Coverage Report

Created: 2025-07-11 06:55

/src/fftw3/rdft/scalar/r2cb/hc2cb2_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 Fri Jul 11 06:54:25 UTC 2025 */
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_hc2c.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 32 -dif -name hc2cb2_32 -include rdft/scalar/hc2cb.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/hc2cb.h"
36
37
static void hc2cb2_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
38
{
39
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
40
     DK(KP198912367, +0.198912367379658006911597622644676228597850501);
41
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
42
     DK(KP668178637, +0.668178637919298919997757686523080761552472251);
43
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
44
     DK(KP414213562, +0.414213562373095048801688724209698078569671875);
45
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
46
     {
47
    INT m;
48
    for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 8, MAKE_VOLATILE_STRIDE(128, rs)) {
49
         E T11, T14, T12, T37, T17, T1b, T39, T3a, T3v, T3d, T3x, T15, T16, T5X, T4p;
50
         E T3G, T78, T7e, T8S, T9s, T8P, T8V, T98, T9m, T7I, T7C, T3y, T4b, T3C, T4g;
51
         E T5u, T6b, T5I, T6e, T1a, T1c, T2O, T4r, T4s, T4W, T3J, T3K, T3Y, T5Z, T60;
52
         E T66, T3i, T3q, T3l, T3e, T7S, T8K, T8m, T8E, T5k, T5U, T5R, T5e, T6i, T7s;
53
         E T6O, T7o;
54
         {
55
        E T77, T9l, T7B, T7d, T9r, T7H, T3b, T5d, T19, T3I;
56
        {
57
       E T13, T3F, T38, T3c;
58
       T11 = W[2];
59
       T14 = W[3];
60
       T12 = W[4];
61
       T37 = W[0];
62
       T13 = T11 * T12;
63
       T3F = T37 * T12;
64
       T38 = T37 * T11;
65
       T3c = T37 * T14;
66
       T17 = W[6];
67
       T77 = T37 * T17;
68
       T9l = T12 * T17;
69
       T7B = T11 * T17;
70
       T1b = W[7];
71
       T7d = T37 * T1b;
72
       T9r = T12 * T1b;
73
       T7H = T11 * T1b;
74
       T39 = W[1];
75
       T3a = FNMS(T39, T14, T38);
76
       T3v = FMA(T39, T14, T38);
77
       T3b = T3a * T12;
78
       T5d = T3v * T12;
79
       T3d = FMA(T39, T11, T3c);
80
       T3x = FNMS(T39, T11, T3c);
81
       T15 = W[5];
82
       T19 = T11 * T15;
83
       T3I = T37 * T15;
84
       T16 = FMA(T14, T15, T13);
85
       T5X = FNMS(T14, T15, T13);
86
       T4p = FMA(T39, T15, T3F);
87
       T3G = FNMS(T39, T15, T3F);
88
        }
89
        T78 = FNMS(T39, T1b, T77);
90
        T7e = FMA(T39, T17, T7d);
91
        T8S = FMA(T14, T17, T7H);
92
        T9s = FNMS(T15, T17, T9r);
93
        T8P = FNMS(T14, T1b, T7B);
94
        T8V = FMA(T39, T1b, T77);
95
        T98 = FNMS(T39, T17, T7d);
96
        T9m = FMA(T15, T1b, T9l);
97
        T7I = FNMS(T14, T17, T7H);
98
        T7C = FMA(T14, T1b, T7B);
99
        {
100
       E T3w, T3B, T5Y, T65;
101
       T3w = T3v * T17;
102
       T3y = FNMS(T3x, T1b, T3w);
103
       T4b = FMA(T3x, T1b, T3w);
104
       T3B = T3v * T1b;
105
       T3C = FMA(T3x, T17, T3B);
106
       T4g = FNMS(T3x, T17, T3B);
107
       {
108
            E T5t, T5H, T18, T2N;
109
            T5t = T3a * T17;
110
            T5u = FMA(T3d, T1b, T5t);
111
            T6b = FNMS(T3d, T1b, T5t);
112
            T5H = T3a * T1b;
113
            T5I = FNMS(T3d, T17, T5H);
114
            T6e = FMA(T3d, T17, T5H);
115
            T18 = T16 * T17;
116
            T2N = T16 * T1b;
117
            T1a = FNMS(T14, T12, T19);
118
            T1c = FMA(T1a, T1b, T18);
119
            T2O = FNMS(T1a, T17, T2N);
120
       }
121
       {
122
            E T4q, T4V, T3H, T3X;
123
            T4q = T4p * T17;
124
            T4V = T4p * T1b;
125
            T4r = FNMS(T39, T12, T3I);
126
            T4s = FMA(T4r, T1b, T4q);
127
            T4W = FNMS(T4r, T17, T4V);
128
            T3H = T3G * T17;
129
            T3X = T3G * T1b;
130
            T3J = FMA(T39, T12, T3I);
131
            T3K = FMA(T3J, T1b, T3H);
132
            T3Y = FNMS(T3J, T17, T3X);
133
       }
134
       T5Y = T5X * T17;
135
       T65 = T5X * T1b;
136
       T5Z = FMA(T14, T12, T19);
137
       T60 = FMA(T5Z, T1b, T5Y);
138
       T66 = FNMS(T5Z, T17, T65);
139
       {
140
            E T8D, T8J, T7R, T8l, T3h;
141
            T3h = T3a * T15;
142
            T3i = FNMS(T3d, T12, T3h);
143
            T3q = FMA(T3d, T12, T3h);
144
            T3l = FNMS(T3d, T15, T3b);
145
            T8D = T3l * T17;
146
            T8J = T3l * T1b;
147
            T3e = FMA(T3d, T15, T3b);
148
            T7R = T3e * T17;
149
            T8l = T3e * T1b;
150
            T7S = FMA(T3i, T1b, T7R);
151
            T8K = FNMS(T3q, T17, T8J);
152
            T8m = FNMS(T3i, T17, T8l);
153
            T8E = FMA(T3q, T1b, T8D);
154
       }
155
       {
156
            E T6h, T6N, T7n, T7r, T5j;
157
            T5j = T3v * T15;
158
            T5k = FMA(T3x, T12, T5j);
159
            T5U = FNMS(T3x, T12, T5j);
160
            T5R = FMA(T3x, T15, T5d);
161
            T6h = T5R * T17;
162
            T6N = T5R * T1b;
163
            T5e = FNMS(T3x, T15, T5d);
164
            T7n = T5e * T17;
165
            T7r = T5e * T1b;
166
            T6i = FMA(T5U, T1b, T6h);
167
            T7s = FNMS(T5k, T17, T7r);
168
            T6O = FNMS(T5U, T17, T6N);
169
            T7o = FMA(T5k, T1b, T7n);
170
       }
171
        }
172
         }
173
         {
174
        E Tf, T6j, T7V, T8W, T8p, T99, T1t, T3L, T2X, T3Z, T4Z, T5J, T6W, T7t, T4v;
175
        E T5v, TZ, T7x, T28, T3S, T91, T9d, T2h, T3R, T4Q, T5B, T8a, T8v, T4N, T5C;
176
        E T6J, T6Z, TK, T7w, T2z, T3P, T94, T9c, T2I, T3O, T4J, T5y, T8h, T8u, T4G;
177
        E T5z, T6A, T6Y, Tu, T6P, T82, T9a, T8s, T8X, T1Q, T40, T30, T3M, T52, T5w;
178
        E T6q, T7u, T4C, T5K;
179
        {
180
       E T3, T1d, T2S, T6Q, T6, T2P, T1g, T6R, Td, T6U, T1r, T2V, Ta, T6T, T1m;
181
       E T2U;
182
       {
183
            E T1, T2, T1e, T1f;
184
            T1 = Rp[0];
185
            T2 = Rm[WS(rs, 15)];
186
            T3 = T1 + T2;
187
            T1d = T1 - T2;
188
            {
189
           E T2Q, T2R, T4, T5;
190
           T2Q = Ip[0];
191
           T2R = Im[WS(rs, 15)];
192
           T2S = T2Q + T2R;
193
           T6Q = T2Q - T2R;
194
           T4 = Rp[WS(rs, 8)];
195
           T5 = Rm[WS(rs, 7)];
196
           T6 = T4 + T5;
197
           T2P = T4 - T5;
198
            }
199
            T1e = Ip[WS(rs, 8)];
200
            T1f = Im[WS(rs, 7)];
201
            T1g = T1e + T1f;
202
            T6R = T1e - T1f;
203
            {
204
           E Tb, Tc, T1n, T1o, T1p, T1q;
205
           Tb = Rm[WS(rs, 3)];
206
           Tc = Rp[WS(rs, 12)];
207
           T1n = Tb - Tc;
208
           T1o = Ip[WS(rs, 12)];
209
           T1p = Im[WS(rs, 3)];
210
           T1q = T1o + T1p;
211
           Td = Tb + Tc;
212
           T6U = T1o - T1p;
213
           T1r = T1n - T1q;
214
           T2V = T1n + T1q;
215
            }
216
            {
217
           E T8, T9, T1i, T1j, T1k, T1l;
218
           T8 = Rp[WS(rs, 4)];
219
           T9 = Rm[WS(rs, 11)];
220
           T1i = T8 - T9;
221
           T1j = Ip[WS(rs, 4)];
222
           T1k = Im[WS(rs, 11)];
223
           T1l = T1j + T1k;
224
           Ta = T8 + T9;
225
           T6T = T1j - T1k;
226
           T1m = T1i - T1l;
227
           T2U = T1i + T1l;
228
            }
229
       }
230
       {
231
            E T7, Te, T7T, T7U;
232
            T7 = T3 + T6;
233
            Te = Ta + Td;
234
            Tf = T7 + Te;
235
            T6j = T7 - Te;
236
            T7T = T3 - T6;
237
            T7U = T6U - T6T;
238
            T7V = T7T - T7U;
239
            T8W = T7T + T7U;
240
       }
241
       {
242
            E T8n, T8o, T1h, T1s;
243
            T8n = T6Q - T6R;
244
            T8o = Ta - Td;
245
            T8p = T8n - T8o;
246
            T99 = T8o + T8n;
247
            T1h = T1d - T1g;
248
            T1s = T1m + T1r;
249
            T1t = FNMS(KP707106781, T1s, T1h);
250
            T3L = FMA(KP707106781, T1s, T1h);
251
       }
252
       {
253
            E T2T, T2W, T4X, T4Y;
254
            T2T = T2P + T2S;
255
            T2W = T2U - T2V;
256
            T2X = FNMS(KP707106781, T2W, T2T);
257
            T3Z = FMA(KP707106781, T2W, T2T);
258
            T4X = T2S - T2P;
259
            T4Y = T1m - T1r;
260
