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 07:04:55 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 -n 64 -name n2fv_64 -with-ostride 2 -include dft/simd/n2f.h -store-multiple 2 */
29 
30 /*
31  * This function contains 456 FP additions, 258 FP multiplications,
32  * (or, 198 additions, 0 multiplications, 258 fused multiply/add),
33  * 120 stack variables, 15 constants, and 160 memory accesses
34  */
35 #include "dft/simd/n2f.h"
36 
n2fv_64(const R * ri,const R * ii,R * ro,R * io,stride is,stride os,INT v,INT ivs,INT ovs)37 static void n2fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38 {
39      DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
40      DVK(KP820678790, +0.820678790828660330972281985331011598767386482);
41      DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
42      DVK(KP534511135, +0.534511135950791641089685961295362908582039528);
43      DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
44      DVK(KP098491403, +0.098491403357164253077197521291327432293052451);
45      DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
46      DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
47      DVK(KP303346683, +0.303346683607342391675883946941299872384187453);
48      DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
49      DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
50      DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
51      DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
52      DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
53      DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
54      {
55 	  INT i;
56 	  const R *xi;
57 	  R *xo;
58 	  xi = ri;
59 	  xo = ro;
60 	  for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) {
61 	       V T7, T26, T47, T69, T5k, T6A, T2V, T3z, Tm, T27, T5n, T6a, T2Y, T3M, T4e;
62 	       V T6B, TC, T2a, T6e, T6E, T3l, T3A, T4o, T5p, TR, T29, T6h, T6D, T3i, T3B;
63 	       V T4x, T5q, T1N, T2x, T6t, T71, T6w, T72, T1W, T2y, T39, T3H, T57, T5N, T5e;
64 	       V T5O, T3c, T3I, T1g, T2u, T6m, T6Y, T6p, T6Z, T1p, T2v, T32, T3E, T4M, T5K;
65 	       V T4T, T5L, T35, T3F;
66 	       {
67 		    V T3, T43, T25, T44, T6, T5i, T22, T45;
68 		    {
69 			 V T1, T2, T23, T24;
70 			 T1 = LD(&(xi[0]), ivs, &(xi[0]));
71 			 T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
72 			 T3 = VADD(T1, T2);
73 			 T43 = VSUB(T1, T2);
74 			 T23 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
75 			 T24 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
76 			 T25 = VADD(T23, T24);
77 			 T44 = VSUB(T23, T24);
78 		    }
79 		    {
80 			 V T4, T5, T20, T21;
81 			 T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
82 			 T5 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
83 			 T6 = VADD(T4, T5);
84 			 T5i = VSUB(T4, T5);
85 			 T20 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
86 			 T21 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
87 			 T22 = VADD(T20, T21);
88 			 T45 = VSUB(T20, T21);
89 		    }
90 		    T7 = VSUB(T3, T6);
91 		    T26 = VSUB(T22, T25);
92 		    {
93 			 V T46, T5j, T2T, T2U;
94 			 T46 = VADD(T44, T45);
95 			 T47 = VFMA(LDK(KP707106781), T46, T43);
96 			 T69 = VFNMS(LDK(KP707106781), T46, T43);
97 			 T5j = VSUB(T45, T44);
98 			 T5k = VFNMS(LDK(KP707106781), T5j, T5i);
99 			 T6A = VFMA(LDK(KP707106781), T5j, T5i);
100 			 T2T = VADD(T3, T6);
101 			 T2U = VADD(T25, T22);
102 			 T2V = VADD(T2T, T2U);
103 			 T3z = VSUB(T2T, T2U);
104 		    }
105 	       }
106 	       {
107 		    V Ta, T48, Tk, T4c, Td, T49, Th, T4b;
108 		    {
109 			 V T8, T9, Ti, Tj;
110 			 T8 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
111 			 T9 = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
112 			 Ta = VADD(T8, T9);
113 			 T48 = VSUB(T8, T9);
114 			 Ti = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
115 			 Tj = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
116 			 Tk = VADD(Ti, Tj);
117 			 T4c = VSUB(Tj, Ti);
118 		    }
119 		    {
120 			 V Tb, Tc, Tf, Tg;
121 			 Tb = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
122 			 Tc = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
123 			 Td = VADD(Tb, Tc);
124 			 T49 = VSUB(Tb, Tc);
125 			 Tf = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
126 			 Tg = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
127 			 Th = VADD(Tf, Tg);
128 			 T4b = VSUB(Tf, Tg);
129 		    }
130 		    {
131 			 V Te, Tl, T5l, T5m;
132 			 Te = VSUB(Ta, Td);
133 			 Tl = VSUB(Th, Tk);
134 			 Tm = VADD(Te, Tl);
135 			 T27 = VSUB(Tl, Te);
136 			 T5l = VFMA(LDK(KP414213562), T48, T49);
137 			 T5m = VFMA(LDK(KP414213562), T4b, T4c);
138 			 T5n = VSUB(T5l, T5m);
139 			 T6a = VADD(T5l, T5m);
140 		    }
141 		    {
142 			 V T2W, T2X, T4a, T4d;
143 			 T2W = VADD(Ta, Td);
144 			 T2X = VADD(Th, Tk);
145 			 T2Y = VADD(T2W, T2X);
146 			 T3M = VSUB(T2X, T2W);
147 			 T4a = VFNMS(LDK(KP414213562), T49, T48);
148 			 T4d = VFNMS(LDK(KP414213562), T4c, T4b);
149 			 T4e = VADD(T4a, T4d);
150 			 T6B = VSUB(T4d, T4a);
151 		    }
152 	       }
153 	       {
154 		    V Tq, T4g, Tt, T4l, Tx, T4m, TA, T4j;
155 		    {
156 			 V To, Tp, Tr, Ts;
157 			 To = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
158 			 Tp = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
159 			 Tq = VADD(To, Tp);
160 			 T4g = VSUB(To, Tp);
161 			 Tr = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
162 			 Ts = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
163 			 Tt = VADD(Tr, Ts);
164 			 T4l = VSUB(Tr, Ts);
165 			 {
166 			      V Tv, Tw, T4h, Ty, Tz, T4i;
167 			      Tv = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
168 			      Tw = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
169 			      T4h = VSUB(Tv, Tw);
170 			      Ty = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
171 			      Tz = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
172 			      T4i = VSUB(Ty, Tz);
173 			      Tx = VADD(Tv, Tw);
174 			      T4m = VSUB(T4h, T4i);
175 			      TA = VADD(Ty, Tz);
176 			      T4j = VADD(T4h, T4i);
177 			 }
178 		    }
179 		    {
180 			 V Tu, TB, T6c, T6d;
181 			 Tu = VSUB(Tq, Tt);
182 			 TB = VSUB(Tx, TA);
183 			 TC = VFNMS(LDK(KP414213562), TB, Tu);
184 			 T2a = VFMA(LDK(KP414213562), Tu, TB);
185 			 T6c = VFNMS(LDK(KP707106781), T4m, T4l);
186 			 T6d = VFNMS(LDK(KP707106781), T4j, T4g);
187 			 T6e = VFNMS(LDK(KP668178637), T6d, T6c);
188 			 T6E = VFMA(LDK(KP668178637), T6c, T6d);
189 		    }
190 		    {
191 			 V T3j, T3k, T4k, T4n;
192 			 T3j = VADD(Tq, Tt);
193 			 T3k = VADD(Tx, TA);
194 			 T3l = VADD(T3j, T3k);
195 			 T3A = VSUB(T3j, T3k);
196 			 T4k = VFMA(LDK(KP707106781), T4j, T4g);
197 			 T4n = VFMA(LDK(KP707106781), T4m, T4l);
198 			 T4o = VFNMS(LDK(KP198912367), T4n, T4k);
199 			 T5p = VFMA(LDK(KP198912367), T4k, T4n);
200 		    }
201 	       }
202 	       {
203 		    V TF, T4p, TI, T4u, TM, T4v, TP, T4s;
204 		    {
205 			 V TD, TE, TG, TH;
206 			 TD = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
207 			 TE = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
208 			 TF = VADD(TD, TE);
209 			 T4p = VSUB(TD, TE);
210 			 TG = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
211 			 TH = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
212 			 TI = VADD(TG, TH);
213 			 T4u = VSUB(TH, TG);
214 			 {
215 			      V TK, TL, T4r, TN, TO, T4q;
216 			      TK = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
217 			      TL = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
218 			      T4r = VSUB(TK, TL);
219 			      TN = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
220 			      TO = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
221 			      T4q = VSUB(TN, TO);
222 			      TM = VADD(TK, TL);
223 			      T4v = VSUB(T4r, T4q);
224 			      TP = VADD(TN, TO);
225 			      T4s = VADD(T4q, T4r);
226 			 }
227 		    }
228 		    {
229 			 V TJ, TQ, T6f, T6g;
230 			 TJ = VSUB(TF, TI);
231 			 TQ = VSUB(TM, TP);
232 			 TR = VFNMS(LDK(KP414213562), TQ, TJ);
233 			 T29 = VFMA(LDK(KP414213562), TJ, TQ);
234 			 T6f = VFNMS(LDK(KP707106781), T4v, T4u);
235 			 T6g = VFNMS(LDK(KP707106781), T4s, T4p);
236 			 T6h = VFNMS(LDK(KP668178637), T6g, T6f);
237 			 T6D = VFMA(LDK(KP668178637), T6f, T6g);
238 		    }
239 		    {
240 			 V T3g, T3h, T4t, T4w;
241 			 T3g = VADD(TF, TI);
242 			 T3h = VADD(TP, TM);
243 			 T3i = VADD(T3g, T3h);
244 			 T3B = VSUB(T3g, T3h);
245 			 T4t = VFMA(LDK(KP707106781), T4s, T4p);
246 			 T4w = VFMA(LDK(KP707106781), T4v, T4u);
247 			 T4x = VFNMS(LDK(KP198912367), T4w, T4t);
248 			 T5q = VFMA(LDK(KP198912367), T4t, T4w);
