comparison src/fftw-3.3.8/dft/simd/common/n1fv_64.c @ 167:bd3cc4d1df30

Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam <cannam@all-day-breakfast.com>
date Tue, 19 Nov 2019 14:52:55 +0000
parents
children
comparison
equal deleted inserted replaced
166:cbd6d7e562c7 167:bd3cc4d1df30
1 /*
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 */
20
21 /* This file was automatically generated --- DO NOT EDIT */
22 /* Generated on Thu May 24 08:04:52 EDT 2018 */
23
24 #include "dft/codelet-dft.h"
25
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27
28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n1fv_64 -include dft/simd/n1f.h */
29
30 /*
31 * This function contains 456 FP additions, 258 FP multiplications,
32 * (or, 198 additions, 0 multiplications, 258 fused multiply/add),
33 * 108 stack variables, 15 constants, and 128 memory accesses
34 */
35 #include "dft/simd/n1f.h"
36
37 static void n1fv_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 T3t, T3x, T3w, T3y;
459 {
460 V T3r, T3s, T3u, T3v;
461 T3r = VADD(T2V, T2Y);
462 T3s = VADD(T3l, T3i);
463 T3t = VADD(T3r, T3s);
464 T3x = VSUB(T3r, T3s);
465 T3u = VADD(T32, T35);
466 T3v = VADD(T39, T3c);
467 T3w = VADD(T3u, T3v);
468 T3y = VSUB(T3v, T3u);
469 }
470 ST(&(xo[WS(os, 32)]), VSUB(T3t, T3w), ovs, &(xo[0]));
471 ST(&(xo[WS(os, 16)]), VFMAI(T3y, T3x), ovs, &(xo[0]));
472 ST(&(xo[0]), VADD(T3t, T3w), ovs, &(xo[0]));
473 ST(&(xo[WS(os, 48)]), VFNMSI(T3y, T3x), ovs, &(xo[0]));
474 }
475 {
476 V T2Z, T3m, T3e, T3n, T36, T3d;
477 T2Z = VSUB(T2V, T2Y);
478 T3m = VSUB(T3i, T3l);
479 T36 = VSUB(T32, T35);
480 T3d = VSUB(T39, T3c);
481 T3e = VADD(T36, T3d);
482 T3n = VSUB(T3d, T36);
483 {
484 V T3f, T3o, T3p, T3q;
485 T3f = VFNMS(LDK(KP707106781), T3e, T2Z);
486 T3o = VFNMS(LDK(KP707106781), T3n, T3m);
487 ST(&(xo[WS(os, 24)]), VFNMSI(T3o, T3f), ovs, &(xo[0]));
488 ST(&(xo[WS(os, 40)]), VFMAI(T3o, T3f), ovs, &(xo[0]));
489 T3p = VFMA(LDK(KP707106781), T3e, T2Z);
490 T3q = VFMA(LDK(KP707106781), T3n, T3m);
491 ST(&(xo[WS(os, 56)]), VFNMSI(T3q, T3p), ovs, &(xo[0]));
492 ST(&(xo[WS(os, 8)]), VFMAI(T3q, T3p), ovs, &(xo[0]));
493 }
494 }
495 {
496 V T3D, T3V, T3O, T3Y, T3K, T3Z, T3R, T3W, T3C, T3N;
497 T3C = VADD(T3A, T3B);
498 T3D = VFMA(LDK(KP707106781), T3C, T3z);
499 T3V = VFNMS(LDK(KP707106781), T3C, T3z);
500 T3N = VSUB(T3B, T3A);
501 T3O = VFMA(LDK(KP707106781), T3N, T3M);
502 T3Y = VFNMS(LDK(KP707106781), T3N, T3M);
503 {
504 V T3G, T3J, T3P, T3Q;
505 T3G = VFNMS(LDK(KP414213562), T3F, T3E);
506 T3J = VFNMS(LDK(KP414213562), T3I, T3H);
507 T3K = VADD(T3G, T3J);
508 T3Z = VSUB(T3J, T3G);
509 T3P = VFMA(LDK(KP414213562), T3H, T3I);
510 T3Q = VFMA(LDK(KP414213562), T3E, T3F);
511 T3R = VSUB(T3P, T3Q);
512 T3W = VADD(T3Q, T3P);
513 }
514 {
515 V T3L, T3S, T41, T42;
516 T3L = VFNMS(LDK(KP923879532), T3K, T3D);
517 T3S = VFNMS(LDK(KP923879532), T3R, T3O);
518 ST(&(xo[WS(os, 28)]), VFNMSI(T3S, T3L), ovs, &(xo[0]));
519 ST(&(xo[WS(os, 36)]), VFMAI(T3S, T3L), ovs, &(xo[0]));
520 T41 = VFMA(LDK(KP923879532), T3W, T3V);
521 T42 = VFNMS(LDK(KP923879532), T3Z, T3Y);
522 ST(&(xo[WS(os, 12)]), VFNMSI(T42, T41), ovs, &(xo[0]));
523 ST(&(xo[WS(os, 52)]), VFMAI(T42, T41), ovs, &(xo[0]));
524 }
525 {
526 V T3T, T3U, T3X, T40;
527 T3T = VFMA(LDK(KP923879532), T3K, T3D);
528 T3U = VFMA(LDK(KP923879532), T3R, T3O);
529 ST(&(xo[WS(os, 60)]), VFNMSI(T3U, T3T), ovs, &(xo[0]));
530 ST(&(xo[WS(os, 4)]), VFMAI(T3U, T3T), ovs, &(xo[0]));
531 T3X = VFNMS(LDK(KP923879532), T3W, T3V);
532 T40 = VFMA(LDK(KP923879532), T3Z, T3Y);
533 ST(&(xo[WS(os, 20)]), VFMAI(T40, T3X), ovs, &(xo[0]));
534 ST(&(xo[WS(os, 44)]), VFNMSI(T40, T3X), ovs, &(xo[0]));
535 }
536 }
537 {
538 V T6X, T7f, T7b, T7g, T74, T7j, T78, T7i;
539 {
540 V T6V, T6W, T79, T7a;
541 T6V = VFMA(LDK(KP923879532), T6a, T69);
542 T6W = VADD(T6E, T6D);
543 T6X = VFMA(LDK(KP831469612), T6W, T6V);
544 T7f = VFNMS(LDK(KP831469612), T6W, T6V);
545 T79 = VFNMS(LDK(KP303346683), T6Y, T6Z);
546 T7a = VFNMS(LDK(KP303346683), T71, T72);
547 T7b = VSUB(T79, T7a);
548 T7g = VADD(T79, T7a);
549 }
550 {
551 V T70, T73, T76, T77;
552 T70 = VFMA(LDK(KP303346683), T6Z, T6Y);
553 T73 = VFMA(LDK(KP303346683), T72, T71);
554 T74 = VADD(T70, T73);
555 T7j = VSUB(T73, T70);
556 T76 = VFMA(LDK(KP923879532), T6B, T6A);
557 T77 = VSUB(T6e, T6h);
558 T78 = VFMA(LDK(KP831469612), T77, T76);
559 T7i = VFNMS(LDK(KP831469612), T77, T76);
560 }
561 {
562 V T75, T7c, T7l, T7m;
563 T75 = VFNMS(LDK(KP956940335), T74, T6X);
564 T7c = VFNMS(LDK(KP956940335), T7b, T78);
565 ST(&(xo[WS(os, 29)]), VFNMSI(T7c, T75), ovs, &(xo[WS(os, 1)]));
566 ST(&(xo[WS(os, 35)]), VFMAI(T7c, T75), ovs, &(xo[WS(os, 1)]));
567 T7l = VFNMS(LDK(KP956940335), T7g, T7f);
568 T7m = VFNMS(LDK(KP956940335), T7j, T7i);
569 ST(&(xo[WS(os, 13)]), VFNMSI(T7m, T7l), ovs, &(xo[WS(os, 1)]));
570 ST(&(xo[WS(os, 51)]), VFMAI(T7m, T7l), ovs, &(xo[WS(os, 1)]));
571 }
572 {
573 V T7d, T7e, T7h, T7k;
574 T7d = VFMA(LDK(KP956940335), T74, T6X);
575 T7e = VFMA(LDK(KP956940335), T7b, T78);
576 ST(&(xo[WS(os, 61)]), VFNMSI(T7e, T7d), ovs, &(xo[WS(os, 1)]));
577 ST(&(xo[WS(os, 3)]), VFMAI(T7e, T7d), ovs, &(xo[WS(os, 1)]));
