2
* Copyright (c) 2003, 2007-8 Matteo Frigo
3
* Copyright (c) 2003, 2007-8 Massachusetts Institute of Technology
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.
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.
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
21
/* This file was automatically generated --- DO NOT EDIT */
22
/* Generated on Sun Jul 12 06:39:13 EDT 2009 */
24
#include "codelet-dft.h"
28
/* Generated by: ../../../genfft/gen_notw_c -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 128 -name n1fv_128 -include n1f.h */
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
37
static void n1fv_128(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
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);
75
for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
76
V T6a, T5J, T6b, T5K, T6B, T6C, T6J, T6A, T6o, T6j, T6r, T68, T6e, T5O, T5R;
79
V Tad, TcZ, T6Z, T8T, T4U, Tr, Tfq, TgG, Ted, Tgf, Td0, Tcc, T9k, T84, Tb6;
80
V Tbt, Td8, TdK, TeK, Tgq, TeV, Tgt, T7q, T94, T3p, T5X, T7B, T97, T2G, T5U;
81
V TbD, Tc0, Tdf, TdN, Tf5, Tgx, Tfg, TgA, T7J, T9b, T4E, T64, T7U, T9e, T3V;
82
V T61, Td2, Td3, T85, T72, T4V, TI, Tcd, Tas, TgH, Tek, Tgg, Tft, T86, T75;
83
V T4W, TZ, TaI, Tcg, Tdr, TdG, Tgi, Tet, Tgj, Teq, T8X, T7a, T5M, T1B, T8W;
84
V T7d, T5N, T1s, TaX, Tcf, Tdo, TdH, Tgl, TeC, Tgm, Tez, T90, T7h, T5P, T2c;
85
V T8Z, T7k, T5Q, T23, T3Y, T49, TdL, Tdb, Tbu, Tbl, Tgu, TeR, Tgr, TeY, Tf6;
86
V TbG, T5V, T3s, T5Y, T3f, T95, T7E, T98, T7x, T4g, T4f, T4q, TbH, T41, TbI;
87
V T44, T4h, T4j, T4k, Tf9, TbN;
89
V Tu, TF, Ty, TL, TW, Tah, Tx, Tag, Tee, Tz, TM, TN, Teh, Tan, TP;
92
V TeG, T2A, Tbq, TeT, Tbp, TeH, T3m, T2x, Td6, T7o, T2q, T3l, T7z, Tbr, T2D;
95
V Ta7, T3, Ta8, T4O, Taa, Tab, Ta, T4P, Te, Tc9, Th, Tca, Tl, Tc6, Tc7;
99
T1 = LD(&(xi[0]), ivs, &(xi[0]));
100
T2 = LD(&(xi[WS(is, 64)]), ivs, &(xi[0]));
101
T4M = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
102
T4N = LD(&(xi[WS(is, 96)]), ivs, &(xi[0]));
105
T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
106
T5 = LD(&(xi[WS(is, 80)]), ivs, &(xi[0]));
107
T7 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0]));
108
T8 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
110
V Tc, T6, T9, Td, Tf, Tg;
111
Tc = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
114
Ta8 = VADD(T4M, T4N);
115
T4O = VSUB(T4M, T4N);
120
Td = LD(&(xi[WS(is, 72)]), ivs, &(xi[0]));
121
Tf = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
122
Tg = LD(&(xi[WS(is, 104)]), ivs, &(xi[0]));
125
Tj = LD(&(xi[WS(is, 120)]), ivs, &(xi[0]));
126
Tk = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
127
Tm = LD(&(xi[WS(is, 88)]), ivs, &(xi[0]));
128
Tn = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
144
V T6X, Tb, Te9, Ta9, Tcb, Tea, T4R, Ti, Tfo, Tac, Tp, T4S, Tc8, Teb, T4Q;
145
T6X = VFNMS(LDK(KP707106781), Ta, T3);
146
Tb = VFMA(LDK(KP707106781), Ta, T3);
147
Te9 = VSUB(Ta7, Ta8);
148
Ta9 = VADD(Ta7, Ta8);
149
Tcb = VADD(Tc9, Tca);
150
Tea = VSUB(Tc9, Tca);
151
T4R = VFMA(LDK(KP414213562), Te, Th);
152
Ti = VFNMS(LDK(KP414213562), Th, Te);
153
Tfo = VSUB(Tab, Taa);
154
Tac = VADD(Taa, Tab);
155
Tp = VFNMS(LDK(KP414213562), To, Tl);
156
T4S = VFMA(LDK(KP414213562), Tl, To);
157
Tc8 = VADD(Tc6, Tc7);
158
Teb = VSUB(Tc6, Tc7);
159
T4Q = VFNMS(LDK(KP707106781), T4P, T4O);
160
T82 = VFMA(LDK(KP707106781), T4P, T4O);
162
V T4T, T6Y, Tq, Tfp, Tec;
163
T4T = VSUB(T4R, T4S);
164
T6Y = VADD(T4R, T4S);
167
Tfp = VSUB(Teb, Tea);
168
Tec = VADD(Tea, Teb);
169
Tad = VSUB(Ta9, Tac);
170
TcZ = VADD(Ta9, Tac);
171
T6Z = VFMA(LDK(KP923879532), T6Y, T6X);
172
T8T = VFNMS(LDK(KP923879532), T6Y, T6X);
173
T4U = VFMA(LDK(KP923879532), T4T, T4Q);
174
T6a = VFNMS(LDK(KP923879532), T4T, T4Q);
175
Tr = VFMA(LDK(KP923879532), Tq, Tb);
176
T5J = VFNMS(LDK(KP923879532), Tq, Tb);
177
Tfq = VFMA(LDK(KP707106781), Tfp, Tfo);
178
TgG = VFNMS(LDK(KP707106781), Tfp, Tfo);
179
Ted = VFMA(LDK(KP707106781), Tec, Te9);
180
Tgf = VFNMS(LDK(KP707106781), Tec, Te9);
181
Td0 = VADD(Tcb, Tc8);
182
Tcc = VSUB(Tc8, Tcb);
187
V T2i, T3j, Tb2, T2B, Tb5, T3k, T2p, T2C;
189
V T2m, Tb0, Tb1, Tb3, T2l, T2n;
191
V T2g, T2h, T3h, T3i, T2j, T2k;
192
T2g = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
193
T2h = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)]));
194
T3h = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
195
T3i = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)]));
196
T2j = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
197
T2k = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)]));
198
T2m = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)]));
199
T9k = VFNMS(LDK(KP923879532), T83, T82);
200
T84 = VFMA(LDK(KP923879532), T83, T82);
201
T2i = VSUB(T2g, T2h);
202
Tb0 = VADD(T2g, T2h);
203
T3j = VSUB(T3h, T3i);
204
Tb1 = VADD(T3h, T3i);
205
Tb3 = VADD(T2j, T2k);
206
T2l = VSUB(T2j, T2k);
207
T2n = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
210
V T2r, T2s, T2u, T2v;
211
T2r = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
212
T2s = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)]));
213
T2u = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
214
T2v = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)]));
215
TeG = VSUB(Tb0, Tb1);
216
Tb2 = VADD(Tb0, Tb1);
218
V T2y, T2z, Tb4, T2o, Tbn, T2t, Tbo, T2w;
219
T2y = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)]));
220
T2z = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
221
Tb4 = VADD(T2m, T2n);
222
T2o = VSUB(T2m, T2n);
223
Tbn = VADD(T2r, T2s);
224
T2t = VSUB(T2r, T2s);
225
Tbo = VADD(T2u, T2v);
226
T2w = VSUB(T2u, T2v);
227
T2B = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
228
T2A = VSUB(T2y, T2z);
229
Tbq = VADD(T2y, T2z);
230
TeT = VSUB(Tb3, Tb4);
231
Tb5 = VADD(Tb3, Tb4);
232
T3k = VSUB(T2l, T2o);
233
T2p = VADD(T2l, T2o);
234
Tbp = VADD(Tbn, Tbo);
235
TeH = VSUB(Tbn, Tbo);
236
T3m = VFMA(LDK(KP414213562), T2t, T2w);
237
T2x = VFNMS(LDK(KP414213562), T2w, T2t);
238
T2C = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)]));
242
Td6 = VADD(Tb2, Tb5);
243
Tb6 = VSUB(Tb2, Tb5);
244
T7o = VFNMS(LDK(KP707106781), T2p, T2i);
245
T2q = VFMA(LDK(KP707106781), T2p, T2i);
246
T3l = VFMA(LDK(KP707106781), T3k, T3j);
247
T7z = VFNMS(LDK(KP707106781), T3k, T3j);
248
Tbr = VADD(T2B, T2C);
249
T2D = VSUB(T2B, T2C);
252
V Tf1, Tfe, Tf2, TbZ, T3M, T4B, Tdd, T3F, T7H, T4A, T7S, TbW, Tf3, T4C, T3T;
254
V T3x, T4y, Tbz, T3Q, TbC, T4z, T3E, T3R, T3P, TbU, TbV, T3S;
256
V T3y, T3z, T3B, T3C;
258
V T3v, T3w, T4w, T4x;
259
T3v = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)]));
260
T3w = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
261
T4w = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)]));
262
T4x = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
263
T3y = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
265
V Tbs, TeI, T3n, T2E, Tbx;
266
Tbs = VADD(Tbq, Tbr);
267
TeI = VSUB(Tbq, Tbr);
268
T3n = VFNMS(LDK(KP414213562), T2A, T2D);
269
T2E = VFMA(LDK(KP414213562), T2D, T2A);
270
T3x = VSUB(T3v, T3w);
271
Tbx = VADD(T3v, T3w);
273
V Tby, Td7, TeJ, TeU;
274
T4y = VSUB(T4w, T4x);
275
Tby = VADD(T4x, T4w);
276
Td7 = VADD(Tbp, Tbs);
277
Tbt = VSUB(Tbp, Tbs);
278
TeJ = VADD(TeH, TeI);
279
TeU = VSUB(TeH, TeI);
281
V T7p, T3o, T7A, T2F;
282
T7p = VSUB(T3m, T3n);
283
T3o = VADD(T3m, T3n);
284
T7A = VSUB(T2x, T2E);
285
T2F = VADD(T2x, T2E);
286
Tbz = VADD(Tbx, Tby);
287
Tf1 = VSUB(Tbx, Tby);
288
Td8 = VADD(Td6, Td7);
289
TdK = VSUB(Td6, Td7);
290
TeK = VFMA(LDK(KP707106781), TeJ, TeG);
291
Tgq = VFNMS(LDK(KP707106781), TeJ, TeG);
292
TeV = VFMA(LDK(KP707106781), TeU, TeT);
293
Tgt = VFNMS(LDK(KP707106781), TeU, TeT);
294
T7q = VFMA(LDK(KP923879532), T7p, T7o);
295
T94 = VFNMS(LDK(KP923879532), T7p, T7o);
296
T3p = VFMA(LDK(KP923879532), T3o, T3l);
297
T5X = VFNMS(LDK(KP923879532), T3o, T3l);
298
T7B = VFNMS(LDK(KP923879532), T7A, T7z);
299
T97 = VFMA(LDK(KP923879532), T7A, T7z);
300
T2G = VFMA(LDK(KP923879532), T2F, T2q);
301
T5U = VFNMS(LDK(KP923879532), T2F, T2q);
302
T3z = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)]));
306
T3B = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)]));
307
T3C = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
310
V T3G, T3H, T3J, T3K;
311
T3G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
312
T3H = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)]));
313
T3J = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)]));
314
T3K = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
316
V T3N, T3A, TbA, T3D, TbB, T3I, TbX, T3L, TbY, T3O;
317
T3N = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)]));
318
T3A = VSUB(T3y, T3z);
319
TbA = VADD(T3y, T3z);
320
T3D = VSUB(T3B, T3C);
321
TbB = VADD(T3B, T3C);
322
T3I = VSUB(T3G, T3H);
323
TbX = VADD(T3G, T3H);
324
T3L = VSUB(T3J, T3K);
325
TbY = VADD(T3K, T3J);
326
T3O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
327
T3Q = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)]));
328
Tfe = VSUB(TbB, TbA);
329
TbC = VADD(TbA, TbB);
330
T4z = VSUB(T3D, T3A);
331
T3E = VADD(T3A, T3D);
332
T3R = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
333
Tf2 = VSUB(TbX, TbY);
334
TbZ = VADD(TbX, TbY);
335
T3M = VFMA(LDK(KP414213562), T3L, T3I);
336
T4B = VFNMS(LDK(KP414213562), T3I, T3L);
337
T3P = VSUB(T3N, T3O);
338
TbU = VADD(T3N, T3O);
342
Tdd = VADD(Tbz, TbC);
343
TbD = VSUB(Tbz, TbC);
344
TbV = VADD(T3R, T3Q);
345
T3S = VSUB(T3Q, T3R);
346
T3F = VFMA(LDK(KP707106781), T3E, T3x);
347
T7H = VFNMS(LDK(KP707106781), T3E, T3x);
348
T4A = VFMA(LDK(KP707106781), T4z, T4y);
349
T7S = VFNMS(LDK(KP707106781), T4z, T4y);
350
TbW = VADD(TbU, TbV);
351
Tf3 = VSUB(TbU, TbV);
352
T4C = VFMA(LDK(KP414213562), T3P, T3S);
353
T3T = VFNMS(LDK(KP414213562), T3S, T3P);
356
V TD, Tae, TE, TJ, TK, TU, TV;
358
V Ts, Tt, Tde, Tf4, Tff;
359
Ts = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
360
Tt = LD(&(xi[WS(is, 68)]), ivs, &(xi[0]));
361
TD = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
362
Tde = VADD(TbZ, TbW);
363
Tc0 = VSUB(TbW, TbZ);
364
Tf4 = VADD(Tf2, Tf3);
365
Tff = VSUB(Tf3, Tf2);
367
V T7I, T4D, T7T, T3U;
368
T7I = VSUB(T4C, T4B);
369
T4D = VADD(T4B, T4C);
370
T7T = VSUB(T3T, T3M);
371
T3U = VADD(T3M, T3T);
374
Tdf = VADD(Tdd, Tde);
375
TdN = VSUB(Tdd, Tde);
376
Tf5 = VFMA(LDK(KP707106781), Tf4, Tf1);
377
Tgx = VFNMS(LDK(KP707106781), Tf4, Tf1);
378
Tfg = VFMA(LDK(KP707106781), Tff, Tfe);
379
TgA = VFNMS(LDK(KP707106781), Tff, Tfe);
380
T7J = VFMA(LDK(KP923879532), T7I, T7H);
381
T9b = VFNMS(LDK(KP923879532), T7I, T7H);
382
T4E = VFMA(LDK(KP923879532), T4D, T4A);
383
T64 = VFNMS(LDK(KP923879532), T4D, T4A);
384
T7U = VFNMS(LDK(KP923879532), T7T, T7S);
385
T9e = VFMA(LDK(KP923879532), T7T, T7S);
386
T3V = VFMA(LDK(KP923879532), T3U, T3F);
387
T61 = VFNMS(LDK(KP923879532), T3U, T3F);
388
TE = LD(&(xi[WS(is, 100)]), ivs, &(xi[0]));
391
TJ = LD(&(xi[WS(is, 124)]), ivs, &(xi[0]));
392
TK = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
393
TU = LD(&(xi[WS(is, 92)]), ivs, &(xi[0]));
394
TV = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
396
V Tal, Tam, Tv, Tw, Taf;
397
Tv = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
398
Tw = LD(&(xi[WS(is, 84)]), ivs, &(xi[0]));
401
Ty = LD(&(xi[WS(is, 116)]), ivs, &(xi[0]));
408
Tag = VADD(Tae, Taf);
409
Tee = VSUB(Tae, Taf);
410
Tz = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
411
TM = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
412
TN = LD(&(xi[WS(is, 76)]), ivs, &(xi[0]));
413
Teh = VSUB(Tal, Tam);
414
Tan = VADD(Tal, Tam);
415
TP = LD(&(xi[WS(is, 108)]), ivs, &(xi[0]));
416
TQ = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
422
V Tev, TeA, Tdm, TaP, Tew, TaV, T1U, T29, T7f, T1N, T28, T7i, Tex, TaS, T21;
425
V Tem, Ter, Ten, TaD, T1j, T1y, TaA, Tdp, T1c, T78, T7b, T1x, TaG, Teo, T1z;
428
V T14, T1v, Taw, Taz, T1b, T1w, T1n, T1o, T1m, TaE, TaF, T1p;
430
V Tau, Tav, T15, T16, T18, T19;
432
V T12, Tai, TA, Tao, TO, T13;
433
T12 = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
438
T13 = LD(&(xi[WS(is, 66)]), ivs, &(xi[0]));
440
V T1t, Tap, TR, Taj, Tef, TG, TB, T1u;
441
T1t = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
444
Taj = VADD(Tah, Tai);
445
Tef = VSUB(Tah, Tai);
448
Tau = VADD(T12, T13);
449
T14 = VSUB(T12, T13);
450
T1u = LD(&(xi[WS(is, 98)]), ivs, &(xi[0]));
452
V Taq, Tei, TX, TS, Tak;
453
Taq = VADD(Tao, Tap);
454
Tei = VSUB(Tap, Tao);
457
Tak = VSUB(Tag, Taj);
458
Td2 = VADD(Tag, Taj);
461
Teg = VFNMS(LDK(KP414213562), Tef, Tee);
462
Tfs = VFMA(LDK(KP414213562), Tee, Tef);
463
T71 = VFNMS(LDK(KP707106781), TG, TF);
464
TH = VFMA(LDK(KP707106781), TG, TF);
466
V T70, TC, Tar, Tej, Tfr;
467
T70 = VFNMS(LDK(KP707106781), TB, Tu);
468
TC = VFMA(LDK(KP707106781), TB, Tu);
469
Tar = VSUB(Tan, Taq);
470
Td3 = VADD(Tan, Taq);
471
Tej = VFNMS(LDK(KP414213562), Tei, Teh);
472
Tfr = VFMA(LDK(KP414213562), Teh, Tei);
475
T74 = VFNMS(LDK(KP707106781), TX, TW);
476
TY = VFMA(LDK(KP707106781), TX, TW);
477
T73 = VFNMS(LDK(KP707106781), TS, TL);
478
TT = VFMA(LDK(KP707106781), TS, TL);
479
T85 = VFNMS(LDK(KP668178637), T70, T71);
480
T72 = VFMA(LDK(KP668178637), T71, T70);
481
T4V = VFMA(LDK(KP198912367), TC, TH);
482
TI = VFNMS(LDK(KP198912367), TH, TC);
483
Tcd = VSUB(Tar, Tak);
484
Tas = VADD(Tak, Tar);
485
TgH = VSUB(Tej, Teg);
486
Tek = VADD(Teg, Tej);
487
Tgg = VADD(Tfs, Tfr);
488
Tft = VSUB(Tfr, Tfs);
489
T86 = VFNMS(LDK(KP668178637), T73, T74);
490
T75 = VFMA(LDK(KP668178637), T74, T73);
491
T4W = VFMA(LDK(KP198912367), TT, TY);
492
TZ = VFNMS(LDK(KP198912367), TY, TT);
493
Tav = VADD(T1t, T1u);
494
T1v = VSUB(T1t, T1u);
501
T15 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
502
T16 = LD(&(xi[WS(is, 82)]), ivs, &(xi[0]));
503
T18 = LD(&(xi[WS(is, 114)]), ivs, &(xi[0]));
504
T19 = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
506
V T1d, T1e, T1g, T1h, Tax, T17, Tay, T1a;
507
T1d = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
508
Taw = VADD(Tau, Tav);
509
Tem = VSUB(Tau, Tav);
510
T1e = LD(&(xi[WS(is, 74)]), ivs, &(xi[0]));
511
T1g = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
512
T1h = LD(&(xi[WS(is, 106)]), ivs, &(xi[0]));
513
Tax = VADD(T15, T16);
514
T17 = VSUB(T15, T16);
515
Tay = VADD(T18, T19);
516
T1a = VSUB(T18, T19);
518
