Coverage Report

Created: 2024-09-08 06:43

/src/fftw3/rdft/scalar/r2cb/hb_15.c
Line
Count
Source (jump to first uncovered line)
1
/*
2
 * Copyright (c) 2003, 2007-14 Matteo Frigo
3
 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
4
 *
5
 * This program is free software; you can redistribute it and/or modify
6
 * it under the terms of the GNU General Public License as published by
7
 * the Free Software Foundation; either version 2 of the License, or
8
 * (at your option) any later version.
9
 *
10
 * This program is distributed in the hope that it will be useful,
11
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13
 * GNU General Public License for more details.
14
 *
15
 * You should have received a copy of the GNU General Public License
16
 * along with this program; if not, write to the Free Software
17
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
18
 *
19
 */
20
21
/* This file was automatically generated --- DO NOT EDIT */
22
/* Generated on Sun Sep  8 06:42:11 UTC 2024 */
23
24
#include "rdft/codelet-rdft.h"
25
26
#if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27
28
/* Generated by: ../../../genfft/gen_hc2hc.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 15 -dif -name hb_15 -include rdft/scalar/hb.h */
29
30
/*
31
 * This function contains 184 FP additions, 140 FP multiplications,
32
 * (or, 72 additions, 28 multiplications, 112 fused multiply/add),
33
 * 78 stack variables, 6 constants, and 60 memory accesses
34
 */
35
#include "rdft/scalar/hb.h"
36
37
static void hb_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
38
{
39
     DK(KP951056516, +0.951056516295153572116439333379382143405698634);
40
     DK(KP559016994, +0.559016994374947424102293417182819058860154590);
41
     DK(KP250000000, +0.250000000000000000000000000000000000000000000);
42
     DK(KP618033988, +0.618033988749894848204586834365638117720309180);
43
     DK(KP866025403, +0.866025403784438646763723170752936183471402627);
44
     DK(KP500000000, +0.500000000000000000000000000000000000000000000);
45
     {
46
    INT m;
47
    for (m = mb, W = W + ((mb - 1) * 28); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
48
         E T5, T11, T1C, T2U, T2f, T3f, TH, T19, T18, TS, T12, T13, T14, T3a, T3g;
49
         E Ts, Tv, T37, T3h, T28, T2h, T21, T2g, T2V, T2W, T2X, T2Y, T2Z, T30, T31;
50
         E T1F, T1I, T1J, T1M, T1P, T1Q, T1R;
51
         {
52
        E T1, TX, T4, T2e, T10, T1B, T1A, T2d;
53
        T1 = cr[0];
54
        TX = ci[WS(rs, 14)];
55
        {
56
       E T2, T3, TY, TZ;
57
       T2 = cr[WS(rs, 5)];
58
       T3 = ci[WS(rs, 4)];
59
       T4 = T2 + T3;
60
       T2e = T2 - T3;
61
       TY = ci[WS(rs, 9)];
62
       TZ = cr[WS(rs, 10)];
63
       T10 = TY - TZ;
64
       T1B = TY + TZ;
65
        }
66
        T5 = T1 + T4;
67
        T11 = TX + T10;
68
        T1A = FNMS(KP500000000, T4, T1);
69
        T1C = FNMS(KP866025403, T1B, T1A);
70
        T2U = FMA(KP866025403, T1B, T1A);
71
        T2d = FNMS(KP500000000, T10, TX);
72
        T2f = FMA(KP866025403, T2e, T2d);
73
        T3f = FNMS(KP866025403, T2e, T2d);
74
         }
75
         {
76
        E Ta, T1W, T1D, Tl, T23, T1K, Tf, T1Z, T1G, TR, T1Y, T1H, Tq, T26, T1N;
77
        E TG, T25, T1O, TM, T1V, T1E, TB, T22, T1L, T38, T39;
78
        {
79
       E T6, T7, T8, T9;
80
       T6 = cr[WS(rs, 3)];
81
       T7 = ci[WS(rs, 6)];
82
       T8 = ci[WS(rs, 1)];
83
