Coverage Report

Created: 2024-09-08 06:43

/src/fftw3/rdft/scalar/r2cf/hf2_32.c
Line
Count
Source (jump to first uncovered line)
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 Sun Sep  8 06:41:22 UTC 2024 */
23
24
#include "rdft/codelet-rdft.h"
25
26
#if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27
28
/* Generated by: ../../../genfft/gen_hc2hc.native -fma -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 32 -dit -name hf2_32 -include rdft/scalar/hf.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 "rdft/scalar/hf.h"
36
37
static void hf2_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
38
{
39
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
40
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
41
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
42
     DK(KP198912367, +0.198912367379658006911597622644676228597850501);
43
     DK(KP668178637, +0.668178637919298919997757686523080761552472251);
44
     DK(KP414213562, +0.414213562373095048801688724209698078569671875);
45
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
46
     {
47
    INT m;
48
    for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - 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, T98, TH, T97, T4b, T8D, TZ, T7g, T4j, T6t, T1g, T7f, T4q;
175
        E T6u, T1v, T1I, T7j, T7k, T7l, T7m, T4z, T6y, T4G, T6x, T22, T2j, T7o, T7p;
176
        E T7q, T7r, T4O, T6B, T4V, T6A, T3G, T7G, T7N, T8n, T5E, T6M, T61, T6P, T2N;
177
        E T7v, T7C, T8i, T55, T6F, T5s, T6I, T43, T7O, T7J, T8o, T5L, T63, T5S, T62;
178
        E T3c, T7D, T7y, T8j, T5c, T5t, T5j, T5u;
179
        {
180
       E T1, T8G, Tk, Tl, To, T8E, Tp, T8F;
181
       T1 = cr[0];
182
       T8G = ci[0];
183
       Tk = cr[WS(rs, 16)];
184
       Tl = Tj * Tk;
185
       To = ci[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
       T98 = T8G - T8F;
193
        }
194
        {
195
       E Tt, Tu, Tx, T47, TB, TC, TF, T49;
196
       Tt = cr[WS(rs, 8)];
197
       Tu = Ts * Tt;
198
       Tx = ci[WS(rs, 8)];
199
       T47 = Ts * Tx;
200
       TB = cr[WS(rs, 24)];
201
       TC = TA * TB;
202
       TF = ci[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
            T97 = 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 = cr[WS(rs, 4)];
221
            TL = TJ * TK;
222
            TN = ci[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 = cr[WS(rs, 20)];
230
            TU = TS * TT;
231
            TX = ci[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
       T7g = T4f + T4h;
238
       T4d = TO - TY;
239
       T4i = T4f - T4h;
240
       T4j = T4d - T4i;
241
       T6t = T4d + T4i;
242
        }
243
        {
244
       E T17, T4m, T1f, T4o, T4k, T4p;
245
       {
246
            E T12, T13, T16, T4l;
247
            T12 = cr[WS(rs, 28)];
248
            T13 = T11 * T12;
249
            T16 = ci[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 = cr[WS(rs, 12)];
257
            T1b = T19 * T1a;
258
            T1e = ci[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
       T7f = 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 = cr[WS(rs, 2)];
275
            T1k = T7 * T1j;
276
            T1l = ci[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 = cr[WS(rs, 26)];
284
            T1E = T1C * T1D;
285
            T1G = ci[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 = cr[WS(rs, 18)];
293
            T1q = T1o * T1p;
294
            T1t = ci[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 = cr[WS(rs, 10)];
302
            T1y = T1w * T1x;
303
            T1A = ci[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
       T7j = T1v - T1I;
311
       T7k = T4u + T4w;
312
       T7l = T4C + T4E;
313
       T7m = T7k - T7l;
314
       {
315
            E T4x, T4y, T4A, T4F;
316
            T4x = T4u - T4w;
317
            T4y = T1B - T1H;
318
            T4z = T4x + T4y;
319
            T6y = T4x - T4y;
320
            T4A = T1m - T1u;
321
            T4F = T4C - T4E;
322
            T4G = T4A - T4F;
323
            T6x = T4A + T4F;
324
       }
325
        }
326
        {
327
       E T1P, T4J, T2i, T4T, T21, T4L, T28, T4R;
328
       {
329
            E T1L, T1M, T1O, T4I;
330
            T1L = cr[WS(rs, 30)];
331
            T1M = T1K * T1L;
332
            T1O = ci[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 = cr[WS(rs, 22)];
340
            T2e = T2c * T2d;
341
            T2h = ci[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 = cr[WS(rs, 14)];
349
            T1X = T1V * T1W;
350
            T20 = ci[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 = cr[WS(rs, 6)];
358
            T25 = T23 * T24;
359
            T27 = ci[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
            T6B = T4M - T4N;
376
            T4P = T1P - T21;
377
            T4U = T4R - T4T;
378
            T4V = T4P - T4U;
379
            T6A = T4P + T4U;
380
       }
381
        }
382
        {
383
       E T3l, T5X, T3E, T5C, T3t, T5Z, T3y, T5A;
384
       {
385
            E T3g, T3h, T3k, T5W;
386
            T3g = cr[WS(rs, 31)];
387
            T3h = T3f * T3g;
388
            T3k = ci[WS(rs, 31)];
389
            T5W = T3f * T3k;
390
            T3l = FMA(T3j, T3k, T3h);
391
            T5X = FNMS(T3j, T3g, T5W);
392
       }
393
       {
394
            E T3A, T3B, T3D, T5B;
395
            T3A = cr[WS(rs, 23)];
396
            T3B = T3z * T3A;
397
            T3D = ci[WS(rs, 23)];
398
            T5B = T3z * T3D;
399
            T3E = FMA(T3C, T3D, T3B);
400
            T5C = FNMS(T3C, T3A, T5B);
401
       }
402
       {
403
            E T3o, T3p, T3s, T5Y;
404
            T3o = cr[WS(rs, 15)];
405
            T3p = T3n * T3o;
406
            T3s = ci[WS(rs, 15)];
407
            T5Y = T3n * T3s;
408
            T3t = FMA(T3r, T3s, T3p);
409
            T5Z = FNMS(T3r, T3o, T5Y);
410
       }
411
       {
412
            E T3v, T3w, T3x, T5z;
413
            T3v = cr[WS(rs, 7)];
414
            T3w = TP * T3v;
415
            T3x = ci[WS(rs, 7)];
416
            T5z = TP * T3x;
417
            T3y = FMA(TR, T3x, T3w);
418
            T5A = FNMS(TR, T3v, T5z);
419
       }
420
       {
421
            E T3u, T3F, T7L, T7M;
422
            T3u = T3l + T3t;
423
            T3F = T3y + T3E;
424
            T3G = T3u + T3F;
425
            T7G = T3u - T3F;
426
            T7L = T5X + T5Z;
427
            T7M = T5A + T5C;
428
            T7N = T7L - T7M;
429
            T8n = T7L + T7M;
430
       }
431
       {
432
            E T5y, T5D, T5V, T60;
433
            T5y = T3l - T3t;
434
            T5D = T5A - T5C;
435
            T5E = T5y - T5D;
436
            T6M = T5y + T5D;
437
            T5V = T3E - T3y;
438
            T60 = T5X - T5Z;
439
            T61 = T5V - T60;
440
            T6P = T60 + T5V;
441
       }
442
        }
443
        {
444
       E T2q, T5n, T2L, T53, T2y, T5p, T2D, T51;
445
       {
446
            E T2n, T2o, T2p, T5m;
447
            T2n = cr[WS(rs, 1)];
448
            T2o = T2 * T2n;
449
            T2p = ci[WS(rs, 1)];
450
            T5m = T2 * T2p;
451
            T2q = FMA(T5, T2p, T2o);
452
            T5n = FNMS(T5, T2n, T5m);
453