            T4Z = FMA(KP707106781, T4Y, T4X);
261
            T5J = FNMS(KP707106781, T4Y, T4X);
262
       }
263
       {
264
            E T6S, T6V, T4t, T4u;
265
            T6S = T6Q + T6R;
266
            T6V = T6T + T6U;
267
            T6W = T6S - T6V;
268
            T7t = T6S + T6V;
269
            T4t = T1d + T1g;
270
            T4u = T2U + T2V;
271
            T4v = FNMS(KP707106781, T4u, T4t);
272
            T5v = FMA(KP707106781, T4u, T4t);
273
       }
274
        }
275
        {
276
       E TR, T87, T1S, T29, T1V, T84, T2c, T6E, TY, T85, T88, T21, T26, T2f, T6H;
277
       E T2e, T86, T89;
278
       {
279
            E TL, TM, TN, TO, TP, TQ;
280
            TL = Rm[0];
281
            TM = Rp[WS(rs, 15)];
282
            TN = TL + TM;
283
            TO = Rp[WS(rs, 7)];
284
            TP = Rm[WS(rs, 8)];
285
            TQ = TO + TP;
286
            TR = TN + TQ;
287
            T87 = TN - TQ;
288
            T1S = TO - TP;
289
            T29 = TL - TM;
290
       }
291
       {
292
            E T1T, T1U, T6C, T2a, T2b, T6D;
293
            T1T = Ip[WS(rs, 15)];
294
            T1U = Im[0];
295
            T6C = T1T - T1U;
296
            T2a = Ip[WS(rs, 7)];
297
            T2b = Im[WS(rs, 8)];
298
            T6D = T2a - T2b;
299
            T1V = T1T + T1U;
300
            T84 = T6C - T6D;
301
            T2c = T2a + T2b;
302
            T6E = T6C + T6D;
303
       }
304
       {
305
            E TU, T1X, T20, T6F, TX, T22, T25, T6G;
306
            {
307
           E TS, TT, T1Y, T1Z;
308
           TS = Rp[WS(rs, 3)];
309
           TT = Rm[WS(rs, 12)];
310
           TU = TS + TT;
311
           T1X = TS - TT;
312
           T1Y = Ip[WS(rs, 3)];
313
           T1Z = Im[WS(rs, 12)];
314
           T20 = T1Y + T1Z;
315
           T6F = T1Y - T1Z;
316
            }
317
            {
318
           E TV, TW, T23, T24;
319
           TV = Rm[WS(rs, 4)];
320
           TW = Rp[WS(rs, 11)];
321
           TX = TV + TW;
322
           T22 = TV - TW;
323
           T23 = Ip[WS(rs, 11)];
324
           T24 = Im[WS(rs, 4)];
325
           T25 = T23 + T24;
326
           T6G = T23 - T24;
327
            }
328
            TY = TU + TX;
329
            T85 = TU - TX;
330
            T88 = T6G - T6F;
331
            T21 = T1X + T20;
332
            T26 = T22 + T25;
333
            T2f = T22 - T25;
334
            T6H = T6F + T6G;
335
            T2e = T1X - T20;
336
       }
337
       TZ = TR + TY;
338
       T7x = T6E + T6H;
339
       {
340
            E T1W, T27, T8Z, T90;
341
            T1W = T1S - T1V;
342
            T27 = T21 - T26;
343
            T28 = FNMS(KP707106781, T27, T1W);
344
            T3S = FMA(KP707106781, T27, T1W);
345
            T8Z = T85 + T84;
346
            T90 = T87 + T88;
347
            T91 = FNMS(KP414213562, T90, T8Z);
348
            T9d = FMA(KP414213562, T8Z, T90);
349
       }
350
       {
351
            E T2d, T2g, T4O, T4P;
352
            T2d = T29 - T2c;
353
            T2g = T2e + T2f;
354
            T2h = FNMS(KP707106781, T2g, T2d);
355
            T3R = FMA(KP707106781, T2g, T2d);
356
            T4O = T1S + T1V;
357
            T4P = T2e - T2f;
358
            T4Q = FNMS(KP707106781, T4P, T4O);
359
            T5B = FMA(KP707106781, T4P, T4O);
360
       }
361
       T86 = T84 - T85;
362
       T89 = T87 - T88;
363
       T8a = FMA(KP414213562, T89, T86);
364
       T8v = FNMS(KP414213562, T86, T89);
365
       {
366
            E T4L, T4M, T6B, T6I;
367
            T4L = T29 + T2c;
368
            T4M = T21 + T26;
369
            T4N = FNMS(KP707106781, T4M, T4L);
370
            T5C = FMA(KP707106781, T4M, T4L);
371
            T6B = TR - TY;
372
            T6I = T6E - T6H;
373
            T6J = T6B + T6I;
374
            T6Z = T6I - T6B;
375
       }
376
        }
377
        {
378
       E TC, T8e, T2j, T2A, T2m, T8b, T2D, T6v, TJ, T8c, T8f, T2s, T2x, T2G, T6y;
379
       E T2F, T8d, T8g;
380
       {
381
            E Tw, Tx, Ty, Tz, TA, TB;
382
            Tw = Rp[WS(rs, 1)];
383
            Tx = Rm[WS(rs, 14)];
384
            Ty = Tw + Tx;
385
            Tz = Rp[WS(rs, 9)];
386
            TA = Rm[WS(rs, 6)];
387
            TB = Tz + TA;
388
            TC = Ty + TB;
389
            T8e = Ty - TB;
390
            T2j = Tz - TA;
391
            T2A = Tw - Tx;
392
       }
393
       {
394
            E T2k, T2l, T6t, T2B, T2C, T6u;
395
            T2k = Ip[WS(rs, 1)];
396
            T2l = Im[WS(rs, 14)];
397
            T6t = T2k - T2l;
398
            T2B = Ip[WS(rs, 9)];
399
            T2C = Im[WS(rs, 6)];
400
            T6u = T2B - T2C;
401
            T2m = T2k + T2l;
402
            T8b = T6t - T6u;
403
            T2D = T2B + T2C;
404
            T6v = T6t + T6u;
405
       }
406
       {
407
            E TF, T2o, T2r, T6w, TI, T2t, T2w, T6x;
408
            {
409
           E TD, TE, T2p, T2q;
410
           TD = Rp[WS(rs, 5)];
411
           TE = Rm[WS(rs, 10)];
412
           TF = TD + TE;
413
           T2o = TD - TE;
414
           T2p = Ip[WS(rs, 5)];
415
           T2q = Im[WS(rs, 10)];
416
           T2r = T2p + T2q;
417
           T6w = T2p - T2q;
418
            }
419
            {
420
           E TG, TH, T2u, T2v;
421
           TG = Rm[WS(rs, 2)];
422
           TH = Rp[WS(rs, 13)];
423
           TI = TG + TH;
424
           T2t = TG - TH;
425
           T2u = Ip[WS(rs, 13)];
426
           T2v = Im[WS(rs, 2)];
427
           T2w = T2u + T2v;
428
           T6x = T2u - T2v;
429
            }
430
            TJ = TF + TI;
431
            T8c = TF - TI;
432
            T8f = T6x - T6w;
433
            T2s = T2o + T2r;
434
            T2x = T2t + T2w;
435
            T2G = T2t - T2w;
436
            T6y = T6w + T6x;
437
            T2F = T2o - T2r;
438
       }
439
       TK = TC + TJ;
440
       T7w = T6v + T6y;
441
       {
442
            E T2n, T2y, T92, T93;
443
            T2n = T2j + T2m;
444
            T2y = T2s - T2x;
445
            T2z = FNMS(KP707106781, T2y, T2n);
446
            T3P = FMA(KP707106781, T2y, T2n);
447
            T92 = T8c + T8b;
448
            T93 = T8e + T8f;
449
            T94 = FMA(KP414213562, T93, T92);
450
            T9c = FNMS(KP414213562, T92, T93);
451
       }
452
       {
453
            E T2E, T2H, T4H, T4I;
454
            T2E = T2A - T2D;
455
            T2H = T2F + T2G;
456
            T2I = FNMS(KP707106781, T2H, T2E);
457
            T3O = FMA(KP707106781, T2H, T2E);
458
            T4H = T2m - T2j;
459
            T4I = T2G - T2F;
460
            T4J = FNMS(KP707106781, T4I, T4H);
461
            T5y = FMA(KP707106781, T4I, T4H);
462
       }
463
       T8d = T8b - T8c;
464
       T8g = T8e - T8f;
465
       T8h = FNMS(KP414213562, T8g, T8d);
466
       T8u = FMA(KP414213562, T8d, T8g);
467
       {
468
            E T4E, T4F, T6s, T6z;
469
            T4E = T2A + T2D;
470
            T4F = T2s + T2x;
471
            T4G = FNMS(KP707106781, T4F, T4E);
472
            T5z = FMA(KP707106781, T4F, T4E);
473
            T6s = TC - TJ;
474
            T6z = T6v - T6y;
475
            T6A = T6s - T6z;
476
            T6Y = T6s + T6z;
477
       }
478
        }
479
        {
480
       E Ti, T6n, Tl, T6o, T1J, T1O, T80, T7Z, T4x, T4w, Tp, T6k, Ts, T6l, T1y;
481
       E T1D, T7X, T7W, T4A, T4z;
482
       {
483
            E T1K, T1I, T1F, T1N;
484
            {
485
           E Tg, Th, T1G, T1H;
486
           Tg = Rp[WS(rs, 2)];
487
           Th = Rm[WS(rs, 13)];
488
           Ti = Tg + Th;
489
           T1K = Tg - Th;
490
           T1G = Ip[WS(rs, 2)];
491
           T1H = Im[WS(rs, 13)];
492
           T1I = T1G + T1H;
493
           T6n = T1G - T1H;
494
            }
495
            {
496
           E Tj, Tk, T1L, T1M;
497
           Tj = Rp[WS(rs, 10)];
498
           Tk = Rm[WS(rs, 5)];
499
           Tl = Tj + Tk;
500
           T1F = Tj - Tk;
501
           T1L = Ip[WS(rs, 10)];
502
           T1M = Im[WS(rs, 5)];
503
           T1N = T1L + T1M;
504
           T6o = T1L - T1M;
505
            }
506
            T1J = T1F + T1I;
507
            T1O = T1K - T1N;
508
            T80 = T6n - T6o;
509
            T7Z = Ti - Tl;
510
            T4x = T1K + T1N;
511
            T4w = T1I - T1F;
512
       }
513
       {
514
            E T1z, T1x, T1u, T1C;
515
            {
516
           E Tn, To, T1v, T1w;
