Coverage Report

Created: 2026-04-01 06:08

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/src/fftw3/dft/scalar/codelets/t2_32.c
Line
Count
Source
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 Wed Apr  1 06:04:44 UTC 2026 */
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 -twiddle-log3 -precompute-twiddles -n 32 -name t2_32 -include dft/scalar/t.h */
29
30
/*
31
 * This function contains 488 FP additions, 350 FP multiplications,
32
 * (or, 236 additions, 98 multiplications, 252 fused multiply/add),
33
 * 164 stack variables, 7 constants, and 128 memory accesses
34
 */
35
#include "dft/scalar/t.h"
36
37
static void t2_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
38
{
39
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
40
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
41
     DK(KP198912367, +0.198912367379658006911597622644676228597850501);
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 * 8); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 8, MAKE_VOLATILE_STRIDE(64, rs)) {
49
         E T2, T8, T3, T6, Te, Ti, T5, T7, TJ, Tb, TM, Tc, Ts, T23, T1w;
50
         E T19, TA, TE, T1s, T1N, T1o, T1C, T1F, T1K, T15, T11, T2F, T31, T2J, T34;
51
         E T3f, T3z, T3j, T3C, Tw, T3M, T3Q, T1z, T2s, T2w, T1d, T3n, T3r, T26, T2T;
52
         E T2X, Th, TR, TP, Td, Tj, TW, Tn, TS, T1U, T2b, T29, T1R, T1V, T2g;
53
         E T1Z, T2c;
54
         {
55
        E Tz, T1n, T10, TD, T1r, T14, T9, T1Q, Tv, T1c;
56
        {
57
       E T4, T18, Ta, Tr;
58
       T2 = W[0];
59
       T8 = W[4];
60
       T3 = W[2];
61
       T6 = W[3];
62
       T4 = T2 * T3;
63
       T18 = T3 * T8;
64
       Ta = T2 * T6;
65
       Tr = T2 * T8;
66
       Te = W[6];
67
       Tz = T3 * Te;
68
       T1n = T8 * Te;
69
       T10 = T2 * Te;
70
       Ti = W[7];
71
       TD = T3 * Ti;
72
       T1r = T8 * Ti;
73
       T14 = T2 * Ti;
74
       T5 = W[1];
75
       T7 = FMA(T5, T6, T4);
76
       TJ = FNMS(T5, T6, T4);
77
       T9 = T7 * T8;
78
       T1Q = TJ * T8;
79
       Tb = FNMS(T5, T3, Ta);
80
       TM = FMA(T5, T3, Ta);
81
       Tc = W[5];
82
       Tv = T2 * Tc;
83
       T1c = T3 * Tc;
84
       Ts = FMA(T5, Tc, Tr);
85
       T23 = FMA(T6, Tc, T18);
86
       T1w = FNMS(T5, Tc, Tr);
87
       T19 = FNMS(T6, Tc, T18);
88
        }
89
        TA = FMA(T6, Ti, Tz);
90
        TE = FNMS(T6, Te, TD);
91
        T1s = FNMS(Tc, Te, T1r);
92
        T1N = FMA(T6, Te, TD);
93
        T1o = FMA(Tc, Ti, T1n);
94
        T1C = FMA(T5, Ti, T10);
95
        T1F = FNMS(T5, Te, T14);
96
        T1K = FNMS(T6, Ti, Tz);
97
        T15 = FMA(T5, Te, T14);
98
        T11 = FNMS(T5, Ti, T10);
99
        {
100
       E T2E, T2I, T2S, T2W;
101
       T2E = T7 * Te;
102
       T2F = FMA(Tb, Ti, T2E);
103
       T31 = FNMS(Tb, Ti, T2E);
104
       T2I = T7 * Ti;
105
       T2J = FNMS(Tb, Te, T2I);
106
       T34 = FMA(Tb, Te, T2I);
107
       {
108
            E T3e, T3i, T3L, T3P;
109
            T3e = TJ * Te;
110
            T3f = FNMS(TM, Ti, T3e);
111
            T3z = FMA(TM, Ti, T3e);
112
            T3i = TJ * Ti;
113
            T3j = FMA(TM, Te, T3i);
114
            T3C = FNMS(TM, Te, T3i);
115
            T3L = Ts * Te;
116
            T3P = Ts * Ti;
117
            Tw = FNMS(T5, T8, Tv);
118
            T3M = FMA(Tw, Ti, T3L);
119
            T3Q = FNMS(Tw, Te, T3P);
120
       }
121
       {
122
            E T2r, T2v, T3m, T3q;
123
            T2r = T1w * Te;
124
            T2v = T1w * Ti;
125
            T1z = FMA(T5, T8, Tv);
126
            T2s = FMA(T1z, Ti, T2r);
127
            T2w = FNMS(T1z, Te, T2v);
128
            T3m = T19 * Te;
129
            T3q = T19 * Ti;
130
            T1d = FMA(T6, T8, T1c);
131
            T3n = FMA(T1d, Ti, T3m);
132
            T3r = FNMS(T1d, Te, T3q);
133
       }
134
       T2S = T23 * Te;
135
       T2W = T23 * Ti;
136
       T26 = FNMS(T6, T8, T1c);
137
       T2T = FMA(T26, Ti, T2S);
138
       T2X = FNMS(T26, Te, T2W);
139
       {
140
            E TQ, TV, Tf, Tm, Tg;
141
            Tg = T7 * Tc;
142
            Th = FMA(Tb, T8, Tg);
143
            TR = FNMS(Tb, T8, Tg);
144
            TP = FMA(Tb, Tc, T9);
145
            TQ = TP * Te;
146
            TV = TP * Ti;
147
            Td = FNMS(Tb, Tc, T9);
148
            Tf = Td * Te;
149
            Tm = Td * Ti;
150
            Tj = FMA(Th, Ti, Tf);
151
            TW = FNMS(TR, Te, TV);
152
            Tn = FNMS(Th, Te, Tm);
153
            TS = FMA(TR, Ti, TQ);
154
       }
155
       {
156
            E T2a, T2f, T1S, T1Y, T1T;
157
            T1T = TJ * Tc;
158
            T1U = FMA(TM, T8, T1T);
159
            T2b = FNMS(TM, T8, T1T);
160
            T29 = FMA(TM, Tc, T1Q);
161
            T2a = T29 * Te;
162
            T2f = T29 * Ti;
163
            T1R = FNMS(TM, Tc, T1Q);
164
            T1S = T1R * Te;
165
            T1Y = T1R * Ti;
166
            T1V = FMA(T1U, Ti, T1S);
167
            T2g = FNMS(T2b, Te, T2f);
168
            T1Z = FNMS(T1U, Te, T1Y);
169
            T2c = FMA(T2b, Ti, T2a);
170
       }
171
        }
172
         }
173
         {
174
        E Tq, T46, T8H, T97, TH, T98, T4b, T8D, TZ, T7f, T4j, T6t, T1g, T7g, T4q;
175
        E T6u, T1v, T1I, T7m, T7j, T7k, T7l, T4z, T6x, T4G, T6y, T22, T2j, T7o, T7p;
176
        E T7q, T7r, T4O, T6A, T4V, T6B, T3G, T7L, T7I, T8n, T5E, T6P, T61, T6M, T2N;
177
        E T7A, T7x, T8i, T55, T6I, T5s, T6F, T43, T7J, T7O, T8o, T5L, T62, T5S, T63;
178
        E T3c, T7y, T7D, T8j, T5c, T5t, T5j, T5u;
179
        {
180
       E T1, T8G, Tk, Tl, To, T8E, Tp, T8F;
181
       T1 = ri[0];
182
       T8G = ii[0];
183
       Tk = ri[WS(rs, 16)];
184
       Tl = Tj * Tk;
185
       To = ii[WS(rs, 16)];
186
       T8E = Tj * To;
187
       Tp = FMA(Tn, To, Tl);
188
       Tq = T1 + Tp;
189
       T46 = T1 - Tp;
190
       T8F = FNMS(Tn, Tk, T8E);
191
       T8H = T8F + T8G;
192
       T97 = T8G - T8F;
193
        }
194
        {
195
       E Tt, Tu, Tx, T47, TB, TC, TF, T49;
196
       Tt = ri[WS(rs, 8)];
197
       Tu = Ts * Tt;
198
       Tx = ii[WS(rs, 8)];
199
       T47 = Ts * Tx;
200
       TB = ri[WS(rs, 24)];
201
       TC = TA * TB;
202
       TF = ii[WS(rs, 24)];
203
       T49 = TA * TF;
204
       {
205
            E Ty, TG, T48, T4a;
206
            Ty = FMA(Tw, Tx, Tu);
207
            TG = FMA(TE, TF, TC);
208
            TH = Ty + TG;
209
            T98 = Ty - TG;
210
            T48 = FNMS(Tw, Tt, T47);
211
            T4a = FNMS(TE, TB, T49);
212
            T4b = T48 - T4a;
213
            T8D = T48 + T4a;
214
       }
215
        }
216
        {
217
       E TO, T4f, TY, T4h, T4d, T4i;
218
       {
219
            E TK, TL, TN, T4e;
220
            TK = ri[WS(rs, 4)];
221
            TL = TJ * TK;
222
            TN = ii[WS(rs, 4)];
223
            T4e = TJ * TN;
224
            TO = FMA(TM, TN, TL);
225
            T4f = FNMS(TM, TK, T4e);
226
       }
227
       {
228
            E TT, TU, TX, T4g;
229
            TT = ri[WS(rs, 20)];
230
            TU = TS * TT;
231
            TX = ii[WS(rs, 20)];
232
            T4g = TS * TX;
233
            TY = FMA(TW, TX, TU);
234
            T4h = FNMS(TW, TT, T4g);
235
       }
236
       TZ = TO + TY;
237
       T7f = T4f + T4h;
238
       T4d = TO - TY;
239
       T4i = T4f - T4h;
240
       T4j = T4d + T4i;
241
       T6t = T4i - T4d;
242
        }
243
        {
244
       E T17, T4m, T1f, T4o, T4k, T4p;
245
       {
246
            E T12, T13, T16, T4l;
247
            T12 = ri[WS(rs, 28)];
248
            T13 = T11 * T12;
249
            T16 = ii[WS(rs, 28)];
250
            T4l = T11 * T16;
251
            T17 = FMA(T15, T16, T13);
252
            T4m = FNMS(T15, T12, T4l);
253
       }
254
       {
255
            E T1a, T1b, T1e, T4n;
256
            T1a = ri[WS(rs, 12)];
257
            T1b = T19 * T1a;
258
            T1e = ii[WS(rs, 12)];
259
            T4n = T19 * T1e;
260
            T1f = FMA(T1d, T1e, T1b);
261
            T4o = FNMS(T1d, T1a, T4n);
262
       }
263
       T1g = T17 + T1f;
264
       T7g = T4m + T4o;
265
       T4k = T17 - T1f;
266
       T4p = T4m - T4o;
267
       T4q = T4k - T4p;
268
       T6u = T4k + T4p;
269
        }
270
        {
271
       E T1m, T4u, T1H, T4E, T1u, T4w, T1B, T4C;
272
       {
273
            E T1j, T1k, T1l, T4t;
274
            T1j = ri[WS(rs, 2)];
275
            T1k = T7 * T1j;
276
            T1l = ii[WS(rs, 2)];
277
