comparison src/fftw-3.3.8/dft/simd/common/n1fv_128.c @ 167:bd3cc4d1df30

Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam <cannam@all-day-breakfast.com>
date Tue, 19 Nov 2019 14:52:55 +0000
parents
children
comparison
equal deleted inserted replaced
166:cbd6d7e562c7 167:bd3cc4d1df30
1 /*
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 */
20
21 /* This file was automatically generated --- DO NOT EDIT */
22 /* Generated on Thu May 24 08:04:53 EDT 2018 */
23
24 #include "dft/codelet-dft.h"
25
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27
28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -n 128 -name n1fv_128 -include dft/simd/n1f.h */
29
30 /*
31 * This function contains 1082 FP additions, 642 FP multiplications,
32 * (or, 440 additions, 0 multiplications, 642 fused multiply/add),
33 * 194 stack variables, 31 constants, and 256 memory accesses
34 */
35 #include "dft/simd/n1f.h"
36
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)
38 {
39 DVK(KP941544065, +0.941544065183020778412509402599502357185589796);
40 DVK(KP903989293, +0.903989293123443331586200297230537048710132025);
41 DVK(KP357805721, +0.357805721314524104672487743774474392487532769);
42 DVK(KP472964775, +0.472964775891319928124438237972992463904131113);
43 DVK(KP970031253, +0.970031253194543992603984207286100251456865962);
44 DVK(KP857728610, +0.857728610000272069902269984284770137042490799);
45 DVK(KP250486960, +0.250486960191305461595702160124721208578685568);
46 DVK(KP599376933, +0.599376933681923766271389869014404232837890546);
47 DVK(KP740951125, +0.740951125354959091175616897495162729728955309);
48 DVK(KP998795456, +0.998795456205172392714771604759100694443203615);
49 DVK(KP906347169, +0.906347169019147157946142717268914412664134293);
50 DVK(KP049126849, +0.049126849769467254105343321271313617079695752);
51 DVK(KP803207531, +0.803207531480644909806676512963141923879569427);
52 DVK(KP989176509, +0.989176509964780973451673738016243063983689533);
53 DVK(KP741650546, +0.741650546272035369581266691172079863842265220);
54 DVK(KP148335987, +0.148335987538347428753676511486911367000625355);
55 DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
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(KP668178637, +0.668178637919298919997757686523080761552472251);
66 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
67 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
68 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
69 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
70 {
71 INT i;
72 const R *xi;
73 R *xo;
74 xi = ri;
75 xo = ro;
76 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(256, is), MAKE_VOLATILE_STRIDE(256, os)) {
77 V Tr, T5J, Ted, Tgf, Tfq, TgG, T4U, T6a, T6Z, T8T, Tad, TcZ, Tcc, Td0, T84;
78 V T9k, Tb6, Tbt, T2G, T5U, TeV, Tgt, T3p, T5X, T7B, T97, TeK, Tgq, T7q, T94;
79 V Td8, TdK, TbD, Tc0, T3V, T61, Tfg, TgA, T4E, T64, T7U, T9e, Tf5, Tgx, T7J;
80 V T9b, Tdf, TdN, Td2, Td3, TI, T4V, Tft, Tgg, TZ, T4W, T75, T86, Tek, TgH;
81 V T72, T85, Tas, Tcd, Tdp, Tdq, TdG, Teq, Tgj, Tet, Tgi, T1s, T5N, T1B, T5M;
82 V T7d, T8W, TaI, Tcg, T7a, T8X, Tdm, Tdn, TdH, Tez, Tgm, TeC, Tgl, T23, T5Q;
83 V T2c, T5P, T7k, T8Z, TaX, Tcf, T7h, T90, Tbl, Tbu, Tdb, TdL, TeY, Tgr, TeR;
84 V Tgu, T7x, T98, T7E, T95, T3f, T5Y, T3s, T5V, TbS, Tc1, Tdi, TdO, Tfj, Tgy;
85 V Tfc, TgB, T7Q, T9f, T7X, T9c, T4u, T65, T4H, T62;
86 {
87 V T3, Ta7, T4O, Ta8, Ta, Tab, T4P, Taa, Tc9, Tca, Ti, Tea, T4R, Tc6, Tc7;
88 V Tp, Teb, T4S;
89 {
90 V T1, T2, T4M, T4N;
91 T1 = LD(&(xi[0]), ivs, &(xi[0]));
92 T2 = LD(&(xi[WS(is, 64)]), ivs, &(xi[0]));
93 T3 = VSUB(T1, T2);
94 Ta7 = VADD(T1, T2);
95 T4M = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
96 T4N = LD(&(xi[WS(is, 96)]), ivs, &(xi[0]));
97 T4O = VSUB(T4M, T4N);
98 Ta8 = VADD(T4M, T4N);
99 }
100 {
101 V T4, T5, T6, T7, T8, T9;
102 T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
103 T5 = LD(&(xi[WS(is, 80)]), ivs, &(xi[0]));
104 T6 = VSUB(T4, T5);
105 T7 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0]));
106 T8 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
107 T9 = VSUB(T7, T8);
108 Ta = VADD(T6, T9);
109 Tab = VADD(T7, T8);
110 T4P = VSUB(T9, T6);
111 Taa = VADD(T4, T5);
112 }
113 {
114 V Te, Th, Tl, To;
115 {
116 V Tc, Td, Tf, Tg;
117 Tc = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
118 Td = LD(&(xi[WS(is, 72)]), ivs, &(xi[0]));
119 Te = VSUB(Tc, Td);
120 Tc9 = VADD(Tc, Td);
121 Tf = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
122 Tg = LD(&(xi[WS(is, 104)]), ivs, &(xi[0]));
123 Th = VSUB(Tf, Tg);
124 Tca = VADD(Tf, Tg);
125 }
126 Ti = VFNMS(LDK(KP414213562), Th, Te);
127 Tea = VSUB(Tc9, Tca);
128 T4R = VFMA(LDK(KP414213562), Te, Th);
129 {
130 V Tj, Tk, Tm, Tn;
131 Tj = LD(&(xi[WS(is, 120)]), ivs, &(xi[0]));
132 Tk = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
133 Tl = VSUB(Tj, Tk);
134 Tc6 = VADD(Tj, Tk);
135 Tm = LD(&(xi[WS(is, 88)]), ivs, &(xi[0]));
136 Tn = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
137 To = VSUB(Tm, Tn);
138 Tc7 = VADD(Tn, Tm);
139 }
140 Tp = VFNMS(LDK(KP414213562), To, Tl);
141 Teb = VSUB(Tc6, Tc7);
142 T4S = VFMA(LDK(KP414213562), Tl, To);
143 }
144 {
145 V Tb, Tq, Te9, Tec;
146 Tb = VFMA(LDK(KP707106781), Ta, T3);
147 Tq = VADD(Ti, Tp);
148 Tr = VFMA(LDK(KP923879532), Tq, Tb);
149 T5J = VFNMS(LDK(KP923879532), Tq, Tb);
150 Te9 = VSUB(Ta7, Ta8);
151 Tec = VADD(Tea, Teb);
152 Ted = VFMA(LDK(KP707106781), Tec, Te9);
153 Tgf = VFNMS(LDK(KP707106781), Tec, Te9);
154 }
155 {
156 V Tfo, Tfp, T4Q, T4T;
157 Tfo = VSUB(Tab, Taa);
158 Tfp = VSUB(Teb, Tea);
159 Tfq = VFMA(LDK(KP707106781), Tfp, Tfo);
160 TgG = VFNMS(LDK(KP707106781), Tfp, Tfo);
161 T4Q = VFNMS(LDK(KP707106781), T4P, T4O);
162 T4T = VSUB(T4R, T4S);
163 T4U = VFMA(LDK(KP923879532), T4T, T4Q);
164 T6a = VFNMS(LDK(KP923879532), T4T, T4Q);
165 }
166 {
167 V T6X, T6Y, Ta9, Tac;
168 T6X = VFNMS(LDK(KP707106781), Ta, T3);
169 T6Y = VADD(T4R, T4S);
170 T6Z = VFMA(LDK(KP923879532), T6Y, T6X);
171 T8T = VFNMS(LDK(KP923879532), T6Y, T6X);
172 Ta9 = VADD(Ta7, Ta8);
173 Tac = VADD(Taa, Tab);
174 Tad = VSUB(Ta9, Tac);
175 TcZ = VADD(Ta9, Tac);
176 }
177 {
178 V Tc8, Tcb, T82, T83;
179 Tc8 = VADD(Tc6, Tc7);
180 Tcb = VADD(Tc9, Tca);
181 Tcc = VSUB(Tc8, Tcb);
182 Td0 = VADD(Tcb, Tc8);
183 T82 = VFMA(LDK(KP707106781), T4P, T4O);
184 T83 = VSUB(Tp, Ti);
185 T84 = VFMA(LDK(KP923879532), T83, T82);
186 T9k = VFNMS(LDK(KP923879532), T83, T82);
187 }
188 }
189 {
190 V Tb0, Tb1, T2i, Tb2, T3j, Tb3, Tb4, T2p, Tb5, T3k, T2x, TeH, T3m, Tbp, T2E;
191 V TeI, T3n, Tbs, T2l, T2o, TeG, TeJ;
192 {
193 V T2g, T2h, T3h, T3i;
194 T2g = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
195 T2h = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)]));
196 Tb0 = VADD(T2g, T2h);
197 T3h = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
198 T3i = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)]));
199 Tb1 = VADD(T3h, T3i);
200 T2i = VSUB(T2g, T2h);
201 Tb2 = VADD(Tb0, Tb1);
202 T3j = VSUB(T3h, T3i);
203 }
204 {
205 V T2j, T2k, T2m, T2n;
206 T2j = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
207 T2k = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)]));
208 T2l = VSUB(T2j, T2k);
209 Tb3 = VADD(T2j, T2k);
210 T2m = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)]));
211 T2n = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
212 T2o = VSUB(T2m, T2n);
213 Tb4 = VADD(T2m, T2n);
214 }
215 T2p = VADD(T2l, T2o);
216 Tb5 = VADD(Tb3, Tb4);
217 T3k = VSUB(T2l, T2o);
218 {
219 V T2t, Tbn, T2w, Tbo;
220 {
221 V T2r, T2s, T2u, T2v;
222 T2r = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
223 T2s = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)]));
224 T2t = VSUB(T2r, T2s);
225 Tbn = VADD(T2r, T2s);
226 T2u = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
227 T2v = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)]));
228 T2w = VSUB(T2u, T2v);
229 Tbo = VADD(T2u, T2v);
230 }
231 T2x = VFNMS(LDK(KP414213562), T2w, T2t);
232 TeH = VSUB(Tbn, Tbo);
233 T3m = VFMA(LDK(KP414213562), T2t, T2w);
234 Tbp = VADD(Tbn, Tbo);
235 }
236 {
237 V T2A, Tbq, T2D, Tbr;
238 {
239 V T2y, T2z, T2B, T2C;
240 T2y = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)]));
241 T2z = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
242 T2A = VSUB(T2y, T2z);
243 Tbq = VADD(T2y, T2z);
244 T2B = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
245 T2C = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)]));
246 T2D = VSUB(T2B, T2C);
247 Tbr = VADD(T2B, T2C);
248 }
249 T2E = VFMA(LDK(KP414213562), T2D, T2A);
250 TeI = VSUB(Tbq, Tbr);
251 T3n = VFNMS(LDK(KP414213562), T2A, T2D);
252 Tbs = VADD(Tbq, Tbr);
253 }
254 Tb6 = VSUB(Tb2, Tb5);
255 Tbt = VSUB(Tbp, Tbs);
256 {
257 V T2q, T2F, TeT, TeU;
258 T2q = VFMA(LDK(KP707106781), T2p, T2i);
259 T2F = VADD(T2x, T2E);
260 T2G = VFMA(LDK(KP923879532), T2F, T2q);
261 T5U = VFNMS(LDK(KP923879532), T2F, T2q);
262 TeT = VSUB(Tb3, Tb4);
263 TeU = VSUB(TeH, TeI);
264 TeV = VFMA(LDK(KP707106781), TeU, TeT);
265 Tgt = VFNMS(LDK(KP707106781), TeU, TeT);
266 }
267 {
268 V T3l, T3o, T7z, T7A;
269 T3l = VFMA(LDK(KP707106781), T3k, T3j);
270 T3o = VADD(T3m, T3n);
271 T3p = VFMA(LDK(KP923879532), T3o, T3l);
272 T5X = VFNMS(LDK(KP923879532), T3o, T3l);
273 T7z = VFNMS(LDK(KP707106781), T3k, T3j);
274 T7A = VSUB(T2x, T2E);
275 T7B = VFNMS(LDK(KP923879532), T7A, T7z);
276 T97 = VFMA(LDK(KP923879532), T7A, T7z);
277 }
278 TeG = VSUB(Tb0, Tb1);
279 TeJ = VADD(TeH, TeI);
280 TeK = VFMA(LDK(KP707106781), TeJ, TeG);
281 Tgq = VFNMS(LDK(KP707106781), TeJ, TeG);
282 {
283 V T7o, T7p, Td6, Td7;
284 T7o = VFNMS(LDK(KP707106781), T2p, T2i);
285 T7p = VSUB(T3m, T3n);
286 T7q = VFMA(LDK(KP923879532), T7p, T7o);
287 T94 = VFNMS(LDK(KP923879532), T7p, T7o);
288 Td6 = VADD(Tb2, Tb5);
289 Td7 = VADD(Tbp, Tbs);
290 Td8 = VADD(Td6, Td7);
291 TdK = VSUB(Td6, Td7);
292 }
293 }
294 {
295 V Tbx, Tby, T3x, Tbz, T4y, TbA, TbB, T3E, TbC, T4z, T3M, Tf2, T4B, TbZ, T3T;
296 V Tf3, T4C, TbW, T3A, T3D, Tf1, Tf4;
297 {
298 V T3v, T3w, T4w, T4x;
299 T3v = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)]));
300 T3w = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
301 Tbx = VADD(T3v, T3w);
302 T4w = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)]));
303 T4x = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
304 Tby = VADD(T4x, T4w);
305 T3x = VSUB(T3v, T3w);
306 Tbz = VADD(Tbx, Tby);
307 T4y = VSUB(T4w, T4x);
308 }
309 {
310 V T3y, T3z, T3B, T3C;
311 T3y = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
312 T3z = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)]));
313 T3A = VSUB(T3y, T3z);
314 TbA = VADD(T3y, T3z);
315 T3B = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)]));
316 T3C = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
317 T3D = VSUB(T3B, T3C);
318 TbB = VADD(T3B, T3C);
319 }
320 T3E = VADD(T3A, T3D);
321 TbC = VADD(TbA, TbB);
322 T4z = VSUB(T3D, T3A);
323 {
324 V T3I, TbX, T3L, TbY;
325 {
326 V T3G, T3H, T3J, T3K;
327 T3G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
328 T3H = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)]));
329 T3I = VSUB(T3G, T3H);
330 TbX = VADD(T3G, T3H);
331 T3J = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)]));
332 T3K = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
333 T3L = VSUB(T3J, T3K);
334 TbY = VADD(T3K, T3J);
335 }
336 T3M = VFMA(LDK(KP414213562), T3L, T3I);
337 Tf2 = VSUB(TbX, TbY);
338 T4B = VFNMS(LDK(KP414213562), T3I, T3L);
339 TbZ = VADD(TbX, TbY);
340 }
341 {
342 V T3P, TbU, T3S, TbV;
343 {
344 V T3N, T3O, T3Q, T3R;
345 T3N = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)]));
346 T3O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
347 T3P = VSUB(T3N, T3O);
348 TbU = VADD(T3N, T3O);
349 T3Q = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)]));
350 T3R = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
351 T3S = VSUB(T3Q, T3R);
352 TbV = VADD(T3R, T3Q);
353 }
354 T3T = VFNMS(LDK(KP414213562), T3S, T3P);
355 Tf3 = VSUB(TbU, TbV);
356 T4C = VFMA(LDK(KP414213562), T3P, T3S);
357 TbW = VADD(TbU, TbV);
358 }
359 TbD = VSUB(Tbz, TbC);
360 Tc0 = VSUB(TbW, TbZ);
361 {
362 V T3F, T3U, Tfe, Tff;
363 T3F = VFMA(LDK(KP707106781), T3E, T3x);
364 T3U = VADD(T3M, T3T);
365 T3V = VFMA(LDK(KP923879532), T3U, T3F);
366 T61 = VFNMS(LDK(KP923879532), T3U, T3F);
367 Tfe = VSUB(TbB, TbA);
368 Tff = VSUB(Tf3, Tf2);
369 Tfg = VFMA(LDK(KP707106781), Tff, Tfe);
370 TgA = VFNMS(LDK(KP707106781), Tff, Tfe);
371 }
372 {
373 V T4A, T4D, T7S, T7T;
374 T4A = VFMA(LDK(KP707106781), T4z, T4y);
375 T4D = VADD(T4B, T4C);
376 T4E = VFMA(LDK(KP923879532), T4D, T4A);
377 T64 = VFNMS(LDK(KP923879532), T4D, T4A);
378 T7S = VFNMS(LDK(KP707106781), T4z, T4y);
379 T7T = VSUB(T3T, T3M);
380 T7U = VFNMS(LDK(KP923879532), T7T, T7S);
381 T9e = VFMA(LDK(KP923879532), T7T, T7S);
382 }
383 Tf1 = VSUB(Tbx, Tby);
384 Tf4 = VADD(Tf2, Tf3);
385 Tf5 = VFMA(LDK(KP707106781), Tf4, Tf1);
386 Tgx = VFNMS(LDK(KP707106781), Tf4, Tf1);
387 {
388 V T7H, T7I, Tdd, Tde;
389 T7H = VFNMS(LDK(KP707106781), T3E, T3x);
390 T7I = VSUB(T4C, T4B);
391 T7J = VFMA(LDK(KP923879532), T7I, T7H);
392 T9b = VFNMS(LDK(KP923879532), T7I, T7H);
393 Tdd = VADD(Tbz, TbC);
394 Tde = VADD(TbZ, TbW);
395 Tdf = VADD(Tdd, Tde);
396 TdN = VSUB(Tdd, Tde);
397 }
398 }
399 {
400 V Tu, Tee, TF, Tag, TL, Teh, TW, Tan, TB, Tef, TG, Taj, TS, Tei, TX;
401 V Taq, Teg, Tej;
402 {
403 V Ts, Tt, Tae, TD, TE, Taf;
404 Ts = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
405 Tt = LD(&(xi[WS(is, 68)]), ivs, &(xi[0]));
406 Tae = VADD(Ts, Tt);
407 TD = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
408 TE = LD(&(xi[WS(is, 100)]), ivs, &(xi[0]));
409 Taf = VADD(TD, TE);
410 Tu = VSUB(Ts, Tt);
411 Tee = VSUB(Tae, Taf);
412 TF = VSUB(TD, TE);
413 Tag = VADD(Tae, Taf);
414 }
415 {
416 V TJ, TK, Tal, TU, TV, Tam;
417 TJ = LD(&(xi[WS(is, 124)]), ivs, &(xi[0]));
418 TK = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
419 Tal = VADD(TJ, TK);
420 TU = LD(&(xi[WS(is, 92)]), ivs, &(xi[0]));
421 TV = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
422 Tam = VADD(TV, TU);
423 TL = VSUB(TJ, TK);
424 Teh = VSUB(Tal, Tam);
425 TW = VSUB(TU, TV);
426 Tan = VADD(Tal, Tam);
427 }
428 {
429 V Tx, Tah, TA, Tai;
430 {
431 V Tv, Tw, Ty, Tz;
432 Tv = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
433 Tw = LD(&(xi[WS(is, 84)]), ivs, &(xi[0]));
434 Tx = VSUB(Tv, Tw);
435 Tah = VADD(Tv, Tw);
436 Ty = LD(&(xi[WS(is, 116)]), ivs, &(xi[0]));
437 Tz = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
438 TA = VSUB(Ty, Tz);
439 Tai = VADD(Ty, Tz);
440 }
441 TB = VADD(Tx, TA);
442 Tef = VSUB(Tah, Tai);
443 TG = VSUB(Tx, TA);
444 Taj = VADD(Tah, Tai);
445 }
446 {
447 V TO, Tao, TR, Tap;
448 {
449 V TM, TN, TP, TQ;
450 TM = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
451 TN = LD(&(xi[WS(is, 76)]), ivs, &(xi[0]));
452 TO = VSUB(TM, TN);
453 Tao = VADD(TM, TN);
454 TP = LD(&(xi[WS(is, 108)]), ivs, &(xi[0]));
455 TQ = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
456 TR = VSUB(TP, TQ);
457 Tap = VADD(TP, TQ);
458 }
459 TS = VADD(TO, TR);
460 Tei = VSUB(Tap, Tao);
461 TX = VSUB(TR, TO);
462 Taq = VADD(Tao, Tap);
463 }
464 Td2 = VADD(Tag, Taj);
465 Td3 = VADD(Tan, Taq);
466 {
467 V TC, TH, Tfr, Tfs;
468 TC = VFMA(LDK(KP707106781), TB, Tu);
469 TH = VFMA(LDK(KP707106781), TG, TF);
470 TI = VFNMS(LDK(KP198912367), TH, TC);
471 T4V = VFMA(LDK(KP198912367), TC, TH);
472 Tfr = VFMA(LDK(KP414213562), Teh, Tei);
473 Tfs = VFMA(LDK(KP414213562), Tee, Tef);
474 Tft = VSUB(Tfr, Tfs);
