comparison src/fftw-3.3.8/dft/scalar/codelets/t1_20.c @ 82:d0c2a83c1364

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 08:04:15 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_twiddle.native -fma -compact -variables 4 -pipeline-latency 4 -n 20 -name t1_20 -include dft/scalar/t.h */
29
30 /*
31 * This function contains 246 FP additions, 148 FP multiplications,
32 * (or, 136 additions, 38 multiplications, 110 fused multiply/add),
33 * 61 stack variables, 4 constants, and 80 memory accesses
34 */
35 #include "dft/scalar/t.h"
36
37 static void t1_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
38 {
39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
42 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
43 {
44 INT m;
45 for (m = mb, W = W + (mb * 38); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 38, MAKE_VOLATILE_STRIDE(40, rs)) {
46 E T8, T4N, T2i, T4r, Tl, T4O, T2n, T4n, TN, T2b, T40, T4b, T2v, T3v, T3i;
47 E T3F, T27, T2f, T3W, T4f, T2R, T3z, T3a, T3J, T1G, T2e, T3T, T4e, T2K, T3y;
48 E T33, T3I, T1e, T2c, T43, T4c, T2C, T3w, T3p, T3G;
49 {
50 E T1, T4q, T3, T6, T4, T4o, T2, T7, T4p, T5;
51 T1 = ri[0];
52 T4q = ii[0];
53 T3 = ri[WS(rs, 10)];
54 T6 = ii[WS(rs, 10)];
55 T2 = W[18];
56 T4 = T2 * T3;
57 T4o = T2 * T6;
58 T5 = W[19];
59 T7 = FMA(T5, T6, T4);
60 T4p = FNMS(T5, T3, T4o);
61 T8 = T1 + T7;
62 T4N = T4q - T4p;
63 T2i = T1 - T7;
64 T4r = T4p + T4q;
65 }
66 {
67 E Ta, Td, Tb, T2j, Tg, Tj, Th, T2l, T9, Tf;
68 Ta = ri[WS(rs, 5)];
69 Td = ii[WS(rs, 5)];
70 T9 = W[8];
71 Tb = T9 * Ta;
72 T2j = T9 * Td;
73 Tg = ri[WS(rs, 15)];
74 Tj = ii[WS(rs, 15)];
75 Tf = W[28];
76 Th = Tf * Tg;
77 T2l = Tf * Tj;
78 {
79 E Te, T2k, Tk, T2m, Tc, Ti;
80 Tc = W[9];
81 Te = FMA(Tc, Td, Tb);
82 T2k = FNMS(Tc, Ta, T2j);
83 Ti = W[29];
84 Tk = FMA(Ti, Tj, Th);
85 T2m = FNMS(Ti, Tg, T2l);
86 Tl = Te + Tk;
87 T4O = Te - Tk;
88 T2n = T2k - T2m;
89 T4n = T2k + T2m;
90 }
91 }
92 {
93 E Ts, T3d, TL, T2t, Ty, T3f, TF, T2r;
94 {
95 E To, Tr, Tp, T3c, Tn, Tq;
96 To = ri[WS(rs, 4)];
97 Tr = ii[WS(rs, 4)];
98 Tn = W[6];
99 Tp = Tn * To;
100 T3c = Tn * Tr;
101 Tq = W[7];
102 Ts = FMA(Tq, Tr, Tp);
103 T3d = FNMS(Tq, To, T3c);
104 }
105 {
106 E TH, TK, TI, T2s, TG, TJ;
107 TH = ri[WS(rs, 19)];
108 TK = ii[WS(rs, 19)];
109 TG = W[36];
110 TI = TG * TH;
111 T2s = TG * TK;
112 TJ = W[37];
113 TL = FMA(TJ, TK, TI);
114 T2t = FNMS(TJ, TH, T2s);
115 }
116 {
117 E Tu, Tx, Tv, T3e, Tt, Tw;
118 Tu = ri[WS(rs, 14)];
119 Tx = ii[WS(rs, 14)];
120 Tt = W[26];
121 Tv = Tt * Tu;
122 T3e = Tt * Tx;
123 Tw = W[27];
124 Ty = FMA(Tw, Tx, Tv);
125 T3f = FNMS(Tw, Tu, T3e);
126 }
127 {
128 E TB, TE, TC, T2q, TA, TD;
129 TB = ri[WS(rs, 9)];
130 TE = ii[WS(rs, 9)];
131 TA = W[16];
132 TC = TA * TB;
133 T2q = TA * TE;
134 TD = W[17];
135 TF = FMA(TD, TE, TC);
136 T2r = FNMS(TD, TB, T2q);
137 }
138 {
139 E Tz, TM, T3Y, T3Z;
140 Tz = Ts + Ty;
141 TM = TF + TL;
142 TN = Tz - TM;
143 T2b = Tz + TM;
144 T3Y = T3d + T3f;
145 T3Z = T2r + T2t;
146 T40 = T3Y - T3Z;
147 T4b = T3Y + T3Z;
148 }
149 {
150 E T2p, T2u, T3g, T3h;
151 T2p = Ts - Ty;
152 T2u = T2r - T2t;
