMS(LDK(KP634393284), Te2, VMUL(LDK(KP773010453), Te1));
 
3012
                              Tef = VFMA(LDK(KP634393284), Te1, VMUL(LDK(KP773010453), Te2));
 
3013
                              Te4 = VSUB(Tdm, Tdn);
 
3014
                              Te5 = VSUB(Tdq, Tdp);
 
3015
                              Te6 = VFMA(LDK(KP773010453), Te4, VMUL(LDK(KP634393284), Te5));
 
3016
                              Teg = VFNMS(LDK(KP634393284), Te4, VMUL(LDK(KP773010453), Te5));
 
3017
                         }
 
3018
                         Te7 = VSUB(Te3, Te6);
 
3019
                         Tep = VADD(Tef, Teg);
 
3020
                         Teh = VSUB(Tef, Teg);
 
3021
                         Tem = VADD(Te3, Te6);
 
3022
                    }
 
3023
                    {
 
3024
                         V TdF, TdQ, Ten, Teq;
 
3025
                         TdF = VBYI(VSUB(Tdt, TdE));
 
3026
                         TdQ = VSUB(TdM, TdP);
 
3027
                         ST(&(xo[WS(os, 34)]), VADD(TdF, TdQ), ovs, &(xo[0]));
 
3028
                         ST(&(xo[WS(os, 94)]), VSUB(TdQ, TdF), ovs, &(xo[0]));
 
3029
                         Ten = VADD(Tel, Tem);
 
3030
                         Teq = VBYI(VADD(Teo, Tep));
 
3031
                         ST(&(xo[WS(os, 114)]), VSUB(Ten, Teq), ovs, &(xo[0]));
 
3032
                         ST(&(xo[WS(os, 14)]), VADD(Ten, Teq), ovs, &(xo[0]));
 
3033
                    }
 
3034
                    {
 
3035
                         V Ter, Tes, TdR, TdS;
 
3036
                         Ter = VSUB(Tel, Tem);
 
3037
                         Tes = VBYI(VSUB(Tep, Teo));
 
3038
                         ST(&(xo[WS(os, 78)]), VSUB(Ter, Tes), ovs, &(xo[0]));
 
3039
                         ST(&(xo[WS(os, 50)]), VADD(Ter, Tes), ovs, &(xo[0]));
 
3040
                         TdR = VBYI(VADD(TdE, Tdt));
 
3041
                         TdS = VADD(TdM, TdP);
 
3042
                         ST(&(xo[WS(os, 30)]), VADD(TdR, TdS), ovs, &(xo[0]));
 
3043
                         ST(&(xo[WS(os, 98)]), VSUB(TdS, TdR), ovs, &(xo[0]));
 
3044
                    }
 
3045
                    {
 
3046
                         V TdV, TdY, Teb, Tei;
 
3047
                         TdV = VADD(TdT, TdU);
 
3048
                         TdY = VBYI(VADD(TdW, TdX));
 
3049
                         ST(&(xo[WS(os, 126)]), VSUB(TdV, TdY), ovs, &(xo[0]));
 
3050
                         ST(&(xo[WS(os, 2)]), VADD(TdV, TdY), ovs, &(xo[0]));
 
3051
                         Teb = VBYI(VSUB(Te7, Tea));
 
3052
                         Tei = VSUB(Tee, Teh);
 
3053
                         ST(&(xo[WS(os, 46)]), VADD(Teb, Tei), ovs, &(xo[0]));
 
3054
                         ST(&(xo[WS(os, 82)]), VSUB(Tei, Teb), ovs, &(xo[0]));
 
3055
                    }
 
3056
                    {
 
3057
                         V Tej, Tek, TdZ, Te0;
 
3058
                         Tej = VBYI(VADD(Tea, Te7));
 
3059
                         Tek = VADD(Tee, Teh);
 
3060
                         ST(&(xo[WS(os, 18)]), VADD(Tej, Tek), ovs, &(xo[0]));
 
3061
                         ST(&(xo[WS(os, 110)]), VSUB(Tek, Tej), ovs, &(xo[0]));
 
3062
                         TdZ = VSUB(TdT, TdU);
 
3063
                         Te0 = VBYI(VSUB(TdX, TdW));
 
3064
                         ST(&(xo[WS(os, 66)]), VSUB(TdZ, Te0), ovs, &(xo[0]));
 
3065
                         ST(&(xo[WS(os, 62)]), VADD(TdZ, Te0), ovs, &(xo[0]));
 
3066
                    }
 
3067
               }
 
3068
               {
 
3069
                    V T7z, T8n, T8f, T8k, T8x, T8P, T8H, T8M, T80, T8L, T8O, T8c, T8j, T8A, T8E;
 
3070
                    V T8m;
 
3071
                    {
 
3072
                         V T7f, T8d, T7y, T8e;
 
3073
                         {
 
3074
                              V T77, T7e, T7q, T7x;
 
3075
                              T77 = VADD(T6Z, T76);
 
3076