475 Tgg = VADD(Tfs, Tfr);
476 }
477 {
478 V TT, TY, T73, T74;
479 TT = VFMA(LDK(KP707106781), TS, TL);
480 TY = VFMA(LDK(KP707106781), TX, TW);
481 TZ = VFNMS(LDK(KP198912367), TY, TT);
482 T4W = VFMA(LDK(KP198912367), TT, TY);
483 T73 = VFNMS(LDK(KP707106781), TS, TL);
484 T74 = VFNMS(LDK(KP707106781), TX, TW);
485 T75 = VFMA(LDK(KP668178637), T74, T73);
486 T86 = VFNMS(LDK(KP668178637), T73, T74);
487 }
488 Teg = VFNMS(LDK(KP414213562), Tef, Tee);
489 Tej = VFNMS(LDK(KP414213562), Tei, Teh);
490 Tek = VADD(Teg, Tej);
491 TgH = VSUB(Tej, Teg);
492 {
493 V T70, T71, Tak, Tar;
494 T70 = VFNMS(LDK(KP707106781), TB, Tu);
495 T71 = VFNMS(LDK(KP707106781), TG, TF);
496 T72 = VFMA(LDK(KP668178637), T71, T70);
497 T85 = VFNMS(LDK(KP668178637), T70, T71);
498 Tak = VSUB(Tag, Taj);
499 Tar = VSUB(Tan, Taq);
500 Tas = VADD(Tak, Tar);
501 Tcd = VSUB(Tar, Tak);
502 }
503 }
504 {
505 V Tau, Tav, T14, Taw, T1v, Tax, Tay, T1b, Taz, T1w, T1j, Ten, T1y, TaD, T1q;
506 V Teo, T1z, TaG, T17, T1a, Tem, Tep;
507 {
508 V T12, T13, T1t, T1u;
509 T12 = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
510 T13 = LD(&(xi[WS(is, 66)]), ivs, &(xi[0]));
511 Tau = VADD(T12, T13);
512 T1t = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
513 T1u = LD(&(xi[WS(is, 98)]), ivs, &(xi[0]));
514 Tav = VADD(T1t, T1u);
515 T14 = VSUB(T12, T13);
516 Taw = VADD(Tau, Tav);
517 T1v = VSUB(T1t, T1u);
518 }
519 {
520 V T15, T16, T18, T19;
521 T15 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
522 T16 = LD(&(xi[WS(is, 82)]), ivs, &(xi[0]));
523 T17 = VSUB(T15, T16);
524 Tax = VADD(T15, T16);
525 T18 = LD(&(xi[WS(is, 114)]), ivs, &(xi[0]));
526 T19 = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
527 T1a = VSUB(T18, T19);
528 Tay = VADD(T18, T19);
529 }
530 T1b = VADD(T17, T1a);
531 Taz = VADD(Tax, Tay);
532 T1w = VSUB(T17, T1a);
533 {
534 V T1f, TaB, T1i, TaC;
535 {
536 V T1d, T1e, T1g, T1h;
537 T1d = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
538 T1e = LD(&(xi[WS(is, 74)]), ivs, &(xi[0]));
539 T1f = VSUB(T1d, T1e);
540 TaB = VADD(T1d, T1e);
541 T1g = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
542 T1h = LD(&(xi[WS(is, 106)]), ivs, &(xi[0]));
543 T1i = VSUB(T1g, T1h);
544 TaC = VADD(T1g, T1h);
545 }
546 T1j = VFNMS(LDK(KP414213562), T1i, T1f);
547 Ten = VSUB(TaB, TaC);
548 T1y = VFMA(LDK(KP414213562), T1f, T1i);
549 TaD = VADD(TaB, TaC);
550 }
551 {
552 V T1m, TaE, T1p, TaF;
553 {
554 V T1k, T1l, T1n, T1o;
555 T1k = LD(&(xi[WS(is, 122)]), ivs, &(xi[0]));
556 T1l = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
557 T1m = VSUB(T1k, T1l);
558 TaE = VADD(T1k, T1l);
559 T1n = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
560 T1o = LD(&(xi[WS(is, 90)]), ivs, &(xi[0]));
561 T1p = VSUB(T1n, T1o);
562 TaF = VADD(T1n, T1o);
563 }
564 T1q = VFMA(LDK(KP414213562), T1p, T1m);
565 Teo = VSUB(TaE, TaF);
566 T1z = VFNMS(LDK(KP414213562), T1m, T1p);
567 TaG = VADD(TaE, TaF);
568 }
569 Tdp = VADD(Taw, Taz);
570 Tdq = VADD(TaD, TaG);
571 TdG = VSUB(Tdp, Tdq);
572 Tem = VSUB(Tau, Tav);
573 Tep = VADD(Ten, Teo);
574 Teq = VFMA(LDK(KP707106781), Tep, Tem);
575 Tgj = VFNMS(LDK(KP707106781), Tep, Tem);
576 {
577 V Ter, Tes, T1c, T1r;
578 Ter = VSUB(Tax, Tay);
579 Tes = VSUB(Ten, Teo);
580 Tet = VFMA(LDK(KP707106781), Tes, Ter);
581 Tgi = VFNMS(LDK(KP707106781), Tes, Ter);
582 T1c = VFMA(LDK(KP707106781), T1b, T14);
583 T1r = VADD(T1j, T1q);
584 T1s = VFMA(LDK(KP923879532), T1r, T1c);
585 T5N = VFNMS(LDK(KP923879532), T1r, T1c);
586 }
587 {
588 V T1x, T1A, T7b, T7c;
589 T1x = VFMA(LDK(KP707106781), T1w, T1v);
590 T1A = VADD(T1y, T1z);
591 T1B = VFMA(LDK(KP923879532), T1A, T1x);
592 T5M = VFNMS(LDK(KP923879532), T1A, T1x);
593 T7b = VFNMS(LDK(KP707106781), T1w, T1v);
594 T7c = VSUB(T1j, T1q);
595 T7d = VFNMS(LDK(KP923879532), T7c, T7b);
596 T8W = VFMA(LDK(KP923879532), T7c, T7b);
597 }
598 {
599 V TaA, TaH, T78, T79;
600 TaA = VSUB(Taw, Taz);
601 TaH = VSUB(TaD, TaG);
602 TaI = VFNMS(LDK(KP414213562), TaH, TaA);
603 Tcg = VFMA(LDK(KP414213562), TaA, TaH);
604 T78 = VFNMS(LDK(KP707106781), T1b, T14);
605 T79 = VSUB(T1y, T1z);
606 T7a = VFMA(LDK(KP923879532), T79, T78);
607 T8X = VFNMS(LDK(KP923879532), T79, T78);
608 }
609 }
610 {
611 V TaJ, TaK, T1F, TaL, T26, TaM, TaN, T1M, TaO, T27, T1U, Tew, T29, TaV, T21;
612 V Tex, T2a, TaS, T1I, T1L, Tev, Tey;
613 {
614 V T1D, T1E, T24, T25;
615 T1D = LD(&(xi[WS(is, 126)]), ivs, &(xi[0]));
616 T1E = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
617 TaJ = VADD(T1D, T1E);
618 T24 = LD(&(xi[WS(is, 94)]), ivs, &(xi[0]));
619 T25 = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
620 TaK = VADD(T25, T24);
621 T1F = VSUB(T1D, T1E);
622 TaL = VADD(TaJ, TaK);
623 T26 = VSUB(T24, T25);
624 }
625 {
626 V T1G, T1H, T1J, T1K;
627 T1G = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
628 T1H = LD(&(xi[WS(is, 78)]), ivs, &(xi[0]));
629 T1I = VSUB(T1G, T1H);
630 TaM = VADD(T1G, T1H);
631 T1J = LD(&(xi[WS(is, 110)]), ivs, &(xi[0]));
632 T1K = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
633 T1L = VSUB(T1J, T1K);
634 TaN = VADD(T1J, T1K);
635 }
636 T1M = VADD(T1I, T1L);
637 TaO = VADD(TaM, TaN);
638 T27 = VSUB(T1L, T1I);
639 {
640 V T1Q, TaT, T1T, TaU;
641 {
642 V T1O, T1P, T1R, T1S;
643 T1O = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
644 T1P = LD(&(xi[WS(is, 70)]), ivs, &(xi[0]));
645 T1Q = VSUB(T1O, T1P);
646 TaT = VADD(T1O, T1P);
647 T1R = LD(&(xi[WS(is, 102)]), ivs, &(xi[0]));
648 T1S = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
649 T1T = VSUB(T1R, T1S);
650 TaU = VADD(T1S, T1R);
651 }
652 T1U = VFMA(LDK(KP414213562), T1T, T1Q);
653 Tew = VSUB(TaT, TaU);
654 T29 = VFNMS(LDK(KP414213562), T1Q, T1T);
655 TaV = VADD(TaT, TaU);
656 }
657 {
658 V T1X, TaQ, T20, TaR;
659 {
660 V T1V, T1W, T1Y, T1Z;
661 T1V = LD(&(xi[WS(is, 118)]), ivs, &(xi[0]));
662 T1W = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
663 T1X = VSUB(T1V, T1W);
664 TaQ = VADD(T1V, T1W);
665 T1Y = LD(&(xi[WS(is, 86)]), ivs, &(xi[0]));
666 T1Z = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
667 T20 = VSUB(T1Y, T1Z);
668 TaR = VADD(T1Z, T1Y);
669 }
670 T21 = VFNMS(LDK(KP414213562), T20, T1X);
671 Tex = VSUB(TaQ, TaR);
672 T2a = VFMA(LDK(KP414213562), T1X, T20);
673 TaS = VADD(TaQ, TaR);
674 }
675 Tdm = VADD(TaL, TaO);
676 Tdn = VADD(TaV, TaS);
677 TdH = VSUB(Tdm, Tdn);
678 Tev = VSUB(TaJ, TaK);
679 Tey = VADD(Tew, Tex);
680 Tez = VFMA(LDK(KP707106781), Tey, Tev);
681 Tgm = VFNMS(LDK(KP707106781), Tey, Tev);
682 {
683 V TeA, TeB, T1N, T22;
684 TeA = VSUB(TaN, TaM);
685 TeB = VSUB(Tex, Tew);
686 TeC = VFMA(LDK(KP707106781), TeB, TeA);
687 Tgl = VFNMS(LDK(KP707106781), TeB, TeA);
688 T1N = VFMA(LDK(KP707106781), T1M, T1F);
689 T22 = VADD(T1U, T21);
690 T23 = VFMA(LDK(KP923879532), T22, T1N);
691 T5Q = VFNMS(LDK(KP923879532), T22, T1N);
692 }
693 {
694 V T28, T2b, T7i, T7j;
695 T28 = VFMA(LDK(KP707106781), T27, T26);
696 T2b = VADD(T29, T2a);
697 T2c = VFMA(LDK(KP923879532), T2b, T28);
698 T5P = VFNMS(LDK(KP923879532), T2b, T28);
699 T7i = VFNMS(LDK(KP707106781), T27, T26);
700 T7j = VSUB(T21, T1U);
701 T7k = VFNMS(LDK(KP923879532), T7j, T7i);
702 T8Z = VFMA(LDK(KP923879532), T7j, T7i);
703 }
704 {
705 V TaP, TaW, T7f, T7g;
706 TaP = VSUB(TaL, TaO);
707 TaW = VSUB(TaS, TaV);
708 TaX = VFNMS(LDK(KP414213562), TaW, TaP);
709 Tcf = VFMA(LDK(KP414213562), TaP, TaW);
710 T7f = VFNMS(LDK(KP707106781), T1M, T1F);
711 T7g = VSUB(T2a, T29);
712 T7h = VFMA(LDK(KP923879532), T7g, T7f);
713 T90 = VFNMS(LDK(KP923879532), T7g, T7f);
714 }
715 }
716 {
717 V T2J, TeL, T2U, Tb9, T30, TeO, T3b, Tbg, T2Q, TeM, T2V, Tbc, T37, TeP, T3c;
718 V Tbj;
719 {
720 V T2H, T2I, Tb7, T2S, T2T, Tb8;
721 T2H = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
722 T2I = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)]));
723 Tb7 = VADD(T2H, T2I);
724 T2S = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
725 T2T = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)]));
726 Tb8 = VADD(T2S, T2T);
727 T2J = VSUB(T2H, T2I);
728 TeL = VSUB(Tb7, Tb8);
729 T2U = VSUB(T2S, T2T);
730 Tb9 = VADD(Tb7, Tb8);
731 }
732 {
733 V T2Y, T2Z, Tbe, T39, T3a, Tbf;
734 T2Y = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)]));
735 T2Z = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
736 Tbe = VADD(T2Y, T2Z);
737 T39 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
738 T3a = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)]));
739 Tbf = VADD(T39, T3a);
740 T30 = VSUB(T2Y, T2Z);
741 TeO = VSUB(Tbe, Tbf);
742 T3b = VSUB(T39, T3a);
743 Tbg = VADD(Tbe, Tbf);
744 }
745 {
746 V T2M, Tba, T2P, Tbb;
747 {
748 V T2K, T2L, T2N, T2O;
749 T2K = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
750 T2L = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)]));
751 T2M = VSUB(T2K, T2L);
752 Tba = VADD(T2K, T2L);
753 T2N = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)]));
754 T2O = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
755 T2P = VSUB(T2N, T2O);
756 Tbb = VADD(T2N, T2O);
757 }
758 T2Q = VADD(T2M, T2P);
759 TeM = VSUB(Tba, Tbb);
760 T2V = VSUB(T2M, T2P);
761 Tbc = VADD(Tba, Tbb);
762 }
763 {
764 V T33, Tbh, T36, Tbi;
765 {
766 V T31, T32, T34, T35;
767 T31 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
768 T32 = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)]));
769 T33 = VSUB(T31, T32);
770 Tbh = VADD(T31, T32);
771 T34 = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)]));
772 T35 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
773 T36 = VSUB(T34, T35);
774 Tbi = VADD(T34, T35);
775 }
776 T37 = VADD(T33, T36);
777 TeP = VSUB(Tbh, Tbi);
778 T3c = VSUB(T33, T36);
779 Tbj = VADD(Tbh, Tbi);
780 }
781 {
782 V Tbd, Tbk, TeN, TeQ;
783 Tbd = VSUB(Tb9, Tbc);
784 Tbk = VSUB(Tbg, Tbj);
785 Tbl = VADD(Tbd, Tbk);
786 Tbu = VSUB(Tbd, Tbk);
787 {
788 V Td9, Tda, TeW, TeX;
789 Td9 = VADD(Tb9, Tbc);
790 Tda = VADD(Tbg, Tbj);
791 Tdb = VADD(Td9, Tda);
792 TdL = VSUB(Td9, Tda);
793 TeW = VFMA(LDK(KP414213562), TeL, TeM);
794 TeX = VFNMS(LDK(KP414213562), TeO, TeP);
795 TeY = VADD(TeW, TeX);
796 Tgr = VSUB(TeW, TeX);
797 }
798 TeN = VFNMS(LDK(KP414213562), TeM, TeL);
799 TeQ = VFMA(LDK(KP414213562), TeP, TeO);
800 TeR = VADD(TeN, TeQ);
801 Tgu = VSUB(TeN, TeQ);
802 {
803 V T7t, T7C, T7w, T7D;
804 {
805 V T7r, T7s, T7u, T7v;
806 T7r = VFNMS(LDK(KP707106781), T2Q, T2J);
807 T7s = VFNMS(LDK(KP707106781), T2V, T2U);
808 T7t = VFMA(LDK(KP668178637), T7s, T7r);
809 T7C = VFNMS(LDK(KP668178637), T7r, T7s);
810 T7u = VFNMS(LDK(KP707106781), T37, T30);
811 T7v = VFNMS(LDK(KP707106781), T3c, T3b);
812 T7w = VFNMS(LDK(KP668178637), T7v, T7u);
813 T7D = VFMA(LDK(KP668178637), T7u, T7v);
814 }
815 T7x = VADD(T7t, T7w);
816 T98 = VSUB(T7t, T7w);
817 T7E = VADD(T7C, T7D);
818 T95 = VSUB(T7D, T7C);
819 }
820 {
821 V T2X, T3q, T3e, T3r;
822 {
823 V T2R, T2W, T38, T3d;
824 T2R = VFMA(LDK(KP707106781), T2Q, T2J);
825 T2W = VFMA(LDK(KP707106781), T2V, T2U);
826 T2X = VFNMS(LDK(KP198912367), T2W, T2R);
827 T3q = VFMA(LDK(KP198912367), T2R, T2W);
828 T38 = VFMA(LDK(KP707106781), T37, T30);
829 T3d = VFMA(LDK(KP707106781), T3c, T3b);
830 T3e = VFMA(LDK(KP198912367), T3d, T38);
831 T3r = VFNMS(LDK(KP198912367), T38, T3d);
832 }
833 T3f = VADD(T2X, T3e);
834 T5Y = VSUB(T2X, T3e);
835 T3s = VADD(T3q, T3r);
836 T5V = VSUB(T3q, T3r);
837 }
838 }
839 }
840 {
841 V T3Y, Tf6, T49, TbG, T4f, Tf9, T4q, TbN, T45, Tf7, T4a, TbJ, T4m, Tfa, T4r;
842 V TbQ;
843 {
844 V T3W, T3X, TbE, T47, T48, TbF;
845 T3W = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
846 T3X = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)]));
847 TbE = VADD(T3W, T3X);
848 T47 = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)]));
849 T48 = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
850 TbF = VADD(T48, T47);
851 T3Y = VSUB(T3W, T3X);
852 Tf6 = VSUB(TbE, TbF);
853 T49 = VSUB(T47, T48);
854 TbG = VADD(TbE, TbF);
855 }
856 {
857 V T4d, T4e, TbL, T4o, T4p, TbM;
858 T4d = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)]));
859 T4e = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
860 TbL = VADD(T4d, T4e);
861 T4o = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)]));
862 T4p = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
863 TbM = VADD(T4p, T4o);
864 T4f = VSUB(T4d, T4e);
865 Tf9 = VSUB(TbL, TbM);
866 T4q = VSUB(T4o, T4p);
867 TbN = VADD(TbL, TbM);
868 }
869 {
870 V T41, TbH, T44, TbI;
871 {
872 V T3Z, T40, T42, T43;
873 T3Z = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
874 T40 = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)]));
875 T41 = VSUB(T3Z, T40);
876 TbH = VADD(T3Z, T40);
877 T42 = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)]));
878 T43 = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
879 T44 = VSUB(T42, T43);
880 TbI = VADD(T42, T43);
881 }
882 T45 = VADD(T41, T44);
883 Tf7 = VSUB(TbI, TbH);
884 T4a = VSUB(T44, T41);
885 TbJ = VADD(TbH, TbI);
886 }
887 {
888 V T4i, TbO, T4l, TbP;
889 {
890 V T4g, T4h, T4j, T4k;
891 T4g = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
892 T4h = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)]));
893 T4i = VSUB(T4g, T4h);
894 TbO = VADD(T4g, T4h);
895 T4j = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)]));
896 T4k = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
897 T4l = VSUB(T4j, T4k);
898 TbP = VADD(T4j, T4k);
899 }
900 T4m = VADD(T4i, T4l);
901 Tfa = VSUB(TbP, TbO);
902 T4r = VSUB(T4l, T4i);
903 TbQ = VADD(TbO, TbP);
904 }
905 {
906 V TbK, TbR, Tf8, Tfb;
907 TbK = VSUB(TbG, TbJ);
908 TbR = VSUB(TbN, TbQ);
909 TbS = VADD(TbK, TbR);
910 Tc1 = VSUB(TbR, TbK);
911 {
912 V Tdg, Tdh, Tfh, Tfi;
913 Tdg = VADD(TbG, TbJ);
914 Tdh = VADD(TbN, TbQ);
915 Tdi = VADD(Tdg, Tdh);
916 TdO = VSUB(Tdh, Tdg);
917 Tfh = VFNMS(LDK(KP414213562), Tf6, Tf7);
918 Tfi = VFMA(LDK(KP414213562), Tf9, Tfa);
919 Tfj = VADD(Tfh, Tfi);
920 Tgy = VSUB(Tfi, Tfh);
921 }
922 Tf8 = VFMA(LDK(KP414213562), Tf7, Tf6);
923 Tfb = VFNMS(LDK(KP414213562), Tfa, Tf9);
924 Tfc = VADD(Tf8, Tfb);
925 TgB = VSUB(Tfb, Tf8);
926 {
927 V T7M, T7V, T7P, T7W;
928 {
929 V T7K, T7L, T7N, T7O;
930 T7K = VFNMS(LDK(KP707106781), T45, T3Y);
931 T7L = VFNMS(LDK(KP707106781), T4a, T49);
932 T7M = VFNMS(LDK(KP668178637), T7L, T7K);
933 T7V = VFMA(LDK(KP668178637), T7K, T7L);
934 T7N = VFNMS(LDK(KP707106781), T4m, T4f);
935 T7O = VFNMS(LDK(KP707106781), T4r, T4q);
936 T7P = VFMA(LDK(KP668178637), T7O, T7N);
937 T7W = VFNMS(LDK(KP668178637), T7N, T7O);
938 }
939 T7Q = VADD(T7M, T7P);
940 T9f = VSUB(T7P, T7M);
941 T7X = VADD(T7V, T7W);
942 T9c = VSUB(T7V, T7W);
943 }
944 {
945 V T4c, T4F, T4t, T4G;
946 {
947 V T46, T4b, T4n, T4s;
948 T46 = VFMA(LDK(KP707106781), T45, T3Y);
949 T4b = VFMA(LDK(KP707106781), T4a, T49);
950 T4c = VFMA(LDK(KP198912367), T4b, T46);
951 T4F = VFNMS(LDK(KP198912367), T46, T4b);
952 T4n = VFMA(LDK(KP707106781), T4m, T4f);
953 T4s = VFMA(LDK(KP707106781), T4r, T4q);
954 T4t = VFNMS(LDK(KP198912367), T4s, T4n);
955 T4G = VFMA(LDK(KP198912367), T4n, T4s);
956 }
957 T4u = VADD(T4c, T4t);
958 T65 = VSUB(T4t, T4c);
959 T4H = VADD(T4F, T4G);
960 T62 = VSUB(T4G, T4F);
961 }
962 }
963 }
964 {
965 V Td5, Tdx, TdC, TdE, Tdk, Tdt, Tds, Tdy, Tdz, TdD;
966 {
967 V Td1, Td4, TdA, TdB;
968 Td1 = VADD(TcZ, Td0);
969 Td4 = VADD(Td2, Td3);
970 Td5 = VSUB(Td1, Td4);
971 Tdx = VADD(Td1, Td4);
972 TdA = VADD(Td8, Tdb);
973 TdB = VADD(Tdf, Tdi);
974 TdC = VADD(TdA, TdB);
975 TdE = VSUB(TdB, TdA);
976 }
977 {
978 V Tdc, Tdj, Tdo, Tdr;
979 Tdc = VSUB(Td8, Tdb);
980 Tdj = VSUB(Tdf, Tdi);
981 Tdk = VADD(Tdc, Tdj);
982 Tdt = VSUB(Tdj, Tdc);