V T1k, T1f, TaB, T1i, TaC, T1l;
519
T1k = LD(&(xi[WS(is, 122)]), ivs, &(xi[0]));
520
T1f = VSUB(T1d, T1e);
521
TaB = VADD(T1d, T1e);
522
T1i = VSUB(T1g, T1h);
523
TaC = VADD(T1g, T1h);
524
T1l = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
525
Taz = VADD(Tax, Tay);
526
Ter = VSUB(Tax, Tay);
527
T1b = VADD(T17, T1a);
528
T1w = VSUB(T17, T1a);
529
T1n = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
530
T1o = LD(&(xi[WS(is, 90)]), ivs, &(xi[0]));
531
Ten = VSUB(TaB, TaC);
532
TaD = VADD(TaB, TaC);
533
T1j = VFNMS(LDK(KP414213562), T1i, T1f);
534
T1y = VFMA(LDK(KP414213562), T1f, T1i);
535
T1m = VSUB(T1k, T1l);
536
TaE = VADD(T1k, T1l);
540
TaA = VSUB(Taw, Taz);
541
Tdp = VADD(Taw, Taz);
542
TaF = VADD(T1n, T1o);
543
T1p = VSUB(T1n, T1o);
544
T1c = VFMA(LDK(KP707106781), T1b, T14);
545
T78 = VFNMS(LDK(KP707106781), T1b, T14);
546
T7b = VFNMS(LDK(KP707106781), T1w, T1v);
547
T1x = VFMA(LDK(KP707106781), T1w, T1v);
548
TaG = VADD(TaE, TaF);
549
Teo = VSUB(TaE, TaF);
550
T1z = VFNMS(LDK(KP414213562), T1m, T1p);
551
T1q = VFMA(LDK(KP414213562), T1p, T1m);
554
V T1F, T26, T1Q, TaT, TaL, TaO, T27, T1M, T1Y, T1Z, TaU, T1T, TaQ, T1X, T20;
557
V T24, TaJ, T25, T1G, T1H, T1J, T1K, T1D, T1E;
558
T1D = LD(&(xi[WS(is, 126)]), ivs, &(xi[0]));
559
T1E = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
560
T24 = LD(&(xi[WS(is, 94)]), ivs, &(xi[0]));
562
V TaH, Tdq, Tes, Tep;
563
TaH = VSUB(TaD, TaG);
564
Tdq = VADD(TaD, TaG);
565
Tes = VSUB(Ten, Teo);
566
Tep = VADD(Ten, Teo);
568
V T79, T1A, T7c, T1r;
569
T79 = VSUB(T1y, T1z);
570
T1A = VADD(T1y, T1z);
571
T7c = VSUB(T1j, T1q);
572
T1r = VADD(T1j, T1q);
573
TaJ = VADD(T1D, T1E);
574
T1F = VSUB(T1D, T1E);
575
TaI = VFNMS(LDK(KP414213562), TaH, TaA);
576
Tcg = VFMA(LDK(KP414213562), TaA, TaH);
577
Tdr = VADD(Tdp, Tdq);
578
TdG = VSUB(Tdp, Tdq);
579
Tgi = VFNMS(LDK(KP707106781), Tes, Ter);
580
Tet = VFMA(LDK(KP707106781), Tes, Ter);
581
Tgj = VFNMS(LDK(KP707106781), Tep, Tem);
582
Teq = VFMA(LDK(KP707106781), Tep, Tem);
583
T8X = VFNMS(LDK(KP923879532), T79, T78);
584
T7a = VFMA(LDK(KP923879532), T79, T78);
585
T5M = VFNMS(LDK(KP923879532), T1A, T1x);
586
T1B = VFMA(LDK(KP923879532), T1A, T1x);
587
T8W = VFMA(LDK(KP923879532), T7c, T7b);
588
T7d = VFNMS(LDK(KP923879532), T7c, T7b);
589
T5N = VFNMS(LDK(KP923879532), T1r, T1c);
590
T1s = VFMA(LDK(KP923879532), T1r, T1c);
591
T25 = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
594
T1G = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
595
T1H = LD(&(xi[WS(is, 78)]), ivs, &(xi[0]));
596
T1J = LD(&(xi[WS(is, 110)]), ivs, &(xi[0]));
597
T1K = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
599
V T1R, T1I, TaM, T1L, TaN, T1S, T1O, T1P, TaK, T1V, T1W;
600
T1O = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
601
T1P = LD(&(xi[WS(is, 70)]), ivs, &(xi[0]));
602
T26 = VSUB(T24, T25);
603
TaK = VADD(T25, T24);
604
T1R = LD(&(xi[WS(is, 102)]), ivs, &(xi[0]));
605
T1I = VSUB(T1G, T1H);
606
TaM = VADD(T1G, T1H);
607
T1L = VSUB(T1J, T1K);
608
TaN = VADD(T1J, T1K);
609
T1Q = VSUB(T1O, T1P);
610
TaT = VADD(T1O, T1P);
611
Tev = VSUB(TaJ, TaK);
612
TaL = VADD(TaJ, TaK);
613
T1S = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
614
T1V = LD(&(xi[WS(is, 118)]), ivs, &(xi[0]));
615
T1W = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
616
TeA = VSUB(TaN, TaM);
617
TaO = VADD(TaM, TaN);
618
T27 = VSUB(T1L, T1I);
619
T1M = VADD(T1I, T1L);
620
T1Y = LD(&(xi[WS(is, 86)]), ivs, &(xi[0]));
621
T1Z = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
622
TaU = VADD(T1S, T1R);
623
T1T = VSUB(T1R, T1S);
624
TaQ = VADD(T1V, T1W);
625
T1X = VSUB(T1V, T1W);
628
Tdm = VADD(TaL, TaO);
629
TaP = VSUB(TaL, TaO);
630
T20 = VSUB(T1Y, T1Z);
631
TaR = VADD(T1Z, T1Y);
632
Tew = VSUB(TaT, TaU);
633
TaV = VADD(TaT, TaU);
634
T1U = VFMA(LDK(KP414213562), T1T, T1Q);
635
T29 = VFNMS(LDK(KP414213562), T1Q, T1T);
636
T7f = VFNMS(LDK(KP707106781), T1M, T1F);
637
T1N = VFMA(LDK(KP707106781), T1M, T1F);
638
T28 = VFMA(LDK(KP707106781), T27, T26);
639
T7i = VFNMS(LDK(KP707106781), T27, T26);
640
Tex = VSUB(TaQ, TaR);
641
TaS = VADD(TaQ, TaR);
642
T21 = VFNMS(LDK(KP414213562), T20, T1X);
643
T2a = VFMA(LDK(KP414213562), T1X, T20);
647
V T2J, T2U, T30, T3b, TeL, Tb9, TeO, Tbg, T2M, Tba, T2P, Tbb, T34, Tbh, T33;
650
V T2H, T2I, T2S, T2T, T2Y, T2Z, T39, T3a;
651
T2H = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
653
V Tdn, TaW, Tey, TeB;
654
Tdn = VADD(TaV, TaS);
655
TaW = VSUB(TaS, TaV);
656
Tey = VADD(Tew, Tex);
657
TeB = VSUB(Tex, Tew);
659
V T2b, T7g, T22, T7j;
660
T2b = VADD(T29, T2a);
661
T7g = VSUB(T2a, T29);
662
T22 = VADD(T1U, T21);
663
T7j = VSUB(T21, T1U);
664
TaX = VFNMS(LDK(KP414213562), TaW, TaP);
665
Tcf = VFMA(LDK(KP414213562), TaP, TaW);
666
Tdo = VADD(Tdm, Tdn);
667
TdH = VSUB(Tdm, Tdn);
668
Tgl = VFNMS(LDK(KP707106781), TeB, TeA);
669
TeC = VFMA(LDK(KP707106781), TeB, TeA);
670
Tgm = VFNMS(LDK(KP707106781), Tey, Tev);
671
Tez = VFMA(LDK(KP707106781), Tey, Tev);
672
T90 = VFNMS(LDK(KP923879532), T7g, T7f);
673
T7h = VFMA(LDK(KP923879532), T7g, T7f);
674
T5P = VFNMS(LDK(KP923879532), T2b, T28);
675
T2c = VFMA(LDK(KP923879532), T2b, T28);
676
T8Z = VFMA(LDK(KP923879532), T7j, T7i);
677
T7k = VFNMS(LDK(KP923879532), T7j, T7i);
678
T5Q = VFNMS(LDK(KP923879532), T22, T1N);
679
T23 = VFMA(LDK(KP923879532), T22, T1N);
680
T2I = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)]));
683
T2S = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
684
T2T = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)]));
685
T2Y = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)]));
686
T2Z = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
687
T39 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
688
T3a = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)]));
690
V T2K, Tbe, Tbf, T2L, T2N, T2O, Tb7, Tb8, T31, T32;
691
T2K = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
692
T2J = VSUB(T2H, T2I);
693
Tb7 = VADD(T2H, T2I);
694
T2U = VSUB(T2S, T2T);
695
Tb8 = VADD(T2S, T2T);
696
T30 = VSUB(T2Y, T2Z);
697
Tbe = VADD(T2Y, T2Z);
698
T3b = VSUB(T39, T3a);
699
Tbf = VADD(T39, T3a);
700
T2L = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)]));
701
T2N = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)]));
702
T2O = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
703
TeL = VSUB(Tb7, Tb8);
704
Tb9 = VADD(Tb7, Tb8);
705
T31 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
706
T32 = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)]));
707
TeO = VSUB(Tbe, Tbf);
708
Tbg = VADD(Tbe, Tbf);
709
T2M = VSUB(T2K, T2L);
710
Tba = VADD(T2K, T2L);
711
T2P = VSUB(T2N, T2O);
712
Tbb = VADD(T2N, T2O);
713
T34 = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)]));
714
Tbh = VADD(T31, T32);
715
T33 = VSUB(T31, T32);
716
T35 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
720
V T4d, T4e, T4o, T4p;
722
V T2X, T3q, T7t, T7C, T3r, T3e, T7D, T7w;
724
V T47, TbE, Tbd, Td9, TeW, TeN, T7s, T2W, T7r, T2R, TeP, Tbj, T37, T3c, T48;
726
V T3W, T3X, TeM, Tbc, T2Q, T2V, Tbi, T36;
727
T3W = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
728
T3X = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)]));
729
TeM = VSUB(Tba, Tbb);
730
Tbc = VADD(Tba, Tbb);
731
T2Q = VADD(T2M, T2P);
732
T2V = VSUB(T2M, T2P);
733
T47 = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)]));
734
Tbi = VADD(T34, T35);
735
T36 = VSUB(T34, T35);
736
TbE = VADD(T3W, T3X);
737
T3Y = VSUB(T3W, T3X);
738
Tbd = VSUB(Tb9, Tbc);
739
Td9 = VADD(Tb9, Tbc);
740
TeW = VFMA(LDK(KP414213562), TeL, TeM);
741
TeN = VFNMS(LDK(KP414213562), TeM, TeL);
742
T7s = VFNMS(LDK(KP707106781), T2V, T2U);
743
T2W = VFMA(LDK(KP707106781), T2V, T2U);
744
T7r = VFNMS(LDK(KP707106781), T2Q, T2J);
745
T2R = VFMA(LDK(KP707106781), T2Q, T2J);
746
TeP = VSUB(Tbh, Tbi);
747
Tbj = VADD(Tbh, Tbi);
748
T37 = VADD(T33, T36);
749
T3c = VSUB(T33, T36);
750
T48 = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
752
T2X = VFNMS(LDK(KP198912367), T2W, T2R);
753
T3q = VFMA(LDK(KP198912367), T2R, T2W);
754
T7t = VFMA(LDK(KP668178637), T7s, T7r);
755
T7C = VFNMS(LDK(KP668178637), T7r, T7s);
757
V Tbk, Tda, TeX, TeQ;
758
Tbk = VSUB(Tbg, Tbj);
759
Tda = VADD(Tbg, Tbj);
760
TeX = VFNMS(LDK(KP414213562), TeO, TeP);
761
TeQ = VFMA(LDK(KP414213562), TeP, TeO);
763
V T7v, T3d, T7u, T38, TbF;
764
T7v = VFNMS(LDK(KP707106781), T3c, T3b);
765
T3d = VFMA(LDK(KP707106781), T3c, T3b);
766
T7u = VFNMS(LDK(KP707106781), T37, T30);
767
T38 = VFMA(LDK(KP707106781), T37, T30);
768
T49 = VSUB(T47, T48);
769
TbF = VADD(T48, T47);
770
TdL = VSUB(Td9, Tda);
771
Tdb = VADD(Td9, Tda);
772
Tbu = VSUB(Tbd, Tbk);
773
Tbl = VADD(Tbd, Tbk);
774
Tgu = VSUB(TeN, TeQ);
775
TeR = VADD(TeN, TeQ);
776
Tgr = VSUB(TeW, TeX);
777
TeY = VADD(TeW, TeX);
778
T3r = VFNMS(LDK(KP198912367), T38, T3d);
779
T3e = VFMA(LDK(KP198912367), T3d, T38);
780
T7D = VFMA(LDK(KP668178637), T7u, T7v);
781
T7w = VFNMS(LDK(KP668178637), T7v, T7u);
782
Tf6 = VSUB(TbE, TbF);
783
TbG = VADD(TbE, TbF);
787
T4d = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)]));
788
T5V = VSUB(T3q, T3r);
789
T3s = VADD(T3q, T3r);
790
T5Y = VSUB(T2X, T3e);
791
T3f = VADD(T2X, T3e);
792
T95 = VSUB(T7D, T7C);
793
T7E = VADD(T7C, T7D);
794
T98 = VSUB(T7t, T7w);
795
T7x = VADD(T7t, T7w);
796
T4e = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
797
T4o = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)]));
798
T4p = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
801
V T3Z, T40, T42, T43, TbL, TbM;
802
T3Z = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
803
T40 = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)]));
804
T42 = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)]));
805
T43 = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
806
T4g = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
807
T4f = VSUB(T4d, T4e);
808
TbL = VADD(T4d, T4e);
809
T4q = VSUB(T4o, T4p);
810
TbM = VADD(T4p, T4o);
811
TbH = VADD(T3Z, T40);
812
T41 = VSUB(T3Z, T40);
813
TbI = VADD(T42, T43);
814
T44 = VSUB(T42, T43);
815
T4h = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)]));
816
T4j = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)]));
817
T4k = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
818
Tf9 = VSUB(TbL, TbM);
819
TbN = VADD(TbL, TbM);
826
V TgB, Tgy, T62, T4H, T65, T4u, T9c, T7X, T9f, T7Q, Tg0, Tga, TfF, TeF, TfT;
827
V TfU, TfP, Tg7, TfI, Tfy, TfA, Tf0, Tfz, Tfl, Tg2, TfS;
829
V Tc1, TbS, Tfc, Tfj, TdX, Te5, TdZ, TdR, Te7, Te3, TdU, Te4;
831
V TdF, TdS, Tdx, Td5, TdO, TdE, TdC, Tdt, Tdk;
833
V Tdc, TdA, T4F, T4c, T7V, T7M, T4G, T4t, T7W, T7P, TdB, Tdj;
835
V Td1, Tdg, TbK, Tf8, Tfh, T4b, T7L, T46, T7K, TbQ, Tfa, T4r, T4m, Td4;
836
TdF = VSUB(TcZ, Td0);
837
Td1 = VADD(TcZ, Td0);
839
V TbJ, Tf7, T4a, T45;
840
TbJ = VADD(TbH, TbI);
841
Tf7 = VSUB(TbI, TbH);
842
T4a = VSUB(T44, T41);
843
T45 = VADD(T41, T44);
845
V TbO, T4i, TbP, T4l;
846
TbO = VADD(T4g, T4h);
847
T4i = VSUB(T4g, T4h);
848
TbP = VADD(T4j, T4k);
849
T4l = VSUB(T4j, T4k);
850
Tdg = VADD(TbG, TbJ);
851
TbK = VSUB(TbG, TbJ);
852
Tf8 = VFMA(LDK(KP414213562), Tf7, Tf6);
853
Tfh = VFNMS(LDK(KP414213562), Tf6, Tf7);
854
T4b = VFMA(LDK(KP707106781), T4a, T49);
855
T7L = VFNMS(LDK(KP707106781), T4a, T49);
856
T46 = VFMA(LDK(KP707106781), T45, T3Y);
857
T7K = VFNMS(LDK(KP707106781), T45, T3Y);
858
TbQ = VADD(TbO, TbP);
859
Tfa = VSUB(TbP, TbO);
860
T4r = VSUB(T4l, T4i);
861
T4m = VADD(T4i, T4l);
862
Td4 = VADD(Td2, Td3);
863
TdS = VSUB(Td3, Td2);
866
Tdc = VSUB(Td8, Tdb);
867
TdA = VADD(Td8, Tdb);
868
T4F = VFNMS(LDK(KP198912367), T46, T4b);
869
T4c = VFMA(LDK(KP198912367), T4b, T46);
870
T7V = VFMA(LDK(KP668178637), T7K, T7L);
871
T7M = VFNMS(LDK(KP668178637), T7L, T7K);
873
V Tdh, TbR, Tfb, Tfi;
874
Tdh = VADD(TbN, TbQ);
875
TbR = VSUB(TbN, TbQ);
876
Tfb = VFNMS(LDK(KP414213562), Tfa, Tf9);
877
Tfi = VFMA(LDK(KP414213562), Tf9, Tfa);
879
V T4s, T7O, T4n, T7N, Tdi;
880
T4s = VFMA(LDK(KP707106781), T4r, T4q);
881
T7O = VFNMS(LDK(KP707106781), T4r, T4q);
882
T4n = VFMA(LDK(KP707106781), T4m, T4f);
883
T7N = VFNMS(LDK(KP707106781), T4m, T4f);
884
Tdx = VADD(Td1, Td4);
885
Td5 = VSUB(Td1, Td4);
886
TdO = VSUB(Tdh, Tdg);
887
Tdi = VADD(Tdg, Tdh);
888
Tc1 = VSUB(TbR, TbK);
889
TbS = VADD(TbK, TbR);
890
TgB = VSUB(Tfb, Tf8);
891
Tfc = VADD(Tf8, Tfb);
892
Tgy = VSUB(Tfi, Tfh);
893
Tfj = VADD(Tfh, Tfi);
894
T4G = VFMA(LDK(KP198912367), T4n, T4s);
895
T4t = VFNMS(LDK(KP198912367), T4s, T4n);
896
T7W = VFNMS(LDK(KP668178637), T7N, T7O);
897
T7P = VFMA(LDK(KP668178637), T7O, T7N);
898
TdB = VADD(Tdf, Tdi);
899
Tdj = VSUB(Tdf, Tdi);
903
T62 = VSUB(T4G, T4F);
904
T4H = VADD(T4F, T4G);
905
T65 = VSUB(T4t, T4c);
906
T4u = VADD(T4c, T4t);
907
T9c = VSUB(T7V, T7W);
908
T7X = VADD(T7V, T7W);
909
T9f = VSUB(T7P, T7M);
910
T7Q = VADD(T7M, T7P);
911
TdE = VSUB(TdB, TdA);
912
TdC = VADD(TdA, TdB);
913
Tdt = VSUB(Tdj, Tdc);
914
Tdk = VADD(Tdc, Tdj);
917
V TdT, Tdl, Tdv, TdJ, Te1, Te2, TdQ, Tdz, TdD, Tdu, Tdw;
919
V TdI, TdP, TdV, TdW, TdM, Tds, Tdy;
920
TdI = VADD(TdG, TdH);
921
TdT = VSUB(TdH, TdG);
922
TdP = VFNMS(LDK(KP414213562), TdO, TdN);
923
TdV = VFMA(LDK(KP414213562), TdN, TdO);
924
TdW = VFMA(LDK(KP414213562), TdK, TdL);
925
TdM = VFNMS(LDK(KP414213562), TdL, TdK);
926
Tdl = VFNMS(LDK(KP707106781), Tdk, Td5);
927
Tdv = VFMA(LDK(KP707106781), Tdk, Td5);
928
Tds = VSUB(Tdo, Tdr);
929
Tdy = VADD(Tdr, Tdo);
930
TdJ = VFMA(LDK(KP707106781), TdI, TdF);
931
Te1 = VFNMS(LDK(KP707106781), TdI, TdF);
932
TdX = VSUB(TdV, TdW);
933
Te2 = VADD(TdW, TdV);
934
Te5 = VSUB(TdP, TdM);
935
TdQ = VADD(TdM, TdP);
936
Tdz = VADD(Tdx, Tdy);
937
TdD = VSUB(Tdx, Tdy);
938
Tdu = VFNMS(LDK(KP707106781), Tdt, Tds);
939
Tdw = VFMA(LDK(KP707106781), Tdt, Tds);
941
TdZ = VFMA(LDK(KP923879532), TdQ, TdJ);
942
TdR = VFNMS(LDK(KP923879532), TdQ, TdJ);
943
Te7 = VFMA(LDK(KP923879532), Te2, Te1);
944
Te3 = VFNMS(LDK(KP923879532), Te2, Te1);
945
ST(&(xo[WS(os, 32)]), VFMAI(TdE, TdD), ovs, &(xo[0]));
946
ST(&(xo[WS(os, 96)]), VFNMSI(TdE, TdD), ovs, &(xo[0]));
947
ST(&(xo[0]), VADD(Tdz, TdC), ovs, &(xo[0]));
948
ST(&(xo[WS(os, 64)]), VSUB(Tdz, TdC), ovs, &(xo[0]));
949
ST(&(xo[WS(os, 16)]), VFMAI(Tdw, Tdv), ovs, &(xo[0]));
950
ST(&(xo[WS(os, 112)]), VFNMSI(Tdw, Tdv), ovs, &(xo[0]));
951
ST(&(xo[WS(os, 80)]), VFMAI(Tdu, Tdl), ovs, &(xo[0]));
952