                              T7e = VADD(T7a, T7d);
 
3077
                              T7f = VFNMS(LDK(KP336889853), T7e, VMUL(LDK(KP941544065), T77));
 
3078
                              T8d = VFMA(LDK(KP336889853), T77, VMUL(LDK(KP941544065), T7e));
 
3079
                              T7q = VADD(T7i, T7p);
 
3080
                              T7x = VADD(T7t, T7w);
 
3081
                              T7y = VFMA(LDK(KP941544065), T7q, VMUL(LDK(KP336889853), T7x));
 
3082
                              T8e = VFNMS(LDK(KP336889853), T7q, VMUL(LDK(KP941544065), T7x));
 
3083
                         }
 
3084
                         T7z = VSUB(T7f, T7y);
 
3085
                         T8n = VADD(T8d, T8e);
 
3086
                         T8f = VSUB(T8d, T8e);
 
3087
                         T8k = VADD(T7f, T7y);
 
3088
                    }
 
3089
                    {
 
3090
                         V T8t, T8F, T8w, T8G;
 
3091
                         {
 
3092
                              V T8r, T8s, T8u, T8v;
 
3093
                              T8r = VSUB(T6Z, T76);
 
3094
                              T8s = VSUB(T7d, T7a);
 
3095
                              T8t = VFNMS(LDK(KP427555093), T8s, VMUL(LDK(KP903989293), T8r));
 
3096
                              T8F = VFMA(LDK(KP427555093), T8r, VMUL(LDK(KP903989293), T8s));
 
3097
                              T8u = VSUB(T7i, T7p);
 
3098
                              T8v = VSUB(T7w, T7t);
 
3099
                              T8w = VFMA(LDK(KP903989293), T8u, VMUL(LDK(KP427555093), T8v));
 
3100
                              T8G = VFNMS(LDK(KP427555093), T8u, VMUL(LDK(KP903989293), T8v));
 
3101
                         }
 
3102
                         T8x = VSUB(T8t, T8w);
 
3103
                         T8P = VADD(T8F, T8G);
 
3104
                         T8H = VSUB(T8F, T8G);
 
3105
                         T8M = VADD(T8t, T8w);
 
3106
                    }
 
3107
                    {
 
3108
                         V T7Z, T8z, T88, T8C, T7O, T8D, T8b, T8y, T7Y, T87;
 
3109
                         T7Y = VSUB(T7U, T7X);
 
3110
                         T7Z = VADD(T7R, T7Y);
 
3111
                         T8z = VSUB(T7Y, T7R);
 
3112
                         T87 = VSUB(T85, T86);
 
3113
                         T88 = VADD(T84, T87);
 
3114
                         T8C = VSUB(T84, T87);
 
3115
                         {
 
3116
                              V T7G, T7N, T89, T8a;
 
3117
                              T7G = VFMA(LDK(KP634393284), T7C, VMUL(LDK(KP773010453), T7F));
 
3118
                              T7N = VFNMS(LDK(KP634393284), T7M, VMUL(LDK(KP773010453), T7J));
 
3119
                              T7O = VADD(T7G, T7N);
 
3120
                              T8D = VSUB(T7G, T7N);
 
3121
                              T89 = VFNMS(LDK(KP634393284), T7F, VMUL(LDK(KP773010453), T7C));
 
3122
                              T8a = VFMA(LDK(KP773010453), T7M, VMUL(LDK(KP634393284), T7J));
 
3123
                              T8b = VADD(T89, T8a);
 
3124
                              T8y = VSUB(T89, T8a);
 
3125
                         }
 
3126
                         T80 = VSUB(T7O, T7Z);
 
3127
                         T8L = VADD(T8C, T8D);
 
3128
                         T8O = VADD(T8z, T8y);
 
3129
                         T8c = VSUB(T88, T8b);
 
3130
                         T8j = VADD(T88, T8b);
 
3131
                         T8A = VSUB(T8y, T8z);
 
3132
                         T8E = VSUB(T8C, T8D);
 
3133
                         T8m = VADD(T7Z, T7O);
 
3134
                    }
 
3135
                    {
 
3136
                         V T81, T8g, T8N, T8Q;
 
3137
                         T81 = VBYI(VSUB(T7z, T80));
 
3138
                         T8g = VSUB(T8c, T8f);
 
3139
                         ST(&(xo[WS(os, 39)]), VADD(T81, T8g), ovs, &(xo[WS(os, 1)]));
 
3140
                         ST(&(xo[WS(os, 89)]), VSUB(T8g, T81), ovs, &(xo[WS(os, 1)]));
 
