comparison src/fftw-3.3.8/rdft/scalar/r2cf/hf_15.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 08:06:30 EDT 2018 */
23
24 #include "rdft/codelet-rdft.h"
25
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -compact -variables 4 -pipeline-latency 4 -n 15 -dit -name hf_15 -include rdft/scalar/hf.h */
29
30 /*
31 * This function contains 184 FP additions, 140 FP multiplications,
32 * (or, 72 additions, 28 multiplications, 112 fused multiply/add),
33 * 51 stack variables, 6 constants, and 60 memory accesses
34 */
35 #include "rdft/scalar/hf.h"
36
37 static void hf_15(R *cr, R *ci, 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 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
44 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
45 {
46 INT m;
47 for (m = mb, W = W + ((mb - 1) * 28); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
48 E T1, T3i, T1G, T3l, Te, T1B, T3j, T3k, T1y, T2i, T2a, T2M, T37, T2Y, Tz;
49 E T2e, T1O, T2t, T39, T2U, TT, T2f, T1V, T2z, T3a, T2V, T1e, T2h, T23, T2G;
50 E T36, T2X;
51 {
52 E T7, T1D, Td, T1F;
53 T1 = cr[0];
54 T3i = ci[0];
55 {
56 E T3, T6, T4, T1C, T2, T5;
57 T3 = cr[WS(rs, 5)];
58 T6 = ci[WS(rs, 5)];
59 T2 = W[8];
60 T4 = T2 * T3;
61 T1C = T2 * T6;
62 T5 = W[9];
63 T7 = FMA(T5, T6, T4);
64 T1D = FNMS(T5, T3, T1C);
65 }
66 {
67 E T9, Tc, Ta, T1E, T8, Tb;
68 T9 = cr[WS(rs, 10)];
69 Tc = ci[WS(rs, 10)];
70 T8 = W[18];
71 Ta = T8 * T9;
72 T1E = T8 * Tc;
73 Tb = W[19];
74 Td = FMA(Tb, Tc, Ta);
75 T1F = FNMS(Tb, T9, T1E);
76 }
77 T1G = T1D - T1F;
78 T3l = Td - T7;
79 Te = T7 + Td;
80 T1B = FNMS(KP500000000, Te, T1);
81 T3j = T1D + T1F;
82 T3k = FNMS(KP500000000, T3j, T3i);
83 }
84 {
85 E T1k, T2I, T1w, T28, T1q, T26;
86 {
87 E T1g, T1j, T1h, T2H, T1f, T1i;
88 T1g = cr[WS(rs, 9)];
89 T1j = ci[WS(rs, 9)];
90 T1f = W[16];
91 T1h = T1f * T1g;
92 T2H = T1f * T1j;
93 T1i = W[17];
94 T1k = FMA(T1i, T1j, T1h);
95 T2I = FNMS(T1i, T1g, T2H);
96 }
97 {
98 E T1s, T1v, T1t, T27, T1r, T1u;
99 T1s = cr[WS(rs, 4)];
100 T1v = ci[WS(rs, 4)];
101 T1r = W[6];
102 T1t = T1r * T1s;
103 T27 = T1r * T1v;
104 T1u = W[7];
105 T1w = FMA(T1u, T1v, T1t);
106 T28 = FNMS(T1u, T1s, T27);
107 }
108 {
109 E T1m, T1p, T1n, T25, T1l, T1o;
110 T1m = cr[WS(rs, 14)];
111 T1p = ci[WS(rs, 14)];
112 T1l = W[26];
113 T1n = T1l * T1m;
114 T25 = T1l * T1p;
115 T1o = W[27];
116 T1q = FMA(T1o, T1p, T1n);
117 T26 = FNMS(T1o, T1m, T25);
118 }
119 {
120 E T29, T1x, T24, T2L, T2J, T2K;
121 T29 = T26 - T28;
122 T1x = T1q + T1w;
