comparison src/fftw-3.3.3/dft/simd/common/n1fv_128.c @ 10:37bf6b4a2645

Add FFTW3
author Chris Cannam
date Wed, 20 Mar 2013 15:35:50 +0000
parents
children
comparison
equal deleted inserted replaced
9:c0fb53affa76 10:37bf6b4a2645
1 /*
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 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 Sun Nov 25 07:36:54 EST 2012 */
23
24 #include "codelet-dft.h"
25
26 #ifdef HAVE_FMA
27
28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 128 -name n1fv_128 -include 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 * 295 stack variables, 31 constants, and 256 memory accesses
34 */
35 #include "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(KP903989293, +0.903989293123443331586200297230537048710132025);
40 DVK(KP941544065, +0.941544065183020778412509402599502357185589796);
41 DVK(KP357805721, +0.357805721314524104672487743774474392487532769);
42 DVK(KP472964775, +0.472964775891319928124438237972992463904131113);
43 DVK(KP857728610, +0.857728610000272069902269984284770137042490799);
44 DVK(KP970031253, +0.970031253194543992603984207286100251456865962);
45 DVK(KP250486960, +0.250486960191305461595702160124721208578685568);
46 DVK(KP998795456, +0.998795456205172392714771604759100694443203615);
47 DVK(KP740951125, +0.740951125354959091175616897495162729728955309);
48 DVK(KP599376933, +0.599376933681923766271389869014404232837890546);
49 DVK(KP906347169, +0.906347169019147157946142717268914412664134293);
50 DVK(KP049126849, +0.049126849769467254105343321271313617079695752);
51 DVK(KP989176509, +0.989176509964780973451673738016243063983689533);
52 DVK(KP803207531, +0.803207531480644909806676512963141923879569427);
53 DVK(KP741650546, +0.741650546272035369581266691172079863842265220);
54 DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
55 DVK(KP148335987, +0.148335987538347428753676511486911367000625355);
56 DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
57 DVK(KP303346683, +0.303346683607342391675883946941299872384187453);
58 DVK(KP534511135, +0.534511135950791641089685961295362908582039528);
59 DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
60 DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
61 DVK(KP820678790, +0.820678790828660330972281985331011598767386482);
62 DVK(KP098491403, +0.098491403357164253077197521291327432293052451);
63 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
64 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
65 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
66 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
67 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
68 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
69 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
70 {
71 INT i;
72 const R *xi;
73 R *xo;
74 xi = 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 T6a, T5J, T6b, T5K, T6B, T6C, T6J, T6A, T6o, T6j, T6r, T68, T6e, T5O, T5R;
78 V T6d, T6D, T6K;
79 {
80 V Tad, TcZ, T6Z, T8T, T4U, Tr, Tfq, TgG, Ted, Tgf, Td0, Tcc, T9k, T84, Tb6;
81 V Tbt, Td8, TdK, TeK, Tgq, TeV, Tgt, T7q, T94, T3p, T5X, T7B, T97, T2G, T5U;
82 V TbD, Tc0, Tdf, TdN, Tf5, Tgx, Tfg, TgA, T7J, T9b, T4E, T64, T7U, T9e, T3V;
83 V T61, Td2, Td3, T85, T72, T4V, TI, Tcd, Tas, TgH, Tek, Tgg, Tft, T86, T75;
84 V T4W, TZ, TaI, Tcg, Tdr, TdG, Tgi, Tet, Tgj, Teq, T8X, T7a, T5M, T1B, T8W;
85 V T7d, T5N, T1s, TaX, Tcf, Tdo, TdH, Tgl, TeC, Tgm, Tez, T90, T7h, T5P, T2c;
86 V T8Z, T7k, T5Q, T23, T3Y, T49, TdL, Tdb, Tbu, Tbl, Tgu, TeR, Tgr, TeY, Tf6;
87 V TbG, T5V, T3s, T5Y, T3f, T95, T7E, T98, T7x, T4g, T4f, T4q, TbH, T41, TbI;
88 V T44, T4h, T4j, T4k, Tf9, TbN;
89 {
90 V Tu, TF, Ty, TL, TW, Tah, Tx, Tag, Tee, Tz, TM, TN, Teh, Tan, TP;
91 V TQ;
92 {
93 V TeG, T2A, Tbq, TeT, Tbp, TeH, T3m, T2x, Td6, T7o, T2q, T3l, T7z, Tbr, T2D;
94 V T82, T83;
95 {
96 V Ta7, T3, Ta8, T4O, Taa, Tab, Ta, T4P, Te, Tc9, Th, Tca, Tl, Tc6, Tc7;
97 V To;
98 {
99 V T1, T2, T4M, T4N;
100 T1 = LD(&(xi[0]), ivs, &(xi[0]));
101 T2 = LD(&(xi[WS(is, 64)]), ivs, &(xi[0]));
102 T4M = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
103 T4N = LD(&(xi[WS(is, 96)]), ivs, &(xi[0]));
104 {
105 V T4, T5, T7, T8;
106 T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
107 T5 = LD(&(xi[WS(is, 80)]), ivs, &(xi[0]));
108 T7 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0]));
109 T8 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
110 {
111 V Tc, T6, T9, Td, Tf, Tg;
112 Tc = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
113 Ta7 = VADD(T1, T2);
114 T3 = VSUB(T1, T2);
115 Ta8 = VADD(T4M, T4N);
116 T4O = VSUB(T4M, T4N);
117 Taa = VADD(T4, T5);
118 T6 = VSUB(T4, T5);
119 Tab = VADD(T7, T8);
120 T9 = VSUB(T7, T8);
121 Td = LD(&(xi[WS(is, 72)]), ivs, &(xi[0]));
122 Tf = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
123 Tg = LD(&(xi[WS(is, 104)]), ivs, &(xi[0]));
124 {
125 V Tj, Tk, Tm, Tn;
126 Tj = LD(&(xi[WS(is, 120)]), ivs, &(xi[0]));
127 Tk = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
128 Tm = LD(&(xi[WS(is, 88)]), ivs, &(xi[0]));
129 Tn = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
130 Ta = VADD(T6, T9);
131 T4P = VSUB(T9, T6);
132 Te = VSUB(Tc, Td);
133 Tc9 = VADD(Tc, Td);
134 Th = VSUB(Tf, Tg);
135 Tca = VADD(Tf, Tg);
136 Tl = VSUB(Tj, Tk);
137 Tc6 = VADD(Tj, Tk);
138 Tc7 = VADD(Tn, Tm);
139 To = VSUB(Tm, Tn);
140 }
141 }
142 }
143 }
144 {
145 V T6X, Tb, Te9, Ta9, Tcb, Tea, T4R, Ti, Tfo, Tac, Tp, T4S, Tc8, Teb, T4Q;
146 T6X = VFNMS(LDK(KP707106781), Ta, T3);
147 Tb = VFMA(LDK(KP707106781), Ta, T3);
148 Te9 = VSUB(Ta7, Ta8);
149 Ta9 = VADD(Ta7, Ta8);
150 Tcb = VADD(Tc9, Tca);
151 Tea = VSUB(Tc9, Tca);
152 T4R = VFMA(LDK(KP414213562), Te, Th);
153 Ti = VFNMS(LDK(KP414213562), Th, Te);
154 Tfo = VSUB(Tab, Taa);
155 Tac = VADD(Taa, Tab);
156 Tp = VFNMS(LDK(KP414213562), To, Tl);
157 T4S = VFMA(LDK(KP414213562), Tl, To);
158 Tc8 = VADD(Tc6, Tc7);
159 Teb = VSUB(Tc6, Tc7);
160 T4Q = VFNMS(LDK(KP707106781), T4P, T4O);
161 T82 = VFMA(LDK(KP707106781), T4P, T4O);
162 {
163 V T4T, T6Y, Tq, Tfp, Tec;
164 T4T = VSUB(T4R, T4S);
165 T6Y = VADD(T4R, T4S);
166 T83 = VSUB(Tp, Ti);
167 Tq = VADD(Ti, Tp);
168 Tfp = VSUB(Teb, Tea);
169 Tec = VADD(Tea, Teb);
170 Tad = VSUB(Ta9, Tac);
171 TcZ = VADD(Ta9, Tac);
172 T6Z = VFMA(LDK(KP923879532), T6Y, T6X);
173 T8T = VFNMS(LDK(KP923879532), T6Y, T6X);
174 T4U = VFMA(LDK(KP923879532), T4T, T4Q);
175 T6a = VFNMS(LDK(KP923879532), T4T, T4Q);
176 Tr = VFMA(LDK(KP923879532), Tq, Tb);
177 T5J = VFNMS(LDK(KP923879532), Tq, Tb);
178 Tfq = VFMA(LDK(KP707106781), Tfp, Tfo);
179 TgG = VFNMS(LDK(KP707106781), Tfp, Tfo);
180 Ted = VFMA(LDK(KP707106781), Tec, Te9);
181 Tgf = VFNMS(LDK(KP707106781), Tec, Te9);
182 Td0 = VADD(Tcb, Tc8);
183 Tcc = VSUB(Tc8, Tcb);
184 }
185 }
186 }
187 {
188 V T2i, T3j, Tb2, T2B, Tb5, T3k, T2p, T2C;
189 {
190 V T2m, Tb0, Tb1, Tb3, T2l, T2n;
191 {
192 V T2g, T2h, T3h, T3i, T2j, T2k;
193 T2g = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
194 T2h = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)]));
195 T3h = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
196 T3i = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)]));
197 T2j = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
198 T2k = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)]));
199 T2m = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)]));
200 T9k = VFNMS(LDK(KP923879532), T83, T82);
201 T84 = VFMA(LDK(KP923879532), T83, T82);
202 T2i = VSUB(T2g, T2h);
203 Tb0 = VADD(T2g, T2h);
204 T3j = VSUB(T3h, T3i);
205 Tb1 = VADD(T3h, T3i);
206 Tb3 = VADD(T2j, T2k);
207 T2l = VSUB(T2j, T2k);
208 T2n = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
209 }
210 {
211 V T2r, T2s, T2u, T2v;
212 T2r = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
213 T2s = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)]));
214 T2u = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
215 T2v = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)]));
216 TeG = VSUB(Tb0, Tb1);
217 Tb2 = VADD(Tb0, Tb1);
218 {
219 V T2y, T2z, Tb4, T2o, Tbn, T2t, Tbo, T2w;
220 T2y = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)]));
221 T2z = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
222 Tb4 = VADD(T2m, T2n);
223 T2o = VSUB(T2m, T2n);
224 Tbn = VADD(T2r, T2s);
225 T2t = VSUB(T2r, T2s);
226 Tbo = VADD(T2u, T2v);
227 T2w = VSUB(T2u, T2v);
228 T2B = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
229 T2A = VSUB(T2y, T2z);
230 Tbq = VADD(T2y, T2z);
231 TeT = VSUB(Tb3, Tb4);
232 Tb5 = VADD(Tb3, Tb4);
233 T3k = VSUB(T2l, T2o);
234 T2p = VADD(T2l, T2o);
235 Tbp = VADD(Tbn, Tbo);
236 TeH = VSUB(Tbn, Tbo);
237 T3m = VFMA(LDK(KP414213562), T2t, T2w);
238 T2x = VFNMS(LDK(KP414213562), T2w, T2t);
239 T2C = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)]));
240 }
241 }
242 }
243 Td6 = VADD(Tb2, Tb5);
244 Tb6 = VSUB(Tb2, Tb5);
245 T7o = VFNMS(LDK(KP707106781), T2p, T2i);
246 T2q = VFMA(LDK(KP707106781), T2p, T2i);
247 T3l = VFMA(LDK(KP707106781), T3k, T3j);
248 T7z = VFNMS(LDK(KP707106781), T3k, T3j);
249 Tbr = VADD(T2B, T2C);
250 T2D = VSUB(T2B, T2C);
251 }
252 {
253 V Tf1, Tfe, Tf2, TbZ, T3M, T4B, Tdd, T3F, T7H, T4A, T7S, TbW, Tf3, T4C, T3T;
254 {
255 V T3x, T4y, Tbz, T3Q, TbC, T4z, T3E, T3R, T3P, TbU, TbV, T3S;
256 {
257 V T3y, T3z, T3B, T3C;
258 {
259 V T3v, T3w, T4w, T4x;
260 T3v = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)]));
261 T3w = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
262 T4w = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)]));
263 T4x = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
264 T3y = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
265 {
266 V Tbs, TeI, T3n, T2E, Tbx;
267 Tbs = VADD(Tbq, Tbr);
268 TeI = VSUB(Tbq, Tbr);
269 T3n = VFNMS(LDK(KP414213562), T2A, T2D);
270 T2E = VFMA(LDK(KP414213562), T2D, T2A);
271 T3x = VSUB(T3v, T3w);
272 Tbx = VADD(T3v, T3w);
273 {
274 V Tby, Td7, TeJ, TeU;
275 T4y = VSUB(T4w, T4x);
276 Tby = VADD(T4x, T4w);
277 Td7 = VADD(Tbp, Tbs);
278 Tbt = VSUB(Tbp, Tbs);
279 TeJ = VADD(TeH, TeI);
280 TeU = VSUB(TeH, TeI);
281 {
282 V T7p, T3o, T7A, T2F;
283 T7p = VSUB(T3m, T3n);
284 T3o = VADD(T3m, T3n);
285 T7A = VSUB(T2x, T2E);
286 T2F = VADD(T2x, T2E);
287 Tbz = VADD(Tbx, Tby);
288 Tf1 = VSUB(Tbx, Tby);
289 Td8 = VADD(Td6, Td7);
290 TdK = VSUB(Td6, Td7);
291 TeK = VFMA(LDK(KP707106781), TeJ, TeG);
292 Tgq = VFNMS(LDK(KP707106781), TeJ, TeG);
293 TeV = VFMA(LDK(KP707106781), TeU, TeT);
294 Tgt = VFNMS(LDK(KP707106781), TeU, TeT);
295 T7q = VFMA(LDK(KP923879532), T7p, T7o);
296 T94 = VFNMS(LDK(KP923879532), T7p, T7o);
297 T3p = VFMA(LDK(KP923879532), T3o, T3l);
298 T5X = VFNMS(LDK(KP923879532), T3o, T3l);
299 T7B = VFNMS(LDK(KP923879532), T7A, T7z);
300 T97 = VFMA(LDK(KP923879532), T7A, T7z);
301 T2G = VFMA(LDK(KP923879532), T2F, T2q);
302 T5U = VFNMS(LDK(KP923879532), T2F, T2q);
303 T3z = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)]));
304 }
305 }
306 }
307 T3B = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)]));
308 T3C = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
309 }
310 {
311 V T3G, T3H, T3J, T3K;
312 T3G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
313 T3H = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)]));
314 T3J = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)]));
315 T3K = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
316 {
317 V T3N, T3A, TbA, T3D, TbB, T3I, TbX, T3L, TbY, T3O;
318 T3N = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)]));
319 T3A = VSUB(T3y, T3z);
320 TbA = VADD(T3y, T3z);
321 T3D = VSUB(T3B, T3C);
322 TbB = VADD(T3B, T3C);
323 T3I = VSUB(T3G, T3H);
324 TbX = VADD(T3G, T3H);
325 T3L = VSUB(T3J, T3K);
326 TbY = VADD(T3K, T3J);
327 T3O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
328 T3Q = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)]));
329 Tfe = VSUB(TbB, TbA);
330 TbC = VADD(TbA, TbB);
331 T4z = VSUB(T3D, T3A);
332 T3E = VADD(T3A, T3D);
333 T3R = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
334 Tf2 = VSUB(TbX, TbY);
335 TbZ = VADD(TbX, TbY);
336 T3M = VFMA(LDK(KP414213562), T3L, T3I);
337 T4B = VFNMS(LDK(KP414213562), T3I, T3L);
338 T3P = VSUB(T3N, T3O);
339 TbU = VADD(T3N, T3O);
340 }
341 }
342 }
343 Tdd = VADD(Tbz, TbC);
344 TbD = VSUB(Tbz, TbC);
345 TbV = VADD(T3R, T3Q);
346 T3S = VSUB(T3Q, T3R);
347 T3F = VFMA(LDK(KP707106781), T3E, T3x);
348 T7H = VFNMS(LDK(KP707106781), T3E, T3x);
349 T4A = VFMA(LDK(KP707106781), T4z, T4y);
350 T7S = VFNMS(LDK(KP707106781), T4z, T4y);
351 TbW = VADD(TbU, TbV);
352 Tf3 = VSUB(TbU, TbV);
353 T4C = VFMA(LDK(KP414213562), T3P, T3S);
354 T3T = VFNMS(LDK(KP414213562), T3S, T3P);
355 }
356 {
357 V TD, Tae, TE, TJ, TK, TU, TV;
358 {
359 V Ts, Tt, Tde, Tf4, Tff;
360 Ts = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
361 Tt = LD(&(xi[WS(is, 68)]), ivs, &(xi[0]));
362 TD = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
363 Tde = VADD(TbZ, TbW);
364 Tc0 = VSUB(TbW, TbZ);
365 Tf4 = VADD(Tf2, Tf3);
366 Tff = VSUB(Tf3, Tf2);
367 {
368 V T7I, T4D, T7T, T3U;
369 T7I = VSUB(T4C, T4B);
370 T4D = VADD(T4B, T4C);
371 T7T = VSUB(T3T, T3M);
372 T3U = VADD(T3M, T3T);
373 Tae = VADD(Ts, Tt);
374 Tu = VSUB(Ts, Tt);
375 Tdf = VADD(Tdd, Tde);
376 TdN = VSUB(Tdd, Tde);
377 Tf5 = VFMA(LDK(KP707106781), Tf4, Tf1);
378 Tgx = VFNMS(LDK(KP707106781), Tf4, Tf1);
379 Tfg = VFMA(LDK(KP707106781), Tff, Tfe);
380 TgA = VFNMS(LDK(KP707106781), Tff, Tfe);
381 T7J = VFMA(LDK(KP923879532), T7I, T7H);
382 T9b = VFNMS(LDK(KP923879532), T7I, T7H);
383 T4E = VFMA(LDK(KP923879532), T4D, T4A);
384 T64 = VFNMS(LDK(KP923879532), T4D, T4A);
385 T7U = VFNMS(LDK(KP923879532), T7T, T7S);
386 T9e = VFMA(LDK(KP923879532), T7T, T7S);
387 T3V = VFMA(LDK(KP923879532), T3U, T3F);
388 T61 = VFNMS(LDK(KP923879532), T3U, T3F);
389 TE = LD(&(xi[WS(is, 100)]), ivs, &(xi[0]));
390 }
391 }
392 TJ = LD(&(xi[WS(is, 124)]), ivs, &(xi[0]));
393 TK = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
394 TU = LD(&(xi[WS(is, 92)]), ivs, &(xi[0]));
395 TV = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
396 {
397 V Tal, Tam, Tv, Tw, Taf;
398 Tv = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
399 Tw = LD(&(xi[WS(is, 84)]), ivs, &(xi[0]));
400 Taf = VADD(TD, TE);
401 TF = VSUB(TD, TE);
402 Ty = LD(&(xi[WS(is, 116)]), ivs, &(xi[0]));
403 TL = VSUB(TJ, TK);
404 Tal = VADD(TJ, TK);
405 TW = VSUB(TU, TV);
406 Tam = VADD(TV, TU);
407 Tah = VADD(Tv, Tw);
408 Tx = VSUB(Tv, Tw);
409 Tag = VADD(Tae, Taf);
410 Tee = VSUB(Tae, Taf);
411 Tz = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
412 TM = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
413 TN = LD(&(xi[WS(is, 76)]), ivs, &(xi[0]));
414 Teh = VSUB(Tal, Tam);
415 Tan = VADD(Tal, Tam);
416 TP = LD(&(xi[WS(is, 108)]), ivs, &(xi[0]));
417 TQ = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
418 }
419 }
420 }
421 }
422 {
423 V Tev, TeA, Tdm, TaP, Tew, TaV, T1U, T29, T7f, T1N, T28, T7i, Tex, TaS, T21;
424 V T2a;
425 {
426 V Tem, Ter, Ten, TaD, T1j, T1y, TaA, Tdp, T1c, T78, T7b, T1x, TaG, Teo, T1z;
427 V T1q;
428 {
429 V T14, T1v, Taw, Taz, T1b, T1w, T1n, T1o, T1m, TaE, TaF, T1p;
430 {
431 V Tau, Tav, T15, T16, T18, T19;
432 {
433 V T12, Tai, TA, Tao, TO, T13;
434 T12 = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
435 Tai = VADD(Ty, Tz);
436 TA = VSUB(Ty, Tz);
437 Tao = VADD(TM, TN);
438 TO = VSUB(TM, TN);
439 T13 = LD(&(xi[WS(is, 66)]), ivs, &(xi[0]));
440 {
441 V T1t, Tap, TR, Taj, Tef, TG, TB, T1u;
