comparison fft/fftw/fftw-3.3.4/rdft/scalar/r2cb/hb_32.c @ 19:26056e866c29

Add FFTW to comparison table
author Chris Cannam
date Tue, 06 Oct 2015 13:08:39 +0100
parents
children
comparison
equal deleted inserted replaced
18:8db794ca3e0b 19:26056e866c29
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 Tue Mar 4 13:50:26 EST 2014 */
23
24 #include "codelet-rdft.h"
25
26 #ifdef HAVE_FMA
27
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -dif -name hb_32 -include hb.h */
29
30 /*
31 * This function contains 434 FP additions, 260 FP multiplications,
32 * (or, 236 additions, 62 multiplications, 198 fused multiply/add),
33 * 135 stack variables, 7 constants, and 128 memory accesses
34 */
35 #include "hb.h"
36
37 static void hb_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
38 {
39 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
40 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
41 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
42 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
43 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
44 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
45 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
46 {
47 INT m;
48 for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 62, MAKE_VOLATILE_STRIDE(64, rs)) {
49 E T5o, T5r, T5q, T5n, T5s, T5p;
50 {
51 E T5K, Tf, T8k, T7k, T8x, T7N, T3i, T1i, T3v, T2L, T5f, T4v, T6T, T6m, T52;
52 E T42, TZ, T6X, T1X, T3p, T8p, T8B, T3o, T26, T58, T4n, T7T, T7z, T59, T4k;
53 E T6p, T6a, TK, T6W, T2o, T3m, T8s, T8A, T3l, T2x, T55, T4g, T7S, T7G, T56;
54 E T4d, T6o, T61, T5Q, T5N, T6f, Tu, T8y, T7r, T8l, T7Q, T3w, T1F, T45, T48;
55 E T3j, T2O, T53, T4y;
56 {
57 E T62, T69, T4j, T4i;
58 {
59 E T6l, T6i, T40, T41;
60 {
61 E T12, T3, T2D, T6, T6g, T2G, T6h, T15, Td, T6k, T1g, T2J, Ta, T17, T1a;
62 E T6j;
63 {
64 E T2E, T2F, T13, T14;
65 {
66 E T1, T2, T4, T5;
67 T1 = cr[0];
68 T2 = ci[WS(rs, 15)];
69 T4 = cr[WS(rs, 8)];
70 T5 = ci[WS(rs, 7)];
71 T2E = ci[WS(rs, 31)];
72 T12 = T1 - T2;
73 T3 = T1 + T2;
74 T2D = T4 - T5;
75 T6 = T4 + T5;
76 T2F = cr[WS(rs, 16)];
77 }
78 T13 = ci[WS(rs, 23)];
79 T14 = cr[WS(rs, 24)];
80 {
81 E Tb, Tc, T1d, T1e;
82 Tb = ci[WS(rs, 3)];
83 T6g = T2E - T2F;
84 T2G = T2E + T2F;
85 T6h = T13 - T14;
86 T15 = T13 + T14;
87 Tc = cr[WS(rs, 12)];
88 T1d = ci[WS(rs, 19)];
89 T1e = cr[WS(rs, 28)];
90 {
91 E T8, T1c, T1f, T9, T18, T19;
92 T8 = cr[WS(rs, 4)];
93 Td = Tb + Tc;
94 T1c = Tb - Tc;
95 T6k = T1d - T1e;
96 T1f = T1d + T1e;
97 T9 = ci[WS(rs, 11)];
98 T18 = ci[WS(rs, 27)];
99 T19 = cr[WS(rs, 20)];
100 T1g = T1c - T1f;
101 T2J = T1c + T1f;
102 Ta = T8 + T9;
103 T17 = T8 - T9;
104 T1a = T18 + T19;
105 T6j = T18 - T19;
106 }
107 }
108 }
109 {
110 E T2I, T7M, T7L, T16, T1h, T4u, T4t, T2H, T2K;
111 {
112 E T7i, T7, T1b, Te, T7j;
113 T7i = T3 - T6;
114 T7 = T3 + T6;
115 T2I = T17 + T1a;
116 T1b = T17 - T1a;
117 Te = Ta + Td;
118 T7M = Ta - Td;
119 T7j = T6k - T6j;
120 T6l = T6j + T6k;
121 T6i = T6g + T6h;
122 T7L = T6g - T6h;
123 T5K = T7 - Te;
124 Tf = T7 + Te;
125 T8k = T7i + T7j;
126 T7k = T7i - T7j;
127 T40 = T12 + T15;
128 T16 = T12 - T15;
129 T1h = T1b + T1g;
130 T4u = T1b - T1g;
131 }
132 T4t = T2G - T2D;
133 T2H = T2D + T2G;
134 T8x = T7M + T7L;
135 T7N = T7L - T7M;
136 T3i = FMA(KP707106781, T1h, T16);
137 T1i = FNMS(KP707106781, T1h, T16);
138 T2K = T2I - T2J;
139 T41 = T2I + T2J;
140 T3v = FMA(KP707106781, T2K, T2H);
141 T2L = FNMS(KP707106781, T2K, T2H);
142 T5f = FNMS(KP707106781, T4u, T4t);
143 T4v = FMA(KP707106781, T4u, T4t);
144 }
145 }
146 {
147 E T1Y, T1H, TR, T7w, T1K, T21, T65, T7t, TV, T1M, TU, T67, T1U, TW, T1N;
148 E T1O;
149 {
150 E TL, TM, TO, TP, T63, T64;
151 TL = ci[0];
152 T6T = T6i + T6l;
153 T6m = T6i - T6l;
154 T52 = FMA(KP707106781, T41, T40);
155 T42 = FNMS(KP707106781, T41, T40);
156 TM = cr[WS(rs, 15)];
157 TO = cr[WS(rs, 7)];
158 TP = ci[WS(rs, 8)];
159 {
160 E T1I, TN, TQ, T1J, T1Z, T20;
161 T1I = ci[WS(rs, 16)];
162 T1Y = TL - TM;
163 TN = TL + TM;
164 T1H = TO - TP;
165 TQ = TO + TP;
166 T1J = cr[WS(rs, 31)];
167 T1Z = ci[WS(rs, 24)];
168 T20 = cr[WS(rs, 23)];
169 TR = TN + TQ;
170 T7w = TN - TQ;
171 T1K = T1I + T1J;
172 T63 = T1I - T1J;
173 T64 = T1Z - T20;
174 T21 = T1Z + T20;
175 }
176 {
177 E TS, TT, T1S, T1T;
178 TS = cr[WS(rs, 3)];
179 T65 = T63 + T64;
180 T7t = T63 - T64;
181 TT = ci[WS(rs, 12)];
182 T1S = ci[WS(rs, 20)];
183 T1T = cr[WS(rs, 27)];
184 TV = ci[WS(rs, 4)];
185 T1M = TS - TT;
186 TU = TS + TT;
187 T67 = T1S - T1T;
188 T1U = T1S + T1T;
189 TW = cr[WS(rs, 11)];
190 T1N = ci[WS(rs, 28)];
191 T1O = cr[WS(rs, 19)];
192 }
193 }
194 {
195 E T4l, T1L, T24, T23, T8n, T7v, T1W, T8o, T7y, T4m, T22, T25;
196 {
197 E T1V, T7u, T7x, T1Q, T1R, TX;
198 T4l = T1H + T1K;
199 T1L = T1H - T1K;
200 T1R = TV - TW;
201 TX = TV + TW;
202 {
203 E T66, T1P, TY, T68;
204 T66 = T1N - T1O;
205 T1P = T1N + T1O;
206 T24 = T1R - T1U;
207 T1V = T1R + T1U;
208 T7u = TU - TX;
209 TY = TU + TX;
210 T68 = T66 + T67;
211 T7x = T67 - T66;
212 T23 = T1M - T1P;
213 T1Q = T1M + T1P;
214 TZ = TR + TY;
215 T62 = TR - TY;
216 T69 = T65 - T68;
217 T6X = T65 + T68;
218 }
219 T8n = T7u + T7t;
220 T7v = T7t - T7u;
221 T4j = T1Q + T1V;
222 T1W = T1Q - T1V;
223 T8o = T7w + T7x;
224 T7y = T7w - T7x;
225 }
226 T4i = T1Y + T21;
227 T22 = T1Y - T21;
228 T25 = T23 + T24;
229 T4m = T23 - T24;
230 T1X = FNMS(KP707106781, T1W, T1L);