123 T24 = FNMS(KP500000000, T1x, T1k);
124 T1y = T1k + T1x;
125 T2i = FMA(KP866025403, T29, T24);
126 T2a = FNMS(KP866025403, T29, T24);
127 T2L = T1q - T1w;
128 T2J = T26 + T28;
129 T2K = FNMS(KP500000000, T2J, T2I);
130 T2M = FNMS(KP866025403, T2L, T2K);
131 T37 = T2I + T2J;
132 T2Y = FMA(KP866025403, T2L, T2K);
133 }
134 }
135 {
136 E Tl, T2p, Tx, T1M, Tr, T1K;
137 {
138 E Th, Tk, Ti, T2o, Tg, Tj;
139 Th = cr[WS(rs, 3)];
140 Tk = ci[WS(rs, 3)];
141 Tg = W[4];
142 Ti = Tg * Th;
143 T2o = Tg * Tk;
144 Tj = W[5];
145 Tl = FMA(Tj, Tk, Ti);
146 T2p = FNMS(Tj, Th, T2o);
147 }
148 {
149 E Tt, Tw, Tu, T1L, Ts, Tv;
150 Tt = cr[WS(rs, 13)];
151 Tw = ci[WS(rs, 13)];
152 Ts = W[24];
153 Tu = Ts * Tt;
154 T1L = Ts * Tw;
155 Tv = W[25];
156 Tx = FMA(Tv, Tw, Tu);
157 T1M = FNMS(Tv, Tt, T1L);
158 }
159 {
160 E Tn, Tq, To, T1J, Tm, Tp;
161 Tn = cr[WS(rs, 8)];
162 Tq = ci[WS(rs, 8)];
163 Tm = W[14];
164 To = Tm * Tn;
165 T1J = Tm * Tq;
166 Tp = W[15];
167 Tr = FMA(Tp, Tq, To);
168 T1K = FNMS(Tp, Tn, T1J);
169 }
170 {
171 E T1N, Ty, T1I, T2s, T2q, T2r;
172 T1N = T1K - T1M;
173 Ty = Tr + Tx;
174 T1I = FNMS(KP500000000, Ty, Tl);
175 Tz = Tl + Ty;
176 T2e = FMA(KP866025403, T1N, T1I);
177 T1O = FNMS(KP866025403, T1N, T1I);
178 T2s = Tr - Tx;
179 T2q = T1K + T1M;
180 T2r = FNMS(KP500000000, T2q, T2p);
181 T2t = FNMS(KP866025403, T2s, T2r);
182 T39 = T2p + T2q;
183 T2U = FMA(KP866025403, T2s, T2r);
184 }
185 }
186 {
187 E TF, T2v, TR, T1T, TL, T1R;
188 {
189 E TB, TE, TC, T2u, TA, TD;
190 TB = cr[WS(rs, 12)];
191 TE = ci[WS(rs, 12)];
192 TA = W[22];
193 TC = TA * TB;
194 T2u = TA * TE;
195 TD = W[23];
196 TF = FMA(TD, TE, TC);
197 T2v = FNMS(TD, TB, T2u);
198 }
199 {
200 E TN, TQ, TO, T1S, TM, TP;
201 TN = cr[WS(rs, 7)];
202 TQ = ci[WS(rs, 7)];
203 TM = W[12];
204 TO = TM * TN;
205 T1S = TM * TQ;
206 TP = W[13];
207 TR = FMA(TP, TQ, TO);
208 T1T = FNMS(TP, TN, T1S);
209 }
210 {
211 E TH, TK, TI, T1Q, TG, TJ;
212 TH = cr[WS(rs, 2)];
213 TK = ci[WS(rs, 2)];
214 TG = W[2];
215 TI = TG * TH;
216 T1Q = TG * TK;
217 TJ = W[3];
218 TL = FMA(TJ, TK, TI);
219 T1R = FNMS(TJ, TH, T1Q);
220 }
221 {
222 E T1U, TS, T1P, T2y, T2w, T2x;
223 T1U = T1R - T1T;
224 TS = TL + TR;
225 T1P = FNMS(KP500000000, TS, TF);
226 TT = TF + TS;
227 T2f = FMA(KP866025403, T1U, T1P);
228 T1V = FNMS(KP866025403, T1U, T1P);
229 T2y = TL - TR;
230 T2w = T1R + T1T;
231 T2x = FNMS(KP500000000, T2w, T2v);
232 T2z = FNMS(KP866025403, T2y, T2x);
233 T3a = T2v + T2w;
234 T2V = FMA(KP866025403, T2y, T2x);
