comparison src/fftw-3.3.8/dft/simd/common/n2sv_32.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 08:05:20 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.native -fma -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n2sv_32 -with-ostride 1 -include dft/simd/n2s.h -store-multiple 4 */
29
30 /*
31 * This function contains 372 FP additions, 136 FP multiplications,
32 * (or, 236 additions, 0 multiplications, 136 fused multiply/add),
33 * 138 stack variables, 7 constants, and 144 memory accesses
34 */
35 #include "dft/simd/n2s.h"
36
37 static void n2sv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38 {
39 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
40 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
41 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
42 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
44 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
45 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
46 {
47 INT i;
48 for (i = v; i > 0; i = i - (2 * VL), ri = ri + ((2 * VL) * ivs), ii = ii + ((2 * VL) * ivs), ro = ro + ((2 * VL) * ovs), io = io + ((2 * VL) * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) {
49 V T7, T4r, T4Z, T18, T1z, T3t, T3T, T2T, Te, T1f, T50, T4s, T2W, T3u, T1G;
50 V T3U, Tm, T1n, T1O, T2Z, T3y, T3X, T4w, T53, Tt, T1u, T1V, T2Y, T3B, T3W;
51 V T4z, T52, T2t, T3L, T3O, T2K, TR, TY, T5F, T5G, T5H, T5I, T4R, T5k, T2E;
52 V T3M, T4W, T5j, T2N, T3P, T22, T3E, T3H, T2j, TC, TJ, T5A, T5B, T5C, T5D;
53 V T4G, T5h, T2d, T3F, T4L, T5g, T2m, T3I;
54 {
55 V T3, T1x, T14, T2R, T6, T2S, T17, T1y;
56 {
57 V T1, T2, T12, T13;
58 T1 = LD(&(ri[0]), ivs, &(ri[0]));
59 T2 = LD(&(ri[WS(is, 16)]), ivs, &(ri[0]));
60 T3 = VADD(T1, T2);
61 T1x = VSUB(T1, T2);
62 T12 = LD(&(ii[0]), ivs, &(ii[0]));
63 T13 = LD(&(ii[WS(is, 16)]), ivs, &(ii[0]));
64 T14 = VADD(T12, T13);
65 T2R = VSUB(T12, T13);
66 }
67 {
68 V T4, T5, T15, T16;
69 T4 = LD(&(ri[WS(is, 8)]), ivs, &(ri[0]));
70 T5 = LD(&(ri[WS(is, 24)]), ivs, &(ri[0]));
71 T6 = VADD(T4, T5);
72 T2S = VSUB(T4, T5);
73 T15 = LD(&(ii[WS(is, 8)]), ivs, &(ii[0]));
74 T16 = LD(&(ii[WS(is, 24)]), ivs, &(ii[0]));
75 T17 = VADD(T15, T16);
76 T1y = VSUB(T15, T16);
77 }
78 T7 = VADD(T3, T6);
79 T4r = VSUB(T3, T6);
80 T4Z = VSUB(T14, T17);
81 T18 = VADD(T14, T17);
82 T1z = VADD(T1x, T1y);
83 T3t = VSUB(T1x, T1y);
84 T3T = VADD(T2S, T2R);
85 T2T = VSUB(T2R, T2S);
86 }
87 {
88 V Ta, T1A, T1b, T1B, Td, T1D, T1e, T1E;
89 {
90 V T8, T9, T19, T1a;
91 T8 = LD(&(ri[WS(is, 4)]), ivs, &(ri[0]));
92 T9 = LD(&(ri[WS(is, 20)]), ivs, &(ri[0]));
93 Ta = VADD(T8, T9);
94 T1A = VSUB(T8, T9);
95 T19 = LD(&(ii[WS(is, 4)]), ivs, &(ii[0]));
96 T1a = LD(&(ii[WS(is, 20)]), ivs, &(ii[0]));
97 T1b = VADD(T19, T1a);
98 T1B = VSUB(T19, T1a);
99 }
100 {
101 V Tb, Tc, T1c, T1d;
102 Tb = LD(&(ri[WS(is, 28)]), ivs, &(ri[0]));
103 Tc = LD(&(ri[WS(is, 12)]), ivs, &(ri[0]));
104 Td = VADD(Tb, Tc);
105 T1D = VSUB(Tb, Tc);
106 T1c = LD(&(ii[WS(is, 28)]), ivs, &(ii[0]));
107 T1d = LD(&(ii[WS(is, 12)]), ivs, &(ii[0]));
108 T1e = VADD(T1c, T1d);
109 T1E = VSUB(T1c, T1d);
110 }
111 Te = VADD(Ta, Td);
112 T1f = VADD(T1b, T1e);
113 T50 = VSUB(Td, Ta);
114 T4s = VSUB(T1b, T1e);
115 {
116 V T2U, T2V, T1C, T1F;
117 T2U = VSUB(T1B, T1A);
118 T2V = VADD(T1D, T1E);
119 T2W = VADD(T2U, T2V);
120 T3u = VSUB(T2U, T2V);
121 T1C = VADD(T1A, T1B);
122 T1F = VSUB(T1D, T1E);
123 T1G = VADD(T1C, T1F);
124 T3U = VSUB(T1F, T1C);
125 }
126 }
127 {
128 V Ti, T1L, T1j, T1I, Tl, T1J, T1m, T1M, T1K, T1N;
129 {
130 V Tg, Th, T1h, T1i;
131 Tg = LD(&(ri[WS(is, 2)]), ivs, &(ri[0]));
132 Th = LD(&(ri[WS(is, 18)]), ivs, &(ri[0]));
133 Ti = VADD(Tg, Th);
134 T1L = VSUB(Tg, Th);
135 T1h = LD(&(ii[WS(is, 2)]), ivs, &(ii[0]));
136 T1i = LD(&(ii[WS(is, 18)]), ivs, &(ii[0]));
137 T1j = VADD(T1h, T1i);
138 T1I = VSUB(T1h, T1i);
139 }
140 {
141 V Tj, Tk, T1k, T1l;
142 Tj = LD(&(ri[WS(is, 10)]), ivs, &(ri[0]));
143 Tk = LD(&(ri[WS(is, 26)]), ivs, &(ri[0]));
144 Tl = VADD(Tj, Tk);
145 T1J = VSUB(Tj, Tk);
146 T1k = LD(&(ii[WS(is, 10)]), ivs, &(ii[0]));
147 T1l = LD(&(ii[WS(is, 26)]), ivs, &(ii[0]));
148 T1m = VADD(T1k, T1l);
149 T1M = VSUB(T1k, T1l);
150 }
151 Tm = VADD(Ti, Tl);
152 T1n = VADD(T1j, T1m);
153 T1K = VSUB(T1I, T1J);
154 T1N = VADD(T1L, T1M);
155 T1O = VFNMS(LDK(KP414213562), T1N, T1K);
156 T2Z = VFMA(LDK(KP414213562), T1K, T1N);
157 {
158 V T3w, T3x, T4u, T4v;
159 T3w = VADD(T1J, T1I);
160 T3x = VSUB(T1L, T1M);
161 T3y = VFMA(LDK(KP414213562), T3x, T3w);
162 T3X = VFNMS(LDK(KP414213562), T3w, T3x);
163 T4u = VSUB(T1j, T1m);
164 T4v = VSUB(Ti, Tl);
165 T4w = VSUB(T4u, T4v);
166 T53 = VADD(T4v, T4u);
167 }
168 }
169 {
170 V Tp, T1S, T1q, T1P, Ts, T1Q, T1t, T1T, T1R, T1U;
171 {
172 V Tn, To, T1o, T1p;
173 Tn = LD(&(ri[WS(is, 30)]), ivs, &(ri[0]));
174 To = LD(&(ri[WS(is, 14)]), ivs, &(ri[0]));
175 Tp = VADD(Tn, To);
176 T1S = VSUB(Tn, To);
177 T1o = LD(&(ii[WS(is, 30)]), ivs, &(ii[0]));
178 T1p = LD(&(ii[WS(is, 14)]), ivs, &(ii[0]));
179 T1q = VADD(T1o, T1p);
180 T1P = VSUB(T1o, T1p);
181 }
182 {
183 V Tq, Tr, T1r, T1s;
184 Tq = LD(&(ri[WS(is, 6)]), ivs, &(ri[0]));
185 Tr = LD(&(ri[WS(is, 22)]), ivs, &(ri[0]));
186 Ts = VADD(Tq, Tr);
187 T1Q = VSUB(Tq, Tr);
188 T1r = LD(&(ii[WS(is, 6)]), ivs, &(ii[0]));
189 T1s = LD(&(ii[WS(is, 22)]), ivs, &(ii[0]));
190 T1t = VADD(T1r, T1s);
191 T1T = VSUB(T1r, T1s);
192 }
193 Tt = VADD(Tp, Ts);
194 T1u = VADD(T1q, T1t);
195 T1R = VSUB(T1P, T1Q);
196 T1U = VADD(T1S, T1T);
197 T1V = VFMA(LDK(KP414213562), T1U, T1R);
198 T2Y = VFNMS(LDK(KP414213562), T1R, T1U);
199 {
200 V T3z, T3A, T4x, T4y;
201 T3z = VADD(T1Q, T1P);
202 T3A = VSUB(T1S, T1T);
203 T3B = VFNMS(LDK(KP414213562), T3A, T3z);
204 T3W = VFMA(LDK(KP414213562), T3z, T3A);
205 T4x = VSUB(Tp, Ts);
206 T4y = VSUB(T1q, T1t);
207 T4z = VADD(T4x, T4y);
208 T52 = VSUB(T4x, T4y);
209 }
210 }
