1 /*
2  * Copyright (c) 2003, 2007-14 Matteo Frigo
3  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License as published by
7  * the Free Software Foundation; either version 2 of the License, or
8  * (at your option) any later version.
9  *
10  * This program is distributed in the hope that it will be useful,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License for more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program; if not, write to the Free Software
17  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
18  *
19  */
20 
21 /* This file was automatically generated --- DO NOT EDIT */
22 /* Generated on Thu Dec 10 07:04:48 EST 2020 */
23 
24 #include "dft/codelet-dft.h"
25 
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27 
28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 128 -name n1bv_128 -include dft/simd/n1b.h */
29 
30 /*
31  * This function contains 1082 FP additions, 642 FP multiplications,
32  * (or, 440 additions, 0 multiplications, 642 fused multiply/add),
33  * 194 stack variables, 31 constants, and 256 memory accesses
34  */
35 #include "dft/simd/n1b.h"
36 
n1bv_128(const R * ri,const R * ii,R * ro,R * io,stride is,stride os,INT v,INT ivs,INT ovs)37 static void n1bv_128(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38 {
39      DVK(KP941544065, +0.941544065183020778412509402599502357185589796);
40      DVK(KP903989293, +0.903989293123443331586200297230537048710132025);
41      DVK(KP357805721, +0.357805721314524104672487743774474392487532769);
42      DVK(KP472964775, +0.472964775891319928124438237972992463904131113);
43      DVK(KP970031253, +0.970031253194543992603984207286100251456865962);
44      DVK(KP857728610, +0.857728610000272069902269984284770137042490799);
45      DVK(KP250486960, +0.250486960191305461595702160124721208578685568);
46      DVK(KP599376933, +0.599376933681923766271389869014404232837890546);
47      DVK(KP740951125, +0.740951125354959091175616897495162729728955309);
48      DVK(KP998795456, +0.998795456205172392714771604759100694443203615);
49      DVK(KP906347169, +0.906347169019147157946142717268914412664134293);
50      DVK(KP049126849, +0.049126849769467254105343321271313617079695752);
51      DVK(KP803207531, +0.803207531480644909806676512963141923879569427);
52      DVK(KP989176509, +0.989176509964780973451673738016243063983689533);
53      DVK(KP741650546, +0.741650546272035369581266691172079863842265220);
54      DVK(KP148335987, +0.148335987538347428753676511486911367000625355);
55      DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
56      DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
57      DVK(KP303346683, +0.303346683607342391675883946941299872384187453);
58      DVK(KP534511135, +0.534511135950791641089685961295362908582039528);
59      DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
60      DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
61      DVK(KP820678790, +0.820678790828660330972281985331011598767386482);
62      DVK(KP098491403, +0.098491403357164253077197521291327432293052451);
63      DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
64      DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
65      DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
66      DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
67      DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
68      DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
69      DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
70      {
71 	  INT i;
72 	  const R *xi;
73 	  R *xo;
74 	  xi = ii;
75 	  xo = io;
76 	  for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(256, is), MAKE_VOLATILE_STRIDE(256, os)) {
77 	       V Tr, T5J, Ted, Tgf, Tfq, TgG, T4U, T6a, T6Z, T8T, Tad, TcZ, Tcc, Td0, T84;
78 	       V T9k, Tb6, Tbt, T2G, T5U, TeV, Tgt, T3p, T5X, T7B, T97, TeK, Tgq, T7q, T94;
79 	       V Td8, TdK, TbD, Tc0, T3V, T61, Tfg, TgA, T4E, T64, T7U, T9e, Tf5, Tgx, T7J;
80 	       V T9b, Tdf, TdN, Td2, Td3, TI, T4V, Tft, Tgg, TZ, T4W, T75, T86, Tek, TgH;
81 	       V T72, T85, Tas, Tcd, Tdm, Tdn, TdG, Teq, Tgj, Tet, Tgi, T1s, T5N, T1B, T5M;
82 	       V T7d, T8W, TaI, Tcf, T7a, T8X, Tdp, Tdq, TdH, Tez, Tgm, TeC, Tgl, T23, T5Q;
83 	       V T2c, T5P, T7k, T8Z, TaX, Tcg, T7h, T90, Tbl, Tbu, Tdb, TdL, TeY, Tgr, TeR;
84 	       V Tgu, T7x, T98, T7E, T95, T3f, T5Y, T3s, T5V, TbS, Tc1, Tdi, TdO, Tfj, Tgy;
85 	       V Tfc, TgB, T7Q, T9f, T7X, T9c, T4u, T65, T4H, T62;
86 	       {
87 		    V T3, Ta7, T4O, Ta8, Ta, Tab, T4P, Taa, Tc6, Tc7, Ti, Tea, T4R, Tc9, Tca;
88 		    V Tp, Teb, T4S;
89 		    {
90 			 V T1, T2, T4M, T4N;
91 			 T1 = LD(&(xi[0]), ivs, &(xi[0]));
92 			 T2 = LD(&(xi[WS(is, 64)]), ivs, &(xi[0]));
93 			 T3 = VSUB(T1, T2);
94 			 Ta7 = VADD(T1, T2);
95 			 T4M = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
96 			 T4N = LD(&(xi[WS(is, 96)]), ivs, &(xi[0]));
97 			 T4O = VSUB(T4M, T4N);
98 			 Ta8 = VADD(T4M, T4N);
99 		    }
100 		    {
101 			 V T4, T5, T6, T7, T8, T9;
102 			 T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
103 			 T5 = LD(&(xi[WS(is, 80)]), ivs, &(xi[0]));
104 			 T6 = VSUB(T4, T5);
105 			 T7 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0]));
106 			 T8 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
107 			 T9 = VSUB(T7, T8);
108 			 Ta = VADD(T6, T9);
109 			 Tab = VADD(T7, T8);
110 			 T4P = VSUB(T6, T9);
111 			 Taa = VADD(T4, T5);
112 		    }
113 		    {
114 			 V Te, Th, Tl, To;
115 			 {
116 			      V Tc, Td, Tf, Tg;
117 			      Tc = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
118 			      Td = LD(&(xi[WS(is, 72)]), ivs, &(xi[0]));
119 			      Te = VSUB(Tc, Td);
120 			      Tc6 = VADD(Tc, Td);
121 			      Tf = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
122 			      Tg = LD(&(xi[WS(is, 104)]), ivs, &(xi[0]));
123 			      Th = VSUB(Tf, Tg);
124 			      Tc7 = VADD(Tf, Tg);
125 			 }
126 			 Ti = VFNMS(LDK(KP414213562), Th, Te);
127 			 Tea = VSUB(Tc6, Tc7);
128 			 T4R = VFMA(LDK(KP414213562), Te, Th);
129 			 {
130 			      V Tj, Tk, Tm, Tn;
131 			      Tj = LD(&(xi[WS(is, 120)]), ivs, &(xi[0]));
132 			      Tk = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
133 			      Tl = VSUB(Tj, Tk);
134 			      Tc9 = VADD(Tj, Tk);
135 			      Tm = LD(&(xi[WS(is, 88)]), ivs, &(xi[0]));
136 			      Tn = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
137 			      To = VSUB(Tm, Tn);
138 			      Tca = VADD(Tn, Tm);
139 			 }
140 			 Tp = VFNMS(LDK(KP414213562), To, Tl);
141 			 Teb = VSUB(Tc9, Tca);
142 			 T4S = VFMA(LDK(KP414213562), Tl, To);
143 		    }
144 		    {
145 			 V Tb, Tq, Te9, Tec;
146 			 Tb = VFMA(LDK(KP707106781), Ta, T3);
147 			 Tq = VADD(Ti, Tp);
148 			 Tr = VFMA(LDK(KP923879532), Tq, Tb);
149 			 T5J = VFNMS(LDK(KP923879532), Tq, Tb);
150 			 Te9 = VSUB(Ta7, Ta8);
151 			 Tec = VADD(Tea, Teb);
152 			 Ted = VFMA(LDK(KP707106781), Tec, Te9);
153 			 Tgf = VFNMS(LDK(KP707106781), Tec, Te9);
154 		    }
155 		    {
156 			 V Tfo, Tfp, T4Q, T4T;
157 			 Tfo = VSUB(Taa, Tab);
158 			 Tfp = VSUB(Tea, Teb);
159 			 Tfq = VFMA(LDK(KP707106781), Tfp, Tfo);
160 			 TgG = VFNMS(LDK(KP707106781), Tfp, Tfo);
161 			 T4Q = VFMA(LDK(KP707106781), T4P, T4O);
162 			 T4T = VSUB(T4R, T4S);
163 			 T4U = VFMA(LDK(KP923879532), T4T, T4Q);
164 			 T6a = VFNMS(LDK(KP923879532), T4T, T4Q);
165 		    }
166 		    {
167 			 V T6X, T6Y, Ta9, Tac;
168 			 T6X = VFNMS(LDK(KP707106781), Ta, T3);
169 			 T6Y = VADD(T4R, T4S);
170 			 T6Z = VFMA(LDK(KP923879532), T6Y, T6X);
171 			 T8T = VFNMS(LDK(KP923879532), T6Y, T6X);
172 			 Ta9 = VADD(Ta7, Ta8);
173 			 Tac = VADD(Taa, Tab);
174 			 Tad = VSUB(Ta9, Tac);
175 			 TcZ = VADD(Ta9, Tac);
176 		    }
177 		    {
178 			 V Tc8, Tcb, T82, T83;
179 			 Tc8 = VADD(Tc6, Tc7);
180 			 Tcb = VADD(Tc9, Tca);
181 			 Tcc = VSUB(Tc8, Tcb);
182 			 Td0 = VADD(Tc8, Tcb);
183 			 T82 = VFNMS(LDK(KP707106781), T4P, T4O);
184 			 T83 = VSUB(Ti, Tp);
185 			 T84 = VFNMS(LDK(KP923879532), T83, T82);
186 			 T9k = VFMA(LDK(KP923879532), T83, T82);
187 		    }
188 	       }
189 	       {
190 		    V Tb0, Tb1, T2i, Tb2, T3j, Tb3, Tb4, T2p, Tb5, T3k, T2x, TeH, T3m, Tbp, T2E;
191 		    V TeI, T3n, Tbs, T2l, T2o, TeG, TeJ;
192 		    {
193 			 V T2g, T2h, T3h, T3i;
194 			 T2g = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
195 			 T2h = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)]));
196 			 Tb0 = VADD(T2g, T2h);
197 			 T3h = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
198 			 T3i = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)]));
199 			 Tb1 = VADD(T3h, T3i);
200 			 T2i = VSUB(T2g, T2h);
201 			 Tb2 = VADD(Tb0, Tb1);
202 			 T3j = VSUB(T3h, T3i);
203 		    }
204 		    {
205 			 V T2j, T2k, T2m, T2n;
206 			 T2j = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
207 			 T2k = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)]));
208 			 T2l = VSUB(T2j, T2k);
209 			 Tb3 = VADD(T2j, T2k);
210 			 T2m = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)]));
211 			 T2n = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
212 			 T2o = VSUB(T2m, T2n);
213 			 Tb4 = VADD(T2m, T2n);
214 		    }
215 		    T2p = VADD(T2l, T2o);
216 		    Tb5 = VADD(Tb3, Tb4);
217 		    T3k = VSUB(T2l, T2o);
218 		    {
219 			 V T2t, Tbn, T2w, Tbo;
220 			 {
221 			      V T2r, T2s, T2u, T2v;
222 			      T2r = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
223 			      T2s = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)]));
224 			      T2t = VSUB(T2r, T2s);
225 			      Tbn = VADD(T2r, T2s);
226 			      T2u = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
227 			      T2v = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)]));
228 			      T2w = VSUB(T2u, T2v);
229 			      Tbo = VADD(T2u, T2v);
230 			 }
231 			 T2x = VFNMS(LDK(KP414213562), T2w, T2t);
232 			 TeH = VSUB(Tbn, Tbo);
233 			 T3m = VFMA(LDK(KP414213562), T2t, T2w);
234 			 Tbp = VADD(Tbn, Tbo);
235 		    }
236 		    {
237 			 V T2A, Tbq, T2D, Tbr;
238 			 {
239 			      V T2y, T2z, T2B, T2C;
240 			      T2y = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)]));
241 			      T2z = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
242 			      T2A = VSUB(T2y, T2z);
243 			      Tbq = VADD(T2y, T2z);
244 			      T2B = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
245 			      T2C = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)]));
246 			      T2D = VSUB(T2B, T2C);
247 			      Tbr = VADD(T2B, T2C);
248 			 }
249 			 T2E = VFMA(LDK(KP414213562), T2D, T2A);
250 			 TeI = VSUB(Tbq, Tbr);
251 			 T3n = VFNMS(LDK(KP414213562), T2A, T2D);
252 			 Tbs = VADD(Tbq, Tbr);
253 		    }
254 		    Tb6 = VSUB(Tb2, Tb5);
255 		    Tbt = VSUB(Tbp, Tbs);
256 		    {
257 			 V T2q, T2F, TeT, TeU;
258 			 T2q = VFMA(LDK(KP707106781), T2p, T2i);
259 			 T2F = VADD(T2x, T2E);
260 			 T2G = VFMA(LDK(KP923879532), T2F, T2q);
261 			 T5U = VFNMS(LDK(KP923879532), T2F, T2q);
262 			 TeT = VSUB(Tb3, Tb4);
263 			 TeU = VSUB(TeH, TeI);
264 			 TeV = VFMA(LDK(KP707106781), TeU, TeT);
265 			 Tgt = VFNMS(LDK(KP707106781), TeU, TeT);
266 		    }
267 		    {
268 			 V T3l, T3o, T7z, T7A;
269 			 T3l = VFMA(LDK(KP707106781), T3k, T3j);
270 			 T3o = VADD(T3m, T3n);
271 			 T3p = VFMA(LDK(KP923879532), T3o, T3l);
272 			 T5X = VFNMS(LDK(KP923879532), T3o, T3l);
273 			 T7z = VFNMS(LDK(KP707106781), T3k, T3j);
274 			 T7A = VSUB(T2x, T2E);
275 			 T7B = VFNMS(LDK(KP923879532), T7A, T7z);
276 			 T97 = VFMA(LDK(KP923879532), T7A, T7z);
277 		    }
278 		    TeG = VSUB(Tb0, Tb1);
279 		    TeJ = VADD(TeH, TeI);
280 		    TeK = VFMA(LDK(KP707106781), TeJ, TeG);
281 		    Tgq = VFNMS(LDK(KP707106781), TeJ, TeG);
282 		    {
283 			 V T7o, T7p, Td6, Td7;
284 			 T7o = VFNMS(LDK(KP707106781), T2p, T2i);
285 			 T7p = VSUB(T3m, T3n);
286 			 T7q = VFMA(LDK(KP923879532), T7p, T7o);
287 			 T94 = VFNMS(LDK(KP923879532), T7p, T7o);
288 			 Td6 = VADD(Tb2, Tb5);
289 			 Td7 = VADD(Tbp, Tbs);
290 			 Td8 = VADD(Td6, Td7);
291 			 TdK = VSUB(Td6, Td7);
292 		    }
293 	       }
294 	       {
295 		    V Tbx, Tby, T3x, Tbz, T4y, TbA, TbB, T3E, TbC, T4z, T3M, Tf2, T4B, TbZ, T3T;
296 		    V Tf3, T4C, TbW, T3A, T3D, Tf1, Tf4;
297 		    {
298 			 V T3v, T3w, T4w, T4x;
299 			 T3v = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)]));
300 			 T3w = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
301 			 Tbx = VADD(T3v, T3w);
302 			 T4w = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)]));
303 			 T4x = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
304 			 Tby = VADD(T4x, T4w);
305 			 T3x = VSUB(T3v, T3w);
306 			 Tbz = VADD(Tbx, Tby);
307 			 T4y = VSUB(T4w, T4x);
308 		    }
309 		    {
310 			 V T3y, T3z, T3B, T3C;
311 			 T3y = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
312 			 T3z = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)]));
313 			 T3A = VSUB(T3y, T3z);
314 			 TbA = VADD(T3y, T3z);
315 			 T3B = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)]));
316 			 T3C = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
317 			 T3D = VSUB(T3B, T3C);
318 			 TbB = VADD(T3B, T3C);
319 		    }
320 		    T3E = VADD(T3A, T3D);
321 		    TbC = VADD(TbA, TbB);
322 		    T4z = VSUB(T3D, T3A);
323 		    {
324 			 V T3I, TbX, T3L, TbY;
325 			 {
326 			      V T3G, T3H, T3J, T3K;
327 			      T3G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
328 			      T3H = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)]));
329 			      T3I = VSUB(T3G, T3H);
330 			      TbX = VADD(T3G, T3H);
331 			      T3J = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)]));
332 			      T3K = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
333 			      T3L = VSUB(T3J, T3K);
334 			      TbY = VADD(T3K, T3J);
335 			 }
336 			 T3M = VFMA(LDK(KP414213562), T3L, T3I);
337 			 Tf2 = VSUB(TbX, TbY);
338 			 T4B = VFNMS(LDK(KP414213562), T3I, T3L);
339 			 TbZ = VADD(TbX, TbY);
340 		    }
341 		    {
342 			 V T3P, TbU, T3S, TbV;
343 			 {
344 			      V T3N, T3O, T3Q, T3R;
345 			      T3N = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)]));
346 			      T3O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
347 			      T3P = VSUB(T3N, T3O);
348 			      TbU = VADD(T3N, T3O);
349 			      T3Q = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)]));
350 			      T3R = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
351 			      T3S = VSUB(T3Q, T3R);
352 			      TbV = VADD(T3R, T3Q);
353 			 }
354 			 T3T = VFNMS(LDK(KP414213562), T3S, T3P);
355 			 Tf3 = VSUB(TbU, TbV);
356 			 T4C = VFMA(LDK(KP414213562), T3P, T3S);
357 			 TbW = VADD(TbU, TbV);
358 		    }
359 		    TbD = VSUB(Tbz, TbC);
360 		    Tc0 = VSUB(TbW, TbZ);
361 		    {
362 			 V T3F, T3U, Tfe, Tff;
363 			 T3F = VFMA(LDK(KP707106781), T3E, T3x);
364 			 T3U = VADD(T3M, T3T);
365 			 T3V = VFMA(LDK(KP923879532), T3U, T3F);
366 			 T61 = VFNMS(LDK(KP923879532), T3U, T3F);
367 			 Tfe = VSUB(TbB, TbA);
368 			 Tff = VSUB(Tf3, Tf2);
369 			 Tfg = VFMA(LDK(KP707106781), Tff, Tfe);
370 			 TgA = VFNMS(LDK(KP707106781), Tff, Tfe);
371 		    }
372 		    {
373 			 V T4A, T4D, T7S, T7T;
374 			 T4A = VFMA(LDK(KP707106781), T4z, T4y);
375 			 T4D = VADD(T4B, T4C);
376 			 T4E = VFMA(LDK(KP923879532), T4D, T4A);
377 			 T64 = VFNMS(LDK(KP923879532), T4D, T4A);
378 			 T7S = VFNMS(LDK(KP707106781), T4z, T4y);
379 			 T7T = VSUB(T3T, T3M);
380 			 T7U = VFNMS(LDK(KP923879532), T7T, T7S);
381 			 T9e = VFMA(LDK(KP923879532), T7T, T7S);
382 		    }
383 		    Tf1 = VSUB(Tbx, Tby);
384 		    Tf4 = VADD(Tf2, Tf3);
385 		    Tf5 = VFMA(LDK(KP707106781), Tf4, Tf1);
386 		    Tgx = VFNMS(LDK(KP707106781), Tf4, Tf1);
387 		    {
388 			 V T7H, T7I, Tdd, Tde;
389 			 T7H = VFNMS(LDK(KP707106781), T3E, T3x);
390 			 T7I = VSUB(T4C, T4B);
391 			 T7J = VFMA(LDK(KP923879532), T7I, T7H);
392 			 T9b = VFNMS(LDK(KP923879532), T7I, T7H);
393 			 Tdd = VADD(Tbz, TbC);
394 			 Tde = VADD(TbZ, TbW);
395 			 Tdf = VADD(Tdd, Tde);
396 			 TdN = VSUB(Tdd, Tde);
397 		    }
398 	       }
399 	       {
400 		    V Tu, Tee, TF, Tag, TL, Teh, TW, Tan, TB, Tef, TG, Taj, TS, Tei, TX;
401 		    V Taq, Teg, Tej;
402 		    {
403 			 V Ts, Tt, Tae, TD, TE, Taf;
404 			 Ts = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
405 			 Tt = LD(&(xi[WS(is, 68)]), ivs, &(xi[0]));
406 			 Tae = VADD(Ts, Tt);
407 			 TD = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
408 			 TE = LD(&(xi[WS(is, 100)]), ivs, &(xi[0]));
409 			 Taf = VADD(TD, TE);
410 			 Tu = VSUB(Ts, Tt);
411 			 Tee = VSUB(Tae, Taf);
412 			 TF = VSUB(TD, TE);
413 			 Tag = VADD(Tae, Taf);
414 		    }
415 		    {
416 			 V TJ, TK, Tal, TU, TV, Tam;
417 			 TJ = LD(&(xi[WS(is, 124)]), ivs, &(xi[0]));
418 			 TK = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
419 			 Tal = VADD(TJ, TK);
420 			 TU = LD(&(xi[WS(is, 92)]), ivs, &(xi[0]));
421 			 TV = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
422 			 Tam = VADD(TV, TU);
423 			 TL = VSUB(TJ, TK);
424 			 Teh = VSUB(Tal, Tam);
425 			 TW = VSUB(TU, TV);
426 			 Tan = VADD(Tal, Tam);
427 		    }
428 		    {
429 			 V Tx, Tah, TA, Tai;
430 			 {
431 			      V Tv, Tw, Ty, Tz;
432 			      Tv = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
433 			      Tw = LD(&(xi[WS(is, 84)]), ivs, &(xi[0]));
434 			      Tx = VSUB(Tv, Tw);
435 			      Tah = VADD(Tv, Tw);
436 			      Ty = LD(&(xi[WS(is, 116)]), ivs, &(xi[0]));
437 			      Tz = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
438 			      TA = VSUB(Ty, Tz);
439 			      Tai = VADD(Ty, Tz);
440 			 }
441 			 TB = VADD(Tx, TA);
442 			 Tef = VSUB(Tah, Tai);
443 			 TG = VSUB(Tx, TA);
444 			 Taj = VADD(Tah, Tai);
445 		    }
446 		    {
447 			 V TO, Tao, TR, Tap;
448 			 {
449 			      V TM, TN, TP, TQ;
450 			      TM = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
451 			      TN = LD(&(xi[WS(is, 76)]), ivs, &(xi[0]));
452 			      TO = VSUB(TM, TN);
453 			      Tao = VADD(TM, TN);
454 			      TP = LD(&(xi[WS(is, 108)]), ivs, &(xi[0]));
455 			      TQ = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