       }
454
       {
455
            E T2G, T2H, T2K, T52;
456
            T2G = cr[WS(rs, 25)];
457
            T2H = T2F * T2G;
458
            T2K = ci[WS(rs, 25)];
459
            T52 = T2F * T2K;
460
            T2L = FMA(T2J, T2K, T2H);
461
            T53 = FNMS(T2J, T2G, T52);
462
       }
463
       {
464
            E T2t, T2u, T2x, T5o;
465
            T2t = cr[WS(rs, 17)];
466
            T2u = T2s * T2t;
467
            T2x = ci[WS(rs, 17)];
468
            T5o = T2s * T2x;
469
            T2y = FMA(T2w, T2x, T2u);
470
            T5p = FNMS(T2w, T2t, T5o);
471
       }
472
       {
473
            E T2A, T2B, T2C, T50;
474
            T2A = cr[WS(rs, 9)];
475
            T2B = T8 * T2A;
476
            T2C = ci[WS(rs, 9)];
477
            T50 = T8 * T2C;
478
            T2D = FMA(Tc, T2C, T2B);
479
            T51 = FNMS(Tc, T2A, T50);
480
       }
481
       {
482
            E T2z, T2M, T7A, T7B;
483
            T2z = T2q + T2y;
484
            T2M = T2D + T2L;
485
            T2N = T2z + T2M;
486
            T7v = T2z - T2M;
487
            T7A = T5n + T5p;
488
            T7B = T51 + T53;
489
            T7C = T7A - T7B;
490
            T8i = T7A + T7B;
491
       }
492
       {
493
            E T4Z, T54, T5q, T5r;
494
            T4Z = T2q - T2y;
495
            T54 = T51 - T53;
496
            T55 = T4Z - T54;
497
            T6F = T4Z + T54;
498
            T5q = T5n - T5p;
499
            T5r = T2D - T2L;
500
            T5s = T5q + T5r;
501
            T6I = T5q - T5r;
502
       }
503
        }
504
        {
505
       E T3K, T5H, T41, T5Q, T3S, T5J, T3X, T5O;
506
       {
507
            E T3H, T3I, T3J, T5G;
508
            T3H = cr[WS(rs, 3)];
509
            T3I = T3 * T3H;
510
            T3J = ci[WS(rs, 3)];
511
            T5G = T3 * T3J;
512
            T3K = FMA(T6, T3J, T3I);
513
            T5H = FNMS(T6, T3H, T5G);
514
       }
515
       {
516
            E T3Y, T3Z, T40, T5P;
517
            T3Y = cr[WS(rs, 11)];
518
            T3Z = Td * T3Y;
519
            T40 = ci[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, T5I;
526
            T3N = cr[WS(rs, 19)];
527
            T3O = T3M * T3N;
528
            T3R = ci[WS(rs, 19)];
529
            T5I = T3M * T3R;
530
            T3S = FMA(T3Q, T3R, T3O);
531
            T5J = FNMS(T3Q, T3N, T5I);
532
       }
533
       {
534
            E T3U, T3V, T3W, T5N;
535
            T3U = cr[WS(rs, 27)];
536
            T3V = Te * T3U;
537
            T3W = ci[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, T7H, T7I;
544
            T3T = T3K + T3S;
545
            T42 = T3X + T41;
546
            T43 = T3T + T42;
547
            T7O = T42 - T3T;
548
            T7H = T5O + T5Q;
549
            T7I = T5H + T5J;
550
            T7J = T7H - T7I;
551
            T8o = T7I + T7H;
552
       }
553
       {
554
            E T5F, T5K, T5M, T5R;
555
            T5F = T3K - T3S;
556
            T5K = T5H - T5J;
557
            T5L = T5F - T5K;
558
            T63 = T5F + T5K;
559
            T5M = T3X - T41;
560
            T5R = T5O - T5Q;
561
            T5S = T5M + T5R;
562
            T62 = T5M - T5R;
563
       }
564
        }
565
        {
566
       E T2R, T58, T3a, T5h, T2Z, T5a, T36, T5f;
567
       {
568
            E T2O, T2P, T2Q, T57;
569
            T2O = cr[WS(rs, 5)];
570
            T2P = T29 * T2O;
571
            T2Q = ci[WS(rs, 5)];
572
            T57 = T29 * T2Q;
573
            T2R = FMA(T2b, T2Q, T2P);
574
            T58 = FNMS(T2b, T2O, T57);
575
       }
576
       {
577
            E T37, T38, T39, T5g;
578
            T37 = cr[WS(rs, 13)];
579
            T38 = T1R * T37;
580
            T39 = ci[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, T59;
587
            T2U = cr[WS(rs, 21)];
588
            T2V = T2T * T2U;
589
            T2Y = ci[WS(rs, 21)];
590
            T59 = T2T * T2Y;
591
            T2Z = FMA(T2X, T2Y, T2V);
592
            T5a = FNMS(T2X, T2U, T59);
593
       }
594
       {
595
            E T32, T33, T35, T5e;
596
            T32 = cr[WS(rs, 29)];
597
            T33 = T31 * T32;
598
            T35 = ci[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, T7w, T7x;
605
            T30 = T2R + T2Z;
606
            T3b = T36 + T3a;
607
            T3c = T30 + T3b;
608
            T7D = T30 - T3b;
609
            T7w = T5f + T5h;
610
            T7x = T58 + T5a;
611
            T7y = T7w - T7x;
612
            T8j = T7x + T7w;
613
       }
614
       {
615
            E T56, T5b, T5d, T5i;
616
            T56 = T2R - T2Z;
617
            T5b = T58 - T5a;
618
            T5c = T56 - T5b;
619
            T5t = T56 + T5b;
620
            T5d = T36 - T3a;
621
            T5i = T5f - T5h;
622
            T5j = T5d + T5i;
623
            T5u = T5i - T5d;
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 = T8n + T8o;
636
            T8y = T8i + T8j;
637
            T8z = T8x - T8y;
638
            T8A = T8y + T8x;
639
       }
640
       {
641
            E T8C, T8I, T1J, T2k;
642
            T8C = T7g + T7f;
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 = T1J - T2k;
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 = T8h - T8k;
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 = T7p + T7q;
669
            T8e = T7k + T7l;
670
            T8f = T8d - T8e;
671
            T8B = T8e + T8d;
672
       }
673
       {
674
            E T2m, T8K, T8M, T8w;
675
            T2m = T1i + T2l;
676
            ci[WS(rs, 15)] = T2m - T45;
677
            cr[0] = T2m + T45;
678
            T8K = T8B + T8J;
679
            cr[WS(rs, 16)] = T8A - T8K;
680
            ci[WS(rs, 31)] = T8A + T8K;
681
            T8M = T8J - T8B;
682
            cr[WS(rs, 24)] = T8L - T8M;
683
            ci[WS(rs, 23)] = T8L + T8M;
684
            T8w = T1i - T2l;
685
            cr[WS(rs, 8)] = T8w - T8z;
686
            ci[WS(rs, 7)] = T8w + T8z;
687
       }
688
       {
689
            E T8g, T8r, T8P, T8Q;
690
            T8g = T8c - T8f;
691
            T8r = T8l + T8q;
692
            ci[WS(rs, 11)] = FNMS(KP707106781, T8r, T8g);
693
            cr[WS(rs, 4)] = FMA(KP707106781, T8r, T8g);
694
            T8P = T8N + T8O;
695
            T8Q = T8q - T8l;
696
            cr[WS(rs, 28)] = FMS(KP707106781, T8Q, T8P);
697
            ci[WS(rs, 19)] = FMA(KP707106781, T8Q, T8P);
698
       }
699
       {
700
            E T8R, T8S, T8s, T8v;
701
            T8R = T8O - T8N;
702
            T8S = T8u - T8t;
703
            cr[WS(rs, 20)] = FMS(KP707106781, T8S, T8R);
704
            ci[WS(rs, 27)] = FMA(KP707106781, T8S, T8R);
705
            T8s = T8c + T8f;
706
            T8v = T8t + T8u;
707
            cr[WS(rs, 12)] = FNMS(KP707106781, T8v, T8s);
708
            ci[WS(rs, 3)] = FMA(KP707106781, T8v, T8s);
709
       }
710
        }
711
        {
712
       E T4s, T6c, T4X, T9c, T9b, T9h, T6f, T9i, T66, T6q, T6a, T6m, T5x, T6p, T69;
713
       E T6j;
714
       {
715
            E T4c, T4r, T6d, T6e;
716
            T4c = T46 - T4b;
717
            T4r = T4j + T4q;
718
            T4s = FNMS(KP707106781, T4r, T4c);
719
            T6c = FMA(KP707106781, T4r, T4c);
720
            {
721
