Coverage Report

Created: 2025-08-26 06:35

/src/fftw3/dft/scalar/codelets/q1_5.c
Line
Count
Source
1
/*
2
 * Copyright (c) 2003, 2007-14 Matteo Frigo
3
 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
4
 *
5
 * This program is free software; you can redistribute it and/or modify
6
 * it under the terms of the GNU General Public License as published by
7
 * the Free Software Foundation; either version 2 of the License, or
8
 * (at your option) any later version.
9
 *
10
 * This program is distributed in the hope that it will be useful,
11
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13
 * GNU General Public License for more details.
14
 *
15
 * You should have received a copy of the GNU General Public License
16
 * along with this program; if not, write to the Free Software
17
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
18
 *
19
 */
20
21
/* This file was automatically generated --- DO NOT EDIT */
22
/* Generated on Tue Aug 26 06:31:47 UTC 2025 */
23
24
#include "dft/codelet-dft.h"
25
26
#if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27
28
/* Generated by: ../../../genfft/gen_twidsq.native -fma -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 5 -name q1_5 -include dft/scalar/q.h */
29
30
/*
31
 * This function contains 200 FP additions, 170 FP multiplications,
32
 * (or, 70 additions, 40 multiplications, 130 fused multiply/add),
33
 * 75 stack variables, 4 constants, and 100 memory accesses
34
 */
35
#include "dft/scalar/q.h"
36
37
static void q1_5(R *rio, R *iio, const R *W, stride rs, stride vs, 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
     {
44
    INT m;
45
    for (m = mb, W = W + (mb * 8); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 8, MAKE_VOLATILE_STRIDE(10, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
46
         E T1, Tb, TM, Tw, T8, Ta, Tn, Tj, TH, Ts, Tq, Tr, TV, T15, T1G;
47
         E T1q, T12, T14, T1h, T1d, T1B, T1m, T1k, T1l, T1P, T1Z, T2A, T2k, T1W, T1Y;
48
         E T2b, T27, T2v, T2g, T2e, T2f, T3Z, T3V, T4j, T44, T42, T43, T3D, T3N, T4o;
49
         E T48, T3K, T3M, T2J, T2T, T3u, T3e, T2Q, T2S, T35, T31, T3p, T3a, T38, T39;
50
         {
51
        E T7, Tv, T4, Tu;
52
        T1 = rio[0];
53
        {
54
       E T5, T6, T2, T3;
55
       T5 = rio[WS(rs, 2)];
56
       T6 = rio[WS(rs, 3)];
57
       T7 = T5 + T6;
58
       Tv = T5 - T6;
59
       T2 = rio[WS(rs, 1)];
60
       T3 = rio[WS(rs, 4)];
61
       T4 = T2 + T3;
62
       Tu = T2 - T3;
63
        }
64
        Tb = T4 - T7;
65
        TM = FNMS(KP618033988, Tu, Tv);
66
        Tw = FMA(KP618033988, Tv, Tu);
67
        T8 = T4 + T7;
68
        Ta = FNMS(KP250000000, T8, T1);
69
         }
70
         {
71
        E Ti, Tp, Tf, To;
72
        Tn = iio[0];
73
        {
74
       E Tg, Th, Td, Te;
75
       Tg = iio[WS(rs, 2)];
76
       Th = iio[WS(rs, 3)];
77
       Ti = Tg - Th;
78
       Tp = Tg + Th;
79
       Td = iio[WS(rs, 1)];
80
       Te = iio[WS(rs, 4)];
81
       Tf = Td - Te;
82
       To = Td + Te;
83
        }
84
        Tj = FMA(KP618033988, Ti, Tf);
85
        TH = FNMS(KP618033988, Tf, Ti);
86
        Ts = To - Tp;
87
        Tq = To + Tp;
88
        Tr = FNMS(KP250000000, Tq, Tn);
89
         }
90
         {
91
        E T11, T1p, TY, T1o;
92
        TV = rio[WS(vs, 1)];
93
        {
94
       E TZ, T10, TW, TX;
95
       TZ = rio[WS(vs, 1) + WS(rs, 2)];
96
       T10 = rio[WS(vs, 1) + WS(rs, 3)];
97
       T11 = TZ + T10;
98
       T1p = TZ - T10;
99
       TW = rio[WS(vs, 1) + WS(rs, 1)];
100
       TX = rio[WS(vs, 1) + WS(rs, 4)];
101
       TY = TW + TX;
102
       T1o = TW - TX;
103
        }
104
        T15 = TY - T11;
105
        T1G = FNMS(KP618033988, T1o, T1p);
106
        T1q = FMA(KP618033988, T1p, T1o);
107
        T12 = TY + T11;
108
        T14 = FNMS(KP250000000, T12, TV);
109
         }
110
         {
111
        E T1c, T1j, T19, T1i;
112
        T1h = iio[WS(vs, 1)];
113
        {
114
       E T1a, T1b, T17, T18;
115
       T1a = iio[WS(vs, 1) + WS(rs, 2)];
116
       T1b = iio[WS(vs, 1) + WS(rs, 3)];
117
       T1c = T1a - T1b;
118
       T1j = T1a + T1b;
119
       T17 = iio[WS(vs, 1) + WS(rs, 1)];
120
       T18 = iio[WS(vs, 1) + WS(rs, 4)];
121
       T19 = T17 - T18;
122
       T1i = T17 + T18;
123
        }
124
        T1d = FMA(KP618033988, T1c, T19);
125
        T1B = FNMS(KP618033988, T19, T1c);
126
        T1m = T1i - T1j;
127
        T1k = T1i + T1j;
