Coverage Report

Created: 2023-09-25 07:08

/src/fftw3/rdft/scalar/r2cf/hf_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 Mon Sep 25 07:06:13 UTC 2023 */
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 -n 32 -dit -name hf_32 -include rdft/scalar/hf.h */
29
30
/*
31
 * This function contains 434 FP additions, 260 FP multiplications,
32
 * (or, 236 additions, 62 multiplications, 198 fused multiply/add),
33
 * 102 stack variables, 7 constants, and 128 memory accesses
34
 */
35
#include "rdft/scalar/hf.h"
36
37
static void hf_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) * 62); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 62, MAKE_VOLATILE_STRIDE(64, rs)) {
49
         E T8, T8y, T3w, T87, Tl, T8x, T3B, T83, Tz, T6G, T3J, T5T, TM, T6F, T3Q;
50
         E T5U, T11, T1e, T6J, T6K, T6L, T6M, T3Z, T5Y, T46, T5X, T1s, T1F, T6O, T6P;
51
         E T6Q, T6R, T4e, T61, T4l, T60, T32, T76, T7d, T7N, T54, T6c, T5r, T6f, T29;
52
         E T6V, T72, T7I, T4v, T65, T4S, T68, T3t, T7e, T79, T7O, T5b, T5t, T5i, T5s;
53
         E T2A, T73, T6Y, T7J, T4C, T4T, T4J, T4U;
54
         {
55
        E T1, T86, T3, T6, T4, T84, T2, T7, T85, T5;
56
        T1 = cr[0];
57
        T86 = ci[0];
58
        T3 = cr[WS(rs, 16)];
59
        T6 = ci[WS(rs, 16)];
60
        T2 = W[30];
61
        T4 = T2 * T3;
62
        T84 = T2 * T6;
63
        T5 = W[31];
64
        T7 = FMA(T5, T6, T4);
65
        T85 = FNMS(T5, T3, T84);
66
        T8 = T1 + T7;
67
        T8y = T86 - T85;
68
        T3w = T1 - T7;
69
        T87 = T85 + T86;
70
         }
71
         {
72
        E Ta, Td, Tb, T3x, Tg, Tj, Th, T3z, T9, Tf;
73
        Ta = cr[WS(rs, 8)];
74
        Td = ci[WS(rs, 8)];
75
        T9 = W[14];
76
        Tb = T9 * Ta;
77
        T3x = T9 * Td;
78
        Tg = cr[WS(rs, 24)];
79
        Tj = ci[WS(rs, 24)];
80
        Tf = W[46];
81
        Th = Tf * Tg;
82
        T3z = Tf * Tj;
83
        {
84
       E Te, T3y, Tk, T3A, Tc, Ti;
85
       Tc = W[15];
86
       Te = FMA(Tc, Td, Tb);
87
       T3y = FNMS(Tc, Ta, T3x);
88
       Ti = W[47];
89
       Tk = FMA(Ti, Tj, Th);
90
       T3A = FNMS(Ti, Tg, T3z);
91
       Tl = Te + Tk;
92
       T8x = Te - Tk;
93
       T3B = T3y - T3A;
94
       T83 = T3y + T3A;
95
        }
96
         }
97
         {
98
        E Ts, T3F, Ty, T3H, T3D, T3I;
99
        {
100
       E To, Tr, Tp, T3E, Tn, Tq;
101
       To = cr[WS(rs, 4)];
102
       Tr = ci[WS(rs, 4)];
103
       Tn = W[6];
104
       Tp = Tn * To;
105
       T3E = Tn * Tr;
106
       Tq = W[7];
107
       Ts = FMA(Tq, Tr, Tp);
108
       T3F = FNMS(Tq, To, T3E);
109
        }
110
        {
111
       E Tu, Tx, Tv, T3G, Tt, Tw;
112
       Tu = cr[WS(rs, 20)];
113
       Tx = ci[WS(rs, 20)];
114
       Tt = W[38];
115
       Tv = Tt * Tu;
116
       T3G = Tt * Tx;
117
       Tw = W[39];
118
       Ty = FMA(Tw, Tx, Tv);
119
       T3H = FNMS(Tw, Tu, T3G);
120
        }
121
        Tz = Ts + Ty;
122
        T6G = T3F + T3H;
123
        T3D = Ts - Ty;
124
        T3I = T3F - T3H;
125
        T3J = T3D - T3I;
126
        T5T = T3D + T3I;
127
         }
128
         {
129
        E TF, T3M, TL, T3O, T3K, T3P;
130
        {
131
       E TB, TE, TC, T3L, TA, TD;
132
       TB = cr[WS(rs, 28)];
133
       TE = ci[WS(rs, 28)];
134
       TA = W[54];
135
       TC = TA * TB;
136
       T3L = TA * TE;
137
       TD = W[55];
138
       TF = FMA(TD, TE, TC);
139
       T3M = FNMS(TD, TB, T3L);
140
        }
141
        {
142
       E TH, TK, TI, T3N, TG, TJ;
143
       TH = cr[WS(rs, 12)];
144
       TK = ci[WS(rs, 12)];
145
       TG = W[22];
146
       TI = TG * TH;
147
       T3N = TG * TK;
148
       TJ = W[23];
149
       TL = FMA(TJ, TK, TI);
150
       T3O = FNMS(TJ, TH, T3N);
151
        }
152
        TM = TF + TL;
153
        T6F = T3M + T3O;
154
        T3K = TF - TL;
155
        T3P = T3M - T3O;
156
        T3Q = T3K + T3P;
157
        T5U = T3K - T3P;
158
         }
159
         {
160
        E TU, T3U, T1d, T44, T10, T3W, T17, T42;
161
        {
162
       E TQ, TT, TR, T3T, TP, TS;
163
       TQ = cr[WS(rs, 2)];
164
       TT = ci[WS(rs, 2)];
165
       TP = W[2];
166
       TR = TP * TQ;
167
       T3T = TP * TT;
168
       TS = W[3];
169
       TU = FMA(TS, TT, TR);
170
       T3U = FNMS(TS, TQ, T3T);
171
        }
172
        {
173
       E T19, T1c, T1a, T43, T18, T1b;
174
       T19 = cr[WS(rs, 26)];
175
       T1c = ci[WS(rs, 26)];
176
       T18 = W[50];
177
       T1a = T18 * T19;
178
       T43 = T18 * T1c;
179
       T1b = W[51];
180
       T1d = FMA(T1b, T1c, T1a);
181
       T44 = FNMS(T1b, T19, T43);
182
        }
183
        {
184
       E TW, TZ, TX, T3V, TV, TY;
185
       TW = cr[WS(rs, 18)];
186
       TZ = ci[WS(rs, 18)];
187
       TV = W[34];
188
       TX = TV * TW;
189
       T3V = TV * TZ;
190
       TY = W[35];
191
       T10 = FMA(TY, TZ, TX);
192
       T3W = FNMS(TY, TW, T3V);
193
        }
194
        {
195
       E T13, T16, T14, T41, T12, T15;
196
       T13 = cr[WS(rs, 10)];
197
       T16 = ci[WS(rs, 10)];
198
       T12 = W[18];
199
       T14 = T12 * T13;
200
       T41 = T12 * T16;
201
       T15 = W[19];
202
       T17 = FMA(T15, T16, T14);
203
       T42 = FNMS(T15, T13, T41);
204
        }
205
        T11 = TU + T10;
206
        T1e = T17 + T1d;
207
        T6J = T11 - T1e;
208
        T6K = T3U + T3W;
209
        T6L = T42 + T44;
210
        T6M = T6K - T6L;
211
        {
212
       E T3X, T3Y, T40, T45;
213
       T3X = T3U - T3W;
214
       T3Y = T17 - T1d;
215
       T3Z = T3X + T3Y;
216
       T5Y = T3X - T3Y;
217
       T40 = TU - T10;
218
       T45 = T42 - T44;
219
       T46 = T40 - T45;
220
       T5X = T40 + T45;
221
        }
222
         }
223
         {
224
        E T1l, T49, T1E, T4j, T1r, T4b, T1y, T4h;
225
        {
226
       E T1h, T1k, T1i, T48, T1g, T1j;
227
       T1h = cr[WS(rs, 30)];
228
       T1k = ci[WS(rs, 30)];
229
       T1g = W[58];
230
       T1i = T1g * T1h;
231
       T48 = T1g * T1k;
232
       T1j = W[59];
233
       T1l = FMA(T1j, T1k, T1i);
234
       T49 = FNMS(T1j, T1h, T48);
235
        }
236
        {
237
       E T1A, T1D, T1B, T4i, T1z, T1C;
238
       T1A = cr[WS(rs, 22)];
239
       T1D = ci[WS(rs, 22)];
240
       T1z = W[42];
241
       T1B = T1z * T1A;
242
       T4i = T1z * T1D;
243
       T1C = W[43];
244
       T1E = FMA(T1C, T1D, T1B);
245
       T4j = FNMS(T1C, T1A, T4i);
246
        }
247
        {
248
       E T1n, T1q, T1o, T4a, T1m, T1p;
249
       T1n = cr[WS(rs, 14)];
250
       T1q = ci[WS(rs, 14)];
251
       T1m = W[26];
252
       T1o = T1m * T1n;
253
       T4a = T1m * T1q;
254
       T1p = W[27];
255
       T1r = FMA(T1p, T1q, T1o);
256
       T4b = FNMS(T1p, T1n, T4a);
257
        }
258
        {
259
       E T1u, T1x, T1v, T4g, T1t, T1w;
260
       T1u = cr[WS(rs, 6)];
261
       T1x = ci[WS(rs, 6)];
262
       T1t = W[10];
263
       T1v = T1t * T1u;
264
       T4g = T1t * T1x;
265
       T1w = W[11];
266
       T1y = FMA(T1w, T1x, T1v);
267
       T4h = FNMS(T1w, T1u, T4g);
268
        }
269
        T1s = T1l + T1r;
270
        T1F = T1y + T1E;
271
        T6O = T1s - T1F;
272
        T6P = T49 + T4b;
273