153 T2v = T2p - T2u;
154 T3v = T2p + T2u;
155 T3g = T3d - T3f;
156 T3h = TF - TL;
157 T3i = T3g + T3h;
158 T3F = T3g - T3h;
159 }
160 }
161 {
162 E T1M, T35, T25, T2P, T1S, T37, T1Z, T2N;
163 {
164 E T1I, T1L, T1J, T34, T1H, T1K;
165 T1I = ri[WS(rs, 12)];
166 T1L = ii[WS(rs, 12)];
167 T1H = W[22];
168 T1J = T1H * T1I;
169 T34 = T1H * T1L;
170 T1K = W[23];
171 T1M = FMA(T1K, T1L, T1J);
172 T35 = FNMS(T1K, T1I, T34);
173 }
174 {
175 E T21, T24, T22, T2O, T20, T23;
176 T21 = ri[WS(rs, 7)];
177 T24 = ii[WS(rs, 7)];
178 T20 = W[12];
179 T22 = T20 * T21;
180 T2O = T20 * T24;
181 T23 = W[13];
182 T25 = FMA(T23, T24, T22);
183 T2P = FNMS(T23, T21, T2O);
184 }
185 {
186 E T1O, T1R, T1P, T36, T1N, T1Q;
187 T1O = ri[WS(rs, 2)];
188 T1R = ii[WS(rs, 2)];
189 T1N = W[2];
190 T1P = T1N * T1O;
191 T36 = T1N * T1R;
192 T1Q = W[3];
193 T1S = FMA(T1Q, T1R, T1P);
194 T37 = FNMS(T1Q, T1O, T36);
195 }
196 {
197 E T1V, T1Y, T1W, T2M, T1U, T1X;
198 T1V = ri[WS(rs, 17)];
199 T1Y = ii[WS(rs, 17)];
200 T1U = W[32];
201 T1W = T1U * T1V;
202 T2M = T1U * T1Y;
203 T1X = W[33];
204 T1Z = FMA(T1X, T1Y, T1W);
205 T2N = FNMS(T1X, T1V, T2M);
206 }
207 {
208 E T1T, T26, T3U, T3V;
209 T1T = T1M + T1S;
210 T26 = T1Z + T25;
211 T27 = T1T - T26;
212 T2f = T1T + T26;
213 T3U = T35 + T37;
214 T3V = T2N + T2P;
215 T3W = T3U - T3V;
216 T4f = T3U + T3V;
217 }
218 {
219 E T2L, T2Q, T38, T39;
220 T2L = T1M - T1S;
221 T2Q = T2N - T2P;
222 T2R = T2L - T2Q;
223 T3z = T2L + T2Q;
224 T38 = T35 - T37;
225 T39 = T1Z - T25;
226 T3a = T38 + T39;
227 T3J = T38 - T39;
228 }
229 }
230 {
231 E T1l, T2Y, T1E, T2I, T1r, T30, T1y, T2G;
232 {
233 E T1h, T1k, T1i, T2X, T1g, T1j;
234 T1h = ri[WS(rs, 8)];
235 T1k = ii[WS(rs, 8)];
236 T1g = W[14];
237 T1i = T1g * T1h;
238 T2X = T1g * T1k;
239 T1j = W[15];
240 T1l = FMA(T1j, T1k, T1i);
241 T2Y = FNMS(T1j, T1h, T2X);
242 }
243 {
244 E T1A, T1D, T1B, T2H, T1z, T1C;
245 T1A = ri[WS(rs, 3)];
246 T1D = ii[WS(rs, 3)];
247 T1z = W[4];
248 T1B = T1z * T1A;
249 T2H = T1z * T1D;
250 T1C = W[5];
251 T1E = FMA(T1C, T1D, T1B);
252 T2I = FNMS(T1C, T1A, T2H);
253 }
254 {
255 E T1n, T1q, T1o, T2Z, T1m, T1p;
256 T1n = ri[WS(rs, 18)];
257 T1q = ii[WS(rs, 18)];
258 T1m = W[34];
259 T1o = T1m * T1n;
260 T2Z = T1m * T1q;
261 T1p = W[35];
262 T1r = FMA(T1p, T1q, T1o);
263 T30 = FNMS(T1p, T1n, T2Z);
264 }
265 {
266 E T1u, T1x, T1v, T2F, T1t, T1w;
267 T1u = ri[WS(rs, 13)];
268 T1x = ii[WS(rs, 13)];
269 T1t = W[24];
270 T1v = T1t * T1u;
271 T2F = T1t * T1x;
272 T1w = W[25];
273 T1y = FMA(T1w, T1x, T1v);
274 T2G = FNMS(T1w, T1u, T2F);
275 }
276 {
277 E T1s, T1F, T3R, T3S;
278 T1s = T1l + T1r;
279 T1F = T1y + T1E;
280 T1G = T1s - T1F;
281 T2e = T1s + T1F;
282 T3R = T2Y + T30;
283 T3S = T2G + T2I;
284 T3T = T3R - T3S;
285 T4e = T3R + T3S;
286 }
287 {
288 E T2E, T2J, T31, T32;
289 T2E = T1l - T1r;
290 T2J = T2G - T2I;
291 T2K = T2E - T2J;
292 T3y = T2E + T2J;
293 T31 = T2Y - T30;
294 T32 = T1y - T1E;
295 T33 = T31 + T32;
296 T3I = T31 - T32;
297 }
298 }
299 {
300 E TT, T3k, T1c, T2A, TZ, T3m, T16, T2y;
301 {
302 E TP, TS, TQ, T3j, TO, TR;
303 TP = ri[WS(rs, 16)];
304 TS = ii[WS(rs, 16)];