       T9 = T7 + T8;
84
       Ta = T6 + T9;
85
       T1W = T7 - T8;
86
       T1D = FNMS(KP500000000, T9, T6);
87
        }
88
        {
89
       E Th, Ti, Tj, Tk;
90
       Th = cr[WS(rs, 6)];
91
       Ti = ci[WS(rs, 3)];
92
       Tj = cr[WS(rs, 1)];
93
       Tk = Ti + Tj;
94
       Tl = Th + Tk;
95
       T23 = Ti - Tj;
96
       T1K = FNMS(KP500000000, Tk, Th);
97
        }
98
        {
99
       E Tb, Tc, Td, Te;
100
       Tb = ci[WS(rs, 2)];
101
       Tc = cr[WS(rs, 2)];
102
       Td = cr[WS(rs, 7)];
103
       Te = Tc + Td;
104
       Tf = Tb + Te;
105
       T1Z = Tc - Td;
106
       T1G = FNMS(KP500000000, Te, Tb);
107
        }
108
        {
109
       E TQ, TN, TO, TP;
110
       TQ = cr[WS(rs, 12)];
111
       TN = ci[WS(rs, 12)];
112
       TO = ci[WS(rs, 7)];
113
       TP = TN + TO;
114
       TR = TP - TQ;
115
       T1Y = FMA(KP500000000, TP, TQ);
116
       T1H = TO - TN;
117
        }
118
        {
119
       E Tm, Tn, To, Tp;
120
       Tm = ci[WS(rs, 5)];
121
       Tn = ci[0];
122
       To = cr[WS(rs, 4)];
123
       Tp = Tn + To;
124
       Tq = Tm + Tp;
125
       T26 = Tn - To;
126
       T1N = FNMS(KP500000000, Tp, Tm);
127
        }
128
        {
129
       E TF, TC, TD, TE;
130
       TF = cr[WS(rs, 9)];
131
       TC = ci[WS(rs, 10)];
132
       TD = cr[WS(rs, 14)];
133
       TE = TC - TD;
134
       TG = TE - TF;
135
       T25 = FMA(KP500000000, TE, TF);
136
       T1O = TC + TD;
137
        }
138
        {
139
       E TI, TJ, TK, TL;
140
       TI = ci[WS(rs, 11)];
141
       TJ = cr[WS(rs, 8)];
142
       TK = cr[WS(rs, 13)];
143
       TL = TJ + TK;
144
       TM = TI - TL;
145
       T1V = FMA(KP500000000, TL, TI);
146
       T1E = TJ - TK;
147
        }
148
        {
149
       E Tx, Ty, Tz, TA;
150
       Tx = ci[WS(rs, 8)];
151
       Ty = ci[WS(rs, 13)];
152
       Tz = cr[WS(rs, 11)];
153
       TA = Ty - Tz;
154
       TB = Tx + TA;
155
       T22 = FNMS(KP500000000, TA, Tx);
156
       T1L = Ty + Tz;
157
        }
158
        TH = TB - TG;
159
        T19 = Ta - Tf;
160
        T18 = Tl - Tq;
161
        TS = TM - TR;
162
        T12 = TM + TR;
163
        T13 = TB + TG;
164
        T14 = T12 + T13;
165
        T38 = FNMS(KP866025403, T1W, T1V);
166
        T39 = FMA(KP866025403, T1Z, T1Y);
167
        T3a = T38 + T39;
168
        T3g = T38 - T39;
169
        {
170
       E Tg, Tr, T1X, T20;
171
       Tg = Ta + Tf;
172
       Tr = Tl + Tq;
173
       Ts = Tg + Tr;
174
       Tv = Tg - Tr;
175
       {
176
            E T35, T36, T24, T27;
177
            T35 = FNMS(KP866025403, T23, T22);
178
            T36 = FMA(KP866025403, T26, T25);
179
            T37 = T35 + T36;
180
            T3h = T35 - T36;
181
            T24 = FMA(KP866025403, T23, T22);
182
            T27 = FNMS(KP866025403, T26, T25);
183
            T28 = T24 + T27;
184
            T2h = T24 - T27;
185
       }
186
       T1X = FMA(KP866025403, T1W, T1V);
187
       T20 = FNMS(KP866025403, T1Z, T1Y);
188
       T21 = T1X + T20;
189
       T2g = T1X - T20;
190
       T2V = FNMS(KP866025403, T1E, T1D);
191
       T2W = FNMS(KP866025403, T1H, T1G);
192
       T2X = T2V + T2W;
193
       T2Y = FNMS(KP866025403, T1L, T1K);
194
       T2Z = FNMS(KP866025403, T1O, T1N);
195
       T30 = T2Y + T2Z;
196
       T31 = T2X + T30;
197
       T1F = FMA(KP866025403, T1E, T1D);
198
       T1I = FMA(KP866025403, T1H, T1G);
199
       T1J = T1F + T1I;