456 			      TR = VSUB(TP, TQ);
457 			      Tap = VADD(TP, TQ);
458 			 }
459 			 TS = VADD(TO, TR);
460 			 Tei = VSUB(Tap, Tao);
461 			 TX = VSUB(TR, TO);
462 			 Taq = VADD(Tao, Tap);
463 		    }
464 		    Td2 = VADD(Tag, Taj);
465 		    Td3 = VADD(Tan, Taq);
466 		    {
467 			 V TC, TH, Tfr, Tfs;
468 			 TC = VFMA(LDK(KP707106781), TB, Tu);
469 			 TH = VFMA(LDK(KP707106781), TG, TF);
470 			 TI = VFNMS(LDK(KP198912367), TH, TC);
471 			 T4V = VFMA(LDK(KP198912367), TC, TH);
472 			 Tfr = VFMA(LDK(KP414213562), Tee, Tef);
473 			 Tfs = VFMA(LDK(KP414213562), Teh, Tei);
474 			 Tft = VSUB(Tfr, Tfs);
475 			 Tgg = VADD(Tfr, Tfs);
476 		    }
477 		    {
478 			 V TT, TY, T73, T74;
479 			 TT = VFMA(LDK(KP707106781), TS, TL);
480 			 TY = VFMA(LDK(KP707106781), TX, TW);
481 			 TZ = VFNMS(LDK(KP198912367), TY, TT);
482 			 T4W = VFMA(LDK(KP198912367), TT, TY);
483 			 T73 = VFNMS(LDK(KP707106781), TS, TL);
484 			 T74 = VFNMS(LDK(KP707106781), TX, TW);
485 			 T75 = VFMA(LDK(KP668178637), T74, T73);
486 			 T86 = VFNMS(LDK(KP668178637), T73, T74);
487 		    }
488 		    Teg = VFNMS(LDK(KP414213562), Tef, Tee);
489 		    Tej = VFNMS(LDK(KP414213562), Tei, Teh);
490 		    Tek = VADD(Teg, Tej);
491 		    TgH = VSUB(Teg, Tej);
492 		    {
493 			 V T70, T71, Tak, Tar;
494 			 T70 = VFNMS(LDK(KP707106781), TB, Tu);
495 			 T71 = VFNMS(LDK(KP707106781), TG, TF);
496 			 T72 = VFMA(LDK(KP668178637), T71, T70);
497 			 T85 = VFNMS(LDK(KP668178637), T70, T71);
498 			 Tak = VSUB(Tag, Taj);
499 			 Tar = VSUB(Tan, Taq);
500 			 Tas = VADD(Tak, Tar);
501 			 Tcd = VSUB(Tak, Tar);
502 		    }
503 	       }
504 	       {
505 		    V Tau, Tav, T14, Taw, T1v, Tax, Tay, T1b, Taz, T1w, T1j, Ten, T1y, TaD, T1q;
506 		    V Teo, T1z, TaG, T17, T1a, Tem, Tep;
507 		    {
508 			 V T12, T13, T1t, T1u;
509 			 T12 = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
510 			 T13 = LD(&(xi[WS(is, 66)]), ivs, &(xi[0]));
511 			 Tau = VADD(T12, T13);
512 			 T1t = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
513 			 T1u = LD(&(xi[WS(is, 98)]), ivs, &(xi[0]));
514 			 Tav = VADD(T1t, T1u);
515 			 T14 = VSUB(T12, T13);
516 			 Taw = VADD(Tau, Tav);
517 			 T1v = VSUB(T1t, T1u);
518 		    }
519 		    {
520 			 V T15, T16, T18, T19;
521 			 T15 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
522 			 T16 = LD(&(xi[WS(is, 82)]), ivs, &(xi[0]));
523 			 T17 = VSUB(T15, T16);
524 			 Tax = VADD(T15, T16);
525 			 T18 = LD(&(xi[WS(is, 114)]), ivs, &(xi[0]));
526 			 T19 = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
527 			 T1a = VSUB(T18, T19);
528 			 Tay = VADD(T18, T19);
529 		    }
530 		    T1b = VADD(T17, T1a);
531 		    Taz = VADD(Tax, Tay);
532 		    T1w = VSUB(T17, T1a);
533 		    {
534 			 V T1f, TaB, T1i, TaC;
535 			 {
536 			      V T1d, T1e, T1g, T1h;
537 			      T1d = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
538 			      T1e = LD(&(xi[WS(is, 74)]), ivs, &(xi[0]));
539 			      T1f = VSUB(T1d, T1e);
540 			      TaB = VADD(T1d, T1e);
541 			      T1g = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
542 			      T1h = LD(&(xi[WS(is, 106)]), ivs, &(xi[0]));
543 			      T1i = VSUB(T1g, T1h);
544 			      TaC = VADD(T1g, T1h);
545 			 }
546 			 T1j = VFNMS(LDK(KP414213562), T1i, T1f);
547 			 Ten = VSUB(TaB, TaC);
548 			 T1y = VFMA(LDK(KP414213562), T1f, T1i);
549 			 TaD = VADD(TaB, TaC);
550 		    }
551 		    {
552 			 V T1m, TaE, T1p, TaF;
553 			 {
554 			      V T1k, T1l, T1n, T1o;
555 			      T1k = LD(&(xi[WS(is, 122)]), ivs, &(xi[0]));
556 			      T1l = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
557 			      T1m = VSUB(T1k, T1l);
558 			      TaE = VADD(T1k, T1l);
559 			      T1n = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
560 			      T1o = LD(&(xi[WS(is, 90)]), ivs, &(xi[0]));
561 			      T1p = VSUB(T1n, T1o);
562 			      TaF = VADD(T1n, T1o);
563 			 }
564 			 T1q = VFMA(LDK(KP414213562), T1p, T1m);
565 			 Teo = VSUB(TaE, TaF);
566 			 T1z = VFNMS(LDK(KP414213562), T1m, T1p);
567 			 TaG = VADD(TaE, TaF);
568 		    }
569 		    Tdm = VADD(Taw, Taz);
570 		    Tdn = VADD(TaD, TaG);
571 		    TdG = VSUB(Tdm, Tdn);
572 		    Tem = VSUB(Tau, Tav);
573 		    Tep = VADD(Ten, Teo);
574 		    Teq = VFMA(LDK(KP707106781), Tep, Tem);
575 		    Tgj = VFNMS(LDK(KP707106781), Tep, Tem);
576 		    {
577 			 V Ter, Tes, T1c, T1r;
578 			 Ter = VSUB(Tax, Tay);
579 			 Tes = VSUB(Ten, Teo);
580 			 Tet = VFMA(LDK(KP707106781), Tes, Ter);
581 			 Tgi = VFNMS(LDK(KP707106781), Tes, Ter);
582 			 T1c = VFMA(LDK(KP707106781), T1b, T14);
583 			 T1r = VADD(T1j, T1q);
584 			 T1s = VFMA(LDK(KP923879532), T1r, T1c);
585 			 T5N = VFNMS(LDK(KP923879532), T1r, T1c);
586 		    }
587 		    {
588 			 V T1x, T1A, T7b, T7c;
589 			 T1x = VFMA(LDK(KP707106781), T1w, T1v);
590 			 T1A = VADD(T1y, T1z);
591 			 T1B = VFMA(LDK(KP923879532), T1A, T1x);
592 			 T5M = VFNMS(LDK(KP923879532), T1A, T1x);
593 			 T7b = VFNMS(LDK(KP707106781), T1w, T1v);
594 			 T7c = VSUB(T1j, T1q);
595 			 T7d = VFNMS(LDK(KP923879532), T7c, T7b);
596 			 T8W = VFMA(LDK(KP923879532), T7c, T7b);
597 		    }
598 		    {
599 			 V TaA, TaH, T78, T79;
600 			 TaA = VSUB(Taw, Taz);
601 			 TaH = VSUB(TaD, TaG);
602 			 TaI = VFNMS(LDK(KP414213562), TaH, TaA);
603 			 Tcf = VFMA(LDK(KP414213562), TaA, TaH);
604 			 T78 = VFNMS(LDK(KP707106781), T1b, T14);
605 			 T79 = VSUB(T1y, T1z);
606 			 T7a = VFMA(LDK(KP923879532), T79, T78);
607 			 T8X = VFNMS(LDK(KP923879532), T79, T78);
608 		    }
609 	       }
610 	       {
611 		    V TaJ, TaK, T1F, TaL, T26, TaM, TaN, T1M, TaO, T27, T1U, Tew, T29, TaV, T21;
612 		    V Tex, T2a, TaS, T1I, T1L, Tev, Tey;
613 		    {
614 			 V T1D, T1E, T24, T25;
615 			 T1D = LD(&(xi[WS(is, 126)]), ivs, &(xi[0]));
616 			 T1E = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
617 			 TaJ = VADD(T1D, T1E);
618 			 T24 = LD(&(xi[WS(is, 94)]), ivs, &(xi[0]));
619 			 T25 = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
620 			 TaK = VADD(T25, T24);
621 			 T1F = VSUB(T1D, T1E);
622 			 TaL = VADD(TaJ, TaK);
623 			 T26 = VSUB(T24, T25);
624 		    }
625 		    {
626 			 V T1G, T1H, T1J, T1K;
627 			 T1G = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
628 			 T1H = LD(&(xi[WS(is, 78)]), ivs, &(xi[0]));
629 			 T1I = VSUB(T1G, T1H);
630 			 TaM = VADD(T1G, T1H);
631 			 T1J = LD(&(xi[WS(is, 110)]), ivs, &(xi[0]));
632 			 T1K = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
633 			 T1L = VSUB(T1J, T1K);
634 			 TaN = VADD(T1J, T1K);
635 		    }
636 		    T1M = VADD(T1I, T1L);
637 		    TaO = VADD(TaM, TaN);
638 		    T27 = VSUB(T1L, T1I);
639 		    {
640 			 V T1Q, TaT, T1T, TaU;
641 			 {
642 			      V T1O, T1P, T1R, T1S;
643 			      T1O = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
644 			      T1P = LD(&(xi[WS(is, 70)]), ivs, &(xi[0]));
645 			      T1Q = VSUB(T1O, T1P);
646 			      TaT = VADD(T1O, T1P);
647 			      T1R = LD(&(xi[WS(is, 102)]), ivs, &(xi[0]));
648 			      T1S = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
649 			      T1T = VSUB(T1R, T1S);
650 			      TaU = VADD(T1S, T1R);
651 			 }
652 			 T1U = VFMA(LDK(KP414213562), T1T, T1Q);
653 			 Tew = VSUB(TaT, TaU);
654 			 T29 = VFNMS(LDK(KP414213562), T1Q, T1T);
655 			 TaV = VADD(TaT, TaU);
656 		    }
657 		    {
658 			 V T1X, TaQ, T20, TaR;
659 			 {
660 			      V T1V, T1W, T1Y, T1Z;
661 			      T1V = LD(&(xi[WS(is, 118)]), ivs, &(xi[0]));
662 			      T1W = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
663 			      T1X = VSUB(T1V, T1W);
664 			      TaQ = VADD(T1V, T1W);
665 			      T1Y = LD(&(xi[WS(is, 86)]), ivs, &(xi[0]));
666 			      T1Z = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
667 			      T20 = VSUB(T1Y, T1Z);
668 			      TaR = VADD(T1Z, T1Y);
669 			 }
670 			 T21 = VFNMS(LDK(KP414213562), T20, T1X);
671 			 Tex = VSUB(TaQ, TaR);
672 			 T2a = VFMA(LDK(KP414213562), T1X, T20);
673 			 TaS = VADD(TaQ, TaR);
674 		    }
675 		    Tdp = VADD(TaL, TaO);
676 		    Tdq = VADD(TaV, TaS);
677 		    TdH = VSUB(Tdp, Tdq);
678 		    Tev = VSUB(TaJ, TaK);
679 		    Tey = VADD(Tew, Tex);
680 		    Tez = VFMA(LDK(KP707106781), Tey, Tev);
681 		    Tgm = VFNMS(LDK(KP707106781), Tey, Tev);
682 		    {
683 			 V TeA, TeB, T1N, T22;
684 			 TeA = VSUB(TaN, TaM);
685 			 TeB = VSUB(Tex, Tew);
686 			 TeC = VFMA(LDK(KP707106781), TeB, TeA);
687 			 Tgl = VFNMS(LDK(KP707106781), TeB, TeA);
688 			 T1N = VFMA(LDK(KP707106781), T1M, T1F);
689 			 T22 = VADD(T1U, T21);
690 			 T23 = VFMA(LDK(KP923879532), T22, T1N);
691 			 T5Q = VFNMS(LDK(KP923879532), T22, T1N);
692 		    }
693 		    {
694 			 V T28, T2b, T7i, T7j;
695 			 T28 = VFMA(LDK(KP707106781), T27, T26);
696 			 T2b = VADD(T29, T2a);
697 			 T2c = VFMA(LDK(KP923879532), T2b, T28);
698 			 T5P = VFNMS(LDK(KP923879532), T2b, T28);
699 			 T7i = VFNMS(LDK(KP707106781), T27, T26);
700 			 T7j = VSUB(T21, T1U);
701 			 T7k = VFNMS(LDK(KP923879532), T7j, T7i);
702 			 T8Z = VFMA(LDK(KP923879532), T7j, T7i);
703 		    }
704 		    {
705 			 V TaP, TaW, T7f, T7g;
706 			 TaP = VSUB(TaL, TaO);
707 			 TaW = VSUB(TaS, TaV);
708 			 TaX = VFNMS(LDK(KP414213562), TaW, TaP);
709 			 Tcg = VFMA(LDK(KP414213562), TaP, TaW);
710 			 T7f = VFNMS(LDK(KP707106781), T1M, T1F);
711 			 T7g = VSUB(T2a, T29);
712 			 T7h = VFMA(LDK(KP923879532), T7g, T7f);
713 			 T90 = VFNMS(LDK(KP923879532), T7g, T7f);
714 		    }
715 	       }
716 	       {
717 		    V T2J, TeL, T2U, Tb9, T30, TeO, T3b, Tbg, T2Q, TeM, T2V, Tbc, T37, TeP, T3c;
718 		    V Tbj;
719 		    {
720 			 V T2H, T2I, Tb7, T2S, T2T, Tb8;
721 			 T2H = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
722 			 T2I = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)]));
723 			 Tb7 = VADD(T2H, T2I);
724 			 T2S = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
725 			 T2T = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)]));
726 			 Tb8 = VADD(T2S, T2T);
727 			 T2J = VSUB(T2H, T2I);
728 			 TeL = VSUB(Tb7, Tb8);
729 			 T2U = VSUB(T2S, T2T);
730 			 Tb9 = VADD(Tb7, Tb8);
731 		    }
732 		    {
733 			 V T2Y, T2Z, Tbe, T39, T3a, Tbf;
734 			 T2Y = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)]));
735 			 T2Z = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
736 			 Tbe = VADD(T2Y, T2Z);
737 			 T39 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
738 			 T3a = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)]));
739 			 Tbf = VADD(T39, T3a);
740 			 T30 = VSUB(T2Y, T2Z);
741 			 TeO = VSUB(Tbe, Tbf);
742 			 T3b = VSUB(T39, T3a);
743 			 Tbg = VADD(Tbe, Tbf);
744 		    }
745 		    {
746 			 V T2M, Tba, T2P, Tbb;
747 			 {
748 			      V T2K, T2L, T2N, T2O;
749 			      T2K = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
750 			      T2L = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)]));
751 			      T2M = VSUB(T2K, T2L);
752 			      Tba = VADD(T2K, T2L);
753 			      T2N = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)]));
754 			      T2O = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
755 			      T2P = VSUB(T2N, T2O);
756 			      Tbb = VADD(T2N, T2O);
757 			 }
758 			 T2Q = VADD(T2M, T2P);
759 			 TeM = VSUB(Tba, Tbb);
760 			 T2V = VSUB(T2M, T2P);
761 			 Tbc = VADD(Tba, Tbb);
762 		    }
763 		    {
764 			 V T33, Tbh, T36, Tbi;
765 			 {
766 			      V T31, T32, T34, T35;
767 			      T31 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
768 			      T32 = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)]));
769 			      T33 = VSUB(T31, T32);
770 			      Tbh = VADD(T31, T32);
771 			      T34 = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)]));
772 			      T35 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
773 			      T36 = VSUB(T34, T35);
774 			      Tbi = VADD(T34, T35);
775 			 }
776 			 T37 = VADD(T33, T36);
777 			 TeP = VSUB(Tbh, Tbi);
778 			 T3c = VSUB(T33, T36);
779 			 Tbj = VADD(Tbh, Tbi);
780 		    }
781 		    {
782 			 V Tbd, Tbk, TeN, TeQ;
783 			 Tbd = VSUB(Tb9, Tbc);
784 			 Tbk = VSUB(Tbg, Tbj);
785 			 Tbl = VADD(Tbd, Tbk);
786 			 Tbu = VSUB(Tbd, Tbk);
787 			 {
788 			      V Td9, Tda, TeW, TeX;
789 			      Td9 = VADD(Tb9, Tbc);
790 			      Tda = VADD(Tbg, Tbj);
791 			      Tdb = VADD(Td9, Tda);
792 			      TdL = VSUB(Td9, Tda);
793 			      TeW = VFMA(LDK(KP414213562), TeL, TeM);
794 			      TeX = VFNMS(LDK(KP414213562), TeO, TeP);
795 			      TeY = VADD(TeW, TeX);
796 			      Tgr = VSUB(TeW, TeX);
797 			 }
798 			 TeN = VFNMS(LDK(KP414213562), TeM, TeL);
799 			 TeQ = VFMA(LDK(KP414213562), TeP, TeO);
800 			 TeR = VADD(TeN, TeQ);
801 			 Tgu = VSUB(TeN, TeQ);
802 			 {
803 			      V T7t, T7C, T7w, T7D;
804 			      {
805 				   V T7r, T7s, T7u, T7v;
806 				   T7r = VFNMS(LDK(KP707106781), T2Q, T2J);
807 				   T7s = VFNMS(LDK(KP707106781), T2V, T2U);
808 				   T7t = VFMA(LDK(KP668178637), T7s, T7r);
809 				   T7C = VFNMS(LDK(KP668178637), T7r, T7s);
810 				   T7u = VFNMS(LDK(KP707106781), T37, T30);
811 				   T7v = VFNMS(LDK(KP707106781), T3c, T3b);
812 				   T7w = VFNMS(LDK(KP668178637), T7v, T7u);
813 				   T7D = VFMA(LDK(KP668178637), T7u, T7v);
814 			      }
815 			      T7x = VADD(T7t, T7w);
816 			      T98 = VSUB(T7t, T7w);
817 			      T7E = VADD(T7C, T7D);
818 			      T95 = VSUB(T7D, T7C);
819 			 }
820 			 {
821 			      V T2X, T3q, T3e, T3r;
822 			      {
823 				   V T2R, T2W, T38, T3d;
824 				   T2R = VFMA(LDK(KP707106781), T2Q, T2J);
825 				   T2W = VFMA(LDK(KP707106781), T2V, T2U);
826 				   T2X = VFNMS(LDK(KP198912367), T2W, T2R);
827 				   T3q = VFMA(LDK(KP198912367), T2R, T2W);
828 				   T38 = VFMA(LDK(KP707106781), T37, T30);
829 				   T3d = VFMA(LDK(KP707106781), T3c, T3b);
830 				   T3e = VFMA(LDK(KP198912367), T3d, T38);
831 				   T3r = VFNMS(LDK(KP198912367), T38, T3d);
832 			      }
833 			      T3f = VADD(T2X, T3e);
834 			      T5Y = VSUB(T2X, T3e);
835 			      T3s = VADD(T3q, T3r);
836 			      T5V = VSUB(T3q, T3r);
837 			 }
838 		    }
839 	       }
840 	       {
841 		    V T3Y, Tf6, T49, TbG, T4f, Tf9, T4q, TbN, T45, Tf7, T4a, TbJ, T4m, Tfa, T4r;
842 		    V TbQ;
843 		    {
844 			 V T3W, T3X, TbE, T47, T48, TbF;
845 			 T3W = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
846 			 T3X = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)]));
847 			 TbE = VADD(T3W, T3X);
848 			 T47 = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)]));
849 			 T48 = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
850 			 TbF = VADD(T48, T47);
851 			 T3Y = VSUB(T3W, T3X);
852 			 Tf6 = VSUB(TbE, TbF);
853 			 T49 = VSUB(T47, T48);
854 			 TbG = VADD(TbE, TbF);
855 		    }
856 		    {
857 			 V T4d, T4e, TbL, T4o, T4p, TbM;
858 			 T4d = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)]));
859 			 T4e = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
860 			 TbL = VADD(T4d, T4e);
861 			 T4o = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)]));
862 			 T4p = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
863 			 TbM = VADD(T4p, T4o);
864 			 T4f = VSUB(T4d, T4e);
865 			 Tf9 = VSUB(TbL, TbM);
866 			 T4q = VSUB(T4o, T4p);
867 			 TbN = VADD(TbL, TbM);
868 		    }
869 		    {
870 			 V T41, TbH, T44, TbI;
871 			 {
872 			      V T3Z, T40, T42, T43;
873 			      T3Z = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
874 			      T40 = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)]));
875 			      T41 = VSUB(T3Z, T40);
876 			      TbH = VADD(T3Z, T40);
877 			      T42 = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)]));
878 			      T43 = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
879 			      T44 = VSUB(T42, T43);
880 			      TbI = VADD(T42, T43);
881 			 }
882 			 T45 = VADD(T41, T44);
883 			 Tf7 = VSUB(TbI, TbH);
884 			 T4a = VSUB(T44, T41);
885 			 TbJ = VADD(TbH, TbI);
886 		    }
887 		    {
888 			 V T4i, TbO, T4l, TbP;
889 			 {
890 			      V T4g, T4h, T4j, T4k;
891 			      T4g = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
892 			      T4h = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)]));
893 			      T4i = VSUB(T4g, T4h);
894 			      TbO = VADD(T4g, T4h);
895 			      T4j = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)]));
896 			      T4k = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
897 			      T4l = VSUB(T4j, T4k);
898 			      TbP = VADD(T4j, T4k);
899 			 }
900 			 T4m = VADD(T4i, T4l);
901 			 Tfa = VSUB(TbP, TbO);
902 			 T4r = VSUB(T4l, T4i);
903 			 TbQ = VADD(TbO, TbP);
904 		    }
905 		    {
906 			 V TbK, TbR, Tf8, Tfb;
907 			 TbK = VSUB(TbG, TbJ);
908 			 TbR = VSUB(TbN, TbQ);
909 			 TbS = VADD(TbK, TbR);
910 			 Tc1 = VSUB(TbR, TbK);
911 			 {
912 			      V Tdg, Tdh, Tfh, Tfi;
913 			      Tdg = VADD(TbG, TbJ);
914 			      Tdh = VADD(TbN, TbQ);
915 			      Tdi = VADD(Tdg, Tdh);
916 			      TdO = VSUB(Tdh, Tdg);
917 			      Tfh = VFNMS(LDK(KP414213562), Tf6, Tf7);
918 			      Tfi = VFMA(LDK(KP414213562), Tf9, Tfa);
919 			      Tfj = VADD(Tfh, Tfi);
920 			      Tgy = VSUB(Tfi, Tfh);
921 			 }
922 			 Tf8 = VFMA(LDK(KP414213562), Tf7, Tf6);
923 			 Tfb = VFNMS(LDK(KP414213562), Tfa, Tf9);
924 			 Tfc = VADD(Tf8, Tfb);
925 			 TgB = VSUB(Tfb, Tf8);
926 			 {
927 			      V T7M, T7V, T7P, T7W;
928 			      {
929 				   V T7K, T7L, T7N, T7O;
930 				   T7K = VFNMS(LDK(KP707106781), T45, T3Y);
931 				   T7L = VFNMS(LDK(KP707106781), T4a, T49);
932 				   T7M = VFNMS(LDK(KP668178637), T7L, T7K);
933 				   T7V = VFMA(LDK(KP668178637), T7K, T7L);
934 				   T7N = VFNMS(LDK(KP707106781), T4m, T4f);
935 				   T7O = VFNMS(LDK(KP707106781), T4r, T4q);
936 				   T7P = VFMA(LDK(KP668178637), T7O, T7N);
937 				   T7W = VFNMS(LDK(KP668178637), T7N, T7O);
938 			      }
939 			      T7Q = VADD(T7M, T7P);
940 			      T9f = VSUB(T7P, T7M);
941 			      T7X = VADD(T7V, T7W);
942 			      T9c = VSUB(T7V, T7W);
943 			 }
944 			 {
945 			      V T4c, T4F, T4t, T4G;
946 			      {
947 				   V T46, T4b, T4n, T4s;
948 				   T46 = VFMA(LDK(KP707106781), T45, T3Y);
949 				   T4b = VFMA(LDK(KP707106781), T4a, T49);
950 				   T4c = VFMA(LDK(KP198912367), T4b, T46);