249 		    }
250 	       }
251 	       {
252 		    V T1t, T4V, T1w, T58, T1Q, T59, T1T, T4Y, T1A, T1D, T1E, T5b, T52, T1H, T1K;
253 		    V T1L, T5c, T55;
254 		    {
255 			 V T1r, T1s, T1u, T1v;
256 			 T1r = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
257 			 T1s = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
258 			 T1t = VADD(T1r, T1s);
259 			 T4V = VSUB(T1r, T1s);
260 			 T1u = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
261 			 T1v = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
262 			 T1w = VADD(T1u, T1v);
263 			 T58 = VSUB(T1v, T1u);
264 		    }
265 		    {
266 			 V T1O, T1P, T4X, T1R, T1S, T4W;
267 			 T1O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
268 			 T1P = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
269 			 T4X = VSUB(T1O, T1P);
270 			 T1R = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
271 			 T1S = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
272 			 T4W = VSUB(T1R, T1S);
273 			 T1Q = VADD(T1O, T1P);
274 			 T59 = VSUB(T4X, T4W);
275 			 T1T = VADD(T1R, T1S);
276 			 T4Y = VADD(T4W, T4X);
277 		    }
278 		    {
279 			 V T50, T51, T53, T54;
280 			 {
281 			      V T1y, T1z, T1B, T1C;
282 			      T1y = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
283 			      T1z = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
284 			      T1A = VADD(T1y, T1z);
285 			      T50 = VSUB(T1y, T1z);
286 			      T1B = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
287 			      T1C = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
288 			      T1D = VADD(T1B, T1C);
289 			      T51 = VSUB(T1C, T1B);
290 			 }
291 			 T1E = VSUB(T1A, T1D);
292 			 T5b = VFNMS(LDK(KP414213562), T50, T51);
293 			 T52 = VFMA(LDK(KP414213562), T51, T50);
294 			 {
295 			      V T1F, T1G, T1I, T1J;
296 			      T1F = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
297 			      T1G = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
298 			      T1H = VADD(T1F, T1G);
299 			      T53 = VSUB(T1F, T1G);
300 			      T1I = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
301 			      T1J = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
302 			      T1K = VADD(T1I, T1J);
303 			      T54 = VSUB(T1J, T1I);
304 			 }
305 			 T1L = VSUB(T1H, T1K);
306 			 T5c = VFMA(LDK(KP414213562), T53, T54);
307 			 T55 = VFNMS(LDK(KP414213562), T54, T53);
308 		    }
309 		    {
310 			 V T1x, T1M, T6r, T6s;
311 			 T1x = VSUB(T1t, T1w);
312 			 T1M = VADD(T1E, T1L);
313 			 T1N = VFMA(LDK(KP707106781), T1M, T1x);
314 			 T2x = VFNMS(LDK(KP707106781), T1M, T1x);
315 			 T6r = VFNMS(LDK(KP707106781), T4Y, T4V);
316 			 T6s = VSUB(T5c, T5b);
317 			 T6t = VFNMS(LDK(KP923879532), T6s, T6r);
318 			 T71 = VFMA(LDK(KP923879532), T6s, T6r);
319 		    }
320 		    {
321 			 V T6u, T6v, T1U, T1V;
322 			 T6u = VFNMS(LDK(KP707106781), T59, T58);
323 			 T6v = VSUB(T55, T52);
324 			 T6w = VFMA(LDK(KP923879532), T6v, T6u);
325 			 T72 = VFNMS(LDK(KP923879532), T6v, T6u);
326 			 T1U = VSUB(T1Q, T1T);
327 			 T1V = VSUB(T1L, T1E);
328 			 T1W = VFMA(LDK(KP707106781), T1V, T1U);
329 			 T2y = VFNMS(LDK(KP707106781), T1V, T1U);
330 		    }
331 		    {
332 			 V T37, T38, T4Z, T56;
333 			 T37 = VADD(T1t, T1w);
334 			 T38 = VADD(T1T, T1Q);
335 			 T39 = VADD(T37, T38);
336 			 T3H = VSUB(T37, T38);
337 			 T4Z = VFMA(LDK(KP707106781), T4Y, T4V);
338 			 T56 = VADD(T52, T55);
339 			 T57 = VFMA(LDK(KP923879532), T56, T4Z);
340 			 T5N = VFNMS(LDK(KP923879532), T56, T4Z);
341 		    }
342 		    {
343 			 V T5a, T5d, T3a, T3b;
344 			 T5a = VFMA(LDK(KP707106781), T59, T58);
345 			 T5d = VADD(T5b, T5c);
346 			 T5e = VFMA(LDK(KP923879532), T5d, T5a);
347 			 T5O = VFNMS(LDK(KP923879532), T5d, T5a);
348 			 T3a = VADD(T1A, T1D);
349 			 T3b = VADD(T1H, T1K);
350 			 T3c = VADD(T3a, T3b);
351 			 T3I = VSUB(T3b, T3a);
352 		    }
353 	       }
354 	       {
355 		    V TW, T4A, TZ, T4N, T1j, T4O, T1m, T4D, T13, T16, T17, T4Q, T4H, T1a, T1d;
356 		    V T1e, T4R, T4K;
357 		    {
358 			 V TU, TV, TX, TY;
359 			 TU = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
360 			 TV = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
361 			 TW = VADD(TU, TV);
362 			 T4A = VSUB(TU, TV);
363 			 TX = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
364 			 TY = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
365 			 TZ = VADD(TX, TY);
366 			 T4N = VSUB(TX, TY);
367 		    }
368 		    {
369 			 V T1h, T1i, T4B, T1k, T1l, T4C;
370 			 T1h = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
371 			 T1i = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
372 			 T4B = VSUB(T1h, T1i);
373 			 T1k = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
374 			 T1l = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
375 			 T4C = VSUB(T1k, T1l);
376 			 T1j = VADD(T1h, T1i);
377 			 T4O = VSUB(T4B, T4C);
378 			 T1m = VADD(T1k, T1l);
379 			 T4D = VADD(T4B, T4C);
380 		    }
381 		    {
382 			 V T4F, T4G, T4I, T4J;
383 			 {
384 			      V T11, T12, T14, T15;
385 			      T11 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
386 			      T12 = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
387 			      T13 = VADD(T11, T12);
388 			      T4F = VSUB(T11, T12);
389 			      T14 = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
390 			      T15 = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
391 			      T16 = VADD(T14, T15);
392 			      T4G = VSUB(T14, T15);
393 			 }
394 			 T17 = VSUB(T13, T16);
395 			 T4Q = VFMA(LDK(KP414213562), T4F, T4G);
396 			 T4H = VFNMS(LDK(KP414213562), T4G, T4F);
397 			 {
398 			      V T18, T19, T1b, T1c;
399 			      T18 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
400 			      T19 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
401 			      T1a = VADD(T18, T19);
402 			      T4I = VSUB(T18, T19);
403 			      T1b = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
404 			      T1c = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
405 			      T1d = VADD(T1b, T1c);
406 			      T4J = VSUB(T1b, T1c);
407 			 }
408 			 T1e = VSUB(T1a, T1d);
409 			 T4R = VFNMS(LDK(KP414213562), T4I, T4J);
410 			 T4K = VFMA(LDK(KP414213562), T4J, T4I);
411 		    }
412 		    {
413 			 V T10, T1f, T6k, T6l;
414 			 T10 = VSUB(TW, TZ);
415 			 T1f = VADD(T17, T1e);
416 			 T1g = VFMA(LDK(KP707106781), T1f, T10);
417 			 T2u = VFNMS(LDK(KP707106781), T1f, T10);
418 			 T6k = VFNMS(LDK(KP707106781), T4D, T4A);
419 			 T6l = VSUB(T4Q, T4R);
420 			 T6m = VFNMS(LDK(KP923879532), T6l, T6k);
421 			 T6Y = VFMA(LDK(KP923879532), T6l, T6k);
422 		    }
423 		    {
424 			 V T6n, T6o, T1n, T1o;
425 			 T6n = VFNMS(LDK(KP707106781), T4O, T4N);
426 			 T6o = VSUB(T4H, T4K);
427 			 T6p = VFMA(LDK(KP923879532), T6o, T6n);
428 			 T6Z = VFNMS(LDK(KP923879532), T6o, T6n);
429 			 T1n = VSUB(T1j, T1m);
430 			 T1o = VSUB(T17, T1e);
431 			 T1p = VFMA(LDK(KP707106781), T1o, T1n);
432 			 T2v = VFNMS(LDK(KP707106781), T1o, T1n);
433 		    }
434 		    {
435 			 V T30, T31, T4E, T4L;
436 			 T30 = VADD(TW, TZ);
437 			 T31 = VADD(T1j, T1m);
438 			 T32 = VADD(T30, T31);
439 			 T3E = VSUB(T30, T31);
440 			 T4E = VFMA(LDK(KP707106781), T4D, T4A);
441 			 T4L = VADD(T4H, T4K);
442 			 T4M = VFMA(LDK(KP923879532), T4L, T4E);
443 			 T5K = VFNMS(LDK(KP923879532), T4L, T4E);
444 		    }
445 		    {
446 			 V T4P, T4S, T33, T34;
447 			 T4P = VFMA(LDK(KP707106781), T4O, T4N);
448 			 T4S = VADD(T4Q, T4R);
449 			 T4T = VFMA(LDK(KP923879532), T4S, T4P);
450 			 T5L = VFNMS(LDK(KP923879532), T4S, T4P);
451 			 T33 = VADD(T13, T16);
452 			 T34 = VADD(T1a, T1d);
453 			 T35 = VADD(T33, T34);
454 			 T3F = VSUB(T33, T34);
455 		    }
456 	       }
457 	       {
458 		    V T7n, T7o, T7p, T7q, T7r, T7s, T7t, T7u, T7w, T7y, T7A, T7B, T7E, T7G, T7I;
459 		    V T7J;
460 		    {
461 			 V T3t, T3x, T3w, T3y;
462 			 {
463 			      V T3r, T3s, T3u, T3v;
464 			      T3r = VADD(T2V, T2Y);
465 			      T3s = VADD(T3l, T3i);
466 			      T3t = VADD(T3r, T3s);
467 			      T3x = VSUB(T3r, T3s);
468 			      T3u = VADD(T32, T35);
469 			      T3v = VADD(T39, T3c);
470 			      T3w = VADD(T3u, T3v);
471 			      T3y = VSUB(T3v, T3u);
472 			 }
473 			 T7n = VSUB(T3t, T3w);
474 			 STM2(&(xo[64]), T7n, ovs, &(xo[0]));
475 			 T7o = VFMAI(T3y, T3x);
476 			 STM2(&(xo[32]), T7o, ovs, &(xo[0]));
477 			 T7p = VADD(T3t, T3w);
478 			 STM2(&(xo[0]), T7p, ovs, &(xo[0]));
479 			 T7q = VFNMSI(T3y, T3x);
480 			 STM2(&(xo[96]), T7q, ovs, &(xo[0]));
481 		    }
482 		    {
483 			 V T2Z, T3m, T3e, T3n, T36, T3d;
484 			 T2Z = VSUB(T2V, T2Y);
485 			 T3m = VSUB(T3i, T3l);
486 			 T36 = VSUB(T32, T35);
487 			 T3d = VSUB(T39, T3c);