211 {
212 V TN, T2G, T2r, T4N, TQ, T2s, T2J, T4O, TU, T2x, T2w, T4T, TX, T2z, T2C;
213 V T4U;
214 {
215 V TL, TM, T2p, T2q;
216 TL = LD(&(ri[WS(is, 31)]), ivs, &(ri[WS(is, 1)]));
217 TM = LD(&(ri[WS(is, 15)]), ivs, &(ri[WS(is, 1)]));
218 TN = VADD(TL, TM);
219 T2G = VSUB(TL, TM);
220 T2p = LD(&(ii[WS(is, 31)]), ivs, &(ii[WS(is, 1)]));
221 T2q = LD(&(ii[WS(is, 15)]), ivs, &(ii[WS(is, 1)]));
222 T2r = VSUB(T2p, T2q);
223 T4N = VADD(T2p, T2q);
224 }
225 {
226 V TO, TP, T2H, T2I;
227 TO = LD(&(ri[WS(is, 7)]), ivs, &(ri[WS(is, 1)]));
228 TP = LD(&(ri[WS(is, 23)]), ivs, &(ri[WS(is, 1)]));
229 TQ = VADD(TO, TP);
230 T2s = VSUB(TO, TP);
231 T2H = LD(&(ii[WS(is, 7)]), ivs, &(ii[WS(is, 1)]));
232 T2I = LD(&(ii[WS(is, 23)]), ivs, &(ii[WS(is, 1)]));
233 T2J = VSUB(T2H, T2I);
234 T4O = VADD(T2H, T2I);
235 }
236 {
237 V TS, TT, T2u, T2v;
238 TS = LD(&(ri[WS(is, 3)]), ivs, &(ri[WS(is, 1)]));
239 TT = LD(&(ri[WS(is, 19)]), ivs, &(ri[WS(is, 1)]));
240 TU = VADD(TS, TT);
241 T2x = VSUB(TS, TT);
242 T2u = LD(&(ii[WS(is, 3)]), ivs, &(ii[WS(is, 1)]));
243 T2v = LD(&(ii[WS(is, 19)]), ivs, &(ii[WS(is, 1)]));
244 T2w = VSUB(T2u, T2v);
245 T4T = VADD(T2u, T2v);
246 }
247 {
248 V TV, TW, T2A, T2B;
249 TV = LD(&(ri[WS(is, 27)]), ivs, &(ri[WS(is, 1)]));
250 TW = LD(&(ri[WS(is, 11)]), ivs, &(ri[WS(is, 1)]));
251 TX = VADD(TV, TW);
252 T2z = VSUB(TV, TW);
253 T2A = LD(&(ii[WS(is, 27)]), ivs, &(ii[WS(is, 1)]));
254 T2B = LD(&(ii[WS(is, 11)]), ivs, &(ii[WS(is, 1)]));
255 T2C = VSUB(T2A, T2B);
256 T4U = VADD(T2A, T2B);
257 }
258 T2t = VSUB(T2r, T2s);
259 T3L = VSUB(T2G, T2J);
260 T3O = VADD(T2s, T2r);
261 T2K = VADD(T2G, T2J);
262 TR = VADD(TN, TQ);
263 TY = VADD(TU, TX);
264 T5F = VSUB(TR, TY);
265 {
266 V T4P, T4Q, T2y, T2D;
267 T5G = VADD(T4N, T4O);
268 T5H = VADD(T4T, T4U);
269 T5I = VSUB(T5G, T5H);
270 T4P = VSUB(T4N, T4O);
271 T4Q = VSUB(TX, TU);
272 T4R = VSUB(T4P, T4Q);
273 T5k = VADD(T4Q, T4P);
274 T2y = VSUB(T2w, T2x);
275 T2D = VADD(T2z, T2C);
276 T2E = VADD(T2y, T2D);
277 T3M = VSUB(T2D, T2y);
278 {
279 V T4S, T4V, T2L, T2M;
280 T4S = VSUB(TN, TQ);
281 T4V = VSUB(T4T, T4U);
282 T4W = VSUB(T4S, T4V);
283 T5j = VADD(T4S, T4V);
284 T2L = VADD(T2x, T2w);
285 T2M = VSUB(T2z, T2C);
286 T2N = VADD(T2L, T2M);
287 T3P = VSUB(T2L, T2M);
288 }
289 }
290 }
291 {
292 V Ty, T2f, T20, T4C, TB, T21, T2i, T4D, TF, T26, T25, T4I, TI, T28, T2b;
293 V T4J;
294 {
295 V Tw, Tx, T1Y, T1Z;
296 Tw = LD(&(ri[WS(is, 1)]), ivs, &(ri[WS(is, 1)]));
297 Tx = LD(&(ri[WS(is, 17)]), ivs, &(ri[WS(is, 1)]));
298 Ty = VADD(Tw, Tx);
299 T2f = VSUB(Tw, Tx);
300 T1Y = LD(&(ii[WS(is, 1)]), ivs, &(ii[WS(is, 1)]));
301 T1Z = LD(&(ii[WS(is, 17)]), ivs, &(ii[WS(is, 1)]));
302 T20 = VSUB(T1Y, T1Z);
303 T4C = VADD(T1Y, T1Z);
304 }
305 {
306 V Tz, TA, T2g, T2h;
307 Tz = LD(&(ri[WS(is, 9)]), ivs, &(ri[WS(is, 1)]));
308 TA = LD(&(ri[WS(is, 25)]), ivs, &(ri[WS(is, 1)]));
309 TB = VADD(Tz, TA);
310 T21 = VSUB(Tz, TA);
311 T2g = LD(&(ii[WS(is, 9)]), ivs, &(ii[WS(is, 1)]));
312 T2h = LD(&(ii[WS(is, 25)]), ivs, &(ii[WS(is, 1)]));
313 T2i = VSUB(T2g, T2h);
314 T4D = VADD(T2g, T2h);
315 }
316 {
317 V TD, TE, T23, T24;
318 TD = LD(&(ri[WS(is, 5)]), ivs, &(ri[WS(is, 1)]));
319 TE = LD(&(ri[WS(is, 21)]), ivs, &(ri[WS(is, 1)]));
320 TF = VADD(TD, TE);
321 T26 = VSUB(TD, TE);
322 T23 = LD(&(ii[WS(is, 5)]), ivs, &(ii[WS(is, 1)]));
323 T24 = LD(&(ii[WS(is, 21)]), ivs, &(ii[WS(is, 1)]));
324 T25 = VSUB(T23, T24);
325 T4I = VADD(T23, T24);
326 }
327 {
328 V TG, TH, T29, T2a;
329 TG = LD(&(ri[WS(is, 29)]), ivs, &(ri[WS(is, 1)]));
330 TH = LD(&(ri[WS(is, 13)]), ivs, &(ri[WS(is, 1)]));
331 TI = VADD(TG, TH);
332 T28 = VSUB(TG, TH);
333 T29 = LD(&(ii[WS(is, 29)]), ivs, &(ii[WS(is, 1)]));
334 T2a = LD(&(ii[WS(is, 13)]), ivs, &(ii[WS(is, 1)]));
335 T2b = VSUB(T29, T2a);
336 T4J = VADD(T29, T2a);
337 }
338 T22 = VSUB(T20, T21);
339 T3E = VSUB(T2f, T2i);
340 T3H = VADD(T21, T20);
341 T2j = VADD(T2f, T2i);
342 TC = VADD(Ty, TB);
343 TJ = VADD(TF, TI);
344 T5A = VSUB(TC, TJ);
345 {
346 V T4E, T4F, T27, T2c;
347 T5B = VADD(T4C, T4D);
348 T5C = VADD(T4I, T4J);
349 T5D = VSUB(T5B, T5C);
350 T4E = VSUB(T4C, T4D);
351 T4F = VSUB(TI, TF);
352 T4G = VSUB(T4E, T4F);
353 T5h = VADD(T4F, T4E);
354 T27 = VSUB(T25, T26);
355 T2c = VADD(T28, T2b);
356 T2d = VADD(T27, T2c);
357 T3F = VSUB(T2c, T27);
358 {
359 V T4H, T4K, T2k, T2l;
360 T4H = VSUB(Ty, TB);
361 T4K = VSUB(T4I, T4J);
362 T4L = VSUB(T4H, T4K);
363 T5g = VADD(T4H, T4K);
364 T2k = VADD(T26, T25);
365 T2l = VSUB(T28, T2b);
366 T2m = VADD(T2k, T2l);
367 T3I = VSUB(T2k, T2l);
368 }
369 }
370 }
371 {
372 V T61, T62, T63, T64, T65, T66, T67, T68, T69, T6a, T6b, T6c, T6d, T6e, T6f;
373 V T6g, T6h, T6i, T6j, T6k, T6l, T6m, T6n, T6o, T6p, T6q, T6r, T6s, T6t, T6u;
374 V T6v, T6w;
375 {
376 V T4B, T5b, T5a, T5c, T4Y, T56, T55, T57;
377 {
378 V T4t, T4A, T58, T59;
379 T4t = VSUB(T4r, T4s);
380 T4A = VSUB(T4w, T4z);
381 T4B = VFMA(LDK(KP707106781), T4A, T4t);
382 T5b = VFNMS(LDK(KP707106781), T4A, T4t);
383 T58 = VFMA(LDK(KP414213562), T4R, T4W);
384 T59 = VFNMS(LDK(KP414213562), T4G, T4L);
385 T5a = VSUB(T58, T59);
386 T5c = VADD(T59, T58);
387 }
388 {
389 V T4M, T4X, T51, T54;
390 T4M = VFMA(LDK(KP414213562), T4L, T4G);
391 T4X = VFNMS(LDK(KP414213562), T4W, T4R);
392 T4Y = VSUB(T4M, T4X);
393 T56 = VADD(T4M, T4X);
394 T51 = VSUB(T4Z, T50);
395 T54 = VSUB(T52, T53);
396 T55 = VFNMS(LDK(KP707106781), T54, T51);
397 T57 = VFMA(LDK(KP707106781), T54, T51);
398 }
399 T61 = VFNMS(LDK(KP923879532), T4Y, T4B);
400 STM4(&(ro[22]), T61, ovs, &(ro[0]));
401 T62 = VFNMS(LDK(KP923879532), T5a, T57);
402 STM4(&(io[22]), T62, ovs, &(io[0]));
403 T63 = VFMA(LDK(KP923879532), T4Y, T4B);
404 STM4(&(ro[6]), T63, ovs, &(ro[0]));
405 T64 = VFMA(LDK(KP923879532), T5a, T57);
406 STM4(&(io[6]), T64, ovs, &(io[0]));
407 T65 = VFNMS(LDK(KP923879532), T56, T55);