983 Tdo = VADD(Tdm, Tdn);
984 Tdr = VADD(Tdp, Tdq);
985 Tds = VSUB(Tdo, Tdr);
986 Tdy = VADD(Tdr, Tdo);
987 }
988 Tdz = VADD(Tdx, Tdy);
989 ST(&(xo[WS(os, 64)]), VSUB(Tdz, TdC), ovs, &(xo[0]));
990 ST(&(xo[0]), VADD(Tdz, TdC), ovs, &(xo[0]));
991 TdD = VSUB(Tdx, Tdy);
992 ST(&(xo[WS(os, 96)]), VFNMSI(TdE, TdD), ovs, &(xo[0]));
993 ST(&(xo[WS(os, 32)]), VFMAI(TdE, TdD), ovs, &(xo[0]));
994 {
995 V Tdl, Tdu, Tdv, Tdw;
996 Tdl = VFNMS(LDK(KP707106781), Tdk, Td5);
997 Tdu = VFNMS(LDK(KP707106781), Tdt, Tds);
998 ST(&(xo[WS(os, 48)]), VFNMSI(Tdu, Tdl), ovs, &(xo[0]));
999 ST(&(xo[WS(os, 80)]), VFMAI(Tdu, Tdl), ovs, &(xo[0]));
1000 Tdv = VFMA(LDK(KP707106781), Tdk, Td5);
1001 Tdw = VFMA(LDK(KP707106781), Tdt, Tds);
1002 ST(&(xo[WS(os, 112)]), VFNMSI(Tdw, Tdv), ovs, &(xo[0]));
1003 ST(&(xo[WS(os, 16)]), VFMAI(Tdw, Tdv), ovs, &(xo[0]));
1004 }
1005 }
1006 {
1007 V TdJ, Te1, TdX, Te2, TdQ, Te5, TdU, Te4;
1008 {
1009 V TdF, TdI, TdV, TdW;
1010 TdF = VSUB(TcZ, Td0);
1011 TdI = VADD(TdG, TdH);
1012 TdJ = VFMA(LDK(KP707106781), TdI, TdF);
1013 Te1 = VFNMS(LDK(KP707106781), TdI, TdF);
1014 TdV = VFMA(LDK(KP414213562), TdN, TdO);
1015 TdW = VFMA(LDK(KP414213562), TdK, TdL);
1016 TdX = VSUB(TdV, TdW);
1017 Te2 = VADD(TdW, TdV);
1018 }
1019 {
1020 V TdM, TdP, TdS, TdT;
1021 TdM = VFNMS(LDK(KP414213562), TdL, TdK);
1022 TdP = VFNMS(LDK(KP414213562), TdO, TdN);
1023 TdQ = VADD(TdM, TdP);
1024 Te5 = VSUB(TdP, TdM);
1025 TdS = VSUB(Td3, Td2);
1026 TdT = VSUB(TdH, TdG);
1027 TdU = VFMA(LDK(KP707106781), TdT, TdS);
1028 Te4 = VFNMS(LDK(KP707106781), TdT, TdS);
1029 }
1030 {
1031 V TdR, TdY, Te7, Te8;
1032 TdR = VFNMS(LDK(KP923879532), TdQ, TdJ);
1033 TdY = VFNMS(LDK(KP923879532), TdX, TdU);
1034 ST(&(xo[WS(os, 56)]), VFNMSI(TdY, TdR), ovs, &(xo[0]));
1035 ST(&(xo[WS(os, 72)]), VFMAI(TdY, TdR), ovs, &(xo[0]));
1036 Te7 = VFMA(LDK(KP923879532), Te2, Te1);
1037 Te8 = VFNMS(LDK(KP923879532), Te5, Te4);
1038 ST(&(xo[WS(os, 24)]), VFNMSI(Te8, Te7), ovs, &(xo[0]));
1039 ST(&(xo[WS(os, 104)]), VFMAI(Te8, Te7), ovs, &(xo[0]));
1040 }
1041 {
1042 V TdZ, Te0, Te3, Te6;
1043 TdZ = VFMA(LDK(KP923879532), TdQ, TdJ);
1044 Te0 = VFMA(LDK(KP923879532), TdX, TdU);
1045 ST(&(xo[WS(os, 120)]), VFNMSI(Te0, TdZ), ovs, &(xo[0]));
1046 ST(&(xo[WS(os, 8)]), VFMAI(Te0, TdZ), ovs, &(xo[0]));
1047 Te3 = VFNMS(LDK(KP923879532), Te2, Te1);
1048 Te6 = VFMA(LDK(KP923879532), Te5, Te4);
1049 ST(&(xo[WS(os, 40)]), VFMAI(Te6, Te3), ovs, &(xo[0]));
1050 ST(&(xo[WS(os, 88)]), VFNMSI(Te6, Te3), ovs, &(xo[0]));
1051 }
1052 }
1053 {
1054 V TaZ, Tcp, Tci, Tcs, Tc4, Tct, Tcl, Tcq;
1055 {
1056 V Tat, TaY, Tce, Tch;
1057 Tat = VFMA(LDK(KP707106781), Tas, Tad);
1058 TaY = VADD(TaI, TaX);
1059 TaZ = VFMA(LDK(KP923879532), TaY, Tat);
1060 Tcp = VFNMS(LDK(KP923879532), TaY, Tat);
1061 Tce = VFMA(LDK(KP707106781), Tcd, Tcc);
1062 Tch = VSUB(Tcf, Tcg);
1063 Tci = VFMA(LDK(KP923879532), Tch, Tce);
1064 Tcs = VFNMS(LDK(KP923879532), Tch, Tce);
1065 {
1066 V Tbw, Tck, Tc3, Tcj;
1067 {
1068 V Tbm, Tbv, TbT, Tc2;
1069 Tbm = VFMA(LDK(KP707106781), Tbl, Tb6);
1070 Tbv = VFMA(LDK(KP707106781), Tbu, Tbt);
1071 Tbw = VFNMS(LDK(KP198912367), Tbv, Tbm);
1072 Tck = VFMA(LDK(KP198912367), Tbm, Tbv);
1073 TbT = VFMA(LDK(KP707106781), TbS, TbD);
1074 Tc2 = VFMA(LDK(KP707106781), Tc1, Tc0);
1075 Tc3 = VFNMS(LDK(KP198912367), Tc2, TbT);
1076 Tcj = VFMA(LDK(KP198912367), TbT, Tc2);
1077 }
1078 Tc4 = VADD(Tbw, Tc3);
1079 Tct = VSUB(Tc3, Tbw);
1080 Tcl = VSUB(Tcj, Tck);
1081 Tcq = VADD(Tck, Tcj);
1082 }
1083 }
1084 {
1085 V Tc5, Tcm, Tcv, Tcw;
1086 Tc5 = VFNMS(LDK(KP980785280), Tc4, TaZ);
1087 Tcm = VFNMS(LDK(KP980785280), Tcl, Tci);
1088 ST(&(xo[WS(os, 60)]), VFNMSI(Tcm, Tc5), ovs, &(xo[0]));
1089 ST(&(xo[WS(os, 68)]), VFMAI(Tcm, Tc5), ovs, &(xo[0]));
1090 Tcv = VFMA(LDK(KP980785280), Tcq, Tcp);
1091 Tcw = VFNMS(LDK(KP980785280), Tct, Tcs);
1092 ST(&(xo[WS(os, 28)]), VFNMSI(Tcw, Tcv), ovs, &(xo[0]));
1093 ST(&(xo[WS(os, 100)]), VFMAI(Tcw, Tcv), ovs, &(xo[0]));
1094 }
1095 {
1096 V Tcn, Tco, Tcr, Tcu;
1097 Tcn = VFMA(LDK(KP980785280), Tc4, TaZ);
1098 Tco = VFMA(LDK(KP980785280), Tcl, Tci);
1099 ST(&(xo[WS(os, 124)]), VFNMSI(Tco, Tcn), ovs, &(xo[0]));
1100 ST(&(xo[WS(os, 4)]), VFMAI(Tco, Tcn), ovs, &(xo[0]));
1101 Tcr = VFNMS(LDK(KP980785280), Tcq, Tcp);
1102 Tcu = VFMA(LDK(KP980785280), Tct, Tcs);
1103 ST(&(xo[WS(os, 36)]), VFMAI(Tcu, Tcr), ovs, &(xo[0]));
1104 ST(&(xo[WS(os, 92)]), VFNMSI(Tcu, Tcr), ovs, &(xo[0]));
1105 }
1106 }
1107 {
1108 V Tcz, TcR, TcK, TcU, TcG, TcV, TcN, TcS;
1109 {
1110 V Tcx, Tcy, TcI, TcJ;
1111 Tcx = VFNMS(LDK(KP707106781), Tas, Tad);
1112 Tcy = VADD(Tcg, Tcf);
1113 Tcz = VFMA(LDK(KP923879532), Tcy, Tcx);
1114 TcR = VFNMS(LDK(KP923879532), Tcy, Tcx);
1115 TcI = VFNMS(LDK(KP707106781), Tcd, Tcc);
1116 TcJ = VSUB(TaX, TaI);
1117 TcK = VFNMS(LDK(KP923879532), TcJ, TcI);
1118 TcU = VFMA(LDK(KP923879532), TcJ, TcI);
1119 {
1120 V TcC, TcM, TcF, TcL;
1121 {
1122 V TcA, TcB, TcD, TcE;
1123 TcA = VFNMS(LDK(KP707106781), Tbl, Tb6);
1124 TcB = VFNMS(LDK(KP707106781), Tbu, Tbt);
1125 TcC = VFMA(LDK(KP668178637), TcB, TcA);
1126 TcM = VFNMS(LDK(KP668178637), TcA, TcB);
1127 TcD = VFNMS(LDK(KP707106781), TbS, TbD);
1128 TcE = VFNMS(LDK(KP707106781), Tc1, Tc0);
1129 TcF = VFMA(LDK(KP668178637), TcE, TcD);
1130 TcL = VFNMS(LDK(KP668178637), TcD, TcE);
1131 }
1132 TcG = VADD(TcC, TcF);
1133 TcV = VSUB(TcF, TcC);
1134 TcN = VSUB(TcL, TcM);
1135 TcS = VADD(TcM, TcL);
1136 }
1137 }
1138 {
1139 V TcH, TcO, TcX, TcY;
1140 TcH = VFNMS(LDK(KP831469612), TcG, Tcz);
1141 TcO = VFNMS(LDK(KP831469612), TcN, TcK);
1142 ST(&(xo[WS(os, 76)]), VFNMSI(TcO, TcH), ovs, &(xo[0]));
1143 ST(&(xo[WS(os, 52)]), VFMAI(TcO, TcH), ovs, &(xo[0]));
1144 TcX = VFNMS(LDK(KP831469612), TcS, TcR);
1145 TcY = VFMA(LDK(KP831469612), TcV, TcU);
1146 ST(&(xo[WS(os, 20)]), VFMAI(TcY, TcX), ovs, &(xo[0]));
1147 ST(&(xo[WS(os, 108)]), VFNMSI(TcY, TcX), ovs, &(xo[0]));
1148 }
1149 {
1150 V TcP, TcQ, TcT, TcW;
1151 TcP = VFMA(LDK(KP831469612), TcG, Tcz);
1152 TcQ = VFMA(LDK(KP831469612), TcN, TcK);
1153 ST(&(xo[WS(os, 12)]), VFNMSI(TcQ, TcP), ovs, &(xo[0]));
1154 ST(&(xo[WS(os, 116)]), VFMAI(TcQ, TcP), ovs, &(xo[0]));
1155 TcT = VFMA(LDK(KP831469612), TcS, TcR);
1156 TcW = VFNMS(LDK(KP831469612), TcV, TcU);
1157 ST(&(xo[WS(os, 44)]), VFNMSI(TcW, TcT), ovs, &(xo[0]));
1158 ST(&(xo[WS(os, 84)]), VFMAI(TcW, TcT), ovs, &(xo[0]));
1159 }
1160 }
1161 {
1162 V TeF, Tga, TfF, Tg0, Tfy, Tg7, TfI, TfP, Tfm, TfJ, TfB, TfG, TfW, Tgb, Tg3;
1163 V Tg8;
1164 {
1165 V Tel, TfY, TeE, TfZ, Teu, TeD;
1166 Tel = VFMA(LDK(KP923879532), Tek, Ted);
1167 TfY = VFNMS(LDK(KP923879532), Tft, Tfq);
1168 Teu = VFNMS(LDK(KP198912367), Tet, Teq);
1169 TeD = VFNMS(LDK(KP198912367), TeC, Tez);
1170 TeE = VADD(Teu, TeD);
1171 TfZ = VSUB(TeD, Teu);
1172 TeF = VFMA(LDK(KP980785280), TeE, Tel);
1173 Tga = VFMA(LDK(KP980785280), TfZ, TfY);
1174 TfF = VFNMS(LDK(KP980785280), TeE, Tel);
1175 Tg0 = VFNMS(LDK(KP980785280), TfZ, TfY);
1176 }
1177 {
1178 V Tfu, TfN, Tfx, TfO, Tfv, Tfw;
1179 Tfu = VFMA(LDK(KP923879532), Tft, Tfq);
1180 TfN = VFNMS(LDK(KP923879532), Tek, Ted);
1181 Tfv = VFMA(LDK(KP198912367), Tez, TeC);
1182 Tfw = VFMA(LDK(KP198912367), Teq, Tet);
1183 Tfx = VSUB(Tfv, Tfw);
1184 TfO = VADD(Tfw, Tfv);
1185 Tfy = VFMA(LDK(KP980785280), Tfx, Tfu);
1186 Tg7 = VFNMS(LDK(KP980785280), TfO, TfN);
1187 TfI = VFNMS(LDK(KP980785280), Tfx, Tfu);
1188 TfP = VFMA(LDK(KP980785280), TfO, TfN);
1189 }
1190 {
1191 V Tf0, TfA, Tfl, Tfz;
1192 {
1193 V TeS, TeZ, Tfd, Tfk;
1194 TeS = VFMA(LDK(KP923879532), TeR, TeK);
1195 TeZ = VFMA(LDK(KP923879532), TeY, TeV);
1196 Tf0 = VFNMS(LDK(KP098491403), TeZ, TeS);
1197 TfA = VFMA(LDK(KP098491403), TeS, TeZ);
1198 Tfd = VFMA(LDK(KP923879532), Tfc, Tf5);
1199 Tfk = VFMA(LDK(KP923879532), Tfj, Tfg);
1200 Tfl = VFNMS(LDK(KP098491403), Tfk, Tfd);
1201 Tfz = VFMA(LDK(KP098491403), Tfd, Tfk);
1202 }
1203 Tfm = VADD(Tf0, Tfl);
1204 TfJ = VSUB(Tfl, Tf0);
1205 TfB = VSUB(Tfz, TfA);
1206 TfG = VADD(TfA, Tfz);
1207 }
1208 {
1209 V TfS, Tg2, TfV, Tg1;
1210 {
1211 V TfQ, TfR, TfT, TfU;
1212 TfQ = VFNMS(LDK(KP923879532), TeR, TeK);
1213 TfR = VFNMS(LDK(KP923879532), TeY, TeV);
1214 TfS = VFMA(LDK(KP820678790), TfR, TfQ);
1215 Tg2 = VFNMS(LDK(KP820678790), TfQ, TfR);
1216 TfT = VFNMS(LDK(KP923879532), Tfc, Tf5);
1217 TfU = VFNMS(LDK(KP923879532), Tfj, Tfg);
1218 TfV = VFMA(LDK(KP820678790), TfU, TfT);
1219 Tg1 = VFNMS(LDK(KP820678790), TfT, TfU);
1220 }
1221 TfW = VADD(TfS, TfV);
1222 Tgb = VSUB(TfV, TfS);
1223 Tg3 = VSUB(Tg1, Tg2);
1224 Tg8 = VADD(Tg2, Tg1);
1225 }
1226 {
1227 V Tfn, TfC, Tg9, Tgc;
1228 Tfn = VFNMS(LDK(KP995184726), Tfm, TeF);
1229 TfC = VFNMS(LDK(KP995184726), TfB, Tfy);
1230 ST(&(xo[WS(os, 62)]), VFNMSI(TfC, Tfn), ovs, &(xo[0]));
1231 ST(&(xo[WS(os, 66)]), VFMAI(TfC, Tfn), ovs, &(xo[0]));
1232 Tg9 = VFMA(LDK(KP773010453), Tg8, Tg7);
1233 Tgc = VFNMS(LDK(KP773010453), Tgb, Tga);
1234 ST(&(xo[WS(os, 46)]), VFNMSI(Tgc, Tg9), ovs, &(xo[0]));
1235 ST(&(xo[WS(os, 82)]), VFMAI(Tgc, Tg9), ovs, &(xo[0]));
1236 }
1237 {
1238 V Tgd, Tge, TfD, TfE;
1239 Tgd = VFNMS(LDK(KP773010453), Tg8, Tg7);
1240 Tge = VFMA(LDK(KP773010453), Tgb, Tga);
1241 ST(&(xo[WS(os, 18)]), VFMAI(Tge, Tgd), ovs, &(xo[0]));
1242 ST(&(xo[WS(os, 110)]), VFNMSI(Tge, Tgd), ovs, &(xo[0]));
1243 TfD = VFMA(LDK(KP995184726), Tfm, TeF);
1244 TfE = VFMA(LDK(KP995184726), TfB, Tfy);
1245 ST(&(xo[WS(os, 126)]), VFNMSI(TfE, TfD), ovs, &(xo[0]));
1246 ST(&(xo[WS(os, 2)]), VFMAI(TfE, TfD), ovs, &(xo[0]));
1247 }
1248 {
1249 V TfH, TfK, TfX, Tg4;
1250 TfH = VFNMS(LDK(KP995184726), TfG, TfF);
1251 TfK = VFMA(LDK(KP995184726), TfJ, TfI);
1252 ST(&(xo[WS(os, 34)]), VFMAI(TfK, TfH), ovs, &(xo[0]));
1253 ST(&(xo[WS(os, 94)]), VFNMSI(TfK, TfH), ovs, &(xo[0]));
1254 TfX = VFNMS(LDK(KP773010453), TfW, TfP);
1255 Tg4 = VFNMS(LDK(KP773010453), Tg3, Tg0);
1256 ST(&(xo[WS(os, 78)]), VFNMSI(Tg4, TfX), ovs, &(xo[0]));
1257 ST(&(xo[WS(os, 50)]), VFMAI(Tg4, TfX), ovs, &(xo[0]));
1258 }
1259 {
1260 V Tg5, Tg6, TfL, TfM;
1261 Tg5 = VFMA(LDK(KP773010453), TfW, TfP);
1262 Tg6 = VFMA(LDK(KP773010453), Tg3, Tg0);
1263 ST(&(xo[WS(os, 14)]), VFNMSI(Tg6, Tg5), ovs, &(xo[0]));
1264 ST(&(xo[WS(os, 114)]), VFMAI(Tg6, Tg5), ovs, &(xo[0]));
1265 TfL = VFMA(LDK(KP995184726), TfG, TfF);
1266 TfM = VFNMS(LDK(KP995184726), TfJ, TfI);
1267 ST(&(xo[WS(os, 30)]), VFNMSI(TfM, TfL), ovs, &(xo[0]));
1268 ST(&(xo[WS(os, 98)]), VFMAI(TfM, TfL), ovs, &(xo[0]));
1269 }
1270 }
1271 {
1272 V Tgp, Tho, TgT, The, TgM, Thl, TgW, Th3, TgE, TgX, TgP, TgU, Tha, Thp, Thh;
1273 V Thm;
1274 {
1275 V Tgh, Thc, Tgo, Thd, Tgk, Tgn;
1276 Tgh = VFNMS(LDK(KP923879532), Tgg, Tgf);
1277 Thc = VFNMS(LDK(KP923879532), TgH, TgG);
1278 Tgk = VFNMS(LDK(KP668178637), Tgj, Tgi);
1279 Tgn = VFNMS(LDK(KP668178637), Tgm, Tgl);
1280 Tgo = VADD(Tgk, Tgn);
1281 Thd = VSUB(Tgn, Tgk);
1282 Tgp = VFNMS(LDK(KP831469612), Tgo, Tgh);
1283 Tho = VFNMS(LDK(KP831469612), Thd, Thc);
1284 TgT = VFMA(LDK(KP831469612), Tgo, Tgh);
1285 The = VFMA(LDK(KP831469612), Thd, Thc);
1286 }
1287 {
1288 V TgI, Th1, TgL, Th2, TgJ, TgK;
1289 TgI = VFMA(LDK(KP923879532), TgH, TgG);
1290 Th1 = VFMA(LDK(KP923879532), Tgg, Tgf);
1291 TgJ = VFMA(LDK(KP668178637), Tgl, Tgm);
1292 TgK = VFMA(LDK(KP668178637), Tgi, Tgj);
1293 TgL = VSUB(TgJ, TgK);
1294 Th2 = VADD(TgK, TgJ);
1295 TgM = VFMA(LDK(KP831469612), TgL, TgI);
1296 Thl = VFNMS(LDK(KP831469612), Th2, Th1);
1297 TgW = VFNMS(LDK(KP831469612), TgL, TgI);
1298 Th3 = VFMA(LDK(KP831469612), Th2, Th1);
1299 }
1300 {
1301 V Tgw, TgO, TgD, TgN;
1302 {
1303 V Tgs, Tgv, Tgz, TgC;
1304 Tgs = VFNMS(LDK(KP923879532), Tgr, Tgq);
1305 Tgv = VFMA(LDK(KP923879532), Tgu, Tgt);
1306 Tgw = VFNMS(LDK(KP534511135), Tgv, Tgs);
1307 TgO = VFMA(LDK(KP534511135), Tgs, Tgv);
1308 Tgz = VFNMS(LDK(KP923879532), Tgy, Tgx);
1309 TgC = VFMA(LDK(KP923879532), TgB, TgA);
1310 TgD = VFNMS(LDK(KP534511135), TgC, Tgz);
1311 TgN = VFMA(LDK(KP534511135), Tgz, TgC);
1312 }
1313 TgE = VADD(Tgw, TgD);
1314 TgX = VSUB(TgD, Tgw);
1315 TgP = VSUB(TgN, TgO);
1316 TgU = VADD(TgO, TgN);
1317 }
1318 {
1319 V Th6, Thg, Th9, Thf;
1320 {
1321 V Th4, Th5, Th7, Th8;
1322 Th4 = VFMA(LDK(KP923879532), Tgr, Tgq);
1323 Th5 = VFNMS(LDK(KP923879532), Tgu, Tgt);
1324 Th6 = VFMA(LDK(KP303346683), Th5, Th4);
1325 Thg = VFNMS(LDK(KP303346683), Th4, Th5);
1326 Th7 = VFMA(LDK(KP923879532), Tgy, Tgx);
1327 Th8 = VFNMS(LDK(KP923879532), TgB, TgA);
1328 Th9 = VFMA(LDK(KP303346683), Th8, Th7);
1329 Thf = VFNMS(LDK(KP303346683), Th7, Th8);
1330 }
1331 Tha = VADD(Th6, Th9);
1332 Thp = VSUB(Th9, Th6);
1333 Thh = VSUB(Thf, Thg);
1334 Thm = VADD(Thg, Thf);
1335 }
1336 {
1337 V TgF, TgQ, Thn, Thq;
1338 TgF = VFNMS(LDK(KP881921264), TgE, Tgp);
1339 TgQ = VFNMS(LDK(KP881921264), TgP, TgM);
1340 ST(&(xo[WS(os, 54)]), VFNMSI(TgQ, TgF), ovs, &(xo[0]));
1341 ST(&(xo[WS(os, 74)]), VFMAI(TgQ, TgF), ovs, &(xo[0]));
1342 Thn = VFMA(LDK(KP956940335), Thm, Thl);
1343 Thq = VFNMS(LDK(KP956940335), Thp, Tho);
1344 ST(&(xo[WS(os, 38)]), VFNMSI(Thq, Thn), ovs, &(xo[0]));
1345 ST(&(xo[WS(os, 90)]), VFMAI(Thq, Thn), ovs, &(xo[0]));
1346 }
1347 {
1348 V Thr, Ths, TgR, TgS;
1349 Thr = VFNMS(LDK(KP956940335), Thm, Thl);
1350 Ths = VFMA(LDK(KP956940335), Thp, Tho);
1351 ST(&(xo[WS(os, 26)]), VFMAI(Ths, Thr), ovs, &(xo[0]));
1352 ST(&(xo[WS(os, 102)]), VFNMSI(Ths, Thr), ovs, &(xo[0]));
1353 TgR = VFMA(LDK(KP881921264), TgE, Tgp);
1354 TgS = VFMA(LDK(KP881921264), TgP, TgM);
1355 ST(&(xo[WS(os, 118)]), VFNMSI(TgS, TgR), ovs, &(xo[0]));
1356 ST(&(xo[WS(os, 10)]), VFMAI(TgS, TgR), ovs, &(xo[0]));
1357 }
1358 {
1359 V TgV, TgY, Thb, Thi;
1360 TgV = VFNMS(LDK(KP881921264), TgU, TgT);
1361 TgY = VFMA(LDK(KP881921264), TgX, TgW);
1362 ST(&(xo[WS(os, 42)]), VFMAI(TgY, TgV), ovs, &(xo[0]));
1363 ST(&(xo[WS(os, 86)]), VFNMSI(TgY, TgV), ovs, &(xo[0]));
1364 Thb = VFNMS(LDK(KP956940335), Tha, Th3);
1365 Thi = VFNMS(LDK(KP956940335), Thh, The);
1366 ST(&(xo[WS(os, 70)]), VFNMSI(Thi, Thb), ovs, &(xo[0]));
1367 ST(&(xo[WS(os, 58)]), VFMAI(Thi, Thb), ovs, &(xo[0]));
1368 }
1369 {
1370 V Thj, Thk, TgZ, Th0;
1371 Thj = VFMA(LDK(KP956940335), Tha, Th3);
1372 Thk = VFMA(LDK(KP956940335), Thh, The);
1373 ST(&(xo[WS(os, 6)]), VFNMSI(Thk, Thj), ovs, &(xo[0]));
1374 ST(&(xo[WS(os, 122)]), VFMAI(Thk, Thj), ovs, &(xo[0]));
1375 TgZ = VFMA(LDK(KP881921264), TgU, TgT);
1376 Th0 = VFNMS(LDK(KP881921264), TgX, TgW);
1377 ST(&(xo[WS(os, 22)]), VFNMSI(Th0, TgZ), ovs, &(xo[0]));
1378 ST(&(xo[WS(os, 106)]), VFMAI(Th0, TgZ), ovs, &(xo[0]));
1379 }
1380 }
1381 {
1382 V T80, T8n, T8f, T8k, T8A, T8P, T8H, T8M, T7n, T8L, T8O, T8c, T8j, T8t, T8E;
1383 V T8m;
1384 {
1385 V T7G, T8d, T7Z, T8e;
1386 {
1387 V T7y, T7F, T7R, T7Y;
1388 T7y = VFMA(LDK(KP831469612), T7x, T7q);
1389 T7F = VFMA(LDK(KP831469612), T7E, T7B);
1390 T7G = VFMA(LDK(KP148335987), T7F, T7y);
1391 T8d = VFNMS(LDK(KP148335987), T7y, T7F);
1392 T7R = VFMA(LDK(KP831469612), T7Q, T7J);
1393 T7Y = VFMA(LDK(KP831469612), T7X, T7U);
1394 T7Z = VFMA(LDK(KP148335987), T7Y, T7R);
1395 T8e = VFNMS(LDK(KP148335987), T7R, T7Y);
1396 }
1397 T80 = VADD(T7G, T7Z);
1398 T8n = VSUB(T7Z, T7G);
1399 T8f = VSUB(T8d, T8e);
1400 T8k = VADD(T8d, T8e);
1401 }
1402 {
1403 V T8w, T8F, T8z, T8G;
1404 {
1405 V T8u, T8v, T8x, T8y;
1406 T8u = VFNMS(LDK(KP831469612), T7x, T7q);
1407 T8v = VFNMS(LDK(KP831469612), T7E, T7B);
1408 T8w = VFNMS(LDK(KP741650546), T8v, T8u);
1409 T8F = VFMA(LDK(KP741650546), T8u, T8v);
1410 T8x = VFNMS(LDK(KP831469612), T7Q, T7J);
1411 T8y = VFNMS(LDK(KP831469612), T7X, T7U);
1412 T8z = VFNMS(LDK(KP741650546), T8y, T8x);
1413 T8G = VFMA(LDK(KP741650546), T8x, T8y);
1414 }