            T4t = T7 * T1l;
278
            T1m = FMA(Tb, T1l, T1k);
279
            T4u = FNMS(Tb, T1j, T4t);
280
       }
281
       {
282
            E T1D, T1E, T1G, T4D;
283
            T1D = ri[WS(rs, 26)];
284
            T1E = T1C * T1D;
285
            T1G = ii[WS(rs, 26)];
286
            T4D = T1C * T1G;
287
            T1H = FMA(T1F, T1G, T1E);
288
            T4E = FNMS(T1F, T1D, T4D);
289
       }
290
       {
291
            E T1p, T1q, T1t, T4v;
292
            T1p = ri[WS(rs, 18)];
293
            T1q = T1o * T1p;
294
            T1t = ii[WS(rs, 18)];
295
            T4v = T1o * T1t;
296
            T1u = FMA(T1s, T1t, T1q);
297
            T4w = FNMS(T1s, T1p, T4v);
298
       }
299
       {
300
            E T1x, T1y, T1A, T4B;
301
            T1x = ri[WS(rs, 10)];
302
            T1y = T1w * T1x;
303
            T1A = ii[WS(rs, 10)];
304
            T4B = T1w * T1A;
305
            T1B = FMA(T1z, T1A, T1y);
306
            T4C = FNMS(T1z, T1x, T4B);
307
       }
308
       T1v = T1m + T1u;
309
       T1I = T1B + T1H;
310
       T7m = T1v - T1I;
311
       T7j = T4u + T4w;
312
       T7k = T4C + T4E;
313
       T7l = T7j - T7k;
314
       {
315
            E T4x, T4y, T4A, T4F;
316
            T4x = T4u - T4w;
317
            T4y = T1B - T1H;
318
            T4z = T4x - T4y;
319
            T6x = T4x + T4y;
320
            T4A = T1m - T1u;
321
            T4F = T4C - T4E;
322
            T4G = T4A + T4F;
323
            T6y = T4A - T4F;
324
       }
325
        }
326
        {
327
       E T1P, T4J, T2i, T4T, T21, T4L, T28, T4R;
328
       {
329
            E T1L, T1M, T1O, T4I;
330
            T1L = ri[WS(rs, 30)];
331
            T1M = T1K * T1L;
332
            T1O = ii[WS(rs, 30)];
333
            T4I = T1K * T1O;
334
            T1P = FMA(T1N, T1O, T1M);
335
            T4J = FNMS(T1N, T1L, T4I);
336
       }
337
       {
338
            E T2d, T2e, T2h, T4S;
339
            T2d = ri[WS(rs, 22)];
340
            T2e = T2c * T2d;
341
            T2h = ii[WS(rs, 22)];
342
            T4S = T2c * T2h;
343
            T2i = FMA(T2g, T2h, T2e);
344
            T4T = FNMS(T2g, T2d, T4S);
345
       }
346
       {
347
            E T1W, T1X, T20, T4K;
348
            T1W = ri[WS(rs, 14)];
349
            T1X = T1V * T1W;
350
            T20 = ii[WS(rs, 14)];
351
            T4K = T1V * T20;
352
            T21 = FMA(T1Z, T20, T1X);
353
            T4L = FNMS(T1Z, T1W, T4K);
354
       }
355
       {
356
            E T24, T25, T27, T4Q;
357
            T24 = ri[WS(rs, 6)];
358
            T25 = T23 * T24;
359
            T27 = ii[WS(rs, 6)];
360
            T4Q = T23 * T27;
361
            T28 = FMA(T26, T27, T25);
362
            T4R = FNMS(T26, T24, T4Q);
363
       }
364
       T22 = T1P + T21;
365
       T2j = T28 + T2i;
366
       T7o = T22 - T2j;
367
       T7p = T4J + T4L;
368
       T7q = T4R + T4T;
369
       T7r = T7p - T7q;
370
       {
371
            E T4M, T4N, T4P, T4U;
372
            T4M = T4J - T4L;
373
            T4N = T28 - T2i;
374
            T4O = T4M - T4N;
375
            T6A = T4M + T4N;
376
            T4P = T1P - T21;
377
            T4U = T4R - T4T;
378
            T4V = T4P + T4U;
379
            T6B = T4P - T4U;
380
       }
381
        }
382
        {
383
       E T3l, T5z, T3E, T5Z, T3t, T5B, T3y, T5X;
384
       {
385
            E T3g, T3h, T3k, T5y;
386
            T3g = ri[WS(rs, 31)];
387
            T3h = T3f * T3g;
388
            T3k = ii[WS(rs, 31)];
389
            T5y = T3f * T3k;
390
            T3l = FMA(T3j, T3k, T3h);
391
            T5z = FNMS(T3j, T3g, T5y);
392
       }
393
       {
394
            E T3A, T3B, T3D, T5Y;
395
            T3A = ri[WS(rs, 23)];
396
            T3B = T3z * T3A;
397
            T3D = ii[WS(rs, 23)];
398
            T5Y = T3z * T3D;
399
            T3E = FMA(T3C, T3D, T3B);
400
            T5Z = FNMS(T3C, T3A, T5Y);
401
       }
402
       {
403
            E T3o, T3p, T3s, T5A;
404
            T3o = ri[WS(rs, 15)];
405
            T3p = T3n * T3o;
406
            T3s = ii[WS(rs, 15)];
407
            T5A = T3n * T3s;
408
            T3t = FMA(T3r, T3s, T3p);
409
            T5B = FNMS(T3r, T3o, T5A);
410
       }
411
       {
412
            E T3v, T3w, T3x, T5W;
413
            T3v = ri[WS(rs, 7)];
414
            T3w = TP * T3v;
415
            T3x = ii[WS(rs, 7)];
416
            T5W = TP * T3x;
417
            T3y = FMA(TR, T3x, T3w);
418
            T5X = FNMS(TR, T3v, T5W);
419
       }
420
       {
421
            E T3u, T3F, T7G, T7H;
422
            T3u = T3l + T3t;
423
            T3F = T3y + T3E;
424
            T3G = T3u + T3F;
425
            T7L = T3u - T3F;
426
            T7G = T5z + T5B;
427
            T7H = T5X + T5Z;
428
            T7I = T7G - T7H;
429
            T8n = T7G + T7H;
430
       }
431
       {
432
            E T5C, T5D, T5V, T60;
433
            T5C = T5z - T5B;
434
            T5D = T3y - T3E;
435
            T5E = T5C - T5D;
436
            T6P = T5C + T5D;
437
            T5V = T3l - T3t;
438
            T60 = T5X - T5Z;
439
            T61 = T5V + T60;
440
            T6M = T5V - T60;
441
       }
442
        }
443
        {
444
       E T2q, T50, T2L, T5q, T2y, T52, T2D, T5o;
445
       {
446
            E T2n, T2o, T2p, T4Z;
447
            T2n = ri[WS(rs, 1)];
448
            T2o = T2 * T2n;
449
            T2p = ii[WS(rs, 1)];
450
            T4Z = T2 * T2p;
451
            T2q = FMA(T5, T2p, T2o);
452
            T50 = FNMS(T5, T2n, T4Z);
453
       }
454
       {
455
            E T2G, T2H, T2K, T5p;
456
            T2G = ri[WS(rs, 25)];
457
            T2H = T2F * T2G;
458
            T2K = ii[WS(rs, 25)];
459
            T5p = T2F * T2K;
460
            T2L = FMA(T2J, T2K, T2H);
461
            T5q = FNMS(T2J, T2G, T5p);
462
       }
463
       {
464
            E T2t, T2u, T2x, T51;
465
            T2t = ri[WS(rs, 17)];
466
            T2u = T2s * T2t;
467
            T2x = ii[WS(rs, 17)];
468
            T51 = T2s * T2x;
469
            T2y = FMA(T2w, T2x, T2u);
470
            T52 = FNMS(T2w, T2t, T51);
471
       }
472
       {
473
            E T2A, T2B, T2C, T5n;
474
            T2A = ri[WS(rs, 9)];
475
            T2B = T8 * T2A;
476
            T2C = ii[WS(rs, 9)];
477
            T5n = T8 * T2C;
478
            T2D = FMA(Tc, T2C, T2B);
479
            T5o = FNMS(Tc, T2A, T5n);
480
       }
481
       {
482
            E T2z, T2M, T7v, T7w;
483
            T2z = T2q + T2y;
484
            T2M = T2D + T2L;
485
            T2N = T2z + T2M;
486
            T7A = T2z - T2M;
487
            T7v = T50 + T52;
488
            T7w = T5o + T5q;
489
            T7x = T7v - T7w;
490
            T8i = T7v + T7w;
491
       }
492
       {
493
            E T53, T54, T5m, T5r;
494
            T53 = T50 - T52;
495
            T54 = T2D - T2L;
496
            T55 = T53 - T54;
497
            T6I = T53 + T54;
498
            T5m = T2q - T2y;
499
            T5r = T5o - T5q;
500
            T5s = T5m + T5r;
501
            T6F = T5m - T5r;
502
       }
503
        }
504
        {
505
       E T3K, T5G, T41, T5Q, T3S, T5I, T3X, T5O;
506
       {
507
            E T3H, T3I, T3J, T5F;
508
            T3H = ri[WS(rs, 3)];
509
            T3I = T3 * T3H;
510
            T3J = ii[WS(rs, 3)];
511
            T5F = T3 * T3J;
512
            T3K = FMA(T6, T3J, T3I);
513
            T5G = FNMS(T6, T3H, T5F);
514
       }
515
       {
516
            E T3Y, T3Z, T40, T5P;
517
            T3Y = ri[WS(rs, 11)];
518
            T3Z = Td * T3Y;
519
            T40 = ii[WS(rs, 11)];
520
            T5P = Td * T40;
521
            T41 = FMA(Th, T40, T3Z);
522
            T5Q = FNMS(Th, T3Y, T5P);
523
       }
524
       {
525
            E T3N, T3O, T3R, T5H;
526
            T3N = ri[WS(rs, 19)];
527
            T3O = T3M * T3N;
528
            T3R = ii[WS(rs, 19)];
529
            T5H = T3M * T3R;
530
            T3S = FMA(T3Q, T3R, T3O);
531
            T5I = FNMS(T3Q, T3N, T5H);
532
       }
533
       {
534
            E T3U, T3V, T3W, T5N;
535
            T3U = ri[WS(rs, 27)];
536
            T3V = Te * T3U;
537
            T3W = ii[WS(rs, 27)];
538
            T5N = Te * T3W;
539
            T3X = FMA(Ti, T3W, T3V);
540
            T5O = FNMS(Ti, T3U, T5N);
541
       }
542
       {
543
            E T3T, T42, T7M, T7N;
544
            T3T = T3K + T3S;
545
            T42 = T3X + T41;
546
            T43 = T3T + T42;