578 T7h = VFMA(LDK(KP956940335), T7g, T7f);
579 T7k = VFMA(LDK(KP956940335), T7j, T7i);
580 ST(&(xo[WS(os, 19)]), VFMAI(T7k, T7h), ovs, &(xo[WS(os, 1)]));
581 ST(&(xo[WS(os, 45)]), VFNMSI(T7k, T7h), ovs, &(xo[WS(os, 1)]));
582 }
583 }
584 {
585 V TT, T2j, T2f, T2k, T1Y, T2n, T2c, T2m;
586 {
587 V Tn, TS, T2d, T2e;
588 Tn = VFMA(LDK(KP707106781), Tm, T7);
589 TS = VADD(TC, TR);
590 TT = VFMA(LDK(KP923879532), TS, Tn);
591 T2j = VFNMS(LDK(KP923879532), TS, Tn);
592 T2d = VFMA(LDK(KP198912367), T1N, T1W);
593 T2e = VFMA(LDK(KP198912367), T1g, T1p);
594 T2f = VSUB(T2d, T2e);
595 T2k = VADD(T2e, T2d);
596 }
597 {
598 V T1q, T1X, T28, T2b;
599 T1q = VFNMS(LDK(KP198912367), T1p, T1g);
600 T1X = VFNMS(LDK(KP198912367), T1W, T1N);
601 T1Y = VADD(T1q, T1X);
602 T2n = VSUB(T1X, T1q);
603 T28 = VFMA(LDK(KP707106781), T27, T26);
604 T2b = VSUB(T29, T2a);
605 T2c = VFMA(LDK(KP923879532), T2b, T28);
606 T2m = VFNMS(LDK(KP923879532), T2b, T28);
607 }
608 {
609 V T1Z, T2g, T2p, T2q;
610 T1Z = VFNMS(LDK(KP980785280), T1Y, TT);
611 T2g = VFNMS(LDK(KP980785280), T2f, T2c);
612 ST(&(xo[WS(os, 30)]), VFNMSI(T2g, T1Z), ovs, &(xo[0]));
613 ST(&(xo[WS(os, 34)]), VFMAI(T2g, T1Z), ovs, &(xo[0]));
614 T2p = VFMA(LDK(KP980785280), T2k, T2j);
615 T2q = VFNMS(LDK(KP980785280), T2n, T2m);
616 ST(&(xo[WS(os, 14)]), VFNMSI(T2q, T2p), ovs, &(xo[0]));
617 ST(&(xo[WS(os, 50)]), VFMAI(T2q, T2p), ovs, &(xo[0]));
618 }
619 {
620 V T2h, T2i, T2l, T2o;
621 T2h = VFMA(LDK(KP980785280), T1Y, TT);
622 T2i = VFMA(LDK(KP980785280), T2f, T2c);
623 ST(&(xo[WS(os, 62)]), VFNMSI(T2i, T2h), ovs, &(xo[0]));
624 ST(&(xo[WS(os, 2)]), VFMAI(T2i, T2h), ovs, &(xo[0]));
625 T2l = VFNMS(LDK(KP980785280), T2k, T2j);
626 T2o = VFMA(LDK(KP980785280), T2n, T2m);
627 ST(&(xo[WS(os, 18)]), VFMAI(T2o, T2l), ovs, &(xo[0]));
628 ST(&(xo[WS(os, 46)]), VFNMSI(T2o, T2l), ovs, &(xo[0]));
629 }
630 }
631 {
632 V T4z, T5z, T5v, T5A, T5g, T5D, T5s, T5C;
633 {
634 V T4f, T4y, T5t, T5u;
635 T4f = VFMA(LDK(KP923879532), T4e, T47);
636 T4y = VADD(T4o, T4x);
637 T4z = VFMA(LDK(KP980785280), T4y, T4f);
638 T5z = VFNMS(LDK(KP980785280), T4y, T4f);
639 T5t = VFMA(LDK(KP098491403), T4M, T4T);
640 T5u = VFMA(LDK(KP098491403), T57, T5e);
641 T5v = VSUB(T5t, T5u);
642 T5A = VADD(T5t, T5u);
643 }
644 {
645 V T4U, T5f, T5o, T5r;
646 T4U = VFNMS(LDK(KP098491403), T4T, T4M);
647 T5f = VFNMS(LDK(KP098491403), T5e, T57);
648 T5g = VADD(T4U, T5f);
649 T5D = VSUB(T5f, T4U);
650 T5o = VFMA(LDK(KP923879532), T5n, T5k);
651 T5r = VSUB(T5p, T5q);
652 T5s = VFMA(LDK(KP980785280), T5r, T5o);
653 T5C = VFNMS(LDK(KP980785280), T5r, T5o);
654 }
655 {
656 V T5h, T5w, T5F, T5G;
657 T5h = VFNMS(LDK(KP995184726), T5g, T4z);
658 T5w = VFNMS(LDK(KP995184726), T5v, T5s);
659 ST(&(xo[WS(os, 33)]), VFNMSI(T5w, T5h), ovs, &(xo[WS(os, 1)]));
660 ST(&(xo[WS(os, 31)]), VFMAI(T5w, T5h), ovs, &(xo[WS(os, 1)]));
661 T5F = VFMA(LDK(KP995184726), T5A, T5z);
662 T5G = VFMA(LDK(KP995184726), T5D, T5C);
663 ST(&(xo[WS(os, 15)]), VFMAI(T5G, T5F), ovs, &(xo[WS(os, 1)]));
664 ST(&(xo[WS(os, 49)]), VFNMSI(T5G, T5F), ovs, &(xo[WS(os, 1)]));
665 }
666 {
667 V T5x, T5y, T5B, T5E;
668 T5x = VFMA(LDK(KP995184726), T5g, T4z);
669 T5y = VFMA(LDK(KP995184726), T5v, T5s);
670 ST(&(xo[WS(os, 1)]), VFNMSI(T5y, T5x), ovs, &(xo[WS(os, 1)]));
671 ST(&(xo[WS(os, 63)]), VFMAI(T5y, T5x), ovs, &(xo[WS(os, 1)]));
672 T5B = VFNMS(LDK(KP995184726), T5A, T5z);
673 T5E = VFNMS(LDK(KP995184726), T5D, T5C);
674 ST(&(xo[WS(os, 17)]), VFNMSI(T5E, T5B), ovs, &(xo[WS(os, 1)]));
675 ST(&(xo[WS(os, 47)]), VFMAI(T5E, T5B), ovs, &(xo[WS(os, 1)]));
676 }
677 }
678 {
679 V T6j, T6N, T6J, T6O, T6y, T6R, T6G, T6Q;
680 {
681 V T6b, T6i, T6H, T6I;
682 T6b = VFNMS(LDK(KP923879532), T6a, T69);
683 T6i = VADD(T6e, T6h);
684 T6j = VFNMS(LDK(KP831469612), T6i, T6b);
685 T6N = VFMA(LDK(KP831469612), T6i, T6b);
686 T6H = VFMA(LDK(KP534511135), T6m, T6p);
687 T6I = VFMA(LDK(KP534511135), T6t, T6w);
688 T6J = VSUB(T6H, T6I);
689 T6O = VADD(T6H, T6I);
690 }
691 {
692 V T6q, T6x, T6C, T6F;
693 T6q = VFNMS(LDK(KP534511135), T6p, T6m);
694 T6x = VFNMS(LDK(KP534511135), T6w, T6t);
695 T6y = VADD(T6q, T6x);
696 T6R = VSUB(T6x, T6q);
697 T6C = VFNMS(LDK(KP923879532), T6B, T6A);
698 T6F = VSUB(T6D, T6E);
699 T6G = VFNMS(LDK(KP831469612), T6F, T6C);
700 T6Q = VFMA(LDK(KP831469612), T6F, T6C);
701 }
702 {
703 V T6z, T6K, T6T, T6U;
704 T6z = VFNMS(LDK(KP881921264), T6y, T6j);
705 T6K = VFNMS(LDK(KP881921264), T6J, T6G);
706 ST(&(xo[WS(os, 37)]), VFNMSI(T6K, T6z), ovs, &(xo[WS(os, 1)]));
707 ST(&(xo[WS(os, 27)]), VFMAI(T6K, T6z), ovs, &(xo[WS(os, 1)]));
708 T6T = VFMA(LDK(KP881921264), T6O, T6N);
709 T6U = VFMA(LDK(KP881921264), T6R, T6Q);
710 ST(&(xo[WS(os, 11)]), VFMAI(T6U, T6T), ovs, &(xo[WS(os, 1)]));
711 ST(&(xo[WS(os, 53)]), VFNMSI(T6U, T6T), ovs, &(xo[WS(os, 1)]));
712 }
713 {
714 V T6L, T6M, T6P, T6S;
715 T6L = VFMA(LDK(KP881921264), T6y, T6j);
716 T6M = VFMA(LDK(KP881921264), T6J, T6G);
717 ST(&(xo[WS(os, 5)]), VFNMSI(T6M, T6L), ovs, &(xo[WS(os, 1)]));
718 ST(&(xo[WS(os, 59)]), VFMAI(T6M, T6L), ovs, &(xo[WS(os, 1)]));
719 T6P = VFNMS(LDK(KP881921264), T6O, T6N);