1415 T8A = VADD(T8w, T8z);
1416 T8P = VSUB(T8z, T8w);
1417 T8H = VSUB(T8F, T8G);
1418 T8M = VADD(T8F, T8G);
1419 }
1420 {
1421 V T77, T8r, T88, T8C, T7m, T8D, T8b, T8s, T76, T87;
1422 T76 = VADD(T72, T75);
1423 T77 = VFMA(LDK(KP831469612), T76, T6Z);
1424 T8r = VFNMS(LDK(KP831469612), T76, T6Z);
1425 T87 = VSUB(T85, T86);
1426 T88 = VFMA(LDK(KP831469612), T87, T84);
1427 T8C = VFNMS(LDK(KP831469612), T87, T84);
1428 {
1429 V T7e, T7l, T89, T8a;
1430 T7e = VFMA(LDK(KP303346683), T7d, T7a);
1431 T7l = VFMA(LDK(KP303346683), T7k, T7h);
1432 T7m = VADD(T7e, T7l);
1433 T8D = VSUB(T7l, T7e);
1434 T89 = VFNMS(LDK(KP303346683), T7a, T7d);
1435 T8a = VFNMS(LDK(KP303346683), T7h, T7k);
1436 T8b = VSUB(T89, T8a);
1437 T8s = VADD(T89, T8a);
1438 }
1439 T7n = VFMA(LDK(KP956940335), T7m, T77);
1440 T8L = VFMA(LDK(KP956940335), T8s, T8r);
1441 T8O = VFMA(LDK(KP956940335), T8D, T8C);
1442 T8c = VFMA(LDK(KP956940335), T8b, T88);
1443 T8j = VFNMS(LDK(KP956940335), T7m, T77);
1444 T8t = VFNMS(LDK(KP956940335), T8s, T8r);
1445 T8E = VFNMS(LDK(KP956940335), T8D, T8C);
1446 T8m = VFNMS(LDK(KP956940335), T8b, T88);
1447 }
1448 {
1449 V T81, T8g, T8N, T8Q;
1450 T81 = VFNMS(LDK(KP989176509), T80, T7n);
1451 T8g = VFNMS(LDK(KP989176509), T8f, T8c);
1452 ST(&(xo[WS(os, 61)]), VFNMSI(T8g, T81), ovs, &(xo[WS(os, 1)]));
1453 ST(&(xo[WS(os, 67)]), VFMAI(T8g, T81), ovs, &(xo[WS(os, 1)]));
1454 T8N = VFNMS(LDK(KP803207531), T8M, T8L);
1455 T8Q = VFNMS(LDK(KP803207531), T8P, T8O);
1456 ST(&(xo[WS(os, 45)]), VFNMSI(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
1457 ST(&(xo[WS(os, 83)]), VFMAI(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
1458 }
1459 {
1460 V T8R, T8S, T8h, T8i;
1461 T8R = VFMA(LDK(KP803207531), T8M, T8L);
1462 T8S = VFMA(LDK(KP803207531), T8P, T8O);
1463 ST(&(xo[WS(os, 19)]), VFMAI(T8S, T8R), ovs, &(xo[WS(os, 1)]));
1464 ST(&(xo[WS(os, 109)]), VFNMSI(T8S, T8R), ovs, &(xo[WS(os, 1)]));
1465 T8h = VFMA(LDK(KP989176509), T80, T7n);
1466 T8i = VFMA(LDK(KP989176509), T8f, T8c);
1467 ST(&(xo[WS(os, 125)]), VFNMSI(T8i, T8h), ovs, &(xo[WS(os, 1)]));
1468 ST(&(xo[WS(os, 3)]), VFMAI(T8i, T8h), ovs, &(xo[WS(os, 1)]));
1469 }
1470 {
1471 V T8l, T8o, T8B, T8I;
1472 T8l = VFMA(LDK(KP989176509), T8k, T8j);
1473 T8o = VFMA(LDK(KP989176509), T8n, T8m);
1474 ST(&(xo[WS(os, 35)]), VFMAI(T8o, T8l), ovs, &(xo[WS(os, 1)]));
1475 ST(&(xo[WS(os, 93)]), VFNMSI(T8o, T8l), ovs, &(xo[WS(os, 1)]));
1476 T8B = VFNMS(LDK(KP803207531), T8A, T8t);
1477 T8I = VFNMS(LDK(KP803207531), T8H, T8E);
1478 ST(&(xo[WS(os, 77)]), VFNMSI(T8I, T8B), ovs, &(xo[WS(os, 1)]));
1479 ST(&(xo[WS(os, 51)]), VFMAI(T8I, T8B), ovs, &(xo[WS(os, 1)]));
1480 }
1481 {
1482 V T8J, T8K, T8p, T8q;
1483 T8J = VFMA(LDK(KP803207531), T8A, T8t);
1484 T8K = VFMA(LDK(KP803207531), T8H, T8E);
1485 ST(&(xo[WS(os, 13)]), VFNMSI(T8K, T8J), ovs, &(xo[WS(os, 1)]));
1486 ST(&(xo[WS(os, 115)]), VFMAI(T8K, T8J), ovs, &(xo[WS(os, 1)]));
1487 T8p = VFNMS(LDK(KP989176509), T8k, T8j);
1488 T8q = VFNMS(LDK(KP989176509), T8n, T8m);
1489 ST(&(xo[WS(os, 29)]), VFNMSI(T8q, T8p), ovs, &(xo[WS(os, 1)]));
1490 ST(&(xo[WS(os, 99)]), VFMAI(T8q, T8p), ovs, &(xo[WS(os, 1)]));
1491 }
1492 }
1493 {
1494 V T4K, T5d, T55, T5a, T5q, T5F, T5x, T5C, T2f, T5B, T5E, T52, T59, T5j, T5u;
1495 V T5c;
1496 {
1497 V T3u, T53, T4J, T54;
1498 {
1499 V T3g, T3t, T4v, T4I;
1500 T3g = VFMA(LDK(KP980785280), T3f, T2G);
1501 T3t = VFMA(LDK(KP980785280), T3s, T3p);
1502 T3u = VFNMS(LDK(KP049126849), T3t, T3g);
1503 T53 = VFMA(LDK(KP049126849), T3g, T3t);
1504 T4v = VFMA(LDK(KP980785280), T4u, T3V);
1505 T4I = VFMA(LDK(KP980785280), T4H, T4E);
1506 T4J = VFNMS(LDK(KP049126849), T4I, T4v);
1507 T54 = VFMA(LDK(KP049126849), T4v, T4I);
1508 }
1509 T4K = VADD(T3u, T4J);
1510 T5d = VSUB(T4J, T3u);
1511 T55 = VSUB(T53, T54);
1512 T5a = VADD(T53, T54);
1513 }
1514 {
1515 V T5m, T5v, T5p, T5w;
1516 {
1517 V T5k, T5l, T5n, T5o;
1518 T5k = VFNMS(LDK(KP980785280), T3f, T2G);
1519 T5l = VFNMS(LDK(KP980785280), T3s, T3p);
1520 T5m = VFMA(LDK(KP906347169), T5l, T5k);
1521 T5v = VFNMS(LDK(KP906347169), T5k, T5l);
1522 T5n = VFNMS(LDK(KP980785280), T4u, T3V);
1523 T5o = VFNMS(LDK(KP980785280), T4H, T4E);
1524 T5p = VFMA(LDK(KP906347169), T5o, T5n);
1525 T5w = VFNMS(LDK(KP906347169), T5n, T5o);
1526 }
1527 T5q = VADD(T5m, T5p);
1528 T5F = VSUB(T5p, T5m);
1529 T5x = VSUB(T5v, T5w);
1530 T5C = VADD(T5v, T5w);
1531 }
1532 {
1533 V T11, T5h, T4Y, T5s, T2e, T5t, T51, T5i, T10, T4X;
1534 T10 = VADD(TI, TZ);
1535 T11 = VFMA(LDK(KP980785280), T10, Tr);
1536 T5h = VFNMS(LDK(KP980785280), T10, Tr);
1537 T4X = VSUB(T4V, T4W);
1538 T4Y = VFMA(LDK(KP980785280), T4X, T4U);
1539 T5s = VFNMS(LDK(KP980785280), T4X, T4U);
1540 {
1541 V T1C, T2d, T4Z, T50;
1542 T1C = VFNMS(LDK(KP098491403), T1B, T1s);
1543 T2d = VFNMS(LDK(KP098491403), T2c, T23);
1544 T2e = VADD(T1C, T2d);
1545 T5t = VSUB(T2d, T1C);
1546 T4Z = VFMA(LDK(KP098491403), T1s, T1B);
1547 T50 = VFMA(LDK(KP098491403), T23, T2c);
1548 T51 = VSUB(T4Z, T50);
1549 T5i = VADD(T4Z, T50);
1550 }
1551 T2f = VFMA(LDK(KP995184726), T2e, T11);
1552 T5B = VFNMS(LDK(KP995184726), T5i, T5h);
1553 T5E = VFNMS(LDK(KP995184726), T5t, T5s);
1554 T52 = VFMA(LDK(KP995184726), T51, T4Y);
1555 T59 = VFNMS(LDK(KP995184726), T2e, T11);
1556 T5j = VFMA(LDK(KP995184726), T5i, T5h);
1557 T5u = VFMA(LDK(KP995184726), T5t, T5s);
1558 T5c = VFNMS(LDK(KP995184726), T51, T4Y);
1559 }
1560 {
1561 V T4L, T56, T5D, T5G;
1562 T4L = VFNMS(LDK(KP998795456), T4K, T2f);
1563 T56 = VFNMS(LDK(KP998795456), T55, T52);
1564 ST(&(xo[WS(os, 65)]), VFNMSI(T56, T4L), ovs, &(xo[WS(os, 1)]));
1565 ST(&(xo[WS(os, 63)]), VFMAI(T56, T4L), ovs, &(xo[WS(os, 1)]));
1566 T5D = VFMA(LDK(KP740951125), T5C, T5B);
1567 T5G = VFMA(LDK(KP740951125), T5F, T5E);
1568 ST(&(xo[WS(os, 47)]), VFMAI(T5G, T5D), ovs, &(xo[WS(os, 1)]));
1569 ST(&(xo[WS(os, 81)]), VFNMSI(T5G, T5D), ovs, &(xo[WS(os, 1)]));
1570 }
1571 {
1572 V T5H, T5I, T57, T58;
1573 T5H = VFNMS(LDK(KP740951125), T5C, T5B);
1574 T5I = VFNMS(LDK(KP740951125), T5F, T5E);
1575 ST(&(xo[WS(os, 17)]), VFNMSI(T5I, T5H), ovs, &(xo[WS(os, 1)]));
1576 ST(&(xo[WS(os, 111)]), VFMAI(T5I, T5H), ovs, &(xo[WS(os, 1)]));
1577 T57 = VFMA(LDK(KP998795456), T4K, T2f);
1578 T58 = VFMA(LDK(KP998795456), T55, T52);
1579 ST(&(xo[WS(os, 1)]), VFNMSI(T58, T57), ovs, &(xo[WS(os, 1)]));
1580 ST(&(xo[WS(os, 127)]), VFMAI(T58, T57), ovs, &(xo[WS(os, 1)]));
1581 }
1582 {
1583 V T5b, T5e, T5r, T5y;
1584 T5b = VFNMS(LDK(KP998795456), T5a, T59);
1585 T5e = VFNMS(LDK(KP998795456), T5d, T5c);
1586 ST(&(xo[WS(os, 33)]), VFNMSI(T5e, T5b), ovs, &(xo[WS(os, 1)]));
1587 ST(&(xo[WS(os, 95)]), VFMAI(T5e, T5b), ovs, &(xo[WS(os, 1)]));
1588 T5r = VFNMS(LDK(KP740951125), T5q, T5j);
1589 T5y = VFNMS(LDK(KP740951125), T5x, T5u);
1590 ST(&(xo[WS(os, 49)]), VFNMSI(T5y, T5r), ovs, &(xo[WS(os, 1)]));
1591 ST(&(xo[WS(os, 79)]), VFMAI(T5y, T5r), ovs, &(xo[WS(os, 1)]));
1592 }
1593 {
1594 V T5z, T5A, T5f, T5g;
1595 T5z = VFMA(LDK(KP740951125), T5q, T5j);
1596 T5A = VFMA(LDK(KP740951125), T5x, T5u);
1597 ST(&(xo[WS(os, 113)]), VFNMSI(T5A, T5z), ovs, &(xo[WS(os, 1)]));
1598 ST(&(xo[WS(os, 15)]), VFMAI(T5A, T5z), ovs, &(xo[WS(os, 1)]));
1599 T5f = VFMA(LDK(KP998795456), T5a, T59);
1600 T5g = VFMA(LDK(KP998795456), T5d, T5c);
1601 ST(&(xo[WS(os, 31)]), VFMAI(T5g, T5f), ovs, &(xo[WS(os, 1)]));
1602 ST(&(xo[WS(os, 97)]), VFNMSI(T5g, T5f), ovs, &(xo[WS(os, 1)]));
1603 }
1604 }
1605 {
1606 V T9i, T9B, T9t, T9y, T9O, Ta3, T9V, Ta0, T93, T9Z, Ta2, T9q, T9x, T9H, T9S;
1607 V T9A;
1608 {
1609 V T9a, T9r, T9h, T9s;
1610 {
1611 V T96, T99, T9d, T9g;
1612 T96 = VFNMS(LDK(KP831469612), T95, T94);
1613 T99 = VFNMS(LDK(KP831469612), T98, T97);
1614 T9a = VFMA(LDK(KP599376933), T99, T96);
1615 T9r = VFNMS(LDK(KP599376933), T96, T99);
1616 T9d = VFNMS(LDK(KP831469612), T9c, T9b);
1617 T9g = VFNMS(LDK(KP831469612), T9f, T9e);
1618 T9h = VFMA(LDK(KP599376933), T9g, T9d);
1619 T9s = VFNMS(LDK(KP599376933), T9d, T9g);
1620 }
1621 T9i = VADD(T9a, T9h);
1622 T9B = VSUB(T9h, T9a);
1623 T9t = VSUB(T9r, T9s);
1624 T9y = VADD(T9r, T9s);
1625 }
1626 {
1627 V T9K, T9T, T9N, T9U;
1628 {
1629 V T9I, T9J, T9L, T9M;
1630 T9I = VFMA(LDK(KP831469612), T95, T94);
1631 T9J = VFMA(LDK(KP831469612), T98, T97);
1632 T9K = VFNMS(LDK(KP250486960), T9J, T9I);
1633 T9T = VFMA(LDK(KP250486960), T9I, T9J);
1634 T9L = VFMA(LDK(KP831469612), T9c, T9b);
1635 T9M = VFMA(LDK(KP831469612), T9f, T9e);
1636 T9N = VFNMS(LDK(KP250486960), T9M, T9L);
1637 T9U = VFMA(LDK(KP250486960), T9L, T9M);
1638 }
1639 T9O = VADD(T9K, T9N);
1640 Ta3 = VSUB(T9N, T9K);
1641 T9V = VSUB(T9T, T9U);
1642 Ta0 = VADD(T9T, T9U);
1643 }
1644 {
1645 V T8V, T9F, T9m, T9Q, T92, T9R, T9p, T9G, T8U, T9l;
1646 T8U = VADD(T85, T86);
1647 T8V = VFMA(LDK(KP831469612), T8U, T8T);
1648 T9F = VFNMS(LDK(KP831469612), T8U, T8T);
1649 T9l = VSUB(T75, T72);
1650 T9m = VFMA(LDK(KP831469612), T9l, T9k);
1651 T9Q = VFNMS(LDK(KP831469612), T9l, T9k);
1652 {
1653 V T8Y, T91, T9n, T9o;
1654 T8Y = VFMA(LDK(KP534511135), T8X, T8W);
1655 T91 = VFMA(LDK(KP534511135), T90, T8Z);
1656 T92 = VADD(T8Y, T91);
1657 T9R = VSUB(T8Y, T91);
1658 T9n = VFNMS(LDK(KP534511135), T8Z, T90);
1659 T9o = VFNMS(LDK(KP534511135), T8W, T8X);
1660 T9p = VSUB(T9n, T9o);
1661 T9G = VADD(T9o, T9n);
1662 }
1663 T93 = VFMA(LDK(KP881921264), T92, T8V);
1664 T9Z = VFNMS(LDK(KP881921264), T9G, T9F);
1665 Ta2 = VFNMS(LDK(KP881921264), T9R, T9Q);
1666 T9q = VFMA(LDK(KP881921264), T9p, T9m);
1667 T9x = VFNMS(LDK(KP881921264), T92, T8V);
1668 T9H = VFMA(LDK(KP881921264), T9G, T9F);
1669 T9S = VFMA(LDK(KP881921264), T9R, T9Q);
1670 T9A = VFNMS(LDK(KP881921264), T9p, T9m);
1671 }
1672 {
1673 V T9j, T9u, Ta1, Ta4;
1674 T9j = VFNMS(LDK(KP857728610), T9i, T93);
1675 T9u = VFNMS(LDK(KP857728610), T9t, T9q);
1676 ST(&(xo[WS(os, 53)]), VFNMSI(T9u, T9j), ovs, &(xo[WS(os, 1)]));
1677 ST(&(xo[WS(os, 75)]), VFMAI(T9u, T9j), ovs, &(xo[WS(os, 1)]));
1678 Ta1 = VFNMS(LDK(KP970031253), Ta0, T9Z);
1679 Ta4 = VFNMS(LDK(KP970031253), Ta3, Ta2);
1680 ST(&(xo[WS(os, 37)]), VFNMSI(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
1681 ST(&(xo[WS(os, 91)]), VFMAI(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
1682 }
1683 {
1684 V Ta5, Ta6, T9v, T9w;
1685 Ta5 = VFMA(LDK(KP970031253), Ta0, T9Z);
1686 Ta6 = VFMA(LDK(KP970031253), Ta3, Ta2);
1687 ST(&(xo[WS(os, 27)]), VFMAI(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
1688 ST(&(xo[WS(os, 101)]), VFNMSI(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
1689 T9v = VFMA(LDK(KP857728610), T9i, T93);
1690 T9w = VFMA(LDK(KP857728610), T9t, T9q);
1691 ST(&(xo[WS(os, 117)]), VFNMSI(T9w, T9v), ovs, &(xo[WS(os, 1)]));
1692 ST(&(xo[WS(os, 11)]), VFMAI(T9w, T9v), ovs, &(xo[WS(os, 1)]));
1693 }
1694 {
1695 V T9z, T9C, T9P, T9W;
1696 T9z = VFMA(LDK(KP857728610), T9y, T9x);
1697 T9C = VFMA(LDK(KP857728610), T9B, T9A);
1698 ST(&(xo[WS(os, 43)]), VFMAI(T9C, T9z), ovs, &(xo[WS(os, 1)]));
1699 ST(&(xo[WS(os, 85)]), VFNMSI(T9C, T9z), ovs, &(xo[WS(os, 1)]));
1700 T9P = VFNMS(LDK(KP970031253), T9O, T9H);
1701 T9W = VFNMS(LDK(KP970031253), T9V, T9S);
1702 ST(&(xo[WS(os, 69)]), VFNMSI(T9W, T9P), ovs, &(xo[WS(os, 1)]));
1703 ST(&(xo[WS(os, 59)]), VFMAI(T9W, T9P), ovs, &(xo[WS(os, 1)]));
1704 }
1705 {
1706 V T9X, T9Y, T9D, T9E;
1707 T9X = VFMA(LDK(KP970031253), T9O, T9H);
1708 T9Y = VFMA(LDK(KP970031253), T9V, T9S);
1709 ST(&(xo[WS(os, 5)]), VFNMSI(T9Y, T9X), ovs, &(xo[WS(os, 1)]));
1710 ST(&(xo[WS(os, 123)]), VFMAI(T9Y, T9X), ovs, &(xo[WS(os, 1)]));
1711 T9D = VFNMS(LDK(KP857728610), T9y, T9x);
1712 T9E = VFNMS(LDK(KP857728610), T9B, T9A);
1713 ST(&(xo[WS(os, 21)]), VFNMSI(T9E, T9D), ovs, &(xo[WS(os, 1)]));
1714 ST(&(xo[WS(os, 107)]), VFMAI(T9E, T9D), ovs, &(xo[WS(os, 1)]));
1715 }
1716 }
1717 {
1718 V T68, T6r, T6j, T6o, T6E, T6T, T6L, T6Q, T5T, T6P, T6S, T6g, T6n, T6x, T6I;
1719 V T6q;
1720 {
1721 V T60, T6h, T67, T6i;
1722 {
1723 V T5W, T5Z, T63, T66;
1724 T5W = VFNMS(LDK(KP980785280), T5V, T5U);
1725 T5Z = VFMA(LDK(KP980785280), T5Y, T5X);
1726 T60 = VFNMS(LDK(KP472964775), T5Z, T5W);
1727 T6h = VFMA(LDK(KP472964775), T5W, T5Z);
1728 T63 = VFNMS(LDK(KP980785280), T62, T61);
1729 T66 = VFMA(LDK(KP980785280), T65, T64);
1730 T67 = VFNMS(LDK(KP472964775), T66, T63);
1731 T6i = VFMA(LDK(KP472964775), T63, T66);
1732 }
1733 T68 = VADD(T60, T67);
1734 T6r = VSUB(T67, T60);
1735 T6j = VSUB(T6h, T6i);
1736 T6o = VADD(T6h, T6i);
1737 }
1738 {
1739 V T6A, T6J, T6D, T6K;
1740 {
1741 V T6y, T6z, T6B, T6C;
1742 T6y = VFMA(LDK(KP980785280), T5V, T5U);
1743 T6z = VFNMS(LDK(KP980785280), T5Y, T5X);
1744 T6A = VFMA(LDK(KP357805721), T6z, T6y);
1745 T6J = VFNMS(LDK(KP357805721), T6y, T6z);
1746 T6B = VFMA(LDK(KP980785280), T62, T61);
1747 T6C = VFNMS(LDK(KP980785280), T65, T64);
1748 T6D = VFMA(LDK(KP357805721), T6C, T6B);
1749 T6K = VFNMS(LDK(KP357805721), T6B, T6C);
1750 }
1751 T6E = VADD(T6A, T6D);
1752 T6T = VSUB(T6D, T6A);
1753 T6L = VSUB(T6J, T6K);
1754 T6Q = VADD(T6J, T6K);
1755 }
1756 {
1757 V T5L, T6v, T6c, T6G, T5S, T6H, T6f, T6w, T5K, T6b;
1758 T5K = VADD(T4V, T4W);
1759 T5L = VFNMS(LDK(KP980785280), T5K, T5J);
1760 T6v = VFMA(LDK(KP980785280), T5K, T5J);
1761 T6b = VSUB(TZ, TI);
1762 T6c = VFNMS(LDK(KP980785280), T6b, T6a);
1763 T6G = VFMA(LDK(KP980785280), T6b, T6a);
1764 {
1765 V T5O, T5R, T6d, T6e;
1766 T5O = VFNMS(LDK(KP820678790), T5N, T5M);
1767 T5R = VFNMS(LDK(KP820678790), T5Q, T5P);
1768 T5S = VADD(T5O, T5R);
1769 T6H = VSUB(T5O, T5R);
1770 T6d = VFMA(LDK(KP820678790), T5P, T5Q);
1771 T6e = VFMA(LDK(KP820678790), T5M, T5N);
1772 T6f = VSUB(T6d, T6e);
1773 T6w = VADD(T6e, T6d);
1774 }
1775 T5T = VFNMS(LDK(KP773010453), T5S, T5L);
1776 T6P = VFNMS(LDK(KP773010453), T6w, T6v);
1777 T6S = VFNMS(LDK(KP773010453), T6H, T6G);
1778 T6g = VFNMS(LDK(KP773010453), T6f, T6c);
1779 T6n = VFMA(LDK(KP773010453), T5S, T5L);
1780 T6x = VFMA(LDK(KP773010453), T6w, T6v);
1781 T6I = VFMA(LDK(KP773010453), T6H, T6G);
1782 T6q = VFMA(LDK(KP773010453), T6f, T6c);
1783 }
1784 {
1785 V T69, T6k, T6R, T6U;
1786 T69 = VFNMS(LDK(KP903989293), T68, T5T);
1787 T6k = VFNMS(LDK(KP903989293), T6j, T6g);
1788 ST(&(xo[WS(os, 73)]), VFNMSI(T6k, T69), ovs, &(xo[WS(os, 1)]));
1789 ST(&(xo[WS(os, 55)]), VFMAI(T6k, T69), ovs, &(xo[WS(os, 1)]));
1790 T6R = VFMA(LDK(KP941544065), T6Q, T6P);
1791 T6U = VFMA(LDK(KP941544065), T6T, T6S);
1792 ST(&(xo[WS(os, 39)]), VFMAI(T6U, T6R), ovs, &(xo[WS(os, 1)]));
1793 ST(&(xo[WS(os, 89)]), VFNMSI(T6U, T6R), ovs, &(xo[WS(os, 1)]));
1794 }
1795 {
1796 V T6V, T6W, T6l, T6m;
1797 T6V = VFNMS(LDK(KP941544065), T6Q, T6P);
1798 T6W = VFNMS(LDK(KP941544065), T6T, T6S);
1799 ST(&(xo[WS(os, 25)]), VFNMSI(T6W, T6V), ovs, &(xo[WS(os, 1)]));
1800 ST(&(xo[WS(os, 103)]), VFMAI(T6W, T6V), ovs, &(xo[WS(os, 1)]));
1801 T6l = VFMA(LDK(KP903989293), T68, T5T);
1802 T6m = VFMA(LDK(KP903989293), T6j, T6g);
1803 ST(&(xo[WS(os, 9)]), VFNMSI(T6m, T6l), ovs, &(xo[WS(os, 1)]));
1804 ST(&(xo[WS(os, 119)]), VFMAI(T6m, T6l), ovs, &(xo[WS(os, 1)]));
1805 }
1806 {
1807 V T6p, T6s, T6F, T6M;
1808 T6p = VFNMS(LDK(KP903989293), T6o, T6n);
1809 T6s = VFNMS(LDK(KP903989293), T6r, T6q);
1810 ST(&(xo[WS(os, 41)]), VFNMSI(T6s, T6p), ovs, &(xo[WS(os, 1)]));
1811 ST(&(xo[WS(os, 87)]), VFMAI(T6s, T6p), ovs, &(xo[WS(os, 1)]));
1812 T6F = VFNMS(LDK(KP941544065), T6E, T6x);
1813 T6M = VFNMS(LDK(KP941544065), T6L, T6I);
1814 ST(&(xo[WS(os, 57)]), VFNMSI(T6M, T6F), ovs, &(xo[WS(os, 1)]));
1815 ST(&(xo[WS(os, 71)]), VFMAI(T6M, T6F), ovs, &(xo[WS(os, 1)]));
1816 }
1817 {
1818 V T6N, T6O, T6t, T6u;
1819 T6N = VFMA(LDK(KP941544065), T6E, T6x);
1820 T6O = VFMA(LDK(KP941544065), T6L, T6I);