128
        T1l = FNMS(KP250000000, T1k, T1h);
129
         }
130
         {
131
        E T1V, T2j, T1S, T2i;
132
        T1P = rio[WS(vs, 2)];
133
        {
134
       E T1T, T1U, T1Q, T1R;
135
       T1T = rio[WS(vs, 2) + WS(rs, 2)];
136
       T1U = rio[WS(vs, 2) + WS(rs, 3)];
137
       T1V = T1T + T1U;
138
       T2j = T1T - T1U;
139
       T1Q = rio[WS(vs, 2) + WS(rs, 1)];
140
       T1R = rio[WS(vs, 2) + WS(rs, 4)];
141
       T1S = T1Q + T1R;
142
       T2i = T1Q - T1R;
143
        }
144
        T1Z = T1S - T1V;
145
        T2A = FNMS(KP618033988, T2i, T2j);
146
        T2k = FMA(KP618033988, T2j, T2i);
147
        T1W = T1S + T1V;
148
        T1Y = FNMS(KP250000000, T1W, T1P);
149
         }
150
         {
151
        E T26, T2d, T23, T2c;
152
        T2b = iio[WS(vs, 2)];
153
        {
154
       E T24, T25, T21, T22;
155
       T24 = iio[WS(vs, 2) + WS(rs, 2)];
156
       T25 = iio[WS(vs, 2) + WS(rs, 3)];
157
       T26 = T24 - T25;
158
       T2d = T24 + T25;
159
       T21 = iio[WS(vs, 2) + WS(rs, 1)];
160
       T22 = iio[WS(vs, 2) + WS(rs, 4)];
161
       T23 = T21 - T22;
162
       T2c = T21 + T22;
163
        }
164
        T27 = FMA(KP618033988, T26, T23);
165
        T2v = FNMS(KP618033988, T23, T26);
166
        T2g = T2c - T2d;
167
        T2e = T2c + T2d;
168
        T2f = FNMS(KP250000000, T2e, T2b);
169
         }
170
         {
171
        E T3U, T41, T3R, T40;
172
        T3Z = iio[WS(vs, 4)];
173
        {
174
       E T3S, T3T, T3P, T3Q;
175
       T3S = iio[WS(vs, 4) + WS(rs, 2)];
176
       T3T = iio[WS(vs, 4) + WS(rs, 3)];
177
       T3U = T3S - T3T;
178
       T41 = T3S + T3T;
179
       T3P = iio[WS(vs, 4) + WS(rs, 1)];
180
       T3Q = iio[WS(vs, 4) + WS(rs, 4)];
181
       T3R = T3P - T3Q;
182
       T40 = T3P + T3Q;
183
        }
184
        T3V = FMA(KP618033988, T3U, T3R);
185
        T4j = FNMS(KP618033988, T3R, T3U);
186
        T44 = T40 - T41;
187
        T42 = T40 + T41;
188
        T43 = FNMS(KP250000000, T42, T3Z);
189
         }
190
         {
191
        E T3J, T47, T3G, T46;
192
        T3D = rio[WS(vs, 4)];
193
        {
194
       E T3H, T3I, T3E, T3F;
195
       T3H = rio[WS(vs, 4) + WS(rs, 2)];
196
       T3I = rio[WS(vs, 4) + WS(rs, 3)];
197
       T3J = T3H + T3I;
198
       T47 = T3H - T3I;
199
       T3E = rio[WS(vs, 4) + WS(rs, 1)];
200
       T3F = rio[WS(vs, 4) + WS(rs, 4)];
201
       T3G = T3E + T3F;
202
       T46 = T3E - T3F;
203
        }
204
        T3N = T3G - T3J;
205
        T4o = FNMS(KP618033988, T46, T47);
206
        T48 = FMA(KP618033988, T47, T46);
207
        T3K = T3G + T3J;
208
        T3M = FNMS(KP250000000, T3K, T3D);
209
         }
210
         {
211
        E T2P, T3d, T2M, T3c;
212
        T2J = rio[WS(vs, 3)];
213
        {
214
       E T2N, T2O, T2K, T2L;
215
       T2N = rio[WS(vs, 3) + WS(rs, 2)];
216
       T2O = rio[WS(vs, 3) + WS(rs, 3)];
217
       T2P = T2N + T2O;
218
       T3d = T2N - T2O;
219
       T2K = rio[WS(vs, 3) + WS(rs, 1)];
220
       T2L = rio[WS(vs, 3) + WS(rs, 4)];
221
       T2M = T2K + T2L;
222
       T3c = T2K - T2L;
223
        }
224
        T2T = T2M - T2P;
225
        T3u = FNMS(KP618033988, T3c, T3d);
226
        T3e = FMA(KP618033988, T3d, T3c);
227
        T2Q = T2M + T2P;
228
        T2S = FNMS(KP250000000, T2Q, T2J);
229
         }
230
         {
231
        E T30, T37, T2X, T36;
232
        T35 = iio[WS(vs, 3)];
233
        {
234
       E T2Y, T2Z, T2V, T2W;
235
       T2Y = iio[WS(vs, 3) + WS(rs, 2)];
236
       T2Z = iio[WS(vs, 3) + WS(rs, 3)];
237
       T30 = T2Y - T2Z;
238
       T37 = T2Y + T2Z;
239
       T2V = iio[WS(vs, 3) + WS(rs, 1)];
240
       T2W = iio[WS(vs, 3) + WS(rs, 4)];
241
       T2X = T2V - T2W;
242
       T36 = T2V + T2W;
243
        }
244
        T31 = FMA(KP618033988, T30, T2X);
245
        T3p = FNMS(KP618033988, T2X, T30);
246
        T3a = T36 - T37;
247
        T38 = T36 + T37;
248
        T39 = FNMS(KP250000000, T38, T35);
249
         }
250
         rio[0] = T1 + T8;
251
         iio[0] = Tn + Tq;
252
         rio[WS(rs, 1)] = TV + T12;
253
         iio[WS(rs, 1)] = T1h + T1k;
254
         rio[WS(rs, 2)] = T1P + T1W;
255
         iio[WS(rs, 2)] = T2b + T2e;
256
         iio[WS(rs, 4)] = T3Z + T42;
257
         rio[WS(rs, 4)] = T3D + T3K;
258
         rio[WS(rs, 3)] = T2J + T2Q;
259
         iio[WS(rs, 3)] = T35 + T38;
260
         {
261
        E Tk, TA, Tx, TD, Tc, Tt;
262
