comparison src/fftw-3.3.3/dft/simd/common/n2sv_32.c @ 10:37bf6b4a2645

Add FFTW3
author Chris Cannam
date Wed, 20 Mar 2013 15:35:50 +0000
parents
children
comparison
equal deleted inserted replaced
9:c0fb53affa76 10:37bf6b4a2645
1 /*
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 */
20
21 /* This file was automatically generated --- DO NOT EDIT */
22 /* Generated on Sun Nov 25 07:37:49 EST 2012 */
23
24 #include "codelet-dft.h"
25
26 #ifdef HAVE_FMA
27
28 /* Generated by: ../../../genfft/gen_notw.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n2sv_32 -with-ostride 1 -include 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 * 194 stack variables, 7 constants, and 144 memory accesses
34 */
35 #include "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 T61, T62, T63, T64, T65, T66, T67, T68, T69, T6a, T6b, T6c, T6d, T6e, T6f;
50 V T6g, T6h, T6i, T6j, T6k, T6l, T6m, T6n, T6o, T6p, T6q, T6r, T6s, T6t, T6u;
51 V T6v, T6w, T3g, T3f, T6x, T6y, T6z, T6A, T6B, T6C, T6D, T6E, T4p, T49, T4l;
52 V T4j, T6F, T6G, T6H, T6I, T6J, T6K, T6L, T6M, T3n, T3b, T3r, T3l, T3o, T3e;
53 V T4q, T4o, T4k, T4g, T3h, T3p;
54 {
55 V T2T, T3T, T4r, T7, T3t, T1z, T18, T4Z, Te, T50, T1f, T4s, T1G, T3U, T2W;
56 V T3u, Tm, T1n, T3X, T3y, T2Z, T1O, T53, T4w, Tt, T1u, T3W, T3B, T2Y, T1V;
57 V T52, T4z, T3O, T2t, T3L, T2K, TZ, T5F, T4R, T5k, T5j, T4W, T5I, T5X, T2E;
58 V T3M, T2N, T3P, T3H, T22, T3E, T2j, T4G, T5h, TK, T5A, T5D, T5W, T2d, T3F;
59 V T4L, T5g, T3I, T2m;
60 {
61 V T1L, T1j, T1k, T1l, T4v, T1K, T3w;
62 {
63 V T1, T2, T12, T13, T4, T5, T15, T16;
64 T1 = LD(&(ri[0]), ivs, &(ri[0]));
65 T2 = LD(&(ri[WS(is, 16)]), ivs, &(ri[0]));
66 T12 = LD(&(ii[0]), ivs, &(ii[0]));
67 T13 = LD(&(ii[WS(is, 16)]), ivs, &(ii[0]));
68 T4 = LD(&(ri[WS(is, 8)]), ivs, &(ri[0]));
69 T5 = LD(&(ri[WS(is, 24)]), ivs, &(ri[0]));
70 T15 = LD(&(ii[WS(is, 8)]), ivs, &(ii[0]));
71 T16 = LD(&(ii[WS(is, 24)]), ivs, &(ii[0]));
72 {
73 V Tb, T1A, Ta, T1B, T1b, Tc, T1c, T1d;
74 {
75 V T8, T1x, T3, T2R, T14, T2S, T6, T1y, T17, T9, T19, T1a;
76 T8 = LD(&(ri[WS(is, 4)]), ivs, &(ri[0]));
77 T1x = VSUB(T1, T2);
78 T3 = VADD(T1, T2);
79 T2R = VSUB(T12, T13);
80 T14 = VADD(T12, T13);
81 T2S = VSUB(T4, T5);
82 T6 = VADD(T4, T5);
83 T1y = VSUB(T15, T16);
84 T17 = VADD(T15, T16);
85 T9 = LD(&(ri[WS(is, 20)]), ivs, &(ri[0]));
86 T19 = LD(&(ii[WS(is, 4)]), ivs, &(ii[0]));
87 T1a = LD(&(ii[WS(is, 20)]), ivs, &(ii[0]));
88 Tb = LD(&(ri[WS(is, 28)]), ivs, &(ri[0]));
89 T2T = VSUB(T2R, T2S);
90 T3T = VADD(T2S, T2R);
91 T4r = VSUB(T3, T6);
92 T7 = VADD(T3, T6);
93 T3t = VSUB(T1x, T1y);
94 T1z = VADD(T1x, T1y);
95 T18 = VADD(T14, T17);
96 T4Z = VSUB(T14, T17);
97 T1A = VSUB(T8, T9);
98 Ta = VADD(T8, T9);
99 T1B = VSUB(T19, T1a);
100 T1b = VADD(T19, T1a);
101 Tc = LD(&(ri[WS(is, 12)]), ivs, &(ri[0]));
102 T1c = LD(&(ii[WS(is, 28)]), ivs, &(ii[0]));
103 T1d = LD(&(ii[WS(is, 12)]), ivs, &(ii[0]));
104 }
105 {
106 V Ti, T1I, T1J, Tl;
107 {
108 V T1h, T1C, T2U, T1D, Td, T1E, T1e, T1i, Tg, Th;
109 Tg = LD(&(ri[WS(is, 2)]), ivs, &(ri[0]));
110 Th = LD(&(ri[WS(is, 18)]), ivs, &(ri[0]));
111 T1h = LD(&(ii[WS(is, 2)]), ivs, &(ii[0]));
112 T1C = VADD(T1A, T1B);
113 T2U = VSUB(T1B, T1A);
114 T1D = VSUB(Tb, Tc);
115 Td = VADD(Tb, Tc);
116 T1E = VSUB(T1c, T1d);
117 T1e = VADD(T1c, T1d);
118 T1L = VSUB(Tg, Th);
119 Ti = VADD(Tg, Th);
120 T1i = LD(&(ii[WS(is, 18)]), ivs, &(ii[0]));
121 {
122 V T2V, T1F, Tj, Tk;
123 Tj = LD(&(ri[WS(is, 10)]), ivs, &(ri[0]));
124 Tk = LD(&(ri[WS(is, 26)]), ivs, &(ri[0]));
125 Te = VADD(Ta, Td);
126 T50 = VSUB(Td, Ta);
127 T2V = VADD(T1D, T1E);
128 T1F = VSUB(T1D, T1E);
129 T1f = VADD(T1b, T1e);
130 T4s = VSUB(T1b, T1e);
131 T1j = VADD(T1h, T1i);
132 T1I = VSUB(T1h, T1i);
133 T1J = VSUB(Tj, Tk);
134 Tl = VADD(Tj, Tk);
135 T1G = VADD(T1C, T1F);
136 T3U = VSUB(T1F, T1C);
137 T2W = VADD(T2U, T2V);
138 T3u = VSUB(T2U, T2V);
139 T1k = LD(&(ii[WS(is, 10)]), ivs, &(ii[0]));
140 T1l = LD(&(ii[WS(is, 26)]), ivs, &(ii[0]));
141 }
142 }
143 T4v = VSUB(Ti, Tl);
144 Tm = VADD(Ti, Tl);
145 T1K = VSUB(T1I, T1J);
146 T3w = VADD(T1J, T1I);
147 }
148 }
149 }
150 {
151 V T1r, T1S, T1q, T1s, T4x, T1R, T3z;
152 {
153 V Tp, T1P, T1Q, Ts;
154 {
155 V Tn, To, T1o, T1M, T1m, T1p;
156 Tn = LD(&(ri[WS(is, 30)]), ivs, &(ri[0]));
157 To = LD(&(ri[WS(is, 14)]), ivs, &(ri[0]));
158 T1o = LD(&(ii[WS(is, 30)]), ivs, &(ii[0]));
159 T1M = VSUB(T1k, T1l);
160 T1m = VADD(T1k, T1l);
161 T1p = LD(&(ii[WS(is, 14)]), ivs, &(ii[0]));
162 {
163 V Tq, Tr, T3x, T1N, T4u;
164 Tq = LD(&(ri[WS(is, 6)]), ivs, &(ri[0]));
165 Tr = LD(&(ri[WS(is, 22)]), ivs, &(ri[0]));
166 T1r = LD(&(ii[WS(is, 6)]), ivs, &(ii[0]));
167 T1S = VSUB(Tn, To);
168 Tp = VADD(Tn, To);
169 T3x = VSUB(T1L, T1M);
170 T1N = VADD(T1L, T1M);
171 T4u = VSUB(T1j, T1m);
172 T1n = VADD(T1j, T1m);
173 T1P = VSUB(T1o, T1p);
174 T1q = VADD(T1o, T1p);
175 T1Q = VSUB(Tq, Tr);
176 Ts = VADD(Tq, Tr);
177 T3X = VFNMS(LDK(KP414213562), T3w, T3x);
178 T3y = VFMA(LDK(KP414213562), T3x, T3w);
179 T2Z = VFMA(LDK(KP414213562), T1K, T1N);
180 T1O = VFNMS(LDK(KP414213562), T1N, T1K);
181 T53 = VADD(T4v, T4u);
182 T4w = VSUB(T4u, T4v);
183 T1s = LD(&(ii[WS(is, 22)]), ivs, &(ii[0]));
184 }
185 }
186 T4x = VSUB(Tp, Ts);
187 Tt = VADD(Tp, Ts);
188 T1R = VSUB(T1P, T1Q);
189 T3z = VADD(T1Q, T1P);
190 }
191 {
192 V T4S, T5G, T2y, T2L, T4V, T5H, T2D, T2M;
193 {
194 V T2G, TN, T4N, T2r, T2s, TQ, T2A, T4O, T2J, T2x, TU, T4T, T2w, T2z, TX;
195 V T2B, T2H, T2I, TR;
196 {
197 V TL, TM, T2p, T1T, T1t, T2q;
198 TL = LD(&(ri[WS(is, 31)]), ivs, &(ri[WS(is, 1)]));
199 TM = LD(&(ri[WS(is, 15)]), ivs, &(ri[WS(is, 1)]));
200 T2p = LD(&(ii[WS(is, 31)]), ivs, &(ii[WS(is, 1)]));
201 T1T = VSUB(T1r, T1s);
202 T1t = VADD(T1r, T1s);