1821 ST(&(xo[WS(os, 121)]), VFNMSI(T6O, T6N), ovs, &(xo[WS(os, 1)]));
1822 ST(&(xo[WS(os, 7)]), VFMAI(T6O, T6N), ovs, &(xo[WS(os, 1)]));
1823 T6t = VFMA(LDK(KP903989293), T6o, T6n);
1824 T6u = VFMA(LDK(KP903989293), T6r, T6q);
1825 ST(&(xo[WS(os, 23)]), VFMAI(T6u, T6t), ovs, &(xo[WS(os, 1)]));
1826 ST(&(xo[WS(os, 105)]), VFNMSI(T6u, T6t), ovs, &(xo[WS(os, 1)]));
1827 }
1828 }
1829 }
1830 }
1831 VLEAVE();
1832 }
1833
1834 static const kdft_desc desc = { 128, XSIMD_STRING("n1fv_128"), {440, 0, 642, 0}, &GENUS, 0, 0, 0, 0 };
1835
1836 void XSIMD(codelet_n1fv_128) (planner *p) {
1837 X(kdft_register) (p, n1fv_128, &desc);
1838 }
1839
1840 #else
1841
1842 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 128 -name n1fv_128 -include dft/simd/n1f.h */
1843
1844 /*
1845 * This function contains 1082 FP additions, 330 FP multiplications,
1846 * (or, 938 additions, 186 multiplications, 144 fused multiply/add),
1847 * 194 stack variables, 31 constants, and 256 memory accesses
1848 */
1849 #include "dft/simd/n1f.h"
1850
1851 static void n1fv_128(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
1852 {
1853 DVK(KP941544065, +0.941544065183020778412509402599502357185589796);
1854 DVK(KP336889853, +0.336889853392220050689253212619147570477766780);
1855 DVK(KP903989293, +0.903989293123443331586200297230537048710132025);
1856 DVK(KP427555093, +0.427555093430282094320966856888798534304578629);
1857 DVK(KP970031253, +0.970031253194543992603984207286100251456865962);
1858 DVK(KP242980179, +0.242980179903263889948274162077471118320990783);
1859 DVK(KP857728610, +0.857728610000272069902269984284770137042490799);
1860 DVK(KP514102744, +0.514102744193221726593693838968815772608049120);
1861 DVK(KP671558954, +0.671558954847018400625376850427421803228750632);
1862 DVK(KP740951125, +0.740951125354959091175616897495162729728955309);
1863 DVK(KP049067674, +0.049067674327418014254954976942682658314745363);
1864 DVK(KP998795456, +0.998795456205172392714771604759100694443203615);
1865 DVK(KP595699304, +0.595699304492433343467036528829969889511926338);
1866 DVK(KP803207531, +0.803207531480644909806676512963141923879569427);
1867 DVK(KP146730474, +0.146730474455361751658850129646717819706215317);
1868 DVK(KP989176509, +0.989176509964780973451673738016243063983689533);
1869 DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
1870 DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
1871 DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
1872 DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
1873 DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
1874 DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
1875 DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
1876 DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
1877 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
1878 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
1879 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
1880 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
1881 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
1882 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
1883 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
1884 {
1885 INT i;
1886 const R *xi;
1887 R *xo;
1888 xi = ri;
1889 xo = ro;
1890 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(256, is), MAKE_VOLATILE_STRIDE(256, os)) {
1891 V Tr, T5J, Ted, Tgf, Tfq, TgH, T4U, T6b, T6Z, T8T, Tad, TcZ, Tcc, Td0, T84;
1892 V T9l, Tb6, Tbt, T2G, T5X, TeV, Tgr, T3p, T5V, T7B, T95, TeK, Tgt, T7q, T97;
1893 V Td8, TdK, TbD, Tc0, T3V, T61, Tfg, TgB, T4E, T65, T7U, T9f, Tf5, Tgx, T7J;
1894 V T9b, Tdf, TdN, Td2, Td3, TI, T4V, Tft, Tgg, TZ, T4W, T75, T86, Tek, TgG;
1895 V T72, T85, Tas, Tcd, Tdp, Tdq, TdG, Teq, Tgm, Tet, Tgl, T1s, T5P, T1B, T5Q;
1896 V T7d, T8Z, TaI, Tcf, T7a, T90, Tdm, Tdn, TdH, Tez, Tgi, TeC, Tgj, T23, T5N;
1897 V T2c, T5M, T7k, T8X, TaX, Tcg, T7h, T8W, Tbl, Tbu, Tdb, TdL, TeY, Tgu, TeR;
1898 V Tgq, T7x, T98, T7E, T94, T3f, T5Y, T3s, T5U, TbS, Tc1, Tdi, TdO, Tfj, Tgy;
1899 V Tfc, TgA, T7Q, T9e, T7X, T9c, T4u, T64, T4H, T62;
1900 {
1901 V T3, Ta7, T4P, Ta8, Ta, Tab, T4M, Taa, Tc9, Tca, Ti, Tea, T4S, Tc6, Tc7;
1902 V Tp, Teb, T4R;
1903 {
1904 V T1, T2, T4N, T4O;
1905 T1 = LD(&(xi[0]), ivs, &(xi[0]));
1906 T2 = LD(&(xi[WS(is, 64)]), ivs, &(xi[0]));
1907 T3 = VSUB(T1, T2);
1908 Ta7 = VADD(T1, T2);
1909 T4N = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
1910 T4O = LD(&(xi[WS(is, 96)]), ivs, &(xi[0]));
1911 T4P = VSUB(T4N, T4O);
1912 Ta8 = VADD(T4N, T4O);
1913 }
1914 {
1915 V T4, T5, T6, T7, T8, T9;
1916 T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
1917 T5 = LD(&(xi[WS(is, 80)]), ivs, &(xi[0]));
1918 T6 = VSUB(T4, T5);
1919 T7 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0]));
1920 T8 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
1921 T9 = VSUB(T7, T8);
1922 Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
1923 Tab = VADD(T7, T8);
1924 T4M = VMUL(LDK(KP707106781), VSUB(T9, T6));
1925 Taa = VADD(T4, T5);
1926 }
1927 {
1928 V Te, Th, Tl, To;
1929 {
1930 V Tc, Td, Tf, Tg;
1931 Tc = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
1932 Td = LD(&(xi[WS(is, 72)]), ivs, &(xi[0]));
1933 Te = VSUB(Tc, Td);
1934 Tc9 = VADD(Tc, Td);
1935 Tf = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
1936 Tg = LD(&(xi[WS(is, 104)]), ivs, &(xi[0]));
1937 Th = VSUB(Tf, Tg);
1938 Tca = VADD(Tf, Tg);
1939 }
1940 Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
1941 Tea = VSUB(Tc9, Tca);
1942 T4S = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
1943 {
1944 V Tj, Tk, Tm, Tn;
1945 Tj = LD(&(xi[WS(is, 120)]), ivs, &(xi[0]));
1946 Tk = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
1947 Tl = VSUB(Tj, Tk);
1948 Tc6 = VADD(Tj, Tk);
1949 Tm = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
1950 Tn = LD(&(xi[WS(is, 88)]), ivs, &(xi[0]));
1951 To = VSUB(Tm, Tn);
1952 Tc7 = VADD(Tm, Tn);
1953 }
1954 Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
1955 Teb = VSUB(Tc6, Tc7);
1956 T4R = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
1957 }
1958 {
1959 V Tb, Tq, Te9, Tec;
1960 Tb = VADD(T3, Ta);
1961 Tq = VADD(Ti, Tp);
1962 Tr = VADD(Tb, Tq);
1963 T5J = VSUB(Tb, Tq);
1964 Te9 = VSUB(Ta7, Ta8);
1965 Tec = VMUL(LDK(KP707106781), VADD(Tea, Teb));
1966 Ted = VADD(Te9, Tec);
1967 Tgf = VSUB(Te9, Tec);
1968 }
1969 {
1970 V Tfo, Tfp, T4Q, T4T;
1971 Tfo = VSUB(Tab, Taa);
1972 Tfp = VMUL(LDK(KP707106781), VSUB(Teb, Tea));
1973 Tfq = VADD(Tfo, Tfp);
1974 TgH = VSUB(Tfp, Tfo);
1975 T4Q = VSUB(T4M, T4P);
1976 T4T = VSUB(T4R, T4S);
1977 T4U = VADD(T4Q, T4T);
1978 T6b = VSUB(T4T, T4Q);
1979 }
1980 {
1981 V T6X, T6Y, Ta9, Tac;
1982 T6X = VSUB(T3, Ta);
1983 T6Y = VADD(T4S, T4R);
1984 T6Z = VADD(T6X, T6Y);
1985 T8T = VSUB(T6X, T6Y);
1986 Ta9 = VADD(Ta7, Ta8);
1987 Tac = VADD(Taa, Tab);
1988 Tad = VSUB(Ta9, Tac);
1989 TcZ = VADD(Ta9, Tac);
1990 }
1991 {
1992 V Tc8, Tcb, T82, T83;
1993 Tc8 = VADD(Tc6, Tc7);
1994 Tcb = VADD(Tc9, Tca);
1995 Tcc = VSUB(Tc8, Tcb);
1996 Td0 = VADD(Tcb, Tc8);
1997 T82 = VADD(T4P, T4M);
1998 T83 = VSUB(Tp, Ti);
1999 T84 = VADD(T82, T83);
2000 T9l = VSUB(T83, T82);
2001 }
2002 }
2003 {
2004 V Tb0, Tb1, T2i, Tb2, T3k, Tb3, Tb4, T2p, Tb5, T3h, T2x, TeH, T3n, Tbs, T2E;
2005 V TeI, T3m, Tbp, T2l, T2o, TeG, TeJ;
2006 {
2007 V T2g, T2h, T3i, T3j;
2008 T2g = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
2009 T2h = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)]));
2010 Tb0 = VADD(T2g, T2h);
2011 T3i = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
2012 T3j = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)]));
2013 Tb1 = VADD(T3i, T3j);
2014 T2i = VSUB(T2g, T2h);
2015 Tb2 = VADD(Tb0, Tb1);
2016 T3k = VSUB(T3i, T3j);
2017 }
2018 {
2019 V T2j, T2k, T2m, T2n;
2020 T2j = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
2021 T2k = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)]));
2022 T2l = VSUB(T2j, T2k);
2023 Tb3 = VADD(T2j, T2k);
2024 T2m = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)]));
2025 T2n = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
2026 T2o = VSUB(T2m, T2n);
2027 Tb4 = VADD(T2m, T2n);
2028 }
2029 T2p = VMUL(LDK(KP707106781), VADD(T2l, T2o));
2030 Tb5 = VADD(Tb3, Tb4);
2031 T3h = VMUL(LDK(KP707106781), VSUB(T2o, T2l));
2032 {
2033 V T2t, Tbq, T2w, Tbr;
2034 {
2035 V T2r, T2s, T2u, T2v;
2036 T2r = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
2037 T2s = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)]));
2038 T2t = VSUB(T2r, T2s);
2039 Tbq = VADD(T2r, T2s);
2040 T2u = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
2041 T2v = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)]));
2042 T2w = VSUB(T2u, T2v);
2043 Tbr = VADD(T2u, T2v);
2044 }
2045 T2x = VFNMS(LDK(KP382683432), T2w, VMUL(LDK(KP923879532), T2t));
2046 TeH = VSUB(Tbq, Tbr);
2047 T3n = VFMA(LDK(KP382683432), T2t, VMUL(LDK(KP923879532), T2w));
2048 Tbs = VADD(Tbq, Tbr);
2049 }
2050 {
2051 V T2A, Tbn, T2D, Tbo;
2052 {
2053 V T2y, T2z, T2B, T2C;
2054 T2y = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)]));
2055 T2z = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
2056 T2A = VSUB(T2y, T2z);
2057 Tbn = VADD(T2y, T2z);
2058 T2B = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
2059 T2C = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)]));
2060 T2D = VSUB(T2B, T2C);
2061 Tbo = VADD(T2B, T2C);
2062 }
2063 T2E = VFMA(LDK(KP923879532), T2A, VMUL(LDK(KP382683432), T2D));
2064 TeI = VSUB(Tbn, Tbo);
2065 T3m = VFNMS(LDK(KP923879532), T2D, VMUL(LDK(KP382683432), T2A));
2066 Tbp = VADD(Tbn, Tbo);
2067 }
2068 Tb6 = VSUB(Tb2, Tb5);
2069 Tbt = VSUB(Tbp, Tbs);
2070 {
2071 V T2q, T2F, TeT, TeU;
2072 T2q = VADD(T2i, T2p);
2073 T2F = VADD(T2x, T2E);
2074 T2G = VADD(T2q, T2F);
2075 T5X = VSUB(T2q, T2F);
2076 TeT = VSUB(Tb4, Tb3);
2077 TeU = VMUL(LDK(KP707106781), VSUB(TeI, TeH));
2078 TeV = VADD(TeT, TeU);
2079 Tgr = VSUB(TeU, TeT);
2080 }
2081 {
2082 V T3l, T3o, T7z, T7A;
2083 T3l = VSUB(T3h, T3k);
2084 T3o = VSUB(T3m, T3n);
2085 T3p = VADD(T3l, T3o);
2086 T5V = VSUB(T3o, T3l);
2087 T7z = VADD(T3k, T3h);
2088 T7A = VSUB(T2E, T2x);
2089 T7B = VADD(T7z, T7A);
2090 T95 = VSUB(T7A, T7z);
2091 }
2092 TeG = VSUB(Tb0, Tb1);
2093 TeJ = VMUL(LDK(KP707106781), VADD(TeH, TeI));
2094 TeK = VADD(TeG, TeJ);
2095 Tgt = VSUB(TeG, TeJ);
2096 {
2097 V T7o, T7p, Td6, Td7;
2098 T7o = VSUB(T2i, T2p);
2099 T7p = VADD(T3n, T3m);
2100 T7q = VADD(T7o, T7p);
2101 T97 = VSUB(T7o, T7p);
2102 Td6 = VADD(Tb2, Tb5);
2103 Td7 = VADD(Tbs, Tbp);
2104 Td8 = VADD(Td6, Td7);
2105 TdK = VSUB(Td6, Td7);
2106 }
2107 }
2108 {
2109 V Tbx, Tby, T3x, Tbz, T4z, TbA, TbB, T3E, TbC, T4w, T3M, Tf2, T4C, TbZ, T3T;
2110 V Tf3, T4B, TbW, T3A, T3D, Tf1, Tf4;
2111 {
2112 V T3v, T3w, T4x, T4y;
2113 T3v = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)]));
2114 T3w = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
2115 Tbx = VADD(T3v, T3w);
2116 T4x = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
2117 T4y = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)]));
2118 Tby = VADD(T4x, T4y);
2119 T3x = VSUB(T3v, T3w);
2120 Tbz = VADD(Tbx, Tby);
2121 T4z = VSUB(T4x, T4y);
2122 }
2123 {
2124 V T3y, T3z, T3B, T3C;
2125 T3y = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
2126 T3z = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)]));
2127 T3A = VSUB(T3y, T3z);
2128 TbA = VADD(T3y, T3z);
2129 T3B = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)]));
2130 T3C = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
2131 T3D = VSUB(T3B, T3C);
2132 TbB = VADD(T3B, T3C);
2133 }
2134 T3E = VMUL(LDK(KP707106781), VADD(T3A, T3D));
2135 TbC = VADD(TbA, TbB);
2136 T4w = VMUL(LDK(KP707106781), VSUB(T3D, T3A));
2137 {
2138 V T3I, TbX, T3L, TbY;
2139 {
2140 V T3G, T3H, T3J, T3K;
2141 T3G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
2142 T3H = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)]));
2143 T3I = VSUB(T3G, T3H);
2144 TbX = VADD(T3G, T3H);
2145 T3J = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
2146 T3K = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)]));
2147 T3L = VSUB(T3J, T3K);
2148 TbY = VADD(T3J, T3K);
2149 }
2150 T3M = VFNMS(LDK(KP382683432), T3L, VMUL(LDK(KP923879532), T3I));
2151 Tf2 = VSUB(TbX, TbY);
2152 T4C = VFMA(LDK(KP382683432), T3I, VMUL(LDK(KP923879532), T3L));
2153 TbZ = VADD(TbX, TbY);
2154 }
2155 {
2156 V T3P, TbU, T3S, TbV;
2157 {
2158 V T3N, T3O, T3Q, T3R;
2159 T3N = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)]));
2160 T3O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
2161 T3P = VSUB(T3N, T3O);
2162 TbU = VADD(T3N, T3O);
2163 T3Q = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
2164 T3R = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)]));
2165 T3S = VSUB(T3Q, T3R);
2166 TbV = VADD(T3Q, T3R);
2167 }
2168 T3T = VFMA(LDK(KP923879532), T3P, VMUL(LDK(KP382683432), T3S));
2169 Tf3 = VSUB(TbU, TbV);
2170 T4B = VFNMS(LDK(KP923879532), T3S, VMUL(LDK(KP382683432), T3P));
2171 TbW = VADD(TbU, TbV);
2172 }
2173 TbD = VSUB(Tbz, TbC);
2174 Tc0 = VSUB(TbW, TbZ);
2175 {
2176 V T3F, T3U, Tfe, Tff;
2177 T3F = VADD(T3x, T3E);
2178 T3U = VADD(T3M, T3T);
2179 T3V = VADD(T3F, T3U);
2180 T61 = VSUB(T3F, T3U);
2181 Tfe = VSUB(TbB, TbA);
2182 Tff = VMUL(LDK(KP707106781), VSUB(Tf3, Tf2));
2183 Tfg = VADD(Tfe, Tff);
2184 TgB = VSUB(Tff, Tfe);
2185 }
2186 {
2187 V T4A, T4D, T7S, T7T;
2188 T4A = VSUB(T4w, T4z);
2189 T4D = VSUB(T4B, T4C);
2190 T4E = VADD(T4A, T4D);
2191 T65 = VSUB(T4D, T4A);
2192 T7S = VADD(T4z, T4w);
2193 T7T = VSUB(T3T, T3M);
2194 T7U = VADD(T7S, T7T);
2195 T9f = VSUB(T7T, T7S);
2196 }
2197 Tf1 = VSUB(Tbx, Tby);
2198 Tf4 = VMUL(LDK(KP707106781), VADD(Tf2, Tf3));
2199 Tf5 = VADD(Tf1, Tf4);
2200 Tgx = VSUB(Tf1, Tf4);
2201 {
2202 V T7H, T7I, Tdd, Tde;
2203 T7H = VSUB(T3x, T3E);
2204 T7I = VADD(T4C, T4B);
2205 T7J = VADD(T7H, T7I);
2206 T9b = VSUB(T7H, T7I);
2207 Tdd = VADD(Tbz, TbC);
2208 Tde = VADD(TbZ, TbW);
2209 Tdf = VADD(Tdd, Tde);
2210 TdN = VSUB(Tdd, Tde);
2211 }
2212 }
2213 {
2214 V Tu, Tee, TG, Tag, TL, Teh, TX, Tan, TB, Tef, TD, Taj, TS, Tei, TU;
2215 V Taq, Teg, Tej;
2216 {
2217 V Ts, Tt, Tae, TE, TF, Taf;
2218 Ts = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
2219 Tt = LD(&(xi[WS(is, 68)]), ivs, &(xi[0]));
2220 Tae = VADD(Ts, Tt);
2221 TE = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
2222 TF = LD(&(xi[WS(is, 100)]), ivs, &(xi[0]));
2223 Taf = VADD(TE, TF);
2224 Tu = VSUB(Ts, Tt);
2225 Tee = VSUB(Tae, Taf);
2226 TG = VSUB(TE, TF);
2227 Tag = VADD(Tae, Taf);
2228 }
2229 {
2230 V TJ, TK, Tal, TV, TW, Tam;
2231 TJ = LD(&(xi[WS(is, 124)]), ivs, &(xi[0]));
2232 TK = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
2233 Tal = VADD(TJ, TK);
2234 TV = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
2235 TW = LD(&(xi[WS(is, 92)]), ivs, &(xi[0]));
2236 Tam = VADD(TV, TW);
2237 TL = VSUB(TJ, TK);
2238 Teh = VSUB(Tal, Tam);
2239 TX = VSUB(TV, TW);
2240 Tan = VADD(Tal, Tam);
2241 }
2242 {
2243 V Tx, Tah, TA, Tai;
2244 {
2245 V Tv, Tw, Ty, Tz;
2246 Tv = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
2247 Tw = LD(&(xi[WS(is, 84)]), ivs, &(xi[0]));
2248 Tx = VSUB(Tv, Tw);
2249 Tah = VADD(Tv, Tw);
2250 Ty = LD(&(xi[WS(is, 116)]), ivs, &(xi[0]));
2251 Tz = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
2252 TA = VSUB(Ty, Tz);
2253 Tai = VADD(Ty, Tz);
2254 }
2255 TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
2256 Tef = VSUB(Tai, Tah);
2257 TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
2258 Taj = VADD(Tah, Tai);
2259 }
2260 {
2261 V TO, Tao, TR, Tap;