        T6Q = T4h + T4j;
274
        T6R = T6P - T6Q;
275
        {
276
       E T4c, T4d, T4f, T4k;
277
       T4c = T49 - T4b;
278
       T4d = T1y - T1E;
279
       T4e = T4c + T4d;
280
       T61 = T4c - T4d;
281
       T4f = T1l - T1r;
282
       T4k = T4h - T4j;
283
       T4l = T4f - T4k;
284
       T60 = T4f + T4k;
285
        }
286
         }
287
         {
288
        E T2H, T5n, T30, T52, T2N, T5p, T2U, T50;
289
        {
290
       E T2D, T2G, T2E, T5m, T2C, T2F;
291
       T2D = cr[WS(rs, 31)];
292
       T2G = ci[WS(rs, 31)];
293
       T2C = W[60];
294
       T2E = T2C * T2D;
295
       T5m = T2C * T2G;
296
       T2F = W[61];
297
       T2H = FMA(T2F, T2G, T2E);
298
       T5n = FNMS(T2F, T2D, T5m);
299
        }
300
        {
301
       E T2W, T2Z, T2X, T51, T2V, T2Y;
302
       T2W = cr[WS(rs, 23)];
303
       T2Z = ci[WS(rs, 23)];
304
       T2V = W[44];
305
       T2X = T2V * T2W;
306
       T51 = T2V * T2Z;
307
       T2Y = W[45];
308
       T30 = FMA(T2Y, T2Z, T2X);
309
       T52 = FNMS(T2Y, T2W, T51);
310
        }
311
        {
312
       E T2J, T2M, T2K, T5o, T2I, T2L;
313
       T2J = cr[WS(rs, 15)];
314
       T2M = ci[WS(rs, 15)];
315
       T2I = W[28];
316
       T2K = T2I * T2J;
317
       T5o = T2I * T2M;
318
       T2L = W[29];
319
       T2N = FMA(T2L, T2M, T2K);
320
       T5p = FNMS(T2L, T2J, T5o);
321
        }
322
        {
323
       E T2Q, T2T, T2R, T4Z, T2P, T2S;
324
       T2Q = cr[WS(rs, 7)];
325
       T2T = ci[WS(rs, 7)];
326
       T2P = W[12];
327
       T2R = T2P * T2Q;
328
       T4Z = T2P * T2T;
329
       T2S = W[13];
330
       T2U = FMA(T2S, T2T, T2R);
331
       T50 = FNMS(T2S, T2Q, T4Z);
332
        }
333
        {
334
       E T2O, T31, T7b, T7c;
335
       T2O = T2H + T2N;
336
       T31 = T2U + T30;
337
       T32 = T2O + T31;
338
       T76 = T2O - T31;
339
       T7b = T5n + T5p;
340
       T7c = T50 + T52;
341
       T7d = T7b - T7c;
342
       T7N = T7b + T7c;
343
        }
344
        {
345
       E T4Y, T53, T5l, T5q;
346
       T4Y = T2H - T2N;
347
       T53 = T50 - T52;
348
       T54 = T4Y - T53;
349
       T6c = T4Y + T53;
350
       T5l = T30 - T2U;
351
       T5q = T5n - T5p;
352
       T5r = T5l - T5q;
353
       T6f = T5q + T5l;
354
        }
355
         }
356
         {
357
        E T1O, T4N, T27, T4t, T1U, T4P, T21, T4r;
358
        {
359
       E T1K, T1N, T1L, T4M, T1J, T1M;
360
       T1K = cr[WS(rs, 1)];
361
       T1N = ci[WS(rs, 1)];
362
       T1J = W[0];
363
       T1L = T1J * T1K;
364
       T4M = T1J * T1N;
365
       T1M = W[1];
366
       T1O = FMA(T1M, T1N, T1L);
367
       T4N = FNMS(T1M, T1K, T4M);
368
        }
369
        {
370
       E T23, T26, T24, T4s, T22, T25;
371
       T23 = cr[WS(rs, 25)];
372
       T26 = ci[WS(rs, 25)];
373
       T22 = W[48];
374
       T24 = T22 * T23;
375
       T4s = T22 * T26;
376
       T25 = W[49];
377
       T27 = FMA(T25, T26, T24);
378
       T4t = FNMS(T25, T23, T4s);
379
        }
380
        {
381
       E T1Q, T1T, T1R, T4O, T1P, T1S;
382
       T1Q = cr[WS(rs, 17)];
383
       T1T = ci[WS(rs, 17)];
384
       T1P = W[32];
385
       T1R = T1P * T1Q;
386
       T4O = T1P * T1T;
387
       T1S = W[33];
388
       T1U = FMA(T1S, T1T, T1R);
389
       T4P = FNMS(T1S, T1Q, T4O);
390
        }
391
        {
392
       E T1X, T20, T1Y, T4q, T1W, T1Z;
393
       T1X = cr[WS(rs, 9)];
394
       T20 = ci[WS(rs, 9)];
395
       T1W = W[16];
396
       T1Y = T1W * T1X;
397
       T4q = T1W * T20;
398
       T1Z = W[17];
399
       T21 = FMA(T1Z, T20, T1Y);
400
       T4r = FNMS(T1Z, T1X, T4q);
401
        }
402
        {
403
       E T1V, T28, T70, T71;
404
       T1V = T1O + T1U;
405
       T28 = T21 + T27;
406
       T29 = T1V + T28;
407
       T6V = T1V - T28;
408
       T70 = T4N + T4P;
409
       T71 = T4r + T4t;
410
       T72 = T70 - T71;
411
       T7I = T70 + T71;
412
        }
413
        {
414
       E T4p, T4u, T4Q, T4R;
415
       T4p = T1O - T1U;
416
       T4u = T4r - T4t;
417
       T4v = T4p - T4u;
418
       T65 = T4p + T4u;
419
       T4Q = T4N - T4P;
420
       T4R = T21 - T27;
421
       T4S = T4Q + T4R;
422
       T68 = T4Q - T4R;
423
        }
424
         }
425
         {
426
        E T38, T57, T3r, T5g, T3e, T59, T3l, T5e;
427
        {
428
       E T34, T37, T35, T56, T33, T36;
429
       T34 = cr[WS(rs, 3)];
430
       T37 = ci[WS(rs, 3)];
431
       T33 = W[4];
432
       T35 = T33 * T34;
433
       T56 = T33 * T37;
434
       T36 = W[5];
435
       T38 = FMA(T36, T37, T35);
436
       T57 = FNMS(T36, T34, T56);
437
        }
438
        {
439
       E T3n, T3q, T3o, T5f, T3m, T3p;
440
       T3n = cr[WS(rs, 11)];
441
       T3q = ci[WS(rs, 11)];
442
       T3m = W[20];
443
       T3o = T3m * T3n;
444
       T5f = T3m * T3q;
445
       T3p = W[21];
446
       T3r = FMA(T3p, T3q, T3o);
447
       T5g = FNMS(T3p, T3n, T5f);
448
        }
449
        {
450
       E T3a, T3d, T3b, T58, T39, T3c;
451
       T3a = cr[WS(rs, 19)];
452
       T3d = ci[WS(rs, 19)];
453
       T39 = W[36];
454
       T3b = T39 * T3a;
455
       T58 = T39 * T3d;
456
       T3c = W[37];
457
       T3e = FMA(T3c, T3d, T3b);
458
       T59 = FNMS(T3c, T3a, T58);
459
        }
460
        {
461
       E T3h, T3k, T3i, T5d, T3g, T3j;
462
       T3h = cr[WS(rs, 27)];
463
       T3k = ci[WS(rs, 27)];
464
       T3g = W[52];
465
       T3i = T3g * T3h;
466
       T5d = T3g * T3k;
467
       T3j = W[53];
468
       T3l = FMA(T3j, T3k, T3i);
469
       T5e = FNMS(T3j, T3h, T5d);
470
        }
471
        {
472
       E T3f, T3s, T77, T78;
473
       T3f = T38 + T3e;
474
       T3s = T3l + T3r;
475
       T3t = T3f + T3s;
476
       T7e = T3s - T3f;
477
       T77 = T5e + T5g;
478
       T78 = T57 + T59;
479
       T79 = T77 - T78;
480
       T7O = T78 + T77;
481
        }
482
        {
483
       E T55, T5a, T5c, T5h;
484
       T55 = T38 - T3e;
485
       T5a = T57 - T59;
486
       T5b = T55 - T5a;
487
       T5t = T55 + T5a;
488
       T5c = T3l - T3r;
489
       T5h = T5e - T5g;
490
       T5i = T5c + T5h;
491
       T5s = T5c - T5h;
492
        }
493
         }
494
         {
495
        E T2f, T4y, T2y, T4H, T2l, T4A, T2s, T4F;
496
        {
497
       E T2b, T2e, T2c, T4x, T2a, T2d;
498
       T2b = cr[WS(rs, 5)];
499
       T2e = ci[WS(rs, 5)];
500
       T2a = W[8];
501
       T2c = T2a * T2b;
502
       T4x = T2a * T2e;
503
       T2d = W[9];
504
       T2f = FMA(T2d, T2e, T2c);
505
       T4y = FNMS(T2d, T2b, T4x);
506
        }
507
        {
508
       E T2u, T2x, T2v, T4G, T2t, T2w;
509
       T2u = cr[WS(rs, 13)];
510
       T2x = ci[WS(rs, 13)];
511
       T2t = W[24];
512
       T2v = T2t * T2u;
513
       T4G = T2t * T2x;
514
       T2w = W[25];
515
       T2y = FMA(T2w, T2x, T2v);
516
       T4H = FNMS(T2w, T2u, T4G);
517
        }
518
        {
519
       E T2h, T2k, T2i, T4z, T2g, T2j;
520
       T2h = cr[WS(rs, 21)];
521
       T2k = ci[WS(rs, 21)];
522
       T2g = W[40];
523
       T2i = T2g * T2h;
524
       T4z = T2g * T2k;
525
       T2j = W[41];
526
       T2l = FMA(T2j, T2k, T2i);
527
       T4A = FNMS(T2j, T2h, T4z);
528
        }
529
        {
530
       E T2o, T2r, T2p, T4E, T2n, T2q;