547
            T7J = T42 - T3T;
548
            T7M = T5G + T5I;
549
            T7N = T5O + T5Q;
550
            T7O = T7M - T7N;
551
            T8o = T7M + T7N;
552
       }
553
       {
554
            E T5J, T5K, T5M, T5R;
555
            T5J = T5G - T5I;
556
            T5K = T3K - T3S;
557
            T5L = T5J - T5K;
558
            T62 = T5K + T5J;
559
            T5M = T3X - T41;
560
            T5R = T5O - T5Q;
561
            T5S = T5M + T5R;
562
            T63 = T5M - T5R;
563
       }
564
        }
565
        {
566
       E T2R, T57, T3a, T5h, T2Z, T59, T36, T5f;
567
       {
568
            E T2O, T2P, T2Q, T56;
569
            T2O = ri[WS(rs, 5)];
570
            T2P = T29 * T2O;
571
            T2Q = ii[WS(rs, 5)];
572
            T56 = T29 * T2Q;
573
            T2R = FMA(T2b, T2Q, T2P);
574
            T57 = FNMS(T2b, T2O, T56);
575
       }
576
       {
577
            E T37, T38, T39, T5g;
578
            T37 = ri[WS(rs, 13)];
579
            T38 = T1R * T37;
580
            T39 = ii[WS(rs, 13)];
581
            T5g = T1R * T39;
582
            T3a = FMA(T1U, T39, T38);
583
            T5h = FNMS(T1U, T37, T5g);
584
       }
585
       {
586
            E T2U, T2V, T2Y, T58;
587
            T2U = ri[WS(rs, 21)];
588
            T2V = T2T * T2U;
589
            T2Y = ii[WS(rs, 21)];
590
            T58 = T2T * T2Y;
591
            T2Z = FMA(T2X, T2Y, T2V);
592
            T59 = FNMS(T2X, T2U, T58);
593
       }
594
       {
595
            E T32, T33, T35, T5e;
596
            T32 = ri[WS(rs, 29)];
597
            T33 = T31 * T32;
598
            T35 = ii[WS(rs, 29)];
599
            T5e = T31 * T35;
600
            T36 = FMA(T34, T35, T33);
601
            T5f = FNMS(T34, T32, T5e);
602
       }
603
       {
604
            E T30, T3b, T7B, T7C;
605
            T30 = T2R + T2Z;
606
            T3b = T36 + T3a;
607
            T3c = T30 + T3b;
608
            T7y = T3b - T30;
609
            T7B = T57 + T59;
610
            T7C = T5f + T5h;
611
            T7D = T7B - T7C;
612
            T8j = T7B + T7C;
613
       }
614
       {
615
            E T5a, T5b, T5d, T5i;
616
            T5a = T57 - T59;
617
            T5b = T2R - T2Z;
618
            T5c = T5a - T5b;
619
            T5t = T5b + T5a;
620
            T5d = T36 - T3a;
621
            T5i = T5f - T5h;
622
            T5j = T5d + T5i;
623
            T5u = T5d - T5i;
624
       }
625
        }
626
        {
627
       E T1i, T8c, T8z, T8A, T8J, T8O, T2l, T8N, T45, T8L, T8l, T8t, T8q, T8u, T8f;
628
       E T8B;
629
       {
630
            E TI, T1h, T8x, T8y;
631
            TI = Tq + TH;
632
            T1h = TZ + T1g;
633
            T1i = TI + T1h;
634
            T8c = TI - T1h;
635
            T8x = T8i + T8j;
636
            T8y = T8n + T8o;
637
            T8z = T8x - T8y;
638
            T8A = T8x + T8y;
639
       }
640
       {
641
            E T8C, T8I, T1J, T2k;
642
            T8C = T7f + T7g;
643
            T8I = T8D + T8H;
644
            T8J = T8C + T8I;
645
            T8O = T8I - T8C;
646
            T1J = T1v + T1I;
647
            T2k = T22 + T2j;
648
            T2l = T1J + T2k;
649
            T8N = T2k - T1J;
650
       }
651
       {
652
            E T3d, T44, T8h, T8k;
653
            T3d = T2N + T3c;
654
            T44 = T3G + T43;
655
            T45 = T3d + T44;
656
            T8L = T44 - T3d;
657
            T8h = T2N - T3c;
658
            T8k = T8i - T8j;
659
            T8l = T8h + T8k;
660
            T8t = T8k - T8h;
661
       }
662
       {
663
            E T8m, T8p, T8d, T8e;
664
            T8m = T3G - T43;
665
            T8p = T8n - T8o;
666
            T8q = T8m - T8p;
667
            T8u = T8m + T8p;
668
            T8d = T7j + T7k;
669
            T8e = T7p + T7q;
670
            T8f = T8d - T8e;
671
            T8B = T8d + T8e;
672
       }
673
       {
674
            E T2m, T8K, T8w, T8M;
675
            T2m = T1i + T2l;
676
            ri[WS(rs, 16)] = T2m - T45;
677
            ri[0] = T2m + T45;
678
            T8K = T8B + T8J;
679
            ii[0] = T8A + T8K;
680
            ii[WS(rs, 16)] = T8K - T8A;
681
            T8w = T1i - T2l;
682
            ri[WS(rs, 24)] = T8w - T8z;
683
            ri[WS(rs, 8)] = T8w + T8z;
684
            T8M = T8J - T8B;
685
            ii[WS(rs, 8)] = T8L + T8M;
686
            ii[WS(rs, 24)] = T8M - T8L;
687
       }
688
       {
689
            E T8g, T8r, T8P, T8Q;
690
            T8g = T8c + T8f;
691
            T8r = T8l + T8q;
692
            ri[WS(rs, 20)] = FNMS(KP707106781, T8r, T8g);
693
            ri[WS(rs, 4)] = FMA(KP707106781, T8r, T8g);
694
            T8P = T8N + T8O;
695
            T8Q = T8t + T8u;
696
            ii[WS(rs, 4)] = FMA(KP707106781, T8Q, T8P);
697
            ii[WS(rs, 20)] = FNMS(KP707106781, T8Q, T8P);
698
       }
699
       {
700
            E T8s, T8v, T8R, T8S;
701
            T8s = T8c - T8f;
702
            T8v = T8t - T8u;
703
            ri[WS(rs, 28)] = FNMS(KP707106781, T8v, T8s);
704
            ri[WS(rs, 12)] = FMA(KP707106781, T8v, T8s);
705
            T8R = T8O - T8N;
706
            T8S = T8q - T8l;
707
            ii[WS(rs, 12)] = FMA(KP707106781, T8S, T8R);
708
            ii[WS(rs, 28)] = FNMS(KP707106781, T8S, T8R);
709
       }
710
        }
711
        {
712
       E T7i, T7W, T86, T8a, T8V, T91, T7t, T8W, T7F, T7T, T7Z, T92, T83, T89, T7Q;
713
       E T7U;
714
       {
715
            E T7e, T7h, T84, T85;
716
            T7e = Tq - TH;
717
            T7h = T7f - T7g;
718
            T7i = T7e - T7h;
719
            T7W = T7e + T7h;
720
            T84 = T7L + T7O;
721
            T85 = T7I + T7J;
722
            T86 = FNMS(KP414213562, T85, T84);
723
            T8a = FMA(KP414213562, T84, T85);
724
       }
725
       {
726
            E T8T, T8U, T7n, T7s;
727
            T8T = T1g - TZ;
728
            T8U = T8H - T8D;
729
            T8V = T8T + T8U;
730
            T91 = T8U - T8T;
731
            T7n = T7l - T7m;
732
            T7s = T7o + T7r;
733
            T7t = T7n - T7s;
734
            T8W = T7n + T7s;
735
       }
736
       {
737
            E T7z, T7E, T7X, T7Y;
738
            T7z = T7x - T7y;
739
            T7E = T7A - T7D;
740
            T7F = FMA(KP414213562, T7E, T7z);
741
            T7T = FNMS(KP414213562, T7z, T7E);
742
            T7X = T7m + T7l;
743
            T7Y = T7o - T7r;
744
            T7Z = T7X + T7Y;
745
            T92 = T7Y - T7X;
746
       }
747
       {
748
            E T81, T82, T7K, T7P;
749
            T81 = T7A + T7D;
750
            T82 = T7x + T7y;
751
            T83 = FMA(KP414213562, T82, T81);
752
            T89 = FNMS(KP414213562, T81, T82);
753
            T7K = T7I - T7J;
754
            T7P = T7L - T7O;
755
            T7Q = FNMS(KP414213562, T7P, T7K);
756
            T7U = FMA(KP414213562, T7K, T7P);
757
       }
758
       {
759
            E T7u, T7R, T93, T94;
760
            T7u = FMA(KP707106781, T7t, T7i);
761
            T7R = T7F - T7Q;
762
            ri[WS(rs, 22)] = FNMS(KP923879532, T7R, T7u);
763
            ri[WS(rs, 6)] = FMA(KP923879532, T7R, T7u);
764
            T93 = FMA(KP707106781, T92, T91);
765
            T94 = T7U - T7T;
766
            ii[WS(rs, 6)] = FMA(KP923879532, T94, T93);
767
            ii[WS(rs, 22)] = FNMS(KP923879532, T94, T93);
768
       }
769
       {
770
            E T7S, T7V, T95, T96;
771
            T7S = FNMS(KP707106781, T7t, T7i);
772
            T7V = T7T + T7U;
773
            ri[WS(rs, 14)] = FNMS(KP923879532, T7V, T7S);
774
            ri[WS(rs, 30)] = FMA(KP923879532, T7V, T7S);
775
            T95 = FNMS(KP707106781, T92, T91);
776
            T96 = T7F + T7Q;
777
            ii[WS(rs, 14)] = FNMS(KP923879532, T96, T95);
778
            ii[WS(rs, 30)] = FMA(KP923879532, T96, T95);
779
       }
780
       {
781
            E T80, T87, T8X, T8Y;
782
            T80 = FMA(KP707106781, T7Z, T7W);
783
            T87 = T83 + T86;
784
            ri[WS(rs, 18)] = FNMS(KP923879532, T87, T80);
785
            ri[WS(rs, 2)] = FMA(KP923879532, T87, T80);
786
            T8X = FMA(KP707106781, T8W, T8V);
787
            T8Y = T89 + T8a;
788
            ii[WS(rs, 2)] = FMA(KP923879532, T8Y, T8X);
789
            ii[WS(rs, 18)] = FNMS(KP923879532, T8Y, T8X);
790
       }
791
       {
792
            E T88, T8b, T8Z, T90;
793
            T88 = FNMS(KP707106781, T7Z, T7W);
794