235 }
236 }
237 {
238 E T10, T2C, T1c, T21, T16, T1Z;
239 {
240 E TW, TZ, TX, T2B, TV, TY;
241 TW = cr[WS(rs, 6)];
242 TZ = ci[WS(rs, 6)];
243 TV = W[10];
244 TX = TV * TW;
245 T2B = TV * TZ;
246 TY = W[11];
247 T10 = FMA(TY, TZ, TX);
248 T2C = FNMS(TY, TW, T2B);
249 }
250 {
251 E T18, T1b, T19, T20, T17, T1a;
252 T18 = cr[WS(rs, 1)];
253 T1b = ci[WS(rs, 1)];
254 T17 = W[0];
255 T19 = T17 * T18;
256 T20 = T17 * T1b;
257 T1a = W[1];
258 T1c = FMA(T1a, T1b, T19);
259 T21 = FNMS(T1a, T18, T20);
260 }
261 {
262 E T12, T15, T13, T1Y, T11, T14;
263 T12 = cr[WS(rs, 11)];
264 T15 = ci[WS(rs, 11)];
265 T11 = W[20];
266 T13 = T11 * T12;
267 T1Y = T11 * T15;
268 T14 = W[21];
269 T16 = FMA(T14, T15, T13);
270 T1Z = FNMS(T14, T12, T1Y);
271 }
272 {
273 E T22, T1d, T1X, T2F, T2D, T2E;
274 T22 = T1Z - T21;
275 T1d = T16 + T1c;
276 T1X = FNMS(KP500000000, T1d, T10);
277 T1e = T10 + T1d;
278 T2h = FMA(KP866025403, T22, T1X);
279 T23 = FNMS(KP866025403, T22, T1X);
280 T2F = T16 - T1c;
281 T2D = T1Z + T21;
282 T2E = FNMS(KP500000000, T2D, T2C);
283 T2G = FNMS(KP866025403, T2F, T2E);
284 T36 = T2C + T2D;
285 T2X = FMA(KP866025403, T2F, T2E);
286 }
287 }
288 {
289 E T3c, T3e, Tf, T1A, T33, T34, T3d, T35;
290 {
291 E T38, T3b, TU, T1z;
292 T38 = T36 - T37;
293 T3b = T39 - T3a;
294 T3c = FNMS(KP618033988, T3b, T38);
295 T3e = FMA(KP618033988, T38, T3b);
296 Tf = T1 + Te;
297 TU = Tz + TT;
298 T1z = T1e + T1y;
299 T1A = TU + T1z;
300 T33 = FNMS(KP250000000, T1A, Tf);
301 T34 = TU - T1z;
302 }
303 cr[0] = Tf + T1A;
304 T3d = FMA(KP559016994, T34, T33);
305 ci[WS(rs, 5)] = FNMS(KP951056516, T3e, T3d);
306 cr[WS(rs, 6)] = FMA(KP951056516, T3e, T3d);
307 T35 = FNMS(KP559016994, T34, T33);
308 ci[WS(rs, 2)] = FNMS(KP951056516, T3c, T35);
309 cr[WS(rs, 3)] = FMA(KP951056516, T3c, T35);
310 }
311 {
312 E T30, T32, T1H, T2c, T2R, T2S, T31, T2T;
313 {
314 E T2W, T2Z, T1W, T2b;
315 T2W = T2U - T2V;
316 T2Z = T2X - T2Y;
317 T30 = FMA(KP618033988, T2Z, T2W);
318 T32 = FNMS(KP618033988, T2W, T2Z);
319 T1H = FNMS(KP866025403, T1G, T1B);
320 T1W = T1O + T1V;
321 T2b = T23 + T2a;
322 T2c = T1W + T2b;
323 T2R = FNMS(KP250000000, T2c, T1H);
324 T2S = T1W - T2b;
325 }
326 cr[WS(rs, 5)] = T1H + T2c;
327 T31 = FNMS(KP559016994, T2S, T2R);
328 cr[WS(rs, 2)] = FNMS(KP951056516, T32, T31);
329 ci[WS(rs, 6)] = FMA(KP951056516, T32, T31);
330 T2T = FMA(KP559016994, T2S, T2R);
331 ci[0] = FNMS(KP951056516, T30, T2T);
332 ci[WS(rs, 3)] = FMA(KP951056516, T30, T2T);