ST(&(xo[WS(os, 48)]), VFNMSI(Tdu, Tdl), ovs, &(xo[0]));
953
TdU = VFMA(LDK(KP707106781), TdT, TdS);
954
Te4 = VFNMS(LDK(KP707106781), TdT, TdS);
958
V Tcx, TcJ, TcI, Tcy, TcA, Tbm, Tcp, TaZ, Tcs, Tci, Tbv, TcB, TcD, TbT, Tc2;
960
Tcx = VFNMS(LDK(KP707106781), Tas, Tad);
961
Tat = VFMA(LDK(KP707106781), Tas, Tad);
962
TaY = VADD(TaI, TaX);
963
TcJ = VSUB(TaX, TaI);
965
V Tce, Tch, Te8, Te6, TdY, Te0;
966
TcI = VFNMS(LDK(KP707106781), Tcd, Tcc);
967
Tce = VFMA(LDK(KP707106781), Tcd, Tcc);
968
Tch = VSUB(Tcf, Tcg);
969
Tcy = VADD(Tcg, Tcf);
970
Te8 = VFNMS(LDK(KP923879532), Te5, Te4);
971
Te6 = VFMA(LDK(KP923879532), Te5, Te4);
972
TdY = VFNMS(LDK(KP923879532), TdX, TdU);
973
Te0 = VFMA(LDK(KP923879532), TdX, TdU);
974
TcA = VFNMS(LDK(KP707106781), Tbl, Tb6);
975
Tbm = VFMA(LDK(KP707106781), Tbl, Tb6);
976
Tcp = VFNMS(LDK(KP923879532), TaY, Tat);
977
TaZ = VFMA(LDK(KP923879532), TaY, Tat);
978
Tcs = VFNMS(LDK(KP923879532), Tch, Tce);
979
Tci = VFMA(LDK(KP923879532), Tch, Tce);
980
ST(&(xo[WS(os, 88)]), VFNMSI(Te6, Te3), ovs, &(xo[0]));
981
ST(&(xo[WS(os, 40)]), VFMAI(Te6, Te3), ovs, &(xo[0]));
982
ST(&(xo[WS(os, 104)]), VFMAI(Te8, Te7), ovs, &(xo[0]));
983
ST(&(xo[WS(os, 24)]), VFNMSI(Te8, Te7), ovs, &(xo[0]));
984
ST(&(xo[WS(os, 8)]), VFMAI(Te0, TdZ), ovs, &(xo[0]));
985
ST(&(xo[WS(os, 120)]), VFNMSI(Te0, TdZ), ovs, &(xo[0]));
986
ST(&(xo[WS(os, 72)]), VFMAI(TdY, TdR), ovs, &(xo[0]));
987
ST(&(xo[WS(os, 56)]), VFNMSI(TdY, TdR), ovs, &(xo[0]));
988
Tbv = VFMA(LDK(KP707106781), Tbu, Tbt);
989
TcB = VFNMS(LDK(KP707106781), Tbu, Tbt);
990
TcD = VFNMS(LDK(KP707106781), TbS, TbD);
991
TbT = VFMA(LDK(KP707106781), TbS, TbD);
992
Tc2 = VFMA(LDK(KP707106781), Tc1, Tc0);
993
TcE = VFNMS(LDK(KP707106781), Tc1, Tc0);
996
V TcR, Tcz, TcU, TcK, Tcq, Tcl, Tct, Tc4;
998
V Tck, Tbw, Tcj, Tc3;
999
Tck = VFMA(LDK(KP198912367), Tbm, Tbv);
1000
Tbw = VFNMS(LDK(KP198912367), Tbv, Tbm);
1001
Tcj = VFMA(LDK(KP198912367), TbT, Tc2);
1002
Tc3 = VFNMS(LDK(KP198912367), Tc2, TbT);
1003
TcR = VFNMS(LDK(KP923879532), Tcy, Tcx);
1004
Tcz = VFMA(LDK(KP923879532), Tcy, Tcx);
1005
TcU = VFMA(LDK(KP923879532), TcJ, TcI);
1006
TcK = VFNMS(LDK(KP923879532), TcJ, TcI);
1007
Tcq = VADD(Tck, Tcj);
1008
Tcl = VSUB(Tcj, Tck);
1009
Tct = VSUB(Tc3, Tbw);
1010
Tc4 = VADD(Tbw, Tc3);
1013
V TfN, Tel, TfY, Tfu, Tfw, Tfv, TcT, TcX, TcQ, TcO, TcW, TcY, TcP, TcH, TfZ;
1016
V Teu, TcS, TcN, TcV, TcG, TeD;
1017
TfN = VFNMS(LDK(KP923879532), Tek, Ted);
1018
Tel = VFMA(LDK(KP923879532), Tek, Ted);
1020
V TcM, TcC, Tcr, Tcv;
1021
TcM = VFNMS(LDK(KP668178637), TcA, TcB);
1022
TcC = VFMA(LDK(KP668178637), TcB, TcA);
1023
Tcr = VFNMS(LDK(KP980785280), Tcq, Tcp);
1024
Tcv = VFMA(LDK(KP980785280), Tcq, Tcp);
1026
V Tco, Tcm, Tcu, Tcw;
1027
Tco = VFMA(LDK(KP980785280), Tcl, Tci);
1028
Tcm = VFNMS(LDK(KP980785280), Tcl, Tci);
1029
Tcu = VFMA(LDK(KP980785280), Tct, Tcs);
1030
Tcw = VFNMS(LDK(KP980785280), Tct, Tcs);
1032
V Tcn, Tc5, TcL, TcF;
1033
Tcn = VFMA(LDK(KP980785280), Tc4, TaZ);
1034
Tc5 = VFNMS(LDK(KP980785280), Tc4, TaZ);
1035
TcL = VFNMS(LDK(KP668178637), TcD, TcE);
1036
TcF = VFMA(LDK(KP668178637), TcE, TcD);
1037
TfY = VFNMS(LDK(KP923879532), Tft, Tfq);
1038
Tfu = VFMA(LDK(KP923879532), Tft, Tfq);
1039
Tfw = VFMA(LDK(KP198912367), Teq, Tet);
1040
Teu = VFNMS(LDK(KP198912367), Tet, Teq);
1041
ST(&(xo[WS(os, 92)]), VFNMSI(Tcu, Tcr), ovs, &(xo[0]));
1042
ST(&(xo[WS(os, 36)]), VFMAI(Tcu, Tcr), ovs, &(xo[0]));
1043
ST(&(xo[WS(os, 100)]), VFMAI(Tcw, Tcv), ovs, &(xo[0]));
1044
ST(&(xo[WS(os, 28)]), VFNMSI(Tcw, Tcv), ovs, &(xo[0]));
1045
ST(&(xo[WS(os, 4)]), VFMAI(Tco, Tcn), ovs, &(xo[0]));
1046
ST(&(xo[WS(os, 124)]), VFNMSI(Tco, Tcn), ovs, &(xo[0]));
1047
ST(&(xo[WS(os, 68)]), VFMAI(Tcm, Tc5), ovs, &(xo[0]));
1048
ST(&(xo[WS(os, 60)]), VFNMSI(Tcm, Tc5), ovs, &(xo[0]));
1049
TcS = VADD(TcM, TcL);
1050
TcN = VSUB(TcL, TcM);
1051
TcV = VSUB(TcF, TcC);
1052
TcG = VADD(TcC, TcF);
1053
TeD = VFNMS(LDK(KP198912367), TeC, Tez);
1054
Tfv = VFMA(LDK(KP198912367), Tez, TeC);
1058
TcT = VFMA(LDK(KP831469612), TcS, TcR);
1059
TcX = VFNMS(LDK(KP831469612), TcS, TcR);
1060
TcQ = VFMA(LDK(KP831469612), TcN, TcK);
1061
TcO = VFNMS(LDK(KP831469612), TcN, TcK);
1062
TcW = VFNMS(LDK(KP831469612), TcV, TcU);
1063
TcY = VFMA(LDK(KP831469612), TcV, TcU);
1064
TcP = VFMA(LDK(KP831469612), TcG, Tcz);
1065
TcH = VFNMS(LDK(KP831469612), TcG, Tcz);
1066
TfZ = VSUB(TeD, Teu);
1067
TeE = VADD(Teu, TeD);
1070
V TfQ, TeS, TfO, Tfx, TeZ, TfR, Tfd, Tfk;
1071
TfQ = VFNMS(LDK(KP923879532), TeR, TeK);
1072
TeS = VFMA(LDK(KP923879532), TeR, TeK);
1073
ST(&(xo[WS(os, 84)]), VFMAI(TcW, TcT), ovs, &(xo[0]));
1074
ST(&(xo[WS(os, 44)]), VFNMSI(TcW, TcT), ovs, &(xo[0]));
1075
ST(&(xo[WS(os, 108)]), VFNMSI(TcY, TcX), ovs, &(xo[0]));
1076
ST(&(xo[WS(os, 20)]), VFMAI(TcY, TcX), ovs, &(xo[0]));
1077
ST(&(xo[WS(os, 116)]), VFMAI(TcQ, TcP), ovs, &(xo[0]));
1078
ST(&(xo[WS(os, 12)]), VFNMSI(TcQ, TcP), ovs, &(xo[0]));
1079
ST(&(xo[WS(os, 52)]), VFMAI(TcO, TcH), ovs, &(xo[0]));
1080
ST(&(xo[WS(os, 76)]), VFNMSI(TcO, TcH), ovs, &(xo[0]));
1081
Tg0 = VFNMS(LDK(KP980785280), TfZ, TfY);
1082
Tga = VFMA(LDK(KP980785280), TfZ, TfY);
1083
TfF = VFNMS(LDK(KP980785280), TeE, Tel);
1084
TeF = VFMA(LDK(KP980785280), TeE, Tel);
1085
TfO = VADD(Tfw, Tfv);
1086
Tfx = VSUB(Tfv, Tfw);
1087
TeZ = VFMA(LDK(KP923879532), TeY, TeV);
1088
TfR = VFNMS(LDK(KP923879532), TeY, TeV);
1089
TfT = VFNMS(LDK(KP923879532), Tfc, Tf5);
1090
Tfd = VFMA(LDK(KP923879532), Tfc, Tf5);
1091
Tfk = VFMA(LDK(KP923879532), Tfj, Tfg);
1092
TfU = VFNMS(LDK(KP923879532), Tfj, Tfg);
1093
TfP = VFMA(LDK(KP980785280), TfO, TfN);
1094
Tg7 = VFNMS(LDK(KP980785280), TfO, TfN);
1095
TfI = VFNMS(LDK(KP980785280), Tfx, Tfu);
1096
Tfy = VFMA(LDK(KP980785280), Tfx, Tfu);
1097
TfA = VFMA(LDK(KP098491403), TeS, TeZ);
1098
Tf0 = VFNMS(LDK(KP098491403), TeZ, TeS);
1099
Tfz = VFMA(LDK(KP098491403), Tfd, Tfk);
1100
Tfl = VFNMS(LDK(KP098491403), Tfk, Tfd);
1101
Tg2 = VFNMS(LDK(KP820678790), TfQ, TfR);
1102
TfS = VFMA(LDK(KP820678790), TfR, TfQ);
1109
V T8x, T8y, T8F, T8w, T8k, T8f, T8n, T80, T9l, T76, T87, T8U, T89, T7e, T7l;
1112
V The, Tho, TgT, Tgp, Th7, Th8, Thg, Th6, Th3, Thl, TgW, TgM, TgU, TgP, TgX;
1115
V Th1, TgI, TgK, TgJ;
1117
V Tgh, Thc, Tgk, TfG, TfB, TfJ, Tfm, Tg1, TfV, Tgn, TfL, TfH;
1118
Th1 = VFMA(LDK(KP923879532), Tgg, Tgf);
1119
Tgh = VFNMS(LDK(KP923879532), Tgg, Tgf);
1120
Thc = VFNMS(LDK(KP923879532), TgH, TgG);
1121
TgI = VFMA(LDK(KP923879532), TgH, TgG);
1122
TgK = VFMA(LDK(KP668178637), Tgi, Tgj);
1123
Tgk = VFNMS(LDK(KP668178637), Tgj, Tgi);
1124
TfG = VADD(TfA, Tfz);
1125
TfB = VSUB(Tfz, TfA);
1126
TfJ = VSUB(Tfl, Tf0);
1127
Tfm = VADD(Tf0, Tfl);
1128
Tg1 = VFNMS(LDK(KP820678790), TfT, TfU);
1129
TfV = VFMA(LDK(KP820678790), TfU, TfT);
1130
Tgn = VFNMS(LDK(KP668178637), Tgm, Tgl);
1131
TgJ = VFMA(LDK(KP668178637), Tgl, Tgm);
1132
TfL = VFMA(LDK(KP995184726), TfG, TfF);
1133
TfH = VFNMS(LDK(KP995184726), TfG, TfF);
1135
V TfE, TfC, TfM, TfK;
1136
TfE = VFMA(LDK(KP995184726), TfB, Tfy);
1137
TfC = VFNMS(LDK(KP995184726), TfB, Tfy);
1138
TfM = VFNMS(LDK(KP995184726), TfJ, TfI);
1139
TfK = VFMA(LDK(KP995184726), TfJ, TfI);
1141
V TfD, Tfn, Tg8, Tg3;
1142
TfD = VFMA(LDK(KP995184726), Tfm, TeF);
1143
Tfn = VFNMS(LDK(KP995184726), Tfm, TeF);
1144
Tg8 = VADD(Tg2, Tg1);
1145
Tg3 = VSUB(Tg1, Tg2);
1147
V Tgb, TfW, Thd, Tgo;
1148
Tgb = VSUB(TfV, TfS);
1149
TfW = VADD(TfS, TfV);
1150
Thd = VSUB(Tgn, Tgk);
1151
Tgo = VADD(Tgk, Tgn);
1152
ST(&(xo[WS(os, 98)]), VFMAI(TfM, TfL), ovs, &(xo[0]));
1153
ST(&(xo[WS(os, 30)]), VFNMSI(TfM, TfL), ovs, &(xo[0]));
1154
ST(&(xo[WS(os, 94)]), VFNMSI(TfK, TfH), ovs, &(xo[0]));
1155
ST(&(xo[WS(os, 34)]), VFMAI(TfK, TfH), ovs, &(xo[0]));
1156
ST(&(xo[WS(os, 2)]), VFMAI(TfE, TfD), ovs, &(xo[0]));
1157
ST(&(xo[WS(os, 126)]), VFNMSI(TfE, TfD), ovs, &(xo[0]));
1158
ST(&(xo[WS(os, 66)]), VFMAI(TfC, Tfn), ovs, &(xo[0]));
1159
ST(&(xo[WS(os, 62)]), VFNMSI(TfC, Tfn), ovs, &(xo[0]));
1161
V Tgd, Tg9, Tg6, Tg4;
1162
Tgd = VFNMS(LDK(KP773010453), Tg8, Tg7);
1163
Tg9 = VFMA(LDK(KP773010453), Tg8, Tg7);
1164
Tg6 = VFMA(LDK(KP773010453), Tg3, Tg0);
1165
Tg4 = VFNMS(LDK(KP773010453), Tg3, Tg0);
1167
V Tge, Tgc, Tg5, TfX;
1168
Tge = VFMA(LDK(KP773010453), Tgb, Tga);
1169
Tgc = VFNMS(LDK(KP773010453), Tgb, Tga);
1170
Tg5 = VFMA(LDK(KP773010453), TfW, TfP);
1171
TfX = VFNMS(LDK(KP773010453), TfW, TfP);
1172
The = VFMA(LDK(KP831469612), Thd, Thc);
1173
Tho = VFNMS(LDK(KP831469612), Thd, Thc);
1174
TgT = VFMA(LDK(KP831469612), Tgo, Tgh);
1175
Tgp = VFNMS(LDK(KP831469612), Tgo, Tgh);
1176
ST(&(xo[WS(os, 110)]), VFNMSI(Tge, Tgd), ovs, &(xo[0]));
1177
ST(&(xo[WS(os, 18)]), VFMAI(Tge, Tgd), ovs, &(xo[0]));
1178
ST(&(xo[WS(os, 82)]), VFMAI(Tgc, Tg9), ovs, &(xo[0]));
1179
ST(&(xo[WS(os, 46)]), VFNMSI(Tgc, Tg9), ovs, &(xo[0]));
1180
ST(&(xo[WS(os, 114)]), VFMAI(Tg6, Tg5), ovs, &(xo[0]));
1181
ST(&(xo[WS(os, 14)]), VFNMSI(Tg6, Tg5), ovs, &(xo[0]));
1182
ST(&(xo[WS(os, 50)]), VFMAI(Tg4, TfX), ovs, &(xo[0]));
1183
ST(&(xo[WS(os, 78)]), VFNMSI(Tg4, TfX), ovs, &(xo[0]));
1191
V Th4, Tgs, Tgv, Th5, Tgz, TgC, Th2, TgL;
1192
Th4 = VFMA(LDK(KP923879532), Tgr, Tgq);
1193
Tgs = VFNMS(LDK(KP923879532), Tgr, Tgq);
1194
Tgv = VFMA(LDK(KP923879532), Tgu, Tgt);
1195
Th5 = VFNMS(LDK(KP923879532), Tgu, Tgt);
1196
Th7 = VFMA(LDK(KP923879532), Tgy, Tgx);
1197
Tgz = VFNMS(LDK(KP923879532), Tgy, Tgx);
1198
TgC = VFMA(LDK(KP923879532), TgB, TgA);
1199
Th8 = VFNMS(LDK(KP923879532), TgB, TgA);
1200
Th2 = VADD(TgK, TgJ);
1201
TgL = VSUB(TgJ, TgK);
1203
V TgO, Tgw, TgN, TgD;
1204
TgO = VFMA(LDK(KP534511135), Tgs, Tgv);
1205
Tgw = VFNMS(LDK(KP534511135), Tgv, Tgs);
1206
TgN = VFMA(LDK(KP534511135), Tgz, TgC);
1207
TgD = VFNMS(LDK(KP534511135), TgC, Tgz);
1208
Thg = VFNMS(LDK(KP303346683), Th4, Th5);
1209
Th6 = VFMA(LDK(KP303346683), Th5, Th4);
1210
Th3 = VFMA(LDK(KP831469612), Th2, Th1);
1211
Thl = VFNMS(LDK(KP831469612), Th2, Th1);
1212
TgW = VFNMS(LDK(KP831469612), TgL, TgI);
1213
TgM = VFMA(LDK(KP831469612), TgL, TgI);
1214
TgU = VADD(TgO, TgN);
1215
TgP = VSUB(TgN, TgO);
1216
TgX = VSUB(TgD, Tgw);
1217
TgE = VADD(Tgw, TgD);
1222
V T8u, T8v, T7R, T8d, T7G, Thm, Thh, Thp, Tha, T7Y, Thr, Thn;
1224
V T7y, T7F, TgZ, TgV;
1225
T8u = VFNMS(LDK(KP831469612), T7x, T7q);
1226
T7y = VFMA(LDK(KP831469612), T7x, T7q);
1227
T7F = VFMA(LDK(KP831469612), T7E, T7B);
1228
T8v = VFNMS(LDK(KP831469612), T7E, T7B);
1229
T8x = VFNMS(LDK(KP831469612), T7Q, T7J);
1230
T7R = VFMA(LDK(KP831469612), T7Q, T7J);
1231
TgZ = VFMA(LDK(KP881921264), TgU, TgT);
1232
TgV = VFNMS(LDK(KP881921264), TgU, TgT);
1234
V TgS, TgQ, Th0, TgY;
1235
TgS = VFMA(LDK(KP881921264), TgP, TgM);
1236
TgQ = VFNMS(LDK(KP881921264), TgP, TgM);
1237
Th0 = VFNMS(LDK(KP881921264), TgX, TgW);
1238
TgY = VFMA(LDK(KP881921264), TgX, TgW);
1240
V TgR, TgF, Thf, Th9;
1241
TgR = VFMA(LDK(KP881921264), TgE, Tgp);
1242
TgF = VFNMS(LDK(KP881921264), TgE, Tgp);
1243
Thf = VFNMS(LDK(KP303346683), Th7, Th8);
1244
Th9 = VFMA(LDK(KP303346683), Th8, Th7);
1245
T8d = VFNMS(LDK(KP148335987), T7y, T7F);
1246
T7G = VFMA(LDK(KP148335987), T7F, T7y);
1247
ST(&(xo[WS(os, 106)]), VFMAI(Th0, TgZ), ovs, &(xo[0]));
1248
ST(&(xo[WS(os, 22)]), VFNMSI(Th0, TgZ), ovs, &(xo[0]));
1249
ST(&(xo[WS(os, 86)]), VFNMSI(TgY, TgV), ovs, &(xo[0]));
1250
ST(&(xo[WS(os, 42)]), VFMAI(TgY, TgV), ovs, &(xo[0]));
1251
ST(&(xo[WS(os, 10)]), VFMAI(TgS, TgR), ovs, &(xo[0]));
1252
ST(&(xo[WS(os, 118)]), VFNMSI(TgS, TgR), ovs, &(xo[0]));
1253
ST(&(xo[WS(os, 74)]), VFMAI(TgQ, TgF), ovs, &(xo[0]));
1254
ST(&(xo[WS(os, 54)]), VFNMSI(TgQ, TgF), ovs, &(xo[0]));
1255
Thm = VADD(Thg, Thf);
1256
Thh = VSUB(Thf, Thg);
1257
Thp = VSUB(Th9, Th6);
1258
Tha = VADD(Th6, Th9);
1259
T7Y = VFMA(LDK(KP831469612), T7X, T7U);
1260
T8y = VFNMS(LDK(KP831469612), T7X, T7U);
1264
Thr = VFNMS(LDK(KP956940335), Thm, Thl);
1265
Thn = VFMA(LDK(KP956940335), Thm, Thl);
1267
V Thk, Thi, Ths, Thq;
1268
Thk = VFMA(LDK(KP956940335), Thh, The);
1269
Thi = VFNMS(LDK(KP956940335), Thh, The);
1270
Ths = VFMA(LDK(KP956940335), Thp, Tho);
1271
Thq = VFNMS(LDK(KP956940335), Thp, Tho);
1273
V Thj, Thb, T8e, T7Z;
1274
Thj = VFMA(LDK(KP956940335), Tha, Th3);
1275
Thb = VFNMS(LDK(KP956940335), Tha, Th3);
1276
T8e = VFNMS(LDK(KP148335987), T7R, T7Y);
1277
T7Z = VFMA(LDK(KP148335987), T7Y, T7R);
1278
T8F = VFMA(LDK(KP741650546), T8u, T8v);
1279
T8w = VFNMS(LDK(KP741650546), T8v, T8u);
1280
ST(&(xo[WS(os, 102)]), VFNMSI(Ths, Thr), ovs, &(xo[0]));
1281
ST(&(xo[WS(os, 26)]), VFMAI(Ths, Thr), ovs, &(xo[0]));
1282
ST(&(xo[WS(os, 90)]), VFMAI(Thq, Thn), ovs, &(xo[0]));
1283
ST(&(xo[WS(os, 38)]), VFNMSI(Thq, Thn), ovs, &(xo[0]));
1284
ST(&(xo[WS(os, 122)]), VFMAI(Thk, Thj), ovs, &(xo[0]));
1285
ST(&(xo[WS(os, 6)]), VFNMSI(Thk, Thj), ovs, &(xo[0]));
1286
ST(&(xo[WS(os, 58)]), VFMAI(Thi, Thb), ovs, &(xo[0]));
1287
ST(&(xo[WS(os, 70)]), VFNMSI(Thi, Thb), ovs, &(xo[0]));
1288
T8k = VADD(T8d, T8e);
1289
T8f = VSUB(T8d, T8e);
1290
T8n = VSUB(T7Z, T7G);
1291
T80 = VADD(T7G, T7Z);
1294
T9l = VSUB(T75, T72);
1295
T76 = VADD(T72, T75);
1296
T87 = VSUB(T85, T86);
1297
T8U = VADD(T85, T86);
1298
T89 = VFNMS(LDK(KP303346683), T7a, T7d);
1299
T7e = VFMA(LDK(KP303346683), T7d, T7a);
1300
T7l = VFMA(LDK(KP303346683), T7k, T7h);
1301
T8a = VFNMS(LDK(KP303346683), T7h, T7k);
1305
V T11, T5h, T5a, T55, T5d, T4K, T5C, T5x, T5F, T5q, T4X, T4Z, T1C, T2d, T50;
1307
V T5k, T3g, T3t, T5l, T5n, T4v, T4I, T5o, T8G, T8z;
1308
T5k = VFNMS(LDK(KP980785280), T3f, T2G);
1309
T3g = VFMA(LDK(KP980785280), T3f, T2G);
1310
T8G = VFMA(LDK(KP741650546), T8x, T8y);
1311
T8z = VFNMS(LDK(KP741650546), T8y, T8x);
1313
V T8r, T77, T8C, T88;
1314
T8r = VFNMS(LDK(KP831469612), T76, T6Z);
1315
T77 = VFMA(LDK(KP831469612), T76, T6Z);
1316
T8C = VFNMS(LDK(KP831469612), T87, T84);
1317
T88 = VFMA(LDK(KP831469612), T87, T84);
1319
V T8D, T7m, T8s, T8b;
1320
T8D = VSUB(T7l, T7e);
1321
T7m = VADD(T7e, T7l);
1322
T8s = VADD(T89, T8a);
1323
T8b = VSUB(T89, T8a);
1325
V T8M, T8H, T8P, T8A;
1326
T8M = VADD(T8F, T8G);
1327
T8H = VSUB(T8F, T8G);
1328
T8P = VSUB(T8z, T8w);
1329
T8A = VADD(T8w, T8z);
1331
V T8E, T8O, T8j, T7n;
1332
T8E = VFNMS(LDK(KP956940335), T8D, T8C);
1333
T8O = VFMA(LDK(KP956940335), T8D, T8C);
1334
T8j = VFNMS(LDK(KP956940335), T7m, T77);
1335
T7n = VFMA(LDK(KP956940335), T7m, T77);
1337
V T8t, T8L, T8m, T8c;
1338