            T8b = T89 - T8a;
795
            ri[WS(rs, 26)] = FNMS(KP923879532, T8b, T88);
796
            ri[WS(rs, 10)] = FMA(KP923879532, T8b, T88);
797
            T8Z = FNMS(KP707106781, T8W, T8V);
798
            T90 = T86 - T83;
799
            ii[WS(rs, 10)] = FMA(KP923879532, T90, T8Z);
800
            ii[WS(rs, 26)] = FNMS(KP923879532, T90, T8Z);
801
       }
802
        }
803
        {
804
       E T4s, T6c, T4X, T9c, T9b, T9h, T6f, T9i, T66, T6q, T6a, T6m, T5x, T6p, T69;
805
       E T6j;
806
       {
807
            E T4c, T4r, T6d, T6e;
808
            T4c = T46 + T4b;
809
            T4r = T4j + T4q;
810
            T4s = FNMS(KP707106781, T4r, T4c);
811
            T6c = FMA(KP707106781, T4r, T4c);
812
            {
813
           E T4H, T4W, T99, T9a;
814
           T4H = FNMS(KP414213562, T4G, T4z);
815
           T4W = FMA(KP414213562, T4V, T4O);
816
           T4X = T4H - T4W;
817
           T9c = T4H + T4W;
818
           T99 = T97 - T98;
819
           T9a = T6t + T6u;
820
           T9b = FMA(KP707106781, T9a, T99);
821
           T9h = FNMS(KP707106781, T9a, T99);
822
            }
823
            T6d = FMA(KP414213562, T4z, T4G);
824
            T6e = FNMS(KP414213562, T4O, T4V);
825
            T6f = T6d + T6e;
826
            T9i = T6e - T6d;
827
            {
828
           E T5U, T6l, T65, T6k, T5T, T64;
829
           T5T = T5L + T5S;
830
           T5U = FNMS(KP707106781, T5T, T5E);
831
           T6l = FMA(KP707106781, T5T, T5E);
832
           T64 = T62 + T63;
833
           T65 = FNMS(KP707106781, T64, T61);
834
           T6k = FMA(KP707106781, T64, T61);
835
           T66 = FNMS(KP668178637, T65, T5U);
836
           T6q = FMA(KP198912367, T6k, T6l);
837
           T6a = FMA(KP668178637, T5U, T65);
838
           T6m = FNMS(KP198912367, T6l, T6k);
839
            }
840
            {
841
           E T5l, T6i, T5w, T6h, T5k, T5v;
842
           T5k = T5c + T5j;
843
           T5l = FNMS(KP707106781, T5k, T55);
844
           T6i = FMA(KP707106781, T5k, T55);
845
           T5v = T5t + T5u;
846
           T5w = FNMS(KP707106781, T5v, T5s);
847
           T6h = FMA(KP707106781, T5v, T5s);
848
           T5x = FMA(KP668178637, T5w, T5l);
849
           T6p = FNMS(KP198912367, T6h, T6i);
850
           T69 = FNMS(KP668178637, T5l, T5w);
851
           T6j = FMA(KP198912367, T6i, T6h);
852
            }
853
       }
854
       {
855
            E T4Y, T67, T9j, T9k;
856
            T4Y = FMA(KP923879532, T4X, T4s);
857
            T67 = T5x - T66;
858
            ri[WS(rs, 21)] = FNMS(KP831469612, T67, T4Y);
859
            ri[WS(rs, 5)] = FMA(KP831469612, T67, T4Y);
860
            T9j = FMA(KP923879532, T9i, T9h);
861
            T9k = T6a - T69;
862
            ii[WS(rs, 5)] = FMA(KP831469612, T9k, T9j);
863
            ii[WS(rs, 21)] = FNMS(KP831469612, T9k, T9j);
864
       }
865
       {
866
            E T68, T6b, T9l, T9m;
867
            T68 = FNMS(KP923879532, T4X, T4s);
868
            T6b = T69 + T6a;
869
            ri[WS(rs, 13)] = FNMS(KP831469612, T6b, T68);
870
            ri[WS(rs, 29)] = FMA(KP831469612, T6b, T68);
871
            T9l = FNMS(KP923879532, T9i, T9h);
872
            T9m = T5x + T66;
873
            ii[WS(rs, 13)] = FNMS(KP831469612, T9m, T9l);
874
            ii[WS(rs, 29)] = FMA(KP831469612, T9m, T9l);
875
       }
876
       {
877
            E T6g, T6n, T9d, T9e;
878
            T6g = FMA(KP923879532, T6f, T6c);
879
            T6n = T6j + T6m;
880
            ri[WS(rs, 17)] = FNMS(KP980785280, T6n, T6g);
881
            ri[WS(rs, 1)] = FMA(KP980785280, T6n, T6g);
882
            T9d = FMA(KP923879532, T9c, T9b);
883
            T9e = T6p + T6q;
884
            ii[WS(rs, 1)] = FMA(KP980785280, T9e, T9d);
885
            ii[WS(rs, 17)] = FNMS(KP980785280, T9e, T9d);
886
       }
887
       {
888
            E T6o, T6r, T9f, T9g;
889
            T6o = FNMS(KP923879532, T6f, T6c);
890
            T6r = T6p - T6q;
891
            ri[WS(rs, 25)] = FNMS(KP980785280, T6r, T6o);
892
            ri[WS(rs, 9)] = FMA(KP980785280, T6r, T6o);
893
            T9f = FNMS(KP923879532, T9c, T9b);
894
            T9g = T6m - T6j;
895
            ii[WS(rs, 9)] = FMA(KP980785280, T9g, T9f);
896
            ii[WS(rs, 25)] = FNMS(KP980785280, T9g, T9f);
897
       }
898
        }
899
        {
900
       E T6w, T6Y, T6D, T9w, T9p, T9v, T71, T9q, T6S, T7c, T6W, T78, T6L, T7b, T6V;
901
       E T75;
902
       {
903
            E T6s, T6v, T6Z, T70;
904
            T6s = T46 - T4b;
905
            T6v = T6t - T6u;
906
            T6w = FMA(KP707106781, T6v, T6s);
907
            T6Y = FNMS(KP707106781, T6v, T6s);
908
            {
909
           E T6z, T6C, T9n, T9o;
910
           T6z = FMA(KP414213562, T6y, T6x);
911
           T6C = FNMS(KP414213562, T6B, T6A);
912
           T6D = T6z - T6C;
913
           T9w = T6z + T6C;
914
           T9n = T98 + T97;
915
           T9o = T4q - T4j;
916
           T9p = FMA(KP707106781, T9o, T9n);
917
           T9v = FNMS(KP707106781, T9o, T9n);
918
            }
919
            T6Z = FNMS(KP414213562, T6x, T6y);
920
            T70 = FMA(KP414213562, T6A, T6B);
921
            T71 = T6Z + T70;
922
            T9q = T70 - T6Z;
923
            {
924
           E T6O, T77, T6R, T76, T6N, T6Q;
925
           T6N = T5S - T5L;
926
           T6O = FNMS(KP707106781, T6N, T6M);
927
           T77 = FMA(KP707106781, T6N, T6M);
928
           T6Q = T62 - T63;
929
           T6R = FNMS(KP707106781, T6Q, T6P);
930
           T76 = FMA(KP707106781, T6Q, T6P);
931
           T6S = FNMS(KP668178637, T6R, T6O);
932
           T7c = FMA(KP198912367, T76, T77);
933
           T6W = FMA(KP668178637, T6O, T6R);
934
           T78 = FNMS(KP198912367, T77, T76);
935
            }
936
            {
937
           E T6H, T74, T6K, T73, T6G, T6J;
938
           T6G = T5j - T5c;
939
           T6H = FNMS(KP707106781, T6G, T6F);
940
           T74 = FMA(KP707106781, T6G, T6F);
941
           T6J = T5t - T5u;
942
           T6K = FNMS(KP707106781, T6J, T6I);
943
           T73 = FMA(KP707106781, T6J, T6I);
944
           T6L = FMA(KP668178637, T6K, T6H);
945
           T7b = FNMS(KP198912367, T73, T74);
946
           T6V = FNMS(KP668178637, T6H, T6K);
947
           T75 = FMA(KP198912367, T74, T73);
948
            }
949
       }
950
       {
951
            E T6E, T6T, T9r, T9s;
952
            T6E = FMA(KP923879532, T6D, T6w);
953
            T6T = T6L + T6S;
954
            ri[WS(rs, 19)] = FNMS(KP831469612, T6T, T6E);
955
            ri[WS(rs, 3)] = FMA(KP831469612, T6T, T6E);
956
            T9r = FMA(KP923879532, T9q, T9p);
957
            T9s = T6V + T6W;
958
            ii[WS(rs, 3)] = FMA(KP831469612, T9s, T9r);
959
            ii[WS(rs, 19)] = FNMS(KP831469612, T9s, T9r);
960
       }
961
       {
962
            E T6U, T6X, T9t, T9u;
963
            T6U = FNMS(KP923879532, T6D, T6w);
964
            T6X = T6V - T6W;
965
            ri[WS(rs, 27)] = FNMS(KP831469612, T6X, T6U);
966
            ri[WS(rs, 11)] = FMA(KP831469612, T6X, T6U);
967
            T9t = FNMS(KP923879532, T9q, T9p);
968
            T9u = T6S - T6L;
969
            ii[WS(rs, 11)] = FMA(KP831469612, T9u, T9t);
970
            ii[WS(rs, 27)] = FNMS(KP831469612, T9u, T9t);
971
       }
972
       {
973
            E T72, T79, T9x, T9y;
974
            T72 = FNMS(KP923879532, T71, T6Y);
975
            T79 = T75 - T78;
976
            ri[WS(rs, 23)] = FNMS(KP980785280, T79, T72);
977
            ri[WS(rs, 7)] = FMA(KP980785280, T79, T72);
978
            T9x = FNMS(KP923879532, T9w, T9v);
979
            T9y = T7c - T7b;
980
            ii[WS(rs, 7)] = FMA(KP980785280, T9y, T9x);
981
            ii[WS(rs, 23)] = FNMS(KP980785280, T9y, T9x);
982
       }
983
       {
984
            E T7a, T7d, T9z, T9A;
985
            T7a = FMA(KP923879532, T71, T6Y);
986
            T7d = T7b + T7c;
987
            ri[WS(rs, 15)] = FNMS(KP980785280, T7d, T7a);
988
            ri[WS(rs, 31)] = FMA(KP980785280, T7d, T7a);
989
            T9z = FMA(KP923879532, T9w, T9v);
990
            T9A = T75 + T78;
991
            ii[WS(rs, 15)] = FNMS(KP980785280, T9A, T9z);
992
            ii[WS(rs, 31)] = FMA(KP980785280, T9A, T9z);
993
       }
994
        }
995
         }
996
    }
997
     }
998
}
999
1000
static const tw_instr twinstr[] = {
1001
     { TW_CEXP, 0, 1 },
1002
     { TW_CEXP, 0, 3 },
1003
     { TW_CEXP, 0, 9 },