333 }
334 {
335 E T2O, T2Q, T2d, T2k, T2l, T2m, T2n, T2P;
336 {
337 E T2A, T2N, T2g, T2j;
338 T2A = T2t - T2z;
339 T2N = T2G - T2M;
340 T2O = FMA(KP618033988, T2N, T2A);
341 T2Q = FNMS(KP618033988, T2A, T2N);
342 T2d = FMA(KP866025403, T1G, T1B);
343 T2g = T2e + T2f;
344 T2j = T2h + T2i;
345 T2k = T2g + T2j;
346 T2l = FNMS(KP250000000, T2k, T2d);
347 T2m = T2g - T2j;
348 }
349 ci[WS(rs, 4)] = T2d + T2k;
350 T2n = FMA(KP559016994, T2m, T2l);
351 cr[WS(rs, 4)] = FNMS(KP951056516, T2O, T2n);
352 cr[WS(rs, 1)] = FMA(KP951056516, T2O, T2n);
353 T2P = FNMS(KP559016994, T2m, T2l);
354 cr[WS(rs, 7)] = FNMS(KP951056516, T2Q, T2P);
355 ci[WS(rs, 1)] = FMA(KP951056516, T2Q, T2P);
356 }
357 {
358 E T3s, T3u, T3m, T3h, T3n, T3o, T3t, T3p;
359 {
360 E T3q, T3r, T3f, T3g;
361 T3q = T2h - T2i;
362 T3r = T2e - T2f;
363 T3s = FNMS(KP618033988, T3r, T3q);
364 T3u = FMA(KP618033988, T3q, T3r);
365 T3m = FMA(KP866025403, T3l, T3k);
366 T3f = T2t + T2z;
367 T3g = T2G + T2M;
368 T3h = T3f + T3g;
369 T3n = FNMS(KP250000000, T3h, T3m);
370 T3o = T3f - T3g;
371 }
372 cr[WS(rs, 10)] = -(T3h + T3m);
373 T3t = FMA(KP559016994, T3o, T3n);
374 ci[WS(rs, 10)] = FMA(KP951056516, T3u, T3t);
375 ci[WS(rs, 13)] = FNMS(KP951056516, T3u, T3t);
376 T3p = FNMS(KP559016994, T3o, T3n);
377 cr[WS(rs, 13)] = FMS(KP951056516, T3s, T3p);
378 ci[WS(rs, 7)] = FMA(KP951056516, T3s, T3p);
379 }
380 {
381 E T3Q, T3S, T3H, T3K, T3L, T3M, T3R, T3N;
382 {
383 E T3O, T3P, T3I, T3J;
384 T3O = TT - Tz;
385 T3P = T1y - T1e;
386 T3Q = FMA(KP618033988, T3P, T3O);
387 T3S = FNMS(KP618033988, T3O, T3P);
388 T3H = T3j + T3i;
389 T3I = T39 + T3a;
390 T3J = T36 + T37;
391 T3K = T3I + T3J;
392 T3L = FNMS(KP250000000, T3K, T3H);
393 T3M = T3I - T3J;
394 }
395 ci[WS(rs, 14)] = T3K + T3H;
396 T3R = FNMS(KP559016994, T3M, T3L);
397 cr[WS(rs, 12)] = FMS(KP951056516, T3S, T3R);
398 ci[WS(rs, 11)] = FMA(KP951056516, T3S, T3R);
399 T3N = FMA(KP559016994, T3M, T3L);
400 cr[WS(rs, 9)] = FMS(KP951056516, T3Q, T3N);
401 ci[WS(rs, 8)] = FMA(KP951056516, T3Q, T3N);
402 }
403 {
404 E T3E, T3G, T3v, T3y, T3z, T3A, T3F, T3B;
405 {
406 E T3C, T3D, T3w, T3x;
407 T3C = T1O - T1V;
408 T3D = T23 - T2a;
409 T3E = FMA(KP618033988, T3D, T3C);
410 T3G = FNMS(KP618033988, T3C, T3D);
411 T3v = FNMS(KP866025403, T3l, T3k);
412 T3w = T2U + T2V;
413 T3x = T2X + T2Y;
414 T3y = T3w + T3x;
415 T3z = FNMS(KP250000000, T3y, T3v);
416 T3A = T3x - T3w;
417 }
418 ci[WS(rs, 9)] = T3y + T3v;
419 T3F = FMA(KP559016994, T3A, T3z);