720 T6S = VFNMS(LDK(KP881921264), T6R, T6Q);
721 ST(&(xo[WS(os, 21)]), VFNMSI(T6S, T6P), ovs, &(xo[WS(os, 1)]));
722 ST(&(xo[WS(os, 43)]), VFMAI(T6S, T6P), ovs, &(xo[WS(os, 1)]));
723 }
724 }
725 {
726 V T2t, T2L, T2H, T2M, T2A, T2P, T2E, T2O;
727 {
728 V T2r, T2s, T2F, T2G;
729 T2r = VFNMS(LDK(KP707106781), Tm, T7);
730 T2s = VADD(T2a, T29);
731 T2t = VFMA(LDK(KP923879532), T2s, T2r);
732 T2L = VFNMS(LDK(KP923879532), T2s, T2r);
733 T2F = VFNMS(LDK(KP668178637), T2x, T2y);
734 T2G = VFNMS(LDK(KP668178637), T2u, T2v);
735 T2H = VSUB(T2F, T2G);
736 T2M = VADD(T2G, T2F);
737 }
738 {
739 V T2w, T2z, T2C, T2D;
740 T2w = VFMA(LDK(KP668178637), T2v, T2u);
741 T2z = VFMA(LDK(KP668178637), T2y, T2x);
742 T2A = VADD(T2w, T2z);
743 T2P = VSUB(T2z, T2w);
744 T2C = VFNMS(LDK(KP707106781), T27, T26);
745 T2D = VSUB(TR, TC);
746 T2E = VFNMS(LDK(KP923879532), T2D, T2C);
747 T2O = VFMA(LDK(KP923879532), T2D, T2C);
748 }
749 {
750 V T2B, T2I, T2R, T2S;
751 T2B = VFNMS(LDK(KP831469612), T2A, T2t);
752 T2I = VFNMS(LDK(KP831469612), T2H, T2E);
753 ST(&(xo[WS(os, 38)]), VFNMSI(T2I, T2B), ovs, &(xo[0]));
754 ST(&(xo[WS(os, 26)]), VFMAI(T2I, T2B), ovs, &(xo[0]));
755 T2R = VFNMS(LDK(KP831469612), T2M, T2L);
756 T2S = VFMA(LDK(KP831469612), T2P, T2O);
757 ST(&(xo[WS(os, 10)]), VFMAI(T2S, T2R), ovs, &(xo[0]));
758 ST(&(xo[WS(os, 54)]), VFNMSI(T2S, T2R), ovs, &(xo[0]));
759 }
760 {
761 V T2J, T2K, T2N, T2Q;
762 T2J = VFMA(LDK(KP831469612), T2A, T2t);
763 T2K = VFMA(LDK(KP831469612), T2H, T2E);
764 ST(&(xo[WS(os, 6)]), VFNMSI(T2K, T2J), ovs, &(xo[0]));
765 ST(&(xo[WS(os, 58)]), VFMAI(T2K, T2J), ovs, &(xo[0]));
766 T2N = VFMA(LDK(KP831469612), T2M, T2L);
767 T2Q = VFNMS(LDK(KP831469612), T2P, T2O);
768 ST(&(xo[WS(os, 22)]), VFNMSI(T2Q, T2N), ovs, &(xo[0]));
769 ST(&(xo[WS(os, 42)]), VFMAI(T2Q, T2N), ovs, &(xo[0]));
770 }
771 }
772 {
773 V T5J, T61, T5X, T62, T5Q, T65, T5U, T64;
774 {
775 V T5H, T5I, T5V, T5W;
776 T5H = VFNMS(LDK(KP923879532), T4e, T47);
777 T5I = VADD(T5p, T5q);
778 T5J = VFMA(LDK(KP980785280), T5I, T5H);
779 T61 = VFNMS(LDK(KP980785280), T5I, T5H);
780 T5V = VFNMS(LDK(KP820678790), T5K, T5L);
781 T5W = VFNMS(LDK(KP820678790), T5N, T5O);
782 T5X = VSUB(T5V, T5W);
783 T62 = VADD(T5V, T5W);
784 }
785 {
786 V T5M, T5P, T5S, T5T;
787 T5M = VFMA(LDK(KP820678790), T5L, T5K);
788 T5P = VFMA(LDK(KP820678790), T5O, T5N);
789 T5Q = VADD(T5M, T5P);
790 T65 = VSUB(T5P, T5M);
791 T5S = VFNMS(LDK(KP923879532), T5n, T5k);
792 T5T = VSUB(T4x, T4o);
793 T5U = VFMA(LDK(KP980785280), T5T, T5S);
794 T64 = VFNMS(LDK(KP980785280), T5T, T5S);
795 }
796 {
797 V T5R, T5Y, T67, T68;
798 T5R = VFNMS(LDK(KP773010453), T5Q, T5J);
799 T5Y = VFNMS(LDK(KP773010453), T5X, T5U);
800 ST(&(xo[WS(os, 25)]), VFNMSI(T5Y, T5R), ovs, &(xo[WS(os, 1)]));
801 ST(&(xo[WS(os, 39)]), VFMAI(T5Y, T5R), ovs, &(xo[WS(os, 1)]));
802 T67 = VFNMS(LDK(KP773010453), T62, T61);
803 T68 = VFNMS(LDK(KP773010453), T65, T64);
804 ST(&(xo[WS(os, 9)]), VFNMSI(T68, T67), ovs, &(xo[WS(os, 1)]));
805 ST(&(xo[WS(os, 55)]), VFMAI(T68, T67), ovs, &(xo[WS(os, 1)]));
806 }
807 {
808 V T5Z, T60, T63, T66;
809 T5Z = VFMA(LDK(KP773010453), T5Q, T5J);
810 T60 = VFMA(LDK(KP773010453), T5X, T5U);
811 ST(&(xo[WS(os, 57)]), VFNMSI(T60, T5Z), ovs, &(xo[WS(os, 1)]));
812 ST(&(xo[WS(os, 7)]), VFMAI(T60, T5Z), ovs, &(xo[WS(os, 1)]));
813 T63 = VFMA(LDK(KP773010453), T62, T61);
814 T66 = VFMA(LDK(KP773010453), T65, T64);
815 ST(&(xo[WS(os, 23)]), VFMAI(T66, T63), ovs, &(xo[WS(os, 1)]));
816 ST(&(xo[WS(os, 41)]), VFNMSI(T66, T63), ovs, &(xo[WS(os, 1)]));
817 }
818 }
819 }
820 }
821 VLEAVE();
822 }
823
824 static const kdft_desc desc = { 64, XSIMD_STRING("n1fv_64"), {198, 0, 258, 0}, &GENUS, 0, 0, 0, 0 };
825
826 void XSIMD(codelet_n1fv_64) (planner *p) {
827 X(kdft_register) (p, n1fv_64, &desc);
828 }
829
830 #else
831
832 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n1fv_64 -include dft/simd/n1f.h */
833
834 /*
835 * This function contains 456 FP additions, 124 FP multiplications,
836 * (or, 404 additions, 72 multiplications, 52 fused multiply/add),
837 * 108 stack variables, 15 constants, and 128 memory accesses
838 */
839 #include "dft/simd/n1f.h"
840
841 static void n1fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
842 {
843 DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
844 DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
845 DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
846 DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
847 DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
848 DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
849 DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
850 DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
851 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
852 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
853 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
854 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
855 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
856 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
857 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