517
           Tn = Rm[WS(rs, 1)];
518
           To = Rp[WS(rs, 14)];
519
           Tp = Tn + To;
520
           T1z = Tn - To;
521
           T1v = Ip[WS(rs, 14)];
522
           T1w = Im[WS(rs, 1)];
523
           T1x = T1v + T1w;
524
           T6k = T1v - T1w;
525
            }
526
            {
527
           E Tq, Tr, T1A, T1B;
528
           Tq = Rp[WS(rs, 6)];
529
           Tr = Rm[WS(rs, 9)];
530
           Ts = Tq + Tr;
531
           T1u = Tq - Tr;
532
           T1A = Ip[WS(rs, 6)];
533
           T1B = Im[WS(rs, 9)];
534
           T1C = T1A + T1B;
535
           T6l = T1A - T1B;
536
            }
537
            T1y = T1u - T1x;
538
            T1D = T1z - T1C;
539
            T7X = Tp - Ts;
540
            T7W = T6k - T6l;
541
            T4A = T1z + T1C;
542
            T4z = T1u + T1x;
543
       }
544
       {
545
            E Tm, Tt, T7Y, T81;
546
            Tm = Ti + Tl;
547
            Tt = Tp + Ts;
548
            Tu = Tm + Tt;
549
            T6P = Tm - Tt;
550
            T7Y = T7W - T7X;
551
            T81 = T7Z + T80;
552
            T82 = T7Y - T81;
553
            T9a = T81 + T7Y;
554
       }
555
       {
556
            E T8q, T8r, T1E, T1P;
557
            T8q = T7Z - T80;
558
            T8r = T7X + T7W;
559
            T8s = T8q - T8r;
560
            T8X = T8q + T8r;
561
            T1E = FNMS(KP414213562, T1D, T1y);
562
            T1P = FMA(KP414213562, T1O, T1J);
563
            T1Q = T1E - T1P;
564
            T40 = T1P + T1E;
565
       }
566
       {
567
            E T2Y, T2Z, T50, T51;
568
            T2Y = FNMS(KP414213562, T1J, T1O);
569
            T2Z = FMA(KP414213562, T1y, T1D);
570
            T30 = T2Y - T2Z;
571
            T3M = T2Y + T2Z;
572
            T50 = FMA(KP414213562, T4w, T4x);
573
            T51 = FMA(KP414213562, T4z, T4A);
574
            T52 = T50 - T51;
575
            T5w = T50 + T51;
576
       }
577
       {
578
            E T6m, T6p, T4y, T4B;
579
            T6m = T6k + T6l;
580
            T6p = T6n + T6o;
581
            T6q = T6m - T6p;
582
            T7u = T6p + T6m;
583
            T4y = FNMS(KP414213562, T4x, T4w);
584
            T4B = FNMS(KP414213562, T4A, T4z);
585
            T4C = T4y + T4B;
586
            T5K = T4B - T4y;
587
       }
588
        }
589
        {
590
       E Tv, T10, T7p, T7v, T7y, T7z, T7q, T7A;
591
       Tv = Tf + Tu;
592
       T10 = TK + TZ;
593
       T7p = Tv - T10;
594
       T7v = T7t + T7u;
595
       T7y = T7w + T7x;
596
       T7z = T7v - T7y;
597
       Rp[0] = Tv + T10;
598
       Rm[0] = T7v + T7y;
599
       T7q = T7o * T7p;
600
       Rp[WS(rs, 8)] = FNMS(T7s, T7z, T7q);
601
       T7A = T7s * T7p;
602
       Rm[WS(rs, 8)] = FMA(T7o, T7z, T7A);
603
        }
604
        {
605
       E T9p, T9x, T9v, T9z;
606
       {
607
            E T9n, T9o, T9t, T9u;
608
            T9n = FMA(KP707106781, T8X, T8W);
609
            T9o = T9c + T9d;
610
            T9p = FNMS(KP923879532, T9o, T9n);
611
            T9x = FMA(KP923879532, T9o, T9n);
612
            T9t = FMA(KP707106781, T9a, T99);
613
            T9u = T94 + T91;
614
            T9v = FNMS(KP923879532, T9u, T9t);
615
            T9z = FMA(KP923879532, T9u, T9t);
616
       }
617
       {
618
            E T9q, T9w, T9y, T9A;
619
            T9q = T9m * T9p;
620
            Rp[WS(rs, 9)] = FNMS(T9s, T9v, T9q);
621
            T9w = T9m * T9v;
622
            Rm[WS(rs, 9)] = FMA(T9s, T9p, T9w);
623
            T9y = T3v * T9x;
624
            Rp[WS(rs, 1)] = FNMS(T3x, T9z, T9y);
625
            T9A = T3v * T9z;
626
            Rm[WS(rs, 1)] = FMA(T3x, T9x, T9A);
627
       }
628
        }
629
        {
630
       E T8H, T8Q, T8N, T8T;
631
       {
632
            E T8F, T8G, T8L, T8M;
633
            T8F = FNMS(KP707106781, T82, T7V);
634
            T8G = T8u + T8v;
635
            T8H = FNMS(KP923879532, T8G, T8F);
636
            T8Q = FMA(KP923879532, T8G, T8F);
637
            T8L = FNMS(KP707106781, T8s, T8p);
638
            T8M = T8h + T8a;
639
            T8N = FNMS(KP923879532, T8M, T8L);
640
            T8T = FMA(KP923879532, T8M, T8L);
641
       }
642
       {
643
            E T8I, T8O, T8R, T8U;
644
            T8I = T8E * T8H;
645
            Rp[WS(rs, 7)] = FNMS(T8K, T8N, T8I);
646
            T8O = T8E * T8N;
647
            Rm[WS(rs, 7)] = FMA(T8K, T8H, T8O);
648
            T8R = T8P * T8Q;
649
            Rp[WS(rs, 15)] = FNMS(T8S, T8T, T8R);
650
            T8U = T8P * T8T;
651
            Rm[WS(rs, 15)] = FMA(T8S, T8Q, T8U);
652
       }
653
        }
654
        {
655
       E T7b, T7j, T7h, T7l;
656
       {
657
            E T79, T7a, T7f, T7g;
658
            T79 = T6j - T6q;
659
            T7a = T6Z - T6Y;
660
            T7b = FNMS(KP707106781, T7a, T79);
661
            T7j = FMA(KP707106781, T7a, T79);
662
            T7f = T6W - T6P;
663
            T7g = T6A - T6J;
664
            T7h = FNMS(KP707106781, T7g, T7f);
665
            T7l = FMA(KP707106781, T7g, T7f);
666
       }
667
       {
668
            E T7c, T7i, T7k, T7m;
669
            T7c = T78 * T7b;
670
            Rp[WS(rs, 14)] = FNMS(T7e, T7h, T7c);
671
            T7i = T78 * T7h;
672
            Rm[WS(rs, 14)] = FMA(T7e, T7b, T7i);
673
            T7k = T5X * T7j;
674
            Rp[WS(rs, 6)] = FNMS(T5Z, T7l, T7k);
675
            T7m = T5X * T7l;
676
            Rm[WS(rs, 6)] = FMA(T5Z, T7j, T7m);
677
       }
678
        }
679
        {
680
       E T96, T9h, T9f, T9j;
681
       {
682
            E T8Y, T95, T9b, T9e;
683
            T8Y = FNMS(KP707106781, T8X, T8W);
684
            T95 = T91 - T94;
685
            T96 = FNMS(KP923879532, T95, T8Y);
686
            T9h = FMA(KP923879532, T95, T8Y);
687
            T9b = FNMS(KP707106781, T9a, T99);
688
            T9e = T9c - T9d;
689
            T9f = FNMS(KP923879532, T9e, T9b);
690
            T9j = FMA(KP923879532, T9e, T9b);
691
       }
692
       {
693
            E T97, T9g, T9i, T9k;
694
            T97 = T8V * T96;
695
            Rp[WS(rs, 13)] = FNMS(T98, T9f, T97);
696
            T9g = T98 * T96;
697
            Rm[WS(rs, 13)] = FMA(T8V, T9f, T9g);
698
            T9i = T3G * T9h;
699
            Rp[WS(rs, 5)] = FNMS(T3J, T9j, T9i);
700
            T9k = T3J * T9h;
701
            Rm[WS(rs, 5)] = FMA(T3G, T9j, T9k);
702
       }
703
        }
704
        {
705
       E T6L, T73, T71, T75;
706
       {
707
            E T6r, T6K, T6X, T70;
708
            T6r = T6j + T6q;
709
            T6K = T6A + T6J;
710
            T6L = FNMS(KP707106781, T6K, T6r);
711
            T73 = FMA(KP707106781, T6K, T6r);
712
            T6X = T6P + T6W;
713
            T70 = T6Y + T6Z;
714
            T71 = FNMS(KP707106781, T70, T6X);
715
            T75 = FMA(KP707106781, T70, T6X);
716
       }
717
       {
718
            E T6M, T72, T74, T76;
719
            T6M = T6i * T6L;
720
            Rp[WS(rs, 10)] = FNMS(T6O, T71, T6M);
721
            T72 = T6O * T6L;
722
            Rm[WS(rs, 10)] = FMA(T6i, T71, T72);
723
            T74 = T3a * T73;
724
            Rp[WS(rs, 2)] = FNMS(T3d, T75, T74);
725
            T76 = T3d * T73;
726
            Rm[WS(rs, 2)] = FMA(T3a, T75, T76);
727
       }
728
        }
729
        {
730
       E T7F, T7N, T7L, T7P;
731
       {
732
            E T7D, T7E, T7J, T7K;
733
            T7D = Tf - Tu;
734
            T7E = T7x - T7w;
735
            T7F = T7D - T7E;
736
            T7N = T7D + T7E;
737
            T7J = T7t - T7u;
738
            T7K = TK - TZ;
739
            T7L = T7J - T7K;
740
            T7P = T7K + T7J;
741
       }
742
       {
743
            E T7G, T7M, T7O, T7Q;
744
            T7G = T7C * T7F;
745
            Rp[WS(rs, 12)] = FNMS(T7I, T7L, T7G);
746
            T7M = T7C * T7L;
747
            Rm[WS(rs, 12)] = FMA(T7I, T7F, T7M);
748
            T7O = T4p * T7N;
749
            Rp[WS(rs, 4)] = FNMS(T4r, T7P, T7O);
750
            T7Q = T4p * T7P;
751
            Rm[WS(rs, 4)] = FMA(T4r, T7N, T7Q);
752
       }
753
        }
754
        {
755
       E T8j, T8z, T8x, T8B;
756
       {
757
            E T83, T8i, T8t, T8w;
758
            T83 = FMA(KP707106781, T82, T7V);
759
            T8i = T8a - T8h;
760
            T8j = FNMS(KP923879532, T8i, T83);
761
            T8z = FMA(KP923879532, T8i, T83);
762
            T8t = FMA(KP707106781, T8s, T8p);