420 cr[WS(rs, 8)] = FMS(KP951056516, T3G, T3F);
421 ci[WS(rs, 12)] = FMA(KP951056516, T3G, T3F);
422 T3B = FNMS(KP559016994, T3A, T3z);
423 cr[WS(rs, 11)] = FMS(KP951056516, T3E, T3B);
424 cr[WS(rs, 14)] = -(FMA(KP951056516, T3E, T3B));
425 }
426 }
427 }
428 }
429
430 static const tw_instr twinstr[] = {
431 {TW_FULL, 1, 15},
432 {TW_NEXT, 1, 0}
433 };
434
435 static const hc2hc_desc desc = { 15, "hf_15", twinstr, &GENUS, {72, 28, 112, 0} };
436
437 void X(codelet_hf_15) (planner *p) {
438 X(khc2hc_register) (p, hf_15, &desc);
439 }
440 #else
441
442 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 15 -dit -name hf_15 -include rdft/scalar/hf.h */
443
444 /*
445 * This function contains 184 FP additions, 112 FP multiplications,
446 * (or, 128 additions, 56 multiplications, 56 fused multiply/add),
447 * 65 stack variables, 6 constants, and 60 memory accesses
448 */
449 #include "rdft/scalar/hf.h"
450
451 static void hf_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
452 {
453 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
454 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
455 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
456 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
457 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
458 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
459 {
460 INT m;
461 for (m = mb, W = W + ((mb - 1) * 28); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
462 E T1q, T2Q, Td, T1n, T2T, T3l, T13, T1k, T1l, T2E, T2F, T3j, T1H, T1T, T2k;
463 E T2w, T2f, T2v, T1M, T1U, Tu, TL, TM, T2H, T2I, T3i, T1w, T1Q, T29, T2t;
464 E T24, T2s, T1B, T1R;
465 {
466 E T1, T2R, T6, T1o, Tb, T1p, Tc, T2S;
467 T1 = cr[0];
468 T2R = ci[0];
469 {
470 E T3, T5, T2, T4;
471 T3 = cr[WS(rs, 5)];
472 T5 = ci[WS(rs, 5)];
473 T2 = W[8];
474 T4 = W[9];
475 T6 = FMA(T2, T3, T4 * T5);
476 T1o = FNMS(T4, T3, T2 * T5);
477 }
478 {
479 E T8, Ta, T7, T9;
480 T8 = cr[WS(rs, 10)];
481 Ta = ci[WS(rs, 10)];
482 T7 = W[18];
483 T9 = W[19];
484 Tb = FMA(T7, T8, T9 * Ta);
485 T1p = FNMS(T9, T8, T7 * Ta);
486 }
487 T1q = KP866025403 * (T1o - T1p);
488 T2Q = KP866025403 * (Tb - T6);
489 Tc = T6 + Tb;
490 Td = T1 + Tc;
491 T1n = FNMS(KP500000000, Tc, T1);
492 T2S = T1o + T1p;
493 T2T = FNMS(KP500000000, T2S, T2R);
494 T3l = T2S + T2R;
495 }
496 {
497 E TR, T2c, T18, T2h, TW, T1E, T11, T1F, T12, T2d, T1d, T1J, T1i, T1K, T1j;
498 E T2i;
499 {
500 E TO, TQ, TN, TP;
501 TO = cr[WS(rs, 6)];
502 TQ = ci[WS(rs, 6)];
