Mercurial > hg > sv-dependency-builds

comparison src/fftw-3.3.8/dft/scalar/codelets/n1_15.c @ 82:d0c2a83c1364

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam
date Tue, 19 Nov 2019 14:52:55 +0000
parents
children
comparison
equal deleted inserted replaced
81:7029a4916348 82:d0c2a83c1364
1 /*
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 */
20
21 /* This file was automatically generated --- DO NOT EDIT */
22 /* Generated on Thu May 24 08:04:10 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 -compact -variables 4 -pipeline-latency 4 -n 15 -name n1_15 -include dft/scalar/n.h */
29
30 /*
31 * This function contains 156 FP additions, 84 FP multiplications,
32 * (or, 72 additions, 0 multiplications, 84 fused multiply/add),
33 * 69 stack variables, 6 constants, and 60 memory accesses
34 */
35 #include "dft/scalar/n.h"
36
37 static void n1_15(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38 {
39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
41 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
42 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
43 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
44 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
45 {
46 INT i;
47 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(60, is), MAKE_VOLATILE_STRIDE(60, os)) {
48 E T5, T2l, Tx, TV, T1z, T1X, Tl, Tq, Tr, TN, TS, TT, T2c, T2d, T2n;
49 E T1O, T1P, T1Z, T1l, T1q, T1B, TZ, T10, T11, Ta, Tf, Tg, TC, TH, TI;
50 E T2f, T2g, T2m, T1R, T1S, T1Y, T1a, T1f, T1A, TW, TX, TY;
51 {
52 E T1, T1v, T4, T1y, Tw, T1w, Tt, T1x;
53 T1 = ri[0];
54 T1v = ii[0];
55 {
56 E T2, T3, Tu, Tv;
57 T2 = ri[WS(is, 5)];
58 T3 = ri[WS(is, 10)];
59 T4 = T2 + T3;
60 T1y = T3 - T2;
61 Tu = ii[WS(is, 5)];
62 Tv = ii[WS(is, 10)];
63 Tw = Tu - Tv;
64 T1w = Tu + Tv;
65 }
66 T5 = T1 + T4;
67 T2l = T1v + T1w;
68 Tt = FNMS(KP500000000, T4, T1);
69 Tx = FNMS(KP866025403, Tw, Tt);
70 TV = FMA(KP866025403, Tw, Tt);
71 T1x = FNMS(KP500000000, T1w, T1v);
72 T1z = FMA(KP866025403, T1y, T1x);
73 T1X = FNMS(KP866025403, T1y, T1x);
74 }
75 {
76 E Th, Tk, TJ, T1k, T1h, T1i, TM, T1j, Tm, Tp, TO, T1p, T1m, T1n, TR;
77 E T1o;
78 {
79 E Ti, Tj, TK, TL;
80 Th = ri[WS(is, 6)];
81 Ti = ri[WS(is, 11)];
82 Tj = ri[WS(is, 1)];
83 Tk = Ti + Tj;
84 TJ = FNMS(KP500000000, Tk, Th);
85 T1k = Tj - Ti;
86 T1h = ii[WS(is, 6)];
87 TK = ii[WS(is, 11)];
88 TL = ii[WS(is, 1)];
89 T1i = TK + TL;
90 TM = TK - TL;
91 T1j = FNMS(KP500000000, T1i, T1h);
92 }
93 {
94 E Tn, To, TP, TQ;
95 Tm = ri[WS(is, 9)];
96 Tn = ri[WS(is, 14)];