531
       T2o = cr[WS(rs, 29)];
532
       T2r = ci[WS(rs, 29)];
533
       T2n = W[56];
534
       T2p = T2n * T2o;
535
       T4E = T2n * T2r;
536
       T2q = W[57];
537
       T2s = FMA(T2q, T2r, T2p);
538
       T4F = FNMS(T2q, T2o, T4E);
539
        }
540
        {
541
       E T2m, T2z, T6W, T6X;
542
       T2m = T2f + T2l;
543
       T2z = T2s + T2y;
544
       T2A = T2m + T2z;
545
       T73 = T2m - T2z;
546
       T6W = T4F + T4H;
547
       T6X = T4y + T4A;
548
       T6Y = T6W - T6X;
549
       T7J = T6X + T6W;
550
        }
551
        {
552
       E T4w, T4B, T4D, T4I;
553
       T4w = T2f - T2l;
554
       T4B = T4y - T4A;
555
       T4C = T4w - T4B;
556
       T4T = T4w + T4B;
557
       T4D = T2s - T2y;
558
       T4I = T4F - T4H;
559
       T4J = T4D + T4I;
560
       T4U = T4I - T4D;
561
        }
562
         }
563
         {
564
        E TO, T7C, T7Z, T80, T89, T8e, T1H, T8d, T3v, T8b, T7L, T7T, T7Q, T7U, T7F;
565
        E T81;
566
        {
567
       E Tm, TN, T7X, T7Y;
568
       Tm = T8 + Tl;
569
       TN = Tz + TM;
570
       TO = Tm + TN;
571
       T7C = Tm - TN;
572
       T7X = T7N + T7O;
573
       T7Y = T7I + T7J;
574
       T7Z = T7X - T7Y;
575
       T80 = T7Y + T7X;
576
        }
577
        {
578
       E T82, T88, T1f, T1G;
579
       T82 = T6G + T6F;
580
       T88 = T83 + T87;
581
       T89 = T82 + T88;
582
       T8e = T88 - T82;
583
       T1f = T11 + T1e;
584
       T1G = T1s + T1F;
585
       T1H = T1f + T1G;
586
       T8d = T1f - T1G;
587
        }
588
        {
589
       E T2B, T3u, T7H, T7K;
590
       T2B = T29 + T2A;
591
       T3u = T32 + T3t;
592
       T3v = T2B + T3u;
593
       T8b = T3u - T2B;
594
       T7H = T29 - T2A;
595
       T7K = T7I - T7J;
596
       T7L = T7H + T7K;
597
       T7T = T7H - T7K;
598
        }
599
        {
600
       E T7M, T7P, T7D, T7E;
601
       T7M = T32 - T3t;
602
       T7P = T7N - T7O;
603
       T7Q = T7M - T7P;
604
       T7U = T7M + T7P;
605
       T7D = T6P + T6Q;
606
       T7E = T6K + T6L;
607
       T7F = T7D - T7E;
608
       T81 = T7E + T7D;
609
        }
610
        {
611
       E T1I, T8a, T8c, T7W;
612
       T1I = TO + T1H;
613
       ci[WS(rs, 15)] = T1I - T3v;
614
       cr[0] = T1I + T3v;
615
       T8a = T81 + T89;
616
       cr[WS(rs, 16)] = T80 - T8a;
617
       ci[WS(rs, 31)] = T80 + T8a;
618
       T8c = T89 - T81;
619
       cr[WS(rs, 24)] = T8b - T8c;
620
       ci[WS(rs, 23)] = T8b + T8c;
621
       T7W = TO - T1H;
622
       cr[WS(rs, 8)] = T7W - T7Z;
623
       ci[WS(rs, 7)] = T7W + T7Z;
624
        }
625
        {
626
       E T7G, T7R, T8f, T8g;
627
       T7G = T7C - T7F;
628
       T7R = T7L + T7Q;
629
       ci[WS(rs, 11)] = FNMS(KP707106781, T7R, T7G);
630
       cr[WS(rs, 4)] = FMA(KP707106781, T7R, T7G);
631
       T8f = T8d + T8e;
632
       T8g = T7Q - T7L;
633
       cr[WS(rs, 28)] = FMS(KP707106781, T8g, T8f);
634
       ci[WS(rs, 19)] = FMA(KP707106781, T8g, T8f);
635
        }
636
        {
637
       E T8h, T8i, T7S, T7V;
638
       T8h = T8e - T8d;
639
       T8i = T7U - T7T;
640
       cr[WS(rs, 20)] = FMS(KP707106781, T8i, T8h);
641
       ci[WS(rs, 27)] = FMA(KP707106781, T8i, T8h);
642
       T7S = T7C + T7F;
643
       T7V = T7T + T7U;
644
       cr[WS(rs, 12)] = FNMS(KP707106781, T7V, T7S);
645
       ci[WS(rs, 3)] = FMA(KP707106781, T7V, T7S);
646
        }
647
         }
648
         {
649
        E T3S, T5C, T4n, T8C, T8B, T8H, T5F, T8I, T5w, T5Q, T5A, T5M, T4X, T5P, T5z;
650
        E T5J;
651
        {
652
       E T3C, T3R, T5D, T5E;
653
       T3C = T3w - T3B;
654
       T3R = T3J + T3Q;
655
       T3S = FNMS(KP707106781, T3R, T3C);
656
       T5C = FMA(KP707106781, T3R, T3C);
657
       {
658
            E T47, T4m, T8z, T8A;
659
            T47 = FMA(KP414213562, T46, T3Z);
660
            T4m = FNMS(KP414213562, T4l, T4e);
661
            T4n = T47 - T4m;
662
            T8C = T47 + T4m;
663
            T8z = T8x + T8y;
664
            T8A = T5T - T5U;
665
            T8B = FMA(KP707106781, T8A, T8z);
666
            T8H = FNMS(KP707106781, T8A, T8z);
667
       }
668
       T5D = FNMS(KP414213562, T3Z, T46);
669
       T5E = FMA(KP414213562, T4e, T4l);
670
       T5F = T5D + T5E;
671
       T8I = T5E - T5D;
672
       {
673
            E T5k, T5K, T5v, T5L, T5j, T5u;
674
            T5j = T5b + T5i;
675
            T5k = FNMS(KP707106781, T5j, T54);
676
            T5K = FMA(KP707106781, T5j, T54);
677
            T5u = T5s - T5t;
678
            T5v = FNMS(KP707106781, T5u, T5r);
679
            T5L = FMA(KP707106781, T5u, T5r);
680
            T5w = FMA(KP668178637, T5v, T5k);
681
            T5Q = FMA(KP198912367, T5K, T5L);
682
            T5A = FNMS(KP668178637, T5k, T5v);
683
            T5M = FNMS(KP198912367, T5L, T5K);
684
       }
685
       {
686
            E T4L, T5H, T4W, T5I, T4K, T4V;
687
            T4K = T4C + T4J;
688
            T4L = FNMS(KP707106781, T4K, T4v);
689
            T5H = FMA(KP707106781, T4K, T4v);
690
            T4V = T4T + T4U;
691
            T4W = FNMS(KP707106781, T4V, T4S);
692
            T5I = FMA(KP707106781, T4V, T4S);
693
            T4X = FMA(KP668178637, T4W, T4L);
694
            T5P = FMA(KP198912367, T5H, T5I);
695
            T5z = FNMS(KP668178637, T4L, T4W);
696
            T5J = FNMS(KP198912367, T5I, T5H);
697
       }
698
        }
699
        {
700
       E T4o, T5x, T8J, T8K;
701
       T4o = FMA(KP923879532, T4n, T3S);
702
       T5x = T4X + T5w;
703
       ci[WS(rs, 12)] = FNMS(KP831469612, T5x, T4o);
704
       cr[WS(rs, 3)] = FMA(KP831469612, T5x, T4o);
705
       T8J = FMA(KP923879532, T8I, T8H);
706
       T8K = T5z - T5A;
707
       cr[WS(rs, 19)] = FMS(KP831469612, T8K, T8J);
708
       ci[WS(rs, 28)] = FMA(KP831469612, T8K, T8J);
709
        }
710
        {
711
       E T8L, T8M, T5y, T5B;
712
       T8L = FNMS(KP923879532, T8I, T8H);
713
       T8M = T5w - T4X;
714
       cr[WS(rs, 27)] = FMS(KP831469612, T8M, T8L);
715
       ci[WS(rs, 20)] = FMA(KP831469612, T8M, T8L);
716
       T5y = FNMS(KP923879532, T4n, T3S);
717
       T5B = T5z + T5A;
718
       cr[WS(rs, 11)] = FMA(KP831469612, T5B, T5y);
719
       ci[WS(rs, 4)] = FNMS(KP831469612, T5B, T5y);
720
        }
721
        {
722
       E T5G, T5N, T8D, T8E;
723
       T5G = FMA(KP923879532, T5F, T5C);
724
       T5N = T5J + T5M;
725
       cr[WS(rs, 15)] = FNMS(KP980785280, T5N, T5G);
726
       ci[0] = FMA(KP980785280, T5N, T5G);
727
       T8D = FMA(KP923879532, T8C, T8B);
728
       T8E = T5Q - T5P;
729
       cr[WS(rs, 31)] = FMS(KP980785280, T8E, T8D);
730
       ci[WS(rs, 16)] = FMA(KP980785280, T8E, T8D);
731
        }
732
        {
733
       E T8F, T8G, T5O, T5R;
734
       T8F = FNMS(KP923879532, T8C, T8B);
735
       T8G = T5M - T5J;
736
       cr[WS(rs, 23)] = FMS(KP980785280, T8G, T8F);
737
       ci[WS(rs, 24)] = FMA(KP980785280, T8G, T8F);
738
       T5O = FNMS(KP923879532, T5F, T5C);
739
       T5R = T5P + T5Q;
740
       ci[WS(rs, 8)] = FNMS(KP980785280, T5R, T5O);
741
       cr[WS(rs, 7)] = FMA(KP980785280, T5R, T5O);
742
        }
743
         }
744
         {
745
        E T6I, T7m, T7w, T7A, T8l, T8r, T6T, T8m, T75, T7k, T7p, T8s, T7t, T7z, T7g;