200
       T1M = FMA(KP866025403, T1L, T1K);
201
       T1P = FMA(KP866025403, T1O, T1N);
202
       T1Q = T1M + T1P;
203
       T1R = T1J + T1Q;
204
        }
205
         }
206
         cr[0] = T5 + Ts;
207
         ci[0] = T11 + T14;
208
         {
209
        E T1a, T1q, T17, T1p, TU, T1u, T1e, T1m, T15, T16;
210
        T1a = FNMS(KP618033988, T19, T18);
211
        T1q = FMA(KP618033988, T18, T19);
212
        T15 = FNMS(KP250000000, T14, T11);
213
        T16 = T12 - T13;
214
        T17 = FNMS(KP559016994, T16, T15);
215
        T1p = FMA(KP559016994, T16, T15);
216
        {
217
       E TT, T1l, Tw, T1k, Tu;
218
       TT = FNMS(KP618033988, TS, TH);
219
       T1l = FMA(KP618033988, TH, TS);
220
       Tu = FNMS(KP250000000, Ts, T5);
221
       Tw = FNMS(KP559016994, Tv, Tu);
222
       T1k = FMA(KP559016994, Tv, Tu);
223
       TU = FNMS(KP951056516, TT, Tw);
224
       T1u = FMA(KP951056516, T1l, T1k);
225
       T1e = FMA(KP951056516, TT, Tw);
226
       T1m = FNMS(KP951056516, T1l, T1k);
227
        }
228
        {
229
       E T1b, TW, T1c, Tt, TV;
230
       T1b = FMA(KP951056516, T1a, T17);
231
       TW = W[5];
232
       T1c = TW * TU;
233
       Tt = W[4];
234
       TV = Tt * TU;
235
       cr[WS(rs, 3)] = FNMS(TW, T1b, TV);
236
       ci[WS(rs, 3)] = FMA(Tt, T1b, T1c);
237
        }
238
        {
239
       E T1x, T1w, T1y, T1t, T1v;
240
       T1x = FNMS(KP951056516, T1q, T1p);
241
       T1w = W[17];
242
       T1y = T1w * T1u;
243
       T1t = W[16];
244
       T1v = T1t * T1u;
245
       cr[WS(rs, 9)] = FNMS(T1w, T1x, T1v);
246
       ci[WS(rs, 9)] = FMA(T1t, T1x, T1y);
247
        }
248
        {
249
       E T1h, T1g, T1i, T1d, T1f;
250
       T1h = FNMS(KP951056516, T1a, T17);
251
       T1g = W[23];
252
       T1i = T1g * T1e;
253
       T1d = W[22];
254
       T1f = T1d * T1e;
255
       cr[WS(rs, 12)] = FNMS(T1g, T1h, T1f);
256
       ci[WS(rs, 12)] = FMA(T1d, T1h, T1i);
257
        }
258
        {
259
       E T1r, T1o, T1s, T1j, T1n;
260
       T1r = FMA(KP951056516, T1q, T1p);
261
       T1o = W[11];
262
       T1s = T1o * T1m;
263
       T1j = W[10];
264
       T1n = T1j * T1m;
265
       cr[WS(rs, 6)] = FNMS(T1o, T1r, T1n);
266
       ci[WS(rs, 6)] = FMA(T1j, T1r, T1s);
267
        }
268
         }
269
         {
270
        E T2o, T2E, T2N, T2P, T2Q, T2S, T2l, T2R, T2D, T2a, T2I, T2s, T2A;
271
        {
272
       E T2m, T2n, T2O, T2k, T2i, T2j;
273
       T2m = T1F - T1I;
274
       T2n = T1M - T1P;
275
       T2o = FMA(KP618033988, T2n, T2m);
276
       T2E = FNMS(KP618033988, T2m, T2n);
277
       T2O = T1C + T1R;
278
       T2N = W[18];
279
       T2P = T2N * T2O;
280
       T2Q = W[19];
281
       T2S = T2Q * T2O;
282
       T2k = T2g - T2h;
283
       T2i = T2g + T2h;
284
       T2j = FNMS(KP250000000, T2i, T2f);
285
       T2l = FMA(KP559016994, T2k, T2j);
286
       T2R = T2f + T2i;
287
       T2D = FNMS(KP559016994, T2k, T2j);
288
       {
289
            E T29, T2z, T1U, T2y, T1S, T1T;
290
            T29 = FMA(KP618033988, T28, T21);
291
            T2z = FNMS(KP618033988, T21, T28);
292
            T1S = FNMS(KP250000000, T1R, T1C);
293
            T1T = T1J - T1Q;
294
            T1U = FMA(KP559016994, T1T, T1S);
295
            T2y = FNMS(KP559016994, T1T, T1S);
296
            T2a = FNMS(KP951056516, T29, T1U);
297
            T2I = FNMS(KP951056516, T2z, T2y);
298
            T2s = FMA(KP951056516, T29, T1U);