231 T3p = FMA(KP707106781, T1W, T1L);
232 T8p = FNMS(KP414213562, T8o, T8n);
233 T8B = FMA(KP414213562, T8n, T8o);
234 T3o = FMA(KP707106781, T25, T22);
235 T26 = FNMS(KP707106781, T25, T22);
236 T58 = FMA(KP707106781, T4m, T4l);
237 T4n = FNMS(KP707106781, T4m, T4l);
238 T7T = FNMS(KP414213562, T7v, T7y);
239 T7z = FMA(KP414213562, T7y, T7v);
240 }
241 }
242 }
243 {
244 E T5T, T60, T4c, T4b;
245 {
246 E T2p, T28, TC, T7D, T2b, T2s, T5W, T7A, TG, T2d, TF, T5Y, T2l, TH, T2e;
247 E T2f;
248 {
249 E Tw, Tx, Tz, TA, T5U, T5V;
250 Tw = cr[WS(rs, 1)];
251 T59 = FMA(KP707106781, T4j, T4i);
252 T4k = FNMS(KP707106781, T4j, T4i);
253 T6p = T69 - T62;
254 T6a = T62 + T69;
255 Tx = ci[WS(rs, 14)];
256 Tz = cr[WS(rs, 9)];
257 TA = ci[WS(rs, 6)];
258 {
259 E T29, Ty, TB, T2a, T2q, T2r;
260 T29 = ci[WS(rs, 30)];
261 T2p = Tw - Tx;
262 Ty = Tw + Tx;
263 T28 = Tz - TA;
264 TB = Tz + TA;
265 T2a = cr[WS(rs, 17)];
266 T2q = ci[WS(rs, 22)];
267 T2r = cr[WS(rs, 25)];
268 TC = Ty + TB;
269 T7D = Ty - TB;
270 T2b = T29 + T2a;
271 T5U = T29 - T2a;
272 T5V = T2q - T2r;
273 T2s = T2q + T2r;
274 }
275 {
276 E TD, TE, T2j, T2k;
277 TD = cr[WS(rs, 5)];
278 T5W = T5U + T5V;
279 T7A = T5U - T5V;
280 TE = ci[WS(rs, 10)];
281 T2j = ci[WS(rs, 18)];
282 T2k = cr[WS(rs, 29)];
283 TG = ci[WS(rs, 2)];
284 T2d = TD - TE;
285 TF = TD + TE;
286 T5Y = T2j - T2k;
287 T2l = T2j + T2k;
288 TH = cr[WS(rs, 13)];
289 T2e = ci[WS(rs, 26)];
290 T2f = cr[WS(rs, 21)];
291 }
292 }
293 {
294 E T4e, T2c, T2v, T2u, T8q, T7C, T2n, T8r, T7F, T4f, T2t, T2w;
295 {
296 E T2m, T7B, T7E, T2h, T2i, TI;
297 T4e = T2b - T28;
298 T2c = T28 + T2b;
299 T2i = TG - TH;
300 TI = TG + TH;
301 {
302 E T5X, T2g, TJ, T5Z;
303 T5X = T2e - T2f;
304 T2g = T2e + T2f;
305 T2v = T2i - T2l;
306 T2m = T2i + T2l;
307 T7B = TF - TI;
308 TJ = TF + TI;
309 T5Z = T5X + T5Y;
310 T7E = T5Y - T5X;
311 T2u = T2d - T2g;
312 T2h = T2d + T2g;
313 TK = TC + TJ;
314 T5T = TC - TJ;
315 T60 = T5W - T5Z;
316 T6W = T5W + T5Z;
317 }
318 T8q = T7B + T7A;
319 T7C = T7A - T7B;
320 T4c = T2h + T2m;
321 T2n = T2h - T2m;
322 T8r = T7D + T7E;
323 T7F = T7D - T7E;
324 }
325 T4b = T2p + T2s;
326 T2t = T2p - T2s;
327 T2w = T2u + T2v;
328 T4f = T2v - T2u;
329 T2o = FNMS(KP707106781, T2n, T2c);
330 T3m = FMA(KP707106781, T2n, T2c);
331 T8s = FMA(KP414213562, T8r, T8q);
332 T8A = FNMS(KP414213562, T8q, T8r);
333 T3l = FMA(KP707106781, T2w, T2t);
334 T2x = FNMS(KP707106781, T2w, T2t);
335 T55 = FMA(KP707106781, T4f, T4e);
336 T4g = FNMS(KP707106781, T4f, T4e);
337 T7S = FMA(KP414213562, T7C, T7F);
338 T7G = FNMS(KP414213562, T7F, T7C);
339 }
340 }
341 {
342 E T44, T1D, Tm, T7o, T7p, T43, T1y, T47, T1s, Tt, T7m, T7l, T46, T1n;
343 {
344 E Tj, T1z, Ti, T5P, T1C, Tk, T1v, T1w;
345 {
346 E Tg, Th, T1A, T1B;
347 Tg = cr[WS(rs, 2)];
348 T56 = FMA(KP707106781, T4c, T4b);
349 T4d = FNMS(KP707106781, T4c, T4b);
350 T6o = T5T + T60;
351 T61 = T5T - T60;
352 Th = ci[WS(rs, 13)];
353 T1A = ci[WS(rs, 21)];
354 T1B = cr[WS(rs, 26)];
355 Tj = cr[WS(rs, 10)];
356 T1z = Tg - Th;
357 Ti = Tg + Th;
358 T5P = T1A - T1B;
359 T1C = T1A + T1B;
360 Tk = ci[WS(rs, 5)];
361 T1v = ci[WS(rs, 29)];
362 T1w = cr[WS(rs, 18)];
363 }
364 {
365 E T1u, Tl, T5O, T1x;
366 T44 = T1z + T1C;
367 T1D = T1z - T1C;
368 T1u = Tj - Tk;
369 Tl = Tj + Tk;
370 T5O = T1v - T1w;
371 T1x = T1v + T1w;
372 Tm = Ti + Tl;
373 T7o = Ti - Tl;
374 T7p = T5O - T5P;
375 T5Q = T5O + T5P;
376 T43 = T1x - T1u;
377 T1y = T1u + T1x;
378 }
379 }
380 {
381 E Tq, T1o, Tp, T5M, T1r, Tr, T1k, T1l;
382 {
383 E Tn, To, T1p, T1q;
384 Tn = ci[WS(rs, 1)];
385 To = cr[WS(rs, 14)];
386 T1p = ci[WS(rs, 25)];
387 T1q = cr[WS(rs, 22)];
388 Tq = cr[WS(rs, 6)];
389 T1o = Tn - To;
390 Tp = Tn + To;
391 T5M = T1p - T1q;
392 T1r = T1p + T1q;
393 Tr = ci[WS(rs, 9)];
394 T1k = ci[WS(rs, 17)];
395 T1l = cr[WS(rs, 30)];
396 }
397 {
398 E T1j, Ts, T5L, T1m;
399 T47 = T1o + T1r;
400 T1s = T1o - T1r;
401 T1j = Tq - Tr;
402 Ts = Tq + Tr;
403 T5L = T1k - T1l;
404 T1m = T1k + T1l;
405 Tt = Tp + Ts;
406 T7m = Tp - Ts;
407 T7l = T5L - T5M;
408 T5N = T5L + T5M;
409 T46 = T1j + T1m;
410 T1n = T1j - T1m;
411 }
412 }
413 {
414 E T7P, T7O, T2N, T1t, T1E, T2M, T7n, T7q, T4w, T4x;
415 T7P = T7m + T7l;
416 T7n = T7l - T7m;
417 T7q = T7o + T7p;
418 T7O = T7o - T7p;
419 T6f = Tm - Tt;
420 Tu = Tm + Tt;
421 T8y = T7q + T7n;
422 T7r = T7n - T7q;
423 T2N = FMA(KP414213562, T1n, T1s);
424 T1t = FNMS(KP414213562, T1s, T1n);
425 T1E = FMA(KP414213562, T1D, T1y);
426 T2M = FNMS(KP414213562, T1y, T1D);
427 T8l = T7O + T7P;
428 T7Q = T7O - T7P;
429 T3w = T1E + T1t;
430 T1F = T1t - T1E;
431 T45 = FNMS(KP414213562, T44, T43);
432 T4w = FMA(KP414213562, T43, T44);
433 T4x = FMA(KP414213562, T46, T47);
434 T48 = FNMS(KP414213562, T47, T46);
435 T3j = T2M + T2N;
436 T2O = T2M - T2N;
437 T53 = T4w + T4x;
438 T4y = T4w - T4x;
439 }
440 }
441 }
442 }
443 {
444 E T72, T5g, T49, T78, T77, T73, T7s, T7U, T7R, T7H, T3f, T3e, T3d;
445 {
446 E T5R, T8m, T8C, T8z, T8t, T8e, T86, T88, T8h, T8f, T8i, T8c, T8g;
447 {
448 E T6P, T6Q, T6Z, T6S, T6R;
449 {
450 E Tv, T10, T6V, T6Y, T6U;
451 T72 = Tf - Tu;
452 Tv = Tf + Tu;
453 T6U = T5Q + T5N;
454 T5R = T5N - T5Q;
455 T5g = T48 - T45;
456 T49 = T45 + T48;
457 T10 = TK + TZ;
458 T78 = TK - TZ;
459 T77 = T6T - T6U;
460 T6V = T6T + T6U;
461 T6Y = T6W + T6X;
462 T73 = T6X - T6W;
463 T6P = W[30];
464 cr[0] = Tv + T10;
465 T6Q = Tv - T10;