305 TO = W[30];
306 TQ = TO * TP;
307 T3j = TO * TS;
308 TR = W[31];
309 TT = FMA(TR, TS, TQ);
310 T3k = FNMS(TR, TP, T3j);
311 }
312 {
313 E T18, T1b, T19, T2z, T17, T1a;
314 T18 = ri[WS(rs, 11)];
315 T1b = ii[WS(rs, 11)];
316 T17 = W[20];
317 T19 = T17 * T18;
318 T2z = T17 * T1b;
319 T1a = W[21];
320 T1c = FMA(T1a, T1b, T19);
321 T2A = FNMS(T1a, T18, T2z);
322 }
323 {
324 E TV, TY, TW, T3l, TU, TX;
325 TV = ri[WS(rs, 6)];
326 TY = ii[WS(rs, 6)];
327 TU = W[10];
328 TW = TU * TV;
329 T3l = TU * TY;
330 TX = W[11];
331 TZ = FMA(TX, TY, TW);
332 T3m = FNMS(TX, TV, T3l);
333 }
334 {
335 E T12, T15, T13, T2x, T11, T14;
336 T12 = ri[WS(rs, 1)];
337 T15 = ii[WS(rs, 1)];
338 T11 = W[0];
339 T13 = T11 * T12;
340 T2x = T11 * T15;
341 T14 = W[1];
342 T16 = FMA(T14, T15, T13);
343 T2y = FNMS(T14, T12, T2x);
344 }
345 {
346 E T10, T1d, T41, T42;
347 T10 = TT + TZ;
348 T1d = T16 + T1c;
349 T1e = T10 - T1d;
350 T2c = T10 + T1d;
351 T41 = T3k + T3m;
352 T42 = T2y + T2A;
353 T43 = T41 - T42;
354 T4c = T41 + T42;
355 }
356 {
357 E T2w, T2B, T3n, T3o;
358 T2w = TT - TZ;
359 T2B = T2y - T2A;
360 T2C = T2w - T2B;
361 T3w = T2w + T2B;
362 T3n = T3k - T3m;
363 T3o = T16 - T1c;
364 T3p = T3n + T3o;
365 T3G = T3n - T3o;
366 }
367 }
368 {
369 E T45, T47, Tm, T29, T3O, T3P, T46, T3Q;
370 {
371 E T3X, T44, T1f, T28;
372 T3X = T3T - T3W;
373 T44 = T40 - T43;
374 T45 = FNMS(KP618033988, T44, T3X);
375 T47 = FMA(KP618033988, T3X, T44);
376 Tm = T8 - Tl;
377 T1f = TN + T1e;
378 T28 = T1G + T27;
379 T29 = T1f + T28;
380 T3O = FNMS(KP250000000, T29, Tm);
381 T3P = T1f - T28;
382 }
383 ri[WS(rs, 10)] = Tm + T29;
384 T46 = FMA(KP559016994, T3P, T3O);
385 ri[WS(rs, 14)] = FNMS(KP951056516, T47, T46);
386 ri[WS(rs, 6)] = FMA(KP951056516, T47, T46);
387 T3Q = FNMS(KP559016994, T3P, T3O);
388 ri[WS(rs, 2)] = FNMS(KP951056516, T45, T3Q);
389 ri[WS(rs, 18)] = FMA(KP951056516, T45, T3Q);
390 }
391 {
392 E T4K, T4M, T4B, T4E, T4F, T4G, T4L, T4H;
393 {
394 E T4I, T4J, T4C, T4D;
395 T4I = T1G - T27;
396 T4J = TN - T1e;
397 T4K = FNMS(KP618033988, T4J, T4I);
398 T4M = FMA(KP618033988, T4I, T4J);
399 T4B = T4r - T4n;
400 T4C = T40 + T43;
401 T4D = T3T + T3W;
402 T4E = T4C + T4D;
403 T4F = FNMS(KP250000000, T4E, T4B);
404 T4G = T4C - T4D;
405 }
406 ii[WS(rs, 10)] = T4E + T4B;
407 T4L = FMA(KP559016994, T4G, T4F);
408 ii[WS(rs, 6)] = FNMS(KP951056516, T4M, T4L);
409 ii[WS(rs, 14)] = FMA(KP951056516, T4M, T4L);
410 T4H = FNMS(KP559016994, T4G, T4F);
411 ii[WS(rs, 2)] = FMA(KP951056516, T4K, T4H);
412 ii[WS(rs, 18)] = FNMS(KP951056516, T4K, T4H);
413 }
414 {
415 E T4h, T4j, T2a, T2h, T48, T49, T4i, T4a;
416 {
417 E T4d, T4g, T2d, T2g;
418 T4d = T4b - T4c;
419 T4g = T4e - T4f;
420 T4h = FMA(KP618033988, T4g, T4d);
421 T4j = FNMS(KP618033988, T4d, T4g);
422 T2a = T8 + Tl;
423 T2d = T2b + T2c;
424 T2g = T2e + T2f;
425 T2h = T2d + T2g;
426 T48 = FNMS(KP250000000, T2h, T2a);
427 T49 = T2d - T2g;
428 }
429 ri[0] = T2a + T2h;
430 T4i = FNMS(KP559016994, T49, T48);
431 ri[WS(rs, 12)] = FNMS(KP951056516, T4j, T4i);
432 ri[WS(rs, 8)] = FMA(KP951056516, T4j, T4i);