T8t = VFNMS(LDK(KP956940335), T8s, T8r);
1339
T8L = VFMA(LDK(KP956940335), T8s, T8r);
1340
T8m = VFNMS(LDK(KP956940335), T8b, T88);
1341
T8c = VFMA(LDK(KP956940335), T8b, T88);
1343
V T8K, T8I, T8S, T8Q;
1344
T8K = VFMA(LDK(KP803207531), T8H, T8E);
1345
T8I = VFNMS(LDK(KP803207531), T8H, T8E);
1346
T8S = VFMA(LDK(KP803207531), T8P, T8O);
1347
T8Q = VFNMS(LDK(KP803207531), T8P, T8O);
1349
V T8p, T8l, T8h, T81;
1350
T8p = VFNMS(LDK(KP989176509), T8k, T8j);
1351
T8l = VFMA(LDK(KP989176509), T8k, T8j);
1352
T8h = VFMA(LDK(KP989176509), T80, T7n);
1353
T81 = VFNMS(LDK(KP989176509), T80, T7n);
1355
V T8J, T8B, T8R, T8N;
1356
T8J = VFMA(LDK(KP803207531), T8A, T8t);
1357
T8B = VFNMS(LDK(KP803207531), T8A, T8t);
1358
T8R = VFMA(LDK(KP803207531), T8M, T8L);
1359
T8N = VFNMS(LDK(KP803207531), T8M, T8L);
1361
V T8q, T8o, T8i, T8g;
1362
T8q = VFNMS(LDK(KP989176509), T8n, T8m);
1363
T8o = VFMA(LDK(KP989176509), T8n, T8m);
1364
T8i = VFMA(LDK(KP989176509), T8f, T8c);
1365
T8g = VFNMS(LDK(KP989176509), T8f, T8c);
1366
ST(&(xo[WS(os, 115)]), VFMAI(T8K, T8J), ovs, &(xo[WS(os, 1)]));
1367
ST(&(xo[WS(os, 13)]), VFNMSI(T8K, T8J), ovs, &(xo[WS(os, 1)]));
1368
ST(&(xo[WS(os, 51)]), VFMAI(T8I, T8B), ovs, &(xo[WS(os, 1)]));
1369
ST(&(xo[WS(os, 77)]), VFNMSI(T8I, T8B), ovs, &(xo[WS(os, 1)]));
1370
ST(&(xo[WS(os, 109)]), VFNMSI(T8S, T8R), ovs, &(xo[WS(os, 1)]));
1371
ST(&(xo[WS(os, 19)]), VFMAI(T8S, T8R), ovs, &(xo[WS(os, 1)]));
1372
ST(&(xo[WS(os, 83)]), VFMAI(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
1373
ST(&(xo[WS(os, 45)]), VFNMSI(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
1374
ST(&(xo[WS(os, 99)]), VFMAI(T8q, T8p), ovs, &(xo[WS(os, 1)]));
1375
ST(&(xo[WS(os, 29)]), VFNMSI(T8q, T8p), ovs, &(xo[WS(os, 1)]));
1376
ST(&(xo[WS(os, 93)]), VFNMSI(T8o, T8l), ovs, &(xo[WS(os, 1)]));
1377
ST(&(xo[WS(os, 35)]), VFMAI(T8o, T8l), ovs, &(xo[WS(os, 1)]));
1378
ST(&(xo[WS(os, 3)]), VFMAI(T8i, T8h), ovs, &(xo[WS(os, 1)]));
1379
ST(&(xo[WS(os, 125)]), VFNMSI(T8i, T8h), ovs, &(xo[WS(os, 1)]));
1380
ST(&(xo[WS(os, 67)]), VFMAI(T8g, T81), ovs, &(xo[WS(os, 1)]));
1381
ST(&(xo[WS(os, 61)]), VFNMSI(T8g, T81), ovs, &(xo[WS(os, 1)]));
1382
T3t = VFMA(LDK(KP980785280), T3s, T3p);
1383
T5l = VFNMS(LDK(KP980785280), T3s, T3p);
1393
T5n = VFNMS(LDK(KP980785280), T4u, T3V);
1394
T4v = VFMA(LDK(KP980785280), T4u, T3V);
1395
T4I = VFMA(LDK(KP980785280), T4H, T4E);
1396
T5o = VFNMS(LDK(KP980785280), T4H, T4E);
1398
V T53, T3u, T54, T4J, T5v, T5m, T5w, T5p, T10;
1401
T53 = VFMA(LDK(KP049126849), T3g, T3t);
1402
T3u = VFNMS(LDK(KP049126849), T3t, T3g);
1403
T54 = VFMA(LDK(KP049126849), T4v, T4I);
1404
T4J = VFNMS(LDK(KP049126849), T4I, T4v);
1405
T5v = VFNMS(LDK(KP906347169), T5k, T5l);
1406
T5m = VFMA(LDK(KP906347169), T5l, T5k);
1407
T5w = VFNMS(LDK(KP906347169), T5n, T5o);
1408
T5p = VFMA(LDK(KP906347169), T5o, T5n);
1409
T11 = VFMA(LDK(KP980785280), T10, Tr);
1410
T5h = VFNMS(LDK(KP980785280), T10, Tr);
1411
T5a = VADD(T53, T54);
1412
T55 = VSUB(T53, T54);
1413
T5d = VSUB(T4J, T3u);
1414
T4K = VADD(T3u, T4J);
1415
T5C = VADD(T5v, T5w);
1416
T5x = VSUB(T5v, T5w);
1417
T5F = VSUB(T5p, T5m);
1418
T5q = VADD(T5m, T5p);
1419
T4X = VSUB(T4V, T4W);
1420
T5K = VADD(T4V, T4W);
1422
T4Z = VFMA(LDK(KP098491403), T1s, T1B);
1423
T1C = VFNMS(LDK(KP098491403), T1B, T1s);
1424
T2d = VFNMS(LDK(KP098491403), T2c, T23);
1425
T50 = VFMA(LDK(KP098491403), T23, T2c);
1428
V T9y, T9t, T9B, T9i, T9o, T9n, T9F, T8V, T9Q, T9m, T9R, T92, Ta0, T9V, Ta3;
1431
V T9I, T9J, T9L, T9d, T5s, T4Y, T5t, T2e, T5i, T51, T9r, T9a, T9g, T9M, T96;
1433
T9I = VFMA(LDK(KP831469612), T95, T94);
1434
T96 = VFNMS(LDK(KP831469612), T95, T94);
1435
T99 = VFNMS(LDK(KP831469612), T98, T97);
1436
T9J = VFMA(LDK(KP831469612), T98, T97);
1437
T9L = VFMA(LDK(KP831469612), T9c, T9b);
1438
T9d = VFNMS(LDK(KP831469612), T9c, T9b);
1439
T5s = VFNMS(LDK(KP980785280), T4X, T4U);
1440
T4Y = VFMA(LDK(KP980785280), T4X, T4U);
1441
T5t = VSUB(T2d, T1C);
1442
T2e = VADD(T1C, T2d);
1443
T5i = VADD(T4Z, T50);
1444
T51 = VSUB(T4Z, T50);
1445
T9r = VFNMS(LDK(KP599376933), T96, T99);
1446
T9a = VFMA(LDK(KP599376933), T99, T96);
1447
T9g = VFNMS(LDK(KP831469612), T9f, T9e);
1448
T9M = VFMA(LDK(KP831469612), T9f, T9e);
1450
V T5u, T5E, T8Y, T91;
1451
T5u = VFMA(LDK(KP995184726), T5t, T5s);
1452
T5E = VFNMS(LDK(KP995184726), T5t, T5s);
1454
V T59, T2f, T5j, T5B;
1455
T59 = VFNMS(LDK(KP995184726), T2e, T11);
1456
T2f = VFMA(LDK(KP995184726), T2e, T11);
1457
T5j = VFMA(LDK(KP995184726), T5i, T5h);
1458
T5B = VFNMS(LDK(KP995184726), T5i, T5h);
1460
V T5c, T52, T9s, T9h;
1461
T5c = VFNMS(LDK(KP995184726), T51, T4Y);
1462
T52 = VFMA(LDK(KP995184726), T51, T4Y);
1463
T9s = VFNMS(LDK(KP599376933), T9d, T9g);
1464
T9h = VFMA(LDK(KP599376933), T9g, T9d);
1466
V T5A, T5y, T5I, T5G;
1467
T5A = VFMA(LDK(KP740951125), T5x, T5u);
1468
T5y = VFNMS(LDK(KP740951125), T5x, T5u);
1469
T5I = VFNMS(LDK(KP740951125), T5F, T5E);
1470
T5G = VFMA(LDK(KP740951125), T5F, T5E);
1472
V T5f, T5b, T57, T4L;
1473
T5f = VFMA(LDK(KP998795456), T5a, T59);
1474
T5b = VFNMS(LDK(KP998795456), T5a, T59);
1475
T57 = VFMA(LDK(KP998795456), T4K, T2f);
1476
T4L = VFNMS(LDK(KP998795456), T4K, T2f);
1478
V T5z, T5r, T5H, T5D;
1479
T5z = VFMA(LDK(KP740951125), T5q, T5j);
1480
T5r = VFNMS(LDK(KP740951125), T5q, T5j);
1481
T5H = VFNMS(LDK(KP740951125), T5C, T5B);
1482
T5D = VFMA(LDK(KP740951125), T5C, T5B);
1484
V T5g, T5e, T58, T56;
1485
T5g = VFMA(LDK(KP998795456), T5d, T5c);
1486
T5e = VFNMS(LDK(KP998795456), T5d, T5c);
1487
T58 = VFMA(LDK(KP998795456), T55, T52);
1488
T56 = VFNMS(LDK(KP998795456), T55, T52);
1489
T9y = VADD(T9r, T9s);
1490
T9t = VSUB(T9r, T9s);
1491
T9B = VSUB(T9h, T9a);
1492
T9i = VADD(T9a, T9h);
1493
ST(&(xo[WS(os, 15)]), VFMAI(T5A, T5z), ovs, &(xo[WS(os, 1)]));
1494
ST(&(xo[WS(os, 113)]), VFNMSI(T5A, T5z), ovs, &(xo[WS(os, 1)]));
1495
ST(&(xo[WS(os, 79)]), VFMAI(T5y, T5r), ovs, &(xo[WS(os, 1)]));
1496
ST(&(xo[WS(os, 49)]), VFNMSI(T5y, T5r), ovs, &(xo[WS(os, 1)]));
1497
ST(&(xo[WS(os, 111)]), VFMAI(T5I, T5H), ovs, &(xo[WS(os, 1)]));
1498
ST(&(xo[WS(os, 17)]), VFNMSI(T5I, T5H), ovs, &(xo[WS(os, 1)]));
1499
ST(&(xo[WS(os, 81)]), VFNMSI(T5G, T5D), ovs, &(xo[WS(os, 1)]));
1500
ST(&(xo[WS(os, 47)]), VFMAI(T5G, T5D), ovs, &(xo[WS(os, 1)]));
1501
ST(&(xo[WS(os, 97)]), VFNMSI(T5g, T5f), ovs, &(xo[WS(os, 1)]));
1502
ST(&(xo[WS(os, 31)]), VFMAI(T5g, T5f), ovs, &(xo[WS(os, 1)]));
1503
ST(&(xo[WS(os, 95)]), VFMAI(T5e, T5b), ovs, &(xo[WS(os, 1)]));
1504
ST(&(xo[WS(os, 33)]), VFNMSI(T5e, T5b), ovs, &(xo[WS(os, 1)]));
1505
ST(&(xo[WS(os, 127)]), VFMAI(T58, T57), ovs, &(xo[WS(os, 1)]));
1506
ST(&(xo[WS(os, 1)]), VFNMSI(T58, T57), ovs, &(xo[WS(os, 1)]));
1507
ST(&(xo[WS(os, 63)]), VFMAI(T56, T4L), ovs, &(xo[WS(os, 1)]));
1508
ST(&(xo[WS(os, 65)]), VFNMSI(T56, T4L), ovs, &(xo[WS(os, 1)]));
1515
T9o = VFNMS(LDK(KP534511135), T8W, T8X);
1516
T8Y = VFMA(LDK(KP534511135), T8X, T8W);
1517
T91 = VFMA(LDK(KP534511135), T90, T8Z);
1518
T9n = VFNMS(LDK(KP534511135), T8Z, T90);
1520
V T9T, T9K, T9U, T9N;
1521
T9T = VFMA(LDK(KP250486960), T9I, T9J);
1522
T9K = VFNMS(LDK(KP250486960), T9J, T9I);
1523
T9U = VFMA(LDK(KP250486960), T9L, T9M);
1524
T9N = VFNMS(LDK(KP250486960), T9M, T9L);
1525
T9F = VFNMS(LDK(KP831469612), T8U, T8T);
1526
T8V = VFMA(LDK(KP831469612), T8U, T8T);
1527
T9Q = VFNMS(LDK(KP831469612), T9l, T9k);
1528
T9m = VFMA(LDK(KP831469612), T9l, T9k);
1529
T9R = VSUB(T8Y, T91);
1530
T92 = VADD(T8Y, T91);
1531
Ta0 = VADD(T9T, T9U);
1532
T9V = VSUB(T9T, T9U);
1533
Ta3 = VSUB(T9N, T9K);
1534
T9O = VADD(T9K, T9N);
1539
V T6y, T6z, T63, T9Y, T9W, Ta6, Ta4, T9D, T9z, T9v, T9j, T6h, T60, T9H, T9Z;
1540
V T9A, T9q, T66, T9X, T9P;
1542
V T5W, T9S, Ta2, T9x, T93, T5Z, T9G, T9p;
1543
T6y = VFMA(LDK(KP980785280), T5V, T5U);
1544
T5W = VFNMS(LDK(KP980785280), T5V, T5U);
1545
T9S = VFMA(LDK(KP881921264), T9R, T9Q);
1546
Ta2 = VFNMS(LDK(KP881921264), T9R, T9Q);
1547
T9x = VFNMS(LDK(KP881921264), T92, T8V);
1548
T93 = VFMA(LDK(KP881921264), T92, T8V);
1549
T5Z = VFMA(LDK(KP980785280), T5Y, T5X);
1550
T6z = VFNMS(LDK(KP980785280), T5Y, T5X);
1551
T6B = VFMA(LDK(KP980785280), T62, T61);
1552
T63 = VFNMS(LDK(KP980785280), T62, T61);
1553
T9G = VADD(T9o, T9n);
1554
T9p = VSUB(T9n, T9o);
1555
T9Y = VFMA(LDK(KP970031253), T9V, T9S);
1556
T9W = VFNMS(LDK(KP970031253), T9V, T9S);
1557
Ta6 = VFMA(LDK(KP970031253), Ta3, Ta2);
1558
Ta4 = VFNMS(LDK(KP970031253), Ta3, Ta2);
1559
T9D = VFNMS(LDK(KP857728610), T9y, T9x);
1560
T9z = VFMA(LDK(KP857728610), T9y, T9x);
1561
T9v = VFMA(LDK(KP857728610), T9i, T93);
1562
T9j = VFNMS(LDK(KP857728610), T9i, T93);
1563
T6h = VFMA(LDK(KP472964775), T5W, T5Z);
1564
T60 = VFNMS(LDK(KP472964775), T5Z, T5W);
1565
T9H = VFMA(LDK(KP881921264), T9G, T9F);
1566
T9Z = VFNMS(LDK(KP881921264), T9G, T9F);
1567
T9A = VFNMS(LDK(KP881921264), T9p, T9m);
1568
T9q = VFMA(LDK(KP881921264), T9p, T9m);
1569
T66 = VFMA(LDK(KP980785280), T65, T64);
1570
T6C = VFNMS(LDK(KP980785280), T65, T64);
1572
T9X = VFMA(LDK(KP970031253), T9O, T9H);
1573
T9P = VFNMS(LDK(KP970031253), T9O, T9H);
1575
V Ta5, Ta1, T9E, T9C;
1576
Ta5 = VFMA(LDK(KP970031253), Ta0, T9Z);
1577
Ta1 = VFNMS(LDK(KP970031253), Ta0, T9Z);
1578
T9E = VFNMS(LDK(KP857728610), T9B, T9A);
1579
T9C = VFMA(LDK(KP857728610), T9B, T9A);
1581
V T9w, T9u, T6i, T67;
1582
T9w = VFMA(LDK(KP857728610), T9t, T9q);
1583
T9u = VFNMS(LDK(KP857728610), T9t, T9q);
1584
T6i = VFMA(LDK(KP472964775), T63, T66);
1585
T67 = VFNMS(LDK(KP472964775), T66, T63);
1586
T6J = VFNMS(LDK(KP357805721), T6y, T6z);
1587
T6A = VFMA(LDK(KP357805721), T6z, T6y);
1588
ST(&(xo[WS(os, 123)]), VFMAI(T9Y, T9X), ovs, &(xo[WS(os, 1)]));
1589
ST(&(xo[WS(os, 5)]), VFNMSI(T9Y, T9X), ovs, &(xo[WS(os, 1)]));
1590
ST(&(xo[WS(os, 59)]), VFMAI(T9W, T9P), ovs, &(xo[WS(os, 1)]));
1591
ST(&(xo[WS(os, 69)]), VFNMSI(T9W, T9P), ovs, &(xo[WS(os, 1)]));
1592
ST(&(xo[WS(os, 101)]), VFNMSI(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
1593
ST(&(xo[WS(os, 27)]), VFMAI(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
1594
ST(&(xo[WS(os, 91)]), VFMAI(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
1595
ST(&(xo[WS(os, 37)]), VFNMSI(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
1596
ST(&(xo[WS(os, 107)]), VFMAI(T9E, T9D), ovs, &(xo[WS(os, 1)]));
1597
ST(&(xo[WS(os, 21)]), VFNMSI(T9E, T9D), ovs, &(xo[WS(os, 1)]));
1598
ST(&(xo[WS(os, 85)]), VFNMSI(T9C, T9z), ovs, &(xo[WS(os, 1)]));
1599
ST(&(xo[WS(os, 43)]), VFMAI(T9C, T9z), ovs, &(xo[WS(os, 1)]));
1600
ST(&(xo[WS(os, 11)]), VFMAI(T9w, T9v), ovs, &(xo[WS(os, 1)]));
1601
ST(&(xo[WS(os, 117)]), VFNMSI(T9w, T9v), ovs, &(xo[WS(os, 1)]));
1602
ST(&(xo[WS(os, 75)]), VFMAI(T9u, T9j), ovs, &(xo[WS(os, 1)]));
1603
ST(&(xo[WS(os, 53)]), VFNMSI(T9u, T9j), ovs, &(xo[WS(os, 1)]));
1604
T6o = VADD(T6h, T6i);
1605
T6j = VSUB(T6h, T6i);
1606
T6r = VSUB(T67, T60);
1607
T68 = VADD(T60, T67);
1610
T6e = VFMA(LDK(KP820678790), T5M, T5N);
1611
T5O = VFNMS(LDK(KP820678790), T5N, T5M);
1612
T5R = VFNMS(LDK(KP820678790), T5Q, T5P);
1613
T6d = VFMA(LDK(KP820678790), T5P, T5Q);
1620
T6D = VFMA(LDK(KP357805721), T6C, T6B);
1621
T6K = VFNMS(LDK(KP357805721), T6B, T6C);
1623
V T5L, T6v, T6c, T6G;
1624
T5L = VFNMS(LDK(KP980785280), T5K, T5J);
1625
T6v = VFMA(LDK(KP980785280), T5K, T5J);
1626
T6c = VFNMS(LDK(KP980785280), T6b, T6a);
1627
T6G = VFMA(LDK(KP980785280), T6b, T6a);
1629
V T5S, T6H, T6f, T6w;
1630
T5S = VADD(T5O, T5R);
1631
T6H = VSUB(T5O, T5R);
1632
T6f = VSUB(T6d, T6e);
1633
T6w = VADD(T6e, T6d);
1635
V T6L, T6Q, T6E, T6T;
1636
T6L = VSUB(T6J, T6K);
1637
T6Q = VADD(T6J, T6K);
1638
T6E = VADD(T6A, T6D);
1639
T6T = VSUB(T6D, T6A);
1641
V T6S, T6I, T5T, T6n;
1642
T6S = VFNMS(LDK(KP773010453), T6H, T6G);
1643
T6I = VFMA(LDK(KP773010453), T6H, T6G);
1644
T5T = VFNMS(LDK(KP773010453), T5S, T5L);
1645
T6n = VFMA(LDK(KP773010453), T5S, T5L);
1647
V T6P, T6x, T6g, T6q;
1648
T6P = VFNMS(LDK(KP773010453), T6w, T6v);
1649
T6x = VFMA(LDK(KP773010453), T6w, T6v);
1650
T6g = VFNMS(LDK(KP773010453), T6f, T6c);
1651
T6q = VFMA(LDK(KP773010453), T6f, T6c);
1653
V T6M, T6O, T6U, T6W;
1654
T6M = VFNMS(LDK(KP941544065), T6L, T6I);
1655
T6O = VFMA(LDK(KP941544065), T6L, T6I);
1656
T6U = VFMA(LDK(KP941544065), T6T, T6S);
1657
T6W = VFNMS(LDK(KP941544065), T6T, T6S);
1659
V T6p, T6t, T69, T6l;
1660
T6p = VFNMS(LDK(KP903989293), T6o, T6n);
1661
T6t = VFMA(LDK(KP903989293), T6o, T6n);
1662
T69 = VFNMS(LDK(KP903989293), T68, T5T);
1663
T6l = VFMA(LDK(KP903989293), T68, T5T);
1665
V T6F, T6N, T6R, T6V;
1666
T6F = VFNMS(LDK(KP941544065), T6E, T6x);
1667
T6N = VFMA(LDK(KP941544065), T6E, T6x);
1668
T6R = VFMA(LDK(KP941544065), T6Q, T6P);
1669
T6V = VFNMS(LDK(KP941544065), T6Q, T6P);
1671
V T6s, T6u, T6k, T6m;
1672
T6s = VFNMS(LDK(KP903989293), T6r, T6q);
1673
T6u = VFMA(LDK(KP903989293), T6r, T6q);
1674
T6k = VFNMS(LDK(KP903989293), T6j, T6g);
1675
T6m = VFMA(LDK(KP903989293), T6j, T6g);
1676
ST(&(xo[WS(os, 7)]), VFMAI(T6O, T6N), ovs, &(xo[WS(os, 1)]));
1677
ST(&(xo[WS(os, 121)]), VFNMSI(T6O, T6N), ovs, &(xo[WS(os, 1)]));
1678
ST(&(xo[WS(os, 71)]), VFMAI(T6M, T6F), ovs, &(xo[WS(os, 1)]));
1679
ST(&(xo[WS(os, 57)]), VFNMSI(T6M, T6F), ovs, &(xo[WS(os, 1)]));
1680
ST(&(xo[WS(os, 103)]), VFMAI(T6W, T6V), ovs, &(xo[WS(os, 1)]));
1681
ST(&(xo[WS(os, 25)]), VFNMSI(T6W, T6V), ovs, &(xo[WS(os, 1)]));
1682
ST(&(xo[WS(os, 89)]), VFNMSI(T6U, T6R), ovs, &(xo[WS(os, 1)]));
1683
ST(&(xo[WS(os, 39)]), VFMAI(T6U, T6R), ovs, &(xo[WS(os, 1)]));
1684
ST(&(xo[WS(os, 105)]), VFNMSI(T6u, T6t), ovs, &(xo[WS(os, 1)]));
1685
ST(&(xo[WS(os, 23)]), VFMAI(T6u, T6t), ovs, &(xo[WS(os, 1)]));
1686
ST(&(xo[WS(os, 87)]), VFMAI(T6s, T6p), ovs, &(xo[WS(os, 1)]));
1687
ST(&(xo[WS(os, 41)]), VFNMSI(T6s, T6p), ovs, &(xo[WS(os, 1)]));
1688
ST(&(xo[WS(os, 119)]), VFMAI(T6m, T6l), ovs, &(xo[WS(os, 1)]));
1689
ST(&(xo[WS(os, 9)]), VFNMSI(T6m, T6l), ovs, &(xo[WS(os, 1)]));
1690
ST(&(xo[WS(os, 55)]), VFMAI(T6k, T69), ovs, &(xo[WS(os, 1)]));
1691
ST(&(xo[WS(os, 73)]), VFNMSI(T6k, T69), ovs, &(xo[WS(os, 1)]));
1704
static const kdft_desc desc = { 128, "n1fv_128", {440, 0, 642, 0}, &GENUS, 0, 0, 0, 0 };
1706
void X(codelet_n1fv_128) (planner *p) {
1707
X(kdft_register) (p, n1fv_128, &desc);
1710
#else /* HAVE_FMA */
1712
/* Generated by: ../../../genfft/gen_notw_c -simd -compact -variables 4 -pipeline-latency 8 -n 128 -name n1fv_128 -include n1f.h */