           E T4H, T4W, T99, T9a;
722
           T4H = FMA(KP414213562, T4G, T4z);
723
           T4W = FNMS(KP414213562, T4V, T4O);
724
           T4X = T4H - T4W;
725
           T9c = T4H + T4W;
726
           T99 = T97 + T98;
727
           T9a = T6t - T6u;
728
           T9b = FMA(KP707106781, T9a, T99);
729
           T9h = FNMS(KP707106781, T9a, T99);
730
            }
731
            T6d = FNMS(KP414213562, T4z, T4G);
732
            T6e = FMA(KP414213562, T4O, T4V);
733
            T6f = T6d + T6e;
734
            T9i = T6e - T6d;
735
            {
736
           E T5U, T6k, T65, T6l, T5T, T64;
737
           T5T = T5L + T5S;
738
           T5U = FNMS(KP707106781, T5T, T5E);
739
           T6k = FMA(KP707106781, T5T, T5E);
740
           T64 = T62 - T63;
741
           T65 = FNMS(KP707106781, T64, T61);
742
           T6l = FMA(KP707106781, T64, T61);
743
           T66 = FMA(KP668178637, T65, T5U);
744
           T6q = FMA(KP198912367, T6k, T6l);
745
           T6a = FNMS(KP668178637, T5U, T65);
746
           T6m = FNMS(KP198912367, T6l, T6k);
747
            }
748
            {
749
           E T5l, T6h, T5w, T6i, T5k, T5v;
750
           T5k = T5c + T5j;
751
           T5l = FNMS(KP707106781, T5k, T55);
752
           T6h = FMA(KP707106781, T5k, T55);
753
           T5v = T5t + T5u;
754
           T5w = FNMS(KP707106781, T5v, T5s);
755
           T6i = FMA(KP707106781, T5v, T5s);
756
           T5x = FMA(KP668178637, T5w, T5l);
757
           T6p = FMA(KP198912367, T6h, T6i);
758
           T69 = FNMS(KP668178637, T5l, T5w);
759
           T6j = FNMS(KP198912367, T6i, T6h);
760
            }
761
       }
762
       {
763
            E T4Y, T67, T9j, T9k;
764
            T4Y = FMA(KP923879532, T4X, T4s);
765
            T67 = T5x + T66;
766
            ci[WS(rs, 12)] = FNMS(KP831469612, T67, T4Y);
767
            cr[WS(rs, 3)] = FMA(KP831469612, T67, T4Y);
768
            T9j = FMA(KP923879532, T9i, T9h);
769
            T9k = T69 - T6a;
770
            cr[WS(rs, 19)] = FMS(KP831469612, T9k, T9j);
771
            ci[WS(rs, 28)] = FMA(KP831469612, T9k, T9j);
772
       }
773
       {
774
            E T9l, T9m, T68, T6b;
775
            T9l = FNMS(KP923879532, T9i, T9h);
776
            T9m = T66 - T5x;
777
            cr[WS(rs, 27)] = FMS(KP831469612, T9m, T9l);
778
            ci[WS(rs, 20)] = FMA(KP831469612, T9m, T9l);
779
            T68 = FNMS(KP923879532, T4X, T4s);
780
            T6b = T69 + T6a;
781
            cr[WS(rs, 11)] = FMA(KP831469612, T6b, T68);
782
            ci[WS(rs, 4)] = FNMS(KP831469612, T6b, T68);
783
       }
784
       {
785
            E T6g, T6n, T9d, T9e;
786
            T6g = FMA(KP923879532, T6f, T6c);
787
            T6n = T6j + T6m;
788
            cr[WS(rs, 15)] = FNMS(KP980785280, T6n, T6g);
789
            ci[0] = FMA(KP980785280, T6n, T6g);
790
            T9d = FMA(KP923879532, T9c, T9b);
791
            T9e = T6q - T6p;
792
            cr[WS(rs, 31)] = FMS(KP980785280, T9e, T9d);
793
            ci[WS(rs, 16)] = FMA(KP980785280, T9e, T9d);
794
       }
795
       {
796
            E T9f, T9g, T6o, T6r;
797
            T9f = FNMS(KP923879532, T9c, T9b);
798
            T9g = T6m - T6j;
799
            cr[WS(rs, 23)] = FMS(KP980785280, T9g, T9f);
800
            ci[WS(rs, 24)] = FMA(KP980785280, T9g, T9f);
801
            T6o = FNMS(KP923879532, T6f, T6c);
802
            T6r = T6p + T6q;
803
            ci[WS(rs, 8)] = FNMS(KP980785280, T6r, T6o);
804
            cr[WS(rs, 7)] = FMA(KP980785280, T6r, T6o);
805
       }
806
        }
807
        {
808
       E T7i, T7W, T86, T8a, T8V, T91, T7t, T8W, T7F, T7U, T7Z, T92, T83, T89, T7Q;
809
       E T7T;
810
       {
811
            E T7e, T7h, T84, T85;
812
            T7e = Tq - TH;
813
            T7h = T7f - T7g;
814
            T7i = T7e - T7h;
815
            T7W = T7e + T7h;
816
            T84 = T7G + T7J;
817
            T85 = T7O - T7N;
818
            T86 = FNMS(KP414213562, T85, T84);
819
            T8a = FMA(KP414213562, T84, T85);
820
       }
821
       {
822
            E T8T, T8U, T7n, T7s;
823
            T8T = TZ - T1g;
824
            T8U = T8H - T8D;
825
            T8V = T8T + T8U;
826
            T91 = T8U - T8T;
827
            T7n = T7j + T7m;
828
            T7s = T7o - T7r;
829
            T7t = T7n + T7s;
830
            T8W = T7n - T7s;
831
       }
832
       {
833
            E T7z, T7E, T7X, T7Y;
834
            T7z = T7v - T7y;
835
            T7E = T7C - T7D;
836
            T7F = FMA(KP414213562, T7E, T7z);
837
            T7U = FNMS(KP414213562, T7z, T7E);
838
            T7X = T7j - T7m;
839
            T7Y = T7o + T7r;
840
            T7Z = T7X + T7Y;
841
            T92 = T7Y - T7X;
842
       }
843
       {
844
            E T81, T82, T7K, T7P;
845
            T81 = T7v + T7y;
846
            T82 = T7C + T7D;
847
            T83 = FNMS(KP414213562, T82, T81);
848
            T89 = FMA(KP414213562, T81, T82);
849
            T7K = T7G - T7J;
850
            T7P = T7N + T7O;
851
            T7Q = FNMS(KP414213562, T7P, T7K);
852
            T7T = FMA(KP414213562, T7K, T7P);
853
       }
854
       {
855
            E T7u, T7R, T93, T94;
856
            T7u = FMA(KP707106781, T7t, T7i);
857
            T7R = T7F + T7Q;
858
            ci[WS(rs, 13)] = FNMS(KP923879532, T7R, T7u);
859
            cr[WS(rs, 2)] = FMA(KP923879532, T7R, T7u);
860
            T93 = FMA(KP707106781, T92, T91);
861
            T94 = T7U + T7T;
862
            cr[WS(rs, 18)] = FMS(KP923879532, T94, T93);
863
            ci[WS(rs, 29)] = FMA(KP923879532, T94, T93);
864
       }
865
       {
866
            E T95, T96, T7S, T7V;
867
            T95 = FNMS(KP707106781, T92, T91);
868
            T96 = T7Q - T7F;
869
            cr[WS(rs, 26)] = FMS(KP923879532, T96, T95);
870
            ci[WS(rs, 21)] = FMA(KP923879532, T96, T95);
871
            T7S = FNMS(KP707106781, T7t, T7i);
872
            T7V = T7T - T7U;
873
            cr[WS(rs, 10)] = FNMS(KP923879532, T7V, T7S);
874
            ci[WS(rs, 5)] = FMA(KP923879532, T7V, T7S);
875
       }
876
       {
877
            E T80, T87, T8X, T8Y;
878
            T80 = FMA(KP707106781, T7Z, T7W);
879
            T87 = T83 + T86;
880
            cr[WS(rs, 14)] = FNMS(KP923879532, T87, T80);
881
            ci[WS(rs, 1)] = FMA(KP923879532, T87, T80);
882
            T8X = FMA(KP707106781, T8W, T8V);
883
            T8Y = T8a - T89;
884
            cr[WS(rs, 30)] = FMS(KP923879532, T8Y, T8X);
885
            ci[WS(rs, 17)] = FMA(KP923879532, T8Y, T8X);
886
       }
887
       {
888
            E T8Z, T90, T88, T8b;
889
            T8Z = FNMS(KP707106781, T8W, T8V);
890
            T90 = T86 - T83;
891
            cr[WS(rs, 22)] = FMS(KP923879532, T90, T8Z);
892
            ci[WS(rs, 25)] = FMA(KP923879532, T90, T8Z);
893
            T88 = FNMS(KP707106781, T7Z, T7W);
894
            T8b = T89 + T8a;
895
            ci[WS(rs, 9)] = FNMS(KP923879532, T8b, T88);
896
            cr[WS(rs, 6)] = FMA(KP923879532, T8b, T88);
897
       }
898
        }
899
        {
900
       E T6w, T6Y, T6D, T9w, T9p, T9v, T71, T9q, T6S, T7c, T6V, T78, T6L, T7b, T6W;
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 = FMA(KP414213562, T6A, T6B);