442 T1t = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
443 Tap = VADD(TP, TQ);
444 TR = VSUB(TP, TQ);
445 Taj = VADD(Tah, Tai);
446 Tef = VSUB(Tah, Tai);
447 TG = VSUB(Tx, TA);
448 TB = VADD(Tx, TA);
449 Tau = VADD(T12, T13);
450 T14 = VSUB(T12, T13);
451 T1u = LD(&(xi[WS(is, 98)]), ivs, &(xi[0]));
452 {
453 V Taq, Tei, TX, TS, Tak;
454 Taq = VADD(Tao, Tap);
455 Tei = VSUB(Tap, Tao);
456 TX = VSUB(TR, TO);
457 TS = VADD(TO, TR);
458 Tak = VSUB(Tag, Taj);
459 Td2 = VADD(Tag, Taj);
460 {
461 V Teg, Tfs, T71, TH;
462 Teg = VFNMS(LDK(KP414213562), Tef, Tee);
463 Tfs = VFMA(LDK(KP414213562), Tee, Tef);
464 T71 = VFNMS(LDK(KP707106781), TG, TF);
465 TH = VFMA(LDK(KP707106781), TG, TF);
466 {
467 V T70, TC, Tar, Tej, Tfr;
468 T70 = VFNMS(LDK(KP707106781), TB, Tu);
469 TC = VFMA(LDK(KP707106781), TB, Tu);
470 Tar = VSUB(Tan, Taq);
471 Td3 = VADD(Tan, Taq);
472 Tej = VFNMS(LDK(KP414213562), Tei, Teh);
473 Tfr = VFMA(LDK(KP414213562), Teh, Tei);
474 {
475 V T74, TY, T73, TT;
476 T74 = VFNMS(LDK(KP707106781), TX, TW);
477 TY = VFMA(LDK(KP707106781), TX, TW);
478 T73 = VFNMS(LDK(KP707106781), TS, TL);
479 TT = VFMA(LDK(KP707106781), TS, TL);
480 T85 = VFNMS(LDK(KP668178637), T70, T71);
481 T72 = VFMA(LDK(KP668178637), T71, T70);
482 T4V = VFMA(LDK(KP198912367), TC, TH);
483 TI = VFNMS(LDK(KP198912367), TH, TC);
484 Tcd = VSUB(Tar, Tak);
485 Tas = VADD(Tak, Tar);
486 TgH = VSUB(Tej, Teg);
487 Tek = VADD(Teg, Tej);
488 Tgg = VADD(Tfs, Tfr);
489 Tft = VSUB(Tfr, Tfs);
490 T86 = VFNMS(LDK(KP668178637), T73, T74);
491 T75 = VFMA(LDK(KP668178637), T74, T73);
492 T4W = VFMA(LDK(KP198912367), TT, TY);
493 TZ = VFNMS(LDK(KP198912367), TY, TT);
494 Tav = VADD(T1t, T1u);
495 T1v = VSUB(T1t, T1u);
496 }
497 }
498 }
499 }
500 }
501 }
502 T15 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
503 T16 = LD(&(xi[WS(is, 82)]), ivs, &(xi[0]));
504 T18 = LD(&(xi[WS(is, 114)]), ivs, &(xi[0]));
505 T19 = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
506 {
507 V T1d, T1e, T1g, T1h, Tax, T17, Tay, T1a;
508 T1d = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
509 Taw = VADD(Tau, Tav);
510 Tem = VSUB(Tau, Tav);
511 T1e = LD(&(xi[WS(is, 74)]), ivs, &(xi[0]));
512 T1g = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
513 T1h = LD(&(xi[WS(is, 106)]), ivs, &(xi[0]));
514 Tax = VADD(T15, T16);
515 T17 = VSUB(T15, T16);
516 Tay = VADD(T18, T19);
517 T1a = VSUB(T18, T19);
518 {
519 V T1k, T1f, TaB, T1i, TaC, T1l;
520 T1k = LD(&(xi[WS(is, 122)]), ivs, &(xi[0]));
521 T1f = VSUB(T1d, T1e);
522 TaB = VADD(T1d, T1e);
523 T1i = VSUB(T1g, T1h);
524 TaC = VADD(T1g, T1h);
525 T1l = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
526 Taz = VADD(Tax, Tay);
527 Ter = VSUB(Tax, Tay);
528 T1b = VADD(T17, T1a);
529 T1w = VSUB(T17, T1a);
530 T1n = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
531 T1o = LD(&(xi[WS(is, 90)]), ivs, &(xi[0]));
532 Ten = VSUB(TaB, TaC);
533 TaD = VADD(TaB, TaC);
534 T1j = VFNMS(LDK(KP414213562), T1i, T1f);
535 T1y = VFMA(LDK(KP414213562), T1f, T1i);
536 T1m = VSUB(T1k, T1l);
537 TaE = VADD(T1k, T1l);
538 }
539 }
540 }
541 TaA = VSUB(Taw, Taz);
542 Tdp = VADD(Taw, Taz);
543 TaF = VADD(T1n, T1o);
544 T1p = VSUB(T1n, T1o);
545 T1c = VFMA(LDK(KP707106781), T1b, T14);
546 T78 = VFNMS(LDK(KP707106781), T1b, T14);
547 T7b = VFNMS(LDK(KP707106781), T1w, T1v);
548 T1x = VFMA(LDK(KP707106781), T1w, T1v);
549 TaG = VADD(TaE, TaF);
550 Teo = VSUB(TaE, TaF);
551 T1z = VFNMS(LDK(KP414213562), T1m, T1p);
552 T1q = VFMA(LDK(KP414213562), T1p, T1m);
553 }
554 {
555 V T1F, T26, T1Q, TaT, TaL, TaO, T27, T1M, T1Y, T1Z, TaU, T1T, TaQ, T1X, T20;
556 V TaR;
557 {
558 V T24, TaJ, T25, T1G, T1H, T1J, T1K, T1D, T1E;
559 T1D = LD(&(xi[WS(is, 126)]), ivs, &(xi[0]));
560 T1E = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
561 T24 = LD(&(xi[WS(is, 94)]), ivs, &(xi[0]));
562 {
563 V TaH, Tdq, Tes, Tep;
564 TaH = VSUB(TaD, TaG);
565 Tdq = VADD(TaD, TaG);
566 Tes = VSUB(Ten, Teo);
567 Tep = VADD(Ten, Teo);
568 {
569 V T79, T1A, T7c, T1r;
570 T79 = VSUB(T1y, T1z);
571 T1A = VADD(T1y, T1z);
572 T7c = VSUB(T1j, T1q);
573 T1r = VADD(T1j, T1q);
574 TaJ = VADD(T1D, T1E);
575 T1F = VSUB(T1D, T1E);
576 TaI = VFNMS(LDK(KP414213562), TaH, TaA);
577 Tcg = VFMA(LDK(KP414213562), TaA, TaH);
578 Tdr = VADD(Tdp, Tdq);
579 TdG = VSUB(Tdp, Tdq);
580 Tgi = VFNMS(LDK(KP707106781), Tes, Ter);
581 Tet = VFMA(LDK(KP707106781), Tes, Ter);
582 Tgj = VFNMS(LDK(KP707106781), Tep, Tem);
583 Teq = VFMA(LDK(KP707106781), Tep, Tem);
584 T8X = VFNMS(LDK(KP923879532), T79, T78);
585 T7a = VFMA(LDK(KP923879532), T79, T78);
586 T5M = VFNMS(LDK(KP923879532), T1A, T1x);
587 T1B = VFMA(LDK(KP923879532), T1A, T1x);
588 T8W = VFMA(LDK(KP923879532), T7c, T7b);
589 T7d = VFNMS(LDK(KP923879532), T7c, T7b);
590 T5N = VFNMS(LDK(KP923879532), T1r, T1c);
591 T1s = VFMA(LDK(KP923879532), T1r, T1c);
592 T25 = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
593 }
594 }
595 T1G = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
596 T1H = LD(&(xi[WS(is, 78)]), ivs, &(xi[0]));
597 T1J = LD(&(xi[WS(is, 110)]), ivs, &(xi[0]));
598 T1K = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
599 {
600 V T1R, T1I, TaM, T1L, TaN, T1S, T1O, T1P, TaK, T1V, T1W;
601 T1O = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
602 T1P = LD(&(xi[WS(is, 70)]), ivs, &(xi[0]));
603 T26 = VSUB(T24, T25);
604 TaK = VADD(T25, T24);
605 T1R = LD(&(xi[WS(is, 102)]), ivs, &(xi[0]));
606 T1I = VSUB(T1G, T1H);
607 TaM = VADD(T1G, T1H);
608 T1L = VSUB(T1J, T1K);
609 TaN = VADD(T1J, T1K);
610 T1Q = VSUB(T1O, T1P);
611 TaT = VADD(T1O, T1P);
612 Tev = VSUB(TaJ, TaK);
613 TaL = VADD(TaJ, TaK);
614 T1S = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
615 T1V = LD(&(xi[WS(is, 118)]), ivs, &(xi[0]));
616 T1W = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
617 TeA = VSUB(TaN, TaM);
618 TaO = VADD(TaM, TaN);
619 T27 = VSUB(T1L, T1I);
620 T1M = VADD(T1I, T1L);
621 T1Y = LD(&(xi[WS(is, 86)]), ivs, &(xi[0]));
622 T1Z = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
623 TaU = VADD(T1S, T1R);
624 T1T = VSUB(T1R, T1S);
625 TaQ = VADD(T1V, T1W);
626 T1X = VSUB(T1V, T1W);
627 }
628 }
629 Tdm = VADD(TaL, TaO);
630 TaP = VSUB(TaL, TaO);
631 T20 = VSUB(T1Y, T1Z);
632 TaR = VADD(T1Z, T1Y);
633 Tew = VSUB(TaT, TaU);
634 TaV = VADD(TaT, TaU);
635 T1U = VFMA(LDK(KP414213562), T1T, T1Q);
636 T29 = VFNMS(LDK(KP414213562), T1Q, T1T);
637 T7f = VFNMS(LDK(KP707106781), T1M, T1F);
638 T1N = VFMA(LDK(KP707106781), T1M, T1F);
639 T28 = VFMA(LDK(KP707106781), T27, T26);
640 T7i = VFNMS(LDK(KP707106781), T27, T26);
641 Tex = VSUB(TaQ, TaR);
642 TaS = VADD(TaQ, TaR);
643 T21 = VFNMS(LDK(KP414213562), T20, T1X);
644 T2a = VFMA(LDK(KP414213562), T1X, T20);
645 }
646 }
647 {
648 V T2J, T2U, T30, T3b, TeL, Tb9, TeO, Tbg, T2M, Tba, T2P, Tbb, T34, Tbh, T33;
649 V T35;
650 {
651 V T2H, T2I, T2S, T2T, T2Y, T2Z, T39, T3a;
652 T2H = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
653 {
654 V Tdn, TaW, Tey, TeB;
655 Tdn = VADD(TaV, TaS);
656 TaW = VSUB(TaS, TaV);
657 Tey = VADD(Tew, Tex);
658 TeB = VSUB(Tex, Tew);
659 {
660 V T2b, T7g, T22, T7j;
661 T2b = VADD(T29, T2a);
662 T7g = VSUB(T2a, T29);
663 T22 = VADD(T1U, T21);
664 T7j = VSUB(T21, T1U);
665 TaX = VFNMS(LDK(KP414213562), TaW, TaP);
666 Tcf = VFMA(LDK(KP414213562), TaP, TaW);
667 Tdo = VADD(Tdm, Tdn);
668 TdH = VSUB(Tdm, Tdn);
669 Tgl = VFNMS(LDK(KP707106781), TeB, TeA);
670 TeC = VFMA(LDK(KP707106781), TeB, TeA);
671 Tgm = VFNMS(LDK(KP707106781), Tey, Tev);
672 Tez = VFMA(LDK(KP707106781), Tey, Tev);
673 T90 = VFNMS(LDK(KP923879532), T7g, T7f);
674 T7h = VFMA(LDK(KP923879532), T7g, T7f);
675 T5P = VFNMS(LDK(KP923879532), T2b, T28);
676 T2c = VFMA(LDK(KP923879532), T2b, T28);
677 T8Z = VFMA(LDK(KP923879532), T7j, T7i);
678 T7k = VFNMS(LDK(KP923879532), T7j, T7i);
679 T5Q = VFNMS(LDK(KP923879532), T22, T1N);
680 T23 = VFMA(LDK(KP923879532), T22, T1N);
681 T2I = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)]));
682 }
683 }
684 T2S = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
685 T2T = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)]));
686 T2Y = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)]));
687 T2Z = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
688 T39 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
689 T3a = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)]));
690 {
691 V T2K, Tbe, Tbf, T2L, T2N, T2O, Tb7, Tb8, T31, T32;
692 T2K = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
693 T2J = VSUB(T2H, T2I);
694 Tb7 = VADD(T2H, T2I);
695 T2U = VSUB(T2S, T2T);
696 Tb8 = VADD(T2S, T2T);
697 T30 = VSUB(T2Y, T2Z);
698 Tbe = VADD(T2Y, T2Z);
699 T3b = VSUB(T39, T3a);
700 Tbf = VADD(T39, T3a);
701 T2L = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)]));
702 T2N = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)]));
703 T2O = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
704 TeL = VSUB(Tb7, Tb8);
705 Tb9 = VADD(Tb7, Tb8);
706 T31 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
707 T32 = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)]));
708 TeO = VSUB(Tbe, Tbf);
709 Tbg = VADD(Tbe, Tbf);
710 T2M = VSUB(T2K, T2L);
711 Tba = VADD(T2K, T2L);
712 T2P = VSUB(T2N, T2O);
713 Tbb = VADD(T2N, T2O);
714 T34 = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)]));
715 Tbh = VADD(T31, T32);
716 T33 = VSUB(T31, T32);
717 T35 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
718 }
719 }
720 {
721 V T4d, T4e, T4o, T4p;
722 {
723 V T2X, T3q, T7t, T7C, T3r, T3e, T7D, T7w;
724 {
725 V T47, TbE, Tbd, Td9, TeW, TeN, T7s, T2W, T7r, T2R, TeP, Tbj, T37, T3c, T48;
726 {
727 V T3W, T3X, TeM, Tbc, T2Q, T2V, Tbi, T36;
728 T3W = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
729 T3X = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)]));
730 TeM = VSUB(Tba, Tbb);
731 Tbc = VADD(Tba, Tbb);
732 T2Q = VADD(T2M, T2P);
733 T2V = VSUB(T2M, T2P);
734 T47 = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)]));
735 Tbi = VADD(T34, T35);
736 T36 = VSUB(T34, T35);
737 TbE = VADD(T3W, T3X);
738 T3Y = VSUB(T3W, T3X);
739 Tbd = VSUB(Tb9, Tbc);
740 Td9 = VADD(Tb9, Tbc);
741 TeW = VFMA(LDK(KP414213562), TeL, TeM);
742 TeN = VFNMS(LDK(KP414213562), TeM, TeL);
743 T7s = VFNMS(LDK(KP707106781), T2V, T2U);
744 T2W = VFMA(LDK(KP707106781), T2V, T2U);
745 T7r = VFNMS(LDK(KP707106781), T2Q, T2J);
746 T2R = VFMA(LDK(KP707106781), T2Q, T2J);
747 TeP = VSUB(Tbh, Tbi);
748 Tbj = VADD(Tbh, Tbi);
749 T37 = VADD(T33, T36);
750 T3c = VSUB(T33, T36);
751 T48 = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
752 }
753 T2X = VFNMS(LDK(KP198912367), T2W, T2R);
754 T3q = VFMA(LDK(KP198912367), T2R, T2W);
755 T7t = VFMA(LDK(KP668178637), T7s, T7r);
756 T7C = VFNMS(LDK(KP668178637), T7r, T7s);
757 {
758 V Tbk, Tda, TeX, TeQ;
759 Tbk = VSUB(Tbg, Tbj);
760 Tda = VADD(Tbg, Tbj);
761 TeX = VFNMS(LDK(KP414213562), TeO, TeP);
762 TeQ = VFMA(LDK(KP414213562), TeP, TeO);
763 {
764 V T7v, T3d, T7u, T38, TbF;
765 T7v = VFNMS(LDK(KP707106781), T3c, T3b);
766 T3d = VFMA(LDK(KP707106781), T3c, T3b);
767 T7u = VFNMS(LDK(KP707106781), T37, T30);
768 T38 = VFMA(LDK(KP707106781), T37, T30);
769 T49 = VSUB(T47, T48);
770 TbF = VADD(T48, T47);
771 TdL = VSUB(Td9, Tda);
772 Tdb = VADD(Td9, Tda);
773 Tbu = VSUB(Tbd, Tbk);
774 Tbl = VADD(Tbd, Tbk);
775 Tgu = VSUB(TeN, TeQ);
776 TeR = VADD(TeN, TeQ);
777 Tgr = VSUB(TeW, TeX);
778 TeY = VADD(TeW, TeX);
779 T3r = VFNMS(LDK(KP198912367), T38, T3d);
780 T3e = VFMA(LDK(KP198912367), T3d, T38);
781 T7D = VFMA(LDK(KP668178637), T7u, T7v);
782 T7w = VFNMS(LDK(KP668178637), T7v, T7u);
783 Tf6 = VSUB(TbE, TbF);
784 TbG = VADD(TbE, TbF);
785 }
786 }
787 }
788 T4d = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)]));
789 T5V = VSUB(T3q, T3r);
790 T3s = VADD(T3q, T3r);
791 T5Y = VSUB(T2X, T3e);
792 T3f = VADD(T2X, T3e);
793 T95 = VSUB(T7D, T7C);
794 T7E = VADD(T7C, T7D);
795 T98 = VSUB(T7t, T7w);
796 T7x = VADD(T7t, T7w);
797 T4e = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
798 T4o = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)]));
799 T4p = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
800 }
801 {
802 V T3Z, T40, T42, T43, TbL, TbM;
803 T3Z = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
804 T40 = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)]));
805 T42 = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)]));
806 T43 = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
807 T4g = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
808 T4f = VSUB(T4d, T4e);
809 TbL = VADD(T4d, T4e);
810 T4q = VSUB(T4o, T4p);
811 TbM = VADD(T4p, T4o);
812 TbH = VADD(T3Z, T40);
813 T41 = VSUB(T3Z, T40);
814 TbI = VADD(T42, T43);
815 T44 = VSUB(T42, T43);
816 T4h = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)]));
817 T4j = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)]));
818 T4k = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
819 Tf9 = VSUB(TbL, TbM);
820 TbN = VADD(TbL, TbM);
821 }
822 }
823 }
824 }
825 }
826 {
827 V TgB, Tgy, T62, T4H, T65, T4u, T9c, T7X, T9f, T7Q, Tg0, Tga, TfF, TeF, TfT;
828 V TfU, TfP, Tg7, TfI, Tfy, TfA, Tf0, Tfz, Tfl, Tg2, TfS;
829 {
830 V Tc1, TbS, Tfc, Tfj, TdX, Te5, TdZ, TdR, Te7, Te3, TdU, Te4;
831 {
832 V TdF, TdS, Tdx, Td5, TdO, TdE, TdC, Tdt, Tdk;
833 {
834 V Tdc, TdA, T4F, T4c, T7V, T7M, T4G, T4t, T7W, T7P, TdB, Tdj;
835 {
836 V Td1, Tdg, TbK, Tf8, Tfh, T4b, T7L, T46, T7K, TbQ, Tfa, T4r, T4m, Td4;
837 TdF = VSUB(TcZ, Td0);
838 Td1 = VADD(TcZ, Td0);
839 {
840 V TbJ, Tf7, T4a, T45;
841 TbJ = VADD(TbH, TbI);
842 Tf7 = VSUB(TbI, TbH);
843 T4a = VSUB(T44, T41);
844 T45 = VADD(T41, T44);
845 {
846 V TbO, T4i, TbP, T4l;
847 TbO = VADD(T4g, T4h);
848 T4i = VSUB(T4g, T4h);
849 TbP = VADD(T4j, T4k);
850 T4l = VSUB(T4j, T4k);
851 Tdg = VADD(TbG, TbJ);
852 TbK = VSUB(TbG, TbJ);
853 Tf8 = VFMA(LDK(KP414213562), Tf7, Tf6);
854 Tfh = VFNMS(LDK(KP414213562), Tf6, Tf7);
855 T4b = VFMA(LDK(KP707106781), T4a, T49);
856 T7L = VFNMS(LDK(KP707106781), T4a, T49);
857 T46 = VFMA(LDK(KP707106781), T45, T3Y);
858 T7K = VFNMS(LDK(KP707106781), T45, T3Y);
859 TbQ = VADD(TbO, TbP);
860 Tfa = VSUB(TbP, TbO);
861 T4r = VSUB(T4l, T4i);
862 T4m = VADD(T4i, T4l);
863 Td4 = VADD(Td2, Td3);
864 TdS = VSUB(Td3, Td2);
865 }
866 }
867 Tdc = VSUB(Td8, Tdb);
868 TdA = VADD(Td8, Tdb);
869 T4F = VFNMS(LDK(KP198912367), T46, T4b);
870 T4c = VFMA(LDK(KP198912367), T4b, T46);
871 T7V = VFMA(LDK(KP668178637), T7K, T7L);
872 T7M = VFNMS(LDK(KP668178637), T7L, T7K);
873 {
874 V Tdh, TbR, Tfb, Tfi;
875 Tdh = VADD(TbN, TbQ);
876 TbR = VSUB(TbN, TbQ);
877 Tfb = VFNMS(LDK(KP414213562), Tfa, Tf9);
878 Tfi = VFMA(LDK(KP414213562), Tf9, Tfa);
879 {
880 V T4s, T7O, T4n, T7N, Tdi;
881 T4s = VFMA(LDK(KP707106781), T4r, T4q);
882 T7O = VFNMS(LDK(KP707106781), T4r, T4q);
883 T4n = VFMA(LDK(KP707106781), T4m, T4f);
884 T7N = VFNMS(LDK(KP707106781), T4m, T4f);
885 Tdx = VADD(Td1, Td4);
886 Td5 = VSUB(Td1, Td4);
887 TdO = VSUB(Tdh, Tdg);
888 Tdi = VADD(Tdg, Tdh);
889 Tc1 = VSUB(TbR, TbK);
890 TbS = VADD(TbK, TbR);
891 TgB = VSUB(Tfb, Tf8);
892 Tfc = VADD(Tf8, Tfb);
893 Tgy = VSUB(Tfi, Tfh);
894 Tfj = VADD(Tfh, Tfi);
895 T4G = VFMA(LDK(KP198912367), T4n, T4s);
896 T4t = VFNMS(LDK(KP198912367), T4s, T4n);
897 T7W = VFNMS(LDK(KP668178637), T7N, T7O);
898 T7P = VFMA(LDK(KP668178637), T7O, T7N);
899 TdB = VADD(Tdf, Tdi);
900 Tdj = VSUB(Tdf, Tdi);
901 }
902 }
903 }
904 T62 = VSUB(T4G, T4F);
905 T4H = VADD(T4F, T4G);
906 T65 = VSUB(T4t, T4c);