488 			 T3e = VADD(T36, T3d);
489 			 T3n = VSUB(T3d, T36);
490 			 {
491 			      V T3f, T3o, T3p, T3q;
492 			      T3f = VFNMS(LDK(KP707106781), T3e, T2Z);
493 			      T3o = VFNMS(LDK(KP707106781), T3n, T3m);
494 			      T7r = VFNMSI(T3o, T3f);
495 			      STM2(&(xo[48]), T7r, ovs, &(xo[0]));
496 			      T7s = VFMAI(T3o, T3f);
497 			      STM2(&(xo[80]), T7s, ovs, &(xo[0]));
498 			      T3p = VFMA(LDK(KP707106781), T3e, T2Z);
499 			      T3q = VFMA(LDK(KP707106781), T3n, T3m);
500 			      T7t = VFNMSI(T3q, T3p);
501 			      STM2(&(xo[112]), T7t, ovs, &(xo[0]));
502 			      T7u = VFMAI(T3q, T3p);
503 			      STM2(&(xo[16]), T7u, ovs, &(xo[0]));
504 			 }
505 		    }
506 		    {
507 			 V T7v, T7x, T7z, T7C;
508 			 {
509 			      V T3D, T3V, T3O, T3Y, T3K, T3Z, T3R, T3W, T3C, T3N;
510 			      T3C = VADD(T3A, T3B);
511 			      T3D = VFMA(LDK(KP707106781), T3C, T3z);
512 			      T3V = VFNMS(LDK(KP707106781), T3C, T3z);
513 			      T3N = VSUB(T3B, T3A);
514 			      T3O = VFMA(LDK(KP707106781), T3N, T3M);
515 			      T3Y = VFNMS(LDK(KP707106781), T3N, T3M);
516 			      {
517 				   V T3G, T3J, T3P, T3Q;
518 				   T3G = VFNMS(LDK(KP414213562), T3F, T3E);
519 				   T3J = VFNMS(LDK(KP414213562), T3I, T3H);
520 				   T3K = VADD(T3G, T3J);
521 				   T3Z = VSUB(T3J, T3G);
522 				   T3P = VFMA(LDK(KP414213562), T3H, T3I);
523 				   T3Q = VFMA(LDK(KP414213562), T3E, T3F);
524 				   T3R = VSUB(T3P, T3Q);
525 				   T3W = VADD(T3Q, T3P);
526 			      }
527 			      {
528 				   V T3L, T3S, T41, T42;
529 				   T3L = VFNMS(LDK(KP923879532), T3K, T3D);
530 				   T3S = VFNMS(LDK(KP923879532), T3R, T3O);
531 				   T7v = VFNMSI(T3S, T3L);
532 				   STM2(&(xo[56]), T7v, ovs, &(xo[0]));
533 				   T7w = VFMAI(T3S, T3L);
534 				   STM2(&(xo[72]), T7w, ovs, &(xo[0]));
535 				   T41 = VFMA(LDK(KP923879532), T3W, T3V);
536 				   T42 = VFNMS(LDK(KP923879532), T3Z, T3Y);
537 				   T7x = VFNMSI(T42, T41);
538 				   STM2(&(xo[24]), T7x, ovs, &(xo[0]));
539 				   T7y = VFMAI(T42, T41);
540 				   STM2(&(xo[104]), T7y, ovs, &(xo[0]));
541 			      }
542 			      {
543 				   V T3T, T3U, T3X, T40;
544 				   T3T = VFMA(LDK(KP923879532), T3K, T3D);
545 				   T3U = VFMA(LDK(KP923879532), T3R, T3O);
546 				   T7z = VFNMSI(T3U, T3T);
547 				   STM2(&(xo[120]), T7z, ovs, &(xo[0]));
548 				   T7A = VFMAI(T3U, T3T);
549 				   STM2(&(xo[8]), T7A, ovs, &(xo[0]));
550 				   T3X = VFNMS(LDK(KP923879532), T3W, T3V);
551 				   T40 = VFMA(LDK(KP923879532), T3Z, T3Y);
552 				   T7B = VFMAI(T40, T3X);
553 				   STM2(&(xo[40]), T7B, ovs, &(xo[0]));
554 				   T7C = VFNMSI(T40, T3X);
555 				   STM2(&(xo[88]), T7C, ovs, &(xo[0]));
556 			      }
557 			 }
558 			 {
559 			      V T6X, T7f, T7b, T7g, T74, T7j, T78, T7i;
560 			      {
561 				   V T6V, T6W, T79, T7a;
562 				   T6V = VFMA(LDK(KP923879532), T6a, T69);
563 				   T6W = VADD(T6E, T6D);
564 				   T6X = VFMA(LDK(KP831469612), T6W, T6V);
565 				   T7f = VFNMS(LDK(KP831469612), T6W, T6V);
566 				   T79 = VFNMS(LDK(KP303346683), T6Y, T6Z);
567 				   T7a = VFNMS(LDK(KP303346683), T71, T72);
568 				   T7b = VSUB(T79, T7a);
569 				   T7g = VADD(T79, T7a);
570 			      }
571 			      {
572 				   V T70, T73, T76, T77;
573 				   T70 = VFMA(LDK(KP303346683), T6Z, T6Y);
574 				   T73 = VFMA(LDK(KP303346683), T72, T71);
575 				   T74 = VADD(T70, T73);
576 				   T7j = VSUB(T73, T70);
577 				   T76 = VFMA(LDK(KP923879532), T6B, T6A);
578 				   T77 = VSUB(T6e, T6h);
579 				   T78 = VFMA(LDK(KP831469612), T77, T76);
580 				   T7i = VFNMS(LDK(KP831469612), T77, T76);
581 			      }
582 			      {
583 				   V T75, T7c, T7D, T7l, T7m, T7F;
584 				   T75 = VFNMS(LDK(KP956940335), T74, T6X);
585 				   T7c = VFNMS(LDK(KP956940335), T7b, T78);
586 				   T7D = VFNMSI(T7c, T75);
587 				   STM2(&(xo[58]), T7D, ovs, &(xo[2]));
588 				   STN2(&(xo[56]), T7v, T7D, ovs);
589 				   T7E = VFMAI(T7c, T75);
590 				   STM2(&(xo[70]), T7E, ovs, &(xo[2]));
591 				   T7l = VFNMS(LDK(KP956940335), T7g, T7f);
592 				   T7m = VFNMS(LDK(KP956940335), T7j, T7i);
593 				   T7F = VFNMSI(T7m, T7l);
594 				   STM2(&(xo[26]), T7F, ovs, &(xo[2]));
595 				   STN2(&(xo[24]), T7x, T7F, ovs);
596 				   T7G = VFMAI(T7m, T7l);
597 				   STM2(&(xo[102]), T7G, ovs, &(xo[2]));
598 			      }
599 			      {
600 				   V T7d, T7e, T7H, T7h, T7k, T7K;
601 				   T7d = VFMA(LDK(KP956940335), T74, T6X);
602 				   T7e = VFMA(LDK(KP956940335), T7b, T78);
603 				   T7H = VFNMSI(T7e, T7d);
604 				   STM2(&(xo[122]), T7H, ovs, &(xo[2]));
605 				   STN2(&(xo[120]), T7z, T7H, ovs);
606 				   T7I = VFMAI(T7e, T7d);
607 				   STM2(&(xo[6]), T7I, ovs, &(xo[2]));
608 				   T7h = VFMA(LDK(KP956940335), T7g, T7f);
609 				   T7k = VFMA(LDK(KP956940335), T7j, T7i);
610 				   T7J = VFMAI(T7k, T7h);
611 				   STM2(&(xo[38]), T7J, ovs, &(xo[2]));
612 				   T7K = VFNMSI(T7k, T7h);
613 				   STM2(&(xo[90]), T7K, ovs, &(xo[2]));
614 				   STN2(&(xo[88]), T7C, T7K, ovs);
615 			      }
616 			 }
617 		    }
618 		    {
619 			 V T7L, T7N, T7P, T7S;
620 			 {
621 			      V TT, T2j, T2f, T2k, T1Y, T2n, T2c, T2m;
622 			      {
623 				   V Tn, TS, T2d, T2e;
624 				   Tn = VFMA(LDK(KP707106781), Tm, T7);
625 				   TS = VADD(TC, TR);
626 				   TT = VFMA(LDK(KP923879532), TS, Tn);
627 				   T2j = VFNMS(LDK(KP923879532), TS, Tn);
628 				   T2d = VFMA(LDK(KP198912367), T1N, T1W);
629 				   T2e = VFMA(LDK(KP198912367), T1g, T1p);
630 				   T2f = VSUB(T2d, T2e);
631 				   T2k = VADD(T2e, T2d);
632 			      }
633 			      {
634 				   V T1q, T1X, T28, T2b;
635 				   T1q = VFNMS(LDK(KP198912367), T1p, T1g);
636 				   T1X = VFNMS(LDK(KP198912367), T1W, T1N);
637 				   T1Y = VADD(T1q, T1X);
638 				   T2n = VSUB(T1X, T1q);
639 				   T28 = VFMA(LDK(KP707106781), T27, T26);
640 				   T2b = VSUB(T29, T2a);
641 				   T2c = VFMA(LDK(KP923879532), T2b, T28);
642 				   T2m = VFNMS(LDK(KP923879532), T2b, T28);
643 			      }
644 			      {
645 				   V T1Z, T2g, T7M, T2p, T2q, T7O;
646 				   T1Z = VFNMS(LDK(KP980785280), T1Y, TT);
647 				   T2g = VFNMS(LDK(KP980785280), T2f, T2c);
648 				   T7L = VFNMSI(T2g, T1Z);
649 				   STM2(&(xo[60]), T7L, ovs, &(xo[0]));
650 				   T7M = VFMAI(T2g, T1Z);
651 				   STM2(&(xo[68]), T7M, ovs, &(xo[0]));
652 				   STN2(&(xo[68]), T7M, T7E, ovs);
653 				   T2p = VFMA(LDK(KP980785280), T2k, T2j);
654 				   T2q = VFNMS(LDK(KP980785280), T2n, T2m);
655 				   T7N = VFNMSI(T2q, T2p);
656 				   STM2(&(xo[28]), T7N, ovs, &(xo[0]));
657 				   T7O = VFMAI(T2q, T2p);
658 				   STM2(&(xo[100]), T7O, ovs, &(xo[0]));
659 				   STN2(&(xo[100]), T7O, T7G, ovs);
660 			      }
661 			      {
662 				   V T2h, T2i, T7Q, T2l, T2o, T7R;
663 				   T2h = VFMA(LDK(KP980785280), T1Y, TT);
664 				   T2i = VFMA(LDK(KP980785280), T2f, T2c);
665 				   T7P = VFNMSI(T2i, T2h);
666 				   STM2(&(xo[124]), T7P, ovs, &(xo[0]));
667 				   T7Q = VFMAI(T2i, T2h);
668 				   STM2(&(xo[4]), T7Q, ovs, &(xo[0]));
669 				   STN2(&(xo[4]), T7Q, T7I, ovs);
670 				   T2l = VFNMS(LDK(KP980785280), T2k, T2j);
671 				   T2o = VFMA(LDK(KP980785280), T2n, T2m);
672 				   T7R = VFMAI(T2o, T2l);
673 				   STM2(&(xo[36]), T7R, ovs, &(xo[0]));
674 				   STN2(&(xo[36]), T7R, T7J, ovs);
675 				   T7S = VFNMSI(T2o, T2l);
676 				   STM2(&(xo[92]), T7S, ovs, &(xo[0]));
677 			      }
678 			 }
679 			 {
680 			      V T4z, T5z, T5v, T5A, T5g, T5D, T5s, T5C;
681 			      {
682 				   V T4f, T4y, T5t, T5u;
683 				   T4f = VFMA(LDK(KP923879532), T4e, T47);
684 				   T4y = VADD(T4o, T4x);
685 				   T4z = VFMA(LDK(KP980785280), T4y, T4f);
686 				   T5z = VFNMS(LDK(KP980785280), T4y, T4f);
687 				   T5t = VFMA(LDK(KP098491403), T4M, T4T);
688 				   T5u = VFMA(LDK(KP098491403), T57, T5e);
689 				   T5v = VSUB(T5t, T5u);
690 				   T5A = VADD(T5t, T5u);
691 			      }
692 			      {
693 				   V T4U, T5f, T5o, T5r;
694 				   T4U = VFNMS(LDK(KP098491403), T4T, T4M);
695 				   T5f = VFNMS(LDK(KP098491403), T5e, T57);
696 				   T5g = VADD(T4U, T5f);
697 				   T5D = VSUB(T5f, T4U);
698 				   T5o = VFMA(LDK(KP923879532), T5n, T5k);
699 				   T5r = VSUB(T5p, T5q);
700 				   T5s = VFMA(LDK(KP980785280), T5r, T5o);
701 				   T5C = VFNMS(LDK(KP980785280), T5r, T5o);
702 			      }
703 			      {
704 				   V T5h, T5w, T7T, T7U;
705 				   T5h = VFNMS(LDK(KP995184726), T5g, T4z);
706 				   T5w = VFNMS(LDK(KP995184726), T5v, T5s);
707 				   T7T = VFNMSI(T5w, T5h);
708 				   STM2(&(xo[66]), T7T, ovs, &(xo[2]));
709 				   STN2(&(xo[64]), T7n, T7T, ovs);
710 				   T7U = VFMAI(T5w, T5h);
711 				   STM2(&(xo[62]), T7U, ovs, &(xo[2]));
712 				   STN2(&(xo[60]), T7L, T7U, ovs);