920
            T70 = FNMS(KP414213562, T6x, T6y);
921
            T71 = T6Z - T70;
922
            T9q = T70 + T6Z;
923
            {
924
           E T6O, T76, T6R, T77, T6N, T6Q;
925
           T6N = T63 + T62;
926
           T6O = FMA(KP707106781, T6N, T6M);
927
           T76 = FNMS(KP707106781, T6N, T6M);
928
           T6Q = T5S - T5L;
929
           T6R = FMA(KP707106781, T6Q, T6P);
930
           T77 = FNMS(KP707106781, T6Q, T6P);
931
           T6S = FNMS(KP198912367, T6R, T6O);
932
           T7c = FNMS(KP668178637, T76, T77);
933
           T6V = FMA(KP198912367, T6O, T6R);
934
           T78 = FMA(KP668178637, T77, T76);
935
            }
936
            {
937
           E T6H, T73, T6K, T74, T6G, T6J;
938
           T6G = T5t - T5u;
939
           T6H = FMA(KP707106781, T6G, T6F);
940
           T73 = FNMS(KP707106781, T6G, T6F);
941
           T6J = T5j - T5c;
942
           T6K = FMA(KP707106781, T6J, T6I);
943
           T74 = FNMS(KP707106781, T6J, T6I);
944
           T6L = FMA(KP198912367, T6K, T6H);
945
           T7b = FMA(KP668178637, T73, T74);
946
           T6W = FNMS(KP198912367, T6H, T6K);
947
           T75 = FNMS(KP668178637, T74, T73);
948
            }
949
       }
950
       {
951
            E T6E, T6T, T9x, T9y;
952
            T6E = FMA(KP923879532, T6D, T6w);
953
            T6T = T6L + T6S;
954
            ci[WS(rs, 14)] = FNMS(KP980785280, T6T, T6E);
955
            cr[WS(rs, 1)] = FMA(KP980785280, T6T, T6E);
956
            T9x = FMA(KP923879532, T9w, T9v);
957
            T9y = T7b + T7c;
958
            cr[WS(rs, 29)] = -(FMA(KP831469612, T9y, T9x));
959
            ci[WS(rs, 18)] = FNMS(KP831469612, T9y, T9x);
960
       }
961
       {
962
            E T9z, T9A, T6U, T6X;
963
            T9z = FNMS(KP923879532, T9w, T9v);
964
            T9A = T78 - T75;
965
            cr[WS(rs, 21)] = FMS(KP831469612, T9A, T9z);
966
            ci[WS(rs, 26)] = FMA(KP831469612, T9A, T9z);
967
            T6U = FNMS(KP923879532, T6D, T6w);
968
            T6X = T6V - T6W;
969
            cr[WS(rs, 9)] = FNMS(KP980785280, T6X, T6U);
970
            ci[WS(rs, 6)] = FMA(KP980785280, T6X, T6U);
971
       }
972
       {
973
            E T72, T79, T9r, T9s;
974
            T72 = FMA(KP923879532, T71, T6Y);
975
            T79 = T75 + T78;
976
            cr[WS(rs, 13)] = FNMS(KP831469612, T79, T72);
977
            ci[WS(rs, 2)] = FMA(KP831469612, T79, T72);
978
            T9r = FMA(KP923879532, T9q, T9p);
979
            T9s = T6W + T6V;
980
            cr[WS(rs, 17)] = FMS(KP980785280, T9s, T9r);
981
            ci[WS(rs, 30)] = FMA(KP980785280, T9s, T9r);
982
       }
983
       {
984
            E T9t, T9u, T7a, T7d;
985
            T9t = FNMS(KP923879532, T9q, T9p);
986
            T9u = T6S - T6L;
987
            cr[WS(rs, 25)] = FMS(KP980785280, T9u, T9t);
988
            ci[WS(rs, 22)] = FMA(KP980785280, T9u, T9t);
989
            T7a = FNMS(KP923879532, T71, T6Y);
990
            T7d = T7b - T7c;
991
            ci[WS(rs, 10)] = FNMS(KP831469612, T7d, T7a);
992
            cr[WS(rs, 5)] = FMA(KP831469612, T7d, T7a);
993
       }
994
        }
995
         }
996
    }
997
     }
998
}
999
1000
static const tw_instr twinstr[] = {
1001
     { TW_CEXP, 1, 1 },
1002
     { TW_CEXP, 1, 3 },
1003
     { TW_CEXP, 1, 9 },
1004
     { TW_CEXP, 1, 27 },
1005
     { TW_NEXT, 1, 0 }
1006
};
1007
1008
static const hc2hc_desc desc = { 32, "hf2_32", twinstr, &GENUS, { 236, 98, 252, 0 } };
1009
1010
void X(codelet_hf2_32) (planner *p) {
1011
     X(khc2hc_register) (p, hf2_32, &desc);
1012
}
1013
#else
1014
1015
/* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 32 -dit -name hf2_32 -include rdft/scalar/hf.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 "rdft/scalar/hf.h"
1023
1024
static void hf2_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
1025
0
{
1026
0
     DK(KP555570233, +0.555570233019602224742830813948532874374937191);
1027
0
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
1028
0
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
1029
0
     DK(KP195090322, +0.195090322016128267848284868477022240927691618);
1030
0
     DK(KP382683432, +0.382683432365089771728459984030398866761344562);
1031
0
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
1032
0
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
1033
0
     {
1034
0
    INT m;
1035
0
    for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 8, MAKE_VOLATILE_STRIDE(64, rs)) {
1036
0
         E T2, T5, T3, T6, T8, TM, TO, Td, T9, Te, Th, Tl, TD, TH, T1y;
1037
0
         E T1H, T15, T1A, T11, T1F, T1n, T1p, T2q, T2I, T2u, T2K, T2V, T3b, T2Z, T3d;
1038
0
         E Tu, Ty, T3l, T3n, T1t, T1v, T2f, T2h, T1a, T1e, T32, T34, T1W, T1Y, T2C;
1039
0
         E T2E, Tg, TR, Tk, TS, Tm, TV, To, TT, T1M, T21, T1P, T22, T1Q, T25;
1040
0
         E T1S, T23;
1041
0
         {
1042
0
        E Ts, T1d, Tx, T18, Tt, T1c, Tw, T19, TB, T14, TG, TZ, TC, T13, TF;
1043
0
        E T10;
1044
0
        {
1045
0
       E T4, Tc, T7, Tb;
1046
0
       T2 = W[0];
1047
0
       T5 = W[1];
1048
0
       T3 = W[2];
1049
0
       T6 = W[3];
1050
0
       T4 = T2 * T3;
1051
0
       Tc = T5 * T3;
1052
0
       T7 = T5 * T6;
1053
0
       Tb = T2 * T6;
1054
0
       T8 = T4 + T7;
1055
0
       TM = T4 - T7;
1056
0
       TO = Tb + Tc;
1057
0
       Td = Tb - Tc;
1058
0
       T9 = W[4];
1059
0
       Ts = T2 * T9;
1060
0
       T1d = T6 * T9;
1061
0
       Tx = T5 * T9;
1062
0
       T18 = T3 * T9;
1063
0
       Te = W[5];
1064
0
       Tt = T5 * Te;
1065
0
       T1c = T3 * Te;
1066
0
       Tw = T2 * Te;
1067
0
       T19 = T6 * Te;
1068
0
       Th = W[6];
1069
0
       TB = T3 * Th;
1070
0
       T14 = T5 * Th;
1071
0
       TG = T6 * Th;
1072
0
       TZ = T2 * Th;
1073
0
       Tl = W[7];
1074
0
       TC = T6 * Tl;
1075
0
       T13 = T2 * Tl;
1076
0
       TF = T3 * Tl;
1077
0
       T10 = T5 * Tl;
1078
0
        }
1079
0
        TD = TB + TC;
1080
0
        TH = TF - TG;
1081
0
        T1y = TZ + T10;
1082
0
        T1H = TF + TG;
1083
0
        T15 = T13 + T14;
1084
0
        T1A = T13 - T14;
1085
0
        T11 = TZ - T10;
1086
0
        T1F = TB - TC;
1087
0
        T1n = FMA(T9, Th, Te * Tl);
1088
0
        T1p = FNMS(Te, Th, T9 * Tl);
1089
0
        {
1090
0
       E T2o, T2p, T2s, T2t;
1091
0
       T2o = T8 * Th;
1092
0
       T2p = Td * Tl;
1093
0
       T2q = T2o + T2p;
1094
0
       T2I = T2o - T2p;
1095
0
       T2s = T8 * Tl;
1096
0
       T2t = Td * Th;
1097
0
       T2u = T2s - T2t;
1098
0
       T2K = T2s + T2t;
1099
0
        }
1100
0
        {
1101
0
       E T2T, T2U, T2X, T2Y;
1102
0
       T2T = TM * Th;
1103
0
       T2U = TO * Tl;
1104
0
       T2V = T2T - T2U;
1105
0
       T3b = T2T + T2U;
1106
0
       T2X = TM * Tl;
1107
0
       T2Y = TO * Th;
1108
0
       T2Z = T2X + T2Y;
1109
0
       T3d = T2X - T2Y;
1110
0
       Tu = Ts + Tt;
1111
0
       Ty = Tw - Tx;
1112
0
       T3l = FMA(Tu, Th, Ty * Tl);
1113
0
       T3n = FNMS(Ty, Th, Tu * Tl);
1114
0
        }
1115
0
        T1t = Ts - Tt;
1116
0
        T1v = Tw + Tx;
1117
0
        T2f = FMA(T1t, Th, T1v * Tl);
1118
0