907 T4u = VADD(T4c, T4t);
908 T9c = VSUB(T7V, T7W);
909 T7X = VADD(T7V, T7W);
910 T9f = VSUB(T7P, T7M);
911 T7Q = VADD(T7M, T7P);
912 TdE = VSUB(TdB, TdA);
913 TdC = VADD(TdA, TdB);
914 Tdt = VSUB(Tdj, Tdc);
915 Tdk = VADD(Tdc, Tdj);
916 }
917 {
918 V TdT, Tdl, Tdv, TdJ, Te1, Te2, TdQ, Tdz, TdD, Tdu, Tdw;
919 {
920 V TdI, TdP, TdV, TdW, TdM, Tds, Tdy;
921 TdI = VADD(TdG, TdH);
922 TdT = VSUB(TdH, TdG);
923 TdP = VFNMS(LDK(KP414213562), TdO, TdN);
924 TdV = VFMA(LDK(KP414213562), TdN, TdO);
925 TdW = VFMA(LDK(KP414213562), TdK, TdL);
926 TdM = VFNMS(LDK(KP414213562), TdL, TdK);
927 Tdl = VFNMS(LDK(KP707106781), Tdk, Td5);
928 Tdv = VFMA(LDK(KP707106781), Tdk, Td5);
929 Tds = VSUB(Tdo, Tdr);
930 Tdy = VADD(Tdr, Tdo);
931 TdJ = VFMA(LDK(KP707106781), TdI, TdF);
932 Te1 = VFNMS(LDK(KP707106781), TdI, TdF);
933 TdX = VSUB(TdV, TdW);
934 Te2 = VADD(TdW, TdV);
935 Te5 = VSUB(TdP, TdM);
936 TdQ = VADD(TdM, TdP);
937 Tdz = VADD(Tdx, Tdy);
938 TdD = VSUB(Tdx, Tdy);
939 Tdu = VFNMS(LDK(KP707106781), Tdt, Tds);
940 Tdw = VFMA(LDK(KP707106781), Tdt, Tds);
941 }
942 TdZ = VFMA(LDK(KP923879532), TdQ, TdJ);
943 TdR = VFNMS(LDK(KP923879532), TdQ, TdJ);
944 Te7 = VFMA(LDK(KP923879532), Te2, Te1);
945 Te3 = VFNMS(LDK(KP923879532), Te2, Te1);
946 ST(&(xo[WS(os, 32)]), VFMAI(TdE, TdD), ovs, &(xo[0]));
947 ST(&(xo[WS(os, 96)]), VFNMSI(TdE, TdD), ovs, &(xo[0]));
948 ST(&(xo[0]), VADD(Tdz, TdC), ovs, &(xo[0]));
949 ST(&(xo[WS(os, 64)]), VSUB(Tdz, TdC), ovs, &(xo[0]));
950 ST(&(xo[WS(os, 16)]), VFMAI(Tdw, Tdv), ovs, &(xo[0]));
951 ST(&(xo[WS(os, 112)]), VFNMSI(Tdw, Tdv), ovs, &(xo[0]));
952 ST(&(xo[WS(os, 80)]), VFMAI(Tdu, Tdl), ovs, &(xo[0]));
953 ST(&(xo[WS(os, 48)]), VFNMSI(Tdu, Tdl), ovs, &(xo[0]));
954 TdU = VFMA(LDK(KP707106781), TdT, TdS);
955 Te4 = VFNMS(LDK(KP707106781), TdT, TdS);
956 }
957 }
958 {
959 V Tcx, TcJ, TcI, Tcy, TcA, Tbm, Tcp, TaZ, Tcs, Tci, Tbv, TcB, TcD, TbT, Tc2;
960 V TcE, Tat, TaY;
961 Tcx = VFNMS(LDK(KP707106781), Tas, Tad);
962 Tat = VFMA(LDK(KP707106781), Tas, Tad);
963 TaY = VADD(TaI, TaX);
964 TcJ = VSUB(TaX, TaI);
965 {
966 V Tce, Tch, Te8, Te6, TdY, Te0;
967 TcI = VFNMS(LDK(KP707106781), Tcd, Tcc);
968 Tce = VFMA(LDK(KP707106781), Tcd, Tcc);
969 Tch = VSUB(Tcf, Tcg);
970 Tcy = VADD(Tcg, Tcf);
971 Te8 = VFNMS(LDK(KP923879532), Te5, Te4);
972 Te6 = VFMA(LDK(KP923879532), Te5, Te4);
973 TdY = VFNMS(LDK(KP923879532), TdX, TdU);
974 Te0 = VFMA(LDK(KP923879532), TdX, TdU);
975 TcA = VFNMS(LDK(KP707106781), Tbl, Tb6);
976 Tbm = VFMA(LDK(KP707106781), Tbl, Tb6);
977 Tcp = VFNMS(LDK(KP923879532), TaY, Tat);
978 TaZ = VFMA(LDK(KP923879532), TaY, Tat);
979 Tcs = VFNMS(LDK(KP923879532), Tch, Tce);
980 Tci = VFMA(LDK(KP923879532), Tch, Tce);
981 ST(&(xo[WS(os, 88)]), VFNMSI(Te6, Te3), ovs, &(xo[0]));
982 ST(&(xo[WS(os, 40)]), VFMAI(Te6, Te3), ovs, &(xo[0]));
983 ST(&(xo[WS(os, 104)]), VFMAI(Te8, Te7), ovs, &(xo[0]));
984 ST(&(xo[WS(os, 24)]), VFNMSI(Te8, Te7), ovs, &(xo[0]));
985 ST(&(xo[WS(os, 8)]), VFMAI(Te0, TdZ), ovs, &(xo[0]));
986 ST(&(xo[WS(os, 120)]), VFNMSI(Te0, TdZ), ovs, &(xo[0]));
987 ST(&(xo[WS(os, 72)]), VFMAI(TdY, TdR), ovs, &(xo[0]));
988 ST(&(xo[WS(os, 56)]), VFNMSI(TdY, TdR), ovs, &(xo[0]));
989 Tbv = VFMA(LDK(KP707106781), Tbu, Tbt);
990 TcB = VFNMS(LDK(KP707106781), Tbu, Tbt);
991 TcD = VFNMS(LDK(KP707106781), TbS, TbD);
992 TbT = VFMA(LDK(KP707106781), TbS, TbD);
993 Tc2 = VFMA(LDK(KP707106781), Tc1, Tc0);
994 TcE = VFNMS(LDK(KP707106781), Tc1, Tc0);
995 }
996 {
997 V TcR, Tcz, TcU, TcK, Tcq, Tcl, Tct, Tc4;
998 {
999 V Tck, Tbw, Tcj, Tc3;
1000 Tck = VFMA(LDK(KP198912367), Tbm, Tbv);
1001 Tbw = VFNMS(LDK(KP198912367), Tbv, Tbm);
1002 Tcj = VFMA(LDK(KP198912367), TbT, Tc2);
1003 Tc3 = VFNMS(LDK(KP198912367), Tc2, TbT);
1004 TcR = VFNMS(LDK(KP923879532), Tcy, Tcx);
1005 Tcz = VFMA(LDK(KP923879532), Tcy, Tcx);
1006 TcU = VFMA(LDK(KP923879532), TcJ, TcI);
1007 TcK = VFNMS(LDK(KP923879532), TcJ, TcI);
1008 Tcq = VADD(Tck, Tcj);
1009 Tcl = VSUB(Tcj, Tck);
1010 Tct = VSUB(Tc3, Tbw);
1011 Tc4 = VADD(Tbw, Tc3);
1012 }
1013 {
1014 V TfN, Tel, TfY, Tfu, Tfw, Tfv, TcT, TcX, TcQ, TcO, TcW, TcY, TcP, TcH, TfZ;
1015 V TeE;
1016 {
1017 V Teu, TcS, TcN, TcV, TcG, TeD;
1018 TfN = VFNMS(LDK(KP923879532), Tek, Ted);
1019 Tel = VFMA(LDK(KP923879532), Tek, Ted);
1020 {
1021 V TcM, TcC, Tcr, Tcv;
1022 TcM = VFNMS(LDK(KP668178637), TcA, TcB);
1023 TcC = VFMA(LDK(KP668178637), TcB, TcA);
1024 Tcr = VFNMS(LDK(KP980785280), Tcq, Tcp);
1025 Tcv = VFMA(LDK(KP980785280), Tcq, Tcp);
1026 {
1027 V Tco, Tcm, Tcu, Tcw;
1028 Tco = VFMA(LDK(KP980785280), Tcl, Tci);
1029 Tcm = VFNMS(LDK(KP980785280), Tcl, Tci);
1030 Tcu = VFMA(LDK(KP980785280), Tct, Tcs);
1031 Tcw = VFNMS(LDK(KP980785280), Tct, Tcs);
1032 {
1033 V Tcn, Tc5, TcL, TcF;
1034 Tcn = VFMA(LDK(KP980785280), Tc4, TaZ);
1035 Tc5 = VFNMS(LDK(KP980785280), Tc4, TaZ);
1036 TcL = VFNMS(LDK(KP668178637), TcD, TcE);
1037 TcF = VFMA(LDK(KP668178637), TcE, TcD);
1038 TfY = VFNMS(LDK(KP923879532), Tft, Tfq);
1039 Tfu = VFMA(LDK(KP923879532), Tft, Tfq);
1040 Tfw = VFMA(LDK(KP198912367), Teq, Tet);
1041 Teu = VFNMS(LDK(KP198912367), Tet, Teq);
1042 ST(&(xo[WS(os, 92)]), VFNMSI(Tcu, Tcr), ovs, &(xo[0]));
1043 ST(&(xo[WS(os, 36)]), VFMAI(Tcu, Tcr), ovs, &(xo[0]));
1044 ST(&(xo[WS(os, 100)]), VFMAI(Tcw, Tcv), ovs, &(xo[0]));
1045 ST(&(xo[WS(os, 28)]), VFNMSI(Tcw, Tcv), ovs, &(xo[0]));
1046 ST(&(xo[WS(os, 4)]), VFMAI(Tco, Tcn), ovs, &(xo[0]));
1047 ST(&(xo[WS(os, 124)]), VFNMSI(Tco, Tcn), ovs, &(xo[0]));
1048 ST(&(xo[WS(os, 68)]), VFMAI(Tcm, Tc5), ovs, &(xo[0]));
1049 ST(&(xo[WS(os, 60)]), VFNMSI(Tcm, Tc5), ovs, &(xo[0]));
1050 TcS = VADD(TcM, TcL);
1051 TcN = VSUB(TcL, TcM);
1052 TcV = VSUB(TcF, TcC);
1053 TcG = VADD(TcC, TcF);
1054 TeD = VFNMS(LDK(KP198912367), TeC, Tez);
1055 Tfv = VFMA(LDK(KP198912367), Tez, TeC);
1056 }
1057 }
1058 }
1059 TcT = VFMA(LDK(KP831469612), TcS, TcR);
1060 TcX = VFNMS(LDK(KP831469612), TcS, TcR);
1061 TcQ = VFMA(LDK(KP831469612), TcN, TcK);
1062 TcO = VFNMS(LDK(KP831469612), TcN, TcK);
1063 TcW = VFNMS(LDK(KP831469612), TcV, TcU);
1064 TcY = VFMA(LDK(KP831469612), TcV, TcU);
1065 TcP = VFMA(LDK(KP831469612), TcG, Tcz);
1066 TcH = VFNMS(LDK(KP831469612), TcG, Tcz);
1067 TfZ = VSUB(TeD, Teu);
1068 TeE = VADD(Teu, TeD);
1069 }
1070 {
1071 V TfQ, TeS, TfO, Tfx, TeZ, TfR, Tfd, Tfk;
1072 TfQ = VFNMS(LDK(KP923879532), TeR, TeK);
1073 TeS = VFMA(LDK(KP923879532), TeR, TeK);
1074 ST(&(xo[WS(os, 84)]), VFMAI(TcW, TcT), ovs, &(xo[0]));
1075 ST(&(xo[WS(os, 44)]), VFNMSI(TcW, TcT), ovs, &(xo[0]));
1076 ST(&(xo[WS(os, 108)]), VFNMSI(TcY, TcX), ovs, &(xo[0]));
1077 ST(&(xo[WS(os, 20)]), VFMAI(TcY, TcX), ovs, &(xo[0]));
1078 ST(&(xo[WS(os, 116)]), VFMAI(TcQ, TcP), ovs, &(xo[0]));
1079 ST(&(xo[WS(os, 12)]), VFNMSI(TcQ, TcP), ovs, &(xo[0]));
1080 ST(&(xo[WS(os, 52)]), VFMAI(TcO, TcH), ovs, &(xo[0]));
1081 ST(&(xo[WS(os, 76)]), VFNMSI(TcO, TcH), ovs, &(xo[0]));
1082 Tg0 = VFNMS(LDK(KP980785280), TfZ, TfY);
1083 Tga = VFMA(LDK(KP980785280), TfZ, TfY);
1084 TfF = VFNMS(LDK(KP980785280), TeE, Tel);
1085 TeF = VFMA(LDK(KP980785280), TeE, Tel);
1086 TfO = VADD(Tfw, Tfv);
1087 Tfx = VSUB(Tfv, Tfw);
1088 TeZ = VFMA(LDK(KP923879532), TeY, TeV);
1089 TfR = VFNMS(LDK(KP923879532), TeY, TeV);
1090 TfT = VFNMS(LDK(KP923879532), Tfc, Tf5);
1091 Tfd = VFMA(LDK(KP923879532), Tfc, Tf5);
1092 Tfk = VFMA(LDK(KP923879532), Tfj, Tfg);
1093 TfU = VFNMS(LDK(KP923879532), Tfj, Tfg);
1094 TfP = VFMA(LDK(KP980785280), TfO, TfN);
1095 Tg7 = VFNMS(LDK(KP980785280), TfO, TfN);
1096 TfI = VFNMS(LDK(KP980785280), Tfx, Tfu);
1097 Tfy = VFMA(LDK(KP980785280), Tfx, Tfu);
1098 TfA = VFMA(LDK(KP098491403), TeS, TeZ);
1099 Tf0 = VFNMS(LDK(KP098491403), TeZ, TeS);
1100 Tfz = VFMA(LDK(KP098491403), Tfd, Tfk);
1101 Tfl = VFNMS(LDK(KP098491403), Tfk, Tfd);
1102 Tg2 = VFNMS(LDK(KP820678790), TfQ, TfR);
1103 TfS = VFMA(LDK(KP820678790), TfR, TfQ);
1104 }
1105 }
1106 }
1107 }
1108 }
1109 {
1110 V T8x, T8y, T8F, T8w, T8k, T8f, T8n, T80, T9l, T76, T87, T8U, T89, T7e, T7l;
1111 V T8a;
1112 {
1113 V The, Tho, TgT, Tgp, Th7, Th8, Thg, Th6, Th3, Thl, TgW, TgM, TgU, TgP, TgX;
1114 V TgE;
1115 {
1116 V Th1, TgI, TgK, TgJ;
1117 {
1118 V Tgh, Thc, Tgk, TfG, TfB, TfJ, Tfm, Tg1, TfV, Tgn, TfL, TfH;
1119 Th1 = VFMA(LDK(KP923879532), Tgg, Tgf);
1120 Tgh = VFNMS(LDK(KP923879532), Tgg, Tgf);
1121 Thc = VFNMS(LDK(KP923879532), TgH, TgG);
1122 TgI = VFMA(LDK(KP923879532), TgH, TgG);
1123 TgK = VFMA(LDK(KP668178637), Tgi, Tgj);
1124 Tgk = VFNMS(LDK(KP668178637), Tgj, Tgi);
1125 TfG = VADD(TfA, Tfz);
1126 TfB = VSUB(Tfz, TfA);
1127 TfJ = VSUB(Tfl, Tf0);
1128 Tfm = VADD(Tf0, Tfl);
1129 Tg1 = VFNMS(LDK(KP820678790), TfT, TfU);
1130 TfV = VFMA(LDK(KP820678790), TfU, TfT);
1131 Tgn = VFNMS(LDK(KP668178637), Tgm, Tgl);
1132 TgJ = VFMA(LDK(KP668178637), Tgl, Tgm);
1133 TfL = VFMA(LDK(KP995184726), TfG, TfF);
1134 TfH = VFNMS(LDK(KP995184726), TfG, TfF);
1135 {
1136 V TfE, TfC, TfM, TfK;
1137 TfE = VFMA(LDK(KP995184726), TfB, Tfy);
1138 TfC = VFNMS(LDK(KP995184726), TfB, Tfy);
1139 TfM = VFNMS(LDK(KP995184726), TfJ, TfI);
1140 TfK = VFMA(LDK(KP995184726), TfJ, TfI);
1141 {
1142 V TfD, Tfn, Tg8, Tg3;
1143 TfD = VFMA(LDK(KP995184726), Tfm, TeF);
1144 Tfn = VFNMS(LDK(KP995184726), Tfm, TeF);
1145 Tg8 = VADD(Tg2, Tg1);
1146 Tg3 = VSUB(Tg1, Tg2);
1147 {
1148 V Tgb, TfW, Thd, Tgo;
1149 Tgb = VSUB(TfV, TfS);
1150 TfW = VADD(TfS, TfV);
1151 Thd = VSUB(Tgn, Tgk);
1152 Tgo = VADD(Tgk, Tgn);
1153 ST(&(xo[WS(os, 98)]), VFMAI(TfM, TfL), ovs, &(xo[0]));
1154 ST(&(xo[WS(os, 30)]), VFNMSI(TfM, TfL), ovs, &(xo[0]));
1155 ST(&(xo[WS(os, 94)]), VFNMSI(TfK, TfH), ovs, &(xo[0]));
1156 ST(&(xo[WS(os, 34)]), VFMAI(TfK, TfH), ovs, &(xo[0]));
1157 ST(&(xo[WS(os, 2)]), VFMAI(TfE, TfD), ovs, &(xo[0]));
1158 ST(&(xo[WS(os, 126)]), VFNMSI(TfE, TfD), ovs, &(xo[0]));
1159 ST(&(xo[WS(os, 66)]), VFMAI(TfC, Tfn), ovs, &(xo[0]));
1160 ST(&(xo[WS(os, 62)]), VFNMSI(TfC, Tfn), ovs, &(xo[0]));
1161 {
1162 V Tgd, Tg9, Tg6, Tg4;
1163 Tgd = VFNMS(LDK(KP773010453), Tg8, Tg7);
1164 Tg9 = VFMA(LDK(KP773010453), Tg8, Tg7);
1165 Tg6 = VFMA(LDK(KP773010453), Tg3, Tg0);
1166 Tg4 = VFNMS(LDK(KP773010453), Tg3, Tg0);
1167 {
1168 V Tge, Tgc, Tg5, TfX;
1169 Tge = VFMA(LDK(KP773010453), Tgb, Tga);
1170 Tgc = VFNMS(LDK(KP773010453), Tgb, Tga);
1171 Tg5 = VFMA(LDK(KP773010453), TfW, TfP);
1172 TfX = VFNMS(LDK(KP773010453), TfW, TfP);
1173 The = VFMA(LDK(KP831469612), Thd, Thc);
1174 Tho = VFNMS(LDK(KP831469612), Thd, Thc);
1175 TgT = VFMA(LDK(KP831469612), Tgo, Tgh);
1176 Tgp = VFNMS(LDK(KP831469612), Tgo, Tgh);
1177 ST(&(xo[WS(os, 110)]), VFNMSI(Tge, Tgd), ovs, &(xo[0]));
1178 ST(&(xo[WS(os, 18)]), VFMAI(Tge, Tgd), ovs, &(xo[0]));
1179 ST(&(xo[WS(os, 82)]), VFMAI(Tgc, Tg9), ovs, &(xo[0]));
1180 ST(&(xo[WS(os, 46)]), VFNMSI(Tgc, Tg9), ovs, &(xo[0]));
1181 ST(&(xo[WS(os, 114)]), VFMAI(Tg6, Tg5), ovs, &(xo[0]));
1182 ST(&(xo[WS(os, 14)]), VFNMSI(Tg6, Tg5), ovs, &(xo[0]));
1183 ST(&(xo[WS(os, 50)]), VFMAI(Tg4, TfX), ovs, &(xo[0]));
1184 ST(&(xo[WS(os, 78)]), VFNMSI(Tg4, TfX), ovs, &(xo[0]));
1185 }
1186 }
1187 }
1188 }
1189 }
1190 }
1191 {
1192 V Th4, Tgs, Tgv, Th5, Tgz, TgC, Th2, TgL;
1193 Th4 = VFMA(LDK(KP923879532), Tgr, Tgq);
1194 Tgs = VFNMS(LDK(KP923879532), Tgr, Tgq);
1195 Tgv = VFMA(LDK(KP923879532), Tgu, Tgt);
1196 Th5 = VFNMS(LDK(KP923879532), Tgu, Tgt);
1197 Th7 = VFMA(LDK(KP923879532), Tgy, Tgx);
1198 Tgz = VFNMS(LDK(KP923879532), Tgy, Tgx);
1199 TgC = VFMA(LDK(KP923879532), TgB, TgA);
1200 Th8 = VFNMS(LDK(KP923879532), TgB, TgA);
1201 Th2 = VADD(TgK, TgJ);
1202 TgL = VSUB(TgJ, TgK);
1203 {
1204 V TgO, Tgw, TgN, TgD;
1205 TgO = VFMA(LDK(KP534511135), Tgs, Tgv);
1206 Tgw = VFNMS(LDK(KP534511135), Tgv, Tgs);
1207 TgN = VFMA(LDK(KP534511135), Tgz, TgC);
1208 TgD = VFNMS(LDK(KP534511135), TgC, Tgz);
1209 Thg = VFNMS(LDK(KP303346683), Th4, Th5);
1210 Th6 = VFMA(LDK(KP303346683), Th5, Th4);
1211 Th3 = VFMA(LDK(KP831469612), Th2, Th1);
1212 Thl = VFNMS(LDK(KP831469612), Th2, Th1);
1213 TgW = VFNMS(LDK(KP831469612), TgL, TgI);
1214 TgM = VFMA(LDK(KP831469612), TgL, TgI);
1215 TgU = VADD(TgO, TgN);
1216 TgP = VSUB(TgN, TgO);
1217 TgX = VSUB(TgD, Tgw);
1218 TgE = VADD(Tgw, TgD);
1219 }
1220 }
1221 }
1222 {
1223 V T8u, T8v, T7R, T8d, T7G, Thm, Thh, Thp, Tha, T7Y, Thr, Thn;
1224 {
1225 V T7y, T7F, TgZ, TgV;
1226 T8u = VFNMS(LDK(KP831469612), T7x, T7q);
1227 T7y = VFMA(LDK(KP831469612), T7x, T7q);
1228 T7F = VFMA(LDK(KP831469612), T7E, T7B);
1229 T8v = VFNMS(LDK(KP831469612), T7E, T7B);
1230 T8x = VFNMS(LDK(KP831469612), T7Q, T7J);
1231 T7R = VFMA(LDK(KP831469612), T7Q, T7J);
1232 TgZ = VFMA(LDK(KP881921264), TgU, TgT);
1233 TgV = VFNMS(LDK(KP881921264), TgU, TgT);
1234 {
1235 V TgS, TgQ, Th0, TgY;
1236 TgS = VFMA(LDK(KP881921264), TgP, TgM);
1237 TgQ = VFNMS(LDK(KP881921264), TgP, TgM);
1238 Th0 = VFNMS(LDK(KP881921264), TgX, TgW);
1239 TgY = VFMA(LDK(KP881921264), TgX, TgW);
1240 {
1241 V TgR, TgF, Thf, Th9;
1242 TgR = VFMA(LDK(KP881921264), TgE, Tgp);
1243 TgF = VFNMS(LDK(KP881921264), TgE, Tgp);
1244 Thf = VFNMS(LDK(KP303346683), Th7, Th8);
1245 Th9 = VFMA(LDK(KP303346683), Th8, Th7);
1246 T8d = VFNMS(LDK(KP148335987), T7y, T7F);
1247 T7G = VFMA(LDK(KP148335987), T7F, T7y);
1248 ST(&(xo[WS(os, 106)]), VFMAI(Th0, TgZ), ovs, &(xo[0]));
1249 ST(&(xo[WS(os, 22)]), VFNMSI(Th0, TgZ), ovs, &(xo[0]));
1250 ST(&(xo[WS(os, 86)]), VFNMSI(TgY, TgV), ovs, &(xo[0]));
1251 ST(&(xo[WS(os, 42)]), VFMAI(TgY, TgV), ovs, &(xo[0]));
1252 ST(&(xo[WS(os, 10)]), VFMAI(TgS, TgR), ovs, &(xo[0]));
1253 ST(&(xo[WS(os, 118)]), VFNMSI(TgS, TgR), ovs, &(xo[0]));
1254 ST(&(xo[WS(os, 74)]), VFMAI(TgQ, TgF), ovs, &(xo[0]));
1255 ST(&(xo[WS(os, 54)]), VFNMSI(TgQ, TgF), ovs, &(xo[0]));
1256 Thm = VADD(Thg, Thf);
1257 Thh = VSUB(Thf, Thg);
1258 Thp = VSUB(Th9, Th6);
1259 Tha = VADD(Th6, Th9);
1260 T7Y = VFMA(LDK(KP831469612), T7X, T7U);
1261 T8y = VFNMS(LDK(KP831469612), T7X, T7U);
1262 }
1263 }
1264 }
1265 Thr = VFNMS(LDK(KP956940335), Thm, Thl);
1266 Thn = VFMA(LDK(KP956940335), Thm, Thl);
1267 {
1268 V Thk, Thi, Ths, Thq;
1269 Thk = VFMA(LDK(KP956940335), Thh, The);
1270 Thi = VFNMS(LDK(KP956940335), Thh, The);
1271 Ths = VFMA(LDK(KP956940335), Thp, Tho);
1272 Thq = VFNMS(LDK(KP956940335), Thp, Tho);
1273 {
1274 V Thj, Thb, T8e, T7Z;
1275 Thj = VFMA(LDK(KP956940335), Tha, Th3);
1276 Thb = VFNMS(LDK(KP956940335), Tha, Th3);
1277 T8e = VFNMS(LDK(KP148335987), T7R, T7Y);
1278 T7Z = VFMA(LDK(KP148335987), T7Y, T7R);
1279 T8F = VFMA(LDK(KP741650546), T8u, T8v);
1280 T8w = VFNMS(LDK(KP741650546), T8v, T8u);
1281 ST(&(xo[WS(os, 102)]), VFNMSI(Ths, Thr), ovs, &(xo[0]));
1282 ST(&(xo[WS(os, 26)]), VFMAI(Ths, Thr), ovs, &(xo[0]));
1283 ST(&(xo[WS(os, 90)]), VFMAI(Thq, Thn), ovs, &(xo[0]));
1284 ST(&(xo[WS(os, 38)]), VFNMSI(Thq, Thn), ovs, &(xo[0]));
1285 ST(&(xo[WS(os, 122)]), VFMAI(Thk, Thj), ovs, &(xo[0]));
1286 ST(&(xo[WS(os, 6)]), VFNMSI(Thk, Thj), ovs, &(xo[0]));
1287 ST(&(xo[WS(os, 58)]), VFMAI(Thi, Thb), ovs, &(xo[0]));
1288 ST(&(xo[WS(os, 70)]), VFNMSI(Thi, Thb), ovs, &(xo[0]));
1289 T8k = VADD(T8d, T8e);
1290 T8f = VSUB(T8d, T8e);
1291 T8n = VSUB(T7Z, T7G);
1292 T80 = VADD(T7G, T7Z);
1293 }
1294 }
1295 T9l = VSUB(T75, T72);
1296 T76 = VADD(T72, T75);
1297 T87 = VSUB(T85, T86);
1298 T8U = VADD(T85, T86);
1299 T89 = VFNMS(LDK(KP303346683), T7a, T7d);
1300 T7e = VFMA(LDK(KP303346683), T7d, T7a);
1301 T7l = VFMA(LDK(KP303346683), T7k, T7h);
1302 T8a = VFNMS(LDK(KP303346683), T7h, T7k);
1303 }
1304 }
1305 {
1306 V T11, T5h, T5a, T55, T5d, T4K, T5C, T5x, T5F, T5q, T4X, T4Z, T1C, T2d, T50;
1307 {