433 T4a = FMA(KP559016994, T49, T48);
434 ri[WS(rs, 4)] = FNMS(KP951056516, T4h, T4a);
435 ri[WS(rs, 16)] = FMA(KP951056516, T4h, T4a);
436 }
437 {
438 E T4y, T4A, T4s, T4m, T4t, T4u, T4z, T4v;
439 {
440 E T4w, T4x, T4k, T4l;
441 T4w = T2b - T2c;
442 T4x = T2e - T2f;
443 T4y = FMA(KP618033988, T4x, T4w);
444 T4A = FNMS(KP618033988, T4w, T4x);
445 T4s = T4n + T4r;
446 T4k = T4b + T4c;
447 T4l = T4e + T4f;
448 T4m = T4k + T4l;
449 T4t = FNMS(KP250000000, T4m, T4s);
450 T4u = T4k - T4l;
451 }
452 ii[0] = T4m + T4s;
453 T4z = FNMS(KP559016994, T4u, T4t);
454 ii[WS(rs, 8)] = FNMS(KP951056516, T4A, T4z);
455 ii[WS(rs, 12)] = FMA(KP951056516, T4A, T4z);
456 T4v = FMA(KP559016994, T4u, T4t);
457 ii[WS(rs, 4)] = FMA(KP951056516, T4y, T4v);
458 ii[WS(rs, 16)] = FNMS(KP951056516, T4y, T4v);
459 }
460 {
461 E T3r, T3t, T2o, T2T, T2U, T2V, T3s, T2W;
462 {
463 E T3b, T3q, T2D, T2S;
464 T3b = T33 - T3a;
465 T3q = T3i - T3p;
466 T3r = FNMS(KP618033988, T3q, T3b);
467 T3t = FMA(KP618033988, T3b, T3q);
468 T2o = T2i - T2n;
469 T2D = T2v + T2C;
470 T2S = T2K + T2R;
471 T2T = T2D + T2S;
472 T2U = FNMS(KP250000000, T2T, T2o);
473 T2V = T2D - T2S;
474 }
475 ri[WS(rs, 15)] = T2o + T2T;
476 T3s = FMA(KP559016994, T2V, T2U);
477 ri[WS(rs, 11)] = FMA(KP951056516, T3t, T3s);
478 ri[WS(rs, 19)] = FNMS(KP951056516, T3t, T3s);
479 T2W = FNMS(KP559016994, T2V, T2U);
480 ri[WS(rs, 3)] = FMA(KP951056516, T3r, T2W);
481 ri[WS(rs, 7)] = FNMS(KP951056516, T3r, T2W);
482 }
483 {
484 E T5a, T5c, T51, T54, T55, T56, T5b, T57;
485 {
486 E T58, T59, T52, T53;
487 T58 = T2K - T2R;
488 T59 = T2v - T2C;
489 T5a = FNMS(KP618033988, T59, T58);
490 T5c = FMA(KP618033988, T58, T59);
491 T51 = T4O + T4N;
492 T52 = T3i + T3p;
493 T53 = T33 + T3a;
494 T54 = T52 + T53;
495 T55 = FNMS(KP250000000, T54, T51);
496 T56 = T52 - T53;
497 }
498 ii[WS(rs, 15)] = T54 + T51;
499 T5b = FMA(KP559016994, T56, T55);
500 ii[WS(rs, 11)] = FNMS(KP951056516, T5c, T5b);
501 ii[WS(rs, 19)] = FMA(KP951056516, T5c, T5b);
502 T57 = FNMS(KP559016994, T56, T55);
503 ii[WS(rs, 3)] = FNMS(KP951056516, T5a, T57);
504 ii[WS(rs, 7)] = FMA(KP951056516, T5a, T57);
505 }
506 {
507 E T3L, T3N, T3u, T3B, T3C, T3D, T3M, T3E;
508 {
509 E T3H, T3K, T3x, T3A;
510 T3H = T3F - T3G;
511 T3K = T3I - T3J;
512 T3L = FMA(KP618033988, T3K, T3H);
513 T3N = FNMS(KP618033988, T3H, T3K);
514 T3u = T2i + T2n;
515 T3x = T3v + T3w;
516 T3A = T3y + T3z;
517 T3B = T3x + T3A;
518 T3C = FNMS(KP250000000, T3B, T3u);
519 T3D = T3x - T3A;
520 }
521 ri[WS(rs, 5)] = T3u + T3B;
522 T3M = FNMS(KP559016994, T3D, T3C);
523 ri[WS(rs, 13)] = FMA(KP951056516, T3N, T3M);
524 ri[WS(rs, 17)] = FNMS(KP951056516, T3N, T3M);
525 T3E = FMA(KP559016994, T3D, T3C);
526 ri[WS(rs, 1)] = FMA(KP951056516, T3L, T3E);
527 ri[WS(rs, 9)] = FNMS(KP951056516, T3L, T3E);
528 }
529 {
530 E T4Y, T50, T4P, T4S, T4T, T4U, T4Z, T4V;
531 {
532 E T4W, T4X, T4Q, T4R;
533 T4W = T3v - T3w;
534 T4X = T3y - T3z;
535 T4Y = FMA(KP618033988, T4X, T4W);
536 T50 = FNMS(KP618033988, T4W, T4X);
537 T4P = T4N - T4O;
538 T4Q = T3F + T3G;
539 T4R = T3I + T3J;
540 T4S = T4Q + T4R;
541 T4T = FNMS(KP250000000, T4S, T4P);