3141
                         T8N = VADD(T8L, T8M);
 
3142
                         T8Q = VBYI(VADD(T8O, T8P));
 
3143
                         ST(&(xo[WS(os, 119)]), VSUB(T8N, T8Q), ovs, &(xo[WS(os, 1)]));
 
3144
                         ST(&(xo[WS(os, 9)]), VADD(T8N, T8Q), ovs, &(xo[WS(os, 1)]));
 
3145
                    }
 
3146
                    {
 
3147
                         V T8R, T8S, T8h, T8i;
 
3148
                         T8R = VSUB(T8L, T8M);
 
3149
                         T8S = VBYI(VSUB(T8P, T8O));
 
3150
                         ST(&(xo[WS(os, 73)]), VSUB(T8R, T8S), ovs, &(xo[WS(os, 1)]));
 
3151
                         ST(&(xo[WS(os, 55)]), VADD(T8R, T8S), ovs, &(xo[WS(os, 1)]));
 
3152
                         T8h = VBYI(VADD(T80, T7z));
 
3153
                         T8i = VADD(T8c, T8f);
 
3154
                         ST(&(xo[WS(os, 25)]), VADD(T8h, T8i), ovs, &(xo[WS(os, 1)]));
 
3155
                         ST(&(xo[WS(os, 103)]), VSUB(T8i, T8h), ovs, &(xo[WS(os, 1)]));
 
3156
                    }
 
3157
                    {
 
3158
                         V T8l, T8o, T8B, T8I;
 
3159
                         T8l = VADD(T8j, T8k);
 
3160
                         T8o = VBYI(VADD(T8m, T8n));
 
3161
                         ST(&(xo[WS(os, 121)]), VSUB(T8l, T8o), ovs, &(xo[WS(os, 1)]));
 
3162
                         ST(&(xo[WS(os, 7)]), VADD(T8l, T8o), ovs, &(xo[WS(os, 1)]));
 
3163
                         T8B = VBYI(VSUB(T8x, T8A));
 
3164
                         T8I = VSUB(T8E, T8H);
 
3165
                         ST(&(xo[WS(os, 41)]), VADD(T8B, T8I), ovs, &(xo[WS(os, 1)]));
 
3166
                         ST(&(xo[WS(os, 87)]), VSUB(T8I, T8B), ovs, &(xo[WS(os, 1)]));
 
3167
                    }
 
3168
                    {
 
3169
                         V T8J, T8K, T8p, T8q;
 
3170
                         T8J = VBYI(VADD(T8A, T8x));
 
3171
                         T8K = VADD(T8E, T8H);
 
3172
                         ST(&(xo[WS(os, 23)]), VADD(T8J, T8K), ovs, &(xo[WS(os, 1)]));
 
3173
                         ST(&(xo[WS(os, 105)]), VSUB(T8K, T8J), ovs, &(xo[WS(os, 1)]));
 
3174
                         T8p = VSUB(T8j, T8k);
 
3175
                         T8q = VBYI(VSUB(T8n, T8m));
 
3176
                         ST(&(xo[WS(os, 71)]), VSUB(T8p, T8q), ovs, &(xo[WS(os, 1)]));
 
3177
                         ST(&(xo[WS(os, 57)]), VADD(T8p, T8q), ovs, &(xo[WS(os, 1)]));
 
3178
                    }
 
3179
               }
 
3180
               {
 
3181
                    V T2v, T5d, T55, T5a, T5n, T5F, T5x, T5C, T4K, T5B, T5E, T52, T59, T5q, T5u;
 
3182
                    V T5c;
 
3183
                    {
 
3184
                         V T1f, T53, T2u, T54;
 
3185
                         {
 
3186
                              V T11, T1e, T2g, T2t;
 
3187
                              T11 = VADD(Tr, T10);
 
3188
                              T1e = VADD(T1a, T1d);
 
3189
                              T1f = VFNMS(LDK(KP242980179), T1e, VMUL(LDK(KP970031253), T11));
 
3190
                              T53 = VFMA(LDK(KP242980179), T11, VMUL(LDK(KP970031253), T1e));
 
3191
                              T2g = VADD(T1G, T2f);
 
3192
                              T2t = VADD(T2p, T2s);
 
3193
                              T2u = VFMA(LDK(KP970031253), T2g, VMUL(LDK(KP242980179), T2t));
 
3194
                              T54 = VFNMS(LDK(KP242980179), T2g, VMUL(LDK(KP970031253), T2t));
 
3195
                         }
 
3196
                         T2v = VSUB(T1f, T2u);
 
3197
                         T5d = VADD(T53, T54);
 
3198
                         T55 = VSUB(T53, T54);
 
3199
                         T5a = VADD(T1f, T2u);
 
3200
                    }
 
3201
                    {
 
3202
                         V T5j, T5v, T5m, T5w;
 
3203
                         {
 
3204