858 {
859 INT i;
860 const R *xi;
861 R *xo;
862 xi = ri;
863 xo = ro;
864 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)) {
865 V T4p, T5q, Tb, T39, T2n, T3A, T6f, T6T, Tq, T3B, T6i, T76, T2i, T3a, T4w;
866 V T5r, TI, T2p, T6C, T6V, T3h, T3E, T4L, T5u, TZ, T2q, T6F, T6U, T3e, T3D;
867 V T4E, T5t, T23, T2N, T6t, T71, T6w, T72, T2c, T2O, T3t, T41, T5f, T5R, T5k;
868 V T5S, T3w, T42, T1s, T2K, T6m, T6Y, T6p, T6Z, T1B, T2L, T3m, T3Y, T4Y, T5O;
869 V T53, T5P, T3p, T3Z;
870 {
871 V T3, T4n, T2m, T4o, T6, T5p, T9, T5o;
872 {
873 V T1, T2, T2k, T2l;
874 T1 = LD(&(xi[0]), ivs, &(xi[0]));
875 T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
876 T3 = VSUB(T1, T2);
877 T4n = VADD(T1, T2);
878 T2k = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
879 T2l = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
880 T2m = VSUB(T2k, T2l);
881 T4o = VADD(T2k, T2l);
882 }
883 {
884 V T4, T5, T7, T8;
885 T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
886 T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
887 T6 = VSUB(T4, T5);
888 T5p = VADD(T4, T5);
889 T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
890 T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
891 T9 = VSUB(T7, T8);
892 T5o = VADD(T7, T8);
893 }
894 T4p = VSUB(T4n, T4o);
895 T5q = VSUB(T5o, T5p);
896 {
897 V Ta, T2j, T6d, T6e;
898 Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
899 Tb = VADD(T3, Ta);
900 T39 = VSUB(T3, Ta);
901 T2j = VMUL(LDK(KP707106781), VSUB(T9, T6));
902 T2n = VSUB(T2j, T2m);
903 T3A = VADD(T2m, T2j);
904 T6d = VADD(T4n, T4o);
905 T6e = VADD(T5p, T5o);
906 T6f = VADD(T6d, T6e);
907 T6T = VSUB(T6d, T6e);
908 }
909 }
910 {
911 V Te, T4q, To, T4u, Th, T4r, Tl, T4t;
912 {
913 V Tc, Td, Tm, Tn;
914 Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
915 Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
916 Te = VSUB(Tc, Td);
917 T4q = VADD(Tc, Td);
918 Tm = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
919 Tn = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
920 To = VSUB(Tm, Tn);
921 T4u = VADD(Tm, Tn);
922 }
923 {
924 V Tf, Tg, Tj, Tk;
925 Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
926 Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
927 Th = VSUB(Tf, Tg);
928 T4r = VADD(Tf, Tg);
929 Tj = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
930 Tk = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
931 Tl = VSUB(Tj, Tk);
932 T4t = VADD(Tj, Tk);
933 }
934 {
935 V Ti, Tp, T6g, T6h;
936 Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
937 Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
938 Tq = VADD(Ti, Tp);
939 T3B = VSUB(Tp, Ti);
940 T6g = VADD(T4q, T4r);
941 T6h = VADD(T4t, T4u);
942 T6i = VADD(T6g, T6h);
943 T76 = VSUB(T6h, T6g);
944 }
945 {
946 V T2g, T2h, T4s, T4v;
947 T2g = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
948 T2h = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
949 T2i = VSUB(T2g, T2h);
950 T3a = VADD(T2h, T2g);
951 T4s = VSUB(T4q, T4r);
952 T4v = VSUB(T4t, T4u);
953 T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v));
954 T5r = VMUL(LDK(KP707106781), VSUB(T4v, T4s));
955 }
956 }
957 {
958 V Tu, T4F, TG, T4G, TB, T4J, TD, T4I;
959 {
960 V Ts, Tt, TE, TF;
961 Ts = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
962 Tt = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
963 Tu = VSUB(Ts, Tt);
964 T4F = VADD(Ts, Tt);
965 TE = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
966 TF = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
967 TG = VSUB(TE, TF);
968 T4G = VADD(TE, TF);
969 {
970 V Tv, Tw, Tx, Ty, Tz, TA;
971 Tv = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
972 Tw = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
973 Tx = VSUB(Tv, Tw);
974 Ty = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
975 Tz = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
976 TA = VSUB(Ty, Tz);
977 TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
978 T4J = VADD(Tv, Tw);
979 TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
980 T4I = VADD(Ty, Tz);
981 }
982 }
983 {
984 V TC, TH, T6A, T6B;
985 TC = VADD(Tu, TB);
986 TH = VSUB(TD, TG);
987 TI = VFMA(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
988 T2p = VFNMS(LDK(KP195090322), TH, VMUL(LDK(KP980785280), TC));
989 T6A = VADD(T4F, T4G);
990 T6B = VADD(T4J, T4I);
991 T6C = VADD(T6A, T6B);
992 T6V = VSUB(T6A, T6B);
993 }
994 {
995 V T3f, T3g, T4H, T4K;
996 T3f = VSUB(Tu, TB);
997 T3g = VADD(TG, TD);
998 T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f));
999 T3E = VFMA(LDK(KP555570233), T3f, VMUL(LDK(KP831469612), T3g));
1000 T4H = VSUB(T4F, T4G);
1001 T4K = VSUB(T4I, T4J);
1002 T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H));
1003 T5u = VFMA(LDK(KP382683432), T4H, VMUL(LDK(KP923879532), T4K));
1004 }
1005 }
1006 {
1007 V TS, T4z, TW, T4y, TP, T4C, TX, T4B;
1008 {
1009 V TQ, TR, TU, TV;
1010 TQ = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
1011 TR = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
1012 TS = VSUB(TQ, TR);
1013 T4z = VADD(TQ, TR);
1014 TU = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