299
            T2A = FMA(KP951056516, T2z, T2y);
300
       }
301
        }
302
        cr[WS(rs, 10)] = FNMS(T2Q, T2R, T2P);
303
        ci[WS(rs, 10)] = FMA(T2N, T2R, T2S);
304
        {
305
       E T2p, T2c, T2q, T1z, T2b;
306
       T2p = FMA(KP951056516, T2o, T2l);
307
       T2c = W[1];
308
       T2q = T2c * T2a;
309
       T1z = W[0];
310
       T2b = T1z * T2a;
311
       cr[WS(rs, 1)] = FNMS(T2c, T2p, T2b);
312
       ci[WS(rs, 1)] = FMA(T1z, T2p, T2q);
313
        }
314
        {
315
       E T2L, T2K, T2M, T2H, T2J;
316
       T2L = FMA(KP951056516, T2E, T2D);
317
       T2K = W[25];
318
       T2M = T2K * T2I;
319
       T2H = W[24];
320
       T2J = T2H * T2I;
321
       cr[WS(rs, 13)] = FNMS(T2K, T2L, T2J);
322
       ci[WS(rs, 13)] = FMA(T2H, T2L, T2M);
323
        }
324
        {
325
       E T2F, T2C, T2G, T2x, T2B;
326
       T2F = FNMS(KP951056516, T2E, T2D);
327
       T2C = W[13];
328
       T2G = T2C * T2A;
329
       T2x = W[12];
330
       T2B = T2x * T2A;
331
       cr[WS(rs, 7)] = FNMS(T2C, T2F, T2B);
332
       ci[WS(rs, 7)] = FMA(T2x, T2F, T2G);
333
        }
334
        {
335
       E T2v, T2u, T2w, T2r, T2t;
336
       T2v = FNMS(KP951056516, T2o, T2l);
337
       T2u = W[7];
338
       T2w = T2u * T2s;
339
       T2r = W[6];
340
       T2t = T2r * T2s;
341
       cr[WS(rs, 4)] = FNMS(T2u, T2v, T2t);
342
       ci[WS(rs, 4)] = FMA(T2r, T2v, T2w);
343
        }
344
         }
345
         {
346
        E T3o, T3E, T3N, T3P, T3Q, T3S, T3l, T3R, T3D, T3c, T3I, T3s, T3A;
347
        {
348
       E T3m, T3n, T3O, T3k, T3i, T3j;
349
       T3m = T2Y - T2Z;
350
       T3n = T2V - T2W;
351
       T3o = FNMS(KP618033988, T3n, T3m);
352
       T3E = FMA(KP618033988, T3m, T3n);
353
       T3O = T2U + T31;
354
       T3N = W[8];
355
       T3P = T3N * T3O;
356
       T3Q = W[9];
357
       T3S = T3Q * T3O;
358
       T3k = T3g - T3h;
359
       T3i = T3g + T3h;
360
       T3j = FNMS(KP250000000, T3i, T3f);
361
       T3l = FNMS(KP559016994, T3k, T3j);
362
       T3R = T3f + T3i;
363
       T3D = FMA(KP559016994, T3k, T3j);
364
       {
365
            E T3b, T3z, T34, T3y, T32, T33;
366
            T3b = FNMS(KP618033988, T3a, T37);
367
            T3z = FMA(KP618033988, T37, T3a);
368
            T32 = FNMS(KP250000000, T31, T2U);
369
            T33 = T2X - T30;
370
            T34 = FNMS(KP559016994, T33, T32);
371
            T3y = FMA(KP559016994, T33, T32);
372
            T3c = FMA(KP951056516, T3b, T34);
373
            T3I = FMA(KP951056516, T3z, T3y);
374
            T3s = FNMS(KP951056516, T3b, T34);
375
            T3A = FNMS(KP951056516, T3z, T3y);
376
       }
377
        }
378
        cr[WS(rs, 5)] = FNMS(T3Q, T3R, T3P);
379
        ci[WS(rs, 5)] = FMA(T3N, T3R, T3S);
380
        {
381
       E T3p, T3e, T3q, T2T, T3d;
382
       T3p = FNMS(KP951056516, T3o, T3l);
383
       T3e = W[3];
384
       T3q = T3e * T3c;
385
       T2T = W[2];
386
       T3d = T2T * T3c;
387
       cr[WS(rs, 2)] = FNMS(T3e, T3p, T3d);
388
       ci[WS(rs, 2)] = FMA(T2T, T3p, T3q);
389
        }
390
        {
391
       E T3L, T3K, T3M, T3H, T3J;
392
       T3L = FNMS(KP951056516, T3E, T3D);
393
       T3K = W[27];
394
       T3M = T3K * T3I;
395
       T3H = W[26];
396
       T3J = T3H * T3I;
397
       cr[WS(rs, 14)] = FNMS(T3K, T3L, T3J);
398
       ci[WS(rs, 14)] = FMA(T3H, T3L, T3M);
399
        }
400
        {
401