        Tc = FMA(KP559016994, Tb, Ta);
263
        Tk = FMA(KP951056516, Tj, Tc);
264
        TA = FNMS(KP951056516, Tj, Tc);
265
        Tt = FMA(KP559016994, Ts, Tr);
266
        Tx = FNMS(KP951056516, Tw, Tt);
267
        TD = FMA(KP951056516, Tw, Tt);
268
        {
269
       E Tl, Ty, T9, Tm;
270
       T9 = W[0];
271
       Tl = T9 * Tk;
272
       Ty = T9 * Tx;
273
       Tm = W[1];
274
       rio[WS(vs, 1)] = FMA(Tm, Tx, Tl);
275
       iio[WS(vs, 1)] = FNMS(Tm, Tk, Ty);
276
        }
277
        {
278
       E TB, TE, Tz, TC;
279
       Tz = W[6];
280
       TB = Tz * TA;
281
       TE = Tz * TD;
282
       TC = W[7];
283
       rio[WS(vs, 4)] = FMA(TC, TD, TB);
284
       iio[WS(vs, 4)] = FNMS(TC, TA, TE);
285
        }
286
         }
287
         {
288
        E TI, TQ, TN, TT, TG, TL;
289
        TG = FNMS(KP559016994, Tb, Ta);
290
        TI = FNMS(KP951056516, TH, TG);
291
        TQ = FMA(KP951056516, TH, TG);
292
        TL = FNMS(KP559016994, Ts, Tr);
293
        TN = FMA(KP951056516, TM, TL);
294
        TT = FNMS(KP951056516, TM, TL);
295
        {
296
       E TJ, TO, TF, TK;
297
       TF = W[2];
298
       TJ = TF * TI;
299
       TO = TF * TN;
300
       TK = W[3];
301
       rio[WS(vs, 2)] = FMA(TK, TN, TJ);
302
       iio[WS(vs, 2)] = FNMS(TK, TI, TO);
303
        }
304
        {
305
       E TR, TU, TP, TS;
306
       TP = W[4];
307
       TR = TP * TQ;
308
       TU = TP * TT;
309
       TS = W[5];
310
       rio[WS(vs, 3)] = FMA(TS, TT, TR);
311
       iio[WS(vs, 3)] = FNMS(TS, TQ, TU);
312
        }
313
         }
314
         {
315
        E T2w, T2E, T2B, T2H, T2u, T2z;
316
        T2u = FNMS(KP559016994, T1Z, T1Y);
317
        T2w = FNMS(KP951056516, T2v, T2u);
318
        T2E = FMA(KP951056516, T2v, T2u);
319
        T2z = FNMS(KP559016994, T2g, T2f);
320
        T2B = FMA(KP951056516, T2A, T2z);
321
        T2H = FNMS(KP951056516, T2A, T2z);
322
        {
323
       E T2x, T2C, T2t, T2y;
324
       T2t = W[2];
325
       T2x = T2t * T2w;
326
       T2C = T2t * T2B;
327
       T2y = W[3];
328
       rio[WS(vs, 2) + WS(rs, 2)] = FMA(T2y, T2B, T2x);
329
       iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T2y, T2w, T2C);
330
        }
331
        {
332
       E T2F, T2I, T2D, T2G;
333
       T2D = W[4];
334
       T2F = T2D * T2E;
335
       T2I = T2D * T2H;
336
       T2G = W[5];
337
       rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2G, T2H, T2F);
338
       iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2G, T2E, T2I);
339
        }
340
         }
341
         {
342
        E T4k, T4s, T4p, T4v, T4i, T4n;
343
        T4i = FNMS(KP559016994, T3N, T3M);
344
        T4k = FNMS(KP951056516, T4j, T4i);
345
        T4s = FMA(KP951056516, T4j, T4i);
346
        T4n = FNMS(KP559016994, T44, T43);
347
        T4p = FMA(KP951056516, T4o, T4n);
348
        T4v = FNMS(KP951056516, T4o, T4n);
349
        {
350
       E T4l, T4q, T4h, T4m;
351
       T4h = W[2];
352
       T4l = T4h * T4k;
353
       T4q = T4h * T4p;
354
       T4m = W[3];
355
       rio[WS(vs, 2) + WS(rs, 4)] = FMA(T4m, T4p, T4l);
356
       iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T4m, T4k, T4q);
357
        }
358
        {
359
       E T4t, T4w, T4r, T4u;
360
       T4r = W[4];
361
       T4t = T4r * T4s;
362
       T4w = T4r * T4v;
363
       T4u = W[5];
364
       rio[WS(vs, 3) + WS(rs, 4)] = FMA(T4u, T4v, T4t);
365
       iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T4u, T4s, T4w);
366
        }
367
         }
368
         {
369
        E T28, T2o, T2l, T2r, T20, T2h;
370
        T20 = FMA(KP559016994, T1Z, T1Y);
371
        T28 = FMA(KP951056516, T27, T20);
372
        T2o = FNMS(KP951056516, T27, T20);
373
        T2h = FMA(KP559016994, T2g, T2f);
374
        T2l = FNMS(KP951056516, T2k, T2h);
375
        T2r = FMA(KP951056516, T2k, T2h);
376
        {
377
       E T29, T2m, T1X, T2a;
378
       T1X = W[0];
379
       T29 = T1X * T28;
380
       T2m = T1X * T2l;
381
       T2a = W[1];
382
       rio[WS(vs, 1) + WS(rs, 2)] = FMA(T2a, T2l, T29);
383
       iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T2a, T28, T2m);
384
        }
385
        {
386
       E T2p, T2s, T2n, T2q;
387
       T2n = W[6];
388
       T2p = T2n * T2o;
389
       T2s = T2n * T2r;
390
       T2q = W[7];
391
       rio[WS(vs, 4) + WS(rs, 2)] = FMA(T2q, T2r, T2p);
392
       iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T2q, T2o, T2s);