1015 TV = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
1016 TW = VSUB(TU, TV);
1017 T4y = VADD(TU, TV);
1018 {
1019 V TJ, TK, TL, TM, TN, TO;
1020 TJ = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
1021 TK = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
1022 TL = VSUB(TJ, TK);
1023 TM = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
1024 TN = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
1025 TO = VSUB(TM, TN);
1026 TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
1027 T4C = VADD(TM, TN);
1028 TX = VMUL(LDK(KP707106781), VADD(TO, TL));
1029 T4B = VADD(TJ, TK);
1030 }
1031 }
1032 {
1033 V TT, TY, T6D, T6E;
1034 TT = VSUB(TP, TS);
1035 TY = VADD(TW, TX);
1036 TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
1037 T2q = VFMA(LDK(KP980785280), TY, VMUL(LDK(KP195090322), TT));
1038 T6D = VADD(T4y, T4z);
1039 T6E = VADD(T4C, T4B);
1040 T6F = VADD(T6D, T6E);
1041 T6U = VSUB(T6D, T6E);
1042 }
1043 {
1044 V T3c, T3d, T4A, T4D;
1045 T3c = VSUB(TW, TX);
1046 T3d = VADD(TS, TP);
1047 T3e = VFMA(LDK(KP831469612), T3c, VMUL(LDK(KP555570233), T3d));
1048 T3D = VFNMS(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d));
1049 T4A = VSUB(T4y, T4z);
1050 T4D = VSUB(T4B, T4C);
1051 T4E = VFMA(LDK(KP923879532), T4A, VMUL(LDK(KP382683432), T4D));
1052 T5t = VFNMS(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D));
1053 }
1054 }
1055 {
1056 V T1F, T55, T2a, T56, T1M, T5h, T27, T5g, T58, T59, T1U, T5a, T25, T5b, T5c;
1057 V T21, T5d, T24;
1058 {
1059 V T1D, T1E, T28, T29;
1060 T1D = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
1061 T1E = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
1062 T1F = VSUB(T1D, T1E);
1063 T55 = VADD(T1D, T1E);
1064 T28 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
1065 T29 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
1066 T2a = VSUB(T28, T29);
1067 T56 = VADD(T28, T29);
1068 }
1069 {
1070 V T1G, T1H, T1I, T1J, T1K, T1L;
1071 T1G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
1072 T1H = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
1073 T1I = VSUB(T1G, T1H);
1074 T1J = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
1075 T1K = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
1076 T1L = VSUB(T1J, T1K);
1077 T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L));
1078 T5h = VADD(T1G, T1H);
1079 T27 = VMUL(LDK(KP707106781), VSUB(T1L, T1I));
1080 T5g = VADD(T1J, T1K);
1081 }
1082 {
1083 V T1Q, T1T, T1X, T20;
1084 {
1085 V T1O, T1P, T1R, T1S;
1086 T1O = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
1087 T1P = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
1088 T1Q = VSUB(T1O, T1P);
1089 T58 = VADD(T1O, T1P);
1090 T1R = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
1091 T1S = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
1092 T1T = VSUB(T1R, T1S);
1093 T59 = VADD(T1R, T1S);
1094 }
1095 T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q));
1096 T5a = VSUB(T58, T59);
1097 T25 = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T));
1098 {
1099 V T1V, T1W, T1Y, T1Z;
1100 T1V = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
1101 T1W = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
1102 T1X = VSUB(T1V, T1W);
1103 T5b = VADD(T1V, T1W);
1104 T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
1105 T1Z = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
1106 T20 = VSUB(T1Y, T1Z);
1107 T5c = VADD(T1Y, T1Z);
1108 }
1109 T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20));
1110 T5d = VSUB(T5b, T5c);
1111 T24 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X));
1112 }
1113 {
1114 V T1N, T22, T6r, T6s;
1115 T1N = VADD(T1F, T1M);
1116 T22 = VADD(T1U, T21);
1117 T23 = VSUB(T1N, T22);
1118 T2N = VADD(T1N, T22);
1119 T6r = VADD(T55, T56);
1120 T6s = VADD(T5h, T5g);
1121 T6t = VADD(T6r, T6s);
1122 T71 = VSUB(T6r, T6s);
1123 }
1124 {
1125 V T6u, T6v, T26, T2b;
1126 T6u = VADD(T58, T59);
1127 T6v = VADD(T5b, T5c);
1128 T6w = VADD(T6u, T6v);
1129 T72 = VSUB(T6v, T6u);
1130 T26 = VSUB(T24, T25);
1131 T2b = VSUB(T27, T2a);
1132 T2c = VSUB(T26, T2b);
1133 T2O = VADD(T2b, T26);
1134 }
1135 {
1136 V T3r, T3s, T57, T5e;
1137 T3r = VSUB(T1F, T1M);
1138 T3s = VADD(T25, T24);
1139 T3t = VADD(T3r, T3s);
1140 T41 = VSUB(T3r, T3s);
1141 T57 = VSUB(T55, T56);
1142 T5e = VMUL(LDK(KP707106781), VADD(T5a, T5d));
1143 T5f = VADD(T57, T5e);
1144 T5R = VSUB(T57, T5e);
1145 }
1146 {
1147 V T5i, T5j, T3u, T3v;
1148 T5i = VSUB(T5g, T5h);
1149 T5j = VMUL(LDK(KP707106781), VSUB(T5d, T5a));
1150 T5k = VADD(T5i, T5j);
1151 T5S = VSUB(T5j, T5i);
1152 T3u = VADD(T2a, T27);
1153 T3v = VSUB(T21, T1U);
1154 T3w = VADD(T3u, T3v);
1155 T42 = VSUB(T3v, T3u);
1156 }
1157 }
1158 {
1159 V T1q, T4P, T1v, T4O, T1n, T50, T1w, T4Z, T4U, T4V, T18, T4W, T1z, T4R, T4S;
1160 V T1f, T4T, T1y;
1161 {
1162 V T1o, T1p, T1t, T1u;
1163 T1o = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
1164 T1p = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
1165 T1q = VSUB(T1o, T1p);
1166 T4P = VADD(T1o, T1p);
1167 T1t = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
1168 T1u = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