1715
* This function contains 1082 FP additions, 330 FP multiplications,
1716
* (or, 938 additions, 186 multiplications, 144 fused multiply/add),
1717
* 194 stack variables, 31 constants, and 256 memory accesses
1721
static void n1fv_128(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
1723
DVK(KP941544065, +0.941544065183020778412509402599502357185589796);
1724
DVK(KP336889853, +0.336889853392220050689253212619147570477766780);
1725
DVK(KP903989293, +0.903989293123443331586200297230537048710132025);
1726
DVK(KP427555093, +0.427555093430282094320966856888798534304578629);
1727
DVK(KP970031253, +0.970031253194543992603984207286100251456865962);
1728
DVK(KP242980179, +0.242980179903263889948274162077471118320990783);
1729
DVK(KP857728610, +0.857728610000272069902269984284770137042490799);
1730
DVK(KP514102744, +0.514102744193221726593693838968815772608049120);
1731
DVK(KP671558954, +0.671558954847018400625376850427421803228750632);
1732
DVK(KP740951125, +0.740951125354959091175616897495162729728955309);
1733
DVK(KP049067674, +0.049067674327418014254954976942682658314745363);
1734
DVK(KP998795456, +0.998795456205172392714771604759100694443203615);
1735
DVK(KP595699304, +0.595699304492433343467036528829969889511926338);
1736
DVK(KP803207531, +0.803207531480644909806676512963141923879569427);
1737
DVK(KP146730474, +0.146730474455361751658850129646717819706215317);
1738
DVK(KP989176509, +0.989176509964780973451673738016243063983689533);
1739
DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
1740
DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
1741
DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
1742
DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
1743
DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
1744
DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
1745
DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
1746
DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
1747
DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
1748
DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
1749
DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
1750
DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
1751
DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
1752
DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
1753
DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
1759
for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
1760
V Tr, T5J, Ted, Tgf, Tfq, TgH, T4U, T6b, T6Z, T8T, Tad, TcZ, Tcc, Td0, T84;
1761
V T9l, Tb6, Tbt, T2G, T5X, TeV, Tgr, T3p, T5V, T7B, T95, TeK, Tgt, T7q, T97;
1762
V Td8, TdK, TbD, Tc0, T3V, T61, Tfg, TgB, T4E, T65, T7U, T9f, Tf5, Tgx, T7J;
1763
V T9b, Tdf, TdN, Td2, Td3, TI, T4V, Tft, Tgg, TZ, T4W, T75, T86, Tek, TgG;
1764
V T72, T85, Tas, Tcd, Tdp, Tdq, TdG, Teq, Tgm, Tet, Tgl, T1s, T5P, T1B, T5Q;
1765
V T7d, T8Z, TaI, Tcf, T7a, T90, Tdm, Tdn, TdH, Tez, Tgi, TeC, Tgj, T23, T5N;
1766
V T2c, T5M, T7k, T8X, TaX, Tcg, T7h, T8W, Tbl, Tbu, Tdb, TdL, TeY, Tgu, TeR;
1767
V Tgq, T7x, T98, T7E, T94, T3f, T5Y, T3s, T5U, TbS, Tc1, Tdi, TdO, Tfj, Tgy;
1768
V Tfc, TgA, T7Q, T9e, T7X, T9c, T4u, T64, T4H, T62;
1770
V T3, Ta7, T4P, Ta8, Ta, Tab, T4M, Taa, Tc9, Tca, Ti, Tea, T4S, Tc6, Tc7;
1774
T1 = LD(&(xi[0]), ivs, &(xi[0]));
1775
T2 = LD(&(xi[WS(is, 64)]), ivs, &(xi[0]));
1778
T4N = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
1779
T4O = LD(&(xi[WS(is, 96)]), ivs, &(xi[0]));
1780
T4P = VSUB(T4N, T4O);
1781
Ta8 = VADD(T4N, T4O);
1784
V T4, T5, T6, T7, T8, T9;
1785
T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
1786
T5 = LD(&(xi[WS(is, 80)]), ivs, &(xi[0]));
1788
T7 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0]));
1789
T8 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
1791
Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
1793
T4M = VMUL(LDK(KP707106781), VSUB(T9, T6));
1800
Tc = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
1801
Td = LD(&(xi[WS(is, 72)]), ivs, &(xi[0]));
1804
Tf = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
1805
Tg = LD(&(xi[WS(is, 104)]), ivs, &(xi[0]));
1809
Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
1810
Tea = VSUB(Tc9, Tca);
1811
T4S = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
1814
Tj = LD(&(xi[WS(is, 120)]), ivs, &(xi[0]));
1815
Tk = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
1818
Tm = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
1819
Tn = LD(&(xi[WS(is, 88)]), ivs, &(xi[0]));
1823
Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
1824
Teb = VSUB(Tc6, Tc7);
1825
T4R = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
1833
Te9 = VSUB(Ta7, Ta8);
1834
Tec = VMUL(LDK(KP707106781), VADD(Tea, Teb));
1835
Ted = VADD(Te9, Tec);
1836
Tgf = VSUB(Te9, Tec);
1839
V Tfo, Tfp, T4Q, T4T;
1840
Tfo = VSUB(Tab, Taa);
1841
Tfp = VMUL(LDK(KP707106781), VSUB(Teb, Tea));
1842
Tfq = VADD(Tfo, Tfp);
1843
TgH = VSUB(Tfp, Tfo);
1844
T4Q = VSUB(T4M, T4P);
1845
T4T = VSUB(T4R, T4S);
1846
T4U = VADD(T4Q, T4T);
1847
T6b = VSUB(T4T, T4Q);
1850
V T6X, T6Y, Ta9, Tac;
1852
T6Y = VADD(T4S, T4R);
1853
T6Z = VADD(T6X, T6Y);
1854
T8T = VSUB(T6X, T6Y);
1855
Ta9 = VADD(Ta7, Ta8);
1856
Tac = VADD(Taa, Tab);
1857
Tad = VSUB(Ta9, Tac);
1858
TcZ = VADD(Ta9, Tac);
1861
V Tc8, Tcb, T82, T83;
1862
Tc8 = VADD(Tc6, Tc7);
1863
Tcb = VADD(Tc9, Tca);
1864
Tcc = VSUB(Tc8, Tcb);
1865
Td0 = VADD(Tcb, Tc8);
1866
T82 = VADD(T4P, T4M);
1868
T84 = VADD(T82, T83);
1869
T9l = VSUB(T83, T82);
1873
V Tb0, Tb1, T2i, Tb2, T3k, Tb3, Tb4, T2p, Tb5, T3h, T2x, TeH, T3n, Tbs, T2E;
1874
V TeI, T3m, Tbp, T2l, T2o, TeG, TeJ;
1876
V T2g, T2h, T3i, T3j;
1877
T2g = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
1878
T2h = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)]));
1879
Tb0 = VADD(T2g, T2h);
1880
T3i = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
1881
T3j = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)]));
1882
Tb1 = VADD(T3i, T3j);
1883
T2i = VSUB(T2g, T2h);
1884
Tb2 = VADD(Tb0, Tb1);
1885
T3k = VSUB(T3i, T3j);
1888
V T2j, T2k, T2m, T2n;
1889
T2j = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
1890
T2k = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)]));
1891
T2l = VSUB(T2j, T2k);
1892
Tb3 = VADD(T2j, T2k);
1893
T2m = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)]));
1894
T2n = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
1895
T2o = VSUB(T2m, T2n);
1896
Tb4 = VADD(T2m, T2n);
1898
T2p = VMUL(LDK(KP707106781), VADD(T2l, T2o));
1899
Tb5 = VADD(Tb3, Tb4);
1900
T3h = VMUL(LDK(KP707106781), VSUB(T2o, T2l));
1902
V T2t, Tbq, T2w, Tbr;
1904
V T2r, T2s, T2u, T2v;
1905
T2r = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
1906
T2s = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)]));
1907
T2t = VSUB(T2r, T2s);
1908
Tbq = VADD(T2r, T2s);
1909
T2u = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
1910
T2v = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)]));
1911
T2w = VSUB(T2u, T2v);
1912
Tbr = VADD(T2u, T2v);
1914
T2x = VFNMS(LDK(KP382683432), T2w, VMUL(LDK(KP923879532), T2t));
1915
TeH = VSUB(Tbq, Tbr);
1916
T3n = VFMA(LDK(KP382683432), T2t, VMUL(LDK(KP923879532), T2w));
1917
Tbs = VADD(Tbq, Tbr);
1920
V T2A, Tbn, T2D, Tbo;
1922
V T2y, T2z, T2B, T2C;
1923
T2y = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)]));
1924
T2z = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
1925
T2A = VSUB(T2y, T2z);
1926
Tbn = VADD(T2y, T2z);
1927
T2B = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
1928
T2C = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)]));
1929
T2D = VSUB(T2B, T2C);
1930
Tbo = VADD(T2B, T2C);
1932
T2E = VFMA(LDK(KP923879532), T2A, VMUL(LDK(KP382683432), T2D));
1933
TeI = VSUB(Tbn, Tbo);
1934
T3m = VFNMS(LDK(KP923879532), T2D, VMUL(LDK(KP382683432), T2A));
1935
Tbp = VADD(Tbn, Tbo);
1937
Tb6 = VSUB(Tb2, Tb5);
1938
Tbt = VSUB(Tbp, Tbs);
1940
V T2q, T2F, TeT, TeU;
1941
T2q = VADD(T2i, T2p);
1942
T2F = VADD(T2x, T2E);
1943
T2G = VADD(T2q, T2F);
1944
T5X = VSUB(T2q, T2F);
1945
TeT = VSUB(Tb4, Tb3);
1946
TeU = VMUL(LDK(KP707106781), VSUB(TeI, TeH));
1947
TeV = VADD(TeT, TeU);
1948
Tgr = VSUB(TeU, TeT);
1951
V T3l, T3o, T7z, T7A;
1952
T3l = VSUB(T3h, T3k);
1953
T3o = VSUB(T3m, T3n);
1954
T3p = VADD(T3l, T3o);
1955
T5V = VSUB(T3o, T3l);
1956
T7z = VADD(T3k, T3h);
1957
T7A = VSUB(T2E, T2x);
1958
T7B = VADD(T7z, T7A);
1959
T95 = VSUB(T7A, T7z);
1961
TeG = VSUB(Tb0, Tb1);
1962
TeJ = VMUL(LDK(KP707106781), VADD(TeH, TeI));
1963
TeK = VADD(TeG, TeJ);
1964
Tgt = VSUB(TeG, TeJ);
1966
V T7o, T7p, Td6, Td7;
1967
T7o = VSUB(T2i, T2p);
1968
T7p = VADD(T3n, T3m);
1969
T7q = VADD(T7o, T7p);
1970
T97 = VSUB(T7o, T7p);
1971
Td6 = VADD(Tb2, Tb5);
1972
Td7 = VADD(Tbs, Tbp);
1973
Td8 = VADD(Td6, Td7);
1974
TdK = VSUB(Td6, Td7);
1978
V Tbx, Tby, T3x, Tbz, T4z, TbA, TbB, T3E, TbC, T4w, T3M, Tf2, T4C, TbZ, T3T;
1979
V Tf3, T4B, TbW, T3A, T3D, Tf1, Tf4;
1981
V T3v, T3w, T4x, T4y;
1982
T3v = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)]));
1983
T3w = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
1984
Tbx = VADD(T3v, T3w);
1985
T4x = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
1986
T4y = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)]));
1987
Tby = VADD(T4x, T4y);
1988
T3x = VSUB(T3v, T3w);
1989
Tbz = VADD(Tbx, Tby);
1990
T4z = VSUB(T4x, T4y);
1993
V T3y, T3z, T3B, T3C;
1994
T3y = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
1995
T3z = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)]));
1996
T3A = VSUB(T3y, T3z);
1997
TbA = VADD(T3y, T3z);
1998
T3B = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)]));
1999
T3C = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
2000
T3D = VSUB(T3B, T3C);
2001
TbB = VADD(T3B, T3C);
2003
T3E = VMUL(LDK(KP707106781), VADD(T3A, T3D));
2004
TbC = VADD(TbA, TbB);
2005
T4w = VMUL(LDK(KP707106781), VSUB(T3D, T3A));
2007
V T3I, TbX, T3L, TbY;
2009
V T3G, T3H, T3J, T3K;
2010
T3G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
2011
T3H = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)]));
2012
T3I = VSUB(T3G, T3H);
2013
TbX = VADD(T3G, T3H);
2014
T3J = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
2015
T3K = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)]));
2016
T3L = VSUB(T3J, T3K);
2017
TbY = VADD(T3J, T3K);
2019
T3M = VFNMS(LDK(KP382683432), T3L, VMUL(LDK(KP923879532), T3I));
2020
Tf2 = VSUB(TbX, TbY);
2021
T4C = VFMA(LDK(KP382683432), T3I, VMUL(LDK(KP923879532), T3L));
2022
TbZ = VADD(TbX, TbY);
2025
V T3P, TbU, T3S, TbV;
2027
V T3N, T3O, T3Q, T3R;
2028
T3N = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)]));
2029
T3O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
2030
T3P = VSUB(T3N, T3O);
2031
TbU = VADD(T3N, T3O);
2032
T3Q = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
2033
T3R = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)]));
2034
T3S = VSUB(T3Q, T3R);
2035
TbV = VADD(T3Q, T3R);
2037
T3T = VFMA(LDK(KP923879532), T3P, VMUL(LDK(KP382683432), T3S));
2038
Tf3 = VSUB(TbU, TbV);
2039
T4B = VFNMS(LDK(KP923879532), T3S, VMUL(LDK(KP382683432), T3P));
2040
TbW = VADD(TbU, TbV);
2042
TbD = VSUB(Tbz, TbC);
2043
Tc0 = VSUB(TbW, TbZ);
2045
V T3F, T3U, Tfe, Tff;
2046
T3F = VADD(T3x, T3E);
2047
T3U = VADD(T3M, T3T);
2048
T3V = VADD(T3F, T3U);
2049
T61 = VSUB(T3F, T3U);
2050
Tfe = VSUB(TbB, TbA);
2051
Tff = VMUL(LDK(KP707106781), VSUB(Tf3, Tf2));
2052
Tfg = VADD(Tfe, Tff);
2053
TgB = VSUB(Tff, Tfe);
2056
V T4A, T4D, T7S, T7T;
2057
T4A = VSUB(T4w, T4z);
2058
T4D = VSUB(T4B, T4C);
2059
T4E = VADD(T4A, T4D);
2060
T65 = VSUB(T4D, T4A);
2061
T7S = VADD(T4z, T4w);
2062
T7T = VSUB(T3T, T3M);
2063
T7U = VADD(T7S, T7T);
2064
T9f = VSUB(T7T, T7S);
2066
Tf1 = VSUB(Tbx, Tby);
2067
Tf4 = VMUL(LDK(KP707106781), VADD(Tf2, Tf3));
2068
Tf5 = VADD(Tf1, Tf4);
2069
Tgx = VSUB(Tf1, Tf4);
2071
V T7H, T7I, Tdd, Tde;
2072
T7H = VSUB(T3x, T3E);
2073
T7I = VADD(T4C, T4B);
2074
T7J = VADD(T7H, T7I);
2075
T9b = VSUB(T7H, T7I);
2076
Tdd = VADD(Tbz, TbC);
2077
Tde = VADD(TbZ, TbW);
2078
Tdf = VADD(Tdd, Tde);
2079
TdN = VSUB(Tdd, Tde);
2083
V Tu, Tee, TG, Tag, TL, Teh, TX, Tan, TB, Tef, TD, Taj, TS, Tei, TU;
2086
V Ts, Tt, Tae, TE, TF, Taf;
2087
Ts = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
2088
Tt = LD(&(xi[WS(is, 68)]), ivs, &(xi[0]));
2090
TE = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
2091
TF = LD(&(xi[WS(is, 100)]), ivs, &(xi[0]));
2094
Tee = VSUB(Tae, Taf);
2096
Tag = VADD(Tae, Taf);
2099
V TJ, TK, Tal, TV, TW, Tam;
2100
TJ = LD(&(xi[WS(is, 124)]), ivs, &(xi[0]));
2101
TK = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
2103
TV = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
2104
TW = LD(&(xi[WS(is, 92)]), ivs, &(xi[0]));
2107
Teh = VSUB(Tal, Tam);
2109
Tan = VADD(Tal, Tam);
2115
Tv = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
2116
Tw = LD(&(xi[WS(is, 84)]), ivs, &(xi[0]));
2119
Ty = LD(&(xi[WS(is, 116)]), ivs, &(xi[0]));
2120
Tz = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
2124
TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
2125
Tef = VSUB(Tai, Tah);
2126
TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
2127
Taj = VADD(Tah, Tai);
2133
TM = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
2134
TN = LD(&(xi[WS(is, 76)]), ivs, &(xi[0]));
2137
TP = LD(&(xi[WS(is, 108)]), ivs, &(xi[0]));
2138
TQ = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
2142
TS = VMUL(LDK(KP707106781), VADD(TO, TR));
2143
Tei = VSUB(Tap, Tao);
2144
TU = VMUL(LDK(KP707106781), VSUB(TR, TO));
2145
Taq = VADD(Tao, Tap);
2147
Td2 = VADD(Tag, Taj);
2148
Td3 = VADD(Tan, Taq);
2153
TI = VFMA(LDK(KP980785280), TC, VMUL(LDK(KP195090322), TH));
2154
T4V = VFNMS(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
2155
Tfr = VFNMS(LDK(KP382683432), Tee, VMUL(LDK(KP923879532), Tef));
2156
Tfs = VFMA(LDK(KP382683432), Teh, VMUL(LDK(KP923879532), Tei));
2157
Tft = VADD(Tfr, Tfs);
2158
Tgg = VSUB(Tfs, Tfr);
2164
TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
2165
T4W = VFMA(LDK(KP195090322), TT, VMUL(LDK(KP980785280), TY));
2168
T75 = VFNMS(LDK(KP555570233), T74, VMUL(LDK(KP831469612), T73));
2169
T86 = VFMA(LDK(KP555570233), T73, VMUL(LDK(KP831469612), T74));
2171
Teg = VFMA(LDK(KP923879532), Tee, VMUL(LDK(KP382683432), Tef));
2172
Tej = VFNMS(LDK(KP382683432), Tei, VMUL(LDK(KP923879532), Teh));
2173
Tek = VADD(Teg, Tej);
2174