97 To = ri[WS(is, 4)];
98 Tp = Tn + To;
99 TO = FNMS(KP500000000, Tp, Tm);
100 T1p = To - Tn;
101 T1m = ii[WS(is, 9)];
102 TP = ii[WS(is, 14)];
103 TQ = ii[WS(is, 4)];
104 T1n = TP + TQ;
105 TR = TP - TQ;
106 T1o = FNMS(KP500000000, T1n, T1m);
107 }
108 Tl = Th + Tk;
109 Tq = Tm + Tp;
110 Tr = Tl + Tq;
111 TN = FNMS(KP866025403, TM, TJ);
112 TS = FNMS(KP866025403, TR, TO);
113 TT = TN + TS;
114 T2c = T1h + T1i;
115 T2d = T1m + T1n;
116 T2n = T2c + T2d;
117 T1O = FNMS(KP866025403, T1k, T1j);
118 T1P = FNMS(KP866025403, T1p, T1o);
119 T1Z = T1O + T1P;
120 T1l = FMA(KP866025403, T1k, T1j);
121 T1q = FMA(KP866025403, T1p, T1o);
122 T1B = T1l + T1q;
123 TZ = FMA(KP866025403, TM, TJ);
124 T10 = FMA(KP866025403, TR, TO);
125 T11 = TZ + T10;
126 }
127 {
128 E T6, T9, Ty, T19, T16, T17, TB, T18, Tb, Te, TD, T1e, T1b, T1c, TG;
129 E T1d;
130 {
131 E T7, T8, Tz, TA;
132 T6 = ri[WS(is, 3)];
133 T7 = ri[WS(is, 8)];
134 T8 = ri[WS(is, 13)];
135 T9 = T7 + T8;
136 Ty = FNMS(KP500000000, T9, T6);
137 T19 = T8 - T7;
138 T16 = ii[WS(is, 3)];
139 Tz = ii[WS(is, 8)];
140 TA = ii[WS(is, 13)];
141 T17 = Tz + TA;
142 TB = Tz - TA;
143 T18 = FNMS(KP500000000, T17, T16);
144 }
145 {
146 E Tc, Td, TE, TF;
147 Tb = ri[WS(is, 12)];
148 Tc = ri[WS(is, 2)];
149 Td = ri[WS(is, 7)];
150 Te = Tc + Td;
151 TD = FNMS(KP500000000, Te, Tb);
152 T1e = Td - Tc;
153 T1b = ii[WS(is, 12)];
154 TE = ii[WS(is, 2)];
155 TF = ii[WS(is, 7)];
156 T1c = TE + TF;
157 TG = TE - TF;
158 T1d = FNMS(KP500000000, T1c, T1b);
159 }
160 Ta = T6 + T9;
161 Tf = Tb + Te;
162 Tg = Ta + Tf;
163 TC = FNMS(KP866025403, TB, Ty);
164 TH = FNMS(KP866025403, TG, TD);
165 TI = TC + TH;
166 T2f = T16 + T17;
167 T2g = T1b + T1c;
168 T2m = T2f + T2g;
169 T1R = FNMS(KP866025403, T19, T18);
170 T1S = FNMS(KP866025403, T1e, T1d);
171 T1Y = T1R + T1S;
172 T1a = FMA(KP866025403, T19, T18);
173 T1f = FMA(KP866025403, T1e, T1d);
174 T1A = T1a + T1f;
175 TW = FMA(KP866025403, TB, Ty);
176 TX = FMA(KP866025403, TG, TD);
177 TY = TW + TX;
178 }
179 {
180 E T2a, Ts, T29, T2i, T2k, T2e, T2h, T2j, T2b;
181 T2a = Tg - Tr;
182 Ts = Tg + Tr;
183 T29 = FNMS(KP250000000, Ts, T5);
184 T2e = T2c - T2d;
185 T2h = T2f - T2g;
186 T2i = FNMS(KP618033988, T2h, T2e);
187 T2k = FMA(KP618033988, T2e, T2h);
188 ro[0] = T5 + Ts;
189 T2j = FMA(KP559016994, T2a, T29);
190 ro[WS(os, 9)] = FNMS(KP951056516, T2k, T2j);
191 ro[WS(os, 6)] = FMA(KP951056516, T2k, T2j);
192 T2b = FNMS(KP559016994, T2a, T29);
193 ro[WS(os, 12)] = FNMS(KP951056516, T2i, T2b);
194 ro[WS(os, 3)] = FMA(KP951056516, T2i, T2b);
195 }
196 {