        T2h = FNMS(T1v, Th, T1t * Tl);
1119
0
        T1a = T18 - T19;
1120
0
        T1e = T1c + T1d;
1121
0
        T32 = FMA(T1a, Th, T1e * Tl);
1122
0
        T34 = FNMS(T1e, Th, T1a * Tl);
1123
0
        T1W = T18 + T19;
1124
0
        T1Y = T1c - T1d;
1125
0
        T2C = FMA(T1W, Th, T1Y * Tl);
1126
0
        T2E = FNMS(T1Y, Th, T1W * Tl);
1127
0
        {
1128
0
       E Ta, Tf, Ti, Tj;
1129
0
       Ta = T8 * T9;
1130
0
       Tf = Td * Te;
1131
0
       Tg = Ta - Tf;
1132
0
       TR = Ta + Tf;
1133
0
       Ti = T8 * Te;
1134
0
       Tj = Td * T9;
1135
0
       Tk = Ti + Tj;
1136
0
       TS = Ti - Tj;
1137
0
        }
1138
0
        Tm = FMA(Tg, Th, Tk * Tl);
1139
0
        TV = FNMS(TS, Th, TR * Tl);
1140
0
        To = FNMS(Tk, Th, Tg * Tl);
1141
0
        TT = FMA(TR, Th, TS * Tl);
1142
0
        {
1143
0
       E T1K, T1L, T1N, T1O;
1144
0
       T1K = TM * T9;
1145
0
       T1L = TO * Te;
1146
0
       T1M = T1K - T1L;
1147
0
       T21 = T1K + T1L;
1148
0
       T1N = TM * Te;
1149
0
       T1O = TO * T9;
1150
0
       T1P = T1N + T1O;
1151
0
       T22 = T1N - T1O;
1152
0
        }
1153
0
        T1Q = FMA(T1M, Th, T1P * Tl);
1154
0
        T25 = FNMS(T22, Th, T21 * Tl);
1155
0
        T1S = FNMS(T1P, Th, T1M * Tl);
1156
0
        T23 = FMA(T21, Th, T22 * Tl);
1157
0
         }
1158
0
         {
1159
0
        E TL, T6f, T8c, T8q, T3F, T5t, T7I, T7W, T2y, T6B, T6y, T7j, T4k, T5G, T4B;
1160
0
        E T5J, T3h, T6H, T6O, T7o, T4L, T5Q, T52, T5N, T1i, T7V, T6i, T7D, T3K, T5u;
1161
0
        E T3P, T5v, T1E, T6k, T6n, T7f, T3W, T5z, T41, T5y, T29, T6p, T6s, T7e, T47;
1162
0
        E T5C, T4c, T5B, T2R, T6z, T6E, T7k, T4v, T5K, T4E, T5H, T3y, T6P, T6K, T7p;
1163
0
        E T4W, T5O, T55, T5R;
1164
0
        {
1165
0
       E T1, T7G, Tq, T7F, TA, T3C, TJ, T3D, Tn, Tp;
1166
0
       T1 = cr[0];
1167
0
       T7G = ci[0];
1168
0
       Tn = cr[WS(rs, 16)];
1169
0
       Tp = ci[WS(rs, 16)];
1170
0
       Tq = FMA(Tm, Tn, To * Tp);
1171
0
       T7F = FNMS(To, Tn, Tm * Tp);
1172
0
       {
1173
0
            E Tv, Tz, TE, TI;
1174
0
            Tv = cr[WS(rs, 8)];
1175
0
            Tz = ci[WS(rs, 8)];
1176
0
            TA = FMA(Tu, Tv, Ty * Tz);
1177
0
            T3C = FNMS(Ty, Tv, Tu * Tz);
1178
0
            TE = cr[WS(rs, 24)];
1179
0
            TI = ci[WS(rs, 24)];
1180
0
            TJ = FMA(TD, TE, TH * TI);
1181
0
            T3D = FNMS(TH, TE, TD * TI);
1182
0
       }
1183
0
       {
1184
0
            E Tr, TK, T8a, T8b;
1185
0
            Tr = T1 + Tq;
1186
0
            TK = TA + TJ;
1187
0
            TL = Tr + TK;
1188
0
            T6f = Tr - TK;
1189
0
            T8a = TA - TJ;
1190
0
            T8b = T7G - T7F;
1191
0
            T8c = T8a + T8b;
1192
0
            T8q = T8b - T8a;
1193
0
       }
1194
0
       {
1195
0
            E T3B, T3E, T7E, T7H;
1196
0
            T3B = T1 - Tq;
1197
0
            T3E = T3C - T3D;
1198
0
            T3F = T3B + T3E;
1199
0
            T5t = T3B - T3E;
1200
0
            T7E = T3C + T3D;
1201
0
            T7H = T7F + T7G;
1202
0
            T7I = T7E + T7H;
1203
0
            T7W = T7H - T7E;
1204
0
       }
1205
0
        }
1206
0
        {
1207
0
       E T2e, T4x, T2w, T4i, T2j, T4y, T2n, T4h;
1208
0
       {
1209
0
            E T2c, T2d, T2r, T2v;
1210
0
            T2c = cr[WS(rs, 1)];
1211
0
            T2d = ci[WS(rs, 1)];
1212
0
            T2e = FMA(T2, T2c, T5 * T2d);
1213
0
            T4x = FNMS(T5, T2c, T2 * T2d);
1214
0
            T2r = cr[WS(rs, 25)];
1215
0
            T2v = ci[WS(rs, 25)];
1216
0
            T2w = FMA(T2q, T2r, T2u * T2v);
1217
0
            T4i = FNMS(T2u, T2r, T2q * T2v);
1218
0
       }
1219
0
       {
1220
0
            E T2g, T2i, T2l, T2m;
1221
0
            T2g = cr[WS(rs, 17)];
1222
0
            T2i = ci[WS(rs, 17)];
1223
0
            T2j = FMA(T2f, T2g, T2h * T2i);
1224
0
            T4y = FNMS(T2h, T2g, T2f * T2i);
1225
0
            T2l = cr[WS(rs, 9)];
1226
0
            T2m = ci[WS(rs, 9)];
1227
0
            T2n = FMA(T9, T2l, Te * T2m);
1228
0
            T4h = FNMS(Te, T2l, T9 * T2m);
1229
0
       }
1230
0
       {
1231
0
            E T2k, T2x, T6w, T6x;
1232
0
            T2k = T2e + T2j;
1233
0
            T2x = T2n + T2w;
1234
0
            T2y = T2k + T2x;
1235
0
            T6B = T2k - T2x;
1236
0
            T6w = T4x + T4y;
1237
0
            T6x = T4h + T4i;
1238
0
            T6y = T6w - T6x;
1239
0
            T7j = T6w + T6x;
1240
0
       }
1241
0
       {
1242
0
            E T4g, T4j, T4z, T4A;
1243
0
            T4g = T2e - T2j;
1244
0
            T4j = T4h - T4i;
1245
0
            T4k = T4g + T4j;
1246
0
            T5G = T4g - T4j;
1247
0
            T4z = T4x - T4y;
1248
0
            T4A = T2n - T2w;
1249
0
            T4B = T4z - T4A;
1250
0
            T5J = T4z + T4A;
1251
0
       }
1252
0
        }
1253
0
        {
1254
0
       E T31, T4H, T3f, T50, T36, T4I, T3a, T4Z;
1255
0
       {
1256
0
            E T2W, T30, T3c, T3e;
1257
0
            T2W = cr[WS(rs, 31)];
1258
0
            T30 = ci[WS(rs, 31)];
1259
0
            T31 = FMA(T2V, T2W, T2Z * T30);
1260
0
            T4H = FNMS(T2Z, T2W, T2V * T30);
1261
0
            T3c = cr[WS(rs, 23)];
1262
0
            T3e = ci[WS(rs, 23)];
1263
0
            T3f = FMA(T3b, T3c, T3d * T3e);
1264
0
            T50 = FNMS(T3d, T3c, T3b * T3e);
1265
0
       }
1266
0
       {
1267
0
            E T33, T35, T38, T39;
1268
0
            T33 = cr[WS(rs, 15)];
1269
0
            T35 = ci[WS(rs, 15)];
1270
0
            T36 = FMA(T32, T33, T34 * T35);
1271
0
            T4I = FNMS(T34, T33, T32 * T35);
1272
0
            T38 = cr[WS(rs, 7)];
1273
0
            T39 = ci[WS(rs, 7)];
1274
0
            T3a = FMA(TR, T38, TS * T39);
1275
0
            T4Z = FNMS(TS, T38, TR * T39);
1276
0
       }
1277
0
       {
1278
0
            E T37, T3g, T6M, T6N;
1279
0
            T37 = T31 + T36;
1280
0
            T3g = T3a + T3f;
1281
0
            T3h = T37 + T3g;
1282
0
            T6H = T37 - T3g;
1283
0
            T6M = T4H + T4I;
1284
0
            T6N = T4Z + T50;
1285
0
            T6O = T6M - T6N;
1286
0
            T7o = T6M + T6N;
1287
0
       }
1288
0
       {
1289
0
            E T4J, T4K, T4Y, T51;
1290
0
            T4J = T4H - T4I;
1291
0
            T4K = T3a - T3f;
1292
0
            T4L = T4J - T4K;
1293
0
            T5Q = T4J + T4K;
1294
0
            T4Y = T31 - T36;
1295
0
            T51 = T4Z - T50;
1296
0
            T52 = T4Y + T51;
1297
0
            T5N = T4Y - T51;
1298
0
       }
1299
0
        }
1300
0
        {
1301
0
       E TQ, T3H, T1g, T3N, TX, T3I, T17, T3M;
1302
0
       {
1303
0
            E TN, TP, T1b, T1f;
1304
0
            TN = cr[WS(rs, 4)];
1305
0
            TP = ci[WS(rs, 4)];
1306
0
            TQ = FMA(TM, TN, TO * TP);
1307
0
            T3H = FNMS(TO, TN, TM * TP);
1308
0
            T1b = cr[WS(rs, 12)];
1309
0
            T1f = ci[WS(rs, 12)];
1310
0
            T1g = FMA(T1a, T1b, T1e * T1f);
1311
0
            T3N = FNMS(T1e, T1b, T1a * T1f);
1312
0
       }
1313
0
       {
1314
0
            E TU, TW, T12, T16;
1315
0
            TU = cr[WS(rs, 20)];
1316
0
            TW = ci[WS(rs, 20)];