408 STM4(&(io[14]), T65, ovs, &(io[0]));
409 T66 = VFNMS(LDK(KP923879532), T5c, T5b);
410 STM4(&(ro[14]), T66, ovs, &(ro[0]));
411 T67 = VFMA(LDK(KP923879532), T56, T55);
412 STM4(&(io[30]), T67, ovs, &(io[0]));
413 T68 = VFMA(LDK(KP923879532), T5c, T5b);
414 STM4(&(ro[30]), T68, ovs, &(ro[0]));
415 }
416 {
417 V T5f, T5r, T5u, T5w, T5m, T5q, T5p, T5v;
418 {
419 V T5d, T5e, T5s, T5t;
420 T5d = VADD(T4r, T4s);
421 T5e = VADD(T53, T52);
422 T5f = VFMA(LDK(KP707106781), T5e, T5d);
423 T5r = VFNMS(LDK(KP707106781), T5e, T5d);
424 T5s = VFNMS(LDK(KP414213562), T5g, T5h);
425 T5t = VFMA(LDK(KP414213562), T5j, T5k);
426 T5u = VSUB(T5s, T5t);
427 T5w = VADD(T5s, T5t);
428 }
429 {
430 V T5i, T5l, T5n, T5o;
431 T5i = VFMA(LDK(KP414213562), T5h, T5g);
432 T5l = VFNMS(LDK(KP414213562), T5k, T5j);
433 T5m = VADD(T5i, T5l);
434 T5q = VSUB(T5l, T5i);
435 T5n = VADD(T50, T4Z);
436 T5o = VADD(T4w, T4z);
437 T5p = VFNMS(LDK(KP707106781), T5o, T5n);
438 T5v = VFMA(LDK(KP707106781), T5o, T5n);
439 }
440 T69 = VFNMS(LDK(KP923879532), T5m, T5f);
441 STM4(&(ro[18]), T69, ovs, &(ro[0]));
442 T6a = VFNMS(LDK(KP923879532), T5w, T5v);
443 STM4(&(io[18]), T6a, ovs, &(io[0]));
444 T6b = VFMA(LDK(KP923879532), T5m, T5f);
445 STM4(&(ro[2]), T6b, ovs, &(ro[0]));
446 T6c = VFMA(LDK(KP923879532), T5w, T5v);
447 STM4(&(io[2]), T6c, ovs, &(io[0]));
448 T6d = VFNMS(LDK(KP923879532), T5q, T5p);
449 STM4(&(io[26]), T6d, ovs, &(io[0]));
450 T6e = VFNMS(LDK(KP923879532), T5u, T5r);
451 STM4(&(ro[26]), T6e, ovs, &(ro[0]));
452 T6f = VFMA(LDK(KP923879532), T5q, T5p);
453 STM4(&(io[10]), T6f, ovs, &(io[0]));
454 T6g = VFMA(LDK(KP923879532), T5u, T5r);
455 STM4(&(ro[10]), T6g, ovs, &(ro[0]));
456 }
457 {
458 V T5z, T5P, T5S, T5U, T5K, T5O, T5N, T5T;
459 {
460 V T5x, T5y, T5Q, T5R;
461 T5x = VSUB(T7, Te);
462 T5y = VSUB(T1n, T1u);
463 T5z = VADD(T5x, T5y);
464 T5P = VSUB(T5x, T5y);
465 T5Q = VSUB(T5D, T5A);
466 T5R = VADD(T5F, T5I);
467 T5S = VSUB(T5Q, T5R);
468 T5U = VADD(T5Q, T5R);
469 }
470 {
471 V T5E, T5J, T5L, T5M;
472 T5E = VADD(T5A, T5D);
473 T5J = VSUB(T5F, T5I);
474 T5K = VADD(T5E, T5J);
475 T5O = VSUB(T5J, T5E);
476 T5L = VSUB(T18, T1f);
477 T5M = VSUB(Tt, Tm);
478 T5N = VSUB(T5L, T5M);
479 T5T = VADD(T5M, T5L);
480 }
481 T6h = VFNMS(LDK(KP707106781), T5K, T5z);
482 STM4(&(ro[20]), T6h, ovs, &(ro[0]));
483 T6i = VFNMS(LDK(KP707106781), T5U, T5T);
484 STM4(&(io[20]), T6i, ovs, &(io[0]));
485 T6j = VFMA(LDK(KP707106781), T5K, T5z);
486 STM4(&(ro[4]), T6j, ovs, &(ro[0]));
487 T6k = VFMA(LDK(KP707106781), T5U, T5T);
488 STM4(&(io[4]), T6k, ovs, &(io[0]));
489 T6l = VFNMS(LDK(KP707106781), T5O, T5N);
490 STM4(&(io[28]), T6l, ovs, &(io[0]));
491 T6m = VFNMS(LDK(KP707106781), T5S, T5P);
492 STM4(&(ro[28]), T6m, ovs, &(ro[0]));
493 T6n = VFMA(LDK(KP707106781), T5O, T5N);
494 STM4(&(io[12]), T6n, ovs, &(io[0]));
495 T6o = VFMA(LDK(KP707106781), T5S, T5P);
496 STM4(&(ro[12]), T6o, ovs, &(ro[0]));
497 }
498 {
499 V Tv, T5V, T5Y, T60, T10, T11, T1w, T5Z;
500 {
501 V Tf, Tu, T5W, T5X;
502 Tf = VADD(T7, Te);
503 Tu = VADD(Tm, Tt);
504 Tv = VADD(Tf, Tu);
505 T5V = VSUB(Tf, Tu);
506 T5W = VADD(T5B, T5C);
507 T5X = VADD(T5G, T5H);
508 T5Y = VSUB(T5W, T5X);
509 T60 = VADD(T5W, T5X);
510 }
511 {
512 V TK, TZ, T1g, T1v;
513 TK = VADD(TC, TJ);
514 TZ = VADD(TR, TY);
515 T10 = VADD(TK, TZ);
516 T11 = VSUB(TZ, TK);
517 T1g = VADD(T18, T1f);
518 T1v = VADD(T1n, T1u);
519 T1w = VSUB(T1g, T1v);
520 T5Z = VADD(T1g, T1v);
521 }
522 T6p = VSUB(Tv, T10);
523 STM4(&(ro[16]), T6p, ovs, &(ro[0]));
524 T6q = VSUB(T5Z, T60);
525 STM4(&(io[16]), T6q, ovs, &(io[0]));
526 T6r = VADD(Tv, T10);
527 STM4(&(ro[0]), T6r, ovs, &(ro[0]));
528 T6s = VADD(T5Z, T60);
529 STM4(&(io[0]), T6s, ovs, &(io[0]));
530 T6t = VADD(T11, T1w);
531 STM4(&(io[8]), T6t, ovs, &(io[0]));
532 T6u = VADD(T5V, T5Y);
533 STM4(&(ro[8]), T6u, ovs, &(ro[0]));
534 T6v = VSUB(T1w, T11);
535 STM4(&(io[24]), T6v, ovs, &(io[0]));
536 T6w = VSUB(T5V, T5Y);
537 STM4(&(ro[24]), T6w, ovs, &(ro[0]));
538 }
539 {
540 V T6x, T6y, T6z, T6A, T6B, T6C, T6D, T6E, T6F, T6G, T6H, T6I, T6J, T6K, T6L;
541 V T6M;
542 {
543 V T1X, T37, T31, T33, T2o, T35, T2P, T34;
544 {
545 V T1H, T1W, T2X, T30;
546 T1H = VFNMS(LDK(KP707106781), T1G, T1z);
547 T1W = VSUB(T1O, T1V);
548 T1X = VFMA(LDK(KP923879532), T1W, T1H);
549 T37 = VFNMS(LDK(KP923879532), T1W, T1H);
550 T2X = VFNMS(LDK(KP707106781), T2W, T2T);
551 T30 = VSUB(T2Y, T2Z);
552 T31 = VFNMS(LDK(KP923879532), T30, T2X);
553 T33 = VFMA(LDK(KP923879532), T30, T2X);
554 }
555 {
556 V T2e, T2n, T2F, T2O;
557 T2e = VFNMS(LDK(KP707106781), T2d, T22);
558 T2n = VFNMS(LDK(KP707106781), T2m, T2j);
559 T2o = VFMA(LDK(KP668178637), T2n, T2e);
560 T35 = VFNMS(LDK(KP668178637), T2e, T2n);
561 T2F = VFNMS(LDK(KP707106781), T2E, T2t);
562 T2O = VFNMS(LDK(KP707106781), T2N, T2K);
563 T2P = VFNMS(LDK(KP668178637), T2O, T2F);
564 T34 = VFMA(LDK(KP668178637), T2F, T2O);
565 }
566 {
567 V T2Q, T36, T32, T38;
568 T2Q = VSUB(T2o, T2P);
569 T6x = VFNMS(LDK(KP831469612), T2Q, T1X);
570 STM4(&(ro[21]), T6x, ovs, &(ro[1]));
571 T6y = VFMA(LDK(KP831469612), T2Q, T1X);
572 STM4(&(ro[5]), T6y, ovs, &(ro[1]));
573 T36 = VSUB(T34, T35);
574 T6z = VFNMS(LDK(KP831469612), T36, T33);
575 STM4(&(io[21]), T6z, ovs, &(io[1]));
576 T6A = VFMA(LDK(KP831469612), T36, T33);
577 STM4(&(io[5]), T6A, ovs, &(io[1]));
578 T32 = VADD(T2o, T2P);
579 T6B = VFNMS(LDK(KP831469612), T32, T31);
580 STM4(&(io[13]), T6B, ovs, &(io[1]));
581 T6C = VFMA(LDK(KP831469612), T32, T31);
582 STM4(&(io[29]), T6C, ovs, &(io[1]));
583 T38 = VADD(T35, T34);
584 T6D = VFNMS(LDK(KP831469612), T38, T37);
585 STM4(&(ro[13]), T6D, ovs, &(ro[1]));
586 T6E = VFMA(LDK(KP831469612), T38, T37);
587 STM4(&(ro[29]), T6E, ovs, &(ro[1]));
588 }
589 }
590 {
591 V T3D, T41, T3Z, T45, T3K, T42, T3R, T43;
592 {
593 V T3v, T3C, T3V, T3Y;
594 T3v = VFMA(LDK(KP707106781), T3u, T3t);