466 ci[0] = T6V + T6Y;
467 T6Z = T6V - T6Y;
468 T6S = W[31];
469 T6R = T6P * T6Q;
470 }
471 {
472 E T8O, T8W, T8Q, T8Z, T8X, T90, T8U, T8Y;
473 {
474 E T8R, T8S, T8M, T8N, T70;
475 T8M = FMA(KP707106781, T8l, T8k);
476 T8m = FNMS(KP707106781, T8l, T8k);
477 T8C = T8A - T8B;
478 T8N = T8A + T8B;
479 T70 = T6S * T6Q;
480 cr[WS(rs, 16)] = FNMS(T6S, T6Z, T6R);
481 T8R = FMA(KP707106781, T8y, T8x);
482 T8z = FNMS(KP707106781, T8y, T8x);
483 T8O = FNMS(KP923879532, T8N, T8M);
484 T8W = FMA(KP923879532, T8N, T8M);
485 ci[WS(rs, 16)] = FMA(T6P, T6Z, T70);
486 T8S = T8s + T8p;
487 T8t = T8p - T8s;
488 {
489 E T8L, T8T, T8P, T8V;
490 T8L = W[34];
491 T8Q = W[35];
492 T8V = W[2];
493 T8Z = FMA(KP923879532, T8S, T8R);
494 T8T = FNMS(KP923879532, T8S, T8R);
495 T8P = T8L * T8O;
496 T8X = T8V * T8W;
497 T90 = T8V * T8Z;
498 T8U = T8L * T8T;
499 cr[WS(rs, 18)] = FNMS(T8Q, T8T, T8P);
500 T8Y = W[3];
501 }
502 }
503 {
504 E T89, T8a, T84, T85;
505 T84 = FNMS(KP707106781, T7r, T7k);
506 T7s = FMA(KP707106781, T7r, T7k);
507 ci[WS(rs, 18)] = FMA(T8Q, T8O, T8U);
508 T85 = T7S + T7T;
509 T7U = T7S - T7T;
510 ci[WS(rs, 2)] = FMA(T8Y, T8W, T90);
511 cr[WS(rs, 2)] = FNMS(T8Y, T8Z, T8X);
512 T7R = FMA(KP707106781, T7Q, T7N);
513 T89 = FNMS(KP707106781, T7Q, T7N);
514 T8e = FMA(KP923879532, T85, T84);
515 T86 = FNMS(KP923879532, T85, T84);
516 T8a = T7G + T7z;
517 T7H = T7z - T7G;
518 {
519 E T83, T8b, T87, T8d;
520 T83 = W[26];
521 T88 = W[27];
522 T8d = W[58];
523 T8h = FMA(KP923879532, T8a, T89);
524 T8b = FNMS(KP923879532, T8a, T89);
525 T87 = T83 * T86;
526 T8f = T8d * T8e;
527 T8i = T8d * T8h;
528 T8c = T83 * T8b;
529 cr[WS(rs, 14)] = FNMS(T88, T8b, T87);
530 T8g = W[59];
531 }
532 }
533 }
534 }
535 {
536 E T5S, T6q, T6n, T6K, T6C, T6b, T6E, T6N, T6L, T6O, T6I, T6M;
537 {
538 E T6F, T6G, T6A, T6B;
539 T6A = T5K - T5R;
540 T5S = T5K + T5R;
541 ci[WS(rs, 14)] = FMA(T88, T86, T8c);
542 T6B = T6p - T6o;
543 T6q = T6o + T6p;
544 ci[WS(rs, 30)] = FMA(T8g, T8e, T8i);
545 cr[WS(rs, 30)] = FNMS(T8g, T8h, T8f);
546 T6n = T6f + T6m;
547 T6F = T6m - T6f;
548 T6K = FMA(KP707106781, T6B, T6A);
549 T6C = FNMS(KP707106781, T6B, T6A);
550 T6G = T61 - T6a;
551 T6b = T61 + T6a;
552 {
553 E T6z, T6H, T6D, T6J;
554 T6z = W[54];
555 T6E = W[55];
556 T6J = W[22];
557 T6N = FMA(KP707106781, T6G, T6F);
558 T6H = FNMS(KP707106781, T6G, T6F);
559 T6D = T6z * T6C;
560 T6L = T6J * T6K;
561 T6O = T6J * T6N;
562 T6I = T6z * T6H;
563 cr[WS(rs, 28)] = FNMS(T6E, T6H, T6D);
564 T6M = W[23];
565 }
566 }
567 {
568 E T8G, T8F, T8J, T8H, T8I, T8u;
569 ci[WS(rs, 28)] = FMA(T6E, T6C, T6I);
570 ci[WS(rs, 12)] = FMA(T6M, T6K, T6O);
571 cr[WS(rs, 12)] = FNMS(T6M, T6N, T6L);
572 T8G = FMA(KP923879532, T8t, T8m);
573 T8u = FNMS(KP923879532, T8t, T8m);
574 {
575 E T8j, T8w, T8D, T8v, T8E;
576 T8j = W[50];
577 T8w = W[51];
578 T8F = W[18];
579 T8J = FMA(KP923879532, T8C, T8z);
580 T8D = FNMS(KP923879532, T8C, T8z);
581 T8v = T8j * T8u;
582 T8E = T8w * T8u;
583 T8H = T8F * T8G;
584 T8I = W[19];
585 cr[WS(rs, 26)] = FNMS(T8w, T8D, T8v);
586 ci[WS(rs, 26)] = FMA(T8j, T8D, T8E);
587 }
588 {
589 E T6c, T6u, T6x, T6r, T8K, T5J, T6e;
590 cr[WS(rs, 10)] = FNMS(T8I, T8J, T8H);
591 T8K = T8I * T8G;
592 ci[WS(rs, 10)] = FMA(T8F, T8J, T8K);
593 T6c = FNMS(KP707106781, T6b, T5S);
594 T6u = FMA(KP707106781, T6b, T5S);
595 T6x = FMA(KP707106781, T6q, T6n);
596 T6r = FNMS(KP707106781, T6q, T6n);
597 T5J = W[38];
598 T6e = W[39];
599 {
600 E T6t, T6w, T6d, T6s, T6v, T6y;
601 T6t = W[6];
602 T6w = W[7];
603 T6d = T5J * T6c;
604 T6s = T6e * T6c;
605 T6v = T6t * T6u;
606 T6y = T6w * T6u;
607 cr[WS(rs, 20)] = FNMS(T6e, T6r, T6d);
608 ci[WS(rs, 20)] = FMA(T5J, T6r, T6s);
609 cr[WS(rs, 4)] = FNMS(T6w, T6x, T6v);
610 ci[WS(rs, 4)] = FMA(T6t, T6x, T6y);
611 }
612 }
613 }
614 }
615 }
616 {
617 E T7c, T7f, T7e, T7g, T7d;
618 {
619 E T71, T74, T79, T76, T75, T7b, T7a;
620 T71 = W[46];
621 T7c = T72 + T73;
622 T74 = T72 - T73;
623 T7f = T78 + T77;
624 T79 = T77 - T78;
625 T76 = W[47];
626 T75 = T71 * T74;
627 T7b = W[14];
628 T7a = T71 * T79;
629 T7e = W[15];
630 cr[WS(rs, 24)] = FNMS(T76, T79, T75);
631 T7g = T7b * T7f;
632 T7d = T7b * T7c;
633 ci[WS(rs, 24)] = FMA(T76, T74, T7a);
634 }
635 {
636 E T81, T7X, T80, T7Z, T82;
637 ci[WS(rs, 8)] = FMA(T7e, T7c, T7g);
638 cr[WS(rs, 8)] = FNMS(T7e, T7f, T7d);
639 {
640 E T7h, T7Y, T7I, T7V, T7K, T7J, T7W;
641 T7h = W[42];
642 T7Y = FMA(KP923879532, T7H, T7s);
643 T7I = FNMS(KP923879532, T7H, T7s);
644 T81 = FMA(KP923879532, T7U, T7R);
645 T7V = FNMS(KP923879532, T7U, T7R);
646 T7K = W[43];
647 T7J = T7h * T7I;
648 T7X = W[10];
649 T80 = W[11];
650 T7W = T7K * T7I;
651 cr[WS(rs, 22)] = FNMS(T7K, T7V, T7J);
652 T7Z = T7X * T7Y;
653 T82 = T80 * T7Y;
654 ci[WS(rs, 22)] = FMA(T7h, T7V, T7W);
655 }
656 {
657 E T2P, T37, T1G, T32, T2R, T2Q, T38, T2z, T27, T2y;
658 T2P = FMA(KP923879532, T2O, T2L);
659 T37 = FNMS(KP923879532, T2O, T2L);
660 cr[WS(rs, 6)] = FNMS(T80, T81, T7Z);
661 ci[WS(rs, 6)] = FMA(T7X, T81, T82);
662 T1G = FMA(KP923879532, T1F, T1i);
663 T32 = FNMS(KP923879532, T1F, T1i);
664 T2R = FNMS(KP668178637, T1X, T26);
665 T27 = FMA(KP668178637, T26, T1X);
666 T2y = FNMS(KP668178637, T2x, T2o);
667 T2Q = FMA(KP668178637, T2o, T2x);
668 T38 = T2y + T27;
669 T2z = T27 - T2y;
670 {
671 E T2C, T2A, T3c, T34, T2U, T39, T36, T31;
672 {
673 E T11, T2W, T2S, T33;
674 T11 = W[40];
675 T2C = W[41];