1004
     { TW_CEXP, 0, 27 },
1005
     { TW_NEXT, 1, 0 }
1006
};
1007
1008
static const ct_desc desc = { 32, "t2_32", twinstr, &GENUS, { 236, 98, 252, 0 }, 0, 0, 0 };
1009
1010
void X(codelet_t2_32) (planner *p) {
1011
     X(kdft_dit_register) (p, t2_32, &desc);
1012
}
1013
#else
1014
1015
/* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 32 -name t2_32 -include dft/scalar/t.h */
1016
1017
/*
1018
 * This function contains 488 FP additions, 280 FP multiplications,
1019
 * (or, 376 additions, 168 multiplications, 112 fused multiply/add),
1020
 * 158 stack variables, 7 constants, and 128 memory accesses
1021
 */
1022
#include "dft/scalar/t.h"
1023
1024
static void t2_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
1025
6
{
1026
6
     DK(KP195090322, +0.195090322016128267848284868477022240927691618);
1027
6
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
1028
6
     DK(KP555570233, +0.555570233019602224742830813948532874374937191);
1029
6
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
1030
6
     DK(KP382683432, +0.382683432365089771728459984030398866761344562);
1031
6
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
1032
6
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
1033
6
     {
1034
6
    INT m;
1035
78
    for (m = mb, W = W + (mb * 8); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 8, MAKE_VOLATILE_STRIDE(64, rs)) {
1036
72
         E T2, T5, T3, T6, T8, TM, TO, Td, T9, Te, Th, Tl, TD, TH, T1y;
1037
72
         E T1H, T15, T1A, T11, T1F, T1n, T1p, T2q, T2I, T2u, T2K, T2V, T3b, T2Z, T3d;
1038
72
         E Tu, Ty, T3l, T3n, T1t, T1v, T2f, T2h, T1a, T1e, T32, T34, T1W, T1Y, T2C;
1039
72
         E T2E, Tg, TR, Tk, TS, Tm, TV, To, TT, T1M, T21, T1P, T22, T1Q, T25;
1040
72
         E T1S, T23;
1041
72
         {
1042
72
        E Ts, T1d, Tx, T18, Tt, T1c, Tw, T19, TB, T14, TG, TZ, TC, T13, TF;
1043
72
        E T10;
1044
72
        {
1045
72
       E T4, Tc, T7, Tb;
1046
72
       T2 = W[0];
1047
72
       T5 = W[1];
1048
72
       T3 = W[2];
1049
72
       T6 = W[3];
1050
72
       T4 = T2 * T3;
1051
72
       Tc = T5 * T3;
1052
72
       T7 = T5 * T6;
1053
72
       Tb = T2 * T6;
1054
72
       T8 = T4 + T7;
1055
72
       TM = T4 - T7;
1056
72
       TO = Tb + Tc;
1057
72
       Td = Tb - Tc;
1058
72
       T9 = W[4];
1059
72
       Ts = T2 * T9;
1060
72
       T1d = T6 * T9;
1061
72
       Tx = T5 * T9;
1062
72
       T18 = T3 * T9;
1063
72
       Te = W[5];
1064
72
       Tt = T5 * Te;
1065
72
       T1c = T3 * Te;
1066
72
       Tw = T2 * Te;
1067
72
       T19 = T6 * Te;
1068
72
       Th = W[6];
1069
72
       TB = T3 * Th;
1070
72
       T14 = T5 * Th;
1071
72
       TG = T6 * Th;
1072
72
       TZ = T2 * Th;
1073
72
       Tl = W[7];
1074
72
       TC = T6 * Tl;
1075
72
       T13 = T2 * Tl;
1076
72
       TF = T3 * Tl;
1077
72
       T10 = T5 * Tl;
1078
72
        }
1079
72
        TD = TB + TC;
1080
72
        TH = TF - TG;
1081
72
        T1y = TZ + T10;
1082
72
        T1H = TF + TG;
1083
72
        T15 = T13 + T14;
1084
72
        T1A = T13 - T14;
1085
72
        T11 = TZ - T10;
1086
72
        T1F = TB - TC;
1087
72
        T1n = FMA(T9, Th, Te * Tl);
1088
72
        T1p = FNMS(Te, Th, T9 * Tl);
1089
72
        {
1090
72
       E T2o, T2p, T2s, T2t;
1091
72
       T2o = T8 * Th;
1092
72
       T2p = Td * Tl;
1093
72
       T2q = T2o + T2p;
1094
72
       T2I = T2o - T2p;
1095
72
       T2s = T8 * Tl;
1096
72
       T2t = Td * Th;
1097
72
       T2u = T2s - T2t;
1098
72
       T2K = T2s + T2t;
1099
72
        }
1100
72
        {
1101
72
       E T2T, T2U, T2X, T2Y;
1102
72
       T2T = TM * Th;
1103
72
       T2U = TO * Tl;
1104
72
       T2V = T2T - T2U;
1105
72
       T3b = T2T + T2U;
1106
72
       T2X = TM * Tl;
1107
72
       T2Y = TO * Th;
1108
72
       T2Z = T2X + T2Y;
1109
72
       T3d = T2X - T2Y;
1110
72
       Tu = Ts + Tt;
1111
72
       Ty = Tw - Tx;
1112
72
       T3l = FMA(Tu, Th, Ty * Tl);
1113
72
       T3n = FNMS(Ty, Th, Tu * Tl);
1114
72
        }
1115
72
        T1t = Ts - Tt;
1116
72
        T1v = Tw + Tx;
1117
72
        T2f = FMA(T1t, Th, T1v * Tl);
1118
72
        T2h = FNMS(T1v, Th, T1t * Tl);
1119
72
        T1a = T18 - T19;
1120
72
        T1e = T1c + T1d;
1121
72
        T32 = FMA(T1a, Th, T1e * Tl);
1122
72
        T34 = FNMS(T1e, Th, T1a * Tl);
1123
72
        T1W = T18 + T19;
1124
72
        T1Y = T1c - T1d;
1125
72
        T2C = FMA(T1W, Th, T1Y * Tl);
1126
72
        T2E = FNMS(T1Y, Th, T1W * Tl);
1127
72
        {
1128
72
       E Ta, Tf, Ti, Tj;
1129
72
       Ta = T8 * T9;
1130
72
       Tf = Td * Te;
1131
72
       Tg = Ta - Tf;
1132
72
       TR = Ta + Tf;
1133
72
       Ti = T8 * Te;
1134
72
       Tj = Td * T9;
1135
72
       Tk = Ti + Tj;
1136
72
       TS = Ti - Tj;
1137
72
        }
1138
72
        Tm = FMA(Tg, Th, Tk * Tl);
1139
72
        TV = FNMS(TS, Th, TR * Tl);
1140
72
        To = FNMS(Tk, Th, Tg * Tl);
1141
72
        TT = FMA(TR, Th, TS * Tl);
1142
72
        {
1143
72
       E T1K, T1L, T1N, T1O;
1144
72
       T1K = TM * T9;
1145
72
       T1L = TO * Te;
1146
72
       T1M = T1K - T1L;
1147
72
       T21 = T1K + T1L;
1148
72
       T1N = TM * Te;
1149
72
       T1O = TO * T9;
1150
72
       T1P = T1N + T1O;
1151
72
       T22 = T1N - T1O;
1152
72
        }
1153
72
        T1Q = FMA(T1M, Th, T1P * Tl);
1154
72
        T25 = FNMS(T22, Th, T21 * Tl);
1155
72
        T1S = FNMS(T1P, Th, T1M * Tl);
1156
72
        T23 = FMA(T21, Th, T22 * Tl);
1157
72
         }
1158
72
         {
1159
72
        E TL, T6f, T8c, T8q, T3F, T5t, T7I, T7W, T2y, T6B, T6y, T7j, T4k, T5J, T4B;
1160
72
        E T5G, T3h, T6H, T6O, T7o, T4L, T5N, T52, T5Q, T1i, T7V, T6i, T7D, T3K, T5u;
1161
72
        E T3P, T5v, T1E, T6n, T6m, T7e, T3W, T5y, T41, T5z, T29, T6p, T6s, T7f, T47;
1162
72
        E T5B, T4c, T5C, T2R, T6z, T6E, T7k, T4v, T5H, T4E, T5K, T3y, T6P, T6K, T7p;
1163
72
        E T4W, T5R, T55, T5O;
1164
72
        {
1165
72
       E T1, T7G, Tq, T7F, TA, T3C, TJ, T3D, Tn, Tp;
1166
72
       T1 = ri[0];
1167
72
       T7G = ii[0];
1168
72
       Tn = ri[WS(rs, 16)];
1169
72
       Tp = ii[WS(rs, 16)];
1170
72
       Tq = FMA(Tm, Tn, To * Tp);
1171
72
       T7F = FNMS(To, Tn, Tm * Tp);
1172
72
       {
1173
72
            E Tv, Tz, TE, TI;
1174
72
            Tv = ri[WS(rs, 8)];
1175
72
            Tz = ii[WS(rs, 8)];
1176
72
            TA = FMA(Tu, Tv, Ty * Tz);
1177
72
            T3C = FNMS(Ty, Tv, Tu * Tz);
1178
72
            TE = ri[WS(rs, 24)];
1179
72
            TI = ii[WS(rs, 24)];
1180
72
            TJ = FMA(TD, TE, TH * TI);
1181
72
            T3D = FNMS(TH, TE, TD * TI);
1182
72
       }
1183
72
       {
1184
72
            E Tr, TK, T8a, T8b;
1185
72
            Tr = T1 + Tq;
1186
72
            TK = TA + TJ;
1187
72
            TL = Tr + TK;
1188
72
            T6f = Tr - TK;
1189
72
            T8a = T7G - T7F;
1190
72
            T8b = TA - TJ;
1191
72
            T8c = T8a - T8b;
1192
72
            T8q = T8b + T8a;
1193
72
       }
1194
72
       {
1195
72
            E T3B, T3E, T7E, T7H;
1196
72
            T3B = T1 - Tq;
1197
72
            T3E = T3C - T3D;
1198
72
            T3F = T3B - T3E;
1199
72
            T5t = T3B + T3E;
1200
72
            T7E = T3C + T3D;
1201
72
            T7H = T7F + T7G;
1202
72
            T7I = T7E + T7H;
1203
72
            T7W = T7H - T7E;
1204
72
       }
1205
72
        }
1206
72
        {
1207
72
       E T2e, T4g, T2w, T4z, T2j, T4h, T2n, T4y;
1208
72
       {
1209
72
            E T2c, T2d, T2r, T2v;
1210
72
            T2c = ri[WS(rs, 1)];
1211
72
            T2d = ii[WS(rs, 1)];
1212
72
            T2e = FMA(T2, T2c, T5 * T2d);
1213
72
            T4g = FNMS(T5, T2c, T2 * T2d);
1214
72
            T2r = ri[WS(rs, 25)];
1215
72
            T2v = ii[WS(rs, 25)];
1216
72
            T2w = FMA(T2q, T2r, T2u * T2v);
1217
72
            T4z = FNMS(T2u, T2r, T2q * T2v);
1218
72
       }
1219
72
       {
1220
72
            E T2g, T2i, T2l, T2m;
1221
72
            T2g = ri[WS(rs, 17)];
1222
72
            T2i = ii[WS(rs, 17)];
1223
72
            T2j = FMA(T2f, T2g, T2h * T2i);
1224
72
            T4h = FNMS(T2h, T2g, T2f * T2i);
1225
72
            T2l = ri[WS(rs, 9)];
1226
72
            T2m = ii[WS(rs, 9)];
1227
72
            T2n = FMA(T9, T2l, Te * T2m);
1228
72
            T4y = FNMS(Te, T2l, T9 * T2m);
1229
72
       }
1230
72
       {
1231
72