TgG = VSUB(Tej, Teg);
2176
V T70, T71, Tak, Tar;
2179
T72 = VFMA(LDK(KP831469612), T70, VMUL(LDK(KP555570233), T71));
2180
T85 = VFNMS(LDK(KP555570233), T70, VMUL(LDK(KP831469612), T71));
2181
Tak = VSUB(Tag, Taj);
2182
Tar = VSUB(Tan, Taq);
2183
Tas = VMUL(LDK(KP707106781), VADD(Tak, Tar));
2184
Tcd = VMUL(LDK(KP707106781), VSUB(Tar, Tak));
2188
V Tav, Tau, T1b, Taw, T1v, Tay, Tax, T18, Taz, T1w, T1j, Teo, T1z, TaD, T1q;
2189
V Ten, T1y, TaG, T14, T17, Tem, Tep;
2191
V T19, T1a, T1t, T1u;
2192
T19 = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
2193
T1a = LD(&(xi[WS(is, 98)]), ivs, &(xi[0]));
2194
Tav = VADD(T19, T1a);
2195
T1t = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
2196
T1u = LD(&(xi[WS(is, 66)]), ivs, &(xi[0]));
2197
Tau = VADD(T1t, T1u);
2198
T1b = VSUB(T19, T1a);
2199
Taw = VADD(Tau, Tav);
2200
T1v = VSUB(T1t, T1u);
2203
V T12, T13, T15, T16;
2204
T12 = LD(&(xi[WS(is, 114)]), ivs, &(xi[0]));
2205
T13 = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
2206
T14 = VSUB(T12, T13);
2207
Tay = VADD(T12, T13);
2208
T15 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
2209
T16 = LD(&(xi[WS(is, 82)]), ivs, &(xi[0]));
2210
T17 = VSUB(T15, T16);
2211
Tax = VADD(T15, T16);
2213
T18 = VMUL(LDK(KP707106781), VSUB(T14, T17));
2214
Taz = VADD(Tax, Tay);
2215
T1w = VMUL(LDK(KP707106781), VADD(T17, T14));
2217
V T1f, TaB, T1i, TaC;
2219
V T1d, T1e, T1g, T1h;
2220
T1d = LD(&(xi[WS(is, 122)]), ivs, &(xi[0]));
2221
T1e = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
2222
T1f = VSUB(T1d, T1e);
2223
TaB = VADD(T1d, T1e);
2224
T1g = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
2225
T1h = LD(&(xi[WS(is, 90)]), ivs, &(xi[0]));
2226
T1i = VSUB(T1g, T1h);
2227
TaC = VADD(T1g, T1h);
2229
T1j = VFNMS(LDK(KP923879532), T1i, VMUL(LDK(KP382683432), T1f));
2230
Teo = VSUB(TaB, TaC);
2231
T1z = VFMA(LDK(KP923879532), T1f, VMUL(LDK(KP382683432), T1i));
2232
TaD = VADD(TaB, TaC);
2235
V T1m, TaE, T1p, TaF;
2237
V T1k, T1l, T1n, T1o;
2238
T1k = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
2239
T1l = LD(&(xi[WS(is, 74)]), ivs, &(xi[0]));
2240
T1m = VSUB(T1k, T1l);
2241
TaE = VADD(T1k, T1l);
2242
T1n = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
2243
T1o = LD(&(xi[WS(is, 106)]), ivs, &(xi[0]));
2244
T1p = VSUB(T1n, T1o);
2245
TaF = VADD(T1n, T1o);
2247
T1q = VFMA(LDK(KP382683432), T1m, VMUL(LDK(KP923879532), T1p));
2248
Ten = VSUB(TaE, TaF);
2249
T1y = VFNMS(LDK(KP382683432), T1p, VMUL(LDK(KP923879532), T1m));
2250
TaG = VADD(TaE, TaF);
2252
Tdp = VADD(Taw, Taz);
2253
Tdq = VADD(TaG, TaD);
2254
TdG = VSUB(Tdp, Tdq);
2255
Tem = VSUB(Tau, Tav);
2256
Tep = VMUL(LDK(KP707106781), VADD(Ten, Teo));
2257
Teq = VADD(Tem, Tep);
2258
Tgm = VSUB(Tem, Tep);
2260
V Ter, Tes, T1c, T1r;
2261
Ter = VSUB(Tay, Tax);
2262
Tes = VMUL(LDK(KP707106781), VSUB(Teo, Ten));
2263
Tet = VADD(Ter, Tes);
2264
Tgl = VSUB(Tes, Ter);
2265
T1c = VSUB(T18, T1b);
2266
T1r = VSUB(T1j, T1q);
2267
T1s = VADD(T1c, T1r);
2268
T5P = VSUB(T1r, T1c);
2271
V T1x, T1A, T7b, T7c;
2272
T1x = VADD(T1v, T1w);
2273
T1A = VADD(T1y, T1z);
2274
T1B = VADD(T1x, T1A);
2275
T5Q = VSUB(T1x, T1A);
2276
T7b = VADD(T1b, T18);
2277
T7c = VSUB(T1z, T1y);
2278
T7d = VADD(T7b, T7c);
2279
T8Z = VSUB(T7c, T7b);
2282
V TaA, TaH, T78, T79;
2283
TaA = VSUB(Taw, Taz);
2284
TaH = VSUB(TaD, TaG);
2285
TaI = VFMA(LDK(KP923879532), TaA, VMUL(LDK(KP382683432), TaH));
2286
Tcf = VFNMS(LDK(KP382683432), TaA, VMUL(LDK(KP923879532), TaH));
2287
T78 = VSUB(T1v, T1w);
2288
T79 = VADD(T1q, T1j);
2289
T7a = VADD(T78, T79);
2290
T90 = VSUB(T78, T79);
2294
V TaJ, TaK, T1F, TaL, T27, TaM, TaN, T1M, TaO, T24, T1U, Tew, T2a, TaV, T21;
2295
V Tex, T29, TaS, T1I, T1L, Tev, Tey;
2297
V T1D, T1E, T25, T26;
2298
T1D = LD(&(xi[WS(is, 126)]), ivs, &(xi[0]));
2299
T1E = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
2300
TaJ = VADD(T1D, T1E);
2301
T25 = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
2302
T26 = LD(&(xi[WS(is, 94)]), ivs, &(xi[0]));
2303
TaK = VADD(T25, T26);
2304
T1F = VSUB(T1D, T1E);
2305
TaL = VADD(TaJ, TaK);
2306
T27 = VSUB(T25, T26);
2309
V T1G, T1H, T1J, T1K;
2310
T1G = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
2311
T1H = LD(&(xi[WS(is, 78)]), ivs, &(xi[0]));
2312
T1I = VSUB(T1G, T1H);
2313
TaM = VADD(T1G, T1H);
2314
T1J = LD(&(xi[WS(is, 110)]), ivs, &(xi[0]));
2315
T1K = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
2316
T1L = VSUB(T1J, T1K);
2317
TaN = VADD(T1J, T1K);
2319
T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L));
2320
TaO = VADD(TaM, TaN);
2321
T24 = VMUL(LDK(KP707106781), VSUB(T1L, T1I));
2323
V T1Q, TaT, T1T, TaU;
2325
V T1O, T1P, T1R, T1S;
2326
T1O = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
2327
T1P = LD(&(xi[WS(is, 70)]), ivs, &(xi[0]));
2328
T1Q = VSUB(T1O, T1P);
2329
TaT = VADD(T1O, T1P);
2330
T1R = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
2331
T1S = LD(&(xi[WS(is, 102)]), ivs, &(xi[0]));
2332
T1T = VSUB(T1R, T1S);
2333
TaU = VADD(T1R, T1S);
2335
T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q));
2336
Tew = VSUB(TaT, TaU);
2337
T2a = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T));
2338
TaV = VADD(TaT, TaU);
2341
V T1X, TaQ, T20, TaR;
2343
V T1V, T1W, T1Y, T1Z;
2344
T1V = LD(&(xi[WS(is, 118)]), ivs, &(xi[0]));
2345
T1W = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
2346
T1X = VSUB(T1V, T1W);
2347
TaQ = VADD(T1V, T1W);
2348
T1Y = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
2349
T1Z = LD(&(xi[WS(is, 86)]), ivs, &(xi[0]));
2350
T20 = VSUB(T1Y, T1Z);
2351
TaR = VADD(T1Y, T1Z);
2353
T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20));
2354
Tex = VSUB(TaQ, TaR);
2355
T29 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X));
2356
TaS = VADD(TaQ, TaR);
2358
Tdm = VADD(TaL, TaO);
2359
Tdn = VADD(TaV, TaS);
2360
TdH = VSUB(Tdm, Tdn);
2361
Tev = VSUB(TaJ, TaK);
2362
Tey = VMUL(LDK(KP707106781), VADD(Tew, Tex));
2363
Tez = VADD(Tev, Tey);
2364
Tgi = VSUB(Tev, Tey);
2366
V TeA, TeB, T1N, T22;
2367
TeA = VSUB(TaN, TaM);
2368
TeB = VMUL(LDK(KP707106781), VSUB(Tex, Tew));
2369
TeC = VADD(TeA, TeB);
2370
Tgj = VSUB(TeB, TeA);
2371
T1N = VADD(T1F, T1M);
2372
T22 = VADD(T1U, T21);
2373
T23 = VADD(T1N, T22);
2374
T5N = VSUB(T1N, T22);
2377
V T28, T2b, T7i, T7j;
2378
T28 = VSUB(T24, T27);
2379
T2b = VSUB(T29, T2a);
2380
T2c = VADD(T28, T2b);
2381
T5M = VSUB(T2b, T28);
2382
T7i = VADD(T27, T24);
2383
T7j = VSUB(T21, T1U);
2384
T7k = VADD(T7i, T7j);
2385
T8X = VSUB(T7j, T7i);
2388
V TaP, TaW, T7f, T7g;
2389
TaP = VSUB(TaL, TaO);
2390
TaW = VSUB(TaS, TaV);
2391
TaX = VFNMS(LDK(KP382683432), TaW, VMUL(LDK(KP923879532), TaP));
2392
Tcg = VFMA(LDK(KP382683432), TaP, VMUL(LDK(KP923879532), TaW));
2393
T7f = VSUB(T1F, T1M);
2394
T7g = VADD(T2a, T29);
2395
T7h = VADD(T7f, T7g);
2396
T8W = VSUB(T7f, T7g);
2400
V T2J, TeL, T2V, Tb9, T30, TeO, T3c, Tbg, T2Q, TeM, T2S, Tbc, T37, TeP, T39;
2403
V T2H, T2I, Tb7, T2T, T2U, Tb8;
2404
T2H = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
2405
T2I = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)]));
2406
Tb7 = VADD(T2H, T2I);
2407
T2T = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
2408
T2U = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)]));
2409
Tb8 = VADD(T2T, T2U);
2410
T2J = VSUB(T2H, T2I);
2411
TeL = VSUB(Tb7, Tb8);
2412
T2V = VSUB(T2T, T2U);
2413
Tb9 = VADD(Tb7, Tb8);
2416
V T2Y, T2Z, Tbe, T3a, T3b, Tbf;
2417
T2Y = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)]));
2418
T2Z = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
2419
Tbe = VADD(T2Y, T2Z);
2420
T3a = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
2421
T3b = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)]));
2422
Tbf = VADD(T3a, T3b);
2423
T30 = VSUB(T2Y, T2Z);
2424
TeO = VSUB(Tbe, Tbf);
2425
T3c = VSUB(T3a, T3b);
2426
Tbg = VADD(Tbe, Tbf);
2429
V T2M, Tba, T2P, Tbb;
2431
V T2K, T2L, T2N, T2O;
2432
T2K = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
2433
T2L = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)]));
2434
T2M = VSUB(T2K, T2L);
2435
Tba = VADD(T2K, T2L);
2436
T2N = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)]));
2437
T2O = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
2438
T2P = VSUB(T2N, T2O);
2439
Tbb = VADD(T2N, T2O);
2441
T2Q = VMUL(LDK(KP707106781), VADD(T2M, T2P));
2442
TeM = VSUB(Tbb, Tba);
2443
T2S = VMUL(LDK(KP707106781), VSUB(T2P, T2M));
2444
Tbc = VADD(Tba, Tbb);
2447
V T33, Tbh, T36, Tbi;
2449
V T31, T32, T34, T35;
2450
T31 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
2451
T32 = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)]));
2452
T33 = VSUB(T31, T32);
2453
Tbh = VADD(T31, T32);
2454
T34 = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)]));
2455
T35 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
2456
T36 = VSUB(T34, T35);
2457
Tbi = VADD(T34, T35);
2459
T37 = VMUL(LDK(KP707106781), VADD(T33, T36));
2460
TeP = VSUB(Tbi, Tbh);
2461
T39 = VMUL(LDK(KP707106781), VSUB(T36, T33));
2462
Tbj = VADD(Tbh, Tbi);
2465
V Tbd, Tbk, TeN, TeQ;
2466
Tbd = VSUB(Tb9, Tbc);
2467
Tbk = VSUB(Tbg, Tbj);
2468
Tbl = VMUL(LDK(KP707106781), VADD(Tbd, Tbk));
2469
Tbu = VMUL(LDK(KP707106781), VSUB(Tbk, Tbd));
2471
V Td9, Tda, TeW, TeX;
2472
Td9 = VADD(Tb9, Tbc);
2473
Tda = VADD(Tbg, Tbj);
2474
Tdb = VADD(Td9, Tda);
2475
TdL = VSUB(Tda, Td9);
2476
TeW = VFNMS(LDK(KP382683432), TeL, VMUL(LDK(KP923879532), TeM));
2477
TeX = VFMA(LDK(KP382683432), TeO, VMUL(LDK(KP923879532), TeP));
2478
TeY = VADD(TeW, TeX);
2479
Tgu = VSUB(TeX, TeW);
2481
TeN = VFMA(LDK(KP923879532), TeL, VMUL(LDK(KP382683432), TeM));
2482
TeQ = VFNMS(LDK(KP382683432), TeP, VMUL(LDK(KP923879532), TeO));
2483
TeR = VADD(TeN, TeQ);
2484
Tgq = VSUB(TeQ, TeN);
2486
V T7t, T7C, T7w, T7D;
2488
V T7r, T7s, T7u, T7v;
2489
T7r = VSUB(T2J, T2Q);
2490
T7s = VADD(T2V, T2S);
2491
T7t = VFMA(LDK(KP831469612), T7r, VMUL(LDK(KP555570233), T7s));
2492
T7C = VFNMS(LDK(KP555570233), T7r, VMUL(LDK(KP831469612), T7s));
2493
T7u = VSUB(T30, T37);
2494
T7v = VADD(T3c, T39);
2495
T7w = VFNMS(LDK(KP555570233), T7v, VMUL(LDK(KP831469612), T7u));
2496
T7D = VFMA(LDK(KP555570233), T7u, VMUL(LDK(KP831469612), T7v));
2498
T7x = VADD(T7t, T7w);
2499
T98 = VSUB(T7D, T7C);
2500
T7E = VADD(T7C, T7D);
2501
T94 = VSUB(T7w, T7t);
2504
V T2X, T3q, T3e, T3r;
2506
V T2R, T2W, T38, T3d;
2507
T2R = VADD(T2J, T2Q);
2508
T2W = VSUB(T2S, T2V);
2509
T2X = VFMA(LDK(KP980785280), T2R, VMUL(LDK(KP195090322), T2W));
2510
T3q = VFNMS(LDK(KP195090322), T2R, VMUL(LDK(KP980785280), T2W));
2511
T38 = VADD(T30, T37);
2512
T3d = VSUB(T39, T3c);
2513
T3e = VFNMS(LDK(KP195090322), T3d, VMUL(LDK(KP980785280), T38));
2514
T3r = VFMA(LDK(KP195090322), T38, VMUL(LDK(KP980785280), T3d));
2516
T3f = VADD(T2X, T3e);
2517
T5Y = VSUB(T3r, T3q);
2518
T3s = VADD(T3q, T3r);
2519
T5U = VSUB(T3e, T2X);
2524
V T3Y, Tf6, T4a, TbG, T4f, Tf9, T4r, TbN, T45, Tf7, T47, TbJ, T4m, Tfa, T4o;
2527
V T3W, T3X, TbE, T48, T49, TbF;
2528
T3W = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
2529
T3X = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)]));
2530
TbE = VADD(T3W, T3X);
2531
T48 = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
2532
T49 = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)]));
2533
TbF = VADD(T48, T49);
2534
T3Y = VSUB(T3W, T3X);
2535
Tf6 = VSUB(TbE, TbF);
2536
T4a = VSUB(T48, T49);
2537
TbG = VADD(TbE, TbF);
2540
V T4d, T4e, TbL, T4p, T4q, TbM;
2541
T4d = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)]));
2542
T4e = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
2543
TbL = VADD(T4d, T4e);
2544
T4p = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
2545
T4q = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)]));
2546
TbM = VADD(T4p, T4q);
2547
T4f = VSUB(T4d, T4e);
2548
Tf9 = VSUB(TbL, TbM);
2549
T4r = VSUB(T4p, T4q);
2550
TbN = VADD(TbL, TbM);
2553
V T41, TbH, T44, TbI;
2555
V T3Z, T40, T42, T43;
2556
T3Z = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
2557
T40 = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)]));
2558
T41 = VSUB(T3Z, T40);
2559
TbH = VADD(T3Z, T40);
2560
T42 = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)]));
2561
T43 = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
2562
T44 = VSUB(T42, T43);
2563
TbI = VADD(T42, T43);
2565
T45 = VMUL(LDK(KP707106781), VADD(T41, T44));
2566
Tf7 = VSUB(TbI, TbH);
2567
T47 = VMUL(LDK(KP707106781), VSUB(T44, T41));
2568
TbJ = VADD(TbH, TbI);
2571
V T4i, TbO, T4l, TbP;
2573
V T4g, T4h, T4j, T4k;
2574
T4g = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
2575
T4h = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)]));
2576
T4i = VSUB(T4g, T4h);
2577
TbO = VADD(T4g, T4h);
2578
T4j = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)]));
2579
T4k = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
2580
T4l = VSUB(T4j, T4k);
2581
TbP = VADD(T4j, T4k);
2583
T4m = VMUL(LDK(KP707106781), VADD(T4i, T4l));
2584
Tfa = VSUB(TbP, TbO);
2585
T4o = VMUL(LDK(KP707106781), VSUB(T4l, T4i));
2586
TbQ = VADD(TbO, TbP);
2589
V TbK, TbR, Tf8, Tfb;
2590
TbK = VSUB(TbG, TbJ);
2591
TbR = VSUB(TbN, TbQ);
2592
TbS = VMUL(LDK(KP707106781), VADD(TbK, TbR));
2593
Tc1 = VMUL(LDK(KP707106781), VSUB(TbR, TbK));
2595
V Tdg, Tdh, Tfh, Tfi;
2596
Tdg = VADD(TbG, TbJ);
2597
Tdh = VADD(TbN, TbQ);
2598
Tdi = VADD(Tdg, Tdh);
2599
TdO = VSUB(Tdh, Tdg);
2600
Tfh = VFNMS(LDK(KP382683432), Tf6, VMUL(LDK(KP923879532), Tf7));
2601
Tfi = VFMA(LDK(KP382683432), Tf9, VMUL(LDK(KP923879532), Tfa));
2602
Tfj = VADD(Tfh, Tfi);
2603
Tgy = VSUB(Tfi, Tfh);
2605
Tf8 = VFMA(LDK(KP923879532), Tf6, VMUL(LDK(KP382683432), Tf7));
2606
Tfb = VFNMS(LDK(KP382683432), Tfa, VMUL(LDK(KP923879532), Tf9));
2607
Tfc = VADD(Tf8, Tfb);
2608
TgA = VSUB(Tfb, Tf8);
2610
V T7M, T7V, T7P, T7W;
2612
V T7K, T7L, T7N, T7O;
2613
T7K = VSUB(T3Y, T45);
2614
T7L = VADD(T4a, T47);
2615
T7M = VFMA(LDK(KP831469612), T7K, VMUL(LDK(KP555570233), T7L));
2616
T7V = VFNMS(LDK(KP555570233), T7K, VMUL(LDK(KP831469612), T7L));
2617
T7N = VSUB(T4f, T4m);
2618
T7O = VADD(T4r, T4o);
2619
T7P = VFNMS(LDK(KP555570233), T7O, VMUL(LDK(KP831469612), T7N));
2620
T7W = VFMA(LDK(KP555570233), T7N, VMUL(LDK(KP831469612), T7O));
2622
T7Q = VADD(T7M, T7P);
2623
T9e = VSUB(T7P, T7M);
2624
T7X = VADD(T7V, T7W);
2625
T9c = VSUB(T7W, T7V);
2628
V T4c, T4F, T4t, T4G;
2630