676 T2A = FNMS(KP831469612, T2z, T1G);
677 T2W = FMA(KP831469612, T2z, T1G);
678 T2S = T2Q - T2R;
679 T33 = T2Q + T2R;
680 {
681 E T2V, T2B, T2T, T2Z, T2X, T2Y, T30;
682 T2V = W[8];
683 T2B = T11 * T2A;
684 T3c = FMA(KP831469612, T33, T32);
685 T34 = FNMS(KP831469612, T33, T32);
686 T2T = FNMS(KP831469612, T2S, T2P);
687 T2Z = FMA(KP831469612, T2S, T2P);
688 T2X = T2V * T2W;
689 T2Y = W[9];
690 T30 = T2V * T2Z;
691 cr[WS(rs, 21)] = FNMS(T2C, T2T, T2B);
692 T2U = T11 * T2T;
693 cr[WS(rs, 5)] = FNMS(T2Y, T2Z, T2X);
694 ci[WS(rs, 5)] = FMA(T2Y, T2W, T30);
695 }
696 }
697 T39 = FNMS(KP831469612, T38, T37);
698 T3f = FMA(KP831469612, T38, T37);
699 ci[WS(rs, 21)] = FMA(T2C, T2A, T2U);
700 T36 = W[25];
701 T31 = W[24];
702 {
703 E T3b, T3g, T3a, T35;
704 T3e = W[57];
705 T3a = T36 * T34;
706 T35 = T31 * T34;
707 T3b = W[56];
708 T3g = T3e * T3c;
709 ci[WS(rs, 13)] = FMA(T31, T39, T3a);
710 cr[WS(rs, 13)] = FNMS(T36, T39, T35);
711 T3d = T3b * T3c;
712 ci[WS(rs, 29)] = FMA(T3b, T3f, T3g);
713 }
714 }
715 }
716 }
717 }
718 {
719 E T4G, T4J, T4I, T4F, T4K;
720 {
721 E T4z, T4R, T4a, T4M, T4h, T4o, T4C, T4N, T4A, T4B;
722 T4z = FMA(KP923879532, T4y, T4v);
723 T4R = FNMS(KP923879532, T4y, T4v);
724 T4a = FNMS(KP923879532, T49, T42);
725 T4M = FMA(KP923879532, T49, T42);
726 cr[WS(rs, 29)] = FNMS(T3e, T3f, T3d);
727 T4h = FNMS(KP668178637, T4g, T4d);
728 T4A = FMA(KP668178637, T4d, T4g);
729 T4B = FMA(KP668178637, T4k, T4n);
730 T4o = FNMS(KP668178637, T4n, T4k);
731 T4C = T4A - T4B;
732 T4N = T4A + T4B;
733 {
734 E T4W, T4Z, T4q, T4X, T50, T4Y;
735 {
736 E T4L, T4Q, T4O, T4p, T4S, T4P, T4U, T4V, T4T;
737 T4L = W[20];
738 T4Q = W[21];
739 T4W = FMA(KP831469612, T4N, T4M);
740 T4O = FNMS(KP831469612, T4N, T4M);
741 T4p = T4h + T4o;
742 T4S = T4h - T4o;
743 T4P = T4L * T4O;
744 T4V = W[52];
745 T4Z = FNMS(KP831469612, T4S, T4R);
746 T4T = FMA(KP831469612, T4S, T4R);
747 T4q = FNMS(KP831469612, T4p, T4a);
748 T4G = FMA(KP831469612, T4p, T4a);
749 cr[WS(rs, 11)] = FNMS(T4Q, T4T, T4P);
750 T4U = T4L * T4T;
751 T4X = T4V * T4W;
752 T50 = T4V * T4Z;
753 T4Y = W[53];
754 ci[WS(rs, 11)] = FMA(T4Q, T4O, T4U);
755 }
756 {
757 E T4D, T4s, T3Z, T4E, T4r;
758 T4J = FMA(KP831469612, T4C, T4z);
759 T4D = FNMS(KP831469612, T4C, T4z);
760 T4s = W[37];
761 ci[WS(rs, 27)] = FMA(T4Y, T4W, T50);
762 cr[WS(rs, 27)] = FNMS(T4Y, T4Z, T4X);
763 T3Z = W[36];
764 T4E = T4s * T4q;
765 T4I = W[5];
766 T4r = T3Z * T4q;
767 ci[WS(rs, 19)] = FMA(T3Z, T4D, T4E);
768 T4F = W[4];
769 T4K = T4I * T4G;
770 cr[WS(rs, 19)] = FNMS(T4s, T4D, T4r);
771 }
772 }
773 }
774 {
775 E T3E, T3H, T3G, T3D, T3I;
776 {
777 E T3x, T3P, T3k, T3K, T3n, T3q, T3A, T3L, T4H, T3y, T3z;
778 T3x = FMA(KP923879532, T3w, T3v);
779 T3P = FNMS(KP923879532, T3w, T3v);
780 T4H = T4F * T4G;
781 ci[WS(rs, 3)] = FMA(T4F, T4J, T4K);
782 T3k = FMA(KP923879532, T3j, T3i);
783 T3K = FNMS(KP923879532, T3j, T3i);
784 T3y = FMA(KP198912367, T3l, T3m);
785 T3n = FNMS(KP198912367, T3m, T3l);
786 cr[WS(rs, 3)] = FNMS(T4I, T4J, T4H);
787 T3z = FNMS(KP198912367, T3o, T3p);
788 T3q = FMA(KP198912367, T3p, T3o);
789 T3A = T3y + T3z;
790 T3L = T3z - T3y;
791 {
792 E T3U, T3X, T3s, T3V, T3Y, T3W;
793 {
794 E T3J, T3O, T3M, T3r, T3Q, T3N, T3S, T3T, T3R;
795 T3J = W[48];
796 T3O = W[49];
797 T3U = FMA(KP980785280, T3L, T3K);
798 T3M = FNMS(KP980785280, T3L, T3K);
799 T3r = T3n + T3q;
800 T3Q = T3n - T3q;
801 T3N = T3J * T3M;
802 T3T = W[16];
803 T3X = FMA(KP980785280, T3Q, T3P);
804 T3R = FNMS(KP980785280, T3Q, T3P);
805 T3s = FNMS(KP980785280, T3r, T3k);
806 T3E = FMA(KP980785280, T3r, T3k);
807 cr[WS(rs, 25)] = FNMS(T3O, T3R, T3N);
808 T3S = T3J * T3R;
809 T3V = T3T * T3U;
810 T3Y = T3T * T3X;
811 T3W = W[17];
812 ci[WS(rs, 25)] = FMA(T3O, T3M, T3S);
813 }
814 {
815 E T3B, T3u, T3h, T3C, T3t;
816 T3H = FMA(KP980785280, T3A, T3x);
817 T3B = FNMS(KP980785280, T3A, T3x);
818 T3u = W[33];
819 ci[WS(rs, 9)] = FMA(T3W, T3U, T3Y);
820 cr[WS(rs, 9)] = FNMS(T3W, T3X, T3V);
821 T3h = W[32];
822 T3C = T3u * T3s;
823 T3G = W[1];
824 T3t = T3h * T3s;
825 ci[WS(rs, 17)] = FMA(T3h, T3B, T3C);
826 T3D = W[0];
827 T3I = T3G * T3E;
828 cr[WS(rs, 17)] = FNMS(T3u, T3B, T3t);
829 }
830 }
831 }
832 {
833 E T5h, T5z, T54, T5u, T57, T5a, T5k, T5v, T3F, T5i, T5j;
834 T5h = FMA(KP923879532, T5g, T5f);
835 T5z = FNMS(KP923879532, T5g, T5f);
836 T3F = T3D * T3E;
837 ci[WS(rs, 1)] = FMA(T3D, T3H, T3I);
838 T54 = FNMS(KP923879532, T53, T52);
839 T5u = FMA(KP923879532, T53, T52);
840 T5i = FMA(KP198912367, T55, T56);
841 T57 = FNMS(KP198912367, T56, T55);
842 cr[WS(rs, 1)] = FNMS(T3G, T3H, T3F);
843 T5j = FMA(KP198912367, T58, T59);
844 T5a = FNMS(KP198912367, T59, T58);
845 T5k = T5i - T5j;
846 T5v = T5i + T5j;
847 {
848 E T5E, T5H, T5c, T5F, T5I, T5G;
849 {
850 E T5t, T5y, T5w, T5b, T5A, T5x, T5C, T5D, T5B;
851 T5t = W[28];
852 T5y = W[29];
853 T5E = FMA(KP980785280, T5v, T5u);
854 T5w = FNMS(KP980785280, T5v, T5u);
855 T5b = T57 + T5a;
856 T5A = T5a - T57;
857 T5x = T5t * T5w;
858 T5D = W[60];
859 T5H = FNMS(KP980785280, T5A, T5z);
860 T5B = FMA(KP980785280, T5A, T5z);
861 T5c = FMA(KP980785280, T5b, T54);
862 T5o = FNMS(KP980785280, T5b, T54);
863 cr[WS(rs, 15)] = FNMS(T5y, T5B, T5x);
864 T5C = T5t * T5B;
865 T5F = T5D * T5E;
866 T5I = T5D * T5H;
867 T5G = W[61];
868 ci[WS(rs, 15)] = FMA(T5y, T5w, T5C);
869 }
870 {
871 E T5l, T5e, T51, T5m, T5d;
872 T5r = FMA(KP980785280, T5k, T5h);