393
        }
394
         }
395
         {
396
        E T32, T3i, T3f, T3l, T2U, T3b;
397
        T2U = FMA(KP559016994, T2T, T2S);
398
        T32 = FMA(KP951056516, T31, T2U);
399
        T3i = FNMS(KP951056516, T31, T2U);
400
        T3b = FMA(KP559016994, T3a, T39);
401
        T3f = FNMS(KP951056516, T3e, T3b);
402
        T3l = FMA(KP951056516, T3e, T3b);
403
        {
404
       E T33, T3g, T2R, T34;
405
       T2R = W[0];
406
       T33 = T2R * T32;
407
       T3g = T2R * T3f;
408
       T34 = W[1];
409
       rio[WS(vs, 1) + WS(rs, 3)] = FMA(T34, T3f, T33);
410
       iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T34, T32, T3g);
411
        }
412
        {
413
       E T3j, T3m, T3h, T3k;
414
       T3h = W[6];
415
       T3j = T3h * T3i;
416
       T3m = T3h * T3l;
417
       T3k = W[7];
418
       rio[WS(vs, 4) + WS(rs, 3)] = FMA(T3k, T3l, T3j);
419
       iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T3k, T3i, T3m);
420
        }
421
         }
422
         {
423
        E T3q, T3y, T3v, T3B, T3o, T3t;
424
        T3o = FNMS(KP559016994, T2T, T2S);
425
        T3q = FNMS(KP951056516, T3p, T3o);
426
        T3y = FMA(KP951056516, T3p, T3o);
427
        T3t = FNMS(KP559016994, T3a, T39);
428
        T3v = FMA(KP951056516, T3u, T3t);
429
        T3B = FNMS(KP951056516, T3u, T3t);
430
        {
431
       E T3r, T3w, T3n, T3s;
432
       T3n = W[2];
433
       T3r = T3n * T3q;
434
       T3w = T3n * T3v;
435
       T3s = W[3];
436
       rio[WS(vs, 2) + WS(rs, 3)] = FMA(T3s, T3v, T3r);
437
       iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T3s, T3q, T3w);
438
        }
439
        {
440
       E T3z, T3C, T3x, T3A;
441
       T3x = W[4];
442
       T3z = T3x * T3y;
443
       T3C = T3x * T3B;
444
       T3A = W[5];
445
       rio[WS(vs, 3) + WS(rs, 3)] = FMA(T3A, T3B, T3z);
446
       iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T3A, T3y, T3C);
447
        }
448
         }
449
         {
450
        E T3W, T4c, T49, T4f, T3O, T45;
451
        T3O = FMA(KP559016994, T3N, T3M);
452
        T3W = FMA(KP951056516, T3V, T3O);
453
        T4c = FNMS(KP951056516, T3V, T3O);
454
        T45 = FMA(KP559016994, T44, T43);
455
        T49 = FNMS(KP951056516, T48, T45);
456
        T4f = FMA(KP951056516, T48, T45);
457
        {
458
       E T3X, T4a, T3L, T3Y;
459
       T3L = W[0];
460
       T3X = T3L * T3W;
461
       T4a = T3L * T49;
462
       T3Y = W[1];
463
       rio[WS(vs, 1) + WS(rs, 4)] = FMA(T3Y, T49, T3X);
464
       iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T3Y, T3W, T4a);
465
        }
466
        {
467
       E T4d, T4g, T4b, T4e;
468
       T4b = W[6];
469
       T4d = T4b * T4c;
470
       T4g = T4b * T4f;
471
       T4e = W[7];
472
       rio[WS(vs, 4) + WS(rs, 4)] = FMA(T4e, T4f, T4d);
473
       iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T4e, T4c, T4g);
474
        }
475
         }
476
         {
477
        E T1C, T1K, T1H, T1N, T1A, T1F;
478
        T1A = FNMS(KP559016994, T15, T14);
479
        T1C = FNMS(KP951056516, T1B, T1A);
480
        T1K = FMA(KP951056516, T1B, T1A);
481
        T1F = FNMS(KP559016994, T1m, T1l);
482
        T1H = FMA(KP951056516, T1G, T1F);
483
        T1N = FNMS(KP951056516, T1G, T1F);
484
        {
485
       E T1D, T1I, T1z, T1E;
486
       T1z = W[2];
487
       T1D = T1z * T1C;
488
       T1I = T1z * T1H;
489
       T1E = W[3];
490
       rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1E, T1H, T1D);
491
       iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1E, T1C, T1I);
492
        }
493
        {
494
       E T1L, T1O, T1J, T1M;
495
       T1J = W[4];
496
       T1L = T1J * T1K;
497
       T1O = T1J * T1N;
498
       T1M = W[5];
499
       rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1M, T1N, T1L);
500
       iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1M, T1K, T1O);
501
        }
502
         }
503
         {
504
        E T1e, T1u, T1r, T1x, T16, T1n;
505
        T16 = FMA(KP559016994, T15, T14);
506
        T1e = FMA(KP951056516, T1d, T16);
507
        T1u = FNMS(KP951056516, T1d, T16);
508
        T1n = FMA(KP559016994, T1m, T1l);
509
        T1r = FNMS(KP951056516, T1q, T1n);
510
        T1x = FMA(KP951056516, T1q, T1n);
511
        {
512
       E T1f, T1s, T13, T1g;
513
       T13 = W[0];
514
       T1f = T13 * T1e;
515
       T1s = T13 * T1r;
516
       T1g = W[1];
517