746
        E T7j;
747
        {
748
       E T6E, T6H, T7u, T7v;
749
       T6E = T8 - Tl;
750
       T6H = T6F - T6G;
751
       T6I = T6E - T6H;
752
       T7m = T6E + T6H;
753
       T7u = T76 + T79;
754
       T7v = T7e - T7d;
755
       T7w = FNMS(KP414213562, T7v, T7u);
756
       T7A = FMA(KP414213562, T7u, T7v);
757
        }
758
        {
759
       E T8j, T8k, T6N, T6S;
760
       T8j = Tz - TM;
761
       T8k = T87 - T83;
762
       T8l = T8j + T8k;
763
       T8r = T8k - T8j;
764
       T6N = T6J + T6M;
765
       T6S = T6O - T6R;
766
       T6T = T6N + T6S;
767
       T8m = T6N - T6S;
768
        }
769
        {
770
       E T6Z, T74, T7n, T7o;
771
       T6Z = T6V - T6Y;
772
       T74 = T72 - T73;
773
       T75 = FMA(KP414213562, T74, T6Z);
774
       T7k = FNMS(KP414213562, T6Z, T74);
775
       T7n = T6J - T6M;
776
       T7o = T6O + T6R;
777
       T7p = T7n + T7o;
778
       T8s = T7o - T7n;
779
        }
780
        {
781
       E T7r, T7s, T7a, T7f;
782
       T7r = T6V + T6Y;
783
       T7s = T72 + T73;
784
       T7t = FNMS(KP414213562, T7s, T7r);
785
       T7z = FMA(KP414213562, T7r, T7s);
786
       T7a = T76 - T79;
787
       T7f = T7d + T7e;
788
       T7g = FNMS(KP414213562, T7f, T7a);
789
       T7j = FMA(KP414213562, T7a, T7f);
790
        }
791
        {
792
       E T6U, T7h, T8t, T8u;
793
       T6U = FMA(KP707106781, T6T, T6I);
794
       T7h = T75 + T7g;
795
       ci[WS(rs, 13)] = FNMS(KP923879532, T7h, T6U);
796
       cr[WS(rs, 2)] = FMA(KP923879532, T7h, T6U);
797
       T8t = FMA(KP707106781, T8s, T8r);
798
       T8u = T7k + T7j;
799
       cr[WS(rs, 18)] = FMS(KP923879532, T8u, T8t);
800
       ci[WS(rs, 29)] = FMA(KP923879532, T8u, T8t);
801
        }
802
        {
803
       E T8v, T8w, T7i, T7l;
804
       T8v = FNMS(KP707106781, T8s, T8r);
805
       T8w = T7g - T75;
806
       cr[WS(rs, 26)] = FMS(KP923879532, T8w, T8v);
807
       ci[WS(rs, 21)] = FMA(KP923879532, T8w, T8v);
808
       T7i = FNMS(KP707106781, T6T, T6I);
809
       T7l = T7j - T7k;
810
       cr[WS(rs, 10)] = FNMS(KP923879532, T7l, T7i);
811
       ci[WS(rs, 5)] = FMA(KP923879532, T7l, T7i);
812
        }
813
        {
814
       E T7q, T7x, T8n, T8o;
815
       T7q = FMA(KP707106781, T7p, T7m);
816
       T7x = T7t + T7w;
817
       cr[WS(rs, 14)] = FNMS(KP923879532, T7x, T7q);
818
       ci[WS(rs, 1)] = FMA(KP923879532, T7x, T7q);
819
       T8n = FMA(KP707106781, T8m, T8l);
820
       T8o = T7A - T7z;
821
       cr[WS(rs, 30)] = FMS(KP923879532, T8o, T8n);
822
       ci[WS(rs, 17)] = FMA(KP923879532, T8o, T8n);
823
        }
824
        {
825
       E T8p, T8q, T7y, T7B;
826
       T8p = FNMS(KP707106781, T8m, T8l);
827
       T8q = T7w - T7t;
828
       cr[WS(rs, 22)] = FMS(KP923879532, T8q, T8p);
829
       ci[WS(rs, 25)] = FMA(KP923879532, T8q, T8p);
830
       T7y = FNMS(KP707106781, T7p, T7m);
831
       T7B = T7z + T7A;
832
       ci[WS(rs, 9)] = FNMS(KP923879532, T7B, T7y);
833
       cr[WS(rs, 6)] = FMA(KP923879532, T7B, T7y);
834
        }
835
         }
836
         {
837
        E T5W, T6o, T63, T8W, T8P, T8V, T6r, T8Q, T6i, T6C, T6l, T6y, T6b, T6B, T6m;
838
        E T6v;
839
        {
840
       E T5S, T5V, T6p, T6q;
841
       T5S = T3w + T3B;
842
       T5V = T5T + T5U;
843
       T5W = FMA(KP707106781, T5V, T5S);
844
       T6o = FNMS(KP707106781, T5V, T5S);
845
       {
846
            E T5Z, T62, T8N, T8O;
847
            T5Z = FMA(KP414213562, T5Y, T5X);
848
            T62 = FNMS(KP414213562, T61, T60);
849
            T63 = T5Z + T62;
850
            T8W = T5Z - T62;
851
            T8N = T8y - T8x;
852
            T8O = T3Q - T3J;
853
            T8P = FMA(KP707106781, T8O, T8N);
854
            T8V = FNMS(KP707106781, T8O, T8N);
855
       }
856
       T6p = FMA(KP414213562, T60, T61);
857
       T6q = FNMS(KP414213562, T5X, T5Y);
858
       T6r = T6p - T6q;
859
       T8Q = T6q + T6p;
860
       {
861
            E T6e, T6w, T6h, T6x, T6d, T6g;
862
            T6d = T5t + T5s;
863
            T6e = FMA(KP707106781, T6d, T6c);
864
            T6w = FNMS(KP707106781, T6d, T6c);
865
            T6g = T5i - T5b;
866
            T6h = FMA(KP707106781, T6g, T6f);
867
            T6x = FNMS(KP707106781, T6g, T6f);
868
            T6i = FNMS(KP198912367, T6h, T6e);
869
            T6C = FNMS(KP668178637, T6w, T6x);
870
            T6l = FMA(KP198912367, T6e, T6h);
871
            T6y = FMA(KP668178637, T6x, T6w);
872
       }
873
       {
874
            E T67, T6t, T6a, T6u, T66, T69;
875
            T66 = T4T - T4U;
876
            T67 = FMA(KP707106781, T66, T65);
877
            T6t = FNMS(KP707106781, T66, T65);
878
            T69 = T4J - T4C;
879
            T6a = FMA(KP707106781, T69, T68);
880
            T6u = FNMS(KP707106781, T69, T68);
881
            T6b = FMA(KP198912367, T6a, T67);
882
            T6B = FMA(KP668178637, T6t, T6u);
883
            T6m = FNMS(KP198912367, T67, T6a);
884
            T6v = FNMS(KP668178637, T6u, T6t);
885
       }
886
        }
887
        {
888
       E T64, T6j, T8X, T8Y;
889
       T64 = FMA(KP923879532, T63, T5W);
890
       T6j = T6b + T6i;
891
       ci[WS(rs, 14)] = FNMS(KP980785280, T6j, T64);
892
       cr[WS(rs, 1)] = FMA(KP980785280, T6j, T64);
893
       T8X = FMA(KP923879532, T8W, T8V);
894
       T8Y = T6B + T6C;
895
       cr[WS(rs, 29)] = -(FMA(KP831469612, T8Y, T8X));
896
       ci[WS(rs, 18)] = FNMS(KP831469612, T8Y, T8X);
897
        }
898
        {
899
       E T8Z, T90, T6k, T6n;
900
       T8Z = FNMS(KP923879532, T8W, T8V);
901
       T90 = T6y - T6v;
902
       cr[WS(rs, 21)] = FMS(KP831469612, T90, T8Z);
903
       ci[WS(rs, 26)] = FMA(KP831469612, T90, T8Z);
904
       T6k = FNMS(KP923879532, T63, T5W);
905
       T6n = T6l - T6m;
906
       cr[WS(rs, 9)] = FNMS(KP980785280, T6n, T6k);
907
       ci[WS(rs, 6)] = FMA(KP980785280, T6n, T6k);
908
        }
909
        {
910
       E T6s, T6z, T8R, T8S;
911
       T6s = FMA(KP923879532, T6r, T6o);
912
       T6z = T6v + T6y;
913
       cr[WS(rs, 13)] = FNMS(KP831469612, T6z, T6s);
914
       ci[WS(rs, 2)] = FMA(KP831469612, T6z, T6s);
915
       T8R = FMA(KP923879532, T8Q, T8P);
916
       T8S = T6m + T6l;
917
       cr[WS(rs, 17)] = FMS(KP980785280, T8S, T8R);
918
       ci[WS(rs, 30)] = FMA(KP980785280, T8S, T8R);
919
        }
920
        {
921
       E T8T, T8U, T6A, T6D;
922
       T8T = FNMS(KP923879532, T8Q, T8P);
923
       T8U = T6i - T6b;
924
       cr[WS(rs, 25)] = FMS(KP980785280, T8U, T8T);
925
       ci[WS(rs, 22)] = FMA(KP980785280, T8U, T8T);
926
       T6A = FNMS(KP923879532, T6r, T6o);
927
       T6D = T6B - T6C;
928
       ci[WS(rs, 10)] = FNMS(KP831469612, T6D, T6A);
929
       cr[WS(rs, 5)] = FMA(KP831469612, T6D, T6A);
930
        }
931
         }
932
    }
933
     }
934
}
935
936
static const tw_instr twinstr[] = {
937
     { TW_FULL, 1, 32 },
938
     { TW_NEXT, 1, 0 }
939
};
940
941
static const hc2hc_desc desc = { 32, "hf_32", twinstr, &GENUS, { 236, 62, 198, 0 } };
942
943
void X(codelet_hf_32) (planner *p) {
944
     X(khc2hc_register) (p, hf_32, &desc);
945
}
946
#else
947
948
/* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 32 -dit -name hf_32 -include rdft/scalar/hf.h */
949
950
/*
951
 * This function contains 434 FP additions, 208 FP multiplications,
952
 * (or, 340 additions, 114 multiplications, 94 fused multiply/add),
953
 * 96 stack variables, 7 constants, and 128 memory accesses
954
 */
955
#include "rdft/scalar/hf.h"
956
957
static void hf_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
958
0
{
959
0
     DK(KP555570233, +0.555570233019602224742830813948532874374937191);
960
0
     DK(KP831469612, +0.831469612302545237078788377617905756738560812);
961
0
     DK(KP980785280, +0.980785280403230449126182236134239036973933731);
962
0
     DK(KP195090322, +0.195090322016128267848284868477022240927691618);
963
0
     DK(KP382683432, +0.382683432365089771728459984030398866761344562);
964
0
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
965
0
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
966
0
     {
967
0
    INT m;
968
0
    for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 62, MAKE_VOLATILE_STRIDE(64, rs)) {
969
0
         E Tj, T5F, T7C, T7Q, T35, T4T, T78, T7m, T1Q, T61, T5Y, T6J, T3K, T56, T41;
970
0
         E T59, T2B, T67, T6e, T6O, T4b, T5g, T4s, T5d, TG, T7l, T5I, T73, T3a, T4U;
971
0
         E T3f, T4V, T14, T5K, T5N, T6F, T3m, T4Z, T3r, T4Y, T1r, T5P, T5S, T6E, T3x;
972
0
         E T52, T3C, T51, T2d, T5Z, T64, T6K, T3V, T5a, T44, T57, T2Y, T6f, T6a, T6P;
973
0
         E T4m, T5e, T4v, T5h;
974
0
         {
975
0
        E T1, T76, T6, T75, Tc, T32, Th, T33;
976
0
        T1 = cr[0];
977
0
        T76 = ci[0];
978
0
        {
979
0
       E T3, T5, T2, T4;
980
0
       T3 = cr[WS(rs, 16)];
981
0
       T5 = ci[WS(rs, 16)];
982
0
       T2 = W[30];
983
0
       T4 = W[31];
984
0
       T6 = FMA(T2, T3, T4 * T5);
985
0
       T75 = FNMS(T4, T3, T2 * T5);
986
0
        }
987
0
        {
988
0
       E T9, Tb, T8, Ta;
989
0
       T9 = cr[WS(rs, 8)];
990
0
       Tb = ci[WS(rs, 8)];
991
0
       T8 = W[14];
992
0
       Ta = W[15];
993
0
       Tc = FMA(T8, T9, Ta * Tb);
994
0
       T32 = FNMS(Ta, T9, T8 * Tb);
995
0
        }
996
0
        {
997
0
       E Te, Tg, Td, Tf;
998
0
       Te = cr[WS(rs, 24)];
999
0
       Tg = ci[WS(rs, 24)];
1000
0
       Td = W[46];
1001
0
       Tf = W[47];
1002
0
       Th = FMA(Td, Te, Tf * Tg);
1003
0
       T33 = FNMS(Tf, Te, Td * Tg);
1004
0
        }
1005
0
        {
1006
0
       E T7, Ti, T7A, T7B;
1007
0
       T7 = T1 + T6;
1008
0
       Ti = Tc + Th;
1009
0
       Tj = T7 + Ti;
1010
0
       T5F = T7 - Ti;
1011
0
       T7A = Tc - Th;
1012
0
       T7B = T76 - T75;
1013
0
       T7C = T7A + T7B;
1014
0
       T7Q = T7B - T7A;
1015
0
        }
1016
0
        {
1017
0
       E T31, T34, T74, T77;
1018
0
       T31 = T1 - T6;
1019
0
       T34 = T32 - T33;
1020
0
       T35 = T31 + T34;
1021
0
       T4T = T31 - T34;
1022
0
       T74 = T32 + T33;
1023
0
       T77 = T75 + T76;
1024
0
       T78 = T74 + T77;
1025
0
       T7m = T77 - T74;
1026
0
        }
1027
0
         }
1028
0
         {
1029
0
        E T1y, T3X, T1O, T3I, T1D, T3Y, T1J, T3H;
1030
0
        {
1031
0
       E T1v, T1x, T1u, T1w;
1032
0
       T1v = cr[WS(rs, 1)];
1033
0
       T1x = ci[WS(rs, 1)];
1034
0
       T1u = W[0];
1035
0
       T1w = W[1];
1036
0
       T1y = FMA(T1u, T1v, T1w * T1x);
1037
0
       T3X = FNMS(T1w, T1v, T1u * T1x);
1038
0
        }
1039
0
        {
1040
0
       E T1L, T1N, T1K, T1M;
1041
0
       T1L = cr[WS(rs, 25)];
1042
0
       T1N = ci[WS(rs, 25)];
1043
0
       T1K = W[48];
1044
0
       T1M = W[49];
1045
0
       T1O = FMA(T1K, T1L, T1M * T1N);
1046
0
       T3I = FNMS(T1M, T1L, T1K * T1N);
1047
0
        }
1048
0
        {
1049
0
       E T1A, T1C, T1z, T1B;
1050
0
       T1A = cr[WS(rs, 17)];
1051
0
       T1C = ci[WS(rs, 17)];
1052
0
       T1z = W[32];
1053
0
       T1B = W[33];
1054
0
       T1D = FMA(T1z, T1A, T1B * T1C);
1055
0
       T3Y = FNMS(T1B, T1A, T1z * T1C);
1056
0
        }
1057
0
        {
1058
0
       E T1G, T1I, T1F, T1H;
1059
0
       T1G = cr[WS(rs, 9)];
1060
0
       T1I = ci[WS(rs, 9)];
1061
0
       T1F = W[16];
1062
0
       T1H = W[17];
1063
0
       T1J = FMA(T1F, T1G, T1H * T1I);
1064
0
       T3H = FNMS(T1H, T1G, T1F * T1I);
1065
0
        }
1066
0
        {
1067
0
       E T1E, T1P, T5W, T5X;
1068
0
       T1E = T1y + T1D;
1069
0
       T1P = T1J + T1O;
1070
0
       T1Q = T1E + T1P;
1071
0
       T61 = T1E - T1P;
1072
0
       T5W = T3X + T3Y;
1073
0
       T5X = T3H + T3I;
1074
0
       T5Y = T5W - T5X;
1075
0
       T6J = T5W + T5X;
1076
0
        }
1077
0
        {
1078
0
       E T3G, T3J, T3Z, T40;
1079
0
       T3G = T1y - T1D;
1080
0
       T3J = T3H - T3I;
1081
0
       T3K = T3G + T3J;
1082
0
       T56 = T3G - T3J;
1083
0
       T3Z = T3X - T3Y;
1084
0
       T40 = T1J - T1O;
1085
0
       T41 = T3Z - T40;
1086
0
       T59 = T3Z + T40;
1087
0
        }
1088
0
         }
1089
0
         {
1090
0
        E T2j, T47, T2z, T4q, T2o, T48, T2u, T4p;
1091
0
        {
1092
0
       E T2g, T2i, T2f, T2h;
1093
0
       T2g = cr[WS(rs, 31)];
1094
0
       T2i = ci[WS(rs, 31)];
1095
0
       T2f = W[60];
1096
0
       T2h = W[61];
1097
0
       T2j = FMA(T2f, T2g, T2h * T2i);
1098
0
       T47 = FNMS(T2h, T2g, T2f * T2i);
1099
0
        }
1100
0
        {
1101
0
       E T2w, T2y, T2v, T2x;
1102
0
       T2w = cr[WS(rs, 23)];
1103
0
       T2y = ci[WS(rs, 23)];
1104
0
       T2v = W[44];
1105
0
       T2x = W[45];
1106
0
       T2z = FMA(T2v, T2w, T2x * T2y);
1107
0
       T4q = FNMS(T2x, T2w, T2v * T2y);
1108
0
        }
1109
0
        {
1110
0
       E T2l, T2n, T2k, T2m;
1111
0
       T2l = cr[WS(rs, 15)];
1112
0
       T2n = ci[WS(rs, 15)];
1113
0
       T2k = W[28];
1114
0
       T2m = W[29];
1115
0
       T2o = FMA(T2k, T2l, T2m * T2n);
1116
0
       T48 = FNMS(T2m, T2l, T2k * T2n);
1117
0
        }
1118
0
        {
1119
0
       E T2r, T2t, T2q, T2s;
1120
0
       T2r = cr[WS(rs, 7)];
1121
0
       T2t = ci[WS(rs, 7)];
1122
0
       T2q = W[12];
1123
0
       T2s = W[13];
1124
0
       T2u = FMA(T2q, T2r, T2s * T2t);
1125
0
       T4p = FNMS(T2s, T2r, T2q * T2t);
1126
0
        }
1127
0
        {
1128
0
       E T2p, T2A, T6c, T6d;
1129
0
       T2p = T2j + T2o;
1130
0
       T2A = T2u + T2z;
1131
0
       T2B = T2p + T2A;
1132
0
       T67 = T2p - T2A;
1133
0
       T6c = T47 + T48;
1134
0
       T6d = T4p + T4q;
1135
0
       T6e = T6c - T6d;
1136
0
       T6O = T6c + T6d;
1137
0
        }
1138
0
        {
1139
0
       E T49, T4a, T4o, T4r;
1140
0
       T49 = T47 - T48;
1141
0
       T4a = T2u - T2z;
1142
0
       T4b = T49 - T4a;
1143
0
       T5g = T49 + T4a;
1144
0
       T4o = T2j - T2o;
1145
0
       T4r = T4p - T4q;
1146
0
       T4s = T4o + T4r;
1147
0
       T5d = T4o - T4r;
1148
0
        }
1149
0
         }
1150
0
         {
1151
0
        E To, T37, TE, T3d, Tt, T38, Tz, T3c;
1152
0
        {
1153
0
       E Tl, Tn, Tk, Tm;
1154
0
       Tl = cr[WS(rs, 4)];
1155
0
       Tn = ci[WS(rs, 4)];
1156
0
       Tk = W[6];
1157
0
       Tm = W[7];
1158
0
       To = FMA(Tk, Tl, Tm * Tn);
1159
0
       T37 = FNMS(Tm, Tl, Tk * Tn);
1160
0
        }
1161
0
        {