203 T2q = LD(&(ii[WS(is, 15)]), ivs, &(ii[WS(is, 1)]));
204 {
205 V TO, TP, T3A, T1U, T4y;
206 TO = LD(&(ri[WS(is, 7)]), ivs, &(ri[WS(is, 1)]));
207 TP = LD(&(ri[WS(is, 23)]), ivs, &(ri[WS(is, 1)]));
208 T2H = LD(&(ii[WS(is, 7)]), ivs, &(ii[WS(is, 1)]));
209 T2G = VSUB(TL, TM);
210 TN = VADD(TL, TM);
211 T3A = VSUB(T1S, T1T);
212 T1U = VADD(T1S, T1T);
213 T4y = VSUB(T1q, T1t);
214 T1u = VADD(T1q, T1t);
215 T4N = VADD(T2p, T2q);
216 T2r = VSUB(T2p, T2q);
217 T2s = VSUB(TO, TP);
218 TQ = VADD(TO, TP);
219 T3W = VFMA(LDK(KP414213562), T3z, T3A);
220 T3B = VFNMS(LDK(KP414213562), T3A, T3z);
221 T2Y = VFNMS(LDK(KP414213562), T1R, T1U);
222 T1V = VFMA(LDK(KP414213562), T1U, T1R);
223 T52 = VSUB(T4x, T4y);
224 T4z = VADD(T4x, T4y);
225 T2I = LD(&(ii[WS(is, 23)]), ivs, &(ii[WS(is, 1)]));
226 }
227 }
228 {
229 V TS, TT, T2u, T2v, TV, TW;
230 TS = LD(&(ri[WS(is, 3)]), ivs, &(ri[WS(is, 1)]));
231 TT = LD(&(ri[WS(is, 19)]), ivs, &(ri[WS(is, 1)]));
232 T2u = LD(&(ii[WS(is, 3)]), ivs, &(ii[WS(is, 1)]));
233 T2v = LD(&(ii[WS(is, 19)]), ivs, &(ii[WS(is, 1)]));
234 TV = LD(&(ri[WS(is, 27)]), ivs, &(ri[WS(is, 1)]));
235 TW = LD(&(ri[WS(is, 11)]), ivs, &(ri[WS(is, 1)]));
236 T2A = LD(&(ii[WS(is, 27)]), ivs, &(ii[WS(is, 1)]));
237 T4O = VADD(T2H, T2I);
238 T2J = VSUB(T2H, T2I);
239 T2x = VSUB(TS, TT);
240 TU = VADD(TS, TT);
241 T4T = VADD(T2u, T2v);
242 T2w = VSUB(T2u, T2v);
243 T2z = VSUB(TV, TW);
244 TX = VADD(TV, TW);
245 T2B = LD(&(ii[WS(is, 11)]), ivs, &(ii[WS(is, 1)]));
246 }
247 T3O = VADD(T2s, T2r);
248 T2t = VSUB(T2r, T2s);
249 T3L = VSUB(T2G, T2J);
250 T2K = VADD(T2G, T2J);
251 T4S = VSUB(TN, TQ);
252 TR = VADD(TN, TQ);
253 {
254 V T4P, T4Q, TY, T4U, T2C;
255 T5G = VADD(T4N, T4O);
256 T4P = VSUB(T4N, T4O);
257 T4Q = VSUB(TX, TU);
258 TY = VADD(TU, TX);
259 T4U = VADD(T2A, T2B);
260 T2C = VSUB(T2A, T2B);
261 T2y = VSUB(T2w, T2x);
262 T2L = VADD(T2x, T2w);
263 TZ = VADD(TR, TY);
264 T5F = VSUB(TR, TY);
265 T4V = VSUB(T4T, T4U);
266 T5H = VADD(T4T, T4U);
267 T2D = VADD(T2z, T2C);
268 T2M = VSUB(T2z, T2C);
269 T4R = VSUB(T4P, T4Q);
270 T5k = VADD(T4Q, T4P);
271 }
272 }
273 {
274 V T2f, Ty, T23, T4C, T20, T21, TB, T4D, T2i, T26, TF, T24, TG, TH, T29;
275 V T2a;
276 {
277 V T1Y, T1Z, Tz, TA, T2g, T2h, Tw, Tx, TD, TE;
278 Tw = LD(&(ri[WS(is, 1)]), ivs, &(ri[WS(is, 1)]));
279 Tx = LD(&(ri[WS(is, 17)]), ivs, &(ri[WS(is, 1)]));
280 T5j = VADD(T4S, T4V);
281 T4W = VSUB(T4S, T4V);
282 T5I = VSUB(T5G, T5H);
283 T5X = VADD(T5G, T5H);
284 T2E = VADD(T2y, T2D);
285 T3M = VSUB(T2D, T2y);
286 T2N = VADD(T2L, T2M);
287 T3P = VSUB(T2L, T2M);
288 T2f = VSUB(Tw, Tx);
289 Ty = VADD(Tw, Tx);
290 T1Y = LD(&(ii[WS(is, 1)]), ivs, &(ii[WS(is, 1)]));
291 T1Z = LD(&(ii[WS(is, 17)]), ivs, &(ii[WS(is, 1)]));
292 Tz = LD(&(ri[WS(is, 9)]), ivs, &(ri[WS(is, 1)]));
293 TA = LD(&(ri[WS(is, 25)]), ivs, &(ri[WS(is, 1)]));
294 T2g = LD(&(ii[WS(is, 9)]), ivs, &(ii[WS(is, 1)]));
295 T2h = LD(&(ii[WS(is, 25)]), ivs, &(ii[WS(is, 1)]));
296 TD = LD(&(ri[WS(is, 5)]), ivs, &(ri[WS(is, 1)]));
297 TE = LD(&(ri[WS(is, 21)]), ivs, &(ri[WS(is, 1)]));
298 T23 = LD(&(ii[WS(is, 5)]), ivs, &(ii[WS(is, 1)]));
299 T4C = VADD(T1Y, T1Z);
300 T20 = VSUB(T1Y, T1Z);
301 T21 = VSUB(Tz, TA);
302 TB = VADD(Tz, TA);
303 T4D = VADD(T2g, T2h);
304 T2i = VSUB(T2g, T2h);
305 T26 = VSUB(TD, TE);
306 TF = VADD(TD, TE);
307 T24 = LD(&(ii[WS(is, 21)]), ivs, &(ii[WS(is, 1)]));
308 TG = LD(&(ri[WS(is, 29)]), ivs, &(ri[WS(is, 1)]));
309 TH = LD(&(ri[WS(is, 13)]), ivs, &(ri[WS(is, 1)]));
310 T29 = LD(&(ii[WS(is, 29)]), ivs, &(ii[WS(is, 1)]));
311 T2a = LD(&(ii[WS(is, 13)]), ivs, &(ii[WS(is, 1)]));
312 }
313 {
314 V T4I, T25, T28, TI, T4J, T2b, T4H, TC, T5B, T4E;
315 T3H = VADD(T21, T20);
316 T22 = VSUB(T20, T21);
317 T3E = VSUB(T2f, T2i);
318 T2j = VADD(T2f, T2i);
319 T4I = VADD(T23, T24);
320 T25 = VSUB(T23, T24);
321 T28 = VSUB(TG, TH);
322 TI = VADD(TG, TH);
323 T4J = VADD(T29, T2a);
324 T2b = VSUB(T29, T2a);
325 T4H = VSUB(Ty, TB);
326 TC = VADD(Ty, TB);
327 T5B = VADD(T4C, T4D);
328 T4E = VSUB(T4C, T4D);
329 {
330 V T27, T2k, TJ, T4F, T4K, T5C, T2c, T2l;
331 T27 = VSUB(T25, T26);
332 T2k = VADD(T26, T25);
333 TJ = VADD(TF, TI);
334 T4F = VSUB(TI, TF);
335 T4K = VSUB(T4I, T4J);
336 T5C = VADD(T4I, T4J);
337 T2c = VADD(T28, T2b);
338 T2l = VSUB(T28, T2b);
339 T4G = VSUB(T4E, T4F);
340 T5h = VADD(T4F, T4E);
341 TK = VADD(TC, TJ);
342 T5A = VSUB(TC, TJ);
343 T5D = VSUB(T5B, T5C);
344 T5W = VADD(T5B, T5C);
345 T2d = VADD(T27, T2c);
346 T3F = VSUB(T2c, T27);
347 T4L = VSUB(T4H, T4K);
348 T5g = VADD(T4H, T4K);
349 T3I = VSUB(T2k, T2l);
350 T2m = VADD(T2k, T2l);
351 }
352 }
353 }
354 }
355 }
356 }
357 {
358 V T1v, T1g, T5V, Tv, T60, T5Y, T11, T10;
359 {
360 V T5o, T5n, T5i, T5r, T5f, T5l, T5w, T5u;
361 {
362 V T5d, T4t, T4A, T4X, T58, T51, T4M, T59, T54, T5e, T5b, T4B;
363 T5d = VADD(T4r, T4s);
364 T4t = VSUB(T4r, T4s);
365 T4A = VSUB(T4w, T4z);
366 T5o = VADD(T4w, T4z);
367 T4X = VFNMS(LDK(KP414213562), T4W, T4R);
368 T58 = VFMA(LDK(KP414213562), T4R, T4W);
369 T5n = VADD(T50, T4Z);
370 T51 = VSUB(T4Z, T50);
371 T4M = VFMA(LDK(KP414213562), T4L, T4G);
372 T59 = VFNMS(LDK(KP414213562), T4G, T4L);
373 T54 = VSUB(T52, T53);
374 T5e = VADD(T53, T52);
375 T5b = VFNMS(LDK(KP707106781), T4A, T4t);
376 T4B = VFMA(LDK(KP707106781), T4A, T4t);
377 {
378 V T5s, T56, T4Y, T5c, T5a, T57, T55, T5t;
379 T5i = VFMA(LDK(KP414213562), T5h, T5g);
380 T5s = VFNMS(LDK(KP414213562), T5g, T5h);
381 T56 = VADD(T4M, T4X);
382 T4Y = VSUB(T4M, T4X);
383 T5c = VADD(T59, T58);
384 T5a = VSUB(T58, T59);
385 T57 = VFMA(LDK(KP707106781), T54, T51);
386 T55 = VFNMS(LDK(KP707106781), T54, T51);
387 T5r = VFNMS(LDK(KP707106781), T5e, T5d);
388 T5f = VFMA(LDK(KP707106781), T5e, T5d);
389 T5t = VFMA(LDK(KP414213562), T5j, T5k);