       rio[WS(vs, 1) + WS(rs, 1)] = FMA(T1g, T1r, T1f);
518
       iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T1g, T1e, T1s);
519
        }
520
        {
521
       E T1v, T1y, T1t, T1w;
522
       T1t = W[6];
523
       T1v = T1t * T1u;
524
       T1y = T1t * T1x;
525
       T1w = W[7];
526
       rio[WS(vs, 4) + WS(rs, 1)] = FMA(T1w, T1x, T1v);
527
       iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T1w, T1u, T1y);
528
        }
529
         }
530
    }
531
     }
532
}
533
534
static const tw_instr twinstr[] = {
535
     { TW_FULL, 0, 5 },
536
     { TW_NEXT, 1, 0 }
537
};
538
539
static const ct_desc desc = { 5, "q1_5", twinstr, &GENUS, { 70, 40, 130, 0 }, 0, 0, 0 };
540
541
void X(codelet_q1_5) (planner *p) {
542
     X(kdft_difsq_register) (p, q1_5, &desc);
543
}
544
#else
545
546
/* Generated by: ../../../genfft/gen_twidsq.native -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 5 -name q1_5 -include dft/scalar/q.h */
547
548
/*
549
 * This function contains 200 FP additions, 140 FP multiplications,
550
 * (or, 130 additions, 70 multiplications, 70 fused multiply/add),
551
 * 75 stack variables, 4 constants, and 100 memory accesses
552
 */
553
#include "dft/scalar/q.h"
554
555
static void q1_5(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
556
11
{
557
11
     DK(KP250000000, +0.250000000000000000000000000000000000000000000);
558
11
     DK(KP587785252, +0.587785252292473129168705954639072768597652438);
559
11
     DK(KP951056516, +0.951056516295153572116439333379382143405698634);
560
11
     DK(KP559016994, +0.559016994374947424102293417182819058860154590);
561
11
     {
562
11
    INT m;
563
110
    for (m = mb, W = W + (mb * 8); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 8, MAKE_VOLATILE_STRIDE(10, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
564
99
         E T1, Ta, TG, Tv, T8, Tb, Tp, Tj, TD, To, Tq, Tr, TN, TW, T1s;
565
99
         E T1h, TU, TX, T1b, T15, T1p, T1a, T1c, T1d, T1z, T1I, T2e, T23, T1G, T1J;
566
99
         E T1X, T1R, T2b, T1W, T1Y, T1Z, T3v, T3p, T3J, T3u, T3w, T3x, T37, T3g, T3M;
567
99
         E T3B, T3e, T3h, T2l, T2u, T30, T2P, T2s, T2v, T2J, T2D, T2X, T2I, T2K, T2L;
568
99
         {
569
99
        E T7, Tu, T4, Tt;
570
99
        T1 = rio[0];
571
99
        {
572
99
       E T5, T6, T2, T3;
573
99
       T5 = rio[WS(rs, 2)];
574
99
       T6 = rio[WS(rs, 3)];
575
99
       T7 = T5 + T6;
576
99
       Tu = T5 - T6;
577
99
       T2 = rio[WS(rs, 1)];
578
99
       T3 = rio[WS(rs, 4)];
579
99
       T4 = T2 + T3;
580
99
       Tt = T2 - T3;
581
99
        }
582
99
        Ta = KP559016994 * (T4 - T7);
583
99
        TG = FNMS(KP587785252, Tt, KP951056516 * Tu);
584
99
        Tv = FMA(KP951056516, Tt, KP587785252 * Tu);
585
99
        T8 = T4 + T7;
586
99
        Tb = FNMS(KP250000000, T8, T1);
587
99
         }
588
99
         {
589
99
        E Ti, Tn, Tf, Tm;
590
99
        Tp = iio[0];
591
99
        {
592
99
       E Tg, Th, Td, Te;
593
99
       Tg = iio[WS(rs, 2)];
594
99
       Th = iio[WS(rs, 3)];
595
99
       Ti = Tg - Th;
596
99
       Tn = Tg + Th;
597
99
       Td = iio[WS(rs, 1)];
598
99
       Te = iio[WS(rs, 4)];
599
99
       Tf = Td - Te;
600
99
       Tm = Td + Te;
601
99
        }
602
99
        Tj = FMA(KP951056516, Tf, KP587785252 * Ti);
603
99
        TD = FNMS(KP587785252, Tf, KP951056516 * Ti);
604
99
        To = KP559016994 * (Tm - Tn);
605
99
        Tq = Tm + Tn;
606
99
        Tr = FNMS(KP250000000, Tq, Tp);
607
99
         }
608
99
         {
609
99
        E TT, T1g, TQ, T1f;
610
99
        TN = rio[WS(vs, 1)];
611
99
        {
612
99
       E TR, TS, TO, TP;
613
99
       TR = rio[WS(vs, 1) + WS(rs, 2)];
614
99
       TS = rio[WS(vs, 1) + WS(rs, 3)];
615
99
       TT = TR + TS;
616
99
       T1g = TR - TS;
617
99
       TO = rio[WS(vs, 1) + WS(rs, 1)];
618
99
       TP = rio[WS(vs, 1) + WS(rs, 4)];
619
99
       TQ = TO + TP;
620
99
       T1f = TO - TP;
621
99
        }
622
99
        TW = KP559016994 * (TQ - TT);
623
99
        T1s = FNMS(KP587785252, T1f, KP951056516 * T1g);
624
99
        T1h = FMA(KP951056516, T1f, KP587785252 * T1g);
625
99
        TU = TQ + TT;
626
99
        TX = FNMS(KP250000000, TU, TN);
627
99
         }
628
99
         {
629
99
        E T14, T19, T11, T18;
630
99
        T1b = iio[WS(vs, 1)];
631
99
        {
632
99
       E T12, T13, TZ, T10;
633
99
       T12 = iio[WS(vs, 1) + WS(rs, 2)];
634
99