1162
0
       E TB, TD, TA, TC;
1163
0
       TB = cr[WS(rs, 12)];
1164
0
       TD = ci[WS(rs, 12)];
1165
0
       TA = W[22];
1166
0
       TC = W[23];
1167
0
       TE = FMA(TA, TB, TC * TD);
1168
0
       T3d = FNMS(TC, TB, TA * TD);
1169
0
        }
1170
0
        {
1171
0
       E Tq, Ts, Tp, Tr;
1172
0
       Tq = cr[WS(rs, 20)];
1173
0
       Ts = ci[WS(rs, 20)];
1174
0
       Tp = W[38];
1175
0
       Tr = W[39];
1176
0
       Tt = FMA(Tp, Tq, Tr * Ts);
1177
0
       T38 = FNMS(Tr, Tq, Tp * Ts);
1178
0
        }
1179
0
        {
1180
0
       E Tw, Ty, Tv, Tx;
1181
0
       Tw = cr[WS(rs, 28)];
1182
0
       Ty = ci[WS(rs, 28)];
1183
0
       Tv = W[54];
1184
0
       Tx = W[55];
1185
0
       Tz = FMA(Tv, Tw, Tx * Ty);
1186
0
       T3c = FNMS(Tx, Tw, Tv * Ty);
1187
0
        }
1188
0
        {
1189
0
       E Tu, TF, T5G, T5H;
1190
0
       Tu = To + Tt;
1191
0
       TF = Tz + TE;
1192
0
       TG = Tu + TF;
1193
0
       T7l = Tu - TF;
1194
0
       T5G = T3c + T3d;
1195
0
       T5H = T37 + T38;
1196
0
       T5I = T5G - T5H;
1197
0
       T73 = T5H + T5G;
1198
0
        }
1199
0
        {
1200
0
       E T36, T39, T3b, T3e;
1201
0
       T36 = To - Tt;
1202
0
       T39 = T37 - T38;
1203
0
       T3a = T36 + T39;
1204
0
       T4U = T36 - T39;
1205
0
       T3b = Tz - TE;
1206
0
       T3e = T3c - T3d;
1207
0
       T3f = T3b - T3e;
1208
0
       T4V = T3b + T3e;
1209
0
        }
1210
0
         }
1211
0
         {
1212
0
        E TM, T3n, T12, T3k, TR, T3o, TX, T3j;
1213
0
        {
1214
0
       E TJ, TL, TI, TK;
1215
0
       TJ = cr[WS(rs, 2)];
1216
0
       TL = ci[WS(rs, 2)];
1217
0
       TI = W[2];
1218
0
       TK = W[3];
1219
0
       TM = FMA(TI, TJ, TK * TL);
1220
0
       T3n = FNMS(TK, TJ, TI * TL);
1221
0
        }
1222
0
        {
1223
0
       E TZ, T11, TY, T10;
1224
0
       TZ = cr[WS(rs, 26)];
1225
0
       T11 = ci[WS(rs, 26)];
1226
0
       TY = W[50];
1227
0
       T10 = W[51];
1228
0
       T12 = FMA(TY, TZ, T10 * T11);
1229
0
       T3k = FNMS(T10, TZ, TY * T11);
1230
0
        }
1231
0
        {
1232
0
       E TO, TQ, TN, TP;
1233
0
       TO = cr[WS(rs, 18)];
1234
0
       TQ = ci[WS(rs, 18)];
1235
0
       TN = W[34];
1236
0
       TP = W[35];
1237
0
       TR = FMA(TN, TO, TP * TQ);
1238
0
       T3o = FNMS(TP, TO, TN * TQ);
1239
0
        }
1240
0
        {
1241
0
       E TU, TW, TT, TV;
1242
0
       TU = cr[WS(rs, 10)];
1243
0
       TW = ci[WS(rs, 10)];
1244
0
       TT = W[18];
1245
0
       TV = W[19];
1246
0
       TX = FMA(TT, TU, TV * TW);
1247
0
       T3j = FNMS(TV, TU, TT * TW);
1248
0
        }
1249
0
        {
1250
0
       E TS, T13, T5L, T5M;
1251
0
       TS = TM + TR;
1252
0
       T13 = TX + T12;
1253
0
       T14 = TS + T13;
1254
0
       T5K = TS - T13;
1255
0
       T5L = T3n + T3o;
1256
0
       T5M = T3j + T3k;
1257
0
       T5N = T5L - T5M;
1258
0
       T6F = T5L + T5M;
1259
0
        }
1260
0
        {
1261
0
       E T3i, T3l, T3p, T3q;
1262
0
       T3i = TM - TR;
1263
0
       T3l = T3j - T3k;
1264
0
       T3m = T3i + T3l;
1265
0
       T4Z = T3i - T3l;
1266
0
       T3p = T3n - T3o;
1267
0
       T3q = TX - T12;
1268
0
       T3r = T3p - T3q;
1269
0
       T4Y = T3p + T3q;
1270
0
        }
1271
0
         }
1272
0
         {
1273
0
        E T19, T3t, T1p, T3A, T1e, T3u, T1k, T3z;
1274
0
        {
1275
0
       E T16, T18, T15, T17;
1276
0
       T16 = cr[WS(rs, 30)];
1277
0
       T18 = ci[WS(rs, 30)];
1278
0
       T15 = W[58];
1279
0
       T17 = W[59];
1280
0
       T19 = FMA(T15, T16, T17 * T18);
1281
0
       T3t = FNMS(T17, T16, T15 * T18);
1282
0
        }
1283
0
        {
1284
0
       E T1m, T1o, T1l, T1n;
1285
0
       T1m = cr[WS(rs, 22)];
1286
0
       T1o = ci[WS(rs, 22)];
1287
0
       T1l = W[42];
1288
0
       T1n = W[43];
1289
0
       T1p = FMA(T1l, T1m, T1n * T1o);
1290
0
       T3A = FNMS(T1n, T1m, T1l * T1o);
1291
0
        }
1292
0
        {
1293
0
       E T1b, T1d, T1a, T1c;
1294
0
       T1b = cr[WS(rs, 14)];
1295
0
       T1d = ci[WS(rs, 14)];
1296
0
       T1a = W[26];
1297
0
       T1c = W[27];
1298
0
       T1e = FMA(T1a, T1b, T1c * T1d);
1299
0
       T3u = FNMS(T1c, T1b, T1a * T1d);
1300
0
        }
1301
0
        {
1302
0
       E T1h, T1j, T1g, T1i;
1303
0
       T1h = cr[WS(rs, 6)];
1304
0
       T1j = ci[WS(rs, 6)];
1305
0
       T1g = W[10];
1306
0
       T1i = W[11];
1307
0
       T1k = FMA(T1g, T1h, T1i * T1j);
1308
0
       T3z = FNMS(T1i, T1h, T1g * T1j);
1309
0
        }
1310
0
        {
1311
0
       E T1f, T1q, T5Q, T5R;
1312
0
       T1f = T19 + T1e;
1313
0
       T1q = T1k + T1p;
1314
0
       T1r = T1f + T1q;
1315
0
       T5P = T1f - T1q;
1316
0
       T5Q = T3t + T3u;
1317
0
       T5R = T3z + T3A;
1318
0
       T5S = T5Q - T5R;
1319
0
       T6E = T5Q + T5R;
1320
0
        }
1321
0
        {
1322
0
       E T3v, T3w, T3y, T3B;
1323
0
       T3v = T3t - T3u;
1324
0
       T3w = T1k - T1p;
1325
0
       T3x = T3v - T3w;
1326
0
       T52 = T3v + T3w;
1327
0
       T3y = T19 - T1e;
1328
0
       T3B = T3z - T3A;
1329
0
       T3C = T3y + T3B;
1330
0
       T51 = T3y - T3B;
1331
0
        }
1332
0
         }
1333
0
         {
1334
0
        E T1V, T3M, T20, T3N, T3L, T3O, T26, T3Q, T2b, T3R, T3S, T3T;
1335
0
        {
1336
0
       E T1S, T1U, T1R, T1T;
1337
0
       T1S = cr[WS(rs, 5)];
1338
0
       T1U = ci[WS(rs, 5)];
1339
0
       T1R = W[8];
1340
0
       T1T = W[9];
1341
0
       T1V = FMA(T1R, T1S, T1T * T1U);
1342
0
       T3M = FNMS(T1T, T1S, T1R * T1U);
1343
0
        }
1344
0
        {
1345
0
       E T1X, T1Z, T1W, T1Y;
1346
0
       T1X = cr[WS(rs, 21)];
1347
0
       T1Z = ci[WS(rs, 21)];
1348
0
       T1W = W[40];
1349
0
       T1Y = W[41];
1350
0
       T20 = FMA(T1W, T1X, T1Y * T1Z);
1351
0
       T3N = FNMS(T1Y, T1X, T1W * T1Z);
1352
0
        }
1353
0
        T3L = T1V - T20;
1354
0
        T3O = T3M - T3N;
1355
0
        {
1356
0
       E T23, T25, T22, T24;
1357
0
       T23 = cr[WS(rs, 29)];
1358
0
       T25 = ci[WS(rs, 29)];
1359
0
       T22 = W[56];
1360
0
       T24 = W[57];
1361
0
       T26 = FMA(T22, T23, T24 * T25);
1362
0
       T3Q = FNMS(T24, T23, T22 * T25);
1363
0
        }
1364
0
        {
1365
0
       E T28, T2a, T27, T29;
1366
0
       T28 = cr[WS(rs, 13)];
1367
0
       T2a = ci[WS(rs, 13)];
1368
0
       T27 = W[24];
1369
0
       T29 = W[25];
1370
0
       T2b = FMA(T27, T28, T29 * T2a);
1371
0
       T3R = FNMS(T29, T28, T27 * T2a);
1372
0
        }
1373
0
        T3S = T3Q - T3R;
1374
0
        T3T = T26 - T2b;
1375
0
        {
1376
0
       E T21, T2c, T62, T63;
1377
0
       T21 = T1V + T20;
1378
0
       T2c = T26 + T2b;
1379
0
       T2d = T21 + T2c;
1380
0
       T5Z = T21 - T2c;
1381
0
       T62 = T3Q + T3R;
1382
0
       T63 = T3M + T3N;
1383
0
       T64 = T62 - T63;
1384
0
       T6K = T63 + T62;
1385
0
        }
1386
0
        {
1387
0
       E T3P, T3U, T42, T43;
1388
0
       T3P = T3L + T3O;
1389
0
       T3U = T3S - T3T;
1390
0
       T3V = KP707106781 * (T3P - T3U);
1391
0
       T5a = KP707106781 * (T3P + T3U);
1392
0
       T42 = T3T + T3S;
1393
0
       T43 = T3L - T3O;
1394
0
       T44 = KP707106781 * (T42 - T43);
1395