595 T3C = VSUB(T3y, T3B);
596 T3D = VFMA(LDK(KP923879532), T3C, T3v);
597 T41 = VFNMS(LDK(KP923879532), T3C, T3v);
598 T3V = VFMA(LDK(KP707106781), T3U, T3T);
599 T3Y = VSUB(T3W, T3X);
600 T3Z = VFNMS(LDK(KP923879532), T3Y, T3V);
601 T45 = VFMA(LDK(KP923879532), T3Y, T3V);
602 }
603 {
604 V T3G, T3J, T3N, T3Q;
605 T3G = VFNMS(LDK(KP707106781), T3F, T3E);
606 T3J = VFNMS(LDK(KP707106781), T3I, T3H);
607 T3K = VFMA(LDK(KP668178637), T3J, T3G);
608 T42 = VFNMS(LDK(KP668178637), T3G, T3J);
609 T3N = VFNMS(LDK(KP707106781), T3M, T3L);
610 T3Q = VFNMS(LDK(KP707106781), T3P, T3O);
611 T3R = VFNMS(LDK(KP668178637), T3Q, T3N);
612 T43 = VFMA(LDK(KP668178637), T3N, T3Q);
613 }
614 {
615 V T3S, T46, T40, T44;
616 T3S = VADD(T3K, T3R);
617 T6F = VFNMS(LDK(KP831469612), T3S, T3D);
618 STM4(&(ro[19]), T6F, ovs, &(ro[1]));
619 T6G = VFMA(LDK(KP831469612), T3S, T3D);
620 STM4(&(ro[3]), T6G, ovs, &(ro[1]));
621 T46 = VADD(T42, T43);
622 T6H = VFNMS(LDK(KP831469612), T46, T45);
623 STM4(&(io[19]), T6H, ovs, &(io[1]));
624 T6I = VFMA(LDK(KP831469612), T46, T45);
625 STM4(&(io[3]), T6I, ovs, &(io[1]));
626 T40 = VSUB(T3R, T3K);
627 T6J = VFNMS(LDK(KP831469612), T40, T3Z);
628 STM4(&(io[27]), T6J, ovs, &(io[1]));
629 T6K = VFMA(LDK(KP831469612), T40, T3Z);
630 STM4(&(io[11]), T6K, ovs, &(io[1]));
631 T44 = VSUB(T42, T43);
632 T6L = VFNMS(LDK(KP831469612), T44, T41);
633 STM4(&(ro[27]), T6L, ovs, &(ro[1]));
634 T6M = VFMA(LDK(KP831469612), T44, T41);
635 STM4(&(ro[11]), T6M, ovs, &(ro[1]));
636 }
637 }
638 {
639 V T49, T4p, T4j, T4l, T4c, T4n, T4f, T4m;
640 {
641 V T47, T48, T4h, T4i;
642 T47 = VFNMS(LDK(KP707106781), T3u, T3t);
643 T48 = VADD(T3X, T3W);
644 T49 = VFNMS(LDK(KP923879532), T48, T47);
645 T4p = VFMA(LDK(KP923879532), T48, T47);
646 T4h = VFNMS(LDK(KP707106781), T3U, T3T);
647 T4i = VADD(T3y, T3B);
648 T4j = VFMA(LDK(KP923879532), T4i, T4h);
649 T4l = VFNMS(LDK(KP923879532), T4i, T4h);
650 }
651 {
652 V T4a, T4b, T4d, T4e;
653 T4a = VFMA(LDK(KP707106781), T3I, T3H);
654 T4b = VFMA(LDK(KP707106781), T3F, T3E);
655 T4c = VFMA(LDK(KP198912367), T4b, T4a);
656 T4n = VFNMS(LDK(KP198912367), T4a, T4b);
657 T4d = VFMA(LDK(KP707106781), T3P, T3O);
658 T4e = VFMA(LDK(KP707106781), T3M, T3L);
659 T4f = VFNMS(LDK(KP198912367), T4e, T4d);
660 T4m = VFMA(LDK(KP198912367), T4d, T4e);
661 }
662 {
663 V T4g, T6N, T6O, T4o, T6P, T6Q;
664 T4g = VSUB(T4c, T4f);
665 T6N = VFNMS(LDK(KP980785280), T4g, T49);
666 STM4(&(ro[23]), T6N, ovs, &(ro[1]));
667 STN4(&(ro[20]), T6h, T6x, T61, T6N, ovs);
668 T6O = VFMA(LDK(KP980785280), T4g, T49);
669 STM4(&(ro[7]), T6O, ovs, &(ro[1]));
670 STN4(&(ro[4]), T6j, T6y, T63, T6O, ovs);
671 T4o = VSUB(T4m, T4n);
672 T6P = VFNMS(LDK(KP980785280), T4o, T4l);
673 STM4(&(io[23]), T6P, ovs, &(io[1]));
674 STN4(&(io[20]), T6i, T6z, T62, T6P, ovs);
675 T6Q = VFMA(LDK(KP980785280), T4o, T4l);
676 STM4(&(io[7]), T6Q, ovs, &(io[1]));
677 STN4(&(io[4]), T6k, T6A, T64, T6Q, ovs);
678 }
679 {
680 V T4k, T6R, T6S, T4q, T6T, T6U;
681 T4k = VADD(T4c, T4f);
682 T6R = VFNMS(LDK(KP980785280), T4k, T4j);
683 STM4(&(io[15]), T6R, ovs, &(io[1]));
684 STN4(&(io[12]), T6n, T6B, T65, T6R, ovs);
685 T6S = VFMA(LDK(KP980785280), T4k, T4j);
686 STM4(&(io[31]), T6S, ovs, &(io[1]));
687 STN4(&(io[28]), T6l, T6C, T67, T6S, ovs);
688 T4q = VADD(T4n, T4m);
689 T6T = VFNMS(LDK(KP980785280), T4q, T4p);
690 STM4(&(ro[15]), T6T, ovs, &(ro[1]));
691 STN4(&(ro[12]), T6o, T6D, T66, T6T, ovs);
692 T6U = VFMA(LDK(KP980785280), T4q, T4p);
693 STM4(&(ro[31]), T6U, ovs, &(ro[1]));
694 STN4(&(ro[28]), T6m, T6E, T68, T6U, ovs);
695 }
696 }
697 {
698 V T3b, T3n, T3l, T3r, T3e, T3o, T3h, T3p;
699 {
700 V T39, T3a, T3j, T3k;
701 T39 = VFMA(LDK(KP707106781), T1G, T1z);
702 T3a = VADD(T2Z, T2Y);
703 T3b = VFMA(LDK(KP923879532), T3a, T39);
704 T3n = VFNMS(LDK(KP923879532), T3a, T39);
705 T3j = VFMA(LDK(KP707106781), T2W, T2T);
706 T3k = VADD(T1O, T1V);
707 T3l = VFNMS(LDK(KP923879532), T3k, T3j);
708 T3r = VFMA(LDK(KP923879532), T3k, T3j);
709 }
710 {
711 V T3c, T3d, T3f, T3g;
712 T3c = VFMA(LDK(KP707106781), T2m, T2j);
713 T3d = VFMA(LDK(KP707106781), T2d, T22);
714 T3e = VFMA(LDK(KP198912367), T3d, T3c);
715 T3o = VFNMS(LDK(KP198912367), T3c, T3d);
716 T3f = VFMA(LDK(KP707106781), T2N, T2K);
717 T3g = VFMA(LDK(KP707106781), T2E, T2t);
718 T3h = VFNMS(LDK(KP198912367), T3g, T3f);
719 T3p = VFMA(LDK(KP198912367), T3f, T3g);
720 }
721 {
722 V T3i, T6V, T6W, T3s, T6X, T6Y;
723 T3i = VADD(T3e, T3h);
724 T6V = VFNMS(LDK(KP980785280), T3i, T3b);
725 STM4(&(ro[17]), T6V, ovs, &(ro[1]));
726 STN4(&(ro[16]), T6p, T6V, T69, T6F, ovs);
727 T6W = VFMA(LDK(KP980785280), T3i, T3b);
728 STM4(&(ro[1]), T6W, ovs, &(ro[1]));
729 STN4(&(ro[0]), T6r, T6W, T6b, T6G, ovs);
730 T3s = VADD(T3o, T3p);
731 T6X = VFNMS(LDK(KP980785280), T3s, T3r);
732 STM4(&(io[17]), T6X, ovs, &(io[1]));
733 STN4(&(io[16]), T6q, T6X, T6a, T6H, ovs);
734 T6Y = VFMA(LDK(KP980785280), T3s, T3r);
735 STM4(&(io[1]), T6Y, ovs, &(io[1]));
736 STN4(&(io[0]), T6s, T6Y, T6c, T6I, ovs);
737 }
738 {
739 V T3m, T6Z, T70, T3q, T71, T72;
740 T3m = VSUB(T3h, T3e);
741 T6Z = VFNMS(LDK(KP980785280), T3m, T3l);
742 STM4(&(io[25]), T6Z, ovs, &(io[1]));
743 STN4(&(io[24]), T6v, T6Z, T6d, T6J, ovs);
744 T70 = VFMA(LDK(KP980785280), T3m, T3l);
745 STM4(&(io[9]), T70, ovs, &(io[1]));
746 STN4(&(io[8]), T6t, T70, T6f, T6K, ovs);
747 T3q = VSUB(T3o, T3p);
748 T71 = VFNMS(LDK(KP980785280), T3q, T3n);
749 STM4(&(ro[25]), T71, ovs, &(ro[1]));
750 STN4(&(ro[24]), T6w, T71, T6e, T6L, ovs);
751 T72 = VFMA(LDK(KP980785280), T3q, T3n);
752 STM4(&(ro[9]), T72, ovs, &(ro[1]));
753 STN4(&(ro[8]), T6u, T72, T6g, T6M, ovs);
754 }
755 }
756 }
757 }
758 }
759 }
760 VLEAVE();
761 }
762
763 static const kdft_desc desc = { 32, XSIMD_STRING("n2sv_32"), {236, 0, 136, 0}, &GENUS, 0, 1, 0, 0 };
764
765 void XSIMD(codelet_n2sv_32) (planner *p) {