1169 T1v = VSUB(T1t, T1u);
1170 T4O = VADD(T1t, T1u);
1171 }
1172 {
1173 V T1h, T1i, T1j, T1k, T1l, T1m;
1174 T1h = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
1175 T1i = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
1176 T1j = VSUB(T1h, T1i);
1177 T1k = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
1178 T1l = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
1179 T1m = VSUB(T1k, T1l);
1180 T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m));
1181 T50 = VADD(T1k, T1l);
1182 T1w = VMUL(LDK(KP707106781), VADD(T1m, T1j));
1183 T4Z = VADD(T1h, T1i);
1184 }
1185 {
1186 V T14, T17, T1b, T1e;
1187 {
1188 V T12, T13, T15, T16;
1189 T12 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
1190 T13 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
1191 T14 = VSUB(T12, T13);
1192 T4U = VADD(T12, T13);
1193 T15 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
1194 T16 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
1195 T17 = VSUB(T15, T16);
1196 T4V = VADD(T15, T16);
1197 }
1198 T18 = VFNMS(LDK(KP923879532), T17, VMUL(LDK(KP382683432), T14));
1199 T4W = VSUB(T4U, T4V);
1200 T1z = VFMA(LDK(KP923879532), T14, VMUL(LDK(KP382683432), T17));
1201 {
1202 V T19, T1a, T1c, T1d;
1203 T19 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
1204 T1a = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
1205 T1b = VSUB(T19, T1a);
1206 T4R = VADD(T19, T1a);
1207 T1c = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
1208 T1d = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
1209 T1e = VSUB(T1c, T1d);
1210 T4S = VADD(T1c, T1d);
1211 }
1212 T1f = VFMA(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e));
1213 T4T = VSUB(T4R, T4S);
1214 T1y = VFNMS(LDK(KP382683432), T1e, VMUL(LDK(KP923879532), T1b));
1215 }
1216 {
1217 V T1g, T1r, T6k, T6l;
1218 T1g = VSUB(T18, T1f);
1219 T1r = VSUB(T1n, T1q);
1220 T1s = VSUB(T1g, T1r);
1221 T2K = VADD(T1r, T1g);
1222 T6k = VADD(T4O, T4P);
1223 T6l = VADD(T50, T4Z);
1224 T6m = VADD(T6k, T6l);
1225 T6Y = VSUB(T6k, T6l);
1226 }
1227 {
1228 V T6n, T6o, T1x, T1A;
1229 T6n = VADD(T4R, T4S);
1230 T6o = VADD(T4U, T4V);
1231 T6p = VADD(T6n, T6o);
1232 T6Z = VSUB(T6o, T6n);
1233 T1x = VADD(T1v, T1w);
1234 T1A = VADD(T1y, T1z);
1235 T1B = VSUB(T1x, T1A);
1236 T2L = VADD(T1x, T1A);
1237 }
1238 {
1239 V T3k, T3l, T4Q, T4X;
1240 T3k = VSUB(T1v, T1w);
1241 T3l = VADD(T1f, T18);
1242 T3m = VADD(T3k, T3l);
1243 T3Y = VSUB(T3k, T3l);
1244 T4Q = VSUB(T4O, T4P);
1245 T4X = VMUL(LDK(KP707106781), VADD(T4T, T4W));
1246 T4Y = VADD(T4Q, T4X);
1247 T5O = VSUB(T4Q, T4X);
1248 }
1249 {
1250 V T51, T52, T3n, T3o;
1251 T51 = VSUB(T4Z, T50);
1252 T52 = VMUL(LDK(KP707106781), VSUB(T4W, T4T));
1253 T53 = VADD(T51, T52);
1254 T5P = VSUB(T52, T51);
1255 T3n = VADD(T1q, T1n);
1256 T3o = VSUB(T1z, T1y);
1257 T3p = VADD(T3n, T3o);
1258 T3Z = VSUB(T3o, T3n);
1259 }
1260 }
1261 {
1262 V T6N, T6R, T6Q, T6S;
1263 {
1264 V T6L, T6M, T6O, T6P;
1265 T6L = VADD(T6f, T6i);
1266 T6M = VADD(T6F, T6C);
1267 T6N = VADD(T6L, T6M);
1268 T6R = VSUB(T6L, T6M);
1269 T6O = VADD(T6m, T6p);
1270 T6P = VADD(T6t, T6w);
1271 T6Q = VADD(T6O, T6P);
1272 T6S = VBYI(VSUB(T6P, T6O));
1273 }
1274 ST(&(xo[WS(os, 32)]), VSUB(T6N, T6Q), ovs, &(xo[0]));
1275 ST(&(xo[WS(os, 16)]), VADD(T6R, T6S), ovs, &(xo[0]));
1276 ST(&(xo[0]), VADD(T6N, T6Q), ovs, &(xo[0]));
1277 ST(&(xo[WS(os, 48)]), VSUB(T6R, T6S), ovs, &(xo[0]));
1278 }
1279 {
1280 V T6j, T6G, T6y, T6H, T6q, T6x;
1281 T6j = VSUB(T6f, T6i);
1282 T6G = VSUB(T6C, T6F);
1283 T6q = VSUB(T6m, T6p);
1284 T6x = VSUB(T6t, T6w);
1285 T6y = VMUL(LDK(KP707106781), VADD(T6q, T6x));
1286 T6H = VMUL(LDK(KP707106781), VSUB(T6x, T6q));
1287 {
1288 V T6z, T6I, T6J, T6K;
1289 T6z = VADD(T6j, T6y);
1290 T6I = VBYI(VADD(T6G, T6H));
1291 ST(&(xo[WS(os, 56)]), VSUB(T6z, T6I), ovs, &(xo[0]));
1292 ST(&(xo[WS(os, 8)]), VADD(T6z, T6I), ovs, &(xo[0]));
1293 T6J = VSUB(T6j, T6y);
1294 T6K = VBYI(VSUB(T6H, T6G));
1295 ST(&(xo[WS(os, 40)]), VSUB(T6J, T6K), ovs, &(xo[0]));
1296 ST(&(xo[WS(os, 24)]), VADD(T6J, T6K), ovs, &(xo[0]));
1297 }
1298 }
1299 {
1300 V T6X, T7i, T78, T7g, T74, T7f, T7b, T7j, T6W, T77;
1301 T6W = VMUL(LDK(KP707106781), VADD(T6U, T6V));
1302 T6X = VADD(T6T, T6W);
1303 T7i = VSUB(T6T, T6W);
1304 T77 = VMUL(LDK(KP707106781), VSUB(T6V, T6U));
1305 T78 = VADD(T76, T77);
1306 T7g = VSUB(T77, T76);
1307 {
1308 V T70, T73, T79, T7a;
1309 T70 = VFMA(LDK(KP923879532), T6Y, VMUL(LDK(KP382683432), T6Z));
1310 T73 = VFNMS(LDK(KP382683432), T72, VMUL(LDK(KP923879532), T71));
1311 T74 = VADD(T70, T73);
1312 T7f = VSUB(T73, T70);
1313 T79 = VFNMS(LDK(KP382683432), T6Y, VMUL(LDK(KP923879532), T6Z));
1314 T7a = VFMA(LDK(KP382683432), T71, VMUL(LDK(KP923879532), T72));
1315 T7b = VADD(T79, T7a);
1316 T7j = VSUB(T7a, T79);
1317 }
1318 {
1319 V T75, T7c, T7l, T7m;
1320 T75 = VADD(T6X, T74);
1321 T7c = VBYI(VADD(T78, T7b));
1322 ST(&(xo[WS(os, 60)]), VSUB(T75, T7c), ovs, &(xo[0]));
1323 ST(&(xo[WS(os, 4)]), VADD(T75, T7c), ovs, &(xo[0]));
1324 T7l = VBYI(VADD(T7g, T7f));
1325 T7m = VADD(T7i, T7j);
1326 ST(&(xo[WS(os, 12)]), VADD(T7l, T7m), ovs, &(xo[0]));