1317
0
            TX = FMA(TT, TU, TV * TW);
1318
0
            T3I = FNMS(TV, TU, TT * TW);
1319
0
            T12 = cr[WS(rs, 28)];
1320
0
            T16 = ci[WS(rs, 28)];
1321
0
            T17 = FMA(T11, T12, T15 * T16);
1322
0
            T3M = FNMS(T15, T12, T11 * T16);
1323
0
       }
1324
0
       {
1325
0
            E TY, T1h, T6g, T6h;
1326
0
            TY = TQ + TX;
1327
0
            T1h = T17 + T1g;
1328
0
            T1i = TY + T1h;
1329
0
            T7V = TY - T1h;
1330
0
            T6g = T3M + T3N;
1331
0
            T6h = T3H + T3I;
1332
0
            T6i = T6g - T6h;
1333
0
            T7D = T6h + T6g;
1334
0
       }
1335
0
       {
1336
0
            E T3G, T3J, T3L, T3O;
1337
0
            T3G = TQ - TX;
1338
0
            T3J = T3H - T3I;
1339
0
            T3K = T3G + T3J;
1340
0
            T5u = T3G - T3J;
1341
0
            T3L = T17 - T1g;
1342
0
            T3O = T3M - T3N;
1343
0
            T3P = T3L - T3O;
1344
0
            T5v = T3L + T3O;
1345
0
       }
1346
0
        }
1347
0
        {
1348
0
       E T1m, T3X, T1C, T3U, T1r, T3Y, T1x, T3T;
1349
0
       {
1350
0
            E T1k, T1l, T1z, T1B;
1351
0
            T1k = cr[WS(rs, 2)];
1352
0
            T1l = ci[WS(rs, 2)];
1353
0
            T1m = FMA(T8, T1k, Td * T1l);
1354
0
            T3X = FNMS(Td, T1k, T8 * T1l);
1355
0
            T1z = cr[WS(rs, 26)];
1356
0
            T1B = ci[WS(rs, 26)];
1357
0
            T1C = FMA(T1y, T1z, T1A * T1B);
1358
0
            T3U = FNMS(T1A, T1z, T1y * T1B);
1359
0
       }
1360
0
       {
1361
0
            E T1o, T1q, T1u, T1w;
1362
0
            T1o = cr[WS(rs, 18)];
1363
0
            T1q = ci[WS(rs, 18)];
1364
0
            T1r = FMA(T1n, T1o, T1p * T1q);
1365
0
            T3Y = FNMS(T1p, T1o, T1n * T1q);
1366
0
            T1u = cr[WS(rs, 10)];
1367
0
            T1w = ci[WS(rs, 10)];
1368
0
            T1x = FMA(T1t, T1u, T1v * T1w);
1369
0
            T3T = FNMS(T1v, T1u, T1t * T1w);
1370
0
       }
1371
0
       {
1372
0
            E T1s, T1D, T6l, T6m;
1373
0
            T1s = T1m + T1r;
1374
0
            T1D = T1x + T1C;
1375
0
            T1E = T1s + T1D;
1376
0
            T6k = T1s - T1D;
1377
0
            T6l = T3X + T3Y;
1378
0
            T6m = T3T + T3U;
1379
0
            T6n = T6l - T6m;
1380
0
            T7f = T6l + T6m;
1381
0
       }
1382
0
       {
1383
0
            E T3S, T3V, T3Z, T40;
1384
0
            T3S = T1m - T1r;
1385
0
            T3V = T3T - T3U;
1386
0
            T3W = T3S + T3V;
1387
0
            T5z = T3S - T3V;
1388
0
            T3Z = T3X - T3Y;
1389
0
            T40 = T1x - T1C;
1390
0
            T41 = T3Z - T40;
1391
0
            T5y = T3Z + T40;
1392
0
       }
1393
0
        }
1394
0
        {
1395
0
       E T1J, T43, T27, T4a, T1U, T44, T20, T49;
1396
0
       {
1397
0
            E T1G, T1I, T24, T26;
1398
0
            T1G = cr[WS(rs, 30)];
1399
0
            T1I = ci[WS(rs, 30)];
1400
0
            T1J = FMA(T1F, T1G, T1H * T1I);
1401
0
            T43 = FNMS(T1H, T1G, T1F * T1I);
1402
0
            T24 = cr[WS(rs, 22)];
1403
0
            T26 = ci[WS(rs, 22)];
1404
0
            T27 = FMA(T23, T24, T25 * T26);
1405
0
            T4a = FNMS(T25, T24, T23 * T26);
1406
0
       }
1407
0
       {
1408
0
            E T1R, T1T, T1X, T1Z;
1409
0
            T1R = cr[WS(rs, 14)];
1410
0
            T1T = ci[WS(rs, 14)];
1411
0
            T1U = FMA(T1Q, T1R, T1S * T1T);
1412
0
            T44 = FNMS(T1S, T1R, T1Q * T1T);
1413
0
            T1X = cr[WS(rs, 6)];
1414
0
            T1Z = ci[WS(rs, 6)];
1415
0
            T20 = FMA(T1W, T1X, T1Y * T1Z);
1416
0
            T49 = FNMS(T1Y, T1X, T1W * T1Z);
1417
0
       }
1418
0
       {
1419
0
            E T1V, T28, T6q, T6r;
1420
0
            T1V = T1J + T1U;
1421
0
            T28 = T20 + T27;
1422
0
            T29 = T1V + T28;
1423
0
            T6p = T1V - T28;
1424
0
            T6q = T43 + T44;
1425
0
            T6r = T49 + T4a;
1426
0
            T6s = T6q - T6r;
1427
0
            T7e = T6q + T6r;
1428
0
       }
1429
0
       {
1430
0
            E T45, T46, T48, T4b;
1431
0
            T45 = T43 - T44;
1432
0
            T46 = T20 - T27;
1433
0
            T47 = T45 - T46;
1434
0
            T5C = T45 + T46;
1435
0
            T48 = T1J - T1U;
1436
0
            T4b = T49 - T4a;
1437
0
            T4c = T48 + T4b;
1438
0
            T5B = T48 - T4b;
1439
0
       }
1440
0
        }
1441
0
        {
1442
0
       E T2B, T4m, T2G, T4n, T4l, T4o, T2M, T4q, T2P, T4r, T4s, T4t;
1443
0
       {
1444
0
            E T2z, T2A, T2D, T2F;
1445
0
            T2z = cr[WS(rs, 5)];
1446
0
            T2A = ci[WS(rs, 5)];
1447
0
            T2B = FMA(T21, T2z, T22 * T2A);
1448
0
            T4m = FNMS(T22, T2z, T21 * T2A);
1449
0
            T2D = cr[WS(rs, 21)];
1450
0
            T2F = ci[WS(rs, 21)];
1451
0
            T2G = FMA(T2C, T2D, T2E * T2F);
1452
0
            T4n = FNMS(T2E, T2D, T2C * T2F);
1453
0
       }
1454
0
       T4l = T2B - T2G;
1455
0
       T4o = T4m - T4n;
1456
0
       {
1457
0
            E T2J, T2L, T2N, T2O;
1458
0
            T2J = cr[WS(rs, 29)];
1459
0
            T2L = ci[WS(rs, 29)];
1460
0
            T2M = FMA(T2I, T2J, T2K * T2L);
1461
0
            T4q = FNMS(T2K, T2J, T2I * T2L);
1462
0
            T2N = cr[WS(rs, 13)];
1463
0
            T2O = ci[WS(rs, 13)];
1464
0
            T2P = FMA(T1M, T2N, T1P * T2O);
1465
0
            T4r = FNMS(T1P, T2N, T1M * T2O);
1466
0
       }
1467
0
       T4s = T4q - T4r;
1468
0
       T4t = T2M - T2P;
1469
0
       {
1470
0
            E T2H, T2Q, T6C, T6D;
1471
0
            T2H = T2B + T2G;
1472
0
            T2Q = T2M + T2P;
1473
0
            T2R = T2H + T2Q;
1474
0
            T6z = T2H - T2Q;
1475
0
            T6C = T4q + T4r;
1476
0
            T6D = T4m + T4n;
1477
0
            T6E = T6C - T6D;
1478
0
            T7k = T6D + T6C;
1479
0
       }
1480
0
       {
1481
0
            E T4p, T4u, T4C, T4D;
1482
0
            T4p = T4l + T4o;
1483
0
            T4u = T4s - T4t;
1484
0
            T4v = KP707106781 * (T4p - T4u);
1485
0
            T5K = KP707106781 * (T4p + T4u);
1486
0
            T4C = T4t + T4s;
1487
0
            T4D = T4l - T4o;
1488
0
            T4E = KP707106781 * (T4C - T4D);
1489
0
            T5H = KP707106781 * (T4D + T4C);
1490
0
       }
1491
0
        }
1492
0
        {
1493
0
       E T3k, T4S, T3p, T4T, T4R, T4U, T3t, T4N, T3w, T4O, T4M, T4P;
1494
0
       {
1495
0
            E T3i, T3j, T3m, T3o;
1496
0
            T3i = cr[WS(rs, 3)];
1497
0
            T3j = ci[WS(rs, 3)];
1498
0
            T3k = FMA(T3, T3i, T6 * T3j);
1499
0
            T4S = FNMS(T6, T3i, T3 * T3j);
1500
0
            T3m = cr[WS(rs, 19)];
1501
0
            T3o = ci[WS(rs, 19)];
1502
0
            T3p = FMA(T3l, T3m, T3n * T3o);
1503
0
            T4T = FNMS(T3n, T3m, T3l * T3o);
1504
0
       }
1505
0
       T4R = T3k - T3p;
1506
0
       T4U = T4S - T4T;
1507
0
       {
1508
0
            E T3r, T3s, T3u, T3v;
1509
0
            T3r = cr[WS(rs, 27)];
1510
0
            T3s = ci[WS(rs, 27)];
1511
0
            T3t = FMA(Th, T3r, Tl * T3s);
1512
0
            T4N = FNMS(Tl, T3r, Th * T3s);
1513
0