763
            T8w = T8u - T8v;
764
            T8x = FNMS(KP923879532, T8w, T8t);
765
            T8B = FMA(KP923879532, T8w, T8t);
766
       }
767
       {
768
            E T8k, T8y, T8A, T8C;
769
            T8k = T7S * T8j;
770
            Rp[WS(rs, 11)] = FNMS(T8m, T8x, T8k);
771
            T8y = T8m * T8j;
772
            Rm[WS(rs, 11)] = FMA(T7S, T8x, T8y);
773
            T8A = T16 * T8z;
774
            Rp[WS(rs, 3)] = FNMS(T1a, T8B, T8A);
775
            T8C = T1a * T8z;
776
            Rm[WS(rs, 3)] = FMA(T16, T8B, T8C);
777
       }
778
        }
779
        {
780
       E T3r, T2L, T3s, T3f, T35, T3z, T3j, T3o;
781
       T3r = FNMS(KP923879532, T30, T2X);
782
       {
783
            E T1R, T2i, T2J, T2K;
784
            T1R = FMA(KP923879532, T1Q, T1t);
785
            T2i = FMA(KP668178637, T2h, T28);
786
            T2J = FNMS(KP668178637, T2I, T2z);
787
            T2K = T2i - T2J;
788
            T2L = FNMS(KP831469612, T2K, T1R);
789
            T3s = T2J + T2i;
790
            T3f = FMA(KP831469612, T2K, T1R);
791
       }
792
       {
793
            E T31, T3m, T34, T3n, T32, T33;
794
            T31 = FMA(KP923879532, T30, T2X);
795
            T3m = FNMS(KP923879532, T1Q, T1t);
796
            T32 = FMA(KP668178637, T2z, T2I);
797
            T33 = FNMS(KP668178637, T28, T2h);
798
            T34 = T32 - T33;
799
            T3n = T32 + T33;
800
            T35 = FNMS(KP831469612, T34, T31);
801
            T3z = FMA(KP831469612, T3n, T3m);
802
            T3j = FMA(KP831469612, T34, T31);
803
            T3o = FNMS(KP831469612, T3n, T3m);
804
       }
805
       {
806
            E T2M, T36, T3g, T3k;
807
            T2M = T1c * T2L;
808
            Ip[WS(rs, 10)] = FNMS(T2O, T35, T2M);
809
            T36 = T1c * T35;
810
            Im[WS(rs, 10)] = FMA(T2O, T2L, T36);
811
            T3g = T3e * T3f;
812
            Ip[WS(rs, 2)] = FNMS(T3i, T3j, T3g);
813
            T3k = T3e * T3j;
814
            Im[WS(rs, 2)] = FMA(T3i, T3f, T3k);
815
            {
816
           E T3A, T3E, T3D, T3p, T3u, T3t;
817
           T3A = T3y * T3z;
818
           T3E = T3C * T3z;
819
           T3D = FMA(KP831469612, T3s, T3r);
820
           Ip[WS(rs, 14)] = FNMS(T3C, T3D, T3A);
821
           Im[WS(rs, 14)] = FMA(T3y, T3D, T3E);
822
           T3p = T3l * T3o;
823
           T3u = T3q * T3o;
824
           T3t = FNMS(KP831469612, T3s, T3r);
825
           Ip[WS(rs, 6)] = FNMS(T3q, T3t, T3p);
826
           Im[WS(rs, 6)] = FMA(T3l, T3t, T3u);
827
            }
828
       }
829
        }
830
        {
831
       E T53, T56, T5p, T5h, T4T, T5r, T59, T5n;
832
       T53 = FMA(KP923879532, T52, T4Z);
833
       {
834
            E T5f, T54, T55, T5g;
835
            T5f = FMA(KP923879532, T4C, T4v);
836
            T54 = FMA(KP668178637, T4G, T4J);
837
            T55 = FMA(KP668178637, T4N, T4Q);
838
            T5g = T54 + T55;
839
            T56 = T54 - T55;
840
            T5p = FMA(KP831469612, T5g, T5f);
841
            T5h = FNMS(KP831469612, T5g, T5f);
842
       }
843
       {
844
            E T4D, T5l, T4S, T5m, T4K, T4R;
845
            T4D = FNMS(KP923879532, T4C, T4v);
846
            T5l = FNMS(KP923879532, T52, T4Z);
847
            T4K = FNMS(KP668178637, T4J, T4G);
848
            T4R = FNMS(KP668178637, T4Q, T4N);
849
            T4S = T4K + T4R;
850
            T5m = T4K - T4R;
851
            T4T = FNMS(KP831469612, T4S, T4D);
852
            T5r = FNMS(KP831469612, T5m, T5l);
853
            T59 = FMA(KP831469612, T4S, T4D);
854
            T5n = FMA(KP831469612, T5m, T5l);
855
       }
856
       {
857
            E T5i, T5o, T5q, T5s;
858
            T5i = T5e * T5h;
859
            Ip[WS(rs, 5)] = FNMS(T5k, T5n, T5i);
860
            T5o = T5e * T5n;
861
            Im[WS(rs, 5)] = FMA(T5k, T5h, T5o);
862
            T5q = T17 * T5p;
863
            Ip[WS(rs, 13)] = FNMS(T1b, T5r, T5q);
864
            T5s = T17 * T5r;
865
            Im[WS(rs, 13)] = FMA(T1b, T5p, T5s);
866
            {
867
           E T5a, T5c, T5b, T4U, T58, T57;
868
           T5a = T11 * T59;
869
           T5c = T14 * T59;
870
           T5b = FMA(KP831469612, T56, T53);
871
           Ip[WS(rs, 1)] = FNMS(T14, T5b, T5a);
872
           Im[WS(rs, 1)] = FMA(T11, T5b, T5c);
873
           T4U = T4s * T4T;
874
           T58 = T4W * T4T;
875
           T57 = FNMS(KP831469612, T56, T53);
876
           Ip[WS(rs, 9)] = FNMS(T4W, T57, T4U);
877
           Im[WS(rs, 9)] = FMA(T4s, T57, T58);
878
            }
879
       }
880
        }
881
        {
882
       E T41, T44, T4l, T4e, T3V, T4n, T47, T4j;
883
       T41 = FMA(KP923879532, T40, T3Z);
884
       {
885
            E T4c, T42, T43, T4d;
886
            T4c = FNMS(KP923879532, T3M, T3L);
887
            T42 = FMA(KP198912367, T3O, T3P);
888
            T43 = FNMS(KP198912367, T3R, T3S);
889
            T4d = T43 - T42;
890
            T44 = T42 + T43;
891
            T4l = FMA(KP980785280, T4d, T4c);
892
            T4e = FNMS(KP980785280, T4d, T4c);
893
       }
894
       {
895
            E T3N, T4h, T3U, T4i, T3Q, T3T;
896
            T3N = FMA(KP923879532, T3M, T3L);
897
            T4h = FNMS(KP923879532, T40, T3Z);
898
            T3Q = FNMS(KP198912367, T3P, T3O);
899
            T3T = FMA(KP198912367, T3S, T3R);
900
            T3U = T3Q + T3T;
901
            T4i = T3Q - T3T;
902
            T3V = FNMS(KP980785280, T3U, T3N);
903
            T4n = FMA(KP980785280, T4i, T4h);
904
            T47 = FMA(KP980785280, T3U, T3N);
905
            T4j = FNMS(KP980785280, T4i, T4h);
906
       }
907
       {
908
            E T4f, T4k, T4m, T4o;
909
            T4f = T4b * T4e;
910
            Ip[WS(rs, 12)] = FNMS(T4g, T4j, T4f);
911
            T4k = T4b * T4j;
912
            Im[WS(rs, 12)] = FMA(T4g, T4e, T4k);
913
            T4m = T12 * T4l;
914
            Ip[WS(rs, 4)] = FNMS(T15, T4n, T4m);
915
            T4o = T12 * T4n;
916
            Im[WS(rs, 4)] = FMA(T15, T4l, T4o);
917
            {
918
           E T48, T4a, T49, T3W, T46, T45;
919
           T48 = T37 * T47;
920
           T4a = T39 * T47;
921
           T49 = FMA(KP980785280, T44, T41);
922
           Ip[0] = FNMS(T39, T49, T48);
923
           Im[0] = FMA(T37, T49, T4a);
924
           T3W = T3K * T3V;
925
           T46 = T3Y * T3V;
926
           T45 = FNMS(KP980785280, T44, T41);
927
           Ip[WS(rs, 8)] = FNMS(T3Y, T45, T3W);
928
           Im[WS(rs, 8)] = FMA(T3K, T45, T46);
929
            }
930
       }
931
        }
932
        {
933
       E T5L, T5O, T6c, T63, T5F, T6f, T5S, T69;
934
       T5L = FMA(KP923879532, T5K, T5J);
935
       {
936
            E T61, T5M, T5N, T62;
937
            T61 = FMA(KP923879532, T5w, T5v);
938
            T5M = FMA(KP198912367, T5y, T5z);
939
            T5N = FMA(KP198912367, T5B, T5C);
940
            T62 = T5M + T5N;
941
            T5O = T5M - T5N;
942
            T6c = FMA(KP980785280, T62, T61);
943
            T63 = FNMS(KP980785280, T62, T61);
944
       }
945
       {
946
            E T5x, T67, T5E, T68, T5A, T5D;
947
            T5x = FNMS(KP923879532, T5w, T5v);
948
            T67 = FNMS(KP923879532, T5K, T5J);
949
            T5A = FNMS(KP198912367, T5z, T5y);
950
            T5D = FNMS(KP198912367, T5C, T5B);
951
            T5E = T5A + T5D;
952
            T68 = T5D - T5A;
953
            T5F = FMA(KP980785280, T5E, T5x);
954
            T6f = FNMS(KP980785280, T68, T67);
955
            T5S = FNMS(KP980785280, T5E, T5x);
956
            T69 = FMA(KP980785280, T68, T67);
957
       }
958
       {
959
            E T64, T6a, T6d, T6g;
960
            T64 = T60 * T63;
961
            Ip[WS(rs, 7)] = FNMS(T66, T69, T64);
962
            T6a = T60 * T69;
963
            Im[WS(rs, 7)] = FMA(T66, T63, T6a);
964
            T6d = T6b * T6c;
965
            Ip[WS(rs, 15)] = FNMS(T6e, T6f, T6d);
966
            T6g = T6b * T6f;
967
            Im[WS(rs, 15)] = FMA(T6e, T6c, T6g);
968
            {
969
           E T5T, T5W, T5V, T5G, T5Q, T5P;
970
           T5T = T5R * T5S;
971
           T5W = T5U * T5S;
972
           T5V = FMA(KP980785280, T5O, T5L);
973
           Ip[WS(rs, 3)] = FNMS(T5U, T5V, T5T);
974
           Im[WS(rs, 3)] = FMA(T5R, T5V, T5W);
975
           T5G = T5u * T5F;
976
           T5Q = T5I * T5F;
977
           T5P = FNMS(KP980785280, T5O, T5L);