951 				   T4F = VFNMS(LDK(KP198912367), T46, T4b);
952 				   T4n = VFMA(LDK(KP707106781), T4m, T4f);
953 				   T4s = VFMA(LDK(KP707106781), T4r, T4q);
954 				   T4t = VFNMS(LDK(KP198912367), T4s, T4n);
955 				   T4G = VFMA(LDK(KP198912367), T4n, T4s);
956 			      }
957 			      T4u = VADD(T4c, T4t);
958 			      T65 = VSUB(T4t, T4c);
959 			      T4H = VADD(T4F, T4G);
960 			      T62 = VSUB(T4G, T4F);
961 			 }
962 		    }
963 	       }
964 	       {
965 		    V Td5, Tdx, TdC, TdE, Tdk, Tdt, Tds, Tdy, Tdz, TdD;
966 		    {
967 			 V Td1, Td4, TdA, TdB;
968 			 Td1 = VADD(TcZ, Td0);
969 			 Td4 = VADD(Td2, Td3);
970 			 Td5 = VSUB(Td1, Td4);
971 			 Tdx = VADD(Td1, Td4);
972 			 TdA = VADD(Td8, Tdb);
973 			 TdB = VADD(Tdf, Tdi);
974 			 TdC = VSUB(TdA, TdB);
975 			 TdE = VADD(TdA, TdB);
976 		    }
977 		    {
978 			 V Tdc, Tdj, Tdo, Tdr;
979 			 Tdc = VSUB(Td8, Tdb);
980 			 Tdj = VSUB(Tdf, Tdi);
981 			 Tdk = VADD(Tdc, Tdj);
982 			 Tdt = VSUB(Tdc, Tdj);
983 			 Tdo = VADD(Tdm, Tdn);
984 			 Tdr = VADD(Tdp, Tdq);
985 			 Tds = VSUB(Tdo, Tdr);
986 			 Tdy = VADD(Tdo, Tdr);
987 		    }
988 		    Tdz = VSUB(Tdx, Tdy);
989 		    ST(&(xo[WS(os, 96)]), VFNMSI(TdC, Tdz), ovs, &(xo[0]));
990 		    ST(&(xo[WS(os, 32)]), VFMAI(TdC, Tdz), ovs, &(xo[0]));
991 		    TdD = VADD(Tdx, Tdy);
992 		    ST(&(xo[WS(os, 64)]), VSUB(TdD, TdE), ovs, &(xo[0]));
993 		    ST(&(xo[0]), VADD(TdD, TdE), ovs, &(xo[0]));
994 		    {
995 			 V Tdl, Tdu, Tdv, Tdw;
996 			 Tdl = VFNMS(LDK(KP707106781), Tdk, Td5);
997 			 Tdu = VFNMS(LDK(KP707106781), Tdt, Tds);
998 			 ST(&(xo[WS(os, 48)]), VFNMSI(Tdu, Tdl), ovs, &(xo[0]));
999 			 ST(&(xo[WS(os, 80)]), VFMAI(Tdu, Tdl), ovs, &(xo[0]));
1000 			 Tdv = VFMA(LDK(KP707106781), Tdk, Td5);
1001 			 Tdw = VFMA(LDK(KP707106781), Tdt, Tds);
1002 			 ST(&(xo[WS(os, 16)]), VFMAI(Tdw, Tdv), ovs, &(xo[0]));
1003 			 ST(&(xo[WS(os, 112)]), VFNMSI(Tdw, Tdv), ovs, &(xo[0]));
1004 		    }
1005 	       }
1006 	       {
1007 		    V TdJ, Te1, TdX, Te2, TdQ, Te5, TdU, Te4;
1008 		    {
1009 			 V TdF, TdI, TdV, TdW;
1010 			 TdF = VSUB(TcZ, Td0);
1011 			 TdI = VADD(TdG, TdH);
1012 			 TdJ = VFMA(LDK(KP707106781), TdI, TdF);
1013 			 Te1 = VFNMS(LDK(KP707106781), TdI, TdF);
1014 			 TdV = VFMA(LDK(KP414213562), TdK, TdL);
1015 			 TdW = VFMA(LDK(KP414213562), TdN, TdO);
1016 			 TdX = VSUB(TdV, TdW);
1017 			 Te2 = VADD(TdV, TdW);
1018 		    }
1019 		    {
1020 			 V TdM, TdP, TdS, TdT;
1021 			 TdM = VFNMS(LDK(KP414213562), TdL, TdK);
1022 			 TdP = VFNMS(LDK(KP414213562), TdO, TdN);
1023 			 TdQ = VADD(TdM, TdP);
1024 			 Te5 = VSUB(TdM, TdP);
1025 			 TdS = VSUB(Td2, Td3);
1026 			 TdT = VSUB(TdG, TdH);
1027 			 TdU = VFMA(LDK(KP707106781), TdT, TdS);
1028 			 Te4 = VFNMS(LDK(KP707106781), TdT, TdS);
1029 		    }
1030 		    {
1031 			 V TdR, TdY, Te7, Te8;
1032 			 TdR = VFNMS(LDK(KP923879532), TdQ, TdJ);
1033 			 TdY = VFNMS(LDK(KP923879532), TdX, TdU);
1034 			 ST(&(xo[WS(os, 56)]), VFNMSI(TdY, TdR), ovs, &(xo[0]));
1035 			 ST(&(xo[WS(os, 72)]), VFMAI(TdY, TdR), ovs, &(xo[0]));
1036 			 Te7 = VFMA(LDK(KP923879532), Te2, Te1);
1037 			 Te8 = VFNMS(LDK(KP923879532), Te5, Te4);
1038 			 ST(&(xo[WS(os, 24)]), VFNMSI(Te8, Te7), ovs, &(xo[0]));
1039 			 ST(&(xo[WS(os, 104)]), VFMAI(Te8, Te7), ovs, &(xo[0]));
1040 		    }
1041 		    {
1042 			 V TdZ, Te0, Te3, Te6;
1043 			 TdZ = VFMA(LDK(KP923879532), TdQ, TdJ);
1044 			 Te0 = VFMA(LDK(KP923879532), TdX, TdU);
1045 			 ST(&(xo[WS(os, 120)]), VFNMSI(Te0, TdZ), ovs, &(xo[0]));
1046 			 ST(&(xo[WS(os, 8)]), VFMAI(Te0, TdZ), ovs, &(xo[0]));
1047 			 Te3 = VFNMS(LDK(KP923879532), Te2, Te1);
1048 			 Te6 = VFMA(LDK(KP923879532), Te5, Te4);
1049 			 ST(&(xo[WS(os, 40)]), VFMAI(Te6, Te3), ovs, &(xo[0]));
1050 			 ST(&(xo[WS(os, 88)]), VFNMSI(Te6, Te3), ovs, &(xo[0]));
1051 		    }
1052 	       }
1053 	       {
1054 		    V TaZ, Tcp, Tci, Tcs, Tc4, Tct, Tcl, Tcq;
1055 		    {
1056 			 V Tat, TaY, Tce, Tch;
1057 			 Tat = VFMA(LDK(KP707106781), Tas, Tad);
1058 			 TaY = VADD(TaI, TaX);
1059 			 TaZ = VFMA(LDK(KP923879532), TaY, Tat);
1060 			 Tcp = VFNMS(LDK(KP923879532), TaY, Tat);
1061 			 Tce = VFMA(LDK(KP707106781), Tcd, Tcc);
1062 			 Tch = VSUB(Tcf, Tcg);
1063 			 Tci = VFMA(LDK(KP923879532), Tch, Tce);
1064 			 Tcs = VFNMS(LDK(KP923879532), Tch, Tce);
1065 			 {
1066 			      V Tbw, Tcj, Tc3, Tck;
1067 			      {
1068 				   V Tbm, Tbv, TbT, Tc2;
1069 				   Tbm = VFMA(LDK(KP707106781), Tbl, Tb6);
1070 				   Tbv = VFMA(LDK(KP707106781), Tbu, Tbt);
1071 				   Tbw = VFNMS(LDK(KP198912367), Tbv, Tbm);
1072 				   Tcj = VFMA(LDK(KP198912367), Tbm, Tbv);
1073 				   TbT = VFMA(LDK(KP707106781), TbS, TbD);
1074 				   Tc2 = VFMA(LDK(KP707106781), Tc1, Tc0);
1075 				   Tc3 = VFNMS(LDK(KP198912367), Tc2, TbT);
1076 				   Tck = VFMA(LDK(KP198912367), TbT, Tc2);
1077 			      }
1078 			      Tc4 = VADD(Tbw, Tc3);
1079 			      Tct = VSUB(Tbw, Tc3);
1080 			      Tcl = VSUB(Tcj, Tck);
1081 			      Tcq = VADD(Tcj, Tck);
1082 			 }
1083 		    }
1084 		    {
1085 			 V Tc5, Tcm, Tcv, Tcw;
1086 			 Tc5 = VFNMS(LDK(KP980785280), Tc4, TaZ);
1087 			 Tcm = VFNMS(LDK(KP980785280), Tcl, Tci);
1088 			 ST(&(xo[WS(os, 60)]), VFNMSI(Tcm, Tc5), ovs, &(xo[0]));
1089 			 ST(&(xo[WS(os, 68)]), VFMAI(Tcm, Tc5), ovs, &(xo[0]));
1090 			 Tcv = VFMA(LDK(KP980785280), Tcq, Tcp);
1091 			 Tcw = VFNMS(LDK(KP980785280), Tct, Tcs);
1092 			 ST(&(xo[WS(os, 28)]), VFNMSI(Tcw, Tcv), ovs, &(xo[0]));
1093 			 ST(&(xo[WS(os, 100)]), VFMAI(Tcw, Tcv), ovs, &(xo[0]));
1094 		    }
1095 		    {
1096 			 V Tcn, Tco, Tcr, Tcu;
1097 			 Tcn = VFMA(LDK(KP980785280), Tc4, TaZ);
1098 			 Tco = VFMA(LDK(KP980785280), Tcl, Tci);
1099 			 ST(&(xo[WS(os, 124)]), VFNMSI(Tco, Tcn), ovs, &(xo[0]));
1100 			 ST(&(xo[WS(os, 4)]), VFMAI(Tco, Tcn), ovs, &(xo[0]));
1101 			 Tcr = VFNMS(LDK(KP980785280), Tcq, Tcp);
1102 			 Tcu = VFMA(LDK(KP980785280), Tct, Tcs);
1103 			 ST(&(xo[WS(os, 36)]), VFMAI(Tcu, Tcr), ovs, &(xo[0]));
1104 			 ST(&(xo[WS(os, 92)]), VFNMSI(Tcu, Tcr), ovs, &(xo[0]));
1105 		    }
1106 	       }
1107 	       {
1108 		    V Tcz, TcR, TcK, TcU, TcG, TcV, TcN, TcS;
1109 		    {
1110 			 V Tcx, Tcy, TcI, TcJ;
1111 			 Tcx = VFNMS(LDK(KP707106781), Tas, Tad);
1112 			 Tcy = VADD(Tcf, Tcg);
1113 			 Tcz = VFMA(LDK(KP923879532), Tcy, Tcx);
1114 			 TcR = VFNMS(LDK(KP923879532), Tcy, Tcx);
1115 			 TcI = VFNMS(LDK(KP707106781), Tcd, Tcc);
1116 			 TcJ = VSUB(TaI, TaX);
1117 			 TcK = VFNMS(LDK(KP923879532), TcJ, TcI);
1118 			 TcU = VFMA(LDK(KP923879532), TcJ, TcI);
1119 			 {
1120 			      V TcC, TcL, TcF, TcM;
1121 			      {
1122 				   V TcA, TcB, TcD, TcE;
1123 				   TcA = VFNMS(LDK(KP707106781), Tbl, Tb6);
1124 				   TcB = VFNMS(LDK(KP707106781), Tbu, Tbt);
1125 				   TcC = VFMA(LDK(KP668178637), TcB, TcA);
1126 				   TcL = VFNMS(LDK(KP668178637), TcA, TcB);
1127 				   TcD = VFNMS(LDK(KP707106781), TbS, TbD);
1128 				   TcE = VFNMS(LDK(KP707106781), Tc1, Tc0);
1129 				   TcF = VFMA(LDK(KP668178637), TcE, TcD);
1130 				   TcM = VFNMS(LDK(KP668178637), TcD, TcE);
1131 			      }
1132 			      TcG = VADD(TcC, TcF);
1133 			      TcV = VSUB(TcC, TcF);
1134 			      TcN = VSUB(TcL, TcM);
1135 			      TcS = VADD(TcL, TcM);
1136 			 }
1137 		    }
1138 		    {
1139 			 V TcH, TcO, TcX, TcY;
1140 			 TcH = VFNMS(LDK(KP831469612), TcG, Tcz);
1141 			 TcO = VFNMS(LDK(KP831469612), TcN, TcK);
1142 			 ST(&(xo[WS(os, 76)]), VFNMSI(TcO, TcH), ovs, &(xo[0]));
1143 			 ST(&(xo[WS(os, 52)]), VFMAI(TcO, TcH), ovs, &(xo[0]));
1144 			 TcX = VFNMS(LDK(KP831469612), TcS, TcR);
1145 			 TcY = VFMA(LDK(KP831469612), TcV, TcU);
1146 			 ST(&(xo[WS(os, 20)]), VFMAI(TcY, TcX), ovs, &(xo[0]));
1147 			 ST(&(xo[WS(os, 108)]), VFNMSI(TcY, TcX), ovs, &(xo[0]));
1148 		    }
1149 		    {
1150 			 V TcP, TcQ, TcT, TcW;
1151 			 TcP = VFMA(LDK(KP831469612), TcG, Tcz);
1152 			 TcQ = VFMA(LDK(KP831469612), TcN, TcK);
1153 			 ST(&(xo[WS(os, 12)]), VFNMSI(TcQ, TcP), ovs, &(xo[0]));
1154 			 ST(&(xo[WS(os, 116)]), VFMAI(TcQ, TcP), ovs, &(xo[0]));
1155 			 TcT = VFMA(LDK(KP831469612), TcS, TcR);
1156 			 TcW = VFNMS(LDK(KP831469612), TcV, TcU);
1157 			 ST(&(xo[WS(os, 44)]), VFNMSI(TcW, TcT), ovs, &(xo[0]));
1158 			 ST(&(xo[WS(os, 84)]), VFMAI(TcW, TcT), ovs, &(xo[0]));
1159 		    }
1160 	       }
1161 	       {
1162 		    V TeF, Tga, TfF, Tg0, Tfy, Tg7, TfI, TfP, Tfm, TfJ, TfB, TfG, TfW, Tgb, Tg3;
1163 		    V Tg8;
1164 		    {
1165 			 V Tel, TfY, TeE, TfZ, Teu, TeD;
1166 			 Tel = VFMA(LDK(KP923879532), Tek, Ted);
1167 			 TfY = VFNMS(LDK(KP923879532), Tft, Tfq);
1168 			 Teu = VFNMS(LDK(KP198912367), Tet, Teq);
1169 			 TeD = VFNMS(LDK(KP198912367), TeC, Tez);
1170 			 TeE = VADD(Teu, TeD);
1171 			 TfZ = VSUB(Teu, TeD);
1172 			 TeF = VFMA(LDK(KP980785280), TeE, Tel);
1173 			 Tga = VFMA(LDK(KP980785280), TfZ, TfY);
1174 			 TfF = VFNMS(LDK(KP980785280), TeE, Tel);
1175 			 Tg0 = VFNMS(LDK(KP980785280), TfZ, TfY);
1176 		    }
1177 		    {
1178 			 V Tfu, TfN, Tfx, TfO, Tfv, Tfw;
1179 			 Tfu = VFMA(LDK(KP923879532), Tft, Tfq);
1180 			 TfN = VFNMS(LDK(KP923879532), Tek, Ted);
1181 			 Tfv = VFMA(LDK(KP198912367), Teq, Tet);
1182 			 Tfw = VFMA(LDK(KP198912367), Tez, TeC);
1183 			 Tfx = VSUB(Tfv, Tfw);
1184 			 TfO = VADD(Tfv, Tfw);
1185 			 Tfy = VFMA(LDK(KP980785280), Tfx, Tfu);
1186 			 Tg7 = VFNMS(LDK(KP980785280), TfO, TfN);
1187 			 TfI = VFNMS(LDK(KP980785280), Tfx, Tfu);
1188 			 TfP = VFMA(LDK(KP980785280), TfO, TfN);
1189 		    }
1190 		    {
1191 			 V Tf0, Tfz, Tfl, TfA;
1192 			 {
1193 			      V TeS, TeZ, Tfd, Tfk;
1194 			      TeS = VFMA(LDK(KP923879532), TeR, TeK);
1195 			      TeZ = VFMA(LDK(KP923879532), TeY, TeV);
1196 			      Tf0 = VFNMS(LDK(KP098491403), TeZ, TeS);
1197 			      Tfz = VFMA(LDK(KP098491403), TeS, TeZ);
1198 			      Tfd = VFMA(LDK(KP923879532), Tfc, Tf5);
1199 			      Tfk = VFMA(LDK(KP923879532), Tfj, Tfg);
1200 			      Tfl = VFNMS(LDK(KP098491403), Tfk, Tfd);
1201 			      TfA = VFMA(LDK(KP098491403), Tfd, Tfk);
1202 			 }
1203 			 Tfm = VADD(Tf0, Tfl);
1204 			 TfJ = VSUB(Tf0, Tfl);
1205 			 TfB = VSUB(Tfz, TfA);
1206 			 TfG = VADD(Tfz, TfA);
1207 		    }
1208 		    {
1209 			 V TfS, Tg1, TfV, Tg2;
1210 			 {
1211 			      V TfQ, TfR, TfT, TfU;
1212 			      TfQ = VFNMS(LDK(KP923879532), TeR, TeK);
1213 			      TfR = VFNMS(LDK(KP923879532), TeY, TeV);
1214 			      TfS = VFMA(LDK(KP820678790), TfR, TfQ);
1215 			      Tg1 = VFNMS(LDK(KP820678790), TfQ, TfR);
1216 			      TfT = VFNMS(LDK(KP923879532), Tfc, Tf5);
1217 			      TfU = VFNMS(LDK(KP923879532), Tfj, Tfg);
1218 			      TfV = VFMA(LDK(KP820678790), TfU, TfT);
1219 			      Tg2 = VFNMS(LDK(KP820678790), TfT, TfU);
1220 			 }
1221 			 TfW = VADD(TfS, TfV);
1222 			 Tgb = VSUB(TfS, TfV);
1223 			 Tg3 = VSUB(Tg1, Tg2);
1224 			 Tg8 = VADD(Tg1, Tg2);
1225 		    }
1226 		    {
1227 			 V Tfn, TfC, Tg9, Tgc;
1228 			 Tfn = VFNMS(LDK(KP995184726), Tfm, TeF);
1229 			 TfC = VFNMS(LDK(KP995184726), TfB, Tfy);
1230 			 ST(&(xo[WS(os, 62)]), VFNMSI(TfC, Tfn), ovs, &(xo[0]));
1231 			 ST(&(xo[WS(os, 66)]), VFMAI(TfC, Tfn), ovs, &(xo[0]));
1232 			 Tg9 = VFMA(LDK(KP773010453), Tg8, Tg7);
1233 			 Tgc = VFNMS(LDK(KP773010453), Tgb, Tga);
1234 			 ST(&(xo[WS(os, 46)]), VFNMSI(Tgc, Tg9), ovs, &(xo[0]));
1235 			 ST(&(xo[WS(os, 82)]), VFMAI(Tgc, Tg9), ovs, &(xo[0]));
1236 		    }
1237 		    {
1238 			 V Tgd, Tge, TfD, TfE;
1239 			 Tgd = VFNMS(LDK(KP773010453), Tg8, Tg7);
1240 			 Tge = VFMA(LDK(KP773010453), Tgb, Tga);
1241 			 ST(&(xo[WS(os, 18)]), VFMAI(Tge, Tgd), ovs, &(xo[0]));
1242 			 ST(&(xo[WS(os, 110)]), VFNMSI(Tge, Tgd), ovs, &(xo[0]));
1243 			 TfD = VFMA(LDK(KP995184726), Tfm, TeF);
1244 			 TfE = VFMA(LDK(KP995184726), TfB, Tfy);
1245 			 ST(&(xo[WS(os, 126)]), VFNMSI(TfE, TfD), ovs, &(xo[0]));
1246 			 ST(&(xo[WS(os, 2)]), VFMAI(TfE, TfD), ovs, &(xo[0]));
1247 		    }
1248 		    {
1249 			 V TfH, TfK, TfX, Tg4;
1250 			 TfH = VFNMS(LDK(KP995184726), TfG, TfF);
1251 			 TfK = VFMA(LDK(KP995184726), TfJ, TfI);
1252 			 ST(&(xo[WS(os, 34)]), VFMAI(TfK, TfH), ovs, &(xo[0]));
1253 			 ST(&(xo[WS(os, 94)]), VFNMSI(TfK, TfH), ovs, &(xo[0]));
1254 			 TfX = VFNMS(LDK(KP773010453), TfW, TfP);
1255 			 Tg4 = VFNMS(LDK(KP773010453), Tg3, Tg0);
1256 			 ST(&(xo[WS(os, 78)]), VFNMSI(Tg4, TfX), ovs, &(xo[0]));
1257 			 ST(&(xo[WS(os, 50)]), VFMAI(Tg4, TfX), ovs, &(xo[0]));
1258 		    }
1259 		    {
1260 			 V Tg5, Tg6, TfL, TfM;
1261 			 Tg5 = VFMA(LDK(KP773010453), TfW, TfP);
1262 			 Tg6 = VFMA(LDK(KP773010453), Tg3, Tg0);
1263 			 ST(&(xo[WS(os, 14)]), VFNMSI(Tg6, Tg5), ovs, &(xo[0]));
1264 			 ST(&(xo[WS(os, 114)]), VFMAI(Tg6, Tg5), ovs, &(xo[0]));
1265 			 TfL = VFMA(LDK(KP995184726), TfG, TfF);
1266 			 TfM = VFNMS(LDK(KP995184726), TfJ, TfI);
1267 			 ST(&(xo[WS(os, 30)]), VFNMSI(TfM, TfL), ovs, &(xo[0]));
1268 			 ST(&(xo[WS(os, 98)]), VFMAI(TfM, TfL), ovs, &(xo[0]));
1269 		    }
1270 	       }
1271 	       {
1272 		    V Tgp, Tho, TgT, The, TgM, Thl, TgW, Th3, TgE, TgX, TgP, TgU, Tha, Thp, Thh;
1273 		    V Thm;
1274 		    {
1275 			 V Tgh, Thc, Tgo, Thd, Tgk, Tgn;
1276 			 Tgh = VFNMS(LDK(KP923879532), Tgg, Tgf);
1277 			 Thc = VFNMS(LDK(KP923879532), TgH, TgG);
1278 			 Tgk = VFNMS(LDK(KP668178637), Tgj, Tgi);
1279 			 Tgn = VFNMS(LDK(KP668178637), Tgm, Tgl);
1280 			 Tgo = VADD(Tgk, Tgn);
1281 			 Thd = VSUB(Tgk, Tgn);
1282 			 Tgp = VFNMS(LDK(KP831469612), Tgo, Tgh);
1283 			 Tho = VFNMS(LDK(KP831469612), Thd, Thc);
1284 			 TgT = VFMA(LDK(KP831469612), Tgo, Tgh);
1285 			 The = VFMA(LDK(KP831469612), Thd, Thc);
1286 		    }
1287 		    {
1288 			 V TgI, Th1, TgL, Th2, TgJ, TgK;
1289 			 TgI = VFMA(LDK(KP923879532), TgH, TgG);
1290 			 Th1 = VFMA(LDK(KP923879532), Tgg, Tgf);
1291 			 TgJ = VFMA(LDK(KP668178637), Tgi, Tgj);
1292 			 TgK = VFMA(LDK(KP668178637), Tgl, Tgm);
1293 			 TgL = VSUB(TgJ, TgK);
1294 			 Th2 = VADD(TgJ, TgK);
1295 			 TgM = VFMA(LDK(KP831469612), TgL, TgI);
1296 			 Thl = VFNMS(LDK(KP831469612), Th2, Th1);
1297 			 TgW = VFNMS(LDK(KP831469612), TgL, TgI);
1298 			 Th3 = VFMA(LDK(KP831469612), Th2, Th1);
1299 		    }
1300 		    {
1301 			 V Tgw, TgN, TgD, TgO;
1302 			 {
1303 			      V Tgs, Tgv, Tgz, TgC;
1304 			      Tgs = VFNMS(LDK(KP923879532), Tgr, Tgq);
1305 			      Tgv = VFMA(LDK(KP923879532), Tgu, Tgt);
1306 			      Tgw = VFNMS(LDK(KP534511135), Tgv, Tgs);
1307 			      TgN = VFMA(LDK(KP534511135), Tgs, Tgv);
1308 			      Tgz = VFNMS(LDK(KP923879532), Tgy, Tgx);
1309 			      TgC = VFMA(LDK(KP923879532), TgB, TgA);
1310 			      TgD = VFNMS(LDK(KP534511135), TgC, Tgz);
1311 			      TgO = VFMA(LDK(KP534511135), Tgz, TgC);
1312 			 }
1313 			 TgE = VADD(Tgw, TgD);
1314 			 TgX = VSUB(Tgw, TgD);
1315 			 TgP = VSUB(TgN, TgO);
1316 			 TgU = VADD(TgN, TgO);
1317 		    }
1318 		    {
1319 			 V Th6, Thf, Th9, Thg;
1320 			 {
1321 			      V Th4, Th5, Th7, Th8;
1322 			      Th4 = VFMA(LDK(KP923879532), Tgr, Tgq);
1323 			      Th5 = VFNMS(LDK(KP923879532), Tgu, Tgt);
1324 			      Th6 = VFMA(LDK(KP303346683), Th5, Th4);
1325 			      Thf = VFNMS(LDK(KP303346683), Th4, Th5);
1326 			      Th7 = VFMA(LDK(KP923879532), Tgy, Tgx);
1327 			      Th8 = VFNMS(LDK(KP923879532), TgB, TgA);
1328 			      Th9 = VFMA(LDK(KP303346683), Th8, Th7);
1329 			      Thg = VFNMS(LDK(KP303346683), Th7, Th8);
1330 			 }
1331 			 Tha = VADD(Th6, Th9);
1332 			 Thp = VSUB(Th6, Th9);
1333 			 Thh = VSUB(Thf, Thg);
1334 			 Thm = VADD(Thf, Thg);
1335 		    }
1336 		    {
1337 			 V TgF, TgQ, Thn, Thq;
1338 			 TgF = VFNMS(LDK(KP881921264), TgE, Tgp);
1339 			 TgQ = VFNMS(LDK(KP881921264), TgP, TgM);
1340 			 ST(&(xo[WS(os, 54)]), VFNMSI(TgQ, TgF), ovs, &(xo[0]));
1341 			 ST(&(xo[WS(os, 74)]), VFMAI(TgQ, TgF), ovs, &(xo[0]));
1342 			 Thn = VFMA(LDK(KP956940335), Thm, Thl);
1343 			 Thq = VFNMS(LDK(KP956940335), Thp, Tho);
1344 			 ST(&(xo[WS(os, 38)]), VFNMSI(Thq, Thn), ovs, &(xo[0]));
1345 			 ST(&(xo[WS(os, 90)]), VFMAI(Thq, Thn), ovs, &(xo[0]));
1346 		    }
1347 		    {
1348 			 V Thr, Ths, TgR, TgS;
1349 			 Thr = VFNMS(LDK(KP956940335), Thm, Thl);
1350 			 Ths = VFMA(LDK(KP956940335), Thp, Tho);
1351 			 ST(&(xo[WS(os, 26)]), VFMAI(Ths, Thr), ovs, &(xo[0]));
1352 			 ST(&(xo[WS(os, 102)]), VFNMSI(Ths, Thr), ovs, &(xo[0]));
1353 			 TgR = VFMA(LDK(KP881921264), TgE, Tgp);
1354 			 TgS = VFMA(LDK(KP881921264), TgP, TgM);
1355 			 ST(&(xo[WS(os, 118)]), VFNMSI(TgS, TgR), ovs, &(xo[0]));
1356 			 ST(&(xo[WS(os, 10)]), VFMAI(TgS, TgR), ovs, &(xo[0]));
1357 		    }
1358 		    {
1359 			 V TgV, TgY, Thb, Thi;
1360 			 TgV = VFNMS(LDK(KP881921264), TgU, TgT);
1361 			 TgY = VFMA(LDK(KP881921264), TgX, TgW);
1362 			 ST(&(xo[WS(os, 42)]), VFMAI(TgY, TgV), ovs, &(xo[0]));
1363 			 ST(&(xo[WS(os, 86)]), VFNMSI(TgY, TgV), ovs, &(xo[0]));
1364 			 Thb = VFNMS(LDK(KP956940335), Tha, Th3);
1365 			 Thi = VFNMS(LDK(KP956940335), Thh, The);
1366 			 ST(&(xo[WS(os, 70)]), VFNMSI(Thi, Thb), ovs, &(xo[0]));
1367 			 ST(&(xo[WS(os, 58)]), VFMAI(Thi, Thb), ovs, &(xo[0]));
1368 		    }
1369 		    {
1370 			 V Thj, Thk, TgZ, Th0;
1371 			 Thj = VFMA(LDK(KP956940335), Tha, Th3);
1372 			 Thk = VFMA(LDK(KP956940335), Thh, The);
1373 			 ST(&(xo[WS(os, 6)]), VFNMSI(Thk, Thj), ovs, &(xo[0]));
1374 			 ST(&(xo[WS(os, 122)]), VFMAI(Thk, Thj), ovs, &(xo[0]));
1375 			 TgZ = VFMA(LDK(KP881921264), TgU, TgT);
1376 			 Th0 = VFNMS(LDK(KP881921264), TgX, TgW);
1377 			 ST(&(xo[WS(os, 22)]), VFNMSI(Th0, TgZ), ovs, &(xo[0]));
1378 			 ST(&(xo[WS(os, 106)]), VFMAI(Th0, TgZ), ovs, &(xo[0]));
1379 		    }
1380 	       }
1381 	       {
1382 		    V T80, T8n, T8f, T8k, T8A, T8P, T8H, T8M, T7n, T8L, T8O, T8c, T8j, T8t, T8E;
1383 		    V T8m;
1384 		    {
1385 			 V T7G, T8d, T7Z, T8e;
1386 			 {
1387 			      V T7y, T7F, T7R, T7Y;
1388 			      T7y = VFMA(LDK(KP831469612), T7x, T7q);
1389 			      T7F = VFMA(LDK(KP831469612), T7E, T7B);
1390 			      T7G = VFMA(LDK(KP148335987), T7F, T7y);
1391 			      T8d = VFNMS(LDK(KP148335987), T7y, T7F);
1392 			      T7R = VFMA(LDK(KP831469612), T7Q, T7J);
1393 			      T7Y = VFMA(LDK(KP831469612), T7X, T7U);
1394 			      T7Z = VFMA(LDK(KP148335987), T7Y, T7R);
1395 			      T8e = VFNMS(LDK(KP148335987), T7R, T7Y);