                              V T5h, T5i, T5k, T5l;
 
3205
                              T5h = VSUB(Tr, T10);
 
3206
                              T5i = VSUB(T1d, T1a);
 
3207
                              T5j = VFNMS(LDK(KP514102744), T5i, VMUL(LDK(KP857728610), T5h));
 
3208
                              T5v = VFMA(LDK(KP514102744), T5h, VMUL(LDK(KP857728610), T5i));
 
3209
                              T5k = VSUB(T1G, T2f);
 
3210
                              T5l = VSUB(T2s, T2p);
 
3211
                              T5m = VFMA(LDK(KP857728610), T5k, VMUL(LDK(KP514102744), T5l));
 
3212
                              T5w = VFNMS(LDK(KP514102744), T5k, VMUL(LDK(KP857728610), T5l));
 
3213
                         }
 
3214
                         T5n = VSUB(T5j, T5m);
 
3215
                         T5F = VADD(T5v, T5w);
 
3216
                         T5x = VSUB(T5v, T5w);
 
3217
                         T5C = VADD(T5j, T5m);
 
3218
                    }
 
3219
                    {
 
3220
                         V T4J, T5p, T4Y, T5s, T3I, T5t, T51, T5o, T4I, T4X;
 
3221
                         T4I = VSUB(T4q, T4H);
 
3222
                         T4J = VADD(T49, T4I);
 
3223
                         T5p = VSUB(T4I, T49);
 
3224
                         T4X = VSUB(T4V, T4W);
 
3225
                         T4Y = VADD(T4U, T4X);
 
3226
                         T5s = VSUB(T4U, T4X);
 
3227
                         {
 
3228
                              V T36, T3H, T4Z, T50;
 
3229
                              T36 = VFMA(LDK(KP881921264), T2W, VMUL(LDK(KP471396736), T35));
 
3230
                              T3H = VFNMS(LDK(KP471396736), T3G, VMUL(LDK(KP881921264), T3x));
 
3231
                              T3I = VADD(T36, T3H);
 
3232
                              T5t = VSUB(T36, T3H);
 
3233
                              T4Z = VFNMS(LDK(KP471396736), T2W, VMUL(LDK(KP881921264), T35));
 
3234
                              T50 = VFMA(LDK(KP471396736), T3x, VMUL(LDK(KP881921264), T3G));
 
3235
                              T51 = VADD(T4Z, T50);
 
3236
                              T5o = VSUB(T4Z, T50);
 
3237
                         }
 
3238
                         T4K = VSUB(T3I, T4J);
 
3239
                         T5B = VADD(T5s, T5t);
 
3240
                         T5E = VADD(T5p, T5o);
 
3241
                         T52 = VSUB(T4Y, T51);
 
3242
                         T59 = VADD(T4Y, T51);
 
3243
                         T5q = VSUB(T5o, T5p);
 
3244
                         T5u = VSUB(T5s, T5t);
 
3245
                         T5c = VADD(T4J, T3I);
 
3246
                    }
 
3247
                    {
 
3248
                         V T4L, T56, T5D, T5G;
 
3249
                         T4L = VBYI(VSUB(T2v, T4K));
 
3250
                         T56 = VSUB(T52, T55);
 
3251
                         ST(&(xo[WS(os, 37)]), VADD(T4L, T56), ovs, &(xo[WS(os, 1)]));
 
3252
                         ST(&(xo[WS(os, 91)]), VSUB(T56, T4L), ovs, &(xo[WS(os, 1)]));
 
3253
                         T5D = VADD(T5B, T5C);
 
3254
                         T5G = VBYI(VADD(T5E, T5F));
 
3255
                         ST(&(xo[WS(os, 117)]), VSUB(T5D, T5G), ovs, &(xo[WS(os, 1)]));
 
3256
                         ST(&(xo[WS(os, 11)]), VADD(T5D, T5G), ovs, &(xo[WS(os, 1)]));
 
3257
                    }
 
3258
                    {
 
3259
                         V T5H, T5I, T57, T58;
 
3260
                         T5H = VSUB(T5B, T5C);
 
3261
                         T5I = VBYI(VSUB(T5F, T5E));
 
3262
                         ST(&(xo[WS(os, 75)]), VSUB(T5H, T5I), ovs, &(xo[WS(os, 1)]));
 
3263
                         ST(&(xo[WS(os, 53)]), VADD(T5H, T5I), ovs, &(xo[WS(os, 1)]));
 
3264
                         T57 = VBYI(VADD(T4K, T2v));
 
3265
                         T58 = VADD(T52, T55);
 
3266
                         ST(&(xo[WS(os, 27)]), VADD(T57, T58), ovs, &(xo[WS(os, 1)]));
 
3267