873 T5l = FNMS(KP980785280, T5k, T5h);
874 T5e = W[45];
875 ci[WS(rs, 31)] = FMA(T5G, T5E, T5I);
876 cr[WS(rs, 31)] = FNMS(T5G, T5H, T5F);
877 T51 = W[44];
878 T5m = T5e * T5c;
879 T5q = W[13];
880 T5d = T51 * T5c;
881 ci[WS(rs, 23)] = FMA(T51, T5l, T5m);
882 T5n = W[12];
883 T5s = T5q * T5o;
884 cr[WS(rs, 23)] = FNMS(T5e, T5l, T5d);
885 }
886 }
887 }
888 }
889 }
890 }
891 }
892 T5p = T5n * T5o;
893 ci[WS(rs, 7)] = FMA(T5n, T5r, T5s);
894 cr[WS(rs, 7)] = FNMS(T5q, T5r, T5p);
895 }
896 }
897 }
898
899 static const tw_instr twinstr[] = {
900 {TW_FULL, 1, 32},
901 {TW_NEXT, 1, 0}
902 };
903
904 static const hc2hc_desc desc = { 32, "hb_32", twinstr, &GENUS, {236, 62, 198, 0} };
905
906 void X(codelet_hb_32) (planner *p) {
907 X(khc2hc_register) (p, hb_32, &desc);
908 }
909 #else /* HAVE_FMA */
910
911 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -dif -name hb_32 -include hb.h */
912
913 /*
914 * This function contains 434 FP additions, 208 FP multiplications,
915 * (or, 340 additions, 114 multiplications, 94 fused multiply/add),
916 * 98 stack variables, 7 constants, and 128 memory accesses
917 */
918 #include "hb.h"
919
920 static void hb_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
921 {
922 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
923 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
924 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
925 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
926 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
927 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
928 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
929 {
930 INT m;
931 for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 62, MAKE_VOLATILE_STRIDE(64, rs)) {
932 E T4o, T6y, T70, T5u, Tf, T12, T5x, T6z, T3m, T3Y, T29, T2y, T4v, T71, T2U;
933 E T3M, Tu, T1U, T6D, T73, T6G, T74, T1h, T2z, T2X, T3o, T4D, T5A, T4K, T5z;
934 E T30, T3n, TK, T1j, T6S, T7w, T6V, T7v, T1y, T2B, T3c, T3S, T4X, T61, T54;
935 E T62, T3f, T3T, TZ, T1A, T6L, T7z, T6O, T7y, T1P, T2C, T35, T3P, T5g, T64;
936 E T5n, T65, T38, T3Q;
937 {
938 E T3, T4m, T24, T4q, T27, T4t, T6, T5s, Ta, T4p, T1X, T5t, T20, T4n, Td;
939 E T4s;
940 {
941 E T1, T2, T22, T23;
942 T1 = cr[0];
943 T2 = ci[WS(rs, 15)];
944 T3 = T1 + T2;
945 T4m = T1 - T2;
946 T22 = ci[WS(rs, 27)];
947 T23 = cr[WS(rs, 20)];
948 T24 = T22 - T23;
949 T4q = T22 + T23;
950 }
951 {
952 E T25, T26, T4, T5;
953 T25 = ci[WS(rs, 19)];
954 T26 = cr[WS(rs, 28)];
955 T27 = T25 - T26;
956 T4t = T25 + T26;
957 T4 = cr[WS(rs, 8)];
958 T5 = ci[WS(rs, 7)];
959 T6 = T4 + T5;
960 T5s = T4 - T5;
961 }
962 {
963 E T8, T9, T1V, T1W;
964 T8 = cr[WS(rs, 4)];
965 T9 = ci[WS(rs, 11)];
966 Ta = T8 + T9;
967 T4p = T8 - T9;
968 T1V = ci[WS(rs, 31)];
969 T1W = cr[WS(rs, 16)];
970 T1X = T1V - T1W;
971 T5t = T1V + T1W;
972 }
973 {
974 E T1Y, T1Z, Tb, Tc;
975 T1Y = ci[WS(rs, 23)];
976 T1Z = cr[WS(rs, 24)];
977 T20 = T1Y - T1Z;
978 T4n = T1Y + T1Z;
979 Tb = ci[WS(rs, 3)];
980 Tc = cr[WS(rs, 12)];
981 Td = Tb + Tc;
982 T4s = Tb - Tc;
983 }
984 {
985 E T7, Te, T21, T28;
986 T4o = T4m - T4n;
987 T6y = T4m + T4n;
988 T70 = T5t - T5s;
989 T5u = T5s + T5t;
990 T7 = T3 + T6;
991 Te = Ta + Td;
992 Tf = T7 + Te;
993 T12 = T7 - Te;
994 {
995 E T5v, T5w, T3k, T3l;
996 T5v = T4p + T4q;
997 T5w = T4s + T4t;
998 T5x = KP707106781 * (T5v - T5w);
999 T6z = KP707106781 * (T5v + T5w);
1000 T3k = T1X - T20;
1001 T3l = Ta - Td;
1002 T3m = T3k - T3l;
1003 T3Y = T3l + T3k;
1004 }
1005 T21 = T1X + T20;
1006 T28 = T24 + T27;
1007 T29 = T21 - T28;
1008 T2y = T21 + T28;
1009 {
1010 E T4r, T4u, T2S, T2T;
1011 T4r = T4p - T4q;
1012 T4u = T4s - T4t;
1013 T4v = KP707106781 * (T4r + T4u);
1014 T71 = KP707106781 * (T4r - T4u);
1015 T2S = T3 - T6;
1016 T2T = T27 - T24;
1017 T2U = T2S - T2T;
1018 T3M = T2S + T2T;
1019 }
1020 }
1021 }
1022 {
1023 E Ti, T4H, T1c, T4F, T1f, T4I, Tl, T4E, Tp, T4A, T15, T4y, T18, T4B, Ts;
1024 E T4x;
1025 {
1026 E Tg, Th, T1a, T1b;
1027 Tg = cr[WS(rs, 2)];
1028 Th = ci[WS(rs, 13)];
1029 Ti = Tg + Th;
1030 T4H = Tg - Th;
1031 T1a = ci[WS(rs, 29)];
1032 T1b = cr[WS(rs, 18)];
1033 T1c = T1a - T1b;
1034 T4F = T1a + T1b;
1035 }
1036 {
1037 E T1d, T1e, Tj, Tk;
1038 T1d = ci[WS(rs, 21)];
1039 T1e = cr[WS(rs, 26)];
1040 T1f = T1d - T1e;
1041 T4I = T1d + T1e;
1042 Tj = cr[WS(rs, 10)];
1043 Tk = ci[WS(rs, 5)];
1044 Tl = Tj + Tk;
1045 T4E = Tj - Tk;
1046 }
1047 {
1048 E Tn, To, T13, T14;
1049 Tn = ci[WS(rs, 1)];
1050 To = cr[WS(rs, 14)];
1051 Tp = Tn + To;
1052 T4A = Tn - To;
1053 T13 = ci[WS(rs, 17)];
1054 T14 = cr[WS(rs, 30)];
1055 T15 = T13 - T14;
1056 T4y = T13 + T14;
1057 }
1058 {
1059 E T16, T17, Tq, Tr;
1060 T16 = ci[WS(rs, 25)];
1061 T17 = cr[WS(rs, 22)];
1062 T18 = T16 - T17;
1063 T4B = T16 + T17;
1064 Tq = cr[WS(rs, 6)];
1065 Tr = ci[WS(rs, 9)];
1066 Ts = Tq + Tr;
1067 T4x = Tq - Tr;
1068 }
1069 {
1070 E Tm, Tt, T6B, T6C;
1071 Tm = Ti + Tl;
1072 Tt = Tp + Ts;
1073 Tu = Tm + Tt;
1074 T1U = Tm - Tt;
1075 T6B = T4H + T4I;
1076 T6C = T4F - T4E;
1077 T6D = FNMS(KP923879532, T6C, KP382683432 * T6B);
1078 T73 = FMA(KP382683432, T6C, KP923879532 * T6B);
1079 }
1080 {
1081 E T6E, T6F, T19, T1g;
1082 T6E = T4A + T4B;
1083 T6F = T4x + T4y;
1084 T6G = FNMS(KP923879532, T6F, KP382683432 * T6E);
1085 T74 = FMA(KP382683432, T6F, KP923879532 * T6E);
1086 T19 = T15 + T18;
1087 T1g = T1c + T1f;
1088 T1h = T19 - T1g;
1089 T2z = T1g + T19;