1396 			 }
1397 			 T80 = VADD(T7G, T7Z);
1398 			 T8n = VSUB(T7G, T7Z);
1399 			 T8f = VSUB(T8d, T8e);
1400 			 T8k = VADD(T8d, T8e);
1401 		    }
1402 		    {
1403 			 V T8w, T8F, T8z, T8G;
1404 			 {
1405 			      V T8u, T8v, T8x, T8y;
1406 			      T8u = VFNMS(LDK(KP831469612), T7x, T7q);
1407 			      T8v = VFNMS(LDK(KP831469612), T7E, T7B);
1408 			      T8w = VFNMS(LDK(KP741650546), T8v, T8u);
1409 			      T8F = VFMA(LDK(KP741650546), T8u, T8v);
1410 			      T8x = VFNMS(LDK(KP831469612), T7Q, T7J);
1411 			      T8y = VFNMS(LDK(KP831469612), T7X, T7U);
1412 			      T8z = VFNMS(LDK(KP741650546), T8y, T8x);
1413 			      T8G = VFMA(LDK(KP741650546), T8x, T8y);
1414 			 }
1415 			 T8A = VADD(T8w, T8z);
1416 			 T8P = VSUB(T8w, T8z);
1417 			 T8H = VSUB(T8F, T8G);
1418 			 T8M = VADD(T8F, T8G);
1419 		    }
1420 		    {
1421 			 V T77, T8r, T88, T8C, T7m, T8D, T8b, T8s, T76, T87;
1422 			 T76 = VADD(T72, T75);
1423 			 T77 = VFMA(LDK(KP831469612), T76, T6Z);
1424 			 T8r = VFNMS(LDK(KP831469612), T76, T6Z);
1425 			 T87 = VSUB(T85, T86);
1426 			 T88 = VFMA(LDK(KP831469612), T87, T84);
1427 			 T8C = VFNMS(LDK(KP831469612), T87, T84);
1428 			 {
1429 			      V T7e, T7l, T89, T8a;
1430 			      T7e = VFMA(LDK(KP303346683), T7d, T7a);
1431 			      T7l = VFMA(LDK(KP303346683), T7k, T7h);
1432 			      T7m = VADD(T7e, T7l);
1433 			      T8D = VSUB(T7e, T7l);
1434 			      T89 = VFNMS(LDK(KP303346683), T7a, T7d);
1435 			      T8a = VFNMS(LDK(KP303346683), T7h, T7k);
1436 			      T8b = VSUB(T89, T8a);
1437 			      T8s = VADD(T89, T8a);
1438 			 }
1439 			 T7n = VFMA(LDK(KP956940335), T7m, T77);
1440 			 T8L = VFMA(LDK(KP956940335), T8s, T8r);
1441 			 T8O = VFNMS(LDK(KP956940335), T8D, T8C);
1442 			 T8c = VFMA(LDK(KP956940335), T8b, T88);
1443 			 T8j = VFNMS(LDK(KP956940335), T7m, T77);
1444 			 T8t = VFNMS(LDK(KP956940335), T8s, T8r);
1445 			 T8E = VFMA(LDK(KP956940335), T8D, T8C);
1446 			 T8m = VFNMS(LDK(KP956940335), T8b, T88);
1447 		    }
1448 		    {
1449 			 V T81, T8g, T8N, T8Q;
1450 			 T81 = VFNMS(LDK(KP989176509), T80, T7n);
1451 			 T8g = VFNMS(LDK(KP989176509), T8f, T8c);
1452 			 ST(&(xo[WS(os, 67)]), VFNMSI(T8g, T81), ovs, &(xo[WS(os, 1)]));
1453 			 ST(&(xo[WS(os, 61)]), VFMAI(T8g, T81), ovs, &(xo[WS(os, 1)]));
1454 			 T8N = VFNMS(LDK(KP803207531), T8M, T8L);
1455 			 T8Q = VFMA(LDK(KP803207531), T8P, T8O);
1456 			 ST(&(xo[WS(os, 45)]), VFMAI(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
1457 			 ST(&(xo[WS(os, 83)]), VFNMSI(T8Q, T8N), ovs, &(xo[WS(os, 1)]));
1458 		    }
1459 		    {
1460 			 V T8R, T8S, T8h, T8i;
1461 			 T8R = VFMA(LDK(KP803207531), T8M, T8L);
1462 			 T8S = VFNMS(LDK(KP803207531), T8P, T8O);
1463 			 ST(&(xo[WS(os, 19)]), VFNMSI(T8S, T8R), ovs, &(xo[WS(os, 1)]));
1464 			 ST(&(xo[WS(os, 109)]), VFMAI(T8S, T8R), ovs, &(xo[WS(os, 1)]));
1465 			 T8h = VFMA(LDK(KP989176509), T80, T7n);
1466 			 T8i = VFMA(LDK(KP989176509), T8f, T8c);
1467 			 ST(&(xo[WS(os, 3)]), VFNMSI(T8i, T8h), ovs, &(xo[WS(os, 1)]));
1468 			 ST(&(xo[WS(os, 125)]), VFMAI(T8i, T8h), ovs, &(xo[WS(os, 1)]));
1469 		    }
1470 		    {
1471 			 V T8l, T8o, T8B, T8I;
1472 			 T8l = VFMA(LDK(KP989176509), T8k, T8j);
1473 			 T8o = VFNMS(LDK(KP989176509), T8n, T8m);
1474 			 ST(&(xo[WS(os, 35)]), VFNMSI(T8o, T8l), ovs, &(xo[WS(os, 1)]));
1475 			 ST(&(xo[WS(os, 93)]), VFMAI(T8o, T8l), ovs, &(xo[WS(os, 1)]));
1476 			 T8B = VFNMS(LDK(KP803207531), T8A, T8t);
1477 			 T8I = VFNMS(LDK(KP803207531), T8H, T8E);
1478 			 ST(&(xo[WS(os, 51)]), VFNMSI(T8I, T8B), ovs, &(xo[WS(os, 1)]));
1479 			 ST(&(xo[WS(os, 77)]), VFMAI(T8I, T8B), ovs, &(xo[WS(os, 1)]));
1480 		    }
1481 		    {
1482 			 V T8J, T8K, T8p, T8q;
1483 			 T8J = VFMA(LDK(KP803207531), T8A, T8t);
1484 			 T8K = VFMA(LDK(KP803207531), T8H, T8E);
1485 			 ST(&(xo[WS(os, 115)]), VFNMSI(T8K, T8J), ovs, &(xo[WS(os, 1)]));
1486 			 ST(&(xo[WS(os, 13)]), VFMAI(T8K, T8J), ovs, &(xo[WS(os, 1)]));
1487 			 T8p = VFNMS(LDK(KP989176509), T8k, T8j);
1488 			 T8q = VFMA(LDK(KP989176509), T8n, T8m);
1489 			 ST(&(xo[WS(os, 29)]), VFMAI(T8q, T8p), ovs, &(xo[WS(os, 1)]));
1490 			 ST(&(xo[WS(os, 99)]), VFNMSI(T8q, T8p), ovs, &(xo[WS(os, 1)]));
1491 		    }
1492 	       }
1493 	       {
1494 		    V T4K, T5d, T55, T5a, T5q, T5F, T5x, T5C, T2f, T5B, T5E, T52, T59, T5j, T5u;
1495 		    V T5c;
1496 		    {
1497 			 V T3u, T53, T4J, T54;
1498 			 {
1499 			      V T3g, T3t, T4v, T4I;
1500 			      T3g = VFMA(LDK(KP980785280), T3f, T2G);
1501 			      T3t = VFMA(LDK(KP980785280), T3s, T3p);
1502 			      T3u = VFNMS(LDK(KP049126849), T3t, T3g);
1503 			      T53 = VFMA(LDK(KP049126849), T3g, T3t);
1504 			      T4v = VFMA(LDK(KP980785280), T4u, T3V);
1505 			      T4I = VFMA(LDK(KP980785280), T4H, T4E);
1506 			      T4J = VFNMS(LDK(KP049126849), T4I, T4v);
1507 			      T54 = VFMA(LDK(KP049126849), T4v, T4I);
1508 			 }
1509 			 T4K = VADD(T3u, T4J);
1510 			 T5d = VSUB(T3u, T4J);
1511 			 T55 = VSUB(T53, T54);
1512 			 T5a = VADD(T53, T54);
1513 		    }
1514 		    {
1515 			 V T5m, T5v, T5p, T5w;
1516 			 {
1517 			      V T5k, T5l, T5n, T5o;
1518 			      T5k = VFNMS(LDK(KP980785280), T3f, T2G);
1519 			      T5l = VFNMS(LDK(KP980785280), T3s, T3p);
1520 			      T5m = VFMA(LDK(KP906347169), T5l, T5k);
1521 			      T5v = VFNMS(LDK(KP906347169), T5k, T5l);
1522 			      T5n = VFNMS(LDK(KP980785280), T4u, T3V);
1523 			      T5o = VFNMS(LDK(KP980785280), T4H, T4E);
1524 			      T5p = VFMA(LDK(KP906347169), T5o, T5n);
1525 			      T5w = VFNMS(LDK(KP906347169), T5n, T5o);
1526 			 }
1527 			 T5q = VADD(T5m, T5p);
1528 			 T5F = VSUB(T5m, T5p);
1529 			 T5x = VSUB(T5v, T5w);
1530 			 T5C = VADD(T5v, T5w);
1531 		    }
1532 		    {
1533 			 V T11, T5h, T4Y, T5s, T2e, T5t, T51, T5i, T10, T4X;
1534 			 T10 = VADD(TI, TZ);
1535 			 T11 = VFMA(LDK(KP980785280), T10, Tr);
1536 			 T5h = VFNMS(LDK(KP980785280), T10, Tr);
1537 			 T4X = VSUB(T4V, T4W);
1538 			 T4Y = VFMA(LDK(KP980785280), T4X, T4U);
1539 			 T5s = VFNMS(LDK(KP980785280), T4X, T4U);
1540 			 {
1541 			      V T1C, T2d, T4Z, T50;
1542 			      T1C = VFNMS(LDK(KP098491403), T1B, T1s);
1543 			      T2d = VFNMS(LDK(KP098491403), T2c, T23);
1544 			      T2e = VADD(T1C, T2d);
1545 			      T5t = VSUB(T1C, T2d);
1546 			      T4Z = VFMA(LDK(KP098491403), T1s, T1B);
1547 			      T50 = VFMA(LDK(KP098491403), T23, T2c);
1548 			      T51 = VSUB(T4Z, T50);
1549 			      T5i = VADD(T4Z, T50);
1550 			 }
1551 			 T2f = VFMA(LDK(KP995184726), T2e, T11);
1552 			 T5B = VFNMS(LDK(KP995184726), T5i, T5h);
1553 			 T5E = VFMA(LDK(KP995184726), T5t, T5s);
1554 			 T52 = VFMA(LDK(KP995184726), T51, T4Y);
1555 			 T59 = VFNMS(LDK(KP995184726), T2e, T11);
1556 			 T5j = VFMA(LDK(KP995184726), T5i, T5h);
1557 			 T5u = VFNMS(LDK(KP995184726), T5t, T5s);
1558 			 T5c = VFNMS(LDK(KP995184726), T51, T4Y);
1559 		    }
1560 		    {
1561 			 V T4L, T56, T5D, T5G;
1562 			 T4L = VFNMS(LDK(KP998795456), T4K, T2f);
1563 			 T56 = VFNMS(LDK(KP998795456), T55, T52);
1564 			 ST(&(xo[WS(os, 63)]), VFNMSI(T56, T4L), ovs, &(xo[WS(os, 1)]));
1565 			 ST(&(xo[WS(os, 65)]), VFMAI(T56, T4L), ovs, &(xo[WS(os, 1)]));
1566 			 T5D = VFMA(LDK(KP740951125), T5C, T5B);
1567 			 T5G = VFNMS(LDK(KP740951125), T5F, T5E);
1568 			 ST(&(xo[WS(os, 47)]), VFNMSI(T5G, T5D), ovs, &(xo[WS(os, 1)]));
1569 			 ST(&(xo[WS(os, 81)]), VFMAI(T5G, T5D), ovs, &(xo[WS(os, 1)]));
1570 		    }
1571 		    {
1572 			 V T5H, T5I, T57, T58;
1573 			 T5H = VFNMS(LDK(KP740951125), T5C, T5B);
1574 			 T5I = VFMA(LDK(KP740951125), T5F, T5E);
1575 			 ST(&(xo[WS(os, 17)]), VFMAI(T5I, T5H), ovs, &(xo[WS(os, 1)]));
1576 			 ST(&(xo[WS(os, 111)]), VFNMSI(T5I, T5H), ovs, &(xo[WS(os, 1)]));
1577 			 T57 = VFMA(LDK(KP998795456), T4K, T2f);
1578 			 T58 = VFMA(LDK(KP998795456), T55, T52);
1579 			 ST(&(xo[WS(os, 127)]), VFNMSI(T58, T57), ovs, &(xo[WS(os, 1)]));
1580 			 ST(&(xo[WS(os, 1)]), VFMAI(T58, T57), ovs, &(xo[WS(os, 1)]));
1581 		    }
1582 		    {
1583 			 V T5b, T5e, T5r, T5y;
1584 			 T5b = VFNMS(LDK(KP998795456), T5a, T59);
1585 			 T5e = VFMA(LDK(KP998795456), T5d, T5c);
1586 			 ST(&(xo[WS(os, 33)]), VFMAI(T5e, T5b), ovs, &(xo[WS(os, 1)]));
1587 			 ST(&(xo[WS(os, 95)]), VFNMSI(T5e, T5b), ovs, &(xo[WS(os, 1)]));
1588 			 T5r = VFNMS(LDK(KP740951125), T5q, T5j);
1589 			 T5y = VFNMS(LDK(KP740951125), T5x, T5u);
1590 			 ST(&(xo[WS(os, 79)]), VFNMSI(T5y, T5r), ovs, &(xo[WS(os, 1)]));
1591 			 ST(&(xo[WS(os, 49)]), VFMAI(T5y, T5r), ovs, &(xo[WS(os, 1)]));
1592 		    }
1593 		    {
1594 			 V T5z, T5A, T5f, T5g;
1595 			 T5z = VFMA(LDK(KP740951125), T5q, T5j);
1596 			 T5A = VFMA(LDK(KP740951125), T5x, T5u);
1597 			 ST(&(xo[WS(os, 15)]), VFNMSI(T5A, T5z), ovs, &(xo[WS(os, 1)]));
1598 			 ST(&(xo[WS(os, 113)]), VFMAI(T5A, T5z), ovs, &(xo[WS(os, 1)]));
1599 			 T5f = VFMA(LDK(KP998795456), T5a, T59);
1600 			 T5g = VFNMS(LDK(KP998795456), T5d, T5c);
1601 			 ST(&(xo[WS(os, 31)]), VFNMSI(T5g, T5f), ovs, &(xo[WS(os, 1)]));
1602 			 ST(&(xo[WS(os, 97)]), VFMAI(T5g, T5f), ovs, &(xo[WS(os, 1)]));
1603 		    }
1604 	       }
1605 	       {
1606 		    V T9i, T9B, T9t, T9y, T9O, Ta3, T9V, Ta0, T93, T9Z, Ta2, T9q, T9x, T9H, T9S;
1607 		    V T9A;
1608 		    {
1609 			 V T9a, T9r, T9h, T9s;
1610 			 {
1611 			      V T96, T99, T9d, T9g;
1612 			      T96 = VFNMS(LDK(KP831469612), T95, T94);
1613 			      T99 = VFNMS(LDK(KP831469612), T98, T97);
1614 			      T9a = VFMA(LDK(KP599376933), T99, T96);
1615 			      T9r = VFNMS(LDK(KP599376933), T96, T99);
1616 			      T9d = VFNMS(LDK(KP831469612), T9c, T9b);
1617 			      T9g = VFNMS(LDK(KP831469612), T9f, T9e);
1618 			      T9h = VFMA(LDK(KP599376933), T9g, T9d);
1619 			      T9s = VFNMS(LDK(KP599376933), T9d, T9g);
1620 			 }
1621 			 T9i = VADD(T9a, T9h);
1622 			 T9B = VSUB(T9a, T9h);
1623 			 T9t = VSUB(T9r, T9s);
1624 			 T9y = VADD(T9r, T9s);
1625 		    }
1626 		    {
1627 			 V T9K, T9T, T9N, T9U;
1628 			 {
1629 			      V T9I, T9J, T9L, T9M;
1630 			      T9I = VFMA(LDK(KP831469612), T95, T94);
1631 			      T9J = VFMA(LDK(KP831469612), T98, T97);
1632 			      T9K = VFNMS(LDK(KP250486960), T9J, T9I);
1633 			      T9T = VFMA(LDK(KP250486960), T9I, T9J);
1634 			      T9L = VFMA(LDK(KP831469612), T9c, T9b);
1635 			      T9M = VFMA(LDK(KP831469612), T9f, T9e);
1636 			      T9N = VFNMS(LDK(KP250486960), T9M, T9L);
1637 			      T9U = VFMA(LDK(KP250486960), T9L, T9M);
1638 			 }
1639 			 T9O = VADD(T9K, T9N);
1640 			 Ta3 = VSUB(T9K, T9N);
1641 			 T9V = VSUB(T9T, T9U);
1642 			 Ta0 = VADD(T9T, T9U);
1643 		    }
1644 		    {
1645 			 V T8V, T9F, T9m, T9Q, T92, T9R, T9p, T9G, T8U, T9l;
1646 			 T8U = VADD(T85, T86);
1647 			 T8V = VFMA(LDK(KP831469612), T8U, T8T);
1648 			 T9F = VFNMS(LDK(KP831469612), T8U, T8T);
1649 			 T9l = VSUB(T72, T75);
1650 			 T9m = VFNMS(LDK(KP831469612), T9l, T9k);
1651 			 T9Q = VFMA(LDK(KP831469612), T9l, T9k);
1652 			 {
1653 			      V T8Y, T91, T9n, T9o;
1654 			      T8Y = VFMA(LDK(KP534511135), T8X, T8W);
1655 			      T91 = VFMA(LDK(KP534511135), T90, T8Z);
1656 			      T92 = VADD(T8Y, T91);
1657 			      T9R = VSUB(T8Y, T91);
1658 			      T9n = VFNMS(LDK(KP534511135), T8W, T8X);
1659 			      T9o = VFNMS(LDK(KP534511135), T8Z, T90);
1660 			      T9p = VSUB(T9n, T9o);
1661 			      T9G = VADD(T9n, T9o);
1662 			 }
1663 			 T93 = VFMA(LDK(KP881921264), T92, T8V);
1664 			 T9Z = VFNMS(LDK(KP881921264), T9G, T9F);
1665 			 Ta2 = VFNMS(LDK(KP881921264), T9R, T9Q);
1666 			 T9q = VFNMS(LDK(KP881921264), T9p, T9m);
1667 			 T9x = VFNMS(LDK(KP881921264), T92, T8V);
1668 			 T9H = VFMA(LDK(KP881921264), T9G, T9F);
1669 			 T9S = VFMA(LDK(KP881921264), T9R, T9Q);
1670 			 T9A = VFMA(LDK(KP881921264), T9p, T9m);
1671 		    }
1672 		    {
1673 			 V T9j, T9u, Ta1, Ta4;
1674 			 T9j = VFNMS(LDK(KP857728610), T9i, T93);
1675 			 T9u = VFNMS(LDK(KP857728610), T9t, T9q);
1676 			 ST(&(xo[WS(os, 75)]), VFNMSI(T9u, T9j), ovs, &(xo[WS(os, 1)]));
1677 			 ST(&(xo[WS(os, 53)]), VFMAI(T9u, T9j), ovs, &(xo[WS(os, 1)]));
1678 			 Ta1 = VFNMS(LDK(KP970031253), Ta0, T9Z);
1679 			 Ta4 = VFMA(LDK(KP970031253), Ta3, Ta2);
1680 			 ST(&(xo[WS(os, 37)]), VFMAI(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
1681 			 ST(&(xo[WS(os, 91)]), VFNMSI(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
1682 		    }
1683 		    {
1684 			 V Ta5, Ta6, T9v, T9w;
1685 			 Ta5 = VFMA(LDK(KP970031253), Ta0, T9Z);
1686 			 Ta6 = VFNMS(LDK(KP970031253), Ta3, Ta2);
1687 			 ST(&(xo[WS(os, 27)]), VFNMSI(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
1688 			 ST(&(xo[WS(os, 101)]), VFMAI(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
1689 			 T9v = VFMA(LDK(KP857728610), T9i, T93);
1690 			 T9w = VFMA(LDK(KP857728610), T9t, T9q);
1691 			 ST(&(xo[WS(os, 11)]), VFNMSI(T9w, T9v), ovs, &(xo[WS(os, 1)]));
1692 			 ST(&(xo[WS(os, 117)]), VFMAI(T9w, T9v), ovs, &(xo[WS(os, 1)]));
1693 		    }
1694 		    {
1695 			 V T9z, T9C, T9P, T9W;
1696 			 T9z = VFMA(LDK(KP857728610), T9y, T9x);
1697 			 T9C = VFNMS(LDK(KP857728610), T9B, T9A);
1698 			 ST(&(xo[WS(os, 43)]), VFNMSI(T9C, T9z), ovs, &(xo[WS(os, 1)]));
1699 			 ST(&(xo[WS(os, 85)]), VFMAI(T9C, T9z), ovs, &(xo[WS(os, 1)]));
1700 			 T9P = VFNMS(LDK(KP970031253), T9O, T9H);
1701 			 T9W = VFNMS(LDK(KP970031253), T9V, T9S);
1702 			 ST(&(xo[WS(os, 59)]), VFNMSI(T9W, T9P), ovs, &(xo[WS(os, 1)]));
1703 			 ST(&(xo[WS(os, 69)]), VFMAI(T9W, T9P), ovs, &(xo[WS(os, 1)]));
1704 		    }
1705 		    {
1706 			 V T9X, T9Y, T9D, T9E;
1707 			 T9X = VFMA(LDK(KP970031253), T9O, T9H);
1708 			 T9Y = VFMA(LDK(KP970031253), T9V, T9S);
1709 			 ST(&(xo[WS(os, 123)]), VFNMSI(T9Y, T9X), ovs, &(xo[WS(os, 1)]));
1710 			 ST(&(xo[WS(os, 5)]), VFMAI(T9Y, T9X), ovs, &(xo[WS(os, 1)]));
1711 			 T9D = VFNMS(LDK(KP857728610), T9y, T9x);
1712 			 T9E = VFMA(LDK(KP857728610), T9B, T9A);
1713 			 ST(&(xo[WS(os, 21)]), VFMAI(T9E, T9D), ovs, &(xo[WS(os, 1)]));
1714 			 ST(&(xo[WS(os, 107)]), VFNMSI(T9E, T9D), ovs, &(xo[WS(os, 1)]));
1715 		    }
1716 	       }
1717 	       {
1718 		    V T68, T6r, T6j, T6o, T6E, T6T, T6L, T6Q, T5T, T6P, T6S, T6g, T6n, T6x, T6I;
1719 		    V T6q;
1720 		    {
1721 			 V T60, T6h, T67, T6i;
1722 			 {
1723 			      V T5W, T5Z, T63, T66;
1724 			      T5W = VFNMS(LDK(KP980785280), T5V, T5U);
1725 			      T5Z = VFMA(LDK(KP980785280), T5Y, T5X);
1726 			      T60 = VFNMS(LDK(KP472964775), T5Z, T5W);
1727 			      T6h = VFMA(LDK(KP472964775), T5W, T5Z);
1728 			      T63 = VFNMS(LDK(KP980785280), T62, T61);
1729 			      T66 = VFMA(LDK(KP980785280), T65, T64);
1730 			      T67 = VFNMS(LDK(KP472964775), T66, T63);
1731 			      T6i = VFMA(LDK(KP472964775), T63, T66);
1732 			 }
1733 			 T68 = VADD(T60, T67);
1734 			 T6r = VSUB(T60, T67);
1735 			 T6j = VSUB(T6h, T6i);
1736 			 T6o = VADD(T6h, T6i);
1737 		    }
1738 		    {
1739 			 V T6A, T6J, T6D, T6K;
1740 			 {
1741 			      V T6y, T6z, T6B, T6C;
1742 			      T6y = VFMA(LDK(KP980785280), T5V, T5U);
1743 			      T6z = VFNMS(LDK(KP980785280), T5Y, T5X);
1744 			      T6A = VFMA(LDK(KP357805721), T6z, T6y);
1745 			      T6J = VFNMS(LDK(KP357805721), T6y, T6z);
1746 			      T6B = VFMA(LDK(KP980785280), T62, T61);
1747 			      T6C = VFNMS(LDK(KP980785280), T65, T64);
1748 			      T6D = VFMA(LDK(KP357805721), T6C, T6B);
1749 			      T6K = VFNMS(LDK(KP357805721), T6B, T6C);
1750 			 }
1751 			 T6E = VADD(T6A, T6D);
1752 			 T6T = VSUB(T6A, T6D);
1753 			 T6L = VSUB(T6J, T6K);
1754 			 T6Q = VADD(T6J, T6K);
1755 		    }
1756 		    {
1757 			 V T5L, T6v, T6c, T6G, T5S, T6H, T6f, T6w, T5K, T6b;
1758 			 T5K = VADD(T4V, T4W);
1759 			 T5L = VFNMS(LDK(KP980785280), T5K, T5J);
1760 			 T6v = VFMA(LDK(KP980785280), T5K, T5J);
1761 			 T6b = VSUB(TI, TZ);
1762 			 T6c = VFMA(LDK(KP980785280), T6b, T6a);
1763 			 T6G = VFNMS(LDK(KP980785280), T6b, T6a);
1764 			 {
1765 			      V T5O, T5R, T6d, T6e;
1766 			      T5O = VFNMS(LDK(KP820678790), T5N, T5M);
1767 			      T5R = VFNMS(LDK(KP820678790), T5Q, T5P);
1768 			      T5S = VADD(T5O, T5R);
1769 			      T6H = VSUB(T5O, T5R);
1770 			      T6d = VFMA(LDK(KP820678790), T5M, T5N);
1771 			      T6e = VFMA(LDK(KP820678790), T5P, T5Q);
1772 			      T6f = VSUB(T6d, T6e);
1773 			      T6w = VADD(T6d, T6e);
1774 			 }
1775 			 T5T = VFNMS(LDK(KP773010453), T5S, T5L);
1776 			 T6P = VFNMS(LDK(KP773010453), T6w, T6v);
1777 			 T6S = VFNMS(LDK(KP773010453), T6H, T6G);
1778 			 T6g = VFMA(LDK(KP773010453), T6f, T6c);
1779 			 T6n = VFMA(LDK(KP773010453), T5S, T5L);
1780 			 T6x = VFMA(LDK(KP773010453), T6w, T6v);
1781 			 T6I = VFMA(LDK(KP773010453), T6H, T6G);
1782 			 T6q = VFNMS(LDK(KP773010453), T6f, T6c);
1783 		    }
1784 		    {
1785 			 V T69, T6k, T6R, T6U;
1786 			 T69 = VFNMS(LDK(KP903989293), T68, T5T);
1787 			 T6k = VFNMS(LDK(KP903989293), T6j, T6g);
1788 			 ST(&(xo[WS(os, 55)]), VFNMSI(T6k, T69), ovs, &(xo[WS(os, 1)]));
1789 			 ST(&(xo[WS(os, 73)]), VFMAI(T6k, T69), ovs, &(xo[WS(os, 1)]));
1790 			 T6R = VFMA(LDK(KP941544065), T6Q, T6P);
1791 			 T6U = VFNMS(LDK(KP941544065), T6T, T6S);
1792 			 ST(&(xo[WS(os, 39)]), VFNMSI(T6U, T6R), ovs, &(xo[WS(os, 1)]));
1793 			 ST(&(xo[WS(os, 89)]), VFMAI(T6U, T6R), ovs, &(xo[WS(os, 1)]));
1794 		    }
1795 		    {
1796 			 V T6V, T6W, T6l, T6m;
1797 			 T6V = VFNMS(LDK(KP941544065), T6Q, T6P);
1798 			 T6W = VFMA(LDK(KP941544065), T6T, T6S);
1799 			 ST(&(xo[WS(os, 25)]), VFMAI(T6W, T6V), ovs, &(xo[WS(os, 1)]));
1800 			 ST(&(xo[WS(os, 103)]), VFNMSI(T6W, T6V), ovs, &(xo[WS(os, 1)]));
1801 			 T6l = VFMA(LDK(KP903989293), T68, T5T);
1802 			 T6m = VFMA(LDK(KP903989293), T6j, T6g);
1803 			 ST(&(xo[WS(os, 119)]), VFNMSI(T6m, T6l), ovs, &(xo[WS(os, 1)]));
1804 			 ST(&(xo[WS(os, 9)]), VFMAI(T6m, T6l), ovs, &(xo[WS(os, 1)]));
1805 		    }
1806 		    {
1807 			 V T6p, T6s, T6F, T6M;
1808 			 T6p = VFNMS(LDK(KP903989293), T6o, T6n);
1809 			 T6s = VFMA(LDK(KP903989293), T6r, T6q);
1810 			 ST(&(xo[WS(os, 41)]), VFMAI(T6s, T6p), ovs, &(xo[WS(os, 1)]));
1811 			 ST(&(xo[WS(os, 87)]), VFNMSI(T6s, T6p), ovs, &(xo[WS(os, 1)]));
1812 			 T6F = VFNMS(LDK(KP941544065), T6E, T6x);