978
           Ip[WS(rs, 11)] = FNMS(T5I, T5P, T5G);
979
           Im[WS(rs, 11)] = FMA(T5u, T5P, T5Q);
980
            }
981
       }
982
        }
983
         }
984
    }
985
     }
986
}
987
988
static const tw_instr twinstr[] = {
989
     { TW_CEXP, 1, 1 },
990
     { TW_CEXP, 1, 3 },
991
     { TW_CEXP, 1, 9 },
992
     { TW_CEXP, 1, 27 },
993
     { TW_NEXT, 1, 0 }
994
};
995
996
static const hc2c_desc desc = { 32, "hc2cb2_32", twinstr, &GENUS, { 236, 98, 252, 0 } };
997
998
void X(codelet_hc2cb2_32) (planner *p) {
999
     X(khc2c_register) (p, hc2cb2_32, &desc, HC2C_VIA_RDFT);
1000
}
1001
#else
1002
1003
/* Generated by: ../../../genfft/gen_hc2c.native -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 32 -dif -name hc2cb2_32 -include rdft/scalar/hc2cb.h */
1004
1005
/*
1006
 * This function contains 488 FP additions, 280 FP multiplications,
1007
 * (or, 376 additions, 168 multiplications, 112 fused multiply/add),
1008
 * 160 stack variables, 7 constants, and 128 memory accesses
1009
 */
1010
#include "rdft/scalar/hc2cb.h"
1011
1012
static void hc2cb2_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
1013
0
{
1014
0
     DK(KP555570233, +0.555570233019602224742830813948532874374937191);
1015
0
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
1016
0
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
1017
0
     DK(KP195090322, +0.195090322016128267848284868477022240927691618);
1018
0
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
1019
0
     DK(KP382683432, +0.382683432365089771728459984030398866761344562);
1020
0
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
1021
0
     {
1022
0
    INT m;
1023
0
    for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 8, MAKE_VOLATILE_STRIDE(128, rs)) {
1024
0
         E T11, T14, T12, T15, T17, T2z, T2B, T1c, T18, T1d, T1g, T1k, T2F, T2L, T3t;
1025
0
         E T4H, T3h, T3V, T3b, T4v, T4T, T4X, T6t, T71, T6z, T75, T81, T8x, T8f, T8z;
1026
0
         E T2R, T2V, T8p, T8t, T4r, T4t, T53, T69, T3n, T3r, T7P, T7T, T4P, T4R, T6F;
1027
0
         E T6R, T1f, T2X, T1j, T2Y, T1l, T31, T2d, T2Z, T49, T4h, T4c, T4i, T4d, T4n;
1028
0
         E T4f, T4j;
1029
0
         {
1030
0
        E T2P, T3q, T2U, T3l, T2Q, T3p, T2T, T3m, T2D, T3g, T2K, T39, T2E, T3f, T2J;
1031
0
        E T3a;
1032
0
        {
1033
0
       E T13, T1b, T16, T1a;
1034
0
       T11 = W[0];
1035
0
       T14 = W[1];
1036
0
       T12 = W[2];
1037
0
       T15 = W[3];
1038
0
       T13 = T11 * T12;
1039
0
       T1b = T14 * T12;
1040
0
       T16 = T14 * T15;
1041
0
       T1a = T11 * T15;
1042
0
       T17 = T13 + T16;
1043
0
       T2z = T13 - T16;
1044
0
       T2B = T1a + T1b;
1045
0
       T1c = T1a - T1b;
1046
0
       T18 = W[4];
1047
0
       T2P = T12 * T18;
1048
0
       T3q = T14 * T18;
1049
0
       T2U = T15 * T18;
1050
0
       T3l = T11 * T18;
1051
0
       T1d = W[5];
1052
0
       T2Q = T15 * T1d;
1053
0
       T3p = T11 * T1d;
1054
0
       T2T = T12 * T1d;
1055
0
       T3m = T14 * T1d;
1056
0
       T1g = W[6];
1057
0
       T2D = T11 * T1g;
1058
0
       T3g = T15 * T1g;
1059
0
       T2K = T14 * T1g;
1060
0
       T39 = T12 * T1g;
1061
0
       T1k = W[7];
1062
0
       T2E = T14 * T1k;
1063
0
       T3f = T12 * T1k;
1064
0
       T2J = T11 * T1k;
1065
0
       T3a = T15 * T1k;
1066
0
        }
1067
0
        T2F = T2D - T2E;
1068
0
        T2L = T2J + T2K;
1069
0
        T3t = T39 - T3a;
1070
0
        T4H = T2J - T2K;
1071
0
        T3h = T3f - T3g;
1072
0
        T3V = T3f + T3g;
1073
0
        T3b = T39 + T3a;
1074
0
        T4v = T2D + T2E;
1075
0
        T4T = FMA(T18, T1g, T1d * T1k);
1076
0
        T4X = FNMS(T1d, T1g, T18 * T1k);
1077
0
        {
1078
0
       E T6r, T6s, T6x, T6y;
1079
0
       T6r = T17 * T1g;
1080
0
       T6s = T1c * T1k;
1081
0
       T6t = T6r - T6s;
1082
0
       T71 = T6r + T6s;
1083
0
       T6x = T17 * T1k;
1084
0
       T6y = T1c * T1g;
1085
0
       T6z = T6x + T6y;
1086
0
       T75 = T6x - T6y;
1087
0
        }
1088
0
        {
1089
0
       E T7Z, T80, T8d, T8e;
1090
0
       T7Z = T2z * T1g;
1091
0
       T80 = T2B * T1k;
1092
0
       T81 = T7Z + T80;
1093
0
       T8x = T7Z - T80;
1094
0
       T8d = T2z * T1k;
1095
0
       T8e = T2B * T1g;
1096
0
       T8f = T8d - T8e;
1097
0
       T8z = T8d + T8e;
1098
0
       T2R = T2P - T2Q;
1099
0
       T2V = T2T + T2U;
1100
0
       T8p = FMA(T2R, T1g, T2V * T1k);
1101
0
       T8t = FNMS(T2V, T1g, T2R * T1k);
1102
0
        }
1103
0
        T4r = T2P + T2Q;
1104
0
        T4t = T2T - T2U;
1105
0
        T53 = FMA(T4r, T1g, T4t * T1k);
1106
0
        T69 = FNMS(T4t, T1g, T4r * T1k);
1107
0
        T3n = T3l + T3m;
1108
0
        T3r = T3p - T3q;
1109
0
        T7P = FMA(T3n, T1g, T3r * T1k);
1110
0
        T7T = FNMS(T3r, T1g, T3n * T1k);
1111
0
        T4P = T3l - T3m;
1112
0
        T4R = T3p + T3q;
1113
0
        T6F = FMA(T4P, T1g, T4R * T1k);
1114
0
        T6R = FNMS(T4R, T1g, T4P * T1k);
1115
0
        {
1116
0
       E T19, T1e, T1h, T1i;
1117
0
       T19 = T17 * T18;
1118
0
       T1e = T1c * T1d;
1119
0
       T1f = T19 + T1e;
1120
0
       T2X = T19 - T1e;
1121
0
       T1h = T17 * T1d;
1122
0
       T1i = T1c * T18;
1123
0
       T1j = T1h - T1i;
1124
0
       T2Y = T1h + T1i;
1125
0
        }
1126
0
        T1l = FMA(T1f, T1g, T1j * T1k);
1127
0
        T31 = FNMS(T2Y, T1g, T2X * T1k);
1128
0
        T2d = FNMS(T1j, T1g, T1f * T1k);
1129
0
        T2Z = FMA(T2X, T1g, T2Y * T1k);
1130
0
        {
1131
0
       E T47, T48, T4a, T4b;
1132
0
       T47 = T2z * T18;
1133
0
       T48 = T2B * T1d;
1134
0
       T49 = T47 - T48;
1135
0
       T4h = T47 + T48;
1136
0
       T4a = T2z * T1d;
1137
0
       T4b = T2B * T18;
1138
0
       T4c = T4a + T4b;
1139
0
       T4i = T4a - T4b;
1140
0
        }
1141
0
        T4d = FMA(T49, T1g, T4c * T1k);
1142
0
        T4n = FNMS(T4i, T1g, T4h * T1k);
1143
0
        T4f = FNMS(T4c, T1g, T49 * T1k);
1144
0
        T4j = FMA(T4h, T1g, T4i * T1k);
1145
0
         }
1146
0
         {
1147
0
        E T56, T7b, T7C, T6c, Tf, T1m, T6f, T7c, T3Y, T4I, T2t, T32, T5d, T7D, T3w;
1148
0
        E T4w, Tu, T2e, T7g, T7F, T7j, T7G, T1B, T33, T3z, T40, T5l, T6i, T5s, T6h;
1149
0
        E T3C, T3Z, TK, T1D, T7v, T86, T7y, T85, T1S, T35, T3O, T4C, T5F, T6J, T5M;
1150
0
        E T6K, T3R, T4D, TZ, T1U, T7o, T89, T7r, T88, T29, T36, T3H, T4z, T5Y, T6M;
1151
0
        E T65, T6N, T3K, T4A;
1152
0
        {
1153
0
       E T3, T54, T2h, T6b, T6, T6a, T2k, T55, Ta, T57, T2o, T58, Td, T5a, T2r;
1154
0
       E T5b;
1155
0
       {
1156
0
            E T1, T2, T2f, T2g;
1157
0
            T1 = Rp[0];
1158
0
            T2 = Rm[WS(rs, 15)];
1159
0
            T3 = T1 + T2;
1160
0
            T54 = T1 - T2;
1161
0
            T2f = Ip[0];
1162
0
            T2g = Im[WS(rs, 15)];
1163
0
            T2h = T2f - T2g;
1164
0
            T6b = T2f + T2g;
1165
0
       }
1166
0
       {
1167
0
            E T4, T5, T2i, T2j;
1168
0
            T4 = Rp[WS(rs, 8)];
1169
0
            T5 = Rm[WS(rs, 7)];
1170
0
            T6 = T4 + T5;
1171
0
            T6a = T4 - T5;
1172
0
            T2i = Ip[WS(rs, 8)];
1173
0
            T2j = Im[WS(rs, 7)];
1174
0
            T2k = T2i - T2j;
1175
0
            T55 = T2i + T2j;
1176
0
       }
1177
0
       {
1178
0
            E T8, T9, T2m, T2n;
1179
0
            T8 = Rp[WS(rs, 4)];
1180
0
            T9 = Rm[WS(rs, 11)];
1181
0
            Ta = T8 + T9;
1182
0
            T57 = T8 - T9;
1183
0
            T2m = Ip[WS(rs, 4)];
1184
0
            T2n = Im[WS(rs, 11)];
1185
0
            T2o = T2m - T2n;
1186
0
            T58 = T2m + T2n;
1187
0
       }
1188
0
       {
1189
0
            E Tb, Tc, T2p, T2q;
1190
0
            Tb = Rm[WS(rs, 3)];
1191
0
            Tc = Rp[WS(rs, 12)];
1192
0
            Td = Tb + Tc;
1193
0
            T5a = Tb - Tc;
1194
0
            T2p = Ip[WS(rs, 12)];
1195
0
            T2q = Im[WS(rs, 3)];
1196
0
            T2r = T2p - T2q;
1197
0
            T5b = T2p + T2q;
1198
0
       }
1199
0
       {
1200
0
            E T7, Te, T2l, T2s;
1201
0
            T56 = T54 - T55;
1202
0
            T7b = T54 + T55;
1203
0
            T7C = T6b - T6a;
1204
0
            T6c = T6a + T6b;
1205
0
            T7 = T3 + T6;
1206
0
            Te = Ta + Td;
1207
0
            Tf = T7 + Te;
1208