V T46, T4b, T4n, T4s;
2631
T46 = VADD(T3Y, T45);
2632
T4b = VSUB(T47, T4a);
2633
T4c = VFMA(LDK(KP980785280), T46, VMUL(LDK(KP195090322), T4b));
2634
T4F = VFNMS(LDK(KP195090322), T46, VMUL(LDK(KP980785280), T4b));
2635
T4n = VADD(T4f, T4m);
2636
T4s = VSUB(T4o, T4r);
2637
T4t = VFNMS(LDK(KP195090322), T4s, VMUL(LDK(KP980785280), T4n));
2638
T4G = VFMA(LDK(KP195090322), T4n, VMUL(LDK(KP980785280), T4s));
2640
T4u = VADD(T4c, T4t);
2641
T64 = VSUB(T4t, T4c);
2642
T4H = VADD(T4F, T4G);
2643
T62 = VSUB(T4G, T4F);
2648
V Td5, Tdx, TdC, TdE, Tdk, Tdt, Tds, Tdy, Tdz, TdD;
2650
V Td1, Td4, TdA, TdB;
2651
Td1 = VADD(TcZ, Td0);
2652
Td4 = VADD(Td2, Td3);
2653
Td5 = VSUB(Td1, Td4);
2654
Tdx = VADD(Td1, Td4);
2655
TdA = VADD(Td8, Tdb);
2656
TdB = VADD(Tdf, Tdi);
2657
TdC = VADD(TdA, TdB);
2658
TdE = VBYI(VSUB(TdB, TdA));
2661
V Tdc, Tdj, Tdo, Tdr;
2662
Tdc = VSUB(Td8, Tdb);
2663
Tdj = VSUB(Tdf, Tdi);
2664
Tdk = VMUL(LDK(KP707106781), VADD(Tdc, Tdj));
2665
Tdt = VMUL(LDK(KP707106781), VSUB(Tdj, Tdc));
2666
Tdo = VADD(Tdm, Tdn);
2667
Tdr = VADD(Tdp, Tdq);
2668
Tds = VSUB(Tdo, Tdr);
2669
Tdy = VADD(Tdr, Tdo);
2671
Tdz = VADD(Tdx, Tdy);
2672
ST(&(xo[WS(os, 64)]), VSUB(Tdz, TdC), ovs, &(xo[0]));
2673
ST(&(xo[0]), VADD(Tdz, TdC), ovs, &(xo[0]));
2674
TdD = VSUB(Tdx, Tdy);
2675
ST(&(xo[WS(os, 96)]), VSUB(TdD, TdE), ovs, &(xo[0]));
2676
ST(&(xo[WS(os, 32)]), VADD(TdD, TdE), ovs, &(xo[0]));
2678
V Tdl, Tdu, Tdv, Tdw;
2679
Tdl = VADD(Td5, Tdk);
2680
Tdu = VBYI(VADD(Tds, Tdt));
2681
ST(&(xo[WS(os, 112)]), VSUB(Tdl, Tdu), ovs, &(xo[0]));
2682
ST(&(xo[WS(os, 16)]), VADD(Tdl, Tdu), ovs, &(xo[0]));
2683
Tdv = VSUB(Td5, Tdk);
2684
Tdw = VBYI(VSUB(Tdt, Tds));
2685
ST(&(xo[WS(os, 80)]), VSUB(Tdv, Tdw), ovs, &(xo[0]));
2686
ST(&(xo[WS(os, 48)]), VADD(Tdv, Tdw), ovs, &(xo[0]));
2690
V TdJ, Te4, TdX, Te5, TdQ, Te1, TdU, Te2;
2692
V TdF, TdI, TdV, TdW;
2693
TdF = VSUB(TcZ, Td0);
2694
TdI = VMUL(LDK(KP707106781), VADD(TdG, TdH));
2695
TdJ = VADD(TdF, TdI);
2696
Te4 = VSUB(TdF, TdI);
2697
TdV = VFNMS(LDK(KP382683432), TdK, VMUL(LDK(KP923879532), TdL));
2698
TdW = VFMA(LDK(KP382683432), TdN, VMUL(LDK(KP923879532), TdO));
2699
TdX = VADD(TdV, TdW);
2700
Te5 = VSUB(TdW, TdV);
2703
V TdM, TdP, TdS, TdT;
2704
TdM = VFMA(LDK(KP923879532), TdK, VMUL(LDK(KP382683432), TdL));
2705
TdP = VFNMS(LDK(KP382683432), TdO, VMUL(LDK(KP923879532), TdN));
2706
TdQ = VADD(TdM, TdP);
2707
Te1 = VSUB(TdP, TdM);
2708
TdS = VSUB(Td3, Td2);
2709
TdT = VMUL(LDK(KP707106781), VSUB(TdH, TdG));
2710
TdU = VADD(TdS, TdT);
2711
Te2 = VSUB(TdT, TdS);
2714
V TdR, TdY, Te7, Te8;
2715
TdR = VADD(TdJ, TdQ);
2716
TdY = VBYI(VADD(TdU, TdX));
2717
ST(&(xo[WS(os, 120)]), VSUB(TdR, TdY), ovs, &(xo[0]));
2718
ST(&(xo[WS(os, 8)]), VADD(TdR, TdY), ovs, &(xo[0]));
2719
Te7 = VBYI(VADD(Te2, Te1));
2720
Te8 = VADD(Te4, Te5);
2721
ST(&(xo[WS(os, 24)]), VADD(Te7, Te8), ovs, &(xo[0]));
2722
ST(&(xo[WS(os, 104)]), VSUB(Te8, Te7), ovs, &(xo[0]));
2725
V TdZ, Te0, Te3, Te6;
2726
TdZ = VSUB(TdJ, TdQ);
2727
Te0 = VBYI(VSUB(TdX, TdU));
2728
ST(&(xo[WS(os, 72)]), VSUB(TdZ, Te0), ovs, &(xo[0]));
2729
ST(&(xo[WS(os, 56)]), VADD(TdZ, Te0), ovs, &(xo[0]));
2730
Te3 = VBYI(VSUB(Te1, Te2));
2731
Te6 = VSUB(Te4, Te5);
2732
ST(&(xo[WS(os, 40)]), VADD(Te3, Te6), ovs, &(xo[0]));
2733
ST(&(xo[WS(os, 88)]), VSUB(Te6, Te3), ovs, &(xo[0]));
2737
V TaZ, Tcs, Tci, Tcq, Tc4, Tct, Tcl, Tcp;
2739
V Tat, TaY, Tce, Tch;
2740
Tat = VADD(Tad, Tas);
2741
TaY = VADD(TaI, TaX);
2742
TaZ = VADD(Tat, TaY);
2743
Tcs = VSUB(Tat, TaY);
2744
Tce = VADD(Tcc, Tcd);
2745
Tch = VADD(Tcf, Tcg);
2746
Tci = VADD(Tce, Tch);
2747
Tcq = VSUB(Tch, Tce);
2749
V Tbw, Tcj, Tc3, Tck;
2751
V Tbm, Tbv, TbT, Tc2;
2752
Tbm = VADD(Tb6, Tbl);
2753
Tbv = VADD(Tbt, Tbu);
2754
Tbw = VFMA(LDK(KP980785280), Tbm, VMUL(LDK(KP195090322), Tbv));
2755
Tcj = VFNMS(LDK(KP195090322), Tbm, VMUL(LDK(KP980785280), Tbv));
2756
TbT = VADD(TbD, TbS);
2757
Tc2 = VADD(Tc0, Tc1);
2758
Tc3 = VFNMS(LDK(KP195090322), Tc2, VMUL(LDK(KP980785280), TbT));
2759
Tck = VFMA(LDK(KP195090322), TbT, VMUL(LDK(KP980785280), Tc2));
2761
Tc4 = VADD(Tbw, Tc3);
2762
Tct = VSUB(Tck, Tcj);
2763
Tcl = VADD(Tcj, Tck);
2764
Tcp = VSUB(Tc3, Tbw);
2768
V Tc5, Tcm, Tcv, Tcw;
2769
Tc5 = VADD(TaZ, Tc4);
2770
Tcm = VBYI(VADD(Tci, Tcl));
2771
ST(&(xo[WS(os, 124)]), VSUB(Tc5, Tcm), ovs, &(xo[0]));
2772
ST(&(xo[WS(os, 4)]), VADD(Tc5, Tcm), ovs, &(xo[0]));
2773
Tcv = VBYI(VADD(Tcq, Tcp));
2774
Tcw = VADD(Tcs, Tct);
2775
ST(&(xo[WS(os, 28)]), VADD(Tcv, Tcw), ovs, &(xo[0]));
2776
ST(&(xo[WS(os, 100)]), VSUB(Tcw, Tcv), ovs, &(xo[0]));
2779
V Tcn, Tco, Tcr, Tcu;
2780
Tcn = VSUB(TaZ, Tc4);
2781
Tco = VBYI(VSUB(Tcl, Tci));
2782
ST(&(xo[WS(os, 68)]), VSUB(Tcn, Tco), ovs, &(xo[0]));
2783
ST(&(xo[WS(os, 60)]), VADD(Tcn, Tco), ovs, &(xo[0]));
2784
Tcr = VBYI(VSUB(Tcp, Tcq));
2785
Tcu = VSUB(Tcs, Tct);
2786
ST(&(xo[WS(os, 36)]), VADD(Tcr, Tcu), ovs, &(xo[0]));
2787
ST(&(xo[WS(os, 92)]), VSUB(Tcu, Tcr), ovs, &(xo[0]));
2791
V Tcz, TcU, TcK, TcS, TcG, TcV, TcN, TcR;
2793
V Tcx, Tcy, TcI, TcJ;
2794
Tcx = VSUB(Tad, Tas);
2795
Tcy = VSUB(Tcg, Tcf);
2796
Tcz = VADD(Tcx, Tcy);
2797
TcU = VSUB(Tcx, Tcy);
2798
TcI = VSUB(Tcd, Tcc);
2799
TcJ = VSUB(TaX, TaI);
2800
TcK = VADD(TcI, TcJ);
2801
TcS = VSUB(TcJ, TcI);
2803
V TcC, TcL, TcF, TcM;
2805
V TcA, TcB, TcD, TcE;
2806
TcA = VSUB(Tb6, Tbl);
2807
TcB = VSUB(Tbu, Tbt);
2808
TcC = VFMA(LDK(KP831469612), TcA, VMUL(LDK(KP555570233), TcB));
2809
TcL = VFNMS(LDK(KP555570233), TcA, VMUL(LDK(KP831469612), TcB));
2810
TcD = VSUB(TbD, TbS);
2811
TcE = VSUB(Tc1, Tc0);
2812
TcF = VFNMS(LDK(KP555570233), TcE, VMUL(LDK(KP831469612), TcD));
2813
TcM = VFMA(LDK(KP555570233), TcD, VMUL(LDK(KP831469612), TcE));
2815
TcG = VADD(TcC, TcF);
2816
TcV = VSUB(TcM, TcL);
2817
TcN = VADD(TcL, TcM);
2818
TcR = VSUB(TcF, TcC);
2822
V TcH, TcO, TcX, TcY;
2823
TcH = VADD(Tcz, TcG);
2824
TcO = VBYI(VADD(TcK, TcN));
2825
ST(&(xo[WS(os, 116)]), VSUB(TcH, TcO), ovs, &(xo[0]));
2826
ST(&(xo[WS(os, 12)]), VADD(TcH, TcO), ovs, &(xo[0]));
2827
TcX = VBYI(VADD(TcS, TcR));
2828
TcY = VADD(TcU, TcV);
2829
ST(&(xo[WS(os, 20)]), VADD(TcX, TcY), ovs, &(xo[0]));
2830
ST(&(xo[WS(os, 108)]), VSUB(TcY, TcX), ovs, &(xo[0]));
2833
V TcP, TcQ, TcT, TcW;
2834
TcP = VSUB(Tcz, TcG);
2835
TcQ = VBYI(VSUB(TcN, TcK));
2836
ST(&(xo[WS(os, 76)]), VSUB(TcP, TcQ), ovs, &(xo[0]));
2837
ST(&(xo[WS(os, 52)]), VADD(TcP, TcQ), ovs, &(xo[0]));
2838
TcT = VBYI(VSUB(TcR, TcS));
2839
TcW = VSUB(TcU, TcV);
2840
ST(&(xo[WS(os, 44)]), VADD(TcT, TcW), ovs, &(xo[0]));
2841
ST(&(xo[WS(os, 84)]), VSUB(TcW, TcT), ovs, &(xo[0]));
2845
V TeF, Tg8, TfI, Tg0, Tfy, Tga, TfG, TfP, Tfm, TfJ, TfB, TfF, TfW, Tgb, Tg3;
2848
V Tel, TfY, TeE, TfZ, Teu, TeD;
2849
Tel = VADD(Ted, Tek);
2850
TfY = VSUB(Tft, Tfq);
2851
Teu = VFMA(LDK(KP980785280), Teq, VMUL(LDK(KP195090322), Tet));
2852
TeD = VFNMS(LDK(KP195090322), TeC, VMUL(LDK(KP980785280), Tez));
2853
TeE = VADD(Teu, TeD);
2854
TfZ = VSUB(TeD, Teu);
2855
TeF = VADD(Tel, TeE);
2856
Tg8 = VSUB(TfZ, TfY);
2857
TfI = VSUB(Tel, TeE);
2858
Tg0 = VADD(TfY, TfZ);
2861
V Tfu, TfN, Tfx, TfO, Tfv, Tfw;
2862
Tfu = VADD(Tfq, Tft);
2863
TfN = VSUB(Ted, Tek);
2864
Tfv = VFNMS(LDK(KP195090322), Teq, VMUL(LDK(KP980785280), Tet));
2865
Tfw = VFMA(LDK(KP195090322), Tez, VMUL(LDK(KP980785280), TeC));
2866
Tfx = VADD(Tfv, Tfw);
2867
TfO = VSUB(Tfw, Tfv);
2868
Tfy = VADD(Tfu, Tfx);
2869
Tga = VSUB(TfN, TfO);
2870
TfG = VSUB(Tfx, Tfu);
2871
TfP = VADD(TfN, TfO);
2874
V Tf0, Tfz, Tfl, TfA;
2876
V TeS, TeZ, Tfd, Tfk;
2877
TeS = VADD(TeK, TeR);
2878
TeZ = VADD(TeV, TeY);
2879
Tf0 = VFMA(LDK(KP995184726), TeS, VMUL(LDK(KP098017140), TeZ));
2880
Tfz = VFNMS(LDK(KP098017140), TeS, VMUL(LDK(KP995184726), TeZ));
2881
Tfd = VADD(Tf5, Tfc);
2882
Tfk = VADD(Tfg, Tfj);
2883
Tfl = VFNMS(LDK(KP098017140), Tfk, VMUL(LDK(KP995184726), Tfd));
2884
TfA = VFMA(LDK(KP098017140), Tfd, VMUL(LDK(KP995184726), Tfk));
2886
Tfm = VADD(Tf0, Tfl);
2887
TfJ = VSUB(TfA, Tfz);
2888
TfB = VADD(Tfz, TfA);
2889
TfF = VSUB(Tfl, Tf0);
2892
V TfS, Tg1, TfV, Tg2;
2894
V TfQ, TfR, TfT, TfU;
2895
TfQ = VSUB(TeK, TeR);
2896
TfR = VSUB(TeY, TeV);
2897
TfS = VFMA(LDK(KP773010453), TfQ, VMUL(LDK(KP634393284), TfR));
2898
Tg1 = VFNMS(LDK(KP634393284), TfQ, VMUL(LDK(KP773010453), TfR));
2899
TfT = VSUB(Tf5, Tfc);
2900
TfU = VSUB(Tfj, Tfg);
2901
TfV = VFNMS(LDK(KP634393284), TfU, VMUL(LDK(KP773010453), TfT));
2902
Tg2 = VFMA(LDK(KP634393284), TfT, VMUL(LDK(KP773010453), TfU));
2904
TfW = VADD(TfS, TfV);
2905
Tgb = VSUB(Tg2, Tg1);
2906
Tg3 = VADD(Tg1, Tg2);
2907
Tg7 = VSUB(TfV, TfS);
2910
V Tfn, TfC, Tg9, Tgc;
2911
Tfn = VADD(TeF, Tfm);
2912
TfC = VBYI(VADD(Tfy, TfB));
2913
ST(&(xo[WS(os, 126)]), VSUB(Tfn, TfC), ovs, &(xo[0]));
2914
ST(&(xo[WS(os, 2)]), VADD(Tfn, TfC), ovs, &(xo[0]));
2915
Tg9 = VBYI(VSUB(Tg7, Tg8));
2916
Tgc = VSUB(Tga, Tgb);
2917
ST(&(xo[WS(os, 46)]), VADD(Tg9, Tgc), ovs, &(xo[0]));
2918
ST(&(xo[WS(os, 82)]), VSUB(Tgc, Tg9), ovs, &(xo[0]));
2921
V Tgd, Tge, TfD, TfE;
2922
Tgd = VBYI(VADD(Tg8, Tg7));
2923
Tge = VADD(Tga, Tgb);
2924
ST(&(xo[WS(os, 18)]), VADD(Tgd, Tge), ovs, &(xo[0]));
2925
ST(&(xo[WS(os, 110)]), VSUB(Tge, Tgd), ovs, &(xo[0]));
2926
TfD = VSUB(TeF, Tfm);
2927
TfE = VBYI(VSUB(TfB, Tfy));
2928
ST(&(xo[WS(os, 66)]), VSUB(TfD, TfE), ovs, &(xo[0]));
2929
ST(&(xo[WS(os, 62)]), VADD(TfD, TfE), ovs, &(xo[0]));
2932
V TfH, TfK, TfX, Tg4;
2933
TfH = VBYI(VSUB(TfF, TfG));
2934
TfK = VSUB(TfI, TfJ);
2935
ST(&(xo[WS(os, 34)]), VADD(TfH, TfK), ovs, &(xo[0]));
2936
ST(&(xo[WS(os, 94)]), VSUB(TfK, TfH), ovs, &(xo[0]));
2937
TfX = VADD(TfP, TfW);
2938
Tg4 = VBYI(VADD(Tg0, Tg3));
2939
ST(&(xo[WS(os, 114)]), VSUB(TfX, Tg4), ovs, &(xo[0]));
2940
ST(&(xo[WS(os, 14)]), VADD(TfX, Tg4), ovs, &(xo[0]));
2943
V Tg5, Tg6, TfL, TfM;
2944
Tg5 = VSUB(TfP, TfW);
2945
Tg6 = VBYI(VSUB(Tg3, Tg0));
2946
ST(&(xo[WS(os, 78)]), VSUB(Tg5, Tg6), ovs, &(xo[0]));
2947
ST(&(xo[WS(os, 50)]), VADD(Tg5, Tg6), ovs, &(xo[0]));
2948
TfL = VBYI(VADD(TfG, TfF));
2949
TfM = VADD(TfI, TfJ);
2950
ST(&(xo[WS(os, 30)]), VADD(TfL, TfM), ovs, &(xo[0]));
2951
ST(&(xo[WS(os, 98)]), VSUB(TfM, TfL), ovs, &(xo[0]));
2955
V Tgp, Thm, TgW, The, TgM, Tho, TgU, Th3, TgE, TgX, TgP, TgT, Tha, Thp, Thh;
2958
V Tgh, Thc, Tgo, Thd, Tgk, Tgn;
2959
Tgh = VSUB(Tgf, Tgg);
2960
Thc = VADD(TgH, TgG);
2961
Tgk = VFMA(LDK(KP555570233), Tgi, VMUL(LDK(KP831469612), Tgj));
2962
Tgn = VFNMS(LDK(KP555570233), Tgm, VMUL(LDK(KP831469612), Tgl));
2963
Tgo = VSUB(Tgk, Tgn);
2964
Thd = VADD(Tgn, Tgk);
2965
Tgp = VADD(Tgh, Tgo);
2966
Thm = VSUB(Thd, Thc);
2967
TgW = VSUB(Tgh, Tgo);
2968
The = VADD(Thc, Thd);
2971
V TgI, Th1, TgL, Th2, TgJ, TgK;
2972
TgI = VSUB(TgG, TgH);
2973
Th1 = VADD(Tgf, Tgg);
2974
TgJ = VFNMS(LDK(KP555570233), Tgj, VMUL(LDK(KP831469612), Tgi));
2975
TgK = VFMA(LDK(KP831469612), Tgm, VMUL(LDK(KP555570233), Tgl));
2976
TgL = VSUB(TgJ, TgK);
2977
Th2 = VADD(TgK, TgJ);
2978
TgM = VADD(TgI, TgL);
2979
Tho = VSUB(Th1, Th2);
2980
TgU = VSUB(TgL, TgI);
2981
Th3 = VADD(Th1, Th2);
2984
V Tgw, TgN, TgD, TgO;
2986
V Tgs, Tgv, Tgz, TgC;
2987
Tgs = VSUB(Tgq, Tgr);
2988
Tgv = VSUB(Tgt, Tgu);
2989
Tgw = VFMA(LDK(KP471396736), Tgs, VMUL(LDK(KP881921264), Tgv));
2990
TgN = VFNMS(LDK(KP471396736), Tgv, VMUL(LDK(KP881921264), Tgs));
2991
Tgz = VSUB(Tgx, Tgy);
2992
TgC = VSUB(TgA, TgB);
2993
TgD = VFNMS(LDK(KP471396736), TgC, VMUL(LDK(KP881921264), Tgz));
2994
TgO = VFMA(LDK(KP881921264), TgC, VMUL(LDK(KP471396736), Tgz));
2996
TgE = VADD(Tgw, TgD);
2997
TgX = VSUB(TgO, TgN);
2998
TgP = VADD(TgN, TgO);
2999
TgT = VSUB(TgD, Tgw);
3002
V Th6, Thf, Th9, Thg;
3004
V Th4, Th5, Th7, Th8;
3005
Th4 = VADD(Tgr, Tgq);
3006
Th5 = VADD(Tgt, Tgu);
3007
Th6 = VFMA(LDK(KP290284677), Th4, VMUL(LDK(KP956940335), Th5));
3008
Thf = VFNMS(LDK(KP290284677), Th5, VMUL(LDK(KP956940335), Th4));
3009
Th7 = VADD(Tgx, Tgy);
3010
Th8 = VADD(TgB, TgA);
3011
Th9 = VFNMS(LDK(KP290284677), Th8, VMUL(LDK(KP956940335), Th7));
3012
Thg = VFMA(LDK(KP956940335), Th8, VMUL(LDK(KP290284677), Th7));
3014
Tha = VADD(Th6, Th9);
3015
Thp = VSUB(Thg, Thf);
3016
Thh = VADD(Thf, Thg);
3017
Thl = VSUB(Th9, Th6);
3020
V TgF, TgQ, Thn, Thq;
3021
TgF = VADD(Tgp, TgE);
3022
TgQ = VBYI(VADD(TgM, TgP));
3023
ST(&(xo[WS(os, 118)]), VSUB(TgF, TgQ), ovs, &(xo[0]));
3024
ST(&(xo[WS(os, 10)]), VADD(TgF, TgQ), ovs, &(xo[0]));
3025
Thn = VBYI(VSUB(Thl, Thm));
3026
Thq = VSUB(Tho, Thp);
3027
ST(&(xo[WS(os, 38)]), VADD(Thn, Thq), ovs, &(xo[0]));
3028
ST(&(xo[WS(os, 90)]), VSUB(Thq, Thn), ovs, &(xo[0]));
3031
V Thr, Ths, TgR, TgS;
3032
Thr = VBYI(VADD(Thm, Thl));
3033
Ths = VADD(Tho, Thp);
3034
ST(&(xo[WS(os, 26)]), VADD(Thr, Ths), ovs, &(xo[0]));
3035
ST(&(xo[WS(os, 102)]), VSUB(Ths, Thr), ovs, &(xo[0]));
3036
TgR = VSUB(Tgp, TgE);
3037
TgS = VBYI(VSUB(TgP, TgM));
3038
ST(&(xo[WS(os, 74)]), VSUB(TgR, TgS), ovs, &(xo[0]));
3039
ST(&(xo[WS(os, 54)]), VADD(TgR, TgS), ovs, &(xo[0]));
3042
V TgV, TgY, Thb, Thi;
3043
TgV = VBYI(VSUB(TgT, TgU));
3044
TgY = VSUB(TgW, TgX);
3045
ST(&(xo[WS(os, 42)]), VADD(TgV, TgY), ovs, &(xo[0]));
3046
ST(&(xo[WS(os, 86)]), VSUB(TgY, TgV), ovs, &(xo[0]));
3047
Thb = VADD(Th3, Tha);
3048
Thi = VBYI(VADD(The, Thh));
3049
ST(&(xo[WS(os, 122)]), VSUB(Thb, Thi), ovs, &(xo[0]));
3050
ST(&(xo[WS(os, 6)]), VADD(Thb, Thi), ovs, &(xo[0]));
3053
V Thj, Thk, TgZ, Th0;
3054
Thj = VSUB(Th3, Tha);
3055
Thk = VBYI(VSUB(Thh, The));
3056
ST(&(xo[WS(os, 70)]), VSUB(Thj, Thk), ovs, &(xo[0]));
3057
ST(&(xo[WS(os, 58)]), VADD(Thj, Thk), ovs, &(xo[0]));
3058
TgZ = VBYI(VADD(TgU, TgT));
3059
Th0 = VADD(TgW, TgX);
3060
ST(&(xo[WS(os, 22)]), VADD(TgZ, Th0), ovs, &(xo[0]));
3061
ST(&(xo[WS(os, 106)]), VSUB(Th0, TgZ), ovs, &(xo[0]));
3065
V T80, T8n, T8f, T8j, T8A, T8P, T8H, T8L, T7n, T8M, T8O, T8c, T8k, T8t, T8E;
3068
V T7G, T8d, T7Z, T8e;
3070
V T7y, T7F, T7R, T7Y;
3071
T7y = VADD(T7q, T7x);
3072
T7F = VADD(T7B, T7E);
3073
T7G = VFMA(LDK(KP989176509), T7y, VMUL(LDK(KP146730474), T7F));
3074
T8d = VFNMS(LDK(KP146730474), T7y, VMUL(LDK(KP989176509), T7F));
3075
T7R = VADD(T7J, T7Q);
3076
T7Y = VADD(T7U, T7X);
3077
T7Z = VFNMS(LDK(KP146730474), T7Y, VMUL(LDK(KP989176509), T7R));
3078
T8e = VFMA(LDK(KP146730474), T7R, VMUL(LDK(KP989176509), T7Y));
3080
T80 = VADD(T7G, T7Z);
3081
T8n = VSUB(T8e, T8d);
3082
T8f = VADD(T8d, T8e);
3083
T8j = VSUB(T7Z, T7G);
3086
V T8w, T8F, T8z, T8G;
3088
V T8u, T8v, T8x, T8y;
3089
T8u = VSUB(T7q, T7x);
3090
T8v = VSUB(T7E, T7B);