542 T4U = T4Q - T4R;
543 }
544 ii[WS(rs, 5)] = T4S + T4P;
545 T4Z = FNMS(KP559016994, T4U, T4T);
546 ii[WS(rs, 13)] = FNMS(KP951056516, T50, T4Z);
547 ii[WS(rs, 17)] = FMA(KP951056516, T50, T4Z);
548 T4V = FMA(KP559016994, T4U, T4T);
549 ii[WS(rs, 1)] = FNMS(KP951056516, T4Y, T4V);
550 ii[WS(rs, 9)] = FMA(KP951056516, T4Y, T4V);
551 }
552 }
553 }
554 }
555
556 static const tw_instr twinstr[] = {
557 {TW_FULL, 0, 20},
558 {TW_NEXT, 1, 0}
559 };
560
561 static const ct_desc desc = { 20, "t1_20", twinstr, &GENUS, {136, 38, 110, 0}, 0, 0, 0 };
562
563 void X(codelet_t1_20) (planner *p) {
564 X(kdft_dit_register) (p, t1_20, &desc);
565 }
566 #else
567
568 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 20 -name t1_20 -include dft/scalar/t.h */
569
570 /*
571 * This function contains 246 FP additions, 124 FP multiplications,
572 * (or, 184 additions, 62 multiplications, 62 fused multiply/add),
573 * 85 stack variables, 4 constants, and 80 memory accesses
574 */
575 #include "dft/scalar/t.h"
576
577 static void t1_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
578 {
579 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
580 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
581 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
582 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
583 {
584 INT m;
585 for (m = mb, W = W + (mb * 38); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 38, MAKE_VOLATILE_STRIDE(40, rs)) {
586 E Tj, T1R, T4g, T4p, T2q, T37, T3Q, T42, T1r, T1O, T1P, T3i, T3l, T44, T3D;
587 E T3E, T3K, T1V, T1W, T1X, T23, T28, T4r, T2W, T2X, T4c, T33, T34, T35, T2G;
588 E T2L, T2M, TG, T13, T14, T3p, T3s, T43, T3A, T3B, T3J, T1S, T1T, T1U, T2e;
589 E T2j, T4q, T2T, T2U, T4b, T30, T31, T32, T2v, T2A, T2B;
590 {
591 E T1, T3O, T6, T3N, Tc, T2n, Th, T2o;
592 T1 = ri[0];
593 T3O = ii[0];
594 {
595 E T3, T5, T2, T4;
596 T3 = ri[WS(rs, 10)];
597 T5 = ii[WS(rs, 10)];
598 T2 = W[18];
599 T4 = W[19];
600 T6 = FMA(T2, T3, T4 * T5);
601 T3N = FNMS(T4, T3, T2 * T5);
602 }
603 {
604 E T9, Tb, T8, Ta;
605 T9 = ri[WS(rs, 5)];
606 Tb = ii[WS(rs, 5)];
607 T8 = W[8];
608 Ta = W[9];
609 Tc = FMA(T8, T9, Ta * Tb);
610 T2n = FNMS(Ta, T9, T8 * Tb);
611 }
612 {
613 E Te, Tg, Td, Tf;
614 Te = ri[WS(rs, 15)];
615 Tg = ii[WS(rs, 15)];
616 Td = W[28];
617 Tf = W[29];
618 Th = FMA(Td, Te, Tf * Tg);
619 T2o = FNMS(Tf, Te, Td * Tg);
620 }
621 {
622 E T7, Ti, T4e, T4f;
623 T7 = T1 + T6;
624 Ti = Tc + Th;
625 Tj = T7 - Ti;
626 T1R = T7 + Ti;
627 T4e = T3O - T3N;
628 T4f = Tc - Th;
629 T4g = T4e - T4f;
630 T4p = T4f + T4e;
631 }
632 {
633 E T2m, T2p, T3M, T3P;
634 T2m = T1 - T6;
635 T2p = T2n - T2o;
636 T2q = T2m - T2p;
637 T37 = T2m + T2p;
638 T3M = T2n + T2o;
639 T3P = T3N + T3O;
640 T3Q = T3M + T3P;
641 T42 = T3P - T3M;
642 }
643 }
644 {
645 E T1f, T3g, T21, T2C, T1N, T3k, T27, T2K, T1q, T3h, T22, T2F, T1C, T3j, T26;
646 E T2H;
647 {
648 E T19, T1Z, T1e, T20;
649 {
650 E T16, T18, T15, T17;
651 T16 = ri[WS(rs, 8)];
652 T18 = ii[WS(rs, 8)];
653 T15 = W[14];
654 T17 = W[15];
655 T19 = FMA(T15, T16, T17 * T18);