713 			      }
714 			      {
715 				   V T5F, T5G, T7V, T7W;
716 				   T5F = VFMA(LDK(KP995184726), T5A, T5z);
717 				   T5G = VFMA(LDK(KP995184726), T5D, T5C);
718 				   T7V = VFMAI(T5G, T5F);
719 				   STM2(&(xo[30]), T7V, ovs, &(xo[2]));
720 				   STN2(&(xo[28]), T7N, T7V, ovs);
721 				   T7W = VFNMSI(T5G, T5F);
722 				   STM2(&(xo[98]), T7W, ovs, &(xo[2]));
723 				   STN2(&(xo[96]), T7q, T7W, ovs);
724 			      }
725 			      {
726 				   V T5x, T5y, T7X, T7Y;
727 				   T5x = VFMA(LDK(KP995184726), T5g, T4z);
728 				   T5y = VFMA(LDK(KP995184726), T5v, T5s);
729 				   T7X = VFNMSI(T5y, T5x);
730 				   STM2(&(xo[2]), T7X, ovs, &(xo[2]));
731 				   STN2(&(xo[0]), T7p, T7X, ovs);
732 				   T7Y = VFMAI(T5y, T5x);
733 				   STM2(&(xo[126]), T7Y, ovs, &(xo[2]));
734 				   STN2(&(xo[124]), T7P, T7Y, ovs);
735 			      }
736 			      {
737 				   V T5B, T5E, T7Z, T80;
738 				   T5B = VFNMS(LDK(KP995184726), T5A, T5z);
739 				   T5E = VFNMS(LDK(KP995184726), T5D, T5C);
740 				   T7Z = VFNMSI(T5E, T5B);
741 				   STM2(&(xo[34]), T7Z, ovs, &(xo[2]));
742 				   STN2(&(xo[32]), T7o, T7Z, ovs);
743 				   T80 = VFMAI(T5E, T5B);
744 				   STM2(&(xo[94]), T80, ovs, &(xo[2]));
745 				   STN2(&(xo[92]), T7S, T80, ovs);
746 			      }
747 			 }
748 		    }
749 		    {
750 			 V T82, T83, T86, T88;
751 			 {
752 			      V T6j, T6N, T6J, T6O, T6y, T6R, T6G, T6Q;
753 			      {
754 				   V T6b, T6i, T6H, T6I;
755 				   T6b = VFNMS(LDK(KP923879532), T6a, T69);
756 				   T6i = VADD(T6e, T6h);
757 				   T6j = VFNMS(LDK(KP831469612), T6i, T6b);
758 				   T6N = VFMA(LDK(KP831469612), T6i, T6b);
759 				   T6H = VFMA(LDK(KP534511135), T6m, T6p);
760 				   T6I = VFMA(LDK(KP534511135), T6t, T6w);
761 				   T6J = VSUB(T6H, T6I);
762 				   T6O = VADD(T6H, T6I);
763 			      }
764 			      {
765 				   V T6q, T6x, T6C, T6F;
766 				   T6q = VFNMS(LDK(KP534511135), T6p, T6m);
767 				   T6x = VFNMS(LDK(KP534511135), T6w, T6t);
768 				   T6y = VADD(T6q, T6x);
769 				   T6R = VSUB(T6x, T6q);
770 				   T6C = VFNMS(LDK(KP923879532), T6B, T6A);
771 				   T6F = VSUB(T6D, T6E);
772 				   T6G = VFNMS(LDK(KP831469612), T6F, T6C);
773 				   T6Q = VFMA(LDK(KP831469612), T6F, T6C);
774 			      }
775 			      {
776 				   V T6z, T6K, T81, T6T, T6U, T84;
777 				   T6z = VFNMS(LDK(KP881921264), T6y, T6j);
778 				   T6K = VFNMS(LDK(KP881921264), T6J, T6G);
779 				   T81 = VFNMSI(T6K, T6z);
780 				   STM2(&(xo[74]), T81, ovs, &(xo[2]));
781 				   STN2(&(xo[72]), T7w, T81, ovs);
782 				   T82 = VFMAI(T6K, T6z);
783 				   STM2(&(xo[54]), T82, ovs, &(xo[2]));
784 				   T6T = VFMA(LDK(KP881921264), T6O, T6N);
785 				   T6U = VFMA(LDK(KP881921264), T6R, T6Q);
786 				   T83 = VFMAI(T6U, T6T);
787 				   STM2(&(xo[22]), T83, ovs, &(xo[2]));
788 				   T84 = VFNMSI(T6U, T6T);
789 				   STM2(&(xo[106]), T84, ovs, &(xo[2]));
790 				   STN2(&(xo[104]), T7y, T84, ovs);
791 			      }
792 			      {
793 				   V T6L, T6M, T85, T6P, T6S, T87;
794 				   T6L = VFMA(LDK(KP881921264), T6y, T6j);
795 				   T6M = VFMA(LDK(KP881921264), T6J, T6G);
796 				   T85 = VFNMSI(T6M, T6L);
797 				   STM2(&(xo[10]), T85, ovs, &(xo[2]));
798 				   STN2(&(xo[8]), T7A, T85, ovs);
799 				   T86 = VFMAI(T6M, T6L);
800 				   STM2(&(xo[118]), T86, ovs, &(xo[2]));
801 				   T6P = VFNMS(LDK(KP881921264), T6O, T6N);
802 				   T6S = VFNMS(LDK(KP881921264), T6R, T6Q);
803 				   T87 = VFNMSI(T6S, T6P);
804 				   STM2(&(xo[42]), T87, ovs, &(xo[2]));
805 				   STN2(&(xo[40]), T7B, T87, ovs);
806 				   T88 = VFMAI(T6S, T6P);
807 				   STM2(&(xo[86]), T88, ovs, &(xo[2]));
808 			      }
809 			 }
810 			 {
811 			      V T89, T8c, T8d, T8f;
812 			      {
813 				   V T2t, T2L, T2H, T2M, T2A, T2P, T2E, T2O;
814 				   {
815 					V T2r, T2s, T2F, T2G;
816 					T2r = VFNMS(LDK(KP707106781), Tm, T7);
817 					T2s = VADD(T2a, T29);
818 					T2t = VFMA(LDK(KP923879532), T2s, T2r);
819 					T2L = VFNMS(LDK(KP923879532), T2s, T2r);
820 					T2F = VFNMS(LDK(KP668178637), T2x, T2y);
821 					T2G = VFNMS(LDK(KP668178637), T2u, T2v);
822 					T2H = VSUB(T2F, T2G);
823 					T2M = VADD(T2G, T2F);
824 				   }
825 				   {
826 					V T2w, T2z, T2C, T2D;
827 					T2w = VFMA(LDK(KP668178637), T2v, T2u);
828 					T2z = VFMA(LDK(KP668178637), T2y, T2x);
829 					T2A = VADD(T2w, T2z);
830 					T2P = VSUB(T2z, T2w);
831 					T2C = VFNMS(LDK(KP707106781), T27, T26);
832 					T2D = VSUB(TR, TC);
833 					T2E = VFNMS(LDK(KP923879532), T2D, T2C);
834 					T2O = VFMA(LDK(KP923879532), T2D, T2C);
835 				   }
836 				   {
837 					V T2B, T2I, T8a, T2R, T2S, T8b;
838 					T2B = VFNMS(LDK(KP831469612), T2A, T2t);
839 					T2I = VFNMS(LDK(KP831469612), T2H, T2E);
840 					T89 = VFNMSI(T2I, T2B);
841 					STM2(&(xo[76]), T89, ovs, &(xo[0]));
842 					T8a = VFMAI(T2I, T2B);
843 					STM2(&(xo[52]), T8a, ovs, &(xo[0]));
844 					STN2(&(xo[52]), T8a, T82, ovs);
845 					T2R = VFNMS(LDK(KP831469612), T2M, T2L);
846 					T2S = VFMA(LDK(KP831469612), T2P, T2O);
847 					T8b = VFMAI(T2S, T2R);
848 					STM2(&(xo[20]), T8b, ovs, &(xo[0]));
849 					STN2(&(xo[20]), T8b, T83, ovs);
850 					T8c = VFNMSI(T2S, T2R);
851 					STM2(&(xo[108]), T8c, ovs, &(xo[0]));
852 				   }
853 				   {
854 					V T2J, T2K, T8e, T2N, T2Q, T8g;
855 					T2J = VFMA(LDK(KP831469612), T2A, T2t);
856 					T2K = VFMA(LDK(KP831469612), T2H, T2E);
857 					T8d = VFNMSI(T2K, T2J);
858 					STM2(&(xo[12]), T8d, ovs, &(xo[0]));
859 					T8e = VFMAI(T2K, T2J);
860 					STM2(&(xo[116]), T8e, ovs, &(xo[0]));
861 					STN2(&(xo[116]), T8e, T86, ovs);
862 					T2N = VFMA(LDK(KP831469612), T2M, T2L);
863 					T2Q = VFNMS(LDK(KP831469612), T2P, T2O);
864 					T8f = VFNMSI(T2Q, T2N);
865 					STM2(&(xo[44]), T8f, ovs, &(xo[0]));
866 					T8g = VFMAI(T2Q, T2N);
867 					STM2(&(xo[84]), T8g, ovs, &(xo[0]));
868 					STN2(&(xo[84]), T8g, T88, ovs);
869 				   }
870 			      }
871 			      {
872 				   V T5J, T61, T5X, T62, T5Q, T65, T5U, T64;
873 				   {
874 					V T5H, T5I, T5V, T5W;
875 					T5H = VFNMS(LDK(KP923879532), T4e, T47);
876 					T5I = VADD(T5p, T5q);
877 					T5J = VFMA(LDK(KP980785280), T5I, T5H);
878 					T61 = VFNMS(LDK(KP980785280), T5I, T5H);
879 					T5V = VFNMS(LDK(KP820678790), T5K, T5L);
880 					T5W = VFNMS(LDK(KP820678790), T5N, T5O);
881 					T5X = VSUB(T5V, T5W);
882 					T62 = VADD(T5V, T5W);
883 				   }
884 				   {
885 					V T5M, T5P, T5S, T5T;
886 					T5M = VFMA(LDK(KP820678790), T5L, T5K);
887 					T5P = VFMA(LDK(KP820678790), T5O, T5N);
888 					T5Q = VADD(T5M, T5P);
889 					T65 = VSUB(T5P, T5M);
890 					T5S = VFNMS(LDK(KP923879532), T5n, T5k);
891 					T5T = VSUB(T4x, T4o);
892 					T5U = VFMA(LDK(KP980785280), T5T, T5S);
893 					T64 = VFNMS(LDK(KP980785280), T5T, T5S);
894 				   }
895 				   {
896 					V T5R, T5Y, T8h, T8i;
897 					T5R = VFNMS(LDK(KP773010453), T5Q, T5J);
898 					T5Y = VFNMS(LDK(KP773010453), T5X, T5U);
899 					T8h = VFNMSI(T5Y, T5R);
900 					STM2(&(xo[50]), T8h, ovs, &(xo[2]));
901 					STN2(&(xo[48]), T7r, T8h, ovs);
902 					T8i = VFMAI(T5Y, T5R);
903 					STM2(&(xo[78]), T8i, ovs, &(xo[2]));
904 					STN2(&(xo[76]), T89, T8i, ovs);
905 				   }
906 				   {
907 					V T67, T68, T8j, T8k;
908 					T67 = VFNMS(LDK(KP773010453), T62, T61);
909 					T68 = VFNMS(LDK(KP773010453), T65, T64);
910 					T8j = VFNMSI(T68, T67);
911 					STM2(&(xo[18]), T8j, ovs, &(xo[2]));
912 					STN2(&(xo[16]), T7u, T8j, ovs);
913 					T8k = VFMAI(T68, T67);
914 					STM2(&(xo[110]), T8k, ovs, &(xo[2]));
915 					STN2(&(xo[108]), T8c, T8k, ovs);
916 				   }
917 				   {
918 					V T5Z, T60, T8l, T8m;
919 					T5Z = VFMA(LDK(KP773010453), T5Q, T5J);
920 					T60 = VFMA(LDK(KP773010453), T5X, T5U);
921 					T8l = VFNMSI(T60, T5Z);
922 					STM2(&(xo[114]), T8l, ovs, &(xo[2]));
923 					STN2(&(xo[112]), T7t, T8l, ovs);
924 					T8m = VFMAI(T60, T5Z);
925 					STM2(&(xo[14]), T8m, ovs, &(xo[2]));
926 					STN2(&(xo[12]), T8d, T8m, ovs);
927 				   }
928 				   {
929 					V T63, T66, T8n, T8o;
930 					T63 = VFMA(LDK(KP773010453), T62, T61);
931 					T66 = VFMA(LDK(KP773010453), T65, T64);
932 					T8n = VFMAI(T66, T63);
933 					STM2(&(xo[46]), T8n, ovs, &(xo[2]));
934 					STN2(&(xo[44]), T8f, T8n, ovs);
935 					T8o = VFNMSI(T66, T63);
936 					STM2(&(xo[82]), T8o, ovs, &(xo[2]));
937 					STN2(&(xo[80]), T7s, T8o, ovs);
938 				   }
939 			      }
940 			 }
941 		    }
942 	       }
943 	  }
944      }
945      VLEAVE();
946 }
947 
948 static const kdft_desc desc = { 64, XSIMD_STRING("n2fv_64"), { 198, 0, 258, 0 }, &GENUS, 0, 2, 0, 0 };
949 
XSIMD(codelet_n2fv_64)950 void XSIMD(codelet_n2fv_64) (planner *p) { X(kdft_register) (p, n2fv_64, &desc);
951 }