1327 ST(&(xo[WS(os, 52)]), VSUB(T7m, T7l), ovs, &(xo[0]));
1328 }
1329 {
1330 V T7d, T7e, T7h, T7k;
1331 T7d = VSUB(T6X, T74);
1332 T7e = VBYI(VSUB(T7b, T78));
1333 ST(&(xo[WS(os, 36)]), VSUB(T7d, T7e), ovs, &(xo[0]));
1334 ST(&(xo[WS(os, 28)]), VADD(T7d, T7e), ovs, &(xo[0]));
1335 T7h = VBYI(VSUB(T7f, T7g));
1336 T7k = VSUB(T7i, T7j);
1337 ST(&(xo[WS(os, 20)]), VADD(T7h, T7k), ovs, &(xo[0]));
1338 ST(&(xo[WS(os, 44)]), VSUB(T7k, T7h), ovs, &(xo[0]));
1339 }
1340 }
1341 {
1342 V T5N, T68, T61, T69, T5U, T65, T5Y, T66;
1343 {
1344 V T5L, T5M, T5Z, T60;
1345 T5L = VSUB(T4p, T4w);
1346 T5M = VSUB(T5u, T5t);
1347 T5N = VADD(T5L, T5M);
1348 T68 = VSUB(T5L, T5M);
1349 T5Z = VFNMS(LDK(KP555570233), T5O, VMUL(LDK(KP831469612), T5P));
1350 T60 = VFMA(LDK(KP555570233), T5R, VMUL(LDK(KP831469612), T5S));
1351 T61 = VADD(T5Z, T60);
1352 T69 = VSUB(T60, T5Z);
1353 }
1354 {
1355 V T5Q, T5T, T5W, T5X;
1356 T5Q = VFMA(LDK(KP831469612), T5O, VMUL(LDK(KP555570233), T5P));
1357 T5T = VFNMS(LDK(KP555570233), T5S, VMUL(LDK(KP831469612), T5R));
1358 T5U = VADD(T5Q, T5T);
1359 T65 = VSUB(T5T, T5Q);
1360 T5W = VSUB(T5r, T5q);
1361 T5X = VSUB(T4L, T4E);
1362 T5Y = VADD(T5W, T5X);
1363 T66 = VSUB(T5X, T5W);
1364 }
1365 {
1366 V T5V, T62, T6b, T6c;
1367 T5V = VADD(T5N, T5U);
1368 T62 = VBYI(VADD(T5Y, T61));
1369 ST(&(xo[WS(os, 58)]), VSUB(T5V, T62), ovs, &(xo[0]));
1370 ST(&(xo[WS(os, 6)]), VADD(T5V, T62), ovs, &(xo[0]));
1371 T6b = VBYI(VADD(T66, T65));
1372 T6c = VADD(T68, T69);
1373 ST(&(xo[WS(os, 10)]), VADD(T6b, T6c), ovs, &(xo[0]));
1374 ST(&(xo[WS(os, 54)]), VSUB(T6c, T6b), ovs, &(xo[0]));
1375 }
1376 {
1377 V T63, T64, T67, T6a;
1378 T63 = VSUB(T5N, T5U);
1379 T64 = VBYI(VSUB(T61, T5Y));
1380 ST(&(xo[WS(os, 38)]), VSUB(T63, T64), ovs, &(xo[0]));
1381 ST(&(xo[WS(os, 26)]), VADD(T63, T64), ovs, &(xo[0]));
1382 T67 = VBYI(VSUB(T65, T66));
1383 T6a = VSUB(T68, T69);
1384 ST(&(xo[WS(os, 22)]), VADD(T67, T6a), ovs, &(xo[0]));
1385 ST(&(xo[WS(os, 42)]), VSUB(T6a, T67), ovs, &(xo[0]));
1386 }
1387 }
1388 {
1389 V T11, T2C, T2v, T2D, T2e, T2z, T2s, T2A;
1390 {
1391 V Tr, T10, T2t, T2u;
1392 Tr = VSUB(Tb, Tq);
1393 T10 = VSUB(TI, TZ);
1394 T11 = VADD(Tr, T10);
1395 T2C = VSUB(Tr, T10);
1396 T2t = VFNMS(LDK(KP634393284), T1B, VMUL(LDK(KP773010453), T1s));
1397 T2u = VFMA(LDK(KP773010453), T2c, VMUL(LDK(KP634393284), T23));
1398 T2v = VADD(T2t, T2u);
1399 T2D = VSUB(T2u, T2t);
1400 }
1401 {
1402 V T1C, T2d, T2o, T2r;
1403 T1C = VFMA(LDK(KP634393284), T1s, VMUL(LDK(KP773010453), T1B));
1404 T2d = VFNMS(LDK(KP634393284), T2c, VMUL(LDK(KP773010453), T23));
1405 T2e = VADD(T1C, T2d);
1406 T2z = VSUB(T2d, T1C);
1407 T2o = VSUB(T2i, T2n);
1408 T2r = VSUB(T2p, T2q);
1409 T2s = VADD(T2o, T2r);
1410 T2A = VSUB(T2r, T2o);
1411 }
1412 {
1413 V T2f, T2w, T2F, T2G;
1414 T2f = VADD(T11, T2e);
1415 T2w = VBYI(VADD(T2s, T2v));
1416 ST(&(xo[WS(os, 57)]), VSUB(T2f, T2w), ovs, &(xo[WS(os, 1)]));
1417 ST(&(xo[WS(os, 7)]), VADD(T2f, T2w), ovs, &(xo[WS(os, 1)]));
1418 T2F = VBYI(VADD(T2A, T2z));
1419 T2G = VADD(T2C, T2D);
1420 ST(&(xo[WS(os, 9)]), VADD(T2F, T2G), ovs, &(xo[WS(os, 1)]));
1421 ST(&(xo[WS(os, 55)]), VSUB(T2G, T2F), ovs, &(xo[WS(os, 1)]));
1422 }
1423 {
1424 V T2x, T2y, T2B, T2E;
1425 T2x = VSUB(T11, T2e);
1426 T2y = VBYI(VSUB(T2v, T2s));
1427 ST(&(xo[WS(os, 39)]), VSUB(T2x, T2y), ovs, &(xo[WS(os, 1)]));
1428 ST(&(xo[WS(os, 25)]), VADD(T2x, T2y), ovs, &(xo[WS(os, 1)]));
1429 T2B = VBYI(VSUB(T2z, T2A));
1430 T2E = VSUB(T2C, T2D);
1431 ST(&(xo[WS(os, 23)]), VADD(T2B, T2E), ovs, &(xo[WS(os, 1)]));
1432 ST(&(xo[WS(os, 41)]), VSUB(T2E, T2B), ovs, &(xo[WS(os, 1)]));
1433 }
1434 }
1435 {
1436 V T3j, T3Q, T3J, T3R, T3y, T3N, T3G, T3O;
1437 {
1438 V T3b, T3i, T3H, T3I;
1439 T3b = VADD(T39, T3a);
1440 T3i = VADD(T3e, T3h);
1441 T3j = VADD(T3b, T3i);
1442 T3Q = VSUB(T3b, T3i);
1443 T3H = VFNMS(LDK(KP290284677), T3m, VMUL(LDK(KP956940335), T3p));
1444 T3I = VFMA(LDK(KP290284677), T3t, VMUL(LDK(KP956940335), T3w));
1445 T3J = VADD(T3H, T3I);
1446 T3R = VSUB(T3I, T3H);
1447 }
1448 {
1449 V T3q, T3x, T3C, T3F;
1450 T3q = VFMA(LDK(KP956940335), T3m, VMUL(LDK(KP290284677), T3p));
1451 T3x = VFNMS(LDK(KP290284677), T3w, VMUL(LDK(KP956940335), T3t));
1452 T3y = VADD(T3q, T3x);
1453 T3N = VSUB(T3x, T3q);
1454 T3C = VADD(T3A, T3B);
1455 T3F = VADD(T3D, T3E);
1456 T3G = VADD(T3C, T3F);
1457 T3O = VSUB(T3F, T3C);
1458 }
1459 {
1460 V T3z, T3K, T3T, T3U;
1461 T3z = VADD(T3j, T3y);
1462 T3K = VBYI(VADD(T3G, T3J));
1463 ST(&(xo[WS(os, 61)]), VSUB(T3z, T3K), ovs, &(xo[WS(os, 1)]));
1464 ST(&(xo[WS(os, 3)]), VADD(T3z, T3K), ovs, &(xo[WS(os, 1)]));
1465 T3T = VBYI(VADD(T3O, T3N));
1466 T3U = VADD(T3Q, T3R);
1467 ST(&(xo[WS(os, 13)]), VADD(T3T, T3U), ovs, &(xo[WS(os, 1)]));
1468 ST(&(xo[WS(os, 51)]), VSUB(T3U, T3T), ovs, &(xo[WS(os, 1)]));
1469 }
1470 {
1471 V T3L, T3M, T3P, T3S;
1472 T3L = VSUB(T3j, T3y);
1473 T3M = VBYI(VSUB(T3J, T3G));
1474 ST(&(xo[WS(os, 35)]), VSUB(T3L, T3M), ovs, &(xo[WS(os, 1)]));
1475 ST(&(xo[WS(os, 29)]), VADD(T3L, T3M), ovs, &(xo[WS(os, 1)]));
1476 T3P = VBYI(VSUB(T3N, T3O));
1477 T3S = VSUB(T3Q, T3R);
1478 ST(&(xo[WS(os, 19)]), VADD(T3P, T3S), ovs, &(xo[WS(os, 1)]));