766 X(kdft_register) (p, n2sv_32, &desc);
767 }
768
769 #else
770
771 /* Generated by: ../../../genfft/gen_notw.native -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n2sv_32 -with-ostride 1 -include dft/simd/n2s.h -store-multiple 4 */
772
773 /*
774 * This function contains 372 FP additions, 84 FP multiplications,
775 * (or, 340 additions, 52 multiplications, 32 fused multiply/add),
776 * 130 stack variables, 7 constants, and 144 memory accesses
777 */
778 #include "dft/simd/n2s.h"
779
780 static void n2sv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
781 {
782 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
783 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
784 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
785 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
786 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
787 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
788 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
789 {
790 INT i;
791 for (i = v; i > 0; i = i - (2 * VL), ri = ri + ((2 * VL) * ivs), ii = ii + ((2 * VL) * ivs), ro = ro + ((2 * VL) * ovs), io = io + ((2 * VL) * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) {
792 V T7, T4r, T4Z, T18, T1z, T3t, T3T, T2T, Te, T1f, T50, T4s, T2W, T3u, T1G;
793 V T3U, Tm, T1n, T1O, T2Z, T3y, T3X, T4w, T53, Tt, T1u, T1V, T2Y, T3B, T3W;
794 V T4z, T52, T2t, T3L, T3O, T2K, TR, TY, T5F, T5G, T5H, T5I, T4R, T5j, T2E;
795 V T3P, T4W, T5k, T2N, T3M, T22, T3E, T3H, T2j, TC, TJ, T5A, T5B, T5C, T5D;
796 V T4G, T5g, T2d, T3F, T4L, T5h, T2m, T3I;
797 {
798 V T3, T1x, T14, T2S, T6, T2R, T17, T1y;
799 {
800 V T1, T2, T12, T13;
801 T1 = LD(&(ri[0]), ivs, &(ri[0]));
802 T2 = LD(&(ri[WS(is, 16)]), ivs, &(ri[0]));
803 T3 = VADD(T1, T2);
804 T1x = VSUB(T1, T2);
805 T12 = LD(&(ii[0]), ivs, &(ii[0]));
806 T13 = LD(&(ii[WS(is, 16)]), ivs, &(ii[0]));
807 T14 = VADD(T12, T13);
808 T2S = VSUB(T12, T13);
809 }
810 {
811 V T4, T5, T15, T16;
812 T4 = LD(&(ri[WS(is, 8)]), ivs, &(ri[0]));
813 T5 = LD(&(ri[WS(is, 24)]), ivs, &(ri[0]));
814 T6 = VADD(T4, T5);
815 T2R = VSUB(T4, T5);
816 T15 = LD(&(ii[WS(is, 8)]), ivs, &(ii[0]));
817 T16 = LD(&(ii[WS(is, 24)]), ivs, &(ii[0]));
818 T17 = VADD(T15, T16);
819 T1y = VSUB(T15, T16);
820 }
821 T7 = VADD(T3, T6);
822 T4r = VSUB(T3, T6);
823 T4Z = VSUB(T14, T17);
824 T18 = VADD(T14, T17);
825 T1z = VSUB(T1x, T1y);
826 T3t = VADD(T1x, T1y);
827 T3T = VSUB(T2S, T2R);
828 T2T = VADD(T2R, T2S);
829 }
830 {
831 V Ta, T1B, T1b, T1A, Td, T1D, T1e, T1E;
832 {
833 V T8, T9, T19, T1a;
834 T8 = LD(&(ri[WS(is, 4)]), ivs, &(ri[0]));
835 T9 = LD(&(ri[WS(is, 20)]), ivs, &(ri[0]));
836 Ta = VADD(T8, T9);
837 T1B = VSUB(T8, T9);
838 T19 = LD(&(ii[WS(is, 4)]), ivs, &(ii[0]));
839 T1a = LD(&(ii[WS(is, 20)]), ivs, &(ii[0]));
840 T1b = VADD(T19, T1a);
841 T1A = VSUB(T19, T1a);
842 }
843 {
844 V Tb, Tc, T1c, T1d;
845 Tb = LD(&(ri[WS(is, 28)]), ivs, &(ri[0]));
846 Tc = LD(&(ri[WS(is, 12)]), ivs, &(ri[0]));
847 Td = VADD(Tb, Tc);
848 T1D = VSUB(Tb, Tc);
849 T1c = LD(&(ii[WS(is, 28)]), ivs, &(ii[0]));
850 T1d = LD(&(ii[WS(is, 12)]), ivs, &(ii[0]));
851 T1e = VADD(T1c, T1d);
852 T1E = VSUB(T1c, T1d);
853 }
854 Te = VADD(Ta, Td);
855 T1f = VADD(T1b, T1e);
856 T50 = VSUB(Td, Ta);
857 T4s = VSUB(T1b, T1e);
858 {
859 V T2U, T2V, T1C, T1F;
860 T2U = VSUB(T1D, T1E);
861 T2V = VADD(T1B, T1A);
862 T2W = VMUL(LDK(KP707106781), VSUB(T2U, T2V));
863 T3u = VMUL(LDK(KP707106781), VADD(T2V, T2U));
864 T1C = VSUB(T1A, T1B);
865 T1F = VADD(T1D, T1E);
866 T1G = VMUL(LDK(KP707106781), VSUB(T1C, T1F));
867 T3U = VMUL(LDK(KP707106781), VADD(T1C, T1F));
868 }
869 }
870 {
871 V Ti, T1L, T1j, T1J, Tl, T1I, T1m, T1M, T1K, T1N;
872 {
873 V Tg, Th, T1h, T1i;
874 Tg = LD(&(ri[WS(is, 2)]), ivs, &(ri[0]));
875 Th = LD(&(ri[WS(is, 18)]), ivs, &(ri[0]));
876 Ti = VADD(Tg, Th);
877 T1L = VSUB(Tg, Th);
878 T1h = LD(&(ii[WS(is, 2)]), ivs, &(ii[0]));
879 T1i = LD(&(ii[WS(is, 18)]), ivs, &(ii[0]));
880 T1j = VADD(T1h, T1i);
881 T1J = VSUB(T1h, T1i);
882 }
883 {
884 V Tj, Tk, T1k, T1l;
885 Tj = LD(&(ri[WS(is, 10)]), ivs, &(ri[0]));
886 Tk = LD(&(ri[WS(is, 26)]), ivs, &(ri[0]));
887 Tl = VADD(Tj, Tk);
888 T1I = VSUB(Tj, Tk);
889 T1k = LD(&(ii[WS(is, 10)]), ivs, &(ii[0]));
890 T1l = LD(&(ii[WS(is, 26)]), ivs, &(ii[0]));
891 T1m = VADD(T1k, T1l);
892 T1M = VSUB(T1k, T1l);
893 }
894 Tm = VADD(Ti, Tl);
895 T1n = VADD(T1j, T1m);
896 T1K = VADD(T1I, T1J);
897 T1N = VSUB(T1L, T1M);
898 T1O = VFNMS(LDK(KP923879532), T1N, VMUL(LDK(KP382683432), T1K));
899 T2Z = VFMA(LDK(KP923879532), T1K, VMUL(LDK(KP382683432), T1N));
900 {
901 V T3w, T3x, T4u, T4v;
902 T3w = VSUB(T1J, T1I);
903 T3x = VADD(T1L, T1M);
904 T3y = VFNMS(LDK(KP382683432), T3x, VMUL(LDK(KP923879532), T3w));
905 T3X = VFMA(LDK(KP382683432), T3w, VMUL(LDK(KP923879532), T3x));
906 T4u = VSUB(T1j, T1m);
907 T4v = VSUB(Ti, Tl);
908 T4w = VSUB(T4u, T4v);
909 T53 = VADD(T4v, T4u);
910 }
911 }
912 {
913 V Tp, T1S, T1q, T1Q, Ts, T1P, T1t, T1T, T1R, T1U;
914 {
915 V Tn, To, T1o, T1p;
916 Tn = LD(&(ri[WS(is, 30)]), ivs, &(ri[0]));
917 To = LD(&(ri[WS(is, 14)]), ivs, &(ri[0]));
918 Tp = VADD(Tn, To);
919 T1S = VSUB(Tn, To);
920 T1o = LD(&(ii[WS(is, 30)]), ivs, &(ii[0]));
921 T1p = LD(&(ii[WS(is, 14)]), ivs, &(ii[0]));
922 T1q = VADD(T1o, T1p);
923 T1Q = VSUB(T1o, T1p);
924 }
925 {
926 V Tq, Tr, T1r, T1s;
927 Tq = LD(&(ri[WS(is, 6)]), ivs, &(ri[0]));
928 Tr = LD(&(ri[WS(is, 22)]), ivs, &(ri[0]));
929 Ts = VADD(Tq, Tr);
930 T1P = VSUB(Tq, Tr);
931 T1r = LD(&(ii[WS(is, 6)]), ivs, &(ii[0]));
932 T1s = LD(&(ii[WS(is, 22)]), ivs, &(ii[0]));
933 T1t = VADD(T1r, T1s);
934 T1T = VSUB(T1r, T1s);
935 }
936 Tt = VADD(Tp, Ts);
937 T1u = VADD(T1q, T1t);
938 T1R = VADD(T1P, T1Q);
939 T1U = VSUB(T1S, T1T);
940 T1V = VFMA(LDK(KP382683432), T1R, VMUL(LDK(KP923879532), T1U));
941 T2Y = VFNMS(LDK(KP923879532), T1R, VMUL(LDK(KP382683432), T1U));