952 
953 #else
954 
955 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n2fv_64 -with-ostride 2 -include dft/simd/n2f.h -store-multiple 2 */
956 
957 /*
958  * This function contains 456 FP additions, 124 FP multiplications,
959  * (or, 404 additions, 72 multiplications, 52 fused multiply/add),
960  * 128 stack variables, 15 constants, and 160 memory accesses
961  */
962 #include "dft/simd/n2f.h"
963 
n2fv_64(const R * ri,const R * ii,R * ro,R * io,stride is,stride os,INT v,INT ivs,INT ovs)964 static void n2fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
965 {
966      DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
967      DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
968      DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
969      DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
970      DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
971      DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
972      DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
973      DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
974      DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
975      DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
976      DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
977      DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
978      DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
979      DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
980      DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
981      {
982 	  INT i;
983 	  const R *xi;
984 	  R *xo;
985 	  xi = ri;
986 	  xo = ro;
987 	  for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) {
988 	       V T4p, T5q, Tb, T39, T2n, T3A, T6f, T6T, Tq, T3B, T6i, T76, T2i, T3a, T4w;
989 	       V T5r, TI, T2p, T6C, T6V, T3h, T3E, T4L, T5u, TZ, T2q, T6F, T6U, T3e, T3D;
990 	       V T4E, T5t, T23, T2N, T6t, T71, T6w, T72, T2c, T2O, T3t, T41, T5f, T5R, T5k;
991 	       V T5S, T3w, T42, T1s, T2K, T6m, T6Y, T6p, T6Z, T1B, T2L, T3m, T3Y, T4Y, T5O;
992 	       V T53, T5P, T3p, T3Z;
993 	       {
994 		    V T3, T4n, T2m, T4o, T6, T5p, T9, T5o;
995 		    {
996 			 V T1, T2, T2k, T2l;
997 			 T1 = LD(&(xi[0]), ivs, &(xi[0]));
998 			 T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
999 			 T3 = VSUB(T1, T2);
1000 			 T4n = VADD(T1, T2);
1001 			 T2k = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
1002 			 T2l = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
1003 			 T2m = VSUB(T2k, T2l);
1004 			 T4o = VADD(T2k, T2l);
1005 		    }
1006 		    {
1007 			 V T4, T5, T7, T8;
1008 			 T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
1009 			 T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
1010 			 T6 = VSUB(T4, T5);
1011 			 T5p = VADD(T4, T5);
1012 			 T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
1013 			 T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
1014 			 T9 = VSUB(T7, T8);
1015 			 T5o = VADD(T7, T8);
1016 		    }
1017 		    T4p = VSUB(T4n, T4o);
1018 		    T5q = VSUB(T5o, T5p);
1019 		    {
1020 			 V Ta, T2j, T6d, T6e;
1021 			 Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
1022 			 Tb = VADD(T3, Ta);
1023 			 T39 = VSUB(T3, Ta);
1024 			 T2j = VMUL(LDK(KP707106781), VSUB(T9, T6));
1025 			 T2n = VSUB(T2j, T2m);
1026 			 T3A = VADD(T2m, T2j);
1027 			 T6d = VADD(T4n, T4o);
1028 			 T6e = VADD(T5p, T5o);
1029 			 T6f = VADD(T6d, T6e);
1030 			 T6T = VSUB(T6d, T6e);
1031 		    }
1032 	       }
1033 	       {
1034 		    V Te, T4q, To, T4u, Th, T4r, Tl, T4t;
1035 		    {
1036 			 V Tc, Td, Tm, Tn;
1037 			 Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
1038 			 Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
1039 			 Te = VSUB(Tc, Td);
1040 			 T4q = VADD(Tc, Td);
1041 			 Tm = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
1042 			 Tn = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
1043 			 To = VSUB(Tm, Tn);
1044 			 T4u = VADD(Tm, Tn);
1045 		    }
1046 		    {
1047 			 V Tf, Tg, Tj, Tk;
1048 			 Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
1049 			 Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
1050 			 Th = VSUB(Tf, Tg);
1051 			 T4r = VADD(Tf, Tg);
1052 			 Tj = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
1053 			 Tk = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
1054 			 Tl = VSUB(Tj, Tk);
1055 			 T4t = VADD(Tj, Tk);
1056 		    }
1057 		    {
1058 			 V Ti, Tp, T6g, T6h;
1059 			 Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
1060 			 Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
1061 			 Tq = VADD(Ti, Tp);
1062 			 T3B = VSUB(Tp, Ti);
1063 			 T6g = VADD(T4q, T4r);
1064 			 T6h = VADD(T4t, T4u);
1065 			 T6i = VADD(T6g, T6h);
1066 			 T76 = VSUB(T6h, T6g);
1067 		    }
1068 		    {
1069 			 V T2g, T2h, T4s, T4v;
1070 			 T2g = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
1071 			 T2h = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
1072 			 T2i = VSUB(T2g, T2h);
1073 			 T3a = VADD(T2h, T2g);
1074 			 T4s = VSUB(T4q, T4r);
1075 			 T4v = VSUB(T4t, T4u);
1076 			 T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v));
1077 			 T5r = VMUL(LDK(KP707106781), VSUB(T4v, T4s));
1078 		    }
1079 	       }
1080 	       {
1081 		    V Tu, T4F, TG, T4G, TB, T4J, TD, T4I;
1082 		    {
1083 			 V Ts, Tt, TE, TF;
1084 			 Ts = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
1085 			 Tt = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
1086 			 Tu = VSUB(Ts, Tt);
1087 			 T4F = VADD(Ts, Tt);
1088 			 TE = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
1089 			 TF = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
1090 			 TG = VSUB(TE, TF);
1091 			 T4G = VADD(TE, TF);
1092 			 {
1093 			      V Tv, Tw, Tx, Ty, Tz, TA;
1094 			      Tv = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
1095 			      Tw = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
1096 			      Tx = VSUB(Tv, Tw);
1097 			      Ty = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
1098 			      Tz = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
1099 			      TA = VSUB(Ty, Tz);
1100 			      TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
1101 			      T4J = VADD(Tv, Tw);
1102 			      TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
1103 			      T4I = VADD(Ty, Tz);
1104 			 }
1105 		    }
1106 		    {
1107 			 V TC, TH, T6A, T6B;
1108 			 TC = VADD(Tu, TB);
1109 			 TH = VSUB(TD, TG);
1110 			 TI = VFMA(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
1111 			 T2p = VFNMS(LDK(KP195090322), TH, VMUL(LDK(KP980785280), TC));
1112 			 T6A = VADD(T4F, T4G);
1113 			 T6B = VADD(T4J, T4I);
1114 			 T6C = VADD(T6A, T6B);
1115 			 T6V = VSUB(T6A, T6B);
1116 		    }
1117 		    {
1118 			 V T3f, T3g, T4H, T4K;
1119 			 T3f = VSUB(Tu, TB);
1120 			 T3g = VADD(TG, TD);
1121 			 T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f));
1122 			 T3E = VFMA(LDK(KP555570233), T3f, VMUL(LDK(KP831469612), T3g));
1123 			 T4H = VSUB(T4F, T4G);
1124 			 T4K = VSUB(T4I, T4J);
1125 			 T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H));
1126 			 T5u = VFMA(LDK(KP382683432), T4H, VMUL(LDK(KP923879532), T4K));
1127 		    }
1128 	       }
1129 	       {
1130 		    V TS, T4z, TW, T4y, TP, T4C, TX, T4B;
1131 		    {
1132 			 V TQ, TR, TU, TV;
1133 			 TQ = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
1134 			 TR = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
1135 			 TS = VSUB(TQ, TR);
1136 			 T4z = VADD(TQ, TR);
1137 			 TU = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
1138 			 TV = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
1139 			 TW = VSUB(TU, TV);
1140 			 T4y = VADD(TU, TV);
1141 			 {
1142 			      V TJ, TK, TL, TM, TN, TO;
1143 			      TJ = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
1144 			      TK = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
1145 			      TL = VSUB(TJ, TK);
1146 			      TM = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
1147 			      TN = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
1148 			      TO = VSUB(TM, TN);
1149 			      TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
1150 			      T4C = VADD(TM, TN);
1151 			      TX = VMUL(LDK(KP707106781), VADD(TO, TL));
1152 			      T4B = VADD(TJ, TK);
1153 			 }
1154 		    }
1155 		    {
1156 			 V TT, TY, T6D, T6E;
1157 			 TT = VSUB(TP, TS);
1158 			 TY = VADD(TW, TX);
1159 			 TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
1160 			 T2q = VFMA(LDK(KP980785280), TY, VMUL(LDK(KP195090322), TT));
1161 			 T6D = VADD(T4y, T4z);
1162 			 T6E = VADD(T4C, T4B);
1163 			 T6F = VADD(T6D, T6E);
1164 			 T6U = VSUB(T6D, T6E);
