comparison src/fftw-3.3.3/rdft/scalar/r2cf/r2cf_32.c @ 10:37bf6b4a2645

Add FFTW3
author Chris Cannam
date Wed, 20 Mar 2013 15:35:50 +0000
parents
children
comparison
equal deleted inserted replaced
9:c0fb53affa76 10:37bf6b4a2645
1 /*
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 */
20
21 /* This file was automatically generated --- DO NOT EDIT */
22 /* Generated on Sun Nov 25 07:39:47 EST 2012 */
23
24 #include "codelet-rdft.h"
25
26 #ifdef HAVE_FMA
27
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 32 -name r2cf_32 -include r2cf.h */
29
30 /*
31 * This function contains 156 FP additions, 68 FP multiplications,
32 * (or, 88 additions, 0 multiplications, 68 fused multiply/add),
33 * 89 stack variables, 7 constants, and 64 memory accesses
34 */
35 #include "r2cf.h"
36
37 static void r2cf_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
38 {
39 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
40 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
41 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
42 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
43 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
44 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
45 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
46 {
47 INT i;
48 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
49 E T1x, T1M, T1I, T1E, T1J, T1H;
50 {
51 E Tv, T1h, T7, T2b, Te, T2n, Ty, T1i, T1l, TF, T2d, Tt, T1k, TC, T2c;
52 E Tm, T2j, T1Z, T2k, T22, TK, T1B, T19, T1C, T1e, TO, TV, T1T, TN, TP;
53 E T2g, T1S;
54 {
55 E TD, Tp, Tq, Tr;
56 {
57 E T1, T2, T4, T5;
58 T1 = R0[0];
59 T2 = R0[WS(rs, 8)];
60 T4 = R0[WS(rs, 4)];
61 T5 = R0[WS(rs, 12)];
62 {
63 E Ta, Tw, Tx, Td, Tn, To;
64 {
65 E T8, T3, T6, T9, Tb, Tc;
66 T8 = R0[WS(rs, 2)];
67 Tv = T1 - T2;
68 T3 = T1 + T2;
69 T1h = T4 - T5;
70 T6 = T4 + T5;
71 T9 = R0[WS(rs, 10)];
72 Tb = R0[WS(rs, 14)];
73 Tc = R0[WS(rs, 6)];
74 T7 = T3 + T6;
75 T2b = T3 - T6;
76 Ta = T8 + T9;
77 Tw = T8 - T9;
78 Tx = Tb - Tc;
79 Td = Tb + Tc;
80 }
81 Tn = R0[WS(rs, 15)];
82 To = R0[WS(rs, 7)];
83 Te = Ta + Td;
84 T2n = Td - Ta;
85 Ty = Tw + Tx;
86 T1i = Tx - Tw;
87 TD = Tn - To;
88 Tp = Tn + To;
89 Tq = R0[WS(rs, 3)];
90 Tr = R0[WS(rs, 11)];
91 }
92 }
93 {
94 E Tj, TA, Ti, Tk;
95 {
96 E Tg, Th, TE, Ts;
97 Tg = R0[WS(rs, 1)];
98 Th = R0[WS(rs, 9)];
99 Tj = R0[WS(rs, 5)];
100 TE = Tq - Tr;
101 Ts = Tq + Tr;
102 TA = Tg - Th;
103 Ti = Tg + Th;
104 T1l = FNMS(KP414213562, TD, TE);
105 TF = FMA(KP414213562, TE, TD);
106 T2d = Tp - Ts;
107 Tt = Tp + Ts;
108 Tk = R0[WS(rs, 13)];