                         ST(&(xo[WS(os, 101)]), VSUB(T58, T57), ovs, &(xo[WS(os, 1)]));
 
3268
                    }
 
3269
                    {
 
3270
                         V T5b, T5e, T5r, T5y;
 
3271
                         T5b = VADD(T59, T5a);
 
3272
                         T5e = VBYI(VADD(T5c, T5d));
 
3273
                         ST(&(xo[WS(os, 123)]), VSUB(T5b, T5e), ovs, &(xo[WS(os, 1)]));
 
3274
                         ST(&(xo[WS(os, 5)]), VADD(T5b, T5e), ovs, &(xo[WS(os, 1)]));
 
3275
                         T5r = VBYI(VSUB(T5n, T5q));
 
3276
                         T5y = VSUB(T5u, T5x);
 
3277
                         ST(&(xo[WS(os, 43)]), VADD(T5r, T5y), ovs, &(xo[WS(os, 1)]));
 
3278
                         ST(&(xo[WS(os, 85)]), VSUB(T5y, T5r), ovs, &(xo[WS(os, 1)]));
 
3279
                    }
 
3280
                    {
 
3281
                         V T5z, T5A, T5f, T5g;
 
3282
                         T5z = VBYI(VADD(T5q, T5n));
 
3283
                         T5A = VADD(T5u, T5x);
 
3284
                         ST(&(xo[WS(os, 21)]), VADD(T5z, T5A), ovs, &(xo[WS(os, 1)]));
 
3285
                         ST(&(xo[WS(os, 107)]), VSUB(T5A, T5z), ovs, &(xo[WS(os, 1)]));
 
3286
                         T5f = VSUB(T59, T5a);
 
3287
                         T5g = VBYI(VSUB(T5d, T5c));
 
3288
                         ST(&(xo[WS(os, 69)]), VSUB(T5f, T5g), ovs, &(xo[WS(os, 1)]));
 
3289
                         ST(&(xo[WS(os, 59)]), VADD(T5f, T5g), ovs, &(xo[WS(os, 1)]));
 
3290
                    }
 
3291
               }
 
3292
               {
 
3293
                    V T9i, T9B, T9t, T9x, T9O, Ta3, T9V, T9Z, T93, Ta0, Ta2, T9q, T9y, T9H, T9S;
 
3294
                    V T9A;
 
3295
                    {
 
3296
                         V T9a, T9r, T9h, T9s;
 
3297
                         {
 
3298
                              V T96, T99, T9d, T9g;
 
3299
                              T96 = VSUB(T94, T95);
 
3300
                              T99 = VSUB(T97, T98);
 
3301
                              T9a = VFMA(LDK(KP740951125), T96, VMUL(LDK(KP671558954), T99));
 
3302
                              T9r = VFNMS(LDK(KP671558954), T96, VMUL(LDK(KP740951125), T99));
 
3303
                              T9d = VSUB(T9b, T9c);
 
3304
                              T9g = VSUB(T9e, T9f);
 
3305
                              T9h = VFNMS(LDK(KP671558954), T9g, VMUL(LDK(KP740951125), T9d));
 
3306
                              T9s = VFMA(LDK(KP671558954), T9d, VMUL(LDK(KP740951125), T9g));
 
3307
                         }
 
3308
                         T9i = VSUB(T9a, T9h);
 
3309
                         T9B = VADD(T9r, T9s);
 
3310
                         T9t = VSUB(T9r, T9s);
 
3311
                         T9x = VADD(T9a, T9h);
 
3312
                    }
 
3313
                    {
 
3314
                         V T9K, T9T, T9N, T9U;
 
3315
                         {
 
3316
                              V T9I, T9J, T9L, T9M;
 
3317
                              T9I = VADD(T95, T94);
 
3318
                              T9J = VADD(T97, T98);
 
3319
                              T9K = VFMA(LDK(KP998795456), T9I, VMUL(LDK(KP049067674), T9J));
 
3320
                              T9T = VFNMS(LDK(KP049067674), T9I, VMUL(LDK(KP998795456), T9J));
 
3321
                              T9L = VADD(T9c, T9b);
 
3322
                              T9M = VADD(T9e, T9f);
 
3323
                              T9N = VFNMS(LDK(KP049067674), T9M, VMUL(LDK(KP998795456), T9L));
 
3324
                              T9U = VFMA(LDK(KP049067674), T9L, VMUL(LDK(KP998795456), T9M));
 
3325
                         }
 
3326
                         T9O = VSUB(T9K, T9N);
 
3327
                         Ta3 = VADD(T9T, T9U);
 
3328
                         T9V = VSUB(T9T, T9U);
 
3329
                         T9Z = VADD(T9K, T9N);
 
3330
                    }
 
3331
                    {
 
3332