1165 		    }
1166 		    {
1167 			 V T3c, T3d, T4A, T4D;
1168 			 T3c = VSUB(TW, TX);
1169 			 T3d = VADD(TS, TP);
1170 			 T3e = VFMA(LDK(KP831469612), T3c, VMUL(LDK(KP555570233), T3d));
1171 			 T3D = VFNMS(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d));
1172 			 T4A = VSUB(T4y, T4z);
1173 			 T4D = VSUB(T4B, T4C);
1174 			 T4E = VFMA(LDK(KP923879532), T4A, VMUL(LDK(KP382683432), T4D));
1175 			 T5t = VFNMS(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D));
1176 		    }
1177 	       }
1178 	       {
1179 		    V T1F, T55, T2a, T56, T1M, T5h, T27, T5g, T58, T59, T1U, T5a, T25, T5b, T5c;
1180 		    V T21, T5d, T24;
1181 		    {
1182 			 V T1D, T1E, T28, T29;
1183 			 T1D = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
1184 			 T1E = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
1185 			 T1F = VSUB(T1D, T1E);
1186 			 T55 = VADD(T1D, T1E);
1187 			 T28 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
1188 			 T29 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
1189 			 T2a = VSUB(T28, T29);
1190 			 T56 = VADD(T28, T29);
1191 		    }
1192 		    {
1193 			 V T1G, T1H, T1I, T1J, T1K, T1L;
1194 			 T1G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
1195 			 T1H = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
1196 			 T1I = VSUB(T1G, T1H);
1197 			 T1J = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
1198 			 T1K = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
1199 			 T1L = VSUB(T1J, T1K);
1200 			 T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L));
1201 			 T5h = VADD(T1G, T1H);
1202 			 T27 = VMUL(LDK(KP707106781), VSUB(T1L, T1I));
1203 			 T5g = VADD(T1J, T1K);
1204 		    }
1205 		    {
1206 			 V T1Q, T1T, T1X, T20;
1207 			 {
1208 			      V T1O, T1P, T1R, T1S;
1209 			      T1O = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
1210 			      T1P = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
1211 			      T1Q = VSUB(T1O, T1P);
1212 			      T58 = VADD(T1O, T1P);
1213 			      T1R = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
1214 			      T1S = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
1215 			      T1T = VSUB(T1R, T1S);
1216 			      T59 = VADD(T1R, T1S);
1217 			 }
1218 			 T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q));
1219 			 T5a = VSUB(T58, T59);
1220 			 T25 = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T));
1221 			 {
1222 			      V T1V, T1W, T1Y, T1Z;
1223 			      T1V = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
1224 			      T1W = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
1225 			      T1X = VSUB(T1V, T1W);
1226 			      T5b = VADD(T1V, T1W);
1227 			      T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
1228 			      T1Z = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
1229 			      T20 = VSUB(T1Y, T1Z);
1230 			      T5c = VADD(T1Y, T1Z);
1231 			 }
1232 			 T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20));
1233 			 T5d = VSUB(T5b, T5c);
1234 			 T24 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X));
1235 		    }
1236 		    {
1237 			 V T1N, T22, T6r, T6s;
1238 			 T1N = VADD(T1F, T1M);
1239 			 T22 = VADD(T1U, T21);
1240 			 T23 = VSUB(T1N, T22);
1241 			 T2N = VADD(T1N, T22);
1242 			 T6r = VADD(T55, T56);
1243 			 T6s = VADD(T5h, T5g);
1244 			 T6t = VADD(T6r, T6s);
1245 			 T71 = VSUB(T6r, T6s);
1246 		    }
1247 		    {
1248 			 V T6u, T6v, T26, T2b;
1249 			 T6u = VADD(T58, T59);
1250 			 T6v = VADD(T5b, T5c);
1251 			 T6w = VADD(T6u, T6v);
1252 			 T72 = VSUB(T6v, T6u);
1253 			 T26 = VSUB(T24, T25);
1254 			 T2b = VSUB(T27, T2a);
1255 			 T2c = VSUB(T26, T2b);
1256 			 T2O = VADD(T2b, T26);
1257 		    }
1258 		    {
1259 			 V T3r, T3s, T57, T5e;
1260 			 T3r = VSUB(T1F, T1M);
1261 			 T3s = VADD(T25, T24);
1262 			 T3t = VADD(T3r, T3s);
1263 			 T41 = VSUB(T3r, T3s);
1264 			 T57 = VSUB(T55, T56);
1265 			 T5e = VMUL(LDK(KP707106781), VADD(T5a, T5d));
1266 			 T5f = VADD(T57, T5e);
1267 			 T5R = VSUB(T57, T5e);
1268 		    }
1269 		    {
1270 			 V T5i, T5j, T3u, T3v;
1271 			 T5i = VSUB(T5g, T5h);
1272 			 T5j = VMUL(LDK(KP707106781), VSUB(T5d, T5a));
1273 			 T5k = VADD(T5i, T5j);
1274 			 T5S = VSUB(T5j, T5i);
1275 			 T3u = VADD(T2a, T27);
1276 			 T3v = VSUB(T21, T1U);
1277 			 T3w = VADD(T3u, T3v);
1278 			 T42 = VSUB(T3v, T3u);
1279 		    }
1280 	       }
1281 	       {
1282 		    V T1q, T4P, T1v, T4O, T1n, T50, T1w, T4Z, T4U, T4V, T18, T4W, T1z, T4R, T4S;
1283 		    V T1f, T4T, T1y;
1284 		    {
1285 			 V T1o, T1p, T1t, T1u;
1286 			 T1o = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
1287 			 T1p = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
1288 			 T1q = VSUB(T1o, T1p);
1289 			 T4P = VADD(T1o, T1p);
1290 			 T1t = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
1291 			 T1u = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
1292 			 T1v = VSUB(T1t, T1u);
1293 			 T4O = VADD(T1t, T1u);
1294 		    }
1295 		    {
1296 			 V T1h, T1i, T1j, T1k, T1l, T1m;
1297 			 T1h = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
1298 			 T1i = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
1299 			 T1j = VSUB(T1h, T1i);
1300 			 T1k = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
1301 			 T1l = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
1302 			 T1m = VSUB(T1k, T1l);
1303 			 T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m));
1304 			 T50 = VADD(T1k, T1l);
1305 			 T1w = VMUL(LDK(KP707106781), VADD(T1m, T1j));
1306 			 T4Z = VADD(T1h, T1i);
1307 		    }
1308 		    {
1309 			 V T14, T17, T1b, T1e;
1310 			 {
1311 			      V T12, T13, T15, T16;
1312 			      T12 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
1313 			      T13 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
1314 			      T14 = VSUB(T12, T13);
1315 			      T4U = VADD(T12, T13);
1316 			      T15 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
1317 			      T16 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
1318 			      T17 = VSUB(T15, T16);
1319 			      T4V = VADD(T15, T16);
1320 			 }
1321 			 T18 = VFNMS(LDK(KP923879532), T17, VMUL(LDK(KP382683432), T14));
1322 			 T4W = VSUB(T4U, T4V);
1323 			 T1z = VFMA(LDK(KP923879532), T14, VMUL(LDK(KP382683432), T17));
1324 			 {
1325 			      V T19, T1a, T1c, T1d;
1326 			      T19 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
1327 			      T1a = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
1328 			      T1b = VSUB(T19, T1a);
1329 			      T4R = VADD(T19, T1a);
1330 			      T1c = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
1331 			      T1d = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
1332 			      T1e = VSUB(T1c, T1d);
1333 			      T4S = VADD(T1c, T1d);
1334 			 }
1335 			 T1f = VFMA(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e));
1336 			 T4T = VSUB(T4R, T4S);
1337 			 T1y = VFNMS(LDK(KP382683432), T1e, VMUL(LDK(KP923879532), T1b));
1338 		    }
1339 		    {
1340 			 V T1g, T1r, T6k, T6l;
1341 			 T1g = VSUB(T18, T1f);
1342 			 T1r = VSUB(T1n, T1q);
1343 			 T1s = VSUB(T1g, T1r);
1344 			 T2K = VADD(T1r, T1g);
1345 			 T6k = VADD(T4O, T4P);
1346 			 T6l = VADD(T50, T4Z);
1347 			 T6m = VADD(T6k, T6l);
1348 			 T6Y = VSUB(T6k, T6l);
1349 		    }
1350 		    {
1351 			 V T6n, T6o, T1x, T1A;
1352 			 T6n = VADD(T4R, T4S);
1353 			 T6o = VADD(T4U, T4V);
1354 			 T6p = VADD(T6n, T6o);
1355 			 T6Z = VSUB(T6o, T6n);
1356 			 T1x = VADD(T1v, T1w);
1357 			 T1A = VADD(T1y, T1z);
1358 			 T1B = VSUB(T1x, T1A);
1359 			 T2L = VADD(T1x, T1A);
1360 		    }
1361 		    {
1362 			 V T3k, T3l, T4Q, T4X;
1363 			 T3k = VSUB(T1v, T1w);
1364 			 T3l = VADD(T1f, T18);
1365 			 T3m = VADD(T3k, T3l);
1366 			 T3Y = VSUB(T3k, T3l);
1367 			 T4Q = VSUB(T4O, T4P);
1368 			 T4X = VMUL(LDK(KP707106781), VADD(T4T, T4W));
1369 			 T4Y = VADD(T4Q, T4X);
1370 			 T5O = VSUB(T4Q, T4X);
1371 		    }
1372 		    {
1373 			 V T51, T52, T3n, T3o;
1374 			 T51 = VSUB(T4Z, T50);
1375 			 T52 = VMUL(LDK(KP707106781), VSUB(T4W, T4T));
1376 			 T53 = VADD(T51, T52);
1377 			 T5P = VSUB(T52, T51);
1378 			 T3n = VADD(T1q, T1n);
1379 			 T3o = VSUB(T1z, T1y);
1380 			 T3p = VADD(T3n, T3o);
1381 			 T3Z = VSUB(T3o, T3n);
1382 		    }
1383 	       }
1384 	       {
1385 		    V T7n, T7o, T7p, T7q, T7r, T7s, T7t, T7u, T7v, T7w, T7x, T7y, T7z, T7A, T7B;
1386 		    V T7C, T7D, T7E, T7F, T7G, T7H, T7I, T7J, T7K;
1387 		    {
1388 			 V T6N, T6R, T6Q, T6S;
1389 			 {
1390 			      V T6L, T6M, T6O, T6P;
1391 			      T6L = VADD(T6f, T6i);
1392 			      T6M = VADD(T6F, T6C);
1393 			      T6N = VADD(T6L, T6M);
1394 			      T6R = VSUB(T6L, T6M);
1395 			      T6O = VADD(T6m, T6p);
1396 			      T6P = VADD(T6t, T6w);