503 TN = W[10];
504 TP = W[11];
505 TR = FMA(TN, TO, TP * TQ);
506 T2c = FNMS(TP, TO, TN * TQ);
507 }
508 {
509 E T15, T17, T14, T16;
510 T15 = cr[WS(rs, 9)];
511 T17 = ci[WS(rs, 9)];
512 T14 = W[16];
513 T16 = W[17];
514 T18 = FMA(T14, T15, T16 * T17);
515 T2h = FNMS(T16, T15, T14 * T17);
516 }
517 {
518 E TT, TV, TS, TU;
519 TT = cr[WS(rs, 11)];
520 TV = ci[WS(rs, 11)];
521 TS = W[20];
522 TU = W[21];
523 TW = FMA(TS, TT, TU * TV);
524 T1E = FNMS(TU, TT, TS * TV);
525 }
526 {
527 E TY, T10, TX, TZ;
528 TY = cr[WS(rs, 1)];
529 T10 = ci[WS(rs, 1)];
530 TX = W[0];
531 TZ = W[1];
532 T11 = FMA(TX, TY, TZ * T10);
533 T1F = FNMS(TZ, TY, TX * T10);
534 }
535 T12 = TW + T11;
536 T2d = T1E + T1F;
537 {
538 E T1a, T1c, T19, T1b;
539 T1a = cr[WS(rs, 14)];
540 T1c = ci[WS(rs, 14)];
541 T19 = W[26];
542 T1b = W[27];
543 T1d = FMA(T19, T1a, T1b * T1c);
544 T1J = FNMS(T1b, T1a, T19 * T1c);
545 }
546 {
547 E T1f, T1h, T1e, T1g;
548 T1f = cr[WS(rs, 4)];
549 T1h = ci[WS(rs, 4)];
550 T1e = W[6];
551 T1g = W[7];
552 T1i = FMA(T1e, T1f, T1g * T1h);
553 T1K = FNMS(T1g, T1f, T1e * T1h);
554 }
555 T1j = T1d + T1i;
556 T2i = T1J + T1K;
557 {
558 E T1D, T1G, T2g, T2j;
559 T13 = TR + T12;
560 T1k = T18 + T1j;
561 T1l = T13 + T1k;
562 T2E = T2c + T2d;
563 T2F = T2h + T2i;
564 T3j = T2E + T2F;
565 T1D = FNMS(KP500000000, T12, TR);
566 T1G = KP866025403 * (T1E - T1F);
567 T1H = T1D - T1G;
568 T1T = T1D + T1G;
569 T2g = KP866025403 * (T1d - T1i);
570 T2j = FNMS(KP500000000, T2i, T2h);
571 T2k = T2g - T2j;
572 T2w = T2g + T2j;
573 {
574 E T2b, T2e, T1I, T1L;
575 T2b = KP866025403 * (T11 - TW);
576 T2e = FNMS(KP500000000, T2d, T2c);
577 T2f = T2b + T2e;
578 T2v = T2e - T2b;
579 T1I = FNMS(KP500000000, T1j, T18);
580 T1L = KP866025403 * (T1J - T1K);
581 T1M = T1I - T1L;
582 T1U = T1I + T1L;
583 }
584 }
585 }
586 {
587 E Ti, T21, Tz, T26, Tn, T1t, Ts, T1u, Tt, T22, TE, T1y, TJ, T1z, TK;
588 E T27;
589 {
590 E Tf, Th, Te, Tg;
591 Tf = cr[WS(rs, 3)];
592 Th = ci[WS(rs, 3)];
593 Te = W[4];
594 Tg = W[5];
595 Ti = FMA(Te, Tf, Tg * Th);
596 T21 = FNMS(Tg, Tf, Te * Th);
597 }
598 {
599 E Tw, Ty, Tv, Tx;
600 Tw = cr[WS(rs, 12)];
601 Ty = ci[WS(rs, 12)];
602 Tv = W[22];
603 Tx = W[23];
604 Tz = FMA(Tv, Tw, Tx * Ty);
605 T26 = FNMS(Tx, Tw, Tv * Ty);
606 }
607 {
608 E Tk, Tm, Tj, Tl;
609 Tk = cr[WS(rs, 8)];
610 Tm = ci[WS(rs, 8)];
611 Tj = W[14];
612 Tl = W[15];
613 Tn = FMA(Tj, Tk, Tl * Tm);
614 T1t = FNMS(Tl, Tk, Tj * Tm);
615 }
616 {
617 E Tp, Tr, To, Tq;