                         V T8V, T9F, T9p, T9R, T92, T9Q, T9m, T9G, T8U, T9n;
 
3333
                         T8U = VADD(T7U, T7X);
 
3334
                         T8V = VSUB(T8T, T8U);
 
3335
                         T9F = VADD(T8T, T8U);
 
3336
                         T9n = VADD(T85, T86);
 
3337
                         T9p = VSUB(T9n, T9o);
 
3338
                         T9R = VADD(T9o, T9n);
 
3339
                         {
 
3340
                              V T8Y, T91, T9k, T9l;
 
3341
                              T8Y = VFMA(LDK(KP098017140), T8W, VMUL(LDK(KP995184726), T8X));
 
3342
                              T91 = VFNMS(LDK(KP098017140), T90, VMUL(LDK(KP995184726), T8Z));
 
3343
                              T92 = VSUB(T8Y, T91);
 
3344
                              T9Q = VADD(T8Y, T91);
 
3345
                              T9k = VFNMS(LDK(KP098017140), T8X, VMUL(LDK(KP995184726), T8W));
 
3346
                              T9l = VFMA(LDK(KP995184726), T90, VMUL(LDK(KP098017140), T8Z));
 
3347
                              T9m = VSUB(T9k, T9l);
 
3348
                              T9G = VADD(T9k, T9l);
 
3349
                         }
 
3350
                         T93 = VSUB(T8V, T92);
 
3351
                         Ta0 = VADD(T9R, T9Q);
 
3352
                         Ta2 = VADD(T9F, T9G);
 
3353
                         T9q = VSUB(T9m, T9p);
 
3354
                         T9y = VADD(T9p, T9m);
 
3355
                         T9H = VSUB(T9F, T9G);
 
3356
                         T9S = VSUB(T9Q, T9R);
 
3357
                         T9A = VADD(T8V, T92);
 
3358
                    }
 
3359
                    {
 
3360
                         V T9j, T9u, Ta1, Ta4;
 
3361
                         T9j = VADD(T93, T9i);
 
3362
                         T9u = VBYI(VADD(T9q, T9t));
 
3363
                         ST(&(xo[WS(os, 111)]), VSUB(T9j, T9u), ovs, &(xo[WS(os, 1)]));
 
3364
                         ST(&(xo[WS(os, 17)]), VADD(T9j, T9u), ovs, &(xo[WS(os, 1)]));
 
3365
                         Ta1 = VBYI(VSUB(T9Z, Ta0));
 
3366
                         Ta4 = VSUB(Ta2, Ta3);
 
3367
                         ST(&(xo[WS(os, 63)]), VADD(Ta1, Ta4), ovs, &(xo[WS(os, 1)]));
 
3368
                         ST(&(xo[WS(os, 65)]), VSUB(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
 
3369
                    }
 
3370
                    {
 
3371
                         V Ta5, Ta6, T9v, T9w;
 
3372
                         Ta5 = VBYI(VADD(Ta0, T9Z));
 
3373
                         Ta6 = VADD(Ta2, Ta3);
 
3374
                         ST(&(xo[WS(os, 1)]), VADD(Ta5, Ta6), ovs, &(xo[WS(os, 1)]));
 
3375
                         ST(&(xo[WS(os, 127)]), VSUB(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
 
3376
                         T9v = VSUB(T93, T9i);
 
3377
                         T9w = VBYI(VSUB(T9t, T9q));
 
3378
                         ST(&(xo[WS(os, 81)]), VSUB(T9v, T9w), ovs, &(xo[WS(os, 1)]));
 
3379
                         ST(&(xo[WS(os, 47)]), VADD(T9v, T9w), ovs, &(xo[WS(os, 1)]));
 
3380
                    }
 
3381
                    {
 
3382
                         V T9z, T9C, T9P, T9W;
 
3383
                         T9z = VBYI(VSUB(T9x, T9y));
 
3384
                         T9C = VSUB(T9A, T9B);
 
3385
                         ST(&(xo[WS(os, 49)]), VADD(T9z, T9C), ovs, &(xo[WS(os, 1)]));
 
3386
                         ST(&(xo[WS(os, 79)]), VSUB(T9C, T9z), ovs, &(xo[WS(os, 1)]));
 
3387
                         T9P = VADD(T9H, T9O);
 
3388
                         T9W = VBYI(VADD(T9S, T9V));
 
3389
                         ST(&(xo[WS(os, 97)]), VSUB(T9P, T9W), ovs, &(xo[WS(os, 1)]));
 
3390
                         ST(&(xo[WS(os, 31)]), VADD(T9P, T9W), ovs, &(xo[WS(os, 1)]));
 
3391
                    }
 
3392
                    {
 
3393
                         V T9X, T9Y, T9D, T9E;
 
3394
                         T9X = VSUB(T9H, T9O);
 