109 }
110 {
111 E T11, T15, T1c, T20, T14, T16, T1X, T1Y, T1Q, T1R;
112 {
113 E T1a, T1b, T12, T13;
114 {
115 E TZ, T10, TB, Tl;
116 TZ = R1[WS(rs, 15)];
117 T10 = R1[WS(rs, 7)];
118 T1a = R1[WS(rs, 11)];
119 TB = Tj - Tk;
120 Tl = Tj + Tk;
121 T1X = TZ + T10;
122 T11 = TZ - T10;
123 T1k = FMA(KP414213562, TA, TB);
124 TC = FNMS(KP414213562, TB, TA);
125 T2c = Ti - Tl;
126 Tm = Ti + Tl;
127 T1b = R1[WS(rs, 3)];
128 }
129 T12 = R1[WS(rs, 1)];
130 T13 = R1[WS(rs, 9)];
131 T15 = R1[WS(rs, 13)];
132 T1Y = T1b + T1a;
133 T1c = T1a - T1b;
134 T20 = T12 + T13;
135 T14 = T12 - T13;
136 T16 = R1[WS(rs, 5)];
137 }
138 T2j = T1X - T1Y;
139 T1Z = T1X + T1Y;
140 {
141 E TT, TU, TL, TM;
142 {
143 E TI, T21, T17, TJ, T18, T1d;
144 TI = R1[0];
145 T21 = T15 + T16;
146 T17 = T15 - T16;
147 TJ = R1[WS(rs, 8)];
148 TT = R1[WS(rs, 4)];
149 T2k = T21 - T20;
150 T22 = T20 + T21;
151 T18 = T14 + T17;
152 T1d = T17 - T14;
153 T1Q = TI + TJ;
154 TK = TI - TJ;
155 T1B = FNMS(KP707106781, T18, T11);
156 T19 = FMA(KP707106781, T18, T11);
157 T1C = FNMS(KP707106781, T1d, T1c);
158 T1e = FMA(KP707106781, T1d, T1c);
159 TU = R1[WS(rs, 12)];
160 }
161 TL = R1[WS(rs, 2)];
162 TM = R1[WS(rs, 10)];
163 TO = R1[WS(rs, 14)];
164 T1R = TT + TU;
165 TV = TT - TU;
166 T1T = TL + TM;
167 TN = TL - TM;
168 TP = R1[WS(rs, 6)];
169 }
170 T2g = T1Q - T1R;
171 T1S = T1Q + T1R;
172 }
173 }
174 }
175 {
176 E T1P, T25, T23, T2h, T1W, T1y, TS, T1z, TX, T27, T2a;
177 {
178 E Tf, Tu, T29, T28;
179 {
180 E T1U, TQ, T1V, TR, TW;
181 T1P = T7 - Te;
182 Tf = T7 + Te;
183 T1U = TO + TP;
184 TQ = TO - TP;
185 Tu = Tm + Tt;
186 T25 = Tt - Tm;
187 T23 = T1Z - T22;
188 T29 = T1Z + T22;
189 T2h = T1U - T1T;
190 T1V = T1T + T1U;
191 TR = TN + TQ;
192 TW = TN - TQ;
193 T27 = Tf + Tu;
194 T1W = T1S - T1V;
195 T28 = T1S + T1V;
196 T1y = FNMS(KP707106781, TR, TK);
197 TS = FMA(KP707106781, TR, TK);
198 T1z = FNMS(KP707106781, TW, TV);
199 TX = FMA(KP707106781, TW, TV);
200 T2a = T28 + T29;
201 }
202 Cr[WS(csr, 8)] = Tf - Tu;
203 Ci[WS(csi, 8)] = T29 - T28;
204 }
205 Cr[0] = T27 + T2a;
206 Cr[WS(csr, 16)] = T27 - T2a;
207 {
208 E T2s, T2i, T2v, T2f, T2r, T2p, T2l, T2t;
209 {
210 E T2o, T2e, T26, T24;
211 T2o = T2d - T2c;
212 T2e = T2c + T2d;
213 T2s = FNMS(KP414213562, T2g, T2h);
214 T2i = FMA(KP414213562, T2h, T2g);
215 T26 = T23 - T1W;
216 T24 = T1W + T23;
217 T2v = FNMS(KP707106781, T2e, T2b);
218 T2f = FMA(KP707106781, T2e, T2b);
219 T2r = FMA(KP707106781, T2o, T2n);
220 T2p = FNMS(KP707106781, T2o, T2n);
221 Ci[WS(csi, 4)] = FMA(KP707106781, T26, T25);