197 E T2q, T2o, T2p, T2u, T2w, T2s, T2t, T2v, T2r;
198 T2q = T2m - T2n;
199 T2o = T2m + T2n;
200 T2p = FNMS(KP250000000, T2o, T2l);
201 T2s = Tl - Tq;
202 T2t = Ta - Tf;
203 T2u = FNMS(KP618033988, T2t, T2s);
204 T2w = FMA(KP618033988, T2s, T2t);
205 io[0] = T2l + T2o;
206 T2v = FMA(KP559016994, T2q, T2p);
207 io[WS(os, 6)] = FNMS(KP951056516, T2w, T2v);
208 io[WS(os, 9)] = FMA(KP951056516, T2w, T2v);
209 T2r = FNMS(KP559016994, T2q, T2p);
210 io[WS(os, 3)] = FNMS(KP951056516, T2u, T2r);
211 io[WS(os, 12)] = FMA(KP951056516, T2u, T2r);
212 }
213 {
214 E T1M, TU, T1L, T1U, T1W, T1Q, T1T, T1V, T1N;
215 T1M = TI - TT;
216 TU = TI + TT;
217 T1L = FNMS(KP250000000, TU, Tx);
218 T1Q = T1O - T1P;
219 T1T = T1R - T1S;
220 T1U = FNMS(KP618033988, T1T, T1Q);
221 T1W = FMA(KP618033988, T1Q, T1T);
222 ro[WS(os, 5)] = Tx + TU;
223 T1V = FMA(KP559016994, T1M, T1L);
224 ro[WS(os, 14)] = FNMS(KP951056516, T1W, T1V);
225 ro[WS(os, 11)] = FMA(KP951056516, T1W, T1V);
226 T1N = FNMS(KP559016994, T1M, T1L);
227 ro[WS(os, 2)] = FNMS(KP951056516, T1U, T1N);
228 ro[WS(os, 8)] = FMA(KP951056516, T1U, T1N);
229 }
230 {
231 E T22, T20, T21, T26, T28, T24, T25, T27, T23;
232 T22 = T1Y - T1Z;
233 T20 = T1Y + T1Z;
234 T21 = FNMS(KP250000000, T20, T1X);
235 T24 = TN - TS;
236 T25 = TC - TH;
237 T26 = FNMS(KP618033988, T25, T24);
238 T28 = FMA(KP618033988, T24, T25);
239 io[WS(os, 5)] = T1X + T20;
240 T27 = FMA(KP559016994, T22, T21);
241 io[WS(os, 11)] = FNMS(KP951056516, T28, T27);
242 io[WS(os, 14)] = FMA(KP951056516, T28, T27);
243 T23 = FNMS(KP559016994, T22, T21);
244 io[WS(os, 2)] = FMA(KP951056516, T26, T23);
245 io[WS(os, 8)] = FNMS(KP951056516, T26, T23);
246 }
247 {
248 E T1E, T1C, T1D, T1I, T1K, T1G, T1H, T1J, T1F;
249 T1E = T1A - T1B;
250 T1C = T1A + T1B;
251 T1D = FNMS(KP250000000, T1C, T1z);
252 T1G = TW - TX;
253 T1H = TZ - T10;
254 T1I = FMA(KP618033988, T1H, T1G);
255 T1K = FNMS(KP618033988, T1G, T1H);
256 io[WS(os, 10)] = T1z + T1C;
257 T1J = FNMS(KP559016994, T1E, T1D);
258 io[WS(os, 7)] = FMA(KP951056516, T1K, T1J);
259 io[WS(os, 13)] = FNMS(KP951056516, T1K, T1J);
260 T1F = FMA(KP559016994, T1E, T1D);
261 io[WS(os, 1)] = FNMS(KP951056516, T1I, T1F);
262 io[WS(os, 4)] = FMA(KP951056516, T1I, T1F);
263 }
264 {
265 E T14, T12, T13, T1s, T1u, T1g, T1r, T1t, T15;
266 T14 = TY - T11;
267 T12 = TY + T11;
268 T13 = FNMS(KP250000000, T12, TV);
269 T1g = T1a - T1f;
270 T1r = T1l - T1q;
271 T1s = FMA(KP618033988, T1r, T1g);
272 T1u = FNMS(KP618033988, T1g, T1r);
273 ro[WS(os, 10)] = TV + T12;
274 T1t = FNMS(KP559016994, T14, T13);