1090 }
1091 {
1092 E T2V, T2W, T4z, T4C;
1093 T2V = T15 - T18;
1094 T2W = Tp - Ts;
1095 T2X = T2V - T2W;
1096 T3o = T2W + T2V;
1097 T4z = T4x - T4y;
1098 T4C = T4A - T4B;
1099 T4D = FNMS(KP382683432, T4C, KP923879532 * T4z);
1100 T5A = FMA(KP382683432, T4z, KP923879532 * T4C);
1101 }
1102 {
1103 E T4G, T4J, T2Y, T2Z;
1104 T4G = T4E + T4F;
1105 T4J = T4H - T4I;
1106 T4K = FMA(KP923879532, T4G, KP382683432 * T4J);
1107 T5z = FNMS(KP382683432, T4G, KP923879532 * T4J);
1108 T2Y = Ti - Tl;
1109 T2Z = T1c - T1f;
1110 T30 = T2Y + T2Z;
1111 T3n = T2Y - T2Z;
1112 }
1113 }
1114 {
1115 E Ty, T4N, TB, T4Y, T1p, T4O, T1m, T4Z, TI, T52, T1w, T4V, TF, T51, T1t;
1116 E T4S;
1117 {
1118 E Tw, Tx, T1k, T1l;
1119 Tw = cr[WS(rs, 1)];
1120 Tx = ci[WS(rs, 14)];
1121 Ty = Tw + Tx;
1122 T4N = Tw - Tx;
1123 {
1124 E Tz, TA, T1n, T1o;
1125 Tz = cr[WS(rs, 9)];
1126 TA = ci[WS(rs, 6)];
1127 TB = Tz + TA;
1128 T4Y = Tz - TA;
1129 T1n = ci[WS(rs, 22)];
1130 T1o = cr[WS(rs, 25)];
1131 T1p = T1n - T1o;
1132 T4O = T1n + T1o;
1133 }
1134 T1k = ci[WS(rs, 30)];
1135 T1l = cr[WS(rs, 17)];
1136 T1m = T1k - T1l;
1137 T4Z = T1k + T1l;
1138 {
1139 E TG, TH, T4T, T1u, T1v, T4U;
1140 TG = ci[WS(rs, 2)];
1141 TH = cr[WS(rs, 13)];
1142 T4T = TG - TH;
1143 T1u = ci[WS(rs, 18)];
1144 T1v = cr[WS(rs, 29)];
1145 T4U = T1u + T1v;
1146 TI = TG + TH;
1147 T52 = T4T + T4U;
1148 T1w = T1u - T1v;
1149 T4V = T4T - T4U;
1150 }
1151 {
1152 E TD, TE, T4Q, T1r, T1s, T4R;
1153 TD = cr[WS(rs, 5)];
1154 TE = ci[WS(rs, 10)];
1155 T4Q = TD - TE;
1156 T1r = ci[WS(rs, 26)];
1157 T1s = cr[WS(rs, 21)];
1158 T4R = T1r + T1s;
1159 TF = TD + TE;
1160 T51 = T4Q + T4R;
1161 T1t = T1r - T1s;
1162 T4S = T4Q - T4R;
1163 }
1164 }
1165 {
1166 E TC, TJ, T6Q, T6R;
1167 TC = Ty + TB;
1168 TJ = TF + TI;
1169 TK = TC + TJ;
1170 T1j = TC - TJ;
1171 T6Q = T4Z - T4Y;
1172 T6R = KP707106781 * (T4S - T4V);
1173 T6S = T6Q + T6R;
1174 T7w = T6Q - T6R;
1175 }
1176 {
1177 E T6T, T6U, T1q, T1x;
1178 T6T = T4N + T4O;
1179 T6U = KP707106781 * (T51 + T52);
1180 T6V = T6T - T6U;
1181 T7v = T6T + T6U;
1182 T1q = T1m + T1p;
1183 T1x = T1t + T1w;
1184 T1y = T1q - T1x;
1185 T2B = T1q + T1x;
1186 }
1187 {
1188 E T3a, T3b, T4P, T4W;
1189 T3a = T1m - T1p;
1190 T3b = TF - TI;
1191 T3c = T3a - T3b;
1192 T3S = T3b + T3a;
1193 T4P = T4N - T4O;
1194 T4W = KP707106781 * (T4S + T4V);
1195 T4X = T4P - T4W;
1196 T61 = T4P + T4W;
1197 }
1198 {
1199 E T50, T53, T3d, T3e;
1200 T50 = T4Y + T4Z;
1201 T53 = KP707106781 * (T51 - T52);
1202 T54 = T50 - T53;
1203 T62 = T50 + T53;
1204 T3d = Ty - TB;
1205 T3e = T1w - T1t;
1206 T3f = T3d - T3e;
1207 T3T = T3d + T3e;
1208 }
1209 }
1210 {
1211 E TN, T56, TQ, T5h, T1G, T57, T1D, T5i, TX, T5l, T1N, T5e, TU, T5k, T1K;
1212 E T5b;
1213 {
1214 E TL, TM, T1B, T1C;
1215 TL = ci[0];
1216 TM = cr[WS(rs, 15)];
1217 TN = TL + TM;
1218 T56 = TL - TM;
1219 {
1220 E TO, TP, T1E, T1F;
1221 TO = cr[WS(rs, 7)];
1222 TP = ci[WS(rs, 8)];
1223 TQ = TO + TP;
1224 T5h = TO - TP;
1225 T1E = ci[WS(rs, 24)];
1226 T1F = cr[WS(rs, 23)];
1227 T1G = T1E - T1F;
1228 T57 = T1E + T1F;
1229 }
1230 T1B = ci[WS(rs, 16)];
1231 T1C = cr[WS(rs, 31)];
1232 T1D = T1B - T1C;
1233 T5i = T1B + T1C;
1234 {
1235 E TV, TW, T5c, T1L, T1M, T5d;
1236 TV = ci[WS(rs, 4)];
1237 TW = cr[WS(rs, 11)];
1238 T5c = TV - TW;
1239 T1L = ci[WS(rs, 20)];
1240 T1M = cr[WS(rs, 27)];
1241 T5d = T1L + T1M;
1242 TX = TV + TW;
1243 T5l = T5c + T5d;
1244 T1N = T1L - T1M;
1245 T5e = T5c - T5d;
1246 }
1247 {
1248 E TS, TT, T59, T1I, T1J, T5a;
1249 TS = cr[WS(rs, 3)];
1250 TT = ci[WS(rs, 12)];
1251 T59 = TS - TT;
1252 T1I = ci[WS(rs, 28)];
1253 T1J = cr[WS(rs, 19)];
1254 T5a = T1I + T1J;
1255 TU = TS + TT;
1256 T5k = T59 + T5a;
1257 T1K = T1I - T1J;
1258 T5b = T59 - T5a;
1259 }
1260 }
1261 {
1262 E TR, TY, T6J, T6K;
1263 TR = TN + TQ;
1264 TY = TU + TX;
1265 TZ = TR + TY;
1266 T1A = TR - TY;
1267 T6J = KP707106781 * (T5b - T5e);
1268 T6K = T5h + T5i;
1269 T6L = T6J - T6K;
1270 T7z = T6K + T6J;
1271 }
1272 {
1273 E T6M, T6N, T1H, T1O;
1274 T6M = T56 + T57;
1275 T6N = KP707106781 * (T5k + T5l);
1276 T6O = T6M - T6N;
1277 T7y = T6M + T6N;
1278 T1H = T1D + T1G;
1279 T1O = T1K + T1N;
1280 T1P = T1H - T1O;
1281 T2C = T1H + T1O;
1282 }
1283 {
1284 E T33, T34, T58, T5f;
1285 T33 = T1D - T1G;
1286 T34 = TU - TX;
1287 T35 = T33 - T34;
1288 T3P = T34 + T33;
1289 T58 = T56 - T57;
1290 T5f = KP707106781 * (T5b + T5e);
1291 T5g = T58 - T5f;
1292 T64 = T58 + T5f;
1293 }
1294 {
1295 E T5j, T5m, T36, T37;
1296 T5j = T5h - T5i;
1297 T5m = KP707106781 * (T5k - T5l);
1298 T5n = T5j - T5m;
1299 T65 = T5j + T5m;
1300 T36 = TN - TQ;
1301 T37 = T1N - T1K;
1302 T38 = T36 - T37;
1303 T3Q = T36 + T37;
1304 }
1305 }
1306 {
1307 E Tv, T10, T2w, T2A, T2D, T2E, T2v, T2x;
1308 Tv = Tf + Tu;
1309 T10 = TK + TZ;
1310 T2w = Tv - T10;
1311 T2A = T2y + T2z;
1312 T2D = T2B + T2C;
1313 T2E = T2A - T2D;
1314 cr[0] = Tv + T10;
1315 ci[0] = T2A + T2D;
1316 T2v = W[30];
1317 T2x = W[31];
1318 cr[WS(rs, 16)] = FNMS(T2x, T2E, T2v * T2w);
1319 ci[WS(rs, 16)] = FMA(T2x, T2w, T2v * T2E);
1320 }
1321 {
1322 E T2I, T2O, T2M, T2Q;
1323 {
1324 E T2G, T2H, T2K, T2L;
1325 T2G = Tf - Tu;
1326 T2H = T2C - T2B;
1327 T2I = T2G - T2H;
1328 T2O = T2G + T2H;
1329 T2K = T2y - T2z;
1330 T2L = TK - TZ;
1331 T2M = T2K - T2L;
1332 T2Q = T2L + T2K;
1333 }
1334 {
1335 E T2F, T2J, T2N, T2P;
1336 T2F = W[46];
1337 T2J = W[47];
1338 cr[WS(rs, 24)] = FNMS(T2J, T2M, T2F * T2I);