1308 V T5k, T3g, T3t, T5l, T5n, T4v, T4I, T5o, T8G, T8z;
1309 T5k = VFNMS(LDK(KP980785280), T3f, T2G);
1310 T3g = VFMA(LDK(KP980785280), T3f, T2G);
1311 T8G = VFMA(LDK(KP741650546), T8x, T8y);
1312 T8z = VFNMS(LDK(KP741650546), T8y, T8x);
1313 {
1314 V T8r, T77, T8C, T88;
1315 T8r = VFNMS(LDK(KP831469612), T76, T6Z);
1316 T77 = VFMA(LDK(KP831469612), T76, T6Z);
1317 T8C = VFNMS(LDK(KP831469612), T87, T84);
1318 T88 = VFMA(LDK(KP831469612), T87, T84);
1319 {
1320 V T8D, T7m, T8s, T8b;
1321 T8D = VSUB(T7l, T7e);
1322 T7m = VADD(T7e, T7l);
1323 T8s = VADD(T89, T8a);
1324 T8b = VSUB(T89, T8a);
1325 {
1326 V T8M, T8H, T8P, T8A;
1327 T8M = VADD(T8F, T8G);
1328 T8H = VSUB(T8F, T8G);
1329 T8P = VSUB(T8z, T8w);
1330 T8A = VADD(T8w, T8z);
1331 {
1332 V T8E, T8O, T8j, T7n;
1333 T8E = VFNMS(LDK(KP956940335), T8D, T8C);
1334 T8O = VFMA(LDK(KP956940335), T8D, T8C);
1335 T8j = VFNMS(LDK(KP956940335), T7m, T77);
1336 T7n = VFMA(LDK(KP956940335), T7m, T77);
1337 {
1338 V T8t, T8L, T8m, T8c;
1339 T8t = VFNMS(LDK(KP956940335), T8s, T8r);
1340 T8L = VFMA(LDK(KP956940335), T8s, T8r);
1341 T8m = VFNMS(LDK(KP956940335), T8b, T88);
1342 T8c = VFMA(LDK(KP956940335), T8b, T88);
1343 {
1344 V T8K, T8I, T8S, T8Q;
1345 T8K = VFMA(LDK(KP803207531), T8H, T8E);
1346 T8I = VFNMS(LDK(KP803207531), T8H, T8E);
1347 T8S = VFMA(LDK(KP803207531), T8P, T8O);
1348 T8Q = VFNMS(LDK(KP803207531), T8P, T8O);
1349 {
1350 V T8p, T8l, T8h, T81;
1351 T8p = VFNMS(LDK(KP989176509), T8k, T8j);
1352 T8l = VFMA(LDK(KP989176509), T8k, T8j);
1353 T8h = VFMA(LDK(KP989176509), T80, T7n);
1354 T81 = VFNMS(LDK(KP989176509), T80, T7n);
1355 {
1356 V T8J, T8B, T8R, T8N;
1357 T8J = VFMA(LDK(KP803207531), T8A, T8t);
1358 T8B = VFNMS(LDK(KP803207531), T8A, T8t);
1359 T8R = VFMA(LDK(KP803207531), T8M, T8L);
1360 T8N = VFNMS(LDK(KP803207531), T8M, T8L);
1361 {
1362 V T8q, T8o, T8i, T8g;
1363 T8q = VFNMS(LDK(KP989176509), T8n, T8m);
1364 T8o = VFMA(LDK(KP989176509), T8n, T8m);
1365 T8i = VFMA(LDK(KP989176509), T8f, T8c);
1366 T8g = VFNMS(LDK(KP989176509), T8f, T8c);
1367 ST(&(xo[WS(os, 115)]), VFMAI(T8K, T8J), ovs, &(xo[WS(os, 1)]));
1368 ST(&(xo[WS(os, 13)]), VFNMSI(T8K, T8J), ovs, &(xo[WS(os, 1)]));
1369 ST(&(xo[WS(os, 51)]), VFMAI(T8I, T8B), ovs, &(xo[WS(os, 1)]));
1370 ST(&(xo[WS(os, 77)]), VFNMSI(T8I, T8B), ovs, &(xo[WS(os, 1)]));
1371 ST(&(xo[WS(os, 109)]), VFNMSI(T8S, T8R), ovs, &(xo[WS(os, 1)]));
1372 ST(&(xo[WS(os, 19)]), VFMAI(T8S, T8R), ovs, &(xo[WS(os, 1)]));
1373 ST(&(xo[WS(os, 83)]), VFMAI(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
1374 ST(&(xo[WS(os, 45)]), VFNMSI(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
1375 ST(&(xo[WS(os, 99)]), VFMAI(T8q, T8p), ovs, &(xo[WS(os, 1)]));
1376 ST(&(xo[WS(os, 29)]), VFNMSI(T8q, T8p), ovs, &(xo[WS(os, 1)]));
1377 ST(&(xo[WS(os, 93)]), VFNMSI(T8o, T8l), ovs, &(xo[WS(os, 1)]));
1378 ST(&(xo[WS(os, 35)]), VFMAI(T8o, T8l), ovs, &(xo[WS(os, 1)]));
1379 ST(&(xo[WS(os, 3)]), VFMAI(T8i, T8h), ovs, &(xo[WS(os, 1)]));
1380 ST(&(xo[WS(os, 125)]), VFNMSI(T8i, T8h), ovs, &(xo[WS(os, 1)]));
1381 ST(&(xo[WS(os, 67)]), VFMAI(T8g, T81), ovs, &(xo[WS(os, 1)]));
1382 ST(&(xo[WS(os, 61)]), VFNMSI(T8g, T81), ovs, &(xo[WS(os, 1)]));
1383 T3t = VFMA(LDK(KP980785280), T3s, T3p);
1384 T5l = VFNMS(LDK(KP980785280), T3s, T3p);
1385 }
1386 }
1387 }
1388 }
1389 }
1390 }
1391 }
1392 }
1393 }
1394 T5n = VFNMS(LDK(KP980785280), T4u, T3V);
1395 T4v = VFMA(LDK(KP980785280), T4u, T3V);
1396 T4I = VFMA(LDK(KP980785280), T4H, T4E);
1397 T5o = VFNMS(LDK(KP980785280), T4H, T4E);
1398 {
1399 V T53, T3u, T54, T4J, T5v, T5m, T5w, T5p, T10;
1400 T6b = VSUB(TZ, TI);
1401 T10 = VADD(TI, TZ);
1402 T53 = VFMA(LDK(KP049126849), T3g, T3t);
1403 T3u = VFNMS(LDK(KP049126849), T3t, T3g);
1404 T54 = VFMA(LDK(KP049126849), T4v, T4I);
1405 T4J = VFNMS(LDK(KP049126849), T4I, T4v);
1406 T5v = VFNMS(LDK(KP906347169), T5k, T5l);
1407 T5m = VFMA(LDK(KP906347169), T5l, T5k);
1408 T5w = VFNMS(LDK(KP906347169), T5n, T5o);
1409 T5p = VFMA(LDK(KP906347169), T5o, T5n);
1410 T11 = VFMA(LDK(KP980785280), T10, Tr);
1411 T5h = VFNMS(LDK(KP980785280), T10, Tr);
1412 T5a = VADD(T53, T54);
1413 T55 = VSUB(T53, T54);
1414 T5d = VSUB(T4J, T3u);
1415 T4K = VADD(T3u, T4J);
1416 T5C = VADD(T5v, T5w);
1417 T5x = VSUB(T5v, T5w);
1418 T5F = VSUB(T5p, T5m);
1419 T5q = VADD(T5m, T5p);
1420 T4X = VSUB(T4V, T4W);
1421 T5K = VADD(T4V, T4W);
1422 }
1423 T4Z = VFMA(LDK(KP098491403), T1s, T1B);
1424 T1C = VFNMS(LDK(KP098491403), T1B, T1s);
1425 T2d = VFNMS(LDK(KP098491403), T2c, T23);
1426 T50 = VFMA(LDK(KP098491403), T23, T2c);
1427 }
1428 {
1429 V T9y, T9t, T9B, T9i, T9o, T9n, T9F, T8V, T9Q, T9m, T9R, T92, Ta0, T9V, Ta3;
1430 V T9O;
1431 {
1432 V T9I, T9J, T9L, T9d, T5s, T4Y, T5t, T2e, T5i, T51, T9r, T9a, T9g, T9M, T96;
1433 V T99;
1434 T9I = VFMA(LDK(KP831469612), T95, T94);
1435 T96 = VFNMS(LDK(KP831469612), T95, T94);
1436 T99 = VFNMS(LDK(KP831469612), T98, T97);
1437 T9J = VFMA(LDK(KP831469612), T98, T97);
1438 T9L = VFMA(LDK(KP831469612), T9c, T9b);
1439 T9d = VFNMS(LDK(KP831469612), T9c, T9b);
1440 T5s = VFNMS(LDK(KP980785280), T4X, T4U);
1441 T4Y = VFMA(LDK(KP980785280), T4X, T4U);
1442 T5t = VSUB(T2d, T1C);
1443 T2e = VADD(T1C, T2d);
1444 T5i = VADD(T4Z, T50);
1445 T51 = VSUB(T4Z, T50);
1446 T9r = VFNMS(LDK(KP599376933), T96, T99);
1447 T9a = VFMA(LDK(KP599376933), T99, T96);
1448 T9g = VFNMS(LDK(KP831469612), T9f, T9e);
1449 T9M = VFMA(LDK(KP831469612), T9f, T9e);
1450 {
1451 V T5u, T5E, T8Y, T91;
1452 T5u = VFMA(LDK(KP995184726), T5t, T5s);
1453 T5E = VFNMS(LDK(KP995184726), T5t, T5s);
1454 {
1455 V T59, T2f, T5j, T5B;
1456 T59 = VFNMS(LDK(KP995184726), T2e, T11);
1457 T2f = VFMA(LDK(KP995184726), T2e, T11);
1458 T5j = VFMA(LDK(KP995184726), T5i, T5h);
1459 T5B = VFNMS(LDK(KP995184726), T5i, T5h);
1460 {
1461 V T5c, T52, T9s, T9h;
1462 T5c = VFNMS(LDK(KP995184726), T51, T4Y);
1463 T52 = VFMA(LDK(KP995184726), T51, T4Y);
1464 T9s = VFNMS(LDK(KP599376933), T9d, T9g);
1465 T9h = VFMA(LDK(KP599376933), T9g, T9d);
1466 {
1467 V T5A, T5y, T5I, T5G;
1468 T5A = VFMA(LDK(KP740951125), T5x, T5u);
1469 T5y = VFNMS(LDK(KP740951125), T5x, T5u);
1470 T5I = VFNMS(LDK(KP740951125), T5F, T5E);
1471 T5G = VFMA(LDK(KP740951125), T5F, T5E);
1472 {
1473 V T5f, T5b, T57, T4L;
1474 T5f = VFMA(LDK(KP998795456), T5a, T59);
1475 T5b = VFNMS(LDK(KP998795456), T5a, T59);
1476 T57 = VFMA(LDK(KP998795456), T4K, T2f);
1477 T4L = VFNMS(LDK(KP998795456), T4K, T2f);
1478 {
1479 V T5z, T5r, T5H, T5D;
1480 T5z = VFMA(LDK(KP740951125), T5q, T5j);
1481 T5r = VFNMS(LDK(KP740951125), T5q, T5j);
1482 T5H = VFNMS(LDK(KP740951125), T5C, T5B);
1483 T5D = VFMA(LDK(KP740951125), T5C, T5B);
1484 {
1485 V T5g, T5e, T58, T56;
1486 T5g = VFMA(LDK(KP998795456), T5d, T5c);
1487 T5e = VFNMS(LDK(KP998795456), T5d, T5c);
1488 T58 = VFMA(LDK(KP998795456), T55, T52);
1489 T56 = VFNMS(LDK(KP998795456), T55, T52);
1490 T9y = VADD(T9r, T9s);
1491 T9t = VSUB(T9r, T9s);
1492 T9B = VSUB(T9h, T9a);
1493 T9i = VADD(T9a, T9h);
1494 ST(&(xo[WS(os, 15)]), VFMAI(T5A, T5z), ovs, &(xo[WS(os, 1)]));
1495 ST(&(xo[WS(os, 113)]), VFNMSI(T5A, T5z), ovs, &(xo[WS(os, 1)]));
1496 ST(&(xo[WS(os, 79)]), VFMAI(T5y, T5r), ovs, &(xo[WS(os, 1)]));
1497 ST(&(xo[WS(os, 49)]), VFNMSI(T5y, T5r), ovs, &(xo[WS(os, 1)]));
1498 ST(&(xo[WS(os, 111)]), VFMAI(T5I, T5H), ovs, &(xo[WS(os, 1)]));
1499 ST(&(xo[WS(os, 17)]), VFNMSI(T5I, T5H), ovs, &(xo[WS(os, 1)]));
1500 ST(&(xo[WS(os, 81)]), VFNMSI(T5G, T5D), ovs, &(xo[WS(os, 1)]));
1501 ST(&(xo[WS(os, 47)]), VFMAI(T5G, T5D), ovs, &(xo[WS(os, 1)]));
1502 ST(&(xo[WS(os, 97)]), VFNMSI(T5g, T5f), ovs, &(xo[WS(os, 1)]));
1503 ST(&(xo[WS(os, 31)]), VFMAI(T5g, T5f), ovs, &(xo[WS(os, 1)]));
1504 ST(&(xo[WS(os, 95)]), VFMAI(T5e, T5b), ovs, &(xo[WS(os, 1)]));
1505 ST(&(xo[WS(os, 33)]), VFNMSI(T5e, T5b), ovs, &(xo[WS(os, 1)]));
1506 ST(&(xo[WS(os, 127)]), VFMAI(T58, T57), ovs, &(xo[WS(os, 1)]));
1507 ST(&(xo[WS(os, 1)]), VFNMSI(T58, T57), ovs, &(xo[WS(os, 1)]));
1508 ST(&(xo[WS(os, 63)]), VFMAI(T56, T4L), ovs, &(xo[WS(os, 1)]));
1509 ST(&(xo[WS(os, 65)]), VFNMSI(T56, T4L), ovs, &(xo[WS(os, 1)]));
1510 }
1511 }
1512 }
1513 }
1514 }
1515 }
1516 T9o = VFNMS(LDK(KP534511135), T8W, T8X);
1517 T8Y = VFMA(LDK(KP534511135), T8X, T8W);
1518 T91 = VFMA(LDK(KP534511135), T90, T8Z);
1519 T9n = VFNMS(LDK(KP534511135), T8Z, T90);
1520 {
1521 V T9T, T9K, T9U, T9N;
1522 T9T = VFMA(LDK(KP250486960), T9I, T9J);
1523 T9K = VFNMS(LDK(KP250486960), T9J, T9I);
1524 T9U = VFMA(LDK(KP250486960), T9L, T9M);
1525 T9N = VFNMS(LDK(KP250486960), T9M, T9L);
1526 T9F = VFNMS(LDK(KP831469612), T8U, T8T);
1527 T8V = VFMA(LDK(KP831469612), T8U, T8T);
1528 T9Q = VFNMS(LDK(KP831469612), T9l, T9k);
1529 T9m = VFMA(LDK(KP831469612), T9l, T9k);
1530 T9R = VSUB(T8Y, T91);
1531 T92 = VADD(T8Y, T91);
1532 Ta0 = VADD(T9T, T9U);
1533 T9V = VSUB(T9T, T9U);
1534 Ta3 = VSUB(T9N, T9K);
1535 T9O = VADD(T9K, T9N);
1536 }
1537 }
1538 }
1539 {
1540 V T6y, T6z, T63, T9Y, T9W, Ta6, Ta4, T9D, T9z, T9v, T9j, T6h, T60, T9H, T9Z;
1541 V T9A, T9q, T66, T9X, T9P;
1542 {
1543 V T5W, T9S, Ta2, T9x, T93, T5Z, T9G, T9p;
1544 T6y = VFMA(LDK(KP980785280), T5V, T5U);
1545 T5W = VFNMS(LDK(KP980785280), T5V, T5U);
1546 T9S = VFMA(LDK(KP881921264), T9R, T9Q);
1547 Ta2 = VFNMS(LDK(KP881921264), T9R, T9Q);
1548 T9x = VFNMS(LDK(KP881921264), T92, T8V);
1549 T93 = VFMA(LDK(KP881921264), T92, T8V);
1550 T5Z = VFMA(LDK(KP980785280), T5Y, T5X);
1551 T6z = VFNMS(LDK(KP980785280), T5Y, T5X);
1552 T6B = VFMA(LDK(KP980785280), T62, T61);
1553 T63 = VFNMS(LDK(KP980785280), T62, T61);
1554 T9G = VADD(T9o, T9n);
1555 T9p = VSUB(T9n, T9o);
1556 T9Y = VFMA(LDK(KP970031253), T9V, T9S);
1557 T9W = VFNMS(LDK(KP970031253), T9V, T9S);
1558 Ta6 = VFMA(LDK(KP970031253), Ta3, Ta2);
1559 Ta4 = VFNMS(LDK(KP970031253), Ta3, Ta2);
1560 T9D = VFNMS(LDK(KP857728610), T9y, T9x);
1561 T9z = VFMA(LDK(KP857728610), T9y, T9x);
1562 T9v = VFMA(LDK(KP857728610), T9i, T93);
1563 T9j = VFNMS(LDK(KP857728610), T9i, T93);
1564 T6h = VFMA(LDK(KP472964775), T5W, T5Z);
1565 T60 = VFNMS(LDK(KP472964775), T5Z, T5W);
1566 T9H = VFMA(LDK(KP881921264), T9G, T9F);
1567 T9Z = VFNMS(LDK(KP881921264), T9G, T9F);
1568 T9A = VFNMS(LDK(KP881921264), T9p, T9m);
1569 T9q = VFMA(LDK(KP881921264), T9p, T9m);
1570 T66 = VFMA(LDK(KP980785280), T65, T64);
1571 T6C = VFNMS(LDK(KP980785280), T65, T64);
1572 }
1573 T9X = VFMA(LDK(KP970031253), T9O, T9H);
1574 T9P = VFNMS(LDK(KP970031253), T9O, T9H);
1575 {
1576 V Ta5, Ta1, T9E, T9C;
1577 Ta5 = VFMA(LDK(KP970031253), Ta0, T9Z);
1578 Ta1 = VFNMS(LDK(KP970031253), Ta0, T9Z);
1579 T9E = VFNMS(LDK(KP857728610), T9B, T9A);
1580 T9C = VFMA(LDK(KP857728610), T9B, T9A);
1581 {
1582 V T9w, T9u, T6i, T67;
1583 T9w = VFMA(LDK(KP857728610), T9t, T9q);
1584 T9u = VFNMS(LDK(KP857728610), T9t, T9q);
1585 T6i = VFMA(LDK(KP472964775), T63, T66);
1586 T67 = VFNMS(LDK(KP472964775), T66, T63);
1587 T6J = VFNMS(LDK(KP357805721), T6y, T6z);
1588 T6A = VFMA(LDK(KP357805721), T6z, T6y);
1589 ST(&(xo[WS(os, 123)]), VFMAI(T9Y, T9X), ovs, &(xo[WS(os, 1)]));
1590 ST(&(xo[WS(os, 5)]), VFNMSI(T9Y, T9X), ovs, &(xo[WS(os, 1)]));
1591 ST(&(xo[WS(os, 59)]), VFMAI(T9W, T9P), ovs, &(xo[WS(os, 1)]));
1592 ST(&(xo[WS(os, 69)]), VFNMSI(T9W, T9P), ovs, &(xo[WS(os, 1)]));
1593 ST(&(xo[WS(os, 101)]), VFNMSI(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
1594 ST(&(xo[WS(os, 27)]), VFMAI(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
1595 ST(&(xo[WS(os, 91)]), VFMAI(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
1596 ST(&(xo[WS(os, 37)]), VFNMSI(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
1597 ST(&(xo[WS(os, 107)]), VFMAI(T9E, T9D), ovs, &(xo[WS(os, 1)]));
1598 ST(&(xo[WS(os, 21)]), VFNMSI(T9E, T9D), ovs, &(xo[WS(os, 1)]));
1599 ST(&(xo[WS(os, 85)]), VFNMSI(T9C, T9z), ovs, &(xo[WS(os, 1)]));
1600 ST(&(xo[WS(os, 43)]), VFMAI(T9C, T9z), ovs, &(xo[WS(os, 1)]));
1601 ST(&(xo[WS(os, 11)]), VFMAI(T9w, T9v), ovs, &(xo[WS(os, 1)]));
1602 ST(&(xo[WS(os, 117)]), VFNMSI(T9w, T9v), ovs, &(xo[WS(os, 1)]));
1603 ST(&(xo[WS(os, 75)]), VFMAI(T9u, T9j), ovs, &(xo[WS(os, 1)]));
1604 ST(&(xo[WS(os, 53)]), VFNMSI(T9u, T9j), ovs, &(xo[WS(os, 1)]));
1605 T6o = VADD(T6h, T6i);
1606 T6j = VSUB(T6h, T6i);
1607 T6r = VSUB(T67, T60);
1608 T68 = VADD(T60, T67);
1609 }
1610 }
1611 T6e = VFMA(LDK(KP820678790), T5M, T5N);
1612 T5O = VFNMS(LDK(KP820678790), T5N, T5M);
1613 T5R = VFNMS(LDK(KP820678790), T5Q, T5P);
1614 T6d = VFMA(LDK(KP820678790), T5P, T5Q);
1615 }
1616 }
1617 }
1618 }
1619 }
1620 }
1621 T6D = VFMA(LDK(KP357805721), T6C, T6B);
1622 T6K = VFNMS(LDK(KP357805721), T6B, T6C);
1623 {
1624 V T5L, T6v, T6c, T6G;
1625 T5L = VFNMS(LDK(KP980785280), T5K, T5J);
1626 T6v = VFMA(LDK(KP980785280), T5K, T5J);
1627 T6c = VFNMS(LDK(KP980785280), T6b, T6a);
1628 T6G = VFMA(LDK(KP980785280), T6b, T6a);
1629 {
1630 V T5S, T6H, T6f, T6w;
1631 T5S = VADD(T5O, T5R);
1632 T6H = VSUB(T5O, T5R);
1633 T6f = VSUB(T6d, T6e);
1634 T6w = VADD(T6e, T6d);
1635 {
1636 V T6L, T6Q, T6E, T6T;
1637 T6L = VSUB(T6J, T6K);
1638 T6Q = VADD(T6J, T6K);
1639 T6E = VADD(T6A, T6D);
1640 T6T = VSUB(T6D, T6A);
1641 {
1642 V T6S, T6I, T5T, T6n;
1643 T6S = VFNMS(LDK(KP773010453), T6H, T6G);
1644 T6I = VFMA(LDK(KP773010453), T6H, T6G);
1645 T5T = VFNMS(LDK(KP773010453), T5S, T5L);
1646 T6n = VFMA(LDK(KP773010453), T5S, T5L);
1647 {
1648 V T6P, T6x, T6g, T6q;
1649 T6P = VFNMS(LDK(KP773010453), T6w, T6v);
1650 T6x = VFMA(LDK(KP773010453), T6w, T6v);
1651 T6g = VFNMS(LDK(KP773010453), T6f, T6c);
1652 T6q = VFMA(LDK(KP773010453), T6f, T6c);
1653 {
1654 V T6M, T6O, T6U, T6W;
1655 T6M = VFNMS(LDK(KP941544065), T6L, T6I);
1656 T6O = VFMA(LDK(KP941544065), T6L, T6I);
1657 T6U = VFMA(LDK(KP941544065), T6T, T6S);
1658 T6W = VFNMS(LDK(KP941544065), T6T, T6S);
1659 {
1660 V T6p, T6t, T69, T6l;
1661 T6p = VFNMS(LDK(KP903989293), T6o, T6n);
1662 T6t = VFMA(LDK(KP903989293), T6o, T6n);
1663 T69 = VFNMS(LDK(KP903989293), T68, T5T);
1664 T6l = VFMA(LDK(KP903989293), T68, T5T);
1665 {
1666 V T6F, T6N, T6R, T6V;
1667 T6F = VFNMS(LDK(KP941544065), T6E, T6x);
1668 T6N = VFMA(LDK(KP941544065), T6E, T6x);
1669 T6R = VFMA(LDK(KP941544065), T6Q, T6P);
1670 T6V = VFNMS(LDK(KP941544065), T6Q, T6P);
1671 {
1672 V T6s, T6u, T6k, T6m;
1673 T6s = VFNMS(LDK(KP903989293), T6r, T6q);
1674 T6u = VFMA(LDK(KP903989293), T6r, T6q);
1675 T6k = VFNMS(LDK(KP903989293), T6j, T6g);
1676 T6m = VFMA(LDK(KP903989293), T6j, T6g);
1677 ST(&(xo[WS(os, 7)]), VFMAI(T6O, T6N), ovs, &(xo[WS(os, 1)]));
1678 ST(&(xo[WS(os, 121)]), VFNMSI(T6O, T6N), ovs, &(xo[WS(os, 1)]));
1679 ST(&(xo[WS(os, 71)]), VFMAI(T6M, T6F), ovs, &(xo[WS(os, 1)]));
1680 ST(&(xo[WS(os, 57)]), VFNMSI(T6M, T6F), ovs, &(xo[WS(os, 1)]));
1681 ST(&(xo[WS(os, 103)]), VFMAI(T6W, T6V), ovs, &(xo[WS(os, 1)]));
1682 ST(&(xo[WS(os, 25)]), VFNMSI(T6W, T6V), ovs, &(xo[WS(os, 1)]));
1683 ST(&(xo[WS(os, 89)]), VFNMSI(T6U, T6R), ovs, &(xo[WS(os, 1)]));
1684 ST(&(xo[WS(os, 39)]), VFMAI(T6U, T6R), ovs, &(xo[WS(os, 1)]));
1685 ST(&(xo[WS(os, 105)]), VFNMSI(T6u, T6t), ovs, &(xo[WS(os, 1)]));
1686 ST(&(xo[WS(os, 23)]), VFMAI(T6u, T6t), ovs, &(xo[WS(os, 1)]));
1687 ST(&(xo[WS(os, 87)]), VFMAI(T6s, T6p), ovs, &(xo[WS(os, 1)]));
1688 ST(&(xo[WS(os, 41)]), VFNMSI(T6s, T6p), ovs, &(xo[WS(os, 1)]));
1689 ST(&(xo[WS(os, 119)]), VFMAI(T6m, T6l), ovs, &(xo[WS(os, 1)]));
1690 ST(&(xo[WS(os, 9)]), VFNMSI(T6m, T6l), ovs, &(xo[WS(os, 1)]));
1691 ST(&(xo[WS(os, 55)]), VFMAI(T6k, T69), ovs, &(xo[WS(os, 1)]));
1692 ST(&(xo[WS(os, 73)]), VFNMSI(T6k, T69), ovs, &(xo[WS(os, 1)]));
1693 }
1694 }
1695 }
1696 }