222 Ci[WS(csi, 12)] = FMS(KP707106781, T26, T25);
223 Cr[WS(csr, 4)] = FMA(KP707106781, T24, T1P);
224 Cr[WS(csr, 12)] = FNMS(KP707106781, T24, T1P);
225 T2l = FNMS(KP414213562, T2k, T2j);
226 T2t = FMA(KP414213562, T2j, T2k);
227 }
228 {
229 E T1v, T1G, TH, T1s, T1F, T1w, T1o, T1g, T1p, T1n;
230 {
231 E T1f, TY, T1t, T1u, T1j, T1m;
232 {
233 E Tz, TG, T1q, T1r;
234 T1v = FNMS(KP707106781, Ty, Tv);
235 Tz = FMA(KP707106781, Ty, Tv);
236 {
237 E T2q, T2m, T2w, T2u;
238 T2q = T2l - T2i;
239 T2m = T2i + T2l;
240 T2w = T2t - T2s;
241 T2u = T2s + T2t;
242 Ci[WS(csi, 10)] = FMA(KP923879532, T2q, T2p);
243 Ci[WS(csi, 6)] = FMS(KP923879532, T2q, T2p);
244 Cr[WS(csr, 2)] = FMA(KP923879532, T2m, T2f);
245 Cr[WS(csr, 14)] = FNMS(KP923879532, T2m, T2f);
246 Cr[WS(csr, 10)] = FNMS(KP923879532, T2w, T2v);
247 Cr[WS(csr, 6)] = FMA(KP923879532, T2w, T2v);
248 Ci[WS(csi, 2)] = FMA(KP923879532, T2u, T2r);
249 Ci[WS(csi, 14)] = FMS(KP923879532, T2u, T2r);
250 TG = TC + TF;
251 T1G = TF - TC;
252 }
253 T1f = FNMS(KP198912367, T1e, T19);
254 T1q = FMA(KP198912367, T19, T1e);
255 T1r = FMA(KP198912367, TS, TX);
256 TY = FNMS(KP198912367, TX, TS);
257 T1t = FNMS(KP923879532, TG, Tz);
258 TH = FMA(KP923879532, TG, Tz);
259 T1u = T1r + T1q;
260 T1s = T1q - T1r;
261 T1F = FMA(KP707106781, T1i, T1h);
262 T1j = FNMS(KP707106781, T1i, T1h);
263 T1m = T1k + T1l;
264 T1w = T1k - T1l;
265 }
266 Cr[WS(csr, 7)] = FMA(KP980785280, T1u, T1t);
267 T1o = T1f - TY;
268 T1g = TY + T1f;
269 T1p = FMA(KP923879532, T1m, T1j);
270 T1n = FNMS(KP923879532, T1m, T1j);
271 Cr[WS(csr, 9)] = FNMS(KP980785280, T1u, T1t);
272 }
273 Cr[WS(csr, 1)] = FMA(KP980785280, T1g, TH);
274 Cr[WS(csr, 15)] = FNMS(KP980785280, T1g, TH);
275 Ci[WS(csi, 1)] = FMS(KP980785280, T1s, T1p);
276 Ci[WS(csi, 15)] = FMA(KP980785280, T1s, T1p);
277 Ci[WS(csi, 9)] = FMS(KP980785280, T1o, T1n);
278 Ci[WS(csi, 7)] = FMA(KP980785280, T1o, T1n);
279 {
280 E T1A, T1D, T1N, T1O, T1K, T1L;
281 T1A = FMA(KP668178637, T1z, T1y);
282 T1K = FNMS(KP668178637, T1y, T1z);
283 T1L = FNMS(KP668178637, T1B, T1C);
284 T1D = FMA(KP668178637, T1C, T1B);
285 T1N = FNMS(KP923879532, T1w, T1v);
286 T1x = FMA(KP923879532, T1w, T1v);
287 T1O = T1K + T1L;
288 T1M = T1K - T1L;
289 Cr[WS(csr, 5)] = FNMS(KP831469612, T1O, T1N);
290 T1I = T1D - T1A;
291 T1E = T1A + T1D;
292 T1J = FMA(KP923879532, T1G, T1F);
293 T1H = FNMS(KP923879532, T1G, T1F);
294 Cr[WS(csr, 11)] = FMA(KP831469612, T1O, T1N);
295 }
296 }
297 }
298 }
299 }
300 Ci[WS(csi, 3)] = FMA(KP831469612, T1M, T1J);
301 Cr[WS(csr, 3)] = FMA(KP831469612, T1E, T1x);