            E T2k, T2x, T6w, T6x;
1232
72
            T2k = T2e + T2j;
1233
72
            T2x = T2n + T2w;
1234
72
            T2y = T2k + T2x;
1235
72
            T6B = T2k - T2x;
1236
72
            T6w = T4g + T4h;
1237
72
            T6x = T4y + T4z;
1238
72
            T6y = T6w - T6x;
1239
72
            T7j = T6w + T6x;
1240
72
       }
1241
72
       {
1242
72
            E T4i, T4j, T4x, T4A;
1243
72
            T4i = T4g - T4h;
1244
72
            T4j = T2n - T2w;
1245
72
            T4k = T4i + T4j;
1246
72
            T5J = T4i - T4j;
1247
72
            T4x = T2e - T2j;
1248
72
            T4A = T4y - T4z;
1249
72
            T4B = T4x - T4A;
1250
72
            T5G = T4x + T4A;
1251
72
       }
1252
72
        }
1253
72
        {
1254
72
       E T31, T4Y, T3f, T4J, T36, T4Z, T3a, T4I;
1255
72
       {
1256
72
            E T2W, T30, T3c, T3e;
1257
72
            T2W = ri[WS(rs, 31)];
1258
72
            T30 = ii[WS(rs, 31)];
1259
72
            T31 = FMA(T2V, T2W, T2Z * T30);
1260
72
            T4Y = FNMS(T2Z, T2W, T2V * T30);
1261
72
            T3c = ri[WS(rs, 23)];
1262
72
            T3e = ii[WS(rs, 23)];
1263
72
            T3f = FMA(T3b, T3c, T3d * T3e);
1264
72
            T4J = FNMS(T3d, T3c, T3b * T3e);
1265
72
       }
1266
72
       {
1267
72
            E T33, T35, T38, T39;
1268
72
            T33 = ri[WS(rs, 15)];
1269
72
            T35 = ii[WS(rs, 15)];
1270
72
            T36 = FMA(T32, T33, T34 * T35);
1271
72
            T4Z = FNMS(T34, T33, T32 * T35);
1272
72
            T38 = ri[WS(rs, 7)];
1273
72
            T39 = ii[WS(rs, 7)];
1274
72
            T3a = FMA(TR, T38, TS * T39);
1275
72
            T4I = FNMS(TS, T38, TR * T39);
1276
72
       }
1277
72
       {
1278
72
            E T37, T3g, T6M, T6N;
1279
72
            T37 = T31 + T36;
1280
72
            T3g = T3a + T3f;
1281
72
            T3h = T37 + T3g;
1282
72
            T6H = T37 - T3g;
1283
72
            T6M = T4Y + T4Z;
1284
72
            T6N = T4I + T4J;
1285
72
            T6O = T6M - T6N;
1286
72
            T7o = T6M + T6N;
1287
72
       }
1288
72
       {
1289
72
            E T4H, T4K, T50, T51;
1290
72
            T4H = T31 - T36;
1291
72
            T4K = T4I - T4J;
1292
72
            T4L = T4H - T4K;
1293
72
            T5N = T4H + T4K;
1294
72
            T50 = T4Y - T4Z;
1295
72
            T51 = T3a - T3f;
1296
72
            T52 = T50 + T51;
1297
72
            T5Q = T50 - T51;
1298
72
       }
1299
72
        }
1300
72
        {
1301
72
       E TQ, T3G, T1g, T3N, TX, T3H, T17, T3M;
1302
72
       {
1303
72
            E TN, TP, T1b, T1f;
1304
72
            TN = ri[WS(rs, 4)];
1305
72
            TP = ii[WS(rs, 4)];
1306
72
            TQ = FMA(TM, TN, TO * TP);
1307
72
            T3G = FNMS(TO, TN, TM * TP);
1308
72
            T1b = ri[WS(rs, 12)];
1309
72
            T1f = ii[WS(rs, 12)];
1310
72
            T1g = FMA(T1a, T1b, T1e * T1f);
1311
72
            T3N = FNMS(T1e, T1b, T1a * T1f);
1312
72
       }
1313
72
       {
1314
72
            E TU, TW, T12, T16;
1315
72
            TU = ri[WS(rs, 20)];
1316
72
            TW = ii[WS(rs, 20)];
1317
72
            TX = FMA(TT, TU, TV * TW);
1318
72
            T3H = FNMS(TV, TU, TT * TW);
1319
72
            T12 = ri[WS(rs, 28)];
1320
72
            T16 = ii[WS(rs, 28)];
1321
72
            T17 = FMA(T11, T12, T15 * T16);
1322
72
            T3M = FNMS(T15, T12, T11 * T16);
1323
72
       }
1324
72
       {
1325
72
            E TY, T1h, T6g, T6h;
1326
72
            TY = TQ + TX;
1327
72
            T1h = T17 + T1g;
1328
72
            T1i = TY + T1h;
1329
72
            T7V = T1h - TY;
1330
72
            T6g = T3G + T3H;
1331
72
            T6h = T3M + T3N;
1332
72
            T6i = T6g - T6h;
1333
72
            T7D = T6g + T6h;
1334
72
       }
1335
72
       {
1336
72
            E T3I, T3J, T3L, T3O;
1337
72
            T3I = T3G - T3H;
1338
72
            T3J = TQ - TX;
1339
72
            T3K = T3I - T3J;
1340
72
            T5u = T3J + T3I;
1341
72
            T3L = T17 - T1g;
1342
72
            T3O = T3M - T3N;
1343
72
            T3P = T3L + T3O;
1344
72
            T5v = T3L - T3O;
1345
72
       }
1346
72
        }
1347
72
        {
1348
72
       E T1m, T3S, T1C, T3Z, T1r, T3T, T1x, T3Y;
1349
72
       {
1350
72
            E T1k, T1l, T1z, T1B;
1351
72
            T1k = ri[WS(rs, 2)];
1352
72
            T1l = ii[WS(rs, 2)];
1353
72
            T1m = FMA(T8, T1k, Td * T1l);
1354
72
            T3S = FNMS(Td, T1k, T8 * T1l);
1355
72
            T1z = ri[WS(rs, 26)];
1356
72
            T1B = ii[WS(rs, 26)];
1357
72
            T1C = FMA(T1y, T1z, T1A * T1B);
1358
72
            T3Z = FNMS(T1A, T1z, T1y * T1B);
1359
72
       }
1360
72
       {
1361
72
            E T1o, T1q, T1u, T1w;
1362
72
            T1o = ri[WS(rs, 18)];
1363
72
            T1q = ii[WS(rs, 18)];
1364
72
            T1r = FMA(T1n, T1o, T1p * T1q);
1365
72
            T3T = FNMS(T1p, T1o, T1n * T1q);
1366
72
            T1u = ri[WS(rs, 10)];
1367
72
            T1w = ii[WS(rs, 10)];
1368
72
            T1x = FMA(T1t, T1u, T1v * T1w);
1369
72
            T3Y = FNMS(T1v, T1u, T1t * T1w);
1370
72
       }
1371
72
       {
1372
72
            E T1s, T1D, T6k, T6l;
1373
72
            T1s = T1m + T1r;
1374
72
            T1D = T1x + T1C;
1375
72
            T1E = T1s + T1D;
1376
72
            T6n = T1s - T1D;
1377
72
            T6k = T3S + T3T;
1378
72
            T6l = T3Y + T3Z;
1379
72
            T6m = T6k - T6l;
1380
72
            T7e = T6k + T6l;
1381
72
       }
1382
72
       {
1383
72
            E T3U, T3V, T3X, T40;
1384
72
            T3U = T3S - T3T;
1385
72
            T3V = T1x - T1C;
1386
72
            T3W = T3U + T3V;
1387
72
            T5y = T3U - T3V;
1388
72
            T3X = T1m - T1r;
1389
72
            T40 = T3Y - T3Z;
1390
72
            T41 = T3X - T40;
1391
72
            T5z = T3X + T40;
1392
72
       }
1393
72
        }
1394
72
        {
1395
72
       E T1J, T43, T27, T4a, T1U, T44, T20, T49;
1396
72
       {
1397
72
            E T1G, T1I, T24, T26;
1398
72
            T1G = ri[WS(rs, 30)];
1399
72
            T1I = ii[WS(rs, 30)];
1400
72
            T1J = FMA(T1F, T1G, T1H * T1I);
1401
72
            T43 = FNMS(T1H, T1G, T1F * T1I);
1402
72
            T24 = ri[WS(rs, 22)];
1403
72
            T26 = ii[WS(rs, 22)];
1404
72
            T27 = FMA(T23, T24, T25 * T26);
1405
72
            T4a = FNMS(T25, T24, T23 * T26);
1406
72
       }
1407
72
       {
1408
72
            E T1R, T1T, T1X, T1Z;
1409
72
            T1R = ri[WS(rs, 14)];
1410
72
            T1T = ii[WS(rs, 14)];
1411
72
            T1U = FMA(T1Q, T1R, T1S * T1T);
1412
72
            T44 = FNMS(T1S, T1R, T1Q * T1T);
1413
72
            T1X = ri[WS(rs, 6)];
1414
72
            T1Z = ii[WS(rs, 6)];
1415
72
            T20 = FMA(T1W, T1X, T1Y * T1Z);
1416
72
            T49 = FNMS(T1Y, T1X, T1W * T1Z);
1417
72
       }
1418
72
       {
1419
72
            E T1V, T28, T6q, T6r;
1420
72
            T1V = T1J + T1U;
1421
72
            T28 = T20 + T27;
1422
72
            T29 = T1V + T28;
1423
72
            T6p = T1V - T28;
1424
72
            T6q = T43 + T44;
1425
72
            T6r = T49 + T4a;
1426
72
            T6s = T6q - T6r;
1427
72
            T7f = T6q + T6r;
1428
72
       }
1429
72
       {
1430
72
            E T45, T46, T48, T4b;
1431
72
            T45 = T43 - T44;
1432
72
            T46 = T20 - T27;
1433
72
            T47 = T45 + T46;
1434
72
            T5B = T45 - T46;
1435
72
            T48 = T1J - T1U;
1436
72
            T4b = T49 - T4a;
1437
72
            T4c = T48 - T4b;
1438
72
            T5C = T48 + T4b;
1439
72
       }
1440
72
        }
1441
72
        {
1442
72
       E T2B, T4r, T2G, T4s, T4q, T4t, T2M, T4m, T2P, T4n, T4l, T4o;
1443
72
       {
1444
72
            E T2z, T2A, T2D, T2F;
1445
72
            T2z = ri[WS(rs, 5)];
1446
72
            T2A = ii[WS(rs, 5)];
1447
72
            T2B = FMA(T21, T2z, T22 * T2A);
1448
72
            T4r = FNMS(T22, T2z, T21 * T2A);
1449
72
            T2D = ri[WS(rs, 21)];
1450
72
            T2F = ii[WS(rs, 21)];
1451
72
            T2G = FMA(T2C, T2D, T2E * T2F);
1452
72
            T4s = FNMS(T2E, T2D, T2C * T2F);
1453
72
       }
1454
72
       T4q = T2B - T2G;
1455
72
       T4t = T4r - T4s;
1456
72
       {
1457
72
            E T2J, T2L, T2N, T2O;
1458
72
            T2J = ri[WS(rs, 29)];
1459
72