275 ro[WS(os, 7)] = FNMS(KP951056516, T1u, T1t);
276 ro[WS(os, 13)] = FMA(KP951056516, T1u, T1t);
277 T15 = FMA(KP559016994, T14, T13);
278 ro[WS(os, 4)] = FNMS(KP951056516, T1s, T15);
279 ro[WS(os, 1)] = FMA(KP951056516, T1s, T15);
280 }
281 }
282 }
283 }
284
285 static const kdft_desc desc = { 15, "n1_15", {72, 0, 84, 0}, &GENUS, 0, 0, 0, 0 };
286
287 void X(codelet_n1_15) (planner *p) {
288 X(kdft_register) (p, n1_15, &desc);
289 }
290
291 #else
292
293 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 15 -name n1_15 -include dft/scalar/n.h */
294
295 /*
296 * This function contains 156 FP additions, 56 FP multiplications,
297 * (or, 128 additions, 28 multiplications, 28 fused multiply/add),
298 * 69 stack variables, 6 constants, and 60 memory accesses
299 */
300 #include "dft/scalar/n.h"
301
302 static void n1_15(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
303 {
304 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
305 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
306 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
307 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
308 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
309 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
310 {
311 INT i;
312 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(60, is), MAKE_VOLATILE_STRIDE(60, os)) {
313 E T5, T2l, Tx, TV, T1C, T20, Tl, Tq, Tr, TN, TS, TT, T2c, T2d, T2n;
314 E T1O, T1P, T22, T1l, T1q, T1w, TZ, T10, T11, Ta, Tf, Tg, TC, TH, TI;
315 E T2f, T2g, T2m, T1R, T1S, T21, T1a, T1f, T1v, TW, TX, TY;
316 {
317 E T1, T1z, T4, T1y, Tw, T1A, Tt, T1B;
318 T1 = ri[0];
319 T1z = ii[0];
320 {
321 E T2, T3, Tu, Tv;
322 T2 = ri[WS(is, 5)];
323 T3 = ri[WS(is, 10)];
324 T4 = T2 + T3;
325 T1y = KP866025403 * (T3 - T2);
326 Tu = ii[WS(is, 5)];
327 Tv = ii[WS(is, 10)];
328 Tw = KP866025403 * (Tu - Tv);
329 T1A = Tu + Tv;
330 }
331 T5 = T1 + T4;
332 T2l = T1z + T1A;
333 Tt = FNMS(KP500000000, T4, T1);
334 Tx = Tt - Tw;
335 TV = Tt + Tw;
336 T1B = FNMS(KP500000000, T1A, T1z);
337 T1C = T1y + T1B;
338 T20 = T1B - T1y;
339 }
340 {
341 E Th, Tk, TJ, T1h, T1i, T1j, TM, T1k, Tm, Tp, TO, T1m, T1n, T1o, TR;
342 E T1p;
343 {
344 E Ti, Tj, TK, TL;
345 Th = ri[WS(is, 6)];
346 Ti = ri[WS(is, 11)];
347 Tj = ri[WS(is, 1)];
348 Tk = Ti + Tj;
349 TJ = FNMS(KP500000000, Tk, Th);
350 T1h = KP866025403 * (Tj - Ti);
351 T1i = ii[WS(is, 6)];
352 TK = ii[WS(is, 11)];
353 TL = ii[WS(is, 1)];
354 T1j = TK + TL;
355 TM = KP866025403 * (TK - TL);
356 T1k = FNMS(KP500000000, T1j, T1i);
357 }