302 Ci[WS(csi, 13)] = FMS(KP831469612, T1M, T1J);
303 Cr[WS(csr, 13)] = FNMS(KP831469612, T1E, T1x);
304 Ci[WS(csi, 11)] = FMA(KP831469612, T1I, T1H);
305 Ci[WS(csi, 5)] = FMS(KP831469612, T1I, T1H);
306 }
307 }
308 }
309
310 static const kr2c_desc desc = { 32, "r2cf_32", {88, 0, 68, 0}, &GENUS };
311
312 void X(codelet_r2cf_32) (planner *p) {
313 X(kr2c_register) (p, r2cf_32, &desc);
314 }
315
316 #else /* HAVE_FMA */
317
318 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 32 -name r2cf_32 -include r2cf.h */
319
320 /*
321 * This function contains 156 FP additions, 42 FP multiplications,
322 * (or, 140 additions, 26 multiplications, 16 fused multiply/add),
323 * 54 stack variables, 7 constants, and 64 memory accesses
324 */
325 #include "r2cf.h"
326
327 static void r2cf_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
328 {
329 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
330 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
331 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
332 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
333 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
334 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
335 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
336 {
337 INT i;
338 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
339 E T7, T2b, Tv, T1l, Te, T2o, Ty, T1k, Tt, T2d, TF, T1h, Tm, T2c, TC;
340 E T1i, T1Z, T22, T2k, T2j, T1e, T1C, T19, T1B, T1S, T1V, T2h, T2g, TX, T1z;
341 E TS, T1y;
342 {
343 E T1, T2, T3, T4, T5, T6;
344 T1 = R0[0];
345 T2 = R0[WS(rs, 8)];
346 T3 = T1 + T2;
347 T4 = R0[WS(rs, 4)];
348 T5 = R0[WS(rs, 12)];
349 T6 = T4 + T5;
350 T7 = T3 + T6;
351 T2b = T3 - T6;
352 Tv = T1 - T2;
353 T1l = T4 - T5;
354 }
355 {
356 E Ta, Tw, Td, Tx;
357 {
358 E T8, T9, Tb, Tc;
359 T8 = R0[WS(rs, 2)];
360 T9 = R0[WS(rs, 10)];
361 Ta = T8 + T9;
362 Tw = T8 - T9;
363 Tb = R0[WS(rs, 14)];
364 Tc = R0[WS(rs, 6)];
365 Td = Tb + Tc;
366 Tx = Tb - Tc;
367 }
368 Te = Ta + Td;
369 T2o = Td - Ta;
370 Ty = KP707106781 * (Tw + Tx);
371 T1k = KP707106781 * (Tx - Tw);
372 }
373 {
374 E Tp, TD, Ts, TE;
375 {
376 E Tn, To, Tq, Tr;
377 Tn = R0[WS(rs, 15)];
378 To = R0[WS(rs, 7)];
379 Tp = Tn + To;
380 TD = Tn - To;
381 Tq = R0[WS(rs, 3)];
382 Tr = R0[WS(rs, 11)];
383 Ts = Tq + Tr;
384 TE = Tq - Tr;
385 }
386 Tt = Tp + Ts;
387 T2d = Tp - Ts;
388 TF = FMA(KP923879532, TD, KP382683432 * TE);
389 T1h = FNMS(KP923879532, TE, KP382683432 * TD);
390 }
391 {
392 E Ti, TA, Tl, TB;
393 {
394 E Tg, Th, Tj, Tk;
395 Tg = R0[WS(rs, 1)];