942 {
943 V T3z, T3A, T4x, T4y;
944 T3z = VSUB(T1Q, T1P);
945 T3A = VADD(T1S, T1T);
946 T3B = VFMA(LDK(KP923879532), T3z, VMUL(LDK(KP382683432), T3A));
947 T3W = VFNMS(LDK(KP382683432), T3z, VMUL(LDK(KP923879532), T3A));
948 T4x = VSUB(Tp, Ts);
949 T4y = VSUB(T1q, T1t);
950 T4z = VADD(T4x, T4y);
951 T52 = VSUB(T4x, T4y);
952 }
953 }
954 {
955 V TN, T2p, T2J, T4S, TQ, T2G, T2s, T4T, TU, T2x, T2w, T4O, TX, T2z, T2C;
956 V T4P;
957 {
958 V TL, TM, T2H, T2I;
959 TL = LD(&(ri[WS(is, 31)]), ivs, &(ri[WS(is, 1)]));
960 TM = LD(&(ri[WS(is, 15)]), ivs, &(ri[WS(is, 1)]));
961 TN = VADD(TL, TM);
962 T2p = VSUB(TL, TM);
963 T2H = LD(&(ii[WS(is, 31)]), ivs, &(ii[WS(is, 1)]));
964 T2I = LD(&(ii[WS(is, 15)]), ivs, &(ii[WS(is, 1)]));
965 T2J = VSUB(T2H, T2I);
966 T4S = VADD(T2H, T2I);
967 }
968 {
969 V TO, TP, T2q, T2r;
970 TO = LD(&(ri[WS(is, 7)]), ivs, &(ri[WS(is, 1)]));
971 TP = LD(&(ri[WS(is, 23)]), ivs, &(ri[WS(is, 1)]));
972 TQ = VADD(TO, TP);
973 T2G = VSUB(TO, TP);
974 T2q = LD(&(ii[WS(is, 7)]), ivs, &(ii[WS(is, 1)]));
975 T2r = LD(&(ii[WS(is, 23)]), ivs, &(ii[WS(is, 1)]));
976 T2s = VSUB(T2q, T2r);
977 T4T = VADD(T2q, T2r);
978 }
979 {
980 V TS, TT, T2u, T2v;
981 TS = LD(&(ri[WS(is, 3)]), ivs, &(ri[WS(is, 1)]));
982 TT = LD(&(ri[WS(is, 19)]), ivs, &(ri[WS(is, 1)]));
983 TU = VADD(TS, TT);
984 T2x = VSUB(TS, TT);
985 T2u = LD(&(ii[WS(is, 3)]), ivs, &(ii[WS(is, 1)]));
986 T2v = LD(&(ii[WS(is, 19)]), ivs, &(ii[WS(is, 1)]));
987 T2w = VSUB(T2u, T2v);
988 T4O = VADD(T2u, T2v);
989 }
990 {
991 V TV, TW, T2A, T2B;
992 TV = LD(&(ri[WS(is, 27)]), ivs, &(ri[WS(is, 1)]));
993 TW = LD(&(ri[WS(is, 11)]), ivs, &(ri[WS(is, 1)]));
994 TX = VADD(TV, TW);
995 T2z = VSUB(TV, TW);
996 T2A = LD(&(ii[WS(is, 27)]), ivs, &(ii[WS(is, 1)]));
997 T2B = LD(&(ii[WS(is, 11)]), ivs, &(ii[WS(is, 1)]));
998 T2C = VSUB(T2A, T2B);
999 T4P = VADD(T2A, T2B);
1000 }
1001 T2t = VSUB(T2p, T2s);
1002 T3L = VADD(T2p, T2s);
1003 T3O = VSUB(T2J, T2G);
1004 T2K = VADD(T2G, T2J);
1005 TR = VADD(TN, TQ);
1006 TY = VADD(TU, TX);
1007 T5F = VSUB(TR, TY);
1008 {
1009 V T4N, T4Q, T2y, T2D;
1010 T5G = VADD(T4S, T4T);
1011 T5H = VADD(T4O, T4P);
1012 T5I = VSUB(T5G, T5H);
1013 T4N = VSUB(TN, TQ);
1014 T4Q = VSUB(T4O, T4P);
1015 T4R = VSUB(T4N, T4Q);
1016 T5j = VADD(T4N, T4Q);
1017 T2y = VSUB(T2w, T2x);
1018 T2D = VADD(T2z, T2C);
1019 T2E = VMUL(LDK(KP707106781), VSUB(T2y, T2D));
1020 T3P = VMUL(LDK(KP707106781), VADD(T2y, T2D));
1021 {
1022 V T4U, T4V, T2L, T2M;
1023 T4U = VSUB(T4S, T4T);
1024 T4V = VSUB(TX, TU);
1025 T4W = VSUB(T4U, T4V);
1026 T5k = VADD(T4V, T4U);
1027 T2L = VSUB(T2z, T2C);
1028 T2M = VADD(T2x, T2w);
1029 T2N = VMUL(LDK(KP707106781), VSUB(T2L, T2M));
1030 T3M = VMUL(LDK(KP707106781), VADD(T2M, T2L));
1031 }
1032 }
1033 }
1034 {
1035 V Ty, T2f, T21, T4C, TB, T1Y, T2i, T4D, TF, T28, T2b, T4I, TI, T23, T26;
1036 V T4J;
1037 {
1038 V Tw, Tx, T1Z, T20;
1039 Tw = LD(&(ri[WS(is, 1)]), ivs, &(ri[WS(is, 1)]));
1040 Tx = LD(&(ri[WS(is, 17)]), ivs, &(ri[WS(is, 1)]));
1041 Ty = VADD(Tw, Tx);
1042 T2f = VSUB(Tw, Tx);
1043 T1Z = LD(&(ii[WS(is, 1)]), ivs, &(ii[WS(is, 1)]));
1044 T20 = LD(&(ii[WS(is, 17)]), ivs, &(ii[WS(is, 1)]));
1045 T21 = VSUB(T1Z, T20);
1046 T4C = VADD(T1Z, T20);
1047 }
1048 {
1049 V Tz, TA, T2g, T2h;
1050 Tz = LD(&(ri[WS(is, 9)]), ivs, &(ri[WS(is, 1)]));
1051 TA = LD(&(ri[WS(is, 25)]), ivs, &(ri[WS(is, 1)]));
1052 TB = VADD(Tz, TA);
1053 T1Y = VSUB(Tz, TA);
1054 T2g = LD(&(ii[WS(is, 9)]), ivs, &(ii[WS(is, 1)]));
1055 T2h = LD(&(ii[WS(is, 25)]), ivs, &(ii[WS(is, 1)]));
1056 T2i = VSUB(T2g, T2h);
1057 T4D = VADD(T2g, T2h);
1058 }
1059 {
1060 V TD, TE, T29, T2a;
1061 TD = LD(&(ri[WS(is, 5)]), ivs, &(ri[WS(is, 1)]));
1062 TE = LD(&(ri[WS(is, 21)]), ivs, &(ri[WS(is, 1)]));
1063 TF = VADD(TD, TE);
1064 T28 = VSUB(TD, TE);
1065 T29 = LD(&(ii[WS(is, 5)]), ivs, &(ii[WS(is, 1)]));
1066 T2a = LD(&(ii[WS(is, 21)]), ivs, &(ii[WS(is, 1)]));
1067 T2b = VSUB(T29, T2a);
1068 T4I = VADD(T29, T2a);
1069 }
1070 {
1071 V TG, TH, T24, T25;
1072 TG = LD(&(ri[WS(is, 29)]), ivs, &(ri[WS(is, 1)]));
1073 TH = LD(&(ri[WS(is, 13)]), ivs, &(ri[WS(is, 1)]));
1074 TI = VADD(TG, TH);
1075 T23 = VSUB(TG, TH);
1076 T24 = LD(&(ii[WS(is, 29)]), ivs, &(ii[WS(is, 1)]));
1077 T25 = LD(&(ii[WS(is, 13)]), ivs, &(ii[WS(is, 1)]));
1078 T26 = VSUB(T24, T25);
1079 T4J = VADD(T24, T25);
1080 }
1081 T22 = VADD(T1Y, T21);
1082 T3E = VADD(T2f, T2i);
1083 T3H = VSUB(T21, T1Y);
1084 T2j = VSUB(T2f, T2i);
1085 TC = VADD(Ty, TB);
1086 TJ = VADD(TF, TI);
1087 T5A = VSUB(TC, TJ);
1088 {
1089 V T4E, T4F, T27, T2c;
1090 T5B = VADD(T4C, T4D);
1091 T5C = VADD(T4I, T4J);
1092 T5D = VSUB(T5B, T5C);
1093 T4E = VSUB(T4C, T4D);
1094 T4F = VSUB(TI, TF);
1095 T4G = VSUB(T4E, T4F);
1096 T5g = VADD(T4F, T4E);
1097 T27 = VSUB(T23, T26);
1098 T2c = VADD(T28, T2b);
1099 T2d = VMUL(LDK(KP707106781), VSUB(T27, T2c));
1100 T3F = VMUL(LDK(KP707106781), VADD(T2c, T27));
1101 {
1102 V T4H, T4K, T2k, T2l;
1103 T4H = VSUB(Ty, TB);
1104 T4K = VSUB(T4I, T4J);
1105 T4L = VSUB(T4H, T4K);
1106 T5h = VADD(T4H, T4K);
1107 T2k = VSUB(T2b, T28);
1108 T2l = VADD(T23, T26);
1109 T2m = VMUL(LDK(KP707106781), VSUB(T2k, T2l));
1110 T3I = VMUL(LDK(KP707106781), VADD(T2k, T2l));
1111 }
1112 }
1113 }
1114 {
1115 V T61, T62, T63, T64, T65, T66, T67, T68, T69, T6a, T6b, T6c, T6d, T6e, T6f;
1116 V T6g, T6h, T6i, T6j, T6k, T6l, T6m, T6n, T6o, T6p, T6q, T6r, T6s, T6t, T6u;
1117 V T6v, T6w;
1118 {
1119 V T4B, T57, T5a, T5c, T4Y, T56, T55, T5b;
1120 {
1121 V T4t, T4A, T58, T59;
1122 T4t = VSUB(T4r, T4s);
1123 T4A = VMUL(LDK(KP707106781), VSUB(T4w, T4z));
1124 T4B = VADD(T4t, T4A);
1125 T57 = VSUB(T4t, T4A);
1126 T58 = VFNMS(LDK(KP923879532), T4L, VMUL(LDK(KP382683432), T4G));
1127 T59 = VFMA(LDK(KP382683432), T4W, VMUL(LDK(KP923879532), T4R));
1128 T5a = VSUB(T58, T59);
1129 T5c = VADD(T58, T59);
1130 }