618 Tp = cr[WS(rs, 13)];
619 Tr = ci[WS(rs, 13)];
620 To = W[24];
621 Tq = W[25];
622 Ts = FMA(To, Tp, Tq * Tr);
623 T1u = FNMS(Tq, Tp, To * Tr);
624 }
625 Tt = Tn + Ts;
626 T22 = T1t + T1u;
627 {
628 E TB, TD, TA, TC;
629 TB = cr[WS(rs, 2)];
630 TD = ci[WS(rs, 2)];
631 TA = W[2];
632 TC = W[3];
633 TE = FMA(TA, TB, TC * TD);
634 T1y = FNMS(TC, TB, TA * TD);
635 }
636 {
637 E TG, TI, TF, TH;
638 TG = cr[WS(rs, 7)];
639 TI = ci[WS(rs, 7)];
640 TF = W[12];
641 TH = W[13];
642 TJ = FMA(TF, TG, TH * TI);
643 T1z = FNMS(TH, TG, TF * TI);
644 }
645 TK = TE + TJ;
646 T27 = T1y + T1z;
647 {
648 E T1s, T1v, T25, T28;
649 Tu = Ti + Tt;
650 TL = Tz + TK;
651 TM = Tu + TL;
652 T2H = T21 + T22;
653 T2I = T26 + T27;
654 T3i = T2H + T2I;
655 T1s = FNMS(KP500000000, Tt, Ti);
656 T1v = KP866025403 * (T1t - T1u);
657 T1w = T1s - T1v;
658 T1Q = T1s + T1v;
659 T25 = KP866025403 * (TJ - TE);
660 T28 = FNMS(KP500000000, T27, T26);
661 T29 = T25 + T28;
662 T2t = T28 - T25;
663 {
664 E T20, T23, T1x, T1A;
665 T20 = KP866025403 * (Ts - Tn);
666 T23 = FNMS(KP500000000, T22, T21);
667 T24 = T20 + T23;
668 T2s = T23 - T20;
669 T1x = FNMS(KP500000000, TK, Tz);
670 T1A = KP866025403 * (T1y - T1z);
671 T1B = T1x - T1A;
672 T1R = T1x + T1A;
673 }
674 }
675 }
676 {
677 E T2C, T1m, T2B, T2K, T2M, T2G, T2J, T2L, T2D;
678 T2C = KP559016994 * (TM - T1l);
679 T1m = TM + T1l;
680 T2B = FNMS(KP250000000, T1m, Td);
681 T2G = T2E - T2F;
682 T2J = T2H - T2I;
683 T2K = FNMS(KP587785252, T2J, KP951056516 * T2G);
684 T2M = FMA(KP951056516, T2J, KP587785252 * T2G);
685 cr[0] = Td + T1m;
686 T2L = T2C + T2B;
687 ci[WS(rs, 5)] = T2L - T2M;
688 cr[WS(rs, 6)] = T2L + T2M;
689 T2D = T2B - T2C;
690 ci[WS(rs, 2)] = T2D - T2K;
691 cr[WS(rs, 3)] = T2D + T2K;
692 }
693 {
694 E T3k, T3m, T3n, T3h, T3p, T3f, T3g, T3q, T3o;
695 T3k = KP559016994 * (T3i - T3j);
696 T3m = T3i + T3j;
697 T3n = FNMS(KP250000000, T3m, T3l);
698 T3f = T1k - T13;
699 T3g = Tu - TL;
700 T3h = FNMS(KP951056516, T3g, KP587785252 * T3f);
701 T3p = FMA(KP587785252, T3g, KP951056516 * T3f);
702 ci[WS(rs, 14)] = T3m + T3l;
703 T3q = T3n - T3k;
704 cr[WS(rs, 12)] = T3p - T3q;
705 ci[WS(rs, 11)] = T3p + T3q;
706 T3o = T3k + T3n;
707 cr[WS(rs, 9)] = T3h - T3o;
708 ci[WS(rs, 8)] = T3h + T3o;
709 }
710 {
711 E T2y, T2A, T1r, T1O, T2p, T2q, T2z, T2r;
712 {
713 E T2u, T2x, T1C, T1N;
714 T2u = T2s - T2t;
715 T2x = T2v - T2w;
716 T2y = FMA(KP951056516, T2u, KP587785252 * T2x);
717 T2A = FNMS(KP587785252, T2u, KP951056516 * T2x);