1813 			 T6M = VFNMS(LDK(KP941544065), T6L, T6I);
1814 			 ST(&(xo[WS(os, 71)]), VFNMSI(T6M, T6F), ovs, &(xo[WS(os, 1)]));
1815 			 ST(&(xo[WS(os, 57)]), VFMAI(T6M, T6F), ovs, &(xo[WS(os, 1)]));
1816 		    }
1817 		    {
1818 			 V T6N, T6O, T6t, T6u;
1819 			 T6N = VFMA(LDK(KP941544065), T6E, T6x);
1820 			 T6O = VFMA(LDK(KP941544065), T6L, T6I);
1821 			 ST(&(xo[WS(os, 7)]), VFNMSI(T6O, T6N), ovs, &(xo[WS(os, 1)]));
1822 			 ST(&(xo[WS(os, 121)]), VFMAI(T6O, T6N), ovs, &(xo[WS(os, 1)]));
1823 			 T6t = VFMA(LDK(KP903989293), T6o, T6n);
1824 			 T6u = VFNMS(LDK(KP903989293), T6r, T6q);
1825 			 ST(&(xo[WS(os, 23)]), VFNMSI(T6u, T6t), ovs, &(xo[WS(os, 1)]));
1826 			 ST(&(xo[WS(os, 105)]), VFMAI(T6u, T6t), ovs, &(xo[WS(os, 1)]));
1827 		    }
1828 	       }
1829 	  }
1830      }
1831      VLEAVE();
1832 }
1833 
1834 static const kdft_desc desc = { 128, XSIMD_STRING("n1bv_128"), { 440, 0, 642, 0 }, &GENUS, 0, 0, 0, 0 };
1835 
XSIMD(codelet_n1bv_128)1836 void XSIMD(codelet_n1bv_128) (planner *p) { X(kdft_register) (p, n1bv_128, &desc);
1837 }
1838 
1839 #else
1840 
1841 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 128 -name n1bv_128 -include dft/simd/n1b.h */
1842 
1843 /*
1844  * This function contains 1082 FP additions, 330 FP multiplications,
1845  * (or, 938 additions, 186 multiplications, 144 fused multiply/add),
1846  * 194 stack variables, 31 constants, and 256 memory accesses
1847  */
1848 #include "dft/simd/n1b.h"
1849 
n1bv_128(const R * ri,const R * ii,R * ro,R * io,stride is,stride os,INT v,INT ivs,INT ovs)1850 static void n1bv_128(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
1851 {
1852      DVK(KP146730474, +0.146730474455361751658850129646717819706215317);
1853      DVK(KP989176509, +0.989176509964780973451673738016243063983689533);
1854      DVK(KP595699304, +0.595699304492433343467036528829969889511926338);
1855      DVK(KP803207531, +0.803207531480644909806676512963141923879569427);
1856      DVK(KP049067674, +0.049067674327418014254954976942682658314745363);
1857      DVK(KP998795456, +0.998795456205172392714771604759100694443203615);
1858      DVK(KP671558954, +0.671558954847018400625376850427421803228750632);
1859      DVK(KP740951125, +0.740951125354959091175616897495162729728955309);
1860      DVK(KP514102744, +0.514102744193221726593693838968815772608049120);
1861      DVK(KP857728610, +0.857728610000272069902269984284770137042490799);
1862      DVK(KP242980179, +0.242980179903263889948274162077471118320990783);
1863      DVK(KP970031253, +0.970031253194543992603984207286100251456865962);
1864      DVK(KP427555093, +0.427555093430282094320966856888798534304578629);
1865      DVK(KP903989293, +0.903989293123443331586200297230537048710132025);
1866      DVK(KP336889853, +0.336889853392220050689253212619147570477766780);
1867      DVK(KP941544065, +0.941544065183020778412509402599502357185589796);
1868      DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
1869      DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
1870      DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
1871      DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
1872      DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
1873      DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
1874      DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
1875      DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
1876      DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
1877      DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
1878      DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
1879      DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
1880      DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
1881      DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
1882      DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
1883      {
1884 	  INT i;
1885 	  const R *xi;
1886 	  R *xo;
1887 	  xi = ii;
1888 	  xo = io;
1889 	  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)) {
1890 	       V T49, T6e, Tev, TgK, TfA, TgL, T4U, T5J, T7R, T9o, Tah, TdG, Tcw, TdB, T84;
1891 	       V T8T, Tfk, Tfo, T1G, T64, Tgs, Th6, T2p, T62, T7t, T9c, Tce, Tdm, T7i, T9e;
1892 	       V Tc8, Tdp, TgF, TgG, T4q, T4V, TeC, Tfx, T4H, T4W, T7X, T86, Tcr, TdH, T7U;
1893 	       V T85, Taw, TdC, Tf3, Tf7, Tr, T5X, Tgl, Th3, T1a, T5V, T7a, T95, TbD, Tdf;
1894 	       V T6Z, T97, Tbx, Tdi, Tgy, Tgz, TgA, TaN, Tdv, TeK, Tfu, T2W, T5M, T35, T5N;
1895 	       V T7F, T8X, TaI, Tdu, T7C, T8W, TgB, TgC, TgD, Tb4, Tdy, TeR, Tfv, T3x, T5P;
1896 	       V T3G, T5Q, T7M, T90, TaZ, Tdx, T7J, T8Z, Tbm, Tdg, TbG, Tdj, Tgo, Th4, Tf0;
1897 	       V Tf8, T76, T98, T7d, T94, T10, T5Y, T1d, T5U, TbX, Tdn, Tch, Tdq, Tgv, Th7;
1898 	       V Tfh, Tfp, T7p, T9f, T7w, T9b, T2f, T65, T2s, T61;
1899 	       {
1900 		    V T47, Ta8, T4O, Ta7, T44, Tcu, T4P, Tct, Taa, Tab, T3P, Tac, T4R, Tad, Tae;
1901 		    V T3W, Taf, T4S;
1902 		    {
1903 			 V T45, T46, T4M, T4N;
1904 			 T45 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
1905 			 T46 = LD(&(xi[WS(is, 96)]), ivs, &(xi[0]));
1906 			 T47 = VSUB(T45, T46);
1907 			 Ta8 = VADD(T45, T46);
1908 			 T4M = LD(&(xi[0]), ivs, &(xi[0]));
1909 			 T4N = LD(&(xi[WS(is, 64)]), ivs, &(xi[0]));
1910 			 T4O = VSUB(T4M, T4N);
1911 			 Ta7 = VADD(T4M, T4N);
1912 		    }
1913 		    {
1914 			 V T3Y, T3Z, T40, T41, T42, T43;
1915 			 T3Y = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
1916 			 T3Z = LD(&(xi[WS(is, 80)]), ivs, &(xi[0]));
1917 			 T40 = VSUB(T3Y, T3Z);
1918 			 T41 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0]));
1919 			 T42 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
1920 			 T43 = VSUB(T41, T42);
1921 			 T44 = VMUL(LDK(KP707106781), VSUB(T40, T43));
1922 			 Tcu = VADD(T41, T42);
1923 			 T4P = VMUL(LDK(KP707106781), VADD(T40, T43));
1924 			 Tct = VADD(T3Y, T3Z);
1925 		    }
1926 		    {
1927 			 V T3L, T3O, T3S, T3V;
1928 			 {
1929 			      V T3J, T3K, T3M, T3N;
1930 			      T3J = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
1931 			      T3K = LD(&(xi[WS(is, 72)]), ivs, &(xi[0]));
1932 			      T3L = VSUB(T3J, T3K);
1933 			      Taa = VADD(T3J, T3K);
1934 			      T3M = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
1935 			      T3N = LD(&(xi[WS(is, 104)]), ivs, &(xi[0]));
1936 			      T3O = VSUB(T3M, T3N);
1937 			      Tab = VADD(T3M, T3N);
1938 			 }
1939 			 T3P = VFNMS(LDK(KP382683432), T3O, VMUL(LDK(KP923879532), T3L));
1940 			 Tac = VSUB(Taa, Tab);
1941 			 T4R = VFMA(LDK(KP382683432), T3L, VMUL(LDK(KP923879532), T3O));
1942 			 {
1943 			      V T3Q, T3R, T3T, T3U;
1944 			      T3Q = LD(&(xi[WS(is, 120)]), ivs, &(xi[0]));
1945 			      T3R = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
1946 			      T3S = VSUB(T3Q, T3R);
1947 			      Tad = VADD(T3Q, T3R);
1948 			      T3T = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
1949 			      T3U = LD(&(xi[WS(is, 88)]), ivs, &(xi[0]));
1950 			      T3V = VSUB(T3T, T3U);
1951 			      Tae = VADD(T3T, T3U);
1952 			 }
1953 			 T3W = VFMA(LDK(KP923879532), T3S, VMUL(LDK(KP382683432), T3V));
1954 			 Taf = VSUB(Tad, Tae);
1955 			 T4S = VFNMS(LDK(KP382683432), T3S, VMUL(LDK(KP923879532), T3V));
1956 		    }
1957 		    {
1958 			 V T3X, T48, Tet, Teu;
1959 			 T3X = VSUB(T3P, T3W);
1960 			 T48 = VSUB(T44, T47);
1961 			 T49 = VSUB(T3X, T48);
1962 			 T6e = VADD(T48, T3X);
1963 			 Tet = VADD(Ta7, Ta8);
1964 			 Teu = VADD(Tct, Tcu);
1965 			 Tev = VSUB(Tet, Teu);
1966 			 TgK = VADD(Tet, Teu);
1967 		    }
1968 		    {
1969 			 V Tfy, Tfz, T4Q, T4T;
1970 			 Tfy = VADD(Taa, Tab);
1971 			 Tfz = VADD(Tad, Tae);
1972 			 TfA = VSUB(Tfy, Tfz);
1973 			 TgL = VADD(Tfy, Tfz);
1974 			 T4Q = VSUB(T4O, T4P);
1975 			 T4T = VSUB(T4R, T4S);
1976 			 T4U = VSUB(T4Q, T4T);
1977 			 T5J = VADD(T4Q, T4T);
1978 		    }
1979 		    {
1980 			 V T7P, T7Q, Ta9, Tag;
1981 			 T7P = VADD(T4R, T4S);
1982 			 T7Q = VADD(T47, T44);
1983 			 T7R = VSUB(T7P, T7Q);
1984 			 T9o = VADD(T7Q, T7P);
1985 			 Ta9 = VSUB(Ta7, Ta8);
1986 			 Tag = VMUL(LDK(KP707106781), VADD(Tac, Taf));
1987 			 Tah = VSUB(Ta9, Tag);
1988 			 TdG = VADD(Ta9, Tag);
1989 		    }
1990 		    {
1991 			 V Tcs, Tcv, T82, T83;
1992 			 Tcs = VMUL(LDK(KP707106781), VSUB(Tac, Taf));
1993 			 Tcv = VSUB(Tct, Tcu);
1994 			 Tcw = VSUB(Tcs, Tcv);
1995 			 TdB = VADD(Tcv, Tcs);
1996 			 T82 = VADD(T4O, T4P);
1997 			 T83 = VADD(T3P, T3W);
1998 			 T84 = VSUB(T82, T83);
1999 			 T8T = VADD(T82, T83);
2000 		    }
2001 	       }
2002 	       {
2003 		    V Tca, Tcb, T1i, Tfm, T2n, Tc5, Tc6, T1p, Tfn, T2k, T1x, Tfi, T2h, Tc0, T1E;
2004 		    V Tfj, T2i, Tc3, T1l, T1o, Tcc, Tcd;
2005 		    {
2006 			 V T1g, T1h, T2l, T2m;
2007 			 T1g = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)]));
2008 			 T1h = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
2009 			 Tca = VADD(T1g, T1h);
2010 			 T2l = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
2011 			 T2m = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)]));
2012 			 Tcb = VADD(T2l, T2m);
2013 			 T1i = VSUB(T1g, T1h);
2014 			 Tfm = VADD(Tca, Tcb);
2015 			 T2n = VSUB(T2l, T2m);
2016 		    }
2017 		    {
2018 			 V T1j, T1k, T1m, T1n;
2019 			 T1j = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
2020 			 T1k = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)]));
2021 			 T1l = VSUB(T1j, T1k);
2022 			 Tc5 = VADD(T1j, T1k);
2023 			 T1m = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)]));
2024 			 T1n = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
2025 			 T1o = VSUB(T1m, T1n);
2026 			 Tc6 = VADD(T1m, T1n);
2027 		    }
2028 		    T1p = VMUL(LDK(KP707106781), VADD(T1l, T1o));
2029 		    Tfn = VADD(Tc5, Tc6);
2030 		    T2k = VMUL(LDK(KP707106781), VSUB(T1l, T1o));
2031 		    {
2032 			 V T1t, TbY, T1w, TbZ;
2033 			 {
2034 			      V T1r, T1s, T1u, T1v;
2035 			      T1r = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
2036 			      T1s = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)]));
2037 			      T1t = VSUB(T1r, T1s);
2038 			      TbY = VADD(T1r, T1s);
2039 			      T1u = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
2040 			      T1v = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)]));
2041 			      T1w = VSUB(T1u, T1v);
2042 			      TbZ = VADD(T1u, T1v);
2043 			 }
2044 			 T1x = VFMA(LDK(KP382683432), T1t, VMUL(LDK(KP923879532), T1w));
2045 			 Tfi = VADD(TbY, TbZ);
2046 			 T2h = VFNMS(LDK(KP382683432), T1w, VMUL(LDK(KP923879532), T1t));
2047 			 Tc0 = VSUB(TbY, TbZ);
2048 		    }
2049 		    {
2050 			 V T1A, Tc2, T1D, Tc1;
2051 			 {
2052 			      V T1y, T1z, T1B, T1C;
2053 			      T1y = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
2054 			      T1z = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)]));
2055 			      T1A = VSUB(T1y, T1z);
2056 			      Tc2 = VADD(T1y, T1z);
2057 			      T1B = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)]));
2058 			      T1C = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
2059 			      T1D = VSUB(T1B, T1C);
2060 			      Tc1 = VADD(T1B, T1C);
2061 			 }
2062 			 T1E = VFNMS(LDK(KP382683432), T1D, VMUL(LDK(KP923879532), T1A));
2063 			 Tfj = VADD(Tc1, Tc2);
2064 			 T2i = VFMA(LDK(KP923879532), T1D, VMUL(LDK(KP382683432), T1A));
2065 			 Tc3 = VSUB(Tc1, Tc2);
2066 		    }
2067 		    Tfk = VSUB(Tfi, Tfj);
2068 		    Tfo = VSUB(Tfm, Tfn);
2069 		    {
2070 			 V T1q, T1F, Tgq, Tgr;
2071 			 T1q = VSUB(T1i, T1p);
2072 			 T1F = VSUB(T1x, T1E);
2073 			 T1G = VSUB(T1q, T1F);
2074 			 T64 = VADD(T1q, T1F);
2075 			 Tgq = VADD(Tfm, Tfn);
2076 			 Tgr = VADD(Tfi, Tfj);
2077 			 Tgs = VSUB(Tgq, Tgr);
2078 			 Th6 = VADD(Tgq, Tgr);
2079 		    }
2080 		    {
2081 			 V T2j, T2o, T7r, T7s;
2082 			 T2j = VSUB(T2h, T2i);
2083 			 T2o = VSUB(T2k, T2n);
2084 			 T2p = VSUB(T2j, T2o);
2085 			 T62 = VADD(T2o, T2j);
2086 			 T7r = VADD(T1x, T1E);
2087 			 T7s = VADD(T2n, T2k);
2088 			 T7t = VSUB(T7r, T7s);
2089 			 T9c = VADD(T7s, T7r);
2090 		    }
2091 		    Tcc = VSUB(Tca, Tcb);
2092 		    Tcd = VMUL(LDK(KP707106781), VADD(Tc0, Tc3));
2093 		    Tce = VSUB(Tcc, Tcd);
2094 		    Tdm = VADD(Tcc, Tcd);
2095 		    {
2096 			 V T7g, T7h, Tc4, Tc7;
2097 			 T7g = VADD(T1i, T1p);
2098 			 T7h = VADD(T2h, T2i);
2099 			 T7i = VSUB(T7g, T7h);
2100 			 T9e = VADD(T7g, T7h);
2101 			 Tc4 = VMUL(LDK(KP707106781), VSUB(Tc0, Tc3));
2102 			 Tc7 = VSUB(Tc5, Tc6);
2103 			 Tc8 = VSUB(Tc4, Tc7);
2104 			 Tdp = VADD(Tc7, Tc4);
2105 		    }
2106 	       }
2107 	       {
2108 		    V T4c, Tew, T4o, Tak, T4A, Tez, T4E, Tau, T4j, Tex, T4l, Tan, T4x, TeA, T4F;
2109 		    V Tar, Tcp, Tcq;
2110 		    {
2111 			 V T4a, T4b, Tai, T4m, T4n, Taj;
2112 			 T4a = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
2113 			 T4b = LD(&(xi[WS(is, 68)]), ivs, &(xi[0]));
2114 			 Tai = VADD(T4a, T4b);
2115 			 T4m = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
2116 			 T4n = LD(&(xi[WS(is, 100)]), ivs, &(xi[0]));
2117 			 Taj = VADD(T4m, T4n);
2118 			 T4c = VSUB(T4a, T4b);
2119 			 Tew = VADD(Tai, Taj);
2120 			 T4o = VSUB(T4m, T4n);
2121 			 Tak = VSUB(Tai, Taj);
2122 		    }
2123 		    {
2124 			 V T4y, T4z, Tat, T4C, T4D, Tas;
2125 			 T4y = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
2126 			 T4z = LD(&(xi[WS(is, 92)]), ivs, &(xi[0]));
2127 			 Tat = VADD(T4y, T4z);
2128 			 T4C = LD(&(xi[WS(is, 124)]), ivs, &(xi[0]));
2129 			 T4D = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
2130 			 Tas = VADD(T4C, T4D);
2131 			 T4A = VSUB(T4y, T4z);
2132 			 Tez = VADD(Tas, Tat);
2133 			 T4E = VSUB(T4C, T4D);
2134 			 Tau = VSUB(Tas, Tat);
2135 		    }
2136 		    {
2137 			 V T4f, Tal, T4i, Tam;
2138 			 {
2139 			      V T4d, T4e, T4g, T4h;
2140 			      T4d = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
2141 			      T4e = LD(&(xi[WS(is, 84)]), ivs, &(xi[0]));
2142 			      T4f = VSUB(T4d, T4e);
2143 			      Tal = VADD(T4d, T4e);
2144 			      T4g = LD(&(xi[WS(is, 116)]), ivs, &(xi[0]));
2145 			      T4h = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
2146 			      T4i = VSUB(T4g, T4h);
2147 			      Tam = VADD(T4g, T4h);
2148 			 }
2149 			 T4j = VMUL(LDK(KP707106781), VADD(T4f, T4i));
2150 			 Tex = VADD(Tal, Tam);
2151 			 T4l = VMUL(LDK(KP707106781), VSUB(T4f, T4i));
2152 			 Tan = VSUB(Tal, Tam);
2153 		    }
2154 		    {
2155 			 V T4t, Tap, T4w, Taq;
2156 			 {
2157 			      V T4r, T4s, T4u, T4v;
2158 			      T4r = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
2159 			      T4s = LD(&(xi[WS(is, 76)]), ivs, &(xi[0]));
2160 			      T4t = VSUB(T4r, T4s);
2161 			      Tap = VADD(T4r, T4s);
2162 			      T4u = LD(&(xi[WS(is, 108)]), ivs, &(xi[0]));
2163 			      T4v = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
2164 			      T4w = VSUB(T4u, T4v);
2165 			      Taq = VADD(T4u, T4v);
2166 			 }
2167 			 T4x = VMUL(LDK(KP707106781), VSUB(T4t, T4w));
2168 			 TeA = VADD(Tap, Taq);
2169 			 T4F = VMUL(LDK(KP707106781), VADD(T4t, T4w));
2170 			 Tar = VSUB(Tap, Taq);
2171 		    }
2172 		    TgF = VADD(Tew, Tex);
2173 		    TgG = VADD(Tez, TeA);
2174 		    {
2175 			 V T4k, T4p, Tey, TeB;
2176 			 T4k = VSUB(T4c, T4j);
2177 			 T4p = VSUB(T4l, T4o);
2178 			 T4q = VFNMS(LDK(KP555570233), T4p, VMUL(LDK(KP831469612), T4k));
2179 			 T4V = VFMA(LDK(KP831469612), T4p, VMUL(LDK(KP555570233), T4k));
2180 			 Tey = VSUB(Tew, Tex);
2181 			 TeB = VSUB(Tez, TeA);
2182 			 TeC = VMUL(LDK(KP707106781), VADD(Tey, TeB));
2183 			 Tfx = VMUL(LDK(KP707106781), VSUB(Tey, TeB));
2184 		    }
2185 		    {
2186 			 V T4B, T4G, T7V, T7W;
2187 			 T4B = VSUB(T4x, T4A);
2188 			 T4G = VSUB(T4E, T4F);
2189 			 T4H = VFMA(LDK(KP555570233), T4B, VMUL(LDK(KP831469612), T4G));
2190 			 T4W = VFNMS(LDK(KP555570233), T4G, VMUL(LDK(KP831469612), T4B));
2191 			 T7V = VADD(T4A, T4x);
2192 			 T7W = VADD(T4E, T4F);
2193 			 T7X = VFMA(LDK(KP195090322), T7V, VMUL(LDK(KP980785280), T7W));
2194 			 T86 = VFNMS(LDK(KP195090322), T7W, VMUL(LDK(KP980785280), T7V));
2195 		    }
2196 		    Tcp = VFNMS(LDK(KP382683432), Tan, VMUL(LDK(KP923879532), Tak));
2197 		    Tcq = VFMA(LDK(KP923879532), Tau, VMUL(LDK(KP382683432), Tar));
2198 		    Tcr = VSUB(Tcp, Tcq);
2199 		    TdH = VADD(Tcp, Tcq);
2200 		    {
2201 			 V T7S, T7T, Tao, Tav;
2202 			 T7S = VADD(T4c, T4j);
2203 			 T7T = VADD(T4o, T4l);
2204 			 T7U = VFNMS(LDK(KP195090322), T7T, VMUL(LDK(KP980785280), T7S));
2205 			 T85 = VFMA(LDK(KP980785280), T7T, VMUL(LDK(KP195090322), T7S));
2206 			 Tao = VFMA(LDK(KP382683432), Tak, VMUL(LDK(KP923879532), Tan));
2207 			 Tav = VFNMS(LDK(KP382683432), Tau, VMUL(LDK(KP923879532), Tar));
2208 			 Taw = VSUB(Tao, Tav);
2209 			 TdC = VADD(Tao, Tav);
2210 		    }
2211 	       }
2212 	       {
2213 		    V Tbz, TbA, T3, Tf5, T18, Tbu, Tbv, Ta, Tf6, T15, Ti, Tf1, T12, Tbp, Tp;
2214 		    V Tf2, T13, Tbs, T6, T9, TbB, TbC;
2215 		    {
2216 			 V T1, T2, T16, T17;
2217 			 T1 = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
2218 			 T2 = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)]));
2219 			 Tbz = VADD(T1, T2);
2220 			 T16 = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
2221 			 T17 = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)]));
2222 			 TbA = VADD(T16, T17);
2223 			 T3 = VSUB(T1, T2);
2224 			 Tf5 = VADD(Tbz, TbA);
2225 			 T18 = VSUB(T16, T17);
2226 		    }
2227 		    {
2228 			 V T4, T5, T7, T8;
2229 			 T4 = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
2230 			 T5 = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)]));
2231 			 T6 = VSUB(T4, T5);
2232 			 Tbu = VADD(T4, T5);
2233 			 T7 = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)]));
2234 			 T8 = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
2235 			 T9 = VSUB(T7, T8);
2236 			 Tbv = VADD(T7, T8);
2237 		    }
2238 		    Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
2239 		    Tf6 = VADD(Tbu, Tbv);
2240 		    T15 = VMUL(LDK(KP707106781), VSUB(T6, T9));
2241 		    {
2242 			 V Te, Tbn, Th, Tbo;
2243 			 {
2244 			      V Tc, Td, Tf, Tg;
2245 			      Tc = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