358 {
359 E Tn, To, TP, TQ;
360 Tm = ri[WS(is, 9)];
361 Tn = ri[WS(is, 14)];
362 To = ri[WS(is, 4)];
363 Tp = Tn + To;
364 TO = FNMS(KP500000000, Tp, Tm);
365 T1m = KP866025403 * (To - Tn);
366 T1n = ii[WS(is, 9)];
367 TP = ii[WS(is, 14)];
368 TQ = ii[WS(is, 4)];
369 T1o = TP + TQ;
370 TR = KP866025403 * (TP - TQ);
371 T1p = FNMS(KP500000000, T1o, T1n);
372 }
373 Tl = Th + Tk;
374 Tq = Tm + Tp;
375 Tr = Tl + Tq;
376 TN = TJ - TM;
377 TS = TO - TR;
378 TT = TN + TS;
379 T2c = T1i + T1j;
380 T2d = T1n + T1o;
381 T2n = T2c + T2d;
382 T1O = T1k - T1h;
383 T1P = T1p - T1m;
384 T22 = T1O + T1P;
385 T1l = T1h + T1k;
386 T1q = T1m + T1p;
387 T1w = T1l + T1q;
388 TZ = TJ + TM;
389 T10 = TO + TR;
390 T11 = TZ + T10;
391 }
392 {
393 E T6, T9, Ty, T16, T17, T18, TB, T19, Tb, Te, TD, T1b, T1c, T1d, TG;
394 E T1e;
395 {
396 E T7, T8, Tz, TA;
397 T6 = ri[WS(is, 3)];
398 T7 = ri[WS(is, 8)];
399 T8 = ri[WS(is, 13)];
400 T9 = T7 + T8;
401 Ty = FNMS(KP500000000, T9, T6);
402 T16 = KP866025403 * (T8 - T7);
403 T17 = ii[WS(is, 3)];
404 Tz = ii[WS(is, 8)];
405 TA = ii[WS(is, 13)];
406 T18 = Tz + TA;
407 TB = KP866025403 * (Tz - TA);
408 T19 = FNMS(KP500000000, T18, T17);
409 }
410 {
411 E Tc, Td, TE, TF;
412 Tb = ri[WS(is, 12)];
413 Tc = ri[WS(is, 2)];
414 Td = ri[WS(is, 7)];
415 Te = Tc + Td;
416 TD = FNMS(KP500000000, Te, Tb);
417 T1b = KP866025403 * (Td - Tc);
418 T1c = ii[WS(is, 12)];
419 TE = ii[WS(is, 2)];
420 TF = ii[WS(is, 7)];
421 T1d = TE + TF;
422 TG = KP866025403 * (TE - TF);
423 T1e = FNMS(KP500000000, T1d, T1c);
424 }
425 Ta = T6 + T9;
426 Tf = Tb + Te;
427 Tg = Ta + Tf;
428 TC = Ty - TB;
429 TH = TD - TG;
430 TI = TC + TH;
431 T2f = T17 + T18;
432 T2g = T1c + T1d;
433 T2m = T2f + T2g;
434 T1R = T19 - T16;
435 T1S = T1e - T1b;
436 T21 = T1R + T1S;
437 T1a = T16 + T19;
438 T1f = T1b + T1e;
439 T1v = T1a + T1f;
440 TW = Ty + TB;
441 TX = TD + TG;
442 TY = TW + TX;
443 }
444 {
445 E T2a, Ts, T29, T2i, T2k, T2e, T2h, T2j, T2b;
446 T2a = KP559016994 * (Tg - Tr);
447 Ts = Tg + Tr;
448 T29 = FNMS(KP250000000, Ts, T5);
449 T2e = T2c - T2d;
450 T2h = T2f - T2g;
451 T2i = FNMS(KP587785252, T2h, KP951056516 * T2e);
452 T2k = FMA(KP951056516, T2h, KP587785252 * T2e);
453 ro[0] = T5 + Ts;
454 T2j = T2a + T29;
455 ro[WS(os, 9)] = T2j - T2k;
456 ro[WS(os, 6)] = T2j + T2k;
457 T2b = T29 - T2a;
458 ro[WS(os, 12)] = T2b - T2i;
459 ro[WS(os, 3)] = T2b + T2i;
460 }
461 {
462 E T2q, T2o, T2p, T2u, T2w, T2s, T2t, T2v, T2r;
463 T2q = KP559016994 * (T2m - T2n);
464 T2o = T2m + T2n;
465 T2p = FNMS(KP250000000, T2o, T2l);
466 T2s = Tl - Tq;