1339 ci[WS(rs, 24)] = FMA(T2F, T2M, T2J * T2I);
1340 T2N = W[14];
1341 T2P = W[15];
1342 cr[WS(rs, 8)] = FNMS(T2P, T2Q, T2N * T2O);
1343 ci[WS(rs, 8)] = FMA(T2N, T2Q, T2P * T2O);
1344 }
1345 }
1346 {
1347 E T1i, T2a, T2o, T2k, T2d, T2l, T1R, T2p;
1348 T1i = T12 + T1h;
1349 T2a = T1U + T29;
1350 T2o = T29 - T1U;
1351 T2k = T12 - T1h;
1352 {
1353 E T2b, T2c, T1z, T1Q;
1354 T2b = T1j + T1y;
1355 T2c = T1P - T1A;
1356 T2d = KP707106781 * (T2b + T2c);
1357 T2l = KP707106781 * (T2c - T2b);
1358 T1z = T1j - T1y;
1359 T1Q = T1A + T1P;
1360 T1R = KP707106781 * (T1z + T1Q);
1361 T2p = KP707106781 * (T1z - T1Q);
1362 }
1363 {
1364 E T1S, T2e, T11, T1T;
1365 T1S = T1i - T1R;
1366 T2e = T2a - T2d;
1367 T11 = W[38];
1368 T1T = W[39];
1369 cr[WS(rs, 20)] = FNMS(T1T, T2e, T11 * T1S);
1370 ci[WS(rs, 20)] = FMA(T1T, T1S, T11 * T2e);
1371 }
1372 {
1373 E T2s, T2u, T2r, T2t;
1374 T2s = T2k + T2l;
1375 T2u = T2o + T2p;
1376 T2r = W[22];
1377 T2t = W[23];
1378 cr[WS(rs, 12)] = FNMS(T2t, T2u, T2r * T2s);
1379 ci[WS(rs, 12)] = FMA(T2r, T2u, T2t * T2s);
1380 }
1381 {
1382 E T2g, T2i, T2f, T2h;
1383 T2g = T1i + T1R;
1384 T2i = T2a + T2d;
1385 T2f = W[6];
1386 T2h = W[7];
1387 cr[WS(rs, 4)] = FNMS(T2h, T2i, T2f * T2g);
1388 ci[WS(rs, 4)] = FMA(T2h, T2g, T2f * T2i);
1389 }
1390 {
1391 E T2m, T2q, T2j, T2n;
1392 T2m = T2k - T2l;
1393 T2q = T2o - T2p;
1394 T2j = W[54];
1395 T2n = W[55];
1396 cr[WS(rs, 28)] = FNMS(T2n, T2q, T2j * T2m);
1397 ci[WS(rs, 28)] = FMA(T2j, T2q, T2n * T2m);
1398 }
1399 }
1400 {
1401 E T3O, T4a, T40, T4e, T3V, T4f, T43, T4b, T3N, T3Z;
1402 T3N = KP707106781 * (T3n + T3o);
1403 T3O = T3M - T3N;
1404 T4a = T3M + T3N;
1405 T3Z = KP707106781 * (T30 + T2X);
1406 T40 = T3Y - T3Z;
1407 T4e = T3Y + T3Z;
1408 {
1409 E T3R, T3U, T41, T42;
1410 T3R = FNMS(KP382683432, T3Q, KP923879532 * T3P);
1411 T3U = FMA(KP923879532, T3S, KP382683432 * T3T);
1412 T3V = T3R - T3U;
1413 T4f = T3U + T3R;
1414 T41 = FNMS(KP382683432, T3S, KP923879532 * T3T);
1415 T42 = FMA(KP382683432, T3P, KP923879532 * T3Q);
1416 T43 = T41 - T42;
1417 T4b = T41 + T42;
1418 }
1419 {
1420 E T3W, T44, T3L, T3X;
1421 T3W = T3O - T3V;
1422 T44 = T40 - T43;
1423 T3L = W[50];
1424 T3X = W[51];
1425 cr[WS(rs, 26)] = FNMS(T3X, T44, T3L * T3W);
1426 ci[WS(rs, 26)] = FMA(T3X, T3W, T3L * T44);
1427 }
1428 {
1429 E T4i, T4k, T4h, T4j;
1430 T4i = T4a + T4b;
1431 T4k = T4e + T4f;
1432 T4h = W[2];
1433 T4j = W[3];
1434 cr[WS(rs, 2)] = FNMS(T4j, T4k, T4h * T4i);
1435 ci[WS(rs, 2)] = FMA(T4h, T4k, T4j * T4i);
1436 }
1437 {
1438 E T46, T48, T45, T47;
1439 T46 = T3O + T3V;
1440 T48 = T40 + T43;
1441 T45 = W[18];
1442 T47 = W[19];
1443 cr[WS(rs, 10)] = FNMS(T47, T48, T45 * T46);
1444 ci[WS(rs, 10)] = FMA(T47, T46, T45 * T48);
1445 }
1446 {
1447 E T4c, T4g, T49, T4d;
1448 T4c = T4a - T4b;
1449 T4g = T4e - T4f;
1450 T49 = W[34];
1451 T4d = W[35];
1452 cr[WS(rs, 18)] = FNMS(T4d, T4g, T49 * T4c);
1453 ci[WS(rs, 18)] = FMA(T49, T4g, T4d * T4c);
1454 }
1455 }
1456 {
1457 E T32, T3A, T3q, T3E, T3h, T3F, T3t, T3B, T31, T3p;
1458 T31 = KP707106781 * (T2X - T30);
1459 T32 = T2U - T31;
1460 T3A = T2U + T31;
1461 T3p = KP707106781 * (T3n - T3o);
1462 T3q = T3m - T3p;
1463 T3E = T3m + T3p;
1464 {
1465 E T39, T3g, T3r, T3s;
1466 T39 = FNMS(KP923879532, T38, KP382683432 * T35);
1467 T3g = FMA(KP382683432, T3c, KP923879532 * T3f);
1468 T3h = T39 - T3g;
1469 T3F = T3g + T39;
1470 T3r = FNMS(KP923879532, T3c, KP382683432 * T3f);
1471 T3s = FMA(KP923879532, T35, KP382683432 * T38);
1472 T3t = T3r - T3s;
1473 T3B = T3r + T3s;
1474 }
1475 {
1476 E T3i, T3u, T2R, T3j;
1477 T3i = T32 - T3h;
1478 T3u = T3q - T3t;
1479 T2R = W[58];
1480 T3j = W[59];
1481 cr[WS(rs, 30)] = FNMS(T3j, T3u, T2R * T3i);
1482 ci[WS(rs, 30)] = FMA(T3j, T3i, T2R * T3u);
1483 }
1484 {
1485 E T3I, T3K, T3H, T3J;
1486 T3I = T3A + T3B;
1487 T3K = T3E + T3F;
1488 T3H = W[10];
1489 T3J = W[11];
1490 cr[WS(rs, 6)] = FNMS(T3J, T3K, T3H * T3I);
1491 ci[WS(rs, 6)] = FMA(T3H, T3K, T3J * T3I);
1492 }
1493 {
1494 E T3w, T3y, T3v, T3x;
1495 T3w = T32 + T3h;
1496 T3y = T3q + T3t;
1497 T3v = W[26];
1498 T3x = W[27];
1499 cr[WS(rs, 14)] = FNMS(T3x, T3y, T3v * T3w);
1500 ci[WS(rs, 14)] = FMA(T3x, T3w, T3v * T3y);
1501 }
1502 {
1503 E T3C, T3G, T3z, T3D;
1504 T3C = T3A - T3B;
1505 T3G = T3E - T3F;
1506 T3z = W[42];
1507 T3D = W[43];
1508 cr[WS(rs, 22)] = FNMS(T3D, T3G, T3z * T3C);
1509 ci[WS(rs, 22)] = FMA(T3z, T3G, T3D * T3C);
1510 }
1511 }
1512 {
1513 E T60, T6m, T6f, T6n, T67, T6r, T6c, T6q;
1514 {
1515 E T5Y, T5Z, T6d, T6e;
1516 T5Y = T4o + T4v;
1517 T5Z = T5z + T5A;
1518 T60 = T5Y + T5Z;
1519 T6m = T5Y - T5Z;
1520 T6d = FMA(KP195090322, T61, KP980785280 * T62);
1521 T6e = FNMS(KP195090322, T64, KP980785280 * T65);
1522 T6f = T6d + T6e;
1523 T6n = T6e - T6d;
1524 }
1525 {
1526 E T63, T66, T6a, T6b;
1527 T63 = FNMS(KP195090322, T62, KP980785280 * T61);
1528 T66 = FMA(KP980785280, T64, KP195090322 * T65);
1529 T67 = T63 + T66;
1530 T6r = T63 - T66;
1531 T6a = T5u + T5x;
1532 T6b = T4K + T4D;
1533 T6c = T6a + T6b;
1534 T6q = T6a - T6b;
1535 }
1536 {
1537 E T68, T6g, T5X, T69;
1538 T68 = T60 - T67;
1539 T6g = T6c - T6f;
1540 T5X = W[32];
1541 T69 = W[33];
1542 cr[WS(rs, 17)] = FNMS(T69, T6g, T5X * T68);
1543 ci[WS(rs, 17)] = FMA(T69, T68, T5X * T6g);
1544 }
1545 {
1546 E T6u, T6w, T6t, T6v;
1547 T6u = T6m + T6n;
1548 T6w = T6q + T6r;
1549 T6t = W[16];
1550 T6v = W[17];