       T13 = iio[WS(vs, 1) + WS(rs, 3)];
635
99
       T14 = T12 - T13;
636
99
       T19 = T12 + T13;
637
99
       TZ = iio[WS(vs, 1) + WS(rs, 1)];
638
99
       T10 = iio[WS(vs, 1) + WS(rs, 4)];
639
99
       T11 = TZ - T10;
640
99
       T18 = TZ + T10;
641
99
        }
642
99
        T15 = FMA(KP951056516, T11, KP587785252 * T14);
643
99
        T1p = FNMS(KP587785252, T11, KP951056516 * T14);
644
99
        T1a = KP559016994 * (T18 - T19);
645
99
        T1c = T18 + T19;
646
99
        T1d = FNMS(KP250000000, T1c, T1b);
647
99
         }
648
99
         {
649
99
        E T1F, T22, T1C, T21;
650
99
        T1z = rio[WS(vs, 2)];
651
99
        {
652
99
       E T1D, T1E, T1A, T1B;
653
99
       T1D = rio[WS(vs, 2) + WS(rs, 2)];
654
99
       T1E = rio[WS(vs, 2) + WS(rs, 3)];
655
99
       T1F = T1D + T1E;
656
99
       T22 = T1D - T1E;
657
99
       T1A = rio[WS(vs, 2) + WS(rs, 1)];
658
99
       T1B = rio[WS(vs, 2) + WS(rs, 4)];
659
99
       T1C = T1A + T1B;
660
99
       T21 = T1A - T1B;
661
99
        }
662
99
        T1I = KP559016994 * (T1C - T1F);
663
99
        T2e = FNMS(KP587785252, T21, KP951056516 * T22);
664
99
        T23 = FMA(KP951056516, T21, KP587785252 * T22);
665
99
        T1G = T1C + T1F;
666
99
        T1J = FNMS(KP250000000, T1G, T1z);
667
99
         }
668
99
         {
669
99
        E T1Q, T1V, T1N, T1U;
670
99
        T1X = iio[WS(vs, 2)];
671
99
        {
672
99
       E T1O, T1P, T1L, T1M;
673
99
       T1O = iio[WS(vs, 2) + WS(rs, 2)];
674
99
       T1P = iio[WS(vs, 2) + WS(rs, 3)];
675
99
       T1Q = T1O - T1P;
676
99
       T1V = T1O + T1P;
677
99
       T1L = iio[WS(vs, 2) + WS(rs, 1)];
678
99
       T1M = iio[WS(vs, 2) + WS(rs, 4)];
679
99
       T1N = T1L - T1M;
680
99
       T1U = T1L + T1M;
681
99
        }
682
99
        T1R = FMA(KP951056516, T1N, KP587785252 * T1Q);
683
99
        T2b = FNMS(KP587785252, T1N, KP951056516 * T1Q);
684
99
        T1W = KP559016994 * (T1U - T1V);
685
99
        T1Y = T1U + T1V;
686
99
        T1Z = FNMS(KP250000000, T1Y, T1X);
687
99
         }
688
99
         {
689
99
        E T3o, T3t, T3l, T3s;
690
99
        T3v = iio[WS(vs, 4)];
691
99
        {
692
99
       E T3m, T3n, T3j, T3k;
693
99
       T3m = iio[WS(vs, 4) + WS(rs, 2)];
694
99
       T3n = iio[WS(vs, 4) + WS(rs, 3)];
695
99
       T3o = T3m - T3n;
696
99
       T3t = T3m + T3n;
697
99
       T3j = iio[WS(vs, 4) + WS(rs, 1)];
698
99
       T3k = iio[WS(vs, 4) + WS(rs, 4)];
699
99
       T3l = T3j - T3k;
700
99
       T3s = T3j + T3k;
701
99
        }
702
99
        T3p = FMA(KP951056516, T3l, KP587785252 * T3o);
703
99
        T3J = FNMS(KP587785252, T3l, KP951056516 * T3o);
704
99
        T3u = KP559016994 * (T3s - T3t);
705
99
        T3w = T3s + T3t;
706
99
        T3x = FNMS(KP250000000, T3w, T3v);
707
99
         }
708
99
         {
709
99
        E T3d, T3A, T3a, T3z;
710
99
        T37 = rio[WS(vs, 4)];
711
99
        {
712
99
       E T3b, T3c, T38, T39;
713
99
       T3b = rio[WS(vs, 4) + WS(rs, 2)];
714
99
       T3c = rio[WS(vs, 4) + WS(rs, 3)];
715
99
       T3d = T3b + T3c;
716
99
       T3A = T3b - T3c;
717
99
       T38 = rio[WS(vs, 4) + WS(rs, 1)];
718
99
       T39 = rio[WS(vs, 4) + WS(rs, 4)];
719
99
       T3a = T38 + T39;
720
99
       T3z = T38 - T39;
721
99
        }
722
99
        T3g = KP559016994 * (T3a - T3d);
723
99
        T3M = FNMS(KP587785252, T3z, KP951056516 * T3A);
724
99
        T3B = FMA(KP951056516, T3z, KP587785252 * T3A);
725
99
        T3e = T3a + T3d;
726
99
        T3h = FNMS(KP250000000, T3e, T37);
727
99
         }
728
99
         {
729
99
        E T2r, T2O, T2o, T2N;
730
99
        T2l = rio[WS(vs, 3)];
731
99
        {
732
99
       E T2p, T2q, T2m, T2n;
733
99
       T2p = rio[WS(vs, 3) + WS(rs, 2)];
734
99
       T2q = rio[WS(vs, 3) + WS(rs, 3)];
735
99
       T2r = T2p + T2q;
736
99
       T2O = T2p - T2q;
737
99
       T2m = rio[WS(vs, 3) + WS(rs, 1)];
738
99
       T2n = rio[WS(vs, 3) + WS(rs, 4)];
739
99
       T2o = T2m + T2n;
740
99
       T2N = T2m - T2n;
741
99
        }
742
99
        T2u = KP559016994 * (T2o - T2r);
743
99
        T30 = FNMS(KP587785252, T2N, KP951056516 * T2O);
744
99
        T2P = FMA(KP951056516, T2N, KP587785252 * T2O);
745
99
        T2s = T2o + T2r;
746
99
        T2v = FNMS(KP250000000, T2s, T2l);
747
99
         }
748
99
         {
749
99
        E T2C, T2H, T2z, T2G;
750
99
        T2J = iio[WS(vs, 3)];
751
99
        {
752
99