2262 {
2263 V TM, TN, TP, TQ;
2264 TM = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
2265 TN = LD(&(xi[WS(is, 76)]), ivs, &(xi[0]));
2266 TO = VSUB(TM, TN);
2267 Tao = VADD(TM, TN);
2268 TP = LD(&(xi[WS(is, 108)]), ivs, &(xi[0]));
2269 TQ = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
2270 TR = VSUB(TP, TQ);
2271 Tap = VADD(TP, TQ);
2272 }
2273 TS = VMUL(LDK(KP707106781), VADD(TO, TR));
2274 Tei = VSUB(Tap, Tao);
2275 TU = VMUL(LDK(KP707106781), VSUB(TR, TO));
2276 Taq = VADD(Tao, Tap);
2277 }
2278 Td2 = VADD(Tag, Taj);
2279 Td3 = VADD(Tan, Taq);
2280 {
2281 V TC, TH, Tfr, Tfs;
2282 TC = VADD(Tu, TB);
2283 TH = VSUB(TD, TG);
2284 TI = VFMA(LDK(KP980785280), TC, VMUL(LDK(KP195090322), TH));
2285 T4V = VFNMS(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
2286 Tfr = VFNMS(LDK(KP382683432), Tee, VMUL(LDK(KP923879532), Tef));
2287 Tfs = VFMA(LDK(KP382683432), Teh, VMUL(LDK(KP923879532), Tei));
2288 Tft = VADD(Tfr, Tfs);
2289 Tgg = VSUB(Tfs, Tfr);
2290 }
2291 {
2292 V TT, TY, T73, T74;
2293 TT = VADD(TL, TS);
2294 TY = VSUB(TU, TX);
2295 TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
2296 T4W = VFMA(LDK(KP195090322), TT, VMUL(LDK(KP980785280), TY));
2297 T73 = VSUB(TL, TS);
2298 T74 = VADD(TX, TU);
2299 T75 = VFNMS(LDK(KP555570233), T74, VMUL(LDK(KP831469612), T73));
2300 T86 = VFMA(LDK(KP555570233), T73, VMUL(LDK(KP831469612), T74));
2301 }
2302 Teg = VFMA(LDK(KP923879532), Tee, VMUL(LDK(KP382683432), Tef));
2303 Tej = VFNMS(LDK(KP382683432), Tei, VMUL(LDK(KP923879532), Teh));
2304 Tek = VADD(Teg, Tej);
2305 TgG = VSUB(Tej, Teg);
2306 {
2307 V T70, T71, Tak, Tar;
2308 T70 = VSUB(Tu, TB);
2309 T71 = VADD(TG, TD);
2310 T72 = VFMA(LDK(KP831469612), T70, VMUL(LDK(KP555570233), T71));
2311 T85 = VFNMS(LDK(KP555570233), T70, VMUL(LDK(KP831469612), T71));
2312 Tak = VSUB(Tag, Taj);
2313 Tar = VSUB(Tan, Taq);
2314 Tas = VMUL(LDK(KP707106781), VADD(Tak, Tar));
2315 Tcd = VMUL(LDK(KP707106781), VSUB(Tar, Tak));
2316 }
2317 }
2318 {
2319 V Tav, Tau, T1b, Taw, T1v, Tay, Tax, T18, Taz, T1w, T1j, Teo, T1z, TaD, T1q;
2320 V Ten, T1y, TaG, T14, T17, Tem, Tep;
2321 {
2322 V T19, T1a, T1t, T1u;
2323 T19 = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
2324 T1a = LD(&(xi[WS(is, 98)]), ivs, &(xi[0]));
2325 Tav = VADD(T19, T1a);
2326 T1t = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
2327 T1u = LD(&(xi[WS(is, 66)]), ivs, &(xi[0]));
2328 Tau = VADD(T1t, T1u);
2329 T1b = VSUB(T19, T1a);
2330 Taw = VADD(Tau, Tav);
2331 T1v = VSUB(T1t, T1u);
2332 }
2333 {
2334 V T12, T13, T15, T16;
2335 T12 = LD(&(xi[WS(is, 114)]), ivs, &(xi[0]));
2336 T13 = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
2337 T14 = VSUB(T12, T13);
2338 Tay = VADD(T12, T13);
2339 T15 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
2340 T16 = LD(&(xi[WS(is, 82)]), ivs, &(xi[0]));
2341 T17 = VSUB(T15, T16);
2342 Tax = VADD(T15, T16);
2343 }
2344 T18 = VMUL(LDK(KP707106781), VSUB(T14, T17));
2345 Taz = VADD(Tax, Tay);
2346 T1w = VMUL(LDK(KP707106781), VADD(T17, T14));
2347 {
2348 V T1f, TaB, T1i, TaC;
2349 {
2350 V T1d, T1e, T1g, T1h;
2351 T1d = LD(&(xi[WS(is, 122)]), ivs, &(xi[0]));
2352 T1e = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
2353 T1f = VSUB(T1d, T1e);
2354 TaB = VADD(T1d, T1e);
2355 T1g = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
2356 T1h = LD(&(xi[WS(is, 90)]), ivs, &(xi[0]));
2357 T1i = VSUB(T1g, T1h);
2358 TaC = VADD(T1g, T1h);
2359 }
2360 T1j = VFNMS(LDK(KP923879532), T1i, VMUL(LDK(KP382683432), T1f));
2361 Teo = VSUB(TaB, TaC);
2362 T1z = VFMA(LDK(KP923879532), T1f, VMUL(LDK(KP382683432), T1i));
2363 TaD = VADD(TaB, TaC);
2364 }
2365 {
2366 V T1m, TaE, T1p, TaF;
2367 {
2368 V T1k, T1l, T1n, T1o;
2369 T1k = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
2370 T1l = LD(&(xi[WS(is, 74)]), ivs, &(xi[0]));
2371 T1m = VSUB(T1k, T1l);
2372 TaE = VADD(T1k, T1l);
2373 T1n = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
2374 T1o = LD(&(xi[WS(is, 106)]), ivs, &(xi[0]));
2375 T1p = VSUB(T1n, T1o);
2376 TaF = VADD(T1n, T1o);
2377 }
2378 T1q = VFMA(LDK(KP382683432), T1m, VMUL(LDK(KP923879532), T1p));
2379 Ten = VSUB(TaE, TaF);
2380 T1y = VFNMS(LDK(KP382683432), T1p, VMUL(LDK(KP923879532), T1m));
2381 TaG = VADD(TaE, TaF);
2382 }
2383 Tdp = VADD(Taw, Taz);
2384 Tdq = VADD(TaG, TaD);
2385 TdG = VSUB(Tdp, Tdq);
2386 Tem = VSUB(Tau, Tav);
2387 Tep = VMUL(LDK(KP707106781), VADD(Ten, Teo));
2388 Teq = VADD(Tem, Tep);
2389 Tgm = VSUB(Tem, Tep);
2390 {
2391 V Ter, Tes, T1c, T1r;
2392 Ter = VSUB(Tay, Tax);
2393 Tes = VMUL(LDK(KP707106781), VSUB(Teo, Ten));
2394 Tet = VADD(Ter, Tes);
2395 Tgl = VSUB(Tes, Ter);
2396 T1c = VSUB(T18, T1b);
2397 T1r = VSUB(T1j, T1q);
2398 T1s = VADD(T1c, T1r);
2399 T5P = VSUB(T1r, T1c);
2400 }
2401 {
2402 V T1x, T1A, T7b, T7c;
2403 T1x = VADD(T1v, T1w);
2404 T1A = VADD(T1y, T1z);
2405 T1B = VADD(T1x, T1A);
2406 T5Q = VSUB(T1x, T1A);
2407 T7b = VADD(T1b, T18);
2408 T7c = VSUB(T1z, T1y);
2409 T7d = VADD(T7b, T7c);
2410 T8Z = VSUB(T7c, T7b);
2411 }
2412 {
2413 V TaA, TaH, T78, T79;
2414 TaA = VSUB(Taw, Taz);
2415 TaH = VSUB(TaD, TaG);
2416 TaI = VFMA(LDK(KP923879532), TaA, VMUL(LDK(KP382683432), TaH));
2417 Tcf = VFNMS(LDK(KP382683432), TaA, VMUL(LDK(KP923879532), TaH));
2418 T78 = VSUB(T1v, T1w);
2419 T79 = VADD(T1q, T1j);
2420 T7a = VADD(T78, T79);
2421 T90 = VSUB(T78, T79);
2422 }
2423 }
2424 {
2425 V TaJ, TaK, T1F, TaL, T27, TaM, TaN, T1M, TaO, T24, T1U, Tew, T2a, TaV, T21;
2426 V Tex, T29, TaS, T1I, T1L, Tev, Tey;
2427 {
2428 V T1D, T1E, T25, T26;
2429 T1D = LD(&(xi[WS(is, 126)]), ivs, &(xi[0]));
2430 T1E = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
2431 TaJ = VADD(T1D, T1E);
2432 T25 = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
2433 T26 = LD(&(xi[WS(is, 94)]), ivs, &(xi[0]));
2434 TaK = VADD(T25, T26);
2435 T1F = VSUB(T1D, T1E);
2436 TaL = VADD(TaJ, TaK);
2437 T27 = VSUB(T25, T26);
2438 }
2439 {
2440 V T1G, T1H, T1J, T1K;
2441 T1G = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
2442 T1H = LD(&(xi[WS(is, 78)]), ivs, &(xi[0]));
2443 T1I = VSUB(T1G, T1H);
2444 TaM = VADD(T1G, T1H);
2445 T1J = LD(&(xi[WS(is, 110)]), ivs, &(xi[0]));
2446 T1K = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
2447 T1L = VSUB(T1J, T1K);
2448 TaN = VADD(T1J, T1K);
2449 }
2450 T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L));
2451 TaO = VADD(TaM, TaN);
2452 T24 = VMUL(LDK(KP707106781), VSUB(T1L, T1I));
2453 {
2454 V T1Q, TaT, T1T, TaU;
2455 {
2456 V T1O, T1P, T1R, T1S;
2457 T1O = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
2458 T1P = LD(&(xi[WS(is, 70)]), ivs, &(xi[0]));
2459 T1Q = VSUB(T1O, T1P);
2460 TaT = VADD(T1O, T1P);
2461 T1R = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
2462 T1S = LD(&(xi[WS(is, 102)]), ivs, &(xi[0]));
2463 T1T = VSUB(T1R, T1S);
2464 TaU = VADD(T1R, T1S);
2465 }
2466 T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q));
2467 Tew = VSUB(TaT, TaU);
2468 T2a = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T));
2469 TaV = VADD(TaT, TaU);
2470 }
2471 {
2472 V T1X, TaQ, T20, TaR;
2473 {
2474 V T1V, T1W, T1Y, T1Z;
2475 T1V = LD(&(xi[WS(is, 118)]), ivs, &(xi[0]));
2476 T1W = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
2477 T1X = VSUB(T1V, T1W);
2478 TaQ = VADD(T1V, T1W);
2479 T1Y = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
2480 T1Z = LD(&(xi[WS(is, 86)]), ivs, &(xi[0]));
2481 T20 = VSUB(T1Y, T1Z);
2482 TaR = VADD(T1Y, T1Z);
2483 }
2484 T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20));
2485 Tex = VSUB(TaQ, TaR);
2486 T29 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X));
2487 TaS = VADD(TaQ, TaR);
2488 }
2489 Tdm = VADD(TaL, TaO);
2490 Tdn = VADD(TaV, TaS);
2491 TdH = VSUB(Tdm, Tdn);
2492 Tev = VSUB(TaJ, TaK);
2493 Tey = VMUL(LDK(KP707106781), VADD(Tew, Tex));
2494 Tez = VADD(Tev, Tey);
2495 Tgi = VSUB(Tev, Tey);
2496 {
2497 V TeA, TeB, T1N, T22;
2498 TeA = VSUB(TaN, TaM);
2499 TeB = VMUL(LDK(KP707106781), VSUB(Tex, Tew));
2500 TeC = VADD(TeA, TeB);
2501 Tgj = VSUB(TeB, TeA);
2502 T1N = VADD(T1F, T1M);
2503 T22 = VADD(T1U, T21);
2504 T23 = VADD(T1N, T22);
2505 T5N = VSUB(T1N, T22);
2506 }
2507 {
2508 V T28, T2b, T7i, T7j;
2509 T28 = VSUB(T24, T27);
2510 T2b = VSUB(T29, T2a);
2511 T2c = VADD(T28, T2b);
2512 T5M = VSUB(T2b, T28);
2513 T7i = VADD(T27, T24);
2514 T7j = VSUB(T21, T1U);
2515 T7k = VADD(T7i, T7j);
2516 T8X = VSUB(T7j, T7i);
2517 }
2518 {
2519 V TaP, TaW, T7f, T7g;
2520 TaP = VSUB(TaL, TaO);
2521 TaW = VSUB(TaS, TaV);
2522 TaX = VFNMS(LDK(KP382683432), TaW, VMUL(LDK(KP923879532), TaP));
2523 Tcg = VFMA(LDK(KP382683432), TaP, VMUL(LDK(KP923879532), TaW));
2524 T7f = VSUB(T1F, T1M);
2525 T7g = VADD(T2a, T29);
2526 T7h = VADD(T7f, T7g);
2527 T8W = VSUB(T7f, T7g);
2528 }
2529 }
2530 {
2531 V T2J, TeL, T2V, Tb9, T30, TeO, T3c, Tbg, T2Q, TeM, T2S, Tbc, T37, TeP, T39;
2532 V Tbj;
2533 {
2534 V T2H, T2I, Tb7, T2T, T2U, Tb8;
2535 T2H = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
2536 T2I = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)]));
2537 Tb7 = VADD(T2H, T2I);
2538 T2T = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
2539 T2U = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)]));
2540 Tb8 = VADD(T2T, T2U);
2541 T2J = VSUB(T2H, T2I);
2542 TeL = VSUB(Tb7, Tb8);
2543 T2V = VSUB(T2T, T2U);
2544 Tb9 = VADD(Tb7, Tb8);
2545 }
2546 {
2547 V T2Y, T2Z, Tbe, T3a, T3b, Tbf;
2548 T2Y = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)]));
2549 T2Z = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
2550 Tbe = VADD(T2Y, T2Z);
2551 T3a = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
2552 T3b = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)]));
2553 Tbf = VADD(T3a, T3b);
2554 T30 = VSUB(T2Y, T2Z);
2555 TeO = VSUB(Tbe, Tbf);
2556 T3c = VSUB(T3a, T3b);
2557 Tbg = VADD(Tbe, Tbf);
2558 }
2559 {
2560 V T2M, Tba, T2P, Tbb;
2561 {
2562 V T2K, T2L, T2N, T2O;
2563 T2K = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
2564 T2L = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)]));
2565 T2M = VSUB(T2K, T2L);
2566 Tba = VADD(T2K, T2L);
2567 T2N = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)]));
2568 T2O = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
2569 T2P = VSUB(T2N, T2O);
2570 Tbb = VADD(T2N, T2O);
2571 }
2572 T2Q = VMUL(LDK(KP707106781), VADD(T2M, T2P));
2573 TeM = VSUB(Tbb, Tba);
2574 T2S = VMUL(LDK(KP707106781), VSUB(T2P, T2M));
2575 Tbc = VADD(Tba, Tbb);
2576 }
2577 {
2578 V T33, Tbh, T36, Tbi;
2579 {
2580 V T31, T32, T34, T35;
2581 T31 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
2582 T32 = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)]));
2583 T33 = VSUB(T31, T32);
2584 Tbh = VADD(T31, T32);
2585 T34 = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)]));
2586 T35 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
2587 T36 = VSUB(T34, T35);
2588 Tbi = VADD(T34, T35);
2589 }
2590 T37 = VMUL(LDK(KP707106781), VADD(T33, T36));
2591 TeP = VSUB(Tbi, Tbh);
2592 T39 = VMUL(LDK(KP707106781), VSUB(T36, T33));
2593 Tbj = VADD(Tbh, Tbi);
2594 }
2595 {
2596 V Tbd, Tbk, TeN, TeQ;
2597 Tbd = VSUB(Tb9, Tbc);
2598 Tbk = VSUB(Tbg, Tbj);
2599 Tbl = VMUL(LDK(KP707106781), VADD(Tbd, Tbk));
2600 Tbu = VMUL(LDK(KP707106781), VSUB(Tbk, Tbd));
2601 {
2602 V Td9, Tda, TeW, TeX;
2603 Td9 = VADD(Tb9, Tbc);
2604 Tda = VADD(Tbg, Tbj);
2605 Tdb = VADD(Td9, Tda);
2606 TdL = VSUB(Tda, Td9);
2607 TeW = VFNMS(LDK(KP382683432), TeL, VMUL(LDK(KP923879532), TeM));
2608 TeX = VFMA(LDK(KP382683432), TeO, VMUL(LDK(KP923879532), TeP));
2609 TeY = VADD(TeW, TeX);
2610 Tgu = VSUB(TeX, TeW);
2611 }
2612 TeN = VFMA(LDK(KP923879532), TeL, VMUL(LDK(KP382683432), TeM));
2613 TeQ = VFNMS(LDK(KP382683432), TeP, VMUL(LDK(KP923879532), TeO));
2614 TeR = VADD(TeN, TeQ);
2615 Tgq = VSUB(TeQ, TeN);
2616 {
2617 V T7t, T7C, T7w, T7D;
2618 {
2619 V T7r, T7s, T7u, T7v;
2620 T7r = VSUB(T2J, T2Q);
2621 T7s = VADD(T2V, T2S);
2622 T7t = VFMA(LDK(KP831469612), T7r, VMUL(LDK(KP555570233), T7s));
2623 T7C = VFNMS(LDK(KP555570233), T7r, VMUL(LDK(KP831469612), T7s));
2624 T7u = VSUB(T30, T37);
2625 T7v = VADD(T3c, T39);
2626 T7w = VFNMS(LDK(KP555570233), T7v, VMUL(LDK(KP831469612), T7u));
2627 T7D = VFMA(LDK(KP555570233), T7u, VMUL(LDK(KP831469612), T7v));
2628 }
2629 T7x = VADD(T7t, T7w);
2630 T98 = VSUB(T7D, T7C);
2631 T7E = VADD(T7C, T7D);
2632 T94 = VSUB(T7w, T7t);
2633 }
2634 {
2635 V T2X, T3q, T3e, T3r;
2636 {
2637 V T2R, T2W, T38, T3d;
2638 T2R = VADD(T2J, T2Q);
2639 T2W = VSUB(T2S, T2V);
2640 T2X = VFMA(LDK(KP980785280), T2R, VMUL(LDK(KP195090322), T2W));
2641 T3q = VFNMS(LDK(KP195090322), T2R, VMUL(LDK(KP980785280), T2W));
2642 T38 = VADD(T30, T37);
2643 T3d = VSUB(T39, T3c);
2644 T3e = VFNMS(LDK(KP195090322), T3d, VMUL(LDK(KP980785280), T38));
2645 T3r = VFMA(LDK(KP195090322), T38, VMUL(LDK(KP980785280), T3d));
2646 }
2647 T3f = VADD(T2X, T3e);
2648 T5Y = VSUB(T3r, T3q);
2649 T3s = VADD(T3q, T3r);
2650 T5U = VSUB(T3e, T2X);
2651 }
2652 }
2653 }
2654 {
2655 V T3Y, Tf6, T4a, TbG, T4f, Tf9, T4r, TbN, T45, Tf7, T47, TbJ, T4m, Tfa, T4o;
2656 V TbQ;
2657 {
2658 V T3W, T3X, TbE, T48, T49, TbF;
2659 T3W = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
2660 T3X = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)]));
2661 TbE = VADD(T3W, T3X);
2662 T48 = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
2663 T49 = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)]));
2664 TbF = VADD(T48, T49);
2665 T3Y = VSUB(T3W, T3X);
2666 Tf6 = VSUB(TbE, TbF);
2667 T4a = VSUB(T48, T49);
2668 TbG = VADD(TbE, TbF);
2669 }
2670 {
2671 V T4d, T4e, TbL, T4p, T4q, TbM;
2672 T4d = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)]));
2673 T4e = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
2674 TbL = VADD(T4d, T4e);
2675 T4p = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
2676 T4q = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)]));
2677 TbM = VADD(T4p, T4q);
2678 T4f = VSUB(T4d, T4e);
2679 Tf9 = VSUB(TbL, TbM);
2680 T4r = VSUB(T4p, T4q);
2681 TbN = VADD(TbL, TbM);
2682 }
2683 {
2684 V T41, TbH, T44, TbI;
2685 {
2686 V T3Z, T40, T42, T43;
2687 T3Z = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
2688 T40 = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)]));
2689 T41 = VSUB(T3Z, T40);
2690 TbH = VADD(T3Z, T40);
2691 T42 = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)]));
2692 T43 = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
2693 T44 = VSUB(T42, T43);
2694 TbI = VADD(T42, T43);
2695 }
2696 T45 = VMUL(LDK(KP707106781), VADD(T41, T44));
2697 Tf7 = VSUB(TbI, TbH);
2698 T47 = VMUL(LDK(KP707106781), VSUB(T44, T41));
2699 TbJ = VADD(TbH, TbI);
2700 }
2701 {
2702 V T4i, TbO, T4l, TbP;
2703 {
2704 V T4g, T4h, T4j, T4k;
2705 T4g = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
2706 T4h = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)]));
2707 T4i = VSUB(T4g, T4h);
2708 TbO = VADD(T4g, T4h);
2709 T4j = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)]));
2710 T4k = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
2711 T4l = VSUB(T4j, T4k);
2712 TbP = VADD(T4j, T4k);
2713 }
2714 T4m = VMUL(LDK(KP707106781), VADD(T4i, T4l));
2715 Tfa = VSUB(TbP, TbO);
2716 T4o = VMUL(LDK(KP707106781), VSUB(T4l, T4i));
2717 TbQ = VADD(TbO, TbP);
2718 }
2719 {
2720 V TbK, TbR, Tf8, Tfb;
2721 TbK = VSUB(TbG, TbJ);
2722 TbR = VSUB(TbN, TbQ);
2723 TbS = VMUL(LDK(KP707106781), VADD(TbK, TbR));
2724 Tc1 = VMUL(LDK(KP707106781), VSUB(TbR, TbK));
2725 {
2726 V Tdg, Tdh, Tfh, Tfi;
2727 Tdg = VADD(TbG, TbJ);
2728 Tdh = VADD(TbN, TbQ);
2729 Tdi = VADD(Tdg, Tdh);