            T2L = ii[WS(rs, 29)];
1460
72
            T2M = FMA(T2I, T2J, T2K * T2L);
1461
72
            T4m = FNMS(T2K, T2J, T2I * T2L);
1462
72
            T2N = ri[WS(rs, 13)];
1463
72
            T2O = ii[WS(rs, 13)];
1464
72
            T2P = FMA(T1M, T2N, T1P * T2O);
1465
72
            T4n = FNMS(T1P, T2N, T1M * T2O);
1466
72
       }
1467
72
       T4l = T2M - T2P;
1468
72
       T4o = T4m - T4n;
1469
72
       {
1470
72
            E T2H, T2Q, T6C, T6D;
1471
72
            T2H = T2B + T2G;
1472
72
            T2Q = T2M + T2P;
1473
72
            T2R = T2H + T2Q;
1474
72
            T6z = T2Q - T2H;
1475
72
            T6C = T4r + T4s;
1476
72
            T6D = T4m + T4n;
1477
72
            T6E = T6C - T6D;
1478
72
            T7k = T6C + T6D;
1479
72
       }
1480
72
       {
1481
72
            E T4p, T4u, T4C, T4D;
1482
72
            T4p = T4l - T4o;
1483
72
            T4u = T4q + T4t;
1484
72
            T4v = KP707106781 * (T4p - T4u);
1485
72
            T5H = KP707106781 * (T4u + T4p);
1486
72
            T4C = T4t - T4q;
1487
72
            T4D = T4l + T4o;
1488
72
            T4E = KP707106781 * (T4C - T4D);
1489
72
            T5K = KP707106781 * (T4C + T4D);
1490
72
       }
1491
72
        }
1492
72
        {
1493
72
       E T3k, T4M, T3p, T4N, T4O, T4P, T3t, T4S, T3w, T4T, T4R, T4U;
1494
72
       {
1495
72
            E T3i, T3j, T3m, T3o;
1496
72
            T3i = ri[WS(rs, 3)];
1497
72
            T3j = ii[WS(rs, 3)];
1498
72
            T3k = FMA(T3, T3i, T6 * T3j);
1499
72
            T4M = FNMS(T6, T3i, T3 * T3j);
1500
72
            T3m = ri[WS(rs, 19)];
1501
72
            T3o = ii[WS(rs, 19)];
1502
72
            T3p = FMA(T3l, T3m, T3n * T3o);
1503
72
            T4N = FNMS(T3n, T3m, T3l * T3o);
1504
72
       }
1505
72
       T4O = T4M - T4N;
1506
72
       T4P = T3k - T3p;
1507
72
       {
1508
72
            E T3r, T3s, T3u, T3v;
1509
72
            T3r = ri[WS(rs, 27)];
1510
72
            T3s = ii[WS(rs, 27)];
1511
72
            T3t = FMA(Th, T3r, Tl * T3s);
1512
72
            T4S = FNMS(Tl, T3r, Th * T3s);
1513
72
            T3u = ri[WS(rs, 11)];
1514
72
            T3v = ii[WS(rs, 11)];
1515
72
            T3w = FMA(Tg, T3u, Tk * T3v);
1516
72
            T4T = FNMS(Tk, T3u, Tg * T3v);
1517
72
       }
1518
72
       T4R = T3t - T3w;
1519
72
       T4U = T4S - T4T;
1520
72
       {
1521
72
            E T3q, T3x, T6I, T6J;
1522
72
            T3q = T3k + T3p;
1523
72
            T3x = T3t + T3w;
1524
72
            T3y = T3q + T3x;
1525
72
            T6P = T3x - T3q;
1526
72
            T6I = T4M + T4N;
1527
72
            T6J = T4S + T4T;
1528
72
            T6K = T6I - T6J;
1529
72
            T7p = T6I + T6J;
1530
72
       }
1531
72
       {
1532
72
            E T4Q, T4V, T53, T54;
1533
72
            T4Q = T4O - T4P;
1534
72
            T4V = T4R + T4U;
1535
72
            T4W = KP707106781 * (T4Q - T4V);
1536
72
            T5R = KP707106781 * (T4Q + T4V);
1537
72
            T53 = T4R - T4U;
1538
72
            T54 = T4P + T4O;
1539
72
            T55 = KP707106781 * (T53 - T54);
1540
72
            T5O = KP707106781 * (T54 + T53);
1541
72
       }
1542
72
        }
1543
72
        {
1544
72
       E T2b, T7x, T7K, T7M, T3A, T7L, T7A, T7B;
1545
72
       {
1546
72
            E T1j, T2a, T7C, T7J;
1547
72
            T1j = TL + T1i;
1548
72
            T2a = T1E + T29;
1549
72
            T2b = T1j + T2a;
1550
72
            T7x = T1j - T2a;
1551
72
            T7C = T7e + T7f;
1552
72
            T7J = T7D + T7I;
1553
72
            T7K = T7C + T7J;
1554
72
            T7M = T7J - T7C;
1555
72
       }
1556
72
       {
1557
72
            E T2S, T3z, T7y, T7z;
1558
72
            T2S = T2y + T2R;
1559
72
            T3z = T3h + T3y;
1560
72
            T3A = T2S + T3z;
1561
72
            T7L = T3z - T2S;
1562
72
            T7y = T7j + T7k;
1563
72
            T7z = T7o + T7p;
1564
72
            T7A = T7y - T7z;
1565
72
            T7B = T7y + T7z;
1566
72
       }
1567
72
       ri[WS(rs, 16)] = T2b - T3A;
1568
72
       ii[WS(rs, 16)] = T7K - T7B;
1569
72
       ri[0] = T2b + T3A;
1570
72
       ii[0] = T7B + T7K;
1571
72
       ri[WS(rs, 24)] = T7x - T7A;
1572
72
       ii[WS(rs, 24)] = T7M - T7L;
1573
72
       ri[WS(rs, 8)] = T7x + T7A;
1574
72
       ii[WS(rs, 8)] = T7L + T7M;
1575
72
        }
1576
72
        {
1577
72
       E T7h, T7t, T7Q, T7S, T7m, T7u, T7r, T7v;
1578
72
       {
1579
72
            E T7d, T7g, T7O, T7P;
1580
72
            T7d = TL - T1i;
1581
72
            T7g = T7e - T7f;
1582
72
            T7h = T7d + T7g;
1583
72
            T7t = T7d - T7g;
1584
72
            T7O = T29 - T1E;
1585
72
            T7P = T7I - T7D;
1586
72
            T7Q = T7O + T7P;
1587
72
            T7S = T7P - T7O;
1588
72
       }
1589
72
       {
1590
72
            E T7i, T7l, T7n, T7q;
1591
72
            T7i = T2y - T2R;
1592
72
            T7l = T7j - T7k;
1593
72
            T7m = T7i + T7l;
1594
72
            T7u = T7l - T7i;
1595
72
            T7n = T3h - T3y;
1596
72
            T7q = T7o - T7p;
1597
72
            T7r = T7n - T7q;
1598
72
            T7v = T7n + T7q;
1599
72
       }
1600
72
       {
1601
72
            E T7s, T7N, T7w, T7R;
1602
72
            T7s = KP707106781 * (T7m + T7r);
1603
72
            ri[WS(rs, 20)] = T7h - T7s;
1604
72
            ri[WS(rs, 4)] = T7h + T7s;
1605
72
            T7N = KP707106781 * (T7u + T7v);
1606
72
            ii[WS(rs, 4)] = T7N + T7Q;
1607
72
            ii[WS(rs, 20)] = T7Q - T7N;
1608
72
            T7w = KP707106781 * (T7u - T7v);
1609
72
            ri[WS(rs, 28)] = T7t - T7w;
1610
72
            ri[WS(rs, 12)] = T7t + T7w;
1611
72
            T7R = KP707106781 * (T7r - T7m);
1612
72
            ii[WS(rs, 12)] = T7R + T7S;
1613
72
            ii[WS(rs, 28)] = T7S - T7R;
1614
72
       }
1615
72
        }
1616
72
        {
1617
72
       E T6j, T7X, T83, T6X, T6u, T7U, T77, T7b, T70, T82, T6G, T6U, T74, T7a, T6R;
1618
72
       E T6V;
1619
72
       {
1620
72
            E T6o, T6t, T6A, T6F;
1621
72
            T6j = T6f - T6i;
1622
72
            T7X = T7V + T7W;
1623
72
            T83 = T7W - T7V;
1624
72
            T6X = T6f + T6i;
1625
72
            T6o = T6m - T6n;
1626
72
            T6t = T6p + T6s;
1627
72
            T6u = KP707106781 * (T6o - T6t);
1628
72
            T7U = KP707106781 * (T6o + T6t);
1629
72
            {
1630
72
           E T75, T76, T6Y, T6Z;
1631
72
           T75 = T6H + T6K;
1632
72
           T76 = T6O + T6P;
1633
72
           T77 = FNMS(KP382683432, T76, KP923879532 * T75);
1634
72
           T7b = FMA(KP923879532, T76, KP382683432 * T75);
1635
72
           T6Y = T6n + T6m;
1636
72
           T6Z = T6p - T6s;
1637
72
           T70 = KP707106781 * (T6Y + T6Z);
1638
72
           T82 = KP707106781 * (T6Z - T6Y);
1639
72
            }
1640
72
            T6A = T6y - T6z;
1641
72
            T6F = T6B - T6E;
1642
72
            T6G = FMA(KP923879532, T6A, KP382683432 * T6F);
1643
72
            T6U = FNMS(KP923879532, T6F, KP382683432 * T6A);
1644
72
            {
1645
72
           E T72, T73, T6L, T6Q;
1646
72
           T72 = T6y + T6z;
1647
72
           T73 = T6B + T6E;
1648
72
           T74 = FMA(KP382683432, T72, KP923879532 * T73);
1649
72
           T7a = FNMS(KP382683432, T73, KP923879532 * T72);
1650
72
           T6L = T6H - T6K;
1651
72
           T6Q = T6O - T6P;
1652
72
           T6R = FNMS(KP923879532, T6Q, KP382683432 * T6L);
1653
72
           T6V = FMA(KP382683432, T6Q, KP923879532 * T6L);
1654
72
            }
1655
72
       }
1656
72
       {
1657
72
            E T6v, T6S, T81, T84;
1658
72
            T6v = T6j + T6u;
1659
72
            T6S = T6G + T6R;
1660
72
            ri[WS(rs, 22)] = T6v - T6S;
1661
72
            ri[WS(rs, 6)] = T6v + T6S;
1662
72
            T81 = T6U + T6V;
1663
72
            T84 = T82 + T83;
1664
72
            ii[WS(rs, 6)] = T81 + T84;
1665
72
            ii[WS(rs, 22)] = T84 - T81;
1666
72
       }
1667
72
       {
1668
72
            E T6T, T6W, T85, T86;
1669
72
            T6T = T6j - T6u;
1670
72
            T6W = T6U - T6V;
1671
72
            ri[WS(rs, 30)] = T6T - T6W;
1672
72
            ri[WS(rs, 14)] = T6T + T6W;
1673
72
            T85 = T6R - T6G;
1674
72
            T86 = T83 - T82;
1675
72
            ii[WS(rs, 14)] = T85 + T86;
1676
72
            ii[WS(rs, 30)] = T86 - T85;
1677
72
       }
1678
72
       {