       E T2A, T2B, T2x, T2y;
753
99
       T2A = iio[WS(vs, 3) + WS(rs, 2)];
754
99
       T2B = iio[WS(vs, 3) + WS(rs, 3)];
755
99
       T2C = T2A - T2B;
756
99
       T2H = T2A + T2B;
757
99
       T2x = iio[WS(vs, 3) + WS(rs, 1)];
758
99
       T2y = iio[WS(vs, 3) + WS(rs, 4)];
759
99
       T2z = T2x - T2y;
760
99
       T2G = T2x + T2y;
761
99
        }
762
99
        T2D = FMA(KP951056516, T2z, KP587785252 * T2C);
763
99
        T2X = FNMS(KP587785252, T2z, KP951056516 * T2C);
764
99
        T2I = KP559016994 * (T2G - T2H);
765
99
        T2K = T2G + T2H;
766
99
        T2L = FNMS(KP250000000, T2K, T2J);
767
99
         }
768
99
         rio[0] = T1 + T8;
769
99
         iio[0] = Tp + Tq;
770
99
         rio[WS(rs, 1)] = TN + TU;
771
99
         iio[WS(rs, 1)] = T1b + T1c;
772
99
         rio[WS(rs, 2)] = T1z + T1G;
773
99
         iio[WS(rs, 2)] = T1X + T1Y;
774
99
         iio[WS(rs, 4)] = T3v + T3w;
775
99
         rio[WS(rs, 4)] = T37 + T3e;
776
99
         rio[WS(rs, 3)] = T2l + T2s;
777
99
         iio[WS(rs, 3)] = T2J + T2K;
778
99
         {
779
99
        E Tk, Ty, Tw, TA, Tc, Ts;
780
99
        Tc = Ta + Tb;
781
99
        Tk = Tc + Tj;
782
99
        Ty = Tc - Tj;
783
99
        Ts = To + Tr;
784
99
        Tw = Ts - Tv;
785
99
        TA = Tv + Ts;
786
99
        {
787
99
       E T9, Tl, Tx, Tz;
788
99
       T9 = W[0];
789
99
       Tl = W[1];
790
99
       rio[WS(vs, 1)] = FMA(T9, Tk, Tl * Tw);
791
99
       iio[WS(vs, 1)] = FNMS(Tl, Tk, T9 * Tw);
792
99
       Tx = W[6];
793
99
       Tz = W[7];
794
99
       rio[WS(vs, 4)] = FMA(Tx, Ty, Tz * TA);
795
99
       iio[WS(vs, 4)] = FNMS(Tz, Ty, Tx * TA);
796
99
        }
797
99
         }
798
99
         {
799
99
        E TE, TK, TI, TM, TC, TH;
800
99
        TC = Tb - Ta;
801
99
        TE = TC - TD;
802
99
        TK = TC + TD;
803
99
        TH = Tr - To;
804
99
        TI = TG + TH;
805
99
        TM = TH - TG;
806
99
        {
807
99
       E TB, TF, TJ, TL;
808
99
       TB = W[2];
809
99
       TF = W[3];
810
99
       rio[WS(vs, 2)] = FMA(TB, TE, TF * TI);
811
99
       iio[WS(vs, 2)] = FNMS(TF, TE, TB * TI);
812
99
       TJ = W[4];
813
99
       TL = W[5];
814
99
       rio[WS(vs, 3)] = FMA(TJ, TK, TL * TM);
815
99
       iio[WS(vs, 3)] = FNMS(TL, TK, TJ * TM);
816
99
        }
817
99
         }
818
99
         {
819
99
        E T2c, T2i, T2g, T2k, T2a, T2f;
820
99
        T2a = T1J - T1I;
821
99
        T2c = T2a - T2b;
822
99
        T2i = T2a + T2b;
823
99
        T2f = T1Z - T1W;
824
99
        T2g = T2e + T2f;
825
99
        T2k = T2f - T2e;
826
99
        {
827
99
       E T29, T2d, T2h, T2j;
828
99
       T29 = W[2];
829
99
       T2d = W[3];
830
99
       rio[WS(vs, 2) + WS(rs, 2)] = FMA(T29, T2c, T2d * T2g);
831
99
       iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T2d, T2c, T29 * T2g);
832
99
       T2h = W[4];
833
99
       T2j = W[5];
834
99
       rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2h, T2i, T2j * T2k);
835
99
       iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2j, T2i, T2h * T2k);
836
99
        }
837
99
         }
838
99
         {
839
99
        E T3K, T3Q, T3O, T3S, T3I, T3N;
840
99
        T3I = T3h - T3g;
841
99
        T3K = T3I - T3J;
842
99
        T3Q = T3I + T3J;
843
99
        T3N = T3x - T3u;
844
99
        T3O = T3M + T3N;
845
99
        T3S = T3N - T3M;
846
99
        {
847
99
       E T3H, T3L, T3P, T3R;
848
99
       T3H = W[2];
849
99
       T3L = W[3];
850
99
       rio[WS(vs, 2) + WS(rs, 4)] = FMA(T3H, T3K, T3L * T3O);
851
99
       iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T3L, T3K, T3H * T3O);
852
99
       T3P = W[4];
853
99
       T3R = W[5];
854
99
       rio[WS(vs, 3) + WS(rs, 4)] = FMA(T3P, T3Q, T3R * T3S);
855
99
       iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T3R, T3Q, T3P * T3S);
856
99
        }
857
99
         }
858
99
         {
859
99
        E T1S, T26, T24, T28, T1K, T20;
860
99
        T1K = T1I + T1J;
861
99
        T1S = T1K + T1R;
862
99
        T26 = T1K - T1R;
863
99
        T20 = T1W + T1Z;
864
99
        T24 = T20 - T23;
865
99
        T28 = T23 + T20;
866
99
        {
867
99
       E T1H, T1T, T25, T27;
868
99
       T1H = W[0];
869
99
       T1T = W[1];
870
99
       rio[WS(vs, 1) + WS(rs, 2)] = FMA(T1H, T1S, T1T * T24);
871
99
       iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T1T, T1S, T1H * T24);
872
99
       T25 = W[6];
873
99