718 T1r = T1n - T1q;
719 T1C = T1w + T1B;
720 T1N = T1H + T1M;
721 T1O = T1C + T1N;
722 T2p = KP559016994 * (T1C - T1N);
723 T2q = FNMS(KP250000000, T1O, T1r);
724 }
725 cr[WS(rs, 5)] = T1r + T1O;
726 T2z = T2q - T2p;
727 cr[WS(rs, 2)] = T2z - T2A;
728 ci[WS(rs, 6)] = T2z + T2A;
729 T2r = T2p + T2q;
730 ci[0] = T2r - T2y;
731 ci[WS(rs, 3)] = T2r + T2y;
732 }
733 {
734 E T35, T3d, T39, T3a, T38, T3b, T3e, T3c;
735 {
736 E T33, T34, T36, T37;
737 T33 = T1w - T1B;
738 T34 = T1H - T1M;
739 T35 = FMA(KP951056516, T33, KP587785252 * T34);
740 T3d = FNMS(KP587785252, T33, KP951056516 * T34);
741 T39 = T2T - T2Q;
742 T36 = T2v + T2w;
743 T37 = T2s + T2t;
744 T3a = T37 + T36;
745 T38 = KP559016994 * (T36 - T37);
746 T3b = FNMS(KP250000000, T3a, T39);
747 }
748 ci[WS(rs, 9)] = T3a + T39;
749 T3e = T38 + T3b;
750 cr[WS(rs, 8)] = T3d - T3e;
751 ci[WS(rs, 12)] = T3d + T3e;
752 T3c = T38 - T3b;
753 cr[WS(rs, 11)] = T35 + T3c;
754 cr[WS(rs, 14)] = T3c - T35;
755 }
756 {
757 E T2X, T31, T2U, T2P, T2Y, T2Z, T32, T30;
758 {
759 E T2V, T2W, T2N, T2O;
760 T2V = T1T - T1U;
761 T2W = T1Q - T1R;
762 T2X = FNMS(KP587785252, T2W, KP951056516 * T2V);
763 T31 = FMA(KP951056516, T2W, KP587785252 * T2V);
764 T2U = T2Q + T2T;
765 T2N = T2k - T2f;
766 T2O = T24 + T29;
767 T2P = T2N - T2O;
768 T2Y = FMA(KP250000000, T2P, T2U);
769 T2Z = KP559016994 * (T2O + T2N);
770 }
771 cr[WS(rs, 10)] = T2P - T2U;
772 T32 = T2Z + T2Y;
773 ci[WS(rs, 10)] = T31 + T32;
774 ci[WS(rs, 13)] = T32 - T31;
775 T30 = T2Y - T2Z;
776 cr[WS(rs, 13)] = T2X - T30;
777 ci[WS(rs, 7)] = T2X + T30;
778 }
779 {
780 E T2m, T2o, T1P, T1W, T1X, T1Y, T1Z, T2n;
781 {
782 E T2a, T2l, T1S, T1V;
783 T2a = T24 - T29;
784 T2l = T2f + T2k;
785 T2m = FMA(KP951056516, T2a, KP587785252 * T2l);
786 T2o = FNMS(KP587785252, T2a, KP951056516 * T2l);
787 T1P = T1n + T1q;
788 T1S = T1Q + T1R;
789 T1V = T1T + T1U;
790 T1W = T1S + T1V;
791 T1X = KP559016994 * (T1S - T1V);
792 T1Y = FNMS(KP250000000, T1W, T1P);
793 }
794 ci[WS(rs, 4)] = T1P + T1W;
795 T1Z = T1X + T1Y;
796 cr[WS(rs, 4)] = T1Z - T2m;
797 cr[WS(rs, 1)] = T1Z + T2m;
798 T2n = T1Y - T1X;
799 cr[WS(rs, 7)] = T2n - T2o;
800 ci[WS(rs, 1)] = T2n + T2o;
801 }
802 }
803 }
804 }
805
806 static const tw_instr twinstr[] = {
807 {TW_FULL, 1, 15},
808 {TW_NEXT, 1, 0}
809 };
810
811 static const hc2hc_desc desc = { 15, "hf_15", twinstr, &GENUS, {128, 56, 56, 0} };
812
813 void X(codelet_hf_15) (planner *p) {
814 X(khc2hc_register) (p, hf_15, &desc);
815 }
816 #endif