390 T5l = VFNMS(LDK(KP414213562), T5k, T5j);
391 T61 = VFMA(LDK(KP923879532), T4Y, T4B);
392 STM4(&(ro[6]), T61, ovs, &(ro[0]));
393 T62 = VFNMS(LDK(KP923879532), T4Y, T4B);
394 STM4(&(ro[22]), T62, ovs, &(ro[0]));
395 T63 = VFMA(LDK(KP923879532), T5c, T5b);
396 STM4(&(ro[30]), T63, ovs, &(ro[0]));
397 T64 = VFNMS(LDK(KP923879532), T5c, T5b);
398 STM4(&(ro[14]), T64, ovs, &(ro[0]));
399 T65 = VFMA(LDK(KP923879532), T5a, T57);
400 STM4(&(io[6]), T65, ovs, &(io[0]));
401 T66 = VFNMS(LDK(KP923879532), T5a, T57);
402 STM4(&(io[22]), T66, ovs, &(io[0]));
403 T67 = VFMA(LDK(KP923879532), T56, T55);
404 STM4(&(io[30]), T67, ovs, &(io[0]));
405 T68 = VFNMS(LDK(KP923879532), T56, T55);
406 STM4(&(io[14]), T68, ovs, &(io[0]));
407 T5w = VADD(T5s, T5t);
408 T5u = VSUB(T5s, T5t);
409 }
410 }
411 {
412 V Tf, T5P, T5z, T5S, T5U, T5O, T5K, T5L, T5M, Tu, T5T, T5N;
413 {
414 V T5E, T5Q, T5q, T5m, T5v, T5p, T5R, T5J, T5x, T5y;
415 Tf = VADD(T7, Te);
416 T5x = VSUB(T7, Te);
417 T5y = VSUB(T1n, T1u);
418 T1v = VADD(T1n, T1u);
419 T69 = VFMA(LDK(KP923879532), T5u, T5r);
420 STM4(&(ro[10]), T69, ovs, &(ro[0]));
421 T6a = VFNMS(LDK(KP923879532), T5u, T5r);
422 STM4(&(ro[26]), T6a, ovs, &(ro[0]));
423 T5E = VADD(T5A, T5D);
424 T5Q = VSUB(T5D, T5A);
425 T5q = VSUB(T5l, T5i);
426 T5m = VADD(T5i, T5l);
427 T5v = VFMA(LDK(KP707106781), T5o, T5n);
428 T5p = VFNMS(LDK(KP707106781), T5o, T5n);
429 T5P = VSUB(T5x, T5y);
430 T5z = VADD(T5x, T5y);
431 T5R = VADD(T5F, T5I);
432 T5J = VSUB(T5F, T5I);
433 T6b = VFMA(LDK(KP923879532), T5m, T5f);
434 STM4(&(ro[2]), T6b, ovs, &(ro[0]));
435 T6c = VFNMS(LDK(KP923879532), T5m, T5f);
436 STM4(&(ro[18]), T6c, ovs, &(ro[0]));
437 T6d = VFMA(LDK(KP923879532), T5w, T5v);
438 STM4(&(io[2]), T6d, ovs, &(io[0]));
439 T6e = VFNMS(LDK(KP923879532), T5w, T5v);
440 STM4(&(io[18]), T6e, ovs, &(io[0]));
441 T6f = VFMA(LDK(KP923879532), T5q, T5p);
442 STM4(&(io[10]), T6f, ovs, &(io[0]));
443 T6g = VFNMS(LDK(KP923879532), T5q, T5p);
444 STM4(&(io[26]), T6g, ovs, &(io[0]));
445 T5S = VSUB(T5Q, T5R);
446 T5U = VADD(T5Q, T5R);
447 T5O = VSUB(T5J, T5E);
448 T5K = VADD(T5E, T5J);
449 T1g = VADD(T18, T1f);
450 T5L = VSUB(T18, T1f);
451 T5M = VSUB(Tt, Tm);
452 Tu = VADD(Tm, Tt);
453 }
454 T6h = VFMA(LDK(KP707106781), T5S, T5P);
455 STM4(&(ro[12]), T6h, ovs, &(ro[0]));
456 T6i = VFNMS(LDK(KP707106781), T5S, T5P);
457 STM4(&(ro[28]), T6i, ovs, &(ro[0]));
458 T6j = VFMA(LDK(KP707106781), T5K, T5z);
459 STM4(&(ro[4]), T6j, ovs, &(ro[0]));
460 T6k = VFNMS(LDK(KP707106781), T5K, T5z);
461 STM4(&(ro[20]), T6k, ovs, &(ro[0]));
462 T5T = VADD(T5M, T5L);
463 T5N = VSUB(T5L, T5M);
464 T5V = VSUB(Tf, Tu);
465 Tv = VADD(Tf, Tu);
466 T6l = VFMA(LDK(KP707106781), T5U, T5T);
467 STM4(&(io[4]), T6l, ovs, &(io[0]));
468 T6m = VFNMS(LDK(KP707106781), T5U, T5T);
469 STM4(&(io[20]), T6m, ovs, &(io[0]));
470 T6n = VFMA(LDK(KP707106781), T5O, T5N);
471 STM4(&(io[12]), T6n, ovs, &(io[0]));
472 T6o = VFNMS(LDK(KP707106781), T5O, T5N);
473 STM4(&(io[28]), T6o, ovs, &(io[0]));
474 T60 = VADD(T5W, T5X);
475 T5Y = VSUB(T5W, T5X);
476 T11 = VSUB(TZ, TK);
477 T10 = VADD(TK, TZ);
478 }
479 }
480 {
481 V T39, T3k, T3j, T3a, T1X, T37, T33, T31, T3d, T3c, T47, T4i, T4h, T48, T4b;
482 V T4a, T4e, T3N, T41, T3D, T45, T3Z, T38, T36, T32, T2Q, T42, T3K, T3Q, T4d;
483 {
484 V T2e, T2n, T2F, T2O, T1w, T5Z;
485 {
486 V T1H, T1W, T2X, T30;
487 T39 = VFMA(LDK(KP707106781), T1G, T1z);
488 T1H = VFNMS(LDK(KP707106781), T1G, T1z);
489 T1W = VSUB(T1O, T1V);
490 T3k = VADD(T1O, T1V);
491 T3j = VFMA(LDK(KP707106781), T2W, T2T);
492 T2X = VFNMS(LDK(KP707106781), T2W, T2T);
493 T30 = VSUB(T2Y, T2Z);
494 T3a = VADD(T2Z, T2Y);
495 T6p = VSUB(T5V, T5Y);
496 STM4(&(ro[24]), T6p, ovs, &(ro[0]));
497 T6q = VADD(T5V, T5Y);
498 STM4(&(ro[8]), T6q, ovs, &(ro[0]));
499 T6r = VADD(Tv, T10);
500 STM4(&(ro[0]), T6r, ovs, &(ro[0]));
501 T6s = VSUB(Tv, T10);
502 STM4(&(ro[16]), T6s, ovs, &(ro[0]));
503 T1w = VSUB(T1g, T1v);
504 T5Z = VADD(T1g, T1v);
505 T1X = VFMA(LDK(KP923879532), T1W, T1H);
506 T37 = VFNMS(LDK(KP923879532), T1W, T1H);
507 T33 = VFMA(LDK(KP923879532), T30, T2X);
508 T31 = VFNMS(LDK(KP923879532), T30, T2X);
509 }
510 T3d = VFMA(LDK(KP707106781), T2d, T22);
511 T2e = VFNMS(LDK(KP707106781), T2d, T22);
512 T2n = VFNMS(LDK(KP707106781), T2m, T2j);
513 T3c = VFMA(LDK(KP707106781), T2m, T2j);
514 T6t = VADD(T5Z, T60);
515 STM4(&(io[0]), T6t, ovs, &(io[0]));
516 T6u = VSUB(T5Z, T60);
517 STM4(&(io[16]), T6u, ovs, &(io[0]));
518 T6v = VSUB(T1w, T11);
519 STM4(&(io[24]), T6v, ovs, &(io[0]));
520 T6w = VADD(T11, T1w);
521 STM4(&(io[8]), T6w, ovs, &(io[0]));
522 T3g = VFMA(LDK(KP707106781), T2E, T2t);
523 T2F = VFNMS(LDK(KP707106781), T2E, T2t);
524 T2O = VFNMS(LDK(KP707106781), T2N, T2K);
525 T3f = VFMA(LDK(KP707106781), T2N, T2K);
526 {
527 V T3v, T35, T2o, T3C, T3V, T3Y;
528 T47 = VFNMS(LDK(KP707106781), T3u, T3t);
529 T3v = VFMA(LDK(KP707106781), T3u, T3t);
530 T35 = VFNMS(LDK(KP668178637), T2e, T2n);
531 T2o = VFMA(LDK(KP668178637), T2n, T2e);
532 T3C = VSUB(T3y, T3B);
533 T4i = VADD(T3y, T3B);
534 T4h = VFNMS(LDK(KP707106781), T3U, T3T);
535 T3V = VFMA(LDK(KP707106781), T3U, T3T);
536 T3Y = VSUB(T3W, T3X);
537 T48 = VADD(T3X, T3W);
538 {
539 V T3G, T34, T2P, T3J;
540 T4b = VFMA(LDK(KP707106781), T3F, T3E);
541 T3G = VFNMS(LDK(KP707106781), T3F, T3E);
542 T34 = VFMA(LDK(KP668178637), T2F, T2O);
543 T2P = VFNMS(LDK(KP668178637), T2O, T2F);
544 T3J = VFNMS(LDK(KP707106781), T3I, T3H);
545 T4a = VFMA(LDK(KP707106781), T3I, T3H);
546 T4e = VFMA(LDK(KP707106781), T3M, T3L);
547 T3N = VFNMS(LDK(KP707106781), T3M, T3L);
548 T41 = VFNMS(LDK(KP923879532), T3C, T3v);
549 T3D = VFMA(LDK(KP923879532), T3C, T3v);
550 T45 = VFMA(LDK(KP923879532), T3Y, T3V);