       E T3F, T3C, T3G, T3x, T3B;
402
       T3F = FMA(KP951056516, T3E, T3D);
403
       T3C = W[21];
404
       T3G = T3C * T3A;
405
       T3x = W[20];
406
       T3B = T3x * T3A;
407
       cr[WS(rs, 11)] = FNMS(T3C, T3F, T3B);
408
       ci[WS(rs, 11)] = FMA(T3x, T3F, T3G);
409
        }
410
        {
411
       E T3v, T3u, T3w, T3r, T3t;
412
       T3v = FMA(KP951056516, T3o, T3l);
413
       T3u = W[15];
414
       T3w = T3u * T3s;
415
       T3r = W[14];
416
       T3t = T3r * T3s;
417
       cr[WS(rs, 8)] = FNMS(T3u, T3v, T3t);
418
       ci[WS(rs, 8)] = FMA(T3r, T3v, T3w);
419
        }
420
         }
421
    }
422
     }
423
}
424
425
static const tw_instr twinstr[] = {
426
     { TW_FULL, 1, 15 },
427
     { TW_NEXT, 1, 0 }
428
};
429
430
static const hc2hc_desc desc = { 15, "hb_15", twinstr, &GENUS, { 72, 28, 112, 0 } };
431
432
void X(codelet_hb_15) (planner *p) {
433
     X(khc2hc_register) (p, hb_15, &desc);
434
}
435
#else
436
437
/* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 15 -dif -name hb_15 -include rdft/scalar/hb.h */
438
439
/*
440
 * This function contains 184 FP additions, 112 FP multiplications,
441
 * (or, 128 additions, 56 multiplications, 56 fused multiply/add),
442
 * 75 stack variables, 6 constants, and 60 memory accesses
443
 */
444
#include "rdft/scalar/hb.h"
445
446
static void hb_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
447
0
{
448
0
     DK(KP559016994, +0.559016994374947424102293417182819058860154590);
449
0
     DK(KP250000000, +0.250000000000000000000000000000000000000000000);
450
0
     DK(KP951056516, +0.951056516295153572116439333379382143405698634);
451
0
     DK(KP587785252, +0.587785252292473129168705954639072768597652438);
452
0
     DK(KP500000000, +0.500000000000000000000000000000000000000000000);
453
0
     DK(KP866025403, +0.866025403784438646763723170752936183471402627);
454
0
     {
455
0
    INT m;
456
0
    for (m = mb, W = W + ((mb - 1) * 28); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
457
0
         E T5, T10, T1J, T2C, T2c, T2M, TH, T18, T17, TS, T2Q, T2R, T2S, Tg, Tr;
458
0
         E Ts, T11, T12, T13, T2N, T2O, T2P, T1u, T1x, T1y, T1W, T1Z, T28, T1P, T1S;
459
0
         E T27, T1B, T1E, T1F, T2G, T2H, T2I, T2D, T2E, T2F;
460
0
         {
461
0
        E T1, TW, T4, T2a, TZ, T1I, T1H, T2b;
462
0
        T1 = cr[0];
463
0
        TW = ci[WS(rs, 14)];
464
0
        {
465
0
       E T2, T3, TX, TY;
466
0
       T2 = cr[WS(rs, 5)];
467
0
       T3 = ci[WS(rs, 4)];
468
0
       T4 = T2 + T3;
469
0
       T2a = KP866025403 * (T2 - T3);
470
0
       TX = ci[WS(rs, 9)];
471
0
       TY = cr[WS(rs, 10)];
472
0
       TZ = TX - TY;
473
0
       T1I = KP866025403 * (TX + TY);
474
0
        }
475
0
        T5 = T1 + T4;
476
0
        T10 = TW + TZ;
477
0
        T1H = FNMS(KP500000000, T4, T1);
478
0
        T1J = T1H - T1I;
479
0
        T2C = T1H + T1I;
480
0
        T2b = FNMS(KP500000000, TZ, TW);
481
0
        T2c = T2a + T2b;
482
0
        T2M = T2b - T2a;
483
0
         }
484
0
         {
485
0
        E Ta, T1N, T1s, Tl, T1U, T1z, Tf, T1Q, T1v, TG, T1R, T1w, Tq, T1X, T1C;
486
0
        E TM, T1V, T1A, TB, T1O, T1t, TR, T1Y, T1D;
487
0
        {
488
0
       E T6, T7, T8, T9;
489
0
       T6 = cr[WS(rs, 3)];
490
0
       T7 = ci[WS(rs, 6)];
491
0
       T8 = ci[WS(rs, 1)];
492
0
       T9 = T7 + T8;
493
0
       Ta = T6 + T9;
494
0
       T1N = KP866025403 * (T7 - T8);
495
0