3091
T8w = VFMA(LDK(KP803207531), T8u, VMUL(LDK(KP595699304), T8v));
3092
T8F = VFNMS(LDK(KP595699304), T8u, VMUL(LDK(KP803207531), T8v));
3093
T8x = VSUB(T7J, T7Q);
3094
T8y = VSUB(T7X, T7U);
3095
T8z = VFNMS(LDK(KP595699304), T8y, VMUL(LDK(KP803207531), T8x));
3096
T8G = VFMA(LDK(KP595699304), T8x, VMUL(LDK(KP803207531), T8y));
3098
T8A = VADD(T8w, T8z);
3099
T8P = VSUB(T8G, T8F);
3100
T8H = VADD(T8F, T8G);
3101
T8L = VSUB(T8z, T8w);
3104
V T77, T8r, T88, T8C, T7m, T8D, T8b, T8s, T76, T87;
3105
T76 = VADD(T72, T75);
3106
T77 = VADD(T6Z, T76);
3107
T8r = VSUB(T6Z, T76);
3108
T87 = VADD(T85, T86);
3109
T88 = VADD(T84, T87);
3110
T8C = VSUB(T87, T84);
3112
V T7e, T7l, T89, T8a;
3113
T7e = VFMA(LDK(KP956940335), T7a, VMUL(LDK(KP290284677), T7d));
3114
T7l = VFNMS(LDK(KP290284677), T7k, VMUL(LDK(KP956940335), T7h));
3115
T7m = VADD(T7e, T7l);
3116
T8D = VSUB(T7l, T7e);
3117
T89 = VFNMS(LDK(KP290284677), T7a, VMUL(LDK(KP956940335), T7d));
3118
T8a = VFMA(LDK(KP290284677), T7h, VMUL(LDK(KP956940335), T7k));
3119
T8b = VADD(T89, T8a);
3120
T8s = VSUB(T8a, T89);
3122
T7n = VADD(T77, T7m);
3123
T8M = VSUB(T8D, T8C);
3124
T8O = VSUB(T8r, T8s);
3125
T8c = VADD(T88, T8b);
3126
T8k = VSUB(T8b, T88);
3127
T8t = VADD(T8r, T8s);
3128
T8E = VADD(T8C, T8D);
3129
T8m = VSUB(T77, T7m);
3132
V T81, T8g, T8N, T8Q;
3133
T81 = VADD(T7n, T80);
3134
T8g = VBYI(VADD(T8c, T8f));
3135
ST(&(xo[WS(os, 125)]), VSUB(T81, T8g), ovs, &(xo[WS(os, 1)]));
3136
ST(&(xo[WS(os, 3)]), VADD(T81, T8g), ovs, &(xo[WS(os, 1)]));
3137
T8N = VBYI(VSUB(T8L, T8M));
3138
T8Q = VSUB(T8O, T8P);
3139
ST(&(xo[WS(os, 45)]), VADD(T8N, T8Q), ovs, &(xo[WS(os, 1)]));
3140
ST(&(xo[WS(os, 83)]), VSUB(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
3143
V T8R, T8S, T8h, T8i;
3144
T8R = VBYI(VADD(T8M, T8L));
3145
T8S = VADD(T8O, T8P);
3146
ST(&(xo[WS(os, 19)]), VADD(T8R, T8S), ovs, &(xo[WS(os, 1)]));
3147
ST(&(xo[WS(os, 109)]), VSUB(T8S, T8R), ovs, &(xo[WS(os, 1)]));
3148
T8h = VSUB(T7n, T80);
3149
T8i = VBYI(VSUB(T8f, T8c));
3150
ST(&(xo[WS(os, 67)]), VSUB(T8h, T8i), ovs, &(xo[WS(os, 1)]));
3151
ST(&(xo[WS(os, 61)]), VADD(T8h, T8i), ovs, &(xo[WS(os, 1)]));
3154
V T8l, T8o, T8B, T8I;
3155
T8l = VBYI(VSUB(T8j, T8k));
3156
T8o = VSUB(T8m, T8n);
3157
ST(&(xo[WS(os, 35)]), VADD(T8l, T8o), ovs, &(xo[WS(os, 1)]));
3158
ST(&(xo[WS(os, 93)]), VSUB(T8o, T8l), ovs, &(xo[WS(os, 1)]));
3159
T8B = VADD(T8t, T8A);
3160
T8I = VBYI(VADD(T8E, T8H));
3161
ST(&(xo[WS(os, 115)]), VSUB(T8B, T8I), ovs, &(xo[WS(os, 1)]));
3162
ST(&(xo[WS(os, 13)]), VADD(T8B, T8I), ovs, &(xo[WS(os, 1)]));
3165
V T8J, T8K, T8p, T8q;
3166
T8J = VSUB(T8t, T8A);
3167
T8K = VBYI(VSUB(T8H, T8E));
3168
ST(&(xo[WS(os, 77)]), VSUB(T8J, T8K), ovs, &(xo[WS(os, 1)]));
3169
ST(&(xo[WS(os, 51)]), VADD(T8J, T8K), ovs, &(xo[WS(os, 1)]));
3170
T8p = VBYI(VADD(T8k, T8j));
3171
T8q = VADD(T8m, T8n);
3172
ST(&(xo[WS(os, 29)]), VADD(T8p, T8q), ovs, &(xo[WS(os, 1)]));
3173
ST(&(xo[WS(os, 99)]), VSUB(T8q, T8p), ovs, &(xo[WS(os, 1)]));
3177
V T4K, T5d, T55, T59, T5q, T5F, T5x, T5B, T2f, T5C, T5E, T52, T5a, T5j, T5u;
3180
V T3u, T53, T4J, T54;
3182
V T3g, T3t, T4v, T4I;
3183
T3g = VADD(T2G, T3f);
3184
T3t = VADD(T3p, T3s);
3185
T3u = VFMA(LDK(KP998795456), T3g, VMUL(LDK(KP049067674), T3t));
3186
T53 = VFNMS(LDK(KP049067674), T3g, VMUL(LDK(KP998795456), T3t));
3187
T4v = VADD(T3V, T4u);
3188
T4I = VADD(T4E, T4H);
3189
T4J = VFNMS(LDK(KP049067674), T4I, VMUL(LDK(KP998795456), T4v));
3190
T54 = VFMA(LDK(KP049067674), T4v, VMUL(LDK(KP998795456), T4I));
3192
T4K = VADD(T3u, T4J);
3193
T5d = VSUB(T54, T53);
3194
T55 = VADD(T53, T54);
3195
T59 = VSUB(T4J, T3u);
3198
V T5m, T5v, T5p, T5w;
3200
V T5k, T5l, T5n, T5o;
3201
T5k = VSUB(T2G, T3f);
3202
T5l = VSUB(T3s, T3p);
3203
T5m = VFMA(LDK(KP740951125), T5k, VMUL(LDK(KP671558954), T5l));
3204
T5v = VFNMS(LDK(KP671558954), T5k, VMUL(LDK(KP740951125), T5l));
3205
T5n = VSUB(T3V, T4u);
3206
T5o = VSUB(T4H, T4E);
3207
T5p = VFNMS(LDK(KP671558954), T5o, VMUL(LDK(KP740951125), T5n));
3208
T5w = VFMA(LDK(KP671558954), T5n, VMUL(LDK(KP740951125), T5o));
3210
T5q = VADD(T5m, T5p);
3211
T5F = VSUB(T5w, T5v);
3212
T5x = VADD(T5v, T5w);
3213
T5B = VSUB(T5p, T5m);
3216
V T11, T5h, T4Y, T5s, T2e, T5t, T51, T5i, T10, T4X;
3218
T11 = VADD(Tr, T10);
3219
T5h = VSUB(Tr, T10);
3220
T4X = VADD(T4V, T4W);
3221
T4Y = VADD(T4U, T4X);
3222
T5s = VSUB(T4X, T4U);
3224
V T1C, T2d, T4Z, T50;
3225
T1C = VFMA(LDK(KP098017140), T1s, VMUL(LDK(KP995184726), T1B));
3226
T2d = VFNMS(LDK(KP098017140), T2c, VMUL(LDK(KP995184726), T23));
3227
T2e = VADD(T1C, T2d);
3228
T5t = VSUB(T2d, T1C);
3229
T4Z = VFNMS(LDK(KP098017140), T1B, VMUL(LDK(KP995184726), T1s));
3230
T50 = VFMA(LDK(KP995184726), T2c, VMUL(LDK(KP098017140), T23));
3231
T51 = VADD(T4Z, T50);
3232
T5i = VSUB(T50, T4Z);
3234
T2f = VADD(T11, T2e);
3235
T5C = VSUB(T5t, T5s);
3236
T5E = VSUB(T5h, T5i);
3237
T52 = VADD(T4Y, T51);
3238
T5a = VSUB(T51, T4Y);
3239
T5j = VADD(T5h, T5i);
3240
T5u = VADD(T5s, T5t);
3241
T5c = VSUB(T11, T2e);
3244
V T4L, T56, T5D, T5G;
3245
T4L = VADD(T2f, T4K);
3246
T56 = VBYI(VADD(T52, T55));
3247
ST(&(xo[WS(os, 127)]), VSUB(T4L, T56), ovs, &(xo[WS(os, 1)]));
3248
ST(&(xo[WS(os, 1)]), VADD(T4L, T56), ovs, &(xo[WS(os, 1)]));
3249
T5D = VBYI(VSUB(T5B, T5C));
3250
T5G = VSUB(T5E, T5F);
3251
ST(&(xo[WS(os, 47)]), VADD(T5D, T5G), ovs, &(xo[WS(os, 1)]));
3252
ST(&(xo[WS(os, 81)]), VSUB(T5G, T5D), ovs, &(xo[WS(os, 1)]));
3255
V T5H, T5I, T57, T58;
3256
T5H = VBYI(VADD(T5C, T5B));
3257
T5I = VADD(T5E, T5F);
3258
ST(&(xo[WS(os, 17)]), VADD(T5H, T5I), ovs, &(xo[WS(os, 1)]));
3259
ST(&(xo[WS(os, 111)]), VSUB(T5I, T5H), ovs, &(xo[WS(os, 1)]));
3260
T57 = VSUB(T2f, T4K);
3261
T58 = VBYI(VSUB(T55, T52));
3262
ST(&(xo[WS(os, 65)]), VSUB(T57, T58), ovs, &(xo[WS(os, 1)]));
3263
ST(&(xo[WS(os, 63)]), VADD(T57, T58), ovs, &(xo[WS(os, 1)]));
3266
V T5b, T5e, T5r, T5y;
3267
T5b = VBYI(VSUB(T59, T5a));
3268
T5e = VSUB(T5c, T5d);
3269
ST(&(xo[WS(os, 33)]), VADD(T5b, T5e), ovs, &(xo[WS(os, 1)]));
3270
ST(&(xo[WS(os, 95)]), VSUB(T5e, T5b), ovs, &(xo[WS(os, 1)]));
3271
T5r = VADD(T5j, T5q);
3272
T5y = VBYI(VADD(T5u, T5x));
3273
ST(&(xo[WS(os, 113)]), VSUB(T5r, T5y), ovs, &(xo[WS(os, 1)]));
3274
ST(&(xo[WS(os, 15)]), VADD(T5r, T5y), ovs, &(xo[WS(os, 1)]));
3277
V T5z, T5A, T5f, T5g;
3278
T5z = VSUB(T5j, T5q);
3279
T5A = VBYI(VSUB(T5x, T5u));
3280
ST(&(xo[WS(os, 79)]), VSUB(T5z, T5A), ovs, &(xo[WS(os, 1)]));
3281
ST(&(xo[WS(os, 49)]), VADD(T5z, T5A), ovs, &(xo[WS(os, 1)]));
3282
T5f = VBYI(VADD(T5a, T59));
3283
T5g = VADD(T5c, T5d);
3284
ST(&(xo[WS(os, 31)]), VADD(T5f, T5g), ovs, &(xo[WS(os, 1)]));
3285
ST(&(xo[WS(os, 97)]), VSUB(T5g, T5f), ovs, &(xo[WS(os, 1)]));
3289
V T9i, T9B, T9t, T9x, T9O, Ta3, T9V, T9Z, T93, Ta0, Ta2, T9q, T9y, T9H, T9S;
3292
V T9a, T9r, T9h, T9s;
3294
V T96, T99, T9d, T9g;
3295
T96 = VSUB(T94, T95);
3296
T99 = VSUB(T97, T98);
3297
T9a = VFMA(LDK(KP514102744), T96, VMUL(LDK(KP857728610), T99));
3298
T9r = VFNMS(LDK(KP514102744), T99, VMUL(LDK(KP857728610), T96));
3299
T9d = VSUB(T9b, T9c);
3300
T9g = VSUB(T9e, T9f);
3301
T9h = VFNMS(LDK(KP514102744), T9g, VMUL(LDK(KP857728610), T9d));
3302
T9s = VFMA(LDK(KP857728610), T9g, VMUL(LDK(KP514102744), T9d));
3304
T9i = VADD(T9a, T9h);
3305
T9B = VSUB(T9s, T9r);
3306
T9t = VADD(T9r, T9s);
3307
T9x = VSUB(T9h, T9a);
3310
V T9K, T9T, T9N, T9U;
3312
V T9I, T9J, T9L, T9M;
3313
T9I = VADD(T95, T94);
3314
T9J = VADD(T97, T98);
3315
T9K = VFMA(LDK(KP242980179), T9I, VMUL(LDK(KP970031253), T9J));
3316
T9T = VFNMS(LDK(KP242980179), T9J, VMUL(LDK(KP970031253), T9I));
3317
T9L = VADD(T9b, T9c);
3318
T9M = VADD(T9f, T9e);
3319
T9N = VFNMS(LDK(KP242980179), T9M, VMUL(LDK(KP970031253), T9L));
3320
T9U = VFMA(LDK(KP970031253), T9M, VMUL(LDK(KP242980179), T9L));
3322
T9O = VADD(T9K, T9N);
3323
Ta3 = VSUB(T9U, T9T);
3324
T9V = VADD(T9T, T9U);
3325
T9Z = VSUB(T9N, T9K);
3328
V T8V, T9F, T9m, T9Q, T92, T9R, T9p, T9G, T8U, T9k;
3329
T8U = VSUB(T86, T85);
3330
T8V = VSUB(T8T, T8U);
3331
T9F = VADD(T8T, T8U);
3332
T9k = VSUB(T75, T72);
3333
T9m = VSUB(T9k, T9l);
3334
T9Q = VADD(T9l, T9k);
3336
V T8Y, T91, T9n, T9o;
3337
T8Y = VFMA(LDK(KP471396736), T8W, VMUL(LDK(KP881921264), T8X));
3338
T91 = VFNMS(LDK(KP471396736), T90, VMUL(LDK(KP881921264), T8Z));
3339
T92 = VSUB(T8Y, T91);
3340
T9R = VADD(T91, T8Y);
3341
T9n = VFNMS(LDK(KP471396736), T8X, VMUL(LDK(KP881921264), T8W));
3342
T9o = VFMA(LDK(KP881921264), T90, VMUL(LDK(KP471396736), T8Z));
3343
T9p = VSUB(T9n, T9o);
3344
T9G = VADD(T9o, T9n);
3346
T93 = VADD(T8V, T92);
3347
Ta0 = VSUB(T9R, T9Q);
3348
Ta2 = VSUB(T9F, T9G);
3349
T9q = VADD(T9m, T9p);
3350
T9y = VSUB(T9p, T9m);
3351
T9H = VADD(T9F, T9G);
3352
T9S = VADD(T9Q, T9R);
3353
T9A = VSUB(T8V, T92);
3356
V T9j, T9u, Ta1, Ta4;
3357
T9j = VADD(T93, T9i);
3358
T9u = VBYI(VADD(T9q, T9t));
3359
ST(&(xo[WS(os, 117)]), VSUB(T9j, T9u), ovs, &(xo[WS(os, 1)]));
3360
ST(&(xo[WS(os, 11)]), VADD(T9j, T9u), ovs, &(xo[WS(os, 1)]));
3361
Ta1 = VBYI(VSUB(T9Z, Ta0));
3362
Ta4 = VSUB(Ta2, Ta3);
3363
ST(&(xo[WS(os, 37)]), VADD(Ta1, Ta4), ovs, &(xo[WS(os, 1)]));
3364
ST(&(xo[WS(os, 91)]), VSUB(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
3367
V Ta5, Ta6, T9v, T9w;
3368
Ta5 = VBYI(VADD(Ta0, T9Z));
3369
Ta6 = VADD(Ta2, Ta3);
3370
ST(&(xo[WS(os, 27)]), VADD(Ta5, Ta6), ovs, &(xo[WS(os, 1)]));
3371
ST(&(xo[WS(os, 101)]), VSUB(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
3372
T9v = VSUB(T93, T9i);
3373
T9w = VBYI(VSUB(T9t, T9q));
3374
ST(&(xo[WS(os, 75)]), VSUB(T9v, T9w), ovs, &(xo[WS(os, 1)]));
3375
ST(&(xo[WS(os, 53)]), VADD(T9v, T9w), ovs, &(xo[WS(os, 1)]));
3378
V T9z, T9C, T9P, T9W;
3379
T9z = VBYI(VSUB(T9x, T9y));
3380
T9C = VSUB(T9A, T9B);
3381
ST(&(xo[WS(os, 43)]), VADD(T9z, T9C), ovs, &(xo[WS(os, 1)]));
3382
ST(&(xo[WS(os, 85)]), VSUB(T9C, T9z), ovs, &(xo[WS(os, 1)]));
3383
T9P = VADD(T9H, T9O);
3384
T9W = VBYI(VADD(T9S, T9V));
3385
ST(&(xo[WS(os, 123)]), VSUB(T9P, T9W), ovs, &(xo[WS(os, 1)]));
3386
ST(&(xo[WS(os, 5)]), VADD(T9P, T9W), ovs, &(xo[WS(os, 1)]));
3389
V T9X, T9Y, T9D, T9E;
3390
T9X = VSUB(T9H, T9O);
3391
T9Y = VBYI(VSUB(T9V, T9S));
3392
ST(&(xo[WS(os, 69)]), VSUB(T9X, T9Y), ovs, &(xo[WS(os, 1)]));
3393
ST(&(xo[WS(os, 59)]), VADD(T9X, T9Y), ovs, &(xo[WS(os, 1)]));
3394
T9D = VBYI(VADD(T9y, T9x));
3395
T9E = VADD(T9A, T9B);
3396
ST(&(xo[WS(os, 21)]), VADD(T9D, T9E), ovs, &(xo[WS(os, 1)]));
3397
ST(&(xo[WS(os, 107)]), VSUB(T9E, T9D), ovs, &(xo[WS(os, 1)]));
3401
V T68, T6r, T6j, T6n, T6E, T6T, T6L, T6P, T5T, T6Q, T6S, T6g, T6o, T6x, T6I;
3404
V T60, T6h, T67, T6i;
3406
V T5W, T5Z, T63, T66;
3407
T5W = VSUB(T5U, T5V);
3408
T5Z = VSUB(T5X, T5Y);
3409
T60 = VFMA(LDK(KP427555093), T5W, VMUL(LDK(KP903989293), T5Z));
3410
T6h = VFNMS(LDK(KP427555093), T5Z, VMUL(LDK(KP903989293), T5W));
3411
T63 = VSUB(T61, T62);
3412
T66 = VSUB(T64, T65);
3413
T67 = VFNMS(LDK(KP427555093), T66, VMUL(LDK(KP903989293), T63));
3414
T6i = VFMA(LDK(KP903989293), T66, VMUL(LDK(KP427555093), T63));
3416
T68 = VADD(T60, T67);
3417
T6r = VSUB(T6i, T6h);
3418
T6j = VADD(T6h, T6i);
3419
T6n = VSUB(T67, T60);
3422
V T6A, T6J, T6D, T6K;
3424
V T6y, T6z, T6B, T6C;
3425
T6y = VADD(T5V, T5U);
3426
T6z = VADD(T5X, T5Y);
3427
T6A = VFMA(LDK(KP336889853), T6y, VMUL(LDK(KP941544065), T6z));
3428
T6J = VFNMS(LDK(KP336889853), T6z, VMUL(LDK(KP941544065), T6y));
3429
T6B = VADD(T61, T62);
3430
T6C = VADD(T65, T64);
3431
T6D = VFNMS(LDK(KP336889853), T6C, VMUL(LDK(KP941544065), T6B));
3432
T6K = VFMA(LDK(KP941544065), T6C, VMUL(LDK(KP336889853), T6B));
3434
T6E = VADD(T6A, T6D);
3435
T6T = VSUB(T6K, T6J);
3436
T6L = VADD(T6J, T6K);
3437
T6P = VSUB(T6D, T6A);
3440
V T5L, T6v, T6c, T6G, T5S, T6H, T6f, T6w, T5K, T6a;
3441
T5K = VSUB(T4W, T4V);
3442
T5L = VSUB(T5J, T5K);
3443
T6v = VADD(T5J, T5K);
3445
T6c = VSUB(T6a, T6b);
3446
T6G = VADD(T6b, T6a);
3448
V T5O, T5R, T6d, T6e;
3449
T5O = VFMA(LDK(KP773010453), T5M, VMUL(LDK(KP634393284), T5N));
3450
T5R = VFNMS(LDK(KP634393284), T5Q, VMUL(LDK(KP773010453), T5P));
3451
T5S = VSUB(T5O, T5R);
3452
T6H = VADD(T5R, T5O);
3453
T6d = VFNMS(LDK(KP634393284), T5M, VMUL(LDK(KP773010453), T5N));
3454
T6e = VFMA(LDK(KP634393284), T5P, VMUL(LDK(KP773010453), T5Q));
3455
T6f = VSUB(T6d, T6e);
3456
T6w = VADD(T6e, T6d);
3458
T5T = VADD(T5L, T5S);
3459
T6Q = VSUB(T6H, T6G);
3460
T6S = VSUB(T6v, T6w);
3461
T6g = VADD(T6c, T6f);
3462
T6o = VSUB(T6f, T6c);
3463
T6x = VADD(T6v, T6w);
3464
T6I = VADD(T6G, T6H);
3465
T6q = VSUB(T5L, T5S);
3468
V T69, T6k, T6R, T6U;
3469
T69 = VADD(T5T, T68);
3470
T6k = VBYI(VADD(T6g, T6j));
3471
ST(&(xo[WS(os, 119)]), VSUB(T69, T6k), ovs, &(xo[WS(os, 1)]));
3472
ST(&(xo[WS(os, 9)]), VADD(T69, T6k), ovs, &(xo[WS(os, 1)]));
3473
T6R = VBYI(VSUB(T6P, T6Q));
3474
T6U = VSUB(T6S, T6T);
3475
ST(&(xo[WS(os, 39)]), VADD(T6R, T6U), ovs, &(xo[WS(os, 1)]));
3476
ST(&(xo[WS(os, 89)]), VSUB(T6U, T6R), ovs, &(xo[WS(os, 1)]));
3479
V T6V, T6W, T6l, T6m;
3480
T6V = VBYI(VADD(T6Q, T6P));
3481
T6W = VADD(T6S, T6T);
3482
ST(&(xo[WS(os, 25)]), VADD(T6V, T6W), ovs, &(xo[WS(os, 1)]));
3483
ST(&(xo[WS(os, 103)]), VSUB(T6W, T6V), ovs, &(xo[WS(os, 1)]));
3484
T6l = VSUB(T5T, T68);
3485
T6m = VBYI(VSUB(T6j, T6g));
3486
ST(&(xo[WS(os, 73)]), VSUB(T6l, T6m), ovs, &(xo[WS(os, 1)]));
3487
ST(&(xo[WS(os, 55)]), VADD(T6l, T6m), ovs, &(xo[WS(os, 1)]));
3490
V T6p, T6s, T6F, T6M;
3491
T6p = VBYI(VSUB(T6n, T6o));
3492
T6s = VSUB(T6q, T6r);
3493
ST(&(xo[WS(os, 41)]), VADD(T6p, T6s), ovs, &(xo[WS(os, 1)]));
3494
ST(&(xo[WS(os, 87)]), VSUB(T6s, T6p), ovs, &(xo[WS(os, 1)]));
3495
T6F = VADD(T6x, T6E);
3496
T6M = VBYI(VADD(T6I, T6L));
3497
ST(&(xo[WS(os, 121)]), VSUB(T6F, T6M), ovs, &(xo[WS(os, 1)]));
3498
ST(&(xo[WS(os, 7)]), VADD(T6F, T6M), ovs, &(xo[WS(os, 1)]));
3501
V T6N, T6O, T6t, T6u;
3502
T6N = VSUB(T6x, T6E);
3503
T6O = VBYI(VSUB(T6L, T6I));
3504
ST(&(xo[WS(os, 71)]), VSUB(T6N, T6O), ovs, &(xo[WS(os, 1)]));
3505
ST(&(xo[WS(os, 57)]), VADD(T6N, T6O), ovs, &(xo[WS(os, 1)]));
3506
T6t = VBYI(VADD(T6o, T6n));
3507
T6u = VADD(T6q, T6r);
3508
ST(&(xo[WS(os, 23)]), VADD(T6t, T6u), ovs, &(xo[WS(os, 1)]));
3509
ST(&(xo[WS(os, 105)]), VSUB(T6u, T6t), ovs, &(xo[WS(os, 1)]));
3515
static const kdft_desc desc = { 128, "n1fv_128", {938, 186, 144, 0}, &GENUS, 0, 0, 0, 0 };
3517
void X(codelet_n1fv_128) (planner *p) {
3518
X(kdft_register) (p, n1fv_128, &desc);
3521
#endif /* HAVE_FMA */