2246 			      Td = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)]));
2247 			      Te = VSUB(Tc, Td);
2248 			      Tbn = VADD(Tc, Td);
2249 			      Tf = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
2250 			      Tg = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)]));
2251 			      Th = VSUB(Tf, Tg);
2252 			      Tbo = VADD(Tf, Tg);
2253 			 }
2254 			 Ti = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
2255 			 Tf1 = VADD(Tbn, Tbo);
2256 			 T12 = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
2257 			 Tbp = VSUB(Tbn, Tbo);
2258 		    }
2259 		    {
2260 			 V Tl, Tbr, To, Tbq;
2261 			 {
2262 			      V Tj, Tk, Tm, Tn;
2263 			      Tj = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
2264 			      Tk = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)]));
2265 			      Tl = VSUB(Tj, Tk);
2266 			      Tbr = VADD(Tj, Tk);
2267 			      Tm = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)]));
2268 			      Tn = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
2269 			      To = VSUB(Tm, Tn);
2270 			      Tbq = VADD(Tm, Tn);
2271 			 }
2272 			 Tp = VFNMS(LDK(KP382683432), To, VMUL(LDK(KP923879532), Tl));
2273 			 Tf2 = VADD(Tbq, Tbr);
2274 			 T13 = VFMA(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
2275 			 Tbs = VSUB(Tbq, Tbr);
2276 		    }
2277 		    Tf3 = VSUB(Tf1, Tf2);
2278 		    Tf7 = VSUB(Tf5, Tf6);
2279 		    {
2280 			 V Tb, Tq, Tgj, Tgk;
2281 			 Tb = VSUB(T3, Ta);
2282 			 Tq = VSUB(Ti, Tp);
2283 			 Tr = VSUB(Tb, Tq);
2284 			 T5X = VADD(Tb, Tq);
2285 			 Tgj = VADD(Tf5, Tf6);
2286 			 Tgk = VADD(Tf1, Tf2);
2287 			 Tgl = VSUB(Tgj, Tgk);
2288 			 Th3 = VADD(Tgj, Tgk);
2289 		    }
2290 		    {
2291 			 V T14, T19, T78, T79;
2292 			 T14 = VSUB(T12, T13);
2293 			 T19 = VSUB(T15, T18);
2294 			 T1a = VSUB(T14, T19);
2295 			 T5V = VADD(T19, T14);
2296 			 T78 = VADD(Ti, Tp);
2297 			 T79 = VADD(T18, T15);
2298 			 T7a = VSUB(T78, T79);
2299 			 T95 = VADD(T79, T78);
2300 		    }
2301 		    TbB = VSUB(Tbz, TbA);
2302 		    TbC = VMUL(LDK(KP707106781), VADD(Tbp, Tbs));
2303 		    TbD = VSUB(TbB, TbC);
2304 		    Tdf = VADD(TbB, TbC);
2305 		    {
2306 			 V T6X, T6Y, Tbt, Tbw;
2307 			 T6X = VADD(T3, Ta);
2308 			 T6Y = VADD(T12, T13);
2309 			 T6Z = VSUB(T6X, T6Y);
2310 			 T97 = VADD(T6X, T6Y);
2311 			 Tbt = VMUL(LDK(KP707106781), VSUB(Tbp, Tbs));
2312 			 Tbw = VSUB(Tbu, Tbv);
2313 			 Tbx = VSUB(Tbt, Tbw);
2314 			 Tdi = VADD(Tbw, Tbt);
2315 		    }
2316 	       }
2317 	       {
2318 		    V TaK, TaJ, T2U, TeE, T2Z, TaF, TaG, T2R, TeF, T30, T2C, TeH, T32, TaA, T2J;
2319 		    V TeI, T33, TaD, T2N, T2Q, TaL, TaM;
2320 		    {
2321 			 V T2S, T2T, T2X, T2Y;
2322 			 T2S = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
2323 			 T2T = LD(&(xi[WS(is, 98)]), ivs, &(xi[0]));
2324 			 TaK = VADD(T2S, T2T);
2325 			 T2X = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
2326 			 T2Y = LD(&(xi[WS(is, 66)]), ivs, &(xi[0]));
2327 			 TaJ = VADD(T2X, T2Y);
2328 			 T2U = VSUB(T2S, T2T);
2329 			 TeE = VADD(TaJ, TaK);
2330 			 T2Z = VSUB(T2X, T2Y);
2331 		    }
2332 		    {
2333 			 V T2L, T2M, T2O, T2P;
2334 			 T2L = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
2335 			 T2M = LD(&(xi[WS(is, 82)]), ivs, &(xi[0]));
2336 			 T2N = VSUB(T2L, T2M);
2337 			 TaF = VADD(T2L, T2M);
2338 			 T2O = LD(&(xi[WS(is, 114)]), ivs, &(xi[0]));
2339 			 T2P = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
2340 			 T2Q = VSUB(T2O, T2P);
2341 			 TaG = VADD(T2O, T2P);
2342 		    }
2343 		    T2R = VMUL(LDK(KP707106781), VSUB(T2N, T2Q));
2344 		    TeF = VADD(TaF, TaG);
2345 		    T30 = VMUL(LDK(KP707106781), VADD(T2N, T2Q));
2346 		    {
2347 			 V T2y, Tay, T2B, Taz;
2348 			 {
2349 			      V T2w, T2x, T2z, T2A;
2350 			      T2w = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
2351 			      T2x = LD(&(xi[WS(is, 74)]), ivs, &(xi[0]));
2352 			      T2y = VSUB(T2w, T2x);
2353 			      Tay = VADD(T2w, T2x);
2354 			      T2z = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
2355 			      T2A = LD(&(xi[WS(is, 106)]), ivs, &(xi[0]));
2356 			      T2B = VSUB(T2z, T2A);
2357 			      Taz = VADD(T2z, T2A);
2358 			 }
2359 			 T2C = VFNMS(LDK(KP382683432), T2B, VMUL(LDK(KP923879532), T2y));
2360 			 TeH = VADD(Tay, Taz);
2361 			 T32 = VFMA(LDK(KP382683432), T2y, VMUL(LDK(KP923879532), T2B));
2362 			 TaA = VSUB(Tay, Taz);
2363 		    }
2364 		    {
2365 			 V T2F, TaB, T2I, TaC;
2366 			 {
2367 			      V T2D, T2E, T2G, T2H;
2368 			      T2D = LD(&(xi[WS(is, 122)]), ivs, &(xi[0]));
2369 			      T2E = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
2370 			      T2F = VSUB(T2D, T2E);
2371 			      TaB = VADD(T2D, T2E);
2372 			      T2G = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
2373 			      T2H = LD(&(xi[WS(is, 90)]), ivs, &(xi[0]));
2374 			      T2I = VSUB(T2G, T2H);
2375 			      TaC = VADD(T2G, T2H);
2376 			 }
2377 			 T2J = VFMA(LDK(KP923879532), T2F, VMUL(LDK(KP382683432), T2I));
2378 			 TeI = VADD(TaB, TaC);
2379 			 T33 = VFNMS(LDK(KP382683432), T2F, VMUL(LDK(KP923879532), T2I));
2380 			 TaD = VSUB(TaB, TaC);
2381 		    }
2382 		    Tgy = VADD(TeE, TeF);
2383 		    Tgz = VADD(TeH, TeI);
2384 		    TgA = VSUB(Tgy, Tgz);
2385 		    TaL = VSUB(TaJ, TaK);
2386 		    TaM = VMUL(LDK(KP707106781), VADD(TaA, TaD));
2387 		    TaN = VSUB(TaL, TaM);
2388 		    Tdv = VADD(TaL, TaM);
2389 		    {
2390 			 V TeG, TeJ, T2K, T2V;
2391 			 TeG = VSUB(TeE, TeF);
2392 			 TeJ = VSUB(TeH, TeI);
2393 			 TeK = VFMA(LDK(KP382683432), TeG, VMUL(LDK(KP923879532), TeJ));
2394 			 Tfu = VFNMS(LDK(KP382683432), TeJ, VMUL(LDK(KP923879532), TeG));
2395 			 T2K = VSUB(T2C, T2J);
2396 			 T2V = VSUB(T2R, T2U);
2397 			 T2W = VSUB(T2K, T2V);
2398 			 T5M = VADD(T2V, T2K);
2399 		    }
2400 		    {
2401 			 V T31, T34, T7D, T7E;
2402 			 T31 = VSUB(T2Z, T30);
2403 			 T34 = VSUB(T32, T33);
2404 			 T35 = VSUB(T31, T34);
2405 			 T5N = VADD(T31, T34);
2406 			 T7D = VADD(T32, T33);
2407 			 T7E = VADD(T2U, T2R);
2408 			 T7F = VSUB(T7D, T7E);
2409 			 T8X = VADD(T7E, T7D);
2410 		    }
2411 		    {
2412 			 V TaE, TaH, T7A, T7B;
2413 			 TaE = VMUL(LDK(KP707106781), VSUB(TaA, TaD));
2414 			 TaH = VSUB(TaF, TaG);
2415 			 TaI = VSUB(TaE, TaH);
2416 			 Tdu = VADD(TaH, TaE);
2417 			 T7A = VADD(T2Z, T30);
2418 			 T7B = VADD(T2C, T2J);
2419 			 T7C = VSUB(T7A, T7B);
2420 			 T8W = VADD(T7A, T7B);
2421 		    }
2422 	       }
2423 	       {
2424 		    V Tb1, Tb0, T3v, TeO, T3A, TaW, TaX, T3s, TeP, T3B, T3d, TeL, T3D, TaR, T3k;
2425 		    V TeM, T3E, TaU, T3o, T3r, Tb2, Tb3;
2426 		    {
2427 			 V T3t, T3u, T3y, T3z;
2428 			 T3t = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
2429 			 T3u = LD(&(xi[WS(is, 94)]), ivs, &(xi[0]));
2430 			 Tb1 = VADD(T3t, T3u);
2431 			 T3y = LD(&(xi[WS(is, 126)]), ivs, &(xi[0]));
2432 			 T3z = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
2433 			 Tb0 = VADD(T3y, T3z);
2434 			 T3v = VSUB(T3t, T3u);
2435 			 TeO = VADD(Tb0, Tb1);
2436 			 T3A = VSUB(T3y, T3z);
2437 		    }
2438 		    {
2439 			 V T3m, T3n, T3p, T3q;
2440 			 T3m = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
2441 			 T3n = LD(&(xi[WS(is, 78)]), ivs, &(xi[0]));
2442 			 T3o = VSUB(T3m, T3n);
2443 			 TaW = VADD(T3m, T3n);
2444 			 T3p = LD(&(xi[WS(is, 110)]), ivs, &(xi[0]));
2445 			 T3q = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
2446 			 T3r = VSUB(T3p, T3q);
2447 			 TaX = VADD(T3p, T3q);
2448 		    }
2449 		    T3s = VMUL(LDK(KP707106781), VSUB(T3o, T3r));
2450 		    TeP = VADD(TaW, TaX);
2451 		    T3B = VMUL(LDK(KP707106781), VADD(T3o, T3r));
2452 		    {
2453 			 V T39, TaP, T3c, TaQ;
2454 			 {
2455 			      V T37, T38, T3a, T3b;
2456 			      T37 = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
2457 			      T38 = LD(&(xi[WS(is, 70)]), ivs, &(xi[0]));
2458 			      T39 = VSUB(T37, T38);
2459 			      TaP = VADD(T37, T38);
2460 			      T3a = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
2461 			      T3b = LD(&(xi[WS(is, 102)]), ivs, &(xi[0]));
2462 			      T3c = VSUB(T3a, T3b);
2463 			      TaQ = VADD(T3a, T3b);
2464 			 }
2465 			 T3d = VFNMS(LDK(KP382683432), T3c, VMUL(LDK(KP923879532), T39));
2466 			 TeL = VADD(TaP, TaQ);
2467 			 T3D = VFMA(LDK(KP382683432), T39, VMUL(LDK(KP923879532), T3c));
2468 			 TaR = VSUB(TaP, TaQ);
2469 		    }
2470 		    {
2471 			 V T3g, TaS, T3j, TaT;
2472 			 {
2473 			      V T3e, T3f, T3h, T3i;
2474 			      T3e = LD(&(xi[WS(is, 118)]), ivs, &(xi[0]));
2475 			      T3f = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
2476 			      T3g = VSUB(T3e, T3f);
2477 			      TaS = VADD(T3e, T3f);
2478 			      T3h = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
2479 			      T3i = LD(&(xi[WS(is, 86)]), ivs, &(xi[0]));
2480 			      T3j = VSUB(T3h, T3i);
2481 			      TaT = VADD(T3h, T3i);
2482 			 }
2483 			 T3k = VFMA(LDK(KP923879532), T3g, VMUL(LDK(KP382683432), T3j));
2484 			 TeM = VADD(TaS, TaT);
2485 			 T3E = VFNMS(LDK(KP382683432), T3g, VMUL(LDK(KP923879532), T3j));
2486 			 TaU = VSUB(TaS, TaT);
2487 		    }
2488 		    TgB = VADD(TeO, TeP);
2489 		    TgC = VADD(TeL, TeM);
2490 		    TgD = VSUB(TgB, TgC);
2491 		    Tb2 = VSUB(Tb0, Tb1);
2492 		    Tb3 = VMUL(LDK(KP707106781), VADD(TaR, TaU));
2493 		    Tb4 = VSUB(Tb2, Tb3);
2494 		    Tdy = VADD(Tb2, Tb3);
2495 		    {
2496 			 V TeN, TeQ, T3l, T3w;
2497 			 TeN = VSUB(TeL, TeM);
2498 			 TeQ = VSUB(TeO, TeP);
2499 			 TeR = VFNMS(LDK(KP382683432), TeQ, VMUL(LDK(KP923879532), TeN));
2500 			 Tfv = VFMA(LDK(KP923879532), TeQ, VMUL(LDK(KP382683432), TeN));
2501 			 T3l = VSUB(T3d, T3k);
2502 			 T3w = VSUB(T3s, T3v);
2503 			 T3x = VSUB(T3l, T3w);
2504 			 T5P = VADD(T3w, T3l);
2505 		    }
2506 		    {
2507 			 V T3C, T3F, T7K, T7L;
2508 			 T3C = VSUB(T3A, T3B);
2509 			 T3F = VSUB(T3D, T3E);
2510 			 T3G = VSUB(T3C, T3F);
2511 			 T5Q = VADD(T3C, T3F);
2512 			 T7K = VADD(T3A, T3B);
2513 			 T7L = VADD(T3d, T3k);
2514 			 T7M = VSUB(T7K, T7L);
2515 			 T90 = VADD(T7K, T7L);
2516 		    }
2517 		    {
2518 			 V TaV, TaY, T7H, T7I;
2519 			 TaV = VMUL(LDK(KP707106781), VSUB(TaR, TaU));
2520 			 TaY = VSUB(TaW, TaX);
2521 			 TaZ = VSUB(TaV, TaY);
2522 			 Tdx = VADD(TaY, TaV);
2523 			 T7H = VADD(T3D, T3E);
2524 			 T7I = VADD(T3v, T3s);
2525 			 T7J = VSUB(T7H, T7I);
2526 			 T8Z = VADD(T7I, T7H);
2527 		    }
2528 	       }
2529 	       {
2530 		    V TB, TeU, TF, Tba, TS, TeX, TW, Tbh, Ty, TeV, TG, Tbd, TP, TeY, TX;
2531 		    V Tbk;
2532 		    {
2533 			 V Tz, TA, Tb9, TD, TE, Tb8;
2534 			 Tz = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
2535 			 TA = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)]));
2536 			 Tb9 = VADD(Tz, TA);
2537 			 TD = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
2538 			 TE = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)]));
2539 			 Tb8 = VADD(TD, TE);
2540 			 TB = VSUB(Tz, TA);
2541 			 TeU = VADD(Tb8, Tb9);
2542 			 TF = VSUB(TD, TE);
2543 			 Tba = VSUB(Tb8, Tb9);
2544 		    }
2545 		    {
2546 			 V TQ, TR, Tbg, TU, TV, Tbf;
2547 			 TQ = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
2548 			 TR = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)]));
2549 			 Tbg = VADD(TQ, TR);
2550 			 TU = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)]));
2551 			 TV = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
2552 			 Tbf = VADD(TU, TV);
2553 			 TS = VSUB(TQ, TR);
2554 			 TeX = VADD(Tbf, Tbg);
2555 			 TW = VSUB(TU, TV);
2556 			 Tbh = VSUB(Tbf, Tbg);
2557 		    }
2558 		    {
2559 			 V Tu, Tbb, Tx, Tbc;
2560 			 {
2561 			      V Ts, Tt, Tv, Tw;
2562 			      Ts = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
2563 			      Tt = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)]));
2564 			      Tu = VSUB(Ts, Tt);
2565 			      Tbb = VADD(Ts, Tt);
2566 			      Tv = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)]));
2567 			      Tw = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
2568 			      Tx = VSUB(Tv, Tw);
2569 			      Tbc = VADD(Tv, Tw);
2570 			 }
2571 			 Ty = VMUL(LDK(KP707106781), VSUB(Tu, Tx));
2572 			 TeV = VADD(Tbb, Tbc);
2573 			 TG = VMUL(LDK(KP707106781), VADD(Tu, Tx));
2574 			 Tbd = VSUB(Tbb, Tbc);
2575 		    }
2576 		    {
2577 			 V TL, Tbi, TO, Tbj;
2578 			 {
2579 			      V TJ, TK, TM, TN;
2580 			      TJ = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
2581 			      TK = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)]));
2582 			      TL = VSUB(TJ, TK);
2583 			      Tbi = VADD(TJ, TK);
2584 			      TM = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)]));
2585 			      TN = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
2586 			      TO = VSUB(TM, TN);
2587 			      Tbj = VADD(TM, TN);
2588 			 }
2589 			 TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
2590 			 TeY = VADD(Tbi, Tbj);
2591 			 TX = VMUL(LDK(KP707106781), VADD(TL, TO));
2592 			 Tbk = VSUB(Tbi, Tbj);
2593 		    }
2594 		    {
2595 			 V Tbe, Tbl, TeW, TeZ;
2596 			 Tbe = VFNMS(LDK(KP382683432), Tbd, VMUL(LDK(KP923879532), Tba));
2597 			 Tbl = VFMA(LDK(KP923879532), Tbh, VMUL(LDK(KP382683432), Tbk));
2598 			 Tbm = VSUB(Tbe, Tbl);
2599 			 Tdg = VADD(Tbe, Tbl);
2600 			 {
2601 			      V TbE, TbF, Tgm, Tgn;
2602 			      TbE = VFMA(LDK(KP382683432), Tba, VMUL(LDK(KP923879532), Tbd));
2603 			      TbF = VFNMS(LDK(KP382683432), Tbh, VMUL(LDK(KP923879532), Tbk));
2604 			      TbG = VSUB(TbE, TbF);
2605 			      Tdj = VADD(TbE, TbF);
2606 			      Tgm = VADD(TeU, TeV);
2607 			      Tgn = VADD(TeX, TeY);
2608 			      Tgo = VSUB(Tgm, Tgn);
2609 			      Th4 = VADD(Tgm, Tgn);
2610 			 }
2611 			 TeW = VSUB(TeU, TeV);
2612 			 TeZ = VSUB(TeX, TeY);
2613 			 Tf0 = VMUL(LDK(KP707106781), VSUB(TeW, TeZ));
2614 			 Tf8 = VMUL(LDK(KP707106781), VADD(TeW, TeZ));
2615 			 {
2616 			      V T72, T7b, T75, T7c;
2617 			      {
2618 				   V T70, T71, T73, T74;
2619 				   T70 = VADD(TB, Ty);
2620 				   T71 = VADD(TF, TG);
2621 				   T72 = VFMA(LDK(KP980785280), T70, VMUL(LDK(KP195090322), T71));
2622 				   T7b = VFNMS(LDK(KP195090322), T70, VMUL(LDK(KP980785280), T71));
2623 				   T73 = VADD(TS, TP);
2624 				   T74 = VADD(TW, TX);
2625 				   T75 = VFNMS(LDK(KP195090322), T74, VMUL(LDK(KP980785280), T73));
2626 				   T7c = VFMA(LDK(KP195090322), T73, VMUL(LDK(KP980785280), T74));
2627 			      }
2628 			      T76 = VSUB(T72, T75);
2629 			      T98 = VADD(T7b, T7c);
2630 			      T7d = VSUB(T7b, T7c);
2631 			      T94 = VADD(T72, T75);
2632 			 }
2633 			 {
2634 			      V TI, T1b, TZ, T1c;
2635 			      {
2636 				   V TC, TH, TT, TY;
2637 				   TC = VSUB(Ty, TB);
2638 				   TH = VSUB(TF, TG);
2639 				   TI = VFMA(LDK(KP831469612), TC, VMUL(LDK(KP555570233), TH));
2640 				   T1b = VFNMS(LDK(KP555570233), TC, VMUL(LDK(KP831469612), TH));
2641 				   TT = VSUB(TP, TS);
2642 				   TY = VSUB(TW, TX);
2643 				   TZ = VFNMS(LDK(KP555570233), TY, VMUL(LDK(KP831469612), TT));
2644 				   T1c = VFMA(LDK(KP555570233), TT, VMUL(LDK(KP831469612), TY));
2645 			      }
2646 			      T10 = VSUB(TI, TZ);
2647 			      T5Y = VADD(T1b, T1c);
2648 			      T1d = VSUB(T1b, T1c);
2649 			      T5U = VADD(TI, TZ);
2650 			 }
2651 		    }
2652 	       }
2653 	       {
2654 		    V T1Q, Tfb, T1U, TbL, T27, Tfe, T2b, TbS, T1N, Tfc, T1V, TbO, T24, Tff, T2c;
2655 		    V TbV;
2656 		    {
2657 			 V T1O, T1P, TbK, T1S, T1T, TbJ;
2658 			 T1O = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
2659 			 T1P = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)]));
2660 			 TbK = VADD(T1O, T1P);
2661 			 T1S = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
2662 			 T1T = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)]));
2663 			 TbJ = VADD(T1S, T1T);
2664 			 T1Q = VSUB(T1O, T1P);
2665 			 Tfb = VADD(TbJ, TbK);
2666 			 T1U = VSUB(T1S, T1T);
2667 			 TbL = VSUB(TbJ, TbK);
2668 		    }
2669 		    {
2670 			 V T25, T26, TbR, T29, T2a, TbQ;
2671 			 T25 = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
2672 			 T26 = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)]));
2673 			 TbR = VADD(T25, T26);
2674 			 T29 = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)]));
2675 			 T2a = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
2676 			 TbQ = VADD(T29, T2a);
2677 			 T27 = VSUB(T25, T26);
2678 			 Tfe = VADD(TbQ, TbR);
2679 			 T2b = VSUB(T29, T2a);
2680 			 TbS = VSUB(TbQ, TbR);
2681 		    }
2682 		    {
2683 			 V T1J, TbM, T1M, TbN;
2684 			 {
2685 			      V T1H, T1I, T1K, T1L;
2686 			      T1H = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
2687 			      T1I = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)]));
2688 			      T1J = VSUB(T1H, T1I);
2689 			      TbM = VADD(T1H, T1I);
2690 			      T1K = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)]));
2691 			      T1L = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
2692 			      T1M = VSUB(T1K, T1L);
2693 			      TbN = VADD(T1K, T1L);
2694 			 }
2695 			 T1N = VMUL(LDK(KP707106781), VSUB(T1J, T1M));
2696 			 Tfc = VADD(TbM, TbN);
2697 			 T1V = VMUL(LDK(KP707106781), VADD(T1J, T1M));
2698 			 TbO = VSUB(TbM, TbN);
2699 		    }
2700 		    {
2701 			 V T20, TbT, T23, TbU;
2702 			 {
2703 			      V T1Y, T1Z, T21, T22;
2704 			      T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
2705 			      T1Z = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)]));
2706 			      T20 = VSUB(T1Y, T1Z);
2707 			      TbT = VADD(T1Y, T1Z);
2708 			      T21 = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)]));
2709 			      T22 = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
2710 			      T23 = VSUB(T21, T22);
2711 			      TbU = VADD(T21, T22);
2712 			 }
2713 			 T24 = VMUL(LDK(KP707106781), VSUB(T20, T23));
2714 			 Tff = VADD(TbT, TbU);
2715 			 T2c = VMUL(LDK(KP707106781), VADD(T20, T23));
2716 			 TbV = VSUB(TbT, TbU);
2717 		    }
2718 		    {
2719 			 V TbP, TbW, Tfd, Tfg;
2720 			 TbP = VFNMS(LDK(KP382683432), TbO, VMUL(LDK(KP923879532), TbL));
2721 			 TbW = VFMA(LDK(KP923879532), TbS, VMUL(LDK(KP382683432), TbV));