0
            T1m = T7 - Te;
1209
0
            {
1210
0
           E T6d, T6e, T3W, T3X;
1211
0
           T6d = T57 + T58;
1212
0
           T6e = T5a + T5b;
1213
0
           T6f = KP707106781 * (T6d - T6e);
1214
0
           T7c = KP707106781 * (T6d + T6e);
1215
0
           T3W = T2h - T2k;
1216
0
           T3X = Ta - Td;
1217
0
           T3Y = T3W - T3X;
1218
0
           T4I = T3X + T3W;
1219
0
            }
1220
0
            T2l = T2h + T2k;
1221
0
            T2s = T2o + T2r;
1222
0
            T2t = T2l - T2s;
1223
0
            T32 = T2l + T2s;
1224
0
            {
1225
0
           E T59, T5c, T3u, T3v;
1226
0
           T59 = T57 - T58;
1227
0
           T5c = T5a - T5b;
1228
0
           T5d = KP707106781 * (T59 + T5c);
1229
0
           T7D = KP707106781 * (T59 - T5c);
1230
0
           T3u = T3 - T6;
1231
0
           T3v = T2r - T2o;
1232
0
           T3w = T3u - T3v;
1233
0
           T4w = T3u + T3v;
1234
0
            }
1235
0
       }
1236
0
        }
1237
0
        {
1238
0
       E Ti, T5p, T1w, T5n, Tl, T5m, T1z, T5q, Tp, T5i, T1p, T5g, Ts, T5f, T1s;
1239
0
       E T5j;
1240
0
       {
1241
0
            E Tg, Th, T1u, T1v;
1242
0
            Tg = Rp[WS(rs, 2)];
1243
0
            Th = Rm[WS(rs, 13)];
1244
0
            Ti = Tg + Th;
1245
0
            T5p = Tg - Th;
1246
0
            T1u = Ip[WS(rs, 2)];
1247
0
            T1v = Im[WS(rs, 13)];
1248
0
            T1w = T1u - T1v;
1249
0
            T5n = T1u + T1v;
1250
0
       }
1251
0
       {
1252
0
            E Tj, Tk, T1x, T1y;
1253
0
            Tj = Rp[WS(rs, 10)];
1254
0
            Tk = Rm[WS(rs, 5)];
1255
0
            Tl = Tj + Tk;
1256
0
            T5m = Tj - Tk;
1257
0
            T1x = Ip[WS(rs, 10)];
1258
0
            T1y = Im[WS(rs, 5)];
1259
0
            T1z = T1x - T1y;
1260
0
            T5q = T1x + T1y;
1261
0
       }
1262
0
       {
1263
0
            E Tn, To, T1n, T1o;
1264
0
            Tn = Rm[WS(rs, 1)];
1265
0
            To = Rp[WS(rs, 14)];
1266
0
            Tp = Tn + To;
1267
0
            T5i = Tn - To;
1268
0
            T1n = Ip[WS(rs, 14)];
1269
0
            T1o = Im[WS(rs, 1)];
1270
0
            T1p = T1n - T1o;
1271
0
            T5g = T1n + T1o;
1272
0
       }
1273
0
       {
1274
0
            E Tq, Tr, T1q, T1r;
1275
0
            Tq = Rp[WS(rs, 6)];
1276
0
            Tr = Rm[WS(rs, 9)];
1277
0
            Ts = Tq + Tr;
1278
0
            T5f = Tq - Tr;
1279
0
            T1q = Ip[WS(rs, 6)];
1280
0
            T1r = Im[WS(rs, 9)];
1281
0
            T1s = T1q - T1r;
1282
0
            T5j = T1q + T1r;
1283
0
       }
1284
0
       {
1285
0
            E Tm, Tt, T7e, T7f;
1286
0
            Tm = Ti + Tl;
1287
0
            Tt = Tp + Ts;
1288
0
            Tu = Tm + Tt;
1289
0
            T2e = Tm - Tt;
1290
0
            T7e = T5p + T5q;
1291
0
            T7f = T5n - T5m;
1292
0
            T7g = FNMS(KP923879532, T7f, KP382683432 * T7e);
1293
0
            T7F = FMA(KP382683432, T7f, KP923879532 * T7e);
1294
0
       }
1295
0
       {
1296
0
            E T7h, T7i, T1t, T1A;
1297
0
            T7h = T5i + T5j;
1298
0
            T7i = T5f + T5g;
1299
0
            T7j = FNMS(KP923879532, T7i, KP382683432 * T7h);
1300
0
            T7G = FMA(KP382683432, T7i, KP923879532 * T7h);
1301
0
            T1t = T1p + T1s;
1302
0
            T1A = T1w + T1z;
1303
0
            T1B = T1t - T1A;
1304
0
            T33 = T1A + T1t;
1305
0
       }
1306
0
       {
1307
0
            E T3x, T3y, T5h, T5k;
1308
0
            T3x = T1p - T1s;
1309
0
            T3y = Tp - Ts;
1310
0
            T3z = T3x - T3y;
1311
0
            T40 = T3y + T3x;
1312
0
            T5h = T5f - T5g;
1313
0
            T5k = T5i - T5j;
1314
0
            T5l = FNMS(KP382683432, T5k, KP923879532 * T5h);
1315
0
            T6i = FMA(KP382683432, T5h, KP923879532 * T5k);
1316
0
       }
1317
0
       {
1318
0
            E T5o, T5r, T3A, T3B;
1319
0
            T5o = T5m + T5n;
1320
0
            T5r = T5p - T5q;
1321
0
            T5s = FMA(KP923879532, T5o, KP382683432 * T5r);
1322
0
            T6h = FNMS(KP382683432, T5o, KP923879532 * T5r);
1323
0
            T3A = Ti - Tl;
1324
0
            T3B = T1w - T1z;
1325
0
            T3C = T3A + T3B;
1326
0
            T3Z = T3A - T3B;
1327
0
       }
1328
0
        }
1329
0
        {
1330
0
       E Ty, T5v, T1G, T5H, TB, T5G, T1J, T5w, TI, T5K, T1Q, T5D, TF, T5J, T1N;
1331
0
       E T5A;
1332
0
       {
1333
0
            E Tw, Tx, T1H, T1I;
1334
0
            Tw = Rp[WS(rs, 1)];
1335
0
            Tx = Rm[WS(rs, 14)];
1336
0
            Ty = Tw + Tx;
1337
0
            T5v = Tw - Tx;
1338
0
            {
1339
0
           E T1E, T1F, Tz, TA;
1340
0
           T1E = Ip[WS(rs, 1)];
1341
0
           T1F = Im[WS(rs, 14)];
1342
0
           T1G = T1E - T1F;
1343
0
           T5H = T1E + T1F;
1344
0
           Tz = Rp[WS(rs, 9)];
1345
0
           TA = Rm[WS(rs, 6)];
1346
0
           TB = Tz + TA;
1347
0
           T5G = Tz - TA;
1348
0
            }
1349
0
            T1H = Ip[WS(rs, 9)];
1350
0
            T1I = Im[WS(rs, 6)];
1351
0
            T1J = T1H - T1I;
1352
0
            T5w = T1H + T1I;
1353
0
            {
1354
0
           E TG, TH, T5B, T1O, T1P, T5C;
1355
0
           TG = Rm[WS(rs, 2)];
1356
0
           TH = Rp[WS(rs, 13)];
1357
0
           T5B = TG - TH;
1358
0
           T1O = Ip[WS(rs, 13)];
1359
0
           T1P = Im[WS(rs, 2)];
1360
0
           T5C = T1O + T1P;
1361
0
           TI = TG + TH;
1362
0
           T5K = T5B + T5C;
1363
0
           T1Q = T1O - T1P;
1364
0
           T5D = T5B - T5C;
1365
0
            }
1366
0
            {
1367
0
           E TD, TE, T5y, T1L, T1M, T5z;
1368
0
           TD = Rp[WS(rs, 5)];
1369
0
           TE = Rm[WS(rs, 10)];
1370
0
           T5y = TD - TE;
1371
0
           T1L = Ip[WS(rs, 5)];
1372
0
           T1M = Im[WS(rs, 10)];
1373
0
           T5z = T1L + T1M;
1374
0
           TF = TD + TE;
1375
0
           T5J = T5y + T5z;
1376
0
           T1N = T1L - T1M;
1377
0
           T5A = T5y - T5z;
1378
0
            }
1379
0
       }
1380
0
       {
1381
0
            E TC, TJ, T7t, T7u;
1382
0
            TC = Ty + TB;
1383
0
            TJ = TF + TI;
1384
0
            TK = TC + TJ;
1385
0
            T1D = TC - TJ;
1386
0
            T7t = T5H - T5G;
1387
0
            T7u = KP707106781 * (T5A - T5D);
1388
0
            T7v = T7t + T7u;
1389
0
            T86 = T7t - T7u;
1390
0
       }
1391
0
       {
1392
0
            E T7w, T7x, T1K, T1R;
1393
0
            T7w = T5v + T5w;
1394
0
            T7x = KP707106781 * (T5J + T5K);
1395
0
            T7y = T7w - T7x;
1396
0
            T85 = T7w + T7x;
1397
0
            T1K = T1G + T1J;
1398
0
            T1R = T1N + T1Q;
1399
0
            T1S = T1K - T1R;
1400
0
            T35 = T1K + T1R;
1401
0
       }
1402
0
       {
1403
0
            E T3M, T3N, T5x, T5E;
1404
0
            T3M = T1G - T1J;
1405
0
            T3N = TF - TI;
1406
0
            T3O = T3M - T3N;
1407
0
            T4C = T3N + T3M;
1408
0
            T5x = T5v - T5w;
1409
0
            T5E = KP707106781 * (T5A + T5D);
1410
0
            T5F = T5x - T5E;
1411
0
            T6J = T5x + T5E;
1412
0
       }
1413
0
       {
1414
0
            E T5I, T5L, T3P, T3Q;
1415
0
            T5I = T5G + T5H;
1416
0
            T5L = KP707106781 * (T5J - T5K);
1417
0
            T5M = T5I - T5L;
1418
0
            T6K = T5I + T5L;
1419
0
            T3P = Ty - TB;
1420
0
            T3Q = T1Q - T1N;
1421
0
            T3R = T3P - T3Q;
1422
0
            T4D = T3P + T3Q;
1423
0
       }
1424
0
        }
1425
0
        {
1426
0
       E TN, T5O, T1X, T60, TQ, T5Z, T20, T5P, TX, T63, T27, T5W, TU, T62, T24;
1427
0
       E T5T;
1428
0
       {
1429
0
            E TL, TM, T1Y, T1Z;
1430
0
            TL = Rm[0];
1431
0
            TM = Rp[WS(rs, 15)];
1432
0
            TN = TL + TM;
1433
0
            T5O = TL - TM;
1434
0
            {
1435
0
           E T1V, T1W, TO, TP;
1436
0
           T1V = Ip[WS(rs, 15)];
1437
0
           T1W = Im[0];
1438
0
           T1X = T1V - T1W;
1439
0
           T60 = T1V + T1W;
1440
0
           TO = Rp[WS(rs, 7)];
1441
0
           TP = Rm[WS(rs, 8)];
1442
0
           TQ = TO + TP;
1443
0
           T5Z = TO - TP;
1444
0
            }
1445
0
            T1Y = Ip[WS(rs, 7)];
1446
0