656 T1Z = FNMS(T17, T16, T15 * T18);
657 }
658 {
659 E T1b, T1d, T1a, T1c;
660 T1b = ri[WS(rs, 18)];
661 T1d = ii[WS(rs, 18)];
662 T1a = W[34];
663 T1c = W[35];
664 T1e = FMA(T1a, T1b, T1c * T1d);
665 T20 = FNMS(T1c, T1b, T1a * T1d);
666 }
667 T1f = T19 + T1e;
668 T3g = T1Z + T20;
669 T21 = T1Z - T20;
670 T2C = T19 - T1e;
671 }
672 {
673 E T1H, T2I, T1M, T2J;
674 {
675 E T1E, T1G, T1D, T1F;
676 T1E = ri[WS(rs, 17)];
677 T1G = ii[WS(rs, 17)];
678 T1D = W[32];
679 T1F = W[33];
680 T1H = FMA(T1D, T1E, T1F * T1G);
681 T2I = FNMS(T1F, T1E, T1D * T1G);
682 }
683 {
684 E T1J, T1L, T1I, T1K;
685 T1J = ri[WS(rs, 7)];
686 T1L = ii[WS(rs, 7)];
687 T1I = W[12];
688 T1K = W[13];
689 T1M = FMA(T1I, T1J, T1K * T1L);
690 T2J = FNMS(T1K, T1J, T1I * T1L);
691 }
692 T1N = T1H + T1M;
693 T3k = T2I + T2J;
694 T27 = T1H - T1M;
695 T2K = T2I - T2J;
696 }
697 {
698 E T1k, T2D, T1p, T2E;
699 {
700 E T1h, T1j, T1g, T1i;
701 T1h = ri[WS(rs, 13)];
702 T1j = ii[WS(rs, 13)];
703 T1g = W[24];
704 T1i = W[25];
705 T1k = FMA(T1g, T1h, T1i * T1j);
706 T2D = FNMS(T1i, T1h, T1g * T1j);
707 }
708 {
709 E T1m, T1o, T1l, T1n;
710 T1m = ri[WS(rs, 3)];
711 T1o = ii[WS(rs, 3)];
712 T1l = W[4];
713 T1n = W[5];
714 T1p = FMA(T1l, T1m, T1n * T1o);
715 T2E = FNMS(T1n, T1m, T1l * T1o);
716 }
717 T1q = T1k + T1p;
718 T3h = T2D + T2E;
719 T22 = T1k - T1p;
720 T2F = T2D - T2E;
721 }
722 {
723 E T1w, T24, T1B, T25;
724 {
725 E T1t, T1v, T1s, T1u;
726 T1t = ri[WS(rs, 12)];
727 T1v = ii[WS(rs, 12)];
728 T1s = W[22];
729 T1u = W[23];
730 T1w = FMA(T1s, T1t, T1u * T1v);
731 T24 = FNMS(T1u, T1t, T1s * T1v);
732 }
733 {
734 E T1y, T1A, T1x, T1z;
735 T1y = ri[WS(rs, 2)];
736 T1A = ii[WS(rs, 2)];
737 T1x = W[2];
738 T1z = W[3];
739 T1B = FMA(T1x, T1y, T1z * T1A);
740 T25 = FNMS(T1z, T1y, T1x * T1A);
741 }
742 T1C = T1w + T1B;
743 T3j = T24 + T25;
744 T26 = T24 - T25;
745 T2H = T1w - T1B;
746 }
747 T1r = T1f - T1q;
748 T1O = T1C - T1N;
749 T1P = T1r + T1O;
750 T3i = T3g - T3h;
751 T3l = T3j - T3k;
752 T44 = T3i + T3l;
753 T3D = T3g + T3h;
754 T3E = T3j + T3k;
755 T3K = T3D + T3E;
756 T1V = T1f + T1q;
757 T1W = T1C + T1N;
758 T1X = T1V + T1W;
759 T23 = T21 + T22;
760 T28 = T26 + T27;
761 T4r = T23 + T28;
762 T2W = T21 - T22;
763 T2X = T26 - T27;
764 T4c = T2W + T2X;
765 T33 = T2C + T2F;
766 T34 = T2H + T2K;
767 T35 = T33 + T34;
768 T2G = T2C - T2F;
769 T2L = T2H - T2K;
770 T2M = T2G + T2L;
771 }
772 {
773 E Tu, T3n, T2c, T2r, T12, T3r, T2i, T2z, TF, T3o, T2d, T2u, TR, T3q, T2h;
774 E T2w;
775 {
776 E To, T2a, Tt, T2b;
777 {
778 E Tl, Tn, Tk, Tm;
779 Tl = ri[WS(rs, 4)];
780 Tn = ii[WS(rs, 4)];
781 Tk = W[6];
782 Tm = W[7];
783 To = FMA(Tk, Tl, Tm * Tn);
784 T2a = FNMS(Tm, Tl, Tk * Tn);
785 }
786 {
787 E Tq, Ts, Tp, Tr;
788 Tq = ri[WS(rs, 14)];
789 Ts = ii[WS(rs, 14)];
790 Tp = W[26];
791 Tr = W[27];
792 Tt = FMA(Tp, Tq, Tr * Ts);
793 T2b = FNMS(Tr, Tq, Tp * Ts);
794 }
795 Tu = To + Tt;
796 T3n = T2a + T2b;
797 T2c = T2a - T2b;
798 T2r = To - Tt;
799 }
800 {
801 E TW, T2x, T11, T2y;
802 {
803 E TT, TV, TS, TU;
804 TT = ri[WS(rs, 1)];
805 TV = ii[WS(rs, 1)];
806 TS = W[0];
807 TU = W[1];
808 TW = FMA(TS, TT, TU * TV);