1697 }
1698 }
1699 }
1700 }
1701 }
1702 }
1703 }
1704 VLEAVE();
1705 }
1706
1707 static const kdft_desc desc = { 128, XSIMD_STRING("n1fv_128"), {440, 0, 642, 0}, &GENUS, 0, 0, 0, 0 };
1708
1709 void XSIMD(codelet_n1fv_128) (planner *p) {
1710 X(kdft_register) (p, n1fv_128, &desc);
1711 }
1712
1713 #else /* HAVE_FMA */
1714
1715 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 128 -name n1fv_128 -include n1f.h */
1716
1717 /*
1718 * This function contains 1082 FP additions, 330 FP multiplications,
1719 * (or, 938 additions, 186 multiplications, 144 fused multiply/add),
1720 * 194 stack variables, 31 constants, and 256 memory accesses
1721 */
1722 #include "n1f.h"
1723
1724 static void n1fv_128(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
1725 {
1726 DVK(KP941544065, +0.941544065183020778412509402599502357185589796);
1727 DVK(KP336889853, +0.336889853392220050689253212619147570477766780);
1728 DVK(KP903989293, +0.903989293123443331586200297230537048710132025);
1729 DVK(KP427555093, +0.427555093430282094320966856888798534304578629);
1730 DVK(KP970031253, +0.970031253194543992603984207286100251456865962);
1731 DVK(KP242980179, +0.242980179903263889948274162077471118320990783);
1732 DVK(KP857728610, +0.857728610000272069902269984284770137042490799);
1733 DVK(KP514102744, +0.514102744193221726593693838968815772608049120);
1734 DVK(KP671558954, +0.671558954847018400625376850427421803228750632);
1735 DVK(KP740951125, +0.740951125354959091175616897495162729728955309);
1736 DVK(KP049067674, +0.049067674327418014254954976942682658314745363);
1737 DVK(KP998795456, +0.998795456205172392714771604759100694443203615);
1738 DVK(KP595699304, +0.595699304492433343467036528829969889511926338);
1739 DVK(KP803207531, +0.803207531480644909806676512963141923879569427);
1740 DVK(KP146730474, +0.146730474455361751658850129646717819706215317);
1741 DVK(KP989176509, +0.989176509964780973451673738016243063983689533);
1742 DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
1743 DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
1744 DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
1745 DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
1746 DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
1747 DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
1748 DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
1749 DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
1750 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
1751 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
1752 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
1753 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
1754 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
1755 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
1756 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
1757 {
1758 INT i;
1759 const R *xi;
1760 R *xo;
1761 xi = ri;
1762 xo = ro;
1763 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)) {
1764 V Tr, T5J, Ted, Tgf, Tfq, TgH, T4U, T6b, T6Z, T8T, Tad, TcZ, Tcc, Td0, T84;
1765 V T9l, Tb6, Tbt, T2G, T5X, TeV, Tgr, T3p, T5V, T7B, T95, TeK, Tgt, T7q, T97;
1766 V Td8, TdK, TbD, Tc0, T3V, T61, Tfg, TgB, T4E, T65, T7U, T9f, Tf5, Tgx, T7J;
1767 V T9b, Tdf, TdN, Td2, Td3, TI, T4V, Tft, Tgg, TZ, T4W, T75, T86, Tek, TgG;
1768 V T72, T85, Tas, Tcd, Tdp, Tdq, TdG, Teq, Tgm, Tet, Tgl, T1s, T5P, T1B, T5Q;
1769 V T7d, T8Z, TaI, Tcf, T7a, T90, Tdm, Tdn, TdH, Tez, Tgi, TeC, Tgj, T23, T5N;
1770 V T2c, T5M, T7k, T8X, TaX, Tcg, T7h, T8W, Tbl, Tbu, Tdb, TdL, TeY, Tgu, TeR;
1771 V Tgq, T7x, T98, T7E, T94, T3f, T5Y, T3s, T5U, TbS, Tc1, Tdi, TdO, Tfj, Tgy;
1772 V Tfc, TgA, T7Q, T9e, T7X, T9c, T4u, T64, T4H, T62;
1773 {
1774 V T3, Ta7, T4P, Ta8, Ta, Tab, T4M, Taa, Tc9, Tca, Ti, Tea, T4S, Tc6, Tc7;
1775 V Tp, Teb, T4R;
1776 {
1777 V T1, T2, T4N, T4O;
1778 T1 = LD(&(xi[0]), ivs, &(xi[0]));
1779 T2 = LD(&(xi[WS(is, 64)]), ivs, &(xi[0]));
1780 T3 = VSUB(T1, T2);
1781 Ta7 = VADD(T1, T2);
1782 T4N = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
1783 T4O = LD(&(xi[WS(is, 96)]), ivs, &(xi[0]));
1784 T4P = VSUB(T4N, T4O);
1785 Ta8 = VADD(T4N, T4O);
1786 }
1787 {
1788 V T4, T5, T6, T7, T8, T9;
1789 T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
1790 T5 = LD(&(xi[WS(is, 80)]), ivs, &(xi[0]));
1791 T6 = VSUB(T4, T5);
1792 T7 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0]));
1793 T8 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
1794 T9 = VSUB(T7, T8);
1795 Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
1796 Tab = VADD(T7, T8);
1797 T4M = VMUL(LDK(KP707106781), VSUB(T9, T6));
1798 Taa = VADD(T4, T5);
1799 }
1800 {
1801 V Te, Th, Tl, To;
1802 {
1803 V Tc, Td, Tf, Tg;
1804 Tc = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
1805 Td = LD(&(xi[WS(is, 72)]), ivs, &(xi[0]));
1806 Te = VSUB(Tc, Td);
1807 Tc9 = VADD(Tc, Td);
1808 Tf = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
1809 Tg = LD(&(xi[WS(is, 104)]), ivs, &(xi[0]));
1810 Th = VSUB(Tf, Tg);
1811 Tca = VADD(Tf, Tg);
1812 }
1813 Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
1814 Tea = VSUB(Tc9, Tca);
1815 T4S = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
1816 {
1817 V Tj, Tk, Tm, Tn;
1818 Tj = LD(&(xi[WS(is, 120)]), ivs, &(xi[0]));
1819 Tk = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
1820 Tl = VSUB(Tj, Tk);
1821 Tc6 = VADD(Tj, Tk);
1822 Tm = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
1823 Tn = LD(&(xi[WS(is, 88)]), ivs, &(xi[0]));
1824 To = VSUB(Tm, Tn);
1825 Tc7 = VADD(Tm, Tn);
1826 }
1827 Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
1828 Teb = VSUB(Tc6, Tc7);
1829 T4R = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
1830 }
1831 {
1832 V Tb, Tq, Te9, Tec;
1833 Tb = VADD(T3, Ta);
1834 Tq = VADD(Ti, Tp);
1835 Tr = VADD(Tb, Tq);
1836 T5J = VSUB(Tb, Tq);
1837 Te9 = VSUB(Ta7, Ta8);
1838 Tec = VMUL(LDK(KP707106781), VADD(Tea, Teb));
1839 Ted = VADD(Te9, Tec);
1840 Tgf = VSUB(Te9, Tec);
1841 }
1842 {
1843 V Tfo, Tfp, T4Q, T4T;
1844 Tfo = VSUB(Tab, Taa);
1845 Tfp = VMUL(LDK(KP707106781), VSUB(Teb, Tea));
1846 Tfq = VADD(Tfo, Tfp);
1847 TgH = VSUB(Tfp, Tfo);
1848 T4Q = VSUB(T4M, T4P);
1849 T4T = VSUB(T4R, T4S);
1850 T4U = VADD(T4Q, T4T);
1851 T6b = VSUB(T4T, T4Q);
1852 }
1853 {
1854 V T6X, T6Y, Ta9, Tac;
1855 T6X = VSUB(T3, Ta);
1856 T6Y = VADD(T4S, T4R);
1857 T6Z = VADD(T6X, T6Y);
1858 T8T = VSUB(T6X, T6Y);
1859 Ta9 = VADD(Ta7, Ta8);
1860 Tac = VADD(Taa, Tab);
1861 Tad = VSUB(Ta9, Tac);
1862 TcZ = VADD(Ta9, Tac);
1863 }
1864 {
1865 V Tc8, Tcb, T82, T83;
1866 Tc8 = VADD(Tc6, Tc7);
1867 Tcb = VADD(Tc9, Tca);
1868 Tcc = VSUB(Tc8, Tcb);
1869 Td0 = VADD(Tcb, Tc8);
1870 T82 = VADD(T4P, T4M);
1871 T83 = VSUB(Tp, Ti);
1872 T84 = VADD(T82, T83);
1873 T9l = VSUB(T83, T82);
1874 }
1875 }
1876 {
1877 V Tb0, Tb1, T2i, Tb2, T3k, Tb3, Tb4, T2p, Tb5, T3h, T2x, TeH, T3n, Tbs, T2E;
1878 V TeI, T3m, Tbp, T2l, T2o, TeG, TeJ;
1879 {
1880 V T2g, T2h, T3i, T3j;
1881 T2g = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
1882 T2h = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)]));
1883 Tb0 = VADD(T2g, T2h);
1884 T3i = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
1885 T3j = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)]));
1886 Tb1 = VADD(T3i, T3j);
1887 T2i = VSUB(T2g, T2h);
1888 Tb2 = VADD(Tb0, Tb1);
1889 T3k = VSUB(T3i, T3j);
1890 }
1891 {
1892 V T2j, T2k, T2m, T2n;
1893 T2j = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
1894 T2k = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)]));
1895 T2l = VSUB(T2j, T2k);
1896 Tb3 = VADD(T2j, T2k);
1897 T2m = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)]));
1898 T2n = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
1899 T2o = VSUB(T2m, T2n);
1900 Tb4 = VADD(T2m, T2n);
1901 }
1902 T2p = VMUL(LDK(KP707106781), VADD(T2l, T2o));
1903 Tb5 = VADD(Tb3, Tb4);
1904 T3h = VMUL(LDK(KP707106781), VSUB(T2o, T2l));
1905 {
1906 V T2t, Tbq, T2w, Tbr;
1907 {
1908 V T2r, T2s, T2u, T2v;
1909 T2r = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
1910 T2s = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)]));
1911 T2t = VSUB(T2r, T2s);
1912 Tbq = VADD(T2r, T2s);
1913 T2u = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
1914 T2v = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)]));
1915 T2w = VSUB(T2u, T2v);
1916 Tbr = VADD(T2u, T2v);
1917 }
1918 T2x = VFNMS(LDK(KP382683432), T2w, VMUL(LDK(KP923879532), T2t));
1919 TeH = VSUB(Tbq, Tbr);
1920 T3n = VFMA(LDK(KP382683432), T2t, VMUL(LDK(KP923879532), T2w));
1921 Tbs = VADD(Tbq, Tbr);
1922 }
1923 {
1924 V T2A, Tbn, T2D, Tbo;
1925 {
1926 V T2y, T2z, T2B, T2C;
1927 T2y = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)]));
1928 T2z = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
1929 T2A = VSUB(T2y, T2z);
1930 Tbn = VADD(T2y, T2z);
1931 T2B = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
1932 T2C = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)]));
1933 T2D = VSUB(T2B, T2C);
1934 Tbo = VADD(T2B, T2C);
1935 }
1936 T2E = VFMA(LDK(KP923879532), T2A, VMUL(LDK(KP382683432), T2D));
1937 TeI = VSUB(Tbn, Tbo);
1938 T3m = VFNMS(LDK(KP923879532), T2D, VMUL(LDK(KP382683432), T2A));
1939 Tbp = VADD(Tbn, Tbo);
1940 }
1941 Tb6 = VSUB(Tb2, Tb5);
1942 Tbt = VSUB(Tbp, Tbs);
1943 {
1944 V T2q, T2F, TeT, TeU;
1945 T2q = VADD(T2i, T2p);
1946 T2F = VADD(T2x, T2E);
1947 T2G = VADD(T2q, T2F);
1948 T5X = VSUB(T2q, T2F);
1949 TeT = VSUB(Tb4, Tb3);
1950 TeU = VMUL(LDK(KP707106781), VSUB(TeI, TeH));
1951 TeV = VADD(TeT, TeU);
1952 Tgr = VSUB(TeU, TeT);
1953 }
1954 {
1955 V T3l, T3o, T7z, T7A;
1956 T3l = VSUB(T3h, T3k);
1957 T3o = VSUB(T3m, T3n);
1958 T3p = VADD(T3l, T3o);
1959 T5V = VSUB(T3o, T3l);
1960 T7z = VADD(T3k, T3h);
1961 T7A = VSUB(T2E, T2x);
1962 T7B = VADD(T7z, T7A);
1963 T95 = VSUB(T7A, T7z);
1964 }
1965 TeG = VSUB(Tb0, Tb1);
1966 TeJ = VMUL(LDK(KP707106781), VADD(TeH, TeI));
1967 TeK = VADD(TeG, TeJ);
1968 Tgt = VSUB(TeG, TeJ);
1969 {
1970 V T7o, T7p, Td6, Td7;
1971 T7o = VSUB(T2i, T2p);
1972 T7p = VADD(T3n, T3m);
1973 T7q = VADD(T7o, T7p);
1974 T97 = VSUB(T7o, T7p);
1975 Td6 = VADD(Tb2, Tb5);
1976 Td7 = VADD(Tbs, Tbp);
1977 Td8 = VADD(Td6, Td7);
1978 TdK = VSUB(Td6, Td7);
1979 }
1980 }
1981 {
1982 V Tbx, Tby, T3x, Tbz, T4z, TbA, TbB, T3E, TbC, T4w, T3M, Tf2, T4C, TbZ, T3T;
1983 V Tf3, T4B, TbW, T3A, T3D, Tf1, Tf4;
1984 {
1985 V T3v, T3w, T4x, T4y;
1986 T3v = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)]));
1987 T3w = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
1988 Tbx = VADD(T3v, T3w);
1989 T4x = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
1990 T4y = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)]));
1991 Tby = VADD(T4x, T4y);
1992 T3x = VSUB(T3v, T3w);
1993 Tbz = VADD(Tbx, Tby);
1994 T4z = VSUB(T4x, T4y);
1995 }
1996 {
1997 V T3y, T3z, T3B, T3C;
1998 T3y = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
1999 T3z = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)]));
2000 T3A = VSUB(T3y, T3z);
2001 TbA = VADD(T3y, T3z);
2002 T3B = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)]));
2003 T3C = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
2004 T3D = VSUB(T3B, T3C);
2005 TbB = VADD(T3B, T3C);
2006 }
2007 T3E = VMUL(LDK(KP707106781), VADD(T3A, T3D));
2008 TbC = VADD(TbA, TbB);
2009 T4w = VMUL(LDK(KP707106781), VSUB(T3D, T3A));
2010 {
2011 V T3I, TbX, T3L, TbY;
2012 {
2013 V T3G, T3H, T3J, T3K;
2014 T3G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
2015 T3H = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)]));
2016 T3I = VSUB(T3G, T3H);
2017 TbX = VADD(T3G, T3H);
2018 T3J = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
2019 T3K = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)]));
2020 T3L = VSUB(T3J, T3K);
2021 TbY = VADD(T3J, T3K);
2022 }
2023 T3M = VFNMS(LDK(KP382683432), T3L, VMUL(LDK(KP923879532), T3I));
2024 Tf2 = VSUB(TbX, TbY);
2025 T4C = VFMA(LDK(KP382683432), T3I, VMUL(LDK(KP923879532), T3L));
2026 TbZ = VADD(TbX, TbY);
2027 }
2028 {
2029 V T3P, TbU, T3S, TbV;
2030 {
2031 V T3N, T3O, T3Q, T3R;
2032 T3N = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)]));
2033 T3O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
2034 T3P = VSUB(T3N, T3O);
2035 TbU = VADD(T3N, T3O);
2036 T3Q = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
2037 T3R = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)]));
2038 T3S = VSUB(T3Q, T3R);
2039 TbV = VADD(T3Q, T3R);
2040 }
2041 T3T = VFMA(LDK(KP923879532), T3P, VMUL(LDK(KP382683432), T3S));
2042 Tf3 = VSUB(TbU, TbV);
2043 T4B = VFNMS(LDK(KP923879532), T3S, VMUL(LDK(KP382683432), T3P));
2044 TbW = VADD(TbU, TbV);
2045 }
2046 TbD = VSUB(Tbz, TbC);
2047 Tc0 = VSUB(TbW, TbZ);
2048 {
2049 V T3F, T3U, Tfe, Tff;
2050 T3F = VADD(T3x, T3E);
2051 T3U = VADD(T3M, T3T);
2052 T3V = VADD(T3F, T3U);
2053 T61 = VSUB(T3F, T3U);
2054 Tfe = VSUB(TbB, TbA);
2055 Tff = VMUL(LDK(KP707106781), VSUB(Tf3, Tf2));
2056 Tfg = VADD(Tfe, Tff);
2057 TgB = VSUB(Tff, Tfe);
2058 }
2059 {
2060 V T4A, T4D, T7S, T7T;
2061 T4A = VSUB(T4w, T4z);
2062 T4D = VSUB(T4B, T4C);
2063 T4E = VADD(T4A, T4D);
2064 T65 = VSUB(T4D, T4A);
2065 T7S = VADD(T4z, T4w);
2066 T7T = VSUB(T3T, T3M);
2067 T7U = VADD(T7S, T7T);
2068 T9f = VSUB(T7T, T7S);
2069 }
2070 Tf1 = VSUB(Tbx, Tby);
2071 Tf4 = VMUL(LDK(KP707106781), VADD(Tf2, Tf3));
2072 Tf5 = VADD(Tf1, Tf4);
2073 Tgx = VSUB(Tf1, Tf4);
2074 {
2075 V T7H, T7I, Tdd, Tde;
2076 T7H = VSUB(T3x, T3E);
2077 T7I = VADD(T4C, T4B);
2078 T7J = VADD(T7H, T7I);
2079 T9b = VSUB(T7H, T7I);
2080 Tdd = VADD(Tbz, TbC);
2081 Tde = VADD(TbZ, TbW);
2082 Tdf = VADD(Tdd, Tde);
2083 TdN = VSUB(Tdd, Tde);
2084 }
2085 }
2086 {
2087 V Tu, Tee, TG, Tag, TL, Teh, TX, Tan, TB, Tef, TD, Taj, TS, Tei, TU;
2088 V Taq, Teg, Tej;
2089 {
2090 V Ts, Tt, Tae, TE, TF, Taf;
2091 Ts = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
2092 Tt = LD(&(xi[WS(is, 68)]), ivs, &(xi[0]));
2093 Tae = VADD(Ts, Tt);
2094 TE = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
2095 TF = LD(&(xi[WS(is, 100)]), ivs, &(xi[0]));
2096 Taf = VADD(TE, TF);
2097 Tu = VSUB(Ts, Tt);
2098 Tee = VSUB(Tae, Taf);
2099 TG = VSUB(TE, TF);
2100 Tag = VADD(Tae, Taf);
2101 }
2102 {
2103 V TJ, TK, Tal, TV, TW, Tam;
2104 TJ = LD(&(xi[WS(is, 124)]), ivs, &(xi[0]));
2105 TK = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
2106 Tal = VADD(TJ, TK);
2107 TV = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
2108 TW = LD(&(xi[WS(is, 92)]), ivs, &(xi[0]));
2109 Tam = VADD(TV, TW);
2110 TL = VSUB(TJ, TK);
2111 Teh = VSUB(Tal, Tam);
2112 TX = VSUB(TV, TW);
2113 Tan = VADD(Tal, Tam);
2114 }
2115 {
2116 V Tx, Tah, TA, Tai;
2117 {
2118 V Tv, Tw, Ty, Tz;
2119 Tv = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
2120 Tw = LD(&(xi[WS(is, 84)]), ivs, &(xi[0]));
2121 Tx = VSUB(Tv, Tw);
2122 Tah = VADD(Tv, Tw);
2123 Ty = LD(&(xi[WS(is, 116)]), ivs, &(xi[0]));
2124 Tz = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
2125 TA = VSUB(Ty, Tz);
2126 Tai = VADD(Ty, Tz);
2127 }
2128 TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
2129 Tef = VSUB(Tai, Tah);
2130 TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
2131 Taj = VADD(Tah, Tai);
2132 }
2133 {
2134 V TO, Tao, TR, Tap;
2135 {
2136 V TM, TN, TP, TQ;
2137 TM = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
2138 TN = LD(&(xi[WS(is, 76)]), ivs, &(xi[0]));
2139 TO = VSUB(TM, TN);
2140 Tao = VADD(TM, TN);
2141 TP = LD(&(xi[WS(is, 108)]), ivs, &(xi[0]));
2142 TQ = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