            T1Z = Im[WS(rs, 8)];
1447
0
            T20 = T1Y - T1Z;
1448
0
            T5P = T1Y + T1Z;
1449
0
            {
1450
0
           E TV, TW, T5U, T25, T26, T5V;
1451
0
           TV = Rm[WS(rs, 4)];
1452
0
           TW = Rp[WS(rs, 11)];
1453
0
           T5U = TV - TW;
1454
0
           T25 = Ip[WS(rs, 11)];
1455
0
           T26 = Im[WS(rs, 4)];
1456
0
           T5V = T25 + T26;
1457
0
           TX = TV + TW;
1458
0
           T63 = T5U + T5V;
1459
0
           T27 = T25 - T26;
1460
0
           T5W = T5U - T5V;
1461
0
            }
1462
0
            {
1463
0
           E TS, TT, T5R, T22, T23, T5S;
1464
0
           TS = Rp[WS(rs, 3)];
1465
0
           TT = Rm[WS(rs, 12)];
1466
0
           T5R = TS - TT;
1467
0
           T22 = Ip[WS(rs, 3)];
1468
0
           T23 = Im[WS(rs, 12)];
1469
0
           T5S = T22 + T23;
1470
0
           TU = TS + TT;
1471
0
           T62 = T5R + T5S;
1472
0
           T24 = T22 - T23;
1473
0
           T5T = T5R - T5S;
1474
0
            }
1475
0
       }
1476
0
       {
1477
0
            E TR, TY, T7m, T7n;
1478
0
            TR = TN + TQ;
1479
0
            TY = TU + TX;
1480
0
            TZ = TR + TY;
1481
0
            T1U = TR - TY;
1482
0
            T7m = KP707106781 * (T5T - T5W);
1483
0
            T7n = T5Z + T60;
1484
0
            T7o = T7m - T7n;
1485
0
            T89 = T7n + T7m;
1486
0
       }
1487
0
       {
1488
0
            E T7p, T7q, T21, T28;
1489
0
            T7p = T5O + T5P;
1490
0
            T7q = KP707106781 * (T62 + T63);
1491
0
            T7r = T7p - T7q;
1492
0
            T88 = T7p + T7q;
1493
0
            T21 = T1X + T20;
1494
0
            T28 = T24 + T27;
1495
0
            T29 = T21 - T28;
1496
0
            T36 = T21 + T28;
1497
0
       }
1498
0
       {
1499
0
            E T3F, T3G, T5Q, T5X;
1500
0
            T3F = T1X - T20;
1501
0
            T3G = TU - TX;
1502
0
            T3H = T3F - T3G;
1503
0
            T4z = T3G + T3F;
1504
0
            T5Q = T5O - T5P;
1505
0
            T5X = KP707106781 * (T5T + T5W);
1506
0
            T5Y = T5Q - T5X;
1507
0
            T6M = T5Q + T5X;
1508
0
       }
1509
0
       {
1510
0
            E T61, T64, T3I, T3J;
1511
0
            T61 = T5Z - T60;
1512
0
            T64 = KP707106781 * (T62 - T63);
1513
0
            T65 = T61 - T64;
1514
0
            T6N = T61 + T64;
1515
0
            T3I = TN - TQ;
1516
0
            T3J = T27 - T24;
1517
0
            T3K = T3I - T3J;
1518
0
            T4A = T3I + T3J;
1519
0
       }
1520
0
        }
1521
0
        {
1522
0
       E Tv, T10, T30, T34, T37, T38;
1523
0
       Tv = Tf + Tu;
1524
0
       T10 = TK + TZ;
1525
0
       T30 = Tv - T10;
1526
0
       T34 = T32 + T33;
1527
0
       T37 = T35 + T36;
1528
0
       T38 = T34 - T37;
1529
0
       Rp[0] = Tv + T10;
1530
0
       Rm[0] = T34 + T37;
1531
0
       Rp[WS(rs, 8)] = FNMS(T31, T38, T2Z * T30);
1532
0
       Rm[WS(rs, 8)] = FMA(T31, T30, T2Z * T38);
1533
0
        }
1534
0
        {
1535
0
       E T3e, T3o, T3k, T3s;
1536
0
       {
1537
0
            E T3c, T3d, T3i, T3j;
1538
0
            T3c = Tf - Tu;
1539
0
            T3d = T36 - T35;
1540
0
            T3e = T3c - T3d;
1541
0
            T3o = T3c + T3d;
1542
0
            T3i = T32 - T33;
1543
0
            T3j = TK - TZ;
1544
0
            T3k = T3i - T3j;
1545
0
            T3s = T3j + T3i;
1546
0
       }
1547
0
       Rp[WS(rs, 12)] = FNMS(T3h, T3k, T3b * T3e);
1548
0
       Rm[WS(rs, 12)] = FMA(T3b, T3k, T3h * T3e);
1549
0
       Rp[WS(rs, 4)] = FNMS(T3r, T3s, T3n * T3o);
1550
0
       Rm[WS(rs, 4)] = FMA(T3n, T3s, T3r * T3o);
1551
0
        }
1552
0
        {
1553
0
       E T1C, T2u, T2M, T2G, T2x, T2H, T2b, T2N;
1554
0
       T1C = T1m + T1B;
1555
0
       T2u = T2e + T2t;
1556
0
       T2M = T2t - T2e;
1557
0
       T2G = T1m - T1B;
1558
0
       {
1559
0
            E T2v, T2w, T1T, T2a;
1560
0
            T2v = T1D + T1S;
1561
0
            T2w = T29 - T1U;
1562
0
            T2x = KP707106781 * (T2v + T2w);
1563
0
            T2H = KP707106781 * (T2w - T2v);
1564
0
            T1T = T1D - T1S;
1565
0
            T2a = T1U + T29;
1566
0
            T2b = KP707106781 * (T1T + T2a);
1567
0
            T2N = KP707106781 * (T1T - T2a);
1568
0
       }
1569
0
       {
1570
0
            E T2c, T2y, T2S, T2W;
1571
0
            T2c = T1C - T2b;
1572
0
            T2y = T2u - T2x;
1573
0
            Rp[WS(rs, 10)] = FNMS(T2d, T2y, T1l * T2c);
1574
0
            Rm[WS(rs, 10)] = FMA(T2d, T2c, T1l * T2y);
1575
0
            T2S = T2G + T2H;
1576
0
            T2W = T2M + T2N;
1577
0
            Rp[WS(rs, 6)] = FNMS(T2V, T2W, T2R * T2S);
1578
0
            Rm[WS(rs, 6)] = FMA(T2R, T2W, T2V * T2S);
1579
0
       }
1580
0
       {
1581
0
            E T2A, T2C, T2I, T2O;
1582
0
            T2A = T1C + T2b;
1583
0
            T2C = T2u + T2x;
1584
0
            Rp[WS(rs, 2)] = FNMS(T2B, T2C, T2z * T2A);
1585
0
            Rm[WS(rs, 2)] = FMA(T2B, T2A, T2z * T2C);
1586
0
            T2I = T2G - T2H;
1587
0
            T2O = T2M - T2N;
1588
0
            Rp[WS(rs, 14)] = FNMS(T2L, T2O, T2F * T2I);
1589
0
            Rm[WS(rs, 14)] = FMA(T2F, T2O, T2L * T2I);
1590
0
       }
1591
0
        }
1592
0
        {
1593
0
       E T4y, T4U, T4K, T4Y, T4F, T4Z, T4N, T4V, T4x, T4J;
1594
0
       T4x = KP707106781 * (T3Z + T40);
1595
0
       T4y = T4w - T4x;
1596
0
       T4U = T4w + T4x;
1597
0
       T4J = KP707106781 * (T3C + T3z);
1598
0
       T4K = T4I - T4J;
1599
0
       T4Y = T4I + T4J;
1600
0
       {
1601
0
            E T4B, T4E, T4L, T4M;
1602
0
            T4B = FNMS(KP382683432, T4A, KP923879532 * T4z);
1603
0
            T4E = FMA(KP923879532, T4C, KP382683432 * T4D);
1604
0
            T4F = T4B - T4E;
1605
0
            T4Z = T4E + T4B;
1606
0
            T4L = FNMS(KP382683432, T4C, KP923879532 * T4D);
1607
0
            T4M = FMA(KP382683432, T4z, KP923879532 * T4A);
1608
0
            T4N = T4L - T4M;
1609
0
            T4V = T4L + T4M;
1610
0
       }
1611
0
       {
1612
0
            E T4G, T4O, T51, T52;
1613
0
            T4G = T4y - T4F;
1614
0
            T4O = T4K - T4N;
1615
0
            Rp[WS(rs, 13)] = FNMS(T4H, T4O, T4v * T4G);
1616
0
            Rm[WS(rs, 13)] = FMA(T4H, T4G, T4v * T4O);
1617
0
            T51 = T4U + T4V;
1618
0
            T52 = T4Y + T4Z;
1619
0
            Rp[WS(rs, 1)] = FNMS(T1c, T52, T17 * T51);
1620
0
            Rm[WS(rs, 1)] = FMA(T17, T52, T1c * T51);
1621
0
       }
1622
0
       {
1623
0
            E T4Q, T4S, T4W, T50;
1624
0
            T4Q = T4y + T4F;
1625
0
            T4S = T4K + T4N;
1626
0
            Rp[WS(rs, 5)] = FNMS(T4R, T4S, T4P * T4Q);
1627
0
            Rm[WS(rs, 5)] = FMA(T4R, T4Q, T4P * T4S);
1628
0
            T4W = T4U - T4V;
1629
0
            T50 = T4Y - T4Z;
1630
0
            Rp[WS(rs, 9)] = FNMS(T4X, T50, T4T * T4W);
1631
0
            Rm[WS(rs, 9)] = FMA(T4T, T50, T4X * T4W);
1632
0
       }
1633
0
        }
1634
0
        {
1635
0
       E T3E, T4k, T42, T4o, T3T, T4p, T45, T4l, T3D, T41;
1636
0
       T3D = KP707106781 * (T3z - T3C);
1637
0
       T3E = T3w - T3D;
1638
0
       T4k = T3w + T3D;
1639
0
       T41 = KP707106781 * (T3Z - T40);
1640
0
       T42 = T3Y - T41;
1641
0
       T4o = T3Y + T41;
1642
0
       {
1643
0
            E T3L, T3S, T43, T44;
1644
0
            T3L = FNMS(KP923879532, T3K, KP382683432 * T3H);
1645
0
            T3S = FMA(KP382683432, T3O, KP923879532 * T3R);
1646
0
            T3T = T3L - T3S;
1647
0
            T4p = T3S + T3L;
1648
0
            T43 = FNMS(KP923879532, T3O, KP382683432 * T3R);
1649
0
            T44 = FMA(KP923879532, T3H, KP382683432 * T3K);
1650
0
            T45 = T43 - T44;
1651
0
            T4l = T43 + T44;
1652
0
       }
1653
0
       {
1654
0
            E T3U, T46, T4s, T4u;
1655
0
            T3U = T3E - T3T;
1656
0
            T46 = T42 - T45;
1657
0
            Rp[WS(rs, 15)] = FNMS(T3V, T46, T3t * T3U);
1658
0
            Rm[WS(rs, 15)] = FMA(T3V, T3U, T3t * T46);
1659
0
            T4s = T4k + T4l;
1660
0
            T4u = T4o + T4p;
1661
0
            Rp[WS(rs, 3)] = FNMS(T4t, T4u, T4r * T4s);
1662
0
            Rm[WS(rs, 3)] = FMA(T4r, T4u, T4t * T4s);
1663
0
       }
1664
0
       {
1665
0
            E T4e, T4g, T4m, T4q;
1666
0
            T4e = T3E + T3T;
1667
0
            T4g = T42 + T45;