1131 {
1132 V T4M, T4X, T51, T54;
1133 T4M = VFMA(LDK(KP923879532), T4G, VMUL(LDK(KP382683432), T4L));
1134 T4X = VFNMS(LDK(KP923879532), T4W, VMUL(LDK(KP382683432), T4R));
1135 T4Y = VADD(T4M, T4X);
1136 T56 = VSUB(T4X, T4M);
1137 T51 = VSUB(T4Z, T50);
1138 T54 = VMUL(LDK(KP707106781), VSUB(T52, T53));
1139 T55 = VSUB(T51, T54);
1140 T5b = VADD(T51, T54);
1141 }
1142 T61 = VSUB(T4B, T4Y);
1143 STM4(&(ro[22]), T61, ovs, &(ro[0]));
1144 T62 = VSUB(T5b, T5c);
1145 STM4(&(io[22]), T62, ovs, &(io[0]));
1146 T63 = VADD(T4B, T4Y);
1147 STM4(&(ro[6]), T63, ovs, &(ro[0]));
1148 T64 = VADD(T5b, T5c);
1149 STM4(&(io[6]), T64, ovs, &(io[0]));
1150 T65 = VSUB(T55, T56);
1151 STM4(&(io[30]), T65, ovs, &(io[0]));
1152 T66 = VSUB(T57, T5a);
1153 STM4(&(ro[30]), T66, ovs, &(ro[0]));
1154 T67 = VADD(T55, T56);
1155 STM4(&(io[14]), T67, ovs, &(io[0]));
1156 T68 = VADD(T57, T5a);
1157 STM4(&(ro[14]), T68, ovs, &(ro[0]));
1158 }
1159 {
1160 V T5f, T5r, T5u, T5w, T5m, T5q, T5p, T5v;
1161 {
1162 V T5d, T5e, T5s, T5t;
1163 T5d = VADD(T4r, T4s);
1164 T5e = VMUL(LDK(KP707106781), VADD(T53, T52));
1165 T5f = VADD(T5d, T5e);
1166 T5r = VSUB(T5d, T5e);
1167 T5s = VFNMS(LDK(KP382683432), T5h, VMUL(LDK(KP923879532), T5g));
1168 T5t = VFMA(LDK(KP923879532), T5k, VMUL(LDK(KP382683432), T5j));
1169 T5u = VSUB(T5s, T5t);
1170 T5w = VADD(T5s, T5t);
1171 }
1172 {
1173 V T5i, T5l, T5n, T5o;
1174 T5i = VFMA(LDK(KP382683432), T5g, VMUL(LDK(KP923879532), T5h));
1175 T5l = VFNMS(LDK(KP382683432), T5k, VMUL(LDK(KP923879532), T5j));
1176 T5m = VADD(T5i, T5l);
1177 T5q = VSUB(T5l, T5i);
1178 T5n = VADD(T50, T4Z);
1179 T5o = VMUL(LDK(KP707106781), VADD(T4w, T4z));
1180 T5p = VSUB(T5n, T5o);
1181 T5v = VADD(T5n, T5o);
1182 }
1183 T69 = VSUB(T5f, T5m);
1184 STM4(&(ro[18]), T69, ovs, &(ro[0]));
1185 T6a = VSUB(T5v, T5w);
1186 STM4(&(io[18]), T6a, ovs, &(io[0]));
1187 T6b = VADD(T5f, T5m);
1188 STM4(&(ro[2]), T6b, ovs, &(ro[0]));
1189 T6c = VADD(T5v, T5w);
1190 STM4(&(io[2]), T6c, ovs, &(io[0]));
1191 T6d = VSUB(T5p, T5q);
1192 STM4(&(io[26]), T6d, ovs, &(io[0]));
1193 T6e = VSUB(T5r, T5u);
1194 STM4(&(ro[26]), T6e, ovs, &(ro[0]));
1195 T6f = VADD(T5p, T5q);
1196 STM4(&(io[10]), T6f, ovs, &(io[0]));
1197 T6g = VADD(T5r, T5u);
1198 STM4(&(ro[10]), T6g, ovs, &(ro[0]));
1199 }
1200 {
1201 V T5z, T5P, T5S, T5U, T5K, T5O, T5N, T5T;
1202 {
1203 V T5x, T5y, T5Q, T5R;
1204 T5x = VSUB(T7, Te);
1205 T5y = VSUB(T1n, T1u);
1206 T5z = VADD(T5x, T5y);
1207 T5P = VSUB(T5x, T5y);
1208 T5Q = VSUB(T5D, T5A);
1209 T5R = VADD(T5F, T5I);
1210 T5S = VMUL(LDK(KP707106781), VSUB(T5Q, T5R));
1211 T5U = VMUL(LDK(KP707106781), VADD(T5Q, T5R));
1212 }
1213 {
1214 V T5E, T5J, T5L, T5M;
1215 T5E = VADD(T5A, T5D);
1216 T5J = VSUB(T5F, T5I);
1217 T5K = VMUL(LDK(KP707106781), VADD(T5E, T5J));
1218 T5O = VMUL(LDK(KP707106781), VSUB(T5J, T5E));
1219 T5L = VSUB(T18, T1f);
1220 T5M = VSUB(Tt, Tm);
1221 T5N = VSUB(T5L, T5M);
1222 T5T = VADD(T5M, T5L);
1223 }
1224 T6h = VSUB(T5z, T5K);
1225 STM4(&(ro[20]), T6h, ovs, &(ro[0]));
1226 T6i = VSUB(T5T, T5U);
1227 STM4(&(io[20]), T6i, ovs, &(io[0]));
1228 T6j = VADD(T5z, T5K);
1229 STM4(&(ro[4]), T6j, ovs, &(ro[0]));
1230 T6k = VADD(T5T, T5U);
1231 STM4(&(io[4]), T6k, ovs, &(io[0]));
1232 T6l = VSUB(T5N, T5O);
1233 STM4(&(io[28]), T6l, ovs, &(io[0]));
1234 T6m = VSUB(T5P, T5S);
1235 STM4(&(ro[28]), T6m, ovs, &(ro[0]));
1236 T6n = VADD(T5N, T5O);
1237 STM4(&(io[12]), T6n, ovs, &(io[0]));
1238 T6o = VADD(T5P, T5S);
1239 STM4(&(ro[12]), T6o, ovs, &(ro[0]));
1240 }
1241 {
1242 V Tv, T5V, T5Y, T60, T10, T11, T1w, T5Z;
1243 {
1244 V Tf, Tu, T5W, T5X;
1245 Tf = VADD(T7, Te);
1246 Tu = VADD(Tm, Tt);
1247 Tv = VADD(Tf, Tu);
1248 T5V = VSUB(Tf, Tu);
1249 T5W = VADD(T5B, T5C);
1250 T5X = VADD(T5G, T5H);
1251 T5Y = VSUB(T5W, T5X);
1252 T60 = VADD(T5W, T5X);
1253 }
1254 {
1255 V TK, TZ, T1g, T1v;
1256 TK = VADD(TC, TJ);
1257 TZ = VADD(TR, TY);
1258 T10 = VADD(TK, TZ);
1259 T11 = VSUB(TZ, TK);
1260 T1g = VADD(T18, T1f);
1261 T1v = VADD(T1n, T1u);
1262 T1w = VSUB(T1g, T1v);
1263 T5Z = VADD(T1g, T1v);
1264 }
1265 T6p = VSUB(Tv, T10);
1266 STM4(&(ro[16]), T6p, ovs, &(ro[0]));
1267 T6q = VSUB(T5Z, T60);
1268 STM4(&(io[16]), T6q, ovs, &(io[0]));
1269 T6r = VADD(Tv, T10);
1270 STM4(&(ro[0]), T6r, ovs, &(ro[0]));
1271 T6s = VADD(T5Z, T60);
1272 STM4(&(io[0]), T6s, ovs, &(io[0]));
1273 T6t = VADD(T11, T1w);
1274 STM4(&(io[8]), T6t, ovs, &(io[0]));
1275 T6u = VADD(T5V, T5Y);
1276 STM4(&(ro[8]), T6u, ovs, &(ro[0]));
1277 T6v = VSUB(T1w, T11);
1278 STM4(&(io[24]), T6v, ovs, &(io[0]));
1279 T6w = VSUB(T5V, T5Y);
1280 STM4(&(ro[24]), T6w, ovs, &(ro[0]));
1281 }
1282 {
1283 V T6x, T6y, T6z, T6A, T6B, T6C, T6D, T6E;
1284 {
1285 V T1X, T33, T31, T37, T2o, T34, T2P, T35;
1286 {
1287 V T1H, T1W, T2X, T30;
1288 T1H = VSUB(T1z, T1G);
1289 T1W = VSUB(T1O, T1V);
1290 T1X = VADD(T1H, T1W);
1291 T33 = VSUB(T1H, T1W);
1292 T2X = VSUB(T2T, T2W);
1293 T30 = VSUB(T2Y, T2Z);
1294 T31 = VSUB(T2X, T30);
1295 T37 = VADD(T2X, T30);
1296 }
1297 {
1298 V T2e, T2n, T2F, T2O;
1299 T2e = VSUB(T22, T2d);
1300 T2n = VSUB(T2j, T2m);
1301 T2o = VFMA(LDK(KP980785280), T2e, VMUL(LDK(KP195090322), T2n));
1302 T34 = VFNMS(LDK(KP980785280), T2n, VMUL(LDK(KP195090322), T2e));
1303 T2F = VSUB(T2t, T2E);
1304 T2O = VSUB(T2K, T2N);
1305 T2P = VFNMS(LDK(KP980785280), T2O, VMUL(LDK(KP195090322), T2F));
1306 T35 = VFMA(LDK(KP195090322), T2O, VMUL(LDK(KP980785280), T2F));
1307 }
1308 {
1309 V T2Q, T38, T32, T36;
1310 T2Q = VADD(T2o, T2P);
1311 T6x = VSUB(T1X, T2Q);
1312 STM4(&(ro[23]), T6x, ovs, &(ro[1]));
1313 T6y = VADD(T1X, T2Q);
1314 STM4(&(ro[7]), T6y, ovs, &(ro[1]));
1315 T38 = VADD(T34, T35);
1316 T6z = VSUB(T37, T38);
1317 STM4(&(io[23]), T6z, ovs, &(io[1]));
1318 T6A = VADD(T37, T38);
1319 STM4(&(io[7]), T6A, ovs, &(io[1]));
1320 T32 = VSUB(T2P, T2o);
1321 T6B = VSUB(T31, T32);
1322 STM4(&(io[31]), T6B, ovs, &(io[1]));