3395
                         T9Y = VBYI(VSUB(T9V, T9S));
 
3396
                         ST(&(xo[WS(os, 95)]), VSUB(T9X, T9Y), ovs, &(xo[WS(os, 1)]));
 
3397
                         ST(&(xo[WS(os, 33)]), VADD(T9X, T9Y), ovs, &(xo[WS(os, 1)]));
 
3398
                         T9D = VBYI(VADD(T9y, T9x));
 
3399
                         T9E = VADD(T9A, T9B);
 
3400
                         ST(&(xo[WS(os, 15)]), VADD(T9D, T9E), ovs, &(xo[WS(os, 1)]));
 
3401
                         ST(&(xo[WS(os, 113)]), VSUB(T9E, T9D), ovs, &(xo[WS(os, 1)]));
 
3402
                    }
 
3403
               }
 
3404
               {
 
3405
                    V T68, T6r, T6j, T6n, T6E, T6T, T6L, T6P, T5T, T6Q, T6S, T6g, T6o, T6x, T6I;
 
3406
                    V T6q;
 
3407
                    {
 
3408
                         V T60, T6h, T67, T6i;
 
3409
                         {
 
3410
                              V T5W, T5Z, T63, T66;
 
3411
                              T5W = VSUB(T5U, T5V);
 
3412
                              T5Z = VSUB(T5X, T5Y);
 
3413
                              T60 = VFMA(LDK(KP803207531), T5W, VMUL(LDK(KP595699304), T5Z));
 
3414
                              T6h = VFNMS(LDK(KP595699304), T5W, VMUL(LDK(KP803207531), T5Z));
 
3415
                              T63 = VSUB(T61, T62);
 
3416
                              T66 = VSUB(T64, T65);
 
3417
                              T67 = VFNMS(LDK(KP595699304), T66, VMUL(LDK(KP803207531), T63));
 
3418
                              T6i = VFMA(LDK(KP595699304), T63, VMUL(LDK(KP803207531), T66));
 
3419
                         }
 
3420
                         T68 = VSUB(T60, T67);
 
3421
                         T6r = VADD(T6h, T6i);
 
3422
                         T6j = VSUB(T6h, T6i);
 
3423
                         T6n = VADD(T60, T67);
 
3424
                    }
 
3425
                    {
 
3426
                         V T6A, T6J, T6D, T6K;
 
3427
                         {
 
3428
                              V T6y, T6z, T6B, T6C;
 
3429
                              T6y = VADD(T5V, T5U);
 
3430
                              T6z = VADD(T5X, T5Y);
 
3431
                              T6A = VFMA(LDK(KP989176509), T6y, VMUL(LDK(KP146730474), T6z));
 
3432
                              T6J = VFNMS(LDK(KP146730474), T6y, VMUL(LDK(KP989176509), T6z));
 
3433
                              T6B = VADD(T62, T61);
 
3434
                              T6C = VADD(T64, T65);
 
3435
                              T6D = VFNMS(LDK(KP146730474), T6C, VMUL(LDK(KP989176509), T6B));
 
3436
                              T6K = VFMA(LDK(KP146730474), T6B, VMUL(LDK(KP989176509), T6C));
 
3437
                         }
 
3438
                         T6E = VSUB(T6A, T6D);
 
3439
                         T6T = VADD(T6J, T6K);
 
3440
                         T6L = VSUB(T6J, T6K);
 
3441
                         T6P = VADD(T6A, T6D);
 
3442
                    }
 
3443
                    {
 
3444
                         V T5L, T6v, T6f, T6H, T5S, T6G, T6c, T6w, T5K, T6d;
 
3445
                         T5K = VADD(T4q, T4H);
 
3446
                         T5L = VSUB(T5J, T5K);
 
3447
                         T6v = VADD(T5J, T5K);
 
3448
                         T6d = VADD(T4V, T4W);
 
3449
                         T6f = VSUB(T6d, T6e);
 
3450
                         T6H = VADD(T6e, T6d);
 
3451
                         {
 
3452
                              V T5O, T5R, T6a, T6b;
 
3453
                              T5O = VFMA(LDK(KP956940335), T5M, VMUL(LDK(KP290284677), T5N));
 
3454
                              T5R = VFNMS(LDK(KP290284677), T5Q, VMUL(LDK(KP956940335), T5P));
 
3455
                              T5S = VSUB(T5O, T5R);
 
3456
                              T6G = VADD(T5O, T5R);
 
3457
                              T6a = VFNMS(LDK(KP290284677), T5M, VMUL(LDK(KP956940335), T5N));
 
3458