809 T2x = FNMS(TU, TT, TS * TV);
810 }
811 {
812 E TY, T10, TX, TZ;
813 TY = ri[WS(rs, 11)];
814 T10 = ii[WS(rs, 11)];
815 TX = W[20];
816 TZ = W[21];
817 T11 = FMA(TX, TY, TZ * T10);
818 T2y = FNMS(TZ, TY, TX * T10);
819 }
820 T12 = TW + T11;
821 T3r = T2x + T2y;
822 T2i = TW - T11;
823 T2z = T2x - T2y;
824 }
825 {
826 E Tz, T2s, TE, T2t;
827 {
828 E Tw, Ty, Tv, Tx;
829 Tw = ri[WS(rs, 9)];
830 Ty = ii[WS(rs, 9)];
831 Tv = W[16];
832 Tx = W[17];
833 Tz = FMA(Tv, Tw, Tx * Ty);
834 T2s = FNMS(Tx, Tw, Tv * Ty);
835 }
836 {
837 E TB, TD, TA, TC;
838 TB = ri[WS(rs, 19)];
839 TD = ii[WS(rs, 19)];
840 TA = W[36];
841 TC = W[37];
842 TE = FMA(TA, TB, TC * TD);
843 T2t = FNMS(TC, TB, TA * TD);
844 }
845 TF = Tz + TE;
846 T3o = T2s + T2t;
847 T2d = Tz - TE;
848 T2u = T2s - T2t;
849 }
850 {
851 E TL, T2f, TQ, T2g;
852 {
853 E TI, TK, TH, TJ;
854 TI = ri[WS(rs, 16)];
855 TK = ii[WS(rs, 16)];
856 TH = W[30];
857 TJ = W[31];
858 TL = FMA(TH, TI, TJ * TK);
859 T2f = FNMS(TJ, TI, TH * TK);
860 }
861 {
862 E TN, TP, TM, TO;
863 TN = ri[WS(rs, 6)];
864 TP = ii[WS(rs, 6)];
865 TM = W[10];
866 TO = W[11];
867 TQ = FMA(TM, TN, TO * TP);
868 T2g = FNMS(TO, TN, TM * TP);
869 }
870 TR = TL + TQ;
871 T3q = T2f + T2g;
872 T2h = T2f - T2g;
873 T2w = TL - TQ;
874 }
875 TG = Tu - TF;
876 T13 = TR - T12;
877 T14 = TG + T13;
878 T3p = T3n - T3o;
879 T3s = T3q - T3r;
880 T43 = T3p + T3s;
881 T3A = T3n + T3o;
882 T3B = T3q + T3r;
883 T3J = T3A + T3B;
884 T1S = Tu + TF;
885 T1T = TR + T12;
886 T1U = T1S + T1T;
887 T2e = T2c + T2d;
888 T2j = T2h + T2i;
889 T4q = T2e + T2j;
890 T2T = T2c - T2d;
891 T2U = T2h - T2i;
892 T4b = T2T + T2U;
893 T30 = T2r + T2u;
894 T31 = T2w + T2z;
895 T32 = T30 + T31;
896 T2v = T2r - T2u;
897 T2A = T2w - T2z;
898 T2B = T2v + T2A;
899 }
900 {
901 E T3e, T1Q, T3d, T3u, T3w, T3m, T3t, T3v, T3f;
902 T3e = KP559016994 * (T14 - T1P);
903 T1Q = T14 + T1P;
904 T3d = FNMS(KP250000000, T1Q, Tj);
905 T3m = T3i - T3l;
906 T3t = T3p - T3s;
907 T3u = FNMS(KP587785252, T3t, KP951056516 * T3m);
908 T3w = FMA(KP951056516, T3t, KP587785252 * T3m);
909 ri[WS(rs, 10)] = Tj + T1Q;
910 T3v = T3e + T3d;
911 ri[WS(rs, 14)] = T3v - T3w;
912 ri[WS(rs, 6)] = T3v + T3w;
913 T3f = T3d - T3e;
914 ri[WS(rs, 2)] = T3f - T3u;
915 ri[WS(rs, 18)] = T3f + T3u;
916 }
917 {
918 E T47, T45, T46, T41, T4a, T3Z, T40, T49, T48;
919 T47 = KP559016994 * (T43 - T44);
920 T45 = T43 + T44;
921 T46 = FNMS(KP250000000, T45, T42);
922 T3Z = T1r - T1O;
923 T40 = TG - T13;
924 T41 = FNMS(KP587785252, T40, KP951056516 * T3Z);
925 T4a = FMA(KP951056516, T40, KP587785252 * T3Z);
926 ii[WS(rs, 10)] = T45 + T42;
927 T49 = T47 + T46;
928 ii[WS(rs, 6)] = T49 - T4a;
929 ii[WS(rs, 14)] = T4a + T49;
930 T48 = T46 - T47;
931 ii[WS(rs, 2)] = T41 + T48;
932 ii[WS(rs, 18)] = T48 - T41;
933 }
934 {
935 E T3x, T1Y, T3y, T3G, T3I, T3C, T3F, T3H, T3z;
936 T3x = KP559016994 * (T1U - T1X);
937 T1Y = T1U + T1X;
938 T3y = FNMS(KP250000000, T1Y, T1R);
939 T3C = T3A - T3B;
940 T3F = T3D - T3E;
941 T3G = FMA(KP951056516, T3C, KP587785252 * T3F);
942 T3I = FNMS(KP587785252, T3C, KP951056516 * T3F);