2730 TdO = VSUB(Tdh, Tdg);
2731 Tfh = VFNMS(LDK(KP382683432), Tf6, VMUL(LDK(KP923879532), Tf7));
2732 Tfi = VFMA(LDK(KP382683432), Tf9, VMUL(LDK(KP923879532), Tfa));
2733 Tfj = VADD(Tfh, Tfi);
2734 Tgy = VSUB(Tfi, Tfh);
2735 }
2736 Tf8 = VFMA(LDK(KP923879532), Tf6, VMUL(LDK(KP382683432), Tf7));
2737 Tfb = VFNMS(LDK(KP382683432), Tfa, VMUL(LDK(KP923879532), Tf9));
2738 Tfc = VADD(Tf8, Tfb);
2739 TgA = VSUB(Tfb, Tf8);
2740 {
2741 V T7M, T7V, T7P, T7W;
2742 {
2743 V T7K, T7L, T7N, T7O;
2744 T7K = VSUB(T3Y, T45);
2745 T7L = VADD(T4a, T47);
2746 T7M = VFMA(LDK(KP831469612), T7K, VMUL(LDK(KP555570233), T7L));
2747 T7V = VFNMS(LDK(KP555570233), T7K, VMUL(LDK(KP831469612), T7L));
2748 T7N = VSUB(T4f, T4m);
2749 T7O = VADD(T4r, T4o);
2750 T7P = VFNMS(LDK(KP555570233), T7O, VMUL(LDK(KP831469612), T7N));
2751 T7W = VFMA(LDK(KP555570233), T7N, VMUL(LDK(KP831469612), T7O));
2752 }
2753 T7Q = VADD(T7M, T7P);
2754 T9e = VSUB(T7P, T7M);
2755 T7X = VADD(T7V, T7W);
2756 T9c = VSUB(T7W, T7V);
2757 }
2758 {
2759 V T4c, T4F, T4t, T4G;
2760 {
2761 V T46, T4b, T4n, T4s;
2762 T46 = VADD(T3Y, T45);
2763 T4b = VSUB(T47, T4a);
2764 T4c = VFMA(LDK(KP980785280), T46, VMUL(LDK(KP195090322), T4b));
2765 T4F = VFNMS(LDK(KP195090322), T46, VMUL(LDK(KP980785280), T4b));
2766 T4n = VADD(T4f, T4m);
2767 T4s = VSUB(T4o, T4r);
2768 T4t = VFNMS(LDK(KP195090322), T4s, VMUL(LDK(KP980785280), T4n));
2769 T4G = VFMA(LDK(KP195090322), T4n, VMUL(LDK(KP980785280), T4s));
2770 }
2771 T4u = VADD(T4c, T4t);
2772 T64 = VSUB(T4t, T4c);
2773 T4H = VADD(T4F, T4G);
2774 T62 = VSUB(T4G, T4F);
2775 }
2776 }
2777 }
2778 {
2779 V Td5, Tdx, TdC, TdE, Tdk, Tdt, Tds, Tdy, Tdz, TdD;
2780 {
2781 V Td1, Td4, TdA, TdB;
2782 Td1 = VADD(TcZ, Td0);
2783 Td4 = VADD(Td2, Td3);
2784 Td5 = VSUB(Td1, Td4);
2785 Tdx = VADD(Td1, Td4);
2786 TdA = VADD(Td8, Tdb);
2787 TdB = VADD(Tdf, Tdi);
2788 TdC = VADD(TdA, TdB);
2789 TdE = VBYI(VSUB(TdB, TdA));
2790 }
2791 {
2792 V Tdc, Tdj, Tdo, Tdr;
2793 Tdc = VSUB(Td8, Tdb);
2794 Tdj = VSUB(Tdf, Tdi);
2795 Tdk = VMUL(LDK(KP707106781), VADD(Tdc, Tdj));
2796 Tdt = VMUL(LDK(KP707106781), VSUB(Tdj, Tdc));
2797 Tdo = VADD(Tdm, Tdn);
2798 Tdr = VADD(Tdp, Tdq);
2799 Tds = VSUB(Tdo, Tdr);
2800 Tdy = VADD(Tdr, Tdo);
2801 }
2802 Tdz = VADD(Tdx, Tdy);
2803 ST(&(xo[WS(os, 64)]), VSUB(Tdz, TdC), ovs, &(xo[0]));
2804 ST(&(xo[0]), VADD(Tdz, TdC), ovs, &(xo[0]));
2805 TdD = VSUB(Tdx, Tdy);
2806 ST(&(xo[WS(os, 96)]), VSUB(TdD, TdE), ovs, &(xo[0]));
2807 ST(&(xo[WS(os, 32)]), VADD(TdD, TdE), ovs, &(xo[0]));
2808 {
2809 V Tdl, Tdu, Tdv, Tdw;
2810 Tdl = VADD(Td5, Tdk);
2811 Tdu = VBYI(VADD(Tds, Tdt));
2812 ST(&(xo[WS(os, 112)]), VSUB(Tdl, Tdu), ovs, &(xo[0]));
2813 ST(&(xo[WS(os, 16)]), VADD(Tdl, Tdu), ovs, &(xo[0]));
2814 Tdv = VSUB(Td5, Tdk);
2815 Tdw = VBYI(VSUB(Tdt, Tds));
2816 ST(&(xo[WS(os, 80)]), VSUB(Tdv, Tdw), ovs, &(xo[0]));
2817 ST(&(xo[WS(os, 48)]), VADD(Tdv, Tdw), ovs, &(xo[0]));
2818 }
2819 }
2820 {
2821 V TdJ, Te4, TdX, Te5, TdQ, Te1, TdU, Te2;
2822 {
2823 V TdF, TdI, TdV, TdW;
2824 TdF = VSUB(TcZ, Td0);
2825 TdI = VMUL(LDK(KP707106781), VADD(TdG, TdH));
2826 TdJ = VADD(TdF, TdI);
2827 Te4 = VSUB(TdF, TdI);
2828 TdV = VFNMS(LDK(KP382683432), TdK, VMUL(LDK(KP923879532), TdL));
2829 TdW = VFMA(LDK(KP382683432), TdN, VMUL(LDK(KP923879532), TdO));
2830 TdX = VADD(TdV, TdW);
2831 Te5 = VSUB(TdW, TdV);
2832 }
2833 {
2834 V TdM, TdP, TdS, TdT;
2835 TdM = VFMA(LDK(KP923879532), TdK, VMUL(LDK(KP382683432), TdL));
2836 TdP = VFNMS(LDK(KP382683432), TdO, VMUL(LDK(KP923879532), TdN));
2837 TdQ = VADD(TdM, TdP);
2838 Te1 = VSUB(TdP, TdM);
2839 TdS = VSUB(Td3, Td2);
2840 TdT = VMUL(LDK(KP707106781), VSUB(TdH, TdG));
2841 TdU = VADD(TdS, TdT);
2842 Te2 = VSUB(TdT, TdS);
2843 }
2844 {
2845 V TdR, TdY, Te7, Te8;
2846 TdR = VADD(TdJ, TdQ);
2847 TdY = VBYI(VADD(TdU, TdX));
2848 ST(&(xo[WS(os, 120)]), VSUB(TdR, TdY), ovs, &(xo[0]));
2849 ST(&(xo[WS(os, 8)]), VADD(TdR, TdY), ovs, &(xo[0]));
2850 Te7 = VBYI(VADD(Te2, Te1));
2851 Te8 = VADD(Te4, Te5);
2852 ST(&(xo[WS(os, 24)]), VADD(Te7, Te8), ovs, &(xo[0]));
2853 ST(&(xo[WS(os, 104)]), VSUB(Te8, Te7), ovs, &(xo[0]));
2854 }
2855 {
2856 V TdZ, Te0, Te3, Te6;
2857 TdZ = VSUB(TdJ, TdQ);
2858 Te0 = VBYI(VSUB(TdX, TdU));
2859 ST(&(xo[WS(os, 72)]), VSUB(TdZ, Te0), ovs, &(xo[0]));
2860 ST(&(xo[WS(os, 56)]), VADD(TdZ, Te0), ovs, &(xo[0]));
2861 Te3 = VBYI(VSUB(Te1, Te2));
2862 Te6 = VSUB(Te4, Te5);
2863 ST(&(xo[WS(os, 40)]), VADD(Te3, Te6), ovs, &(xo[0]));
2864 ST(&(xo[WS(os, 88)]), VSUB(Te6, Te3), ovs, &(xo[0]));
2865 }
2866 }
2867 {
2868 V TaZ, Tcs, Tci, Tcq, Tc4, Tct, Tcl, Tcp;
2869 {
2870 V Tat, TaY, Tce, Tch;
2871 Tat = VADD(Tad, Tas);
2872 TaY = VADD(TaI, TaX);
2873 TaZ = VADD(Tat, TaY);
2874 Tcs = VSUB(Tat, TaY);
2875 Tce = VADD(Tcc, Tcd);
2876 Tch = VADD(Tcf, Tcg);
2877 Tci = VADD(Tce, Tch);
2878 Tcq = VSUB(Tch, Tce);
2879 {
2880 V Tbw, Tcj, Tc3, Tck;
2881 {
2882 V Tbm, Tbv, TbT, Tc2;
2883 Tbm = VADD(Tb6, Tbl);
2884 Tbv = VADD(Tbt, Tbu);
2885 Tbw = VFMA(LDK(KP980785280), Tbm, VMUL(LDK(KP195090322), Tbv));
2886 Tcj = VFNMS(LDK(KP195090322), Tbm, VMUL(LDK(KP980785280), Tbv));
2887 TbT = VADD(TbD, TbS);
2888 Tc2 = VADD(Tc0, Tc1);
2889 Tc3 = VFNMS(LDK(KP195090322), Tc2, VMUL(LDK(KP980785280), TbT));
2890 Tck = VFMA(LDK(KP195090322), TbT, VMUL(LDK(KP980785280), Tc2));
2891 }
2892 Tc4 = VADD(Tbw, Tc3);
2893 Tct = VSUB(Tck, Tcj);
2894 Tcl = VADD(Tcj, Tck);
2895 Tcp = VSUB(Tc3, Tbw);
2896 }
2897 }
2898 {
2899 V Tc5, Tcm, Tcv, Tcw;
2900 Tc5 = VADD(TaZ, Tc4);
2901 Tcm = VBYI(VADD(Tci, Tcl));
2902 ST(&(xo[WS(os, 124)]), VSUB(Tc5, Tcm), ovs, &(xo[0]));
2903 ST(&(xo[WS(os, 4)]), VADD(Tc5, Tcm), ovs, &(xo[0]));
2904 Tcv = VBYI(VADD(Tcq, Tcp));
2905 Tcw = VADD(Tcs, Tct);
2906 ST(&(xo[WS(os, 28)]), VADD(Tcv, Tcw), ovs, &(xo[0]));
2907 ST(&(xo[WS(os, 100)]), VSUB(Tcw, Tcv), ovs, &(xo[0]));
2908 }
2909 {
2910 V Tcn, Tco, Tcr, Tcu;
2911 Tcn = VSUB(TaZ, Tc4);
2912 Tco = VBYI(VSUB(Tcl, Tci));
2913 ST(&(xo[WS(os, 68)]), VSUB(Tcn, Tco), ovs, &(xo[0]));
2914 ST(&(xo[WS(os, 60)]), VADD(Tcn, Tco), ovs, &(xo[0]));
2915 Tcr = VBYI(VSUB(Tcp, Tcq));
2916 Tcu = VSUB(Tcs, Tct);
2917 ST(&(xo[WS(os, 36)]), VADD(Tcr, Tcu), ovs, &(xo[0]));
2918 ST(&(xo[WS(os, 92)]), VSUB(Tcu, Tcr), ovs, &(xo[0]));
2919 }
2920 }
2921 {
2922 V Tcz, TcU, TcK, TcS, TcG, TcV, TcN, TcR;
2923 {
2924 V Tcx, Tcy, TcI, TcJ;
2925 Tcx = VSUB(Tad, Tas);
2926 Tcy = VSUB(Tcg, Tcf);
2927 Tcz = VADD(Tcx, Tcy);
2928 TcU = VSUB(Tcx, Tcy);
2929 TcI = VSUB(Tcd, Tcc);
2930 TcJ = VSUB(TaX, TaI);
2931 TcK = VADD(TcI, TcJ);
2932 TcS = VSUB(TcJ, TcI);
2933 {
2934 V TcC, TcL, TcF, TcM;
2935 {
2936 V TcA, TcB, TcD, TcE;
2937 TcA = VSUB(Tb6, Tbl);
2938 TcB = VSUB(Tbu, Tbt);
2939 TcC = VFMA(LDK(KP831469612), TcA, VMUL(LDK(KP555570233), TcB));
2940 TcL = VFNMS(LDK(KP555570233), TcA, VMUL(LDK(KP831469612), TcB));
2941 TcD = VSUB(TbD, TbS);
2942 TcE = VSUB(Tc1, Tc0);
2943 TcF = VFNMS(LDK(KP555570233), TcE, VMUL(LDK(KP831469612), TcD));
2944 TcM = VFMA(LDK(KP555570233), TcD, VMUL(LDK(KP831469612), TcE));
2945 }
2946 TcG = VADD(TcC, TcF);
2947 TcV = VSUB(TcM, TcL);
2948 TcN = VADD(TcL, TcM);
2949 TcR = VSUB(TcF, TcC);
2950 }
2951 }
2952 {
2953 V TcH, TcO, TcX, TcY;
2954 TcH = VADD(Tcz, TcG);
2955 TcO = VBYI(VADD(TcK, TcN));
2956 ST(&(xo[WS(os, 116)]), VSUB(TcH, TcO), ovs, &(xo[0]));
2957 ST(&(xo[WS(os, 12)]), VADD(TcH, TcO), ovs, &(xo[0]));
2958 TcX = VBYI(VADD(TcS, TcR));
2959 TcY = VADD(TcU, TcV);
2960 ST(&(xo[WS(os, 20)]), VADD(TcX, TcY), ovs, &(xo[0]));
2961 ST(&(xo[WS(os, 108)]), VSUB(TcY, TcX), ovs, &(xo[0]));
2962 }
2963 {
2964 V TcP, TcQ, TcT, TcW;
2965 TcP = VSUB(Tcz, TcG);
2966 TcQ = VBYI(VSUB(TcN, TcK));
2967 ST(&(xo[WS(os, 76)]), VSUB(TcP, TcQ), ovs, &(xo[0]));
2968 ST(&(xo[WS(os, 52)]), VADD(TcP, TcQ), ovs, &(xo[0]));
2969 TcT = VBYI(VSUB(TcR, TcS));
2970 TcW = VSUB(TcU, TcV);
2971 ST(&(xo[WS(os, 44)]), VADD(TcT, TcW), ovs, &(xo[0]));
2972 ST(&(xo[WS(os, 84)]), VSUB(TcW, TcT), ovs, &(xo[0]));
2973 }
2974 }
2975 {
2976 V TeF, Tg8, TfI, Tg0, Tfy, Tga, TfG, TfP, Tfm, TfJ, TfB, TfF, TfW, Tgb, Tg3;
2977 V Tg7;
2978 {
2979 V Tel, TfY, TeE, TfZ, Teu, TeD;
2980 Tel = VADD(Ted, Tek);
2981 TfY = VSUB(Tft, Tfq);
2982 Teu = VFMA(LDK(KP980785280), Teq, VMUL(LDK(KP195090322), Tet));
2983 TeD = VFNMS(LDK(KP195090322), TeC, VMUL(LDK(KP980785280), Tez));
2984 TeE = VADD(Teu, TeD);
2985 TfZ = VSUB(TeD, Teu);
2986 TeF = VADD(Tel, TeE);
2987 Tg8 = VSUB(TfZ, TfY);
2988 TfI = VSUB(Tel, TeE);
2989 Tg0 = VADD(TfY, TfZ);
2990 }
2991 {
2992 V Tfu, TfN, Tfx, TfO, Tfv, Tfw;
2993 Tfu = VADD(Tfq, Tft);
2994 TfN = VSUB(Ted, Tek);
2995 Tfv = VFNMS(LDK(KP195090322), Teq, VMUL(LDK(KP980785280), Tet));
2996 Tfw = VFMA(LDK(KP195090322), Tez, VMUL(LDK(KP980785280), TeC));
2997 Tfx = VADD(Tfv, Tfw);
2998 TfO = VSUB(Tfw, Tfv);
2999 Tfy = VADD(Tfu, Tfx);
3000 Tga = VSUB(TfN, TfO);
3001 TfG = VSUB(Tfx, Tfu);
3002 TfP = VADD(TfN, TfO);
3003 }
3004 {
3005 V Tf0, Tfz, Tfl, TfA;
3006 {
3007 V TeS, TeZ, Tfd, Tfk;
3008 TeS = VADD(TeK, TeR);
3009 TeZ = VADD(TeV, TeY);
3010 Tf0 = VFMA(LDK(KP995184726), TeS, VMUL(LDK(KP098017140), TeZ));
3011 Tfz = VFNMS(LDK(KP098017140), TeS, VMUL(LDK(KP995184726), TeZ));
3012 Tfd = VADD(Tf5, Tfc);
3013 Tfk = VADD(Tfg, Tfj);
3014 Tfl = VFNMS(LDK(KP098017140), Tfk, VMUL(LDK(KP995184726), Tfd));
3015 TfA = VFMA(LDK(KP098017140), Tfd, VMUL(LDK(KP995184726), Tfk));
3016 }
3017 Tfm = VADD(Tf0, Tfl);
3018 TfJ = VSUB(TfA, Tfz);
3019 TfB = VADD(Tfz, TfA);
3020 TfF = VSUB(Tfl, Tf0);
3021 }
3022 {
3023 V TfS, Tg1, TfV, Tg2;
3024 {
3025 V TfQ, TfR, TfT, TfU;
3026 TfQ = VSUB(TeK, TeR);
3027 TfR = VSUB(TeY, TeV);
3028 TfS = VFMA(LDK(KP773010453), TfQ, VMUL(LDK(KP634393284), TfR));
3029 Tg1 = VFNMS(LDK(KP634393284), TfQ, VMUL(LDK(KP773010453), TfR));
3030 TfT = VSUB(Tf5, Tfc);
3031 TfU = VSUB(Tfj, Tfg);
3032 TfV = VFNMS(LDK(KP634393284), TfU, VMUL(LDK(KP773010453), TfT));
3033 Tg2 = VFMA(LDK(KP634393284), TfT, VMUL(LDK(KP773010453), TfU));
3034 }
3035 TfW = VADD(TfS, TfV);
3036 Tgb = VSUB(Tg2, Tg1);
3037 Tg3 = VADD(Tg1, Tg2);
3038 Tg7 = VSUB(TfV, TfS);
3039 }
3040 {
3041 V Tfn, TfC, Tg9, Tgc;
3042 Tfn = VADD(TeF, Tfm);
3043 TfC = VBYI(VADD(Tfy, TfB));
3044 ST(&(xo[WS(os, 126)]), VSUB(Tfn, TfC), ovs, &(xo[0]));
3045 ST(&(xo[WS(os, 2)]), VADD(Tfn, TfC), ovs, &(xo[0]));
3046 Tg9 = VBYI(VSUB(Tg7, Tg8));
3047 Tgc = VSUB(Tga, Tgb);
3048 ST(&(xo[WS(os, 46)]), VADD(Tg9, Tgc), ovs, &(xo[0]));
3049 ST(&(xo[WS(os, 82)]), VSUB(Tgc, Tg9), ovs, &(xo[0]));
3050 }
3051 {
3052 V Tgd, Tge, TfD, TfE;
3053 Tgd = VBYI(VADD(Tg8, Tg7));
3054 Tge = VADD(Tga, Tgb);
3055 ST(&(xo[WS(os, 18)]), VADD(Tgd, Tge), ovs, &(xo[0]));
3056 ST(&(xo[WS(os, 110)]), VSUB(Tge, Tgd), ovs, &(xo[0]));
3057 TfD = VSUB(TeF, Tfm);
3058 TfE = VBYI(VSUB(TfB, Tfy));
3059 ST(&(xo[WS(os, 66)]), VSUB(TfD, TfE), ovs, &(xo[0]));
3060 ST(&(xo[WS(os, 62)]), VADD(TfD, TfE), ovs, &(xo[0]));
3061 }
3062 {
3063 V TfH, TfK, TfX, Tg4;
3064 TfH = VBYI(VSUB(TfF, TfG));
3065 TfK = VSUB(TfI, TfJ);
3066 ST(&(xo[WS(os, 34)]), VADD(TfH, TfK), ovs, &(xo[0]));
3067 ST(&(xo[WS(os, 94)]), VSUB(TfK, TfH), ovs, &(xo[0]));
3068 TfX = VADD(TfP, TfW);
3069 Tg4 = VBYI(VADD(Tg0, Tg3));
3070 ST(&(xo[WS(os, 114)]), VSUB(TfX, Tg4), ovs, &(xo[0]));
3071 ST(&(xo[WS(os, 14)]), VADD(TfX, Tg4), ovs, &(xo[0]));
3072 }
3073 {
3074 V Tg5, Tg6, TfL, TfM;
3075 Tg5 = VSUB(TfP, TfW);
3076 Tg6 = VBYI(VSUB(Tg3, Tg0));
3077 ST(&(xo[WS(os, 78)]), VSUB(Tg5, Tg6), ovs, &(xo[0]));
3078 ST(&(xo[WS(os, 50)]), VADD(Tg5, Tg6), ovs, &(xo[0]));
3079 TfL = VBYI(VADD(TfG, TfF));
3080 TfM = VADD(TfI, TfJ);
3081 ST(&(xo[WS(os, 30)]), VADD(TfL, TfM), ovs, &(xo[0]));
3082 ST(&(xo[WS(os, 98)]), VSUB(TfM, TfL), ovs, &(xo[0]));
3083 }
3084 }
3085 {
3086 V Tgp, Thm, TgW, The, TgM, Tho, TgU, Th3, TgE, TgX, TgP, TgT, Tha, Thp, Thh;
3087 V Thl;
3088 {
3089 V Tgh, Thc, Tgo, Thd, Tgk, Tgn;
3090 Tgh = VSUB(Tgf, Tgg);
3091 Thc = VADD(TgH, TgG);
3092 Tgk = VFMA(LDK(KP555570233), Tgi, VMUL(LDK(KP831469612), Tgj));
3093 Tgn = VFNMS(LDK(KP555570233), Tgm, VMUL(LDK(KP831469612), Tgl));
3094 Tgo = VSUB(Tgk, Tgn);
3095 Thd = VADD(Tgn, Tgk);
3096 Tgp = VADD(Tgh, Tgo);
3097 Thm = VSUB(Thd, Thc);
3098 TgW = VSUB(Tgh, Tgo);
3099 The = VADD(Thc, Thd);
3100 }
3101 {
3102 V TgI, Th1, TgL, Th2, TgJ, TgK;
3103 TgI = VSUB(TgG, TgH);
3104 Th1 = VADD(Tgf, Tgg);
3105 TgJ = VFNMS(LDK(KP555570233), Tgj, VMUL(LDK(KP831469612), Tgi));
3106 TgK = VFMA(LDK(KP831469612), Tgm, VMUL(LDK(KP555570233), Tgl));
3107 TgL = VSUB(TgJ, TgK);
3108 Th2 = VADD(TgK, TgJ);
3109 TgM = VADD(TgI, TgL);
3110 Tho = VSUB(Th1, Th2);
3111 TgU = VSUB(TgL, TgI);
3112 Th3 = VADD(Th1, Th2);
3113 }
3114 {
3115 V Tgw, TgN, TgD, TgO;
3116 {
3117 V Tgs, Tgv, Tgz, TgC;
3118 Tgs = VSUB(Tgq, Tgr);
3119 Tgv = VSUB(Tgt, Tgu);
3120 Tgw = VFMA(LDK(KP471396736), Tgs, VMUL(LDK(KP881921264), Tgv));
3121 TgN = VFNMS(LDK(KP471396736), Tgv, VMUL(LDK(KP881921264), Tgs));
3122 Tgz = VSUB(Tgx, Tgy);
3123 TgC = VSUB(TgA, TgB);
3124 TgD = VFNMS(LDK(KP471396736), TgC, VMUL(LDK(KP881921264), Tgz));
3125 TgO = VFMA(LDK(KP881921264), TgC, VMUL(LDK(KP471396736), Tgz));
3126 }
3127 TgE = VADD(Tgw, TgD);
3128 TgX = VSUB(TgO, TgN);
3129 TgP = VADD(TgN, TgO);
3130 TgT = VSUB(TgD, Tgw);
3131 }
3132 {
3133 V Th6, Thf, Th9, Thg;
3134 {
3135 V Th4, Th5, Th7, Th8;
3136 Th4 = VADD(Tgr, Tgq);
3137 Th5 = VADD(Tgt, Tgu);
3138 Th6 = VFMA(LDK(KP290284677), Th4, VMUL(LDK(KP956940335), Th5));
3139 Thf = VFNMS(LDK(KP290284677), Th5, VMUL(LDK(KP956940335), Th4));
3140 Th7 = VADD(Tgx, Tgy);
3141 Th8 = VADD(TgB, TgA);
3142 Th9 = VFNMS(LDK(KP290284677), Th8, VMUL(LDK(KP956940335), Th7));
3143 Thg = VFMA(LDK(KP956940335), Th8, VMUL(LDK(KP290284677), Th7));
3144 }
3145 Tha = VADD(Th6, Th9);
3146 Thp = VSUB(Thg, Thf);
3147 Thh = VADD(Thf, Thg);
3148 Thl = VSUB(Th9, Th6);
3149 }
3150 {
3151 V TgF, TgQ, Thn, Thq;
3152 TgF = VADD(Tgp, TgE);
3153 TgQ = VBYI(VADD(TgM, TgP));
3154 ST(&(xo[WS(os, 118)]), VSUB(TgF, TgQ), ovs, &(xo[0]));
3155 ST(&(xo[WS(os, 10)]), VADD(TgF, TgQ), ovs, &(xo[0]));
3156 Thn = VBYI(VSUB(Thl, Thm));
3157 Thq = VSUB(Tho, Thp);
3158 ST(&(xo[WS(os, 38)]), VADD(Thn, Thq), ovs, &(xo[0]));
3159 ST(&(xo[WS(os, 90)]), VSUB(Thq, Thn), ovs, &(xo[0]));
3160 }
3161 {
3162 V Thr, Ths, TgR, TgS;
3163 Thr = VBYI(VADD(Thm, Thl));
3164 Ths = VADD(Tho, Thp);
3165 ST(&(xo[WS(os, 26)]), VADD(Thr, Ths), ovs, &(xo[0]));
3166 ST(&(xo[WS(os, 102)]), VSUB(Ths, Thr), ovs, &(xo[0]));
3167 TgR = VSUB(Tgp, TgE);
3168 TgS = VBYI(VSUB(TgP, TgM));
3169 ST(&(xo[WS(os, 74)]), VSUB(TgR, TgS), ovs, &(xo[0]));
3170 ST(&(xo[WS(os, 54)]), VADD(TgR, TgS), ovs, &(xo[0]));
3171 }
3172 {
3173 V TgV, TgY, Thb, Thi;
3174 TgV = VBYI(VSUB(TgT, TgU));
3175 TgY = VSUB(TgW, TgX);
3176 ST(&(xo[WS(os, 42)]), VADD(TgV, TgY), ovs, &(xo[0]));
3177 ST(&(xo[WS(os, 86)]), VSUB(TgY, TgV), ovs, &(xo[0]));
3178 Thb = VADD(Th3, Tha);
3179 Thi = VBYI(VADD(The, Thh));
3180 ST(&(xo[WS(os, 122)]), VSUB(Thb, Thi), ovs, &(xo[0]));
3181 ST(&(xo[WS(os, 6)]), VADD(Thb, Thi), ovs, &(xo[0]));
3182 }
3183 {
3184 V Thj, Thk, TgZ, Th0;
3185 Thj = VSUB(Th3, Tha);
3186 Thk = VBYI(VSUB(Thh, The));
3187 ST(&(xo[WS(os, 70)]), VSUB(Thj, Thk), ovs, &(xo[0]));
3188 ST(&(xo[WS(os, 58)]), VADD(Thj, Thk), ovs, &(xo[0]));
3189 TgZ = VBYI(VADD(TgU, TgT));
3190 Th0 = VADD(TgW, TgX);
3191 ST(&(xo[WS(os, 22)]), VADD(TgZ, Th0), ovs, &(xo[0]));
3192 ST(&(xo[WS(os, 106)]), VSUB(Th0, TgZ), ovs, &(xo[0]));
3193 }
3194 }
3195 {
3196 V T80, T8n, T8f, T8j, T8A, T8P, T8H, T8L, T7n, T8M, T8O, T8c, T8k, T8t, T8E;
3197 V T8m;
3198 {
3199 V T7G, T8d, T7Z, T8e;
3200 {
3201 V T7y, T7F, T7R, T7Y;
3202 T7y = VADD(T7q, T7x);
3203 T7F = VADD(T7B, T7E);
3204 T7G = VFMA(LDK(KP989176509), T7y, VMUL(LDK(KP146730474), T7F));