2143 TR = VSUB(TP, TQ);
2144 Tap = VADD(TP, TQ);
2145 }
2146 TS = VMUL(LDK(KP707106781), VADD(TO, TR));
2147 Tei = VSUB(Tap, Tao);
2148 TU = VMUL(LDK(KP707106781), VSUB(TR, TO));
2149 Taq = VADD(Tao, Tap);
2150 }
2151 Td2 = VADD(Tag, Taj);
2152 Td3 = VADD(Tan, Taq);
2153 {
2154 V TC, TH, Tfr, Tfs;
2155 TC = VADD(Tu, TB);
2156 TH = VSUB(TD, TG);
2157 TI = VFMA(LDK(KP980785280), TC, VMUL(LDK(KP195090322), TH));
2158 T4V = VFNMS(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
2159 Tfr = VFNMS(LDK(KP382683432), Tee, VMUL(LDK(KP923879532), Tef));
2160 Tfs = VFMA(LDK(KP382683432), Teh, VMUL(LDK(KP923879532), Tei));
2161 Tft = VADD(Tfr, Tfs);
2162 Tgg = VSUB(Tfs, Tfr);
2163 }
2164 {
2165 V TT, TY, T73, T74;
2166 TT = VADD(TL, TS);
2167 TY = VSUB(TU, TX);
2168 TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
2169 T4W = VFMA(LDK(KP195090322), TT, VMUL(LDK(KP980785280), TY));
2170 T73 = VSUB(TL, TS);
2171 T74 = VADD(TX, TU);
2172 T75 = VFNMS(LDK(KP555570233), T74, VMUL(LDK(KP831469612), T73));
2173 T86 = VFMA(LDK(KP555570233), T73, VMUL(LDK(KP831469612), T74));
2174 }
2175 Teg = VFMA(LDK(KP923879532), Tee, VMUL(LDK(KP382683432), Tef));
2176 Tej = VFNMS(LDK(KP382683432), Tei, VMUL(LDK(KP923879532), Teh));
2177 Tek = VADD(Teg, Tej);
2178 TgG = VSUB(Tej, Teg);
2179 {
2180 V T70, T71, Tak, Tar;
2181 T70 = VSUB(Tu, TB);
2182 T71 = VADD(TG, TD);
2183 T72 = VFMA(LDK(KP831469612), T70, VMUL(LDK(KP555570233), T71));
2184 T85 = VFNMS(LDK(KP555570233), T70, VMUL(LDK(KP831469612), T71));
2185 Tak = VSUB(Tag, Taj);
2186 Tar = VSUB(Tan, Taq);
2187 Tas = VMUL(LDK(KP707106781), VADD(Tak, Tar));
2188 Tcd = VMUL(LDK(KP707106781), VSUB(Tar, Tak));
2189 }
2190 }
2191 {
2192 V Tav, Tau, T1b, Taw, T1v, Tay, Tax, T18, Taz, T1w, T1j, Teo, T1z, TaD, T1q;
2193 V Ten, T1y, TaG, T14, T17, Tem, Tep;
2194 {
2195 V T19, T1a, T1t, T1u;
2196 T19 = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
2197 T1a = LD(&(xi[WS(is, 98)]), ivs, &(xi[0]));
2198 Tav = VADD(T19, T1a);
2199 T1t = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
2200 T1u = LD(&(xi[WS(is, 66)]), ivs, &(xi[0]));
2201 Tau = VADD(T1t, T1u);
2202 T1b = VSUB(T19, T1a);
2203 Taw = VADD(Tau, Tav);
2204 T1v = VSUB(T1t, T1u);
2205 }
2206 {
2207 V T12, T13, T15, T16;
2208 T12 = LD(&(xi[WS(is, 114)]), ivs, &(xi[0]));
2209 T13 = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
2210 T14 = VSUB(T12, T13);
2211 Tay = VADD(T12, T13);
2212 T15 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
2213 T16 = LD(&(xi[WS(is, 82)]), ivs, &(xi[0]));
2214 T17 = VSUB(T15, T16);
2215 Tax = VADD(T15, T16);
2216 }
2217 T18 = VMUL(LDK(KP707106781), VSUB(T14, T17));
2218 Taz = VADD(Tax, Tay);
2219 T1w = VMUL(LDK(KP707106781), VADD(T17, T14));
2220 {
2221 V T1f, TaB, T1i, TaC;
2222 {
2223 V T1d, T1e, T1g, T1h;
2224 T1d = LD(&(xi[WS(is, 122)]), ivs, &(xi[0]));
2225 T1e = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
2226 T1f = VSUB(T1d, T1e);
2227 TaB = VADD(T1d, T1e);
2228 T1g = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
2229 T1h = LD(&(xi[WS(is, 90)]), ivs, &(xi[0]));
2230 T1i = VSUB(T1g, T1h);
2231 TaC = VADD(T1g, T1h);
2232 }
2233 T1j = VFNMS(LDK(KP923879532), T1i, VMUL(LDK(KP382683432), T1f));
2234 Teo = VSUB(TaB, TaC);
2235 T1z = VFMA(LDK(KP923879532), T1f, VMUL(LDK(KP382683432), T1i));
2236 TaD = VADD(TaB, TaC);
2237 }
2238 {
2239 V T1m, TaE, T1p, TaF;
2240 {
2241 V T1k, T1l, T1n, T1o;
2242 T1k = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
2243 T1l = LD(&(xi[WS(is, 74)]), ivs, &(xi[0]));
2244 T1m = VSUB(T1k, T1l);
2245 TaE = VADD(T1k, T1l);
2246 T1n = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
2247 T1o = LD(&(xi[WS(is, 106)]), ivs, &(xi[0]));
2248 T1p = VSUB(T1n, T1o);
2249 TaF = VADD(T1n, T1o);
2250 }
2251 T1q = VFMA(LDK(KP382683432), T1m, VMUL(LDK(KP923879532), T1p));
2252 Ten = VSUB(TaE, TaF);
2253 T1y = VFNMS(LDK(KP382683432), T1p, VMUL(LDK(KP923879532), T1m));
2254 TaG = VADD(TaE, TaF);
2255 }
2256 Tdp = VADD(Taw, Taz);
2257 Tdq = VADD(TaG, TaD);
2258 TdG = VSUB(Tdp, Tdq);
2259 Tem = VSUB(Tau, Tav);
2260 Tep = VMUL(LDK(KP707106781), VADD(Ten, Teo));
2261 Teq = VADD(Tem, Tep);
2262 Tgm = VSUB(Tem, Tep);
2263 {
2264 V Ter, Tes, T1c, T1r;
2265 Ter = VSUB(Tay, Tax);
2266 Tes = VMUL(LDK(KP707106781), VSUB(Teo, Ten));
2267 Tet = VADD(Ter, Tes);
2268 Tgl = VSUB(Tes, Ter);
2269 T1c = VSUB(T18, T1b);
2270 T1r = VSUB(T1j, T1q);
2271 T1s = VADD(T1c, T1r);
2272 T5P = VSUB(T1r, T1c);
2273 }
2274 {
2275 V T1x, T1A, T7b, T7c;
2276 T1x = VADD(T1v, T1w);
2277 T1A = VADD(T1y, T1z);
2278 T1B = VADD(T1x, T1A);
2279 T5Q = VSUB(T1x, T1A);
2280 T7b = VADD(T1b, T18);
2281 T7c = VSUB(T1z, T1y);
2282 T7d = VADD(T7b, T7c);
2283 T8Z = VSUB(T7c, T7b);
2284 }
2285 {
2286 V TaA, TaH, T78, T79;
2287 TaA = VSUB(Taw, Taz);
2288 TaH = VSUB(TaD, TaG);
2289 TaI = VFMA(LDK(KP923879532), TaA, VMUL(LDK(KP382683432), TaH));
2290 Tcf = VFNMS(LDK(KP382683432), TaA, VMUL(LDK(KP923879532), TaH));
2291 T78 = VSUB(T1v, T1w);
2292 T79 = VADD(T1q, T1j);
2293 T7a = VADD(T78, T79);
2294 T90 = VSUB(T78, T79);
2295 }
2296 }
2297 {
2298 V TaJ, TaK, T1F, TaL, T27, TaM, TaN, T1M, TaO, T24, T1U, Tew, T2a, TaV, T21;
2299 V Tex, T29, TaS, T1I, T1L, Tev, Tey;
2300 {
2301 V T1D, T1E, T25, T26;
2302 T1D = LD(&(xi[WS(is, 126)]), ivs, &(xi[0]));
2303 T1E = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
2304 TaJ = VADD(T1D, T1E);
2305 T25 = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
2306 T26 = LD(&(xi[WS(is, 94)]), ivs, &(xi[0]));
2307 TaK = VADD(T25, T26);
2308 T1F = VSUB(T1D, T1E);
2309 TaL = VADD(TaJ, TaK);
2310 T27 = VSUB(T25, T26);
2311 }
2312 {
2313 V T1G, T1H, T1J, T1K;
2314 T1G = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
2315 T1H = LD(&(xi[WS(is, 78)]), ivs, &(xi[0]));
2316 T1I = VSUB(T1G, T1H);
2317 TaM = VADD(T1G, T1H);
2318 T1J = LD(&(xi[WS(is, 110)]), ivs, &(xi[0]));
2319 T1K = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
2320 T1L = VSUB(T1J, T1K);
2321 TaN = VADD(T1J, T1K);
2322 }
2323 T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L));
2324 TaO = VADD(TaM, TaN);
2325 T24 = VMUL(LDK(KP707106781), VSUB(T1L, T1I));
2326 {
2327 V T1Q, TaT, T1T, TaU;
2328 {
2329 V T1O, T1P, T1R, T1S;
2330 T1O = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
2331 T1P = LD(&(xi[WS(is, 70)]), ivs, &(xi[0]));
2332 T1Q = VSUB(T1O, T1P);
2333 TaT = VADD(T1O, T1P);
2334 T1R = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
2335 T1S = LD(&(xi[WS(is, 102)]), ivs, &(xi[0]));
2336 T1T = VSUB(T1R, T1S);
2337 TaU = VADD(T1R, T1S);
2338 }
2339 T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q));
2340 Tew = VSUB(TaT, TaU);
2341 T2a = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T));
2342 TaV = VADD(TaT, TaU);
2343 }
2344 {
2345 V T1X, TaQ, T20, TaR;
2346 {
2347 V T1V, T1W, T1Y, T1Z;
2348 T1V = LD(&(xi[WS(is, 118)]), ivs, &(xi[0]));
2349 T1W = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
2350 T1X = VSUB(T1V, T1W);
2351 TaQ = VADD(T1V, T1W);
2352 T1Y = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
2353 T1Z = LD(&(xi[WS(is, 86)]), ivs, &(xi[0]));
2354 T20 = VSUB(T1Y, T1Z);
2355 TaR = VADD(T1Y, T1Z);
2356 }
2357 T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20));
2358 Tex = VSUB(TaQ, TaR);
2359 T29 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X));
2360 TaS = VADD(TaQ, TaR);
2361 }
2362 Tdm = VADD(TaL, TaO);
2363 Tdn = VADD(TaV, TaS);
2364 TdH = VSUB(Tdm, Tdn);
2365 Tev = VSUB(TaJ, TaK);
2366 Tey = VMUL(LDK(KP707106781), VADD(Tew, Tex));
2367 Tez = VADD(Tev, Tey);
2368 Tgi = VSUB(Tev, Tey);
2369 {
2370 V TeA, TeB, T1N, T22;
2371 TeA = VSUB(TaN, TaM);
2372 TeB = VMUL(LDK(KP707106781), VSUB(Tex, Tew));
2373 TeC = VADD(TeA, TeB);
2374 Tgj = VSUB(TeB, TeA);
2375 T1N = VADD(T1F, T1M);
2376 T22 = VADD(T1U, T21);
2377 T23 = VADD(T1N, T22);
2378 T5N = VSUB(T1N, T22);
2379 }
2380 {
2381 V T28, T2b, T7i, T7j;
2382 T28 = VSUB(T24, T27);
2383 T2b = VSUB(T29, T2a);
2384 T2c = VADD(T28, T2b);
2385 T5M = VSUB(T2b, T28);
2386 T7i = VADD(T27, T24);
2387 T7j = VSUB(T21, T1U);
2388 T7k = VADD(T7i, T7j);
2389 T8X = VSUB(T7j, T7i);
2390 }
2391 {
2392 V TaP, TaW, T7f, T7g;
2393 TaP = VSUB(TaL, TaO);
2394 TaW = VSUB(TaS, TaV);
2395 TaX = VFNMS(LDK(KP382683432), TaW, VMUL(LDK(KP923879532), TaP));
2396 Tcg = VFMA(LDK(KP382683432), TaP, VMUL(LDK(KP923879532), TaW));
2397 T7f = VSUB(T1F, T1M);
2398 T7g = VADD(T2a, T29);
2399 T7h = VADD(T7f, T7g);
2400 T8W = VSUB(T7f, T7g);
2401 }
2402 }
2403 {
2404 V T2J, TeL, T2V, Tb9, T30, TeO, T3c, Tbg, T2Q, TeM, T2S, Tbc, T37, TeP, T39;
2405 V Tbj;
2406 {
2407 V T2H, T2I, Tb7, T2T, T2U, Tb8;
2408 T2H = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
2409 T2I = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)]));
2410 Tb7 = VADD(T2H, T2I);
2411 T2T = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
2412 T2U = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)]));
2413 Tb8 = VADD(T2T, T2U);
2414 T2J = VSUB(T2H, T2I);
2415 TeL = VSUB(Tb7, Tb8);
2416 T2V = VSUB(T2T, T2U);
2417 Tb9 = VADD(Tb7, Tb8);
2418 }
2419 {
2420 V T2Y, T2Z, Tbe, T3a, T3b, Tbf;
2421 T2Y = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)]));
2422 T2Z = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
2423 Tbe = VADD(T2Y, T2Z);
2424 T3a = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
2425 T3b = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)]));
2426 Tbf = VADD(T3a, T3b);
2427 T30 = VSUB(T2Y, T2Z);
2428 TeO = VSUB(Tbe, Tbf);
2429 T3c = VSUB(T3a, T3b);
2430 Tbg = VADD(Tbe, Tbf);
2431 }
2432 {
2433 V T2M, Tba, T2P, Tbb;
2434 {
2435 V T2K, T2L, T2N, T2O;
2436 T2K = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
2437 T2L = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)]));
2438 T2M = VSUB(T2K, T2L);
2439 Tba = VADD(T2K, T2L);
2440 T2N = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)]));
2441 T2O = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
2442 T2P = VSUB(T2N, T2O);
2443 Tbb = VADD(T2N, T2O);
2444 }
2445 T2Q = VMUL(LDK(KP707106781), VADD(T2M, T2P));
2446 TeM = VSUB(Tbb, Tba);
2447 T2S = VMUL(LDK(KP707106781), VSUB(T2P, T2M));
2448 Tbc = VADD(Tba, Tbb);
2449 }
2450 {
2451 V T33, Tbh, T36, Tbi;
2452 {
2453 V T31, T32, T34, T35;
2454 T31 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
2455 T32 = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)]));
2456 T33 = VSUB(T31, T32);
2457 Tbh = VADD(T31, T32);
2458 T34 = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)]));
2459 T35 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
2460 T36 = VSUB(T34, T35);
2461 Tbi = VADD(T34, T35);
2462 }
2463 T37 = VMUL(LDK(KP707106781), VADD(T33, T36));
2464 TeP = VSUB(Tbi, Tbh);
2465 T39 = VMUL(LDK(KP707106781), VSUB(T36, T33));
2466 Tbj = VADD(Tbh, Tbi);
2467 }
2468 {
2469 V Tbd, Tbk, TeN, TeQ;
2470 Tbd = VSUB(Tb9, Tbc);
2471 Tbk = VSUB(Tbg, Tbj);
2472 Tbl = VMUL(LDK(KP707106781), VADD(Tbd, Tbk));
2473 Tbu = VMUL(LDK(KP707106781), VSUB(Tbk, Tbd));
2474 {
2475 V Td9, Tda, TeW, TeX;
2476 Td9 = VADD(Tb9, Tbc);
2477 Tda = VADD(Tbg, Tbj);
2478 Tdb = VADD(Td9, Tda);
2479 TdL = VSUB(Tda, Td9);
2480 TeW = VFNMS(LDK(KP382683432), TeL, VMUL(LDK(KP923879532), TeM));
2481 TeX = VFMA(LDK(KP382683432), TeO, VMUL(LDK(KP923879532), TeP));
2482 TeY = VADD(TeW, TeX);
2483 Tgu = VSUB(TeX, TeW);
2484 }
2485 TeN = VFMA(LDK(KP923879532), TeL, VMUL(LDK(KP382683432), TeM));
2486 TeQ = VFNMS(LDK(KP382683432), TeP, VMUL(LDK(KP923879532), TeO));
2487 TeR = VADD(TeN, TeQ);
2488 Tgq = VSUB(TeQ, TeN);
2489 {
2490 V T7t, T7C, T7w, T7D;
2491 {
2492 V T7r, T7s, T7u, T7v;
2493 T7r = VSUB(T2J, T2Q);
2494 T7s = VADD(T2V, T2S);
2495 T7t = VFMA(LDK(KP831469612), T7r, VMUL(LDK(KP555570233), T7s));
2496 T7C = VFNMS(LDK(KP555570233), T7r, VMUL(LDK(KP831469612), T7s));
2497 T7u = VSUB(T30, T37);
2498 T7v = VADD(T3c, T39);
2499 T7w = VFNMS(LDK(KP555570233), T7v, VMUL(LDK(KP831469612), T7u));
2500 T7D = VFMA(LDK(KP555570233), T7u, VMUL(LDK(KP831469612), T7v));
2501 }
2502 T7x = VADD(T7t, T7w);
2503 T98 = VSUB(T7D, T7C);
2504 T7E = VADD(T7C, T7D);
2505 T94 = VSUB(T7w, T7t);
2506 }
2507 {
2508 V T2X, T3q, T3e, T3r;
2509 {
2510 V T2R, T2W, T38, T3d;
2511 T2R = VADD(T2J, T2Q);
2512 T2W = VSUB(T2S, T2V);
2513 T2X = VFMA(LDK(KP980785280), T2R, VMUL(LDK(KP195090322), T2W));
2514 T3q = VFNMS(LDK(KP195090322), T2R, VMUL(LDK(KP980785280), T2W));
2515 T38 = VADD(T30, T37);
2516 T3d = VSUB(T39, T3c);
2517 T3e = VFNMS(LDK(KP195090322), T3d, VMUL(LDK(KP980785280), T38));
2518 T3r = VFMA(LDK(KP195090322), T38, VMUL(LDK(KP980785280), T3d));
2519 }
2520 T3f = VADD(T2X, T3e);
2521 T5Y = VSUB(T3r, T3q);
2522 T3s = VADD(T3q, T3r);
2523 T5U = VSUB(T3e, T2X);
2524 }
2525 }
2526 }
2527 {
2528 V T3Y, Tf6, T4a, TbG, T4f, Tf9, T4r, TbN, T45, Tf7, T47, TbJ, T4m, Tfa, T4o;
2529 V TbQ;
2530 {
2531 V T3W, T3X, TbE, T48, T49, TbF;
2532 T3W = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
2533 T3X = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)]));
2534 TbE = VADD(T3W, T3X);
2535 T48 = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
2536 T49 = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)]));
2537 TbF = VADD(T48, T49);
2538 T3Y = VSUB(T3W, T3X);
2539 Tf6 = VSUB(TbE, TbF);
2540 T4a = VSUB(T48, T49);
2541 TbG = VADD(TbE, TbF);
2542 }
2543 {
2544 V T4d, T4e, TbL, T4p, T4q, TbM;
2545 T4d = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)]));
2546 T4e = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
2547 TbL = VADD(T4d, T4e);
2548 T4p = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
2549 T4q = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)]));
2550 TbM = VADD(T4p, T4q);
2551 T4f = VSUB(T4d, T4e);
2552 Tf9 = VSUB(TbL, TbM);
2553 T4r = VSUB(T4p, T4q);
2554 TbN = VADD(TbL, TbM);
2555 }
2556 {
2557 V T41, TbH, T44, TbI;
2558 {
2559 V T3Z, T40, T42, T43;
2560 T3Z = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
2561 T40 = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)]));
2562 T41 = VSUB(T3Z, T40);
2563 TbH = VADD(T3Z, T40);
2564 T42 = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)]));
2565 T43 = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
2566 T44 = VSUB(T42, T43);
2567 TbI = VADD(T42, T43);
2568 }
2569 T45 = VMUL(LDK(KP707106781), VADD(T41, T44));
2570 Tf7 = VSUB(TbI, TbH);
2571 T47 = VMUL(LDK(KP707106781), VSUB(T44, T41));
2572 TbJ = VADD(TbH, TbI);
2573 }
2574 {
2575 V T4i, TbO, T4l, TbP;
2576 {
2577 V T4g, T4h, T4j, T4k;
2578 T4g = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
2579 T4h = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)]));
2580 T4i = VSUB(T4g, T4h);
2581 TbO = VADD(T4g, T4h);
2582 T4j = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)]));
2583 T4k = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
2584 T4l = VSUB(T4j, T4k);
2585 TbP = VADD(T4j, T4k);
2586 }
2587 T4m = VMUL(LDK(KP707106781), VADD(T4i, T4l));
2588 Tfa = VSUB(TbP, TbO);
2589 T4o = VMUL(LDK(KP707106781), VSUB(T4l, T4i));
2590 TbQ = VADD(TbO, TbP);
2591 }
2592 {
2593 V TbK, TbR, Tf8, Tfb;
2594 TbK = VSUB(TbG, TbJ);
2595 TbR = VSUB(TbN, TbQ);
2596 TbS = VMUL(LDK(KP707106781), VADD(TbK, TbR));
2597 Tc1 = VMUL(LDK(KP707106781), VSUB(TbR, TbK));
2598 {
2599 V Tdg, Tdh, Tfh, Tfi;
2600 Tdg = VADD(TbG, TbJ);
2601 Tdh = VADD(TbN, TbQ);
2602 Tdi = VADD(Tdg, Tdh);
2603 TdO = VSUB(Tdh, Tdg);
2604 Tfh = VFNMS(LDK(KP382683432), Tf6, VMUL(LDK(KP923879532), Tf7));
2605 Tfi = VFMA(LDK(KP382683432), Tf9, VMUL(LDK(KP923879532), Tfa));
2606 Tfj = VADD(Tfh, Tfi);
2607 Tgy = VSUB(Tfi, Tfh);