1323 T6C = VADD(T31, T32);
1324 STM4(&(io[15]), T6C, ovs, &(io[1]));
1325 T36 = VSUB(T34, T35);
1326 T6D = VSUB(T33, T36);
1327 STM4(&(ro[31]), T6D, ovs, &(ro[1]));
1328 T6E = VADD(T33, T36);
1329 STM4(&(ro[15]), T6E, ovs, &(ro[1]));
1330 }
1331 }
1332 {
1333 V T3D, T41, T3Z, T45, T3K, T42, T3R, T43;
1334 {
1335 V T3v, T3C, T3V, T3Y;
1336 T3v = VSUB(T3t, T3u);
1337 T3C = VSUB(T3y, T3B);
1338 T3D = VADD(T3v, T3C);
1339 T41 = VSUB(T3v, T3C);
1340 T3V = VSUB(T3T, T3U);
1341 T3Y = VSUB(T3W, T3X);
1342 T3Z = VSUB(T3V, T3Y);
1343 T45 = VADD(T3V, T3Y);
1344 }
1345 {
1346 V T3G, T3J, T3N, T3Q;
1347 T3G = VSUB(T3E, T3F);
1348 T3J = VSUB(T3H, T3I);
1349 T3K = VFMA(LDK(KP555570233), T3G, VMUL(LDK(KP831469612), T3J));
1350 T42 = VFNMS(LDK(KP831469612), T3G, VMUL(LDK(KP555570233), T3J));
1351 T3N = VSUB(T3L, T3M);
1352 T3Q = VSUB(T3O, T3P);
1353 T3R = VFNMS(LDK(KP831469612), T3Q, VMUL(LDK(KP555570233), T3N));
1354 T43 = VFMA(LDK(KP831469612), T3N, VMUL(LDK(KP555570233), T3Q));
1355 }
1356 {
1357 V T3S, T6F, T6G, T46, T6H, T6I;
1358 T3S = VADD(T3K, T3R);
1359 T6F = VSUB(T3D, T3S);
1360 STM4(&(ro[21]), T6F, ovs, &(ro[1]));
1361 STN4(&(ro[20]), T6h, T6F, T61, T6x, ovs);
1362 T6G = VADD(T3D, T3S);
1363 STM4(&(ro[5]), T6G, ovs, &(ro[1]));
1364 STN4(&(ro[4]), T6j, T6G, T63, T6y, ovs);
1365 T46 = VADD(T42, T43);
1366 T6H = VSUB(T45, T46);
1367 STM4(&(io[21]), T6H, ovs, &(io[1]));
1368 STN4(&(io[20]), T6i, T6H, T62, T6z, ovs);
1369 T6I = VADD(T45, T46);
1370 STM4(&(io[5]), T6I, ovs, &(io[1]));
1371 STN4(&(io[4]), T6k, T6I, T64, T6A, ovs);
1372 }
1373 {
1374 V T40, T6J, T6K, T44, T6L, T6M;
1375 T40 = VSUB(T3R, T3K);
1376 T6J = VSUB(T3Z, T40);
1377 STM4(&(io[29]), T6J, ovs, &(io[1]));
1378 STN4(&(io[28]), T6l, T6J, T65, T6B, ovs);
1379 T6K = VADD(T3Z, T40);
1380 STM4(&(io[13]), T6K, ovs, &(io[1]));
1381 STN4(&(io[12]), T6n, T6K, T67, T6C, ovs);
1382 T44 = VSUB(T42, T43);
1383 T6L = VSUB(T41, T44);
1384 STM4(&(ro[29]), T6L, ovs, &(ro[1]));
1385 STN4(&(ro[28]), T6m, T6L, T66, T6D, ovs);
1386 T6M = VADD(T41, T44);
1387 STM4(&(ro[13]), T6M, ovs, &(ro[1]));
1388 STN4(&(ro[12]), T6o, T6M, T68, T6E, ovs);
1389 }
1390 }
1391 }
1392 {
1393 V T6N, T6O, T6P, T6Q, T6R, T6S, T6T, T6U;
1394 {
1395 V T49, T4l, T4j, T4p, T4c, T4m, T4f, T4n;
1396 {
1397 V T47, T48, T4h, T4i;
1398 T47 = VADD(T3t, T3u);
1399 T48 = VADD(T3X, T3W);
1400 T49 = VADD(T47, T48);
1401 T4l = VSUB(T47, T48);
1402 T4h = VADD(T3T, T3U);
1403 T4i = VADD(T3y, T3B);
1404 T4j = VSUB(T4h, T4i);
1405 T4p = VADD(T4h, T4i);
1406 }
1407 {
1408 V T4a, T4b, T4d, T4e;
1409 T4a = VADD(T3E, T3F);
1410 T4b = VADD(T3H, T3I);
1411 T4c = VFMA(LDK(KP980785280), T4a, VMUL(LDK(KP195090322), T4b));
1412 T4m = VFNMS(LDK(KP195090322), T4a, VMUL(LDK(KP980785280), T4b));
1413 T4d = VADD(T3L, T3M);
1414 T4e = VADD(T3O, T3P);
1415 T4f = VFNMS(LDK(KP195090322), T4e, VMUL(LDK(KP980785280), T4d));
1416 T4n = VFMA(LDK(KP195090322), T4d, VMUL(LDK(KP980785280), T4e));
1417 }
1418 {
1419 V T4g, T4q, T4k, T4o;
1420 T4g = VADD(T4c, T4f);
1421 T6N = VSUB(T49, T4g);
1422 STM4(&(ro[17]), T6N, ovs, &(ro[1]));
1423 T6O = VADD(T49, T4g);
1424 STM4(&(ro[1]), T6O, ovs, &(ro[1]));
1425 T4q = VADD(T4m, T4n);
1426 T6P = VSUB(T4p, T4q);
1427 STM4(&(io[17]), T6P, ovs, &(io[1]));
1428 T6Q = VADD(T4p, T4q);
1429 STM4(&(io[1]), T6Q, ovs, &(io[1]));
1430 T4k = VSUB(T4f, T4c);
1431 T6R = VSUB(T4j, T4k);
1432 STM4(&(io[25]), T6R, ovs, &(io[1]));
1433 T6S = VADD(T4j, T4k);
1434 STM4(&(io[9]), T6S, ovs, &(io[1]));
1435 T4o = VSUB(T4m, T4n);
1436 T6T = VSUB(T4l, T4o);
1437 STM4(&(ro[25]), T6T, ovs, &(ro[1]));
1438 T6U = VADD(T4l, T4o);
1439 STM4(&(ro[9]), T6U, ovs, &(ro[1]));
1440 }
1441 }
1442 {
1443 V T3b, T3n, T3l, T3r, T3e, T3o, T3h, T3p;
1444 {
1445 V T39, T3a, T3j, T3k;
1446 T39 = VADD(T1z, T1G);
1447 T3a = VADD(T2Z, T2Y);
1448 T3b = VADD(T39, T3a);
1449 T3n = VSUB(T39, T3a);
1450 T3j = VADD(T2T, T2W);
1451 T3k = VADD(T1O, T1V);
1452 T3l = VSUB(T3j, T3k);
1453 T3r = VADD(T3j, T3k);
1454 }
1455 {
1456 V T3c, T3d, T3f, T3g;
1457 T3c = VADD(T22, T2d);
1458 T3d = VADD(T2j, T2m);
1459 T3e = VFMA(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d));
1460 T3o = VFNMS(LDK(KP555570233), T3d, VMUL(LDK(KP831469612), T3c));
1461 T3f = VADD(T2t, T2E);
1462 T3g = VADD(T2K, T2N);
1463 T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f));
1464 T3p = VFMA(LDK(KP831469612), T3g, VMUL(LDK(KP555570233), T3f));
1465 }
1466 {
1467 V T3i, T6V, T6W, T3s, T6X, T6Y;
1468 T3i = VADD(T3e, T3h);
1469 T6V = VSUB(T3b, T3i);
1470 STM4(&(ro[19]), T6V, ovs, &(ro[1]));
1471 STN4(&(ro[16]), T6p, T6N, T69, T6V, ovs);
1472 T6W = VADD(T3b, T3i);
1473 STM4(&(ro[3]), T6W, ovs, &(ro[1]));
1474 STN4(&(ro[0]), T6r, T6O, T6b, T6W, ovs);
1475 T3s = VADD(T3o, T3p);
1476 T6X = VSUB(T3r, T3s);
1477 STM4(&(io[19]), T6X, ovs, &(io[1]));
1478 STN4(&(io[16]), T6q, T6P, T6a, T6X, ovs);
1479 T6Y = VADD(T3r, T3s);
1480 STM4(&(io[3]), T6Y, ovs, &(io[1]));
1481 STN4(&(io[0]), T6s, T6Q, T6c, T6Y, ovs);
1482 }
1483 {
1484 V T3m, T6Z, T70, T3q, T71, T72;
1485 T3m = VSUB(T3h, T3e);
1486 T6Z = VSUB(T3l, T3m);
1487 STM4(&(io[27]), T6Z, ovs, &(io[1]));
1488 STN4(&(io[24]), T6v, T6R, T6d, T6Z, ovs);
1489 T70 = VADD(T3l, T3m);
1490 STM4(&(io[11]), T70, ovs, &(io[1]));
1491 STN4(&(io[8]), T6t, T6S, T6f, T70, ovs);
1492 T3q = VSUB(T3o, T3p);
1493 T71 = VSUB(T3n, T3q);
1494 STM4(&(ro[27]), T71, ovs, &(ro[1]));
1495 STN4(&(ro[24]), T6w, T6T, T6e, T71, ovs);
1496 T72 = VADD(T3n, T3q);
1497 STM4(&(ro[11]), T72, ovs, &(ro[1]));
1498 STN4(&(ro[8]), T6u, T6U, T6g, T72, ovs);
1499 }
1500 }
1501 }
1502 }
1503 }
1504 }
1505 VLEAVE();
1506 }
1507
1508 static const kdft_desc desc = { 32, XSIMD_STRING("n2sv_32"), {340, 52, 32, 0}, &GENUS, 0, 1, 0, 0 };
1509
1510 void XSIMD(codelet_n2sv_32) (planner *p) {
1511 X(kdft_register) (p, n2sv_32, &desc);
1512 }
1513
1514 #endif