                              T6b = VFMA(LDK(KP290284677), T5P, VMUL(LDK(KP956940335), T5Q));
 
3459
                              T6c = VSUB(T6a, T6b);
 
3460
                              T6w = VADD(T6a, T6b);
 
3461
                         }
 
3462
                         T5T = VSUB(T5L, T5S);
 
3463
                         T6Q = VADD(T6H, T6G);
 
3464
                         T6S = VADD(T6v, T6w);
 
3465
                         T6g = VSUB(T6c, T6f);
 
3466
                         T6o = VADD(T6f, T6c);
 
3467
                         T6x = VSUB(T6v, T6w);
 
3468
                         T6I = VSUB(T6G, T6H);
 
3469
                         T6q = VADD(T5L, T5S);
 
3470
                    }
 
3471
                    {
 
3472
                         V T69, T6k, T6R, T6U;
 
3473
                         T69 = VADD(T5T, T68);
 
3474
                         T6k = VBYI(VADD(T6g, T6j));
 
3475
                         ST(&(xo[WS(os, 109)]), VSUB(T69, T6k), ovs, &(xo[WS(os, 1)]));
 
3476
                         ST(&(xo[WS(os, 19)]), VADD(T69, T6k), ovs, &(xo[WS(os, 1)]));
 
3477
                         T6R = VBYI(VSUB(T6P, T6Q));
 
3478
                         T6U = VSUB(T6S, T6T);
 
3479
                         ST(&(xo[WS(os, 61)]), VADD(T6R, T6U), ovs, &(xo[WS(os, 1)]));
 
3480
                         ST(&(xo[WS(os, 67)]), VSUB(T6U, T6R), ovs, &(xo[WS(os, 1)]));
 
3481
                    }
 
3482
                    {
 
3483
                         V T6V, T6W, T6l, T6m;
 
3484
                         T6V = VBYI(VADD(T6Q, T6P));
 
3485
                         T6W = VADD(T6S, T6T);
 
3486
                         ST(&(xo[WS(os, 3)]), VADD(T6V, T6W), ovs, &(xo[WS(os, 1)]));
 
3487
                         ST(&(xo[WS(os, 125)]), VSUB(T6W, T6V), ovs, &(xo[WS(os, 1)]));
 
3488
                         T6l = VSUB(T5T, T68);
 
3489
                         T6m = VBYI(VSUB(T6j, T6g));
 
3490
                         ST(&(xo[WS(os, 83)]), VSUB(T6l, T6m), ovs, &(xo[WS(os, 1)]));
 
3491
                         ST(&(xo[WS(os, 45)]), VADD(T6l, T6m), ovs, &(xo[WS(os, 1)]));
 
3492
                    }
 
3493
                    {
 
3494
                         V T6p, T6s, T6F, T6M;
 
3495
                         T6p = VBYI(VSUB(T6n, T6o));
 
3496
                         T6s = VSUB(T6q, T6r);
 
3497
                         ST(&(xo[WS(os, 51)]), VADD(T6p, T6s), ovs, &(xo[WS(os, 1)]));
 
3498
                         ST(&(xo[WS(os, 77)]), VSUB(T6s, T6p), ovs, &(xo[WS(os, 1)]));
 
3499
                         T6F = VADD(T6x, T6E);
 
3500
                         T6M = VBYI(VADD(T6I, T6L));
 
3501
                         ST(&(xo[WS(os, 99)]), VSUB(T6F, T6M), ovs, &(xo[WS(os, 1)]));
 
3502
                         ST(&(xo[WS(os, 29)]), VADD(T6F, T6M), ovs, &(xo[WS(os, 1)]));
 
3503
                    }
 
3504
                    {
 
3505
                         V T6N, T6O, T6t, T6u;
 
3506
                         T6N = VSUB(T6x, T6E);
 
3507
                         T6O = VBYI(VSUB(T6L, T6I));
 
3508
                         ST(&(xo[WS(os, 93)]), VSUB(T6N, T6O), ovs, &(xo[WS(os, 1)]));
 
3509
                         ST(&(xo[WS(os, 35)]), VADD(T6N, T6O), ovs, &(xo[WS(os, 1)]));
 
3510
                         T6t = VBYI(VADD(T6o, T6n));
 
3511
                         T6u = VADD(T6q, T6r);
 
3512
                         ST(&(xo[WS(os, 13)]), VADD(T6t, T6u), ovs, &(xo[WS(os, 1)]));
 
3513
                         ST(&(xo[WS(os, 115)]), VSUB(T6u, T6t), ovs, &(xo[WS(os, 1)]));
 
3514
                    }
 
3515
               }
 
3516
          }
 
3517
     }
 
3518
     VLEAVE();
 
3519
}
 
3520
 
 
3521
static const kdft_desc desc = { 128, XSIMD_STRING("n1bv_128"), {938, 186, 144, 0}, &GENUS, 0, 0, 0, 0 };
 
3522
 
 
3523
void XSIMD(codelet_n1bv_128) (planner *p) {
 
3524
     X(kdft_register) (p, n1bv_128, &desc);
 
3525
}
 
3526
 
 
3527
#endif                          /* HAVE_FMA */