551 T3Z = VFNMS(LDK(KP923879532), T3Y, T3V);
552 T38 = VADD(T35, T34);
553 T36 = VSUB(T34, T35);
554 T32 = VADD(T2o, T2P);
555 T2Q = VSUB(T2o, T2P);
556 T42 = VFNMS(LDK(KP668178637), T3G, T3J);
557 T3K = VFMA(LDK(KP668178637), T3J, T3G);
558 T3Q = VFNMS(LDK(KP707106781), T3P, T3O);
559 T4d = VFMA(LDK(KP707106781), T3P, T3O);
560 }
561 }
562 }
563 {
564 V T4n, T4c, T43, T3R, T4m, T4f;
565 T6x = VFMA(LDK(KP831469612), T38, T37);
566 STM4(&(ro[29]), T6x, ovs, &(ro[1]));
567 T6y = VFNMS(LDK(KP831469612), T38, T37);
568 STM4(&(ro[13]), T6y, ovs, &(ro[1]));
569 T6z = VFMA(LDK(KP831469612), T36, T33);
570 STM4(&(io[5]), T6z, ovs, &(io[1]));
571 T6A = VFNMS(LDK(KP831469612), T36, T33);
572 STM4(&(io[21]), T6A, ovs, &(io[1]));
573 T6B = VFMA(LDK(KP831469612), T32, T31);
574 STM4(&(io[29]), T6B, ovs, &(io[1]));
575 T6C = VFNMS(LDK(KP831469612), T32, T31);
576 STM4(&(io[13]), T6C, ovs, &(io[1]));
577 T6D = VFMA(LDK(KP831469612), T2Q, T1X);
578 STM4(&(ro[5]), T6D, ovs, &(ro[1]));
579 T6E = VFNMS(LDK(KP831469612), T2Q, T1X);
580 STM4(&(ro[21]), T6E, ovs, &(ro[1]));
581 T43 = VFMA(LDK(KP668178637), T3N, T3Q);
582 T3R = VFNMS(LDK(KP668178637), T3Q, T3N);
583 {
584 V T44, T46, T40, T3S;
585 T44 = VSUB(T42, T43);
586 T46 = VADD(T42, T43);
587 T40 = VSUB(T3R, T3K);
588 T3S = VADD(T3K, T3R);
589 T4p = VFMA(LDK(KP923879532), T48, T47);
590 T49 = VFNMS(LDK(KP923879532), T48, T47);
591 T4l = VFNMS(LDK(KP923879532), T4i, T4h);
592 T4j = VFMA(LDK(KP923879532), T4i, T4h);
593 T4n = VFNMS(LDK(KP198912367), T4a, T4b);
594 T4c = VFMA(LDK(KP198912367), T4b, T4a);
595 T6F = VFMA(LDK(KP831469612), T44, T41);
596 STM4(&(ro[11]), T6F, ovs, &(ro[1]));
597 T6G = VFNMS(LDK(KP831469612), T44, T41);
598 STM4(&(ro[27]), T6G, ovs, &(ro[1]));
599 T6H = VFMA(LDK(KP831469612), T46, T45);
600 STM4(&(io[3]), T6H, ovs, &(io[1]));
601 T6I = VFNMS(LDK(KP831469612), T46, T45);
602 STM4(&(io[19]), T6I, ovs, &(io[1]));
603 T6J = VFMA(LDK(KP831469612), T40, T3Z);
604 STM4(&(io[11]), T6J, ovs, &(io[1]));
605 T6K = VFNMS(LDK(KP831469612), T40, T3Z);
606 STM4(&(io[27]), T6K, ovs, &(io[1]));
607 T6L = VFMA(LDK(KP831469612), T3S, T3D);
608 STM4(&(ro[3]), T6L, ovs, &(ro[1]));
609 T6M = VFNMS(LDK(KP831469612), T3S, T3D);
610 STM4(&(ro[19]), T6M, ovs, &(ro[1]));
611 }
612 T4m = VFMA(LDK(KP198912367), T4d, T4e);
613 T4f = VFNMS(LDK(KP198912367), T4e, T4d);
614 T3n = VFNMS(LDK(KP923879532), T3a, T39);
615 T3b = VFMA(LDK(KP923879532), T3a, T39);
616 T3r = VFMA(LDK(KP923879532), T3k, T3j);
617 T3l = VFNMS(LDK(KP923879532), T3k, T3j);
618 T3o = VFNMS(LDK(KP198912367), T3c, T3d);
619 T3e = VFMA(LDK(KP198912367), T3d, T3c);
620 T4q = VADD(T4n, T4m);
621 T4o = VSUB(T4m, T4n);
622 T4k = VADD(T4c, T4f);
623 T4g = VSUB(T4c, T4f);
624 }
625 }
626 }
627 }
628 {
629 V T6N, T6O, T6P, T6Q;
630 T6N = VFMA(LDK(KP980785280), T4q, T4p);
631 STM4(&(ro[31]), T6N, ovs, &(ro[1]));
632 STN4(&(ro[28]), T6i, T6x, T63, T6N, ovs);
633 T6O = VFNMS(LDK(KP980785280), T4q, T4p);
634 STM4(&(ro[15]), T6O, ovs, &(ro[1]));
635 STN4(&(ro[12]), T6h, T6y, T64, T6O, ovs);
636 T6P = VFMA(LDK(KP980785280), T4o, T4l);
637 STM4(&(io[7]), T6P, ovs, &(io[1]));
638 STN4(&(io[4]), T6l, T6z, T65, T6P, ovs);
639 T6Q = VFNMS(LDK(KP980785280), T4o, T4l);
640 STM4(&(io[23]), T6Q, ovs, &(io[1]));
641 STN4(&(io[20]), T6m, T6A, T66, T6Q, ovs);
642 {
643 V T6R, T6S, T6T, T6U;
644 T6R = VFMA(LDK(KP980785280), T4k, T4j);
645 STM4(&(io[31]), T6R, ovs, &(io[1]));
646 STN4(&(io[28]), T6o, T6B, T67, T6R, ovs);
647 T6S = VFNMS(LDK(KP980785280), T4k, T4j);
648 STM4(&(io[15]), T6S, ovs, &(io[1]));
649 STN4(&(io[12]), T6n, T6C, T68, T6S, ovs);
650 T6T = VFMA(LDK(KP980785280), T4g, T49);
651 STM4(&(ro[7]), T6T, ovs, &(ro[1]));
652 STN4(&(ro[4]), T6j, T6D, T61, T6T, ovs);
653 T6U = VFNMS(LDK(KP980785280), T4g, T49);
654 STM4(&(ro[23]), T6U, ovs, &(ro[1]));
655 STN4(&(ro[20]), T6k, T6E, T62, T6U, ovs);
656 T3h = VFNMS(LDK(KP198912367), T3g, T3f);
657 T3p = VFMA(LDK(KP198912367), T3f, T3g);
658 }
659 }
660 {
661 V T3s, T3q, T3i, T3m;
662 T3s = VADD(T3o, T3p);
663 T3q = VSUB(T3o, T3p);
664 T3i = VADD(T3e, T3h);
665 T3m = VSUB(T3h, T3e);
666 {
667 V T6V, T6W, T6X, T6Y;
668 T6V = VFMA(LDK(KP980785280), T3q, T3n);
669 STM4(&(ro[9]), T6V, ovs, &(ro[1]));
670 STN4(&(ro[8]), T6q, T6V, T69, T6F, ovs);
671 T6W = VFNMS(LDK(KP980785280), T3q, T3n);
672 STM4(&(ro[25]), T6W, ovs, &(ro[1]));
673 STN4(&(ro[24]), T6p, T6W, T6a, T6G, ovs);
674 T6X = VFMA(LDK(KP980785280), T3s, T3r);
675 STM4(&(io[1]), T6X, ovs, &(io[1]));
676 STN4(&(io[0]), T6t, T6X, T6d, T6H, ovs);
677 T6Y = VFNMS(LDK(KP980785280), T3s, T3r);
678 STM4(&(io[17]), T6Y, ovs, &(io[1]));
679 STN4(&(io[16]), T6u, T6Y, T6e, T6I, ovs);
680 {
681 V T6Z, T70, T71, T72;
682 T6Z = VFMA(LDK(KP980785280), T3m, T3l);
683 STM4(&(io[9]), T6Z, ovs, &(io[1]));
684 STN4(&(io[8]), T6w, T6Z, T6f, T6J, ovs);
685 T70 = VFNMS(LDK(KP980785280), T3m, T3l);
686 STM4(&(io[25]), T70, ovs, &(io[1]));
687 STN4(&(io[24]), T6v, T70, T6g, T6K, ovs);
688 T71 = VFMA(LDK(KP980785280), T3i, T3b);
689 STM4(&(ro[1]), T71, ovs, &(ro[1]));
690 STN4(&(ro[0]), T6r, T71, T6b, T6L, ovs);
691 T72 = VFNMS(LDK(KP980785280), T3i, T3b);
692 STM4(&(ro[17]), T72, ovs, &(ro[1]));
693 STN4(&(ro[16]), T6s, T72, T6c, T6M, ovs);
694 }
695 }
696 }
697 }
698 }
699 VLEAVE();
700 }
701
702 static const kdft_desc desc = { 32, XSIMD_STRING("n2sv_32"), {236, 0, 136, 0}, &GENUS, 0, 1, 0, 0 };
703
704 void XSIMD(codelet_n2sv_32) (planner *p) {
705 X(kdft_register) (p, n2sv_32, &desc);
706 }
707
708 #else /* HAVE_FMA */
709
710 /* Generated by: ../../../genfft/gen_notw.native -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n2sv_32 -with-ostride 1 -include n2s.h -store-multiple 4 */
711
712 /*