       T1s = FNMS(KP500000000, T9, T6);
496
0
        }
497
0
        {
498
0
       E Th, Ti, Tj, Tk;
499
0
       Th = cr[WS(rs, 6)];
500
0
       Ti = ci[WS(rs, 3)];
501
0
       Tj = cr[WS(rs, 1)];
502
0
       Tk = Ti + Tj;
503
0
       Tl = Th + Tk;
504
0
       T1U = KP866025403 * (Ti - Tj);
505
0
       T1z = FNMS(KP500000000, Tk, Th);
506
0
        }
507
0
        {
508
0
       E Tb, Tc, Td, Te;
509
0
       Tb = ci[WS(rs, 2)];
510
0
       Tc = cr[WS(rs, 2)];
511
0
       Td = cr[WS(rs, 7)];
512
0
       Te = Tc + Td;
513
0
       Tf = Tb + Te;
514
0
       T1Q = KP866025403 * (Tc - Td);
515
0
       T1v = FNMS(KP500000000, Te, Tb);
516
0
        }
517
0
        {
518
0
       E TF, TC, TD, TE;
519
0
       TF = cr[WS(rs, 12)];
520
0
       TC = ci[WS(rs, 12)];
521
0
       TD = ci[WS(rs, 7)];
522
0
       TE = TC + TD;
523
0
       TG = TE - TF;
524
0
       T1R = FMA(KP500000000, TE, TF);
525
0
       T1w = KP866025403 * (TD - TC);
526
0
        }
527
0
        {
528
0
       E Tm, Tn, To, Tp;
529
0
       Tm = ci[WS(rs, 5)];
530
0
       Tn = ci[0];
531
0
       To = cr[WS(rs, 4)];
532
0
       Tp = Tn + To;
533
0
       Tq = Tm + Tp;
534
0
       T1X = KP866025403 * (Tn - To);
535
0
       T1C = FNMS(KP500000000, Tp, Tm);
536
0
        }
537
0
        {
538
0
       E TI, TJ, TK, TL;
539
0
       TI = ci[WS(rs, 8)];
540
0
       TJ = ci[WS(rs, 13)];
541
0
       TK = cr[WS(rs, 11)];
542
0
       TL = TJ - TK;
543
0
       TM = TI + TL;
544
0
       T1V = FNMS(KP500000000, TL, TI);
545
0
       T1A = KP866025403 * (TJ + TK);
546
0
        }
547
0
        {
548
0
       E Tx, Ty, Tz, TA;
549
0
       Tx = ci[WS(rs, 11)];
550
0
       Ty = cr[WS(rs, 8)];
551
0
       Tz = cr[WS(rs, 13)];
552
0
       TA = Ty + Tz;
553
0
       TB = Tx - TA;
554
0
       T1O = FMA(KP500000000, TA, Tx);
555
0
       T1t = KP866025403 * (Ty - Tz);
556
0
        }
557
0
        {
558
0
       E TQ, TN, TO, TP;
559
0
       TQ = cr[WS(rs, 9)];
560
0
       TN = ci[WS(rs, 10)];
561
0
       TO = cr[WS(rs, 14)];
562
0
       TP = TN - TO;
563
0
       TR = TP - TQ;
564
0
       T1Y = FMA(KP500000000, TP, TQ);
565
0
       T1D = KP866025403 * (TN + TO);
566
0
        }
567
0
        TH = TB - TG;
568
0
        T18 = Tl - Tq;
569
0
        T17 = Ta - Tf;
570
0
        TS = TM - TR;
571
0
        T2Q = T1V - T1U;
572
0
        T2R = T1X + T1Y;
573
0
        T2S = T2Q - T2R;
574
0
        Tg = Ta + Tf;
575
0
        Tr = Tl + Tq;
576
0
        Ts = Tg + Tr;
577
0
        T11 = TB + TG;
578
0
        T12 = TM + TR;
579
0
        T13 = T11 + T12;
580
0
        T2N = T1O - T1N;
581
0
        T2O = T1Q + T1R;
582
0
        T2P = T2N - T2O;
583
0
        T1u = T1s + T1t;
584
0
        T1x = T1v + T1w;
585
0
        T1y = T1u + T1x;
586
0
        T1W = T1U + T1V;
587
0
        T1Z = T1X - T1Y;
588
0
        T28 = T1W + T1Z;
589
0
        T1P = T1N + T1O;
590
0
        T1S = T1Q - T1R;
591
0
        T27 = T1P + T1S;
592
0
        T1B = T1z + T1A;
593
0
        T1E = T1C + T1D;
594
0
        T1F = T1B + T1E;
595
0
        T2G = T1z - T1A;
596
0
        T2H = T1C - T1D;
597
0
        T2I = T2G + T2H;
598
0
        T2D = T1s - T1t;
599
0
        T2E = T1v - T1w;
600
0
        T2F = T2D + T2E;
601
0
         }
602
0
         cr[0] = T5 + Ts;
603
0
         ci[0] = T10 + T13;
604
0
         {
605
0
        E TT, T19, T1k, T1h, T16, T1l, Tw, T1g;
606
0
        TT = FNMS(KP951056516, TS, KP587785252 * TH);