            T3u = cr[WS(rs, 11)];
1514
0
            T3v = ci[WS(rs, 11)];
1515
0
            T3w = FMA(Tg, T3u, Tk * T3v);
1516
0
            T4O = FNMS(Tk, T3u, Tg * T3v);
1517
0
       }
1518
0
       T4M = T3t - T3w;
1519
0
       T4P = T4N - T4O;
1520
0
       {
1521
0
            E T3q, T3x, T6I, T6J;
1522
0
            T3q = T3k + T3p;
1523
0
            T3x = T3t + T3w;
1524
0
            T3y = T3q + T3x;
1525
0
            T6P = T3q - T3x;
1526
0
            T6I = T4N + T4O;
1527
0
            T6J = T4S + T4T;
1528
0
            T6K = T6I - T6J;
1529
0
            T7p = T6J + T6I;
1530
0
       }
1531
0
       {
1532
0
            E T4Q, T4V, T53, T54;
1533
0
            T4Q = T4M + T4P;
1534
0
            T4V = T4R - T4U;
1535
0
            T4W = KP707106781 * (T4Q - T4V);
1536
0
            T5O = KP707106781 * (T4V + T4Q);
1537
0
            T53 = T4R + T4U;
1538
0
            T54 = T4P - T4M;
1539
0
            T55 = KP707106781 * (T53 - T54);
1540
0
            T5R = KP707106781 * (T53 + T54);
1541
0
       }
1542
0
        }
1543
0
        {
1544
0
       E T2b, T7x, T7K, T7M, T3A, T7L, T7A, T7B;
1545
0
       {
1546
0
            E T1j, T2a, T7C, T7J;
1547
0
            T1j = TL + T1i;
1548
0
            T2a = T1E + T29;
1549
0
            T2b = T1j + T2a;
1550
0
            T7x = T1j - T2a;
1551
0
            T7C = T7f + T7e;
1552
0
            T7J = T7D + T7I;
1553
0
            T7K = T7C + T7J;
1554
0
            T7M = T7J - T7C;
1555
0
       }
1556
0
       {
1557
0
            E T2S, T3z, T7y, T7z;
1558
0
            T2S = T2y + T2R;
1559
0
            T3z = T3h + T3y;
1560
0
            T3A = T2S + T3z;
1561
0
            T7L = T3z - T2S;
1562
0
            T7y = T7o + T7p;
1563
0
            T7z = T7j + T7k;
1564
0
            T7A = T7y - T7z;
1565
0
            T7B = T7z + T7y;
1566
0
       }
1567
0
       ci[WS(rs, 15)] = T2b - T3A;
1568
0
       cr[WS(rs, 24)] = T7L - T7M;
1569
0
       ci[WS(rs, 23)] = T7L + T7M;
1570
0
       cr[0] = T2b + T3A;
1571
0
       cr[WS(rs, 8)] = T7x - T7A;
1572
0
       cr[WS(rs, 16)] = T7B - T7K;
1573
0
       ci[WS(rs, 31)] = T7B + T7K;
1574
0
       ci[WS(rs, 7)] = T7x + T7A;
1575
0
        }
1576
0
        {
1577
0
       E T5x, T5Z, T8d, T8j, T5E, T88, T69, T6d, T5M, T5W, T62, T8i, T66, T6c, T5T;
1578
0
       E T5X, T5w, T89;
1579
0
       T5w = KP707106781 * (T5u + T5v);
1580
0
       T5x = T5t - T5w;
1581
0
       T5Z = T5t + T5w;
1582
0
       T89 = KP707106781 * (T3K - T3P);
1583
0
       T8d = T89 + T8c;
1584
0
       T8j = T8c - T89;
1585
0
       {
1586
0
            E T5A, T5D, T67, T68;
1587
0
            T5A = FMA(KP923879532, T5y, KP382683432 * T5z);
1588
0
            T5D = FNMS(KP923879532, T5C, KP382683432 * T5B);
1589
0
            T5E = T5A + T5D;
1590
0
            T88 = T5A - T5D;
1591
0
            T67 = T5N + T5O;
1592
0
            T68 = T5Q + T5R;
1593
0
            T69 = FNMS(KP980785280, T68, KP195090322 * T67);
1594
0
            T6d = FMA(KP980785280, T67, KP195090322 * T68);
1595
0
       }
1596
0
       {
1597
0
            E T5I, T5L, T60, T61;
1598
0
            T5I = T5G - T5H;
1599
0
            T5L = T5J - T5K;
1600
0
            T5M = FMA(KP831469612, T5I, KP555570233 * T5L);
1601
0
            T5W = FNMS(KP831469612, T5L, KP555570233 * T5I);
1602
0
            T60 = FNMS(KP382683432, T5y, KP923879532 * T5z);
1603
0
            T61 = FMA(KP382683432, T5C, KP923879532 * T5B);
1604
0
            T62 = T60 + T61;
1605
0
            T8i = T61 - T60;
1606
0
       }
1607
0
       {
1608
0
            E T64, T65, T5P, T5S;
1609
0
            T64 = T5G + T5H;
1610
0
            T65 = T5J + T5K;
1611
0
            T66 = FMA(KP195090322, T64, KP980785280 * T65);
1612
0
            T6c = FNMS(KP195090322, T65, KP980785280 * T64);
1613
0
            T5P = T5N - T5O;
1614
0
            T5S = T5Q - T5R;
1615
0
            T5T = FNMS(KP555570233, T5S, KP831469612 * T5P);
1616
0
            T5X = FMA(KP555570233, T5P, KP831469612 * T5S);
1617
0
       }
1618
0
       {
1619
0
            E T5F, T5U, T8h, T8k;
1620
0
            T5F = T5x + T5E;
1621
0
            T5U = T5M + T5T;
1622
0
            ci[WS(rs, 12)] = T5F - T5U;
1623
0
            cr[WS(rs, 3)] = T5F + T5U;
1624
0
            T8h = T5X - T5W;
1625
0
            T8k = T8i + T8j;
1626
0
            cr[WS(rs, 19)] = T8h - T8k;
1627
0
            ci[WS(rs, 28)] = T8h + T8k;
1628
0
       }
1629
0
       {
1630
0
            E T8l, T8m, T5V, T5Y;
1631
0
            T8l = T5T - T5M;
1632
0
            T8m = T8j - T8i;
1633
0
            cr[WS(rs, 27)] = T8l - T8m;
1634
0
            ci[WS(rs, 20)] = T8l + T8m;
1635
0
            T5V = T5x - T5E;
1636
0
            T5Y = T5W + T5X;
1637
0
            cr[WS(rs, 11)] = T5V - T5Y;
1638
0
            ci[WS(rs, 4)] = T5V + T5Y;
1639
0
       }
1640
0
       {
1641
0
            E T63, T6a, T87, T8e;
1642
0
            T63 = T5Z - T62;
1643
0
            T6a = T66 + T69;
1644
0
            ci[WS(rs, 8)] = T63 - T6a;
1645
0
            cr[WS(rs, 7)] = T63 + T6a;
1646
0
            T87 = T69 - T66;
1647
0
            T8e = T88 + T8d;
1648
0
            cr[WS(rs, 31)] = T87 - T8e;
1649
0
            ci[WS(rs, 16)] = T87 + T8e;
1650
0
       }
1651
0
       {
1652
0
            E T8f, T8g, T6b, T6e;
1653
0
            T8f = T6d - T6c;
1654
0
            T8g = T8d - T88;
1655
0
            cr[WS(rs, 23)] = T8f - T8g;
1656
0
            ci[WS(rs, 24)] = T8f + T8g;
1657
0
            T6b = T5Z + T62;
1658
0
            T6e = T6c + T6d;
1659
0
            cr[WS(rs, 15)] = T6b - T6e;
1660
0
            ci[0] = T6b + T6e;
1661
0
       }
1662
0
        }
1663
0
        {
1664
0
       E T7h, T7t, T7Q, T7S, T7m, T7u, T7r, T7v;
1665
0
       {
1666
0
            E T7d, T7g, T7O, T7P;
1667
0
            T7d = TL - T1i;
1668
0
            T7g = T7e - T7f;
1669
0
            T7h = T7d - T7g;
1670
0
            T7t = T7d + T7g;
1671
0
            T7O = T1E - T29;
1672
0
            T7P = T7I - T7D;
1673
0
            T7Q = T7O + T7P;
1674
0
            T7S = T7P - T7O;
1675
0
       }
1676
0
       {
1677
0
            E T7i, T7l, T7n, T7q;
1678
0
            T7i = T2y - T2R;
1679
0
            T7l = T7j - T7k;
1680
0
            T7m = T7i + T7l;
1681
0
            T7u = T7i - T7l;
1682
0
            T7n = T3h - T3y;
1683
0
            T7q = T7o - T7p;
1684
0
            T7r = T7n - T7q;
1685
0
            T7v = T7n + T7q;
1686
0
       }
1687
0
       {
1688
0
            E T7s, T7R, T7w, T7N;
1689
0
            T7s = KP707106781 * (T7m + T7r);
1690
0
            ci[WS(rs, 11)] = T7h - T7s;
1691
0
            cr[WS(rs, 4)] = T7h + T7s;
1692
0
            T7R = KP707106781 * (T7v - T7u);
1693
0
            cr[WS(rs, 20)] = T7R - T7S;
1694
0
            ci[WS(rs, 27)] = T7R + T7S;
1695
0
            T7w = KP707106781 * (T7u + T7v);
1696
0
            cr[WS(rs, 12)] = T7t - T7w;
1697
0
            ci[WS(rs, 3)] = T7t + T7w;
1698
0
            T7N = KP707106781 * (T7r - T7m);
1699
0
            cr[WS(rs, 28)] = T7N - T7Q;
1700
0
            ci[WS(rs, 19)] = T7N + T7Q;
1701
0
       }
1702
0
        }
1703
0
        {
1704
0