713 * This function contains 372 FP additions, 84 FP multiplications,
714 * (or, 340 additions, 52 multiplications, 32 fused multiply/add),
715 * 130 stack variables, 7 constants, and 144 memory accesses
716 */
717 #include "n2s.h"
718
719 static void n2sv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
720 {
721 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
722 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
723 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
724 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
725 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
726 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
727 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
728 {
729 INT i;
730 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)) {
731 V T7, T4r, T4Z, T18, T1z, T3t, T3T, T2T, Te, T1f, T50, T4s, T2W, T3u, T1G;
732 V T3U, Tm, T1n, T1O, T2Z, T3y, T3X, T4w, T53, Tt, T1u, T1V, T2Y, T3B, T3W;
733 V T4z, T52, T2t, T3L, T3O, T2K, TR, TY, T5F, T5G, T5H, T5I, T4R, T5j, T2E;
734 V T3P, T4W, T5k, T2N, T3M, T22, T3E, T3H, T2j, TC, TJ, T5A, T5B, T5C, T5D;
735 V T4G, T5g, T2d, T3F, T4L, T5h, T2m, T3I;
736 {
737 V T3, T1x, T14, T2S, T6, T2R, T17, T1y;
738 {
739 V T1, T2, T12, T13;
740 T1 = LD(&(ri[0]), ivs, &(ri[0]));
741 T2 = LD(&(ri[WS(is, 16)]), ivs, &(ri[0]));
742 T3 = VADD(T1, T2);
743 T1x = VSUB(T1, T2);
744 T12 = LD(&(ii[0]), ivs, &(ii[0]));
745 T13 = LD(&(ii[WS(is, 16)]), ivs, &(ii[0]));
746 T14 = VADD(T12, T13);
747 T2S = VSUB(T12, T13);
748 }
749 {
750 V T4, T5, T15, T16;
751 T4 = LD(&(ri[WS(is, 8)]), ivs, &(ri[0]));
752 T5 = LD(&(ri[WS(is, 24)]), ivs, &(ri[0]));
753 T6 = VADD(T4, T5);
754 T2R = VSUB(T4, T5);
755 T15 = LD(&(ii[WS(is, 8)]), ivs, &(ii[0]));
756 T16 = LD(&(ii[WS(is, 24)]), ivs, &(ii[0]));
757 T17 = VADD(T15, T16);
758 T1y = VSUB(T15, T16);
759 }
760 T7 = VADD(T3, T6);
761 T4r = VSUB(T3, T6);
762 T4Z = VSUB(T14, T17);
763 T18 = VADD(T14, T17);
764 T1z = VSUB(T1x, T1y);
765 T3t = VADD(T1x, T1y);
766 T3T = VSUB(T2S, T2R);
767 T2T = VADD(T2R, T2S);
768 }
769 {
770 V Ta, T1B, T1b, T1A, Td, T1D, T1e, T1E;
771 {
772 V T8, T9, T19, T1a;
773 T8 = LD(&(ri[WS(is, 4)]), ivs, &(ri[0]));
774 T9 = LD(&(ri[WS(is, 20)]), ivs, &(ri[0]));
775 Ta = VADD(T8, T9);
776 T1B = VSUB(T8, T9);
777 T19 = LD(&(ii[WS(is, 4)]), ivs, &(ii[0]));
778 T1a = LD(&(ii[WS(is, 20)]), ivs, &(ii[0]));
779 T1b = VADD(T19, T1a);
780 T1A = VSUB(T19, T1a);
781 }
782 {
783 V Tb, Tc, T1c, T1d;
784 Tb = LD(&(ri[WS(is, 28)]), ivs, &(ri[0]));
785 Tc = LD(&(ri[WS(is, 12)]), ivs, &(ri[0]));
786 Td = VADD(Tb, Tc);
787 T1D = VSUB(Tb, Tc);
788 T1c = LD(&(ii[WS(is, 28)]), ivs, &(ii[0]));
789 T1d = LD(&(ii[WS(is, 12)]), ivs, &(ii[0]));
790 T1e = VADD(T1c, T1d);
791 T1E = VSUB(T1c, T1d);
792 }
793 Te = VADD(Ta, Td);
794 T1f = VADD(T1b, T1e);
795 T50 = VSUB(Td, Ta);
796 T4s = VSUB(T1b, T1e);
797 {
798 V T2U, T2V, T1C, T1F;
799 T2U = VSUB(T1D, T1E);
800 T2V = VADD(T1B, T1A);
801 T2W = VMUL(LDK(KP707106781), VSUB(T2U, T2V));
802 T3u = VMUL(LDK(KP707106781), VADD(T2V, T2U));
803 T1C = VSUB(T1A, T1B);
804 T1F = VADD(T1D, T1E);
805 T1G = VMUL(LDK(KP707106781), VSUB(T1C, T1F));
806 T3U = VMUL(LDK(KP707106781), VADD(T1C, T1F));
807 }
808 }
809 {
810 V Ti, T1L, T1j, T1J, Tl, T1I, T1m, T1M, T1K, T1N;
811 {
812 V Tg, Th, T1h, T1i;
813 Tg = LD(&(ri[WS(is, 2)]), ivs, &(ri[0]));
814 Th = LD(&(ri[WS(is, 18)]), ivs, &(ri[0]));
815 Ti = VADD(Tg, Th);
816 T1L = VSUB(Tg, Th);
817 T1h = LD(&(ii[WS(is, 2)]), ivs, &(ii[0]));
818 T1i = LD(&(ii[WS(is, 18)]), ivs, &(ii[0]));
819 T1j = VADD(T1h, T1i);
820 T1J = VSUB(T1h, T1i);
821 }
822 {
823 V Tj, Tk, T1k, T1l;
824 Tj = LD(&(ri[WS(is, 10)]), ivs, &(ri[0]));
825 Tk = LD(&(ri[WS(is, 26)]), ivs, &(ri[0]));
826 Tl = VADD(Tj, Tk);
827 T1I = VSUB(Tj, Tk);
828 T1k = LD(&(ii[WS(is, 10)]), ivs, &(ii[0]));
829 T1l = LD(&(ii[WS(is, 26)]), ivs, &(ii[0]));
830 T1m = VADD(T1k, T1l);
831 T1M = VSUB(T1k, T1l);
832 }
833 Tm = VADD(Ti, Tl);
834 T1n = VADD(T1j, T1m);
835 T1K = VADD(T1I, T1J);
836 T1N = VSUB(T1L, T1M);
837 T1O = VFNMS(LDK(KP923879532), T1N, VMUL(LDK(KP382683432), T1K));
838 T2Z = VFMA(LDK(KP923879532), T1K, VMUL(LDK(KP382683432), T1N));
839 {
840 V T3w, T3x, T4u, T4v;
841 T3w = VSUB(T1J, T1I);
842 T3x = VADD(T1L, T1M);
843 T3y = VFNMS(LDK(KP382683432), T3x, VMUL(LDK(KP923879532), T3w));
844 T3X = VFMA(LDK(KP382683432), T3w, VMUL(LDK(KP923879532), T3x));
845 T4u = VSUB(T1j, T1m);
846 T4v = VSUB(Ti, Tl);
847 T4w = VSUB(T4u, T4v);
848 T53 = VADD(T4v, T4u);
849 }
850 }
851 {
852 V Tp, T1S, T1q, T1Q, Ts, T1P, T1t, T1T, T1R, T1U;
853 {
854 V Tn, To, T1o, T1p;
855 Tn = LD(&(ri[WS(is, 30)]), ivs, &(ri[0]));
856 To = LD(&(ri[WS(is, 14)]), ivs, &(ri[0]));
857 Tp = VADD(Tn, To);
858 T1S = VSUB(Tn, To);
859 T1o = LD(&(ii[WS(is, 30)]), ivs, &(ii[0]));
860 T1p = LD(&(ii[WS(is, 14)]), ivs, &(ii[0]));
861 T1q = VADD(T1o, T1p);
862 T1Q = VSUB(T1o, T1p);
863 }
864 {
865 V Tq, Tr, T1r, T1s;
866 Tq = LD(&(ri[WS(is, 6)]), ivs, &(ri[0]));
867 Tr = LD(&(ri[WS(is, 22)]), ivs, &(ri[0]));
868 Ts = VADD(Tq, Tr);
869 T1P = VSUB(Tq, Tr);
870 T1r = LD(&(ii[WS(is, 6)]), ivs, &(ii[0]));
871 T1s = LD(&(ii[WS(is, 22)]), ivs, &(ii[0]));
872 T1t = VADD(T1r, T1s);
873 T1T = VSUB(T1r, T1s);
874 }
875 Tt = VADD(Tp, Ts);
876 T1u = VADD(T1q, T1t);
877 T1R = VADD(T1P, T1Q);
878 T1U = VSUB(T1S, T1T);
879 T1V = VFMA(LDK(KP382683432), T1R, VMUL(LDK(KP923879532), T1U));
880 T2Y = VFNMS(LDK(KP923879532), T1R, VMUL(LDK(KP382683432), T1U));
881 {
882 V T3z, T3A, T4x, T4y;
883 T3z = VSUB(T1Q, T1P);
884 T3A = VADD(T1S, T1T);
885 T3B = VFMA(LDK(KP923879532), T3z, VMUL(LDK(KP382683432), T3A));
886 T3W = VFNMS(LDK(KP382683432), T3z, VMUL(LDK(KP923879532), T3A));
887 T4x = VSUB(Tp, Ts);
888 T4y = VSUB(T1q, T1t);
889 T4z = VADD(T4x, T4y);