3205 T8d = VFNMS(LDK(KP146730474), T7y, VMUL(LDK(KP989176509), T7F));
3206 T7R = VADD(T7J, T7Q);
3207 T7Y = VADD(T7U, T7X);
3208 T7Z = VFNMS(LDK(KP146730474), T7Y, VMUL(LDK(KP989176509), T7R));
3209 T8e = VFMA(LDK(KP146730474), T7R, VMUL(LDK(KP989176509), T7Y));
3210 }
3211 T80 = VADD(T7G, T7Z);
3212 T8n = VSUB(T8e, T8d);
3213 T8f = VADD(T8d, T8e);
3214 T8j = VSUB(T7Z, T7G);
3215 }
3216 {
3217 V T8w, T8F, T8z, T8G;
3218 {
3219 V T8u, T8v, T8x, T8y;
3220 T8u = VSUB(T7q, T7x);
3221 T8v = VSUB(T7E, T7B);
3222 T8w = VFMA(LDK(KP803207531), T8u, VMUL(LDK(KP595699304), T8v));
3223 T8F = VFNMS(LDK(KP595699304), T8u, VMUL(LDK(KP803207531), T8v));
3224 T8x = VSUB(T7J, T7Q);
3225 T8y = VSUB(T7X, T7U);
3226 T8z = VFNMS(LDK(KP595699304), T8y, VMUL(LDK(KP803207531), T8x));
3227 T8G = VFMA(LDK(KP595699304), T8x, VMUL(LDK(KP803207531), T8y));
3228 }
3229 T8A = VADD(T8w, T8z);
3230 T8P = VSUB(T8G, T8F);
3231 T8H = VADD(T8F, T8G);
3232 T8L = VSUB(T8z, T8w);
3233 }
3234 {
3235 V T77, T8r, T88, T8C, T7m, T8D, T8b, T8s, T76, T87;
3236 T76 = VADD(T72, T75);
3237 T77 = VADD(T6Z, T76);
3238 T8r = VSUB(T6Z, T76);
3239 T87 = VADD(T85, T86);
3240 T88 = VADD(T84, T87);
3241 T8C = VSUB(T87, T84);
3242 {
3243 V T7e, T7l, T89, T8a;
3244 T7e = VFMA(LDK(KP956940335), T7a, VMUL(LDK(KP290284677), T7d));
3245 T7l = VFNMS(LDK(KP290284677), T7k, VMUL(LDK(KP956940335), T7h));
3246 T7m = VADD(T7e, T7l);
3247 T8D = VSUB(T7l, T7e);
3248 T89 = VFNMS(LDK(KP290284677), T7a, VMUL(LDK(KP956940335), T7d));
3249 T8a = VFMA(LDK(KP290284677), T7h, VMUL(LDK(KP956940335), T7k));
3250 T8b = VADD(T89, T8a);
3251 T8s = VSUB(T8a, T89);
3252 }
3253 T7n = VADD(T77, T7m);
3254 T8M = VSUB(T8D, T8C);
3255 T8O = VSUB(T8r, T8s);
3256 T8c = VADD(T88, T8b);
3257 T8k = VSUB(T8b, T88);
3258 T8t = VADD(T8r, T8s);
3259 T8E = VADD(T8C, T8D);
3260 T8m = VSUB(T77, T7m);
3261 }
3262 {
3263 V T81, T8g, T8N, T8Q;
3264 T81 = VADD(T7n, T80);
3265 T8g = VBYI(VADD(T8c, T8f));
3266 ST(&(xo[WS(os, 125)]), VSUB(T81, T8g), ovs, &(xo[WS(os, 1)]));
3267 ST(&(xo[WS(os, 3)]), VADD(T81, T8g), ovs, &(xo[WS(os, 1)]));
3268 T8N = VBYI(VSUB(T8L, T8M));
3269 T8Q = VSUB(T8O, T8P);
3270 ST(&(xo[WS(os, 45)]), VADD(T8N, T8Q), ovs, &(xo[WS(os, 1)]));
3271 ST(&(xo[WS(os, 83)]), VSUB(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
3272 }
3273 {
3274 V T8R, T8S, T8h, T8i;
3275 T8R = VBYI(VADD(T8M, T8L));
3276 T8S = VADD(T8O, T8P);
3277 ST(&(xo[WS(os, 19)]), VADD(T8R, T8S), ovs, &(xo[WS(os, 1)]));
3278 ST(&(xo[WS(os, 109)]), VSUB(T8S, T8R), ovs, &(xo[WS(os, 1)]));
3279 T8h = VSUB(T7n, T80);
3280 T8i = VBYI(VSUB(T8f, T8c));
3281 ST(&(xo[WS(os, 67)]), VSUB(T8h, T8i), ovs, &(xo[WS(os, 1)]));
3282 ST(&(xo[WS(os, 61)]), VADD(T8h, T8i), ovs, &(xo[WS(os, 1)]));
3283 }
3284 {
3285 V T8l, T8o, T8B, T8I;
3286 T8l = VBYI(VSUB(T8j, T8k));
3287 T8o = VSUB(T8m, T8n);
3288 ST(&(xo[WS(os, 35)]), VADD(T8l, T8o), ovs, &(xo[WS(os, 1)]));
3289 ST(&(xo[WS(os, 93)]), VSUB(T8o, T8l), ovs, &(xo[WS(os, 1)]));
3290 T8B = VADD(T8t, T8A);
3291 T8I = VBYI(VADD(T8E, T8H));
3292 ST(&(xo[WS(os, 115)]), VSUB(T8B, T8I), ovs, &(xo[WS(os, 1)]));
3293 ST(&(xo[WS(os, 13)]), VADD(T8B, T8I), ovs, &(xo[WS(os, 1)]));
3294 }
3295 {
3296 V T8J, T8K, T8p, T8q;
3297 T8J = VSUB(T8t, T8A);
3298 T8K = VBYI(VSUB(T8H, T8E));
3299 ST(&(xo[WS(os, 77)]), VSUB(T8J, T8K), ovs, &(xo[WS(os, 1)]));
3300 ST(&(xo[WS(os, 51)]), VADD(T8J, T8K), ovs, &(xo[WS(os, 1)]));
3301 T8p = VBYI(VADD(T8k, T8j));
3302 T8q = VADD(T8m, T8n);
3303 ST(&(xo[WS(os, 29)]), VADD(T8p, T8q), ovs, &(xo[WS(os, 1)]));
3304 ST(&(xo[WS(os, 99)]), VSUB(T8q, T8p), ovs, &(xo[WS(os, 1)]));
3305 }
3306 }
3307 {
3308 V T4K, T5d, T55, T59, T5q, T5F, T5x, T5B, T2f, T5C, T5E, T52, T5a, T5j, T5u;
3309 V T5c;
3310 {
3311 V T3u, T53, T4J, T54;
3312 {
3313 V T3g, T3t, T4v, T4I;
3314 T3g = VADD(T2G, T3f);
3315 T3t = VADD(T3p, T3s);
3316 T3u = VFMA(LDK(KP998795456), T3g, VMUL(LDK(KP049067674), T3t));
3317 T53 = VFNMS(LDK(KP049067674), T3g, VMUL(LDK(KP998795456), T3t));
3318 T4v = VADD(T3V, T4u);
3319 T4I = VADD(T4E, T4H);
3320 T4J = VFNMS(LDK(KP049067674), T4I, VMUL(LDK(KP998795456), T4v));
3321 T54 = VFMA(LDK(KP049067674), T4v, VMUL(LDK(KP998795456), T4I));
3322 }
3323 T4K = VADD(T3u, T4J);
3324 T5d = VSUB(T54, T53);
3325 T55 = VADD(T53, T54);
3326 T59 = VSUB(T4J, T3u);
3327 }
3328 {
3329 V T5m, T5v, T5p, T5w;
3330 {
3331 V T5k, T5l, T5n, T5o;
3332 T5k = VSUB(T2G, T3f);
3333 T5l = VSUB(T3s, T3p);
3334 T5m = VFMA(LDK(KP740951125), T5k, VMUL(LDK(KP671558954), T5l));
3335 T5v = VFNMS(LDK(KP671558954), T5k, VMUL(LDK(KP740951125), T5l));
3336 T5n = VSUB(T3V, T4u);
3337 T5o = VSUB(T4H, T4E);
3338 T5p = VFNMS(LDK(KP671558954), T5o, VMUL(LDK(KP740951125), T5n));
3339 T5w = VFMA(LDK(KP671558954), T5n, VMUL(LDK(KP740951125), T5o));
3340 }
3341 T5q = VADD(T5m, T5p);
3342 T5F = VSUB(T5w, T5v);
3343 T5x = VADD(T5v, T5w);
3344 T5B = VSUB(T5p, T5m);
3345 }
3346 {
3347 V T11, T5h, T4Y, T5s, T2e, T5t, T51, T5i, T10, T4X;
3348 T10 = VADD(TI, TZ);
3349 T11 = VADD(Tr, T10);
3350 T5h = VSUB(Tr, T10);
3351 T4X = VADD(T4V, T4W);
3352 T4Y = VADD(T4U, T4X);
3353 T5s = VSUB(T4X, T4U);
3354 {
3355 V T1C, T2d, T4Z, T50;
3356 T1C = VFMA(LDK(KP098017140), T1s, VMUL(LDK(KP995184726), T1B));
3357 T2d = VFNMS(LDK(KP098017140), T2c, VMUL(LDK(KP995184726), T23));
3358 T2e = VADD(T1C, T2d);
3359 T5t = VSUB(T2d, T1C);
3360 T4Z = VFNMS(LDK(KP098017140), T1B, VMUL(LDK(KP995184726), T1s));
3361 T50 = VFMA(LDK(KP995184726), T2c, VMUL(LDK(KP098017140), T23));
3362 T51 = VADD(T4Z, T50);
3363 T5i = VSUB(T50, T4Z);
3364 }
3365 T2f = VADD(T11, T2e);
3366 T5C = VSUB(T5t, T5s);
3367 T5E = VSUB(T5h, T5i);
3368 T52 = VADD(T4Y, T51);
3369 T5a = VSUB(T51, T4Y);
3370 T5j = VADD(T5h, T5i);
3371 T5u = VADD(T5s, T5t);
3372 T5c = VSUB(T11, T2e);
3373 }
3374 {
3375 V T4L, T56, T5D, T5G;
3376 T4L = VADD(T2f, T4K);
3377 T56 = VBYI(VADD(T52, T55));
3378 ST(&(xo[WS(os, 127)]), VSUB(T4L, T56), ovs, &(xo[WS(os, 1)]));
3379 ST(&(xo[WS(os, 1)]), VADD(T4L, T56), ovs, &(xo[WS(os, 1)]));
3380 T5D = VBYI(VSUB(T5B, T5C));
3381 T5G = VSUB(T5E, T5F);
3382 ST(&(xo[WS(os, 47)]), VADD(T5D, T5G), ovs, &(xo[WS(os, 1)]));
3383 ST(&(xo[WS(os, 81)]), VSUB(T5G, T5D), ovs, &(xo[WS(os, 1)]));
3384 }
3385 {
3386 V T5H, T5I, T57, T58;
3387 T5H = VBYI(VADD(T5C, T5B));
3388 T5I = VADD(T5E, T5F);
3389 ST(&(xo[WS(os, 17)]), VADD(T5H, T5I), ovs, &(xo[WS(os, 1)]));
3390 ST(&(xo[WS(os, 111)]), VSUB(T5I, T5H), ovs, &(xo[WS(os, 1)]));
3391 T57 = VSUB(T2f, T4K);
3392 T58 = VBYI(VSUB(T55, T52));
3393 ST(&(xo[WS(os, 65)]), VSUB(T57, T58), ovs, &(xo[WS(os, 1)]));
3394 ST(&(xo[WS(os, 63)]), VADD(T57, T58), ovs, &(xo[WS(os, 1)]));
3395 }
3396 {
3397 V T5b, T5e, T5r, T5y;
3398 T5b = VBYI(VSUB(T59, T5a));
3399 T5e = VSUB(T5c, T5d);
3400 ST(&(xo[WS(os, 33)]), VADD(T5b, T5e), ovs, &(xo[WS(os, 1)]));
3401 ST(&(xo[WS(os, 95)]), VSUB(T5e, T5b), ovs, &(xo[WS(os, 1)]));
3402 T5r = VADD(T5j, T5q);
3403 T5y = VBYI(VADD(T5u, T5x));
3404 ST(&(xo[WS(os, 113)]), VSUB(T5r, T5y), ovs, &(xo[WS(os, 1)]));
3405 ST(&(xo[WS(os, 15)]), VADD(T5r, T5y), ovs, &(xo[WS(os, 1)]));
3406 }
3407 {
3408 V T5z, T5A, T5f, T5g;
3409 T5z = VSUB(T5j, T5q);
3410 T5A = VBYI(VSUB(T5x, T5u));
3411 ST(&(xo[WS(os, 79)]), VSUB(T5z, T5A), ovs, &(xo[WS(os, 1)]));
3412 ST(&(xo[WS(os, 49)]), VADD(T5z, T5A), ovs, &(xo[WS(os, 1)]));
3413 T5f = VBYI(VADD(T5a, T59));
3414 T5g = VADD(T5c, T5d);
3415 ST(&(xo[WS(os, 31)]), VADD(T5f, T5g), ovs, &(xo[WS(os, 1)]));
3416 ST(&(xo[WS(os, 97)]), VSUB(T5g, T5f), ovs, &(xo[WS(os, 1)]));
3417 }
3418 }
3419 {
3420 V T9i, T9B, T9t, T9x, T9O, Ta3, T9V, T9Z, T93, Ta0, Ta2, T9q, T9y, T9H, T9S;
3421 V T9A;
3422 {
3423 V T9a, T9r, T9h, T9s;
3424 {
3425 V T96, T99, T9d, T9g;
3426 T96 = VSUB(T94, T95);
3427 T99 = VSUB(T97, T98);
3428 T9a = VFMA(LDK(KP514102744), T96, VMUL(LDK(KP857728610), T99));
3429 T9r = VFNMS(LDK(KP514102744), T99, VMUL(LDK(KP857728610), T96));
3430 T9d = VSUB(T9b, T9c);
3431 T9g = VSUB(T9e, T9f);
3432 T9h = VFNMS(LDK(KP514102744), T9g, VMUL(LDK(KP857728610), T9d));
3433 T9s = VFMA(LDK(KP857728610), T9g, VMUL(LDK(KP514102744), T9d));
3434 }
3435 T9i = VADD(T9a, T9h);
3436 T9B = VSUB(T9s, T9r);
3437 T9t = VADD(T9r, T9s);
3438 T9x = VSUB(T9h, T9a);
3439 }
3440 {
3441 V T9K, T9T, T9N, T9U;
3442 {
3443 V T9I, T9J, T9L, T9M;
3444 T9I = VADD(T95, T94);
3445 T9J = VADD(T97, T98);
3446 T9K = VFMA(LDK(KP242980179), T9I, VMUL(LDK(KP970031253), T9J));
3447 T9T = VFNMS(LDK(KP242980179), T9J, VMUL(LDK(KP970031253), T9I));
3448 T9L = VADD(T9b, T9c);
3449 T9M = VADD(T9f, T9e);
3450 T9N = VFNMS(LDK(KP242980179), T9M, VMUL(LDK(KP970031253), T9L));
3451 T9U = VFMA(LDK(KP970031253), T9M, VMUL(LDK(KP242980179), T9L));
3452 }
3453 T9O = VADD(T9K, T9N);
3454 Ta3 = VSUB(T9U, T9T);
3455 T9V = VADD(T9T, T9U);
3456 T9Z = VSUB(T9N, T9K);
3457 }
3458 {
3459 V T8V, T9F, T9m, T9Q, T92, T9R, T9p, T9G, T8U, T9k;
3460 T8U = VSUB(T86, T85);
3461 T8V = VSUB(T8T, T8U);
3462 T9F = VADD(T8T, T8U);
3463 T9k = VSUB(T75, T72);
3464 T9m = VSUB(T9k, T9l);
3465 T9Q = VADD(T9l, T9k);
3466 {
3467 V T8Y, T91, T9n, T9o;
3468 T8Y = VFMA(LDK(KP471396736), T8W, VMUL(LDK(KP881921264), T8X));
3469 T91 = VFNMS(LDK(KP471396736), T90, VMUL(LDK(KP881921264), T8Z));
3470 T92 = VSUB(T8Y, T91);
3471 T9R = VADD(T91, T8Y);
3472 T9n = VFNMS(LDK(KP471396736), T8X, VMUL(LDK(KP881921264), T8W));
3473 T9o = VFMA(LDK(KP881921264), T90, VMUL(LDK(KP471396736), T8Z));
3474 T9p = VSUB(T9n, T9o);
3475 T9G = VADD(T9o, T9n);
3476 }
3477 T93 = VADD(T8V, T92);
3478 Ta0 = VSUB(T9R, T9Q);
3479 Ta2 = VSUB(T9F, T9G);
3480 T9q = VADD(T9m, T9p);
3481 T9y = VSUB(T9p, T9m);
3482 T9H = VADD(T9F, T9G);
3483 T9S = VADD(T9Q, T9R);
3484 T9A = VSUB(T8V, T92);
3485 }
3486 {
3487 V T9j, T9u, Ta1, Ta4;
3488 T9j = VADD(T93, T9i);
3489 T9u = VBYI(VADD(T9q, T9t));
3490 ST(&(xo[WS(os, 117)]), VSUB(T9j, T9u), ovs, &(xo[WS(os, 1)]));
3491 ST(&(xo[WS(os, 11)]), VADD(T9j, T9u), ovs, &(xo[WS(os, 1)]));
3492 Ta1 = VBYI(VSUB(T9Z, Ta0));
3493 Ta4 = VSUB(Ta2, Ta3);
3494 ST(&(xo[WS(os, 37)]), VADD(Ta1, Ta4), ovs, &(xo[WS(os, 1)]));
3495 ST(&(xo[WS(os, 91)]), VSUB(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
3496 }
3497 {
3498 V Ta5, Ta6, T9v, T9w;
3499 Ta5 = VBYI(VADD(Ta0, T9Z));
3500 Ta6 = VADD(Ta2, Ta3);
3501 ST(&(xo[WS(os, 27)]), VADD(Ta5, Ta6), ovs, &(xo[WS(os, 1)]));
3502 ST(&(xo[WS(os, 101)]), VSUB(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
3503 T9v = VSUB(T93, T9i);
3504 T9w = VBYI(VSUB(T9t, T9q));
3505 ST(&(xo[WS(os, 75)]), VSUB(T9v, T9w), ovs, &(xo[WS(os, 1)]));
3506 ST(&(xo[WS(os, 53)]), VADD(T9v, T9w), ovs, &(xo[WS(os, 1)]));
3507 }
3508 {
3509 V T9z, T9C, T9P, T9W;
3510 T9z = VBYI(VSUB(T9x, T9y));
3511 T9C = VSUB(T9A, T9B);
3512 ST(&(xo[WS(os, 43)]), VADD(T9z, T9C), ovs, &(xo[WS(os, 1)]));
3513 ST(&(xo[WS(os, 85)]), VSUB(T9C, T9z), ovs, &(xo[WS(os, 1)]));
3514 T9P = VADD(T9H, T9O);
3515 T9W = VBYI(VADD(T9S, T9V));
3516 ST(&(xo[WS(os, 123)]), VSUB(T9P, T9W), ovs, &(xo[WS(os, 1)]));
3517 ST(&(xo[WS(os, 5)]), VADD(T9P, T9W), ovs, &(xo[WS(os, 1)]));
3518 }
3519 {
3520 V T9X, T9Y, T9D, T9E;
3521 T9X = VSUB(T9H, T9O);
3522 T9Y = VBYI(VSUB(T9V, T9S));
3523 ST(&(xo[WS(os, 69)]), VSUB(T9X, T9Y), ovs, &(xo[WS(os, 1)]));
3524 ST(&(xo[WS(os, 59)]), VADD(T9X, T9Y), ovs, &(xo[WS(os, 1)]));
3525 T9D = VBYI(VADD(T9y, T9x));
3526 T9E = VADD(T9A, T9B);
3527 ST(&(xo[WS(os, 21)]), VADD(T9D, T9E), ovs, &(xo[WS(os, 1)]));
3528 ST(&(xo[WS(os, 107)]), VSUB(T9E, T9D), ovs, &(xo[WS(os, 1)]));
3529 }
3530 }
3531 {
3532 V T68, T6r, T6j, T6n, T6E, T6T, T6L, T6P, T5T, T6Q, T6S, T6g, T6o, T6x, T6I;
3533 V T6q;
3534 {
3535 V T60, T6h, T67, T6i;
3536 {
3537 V T5W, T5Z, T63, T66;
3538 T5W = VSUB(T5U, T5V);
3539 T5Z = VSUB(T5X, T5Y);
3540 T60 = VFMA(LDK(KP427555093), T5W, VMUL(LDK(KP903989293), T5Z));
3541 T6h = VFNMS(LDK(KP427555093), T5Z, VMUL(LDK(KP903989293), T5W));
3542 T63 = VSUB(T61, T62);
3543 T66 = VSUB(T64, T65);
3544 T67 = VFNMS(LDK(KP427555093), T66, VMUL(LDK(KP903989293), T63));
3545 T6i = VFMA(LDK(KP903989293), T66, VMUL(LDK(KP427555093), T63));
3546 }
3547 T68 = VADD(T60, T67);
3548 T6r = VSUB(T6i, T6h);
3549 T6j = VADD(T6h, T6i);
3550 T6n = VSUB(T67, T60);
3551 }
3552 {
3553 V T6A, T6J, T6D, T6K;
3554 {
3555 V T6y, T6z, T6B, T6C;
3556 T6y = VADD(T5V, T5U);
3557 T6z = VADD(T5X, T5Y);
3558 T6A = VFMA(LDK(KP336889853), T6y, VMUL(LDK(KP941544065), T6z));
3559 T6J = VFNMS(LDK(KP336889853), T6z, VMUL(LDK(KP941544065), T6y));
3560 T6B = VADD(T61, T62);
3561 T6C = VADD(T65, T64);
3562 T6D = VFNMS(LDK(KP336889853), T6C, VMUL(LDK(KP941544065), T6B));
3563 T6K = VFMA(LDK(KP941544065), T6C, VMUL(LDK(KP336889853), T6B));
3564 }
3565 T6E = VADD(T6A, T6D);
3566 T6T = VSUB(T6K, T6J);
3567 T6L = VADD(T6J, T6K);
3568 T6P = VSUB(T6D, T6A);
3569 }
3570 {
3571 V T5L, T6v, T6c, T6G, T5S, T6H, T6f, T6w, T5K, T6a;
3572 T5K = VSUB(T4W, T4V);
3573 T5L = VSUB(T5J, T5K);
3574 T6v = VADD(T5J, T5K);
3575 T6a = VSUB(TZ, TI);
3576 T6c = VSUB(T6a, T6b);
3577 T6G = VADD(T6b, T6a);
3578 {
3579 V T5O, T5R, T6d, T6e;
3580 T5O = VFMA(LDK(KP773010453), T5M, VMUL(LDK(KP634393284), T5N));
3581 T5R = VFNMS(LDK(KP634393284), T5Q, VMUL(LDK(KP773010453), T5P));
3582 T5S = VSUB(T5O, T5R);
3583 T6H = VADD(T5R, T5O);
3584 T6d = VFNMS(LDK(KP634393284), T5M, VMUL(LDK(KP773010453), T5N));
3585 T6e = VFMA(LDK(KP634393284), T5P, VMUL(LDK(KP773010453), T5Q));
3586 T6f = VSUB(T6d, T6e);
3587 T6w = VADD(T6e, T6d);
3588 }
3589 T5T = VADD(T5L, T5S);
3590 T6Q = VSUB(T6H, T6G);
3591 T6S = VSUB(T6v, T6w);
3592 T6g = VADD(T6c, T6f);
3593 T6o = VSUB(T6f, T6c);
3594 T6x = VADD(T6v, T6w);
3595 T6I = VADD(T6G, T6H);
3596 T6q = VSUB(T5L, T5S);
3597 }
3598 {
3599 V T69, T6k, T6R, T6U;
3600 T69 = VADD(T5T, T68);
3601 T6k = VBYI(VADD(T6g, T6j));
3602 ST(&(xo[WS(os, 119)]), VSUB(T69, T6k), ovs, &(xo[WS(os, 1)]));
3603 ST(&(xo[WS(os, 9)]), VADD(T69, T6k), ovs, &(xo[WS(os, 1)]));
3604 T6R = VBYI(VSUB(T6P, T6Q));
3605 T6U = VSUB(T6S, T6T);
3606 ST(&(xo[WS(os, 39)]), VADD(T6R, T6U), ovs, &(xo[WS(os, 1)]));
3607 ST(&(xo[WS(os, 89)]), VSUB(T6U, T6R), ovs, &(xo[WS(os, 1)]));
3608 }
3609 {
3610 V T6V, T6W, T6l, T6m;
3611 T6V = VBYI(VADD(T6Q, T6P));
3612 T6W = VADD(T6S, T6T);
3613 ST(&(xo[WS(os, 25)]), VADD(T6V, T6W), ovs, &(xo[WS(os, 1)]));
3614 ST(&(xo[WS(os, 103)]), VSUB(T6W, T6V), ovs, &(xo[WS(os, 1)]));
3615 T6l = VSUB(T5T, T68);
3616 T6m = VBYI(VSUB(T6j, T6g));
3617 ST(&(xo[WS(os, 73)]), VSUB(T6l, T6m), ovs, &(xo[WS(os, 1)]));
3618 ST(&(xo[WS(os, 55)]), VADD(T6l, T6m), ovs, &(xo[WS(os, 1)]));
3619 }
3620 {
3621 V T6p, T6s, T6F, T6M;
3622 T6p = VBYI(VSUB(T6n, T6o));
3623 T6s = VSUB(T6q, T6r);
3624 ST(&(xo[WS(os, 41)]), VADD(T6p, T6s), ovs, &(xo[WS(os, 1)]));
3625 ST(&(xo[WS(os, 87)]), VSUB(T6s, T6p), ovs, &(xo[WS(os, 1)]));
3626 T6F = VADD(T6x, T6E);
3627 T6M = VBYI(VADD(T6I, T6L));
3628 ST(&(xo[WS(os, 121)]), VSUB(T6F, T6M), ovs, &(xo[WS(os, 1)]));
3629 ST(&(xo[WS(os, 7)]), VADD(T6F, T6M), ovs, &(xo[WS(os, 1)]));
3630 }
3631 {
3632 V T6N, T6O, T6t, T6u;
3633 T6N = VSUB(T6x, T6E);
3634 T6O = VBYI(VSUB(T6L, T6I));
3635 ST(&(xo[WS(os, 71)]), VSUB(T6N, T6O), ovs, &(xo[WS(os, 1)]));
3636 ST(&(xo[WS(os, 57)]), VADD(T6N, T6O), ovs, &(xo[WS(os, 1)]));
3637 T6t = VBYI(VADD(T6o, T6n));
3638 T6u = VADD(T6q, T6r);
3639 ST(&(xo[WS(os, 23)]), VADD(T6t, T6u), ovs, &(xo[WS(os, 1)]));
3640 ST(&(xo[WS(os, 105)]), VSUB(T6u, T6t), ovs, &(xo[WS(os, 1)]));
3641 }
3642 }
3643 }
3644 }
3645 VLEAVE();
3646 }
3647
3648 static const kdft_desc desc = { 128, XSIMD_STRING("n1fv_128"), {938, 186, 144, 0}, &GENUS, 0, 0, 0, 0 };
3649
3650 void XSIMD(codelet_n1fv_128) (planner *p) {
3651 X(kdft_register) (p, n1fv_128, &desc);
3652 }
3653
3654 #endif