1397 			      T6Q = VADD(T6O, T6P);
1398 			      T6S = VBYI(VSUB(T6P, T6O));
1399 			 }
1400 			 T7n = VSUB(T6N, T6Q);
1401 			 STM2(&(xo[64]), T7n, ovs, &(xo[0]));
1402 			 T7o = VADD(T6R, T6S);
1403 			 STM2(&(xo[32]), T7o, ovs, &(xo[0]));
1404 			 T7p = VADD(T6N, T6Q);
1405 			 STM2(&(xo[0]), T7p, ovs, &(xo[0]));
1406 			 T7q = VSUB(T6R, T6S);
1407 			 STM2(&(xo[96]), T7q, ovs, &(xo[0]));
1408 		    }
1409 		    {
1410 			 V T6j, T6G, T6y, T6H, T6q, T6x;
1411 			 T6j = VSUB(T6f, T6i);
1412 			 T6G = VSUB(T6C, T6F);
1413 			 T6q = VSUB(T6m, T6p);
1414 			 T6x = VSUB(T6t, T6w);
1415 			 T6y = VMUL(LDK(KP707106781), VADD(T6q, T6x));
1416 			 T6H = VMUL(LDK(KP707106781), VSUB(T6x, T6q));
1417 			 {
1418 			      V T6z, T6I, T6J, T6K;
1419 			      T6z = VADD(T6j, T6y);
1420 			      T6I = VBYI(VADD(T6G, T6H));
1421 			      T7r = VSUB(T6z, T6I);
1422 			      STM2(&(xo[112]), T7r, ovs, &(xo[0]));
1423 			      T7s = VADD(T6z, T6I);
1424 			      STM2(&(xo[16]), T7s, ovs, &(xo[0]));
1425 			      T6J = VSUB(T6j, T6y);
1426 			      T6K = VBYI(VSUB(T6H, T6G));
1427 			      T7t = VSUB(T6J, T6K);
1428 			      STM2(&(xo[80]), T7t, ovs, &(xo[0]));
1429 			      T7u = VADD(T6J, T6K);
1430 			      STM2(&(xo[48]), T7u, ovs, &(xo[0]));
1431 			 }
1432 		    }
1433 		    {
1434 			 V T6X, T7i, T78, T7g, T74, T7f, T7b, T7j, T6W, T77;
1435 			 T6W = VMUL(LDK(KP707106781), VADD(T6U, T6V));
1436 			 T6X = VADD(T6T, T6W);
1437 			 T7i = VSUB(T6T, T6W);
1438 			 T77 = VMUL(LDK(KP707106781), VSUB(T6V, T6U));
1439 			 T78 = VADD(T76, T77);
1440 			 T7g = VSUB(T77, T76);
1441 			 {
1442 			      V T70, T73, T79, T7a;
1443 			      T70 = VFMA(LDK(KP923879532), T6Y, VMUL(LDK(KP382683432), T6Z));
1444 			      T73 = VFNMS(LDK(KP382683432), T72, VMUL(LDK(KP923879532), T71));
1445 			      T74 = VADD(T70, T73);
1446 			      T7f = VSUB(T73, T70);
1447 			      T79 = VFNMS(LDK(KP382683432), T6Y, VMUL(LDK(KP923879532), T6Z));
1448 			      T7a = VFMA(LDK(KP382683432), T71, VMUL(LDK(KP923879532), T72));
1449 			      T7b = VADD(T79, T7a);
1450 			      T7j = VSUB(T7a, T79);
1451 			 }
1452 			 {
1453 			      V T75, T7c, T7l, T7m;
1454 			      T75 = VADD(T6X, T74);
1455 			      T7c = VBYI(VADD(T78, T7b));
1456 			      T7v = VSUB(T75, T7c);
1457 			      STM2(&(xo[120]), T7v, ovs, &(xo[0]));
1458 			      T7w = VADD(T75, T7c);
1459 			      STM2(&(xo[8]), T7w, ovs, &(xo[0]));
1460 			      T7l = VBYI(VADD(T7g, T7f));
1461 			      T7m = VADD(T7i, T7j);
1462 			      T7x = VADD(T7l, T7m);
1463 			      STM2(&(xo[24]), T7x, ovs, &(xo[0]));
1464 			      T7y = VSUB(T7m, T7l);
1465 			      STM2(&(xo[104]), T7y, ovs, &(xo[0]));
1466 			 }
1467 			 {
1468 			      V T7d, T7e, T7h, T7k;
1469 			      T7d = VSUB(T6X, T74);
1470 			      T7e = VBYI(VSUB(T7b, T78));
1471 			      T7z = VSUB(T7d, T7e);
1472 			      STM2(&(xo[72]), T7z, ovs, &(xo[0]));
1473 			      T7A = VADD(T7d, T7e);
1474 			      STM2(&(xo[56]), T7A, ovs, &(xo[0]));
1475 			      T7h = VBYI(VSUB(T7f, T7g));
1476 			      T7k = VSUB(T7i, T7j);
1477 			      T7B = VADD(T7h, T7k);
1478 			      STM2(&(xo[40]), T7B, ovs, &(xo[0]));
1479 			      T7C = VSUB(T7k, T7h);
1480 			      STM2(&(xo[88]), T7C, ovs, &(xo[0]));
1481 			 }
1482 		    }
1483 		    {
1484 			 V T5N, T68, T61, T69, T5U, T65, T5Y, T66;
1485 			 {
1486 			      V T5L, T5M, T5Z, T60;
1487 			      T5L = VSUB(T4p, T4w);
1488 			      T5M = VSUB(T5u, T5t);
1489 			      T5N = VADD(T5L, T5M);
1490 			      T68 = VSUB(T5L, T5M);
1491 			      T5Z = VFNMS(LDK(KP555570233), T5O, VMUL(LDK(KP831469612), T5P));
1492 			      T60 = VFMA(LDK(KP555570233), T5R, VMUL(LDK(KP831469612), T5S));
1493 			      T61 = VADD(T5Z, T60);
1494 			      T69 = VSUB(T60, T5Z);
1495 			 }
1496 			 {
1497 			      V T5Q, T5T, T5W, T5X;
1498 			      T5Q = VFMA(LDK(KP831469612), T5O, VMUL(LDK(KP555570233), T5P));
1499 			      T5T = VFNMS(LDK(KP555570233), T5S, VMUL(LDK(KP831469612), T5R));
1500 			      T5U = VADD(T5Q, T5T);
1501 			      T65 = VSUB(T5T, T5Q);
1502 			      T5W = VSUB(T5r, T5q);
1503 			      T5X = VSUB(T4L, T4E);
1504 			      T5Y = VADD(T5W, T5X);
1505 			      T66 = VSUB(T5X, T5W);
1506 			 }
1507 			 {
1508 			      V T5V, T62, T6b, T6c;
1509 			      T5V = VADD(T5N, T5U);
1510 			      T62 = VBYI(VADD(T5Y, T61));
1511 			      T7D = VSUB(T5V, T62);
1512 			      STM2(&(xo[116]), T7D, ovs, &(xo[0]));
1513 			      T7E = VADD(T5V, T62);
1514 			      STM2(&(xo[12]), T7E, ovs, &(xo[0]));
1515 			      T6b = VBYI(VADD(T66, T65));
1516 			      T6c = VADD(T68, T69);
1517 			      T7F = VADD(T6b, T6c);
1518 			      STM2(&(xo[20]), T7F, ovs, &(xo[0]));
1519 			      T7G = VSUB(T6c, T6b);
1520 			      STM2(&(xo[108]), T7G, ovs, &(xo[0]));
1521 			 }
1522 			 {
1523 			      V T63, T64, T67, T6a;
1524 			      T63 = VSUB(T5N, T5U);
1525 			      T64 = VBYI(VSUB(T61, T5Y));
1526 			      T7H = VSUB(T63, T64);
1527 			      STM2(&(xo[76]), T7H, ovs, &(xo[0]));
1528 			      T7I = VADD(T63, T64);
1529 			      STM2(&(xo[52]), T7I, ovs, &(xo[0]));
1530 			      T67 = VBYI(VSUB(T65, T66));
1531 			      T6a = VSUB(T68, T69);
1532 			      T7J = VADD(T67, T6a);
1533 			      STM2(&(xo[44]), T7J, ovs, &(xo[0]));
1534 			      T7K = VSUB(T6a, T67);
1535 			      STM2(&(xo[84]), T7K, ovs, &(xo[0]));
1536 			 }
1537 		    }
1538 		    {
1539 			 V T7U, T7W, T7X, T7Z;
1540 			 {
1541 			      V T11, T2C, T2v, T2D, T2e, T2z, T2s, T2A;
1542 			      {
1543 				   V Tr, T10, T2t, T2u;
1544 				   Tr = VSUB(Tb, Tq);
1545 				   T10 = VSUB(TI, TZ);
1546 				   T11 = VADD(Tr, T10);
1547 				   T2C = VSUB(Tr, T10);
1548 				   T2t = VFNMS(LDK(KP634393284), T1B, VMUL(LDK(KP773010453), T1s));
1549 				   T2u = VFMA(LDK(KP773010453), T2c, VMUL(LDK(KP634393284), T23));
1550 				   T2v = VADD(T2t, T2u);
1551 				   T2D = VSUB(T2u, T2t);
1552 			      }
1553 			      {
1554 				   V T1C, T2d, T2o, T2r;
1555 				   T1C = VFMA(LDK(KP634393284), T1s, VMUL(LDK(KP773010453), T1B));
1556 				   T2d = VFNMS(LDK(KP634393284), T2c, VMUL(LDK(KP773010453), T23));
1557 				   T2e = VADD(T1C, T2d);
1558 				   T2z = VSUB(T2d, T1C);
1559 				   T2o = VSUB(T2i, T2n);
1560 				   T2r = VSUB(T2p, T2q);
1561 				   T2s = VADD(T2o, T2r);
1562 				   T2A = VSUB(T2r, T2o);
1563 			      }
1564 			      {
1565 				   V T2f, T2w, T7L, T7M;
1566 				   T2f = VADD(T11, T2e);
1567 				   T2w = VBYI(VADD(T2s, T2v));
1568 				   T7L = VSUB(T2f, T2w);
1569 				   STM2(&(xo[114]), T7L, ovs, &(xo[2]));
1570 				   STN2(&(xo[112]), T7r, T7L, ovs);
1571 				   T7M = VADD(T2f, T2w);
1572 				   STM2(&(xo[14]), T7M, ovs, &(xo[2]));
1573 				   STN2(&(xo[12]), T7E, T7M, ovs);
1574 			      }
1575 			      {
1576 				   V T2F, T2G, T7N, T7O;
1577 				   T2F = VBYI(VADD(T2A, T2z));
1578 				   T2G = VADD(T2C, T2D);
1579 				   T7N = VADD(T2F, T2G);
1580 				   STM2(&(xo[18]), T7N, ovs, &(xo[2]));
1581 				   STN2(&(xo[16]), T7s, T7N, ovs);
1582 				   T7O = VSUB(T2G, T2F);
1583 				   STM2(&(xo[110]), T7O, ovs, &(xo[2]));
1584 				   STN2(&(xo[108]), T7G, T7O, ovs);
1585 			      }
1586 			      {
1587 				   V T2x, T2y, T7P, T7Q;
1588 				   T2x = VSUB(T11, T2e);
1589 				   T2y = VBYI(VSUB(T2v, T2s));
1590 				   T7P = VSUB(T2x, T2y);
1591 				   STM2(&(xo[78]), T7P, ovs, &(xo[2]));
1592 				   STN2(&(xo[76]), T7H, T7P, ovs);
1593 				   T7Q = VADD(T2x, T2y);
1594 				   STM2(&(xo[50]), T7Q, ovs, &(xo[2]));
1595 				   STN2(&(xo[48]), T7u, T7Q, ovs);
1596 			      }
1597 			      {
1598 				   V T2B, T2E, T7R, T7S;
1599 				   T2B = VBYI(VSUB(T2z, T2A));
1600 				   T2E = VSUB(T2C, T2D);
1601 				   T7R = VADD(T2B, T2E);
1602 				   STM2(&(xo[46]), T7R, ovs, &(xo[2]));
1603 				   STN2(&(xo[44]), T7J, T7R, ovs);
1604 				   T7S = VSUB(T2E, T2B);
1605 				   STM2(&(xo[82]), T7S, ovs, &(xo[2]));
1606 				   STN2(&(xo[80]), T7t, T7S, ovs);
1607 			      }
1608 			 }
1609 			 {
1610 			      V T3j, T3Q, T3J, T3R, T3y, T3N, T3G, T3O;
1611 			      {
1612 				   V T3b, T3i, T3H, T3I;
1613 				   T3b = VADD(T39, T3a);
1614 				   T3i = VADD(T3e, T3h);
1615 				   T3j = VADD(T3b, T3i);
1616 				   T3Q = VSUB(T3b, T3i);
1617 				   T3H = VFNMS(LDK(KP290284677), T3m, VMUL(LDK(KP956940335), T3p));
1618 				   T3I = VFMA(LDK(KP290284677), T3t, VMUL(LDK(KP956940335), T3w));
1619 				   T3J = VADD(T3H, T3I);
1620 				   T3R = VSUB(T3I, T3H);
1621 			      }
1622 			      {
1623 				   V T3q, T3x, T3C, T3F;
1624 				   T3q = VFMA(LDK(KP956940335), T3m, VMUL(LDK(KP290284677), T3p));
1625 				   T3x = VFNMS(LDK(KP290284677), T3w, VMUL(LDK(KP956940335), T3t));
1626 				   T3y = VADD(T3q, T3x);
1627 				   T3N = VSUB(T3x, T3q);
1628 				   T3C = VADD(T3A, T3B);