890 T52 = VSUB(T4x, T4y);
891 }
892 }
893 {
894 V TN, T2p, T2J, T4S, TQ, T2G, T2s, T4T, TU, T2x, T2w, T4O, TX, T2z, T2C;
895 V T4P;
896 {
897 V TL, TM, T2H, T2I;
898 TL = LD(&(ri[WS(is, 31)]), ivs, &(ri[WS(is, 1)]));
899 TM = LD(&(ri[WS(is, 15)]), ivs, &(ri[WS(is, 1)]));
900 TN = VADD(TL, TM);
901 T2p = VSUB(TL, TM);
902 T2H = LD(&(ii[WS(is, 31)]), ivs, &(ii[WS(is, 1)]));
903 T2I = LD(&(ii[WS(is, 15)]), ivs, &(ii[WS(is, 1)]));
904 T2J = VSUB(T2H, T2I);
905 T4S = VADD(T2H, T2I);
906 }
907 {
908 V TO, TP, T2q, T2r;
909 TO = LD(&(ri[WS(is, 7)]), ivs, &(ri[WS(is, 1)]));
910 TP = LD(&(ri[WS(is, 23)]), ivs, &(ri[WS(is, 1)]));
911 TQ = VADD(TO, TP);
912 T2G = VSUB(TO, TP);
913 T2q = LD(&(ii[WS(is, 7)]), ivs, &(ii[WS(is, 1)]));
914 T2r = LD(&(ii[WS(is, 23)]), ivs, &(ii[WS(is, 1)]));
915 T2s = VSUB(T2q, T2r);
916 T4T = VADD(T2q, T2r);
917 }
918 {
919 V TS, TT, T2u, T2v;
920 TS = LD(&(ri[WS(is, 3)]), ivs, &(ri[WS(is, 1)]));
921 TT = LD(&(ri[WS(is, 19)]), ivs, &(ri[WS(is, 1)]));
922 TU = VADD(TS, TT);
923 T2x = VSUB(TS, TT);
924 T2u = LD(&(ii[WS(is, 3)]), ivs, &(ii[WS(is, 1)]));
925 T2v = LD(&(ii[WS(is, 19)]), ivs, &(ii[WS(is, 1)]));
926 T2w = VSUB(T2u, T2v);
927 T4O = VADD(T2u, T2v);
928 }
929 {
930 V TV, TW, T2A, T2B;
931 TV = LD(&(ri[WS(is, 27)]), ivs, &(ri[WS(is, 1)]));
932 TW = LD(&(ri[WS(is, 11)]), ivs, &(ri[WS(is, 1)]));
933 TX = VADD(TV, TW);
934 T2z = VSUB(TV, TW);
935 T2A = LD(&(ii[WS(is, 27)]), ivs, &(ii[WS(is, 1)]));
936 T2B = LD(&(ii[WS(is, 11)]), ivs, &(ii[WS(is, 1)]));
937 T2C = VSUB(T2A, T2B);
938 T4P = VADD(T2A, T2B);
939 }
940 T2t = VSUB(T2p, T2s);
941 T3L = VADD(T2p, T2s);
942 T3O = VSUB(T2J, T2G);
943 T2K = VADD(T2G, T2J);
944 TR = VADD(TN, TQ);
945 TY = VADD(TU, TX);
946 T5F = VSUB(TR, TY);
947 {
948 V T4N, T4Q, T2y, T2D;
949 T5G = VADD(T4S, T4T);
950 T5H = VADD(T4O, T4P);
951 T5I = VSUB(T5G, T5H);
952 T4N = VSUB(TN, TQ);
953 T4Q = VSUB(T4O, T4P);
954 T4R = VSUB(T4N, T4Q);
955 T5j = VADD(T4N, T4Q);
956 T2y = VSUB(T2w, T2x);
957 T2D = VADD(T2z, T2C);
958 T2E = VMUL(LDK(KP707106781), VSUB(T2y, T2D));
959 T3P = VMUL(LDK(KP707106781), VADD(T2y, T2D));
960 {
961 V T4U, T4V, T2L, T2M;
962 T4U = VSUB(T4S, T4T);
963 T4V = VSUB(TX, TU);
964 T4W = VSUB(T4U, T4V);
965 T5k = VADD(T4V, T4U);
966 T2L = VSUB(T2z, T2C);
967 T2M = VADD(T2x, T2w);
968 T2N = VMUL(LDK(KP707106781), VSUB(T2L, T2M));
969 T3M = VMUL(LDK(KP707106781), VADD(T2M, T2L));
970 }
971 }
972 }
973 {
974 V Ty, T2f, T21, T4C, TB, T1Y, T2i, T4D, TF, T28, T2b, T4I, TI, T23, T26;
975 V T4J;
976 {
977 V Tw, Tx, T1Z, T20;
978 Tw = LD(&(ri[WS(is, 1)]), ivs, &(ri[WS(is, 1)]));
979 Tx = LD(&(ri[WS(is, 17)]), ivs, &(ri[WS(is, 1)]));
980 Ty = VADD(Tw, Tx);
981 T2f = VSUB(Tw, Tx);
982 T1Z = LD(&(ii[WS(is, 1)]), ivs, &(ii[WS(is, 1)]));
983 T20 = LD(&(ii[WS(is, 17)]), ivs, &(ii[WS(is, 1)]));
984 T21 = VSUB(T1Z, T20);
985 T4C = VADD(T1Z, T20);
986 }
987 {
988 V Tz, TA, T2g, T2h;
989 Tz = LD(&(ri[WS(is, 9)]), ivs, &(ri[WS(is, 1)]));
990 TA = LD(&(ri[WS(is, 25)]), ivs, &(ri[WS(is, 1)]));
991 TB = VADD(Tz, TA);
992 T1Y = VSUB(Tz, TA);
993 T2g = LD(&(ii[WS(is, 9)]), ivs, &(ii[WS(is, 1)]));
994 T2h = LD(&(ii[WS(is, 25)]), ivs, &(ii[WS(is, 1)]));
995 T2i = VSUB(T2g, T2h);
996 T4D = VADD(T2g, T2h);
997 }
998 {
999 V TD, TE, T29, T2a;
1000 TD = LD(&(ri[WS(is, 5)]), ivs, &(ri[WS(is, 1)]));
1001 TE = LD(&(ri[WS(is, 21)]), ivs, &(ri[WS(is, 1)]));
1002 TF = VADD(TD, TE);
1003 T28 = VSUB(TD, TE);
1004 T29 = LD(&(ii[WS(is, 5)]), ivs, &(ii[WS(is, 1)]));
1005 T2a = LD(&(ii[WS(is, 21)]), ivs, &(ii[WS(is, 1)]));
1006 T2b = VSUB(T29, T2a);
1007 T4I = VADD(T29, T2a);
1008 }
1009 {
1010 V TG, TH, T24, T25;
1011 TG = LD(&(ri[WS(is, 29)]), ivs, &(ri[WS(is, 1)]));
1012 TH = LD(&(ri[WS(is, 13)]), ivs, &(ri[WS(is, 1)]));
1013 TI = VADD(TG, TH);
1014 T23 = VSUB(TG, TH);
1015 T24 = LD(&(ii[WS(is, 29)]), ivs, &(ii[WS(is, 1)]));
1016 T25 = LD(&(ii[WS(is, 13)]), ivs, &(ii[WS(is, 1)]));
1017 T26 = VSUB(T24, T25);
1018 T4J = VADD(T24, T25);
1019 }
1020 T22 = VADD(T1Y, T21);
1021 T3E = VADD(T2f, T2i);
1022 T3H = VSUB(T21, T1Y);
1023 T2j = VSUB(T2f, T2i);
1024 TC = VADD(Ty, TB);
1025 TJ = VADD(TF, TI);
1026 T5A = VSUB(TC, TJ);
1027 {
1028 V T4E, T4F, T27, T2c;
1029 T5B = VADD(T4C, T4D);
1030 T5C = VADD(T4I, T4J);
1031 T5D = VSUB(T5B, T5C);
1032 T4E = VSUB(T4C, T4D);
1033 T4F = VSUB(TI, TF);
1034 T4G = VSUB(T4E, T4F);
1035 T5g = VADD(T4F, T4E);
1036 T27 = VSUB(T23, T26);
1037 T2c = VADD(T28, T2b);
1038 T2d = VMUL(LDK(KP707106781), VSUB(T27, T2c));
1039 T3F = VMUL(LDK(KP707106781), VADD(T2c, T27));
1040 {
1041 V T4H, T4K, T2k, T2l;
1042 T4H = VSUB(Ty, TB);
1043 T4K = VSUB(T4I, T4J);
1044 T4L = VSUB(T4H, T4K);
1045 T5h = VADD(T4H, T4K);
1046 T2k = VSUB(T2b, T28);
1047 T2l = VADD(T23, T26);
1048 T2m = VMUL(LDK(KP707106781), VSUB(T2k, T2l));
1049 T3I = VMUL(LDK(KP707106781), VADD(T2k, T2l));
1050 }
1051 }
1052 }
1053 {
1054 V T61, T62, T63, T64, T65, T66, T67, T68, T69, T6a, T6b, T6c, T6d, T6e, T6f;
1055 V T6g, T6h, T6i, T6j, T6k, T6l, T6m, T6n, T6o, T6p, T6q, T6r, T6s, T6t, T6u;
1056 V T6v, T6w;
1057 {
1058 V T4B, T57, T5a, T5c, T4Y, T56, T55, T5b;
1059 {
1060 V T4t, T4A, T58, T59;
1061 T4t = VSUB(T4r, T4s);
1062 T4A = VMUL(LDK(KP707106781), VSUB(T4w, T4z));
1063 T4B = VADD(T4t, T4A);
1064 T57 = VSUB(T4t, T4A);
1065 T58 = VFNMS(LDK(KP923879532), T4L, VMUL(LDK(KP382683432), T4G));
1066 T59 = VFMA(LDK(KP382683432), T4W, VMUL(LDK(KP923879532), T4R));
1067 T5a = VSUB(T58, T59);
1068 T5c = VADD(T58, T59);
1069 }
1070 {
1071 V T4M, T4X, T51, T54;
1072 T4M = VFMA(LDK(KP923879532), T4G, VMUL(LDK(KP382683432), T4L));
1073 T4X = VFNMS(LDK(KP923879532), T4W, VMUL(LDK(KP382683432), T4R));
1074 T4Y = VADD(T4M, T4X);
1075 T56 = VSUB(T4X, T4M);
1076 T51 = VSUB(T4Z, T50);