       T27 = W[7];
874
99
       rio[WS(vs, 4) + WS(rs, 2)] = FMA(T25, T26, T27 * T28);
875
99
       iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T27, T26, T25 * T28);
876
99
        }
877
99
         }
878
99
         {
879
99
        E T2E, T2S, T2Q, T2U, T2w, T2M;
880
99
        T2w = T2u + T2v;
881
99
        T2E = T2w + T2D;
882
99
        T2S = T2w - T2D;
883
99
        T2M = T2I + T2L;
884
99
        T2Q = T2M - T2P;
885
99
        T2U = T2P + T2M;
886
99
        {
887
99
       E T2t, T2F, T2R, T2T;
888
99
       T2t = W[0];
889
99
       T2F = W[1];
890
99
       rio[WS(vs, 1) + WS(rs, 3)] = FMA(T2t, T2E, T2F * T2Q);
891
99
       iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T2F, T2E, T2t * T2Q);
892
99
       T2R = W[6];
893
99
       T2T = W[7];
894
99
       rio[WS(vs, 4) + WS(rs, 3)] = FMA(T2R, T2S, T2T * T2U);
895
99
       iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T2T, T2S, T2R * T2U);
896
99
        }
897
99
         }
898
99
         {
899
99
        E T2Y, T34, T32, T36, T2W, T31;
900
99
        T2W = T2v - T2u;
901
99
        T2Y = T2W - T2X;
902
99
        T34 = T2W + T2X;
903
99
        T31 = T2L - T2I;
904
99
        T32 = T30 + T31;
905
99
        T36 = T31 - T30;
906
99
        {
907
99
       E T2V, T2Z, T33, T35;
908
99
       T2V = W[2];
909
99
       T2Z = W[3];
910
99
       rio[WS(vs, 2) + WS(rs, 3)] = FMA(T2V, T2Y, T2Z * T32);
911
99
       iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T2Z, T2Y, T2V * T32);
912
99
       T33 = W[4];
913
99
       T35 = W[5];
914
99
       rio[WS(vs, 3) + WS(rs, 3)] = FMA(T33, T34, T35 * T36);
915
99
       iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T35, T34, T33 * T36);
916
99
        }
917
99
         }
918
99
         {
919
99
        E T3q, T3E, T3C, T3G, T3i, T3y;
920
99
        T3i = T3g + T3h;
921
99
        T3q = T3i + T3p;
922
99
        T3E = T3i - T3p;
923
99
        T3y = T3u + T3x;
924
99
        T3C = T3y - T3B;
925
99
        T3G = T3B + T3y;
926
99
        {
927
99
       E T3f, T3r, T3D, T3F;
928
99
       T3f = W[0];
929
99
       T3r = W[1];
930
99
       rio[WS(vs, 1) + WS(rs, 4)] = FMA(T3f, T3q, T3r * T3C);
931
99
       iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T3r, T3q, T3f * T3C);
932
99
       T3D = W[6];
933
99
       T3F = W[7];
934
99
       rio[WS(vs, 4) + WS(rs, 4)] = FMA(T3D, T3E, T3F * T3G);
935
99
       iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T3F, T3E, T3D * T3G);
936
99
        }
937
99
         }
938
99
         {
939
99
        E T1q, T1w, T1u, T1y, T1o, T1t;
940
99
        T1o = TX - TW;
941
99
        T1q = T1o - T1p;
942
99
        T1w = T1o + T1p;
943
99
        T1t = T1d - T1a;
944
99
        T1u = T1s + T1t;
945
99
        T1y = T1t - T1s;
946
99
        {
947
99
       E T1n, T1r, T1v, T1x;
948
99
       T1n = W[2];
949
99
       T1r = W[3];
950
99
       rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1n, T1q, T1r * T1u);
951
99
       iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1r, T1q, T1n * T1u);
952
99
       T1v = W[4];
953
99
       T1x = W[5];
954
99
       rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1v, T1w, T1x * T1y);
955
99
       iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1x, T1w, T1v * T1y);
956
99
        }
957
99
         }
958
99
         {
959
99
        E T16, T1k, T1i, T1m, TY, T1e;
960
99
        TY = TW + TX;
961
99
        T16 = TY + T15;
962
99
        T1k = TY - T15;
963
99
        T1e = T1a + T1d;
964
99
        T1i = T1e - T1h;
965
99
        T1m = T1h + T1e;
966
99
        {
967
99
       E TV, T17, T1j, T1l;
968
99
       TV = W[0];
969
99
       T17 = W[1];
970
99
       rio[WS(vs, 1) + WS(rs, 1)] = FMA(TV, T16, T17 * T1i);
971
99
       iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T17, T16, TV * T1i);
972
99
       T1j = W[6];
973
99
       T1l = W[7];
974
99
       rio[WS(vs, 4) + WS(rs, 1)] = FMA(T1j, T1k, T1l * T1m);
975
99
       iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T1l, T1k, T1j * T1m);
976
99
        }
977
99
         }
978
99
    }
979
11
     }
980
11
}
981
982
static const tw_instr twinstr[] = {
983
     { TW_FULL, 0, 5 },
984
     { TW_NEXT, 1, 0 }
985
};
986
987
static const ct_desc desc = { 5, "q1_5", twinstr, &GENUS, { 130, 70, 70, 0 }, 0, 0, 0 };
988
989
1
void X(codelet_q1_5) (planner *p) {
990
1
     X(kdft_difsq_register) (p, q1_5, &desc);
991
1
}
992
#endif