1551 cr[WS(rs, 9)] = FNMS(T6v, T6w, T6t * T6u);
1552 ci[WS(rs, 9)] = FMA(T6t, T6w, T6v * T6u);
1553 }
1554 {
1555 E T6i, T6k, T6h, T6j;
1556 T6i = T60 + T67;
1557 T6k = T6c + T6f;
1558 T6h = W[0];
1559 T6j = W[1];
1560 cr[WS(rs, 1)] = FNMS(T6j, T6k, T6h * T6i);
1561 ci[WS(rs, 1)] = FMA(T6j, T6i, T6h * T6k);
1562 }
1563 {
1564 E T6o, T6s, T6l, T6p;
1565 T6o = T6m - T6n;
1566 T6s = T6q - T6r;
1567 T6l = W[48];
1568 T6p = W[49];
1569 cr[WS(rs, 25)] = FNMS(T6p, T6s, T6l * T6o);
1570 ci[WS(rs, 25)] = FMA(T6l, T6s, T6p * T6o);
1571 }
1572 }
1573 {
1574 E T7u, T7Q, T7J, T7R, T7B, T7V, T7G, T7U;
1575 {
1576 E T7s, T7t, T7H, T7I;
1577 T7s = T6y + T6z;
1578 T7t = T73 + T74;
1579 T7u = T7s - T7t;
1580 T7Q = T7s + T7t;
1581 T7H = FMA(KP195090322, T7w, KP980785280 * T7v);
1582 T7I = FMA(KP195090322, T7z, KP980785280 * T7y);
1583 T7J = T7H - T7I;
1584 T7R = T7H + T7I;
1585 }
1586 {
1587 E T7x, T7A, T7E, T7F;
1588 T7x = FNMS(KP980785280, T7w, KP195090322 * T7v);
1589 T7A = FNMS(KP980785280, T7z, KP195090322 * T7y);
1590 T7B = T7x + T7A;
1591 T7V = T7x - T7A;
1592 T7E = T70 - T71;
1593 T7F = T6D - T6G;
1594 T7G = T7E + T7F;
1595 T7U = T7E - T7F;
1596 }
1597 {
1598 E T7C, T7K, T7r, T7D;
1599 T7C = T7u - T7B;
1600 T7K = T7G - T7J;
1601 T7r = W[44];
1602 T7D = W[45];
1603 cr[WS(rs, 23)] = FNMS(T7D, T7K, T7r * T7C);
1604 ci[WS(rs, 23)] = FMA(T7D, T7C, T7r * T7K);
1605 }
1606 {
1607 E T7Y, T80, T7X, T7Z;
1608 T7Y = T7Q + T7R;
1609 T80 = T7U - T7V;
1610 T7X = W[60];
1611 T7Z = W[61];
1612 cr[WS(rs, 31)] = FNMS(T7Z, T80, T7X * T7Y);
1613 ci[WS(rs, 31)] = FMA(T7X, T80, T7Z * T7Y);
1614 }
1615 {
1616 E T7M, T7O, T7L, T7N;
1617 T7M = T7u + T7B;
1618 T7O = T7G + T7J;
1619 T7L = W[12];
1620 T7N = W[13];
1621 cr[WS(rs, 7)] = FNMS(T7N, T7O, T7L * T7M);
1622 ci[WS(rs, 7)] = FMA(T7N, T7M, T7L * T7O);
1623 }
1624 {
1625 E T7S, T7W, T7P, T7T;
1626 T7S = T7Q - T7R;
1627 T7W = T7U + T7V;
1628 T7P = W[28];
1629 T7T = W[29];
1630 cr[WS(rs, 15)] = FNMS(T7T, T7W, T7P * T7S);
1631 ci[WS(rs, 15)] = FMA(T7P, T7W, T7T * T7S);
1632 }
1633 }
1634 {
1635 E T4M, T5M, T5F, T5N, T5p, T5R, T5C, T5Q;
1636 {
1637 E T4w, T4L, T5D, T5E;
1638 T4w = T4o - T4v;
1639 T4L = T4D - T4K;
1640 T4M = T4w + T4L;
1641 T5M = T4w - T4L;
1642 T5D = FMA(KP831469612, T4X, KP555570233 * T54);
1643 T5E = FNMS(KP831469612, T5g, KP555570233 * T5n);
1644 T5F = T5D + T5E;
1645 T5N = T5E - T5D;
1646 }
1647 {
1648 E T55, T5o, T5y, T5B;
1649 T55 = FNMS(KP831469612, T54, KP555570233 * T4X);
1650 T5o = FMA(KP555570233, T5g, KP831469612 * T5n);
1651 T5p = T55 + T5o;
1652 T5R = T55 - T5o;
1653 T5y = T5u - T5x;
1654 T5B = T5z - T5A;
1655 T5C = T5y + T5B;
1656 T5Q = T5y - T5B;
1657 }
1658 {
1659 E T5q, T5G, T4l, T5r;
1660 T5q = T4M - T5p;
1661 T5G = T5C - T5F;
1662 T4l = W[40];
1663 T5r = W[41];
1664 cr[WS(rs, 21)] = FNMS(T5r, T5G, T4l * T5q);
1665 ci[WS(rs, 21)] = FMA(T5r, T5q, T4l * T5G);
1666 }
1667 {
1668 E T5U, T5W, T5T, T5V;
1669 T5U = T5M + T5N;
1670 T5W = T5Q + T5R;
1671 T5T = W[24];
1672 T5V = W[25];
1673 cr[WS(rs, 13)] = FNMS(T5V, T5W, T5T * T5U);
1674 ci[WS(rs, 13)] = FMA(T5T, T5W, T5V * T5U);
1675 }
1676 {
1677 E T5I, T5K, T5H, T5J;
1678 T5I = T4M + T5p;
1679 T5K = T5C + T5F;
1680 T5H = W[8];
1681 T5J = W[9];
1682 cr[WS(rs, 5)] = FNMS(T5J, T5K, T5H * T5I);
1683 ci[WS(rs, 5)] = FMA(T5J, T5I, T5H * T5K);
1684 }
1685 {
1686 E T5O, T5S, T5L, T5P;
1687 T5O = T5M - T5N;
1688 T5S = T5Q - T5R;
1689 T5L = W[56];
1690 T5P = W[57];
1691 cr[WS(rs, 29)] = FNMS(T5P, T5S, T5L * T5O);
1692 ci[WS(rs, 29)] = FMA(T5L, T5S, T5P * T5O);
1693 }
1694 }
1695 {
1696 E T6I, T7g, T79, T7h, T6X, T7l, T76, T7k;
1697 {
1698 E T6A, T6H, T77, T78;
1699 T6A = T6y - T6z;
1700 T6H = T6D + T6G;
1701 T6I = T6A - T6H;
1702 T7g = T6A + T6H;
1703 T77 = FNMS(KP555570233, T6S, KP831469612 * T6V);
1704 T78 = FMA(KP555570233, T6L, KP831469612 * T6O);
1705 T79 = T77 - T78;
1706 T7h = T77 + T78;
1707 }
1708 {
1709 E T6P, T6W, T72, T75;
1710 T6P = FNMS(KP555570233, T6O, KP831469612 * T6L);
1711 T6W = FMA(KP831469612, T6S, KP555570233 * T6V);
1712 T6X = T6P - T6W;
1713 T7l = T6W + T6P;
1714 T72 = T70 + T71;
1715 T75 = T73 - T74;
1716 T76 = T72 - T75;
1717 T7k = T72 + T75;
1718 }
1719 {
1720 E T6Y, T7a, T6x, T6Z;
1721 T6Y = T6I - T6X;
1722 T7a = T76 - T79;
1723 T6x = W[52];
1724 T6Z = W[53];
1725 cr[WS(rs, 27)] = FNMS(T6Z, T7a, T6x * T6Y);
1726 ci[WS(rs, 27)] = FMA(T6Z, T6Y, T6x * T7a);
1727 }
1728 {
1729 E T7o, T7q, T7n, T7p;
1730 T7o = T7g + T7h;
1731 T7q = T7k + T7l;
1732 T7n = W[4];
1733 T7p = W[5];
1734 cr[WS(rs, 3)] = FNMS(T7p, T7q, T7n * T7o);
1735 ci[WS(rs, 3)] = FMA(T7n, T7q, T7p * T7o);
1736 }
1737 {
1738 E T7c, T7e, T7b, T7d;
1739 T7c = T6I + T6X;
1740 T7e = T76 + T79;
1741 T7b = W[20];
1742 T7d = W[21];
1743 cr[WS(rs, 11)] = FNMS(T7d, T7e, T7b * T7c);
1744 ci[WS(rs, 11)] = FMA(T7d, T7c, T7b * T7e);
1745 }
1746 {
1747 E T7i, T7m, T7f, T7j;
1748 T7i = T7g - T7h;
1749 T7m = T7k - T7l;
1750 T7f = W[36];
1751 T7j = W[37];
1752 cr[WS(rs, 19)] = FNMS(T7j, T7m, T7f * T7i);
1753 ci[WS(rs, 19)] = FMA(T7f, T7m, T7j * T7i);
1754 }
1755 }
1756 }
1757 }
1758 }
1759
1760 static const tw_instr twinstr[] = {
1761 {TW_FULL, 1, 32},
1762 {TW_NEXT, 1, 0}
1763 };
1764
1765 static const hc2hc_desc desc = { 32, "hb_32", twinstr, &GENUS, {340, 114, 94, 0} };
1766
1767 void X(codelet_hb_32) (planner *p) {
1768 X(khc2hc_register) (p, hb_32, &desc);
1769 }
1770 #endif /* HAVE_FMA */