1077 T54 = VMUL(LDK(KP707106781), VSUB(T52, T53));
1078 T55 = VSUB(T51, T54);
1079 T5b = VADD(T51, T54);
1080 }
1081 T61 = VSUB(T4B, T4Y);
1082 STM4(&(ro[22]), T61, ovs, &(ro[0]));
1083 T62 = VSUB(T5b, T5c);
1084 STM4(&(io[22]), T62, ovs, &(io[0]));
1085 T63 = VADD(T4B, T4Y);
1086 STM4(&(ro[6]), T63, ovs, &(ro[0]));
1087 T64 = VADD(T5b, T5c);
1088 STM4(&(io[6]), T64, ovs, &(io[0]));
1089 T65 = VSUB(T55, T56);
1090 STM4(&(io[30]), T65, ovs, &(io[0]));
1091 T66 = VSUB(T57, T5a);
1092 STM4(&(ro[30]), T66, ovs, &(ro[0]));
1093 T67 = VADD(T55, T56);
1094 STM4(&(io[14]), T67, ovs, &(io[0]));
1095 T68 = VADD(T57, T5a);
1096 STM4(&(ro[14]), T68, ovs, &(ro[0]));
1097 }
1098 {
1099 V T5f, T5r, T5u, T5w, T5m, T5q, T5p, T5v;
1100 {
1101 V T5d, T5e, T5s, T5t;
1102 T5d = VADD(T4r, T4s);
1103 T5e = VMUL(LDK(KP707106781), VADD(T53, T52));
1104 T5f = VADD(T5d, T5e);
1105 T5r = VSUB(T5d, T5e);
1106 T5s = VFNMS(LDK(KP382683432), T5h, VMUL(LDK(KP923879532), T5g));
1107 T5t = VFMA(LDK(KP923879532), T5k, VMUL(LDK(KP382683432), T5j));
1108 T5u = VSUB(T5s, T5t);
1109 T5w = VADD(T5s, T5t);
1110 }
1111 {
1112 V T5i, T5l, T5n, T5o;
1113 T5i = VFMA(LDK(KP382683432), T5g, VMUL(LDK(KP923879532), T5h));
1114 T5l = VFNMS(LDK(KP382683432), T5k, VMUL(LDK(KP923879532), T5j));
1115 T5m = VADD(T5i, T5l);
1116 T5q = VSUB(T5l, T5i);
1117 T5n = VADD(T50, T4Z);
1118 T5o = VMUL(LDK(KP707106781), VADD(T4w, T4z));
1119 T5p = VSUB(T5n, T5o);
1120 T5v = VADD(T5n, T5o);
1121 }
1122 T69 = VSUB(T5f, T5m);
1123 STM4(&(ro[18]), T69, ovs, &(ro[0]));
1124 T6a = VSUB(T5v, T5w);
1125 STM4(&(io[18]), T6a, ovs, &(io[0]));
1126 T6b = VADD(T5f, T5m);
1127 STM4(&(ro[2]), T6b, ovs, &(ro[0]));
1128 T6c = VADD(T5v, T5w);
1129 STM4(&(io[2]), T6c, ovs, &(io[0]));
1130 T6d = VSUB(T5p, T5q);
1131 STM4(&(io[26]), T6d, ovs, &(io[0]));
1132 T6e = VSUB(T5r, T5u);
1133 STM4(&(ro[26]), T6e, ovs, &(ro[0]));
1134 T6f = VADD(T5p, T5q);
1135 STM4(&(io[10]), T6f, ovs, &(io[0]));
1136 T6g = VADD(T5r, T5u);
1137 STM4(&(ro[10]), T6g, ovs, &(ro[0]));
1138 }
1139 {
1140 V T5z, T5P, T5S, T5U, T5K, T5O, T5N, T5T;
1141 {
1142 V T5x, T5y, T5Q, T5R;
1143 T5x = VSUB(T7, Te);
1144 T5y = VSUB(T1n, T1u);
1145 T5z = VADD(T5x, T5y);
1146 T5P = VSUB(T5x, T5y);
1147 T5Q = VSUB(T5D, T5A);
1148 T5R = VADD(T5F, T5I);
1149 T5S = VMUL(LDK(KP707106781), VSUB(T5Q, T5R));
1150 T5U = VMUL(LDK(KP707106781), VADD(T5Q, T5R));
1151 }
1152 {
1153 V T5E, T5J, T5L, T5M;
1154 T5E = VADD(T5A, T5D);
1155 T5J = VSUB(T5F, T5I);
1156 T5K = VMUL(LDK(KP707106781), VADD(T5E, T5J));
1157 T5O = VMUL(LDK(KP707106781), VSUB(T5J, T5E));
1158 T5L = VSUB(T18, T1f);
1159 T5M = VSUB(Tt, Tm);
1160 T5N = VSUB(T5L, T5M);
1161 T5T = VADD(T5M, T5L);
1162 }
1163 T6h = VSUB(T5z, T5K);
1164 STM4(&(ro[20]), T6h, ovs, &(ro[0]));
1165 T6i = VSUB(T5T, T5U);
1166 STM4(&(io[20]), T6i, ovs, &(io[0]));
1167 T6j = VADD(T5z, T5K);
1168 STM4(&(ro[4]), T6j, ovs, &(ro[0]));
1169 T6k = VADD(T5T, T5U);
1170 STM4(&(io[4]), T6k, ovs, &(io[0]));
1171 T6l = VSUB(T5N, T5O);
1172 STM4(&(io[28]), T6l, ovs, &(io[0]));
1173 T6m = VSUB(T5P, T5S);
1174 STM4(&(ro[28]), T6m, ovs, &(ro[0]));
1175 T6n = VADD(T5N, T5O);
1176 STM4(&(io[12]), T6n, ovs, &(io[0]));
1177 T6o = VADD(T5P, T5S);
1178 STM4(&(ro[12]), T6o, ovs, &(ro[0]));
1179 }
1180 {
1181 V Tv, T5V, T5Y, T60, T10, T11, T1w, T5Z;
1182 {
1183 V Tf, Tu, T5W, T5X;
1184 Tf = VADD(T7, Te);
1185 Tu = VADD(Tm, Tt);
1186 Tv = VADD(Tf, Tu);
1187 T5V = VSUB(Tf, Tu);
1188 T5W = VADD(T5B, T5C);
1189 T5X = VADD(T5G, T5H);
1190 T5Y = VSUB(T5W, T5X);
1191 T60 = VADD(T5W, T5X);
1192 }
1193 {
1194 V TK, TZ, T1g, T1v;
1195 TK = VADD(TC, TJ);
1196 TZ = VADD(TR, TY);
1197 T10 = VADD(TK, TZ);
1198 T11 = VSUB(TZ, TK);
1199 T1g = VADD(T18, T1f);
1200 T1v = VADD(T1n, T1u);
1201 T1w = VSUB(T1g, T1v);
1202 T5Z = VADD(T1g, T1v);
1203 }
1204 T6p = VSUB(Tv, T10);
1205 STM4(&(ro[16]), T6p, ovs, &(ro[0]));
1206 T6q = VSUB(T5Z, T60);
1207 STM4(&(io[16]), T6q, ovs, &(io[0]));
1208 T6r = VADD(Tv, T10);
1209 STM4(&(ro[0]), T6r, ovs, &(ro[0]));
1210 T6s = VADD(T5Z, T60);
1211 STM4(&(io[0]), T6s, ovs, &(io[0]));
1212 T6t = VADD(T11, T1w);
1213 STM4(&(io[8]), T6t, ovs, &(io[0]));
1214 T6u = VADD(T5V, T5Y);
1215 STM4(&(ro[8]), T6u, ovs, &(ro[0]));
1216 T6v = VSUB(T1w, T11);
1217 STM4(&(io[24]), T6v, ovs, &(io[0]));
1218 T6w = VSUB(T5V, T5Y);
1219 STM4(&(ro[24]), T6w, ovs, &(ro[0]));
1220 }
1221 {
1222 V T6x, T6y, T6z, T6A, T6B, T6C, T6D, T6E;
1223 {
1224 V T1X, T33, T31, T37, T2o, T34, T2P, T35;
1225 {
1226 V T1H, T1W, T2X, T30;
1227 T1H = VSUB(T1z, T1G);
1228 T1W = VSUB(T1O, T1V);
1229 T1X = VADD(T1H, T1W);
1230 T33 = VSUB(T1H, T1W);
1231 T2X = VSUB(T2T, T2W);
1232 T30 = VSUB(T2Y, T2Z);
1233 T31 = VSUB(T2X, T30);
1234 T37 = VADD(T2X, T30);
1235 }
1236 {
1237 V T2e, T2n, T2F, T2O;
1238 T2e = VSUB(T22, T2d);
1239 T2n = VSUB(T2j, T2m);
1240 T2o = VFMA(LDK(KP980785280), T2e, VMUL(LDK(KP195090322), T2n));
1241 T34 = VFNMS(LDK(KP980785280), T2n, VMUL(LDK(KP195090322), T2e));
1242 T2F = VSUB(T2t, T2E);
1243 T2O = VSUB(T2K, T2N);
1244 T2P = VFNMS(LDK(KP980785280), T2O, VMUL(LDK(KP195090322), T2F));
1245 T35 = VFMA(LDK(KP195090322), T2O, VMUL(LDK(KP980785280), T2F));
1246 }
1247 {
1248 V T2Q, T38, T32, T36;
1249 T2Q = VADD(T2o, T2P);
1250 T6x = VSUB(T1X, T2Q);
1251 STM4(&(ro[23]), T6x, ovs, &(ro[1]));
1252 T6y = VADD(T1X, T2Q);
1253 STM4(&(ro[7]), T6y, ovs, &(ro[1]));
1254 T38 = VADD(T34, T35);
1255 T6z = VSUB(T37, T38);
1256 STM4(&(io[23]), T6z, ovs, &(io[1]));
1257 T6A = VADD(T37, T38);
1258 STM4(&(io[7]), T6A, ovs, &(io[1]));
1259 T32 = VSUB(T2P, T2o);
1260 T6B = VSUB(T31, T32);
1261 STM4(&(io[31]), T6B, ovs, &(io[1]));
1262 T6C = VADD(T31, T32);
1263 STM4(&(io[15]), T6C, ovs, &(io[1]));
1264 T36 = VSUB(T34, T35);
1265 T6D = VSUB(T33, T36);