1479 ST(&(xo[WS(os, 45)]), VSUB(T3S, T3P), ovs, &(xo[WS(os, 1)]));
1480 }
1481 }
1482 {
1483 V T4N, T5G, T5z, T5H, T5m, T5D, T5w, T5E;
1484 {
1485 V T4x, T4M, T5x, T5y;
1486 T4x = VADD(T4p, T4w);
1487 T4M = VADD(T4E, T4L);
1488 T4N = VADD(T4x, T4M);
1489 T5G = VSUB(T4x, T4M);
1490 T5x = VFNMS(LDK(KP195090322), T4Y, VMUL(LDK(KP980785280), T53));
1491 T5y = VFMA(LDK(KP195090322), T5f, VMUL(LDK(KP980785280), T5k));
1492 T5z = VADD(T5x, T5y);
1493 T5H = VSUB(T5y, T5x);
1494 }
1495 {
1496 V T54, T5l, T5s, T5v;
1497 T54 = VFMA(LDK(KP980785280), T4Y, VMUL(LDK(KP195090322), T53));
1498 T5l = VFNMS(LDK(KP195090322), T5k, VMUL(LDK(KP980785280), T5f));
1499 T5m = VADD(T54, T5l);
1500 T5D = VSUB(T5l, T54);
1501 T5s = VADD(T5q, T5r);
1502 T5v = VADD(T5t, T5u);
1503 T5w = VADD(T5s, T5v);
1504 T5E = VSUB(T5v, T5s);
1505 }
1506 {
1507 V T5n, T5A, T5J, T5K;
1508 T5n = VADD(T4N, T5m);
1509 T5A = VBYI(VADD(T5w, T5z));
1510 ST(&(xo[WS(os, 62)]), VSUB(T5n, T5A), ovs, &(xo[0]));
1511 ST(&(xo[WS(os, 2)]), VADD(T5n, T5A), ovs, &(xo[0]));
1512 T5J = VBYI(VADD(T5E, T5D));
1513 T5K = VADD(T5G, T5H);
1514 ST(&(xo[WS(os, 14)]), VADD(T5J, T5K), ovs, &(xo[0]));
1515 ST(&(xo[WS(os, 50)]), VSUB(T5K, T5J), ovs, &(xo[0]));
1516 }
1517 {
1518 V T5B, T5C, T5F, T5I;
1519 T5B = VSUB(T4N, T5m);
1520 T5C = VBYI(VSUB(T5z, T5w));
1521 ST(&(xo[WS(os, 34)]), VSUB(T5B, T5C), ovs, &(xo[0]));
1522 ST(&(xo[WS(os, 30)]), VADD(T5B, T5C), ovs, &(xo[0]));
1523 T5F = VBYI(VSUB(T5D, T5E));
1524 T5I = VSUB(T5G, T5H);
1525 ST(&(xo[WS(os, 18)]), VADD(T5F, T5I), ovs, &(xo[0]));
1526 ST(&(xo[WS(os, 46)]), VSUB(T5I, T5F), ovs, &(xo[0]));
1527 }
1528 }
1529 {
1530 V T2J, T34, T2X, T35, T2Q, T31, T2U, T32;
1531 {
1532 V T2H, T2I, T2V, T2W;
1533 T2H = VADD(Tb, Tq);
1534 T2I = VADD(T2q, T2p);
1535 T2J = VADD(T2H, T2I);
1536 T34 = VSUB(T2H, T2I);
1537 T2V = VFNMS(LDK(KP098017140), T2L, VMUL(LDK(KP995184726), T2K));
1538 T2W = VFMA(LDK(KP995184726), T2O, VMUL(LDK(KP098017140), T2N));
1539 T2X = VADD(T2V, T2W);
1540 T35 = VSUB(T2W, T2V);
1541 }
1542 {
1543 V T2M, T2P, T2S, T2T;
1544 T2M = VFMA(LDK(KP098017140), T2K, VMUL(LDK(KP995184726), T2L));
1545 T2P = VFNMS(LDK(KP098017140), T2O, VMUL(LDK(KP995184726), T2N));
1546 T2Q = VADD(T2M, T2P);
1547 T31 = VSUB(T2P, T2M);
1548 T2S = VADD(T2n, T2i);
1549 T2T = VADD(TZ, TI);
1550 T2U = VADD(T2S, T2T);
1551 T32 = VSUB(T2T, T2S);
1552 }
1553 {
1554 V T2R, T2Y, T37, T38;
1555 T2R = VADD(T2J, T2Q);
1556 T2Y = VBYI(VADD(T2U, T2X));
1557 ST(&(xo[WS(os, 63)]), VSUB(T2R, T2Y), ovs, &(xo[WS(os, 1)]));
1558 ST(&(xo[WS(os, 1)]), VADD(T2R, T2Y), ovs, &(xo[WS(os, 1)]));
1559 T37 = VBYI(VADD(T32, T31));
1560 T38 = VADD(T34, T35);
1561 ST(&(xo[WS(os, 15)]), VADD(T37, T38), ovs, &(xo[WS(os, 1)]));
1562 ST(&(xo[WS(os, 49)]), VSUB(T38, T37), ovs, &(xo[WS(os, 1)]));
1563 }
1564 {
1565 V T2Z, T30, T33, T36;
1566 T2Z = VSUB(T2J, T2Q);
1567 T30 = VBYI(VSUB(T2X, T2U));
1568 ST(&(xo[WS(os, 33)]), VSUB(T2Z, T30), ovs, &(xo[WS(os, 1)]));
1569 ST(&(xo[WS(os, 31)]), VADD(T2Z, T30), ovs, &(xo[WS(os, 1)]));
1570 T33 = VBYI(VSUB(T31, T32));
1571 T36 = VSUB(T34, T35);
1572 ST(&(xo[WS(os, 17)]), VADD(T33, T36), ovs, &(xo[WS(os, 1)]));
1573 ST(&(xo[WS(os, 47)]), VSUB(T36, T33), ovs, &(xo[WS(os, 1)]));
1574 }
1575 }
1576 {
1577 V T3X, T4i, T4b, T4j, T44, T4f, T48, T4g;
1578 {
1579 V T3V, T3W, T49, T4a;
1580 T3V = VSUB(T39, T3a);
1581 T3W = VSUB(T3E, T3D);
1582 T3X = VADD(T3V, T3W);
1583 T4i = VSUB(T3V, T3W);
1584 T49 = VFNMS(LDK(KP471396736), T3Y, VMUL(LDK(KP881921264), T3Z));
1585 T4a = VFMA(LDK(KP471396736), T41, VMUL(LDK(KP881921264), T42));
1586 T4b = VADD(T49, T4a);
1587 T4j = VSUB(T4a, T49);
1588 }
1589 {
1590 V T40, T43, T46, T47;
1591 T40 = VFMA(LDK(KP881921264), T3Y, VMUL(LDK(KP471396736), T3Z));
1592 T43 = VFNMS(LDK(KP471396736), T42, VMUL(LDK(KP881921264), T41));
1593 T44 = VADD(T40, T43);
1594 T4f = VSUB(T43, T40);
1595 T46 = VSUB(T3B, T3A);
1596 T47 = VSUB(T3h, T3e);
1597 T48 = VADD(T46, T47);
1598 T4g = VSUB(T47, T46);
1599 }
1600 {
1601 V T45, T4c, T4l, T4m;
1602 T45 = VADD(T3X, T44);
1603 T4c = VBYI(VADD(T48, T4b));
1604 ST(&(xo[WS(os, 59)]), VSUB(T45, T4c), ovs, &(xo[WS(os, 1)]));
1605 ST(&(xo[WS(os, 5)]), VADD(T45, T4c), ovs, &(xo[WS(os, 1)]));
1606 T4l = VBYI(VADD(T4g, T4f));
1607 T4m = VADD(T4i, T4j);
1608 ST(&(xo[WS(os, 11)]), VADD(T4l, T4m), ovs, &(xo[WS(os, 1)]));
1609 ST(&(xo[WS(os, 53)]), VSUB(T4m, T4l), ovs, &(xo[WS(os, 1)]));
1610 }
1611 {
1612 V T4d, T4e, T4h, T4k;
1613 T4d = VSUB(T3X, T44);
1614 T4e = VBYI(VSUB(T4b, T48));
1615 ST(&(xo[WS(os, 37)]), VSUB(T4d, T4e), ovs, &(xo[WS(os, 1)]));
1616 ST(&(xo[WS(os, 27)]), VADD(T4d, T4e), ovs, &(xo[WS(os, 1)]));
1617 T4h = VBYI(VSUB(T4f, T4g));
1618 T4k = VSUB(T4i, T4j);
1619 ST(&(xo[WS(os, 21)]), VADD(T4h, T4k), ovs, &(xo[WS(os, 1)]));
1620 ST(&(xo[WS(os, 43)]), VSUB(T4k, T4h), ovs, &(xo[WS(os, 1)]));
1621 }
1622 }
1623 }
1624 }
1625 VLEAVE();
1626 }
1627
1628 static const kdft_desc desc = { 64, XSIMD_STRING("n1fv_64"), {404, 72, 52, 0}, &GENUS, 0, 0, 0, 0 };
1629
1630 void XSIMD(codelet_n1fv_64) (planner *p) {
1631 X(kdft_register) (p, n1fv_64, &desc);
1632 }
1633
1634 #endif