607
0
        T19 = FNMS(KP951056516, T18, KP587785252 * T17);
608
0
        T1k = FMA(KP951056516, T17, KP587785252 * T18);
609
0
        T1h = FMA(KP951056516, TH, KP587785252 * TS);
610
0
        {
611
0
       E T14, T15, Tu, Tv;
612
0
       T14 = FNMS(KP250000000, T13, T10);
613
0
       T15 = KP559016994 * (T11 - T12);
614
0
       T16 = T14 - T15;
615
0
       T1l = T15 + T14;
616
0
       Tu = FNMS(KP250000000, Ts, T5);
617
0
       Tv = KP559016994 * (Tg - Tr);
618
0
       Tw = Tu - Tv;
619
0
       T1g = Tv + Tu;
620
0
        }
621
0
        {
622
0
       E TU, T1a, Tt, TV;
623
0
       TU = Tw + TT;
624
0
       T1a = T16 - T19;
625
0
       Tt = W[4];
626
0
       TV = W[5];
627
0
       cr[WS(rs, 3)] = FNMS(TV, T1a, Tt * TU);
628
0
       ci[WS(rs, 3)] = FMA(TV, TU, Tt * T1a);
629
0
        }
630
0
        {
631
0
       E T1o, T1q, T1n, T1p;
632
0
       T1o = T1g + T1h;
633
0
       T1q = T1l - T1k;
634
0
       T1n = W[16];
635
0
       T1p = W[17];
636
0
       cr[WS(rs, 9)] = FNMS(T1p, T1q, T1n * T1o);
637
0
       ci[WS(rs, 9)] = FMA(T1p, T1o, T1n * T1q);
638
0
        }
639
0
        {
640
0
       E T1c, T1e, T1b, T1d;
641
0
       T1c = Tw - TT;
642
0
       T1e = T19 + T16;
643
0
       T1b = W[22];
644
0
       T1d = W[23];
645
0
       cr[WS(rs, 12)] = FNMS(T1d, T1e, T1b * T1c);
646
0
       ci[WS(rs, 12)] = FMA(T1d, T1c, T1b * T1e);
647
0
        }
648
0
        {
649
0
       E T1i, T1m, T1f, T1j;
650
0
       T1i = T1g - T1h;
651
0
       T1m = T1k + T1l;
652
0
       T1f = W[10];
653
0
       T1j = W[11];
654
0
       cr[WS(rs, 6)] = FNMS(T1j, T1m, T1f * T1i);
655
0
       ci[WS(rs, 6)] = FMA(T1j, T1i, T1f * T1m);
656
0
        }
657
0
         }
658
0
         {
659
0
        E T21, T2n, T26, T2q, T1M, T2y, T2m, T2f, T2A, T2r, T2x, T2z;
660
0
        {
661
0
       E T1T, T20, T24, T25;
662
0
       T1T = T1P - T1S;
663
0
       T20 = T1W - T1Z;
664
0
       T21 = FMA(KP951056516, T1T, KP587785252 * T20);
665
0
       T2n = FNMS(KP951056516, T20, KP587785252 * T1T);
666
0
       T24 = T1u - T1x;
667
0
       T25 = T1B - T1E;
668
0
       T26 = FMA(KP951056516, T24, KP587785252 * T25);
669
0
       T2q = FNMS(KP951056516, T25, KP587785252 * T24);
670
0
        }
671
0
        {
672
0
       E T1G, T1K, T1L, T29, T2d, T2e;
673
0
       T1G = KP559016994 * (T1y - T1F);
674
0
       T1K = T1y + T1F;
675
0
       T1L = FNMS(KP250000000, T1K, T1J);
676
0
       T1M = T1G + T1L;
677
0
       T2y = T1J + T1K;
678
0
       T2m = T1L - T1G;
679
0
       T29 = KP559016994 * (T27 - T28);
680
0
       T2d = T27 + T28;
681
0
       T2e = FNMS(KP250000000, T2d, T2c);
682
0
       T2f = T29 + T2e;
683
0
       T2A = T2c + T2d;
684
0
       T2r = T2e - T29;
685
0
        }
686
0
        T2x = W[18];
687
0
        T2z = W[19];
688
0
        cr[WS(rs, 10)] = FNMS(T2z, T2A, T2x * T2y);
689
0
        ci[WS(rs, 10)] = FMA(T2z, T2y, T2x * T2A);
690
0
        {
691
0
       E T2u, T2w, T2t, T2v;
692
0
       T2u = T2m + T2n;
693
0
       T2w = T2r - T2q;
694
0
       T2t = W[24];
695
0
       T2v = W[25];
696
0
       cr[WS(rs, 13)] = FNMS(T2v, T2w, T2t * T2u);
697
0
       ci[WS(rs, 13)] = FMA(T2v, T2u, T2t * T2w);
698
0
        }
699
0
        {
700
0
       E T22, T2g, T1r, T23;
701
0
       T22 = T1M - T21;
702
0
       T2g = T26 + T2f;
703
0
       T1r = W[0];
704
0
       T23 = W[1];
705
0