2722 			 TbX = VSUB(TbP, TbW);
2723 			 Tdn = VADD(TbP, TbW);
2724 			 {
2725 			      V Tcf, Tcg, Tgt, Tgu;
2726 			      Tcf = VFMA(LDK(KP382683432), TbL, VMUL(LDK(KP923879532), TbO));
2727 			      Tcg = VFNMS(LDK(KP382683432), TbS, VMUL(LDK(KP923879532), TbV));
2728 			      Tch = VSUB(Tcf, Tcg);
2729 			      Tdq = VADD(Tcf, Tcg);
2730 			      Tgt = VADD(Tfb, Tfc);
2731 			      Tgu = VADD(Tfe, Tff);
2732 			      Tgv = VSUB(Tgt, Tgu);
2733 			      Th7 = VADD(Tgt, Tgu);
2734 			 }
2735 			 Tfd = VSUB(Tfb, Tfc);
2736 			 Tfg = VSUB(Tfe, Tff);
2737 			 Tfh = VMUL(LDK(KP707106781), VSUB(Tfd, Tfg));
2738 			 Tfp = VMUL(LDK(KP707106781), VADD(Tfd, Tfg));
2739 			 {
2740 			      V T7l, T7u, T7o, T7v;
2741 			      {
2742 				   V T7j, T7k, T7m, T7n;
2743 				   T7j = VADD(T1Q, T1N);
2744 				   T7k = VADD(T1U, T1V);
2745 				   T7l = VFMA(LDK(KP980785280), T7j, VMUL(LDK(KP195090322), T7k));
2746 				   T7u = VFNMS(LDK(KP195090322), T7j, VMUL(LDK(KP980785280), T7k));
2747 				   T7m = VADD(T27, T24);
2748 				   T7n = VADD(T2b, T2c);
2749 				   T7o = VFNMS(LDK(KP195090322), T7n, VMUL(LDK(KP980785280), T7m));
2750 				   T7v = VFMA(LDK(KP195090322), T7m, VMUL(LDK(KP980785280), T7n));
2751 			      }
2752 			      T7p = VSUB(T7l, T7o);
2753 			      T9f = VADD(T7u, T7v);
2754 			      T7w = VSUB(T7u, T7v);
2755 			      T9b = VADD(T7l, T7o);
2756 			 }
2757 			 {
2758 			      V T1X, T2q, T2e, T2r;
2759 			      {
2760 				   V T1R, T1W, T28, T2d;
2761 				   T1R = VSUB(T1N, T1Q);
2762 				   T1W = VSUB(T1U, T1V);
2763 				   T1X = VFMA(LDK(KP831469612), T1R, VMUL(LDK(KP555570233), T1W));
2764 				   T2q = VFNMS(LDK(KP555570233), T1R, VMUL(LDK(KP831469612), T1W));
2765 				   T28 = VSUB(T24, T27);
2766 				   T2d = VSUB(T2b, T2c);
2767 				   T2e = VFNMS(LDK(KP555570233), T2d, VMUL(LDK(KP831469612), T28));
2768 				   T2r = VFMA(LDK(KP555570233), T28, VMUL(LDK(KP831469612), T2d));
2769 			      }
2770 			      T2f = VSUB(T1X, T2e);
2771 			      T65 = VADD(T2q, T2r);
2772 			      T2s = VSUB(T2q, T2r);
2773 			      T61 = VADD(T1X, T2e);
2774 			 }
2775 		    }
2776 	       }
2777 	       {
2778 		    V Tgx, TgW, TgR, TgZ, TgI, TgY, TgO, TgV;
2779 		    {
2780 			 V Tgp, Tgw, TgP, TgQ;
2781 			 Tgp = VFNMS(LDK(KP382683432), Tgo, VMUL(LDK(KP923879532), Tgl));
2782 			 Tgw = VFMA(LDK(KP923879532), Tgs, VMUL(LDK(KP382683432), Tgv));
2783 			 Tgx = VSUB(Tgp, Tgw);
2784 			 TgW = VADD(Tgp, Tgw);
2785 			 TgP = VFMA(LDK(KP382683432), Tgl, VMUL(LDK(KP923879532), Tgo));
2786 			 TgQ = VFNMS(LDK(KP382683432), Tgs, VMUL(LDK(KP923879532), Tgv));
2787 			 TgR = VSUB(TgP, TgQ);
2788 			 TgZ = VADD(TgP, TgQ);
2789 		    }
2790 		    {
2791 			 V TgE, TgH, TgM, TgN;
2792 			 TgE = VMUL(LDK(KP707106781), VSUB(TgA, TgD));
2793 			 TgH = VSUB(TgF, TgG);
2794 			 TgI = VSUB(TgE, TgH);
2795 			 TgY = VADD(TgH, TgE);
2796 			 TgM = VSUB(TgK, TgL);
2797 			 TgN = VMUL(LDK(KP707106781), VADD(TgA, TgD));
2798 			 TgO = VSUB(TgM, TgN);
2799 			 TgV = VADD(TgM, TgN);
2800 		    }
2801 		    {
2802 			 V TgJ, TgS, Th1, Th2;
2803 			 TgJ = VBYI(VSUB(Tgx, TgI));
2804 			 TgS = VSUB(TgO, TgR);
2805 			 ST(&(xo[WS(os, 40)]), VADD(TgJ, TgS), ovs, &(xo[0]));
2806 			 ST(&(xo[WS(os, 88)]), VSUB(TgS, TgJ), ovs, &(xo[0]));
2807 			 Th1 = VSUB(TgV, TgW);
2808 			 Th2 = VBYI(VSUB(TgZ, TgY));
2809 			 ST(&(xo[WS(os, 72)]), VSUB(Th1, Th2), ovs, &(xo[0]));
2810 			 ST(&(xo[WS(os, 56)]), VADD(Th1, Th2), ovs, &(xo[0]));
2811 		    }
2812 		    {
2813 			 V TgT, TgU, TgX, Th0;
2814 			 TgT = VBYI(VADD(TgI, Tgx));
2815 			 TgU = VADD(TgO, TgR);
2816 			 ST(&(xo[WS(os, 24)]), VADD(TgT, TgU), ovs, &(xo[0]));
2817 			 ST(&(xo[WS(os, 104)]), VSUB(TgU, TgT), ovs, &(xo[0]));
2818 			 TgX = VADD(TgV, TgW);
2819 			 Th0 = VBYI(VADD(TgY, TgZ));
2820 			 ST(&(xo[WS(os, 120)]), VSUB(TgX, Th0), ovs, &(xo[0]));
2821 			 ST(&(xo[WS(os, 8)]), VADD(TgX, Th0), ovs, &(xo[0]));
2822 		    }
2823 	       }
2824 	       {
2825 		    V Th9, Thh, Thq, Ths, Thc, Thm, Thg, Thl, Thn, Thr;
2826 		    {
2827 			 V Th5, Th8, Tho, Thp;
2828 			 Th5 = VSUB(Th3, Th4);
2829 			 Th8 = VSUB(Th6, Th7);
2830 			 Th9 = VMUL(LDK(KP707106781), VSUB(Th5, Th8));
2831 			 Thh = VMUL(LDK(KP707106781), VADD(Th5, Th8));
2832 			 Tho = VADD(Th3, Th4);
2833 			 Thp = VADD(Th6, Th7);
2834 			 Thq = VBYI(VSUB(Tho, Thp));
2835 			 Ths = VADD(Tho, Thp);
2836 		    }
2837 		    {
2838 			 V Tha, Thb, The, Thf;
2839 			 Tha = VADD(Tgy, Tgz);
2840 			 Thb = VADD(TgB, TgC);
2841 			 Thc = VSUB(Tha, Thb);
2842 			 Thm = VADD(Tha, Thb);
2843 			 The = VADD(TgK, TgL);
2844 			 Thf = VADD(TgF, TgG);
2845 			 Thg = VSUB(The, Thf);
2846 			 Thl = VADD(The, Thf);
2847 		    }
2848 		    Thn = VSUB(Thl, Thm);
2849 		    ST(&(xo[WS(os, 96)]), VSUB(Thn, Thq), ovs, &(xo[0]));
2850 		    ST(&(xo[WS(os, 32)]), VADD(Thn, Thq), ovs, &(xo[0]));
2851 		    Thr = VADD(Thl, Thm);
2852 		    ST(&(xo[WS(os, 64)]), VSUB(Thr, Ths), ovs, &(xo[0]));
2853 		    ST(&(xo[0]), VADD(Thr, Ths), ovs, &(xo[0]));
2854 		    {
2855 			 V Thd, Thi, Thj, Thk;
2856 			 Thd = VBYI(VSUB(Th9, Thc));
2857 			 Thi = VSUB(Thg, Thh);
2858 			 ST(&(xo[WS(os, 48)]), VADD(Thd, Thi), ovs, &(xo[0]));
2859 			 ST(&(xo[WS(os, 80)]), VSUB(Thi, Thd), ovs, &(xo[0]));
2860 			 Thj = VBYI(VADD(Thc, Th9));
2861 			 Thk = VADD(Thg, Thh);
2862 			 ST(&(xo[WS(os, 16)]), VADD(Thj, Thk), ovs, &(xo[0]));
2863 			 ST(&(xo[WS(os, 112)]), VSUB(Thk, Thj), ovs, &(xo[0]));
2864 		    }
2865 	       }
2866 	       {
2867 		    V TeT, TfM, TfC, TfK, Tfs, TfN, TfF, TfJ;
2868 		    {
2869 			 V TeD, TeS, Tfw, TfB;
2870 			 TeD = VSUB(Tev, TeC);
2871 			 TeS = VSUB(TeK, TeR);
2872 			 TeT = VSUB(TeD, TeS);
2873 			 TfM = VADD(TeD, TeS);
2874 			 Tfw = VSUB(Tfu, Tfv);
2875 			 TfB = VSUB(Tfx, TfA);
2876 			 TfC = VSUB(Tfw, TfB);
2877 			 TfK = VADD(TfB, Tfw);
2878 			 {
2879 			      V Tfa, TfD, Tfr, TfE;
2880 			      {
2881 				   V Tf4, Tf9, Tfl, Tfq;
2882 				   Tf4 = VSUB(Tf0, Tf3);
2883 				   Tf9 = VSUB(Tf7, Tf8);
2884 				   Tfa = VFMA(LDK(KP831469612), Tf4, VMUL(LDK(KP555570233), Tf9));
2885 				   TfD = VFNMS(LDK(KP555570233), Tf4, VMUL(LDK(KP831469612), Tf9));
2886 				   Tfl = VSUB(Tfh, Tfk);
2887 				   Tfq = VSUB(Tfo, Tfp);
2888 				   Tfr = VFNMS(LDK(KP555570233), Tfq, VMUL(LDK(KP831469612), Tfl));
2889 				   TfE = VFMA(LDK(KP555570233), Tfl, VMUL(LDK(KP831469612), Tfq));
2890 			      }
2891 			      Tfs = VSUB(Tfa, Tfr);
2892 			      TfN = VADD(TfD, TfE);
2893 			      TfF = VSUB(TfD, TfE);
2894 			      TfJ = VADD(Tfa, Tfr);
2895 			 }
2896 		    }
2897 		    {
2898 			 V Tft, TfG, TfP, TfQ;
2899 			 Tft = VADD(TeT, Tfs);
2900 			 TfG = VBYI(VADD(TfC, TfF));
2901 			 ST(&(xo[WS(os, 108)]), VSUB(Tft, TfG), ovs, &(xo[0]));
2902 			 ST(&(xo[WS(os, 20)]), VADD(Tft, TfG), ovs, &(xo[0]));
2903 			 TfP = VBYI(VADD(TfK, TfJ));
2904 			 TfQ = VADD(TfM, TfN);
2905 			 ST(&(xo[WS(os, 12)]), VADD(TfP, TfQ), ovs, &(xo[0]));
2906 			 ST(&(xo[WS(os, 116)]), VSUB(TfQ, TfP), ovs, &(xo[0]));
2907 		    }
2908 		    {
2909 			 V TfH, TfI, TfL, TfO;
2910 			 TfH = VSUB(TeT, Tfs);
2911 			 TfI = VBYI(VSUB(TfF, TfC));
2912 			 ST(&(xo[WS(os, 84)]), VSUB(TfH, TfI), ovs, &(xo[0]));
2913 			 ST(&(xo[WS(os, 44)]), VADD(TfH, TfI), ovs, &(xo[0]));
2914 			 TfL = VBYI(VSUB(TfJ, TfK));
2915 			 TfO = VSUB(TfM, TfN);
2916 			 ST(&(xo[WS(os, 52)]), VADD(TfL, TfO), ovs, &(xo[0]));
2917 			 ST(&(xo[WS(os, 76)]), VSUB(TfO, TfL), ovs, &(xo[0]));
2918 		    }
2919 	       }
2920 	       {
2921 		    V TfT, Tge, Tg4, Tgc, Tg0, Tgf, Tg7, Tgb;
2922 		    {
2923 			 V TfR, TfS, Tg2, Tg3;
2924 			 TfR = VADD(Tev, TeC);
2925 			 TfS = VADD(Tfu, Tfv);
2926 			 TfT = VSUB(TfR, TfS);
2927 			 Tge = VADD(TfR, TfS);
2928 			 Tg2 = VADD(TeK, TeR);
2929 			 Tg3 = VADD(TfA, Tfx);
2930 			 Tg4 = VSUB(Tg2, Tg3);
2931 			 Tgc = VADD(Tg3, Tg2);
2932 			 {
2933 			      V TfW, Tg5, TfZ, Tg6;
2934 			      {
2935 				   V TfU, TfV, TfX, TfY;
2936 				   TfU = VADD(Tf3, Tf0);
2937 				   TfV = VADD(Tf7, Tf8);
2938 				   TfW = VFMA(LDK(KP980785280), TfU, VMUL(LDK(KP195090322), TfV));
2939 				   Tg5 = VFNMS(LDK(KP195090322), TfU, VMUL(LDK(KP980785280), TfV));
2940 				   TfX = VADD(Tfk, Tfh);
2941 				   TfY = VADD(Tfo, Tfp);
2942 				   TfZ = VFNMS(LDK(KP195090322), TfY, VMUL(LDK(KP980785280), TfX));
2943 				   Tg6 = VFMA(LDK(KP195090322), TfX, VMUL(LDK(KP980785280), TfY));
2944 			      }
2945 			      Tg0 = VSUB(TfW, TfZ);
2946 			      Tgf = VADD(Tg5, Tg6);
2947 			      Tg7 = VSUB(Tg5, Tg6);
2948 			      Tgb = VADD(TfW, TfZ);
2949 			 }
2950 		    }
2951 		    {
2952 			 V Tg1, Tg8, Tgh, Tgi;
2953 			 Tg1 = VADD(TfT, Tg0);
2954 			 Tg8 = VBYI(VADD(Tg4, Tg7));
2955 			 ST(&(xo[WS(os, 100)]), VSUB(Tg1, Tg8), ovs, &(xo[0]));
2956 			 ST(&(xo[WS(os, 28)]), VADD(Tg1, Tg8), ovs, &(xo[0]));
2957 			 Tgh = VBYI(VADD(Tgc, Tgb));
2958 			 Tgi = VADD(Tge, Tgf);
2959 			 ST(&(xo[WS(os, 4)]), VADD(Tgh, Tgi), ovs, &(xo[0]));
2960 			 ST(&(xo[WS(os, 124)]), VSUB(Tgi, Tgh), ovs, &(xo[0]));
2961 		    }
2962 		    {
2963 			 V Tg9, Tga, Tgd, Tgg;
2964 			 Tg9 = VSUB(TfT, Tg0);
2965 			 Tga = VBYI(VSUB(Tg7, Tg4));
2966 			 ST(&(xo[WS(os, 92)]), VSUB(Tg9, Tga), ovs, &(xo[0]));
2967 			 ST(&(xo[WS(os, 36)]), VADD(Tg9, Tga), ovs, &(xo[0]));
2968 			 Tgd = VBYI(VSUB(Tgb, Tgc));
2969 			 Tgg = VSUB(Tge, Tgf);
2970 			 ST(&(xo[WS(os, 60)]), VADD(Tgd, Tgg), ovs, &(xo[0]));
2971 			 ST(&(xo[WS(os, 68)]), VSUB(Tgg, Tgd), ovs, &(xo[0]));
2972 		    }
2973 	       }
2974 	       {
2975 		    V Tb7, Td8, TcI, Td0, Tcy, Tda, TcG, TcP, Tck, TcJ, TcB, TcF, TcW, Tdb, Td3;
2976 		    V Td7;
2977 		    {
2978 			 V Tax, TcZ, Tb6, TcY, TaO, Tb5;
2979 			 Tax = VSUB(Tah, Taw);
2980 			 TcZ = VADD(Tcw, Tcr);
2981 			 TaO = VFMA(LDK(KP831469612), TaI, VMUL(LDK(KP555570233), TaN));
2982 			 Tb5 = VFNMS(LDK(KP555570233), Tb4, VMUL(LDK(KP831469612), TaZ));
2983 			 Tb6 = VSUB(TaO, Tb5);
2984 			 TcY = VADD(TaO, Tb5);
2985 			 Tb7 = VSUB(Tax, Tb6);
2986 			 Td8 = VADD(TcZ, TcY);
2987 			 TcI = VADD(Tax, Tb6);
2988 			 Td0 = VSUB(TcY, TcZ);
2989 		    }
2990 		    {
2991 			 V Tcx, TcN, Tco, TcO, Tcm, Tcn;
2992 			 Tcx = VSUB(Tcr, Tcw);
2993 			 TcN = VADD(Tah, Taw);
2994 			 Tcm = VFNMS(LDK(KP555570233), TaI, VMUL(LDK(KP831469612), TaN));
2995 			 Tcn = VFMA(LDK(KP555570233), TaZ, VMUL(LDK(KP831469612), Tb4));
2996 			 Tco = VSUB(Tcm, Tcn);
2997 			 TcO = VADD(Tcm, Tcn);
2998 			 Tcy = VSUB(Tco, Tcx);
2999 			 Tda = VADD(TcN, TcO);
3000 			 TcG = VADD(Tcx, Tco);
3001 			 TcP = VSUB(TcN, TcO);
3002 		    }
3003 		    {
3004 			 V TbI, Tcz, Tcj, TcA;
3005 			 {
3006 			      V Tby, TbH, Tc9, Tci;
3007 			      Tby = VSUB(Tbm, Tbx);
3008 			      TbH = VSUB(TbD, TbG);
3009 			      TbI = VFMA(LDK(KP881921264), Tby, VMUL(LDK(KP471396736), TbH));
3010 			      Tcz = VFNMS(LDK(KP471396736), Tby, VMUL(LDK(KP881921264), TbH));
3011 			      Tc9 = VSUB(TbX, Tc8);
3012 			      Tci = VSUB(Tce, Tch);
3013 			      Tcj = VFNMS(LDK(KP471396736), Tci, VMUL(LDK(KP881921264), Tc9));
3014 			      TcA = VFMA(LDK(KP471396736), Tc9, VMUL(LDK(KP881921264), Tci));
3015 			 }
3016 			 Tck = VSUB(TbI, Tcj);
3017 			 TcJ = VADD(Tcz, TcA);
3018 			 TcB = VSUB(Tcz, TcA);
3019 			 TcF = VADD(TbI, Tcj);
3020 		    }
3021 		    {
3022 			 V TcS, Td1, TcV, Td2;
3023 			 {
3024 			      V TcQ, TcR, TcT, TcU;
3025 			      TcQ = VADD(Tbx, Tbm);
3026 			      TcR = VADD(TbD, TbG);
3027 			      TcS = VFMA(LDK(KP956940335), TcQ, VMUL(LDK(KP290284677), TcR));
3028 			      Td1 = VFNMS(LDK(KP290284677), TcQ, VMUL(LDK(KP956940335), TcR));
3029 			      TcT = VADD(Tc8, TbX);
3030 			      TcU = VADD(Tce, Tch);
3031 			      TcV = VFNMS(LDK(KP290284677), TcU, VMUL(LDK(KP956940335), TcT));
3032 			      Td2 = VFMA(LDK(KP290284677), TcT, VMUL(LDK(KP956940335), TcU));
3033 			 }
3034 			 TcW = VSUB(TcS, TcV);
3035 			 Tdb = VADD(Td1, Td2);
3036 			 Td3 = VSUB(Td1, Td2);
3037 			 Td7 = VADD(TcS, TcV);
3038 		    }
3039 		    {
3040 			 V Tcl, TcC, Td9, Tdc;
3041 			 Tcl = VADD(Tb7, Tck);
3042 			 TcC = VBYI(VADD(Tcy, TcB));
3043 			 ST(&(xo[WS(os, 106)]), VSUB(Tcl, TcC), ovs, &(xo[0]));
3044 			 ST(&(xo[WS(os, 22)]), VADD(Tcl, TcC), ovs, &(xo[0]));
3045 			 Td9 = VBYI(VSUB(Td7, Td8));
3046 			 Tdc = VSUB(Tda, Tdb);
3047 			 ST(&(xo[WS(os, 58)]), VADD(Td9, Tdc), ovs, &(xo[0]));
3048 			 ST(&(xo[WS(os, 70)]), VSUB(Tdc, Td9), ovs, &(xo[0]));
3049 		    }
3050 		    {
3051 			 V Tdd, Tde, TcD, TcE;
3052 			 Tdd = VBYI(VADD(Td8, Td7));
3053 			 Tde = VADD(Tda, Tdb);
3054 			 ST(&(xo[WS(os, 6)]), VADD(Tdd, Tde), ovs, &(xo[0]));
3055 			 ST(&(xo[WS(os, 122)]), VSUB(Tde, Tdd), ovs, &(xo[0]));
3056 			 TcD = VSUB(Tb7, Tck);
3057 			 TcE = VBYI(VSUB(TcB, Tcy));
3058 			 ST(&(xo[WS(os, 86)]), VSUB(TcD, TcE), ovs, &(xo[0]));
3059 			 ST(&(xo[WS(os, 42)]), VADD(TcD, TcE), ovs, &(xo[0]));
3060 		    }
3061 		    {
3062 			 V TcH, TcK, TcX, Td4;
3063 			 TcH = VBYI(VSUB(TcF, TcG));
3064 			 TcK = VSUB(TcI, TcJ);
3065 			 ST(&(xo[WS(os, 54)]), VADD(TcH, TcK), ovs, &(xo[0]));
3066 			 ST(&(xo[WS(os, 74)]), VSUB(TcK, TcH), ovs, &(xo[0]));
3067 			 TcX = VADD(TcP, TcW);
3068 			 Td4 = VBYI(VADD(Td0, Td3));
3069 			 ST(&(xo[WS(os, 102)]), VSUB(TcX, Td4), ovs, &(xo[0]));
3070 			 ST(&(xo[WS(os, 26)]), VADD(TcX, Td4), ovs, &(xo[0]));
3071 		    }
3072 		    {
3073 			 V Td5, Td6, TcL, TcM;
3074 			 Td5 = VSUB(TcP, TcW);
3075 			 Td6 = VBYI(VSUB(Td3, Td0));
3076 			 ST(&(xo[WS(os, 90)]), VSUB(Td5, Td6), ovs, &(xo[0]));
3077 			 ST(&(xo[WS(os, 38)]), VADD(Td5, Td6), ovs, &(xo[0]));
3078 			 TcL = VBYI(VADD(TcG, TcF));
3079 			 TcM = VADD(TcI, TcJ);
3080 			 ST(&(xo[WS(os, 10)]), VADD(TcL, TcM), ovs, &(xo[0]));
3081 			 ST(&(xo[WS(os, 118)]), VSUB(TcM, TcL), ovs, &(xo[0]));
3082 		    }
3083 	       }
3084 	       {
3085 		    V TdE, Tel, TdW, Tee, TdM, Teo, TdT, Tea, Tdt, TdX, TdP, TdU, Te7, Tep, Teh;
3086 		    V Tem;
3087 		    {
3088 			 V TdD, Tec, TdA, Ted, Tdw, Tdz;
3089 			 TdD = VADD(TdB, TdC);
3090 			 Tec = VSUB(TdG, TdH);
3091 			 Tdw = VFMA(LDK(KP980785280), Tdu, VMUL(LDK(KP195090322), Tdv));
3092 			 Tdz = VFNMS(LDK(KP195090322), Tdy, VMUL(LDK(KP980785280), Tdx));
3093 			 TdA = VADD(Tdw, Tdz);
3094 			 Ted = VSUB(Tdw, Tdz);
3095 			 TdE = VSUB(TdA, TdD);
3096 			 Tel = VADD(Tec, Ted);
3097 			 TdW = VADD(TdD, TdA);
3098 			 Tee = VSUB(Tec, Ted);
3099 		    }
3100 		    {
3101 			 V TdI, Te9, TdL, Te8, TdJ, TdK;
3102 			 TdI = VADD(TdG, TdH);
3103 			 Te9 = VSUB(TdC, TdB);
3104 			 TdJ = VFNMS(LDK(KP195090322), Tdu, VMUL(LDK(KP980785280), Tdv));
3105 			 TdK = VFMA(LDK(KP195090322), Tdx, VMUL(LDK(KP980785280), Tdy));
3106 			 TdL = VADD(TdJ, TdK);
3107 			 Te8 = VSUB(TdJ, TdK);
3108 			 TdM = VSUB(TdI, TdL);
3109 			 Teo = VADD(Te9, Te8);
3110 			 TdT = VADD(TdI, TdL);
3111 			 Tea = VSUB(Te8, Te9);
3112 		    }
3113 		    {
3114 			 V Tdl, TdN, Tds, TdO;
3115 			 {
3116 			      V Tdh, Tdk, Tdo, Tdr;
3117 			      Tdh = VADD(Tdf, Tdg);
3118 			      Tdk = VADD(Tdi, Tdj);
3119 			      Tdl = VFNMS(LDK(KP098017140), Tdk, VMUL(LDK(KP995184726), Tdh));
3120 			      TdN = VFMA(LDK(KP098017140), Tdh, VMUL(LDK(KP995184726), Tdk));
3121 			      Tdo = VADD(Tdm, Tdn);
3122 			      Tdr = VADD(Tdp, Tdq);
3123 			      Tds = VFMA(LDK(KP995184726), Tdo, VMUL(LDK(KP098017140), Tdr));
3124 			      TdO = VFNMS(LDK(KP098017140), Tdo, VMUL(LDK(KP995184726), Tdr));
3125 			 }
3126 			 Tdt = VSUB(Tdl, Tds);
3127 			 TdX = VADD(TdN, TdO);
3128 			 TdP = VSUB(TdN, TdO);
3129 			 TdU = VADD(Tdl, Tds);
3130 		    }
3131 		    {
3132 			 V Te3, Tef, Te6, Teg;
3133 			 {
3134 			      V Te1, Te2, Te4, Te5;
3135 			      Te1 = VSUB(Tdf, Tdg);
3136 			      Te2 = VSUB(Tdj, Tdi);
3137 			      Te3 = VFNMS(LDK(KP634393284), Te2, VMUL(LDK(KP773010453), Te1));
3138 			      Tef = VFMA(LDK(KP634393284), Te1, VMUL(LDK(KP773010453), Te2));
3139 			      Te4 = VSUB(Tdm, Tdn);
3140 			      Te5 = VSUB(Tdq, Tdp);
3141 			      Te6 = VFMA(LDK(KP773010453), Te4, VMUL(LDK(KP634393284), Te5));
3142 			      Teg = VFNMS(LDK(KP634393284), Te4, VMUL(LDK(KP773010453), Te5));
3143 			 }
3144 			 Te7 = VSUB(Te3, Te6);
3145 			 Tep = VADD(Tef, Teg);
3146 			 Teh = VSUB(Tef, Teg);
3147 			 Tem = VADD(Te3, Te6);
3148 		    }
3149 		    {
3150 			 V TdF, TdQ, Ten, Teq;
3151 			 TdF = VBYI(VSUB(Tdt, TdE));
3152 			 TdQ = VSUB(TdM, TdP);
3153 			 ST(&(xo[WS(os, 34)]), VADD(TdF, TdQ), ovs, &(xo[0]));
3154 			 ST(&(xo[WS(os, 94)]), VSUB(TdQ, TdF), ovs, &(xo[0]));
3155 			 Ten = VADD(Tel, Tem);
3156 			 Teq = VBYI(VADD(Teo, Tep));
3157 			 ST(&(xo[WS(os, 114)]), VSUB(Ten, Teq), ovs, &(xo[0]));
3158 			 ST(&(xo[WS(os, 14)]), VADD(Ten, Teq), ovs, &(xo[0]));
3159 		    }
3160 		    {
3161 			 V Ter, Tes, TdR, TdS;
3162 			 Ter = VSUB(Tel, Tem);
3163 			 Tes = VBYI(VSUB(Tep, Teo));
3164 			 ST(&(xo[WS(os, 78)]), VSUB(Ter, Tes), ovs, &(xo[0]));
3165 			 ST(&(xo[WS(os, 50)]), VADD(Ter, Tes), ovs, &(xo[0]));
3166 			 TdR = VBYI(VADD(TdE, Tdt));
3167 			 TdS = VADD(TdM, TdP);
3168 			 ST(&(xo[WS(os, 30)]), VADD(TdR, TdS), ovs, &(xo[0]));
3169 			 ST(&(xo[WS(os, 98)]), VSUB(TdS, TdR), ovs, &(xo[0]));
3170 		    }
3171 		    {
3172 			 V TdV, TdY, Teb, Tei;
3173 			 TdV = VADD(TdT, TdU);
3174 			 TdY = VBYI(VADD(TdW, TdX));
3175 			 ST(&(xo[WS(os, 126)]), VSUB(TdV, TdY), ovs, &(xo[0]));
3176 			 ST(&(xo[WS(os, 2)]), VADD(TdV, TdY), ovs, &(xo[0]));
3177 			 Teb = VBYI(VSUB(Te7, Tea));
3178 			 Tei = VSUB(Tee, Teh);
3179 			 ST(&(xo[WS(os, 46)]), VADD(Teb, Tei), ovs, &(xo[0]));
3180 			 ST(&(xo[WS(os, 82)]), VSUB(Tei, Teb), ovs, &(xo[0]));
3181 		    }
3182 		    {
3183 			 V Tej, Tek, TdZ, Te0;
3184 			 Tej = VBYI(VADD(Tea, Te7));
3185 			 Tek = VADD(Tee, Teh);
3186 			 ST(&(xo[WS(os, 18)]), VADD(Tej, Tek), ovs, &(xo[0]));
3187 			 ST(&(xo[WS(os, 110)]), VSUB(Tek, Tej), ovs, &(xo[0]));
3188 			 TdZ = VSUB(TdT, TdU);
3189 			 Te0 = VBYI(VSUB(TdX, TdW));
3190 			 ST(&(xo[WS(os, 66)]), VSUB(TdZ, Te0), ovs, &(xo[0]));
3191 			 ST(&(xo[WS(os, 62)]), VADD(TdZ, Te0), ovs, &(xo[0]));
3192 		    }
3193 	       }
3194 	       {
3195 		    V T7z, T8n, T8f, T8k, T8x, T8P, T8H, T8M, T80, T8L, T8O, T8c, T8j, T8A, T8E;
3196 		    V T8m;
3197 		    {
3198 			 V T7f, T8d, T7y, T8e;
3199 			 {