0
       T57 = KP707106781 * (T43 + T42);
1396
0
        }
1397
0
         }
1398
0
         {
1399
0
        E T2G, T4i, T2L, T4j, T4h, T4k, T2R, T4d, T2W, T4e, T4c, T4f;
1400
0
        {
1401
0
       E T2D, T2F, T2C, T2E;
1402
0
       T2D = cr[WS(rs, 3)];
1403
0
       T2F = ci[WS(rs, 3)];
1404
0
       T2C = W[4];
1405
0
       T2E = W[5];
1406
0
       T2G = FMA(T2C, T2D, T2E * T2F);
1407
0
       T4i = FNMS(T2E, T2D, T2C * T2F);
1408
0
        }
1409
0
        {
1410
0
       E T2I, T2K, T2H, T2J;
1411
0
       T2I = cr[WS(rs, 19)];
1412
0
       T2K = ci[WS(rs, 19)];
1413
0
       T2H = W[36];
1414
0
       T2J = W[37];
1415
0
       T2L = FMA(T2H, T2I, T2J * T2K);
1416
0
       T4j = FNMS(T2J, T2I, T2H * T2K);
1417
0
        }
1418
0
        T4h = T2G - T2L;
1419
0
        T4k = T4i - T4j;
1420
0
        {
1421
0
       E T2O, T2Q, T2N, T2P;
1422
0
       T2O = cr[WS(rs, 27)];
1423
0
       T2Q = ci[WS(rs, 27)];
1424
0
       T2N = W[52];
1425
0
       T2P = W[53];
1426
0
       T2R = FMA(T2N, T2O, T2P * T2Q);
1427
0
       T4d = FNMS(T2P, T2O, T2N * T2Q);
1428
0
        }
1429
0
        {
1430
0
       E T2T, T2V, T2S, T2U;
1431
0
       T2T = cr[WS(rs, 11)];
1432
0
       T2V = ci[WS(rs, 11)];
1433
0
       T2S = W[20];
1434
0
       T2U = W[21];
1435
0
       T2W = FMA(T2S, T2T, T2U * T2V);
1436
0
       T4e = FNMS(T2U, T2T, T2S * T2V);
1437
0
        }
1438
0
        T4c = T2R - T2W;
1439
0
        T4f = T4d - T4e;
1440
0
        {
1441
0
       E T2M, T2X, T68, T69;
1442
0
       T2M = T2G + T2L;
1443
0
       T2X = T2R + T2W;
1444
0
       T2Y = T2M + T2X;
1445
0
       T6f = T2M - T2X;
1446
0
       T68 = T4d + T4e;
1447
0
       T69 = T4i + T4j;
1448
0
       T6a = T68 - T69;
1449
0
       T6P = T69 + T68;
1450
0
        }
1451
0
        {
1452
0
       E T4g, T4l, T4t, T4u;
1453
0
       T4g = T4c + T4f;
1454
0
       T4l = T4h - T4k;
1455
0
       T4m = KP707106781 * (T4g - T4l);
1456
0
       T5e = KP707106781 * (T4l + T4g);
1457
0
       T4t = T4h + T4k;
1458
0
       T4u = T4f - T4c;
1459
0
       T4v = KP707106781 * (T4t - T4u);
1460
0
       T5h = KP707106781 * (T4t + T4u);
1461
0
        }
1462
0
         }
1463
0
         {
1464
0
        E T1t, T6X, T7a, T7c, T30, T7b, T70, T71;
1465
0
        {
1466
0
       E TH, T1s, T72, T79;
1467
0
       TH = Tj + TG;
1468
0
       T1s = T14 + T1r;
1469
0
       T1t = TH + T1s;
1470
0
       T6X = TH - T1s;
1471
0
       T72 = T6F + T6E;
1472
0
       T79 = T73 + T78;
1473
0
       T7a = T72 + T79;
1474
0
       T7c = T79 - T72;
1475
0
        }
1476
0
        {
1477
0
       E T2e, T2Z, T6Y, T6Z;
1478
0
       T2e = T1Q + T2d;
1479
0
       T2Z = T2B + T2Y;
1480
0
       T30 = T2e + T2Z;
1481
0
       T7b = T2Z - T2e;
1482
0
       T6Y = T6O + T6P;
1483
0
       T6Z = T6J + T6K;
1484
0
       T70 = T6Y - T6Z;
1485
0
       T71 = T6Z + T6Y;
1486
0
        }
1487
0
        ci[WS(rs, 15)] = T1t - T30;
1488
0
        cr[WS(rs, 24)] = T7b - T7c;
1489
0
        ci[WS(rs, 23)] = T7b + T7c;
1490
0
        cr[0] = T1t + T30;
1491
0
        cr[WS(rs, 8)] = T6X - T70;
1492
0
        cr[WS(rs, 16)] = T71 - T7a;
1493
0
        ci[WS(rs, 31)] = T71 + T7a;
1494
0
        ci[WS(rs, 7)] = T6X + T70;
1495
0
         }
1496
0
         {
1497
0
        E T4X, T5p, T7D, T7J, T54, T7y, T5z, T5D, T5c, T5m, T5s, T7I, T5w, T5C, T5j;
1498
0
        E T5n, T4W, T7z;
1499
0
        T4W = KP707106781 * (T4U + T4V);
1500
0
        T4X = T4T - T4W;
1501
0
        T5p = T4T + T4W;
1502
0
        T7z = KP707106781 * (T3a - T3f);
1503
0
        T7D = T7z + T7C;
1504
0
        T7J = T7C - T7z;
1505
0
        {
1506
0
       E T50, T53, T5x, T5y;
1507
0
       T50 = FMA(KP923879532, T4Y, KP382683432 * T4Z);
1508
0
       T53 = FNMS(KP923879532, T52, KP382683432 * T51);
1509
0
       T54 = T50 + T53;
1510
0
       T7y = T50 - T53;
1511
0
       T5x = T5d + T5e;
1512
0
       T5y = T5g + T5h;
1513
0
       T5z = FNMS(KP980785280, T5y, KP195090322 * T5x);
1514
0
       T5D = FMA(KP980785280, T5x, KP195090322 * T5y);
1515
0
        }
1516
0
        {
1517
0
       E T58, T5b, T5q, T5r;
1518
0
       T58 = T56 - T57;
1519
0
       T5b = T59 - T5a;
1520
0
       T5c = FMA(KP831469612, T58, KP555570233 * T5b);
1521
0
       T5m = FNMS(KP831469612, T5b, KP555570233 * T58);
1522
0
       T5q = FNMS(KP382683432, T4Y, KP923879532 * T4Z);
1523
0
       T5r = FMA(KP382683432, T52, KP923879532 * T51);
1524
0
       T5s = T5q + T5r;
1525
0
       T7I = T5r - T5q;
1526
0
        }
1527
0
        {
1528
0
       E T5u, T5v, T5f, T5i;
1529
0
       T5u = T56 + T57;
1530
0
       T5v = T59 + T5a;
1531
0
       T5w = FMA(KP195090322, T5u, KP980785280 * T5v);
1532
0
       T5C = FNMS(KP195090322, T5v, KP980785280 * T5u);
1533
0
       T5f = T5d - T5e;
1534
0
       T5i = T5g - T5h;
1535
0
       T5j = FNMS(KP555570233, T5i, KP831469612 * T5f);
1536
0
       T5n = FMA(KP555570233, T5f, KP831469612 * T5i);
1537
0
        }
1538
0
        {
1539
0
       E T55, T5k, T7H, T7K;
1540
0
       T55 = T4X + T54;
1541
0
       T5k = T5c + T5j;
1542
0
       ci[WS(rs, 12)] = T55 - T5k;
1543
0
       cr[WS(rs, 3)] = T55 + T5k;
1544
0
       T7H = T5n - T5m;
1545
0
       T7K = T7I + T7J;
1546
0
       cr[WS(rs, 19)] = T7H - T7K;
1547
0
       ci[WS(rs, 28)] = T7H + T7K;
1548
0
        }
1549
0
        {
1550
0
       E T7L, T7M, T5l, T5o;
1551
0
       T7L = T5j - T5c;
1552
0
       T7M = T7J - T7I;
1553
0
       cr[WS(rs, 27)] = T7L - T7M;
1554
0
       ci[WS(rs, 20)] = T7L + T7M;
1555
0
       T5l = T4X - T54;
1556
0
       T5o = T5m + T5n;
1557
0
       cr[WS(rs, 11)] = T5l - T5o;
1558
0
       ci[WS(rs, 4)] = T5l + T5o;
1559
0
        }
1560
0
        {
1561
0
       E T5t, T5A, T7x, T7E;
1562
0
       T5t = T5p - T5s;
1563
0
       T5A = T5w + T5z;
1564
0
       ci[WS(rs, 8)] = T5t - T5A;
1565
0
       cr[WS(rs, 7)] = T5t + T5A;
1566
0
       T7x = T5z - T5w;
1567
0
       T7E = T7y + T7D;
1568
0
       cr[WS(rs, 31)] = T7x - T7E;
1569
0
       ci[WS(rs, 16)] = T7x + T7E;
1570
0
        }
1571
0
        {
1572
0
       E T7F, T7G, T5B, T5E;
1573
0
       T7F = T5D - T5C;
1574
0
       T7G = T7D - T7y;
1575
0
       cr[WS(rs, 23)] = T7F - T7G;
1576
0
       ci[WS(rs, 24)] = T7F + T7G;
1577
0
       T5B = T5p + T5s;
1578
0
       T5E = T5C + T5D;
1579
0
       cr[WS(rs, 15)] = T5B - T5E;
1580
0
       ci[0] = T5B + T5E;
1581
0
        }
1582
0
         }
1583
0
         {
1584
0
        E T6H, T6T, T7g, T7i, T6M, T6U, T6R, T6V;
1585
0
        {
1586
0
       E T6D, T6G, T7e, T7f;
1587
0
       T6D = Tj - TG;
1588
0
       T6G = T6E - T6F;
1589
0
       T6H = T6D - T6G;
1590
0
       T6T = T6D + T6G;
1591
0
       T7e = T14 - T1r;
1592
0
       T7f = T78 - T73;
1593
0
       T7g = T7e + T7f;
1594
0
       T7i = T7f - T7e;
1595
0
        }
1596
0
        {
1597
0
       E T6I, T6L, T6N, T6Q;
1598
0
       T6I = T1Q - T2d;
1599
0
       T6L = T6J - T6K;
1600
0
       T6M = T6I + T6L;
1601
0
       T6U = T6I - T6L;
1602
0
       T6N = T2B - T2Y;
1603
0
       T6Q = T6O - T6P;
1604
0
       T6R = T6N - T6Q;
1605
0
       T6V = T6N + T6Q;
1606
0
        }
1607
0
        {
1608
0
       E T6S, T7h, T6W, T7d;
1609