943 ri[0] = T1R + T1Y;
944 T3H = T3y - T3x;
945 ri[WS(rs, 12)] = T3H - T3I;
946 ri[WS(rs, 8)] = T3H + T3I;
947 T3z = T3x + T3y;
948 ri[WS(rs, 4)] = T3z - T3G;
949 ri[WS(rs, 16)] = T3z + T3G;
950 }
951 {
952 E T3U, T3L, T3V, T3T, T3Y, T3R, T3S, T3X, T3W;
953 T3U = KP559016994 * (T3J - T3K);
954 T3L = T3J + T3K;
955 T3V = FNMS(KP250000000, T3L, T3Q);
956 T3R = T1S - T1T;
957 T3S = T1V - T1W;
958 T3T = FMA(KP951056516, T3R, KP587785252 * T3S);
959 T3Y = FNMS(KP587785252, T3R, KP951056516 * T3S);
960 ii[0] = T3L + T3Q;
961 T3X = T3V - T3U;
962 ii[WS(rs, 8)] = T3X - T3Y;
963 ii[WS(rs, 12)] = T3Y + T3X;
964 T3W = T3U + T3V;
965 ii[WS(rs, 4)] = T3T + T3W;
966 ii[WS(rs, 16)] = T3W - T3T;
967 }
968 {
969 E T2P, T2N, T2O, T2l, T2R, T29, T2k, T2S, T2Q;
970 T2P = KP559016994 * (T2B - T2M);
971 T2N = T2B + T2M;
972 T2O = FNMS(KP250000000, T2N, T2q);
973 T29 = T23 - T28;
974 T2k = T2e - T2j;
975 T2l = FNMS(KP587785252, T2k, KP951056516 * T29);
976 T2R = FMA(KP951056516, T2k, KP587785252 * T29);
977 ri[WS(rs, 15)] = T2q + T2N;
978 T2S = T2P + T2O;
979 ri[WS(rs, 11)] = T2R + T2S;
980 ri[WS(rs, 19)] = T2S - T2R;
981 T2Q = T2O - T2P;
982 ri[WS(rs, 3)] = T2l + T2Q;
983 ri[WS(rs, 7)] = T2Q - T2l;
984 }
985 {
986 E T4u, T4s, T4t, T4y, T4A, T4w, T4x, T4z, T4v;
987 T4u = KP559016994 * (T4q - T4r);
988 T4s = T4q + T4r;
989 T4t = FNMS(KP250000000, T4s, T4p);
990 T4w = T2G - T2L;
991 T4x = T2v - T2A;
992 T4y = FNMS(KP587785252, T4x, KP951056516 * T4w);
993 T4A = FMA(KP951056516, T4x, KP587785252 * T4w);
994 ii[WS(rs, 15)] = T4s + T4p;
995 T4z = T4u + T4t;
996 ii[WS(rs, 11)] = T4z - T4A;
997 ii[WS(rs, 19)] = T4A + T4z;
998 T4v = T4t - T4u;
999 ii[WS(rs, 3)] = T4v - T4y;
1000 ii[WS(rs, 7)] = T4y + T4v;
1001 }
1002 {
1003 E T36, T38, T39, T2Z, T3b, T2V, T2Y, T3c, T3a;
1004 T36 = KP559016994 * (T32 - T35);
1005 T38 = T32 + T35;
1006 T39 = FNMS(KP250000000, T38, T37);
1007 T2V = T2T - T2U;
1008 T2Y = T2W - T2X;
1009 T2Z = FMA(KP951056516, T2V, KP587785252 * T2Y);
1010 T3b = FNMS(KP587785252, T2V, KP951056516 * T2Y);
1011 ri[WS(rs, 5)] = T37 + T38;
1012 T3c = T39 - T36;
1013 ri[WS(rs, 13)] = T3b + T3c;
1014 ri[WS(rs, 17)] = T3c - T3b;
1015 T3a = T36 + T39;
1016 ri[WS(rs, 1)] = T2Z + T3a;
1017 ri[WS(rs, 9)] = T3a - T2Z;
1018 }
1019 {
1020 E T4d, T4h, T4i, T4m, T4o, T4k, T4l, T4n, T4j;
1021 T4d = KP559016994 * (T4b - T4c);
1022 T4h = T4b + T4c;
1023 T4i = FNMS(KP250000000, T4h, T4g);
1024 T4k = T30 - T31;
1025 T4l = T33 - T34;
1026 T4m = FMA(KP951056516, T4k, KP587785252 * T4l);
1027 T4o = FNMS(KP587785252, T4k, KP951056516 * T4l);
1028 ii[WS(rs, 5)] = T4h + T4g;
1029 T4n = T4i - T4d;
1030 ii[WS(rs, 13)] = T4n - T4o;
1031 ii[WS(rs, 17)] = T4o + T4n;
1032 T4j = T4d + T4i;
1033 ii[WS(rs, 1)] = T4j - T4m;
1034 ii[WS(rs, 9)] = T4m + T4j;
1035 }
1036 }
1037 }
1038 }
1039
1040 static const tw_instr twinstr[] = {
1041 {TW_FULL, 0, 20},
1042 {TW_NEXT, 1, 0}
1043 };
1044
1045 static const ct_desc desc = { 20, "t1_20", twinstr, &GENUS, {184, 62, 62, 0}, 0, 0, 0 };
1046
1047 void X(codelet_t1_20) (planner *p) {
1048 X(kdft_dit_register) (p, t1_20, &desc);
1049 }
1050 #endif