3200 			      V T77, T7e, T7q, T7x;
3201 			      T77 = VADD(T6Z, T76);
3202 			      T7e = VADD(T7a, T7d);
3203 			      T7f = VFNMS(LDK(KP336889853), T7e, VMUL(LDK(KP941544065), T77));
3204 			      T8d = VFMA(LDK(KP336889853), T77, VMUL(LDK(KP941544065), T7e));
3205 			      T7q = VADD(T7i, T7p);
3206 			      T7x = VADD(T7t, T7w);
3207 			      T7y = VFMA(LDK(KP941544065), T7q, VMUL(LDK(KP336889853), T7x));
3208 			      T8e = VFNMS(LDK(KP336889853), T7q, VMUL(LDK(KP941544065), T7x));
3209 			 }
3210 			 T7z = VSUB(T7f, T7y);
3211 			 T8n = VADD(T8d, T8e);
3212 			 T8f = VSUB(T8d, T8e);
3213 			 T8k = VADD(T7f, T7y);
3214 		    }
3215 		    {
3216 			 V T8t, T8F, T8w, T8G;
3217 			 {
3218 			      V T8r, T8s, T8u, T8v;
3219 			      T8r = VSUB(T6Z, T76);
3220 			      T8s = VSUB(T7d, T7a);
3221 			      T8t = VFNMS(LDK(KP427555093), T8s, VMUL(LDK(KP903989293), T8r));
3222 			      T8F = VFMA(LDK(KP427555093), T8r, VMUL(LDK(KP903989293), T8s));
3223 			      T8u = VSUB(T7i, T7p);
3224 			      T8v = VSUB(T7w, T7t);
3225 			      T8w = VFMA(LDK(KP903989293), T8u, VMUL(LDK(KP427555093), T8v));
3226 			      T8G = VFNMS(LDK(KP427555093), T8u, VMUL(LDK(KP903989293), T8v));
3227 			 }
3228 			 T8x = VSUB(T8t, T8w);
3229 			 T8P = VADD(T8F, T8G);
3230 			 T8H = VSUB(T8F, T8G);
3231 			 T8M = VADD(T8t, T8w);
3232 		    }
3233 		    {
3234 			 V T7Z, T8z, T88, T8C, T7O, T8D, T8b, T8y, T7Y, T87;
3235 			 T7Y = VSUB(T7U, T7X);
3236 			 T7Z = VADD(T7R, T7Y);
3237 			 T8z = VSUB(T7Y, T7R);
3238 			 T87 = VSUB(T85, T86);
3239 			 T88 = VADD(T84, T87);
3240 			 T8C = VSUB(T84, T87);
3241 			 {
3242 			      V T7G, T7N, T89, T8a;
3243 			      T7G = VFMA(LDK(KP634393284), T7C, VMUL(LDK(KP773010453), T7F));
3244 			      T7N = VFNMS(LDK(KP634393284), T7M, VMUL(LDK(KP773010453), T7J));
3245 			      T7O = VADD(T7G, T7N);
3246 			      T8D = VSUB(T7G, T7N);
3247 			      T89 = VFNMS(LDK(KP634393284), T7F, VMUL(LDK(KP773010453), T7C));
3248 			      T8a = VFMA(LDK(KP773010453), T7M, VMUL(LDK(KP634393284), T7J));
3249 			      T8b = VADD(T89, T8a);
3250 			      T8y = VSUB(T89, T8a);
3251 			 }
3252 			 T80 = VSUB(T7O, T7Z);
3253 			 T8L = VADD(T8C, T8D);
3254 			 T8O = VADD(T8z, T8y);
3255 			 T8c = VSUB(T88, T8b);
3256 			 T8j = VADD(T88, T8b);
3257 			 T8A = VSUB(T8y, T8z);
3258 			 T8E = VSUB(T8C, T8D);
3259 			 T8m = VADD(T7Z, T7O);
3260 		    }
3261 		    {
3262 			 V T81, T8g, T8N, T8Q;
3263 			 T81 = VBYI(VSUB(T7z, T80));
3264 			 T8g = VSUB(T8c, T8f);
3265 			 ST(&(xo[WS(os, 39)]), VADD(T81, T8g), ovs, &(xo[WS(os, 1)]));
3266 			 ST(&(xo[WS(os, 89)]), VSUB(T8g, T81), ovs, &(xo[WS(os, 1)]));
3267 			 T8N = VADD(T8L, T8M);
3268 			 T8Q = VBYI(VADD(T8O, T8P));
3269 			 ST(&(xo[WS(os, 119)]), VSUB(T8N, T8Q), ovs, &(xo[WS(os, 1)]));
3270 			 ST(&(xo[WS(os, 9)]), VADD(T8N, T8Q), ovs, &(xo[WS(os, 1)]));
3271 		    }
3272 		    {
3273 			 V T8R, T8S, T8h, T8i;
3274 			 T8R = VSUB(T8L, T8M);
3275 			 T8S = VBYI(VSUB(T8P, T8O));
3276 			 ST(&(xo[WS(os, 73)]), VSUB(T8R, T8S), ovs, &(xo[WS(os, 1)]));
3277 			 ST(&(xo[WS(os, 55)]), VADD(T8R, T8S), ovs, &(xo[WS(os, 1)]));
3278 			 T8h = VBYI(VADD(T80, T7z));
3279 			 T8i = VADD(T8c, T8f);
3280 			 ST(&(xo[WS(os, 25)]), VADD(T8h, T8i), ovs, &(xo[WS(os, 1)]));
3281 			 ST(&(xo[WS(os, 103)]), VSUB(T8i, T8h), ovs, &(xo[WS(os, 1)]));
3282 		    }
3283 		    {
3284 			 V T8l, T8o, T8B, T8I;
3285 			 T8l = VADD(T8j, T8k);
3286 			 T8o = VBYI(VADD(T8m, T8n));
3287 			 ST(&(xo[WS(os, 121)]), VSUB(T8l, T8o), ovs, &(xo[WS(os, 1)]));
3288 			 ST(&(xo[WS(os, 7)]), VADD(T8l, T8o), ovs, &(xo[WS(os, 1)]));
3289 			 T8B = VBYI(VSUB(T8x, T8A));
3290 			 T8I = VSUB(T8E, T8H);
3291 			 ST(&(xo[WS(os, 41)]), VADD(T8B, T8I), ovs, &(xo[WS(os, 1)]));
3292 			 ST(&(xo[WS(os, 87)]), VSUB(T8I, T8B), ovs, &(xo[WS(os, 1)]));
3293 		    }
3294 		    {
3295 			 V T8J, T8K, T8p, T8q;
3296 			 T8J = VBYI(VADD(T8A, T8x));
3297 			 T8K = VADD(T8E, T8H);
3298 			 ST(&(xo[WS(os, 23)]), VADD(T8J, T8K), ovs, &(xo[WS(os, 1)]));
3299 			 ST(&(xo[WS(os, 105)]), VSUB(T8K, T8J), ovs, &(xo[WS(os, 1)]));
3300 			 T8p = VSUB(T8j, T8k);
3301 			 T8q = VBYI(VSUB(T8n, T8m));
3302 			 ST(&(xo[WS(os, 71)]), VSUB(T8p, T8q), ovs, &(xo[WS(os, 1)]));
3303 			 ST(&(xo[WS(os, 57)]), VADD(T8p, T8q), ovs, &(xo[WS(os, 1)]));
3304 		    }
3305 	       }
3306 	       {
3307 		    V T2v, T5d, T55, T5a, T5n, T5F, T5x, T5C, T4K, T5B, T5E, T52, T59, T5q, T5u;
3308 		    V T5c;
3309 		    {
3310 			 V T1f, T53, T2u, T54;
3311 			 {
3312 			      V T11, T1e, T2g, T2t;
3313 			      T11 = VADD(Tr, T10);
3314 			      T1e = VADD(T1a, T1d);
3315 			      T1f = VFNMS(LDK(KP242980179), T1e, VMUL(LDK(KP970031253), T11));
3316 			      T53 = VFMA(LDK(KP242980179), T11, VMUL(LDK(KP970031253), T1e));
3317 			      T2g = VADD(T1G, T2f);
3318 			      T2t = VADD(T2p, T2s);
3319 			      T2u = VFMA(LDK(KP970031253), T2g, VMUL(LDK(KP242980179), T2t));
3320 			      T54 = VFNMS(LDK(KP242980179), T2g, VMUL(LDK(KP970031253), T2t));
3321 			 }
3322 			 T2v = VSUB(T1f, T2u);
3323 			 T5d = VADD(T53, T54);
3324 			 T55 = VSUB(T53, T54);
3325 			 T5a = VADD(T1f, T2u);
3326 		    }
3327 		    {
3328 			 V T5j, T5v, T5m, T5w;
3329 			 {
3330 			      V T5h, T5i, T5k, T5l;
3331 			      T5h = VSUB(Tr, T10);
3332 			      T5i = VSUB(T1d, T1a);
3333 			      T5j = VFNMS(LDK(KP514102744), T5i, VMUL(LDK(KP857728610), T5h));
3334 			      T5v = VFMA(LDK(KP514102744), T5h, VMUL(LDK(KP857728610), T5i));
3335 			      T5k = VSUB(T1G, T2f);
3336 			      T5l = VSUB(T2s, T2p);
3337 			      T5m = VFMA(LDK(KP857728610), T5k, VMUL(LDK(KP514102744), T5l));
3338 			      T5w = VFNMS(LDK(KP514102744), T5k, VMUL(LDK(KP857728610), T5l));
3339 			 }
3340 			 T5n = VSUB(T5j, T5m);
3341 			 T5F = VADD(T5v, T5w);
3342 			 T5x = VSUB(T5v, T5w);
3343 			 T5C = VADD(T5j, T5m);
3344 		    }
3345 		    {
3346 			 V T4J, T5p, T4Y, T5s, T3I, T5t, T51, T5o, T4I, T4X;
3347 			 T4I = VSUB(T4q, T4H);
3348 			 T4J = VADD(T49, T4I);
3349 			 T5p = VSUB(T4I, T49);
3350 			 T4X = VSUB(T4V, T4W);
3351 			 T4Y = VADD(T4U, T4X);
3352 			 T5s = VSUB(T4U, T4X);
3353 			 {
3354 			      V T36, T3H, T4Z, T50;
3355 			      T36 = VFMA(LDK(KP881921264), T2W, VMUL(LDK(KP471396736), T35));
3356 			      T3H = VFNMS(LDK(KP471396736), T3G, VMUL(LDK(KP881921264), T3x));
3357 			      T3I = VADD(T36, T3H);
3358 			      T5t = VSUB(T36, T3H);
3359 			      T4Z = VFNMS(LDK(KP471396736), T2W, VMUL(LDK(KP881921264), T35));
3360 			      T50 = VFMA(LDK(KP471396736), T3x, VMUL(LDK(KP881921264), T3G));
3361 			      T51 = VADD(T4Z, T50);
3362 			      T5o = VSUB(T4Z, T50);
3363 			 }
3364 			 T4K = VSUB(T3I, T4J);
3365 			 T5B = VADD(T5s, T5t);
3366 			 T5E = VADD(T5p, T5o);
3367 			 T52 = VSUB(T4Y, T51);
3368 			 T59 = VADD(T4Y, T51);
3369 			 T5q = VSUB(T5o, T5p);
3370 			 T5u = VSUB(T5s, T5t);
3371 			 T5c = VADD(T4J, T3I);
3372 		    }
3373 		    {
3374 			 V T4L, T56, T5D, T5G;
3375 			 T4L = VBYI(VSUB(T2v, T4K));
3376 			 T56 = VSUB(T52, T55);
3377 			 ST(&(xo[WS(os, 37)]), VADD(T4L, T56), ovs, &(xo[WS(os, 1)]));
3378 			 ST(&(xo[WS(os, 91)]), VSUB(T56, T4L), ovs, &(xo[WS(os, 1)]));
3379 			 T5D = VADD(T5B, T5C);
3380 			 T5G = VBYI(VADD(T5E, T5F));
3381 			 ST(&(xo[WS(os, 117)]), VSUB(T5D, T5G), ovs, &(xo[WS(os, 1)]));
3382 			 ST(&(xo[WS(os, 11)]), VADD(T5D, T5G), ovs, &(xo[WS(os, 1)]));
3383 		    }
3384 		    {
3385 			 V T5H, T5I, T57, T58;
3386 			 T5H = VSUB(T5B, T5C);
3387 			 T5I = VBYI(VSUB(T5F, T5E));
3388 			 ST(&(xo[WS(os, 75)]), VSUB(T5H, T5I), ovs, &(xo[WS(os, 1)]));
3389 			 ST(&(xo[WS(os, 53)]), VADD(T5H, T5I), ovs, &(xo[WS(os, 1)]));
3390 			 T57 = VBYI(VADD(T4K, T2v));
3391 			 T58 = VADD(T52, T55);
3392 			 ST(&(xo[WS(os, 27)]), VADD(T57, T58), ovs, &(xo[WS(os, 1)]));
3393 			 ST(&(xo[WS(os, 101)]), VSUB(T58, T57), ovs, &(xo[WS(os, 1)]));
3394 		    }
3395 		    {
3396 			 V T5b, T5e, T5r, T5y;
3397 			 T5b = VADD(T59, T5a);
3398 			 T5e = VBYI(VADD(T5c, T5d));
3399 			 ST(&(xo[WS(os, 123)]), VSUB(T5b, T5e), ovs, &(xo[WS(os, 1)]));
3400 			 ST(&(xo[WS(os, 5)]), VADD(T5b, T5e), ovs, &(xo[WS(os, 1)]));
3401 			 T5r = VBYI(VSUB(T5n, T5q));
3402 			 T5y = VSUB(T5u, T5x);
3403 			 ST(&(xo[WS(os, 43)]), VADD(T5r, T5y), ovs, &(xo[WS(os, 1)]));
3404 			 ST(&(xo[WS(os, 85)]), VSUB(T5y, T5r), ovs, &(xo[WS(os, 1)]));
3405 		    }
3406 		    {
3407 			 V T5z, T5A, T5f, T5g;
3408 			 T5z = VBYI(VADD(T5q, T5n));
3409 			 T5A = VADD(T5u, T5x);
3410 			 ST(&(xo[WS(os, 21)]), VADD(T5z, T5A), ovs, &(xo[WS(os, 1)]));
3411 			 ST(&(xo[WS(os, 107)]), VSUB(T5A, T5z), ovs, &(xo[WS(os, 1)]));
3412 			 T5f = VSUB(T59, T5a);
3413 			 T5g = VBYI(VSUB(T5d, T5c));
3414 			 ST(&(xo[WS(os, 69)]), VSUB(T5f, T5g), ovs, &(xo[WS(os, 1)]));
3415 			 ST(&(xo[WS(os, 59)]), VADD(T5f, T5g), ovs, &(xo[WS(os, 1)]));
3416 		    }
3417 	       }
3418 	       {
3419 		    V T9i, T9B, T9t, T9x, T9O, Ta3, T9V, T9Z, T93, Ta0, Ta2, T9q, T9y, T9H, T9S;
3420 		    V T9A;
3421 		    {
3422 			 V T9a, T9r, T9h, T9s;
3423 			 {
3424 			      V T96, T99, T9d, T9g;
3425 			      T96 = VSUB(T94, T95);
3426 			      T99 = VSUB(T97, T98);
3427 			      T9a = VFMA(LDK(KP740951125), T96, VMUL(LDK(KP671558954), T99));
3428 			      T9r = VFNMS(LDK(KP671558954), T96, VMUL(LDK(KP740951125), T99));
3429 			      T9d = VSUB(T9b, T9c);
3430 			      T9g = VSUB(T9e, T9f);
3431 			      T9h = VFNMS(LDK(KP671558954), T9g, VMUL(LDK(KP740951125), T9d));
3432 			      T9s = VFMA(LDK(KP671558954), T9d, VMUL(LDK(KP740951125), T9g));
3433 			 }
3434 			 T9i = VSUB(T9a, T9h);
3435 			 T9B = VADD(T9r, T9s);
3436 			 T9t = VSUB(T9r, T9s);
3437 			 T9x = VADD(T9a, T9h);
3438 		    }
3439 		    {
3440 			 V T9K, T9T, T9N, T9U;
3441 			 {
3442 			      V T9I, T9J, T9L, T9M;
3443 			      T9I = VADD(T95, T94);
3444 			      T9J = VADD(T97, T98);
3445 			      T9K = VFMA(LDK(KP998795456), T9I, VMUL(LDK(KP049067674), T9J));
3446 			      T9T = VFNMS(LDK(KP049067674), T9I, VMUL(LDK(KP998795456), T9J));
3447 			      T9L = VADD(T9c, T9b);
3448 			      T9M = VADD(T9e, T9f);
3449 			      T9N = VFNMS(LDK(KP049067674), T9M, VMUL(LDK(KP998795456), T9L));
3450 			      T9U = VFMA(LDK(KP049067674), T9L, VMUL(LDK(KP998795456), T9M));
3451 			 }
3452 			 T9O = VSUB(T9K, T9N);
3453 			 Ta3 = VADD(T9T, T9U);
3454 			 T9V = VSUB(T9T, T9U);
3455 			 T9Z = VADD(T9K, T9N);
3456 		    }
3457 		    {
3458 			 V T8V, T9F, T9p, T9R, T92, T9Q, T9m, T9G, T8U, T9n;
3459 			 T8U = VADD(T7U, T7X);
3460 			 T8V = VSUB(T8T, T8U);
3461 			 T9F = VADD(T8T, T8U);
3462 			 T9n = VADD(T85, T86);
3463 			 T9p = VSUB(T9n, T9o);
3464 			 T9R = VADD(T9o, T9n);
3465 			 {
3466 			      V T8Y, T91, T9k, T9l;
3467 			      T8Y = VFMA(LDK(KP098017140), T8W, VMUL(LDK(KP995184726), T8X));
3468 			      T91 = VFNMS(LDK(KP098017140), T90, VMUL(LDK(KP995184726), T8Z));
3469 			      T92 = VSUB(T8Y, T91);
3470 			      T9Q = VADD(T8Y, T91);
3471 			      T9k = VFNMS(LDK(KP098017140), T8X, VMUL(LDK(KP995184726), T8W));
3472 			      T9l = VFMA(LDK(KP995184726), T90, VMUL(LDK(KP098017140), T8Z));
3473 			      T9m = VSUB(T9k, T9l);
3474 			      T9G = VADD(T9k, T9l);
3475 			 }
3476 			 T93 = VSUB(T8V, T92);
3477 			 Ta0 = VADD(T9R, T9Q);
3478 			 Ta2 = VADD(T9F, T9G);
3479 			 T9q = VSUB(T9m, T9p);
3480 			 T9y = VADD(T9p, T9m);
3481 			 T9H = VSUB(T9F, T9G);
3482 			 T9S = VSUB(T9Q, T9R);
3483 			 T9A = VADD(T8V, T92);
3484 		    }
3485 		    {
3486 			 V T9j, T9u, Ta1, Ta4;
3487 			 T9j = VADD(T93, T9i);
3488 			 T9u = VBYI(VADD(T9q, T9t));
3489 			 ST(&(xo[WS(os, 111)]), VSUB(T9j, T9u), ovs, &(xo[WS(os, 1)]));
3490 			 ST(&(xo[WS(os, 17)]), VADD(T9j, T9u), ovs, &(xo[WS(os, 1)]));
3491 			 Ta1 = VBYI(VSUB(T9Z, Ta0));
3492 			 Ta4 = VSUB(Ta2, Ta3);
3493 			 ST(&(xo[WS(os, 63)]), VADD(Ta1, Ta4), ovs, &(xo[WS(os, 1)]));
3494 			 ST(&(xo[WS(os, 65)]), VSUB(Ta4, Ta1), ovs, &(xo[WS(os, 1)]));
3495 		    }
3496 		    {
3497 			 V Ta5, Ta6, T9v, T9w;
3498 			 Ta5 = VBYI(VADD(Ta0, T9Z));
3499 			 Ta6 = VADD(Ta2, Ta3);
3500 			 ST(&(xo[WS(os, 1)]), VADD(Ta5, Ta6), ovs, &(xo[WS(os, 1)]));
3501 			 ST(&(xo[WS(os, 127)]), VSUB(Ta6, Ta5), ovs, &(xo[WS(os, 1)]));
3502 			 T9v = VSUB(T93, T9i);
3503 			 T9w = VBYI(VSUB(T9t, T9q));
3504 			 ST(&(xo[WS(os, 81)]), VSUB(T9v, T9w), ovs, &(xo[WS(os, 1)]));
3505 			 ST(&(xo[WS(os, 47)]), VADD(T9v, T9w), ovs, &(xo[WS(os, 1)]));
3506 		    }
3507 		    {
3508 			 V T9z, T9C, T9P, T9W;
3509 			 T9z = VBYI(VSUB(T9x, T9y));
3510 			 T9C = VSUB(T9A, T9B);
3511 			 ST(&(xo[WS(os, 49)]), VADD(T9z, T9C), ovs, &(xo[WS(os, 1)]));
3512 			 ST(&(xo[WS(os, 79)]), VSUB(T9C, T9z), ovs, &(xo[WS(os, 1)]));
3513 			 T9P = VADD(T9H, T9O);
3514 			 T9W = VBYI(VADD(T9S, T9V));
3515 			 ST(&(xo[WS(os, 97)]), VSUB(T9P, T9W), ovs, &(xo[WS(os, 1)]));
3516 			 ST(&(xo[WS(os, 31)]), VADD(T9P, T9W), ovs, &(xo[WS(os, 1)]));
3517 		    }
3518 		    {
3519 			 V T9X, T9Y, T9D, T9E;
3520 			 T9X = VSUB(T9H, T9O);
3521 			 T9Y = VBYI(VSUB(T9V, T9S));
3522 			 ST(&(xo[WS(os, 95)]), VSUB(T9X, T9Y), ovs, &(xo[WS(os, 1)]));
3523 			 ST(&(xo[WS(os, 33)]), VADD(T9X, T9Y), ovs, &(xo[WS(os, 1)]));
3524 			 T9D = VBYI(VADD(T9y, T9x));
3525 			 T9E = VADD(T9A, T9B);
3526 			 ST(&(xo[WS(os, 15)]), VADD(T9D, T9E), ovs, &(xo[WS(os, 1)]));
3527 			 ST(&(xo[WS(os, 113)]), VSUB(T9E, T9D), ovs, &(xo[WS(os, 1)]));
3528 		    }
3529 	       }
3530 	       {
3531 		    V T68, T6r, T6j, T6n, T6E, T6T, T6L, T6P, T5T, T6Q, T6S, T6g, T6o, T6x, T6I;
3532 		    V T6q;
3533 		    {
3534 			 V T60, T6h, T67, T6i;
3535 			 {
3536 			      V T5W, T5Z, T63, T66;
3537 			      T5W = VSUB(T5U, T5V);
3538 			      T5Z = VSUB(T5X, T5Y);
3539 			      T60 = VFMA(LDK(KP803207531), T5W, VMUL(LDK(KP595699304), T5Z));
3540 			      T6h = VFNMS(LDK(KP595699304), T5W, VMUL(LDK(KP803207531), T5Z));
3541 			      T63 = VSUB(T61, T62);
3542 			      T66 = VSUB(T64, T65);
3543 			      T67 = VFNMS(LDK(KP595699304), T66, VMUL(LDK(KP803207531), T63));
3544 			      T6i = VFMA(LDK(KP595699304), T63, VMUL(LDK(KP803207531), T66));
3545 			 }
3546 			 T68 = VSUB(T60, T67);
3547 			 T6r = VADD(T6h, T6i);
3548 			 T6j = VSUB(T6h, T6i);
3549 			 T6n = VADD(T60, T67);
3550 		    }
3551 		    {
3552 			 V T6A, T6J, T6D, T6K;
3553 			 {
3554 			      V T6y, T6z, T6B, T6C;
3555 			      T6y = VADD(T5V, T5U);
3556 			      T6z = VADD(T5X, T5Y);
3557 			      T6A = VFMA(LDK(KP989176509), T6y, VMUL(LDK(KP146730474), T6z));
3558 			      T6J = VFNMS(LDK(KP146730474), T6y, VMUL(LDK(KP989176509), T6z));
3559 			      T6B = VADD(T62, T61);
3560 			      T6C = VADD(T64, T65);
3561 			      T6D = VFNMS(LDK(KP146730474), T6C, VMUL(LDK(KP989176509), T6B));
3562 			      T6K = VFMA(LDK(KP146730474), T6B, VMUL(LDK(KP989176509), T6C));
3563 			 }
3564 			 T6E = VSUB(T6A, T6D);
3565 			 T6T = VADD(T6J, T6K);
3566 			 T6L = VSUB(T6J, T6K);
3567 			 T6P = VADD(T6A, T6D);
3568 		    }
3569 		    {
3570 			 V T5L, T6v, T6f, T6H, T5S, T6G, T6c, T6w, T5K, T6d;
3571 			 T5K = VADD(T4q, T4H);
3572 			 T5L = VSUB(T5J, T5K);
3573 			 T6v = VADD(T5J, T5K);
3574 			 T6d = VADD(T4V, T4W);
3575 			 T6f = VSUB(T6d, T6e);
3576 			 T6H = VADD(T6e, T6d);
3577 			 {
3578 			      V T5O, T5R, T6a, T6b;
3579 			      T5O = VFMA(LDK(KP956940335), T5M, VMUL(LDK(KP290284677), T5N));
3580 			      T5R = VFNMS(LDK(KP290284677), T5Q, VMUL(LDK(KP956940335), T5P));
3581 			      T5S = VSUB(T5O, T5R);
3582 			      T6G = VADD(T5O, T5R);
3583 			      T6a = VFNMS(LDK(KP290284677), T5M, VMUL(LDK(KP956940335), T5N));
3584 			      T6b = VFMA(LDK(KP290284677), T5P, VMUL(LDK(KP956940335), T5Q));
3585 			      T6c = VSUB(T6a, T6b);
3586 			      T6w = VADD(T6a, T6b);
3587 			 }
3588 			 T5T = VSUB(T5L, T5S);
3589 			 T6Q = VADD(T6H, T6G);
3590 			 T6S = VADD(T6v, T6w);
3591 			 T6g = VSUB(T6c, T6f);
3592 			 T6o = VADD(T6f, T6c);
3593 			 T6x = VSUB(T6v, T6w);
3594 			 T6I = VSUB(T6G, T6H);
3595 			 T6q = VADD(T5L, T5S);
3596 		    }
3597 		    {
3598 			 V T69, T6k, T6R, T6U;
3599 			 T69 = VADD(T5T, T68);
3600 			 T6k = VBYI(VADD(T6g, T6j));
3601 			 ST(&(xo[WS(os, 109)]), VSUB(T69, T6k), ovs, &(xo[WS(os, 1)]));
3602 			 ST(&(xo[WS(os, 19)]), VADD(T69, T6k), ovs, &(xo[WS(os, 1)]));
3603 			 T6R = VBYI(VSUB(T6P, T6Q));
3604 			 T6U = VSUB(T6S, T6T);
3605 			 ST(&(xo[WS(os, 61)]), VADD(T6R, T6U), ovs, &(xo[WS(os, 1)]));
3606 			 ST(&(xo[WS(os, 67)]), VSUB(T6U, T6R), ovs, &(xo[WS(os, 1)]));
3607 		    }
3608 		    {
3609 			 V T6V, T6W, T6l, T6m;
3610 			 T6V = VBYI(VADD(T6Q, T6P));
3611 			 T6W = VADD(T6S, T6T);
3612 			 ST(&(xo[WS(os, 3)]), VADD(T6V, T6W), ovs, &(xo[WS(os, 1)]));
3613 			 ST(&(xo[WS(os, 125)]), VSUB(T6W, T6V), ovs, &(xo[WS(os, 1)]));
3614 			 T6l = VSUB(T5T, T68);
3615 			 T6m = VBYI(VSUB(T6j, T6g));
3616 			 ST(&(xo[WS(os, 83)]), VSUB(T6l, T6m), ovs, &(xo[WS(os, 1)]));
3617 			 ST(&(xo[WS(os, 45)]), VADD(T6l, T6m), ovs, &(xo[WS(os, 1)]));
3618 		    }
3619 		    {
3620 			 V T6p, T6s, T6F, T6M;
3621 			 T6p = VBYI(VSUB(T6n, T6o));
3622 			 T6s = VSUB(T6q, T6r);
3623 			 ST(&(xo[WS(os, 51)]), VADD(T6p, T6s), ovs, &(xo[WS(os, 1)]));
3624 			 ST(&(xo[WS(os, 77)]), VSUB(T6s, T6p), ovs, &(xo[WS(os, 1)]));
3625 			 T6F = VADD(T6x, T6E);
3626 			 T6M = VBYI(VADD(T6I, T6L));
3627 			 ST(&(xo[WS(os, 99)]), VSUB(T6F, T6M), ovs, &(xo[WS(os, 1)]));
3628 			 ST(&(xo[WS(os, 29)]), VADD(T6F, T6M), ovs, &(xo[WS(os, 1)]));
3629 		    }
3630 		    {
3631 			 V T6N, T6O, T6t, T6u;
3632 			 T6N = VSUB(T6x, T6E);
3633 			 T6O = VBYI(VSUB(T6L, T6I));
3634 			 ST(&(xo[WS(os, 93)]), VSUB(T6N, T6O), ovs, &(xo[WS(os, 1)]));
3635 			 ST(&(xo[WS(os, 35)]), VADD(T6N, T6O), ovs, &(xo[WS(os, 1)]));
3636 			 T6t = VBYI(VADD(T6o, T6n));
3637 			 T6u = VADD(T6q, T6r);
3638 			 ST(&(xo[WS(os, 13)]), VADD(T6t, T6u), ovs, &(xo[WS(os, 1)]));
3639 			 ST(&(xo[WS(os, 115)]), VSUB(T6u, T6t), ovs, &(xo[WS(os, 1)]));
3640 		    }
3641 	       }
3642 	  }
3643      }
3644      VLEAVE();
3645 }
3646 
3647 static const kdft_desc desc = { 128, XSIMD_STRING("n1bv_128"), { 938, 186, 144, 0 }, &GENUS, 0, 0, 0, 0 };
3648 
XSIMD(codelet_n1bv_128)3649 void XSIMD(codelet_n1bv_128) (planner *p) { X(kdft_register) (p, n1bv_128, &desc);
3650 }
3651 
3652 #endif
3653