2608 }
2609 Tf8 = VFMA(LDK(KP923879532), Tf6, VMUL(LDK(KP382683432), Tf7));
2610 Tfb = VFNMS(LDK(KP382683432), Tfa, VMUL(LDK(KP923879532), Tf9));
2611 Tfc = VADD(Tf8, Tfb);
2612 TgA = VSUB(Tfb, Tf8);
2613 {
2614 V T7M, T7V, T7P, T7W;
2615 {
2616 V T7K, T7L, T7N, T7O;
2617 T7K = VSUB(T3Y, T45);
2618 T7L = VADD(T4a, T47);
2619 T7M = VFMA(LDK(KP831469612), T7K, VMUL(LDK(KP555570233), T7L));
2620 T7V = VFNMS(LDK(KP555570233), T7K, VMUL(LDK(KP831469612), T7L));
2621 T7N = VSUB(T4f, T4m);
2622 T7O = VADD(T4r, T4o);
2623 T7P = VFNMS(LDK(KP555570233), T7O, VMUL(LDK(KP831469612), T7N));
2624 T7W = VFMA(LDK(KP555570233), T7N, VMUL(LDK(KP831469612), T7O));
2625 }
2626 T7Q = VADD(T7M, T7P);
2627 T9e = VSUB(T7P, T7M);
2628 T7X = VADD(T7V, T7W);
2629 T9c = VSUB(T7W, T7V);
2630 }
2631 {
2632 V T4c, T4F, T4t, T4G;
2633 {
2634 V T46, T4b, T4n, T4s;
2635 T46 = VADD(T3Y, T45);
2636 T4b = VSUB(T47, T4a);
2637 T4c = VFMA(LDK(KP980785280), T46, VMUL(LDK(KP195090322), T4b));
2638 T4F = VFNMS(LDK(KP195090322), T46, VMUL(LDK(KP980785280), T4b));
2639 T4n = VADD(T4f, T4m);
2640 T4s = VSUB(T4o, T4r);
2641 T4t = VFNMS(LDK(KP195090322), T4s, VMUL(LDK(KP980785280), T4n));
2642 T4G = VFMA(LDK(KP195090322), T4n, VMUL(LDK(KP980785280), T4s));
2643 }
2644 T4u = VADD(T4c, T4t);
2645 T64 = VSUB(T4t, T4c);
2646 T4H = VADD(T4F, T4G);
2647 T62 = VSUB(T4G, T4F);
2648 }
2649 }
2650 }
2651 {
2652 V Td5, Tdx, TdC, TdE, Tdk, Tdt, Tds, Tdy, Tdz, TdD;
2653 {
2654 V Td1, Td4, TdA, TdB;
2655 Td1 = VADD(TcZ, Td0);
2656 Td4 = VADD(Td2, Td3);
2657 Td5 = VSUB(Td1, Td4);
2658 Tdx = VADD(Td1, Td4);
2659 TdA = VADD(Td8, Tdb);
2660 TdB = VADD(Tdf, Tdi);
2661 TdC = VADD(TdA, TdB);
2662 TdE = VBYI(VSUB(TdB, TdA));
2663 }
2664 {
2665 V Tdc, Tdj, Tdo, Tdr;
2666 Tdc = VSUB(Td8, Tdb);
2667 Tdj = VSUB(Tdf, Tdi);
2668 Tdk = VMUL(LDK(KP707106781), VADD(Tdc, Tdj));
2669 Tdt = VMUL(LDK(KP707106781), VSUB(Tdj, Tdc));
2670 Tdo = VADD(Tdm, Tdn);
2671 Tdr = VADD(Tdp, Tdq);
2672 Tds = VSUB(Tdo, Tdr);
2673 Tdy = VADD(Tdr, Tdo);
2674 }
2675 Tdz = VADD(Tdx, Tdy);
2676 ST(&(xo[WS(os, 64)]), VSUB(Tdz, TdC), ovs, &(xo[0]));
2677 ST(&(xo[0]), VADD(Tdz, TdC), ovs, &(xo[0]));
2678 TdD = VSUB(Tdx, Tdy);
2679 ST(&(xo[WS(os, 96)]), VSUB(TdD, TdE), ovs, &(xo[0]));
2680 ST(&(xo[WS(os, 32)]), VADD(TdD, TdE), ovs, &(xo[0]));
2681 {
2682 V Tdl, Tdu, Tdv, Tdw;
2683 Tdl = VADD(Td5, Tdk);
2684 Tdu = VBYI(VADD(Tds, Tdt));
2685 ST(&(xo[WS(os, 112)]), VSUB(Tdl, Tdu), ovs, &(xo[0]));
2686 ST(&(xo[WS(os, 16)]), VADD(Tdl, Tdu), ovs, &(xo[0]));
2687 Tdv = VSUB(Td5, Tdk);
2688 Tdw = VBYI(VSUB(Tdt, Tds));
2689 ST(&(xo[WS(os, 80)]), VSUB(Tdv, Tdw), ovs, &(xo[0]));
2690 ST(&(xo[WS(os, 48)]), VADD(Tdv, Tdw), ovs, &(xo[0]));
2691 }
2692 }
2693 {
2694 V TdJ, Te4, TdX, Te5, TdQ, Te1, TdU, Te2;
2695 {
2696 V TdF, TdI, TdV, TdW;
2697 TdF = VSUB(TcZ, Td0);
2698 TdI = VMUL(LDK(KP707106781), VADD(TdG, TdH));
2699 TdJ = VADD(TdF, TdI);
2700 Te4 = VSUB(TdF, TdI);
2701 TdV = VFNMS(LDK(KP382683432), TdK, VMUL(LDK(KP923879532), TdL));
2702 TdW = VFMA(LDK(KP382683432), TdN, VMUL(LDK(KP923879532), TdO));
2703 TdX = VADD(TdV, TdW);
2704 Te5 = VSUB(TdW, TdV);
2705 }
2706 {
2707 V TdM, TdP, TdS, TdT;
2708 TdM = VFMA(LDK(KP923879532), TdK, VMUL(LDK(KP382683432), TdL));
2709 TdP = VFNMS(LDK(KP382683432), TdO, VMUL(LDK(KP923879532), TdN));
2710 TdQ = VADD(TdM, TdP);
2711 Te1 = VSUB(TdP, TdM);
2712 TdS = VSUB(Td3, Td2);
2713 TdT = VMUL(LDK(KP707106781), VSUB(TdH, TdG));
2714 TdU = VADD(TdS, TdT);
2715 Te2 = VSUB(TdT, TdS);
2716 }
2717 {
2718 V TdR, TdY, Te7, Te8;
2719 TdR = VADD(TdJ, TdQ);
2720 TdY = VBYI(VADD(TdU, TdX));
2721 ST(&(xo[WS(os, 120)]), VSUB(TdR, TdY), ovs, &(xo[0]));
2722 ST(&(xo[WS(os, 8)]), VADD(TdR, TdY), ovs, &(xo[0]));
2723 Te7 = VBYI(VADD(Te2, Te1));
2724 Te8 = VADD(Te4, Te5);
2725 ST(&(xo[WS(os, 24)]), VADD(Te7, Te8), ovs, &(xo[0]));
2726 ST(&(xo[WS(os, 104)]), VSUB(Te8, Te7), ovs, &(xo[0]));
2727 }
2728 {
2729 V TdZ, Te0, Te3, Te6;
2730 TdZ = VSUB(TdJ, TdQ);
2731 Te0 = VBYI(VSUB(TdX, TdU));
2732 ST(&(xo[WS(os, 72)]), VSUB(TdZ, Te0), ovs, &(xo[0]));
2733 ST(&(xo[WS(os, 56)]), VADD(TdZ, Te0), ovs, &(xo[0]));
2734 Te3 = VBYI(VSUB(Te1, Te2));
2735 Te6 = VSUB(Te4, Te5);
2736 ST(&(xo[WS(os, 40)]), VADD(Te3, Te6), ovs, &(xo[0]));
2737 ST(&(xo[WS(os, 88)]), VSUB(Te6, Te3), ovs, &(xo[0]));
2738 }
2739 }
2740 {
2741 V TaZ, Tcs, Tci, Tcq, Tc4, Tct, Tcl, Tcp;
2742 {
2743 V Tat, TaY, Tce, Tch;
2744 Tat = VADD(Tad, Tas);
2745 TaY = VADD(TaI, TaX);
2746 TaZ = VADD(Tat, TaY);
2747 Tcs = VSUB(Tat, TaY);
2748 Tce = VADD(Tcc, Tcd);
2749 Tch = VADD(Tcf, Tcg);
2750 Tci = VADD(Tce, Tch);
2751 Tcq = VSUB(Tch, Tce);
2752 {
2753 V Tbw, Tcj, Tc3, Tck;
2754 {
2755 V Tbm, Tbv, TbT, Tc2;
2756 Tbm = VADD(Tb6, Tbl);
2757 Tbv = VADD(Tbt, Tbu);
2758 Tbw = VFMA(LDK(KP980785280), Tbm, VMUL(LDK(KP195090322), Tbv));
2759 Tcj = VFNMS(LDK(KP195090322), Tbm, VMUL(LDK(KP980785280), Tbv));
2760 TbT = VADD(TbD, TbS);
2761 Tc2 = VADD(Tc0, Tc1);
2762 Tc3 = VFNMS(LDK(KP195090322), Tc2, VMUL(LDK(KP980785280), TbT));
2763 Tck = VFMA(LDK(KP195090322), TbT, VMUL(LDK(KP980785280), Tc2));
2764 }
2765 Tc4 = VADD(Tbw, Tc3);
2766 Tct = VSUB(Tck, Tcj);
2767 Tcl = VADD(Tcj, Tck);
2768 Tcp = VSUB(Tc3, Tbw);
2769 }
2770 }
2771 {
2772 V Tc5, Tcm, Tcv, Tcw;
2773 Tc5 = VADD(TaZ, Tc4);
2774 Tcm = VBYI(VADD(Tci, Tcl));
2775 ST(&(xo[WS(os, 124)]), VSUB(Tc5, Tcm), ovs, &(xo[0]));
2776 ST(&(xo[WS(os, 4)]), VADD(Tc5, Tcm), ovs, &(xo[0]));
2777 Tcv = VBYI(VADD(Tcq, Tcp));
2778 Tcw = VADD(Tcs, Tct);
2779 ST(&(xo[WS(os, 28)]), VADD(Tcv, Tcw), ovs, &(xo[0]));
2780 ST(&(xo[WS(os, 100)]), VSUB(Tcw, Tcv), ovs, &(xo[0]));
2781 }
2782 {
2783 V Tcn, Tco, Tcr, Tcu;
2784 Tcn = VSUB(TaZ, Tc4);
2785 Tco = VBYI(VSUB(Tcl, Tci));
2786 ST(&(xo[WS(os, 68)]), VSUB(Tcn, Tco), ovs, &(xo[0]));
2787 ST(&(xo[WS(os, 60)]), VADD(Tcn, Tco), ovs, &(xo[0]));
2788 Tcr = VBYI(VSUB(Tcp, Tcq));
2789 Tcu = VSUB(Tcs, Tct);
2790 ST(&(xo[WS(os, 36)]), VADD(Tcr, Tcu), ovs, &(xo[0]));
2791 ST(&(xo[WS(os, 92)]), VSUB(Tcu, Tcr), ovs, &(xo[0]));
2792 }
2793 }
2794 {
2795 V Tcz, TcU, TcK, TcS, TcG, TcV, TcN, TcR;
2796 {
2797 V Tcx, Tcy, TcI, TcJ;
2798 Tcx = VSUB(Tad, Tas);
2799 Tcy = VSUB(Tcg, Tcf);
2800 Tcz = VADD(Tcx, Tcy);
2801 TcU = VSUB(Tcx, Tcy);
2802 TcI = VSUB(Tcd, Tcc);
2803 TcJ = VSUB(TaX, TaI);
2804 TcK = VADD(TcI, TcJ);
2805 TcS = VSUB(TcJ, TcI);
2806 {
2807 V TcC, TcL, TcF, TcM;
2808 {
2809 V TcA, TcB, TcD, TcE;
2810 TcA = VSUB(Tb6, Tbl);
2811 TcB = VSUB(Tbu, Tbt);
2812 TcC = VFMA(LDK(KP831469612), TcA, VMUL(LDK(KP555570233), TcB));
2813 TcL = VFNMS(LDK(KP555570233), TcA, VMUL(LDK(KP831469612), TcB));
2814 TcD = VSUB(TbD, TbS);
2815 TcE = VSUB(Tc1, Tc0);
2816 TcF = VFNMS(LDK(KP555570233), TcE, VMUL(LDK(KP831469612), TcD));
2817 TcM = VFMA(LDK(KP555570233), TcD, VMUL(LDK(KP831469612), TcE));
2818 }
2819 TcG = VADD(TcC, TcF);
2820 TcV = VSUB(TcM, TcL);
2821 TcN = VADD(TcL, TcM);
2822 TcR = VSUB(TcF, TcC);
2823 }
2824 }
2825 {
2826 V TcH, TcO, TcX, TcY;
2827 TcH = VADD(Tcz, TcG);
2828 TcO = VBYI(VADD(TcK, TcN));
2829 ST(&(xo[WS(os, 116)]), VSUB(TcH, TcO), ovs, &(xo[0]));
2830 ST(&(xo[WS(os, 12)]), VADD(TcH, TcO), ovs, &(xo[0]));
2831 TcX = VBYI(VADD(TcS, TcR));
2832 TcY = VADD(TcU, TcV);
2833 ST(&(xo[WS(os, 20)]), VADD(TcX, TcY), ovs, &(xo[0]));
2834 ST(&(xo[WS(os, 108)]), VSUB(TcY, TcX), ovs, &(xo[0]));
2835 }
2836 {
2837 V TcP, TcQ, TcT, TcW;
2838 TcP = VSUB(Tcz, TcG);
2839 TcQ = VBYI(VSUB(TcN, TcK));
2840 ST(&(xo[WS(os, 76)]), VSUB(TcP, TcQ), ovs, &(xo[0]));
2841 ST(&(xo[WS(os, 52)]), VADD(TcP, TcQ), ovs, &(xo[0]));
2842 TcT = VBYI(VSUB(TcR, TcS));
2843 TcW = VSUB(TcU, TcV);
2844 ST(&(xo[WS(os, 44)]), VADD(TcT, TcW), ovs, &(xo[0]));
2845 ST(&(xo[WS(os, 84)]), VSUB(TcW, TcT), ovs, &(xo[0]));
2846 }
2847 }
2848 {
2849 V TeF, Tg8, TfI, Tg0, Tfy, Tga, TfG, TfP, Tfm, TfJ, TfB, TfF, TfW, Tgb, Tg3;
2850 V Tg7;
2851 {
2852 V Tel, TfY, TeE, TfZ, Teu, TeD;
2853 Tel = VADD(Ted, Tek);
2854 TfY = VSUB(Tft, Tfq);
2855 Teu = VFMA(LDK(KP980785280), Teq, VMUL(LDK(KP195090322), Tet));
2856 TeD = VFNMS(LDK(KP195090322), TeC, VMUL(LDK(KP980785280), Tez));
2857 TeE = VADD(Teu, TeD);
2858 TfZ = VSUB(TeD, Teu);
2859 TeF = VADD(Tel, TeE);
2860 Tg8 = VSUB(TfZ, TfY);
2861 TfI = VSUB(Tel, TeE);
2862 Tg0 = VADD(TfY, TfZ);
2863 }
2864 {
2865 V Tfu, TfN, Tfx, TfO, Tfv, Tfw;
2866 Tfu = VADD(Tfq, Tft);
2867 TfN = VSUB(Ted, Tek);
2868 Tfv = VFNMS(LDK(KP195090322), Teq, VMUL(LDK(KP980785280), Tet));
2869 Tfw = VFMA(LDK(KP195090322), Tez, VMUL(LDK(KP980785280), TeC));
2870 Tfx = VADD(Tfv, Tfw);
2871 TfO = VSUB(Tfw, Tfv);
2872 Tfy = VADD(Tfu, Tfx);
2873 Tga = VSUB(TfN, TfO);
2874 TfG = VSUB(Tfx, Tfu);
2875 TfP = VADD(TfN, TfO);
2876 }
2877 {
2878 V Tf0, Tfz, Tfl, TfA;
2879 {
2880 V TeS, TeZ, Tfd, Tfk;
2881 TeS = VADD(TeK, TeR);
2882 TeZ = VADD(TeV, TeY);
2883 Tf0 = VFMA(LDK(KP995184726), TeS, VMUL(LDK(KP098017140), TeZ));
2884 Tfz = VFNMS(LDK(KP098017140), TeS, VMUL(LDK(KP995184726), TeZ));
2885 Tfd = VADD(Tf5, Tfc);
2886 Tfk = VADD(Tfg, Tfj);
2887 Tfl = VFNMS(LDK(KP098017140), Tfk, VMUL(LDK(KP995184726), Tfd));
2888 TfA = VFMA(LDK(KP098017140), Tfd, VMUL(LDK(KP995184726), Tfk));
2889 }
2890 Tfm = VADD(Tf0, Tfl);
2891 TfJ = VSUB(TfA, Tfz);
2892 TfB = VADD(Tfz, TfA);
2893 TfF = VSUB(Tfl, Tf0);
2894 }
2895 {
2896 V TfS, Tg1, TfV, Tg2;
2897 {
2898 V TfQ, TfR, TfT, TfU;
2899 TfQ = VSUB(TeK, TeR);
2900 TfR = VSUB(TeY, TeV);
2901 TfS = VFMA(LDK(KP773010453), TfQ, VMUL(LDK(KP634393284), TfR));
2902 Tg1 = VFNMS(LDK(KP634393284), TfQ, VMUL(LDK(KP773010453), TfR));
2903 TfT = VSUB(Tf5, Tfc);
2904 TfU = VSUB(Tfj, Tfg);
2905 TfV = VFNMS(LDK(KP634393284), TfU, VMUL(LDK(KP773010453), TfT));
2906 Tg2 = VFMA(LDK(KP634393284), TfT, VMUL(LDK(KP773010453), TfU));
2907 }
2908 TfW = VADD(TfS, TfV);
2909 Tgb = VSUB(Tg2, Tg1);
2910 Tg3 = VADD(Tg1, Tg2);
2911 Tg7 = VSUB(TfV, TfS);
2912 }
2913 {
2914 V Tfn, TfC, Tg9, Tgc;
2915 Tfn = VADD(TeF, Tfm);
2916 TfC = VBYI(VADD(Tfy, TfB));
2917 ST(&(xo[WS(os, 126)]), VSUB(Tfn, TfC), ovs, &(xo[0]));
2918 ST(&(xo[WS(os, 2)]), VADD(Tfn, TfC), ovs, &(xo[0]));
2919 Tg9 = VBYI(VSUB(Tg7, Tg8));
2920 Tgc = VSUB(Tga, Tgb);
2921 ST(&(xo[WS(os, 46)]), VADD(Tg9, Tgc), ovs, &(xo[0]));
2922 ST(&(xo[WS(os, 82)]), VSUB(Tgc, Tg9), ovs, &(xo[0]));
2923 }
2924 {
2925 V Tgd, Tge, TfD, TfE;
2926 Tgd = VBYI(VADD(Tg8, Tg7));
2927 Tge = VADD(Tga, Tgb);
2928 ST(&(xo[WS(os, 18)]), VADD(Tgd, Tge), ovs, &(xo[0]));
2929 ST(&(xo[WS(os, 110)]), VSUB(Tge, Tgd), ovs, &(xo[0]));
2930 TfD = VSUB(TeF, Tfm);
2931 TfE = VBYI(VSUB(TfB, Tfy));
2932 ST(&(xo[WS(os, 66)]), VSUB(TfD, TfE), ovs, &(xo[0]));
2933 ST(&(xo[WS(os, 62)]), VADD(TfD, TfE), ovs, &(xo[0]));
2934 }
2935 {
2936 V TfH, TfK, TfX, Tg4;
2937 TfH = VBYI(VSUB(TfF, TfG));
2938 TfK = VSUB(TfI, TfJ);
2939 ST(&(xo[WS(os, 34)]), VADD(TfH, TfK), ovs, &(xo[0]));
2940 ST(&(xo[WS(os, 94)]), VSUB(TfK, TfH), ovs, &(xo[0]));
2941 TfX = VADD(TfP, TfW);
2942 Tg4 = VBYI(VADD(Tg0, Tg3));
2943 ST(&(xo[WS(os, 114)]), VSUB(TfX, Tg4), ovs, &(xo[0]));
2944 ST(&(xo[WS(os, 14)]), VADD(TfX, Tg4), ovs, &(xo[0]));
2945 }
2946 {
2947 V Tg5, Tg6, TfL, TfM;
2948 Tg5 = VSUB(TfP, TfW);
2949 Tg6 = VBYI(VSUB(Tg3, Tg0));
2950 ST(&(xo[WS(os, 78)]), VSUB(Tg5, Tg6), ovs, &(xo[0]));
2951 ST(&(xo[WS(os, 50)]), VADD(Tg5, Tg6), ovs, &(xo[0]));
2952 TfL = VBYI(VADD(TfG, TfF));
2953 TfM = VADD(TfI, TfJ);
2954 ST(&(xo[WS(os, 30)]), VADD(TfL, TfM), ovs, &(xo[0]));
2955 ST(&(xo[WS(os, 98)]), VSUB(TfM, TfL), ovs, &(xo[0]));
2956 }
2957 }
2958 {
2959 V Tgp, Thm, TgW, The, TgM, Tho, TgU, Th3, TgE, TgX, TgP, TgT, Tha, Thp, Thh;
2960 V Thl;
2961 {
2962 V Tgh, Thc, Tgo, Thd, Tgk, Tgn;
2963 Tgh = VSUB(Tgf, Tgg);
2964 Thc = VADD(TgH, TgG);
2965 Tgk = VFMA(LDK(KP555570233), Tgi, VMUL(LDK(KP831469612), Tgj));
2966 Tgn = VFNMS(LDK(KP555570233), Tgm, VMUL(LDK(KP831469612), Tgl));
2967 Tgo = VSUB(Tgk, Tgn);
2968 Thd = VADD(Tgn, Tgk);
2969 Tgp = VADD(Tgh, Tgo);
2970 Thm = VSUB(Thd, Thc);
2971 TgW = VSUB(Tgh, Tgo);
2972 The = VADD(Thc, Thd);
2973 }
2974 {
2975 V TgI, Th1, TgL, Th2, TgJ, TgK;
2976 TgI = VSUB(TgG, TgH);
2977 Th1 = VADD(Tgf, Tgg);
2978 TgJ = VFNMS(LDK(KP555570233), Tgj, VMUL(LDK(KP831469612), Tgi));
2979 TgK = VFMA(LDK(KP831469612), Tgm, VMUL(LDK(KP555570233), Tgl));
2980 TgL = VSUB(TgJ, TgK);
2981 Th2 = VADD(TgK, TgJ);
2982 TgM = VADD(TgI, TgL);
2983 Tho = VSUB(Th1, Th2);
2984 TgU = VSUB(TgL, TgI);
2985 Th3 = VADD(Th1, Th2);
2986 }
2987 {
2988 V Tgw, TgN, TgD, TgO;
2989 {
2990 V Tgs, Tgv, Tgz, TgC;
2991 Tgs = VSUB(Tgq, Tgr);
2992 Tgv = VSUB(Tgt, Tgu);
2993 Tgw = VFMA(LDK(KP471396736), Tgs, VMUL(LDK(KP881921264), Tgv));
2994 TgN = VFNMS(LDK(KP471396736), Tgv, VMUL(LDK(KP881921264), Tgs));
2995 Tgz = VSUB(Tgx, Tgy);
2996 TgC = VSUB(TgA, TgB);
2997 TgD = VFNMS(LDK(KP471396736), TgC, VMUL(LDK(KP881921264), Tgz));
2998 TgO = VFMA(LDK(KP881921264), TgC, VMUL(LDK(KP471396736), Tgz));
2999 }
3000 TgE = VADD(Tgw, TgD);
3001 TgX = VSUB(TgO, TgN);
3002 TgP = VADD(TgN, TgO);
3003 TgT = VSUB(TgD, Tgw);
3004 }
3005 {
3006 V Th6, Thf, Th9, Thg;
3007 {
3008 V Th4, Th5, Th7, Th8;
3009 Th4 = VADD(Tgr, Tgq);
3010 Th5 = VADD(Tgt, Tgu);
3011 Th6 = VFMA(LDK(KP290284677), Th4, VMUL(LDK(KP956940335), Th5));
3012 Thf = VFNMS(LDK(KP290284677), Th5, VMUL(LDK(KP956940335), Th4));
3013 Th7 = VADD(Tgx, Tgy);
3014 Th8 = VADD(TgB, TgA);
3015 Th9 = VFNMS(LDK(KP290284677), Th8, VMUL(LDK(KP956940335), Th7));
3016 Thg = VFMA(LDK(KP956940335), Th8, VMUL(LDK(KP290284677), Th7));
3017 }
3018 Tha = VADD(Th6, Th9);
3019 Thp = VSUB(Thg, Thf);
3020 Thh = VADD(Thf, Thg);
3021 Thl = VSUB(Th9, Th6);
3022 }
3023 {
3024 V TgF, TgQ, Thn, Thq;
3025 TgF = VADD(Tgp, TgE);
3026 TgQ = VBYI(VADD(TgM, TgP));
3027 ST(&(xo[WS(os, 118)]), VSUB(TgF, TgQ), ovs, &(xo[0]));
3028 ST(&(xo[WS(os, 10)]), VADD(TgF, TgQ), ovs, &(xo[0]));
3029 Thn = VBYI(VSUB(Thl, Thm));
3030 Thq = VSUB(Tho, Thp);
3031 ST(&(xo[WS(os, 38)]), VADD(Thn, Thq), ovs, &(xo[0]));
3032 ST(&(xo[WS(os, 90)]), VSUB(Thq, Thn), ovs, &(xo[0]));
3033 }
3034 {
3035 V Thr, Ths, TgR, TgS;
3036 Thr = VBYI(VADD(Thm, Thl));
3037 Ths = VADD(Tho, Thp);
3038 ST(&(xo[WS(os, 26)]), VADD(Thr, Ths), ovs, &(xo[0]));
3039 ST(&(xo[WS(os, 102)]), VSUB(Ths, Thr), ovs, &(xo[0]));
3040 TgR = VSUB(Tgp, TgE);
3041 TgS = VBYI(VSUB(TgP, TgM));
3042 ST(&(xo[WS(os, 74)]), VSUB(TgR, TgS), ovs, &(xo[0]));
3043 ST(&(xo[WS(os, 54)]), VADD(TgR, TgS), ovs, &(xo[0]));
3044 }
3045 {
3046 V TgV, TgY, Thb, Thi;
3047 TgV = VBYI(VSUB(TgT, TgU));
3048 TgY = VSUB(TgW, TgX);
3049 ST(&(xo[WS(os, 42)]), VADD(TgV, TgY), ovs, &(xo[0]));
3050 ST(&(xo[WS(os, 86)]), VSUB(TgY, TgV), ovs, &(xo[0]));
3051 Thb = VADD(Th3, Tha);
3052 Thi = VBYI(VADD(The, Thh));
3053 ST(&(xo[WS(os, 122)]), VSUB(Thb, Thi), ovs, &(xo[0]));
3054 ST(&(xo[WS(os, 6)]), VADD(Thb, Thi), ovs, &(xo[0]));
3055 }
3056 {
3057 V Thj, Thk, TgZ, Th0;
3058 Thj = VSUB(Th3, Tha);
3059 Thk = VBYI(VSUB(Thh, The));
3060 ST(&(xo[WS(os, 70)]), VSUB(Thj, Thk), ovs, &(xo[0]));
3061 ST(&(xo[WS(os, 58)]), VADD(Thj, Thk), ovs, &(xo[0]));
3062 TgZ = VBYI(VADD(TgU, TgT));
3063 Th0 = VADD(TgW, TgX);
3064 ST(&(xo[WS(os, 22)]), VADD(TgZ, Th0), ovs, &(xo[0]));
3065 ST(&(xo[WS(os, 106)]), VSUB(Th0, TgZ), ovs, &(xo[0]));
3066 }
3067 }
3068 {
3069 V T80, T8n, T8f, T8j, T8A, T8P, T8H, T8L, T7n, T8M, T8O, T8c, T8k, T8t, T8E;
3070 V T8m;
3071 {
3072 V T7G, T8d, T7Z, T8e;
3073 {
3074 V T7y, T7F, T7R, T7Y;
3075 T7y = VADD(T7q, T7x);
3076 T7F = VADD(T7B, T7E);
3077 T7G = VFMA(LDK(KP989176509), T7y, VMUL(LDK(KP146730474), T7F));
3078 T8d = VFNMS(LDK(KP146730474), T7y, VMUL(LDK(KP989176509), T7F));
3079 T7R = VADD(T7J, T7Q);