1266 STM4(&(ro[31]), T6D, ovs, &(ro[1]));
1267 T6E = VADD(T33, T36);
1268 STM4(&(ro[15]), T6E, ovs, &(ro[1]));
1269 }
1270 }
1271 {
1272 V T3D, T41, T3Z, T45, T3K, T42, T3R, T43;
1273 {
1274 V T3v, T3C, T3V, T3Y;
1275 T3v = VSUB(T3t, T3u);
1276 T3C = VSUB(T3y, T3B);
1277 T3D = VADD(T3v, T3C);
1278 T41 = VSUB(T3v, T3C);
1279 T3V = VSUB(T3T, T3U);
1280 T3Y = VSUB(T3W, T3X);
1281 T3Z = VSUB(T3V, T3Y);
1282 T45 = VADD(T3V, T3Y);
1283 }
1284 {
1285 V T3G, T3J, T3N, T3Q;
1286 T3G = VSUB(T3E, T3F);
1287 T3J = VSUB(T3H, T3I);
1288 T3K = VFMA(LDK(KP555570233), T3G, VMUL(LDK(KP831469612), T3J));
1289 T42 = VFNMS(LDK(KP831469612), T3G, VMUL(LDK(KP555570233), T3J));
1290 T3N = VSUB(T3L, T3M);
1291 T3Q = VSUB(T3O, T3P);
1292 T3R = VFNMS(LDK(KP831469612), T3Q, VMUL(LDK(KP555570233), T3N));
1293 T43 = VFMA(LDK(KP831469612), T3N, VMUL(LDK(KP555570233), T3Q));
1294 }
1295 {
1296 V T3S, T6F, T6G, T46, T6H, T6I;
1297 T3S = VADD(T3K, T3R);
1298 T6F = VSUB(T3D, T3S);
1299 STM4(&(ro[21]), T6F, ovs, &(ro[1]));
1300 STN4(&(ro[20]), T6h, T6F, T61, T6x, ovs);
1301 T6G = VADD(T3D, T3S);
1302 STM4(&(ro[5]), T6G, ovs, &(ro[1]));
1303 STN4(&(ro[4]), T6j, T6G, T63, T6y, ovs);
1304 T46 = VADD(T42, T43);
1305 T6H = VSUB(T45, T46);
1306 STM4(&(io[21]), T6H, ovs, &(io[1]));
1307 STN4(&(io[20]), T6i, T6H, T62, T6z, ovs);
1308 T6I = VADD(T45, T46);
1309 STM4(&(io[5]), T6I, ovs, &(io[1]));
1310 STN4(&(io[4]), T6k, T6I, T64, T6A, ovs);
1311 }
1312 {
1313 V T40, T6J, T6K, T44, T6L, T6M;
1314 T40 = VSUB(T3R, T3K);
1315 T6J = VSUB(T3Z, T40);
1316 STM4(&(io[29]), T6J, ovs, &(io[1]));
1317 STN4(&(io[28]), T6l, T6J, T65, T6B, ovs);
1318 T6K = VADD(T3Z, T40);
1319 STM4(&(io[13]), T6K, ovs, &(io[1]));
1320 STN4(&(io[12]), T6n, T6K, T67, T6C, ovs);
1321 T44 = VSUB(T42, T43);
1322 T6L = VSUB(T41, T44);
1323 STM4(&(ro[29]), T6L, ovs, &(ro[1]));
1324 STN4(&(ro[28]), T6m, T6L, T66, T6D, ovs);
1325 T6M = VADD(T41, T44);
1326 STM4(&(ro[13]), T6M, ovs, &(ro[1]));
1327 STN4(&(ro[12]), T6o, T6M, T68, T6E, ovs);
1328 }
1329 }
1330 }
1331 {
1332 V T6N, T6O, T6P, T6Q, T6R, T6S, T6T, T6U;
1333 {
1334 V T49, T4l, T4j, T4p, T4c, T4m, T4f, T4n;
1335 {
1336 V T47, T48, T4h, T4i;
1337 T47 = VADD(T3t, T3u);
1338 T48 = VADD(T3X, T3W);
1339 T49 = VADD(T47, T48);
1340 T4l = VSUB(T47, T48);
1341 T4h = VADD(T3T, T3U);
1342 T4i = VADD(T3y, T3B);
1343 T4j = VSUB(T4h, T4i);
1344 T4p = VADD(T4h, T4i);
1345 }
1346 {
1347 V T4a, T4b, T4d, T4e;
1348 T4a = VADD(T3E, T3F);
1349 T4b = VADD(T3H, T3I);
1350 T4c = VFMA(LDK(KP980785280), T4a, VMUL(LDK(KP195090322), T4b));
1351 T4m = VFNMS(LDK(KP195090322), T4a, VMUL(LDK(KP980785280), T4b));
1352 T4d = VADD(T3L, T3M);
1353 T4e = VADD(T3O, T3P);
1354 T4f = VFNMS(LDK(KP195090322), T4e, VMUL(LDK(KP980785280), T4d));
1355 T4n = VFMA(LDK(KP195090322), T4d, VMUL(LDK(KP980785280), T4e));
1356 }
1357 {
1358 V T4g, T4q, T4k, T4o;
1359 T4g = VADD(T4c, T4f);
1360 T6N = VSUB(T49, T4g);
1361 STM4(&(ro[17]), T6N, ovs, &(ro[1]));
1362 T6O = VADD(T49, T4g);
1363 STM4(&(ro[1]), T6O, ovs, &(ro[1]));
1364 T4q = VADD(T4m, T4n);
1365 T6P = VSUB(T4p, T4q);
1366 STM4(&(io[17]), T6P, ovs, &(io[1]));
1367 T6Q = VADD(T4p, T4q);
1368 STM4(&(io[1]), T6Q, ovs, &(io[1]));
1369 T4k = VSUB(T4f, T4c);
1370 T6R = VSUB(T4j, T4k);
1371 STM4(&(io[25]), T6R, ovs, &(io[1]));
1372 T6S = VADD(T4j, T4k);
1373 STM4(&(io[9]), T6S, ovs, &(io[1]));
1374 T4o = VSUB(T4m, T4n);
1375 T6T = VSUB(T4l, T4o);
1376 STM4(&(ro[25]), T6T, ovs, &(ro[1]));
1377 T6U = VADD(T4l, T4o);
1378 STM4(&(ro[9]), T6U, ovs, &(ro[1]));
1379 }
1380 }
1381 {
1382 V T3b, T3n, T3l, T3r, T3e, T3o, T3h, T3p;
1383 {
1384 V T39, T3a, T3j, T3k;
1385 T39 = VADD(T1z, T1G);
1386 T3a = VADD(T2Z, T2Y);
1387 T3b = VADD(T39, T3a);
1388 T3n = VSUB(T39, T3a);
1389 T3j = VADD(T2T, T2W);
1390 T3k = VADD(T1O, T1V);
1391 T3l = VSUB(T3j, T3k);
1392 T3r = VADD(T3j, T3k);
1393 }
1394 {
1395 V T3c, T3d, T3f, T3g;
1396 T3c = VADD(T22, T2d);
1397 T3d = VADD(T2j, T2m);
1398 T3e = VFMA(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d));
1399 T3o = VFNMS(LDK(KP555570233), T3d, VMUL(LDK(KP831469612), T3c));
1400 T3f = VADD(T2t, T2E);
1401 T3g = VADD(T2K, T2N);
1402 T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f));
1403 T3p = VFMA(LDK(KP831469612), T3g, VMUL(LDK(KP555570233), T3f));
1404 }
1405 {
1406 V T3i, T6V, T6W, T3s, T6X, T6Y;
1407 T3i = VADD(T3e, T3h);
1408 T6V = VSUB(T3b, T3i);
1409 STM4(&(ro[19]), T6V, ovs, &(ro[1]));
1410 STN4(&(ro[16]), T6p, T6N, T69, T6V, ovs);
1411 T6W = VADD(T3b, T3i);
1412 STM4(&(ro[3]), T6W, ovs, &(ro[1]));
1413 STN4(&(ro[0]), T6r, T6O, T6b, T6W, ovs);
1414 T3s = VADD(T3o, T3p);
1415 T6X = VSUB(T3r, T3s);
1416 STM4(&(io[19]), T6X, ovs, &(io[1]));
1417 STN4(&(io[16]), T6q, T6P, T6a, T6X, ovs);
1418 T6Y = VADD(T3r, T3s);
1419 STM4(&(io[3]), T6Y, ovs, &(io[1]));
1420 STN4(&(io[0]), T6s, T6Q, T6c, T6Y, ovs);
1421 }
1422 {
1423 V T3m, T6Z, T70, T3q, T71, T72;
1424 T3m = VSUB(T3h, T3e);
1425 T6Z = VSUB(T3l, T3m);
1426 STM4(&(io[27]), T6Z, ovs, &(io[1]));
1427 STN4(&(io[24]), T6v, T6R, T6d, T6Z, ovs);
1428 T70 = VADD(T3l, T3m);
1429 STM4(&(io[11]), T70, ovs, &(io[1]));
1430 STN4(&(io[8]), T6t, T6S, T6f, T70, ovs);
1431 T3q = VSUB(T3o, T3p);
1432 T71 = VSUB(T3n, T3q);
1433 STM4(&(ro[27]), T71, ovs, &(ro[1]));
1434 STN4(&(ro[24]), T6w, T6T, T6e, T71, ovs);
1435 T72 = VADD(T3n, T3q);
1436 STM4(&(ro[11]), T72, ovs, &(ro[1]));
1437 STN4(&(ro[8]), T6u, T6U, T6g, T72, ovs);
1438 }
1439 }
1440 }
1441 }
1442 }
1443 }
1444 VLEAVE();
1445 }
1446
1447 static const kdft_desc desc = { 32, XSIMD_STRING("n2sv_32"), {340, 52, 32, 0}, &GENUS, 0, 1, 0, 0 };
1448
1449 void XSIMD(codelet_n2sv_32) (planner *p) {
1450 X(kdft_register) (p, n2sv_32, &desc);
1451 }
1452
1453 #endif /* HAVE_FMA */