1668
0
            Rp[WS(rs, 7)] = FNMS(T4f, T4g, T4d * T4e);
1669
0
            Rm[WS(rs, 7)] = FMA(T4f, T4e, T4d * T4g);
1670
0
            T4m = T4k - T4l;
1671
0
            T4q = T4o - T4p;
1672
0
            Rp[WS(rs, 11)] = FNMS(T4n, T4q, T4j * T4m);
1673
0
            Rm[WS(rs, 11)] = FMA(T4j, T4q, T4n * T4m);
1674
0
       }
1675
0
        }
1676
0
        {
1677
0
       E T6I, T72, T6X, T73, T6P, T77, T6U, T76;
1678
0
       {
1679
0
            E T6G, T6H, T6V, T6W;
1680
0
            T6G = T56 + T5d;
1681
0
            T6H = T6h + T6i;
1682
0
            T6I = T6G + T6H;
1683
0
            T72 = T6G - T6H;
1684
0
            T6V = FMA(KP195090322, T6J, KP980785280 * T6K);
1685
0
            T6W = FNMS(KP195090322, T6M, KP980785280 * T6N);
1686
0
            T6X = T6V + T6W;
1687
0
            T73 = T6W - T6V;
1688
0
       }
1689
0
       {
1690
0
            E T6L, T6O, T6S, T6T;
1691
0
            T6L = FNMS(KP195090322, T6K, KP980785280 * T6J);
1692
0
            T6O = FMA(KP980785280, T6M, KP195090322 * T6N);
1693
0
            T6P = T6L + T6O;
1694
0
            T77 = T6L - T6O;
1695
0
            T6S = T6c + T6f;
1696
0
            T6T = T5s + T5l;
1697
0
            T6U = T6S + T6T;
1698
0
            T76 = T6S - T6T;
1699
0
       }
1700
0
       {
1701
0
            E T6Q, T6Y, T79, T7a;
1702
0
            T6Q = T6I - T6P;
1703
0
            T6Y = T6U - T6X;
1704
0
            Ip[WS(rs, 8)] = FNMS(T6R, T6Y, T6F * T6Q);
1705
0
            Im[WS(rs, 8)] = FMA(T6R, T6Q, T6F * T6Y);
1706
0
            T79 = T72 + T73;
1707
0
            T7a = T76 + T77;
1708
0
            Ip[WS(rs, 4)] = FNMS(T1d, T7a, T18 * T79);
1709
0
            Im[WS(rs, 4)] = FMA(T18, T7a, T1d * T79);
1710
0
       }
1711
0
       {
1712
0
            E T6Z, T70, T74, T78;
1713
0
            T6Z = T6I + T6P;
1714
0
            T70 = T6U + T6X;
1715
0
            Ip[0] = FNMS(T14, T70, T11 * T6Z);
1716
0
            Im[0] = FMA(T14, T6Z, T11 * T70);
1717
0
            T74 = T72 - T73;
1718
0
            T78 = T76 - T77;
1719
0
            Ip[WS(rs, 12)] = FNMS(T75, T78, T71 * T74);
1720
0
            Im[WS(rs, 12)] = FMA(T71, T78, T75 * T74);
1721
0
       }
1722
0
        }
1723
0
        {
1724
0
       E T84, T8q, T8l, T8r, T8b, T8v, T8i, T8u;
1725
0
       {
1726
0
            E T82, T83, T8j, T8k;
1727
0
            T82 = T7b + T7c;
1728
0
            T83 = T7F + T7G;
1729
0
            T84 = T82 - T83;
1730
0
            T8q = T82 + T83;
1731
0
            T8j = FMA(KP195090322, T86, KP980785280 * T85);
1732
0
            T8k = FMA(KP195090322, T89, KP980785280 * T88);
1733
0
            T8l = T8j - T8k;
1734
0
            T8r = T8j + T8k;
1735
0
       }
1736
0
       {
1737
0
            E T87, T8a, T8g, T8h;
1738
0
            T87 = FNMS(KP980785280, T86, KP195090322 * T85);
1739
0
            T8a = FNMS(KP980785280, T89, KP195090322 * T88);
1740
0
            T8b = T87 + T8a;
1741
0
            T8v = T87 - T8a;
1742
0
            T8g = T7C - T7D;
1743
0
            T8h = T7g - T7j;
1744
0
            T8i = T8g + T8h;
1745
0
            T8u = T8g - T8h;
1746
0
       }
1747
0
       {
1748
0
            E T8c, T8m, T8y, T8A;
1749
0
            T8c = T84 - T8b;
1750
0
            T8m = T8i - T8l;
1751
0
            Ip[WS(rs, 11)] = FNMS(T8f, T8m, T81 * T8c);
1752
0
            Im[WS(rs, 11)] = FMA(T8f, T8c, T81 * T8m);
1753
0
            T8y = T8q + T8r;
1754
0
            T8A = T8u - T8v;
1755
0
            Ip[WS(rs, 15)] = FNMS(T8z, T8A, T8x * T8y);
1756
0
            Im[WS(rs, 15)] = FMA(T8x, T8A, T8z * T8y);
1757
0
       }
1758
0
       {
1759
0
            E T8n, T8o, T8s, T8w;
1760
0
            T8n = T84 + T8b;
1761
0
            T8o = T8i + T8l;
1762
0
            Ip[WS(rs, 3)] = FNMS(T1j, T8o, T1f * T8n);
1763
0
            Im[WS(rs, 3)] = FMA(T1j, T8n, T1f * T8o);
1764
0
            T8s = T8q - T8r;
1765
0
            T8w = T8u + T8v;
1766
0
            Ip[WS(rs, 7)] = FNMS(T8t, T8w, T8p * T8s);
1767
0
            Im[WS(rs, 7)] = FMA(T8p, T8w, T8t * T8s);
1768
0
       }
1769
0
        }
1770
0
        {
1771
0
       E T5u, T6u, T6n, T6v, T67, T6B, T6k, T6A;
1772
0
       {
1773
0
            E T5e, T5t, T6l, T6m;
1774
0
            T5e = T56 - T5d;
1775
0
            T5t = T5l - T5s;
1776
0
            T5u = T5e + T5t;
1777
0
            T6u = T5e - T5t;
1778
0
            T6l = FMA(KP831469612, T5F, KP555570233 * T5M);
1779
0
            T6m = FNMS(KP831469612, T5Y, KP555570233 * T65);
1780
0
            T6n = T6l + T6m;
1781
0
            T6v = T6m - T6l;
1782
0
       }
1783
0
       {
1784
0
            E T5N, T66, T6g, T6j;
1785
0
            T5N = FNMS(KP831469612, T5M, KP555570233 * T5F);
1786
0
            T66 = FMA(KP555570233, T5Y, KP831469612 * T65);
1787
0
            T67 = T5N + T66;
1788
0
            T6B = T5N - T66;
1789
0
            T6g = T6c - T6f;
1790
0
            T6j = T6h - T6i;
1791
0
            T6k = T6g + T6j;
1792
0
            T6A = T6g - T6j;
1793
0
       }
1794
0
       {
1795
0
            E T68, T6o, T6D, T6E;
1796
0
            T68 = T5u - T67;
1797
0
            T6o = T6k - T6n;
1798
0
            Ip[WS(rs, 10)] = FNMS(T69, T6o, T53 * T68);
1799
0
            Im[WS(rs, 10)] = FMA(T69, T68, T53 * T6o);
1800
0
            T6D = T6u + T6v;
1801
0
            T6E = T6A + T6B;
1802
0
            Ip[WS(rs, 6)] = FNMS(T4c, T6E, T49 * T6D);
1803
0
            Im[WS(rs, 6)] = FMA(T49, T6E, T4c * T6D);
1804
0
       }
1805
0
       {
1806
0
            E T6p, T6q, T6w, T6C;
1807
0
            T6p = T5u + T67;
1808
0
            T6q = T6k + T6n;
1809
0
            Ip[WS(rs, 2)] = FNMS(T4i, T6q, T4h * T6p);
1810
0
            Im[WS(rs, 2)] = FMA(T4i, T6p, T4h * T6q);
1811
0
            T6w = T6u - T6v;
1812
0
            T6C = T6A - T6B;
1813
0
            Ip[WS(rs, 14)] = FNMS(T6z, T6C, T6t * T6w);
1814
0
            Im[WS(rs, 14)] = FMA(T6t, T6C, T6z * T6w);
1815
0
       }
1816
0
        }
1817
0
        {
1818
0
       E T7l, T7Q, T7L, T7R, T7A, T7V, T7I, T7U;
1819
0
       {
1820
0
            E T7d, T7k, T7J, T7K;
1821
0
            T7d = T7b - T7c;
1822
0
            T7k = T7g + T7j;
1823
0
            T7l = T7d - T7k;
1824
0
            T7Q = T7d + T7k;
1825
0
            T7J = FNMS(KP555570233, T7v, KP831469612 * T7y);
1826
0
            T7K = FMA(KP555570233, T7o, KP831469612 * T7r);
1827
0
            T7L = T7J - T7K;
1828
0
            T7R = T7J + T7K;
1829
0
       }
1830
0
       {
1831
0
            E T7s, T7z, T7E, T7H;
1832
0
            T7s = FNMS(KP555570233, T7r, KP831469612 * T7o);
1833
0
            T7z = FMA(KP831469612, T7v, KP555570233 * T7y);
1834
0
            T7A = T7s - T7z;
1835
0
            T7V = T7z + T7s;
1836
0
            T7E = T7C + T7D;
1837
0
            T7H = T7F - T7G;
1838
0
            T7I = T7E - T7H;
1839
0
            T7U = T7E + T7H;
1840
0
       }
1841
0
       {
1842
0
            E T7B, T7M, T7X, T7Y;
1843
0
            T7B = T7l - T7A;
1844
0
            T7M = T7I - T7L;
1845
0
            Ip[WS(rs, 13)] = FNMS(T1k, T7M, T1g * T7B);
1846
0
            Im[WS(rs, 13)] = FMA(T1k, T7B, T1g * T7M);
1847
0
            T7X = T7Q + T7R;
1848
0
            T7Y = T7U + T7V;
1849
0
            Ip[WS(rs, 1)] = FNMS(T15, T7Y, T12 * T7X);
1850
0
            Im[WS(rs, 1)] = FMA(T12, T7Y, T15 * T7X);
1851
0
       }
1852
0
       {
1853
0
            E T7N, T7O, T7S, T7W;
1854
0
            T7N = T7l + T7A;
1855
0
            T7O = T7I + T7L;
1856
0
            Ip[WS(rs, 5)] = FNMS(T2Y, T7O, T2X * T7N);
1857
0
            Im[WS(rs, 5)] = FMA(T2Y, T7N, T2X * T7O);
1858
0
            T7S = T7Q - T7R;
1859
0
            T7W = T7U - T7V;
1860
0
            Ip[WS(rs, 9)] = FNMS(T7T, T7W, T7P * T7S);
1861
0
            Im[WS(rs, 9)] = FMA(T7P, T7W, T7T * T7S);
1862
0
       }
1863
0
        }
1864
0
         }
1865
0
    }
1866
0
     }
1867
0
}
1868
1869
static const tw_instr twinstr[] = {
1870
     { TW_CEXP, 1, 1 },
1871
     { TW_CEXP, 1, 3 },
1872
     { TW_CEXP, 1, 9 },
1873
     { TW_CEXP, 1, 27 },
1874
     { TW_NEXT, 1, 0 }
1875
};
1876
1877
static const hc2c_desc desc = { 32, "hc2cb2_32", twinstr, &GENUS, { 376, 168, 112, 0 } };
1878
1879
1
void X(codelet_hc2cb2_32) (planner *p) {
1880
1
     X(khc2c_register) (p, hc2cb2_32, &desc, HC2C_VIA_RDFT);
1881
1
}
1882
#endif