396 Th = R0[WS(rs, 9)];
397 Ti = Tg + Th;
398 TA = Tg - Th;
399 Tj = R0[WS(rs, 5)];
400 Tk = R0[WS(rs, 13)];
401 Tl = Tj + Tk;
402 TB = Tj - Tk;
403 }
404 Tm = Ti + Tl;
405 T2c = Ti - Tl;
406 TC = FNMS(KP382683432, TB, KP923879532 * TA);
407 T1i = FMA(KP382683432, TA, KP923879532 * TB);
408 }
409 {
410 E T11, T1X, T1d, T1Y, T14, T20, T17, T21, T1a, T18;
411 {
412 E TZ, T10, T1b, T1c;
413 TZ = R1[WS(rs, 15)];
414 T10 = R1[WS(rs, 7)];
415 T11 = TZ - T10;
416 T1X = TZ + T10;
417 T1b = R1[WS(rs, 3)];
418 T1c = R1[WS(rs, 11)];
419 T1d = T1b - T1c;
420 T1Y = T1b + T1c;
421 }
422 {
423 E T12, T13, T15, T16;
424 T12 = R1[WS(rs, 1)];
425 T13 = R1[WS(rs, 9)];
426 T14 = T12 - T13;
427 T20 = T12 + T13;
428 T15 = R1[WS(rs, 13)];
429 T16 = R1[WS(rs, 5)];
430 T17 = T15 - T16;
431 T21 = T15 + T16;
432 }
433 T1Z = T1X + T1Y;
434 T22 = T20 + T21;
435 T2k = T21 - T20;
436 T2j = T1X - T1Y;
437 T1a = KP707106781 * (T17 - T14);
438 T1e = T1a - T1d;
439 T1C = T1d + T1a;
440 T18 = KP707106781 * (T14 + T17);
441 T19 = T11 + T18;
442 T1B = T11 - T18;
443 }
444 {
445 E TK, T1Q, TW, T1R, TN, T1T, TQ, T1U, TT, TR;
446 {
447 E TI, TJ, TU, TV;
448 TI = R1[0];
449 TJ = R1[WS(rs, 8)];
450 TK = TI - TJ;
451 T1Q = TI + TJ;
452 TU = R1[WS(rs, 4)];
453 TV = R1[WS(rs, 12)];
454 TW = TU - TV;
455 T1R = TU + TV;
456 }
457 {
458 E TL, TM, TO, TP;
459 TL = R1[WS(rs, 2)];
460 TM = R1[WS(rs, 10)];
461 TN = TL - TM;
462 T1T = TL + TM;
463 TO = R1[WS(rs, 14)];
464 TP = R1[WS(rs, 6)];
465 TQ = TO - TP;
466 T1U = TO + TP;
467 }
468 T1S = T1Q + T1R;
469 T1V = T1T + T1U;
470 T2h = T1U - T1T;
471 T2g = T1Q - T1R;
472 TT = KP707106781 * (TQ - TN);
473 TX = TT - TW;
474 T1z = TW + TT;
475 TR = KP707106781 * (TN + TQ);
476 TS = TK + TR;
477 T1y = TK - TR;
478 }
479 {
480 E Tf, Tu, T27, T28, T29, T2a;
481 Tf = T7 + Te;
482 Tu = Tm + Tt;
483 T27 = Tf + Tu;
484 T28 = T1S + T1V;
485 T29 = T1Z + T22;
486 T2a = T28 + T29;
487 Cr[WS(csr, 8)] = Tf - Tu;
488 Ci[WS(csi, 8)] = T29 - T28;
489 Cr[WS(csr, 16)] = T27 - T2a;
490 Cr[0] = T27 + T2a;
491 }
492 {
493 E T1P, T25, T24, T26, T1W, T23;
494 T1P = T7 - Te;
495 T25 = Tt - Tm;
496 T1W = T1S - T1V;
497 T23 = T1Z - T22;
498 T24 = KP707106781 * (T1W + T23);
499 T26 = KP707106781 * (T23 - T1W);
500 Cr[WS(csr, 12)] = T1P - T24;
501 Ci[WS(csi, 12)] = T26 - T25;
502 Cr[WS(csr, 4)] = T1P + T24;
503 Ci[WS(csi, 4)] = T25 + T26;
504 }
505 {
506 E T2f, T2v, T2p, T2r, T2m, T2q, T2u, T2w, T2e, T2n;
507 T2e = KP707106781 * (T2c + T2d);
508 T2f = T2b + T2e;
509 T2v = T2b - T2e;
510 T2n = KP707106781 * (T2d - T2c);