1629 				   T3F = VADD(T3D, T3E);
1630 				   T3G = VADD(T3C, T3F);
1631 				   T3O = VSUB(T3F, T3C);
1632 			      }
1633 			      {
1634 				   V T3z, T3K, T7T, T3T, T3U, T7V;
1635 				   T3z = VADD(T3j, T3y);
1636 				   T3K = VBYI(VADD(T3G, T3J));
1637 				   T7T = VSUB(T3z, T3K);
1638 				   STM2(&(xo[122]), T7T, ovs, &(xo[2]));
1639 				   STN2(&(xo[120]), T7v, T7T, ovs);
1640 				   T7U = VADD(T3z, T3K);
1641 				   STM2(&(xo[6]), T7U, ovs, &(xo[2]));
1642 				   T3T = VBYI(VADD(T3O, T3N));
1643 				   T3U = VADD(T3Q, T3R);
1644 				   T7V = VADD(T3T, T3U);
1645 				   STM2(&(xo[26]), T7V, ovs, &(xo[2]));
1646 				   STN2(&(xo[24]), T7x, T7V, ovs);
1647 				   T7W = VSUB(T3U, T3T);
1648 				   STM2(&(xo[102]), T7W, ovs, &(xo[2]));
1649 			      }
1650 			      {
1651 				   V T3L, T3M, T7Y, T3P, T3S, T80;
1652 				   T3L = VSUB(T3j, T3y);
1653 				   T3M = VBYI(VSUB(T3J, T3G));
1654 				   T7X = VSUB(T3L, T3M);
1655 				   STM2(&(xo[70]), T7X, ovs, &(xo[2]));
1656 				   T7Y = VADD(T3L, T3M);
1657 				   STM2(&(xo[58]), T7Y, ovs, &(xo[2]));
1658 				   STN2(&(xo[56]), T7A, T7Y, ovs);
1659 				   T3P = VBYI(VSUB(T3N, T3O));
1660 				   T3S = VSUB(T3Q, T3R);
1661 				   T7Z = VADD(T3P, T3S);
1662 				   STM2(&(xo[38]), T7Z, ovs, &(xo[2]));
1663 				   T80 = VSUB(T3S, T3P);
1664 				   STM2(&(xo[90]), T80, ovs, &(xo[2]));
1665 				   STN2(&(xo[88]), T7C, T80, ovs);
1666 			      }
1667 			 }
1668 			 {
1669 			      V T81, T83, T86, T88;
1670 			      {
1671 				   V T4N, T5G, T5z, T5H, T5m, T5D, T5w, T5E;
1672 				   {
1673 					V T4x, T4M, T5x, T5y;
1674 					T4x = VADD(T4p, T4w);
1675 					T4M = VADD(T4E, T4L);
1676 					T4N = VADD(T4x, T4M);
1677 					T5G = VSUB(T4x, T4M);
1678 					T5x = VFNMS(LDK(KP195090322), T4Y, VMUL(LDK(KP980785280), T53));
1679 					T5y = VFMA(LDK(KP195090322), T5f, VMUL(LDK(KP980785280), T5k));
1680 					T5z = VADD(T5x, T5y);
1681 					T5H = VSUB(T5y, T5x);
1682 				   }
1683 				   {
1684 					V T54, T5l, T5s, T5v;
1685 					T54 = VFMA(LDK(KP980785280), T4Y, VMUL(LDK(KP195090322), T53));
1686 					T5l = VFNMS(LDK(KP195090322), T5k, VMUL(LDK(KP980785280), T5f));
1687 					T5m = VADD(T54, T5l);
1688 					T5D = VSUB(T5l, T54);
1689 					T5s = VADD(T5q, T5r);
1690 					T5v = VADD(T5t, T5u);
1691 					T5w = VADD(T5s, T5v);
1692 					T5E = VSUB(T5v, T5s);
1693 				   }
1694 				   {
1695 					V T5n, T5A, T82, T5J, T5K, T84;
1696 					T5n = VADD(T4N, T5m);
1697 					T5A = VBYI(VADD(T5w, T5z));
1698 					T81 = VSUB(T5n, T5A);
1699 					STM2(&(xo[124]), T81, ovs, &(xo[0]));
1700 					T82 = VADD(T5n, T5A);
1701 					STM2(&(xo[4]), T82, ovs, &(xo[0]));
1702 					STN2(&(xo[4]), T82, T7U, ovs);
1703 					T5J = VBYI(VADD(T5E, T5D));
1704 					T5K = VADD(T5G, T5H);
1705 					T83 = VADD(T5J, T5K);
1706 					STM2(&(xo[28]), T83, ovs, &(xo[0]));
1707 					T84 = VSUB(T5K, T5J);
1708 					STM2(&(xo[100]), T84, ovs, &(xo[0]));
1709 					STN2(&(xo[100]), T84, T7W, ovs);
1710 				   }
1711 				   {
1712 					V T5B, T5C, T85, T5F, T5I, T87;
1713 					T5B = VSUB(T4N, T5m);
1714 					T5C = VBYI(VSUB(T5z, T5w));
1715 					T85 = VSUB(T5B, T5C);
1716 					STM2(&(xo[68]), T85, ovs, &(xo[0]));
1717 					STN2(&(xo[68]), T85, T7X, ovs);
1718 					T86 = VADD(T5B, T5C);
1719 					STM2(&(xo[60]), T86, ovs, &(xo[0]));
1720 					T5F = VBYI(VSUB(T5D, T5E));
1721 					T5I = VSUB(T5G, T5H);
1722 					T87 = VADD(T5F, T5I);
1723 					STM2(&(xo[36]), T87, ovs, &(xo[0]));
1724 					STN2(&(xo[36]), T87, T7Z, ovs);
1725 					T88 = VSUB(T5I, T5F);
1726 					STM2(&(xo[92]), T88, ovs, &(xo[0]));
1727 				   }
1728 			      }
1729 			      {
1730 				   V T2J, T34, T2X, T35, T2Q, T31, T2U, T32;
1731 				   {
1732 					V T2H, T2I, T2V, T2W;
1733 					T2H = VADD(Tb, Tq);
1734 					T2I = VADD(T2q, T2p);
1735 					T2J = VADD(T2H, T2I);
1736 					T34 = VSUB(T2H, T2I);
1737 					T2V = VFNMS(LDK(KP098017140), T2L, VMUL(LDK(KP995184726), T2K));
1738 					T2W = VFMA(LDK(KP995184726), T2O, VMUL(LDK(KP098017140), T2N));
1739 					T2X = VADD(T2V, T2W);
1740 					T35 = VSUB(T2W, T2V);
1741 				   }
1742 				   {
1743 					V T2M, T2P, T2S, T2T;
1744 					T2M = VFMA(LDK(KP098017140), T2K, VMUL(LDK(KP995184726), T2L));
1745 					T2P = VFNMS(LDK(KP098017140), T2O, VMUL(LDK(KP995184726), T2N));
1746 					T2Q = VADD(T2M, T2P);
1747 					T31 = VSUB(T2P, T2M);
1748 					T2S = VADD(T2n, T2i);
1749 					T2T = VADD(TZ, TI);
1750 					T2U = VADD(T2S, T2T);
1751 					T32 = VSUB(T2T, T2S);
1752 				   }
1753 				   {
1754 					V T2R, T2Y, T89, T8a;
1755 					T2R = VADD(T2J, T2Q);
1756 					T2Y = VBYI(VADD(T2U, T2X));
1757 					T89 = VSUB(T2R, T2Y);
1758 					STM2(&(xo[126]), T89, ovs, &(xo[2]));
1759 					STN2(&(xo[124]), T81, T89, ovs);
1760 					T8a = VADD(T2R, T2Y);
1761 					STM2(&(xo[2]), T8a, ovs, &(xo[2]));
1762 					STN2(&(xo[0]), T7p, T8a, ovs);
1763 				   }
1764 				   {
1765 					V T37, T38, T8b, T8c;
1766 					T37 = VBYI(VADD(T32, T31));
1767 					T38 = VADD(T34, T35);
1768 					T8b = VADD(T37, T38);
1769 					STM2(&(xo[30]), T8b, ovs, &(xo[2]));
1770 					STN2(&(xo[28]), T83, T8b, ovs);
1771 					T8c = VSUB(T38, T37);
1772 					STM2(&(xo[98]), T8c, ovs, &(xo[2]));
1773 					STN2(&(xo[96]), T7q, T8c, ovs);
1774 				   }
1775 				   {
1776 					V T2Z, T30, T8d, T8e;
1777 					T2Z = VSUB(T2J, T2Q);
1778 					T30 = VBYI(VSUB(T2X, T2U));
1779 					T8d = VSUB(T2Z, T30);
1780 					STM2(&(xo[66]), T8d, ovs, &(xo[2]));
1781 					STN2(&(xo[64]), T7n, T8d, ovs);
1782 					T8e = VADD(T2Z, T30);
1783 					STM2(&(xo[62]), T8e, ovs, &(xo[2]));
1784 					STN2(&(xo[60]), T86, T8e, ovs);
1785 				   }
1786 				   {
1787 					V T33, T36, T8f, T8g;
1788 					T33 = VBYI(VSUB(T31, T32));
1789 					T36 = VSUB(T34, T35);
1790 					T8f = VADD(T33, T36);
1791 					STM2(&(xo[34]), T8f, ovs, &(xo[2]));
1792 					STN2(&(xo[32]), T7o, T8f, ovs);
1793 					T8g = VSUB(T36, T33);
1794 					STM2(&(xo[94]), T8g, ovs, &(xo[2]));
1795 					STN2(&(xo[92]), T88, T8g, ovs);
1796 				   }
1797 			      }
1798 			      {
1799 				   V T3X, T4i, T4b, T4j, T44, T4f, T48, T4g;
1800 				   {
1801 					V T3V, T3W, T49, T4a;
1802 					T3V = VSUB(T39, T3a);
1803 					T3W = VSUB(T3E, T3D);
1804 					T3X = VADD(T3V, T3W);
1805 					T4i = VSUB(T3V, T3W);
1806 					T49 = VFNMS(LDK(KP471396736), T3Y, VMUL(LDK(KP881921264), T3Z));
1807 					T4a = VFMA(LDK(KP471396736), T41, VMUL(LDK(KP881921264), T42));
1808 					T4b = VADD(T49, T4a);
1809 					T4j = VSUB(T4a, T49);
1810 				   }
1811 				   {
1812 					V T40, T43, T46, T47;
1813 					T40 = VFMA(LDK(KP881921264), T3Y, VMUL(LDK(KP471396736), T3Z));
1814 					T43 = VFNMS(LDK(KP471396736), T42, VMUL(LDK(KP881921264), T41));
1815 					T44 = VADD(T40, T43);
1816 					T4f = VSUB(T43, T40);
1817 					T46 = VSUB(T3B, T3A);
1818 					T47 = VSUB(T3h, T3e);
1819 					T48 = VADD(T46, T47);
1820 					T4g = VSUB(T47, T46);
1821 				   }
1822 				   {
1823 					V T45, T4c, T8h, T8i;
1824 					T45 = VADD(T3X, T44);
1825 					T4c = VBYI(VADD(T48, T4b));
1826 					T8h = VSUB(T45, T4c);
1827 					STM2(&(xo[118]), T8h, ovs, &(xo[2]));
1828 					STN2(&(xo[116]), T7D, T8h, ovs);
1829 					T8i = VADD(T45, T4c);
1830 					STM2(&(xo[10]), T8i, ovs, &(xo[2]));
1831 					STN2(&(xo[8]), T7w, T8i, ovs);
1832 				   }
1833 				   {
1834 					V T4l, T4m, T8j, T8k;
1835 					T4l = VBYI(VADD(T4g, T4f));
1836 					T4m = VADD(T4i, T4j);
1837 					T8j = VADD(T4l, T4m);
1838 					STM2(&(xo[22]), T8j, ovs, &(xo[2]));
1839 					STN2(&(xo[20]), T7F, T8j, ovs);
1840 					T8k = VSUB(T4m, T4l);
1841 					STM2(&(xo[106]), T8k, ovs, &(xo[2]));
1842 					STN2(&(xo[104]), T7y, T8k, ovs);
1843 				   }
1844 				   {
1845 					V T4d, T4e, T8l, T8m;
1846 					T4d = VSUB(T3X, T44);
1847 					T4e = VBYI(VSUB(T4b, T48));
1848 					T8l = VSUB(T4d, T4e);
1849 					STM2(&(xo[74]), T8l, ovs, &(xo[2]));
1850 					STN2(&(xo[72]), T7z, T8l, ovs);
1851 					T8m = VADD(T4d, T4e);
1852 					STM2(&(xo[54]), T8m, ovs, &(xo[2]));
1853 					STN2(&(xo[52]), T7I, T8m, ovs);
1854 				   }
1855 				   {
1856 					V T4h, T4k, T8n, T8o;
1857 					T4h = VBYI(VSUB(T4f, T4g));
1858 					T4k = VSUB(T4i, T4j);
1859 					T8n = VADD(T4h, T4k);
1860 					STM2(&(xo[42]), T8n, ovs, &(xo[2]));
1861 					STN2(&(xo[40]), T7B, T8n, ovs);
1862 					T8o = VSUB(T4k, T4h);
1863 					STM2(&(xo[86]), T8o, ovs, &(xo[2]));
1864 					STN2(&(xo[84]), T7K, T8o, ovs);
1865 				   }
1866 			      }
1867 			 }
1868 		    }
1869 	       }
1870 	  }
1871      }
1872      VLEAVE();
1873 }
1874 
1875 static const kdft_desc desc = { 64, XSIMD_STRING("n2fv_64"), { 404, 72, 52, 0 }, &GENUS, 0, 2, 0, 0 };
1876 
XSIMD(codelet_n2fv_64)1877 void XSIMD(codelet_n2fv_64) (planner *p) { X(kdft_register) (p, n2fv_64, &desc);
1878 }
1879 
1880 #endif
1881