1679
72
            E T71, T78, T7T, T7Y;
1680
72
            T71 = T6X + T70;
1681
72
            T78 = T74 + T77;
1682
72
            ri[WS(rs, 18)] = T71 - T78;
1683
72
            ri[WS(rs, 2)] = T71 + T78;
1684
72
            T7T = T7a + T7b;
1685
72
            T7Y = T7U + T7X;
1686
72
            ii[WS(rs, 2)] = T7T + T7Y;
1687
72
            ii[WS(rs, 18)] = T7Y - T7T;
1688
72
       }
1689
72
       {
1690
72
            E T79, T7c, T7Z, T80;
1691
72
            T79 = T6X - T70;
1692
72
            T7c = T7a - T7b;
1693
72
            ri[WS(rs, 26)] = T79 - T7c;
1694
72
            ri[WS(rs, 10)] = T79 + T7c;
1695
72
            T7Z = T77 - T74;
1696
72
            T80 = T7X - T7U;
1697
72
            ii[WS(rs, 10)] = T7Z + T80;
1698
72
            ii[WS(rs, 26)] = T80 - T7Z;
1699
72
       }
1700
72
        }
1701
72
        {
1702
72
       E T3R, T5d, T8r, T8x, T4e, T8o, T5n, T5r, T4G, T5a, T5g, T8w, T5k, T5q, T57;
1703
72
       E T5b, T3Q, T8p;
1704
72
       T3Q = KP707106781 * (T3K - T3P);
1705
72
       T3R = T3F - T3Q;
1706
72
       T5d = T3F + T3Q;
1707
72
       T8p = KP707106781 * (T5v - T5u);
1708
72
       T8r = T8p + T8q;
1709
72
       T8x = T8q - T8p;
1710
72
       {
1711
72
            E T42, T4d, T5l, T5m;
1712
72
            T42 = FNMS(KP923879532, T41, KP382683432 * T3W);
1713
72
            T4d = FMA(KP382683432, T47, KP923879532 * T4c);
1714
72
            T4e = T42 - T4d;
1715
72
            T8o = T42 + T4d;
1716
72
            T5l = T4L + T4W;
1717
72
            T5m = T52 + T55;
1718
72
            T5n = FNMS(KP555570233, T5m, KP831469612 * T5l);
1719
72
            T5r = FMA(KP831469612, T5m, KP555570233 * T5l);
1720
72
       }
1721
72
       {
1722
72
            E T4w, T4F, T5e, T5f;
1723
72
            T4w = T4k - T4v;
1724
72
            T4F = T4B - T4E;
1725
72
            T4G = FMA(KP980785280, T4w, KP195090322 * T4F);
1726
72
            T5a = FNMS(KP980785280, T4F, KP195090322 * T4w);
1727
72
            T5e = FMA(KP923879532, T3W, KP382683432 * T41);
1728
72
            T5f = FNMS(KP923879532, T47, KP382683432 * T4c);
1729
72
            T5g = T5e + T5f;
1730
72
            T8w = T5f - T5e;
1731
72
       }
1732
72
       {
1733
72
            E T5i, T5j, T4X, T56;
1734
72
            T5i = T4k + T4v;
1735
72
            T5j = T4B + T4E;
1736
72
            T5k = FMA(KP555570233, T5i, KP831469612 * T5j);
1737
72
            T5q = FNMS(KP555570233, T5j, KP831469612 * T5i);
1738
72
            T4X = T4L - T4W;
1739
72
            T56 = T52 - T55;
1740
72
            T57 = FNMS(KP980785280, T56, KP195090322 * T4X);
1741
72
            T5b = FMA(KP195090322, T56, KP980785280 * T4X);
1742
72
       }
1743
72
       {
1744
72
            E T4f, T58, T8v, T8y;
1745
72
            T4f = T3R + T4e;
1746
72
            T58 = T4G + T57;
1747
72
            ri[WS(rs, 23)] = T4f - T58;
1748
72
            ri[WS(rs, 7)] = T4f + T58;
1749
72
            T8v = T5a + T5b;
1750
72
            T8y = T8w + T8x;
1751
72
            ii[WS(rs, 7)] = T8v + T8y;
1752
72
            ii[WS(rs, 23)] = T8y - T8v;
1753
72
       }
1754
72
       {
1755
72
            E T59, T5c, T8z, T8A;
1756
72
            T59 = T3R - T4e;
1757
72
            T5c = T5a - T5b;
1758
72
            ri[WS(rs, 31)] = T59 - T5c;
1759
72
            ri[WS(rs, 15)] = T59 + T5c;
1760
72
            T8z = T57 - T4G;
1761
72
            T8A = T8x - T8w;
1762
72
            ii[WS(rs, 15)] = T8z + T8A;
1763
72
            ii[WS(rs, 31)] = T8A - T8z;
1764
72
       }
1765
72
       {
1766
72
            E T5h, T5o, T8n, T8s;
1767
72
            T5h = T5d + T5g;
1768
72
            T5o = T5k + T5n;
1769
72
            ri[WS(rs, 19)] = T5h - T5o;
1770
72
            ri[WS(rs, 3)] = T5h + T5o;
1771
72
            T8n = T5q + T5r;
1772
72
            T8s = T8o + T8r;
1773
72
            ii[WS(rs, 3)] = T8n + T8s;
1774
72
            ii[WS(rs, 19)] = T8s - T8n;
1775
72
       }
1776
72
       {
1777
72
            E T5p, T5s, T8t, T8u;
1778
72
            T5p = T5d - T5g;
1779
72
            T5s = T5q - T5r;
1780
72
            ri[WS(rs, 27)] = T5p - T5s;
1781
72
            ri[WS(rs, 11)] = T5p + T5s;
1782
72
            T8t = T5n - T5k;
1783
72
            T8u = T8r - T8o;
1784
72
            ii[WS(rs, 11)] = T8t + T8u;
1785
72
            ii[WS(rs, 27)] = T8u - T8t;
1786
72
       }
1787
72
        }
1788
72
        {
1789
72
       E T5x, T5Z, T8d, T8j, T5E, T88, T69, T6d, T5M, T5W, T62, T8i, T66, T6c, T5T;
1790
72
       E T5X, T5w, T89;
1791
72
       T5w = KP707106781 * (T5u + T5v);
1792
72
       T5x = T5t - T5w;
1793
72
       T5Z = T5t + T5w;
1794
72
       T89 = KP707106781 * (T3K + T3P);
1795
72
       T8d = T89 + T8c;
1796
72
       T8j = T8c - T89;
1797
72
       {
1798
72
            E T5A, T5D, T67, T68;
1799
72
            T5A = FNMS(KP382683432, T5z, KP923879532 * T5y);
1800
72
            T5D = FMA(KP923879532, T5B, KP382683432 * T5C);
1801
72
            T5E = T5A - T5D;
1802
72
            T88 = T5A + T5D;
1803
72
            T67 = T5N + T5O;
1804
72
            T68 = T5Q + T5R;
1805
72
            T69 = FNMS(KP195090322, T68, KP980785280 * T67);
1806
72
            T6d = FMA(KP195090322, T67, KP980785280 * T68);
1807
72
       }
1808
72
       {
1809
72
            E T5I, T5L, T60, T61;
1810
72
            T5I = T5G - T5H;
1811
72
            T5L = T5J - T5K;
1812
72
            T5M = FMA(KP555570233, T5I, KP831469612 * T5L);
1813
72
            T5W = FNMS(KP831469612, T5I, KP555570233 * T5L);
1814
72
            T60 = FMA(KP382683432, T5y, KP923879532 * T5z);
1815
72
            T61 = FNMS(KP382683432, T5B, KP923879532 * T5C);
1816
72
            T62 = T60 + T61;
1817
72
            T8i = T61 - T60;
1818
72
       }
1819
72
       {
1820
72
            E T64, T65, T5P, T5S;
1821
72
            T64 = T5G + T5H;
1822
72
            T65 = T5J + T5K;
1823
72
            T66 = FMA(KP980785280, T64, KP195090322 * T65);
1824
72
            T6c = FNMS(KP195090322, T64, KP980785280 * T65);
1825
72
            T5P = T5N - T5O;
1826
72
            T5S = T5Q - T5R;
1827
72
            T5T = FNMS(KP831469612, T5S, KP555570233 * T5P);
1828
72
            T5X = FMA(KP831469612, T5P, KP555570233 * T5S);
1829
72
       }
1830
72
       {
1831
72
            E T5F, T5U, T8h, T8k;
1832
72
            T5F = T5x + T5E;
1833
72
            T5U = T5M + T5T;
1834
72
            ri[WS(rs, 21)] = T5F - T5U;
1835
72
            ri[WS(rs, 5)] = T5F + T5U;
1836
72
            T8h = T5W + T5X;
1837
72
            T8k = T8i + T8j;
1838
72
            ii[WS(rs, 5)] = T8h + T8k;
1839
72
            ii[WS(rs, 21)] = T8k - T8h;
1840
72
       }
1841
72
       {
1842
72
            E T5V, T5Y, T8l, T8m;
1843
72
            T5V = T5x - T5E;
1844
72
            T5Y = T5W - T5X;
1845
72
            ri[WS(rs, 29)] = T5V - T5Y;
1846
72
            ri[WS(rs, 13)] = T5V + T5Y;
1847
72
            T8l = T5T - T5M;
1848
72
            T8m = T8j - T8i;
1849
72
            ii[WS(rs, 13)] = T8l + T8m;
1850
72
            ii[WS(rs, 29)] = T8m - T8l;
1851
72
       }
1852
72
       {
1853
72
            E T63, T6a, T87, T8e;
1854
72
            T63 = T5Z + T62;
1855
72
            T6a = T66 + T69;
1856
72
            ri[WS(rs, 17)] = T63 - T6a;
1857
72
            ri[WS(rs, 1)] = T63 + T6a;
1858
72
            T87 = T6c + T6d;
1859
72
            T8e = T88 + T8d;
1860
72
            ii[WS(rs, 1)] = T87 + T8e;
1861
72
            ii[WS(rs, 17)] = T8e - T87;
1862
72
       }
1863
72
       {
1864
72
            E T6b, T6e, T8f, T8g;
1865
72
            T6b = T5Z - T62;
1866
72
            T6e = T6c - T6d;
1867
72
            ri[WS(rs, 25)] = T6b - T6e;
1868
72
            ri[WS(rs, 9)] = T6b + T6e;
1869
72
            T8f = T69 - T66;
1870
72
            T8g = T8d - T88;
1871
72
            ii[WS(rs, 9)] = T8f + T8g;
1872
72
            ii[WS(rs, 25)] = T8g - T8f;
1873
72
       }
1874
72
        }
1875
72
         }
1876
72
    }
1877
6
     }
1878
6
}
1879
1880
static const tw_instr twinstr[] = {
1881
     { TW_CEXP, 0, 1 },
1882
     { TW_CEXP, 0, 3 },
1883
     { TW_CEXP, 0, 9 },
1884
     { TW_CEXP, 0, 27 },
1885
     { TW_NEXT, 1, 0 }
1886
};
1887
1888
static const ct_desc desc = { 32, "t2_32", twinstr, &GENUS, { 376, 168, 112, 0 }, 0, 0, 0 };
1889
1890
1
void X(codelet_t2_32) (planner *p) {
1891
1
     X(kdft_dit_register) (p, t2_32, &desc);
1892
1
}
1893
#endif