511 T2p = T2n - T2o;
512 T2r = T2o + T2n;
513 {
514 E T2i, T2l, T2s, T2t;
515 T2i = FMA(KP923879532, T2g, KP382683432 * T2h);
516 T2l = FNMS(KP382683432, T2k, KP923879532 * T2j);
517 T2m = T2i + T2l;
518 T2q = T2l - T2i;
519 T2s = FNMS(KP382683432, T2g, KP923879532 * T2h);
520 T2t = FMA(KP382683432, T2j, KP923879532 * T2k);
521 T2u = T2s + T2t;
522 T2w = T2t - T2s;
523 }
524 Cr[WS(csr, 14)] = T2f - T2m;
525 Ci[WS(csi, 14)] = T2u - T2r;
526 Cr[WS(csr, 2)] = T2f + T2m;
527 Ci[WS(csi, 2)] = T2r + T2u;
528 Ci[WS(csi, 6)] = T2p + T2q;
529 Cr[WS(csr, 6)] = T2v + T2w;
530 Ci[WS(csi, 10)] = T2q - T2p;
531 Cr[WS(csr, 10)] = T2v - T2w;
532 }
533 {
534 E TH, T1t, T1s, T1u, T1g, T1o, T1n, T1p;
535 {
536 E Tz, TG, T1q, T1r;
537 Tz = Tv + Ty;
538 TG = TC + TF;
539 TH = Tz + TG;
540 T1t = Tz - TG;
541 T1q = FNMS(KP195090322, TS, KP980785280 * TX);
542 T1r = FMA(KP195090322, T19, KP980785280 * T1e);
543 T1s = T1q + T1r;
544 T1u = T1r - T1q;
545 }
546 {
547 E TY, T1f, T1j, T1m;
548 TY = FMA(KP980785280, TS, KP195090322 * TX);
549 T1f = FNMS(KP195090322, T1e, KP980785280 * T19);
550 T1g = TY + T1f;
551 T1o = T1f - TY;
552 T1j = T1h - T1i;
553 T1m = T1k - T1l;
554 T1n = T1j - T1m;
555 T1p = T1m + T1j;
556 }
557 Cr[WS(csr, 15)] = TH - T1g;
558 Ci[WS(csi, 15)] = T1s - T1p;
559 Cr[WS(csr, 1)] = TH + T1g;
560 Ci[WS(csi, 1)] = T1p + T1s;
561 Ci[WS(csi, 7)] = T1n + T1o;
562 Cr[WS(csr, 7)] = T1t + T1u;
563 Ci[WS(csi, 9)] = T1o - T1n;
564 Cr[WS(csr, 9)] = T1t - T1u;
565 }
566 {
567 E T1x, T1N, T1M, T1O, T1E, T1I, T1H, T1J;
568 {
569 E T1v, T1w, T1K, T1L;
570 T1v = Tv - Ty;
571 T1w = T1i + T1h;
572 T1x = T1v + T1w;
573 T1N = T1v - T1w;
574 T1K = FNMS(KP555570233, T1y, KP831469612 * T1z);
575 T1L = FMA(KP555570233, T1B, KP831469612 * T1C);
576 T1M = T1K + T1L;
577 T1O = T1L - T1K;
578 }
579 {
580 E T1A, T1D, T1F, T1G;
581 T1A = FMA(KP831469612, T1y, KP555570233 * T1z);
582 T1D = FNMS(KP555570233, T1C, KP831469612 * T1B);
583 T1E = T1A + T1D;
584 T1I = T1D - T1A;
585 T1F = TF - TC;
586 T1G = T1l + T1k;
587 T1H = T1F - T1G;
588 T1J = T1G + T1F;
589 }
590 Cr[WS(csr, 13)] = T1x - T1E;
591 Ci[WS(csi, 13)] = T1M - T1J;
592 Cr[WS(csr, 3)] = T1x + T1E;
593 Ci[WS(csi, 3)] = T1J + T1M;
594 Ci[WS(csi, 5)] = T1H + T1I;
595 Cr[WS(csr, 5)] = T1N + T1O;
596 Ci[WS(csi, 11)] = T1I - T1H;
597 Cr[WS(csr, 11)] = T1N - T1O;
598 }
599 }
600 }
601 }
602
603 static const kr2c_desc desc = { 32, "r2cf_32", {140, 26, 16, 0}, &GENUS };
604
605 void X(codelet_r2cf_32) (planner *p) {
606 X(kr2c_register) (p, r2cf_32, &desc);
607 }
608
609 #endif /* HAVE_FMA */