       E T6j, T7X, T83, T6X, T6u, T7U, T77, T7b, T70, T82, T6G, T6U, T74, T7a, T6R;
1705
0
       E T6V;
1706
0
       {
1707
0
            E T6o, T6t, T6A, T6F;
1708
0
            T6j = T6f - T6i;
1709
0
            T7X = T7V + T7W;
1710
0
            T83 = T7W - T7V;
1711
0
            T6X = T6f + T6i;
1712
0
            T6o = T6k + T6n;
1713
0
            T6t = T6p - T6s;
1714
0
            T6u = KP707106781 * (T6o + T6t);
1715
0
            T7U = KP707106781 * (T6o - T6t);
1716
0
            {
1717
0
           E T75, T76, T6Y, T6Z;
1718
0
           T75 = T6O + T6P;
1719
0
           T76 = T6H + T6K;
1720
0
           T77 = FMA(KP382683432, T75, KP923879532 * T76);
1721
0
           T7b = FNMS(KP923879532, T75, KP382683432 * T76);
1722
0
           T6Y = T6k - T6n;
1723
0
           T6Z = T6p + T6s;
1724
0
           T70 = KP707106781 * (T6Y + T6Z);
1725
0
           T82 = KP707106781 * (T6Z - T6Y);
1726
0
            }
1727
0
            T6A = T6y - T6z;
1728
0
            T6F = T6B - T6E;
1729
0
            T6G = FMA(KP382683432, T6A, KP923879532 * T6F);
1730
0
            T6U = FNMS(KP923879532, T6A, KP382683432 * T6F);
1731
0
            {
1732
0
           E T72, T73, T6L, T6Q;
1733
0
           T72 = T6B + T6E;
1734
0
           T73 = T6y + T6z;
1735
0
           T74 = FNMS(KP382683432, T73, KP923879532 * T72);
1736
0
           T7a = FMA(KP923879532, T73, KP382683432 * T72);
1737
0
           T6L = T6H - T6K;
1738
0
           T6Q = T6O - T6P;
1739
0
           T6R = FNMS(KP382683432, T6Q, KP923879532 * T6L);
1740
0
           T6V = FMA(KP923879532, T6Q, KP382683432 * T6L);
1741
0
            }
1742
0
       }
1743
0
       {
1744
0
            E T6v, T6S, T81, T84;
1745
0
            T6v = T6j + T6u;
1746
0
            T6S = T6G + T6R;
1747
0
            ci[WS(rs, 13)] = T6v - T6S;
1748
0
            cr[WS(rs, 2)] = T6v + T6S;
1749
0
            T81 = T6V - T6U;
1750
0
            T84 = T82 + T83;
1751
0
            cr[WS(rs, 18)] = T81 - T84;
1752
0
            ci[WS(rs, 29)] = T81 + T84;
1753
0
       }
1754
0
       {
1755
0
            E T85, T86, T6T, T6W;
1756
0
            T85 = T6R - T6G;
1757
0
            T86 = T83 - T82;
1758
0
            cr[WS(rs, 26)] = T85 - T86;
1759
0
            ci[WS(rs, 21)] = T85 + T86;
1760
0
            T6T = T6j - T6u;
1761
0
            T6W = T6U + T6V;
1762
0
            cr[WS(rs, 10)] = T6T - T6W;
1763
0
            ci[WS(rs, 5)] = T6T + T6W;
1764
0
       }
1765
0
       {
1766
0
            E T71, T78, T7T, T7Y;
1767
0
            T71 = T6X + T70;
1768
0
            T78 = T74 + T77;
1769
0
            cr[WS(rs, 14)] = T71 - T78;
1770
0
            ci[WS(rs, 1)] = T71 + T78;
1771
0
            T7T = T7b - T7a;
1772
0
            T7Y = T7U + T7X;
1773
0
            cr[WS(rs, 30)] = T7T - T7Y;
1774
0
            ci[WS(rs, 17)] = T7T + T7Y;
1775
0
       }
1776
0
       {
1777
0
            E T7Z, T80, T79, T7c;
1778
0
            T7Z = T77 - T74;
1779
0
            T80 = T7X - T7U;
1780
0
            cr[WS(rs, 22)] = T7Z - T80;
1781
0
            ci[WS(rs, 25)] = T7Z + T80;
1782
0
            T79 = T6X - T70;
1783
0
            T7c = T7a + T7b;
1784
0
            ci[WS(rs, 9)] = T79 - T7c;
1785
0
            cr[WS(rs, 6)] = T79 + T7c;
1786
0
       }
1787
0
        }
1788
0
        {
1789
0
       E T3R, T5d, T8r, T8x, T4e, T8o, T5n, T5r, T4G, T5a, T5g, T8w, T5k, T5q, T57;
1790
0
       E T5b, T3Q, T8p;
1791
0
       T3Q = KP707106781 * (T3K + T3P);
1792
0
       T3R = T3F - T3Q;
1793
0
       T5d = T3F + T3Q;
1794
0
       T8p = KP707106781 * (T5v - T5u);
1795
0
       T8r = T8p + T8q;
1796
0
       T8x = T8q - T8p;
1797
0
       {
1798
0
            E T42, T4d, T5l, T5m;
1799
0
            T42 = FNMS(KP923879532, T41, KP382683432 * T3W);
1800
0
            T4d = FMA(KP923879532, T47, KP382683432 * T4c);
1801
0
            T4e = T42 + T4d;
1802
0
            T8o = T4d - T42;
1803
0
            T5l = T52 + T55;
1804
0
            T5m = T4L + T4W;
1805
0
            T5n = FNMS(KP195090322, T5m, KP980785280 * T5l);
1806
0
            T5r = FMA(KP980785280, T5m, KP195090322 * T5l);
1807
0
       }
1808
0
       {
1809
0
            E T4w, T4F, T5e, T5f;
1810
0
            T4w = T4k - T4v;
1811
0
            T4F = T4B - T4E;
1812
0
            T4G = FNMS(KP555570233, T4F, KP831469612 * T4w);
1813
0
            T5a = FMA(KP831469612, T4F, KP555570233 * T4w);
1814
0
            T5e = FMA(KP382683432, T41, KP923879532 * T3W);
1815
0
            T5f = FNMS(KP382683432, T47, KP923879532 * T4c);
1816
0
            T5g = T5e + T5f;
1817
0
            T8w = T5e - T5f;
1818
0
       }
1819
0
       {
1820
0
            E T5i, T5j, T4X, T56;
1821
0
            T5i = T4B + T4E;
1822
0
            T5j = T4k + T4v;
1823
0
            T5k = FMA(KP195090322, T5i, KP980785280 * T5j);
1824
0
            T5q = FNMS(KP980785280, T5i, KP195090322 * T5j);
1825
0
            T4X = T4L - T4W;
1826
0
            T56 = T52 - T55;
1827
0
            T57 = FMA(KP555570233, T4X, KP831469612 * T56);
1828
0
            T5b = FNMS(KP831469612, T4X, KP555570233 * T56);
1829
0
       }
1830
0
       {
1831
0
            E T4f, T58, T8v, T8y;
1832
0
            T4f = T3R + T4e;
1833
0
            T58 = T4G + T57;
1834
0
            cr[WS(rs, 13)] = T4f - T58;
1835
0
            ci[WS(rs, 2)] = T4f + T58;
1836
0
            T8v = T5b - T5a;
1837
0
            T8y = T8w + T8x;
1838
0
            cr[WS(rs, 29)] = T8v - T8y;
1839
0
            ci[WS(rs, 18)] = T8v + T8y;
1840
0
       }
1841
0
       {
1842
0
            E T8z, T8A, T59, T5c;
1843
0
            T8z = T57 - T4G;
1844
0
            T8A = T8x - T8w;
1845
0
            cr[WS(rs, 21)] = T8z - T8A;
1846
0
            ci[WS(rs, 26)] = T8z + T8A;
1847
0
            T59 = T3R - T4e;
1848
0
            T5c = T5a + T5b;
1849
0
            ci[WS(rs, 10)] = T59 - T5c;
1850
0
            cr[WS(rs, 5)] = T59 + T5c;
1851
0
       }
1852
0
       {
1853
0
            E T5h, T5o, T8n, T8s;
1854
0
            T5h = T5d + T5g;
1855
0
            T5o = T5k + T5n;
1856
0
            ci[WS(rs, 14)] = T5h - T5o;
1857
0
            cr[WS(rs, 1)] = T5h + T5o;
1858
0
            T8n = T5r - T5q;
1859
0
            T8s = T8o + T8r;
1860
0
            cr[WS(rs, 17)] = T8n - T8s;
1861
0
            ci[WS(rs, 30)] = T8n + T8s;
1862
0
       }
1863
0
       {
1864
0
            E T8t, T8u, T5p, T5s;
1865
0
            T8t = T5n - T5k;
1866
0
            T8u = T8r - T8o;
1867
0
            cr[WS(rs, 25)] = T8t - T8u;
1868
0
            ci[WS(rs, 22)] = T8t + T8u;
1869
0
            T5p = T5d - T5g;
1870
0
            T5s = T5q + T5r;
1871
0
            cr[WS(rs, 9)] = T5p - T5s;
1872
0
            ci[WS(rs, 6)] = T5p + T5s;
1873
0
       }
1874
0
        }
1875
0
         }
1876
0
    }
1877
0
     }
1878
0
}
1879
1880
static const tw_instr twinstr[] = {
1881
     { TW_CEXP, 1, 1 },
1882
     { TW_CEXP, 1, 3 },
1883
     { TW_CEXP, 1, 9 },
1884
     { TW_CEXP, 1, 27 },
1885
     { TW_NEXT, 1, 0 }
1886
};
1887
1888
static const hc2hc_desc desc = { 32, "hf2_32", twinstr, &GENUS, { 376, 168, 112, 0 } };
1889
1890
1
void X(codelet_hf2_32) (planner *p) {
1891
1
     X(khc2hc_register) (p, hf2_32, &desc);
1892
1
}
1893
#endif