0
       T6S = KP707106781 * (T6M + T6R);
1610
0
       ci[WS(rs, 11)] = T6H - T6S;
1611
0
       cr[WS(rs, 4)] = T6H + T6S;
1612
0
       T7h = KP707106781 * (T6V - T6U);
1613
0
       cr[WS(rs, 20)] = T7h - T7i;
1614
0
       ci[WS(rs, 27)] = T7h + T7i;
1615
0
       T6W = KP707106781 * (T6U + T6V);
1616
0
       cr[WS(rs, 12)] = T6T - T6W;
1617
0
       ci[WS(rs, 3)] = T6T + T6W;
1618
0
       T7d = KP707106781 * (T6R - T6M);
1619
0
       cr[WS(rs, 28)] = T7d - T7g;
1620
0
       ci[WS(rs, 19)] = T7d + T7g;
1621
0
        }
1622
0
         }
1623
0
         {
1624
0
        E T5J, T7n, T7t, T6n, T5U, T7k, T6x, T6B, T6q, T7s, T66, T6k, T6u, T6A, T6h;
1625
0
        E T6l;
1626
0
        {
1627
0
       E T5O, T5T, T60, T65;
1628
0
       T5J = T5F - T5I;
1629
0
       T7n = T7l + T7m;
1630
0
       T7t = T7m - T7l;
1631
0
       T6n = T5F + T5I;
1632
0
       T5O = T5K + T5N;
1633
0
       T5T = T5P - T5S;
1634
0
       T5U = KP707106781 * (T5O + T5T);
1635
0
       T7k = KP707106781 * (T5O - T5T);
1636
0
       {
1637
0
            E T6v, T6w, T6o, T6p;
1638
0
            T6v = T6e + T6f;
1639
0
            T6w = T67 + T6a;
1640
0
            T6x = FMA(KP382683432, T6v, KP923879532 * T6w);
1641
0
            T6B = FNMS(KP923879532, T6v, KP382683432 * T6w);
1642
0
            T6o = T5K - T5N;
1643
0
            T6p = T5P + T5S;
1644
0
            T6q = KP707106781 * (T6o + T6p);
1645
0
            T7s = KP707106781 * (T6p - T6o);
1646
0
       }
1647
0
       T60 = T5Y - T5Z;
1648
0
       T65 = T61 - T64;
1649
0
       T66 = FMA(KP382683432, T60, KP923879532 * T65);
1650
0
       T6k = FNMS(KP923879532, T60, KP382683432 * T65);
1651
0
       {
1652
0
            E T6s, T6t, T6b, T6g;
1653
0
            T6s = T61 + T64;
1654
0
            T6t = T5Y + T5Z;
1655
0
            T6u = FNMS(KP382683432, T6t, KP923879532 * T6s);
1656
0
            T6A = FMA(KP923879532, T6t, KP382683432 * T6s);
1657
0
            T6b = T67 - T6a;
1658
0
            T6g = T6e - T6f;
1659
0
            T6h = FNMS(KP382683432, T6g, KP923879532 * T6b);
1660
0
            T6l = FMA(KP923879532, T6g, KP382683432 * T6b);
1661
0
       }
1662
0
        }
1663
0
        {
1664
0
       E T5V, T6i, T7r, T7u;
1665
0
       T5V = T5J + T5U;
1666
0
       T6i = T66 + T6h;
1667
0
       ci[WS(rs, 13)] = T5V - T6i;
1668
0
       cr[WS(rs, 2)] = T5V + T6i;
1669
0
       T7r = T6l - T6k;
1670
0
       T7u = T7s + T7t;
1671
0
       cr[WS(rs, 18)] = T7r - T7u;
1672
0
       ci[WS(rs, 29)] = T7r + T7u;
1673
0
        }
1674
0
        {
1675
0
       E T7v, T7w, T6j, T6m;
1676
0
       T7v = T6h - T66;
1677
0
       T7w = T7t - T7s;
1678
0
       cr[WS(rs, 26)] = T7v - T7w;
1679
0
       ci[WS(rs, 21)] = T7v + T7w;
1680
0
       T6j = T5J - T5U;
1681
0
       T6m = T6k + T6l;
1682
0
       cr[WS(rs, 10)] = T6j - T6m;
1683
0
       ci[WS(rs, 5)] = T6j + T6m;
1684
0
        }
1685
0
        {
1686
0
       E T6r, T6y, T7j, T7o;
1687
0
       T6r = T6n + T6q;
1688
0
       T6y = T6u + T6x;
1689
0
       cr[WS(rs, 14)] = T6r - T6y;
1690
0
       ci[WS(rs, 1)] = T6r + T6y;
1691
0
       T7j = T6B - T6A;
1692
0
       T7o = T7k + T7n;
1693
0
       cr[WS(rs, 30)] = T7j - T7o;
1694
0
       ci[WS(rs, 17)] = T7j + T7o;
1695
0
        }
1696
0
        {
1697
0
       E T7p, T7q, T6z, T6C;
1698
0
       T7p = T6x - T6u;
1699
0
       T7q = T7n - T7k;
1700
0
       cr[WS(rs, 22)] = T7p - T7q;
1701
0
       ci[WS(rs, 25)] = T7p + T7q;
1702
0
       T6z = T6n - T6q;
1703
0
       T6C = T6A + T6B;
1704
0
       ci[WS(rs, 9)] = T6z - T6C;
1705
0
       cr[WS(rs, 6)] = T6z + T6C;
1706
0
        }
1707
0
         }
1708
0
         {
1709
0
        E T3h, T4D, T7R, T7X, T3E, T7O, T4N, T4R, T46, T4A, T4G, T7W, T4K, T4Q, T4x;
1710
0
        E T4B, T3g, T7P;
1711
0
        T3g = KP707106781 * (T3a + T3f);
1712
0
        T3h = T35 - T3g;
1713
0
        T4D = T35 + T3g;
1714
0
        T7P = KP707106781 * (T4V - T4U);
1715
0
        T7R = T7P + T7Q;
1716
0
        T7X = T7Q - T7P;
1717
0
        {
1718
0
       E T3s, T3D, T4L, T4M;
1719
0
       T3s = FNMS(KP923879532, T3r, KP382683432 * T3m);
1720
0
       T3D = FMA(KP923879532, T3x, KP382683432 * T3C);
1721
0
       T3E = T3s + T3D;
1722
0
       T7O = T3D - T3s;
1723
0
       T4L = T4s + T4v;
1724
0
       T4M = T4b + T4m;
1725
0
       T4N = FNMS(KP195090322, T4M, KP980785280 * T4L);
1726
0
       T4R = FMA(KP980785280, T4M, KP195090322 * T4L);
1727
0
        }
1728
0
        {
1729
0
       E T3W, T45, T4E, T4F;
1730
0
       T3W = T3K - T3V;
1731
0
       T45 = T41 - T44;
1732
0
       T46 = FNMS(KP555570233, T45, KP831469612 * T3W);
1733
0
       T4A = FMA(KP831469612, T45, KP555570233 * T3W);
1734
0
       T4E = FMA(KP382683432, T3r, KP923879532 * T3m);
1735
0
       T4F = FNMS(KP382683432, T3x, KP923879532 * T3C);
1736
0
       T4G = T4E + T4F;
1737
0
       T7W = T4E - T4F;
1738
0
        }
1739
0
        {
1740
0
       E T4I, T4J, T4n, T4w;
1741
0
       T4I = T41 + T44;
1742
0
       T4J = T3K + T3V;
1743
0
       T4K = FMA(KP195090322, T4I, KP980785280 * T4J);
1744
0
       T4Q = FNMS(KP980785280, T4I, KP195090322 * T4J);
1745
0
       T4n = T4b - T4m;
1746
0
       T4w = T4s - T4v;
1747
0
       T4x = FMA(KP555570233, T4n, KP831469612 * T4w);
1748
0
       T4B = FNMS(KP831469612, T4n, KP555570233 * T4w);
1749
0
        }
1750
0
        {
1751
0
       E T3F, T4y, T7V, T7Y;
1752
0
       T3F = T3h + T3E;
1753
0
       T4y = T46 + T4x;
1754
0
       cr[WS(rs, 13)] = T3F - T4y;
1755
0
       ci[WS(rs, 2)] = T3F + T4y;
1756
0
       T7V = T4B - T4A;
1757
0
       T7Y = T7W + T7X;
1758
0
       cr[WS(rs, 29)] = T7V - T7Y;
1759
0
       ci[WS(rs, 18)] = T7V + T7Y;
1760
0
        }
1761
0
        {
1762
0
       E T7Z, T80, T4z, T4C;
1763
0
       T7Z = T4x - T46;
1764
0
       T80 = T7X - T7W;
1765
0
       cr[WS(rs, 21)] = T7Z - T80;
1766
0
       ci[WS(rs, 26)] = T7Z + T80;
1767
0
       T4z = T3h - T3E;
1768
0
       T4C = T4A + T4B;
1769
0
       ci[WS(rs, 10)] = T4z - T4C;
1770
0
       cr[WS(rs, 5)] = T4z + T4C;
1771
0
        }
1772
0
        {
1773
0
       E T4H, T4O, T7N, T7S;
1774
0
       T4H = T4D + T4G;
1775
0
       T4O = T4K + T4N;
1776
0
       ci[WS(rs, 14)] = T4H - T4O;
1777
0
       cr[WS(rs, 1)] = T4H + T4O;
1778
0
       T7N = T4R - T4Q;
1779
0
       T7S = T7O + T7R;
1780
0
       cr[WS(rs, 17)] = T7N - T7S;
1781
0
       ci[WS(rs, 30)] = T7N + T7S;
1782
0
        }
1783
0
        {
1784
0
       E T7T, T7U, T4P, T4S;
1785
0
       T7T = T4N - T4K;
1786
0
       T7U = T7R - T7O;
1787
0
       cr[WS(rs, 25)] = T7T - T7U;
1788
0
       ci[WS(rs, 22)] = T7T + T7U;
1789
0
       T4P = T4D - T4G;
1790
0
       T4S = T4Q + T4R;
1791
0
       cr[WS(rs, 9)] = T4P - T4S;
1792
0
       ci[WS(rs, 6)] = T4P + T4S;
1793
0
        }
1794
0
         }
1795
0
    }
1796
0
     }
1797
0
}
1798
1799
static const tw_instr twinstr[] = {
1800
     { TW_FULL, 1, 32 },
1801
     { TW_NEXT, 1, 0 }
1802
};
1803
1804
static const hc2hc_desc desc = { 32, "hf_32", twinstr, &GENUS, { 340, 114, 94, 0 } };
1805
1806
1
void X(codelet_hf_32) (planner *p) {
1807
1
     X(khc2hc_register) (p, hf_32, &desc);
1808
1
}
1809
#endif