467 T2t = Ta - Tf;
468 T2u = FNMS(KP587785252, T2t, KP951056516 * T2s);
469 T2w = FMA(KP951056516, T2t, KP587785252 * T2s);
470 io[0] = T2l + T2o;
471 T2v = T2q + T2p;
472 io[WS(os, 6)] = T2v - T2w;
473 io[WS(os, 9)] = T2w + T2v;
474 T2r = T2p - T2q;
475 io[WS(os, 3)] = T2r - T2u;
476 io[WS(os, 12)] = T2u + T2r;
477 }
478 {
479 E T1M, TU, T1L, T1U, T1W, T1Q, T1T, T1V, T1N;
480 T1M = KP559016994 * (TI - TT);
481 TU = TI + TT;
482 T1L = FNMS(KP250000000, TU, Tx);
483 T1Q = T1O - T1P;
484 T1T = T1R - T1S;
485 T1U = FNMS(KP587785252, T1T, KP951056516 * T1Q);
486 T1W = FMA(KP951056516, T1T, KP587785252 * T1Q);
487 ro[WS(os, 5)] = Tx + TU;
488 T1V = T1M + T1L;
489 ro[WS(os, 14)] = T1V - T1W;
490 ro[WS(os, 11)] = T1V + T1W;
491 T1N = T1L - T1M;
492 ro[WS(os, 2)] = T1N - T1U;
493 ro[WS(os, 8)] = T1N + T1U;
494 }
495 {
496 E T25, T23, T24, T1Z, T28, T1X, T1Y, T27, T26;
497 T25 = KP559016994 * (T21 - T22);
498 T23 = T21 + T22;
499 T24 = FNMS(KP250000000, T23, T20);
500 T1X = TN - TS;
501 T1Y = TC - TH;
502 T1Z = FNMS(KP587785252, T1Y, KP951056516 * T1X);
503 T28 = FMA(KP951056516, T1Y, KP587785252 * T1X);
504 io[WS(os, 5)] = T20 + T23;
505 T27 = T25 + T24;
506 io[WS(os, 11)] = T27 - T28;
507 io[WS(os, 14)] = T28 + T27;
508 T26 = T24 - T25;
509 io[WS(os, 2)] = T1Z + T26;
510 io[WS(os, 8)] = T26 - T1Z;
511 }
512 {
513 E T1x, T1D, T1E, T1I, T1J, T1G, T1H, T1K, T1F;
514 T1x = KP559016994 * (T1v - T1w);
515 T1D = T1v + T1w;
516 T1E = FNMS(KP250000000, T1D, T1C);
517 T1G = TW - TX;
518 T1H = TZ - T10;
519 T1I = FMA(KP951056516, T1G, KP587785252 * T1H);
520 T1J = FNMS(KP587785252, T1G, KP951056516 * T1H);
521 io[WS(os, 10)] = T1C + T1D;
522 T1K = T1E - T1x;
523 io[WS(os, 7)] = T1J + T1K;
524 io[WS(os, 13)] = T1K - T1J;
525 T1F = T1x + T1E;
526 io[WS(os, 1)] = T1F - T1I;
527 io[WS(os, 4)] = T1I + T1F;
528 }
529 {
530 E T13, T12, T14, T1s, T1u, T1g, T1r, T1t, T15;
531 T13 = KP559016994 * (TY - T11);
532 T12 = TY + T11;
533 T14 = FNMS(KP250000000, T12, TV);
534 T1g = T1a - T1f;
535 T1r = T1l - T1q;
536 T1s = FMA(KP951056516, T1g, KP587785252 * T1r);
537 T1u = FNMS(KP587785252, T1g, KP951056516 * T1r);
538 ro[WS(os, 10)] = TV + T12;
539 T1t = T14 - T13;
540 ro[WS(os, 7)] = T1t - T1u;
541 ro[WS(os, 13)] = T1t + T1u;
542 T15 = T13 + T14;
543 ro[WS(os, 4)] = T15 - T1s;
544 ro[WS(os, 1)] = T15 + T1s;
545 }
546 }
547 }
548 }
549
550 static const kdft_desc desc = { 15, "n1_15", {128, 28, 28, 0}, &GENUS, 0, 0, 0, 0 };
551
552 void X(codelet_n1_15) (planner *p) {
553 X(kdft_register) (p, n1_15, &desc);
554 }
555
556 #endif