       cr[WS(rs, 1)] = FNMS(T23, T2g, T1r * T22);
706
0
       ci[WS(rs, 1)] = FMA(T23, T22, T1r * T2g);
707
0
        }
708
0
        {
709
0
       E T2i, T2k, T2h, T2j;
710
0
       T2i = T1M + T21;
711
0
       T2k = T2f - T26;
712
0
       T2h = W[6];
713
0
       T2j = W[7];
714
0
       cr[WS(rs, 4)] = FNMS(T2j, T2k, T2h * T2i);
715
0
       ci[WS(rs, 4)] = FMA(T2j, T2i, T2h * T2k);
716
0
        }
717
0
        {
718
0
       E T2o, T2s, T2l, T2p;
719
0
       T2o = T2m - T2n;
720
0
       T2s = T2q + T2r;
721
0
       T2l = W[12];
722
0
       T2p = W[13];
723
0
       cr[WS(rs, 7)] = FNMS(T2p, T2s, T2l * T2o);
724
0
       ci[WS(rs, 7)] = FMA(T2p, T2o, T2l * T2s);
725
0
        }
726
0
         }
727
0
         {
728
0
        E T31, T3h, T36, T3k, T2K, T3g, T2Y, T2U, T3l, T39, T2B, T2L;
729
0
        {
730
0
       E T2Z, T30, T34, T35;
731
0
       T2Z = T2N + T2O;
732
0
       T30 = T2Q + T2R;
733
0
       T31 = FNMS(KP951056516, T30, KP587785252 * T2Z);
734
0
       T3h = FMA(KP951056516, T2Z, KP587785252 * T30);
735
0
       T34 = T2D - T2E;
736
0
       T35 = T2G - T2H;
737
0
       T36 = FNMS(KP951056516, T35, KP587785252 * T34);
738
0
       T3k = FMA(KP951056516, T34, KP587785252 * T35);
739
0
        }
740
0
        {
741
0
       E T2X, T2J, T2W, T38, T2T, T37;
742
0
       T2X = KP559016994 * (T2F - T2I);
743
0
       T2J = T2F + T2I;
744
0
       T2W = FNMS(KP250000000, T2J, T2C);
745
0
       T2K = T2C + T2J;
746
0
       T3g = T2X + T2W;
747
0
       T2Y = T2W - T2X;
748
0
       T38 = KP559016994 * (T2P - T2S);
749
0
       T2T = T2P + T2S;
750
0
       T37 = FNMS(KP250000000, T2T, T2M);
751
0
       T2U = T2M + T2T;
752
0
       T3l = T38 + T37;
753
0
       T39 = T37 - T38;
754
0
        }
755
0
        T2B = W[8];
756
0
        T2L = W[9];
757
0
        cr[WS(rs, 5)] = FNMS(T2L, T2U, T2B * T2K);
758
0
        ci[WS(rs, 5)] = FMA(T2L, T2K, T2B * T2U);
759
0
        {
760
0
       E T3o, T3q, T3n, T3p;
761
0
       T3o = T3g + T3h;
762
0
       T3q = T3l - T3k;
763
0
       T3n = W[26];
764
0
       T3p = W[27];
765
0
       cr[WS(rs, 14)] = FNMS(T3p, T3q, T3n * T3o);
766
0
       ci[WS(rs, 14)] = FMA(T3n, T3q, T3p * T3o);
767
0
        }
768
0
        {
769
0
       E T32, T3a, T2V, T33;
770
0
       T32 = T2Y - T31;
771
0
       T3a = T36 + T39;
772
0
       T2V = W[2];
773
0
       T33 = W[3];
774
0
       cr[WS(rs, 2)] = FNMS(T33, T3a, T2V * T32);
775
0
       ci[WS(rs, 2)] = FMA(T2V, T3a, T33 * T32);
776
0
        }
777
0
        {
778
0
       E T3c, T3e, T3b, T3d;
779
0
       T3c = T2Y + T31;
780
0
       T3e = T39 - T36;
781
0
       T3b = W[14];
782
0
       T3d = W[15];
783
0
       cr[WS(rs, 8)] = FNMS(T3d, T3e, T3b * T3c);
784
0
       ci[WS(rs, 8)] = FMA(T3b, T3e, T3d * T3c);
785
0
        }
786
0
        {
787
0
       E T3i, T3m, T3f, T3j;
788
0
       T3i = T3g - T3h;
789
0
       T3m = T3k + T3l;
790
0
       T3f = W[20];
791
0
       T3j = W[21];
792
0
       cr[WS(rs, 11)] = FNMS(T3j, T3m, T3f * T3i);
793
0
       ci[WS(rs, 11)] = FMA(T3f, T3m, T3j * T3i);
794
0
        }
795
0
         }
796
0
    }
797
0
     }
798
0
}
799
800
static const tw_instr twinstr[] = {
801
     { TW_FULL, 1, 15 },
802
     { TW_NEXT, 1, 0 }
803
};
804
805
static const hc2hc_desc desc = { 15, "hb_15", twinstr, &GENUS, { 128, 56, 56, 0 } };
806
807
1
void X(codelet_hb_15) (planner *p) {
808
1
     X(khc2hc_register) (p, hb_15, &desc);
809
1
}
810
#endif