3080 T7Y = VADD(T7U, T7X);
3081 T7Z = VFNMS(LDK(KP146730474), T7Y, VMUL(LDK(KP989176509), T7R));
3082 T8e = VFMA(LDK(KP146730474), T7R, VMUL(LDK(KP989176509), T7Y));
3083 }
3084 T80 = VADD(T7G, T7Z);
3085 T8n = VSUB(T8e, T8d);
3086 T8f = VADD(T8d, T8e);
3087 T8j = VSUB(T7Z, T7G);
3088 }
3089 {
3090 V T8w, T8F, T8z, T8G;
3091 {
3092 V T8u, T8v, T8x, T8y;
3093 T8u = VSUB(T7q, T7x);
3094 T8v = VSUB(T7E, T7B);
3095 T8w = VFMA(LDK(KP803207531), T8u, VMUL(LDK(KP595699304), T8v));
3096 T8F = VFNMS(LDK(KP595699304), T8u, VMUL(LDK(KP803207531), T8v));
3097 T8x = VSUB(T7J, T7Q);
3098 T8y = VSUB(T7X, T7U);
3099 T8z = VFNMS(LDK(KP595699304), T8y, VMUL(LDK(KP803207531), T8x));
3100 T8G = VFMA(LDK(KP595699304), T8x, VMUL(LDK(KP803207531), T8y));
3101 }
3102 T8A = VADD(T8w, T8z);
3103 T8P = VSUB(T8G, T8F);
3104 T8H = VADD(T8F, T8G);
3105 T8L = VSUB(T8z, T8w);
3106 }
3107 {
3108 V T77, T8r, T88, T8C, T7m, T8D, T8b, T8s, T76, T87;
3109 T76 = VADD(T72, T75);
3110 T77 = VADD(T6Z, T76);
3111 T8r = VSUB(T6Z, T76);
3112 T87 = VADD(T85, T86);
3113 T88 = VADD(T84, T87);
3114 T8C = VSUB(T87, T84);
3115 {
3116 V T7e, T7l, T89, T8a;
3117 T7e = VFMA(LDK(KP956940335), T7a, VMUL(LDK(KP290284677), T7d));
3118 T7l = VFNMS(LDK(KP290284677), T7k, VMUL(LDK(KP956940335), T7h));
3119 T7m = VADD(T7e, T7l);
3120 T8D = VSUB(T7l, T7e);
3121 T89 = VFNMS(LDK(KP290284677), T7a, VMUL(LDK(KP956940335), T7d));
3122 T8a = VFMA(LDK(KP290284677), T7h, VMUL(LDK(KP956940335), T7k));
3123 T8b = VADD(T89, T8a);
3124 T8s = VSUB(T8a, T89);
3125 }
3126 T7n = VADD(T77, T7m);
3127 T8M = VSUB(T8D, T8C);
3128 T8O = VSUB(T8r, T8s);
3129 T8c = VADD(T88, T8b);
3130 T8k = VSUB(T8b, T88);
3131 T8t = VADD(T8r, T8s);
3132 T8E = VADD(T8C, T8D);
3133 T8m = VSUB(T77, T7m);
3134 }
3135 {
3136 V T81, T8g, T8N, T8Q;
3137 T81 = VADD(T7n, T80);
3138 T8g = VBYI(VADD(T8c, T8f));
3139 ST(&(xo[WS(os, 125)]), VSUB(T81, T8g), ovs, &(xo[WS(os, 1)]));
3140 ST(&(xo[WS(os, 3)]), VADD(T81, T8g), ovs, &(xo[WS(os, 1)]));
3141 T8N = VBYI(VSUB(T8L, T8M));
3142 T8Q = VSUB(T8O, T8P);
3143 ST(&(xo[WS(os, 45)]), VADD(T8N, T8Q), ovs, &(xo[WS(os, 1)]));
3144 ST(&(xo[WS(os, 83)]), VSUB(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
3145 }
3146 {
3147 V T8R, T8S, T8h, T8i;
3148 T8R = VBYI(VADD(T8M, T8L));
3149 T8S = VADD(T8O, T8P);
3150 ST(&(xo[WS(os, 19)]), VADD(T8R, T8S), ovs, &(xo[WS(os, 1)]));
3151 ST(&(xo[WS(os, 109)]), VSUB(T8S, T8R), ovs, &(xo[WS(os, 1)]));
3152 T8h = VSUB(T7n, T80);
3153 T8i = VBYI(VSUB(T8f, T8c));
3154 ST(&(xo[WS(os, 67)]), VSUB(T8h, T8i), ovs, &(xo[WS(os, 1)]));
3155 ST(&(xo[WS(os, 61)]), VADD(T8h, T8i), ovs, &(xo[WS(os, 1)]));
3156 }
3157 {
3158 V T8l, T8o, T8B, T8I;
3159 T8l = VBYI(VSUB(T8j, T8k));
3160 T8o = VSUB(T8m, T8n);
3161 ST(&(xo[WS(os, 35)]), VADD(T8l, T8o), ovs, &(xo[WS(os, 1)]));
3162 ST(&(xo[WS(os, 93)]), VSUB(T8o, T8l), ovs, &(xo[WS(os, 1)]));
3163 T8B = VADD(T8t, T8A);
3164 T8I = VBYI(VADD(T8E, T8H));
3165 ST(&(xo[WS(os, 115)]), VSUB(T8B, T8I), ovs, &(xo[WS(os, 1)]));
3166 ST(&(xo[WS(os, 13)]), VADD(T8B, T8I), ovs, &(xo[WS(os, 1)]));
3167 }
3168 {
3169 V T8J, T8K, T8p, T8q;
3170 T8J = VSUB(T8t, T8A);
3171 T8K = VBYI(VSUB(T8H, T8E));
3172 ST(&(xo[WS(os, 77)]), VSUB(T8J, T8K), ovs, &(xo[WS(os, 1)]));
3173 ST(&(xo[WS(os, 51)]), VADD(T8J, T8K), ovs, &(xo[WS(os, 1)]));
3174 T8p = VBYI(VADD(T8k, T8j));
3175 T8q = VADD(T8m, T8n);
3176 ST(&(xo[WS(os, 29)]), VADD(T8p, T8q), ovs, &(xo[WS(os, 1)]));
3177 ST(&(xo[WS(os, 99)]), VSUB(T8q, T8p), ovs, &(xo[WS(os, 1)]));
3178 }
3179 }
3180 {
3181 V T4K, T5d, T55, T59, T5q, T5F, T5x, T5B, T2f, T5C, T5E, T52, T5a, T5j, T5u;
3182 V T5c;
3183 {
3184 V T3u, T53, T4J, T54;
3185 {
3186 V T3g, T3t, T4v, T4I;
3187 T3g = VADD(T2G, T3f);
3188 T3t = VADD(T3p, T3s);
3189 T3u = VFMA(LDK(KP998795456), T3g, VMUL(LDK(KP049067674), T3t));
3190 T53 = VFNMS(LDK(KP049067674), T3g, VMUL(LDK(KP998795456), T3t));
3191 T4v = VADD(T3V, T4u);
3192 T4I = VADD(T4E, T4H);
3193 T4J = VFNMS(LDK(KP049067674), T4I, VMUL(LDK(KP998795456), T4v));
3194 T54 = VFMA(LDK(KP049067674), T4v, VMUL(LDK(KP998795456), T4I));
3195 }
3196 T4K = VADD(T3u, T4J);
3197 T5d = VSUB(T54, T53);
3198 T55 = VADD(T53, T54);
3199 T59 = VSUB(T4J, T3u);
3200 }
3201 {
3202 V T5m, T5v, T5p, T5w;
3203 {
3204 V T5k, T5l, T5n, T5o;
3205 T5k = VSUB(T2G, T3f);
3206 T5l = VSUB(T3s, T3p);
3207 T5m = VFMA(LDK(KP740951125), T5k, VMUL(LDK(KP671558954), T5l));
3208 T5v = VFNMS(LDK(KP671558954), T5k, VMUL(LDK(KP740951125), T5l));
3209 T5n = VSUB(T3V, T4u);
3210 T5o = VSUB(T4H, T4E);
3211 T5p = VFNMS(LDK(KP671558954), T5o, VMUL(LDK(KP740951125), T5n));
3212 T5w = VFMA(LDK(KP671558954), T5n, VMUL(LDK(KP740951125), T5o));
3213 }
3214 T5q = VADD(T5m, T5p);
3215 T5F = VSUB(T5w, T5v);
3216 T5x = VADD(T5v, T5w);
3217 T5B = VSUB(T5p, T5m);
3218 }
3219 {
3220 V T11, T5h, T4Y, T5s, T2e, T5t, T51, T5i, T10, T4X;
3221 T10 = VADD(TI, TZ);
3222 T11 = VADD(Tr, T10);
3223 T5h = VSUB(Tr, T10);
3224 T4X = VADD(T4V, T4W);
3225 T4Y = VADD(T4U, T4X);
3226 T5s = VSUB(T4X, T4U);
3227 {
3228 V T1C, T2d, T4Z, T50;
3229 T1C = VFMA(LDK(KP098017140), T1s, VMUL(LDK(KP995184726), T1B));
3230 T2d = VFNMS(LDK(KP098017140), T2c, VMUL(LDK(KP995184726), T23));
3231 T2e = VADD(T1C, T2d);
3232 T5t = VSUB(T2d, T1C);
3233 T4Z = VFNMS(LDK(KP098017140), T1B, VMUL(LDK(KP995184726), T1s));
3234 T50 = VFMA(LDK(KP995184726), T2c, VMUL(LDK(KP098017140), T23));
3235 T51 = VADD(T4Z, T50);
3236 T5i = VSUB(T50, T4Z);
3237 }
3238 T2f = VADD(T11, T2e);
3239 T5C = VSUB(T5t, T5s);
3240 T5E = VSUB(T5h, T5i);
3241 T52 = VADD(T4Y, T51);
3242 T5a = VSUB(T51, T4Y);
3243 T5j = VADD(T5h, T5i);
3244 T5u = VADD(T5s, T5t);
3245 T5c = VSUB(T11, T2e);
3246 }
3247 {
3248 V T4L, T56, T5D, T5G;
3249 T4L = VADD(T2f, T4K);
3250 T56 = VBYI(VADD(T52, T55));
3251 ST(&(xo[WS(os, 127)]), VSUB(T4L, T56), ovs, &(xo[WS(os, 1)]));
3252 ST(&(xo[WS(os, 1)]), VADD(T4L, T56), ovs, &(xo[WS(os, 1)]));
3253 T5D = VBYI(VSUB(T5B, T5C));
3254 T5G = VSUB(T5E, T5F);
3255 ST(&(xo[WS(os, 47)]), VADD(T5D, T5G), ovs, &(xo[WS(os, 1)]));
3256 ST(&(xo[WS(os, 81)]), VSUB(T5G, T5D), ovs, &(xo[WS(os, 1)]));
3257 }
3258 {
3259 V T5H, T5I, T57, T58;
3260 T5H = VBYI(VADD(T5C, T5B));
3261 T5I = VADD(T5E, T5F);
3262 ST(&(xo[WS(os, 17)]), VADD(T5H, T5I), ovs, &(xo[WS(os, 1)]));
3263 ST(&(xo[WS(os, 111)]), VSUB(T5I, T5H), ovs, &(xo[WS(os, 1)]));
3264 T57 = VSUB(T2f, T4K);
3265 T58 = VBYI(VSUB(T55, T52));
3266 ST(&(xo[WS(os, 65)]), VSUB(T57, T58), ovs, &(xo[WS(os, 1)]));
3267 ST(&(xo[WS(os, 63)]), VADD(T57, T58), ovs, &(xo[WS(os, 1)]));
3268 }
3269 {
3270 V T5b, T5e, T5r, T5y;
3271 T5b = VBYI(VSUB(T59, T5a));
3272 T5e = VSUB(T5c, T5d);
3273 ST(&(xo[WS(os, 33)]), VADD(T5b, T5e), ovs, &(xo[WS(os, 1)]));
3274 ST(&(xo[WS(os, 95)]), VSUB(T5e, T5b), ovs, &(xo[WS(os, 1)]));
3275 T5r = VADD(T5j, T5q);
3276 T5y = VBYI(VADD(T5u, T5x));
3277 ST(&(xo[WS(os, 113)]), VSUB(T5r, T5y), ovs, &(xo[WS(os, 1)]));
3278 ST(&(xo[WS(os, 15)]), VADD(T5r, T5y), ovs, &(xo[WS(os, 1)]));
3279 }
3280 {
3281 V T5z, T5A, T5f, T5g;
3282 T5z = VSUB(T5j, T5q);
3283 T5A = VBYI(VSUB(T5x, T5u));
3284 ST(&(xo[WS(os, 79)]), VSUB(T5z, T5A), ovs, &(xo[WS(os, 1)]));
3285 ST(&(xo[WS(os, 49)]), VADD(T5z, T5A), ovs, &(xo[WS(os, 1)]));
3286 T5f = VBYI(VADD(T5a, T59));
3287 T5g = VADD(T5c, T5d);
3288 ST(&(xo[WS(os, 31)]), VADD(T5f, T5g), ovs, &(xo[WS(os, 1)]));
3289 ST(&(xo[WS(os, 97)]), VSUB(T5g, T5f), ovs, &(xo[WS(os, 1)]));
3290 }
3291 }
3292 {
3293 V T9i, T9B, T9t, T9x, T9O, Ta3, T9V, T9Z, T93, Ta0, Ta2, T9q, T9y, T9H, T9S;
3294 V T9A;
3295 {
3296 V T9a, T9r, T9h, T9s;
3297 {
3298 V T96, T99, T9d, T9g;
3299 T96 = VSUB(T94, T95);
3300 T99 = VSUB(T97, T98);
3301 T9a = VFMA(LDK(KP514102744), T96, VMUL(LDK(KP857728610), T99));
3302 T9r = VFNMS(LDK(KP514102744), T99, VMUL(LDK(KP857728610), T96));
3303 T9d = VSUB(T9b, T9c);
3304 T9g = VSUB(T9e, T9f);
3305 T9h = VFNMS(LDK(KP514102744), T9g, VMUL(LDK(KP857728610), T9d));
3306 T9s = VFMA(LDK(KP857728610), T9g, VMUL(LDK(KP514102744), T9d));
3307 }
3308 T9i = VADD(T9a, T9h);
3309 T9B = VSUB(T9s, T9r);
3310 T9t = VADD(T9r, T9s);
3311 T9x = VSUB(T9h, T9a);
3312 }
3313 {
3314 V T9K, T9T, T9N, T9U;
3315 {
3316 V T9I, T9J, T9L, T9M;
3317 T9I = VADD(T95, T94);
3318 T9J = VADD(T97, T98);
3319 T9K = VFMA(LDK(KP242980179), T9I, VMUL(LDK(KP970031253), T9J));
3320 T9T = VFNMS(LDK(KP242980179), T9J, VMUL(LDK(KP970031253), T9I));
3321 T9L = VADD(T9b, T9c);
3322 T9M = VADD(T9f, T9e);
3323 T9N = VFNMS(LDK(KP242980179), T9M, VMUL(LDK(KP970031253), T9L));
3324 T9U = VFMA(LDK(KP970031253), T9M, VMUL(LDK(KP242980179), T9L));
3325 }
3326 T9O = VADD(T9K, T9N);
3327 Ta3 = VSUB(T9U, T9T);
3328 T9V = VADD(T9T, T9U);
3329 T9Z = VSUB(T9N, T9K);
3330 }
3331 {
3332 V T8V, T9F, T9m, T9Q, T92, T9R, T9p, T9G, T8U, T9k;
3333 T8U = VSUB(T86, T85);
3334 T8V = VSUB(T8T, T8U);
3335 T9F = VADD(T8T, T8U);
3336 T9k = VSUB(T75, T72);
3337 T9m = VSUB(T9k, T9l);
3338 T9Q = VADD(T9l, T9k);
3339 {
3340 V T8Y, T91, T9n, T9o;
3341 T8Y = VFMA(LDK(KP471396736), T8W, VMUL(LDK(KP881921264), T8X));
3342 T91 = VFNMS(LDK(KP471396736), T90, VMUL(LDK(KP881921264), T8Z));
3343 T92 = VSUB(T8Y, T91);
3344 T9R = VADD(T91, T8Y);
3345 T9n = VFNMS(LDK(KP471396736), T8X, VMUL(LDK(KP881921264), T8W));
3346 T9o = VFMA(LDK(KP881921264), T90, VMUL(LDK(KP471396736), T8Z));
3347 T9p = VSUB(T9n, T9o);
3348 T9G = VADD(T9o, T9n);
3349 }
3350 T93 = VADD(T8V, T92);
3351 Ta0 = VSUB(T9R, T9Q);
3352 Ta2 = VSUB(T9F, T9G);
3353 T9q = VADD(T9m, T9p);
3354 T9y = VSUB(T9p, T9m);
3355 T9H = VADD(T9F, T9G);
3356 T9S = VADD(T9Q, T9R);
3357 T9A = VSUB(T8V, T92);
3358 }
3359 {
3360 V T9j, T9u, Ta1, Ta4;
3361 T9j = VADD(T93, T9i);
3362 T9u = VBYI(VADD(T9q, T9t));
3363 ST(&(xo[WS(os, 117)]), VSUB(T9j, T9u), ovs, &(xo[WS(os, 1)]));
3364 ST(&(xo[WS(os, 11)]), VADD(T9j, T9u), ovs, &(xo[WS(os, 1)]));
3365 Ta1 = VBYI(VSUB(T9Z, Ta0));
3366 Ta4 = VSUB(Ta2, Ta3);
3367 ST(&(xo[WS(os, 37)]), VADD(Ta1, Ta4), ovs, &(xo[WS(os, 1)]));
3368 ST(&(xo[WS(os, 91)]), VSUB(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
3369 }
3370 {
3371 V Ta5, Ta6, T9v, T9w;
3372 Ta5 = VBYI(VADD(Ta0, T9Z));
3373 Ta6 = VADD(Ta2, Ta3);
3374 ST(&(xo[WS(os, 27)]), VADD(Ta5, Ta6), ovs, &(xo[WS(os, 1)]));
3375 ST(&(xo[WS(os, 101)]), VSUB(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
3376 T9v = VSUB(T93, T9i);
3377 T9w = VBYI(VSUB(T9t, T9q));
3378 ST(&(xo[WS(os, 75)]), VSUB(T9v, T9w), ovs, &(xo[WS(os, 1)]));
3379 ST(&(xo[WS(os, 53)]), VADD(T9v, T9w), ovs, &(xo[WS(os, 1)]));
3380 }
3381 {
3382 V T9z, T9C, T9P, T9W;
3383 T9z = VBYI(VSUB(T9x, T9y));
3384 T9C = VSUB(T9A, T9B);
3385 ST(&(xo[WS(os, 43)]), VADD(T9z, T9C), ovs, &(xo[WS(os, 1)]));
3386 ST(&(xo[WS(os, 85)]), VSUB(T9C, T9z), ovs, &(xo[WS(os, 1)]));
3387 T9P = VADD(T9H, T9O);
3388 T9W = VBYI(VADD(T9S, T9V));
3389 ST(&(xo[WS(os, 123)]), VSUB(T9P, T9W), ovs, &(xo[WS(os, 1)]));
3390 ST(&(xo[WS(os, 5)]), VADD(T9P, T9W), ovs, &(xo[WS(os, 1)]));
3391 }
3392 {
3393 V T9X, T9Y, T9D, T9E;
3394 T9X = VSUB(T9H, T9O);
3395 T9Y = VBYI(VSUB(T9V, T9S));
3396 ST(&(xo[WS(os, 69)]), VSUB(T9X, T9Y), ovs, &(xo[WS(os, 1)]));
3397 ST(&(xo[WS(os, 59)]), VADD(T9X, T9Y), ovs, &(xo[WS(os, 1)]));
3398 T9D = VBYI(VADD(T9y, T9x));
3399 T9E = VADD(T9A, T9B);
3400 ST(&(xo[WS(os, 21)]), VADD(T9D, T9E), ovs, &(xo[WS(os, 1)]));
3401 ST(&(xo[WS(os, 107)]), VSUB(T9E, T9D), ovs, &(xo[WS(os, 1)]));
3402 }
3403 }
3404 {
3405 V T68, T6r, T6j, T6n, T6E, T6T, T6L, T6P, T5T, T6Q, T6S, T6g, T6o, T6x, T6I;
3406 V T6q;
3407 {
3408 V T60, T6h, T67, T6i;
3409 {
3410 V T5W, T5Z, T63, T66;
3411 T5W = VSUB(T5U, T5V);
3412 T5Z = VSUB(T5X, T5Y);
3413 T60 = VFMA(LDK(KP427555093), T5W, VMUL(LDK(KP903989293), T5Z));
3414 T6h = VFNMS(LDK(KP427555093), T5Z, VMUL(LDK(KP903989293), T5W));
3415 T63 = VSUB(T61, T62);
3416 T66 = VSUB(T64, T65);
3417 T67 = VFNMS(LDK(KP427555093), T66, VMUL(LDK(KP903989293), T63));
3418 T6i = VFMA(LDK(KP903989293), T66, VMUL(LDK(KP427555093), T63));
3419 }
3420 T68 = VADD(T60, T67);
3421 T6r = VSUB(T6i, T6h);
3422 T6j = VADD(T6h, T6i);
3423 T6n = VSUB(T67, T60);
3424 }
3425 {
3426 V T6A, T6J, T6D, T6K;
3427 {
3428 V T6y, T6z, T6B, T6C;
3429 T6y = VADD(T5V, T5U);
3430 T6z = VADD(T5X, T5Y);
3431 T6A = VFMA(LDK(KP336889853), T6y, VMUL(LDK(KP941544065), T6z));
3432 T6J = VFNMS(LDK(KP336889853), T6z, VMUL(LDK(KP941544065), T6y));
3433 T6B = VADD(T61, T62);
3434 T6C = VADD(T65, T64);
3435 T6D = VFNMS(LDK(KP336889853), T6C, VMUL(LDK(KP941544065), T6B));
3436 T6K = VFMA(LDK(KP941544065), T6C, VMUL(LDK(KP336889853), T6B));
3437 }
3438 T6E = VADD(T6A, T6D);
3439 T6T = VSUB(T6K, T6J);
3440 T6L = VADD(T6J, T6K);
3441 T6P = VSUB(T6D, T6A);
3442 }
3443 {
3444 V T5L, T6v, T6c, T6G, T5S, T6H, T6f, T6w, T5K, T6a;
3445 T5K = VSUB(T4W, T4V);
3446 T5L = VSUB(T5J, T5K);
3447 T6v = VADD(T5J, T5K);
3448 T6a = VSUB(TZ, TI);
3449 T6c = VSUB(T6a, T6b);
3450 T6G = VADD(T6b, T6a);
3451 {
3452 V T5O, T5R, T6d, T6e;
3453 T5O = VFMA(LDK(KP773010453), T5M, VMUL(LDK(KP634393284), T5N));
3454 T5R = VFNMS(LDK(KP634393284), T5Q, VMUL(LDK(KP773010453), T5P));
3455 T5S = VSUB(T5O, T5R);
3456 T6H = VADD(T5R, T5O);
3457 T6d = VFNMS(LDK(KP634393284), T5M, VMUL(LDK(KP773010453), T5N));
3458 T6e = VFMA(LDK(KP634393284), T5P, VMUL(LDK(KP773010453), T5Q));
3459 T6f = VSUB(T6d, T6e);
3460 T6w = VADD(T6e, T6d);
3461 }
3462 T5T = VADD(T5L, T5S);
3463 T6Q = VSUB(T6H, T6G);
3464 T6S = VSUB(T6v, T6w);
3465 T6g = VADD(T6c, T6f);
3466 T6o = VSUB(T6f, T6c);
3467 T6x = VADD(T6v, T6w);
3468 T6I = VADD(T6G, T6H);
3469 T6q = VSUB(T5L, T5S);
3470 }
3471 {
3472 V T69, T6k, T6R, T6U;
3473 T69 = VADD(T5T, T68);
3474 T6k = VBYI(VADD(T6g, T6j));
3475 ST(&(xo[WS(os, 119)]), VSUB(T69, T6k), ovs, &(xo[WS(os, 1)]));
3476 ST(&(xo[WS(os, 9)]), VADD(T69, T6k), ovs, &(xo[WS(os, 1)]));
3477 T6R = VBYI(VSUB(T6P, T6Q));
3478 T6U = VSUB(T6S, T6T);
3479 ST(&(xo[WS(os, 39)]), VADD(T6R, T6U), ovs, &(xo[WS(os, 1)]));
3480 ST(&(xo[WS(os, 89)]), VSUB(T6U, T6R), ovs, &(xo[WS(os, 1)]));
3481 }
3482 {
3483 V T6V, T6W, T6l, T6m;
3484 T6V = VBYI(VADD(T6Q, T6P));
3485 T6W = VADD(T6S, T6T);
3486 ST(&(xo[WS(os, 25)]), VADD(T6V, T6W), ovs, &(xo[WS(os, 1)]));
3487 ST(&(xo[WS(os, 103)]), VSUB(T6W, T6V), ovs, &(xo[WS(os, 1)]));
3488 T6l = VSUB(T5T, T68);
3489 T6m = VBYI(VSUB(T6j, T6g));
3490 ST(&(xo[WS(os, 73)]), VSUB(T6l, T6m), ovs, &(xo[WS(os, 1)]));
3491 ST(&(xo[WS(os, 55)]), VADD(T6l, T6m), ovs, &(xo[WS(os, 1)]));
3492 }
3493 {
3494 V T6p, T6s, T6F, T6M;
3495 T6p = VBYI(VSUB(T6n, T6o));
3496 T6s = VSUB(T6q, T6r);
3497 ST(&(xo[WS(os, 41)]), VADD(T6p, T6s), ovs, &(xo[WS(os, 1)]));
3498 ST(&(xo[WS(os, 87)]), VSUB(T6s, T6p), ovs, &(xo[WS(os, 1)]));
3499 T6F = VADD(T6x, T6E);
3500 T6M = VBYI(VADD(T6I, T6L));
3501 ST(&(xo[WS(os, 121)]), VSUB(T6F, T6M), ovs, &(xo[WS(os, 1)]));
3502 ST(&(xo[WS(os, 7)]), VADD(T6F, T6M), ovs, &(xo[WS(os, 1)]));
3503 }
3504 {
3505 V T6N, T6O, T6t, T6u;
3506 T6N = VSUB(T6x, T6E);
3507 T6O = VBYI(VSUB(T6L, T6I));
3508 ST(&(xo[WS(os, 71)]), VSUB(T6N, T6O), ovs, &(xo[WS(os, 1)]));
3509 ST(&(xo[WS(os, 57)]), VADD(T6N, T6O), ovs, &(xo[WS(os, 1)]));
3510 T6t = VBYI(VADD(T6o, T6n));
3511 T6u = VADD(T6q, T6r);
3512 ST(&(xo[WS(os, 23)]), VADD(T6t, T6u), ovs, &(xo[WS(os, 1)]));
3513 ST(&(xo[WS(os, 105)]), VSUB(T6u, T6t), ovs, &(xo[WS(os, 1)]));
3514 }
3515 }
3516 }
3517 }
3518 VLEAVE();
3519 }
3520
3521 static const kdft_desc desc = { 128, XSIMD_STRING("n1fv_128"), {938, 186, 144, 0}, &GENUS, 0, 0, 0, 0 };
3522
3523 void XSIMD(codelet_n1fv_128) (planner *p) {
3524 X(kdft_register) (p, n1fv_128, &desc);
3525 }
3526
3527 #endif /* HAVE_FMA */