comparison src/fftw-3.3.8/dft/simd/common/t3bv_32.c @ 167:bd3cc4d1df30

Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam <cannam@all-day-breakfast.com>
date Tue, 19 Nov 2019 14:52:55 +0000
parents
children
comparison
equal deleted inserted replaced
166:cbd6d7e562c7 167:bd3cc4d1df30
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 Thu May 24 08:06:06 EDT 2018 */
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_twiddle_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -no-generate-bytw -n 32 -name t3bv_32 -include dft/simd/t3b.h -sign 1 */
29
30 /*
31 * This function contains 244 FP additions, 214 FP multiplications,
32 * (or, 146 additions, 116 multiplications, 98 fused multiply/add),
33 * 90 stack variables, 7 constants, and 64 memory accesses
34 */
35 #include "dft/simd/t3b.h"
36
37 static void t3bv_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
38 {
39 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
40 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
41 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
42 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
44 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
45 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
46 {
47 INT m;
48 R *x;
49 x = ii;
50 for (m = mb, W = W + (mb * ((TWVL / VL) * 8)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 8), MAKE_VOLATILE_STRIDE(32, rs)) {
51 V T2, T5, T3, T4, Tc, T1C, TT, Tz, Tn, T6, TP, Tf, TK, T7, T8;
52 V Tv, T1w, T21, Tg, Tk, T1D, T1O, TC, T18, T12, T1t, TH, TL, TQ, T1m;
53 V T1c;
54 T2 = LDW(&(W[0]));
55 T5 = LDW(&(W[TWVL * 4]));
56 T3 = LDW(&(W[TWVL * 2]));
57 T4 = VZMULJ(T2, T3);
58 Tc = VZMUL(T2, T3);
59 T1C = VZMULJ(T2, T5);
60 TT = VZMULJ(T3, T5);
61 Tz = VZMUL(T2, T5);
62 Tn = VZMUL(T3, T5);
63 T6 = VZMUL(T4, T5);
64 TP = VZMULJ(Tc, T5);
65 Tf = VZMULJ(T4, T5);
66 TK = VZMUL(Tc, T5);
67 T7 = LDW(&(W[TWVL * 6]));
68 T8 = VZMULJ(T6, T7);
69 Tv = VZMULJ(T5, T7);
70 T1w = VZMULJ(Tn, T7);
71 T21 = VZMULJ(T3, T7);
72 Tg = VZMULJ(Tf, T7);
73 Tk = VZMUL(T2, T7);
74 T1D = VZMULJ(T1C, T7);
75 T1O = VZMULJ(Tc, T7);
76 TC = VZMULJ(T2, T7);
77 T18 = VZMULJ(TT, T7);
78 T12 = VZMULJ(Tz, T7);
79 T1t = VZMUL(Tc, T7);
80 TH = VZMUL(T3, T7);
81 TL = VZMULJ(TK, T7);
82 TQ = VZMULJ(TP, T7);
83 T1m = VZMULJ(T4, T7);
84 T1c = VZMUL(T4, T7);
85 {
86 V Tb, T24, T2T, T3x, Tr, T25, T2W, T3K, TX, T28, T3j, T3z, TG, T27, T3g;
87 V T3y, T1N, T2v, T3a, T3G, T1V, T2w, T37, T3F, T1j, T2s, T33, T3D, T1r, T2t;
88 V T30, T3C;
89 {
90 V T1, T23, Ta, T20, T22, T9, T1Z, T2R, T2S;
91 T1 = LD(&(x[0]), ms, &(x[0]));
92 T22 = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
93 T23 = VZMUL(T21, T22);
94 T9 = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
95 Ta = VZMUL(T8, T9);
96 T1Z = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
97 T20 = VZMUL(T1C, T1Z);
98 Tb = VSUB(T1, Ta);
99 T24 = VSUB(T20, T23);
100 T2R = VADD(T1, Ta);
101 T2S = VADD(T20, T23);
102 T2T = VADD(T2R, T2S);
103 T3x = VSUB(T2R, T2S);
104 }
105 {
106 V Te, Tp, Ti, Tm;
107 {
108 V Td, To, Th, Tl;
109 Td = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
110 Te = VZMUL(Tc, Td);
111 To = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
112 Tp = VZMUL(Tn, To);
113 Th = LD(&(x[WS(rs, 20)]), ms, &(x[0]));
114 Ti = VZMUL(Tg, Th);
115 Tl = LD(&(x[WS(rs, 28)]), ms, &(x[0]));
116 Tm = VZMUL(Tk, Tl);
117 }
118 {
119 V Tj, Tq, T2U, T2V;
120 Tj = VSUB(Te, Ti);
121 Tq = VSUB(Tm, Tp);
122 Tr = VADD(Tj, Tq);
123 T25 = VSUB(Tj, Tq);
124 T2U = VADD(Te, Ti);
125 T2V = VADD(Tm, Tp);
126 T2W = VADD(T2U, T2V);
127 T3K = VSUB(T2U, T2V);
128 }
129 }
130 {
131 V TJ, TV, TN, TS;
132 {
133 V TI, TU, TM, TR;
134 TI = LD(&(x[WS(rs, 30)]), ms, &(x[0]));
135 TJ = VZMUL(TH, TI);
136 TU = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
137 TV = VZMUL(TT, TU);
138 TM = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
139 TN = VZMUL(TL, TM);
140 TR = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
141 TS = VZMUL(TQ, TR);
142 }
143 {
144 V TO, TW, T3h, T3i;
145 TO = VSUB(TJ, TN);
146 TW = VSUB(TS, TV);
147 TX = VFNMS(LDK(KP414213562), TW, TO);
148 T28 = VFMA(LDK(KP414213562), TO, TW);
149 T3h = VADD(TJ, TN);
150 T3i = VADD(TV, TS);
151 T3j = VADD(T3h, T3i);
152 T3z = VSUB(T3h, T3i);
153 }
154 }
155 {
156 V Tu, TE, Tx, TB;
157 {
158 V Tt, TD, Tw, TA;
159 Tt = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
160 Tu = VZMUL(T4, Tt);
161 TD = LD(&(x[WS(rs, 26)]), ms, &(x[0]));
162 TE = VZMUL(TC, TD);
163 Tw = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
164 Tx = VZMUL(Tv, Tw);
165 TA = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
166 TB = VZMUL(Tz, TA);
167 }
168 {
169 V Ty, TF, T3e, T3f;
170 Ty = VSUB(Tu, Tx);
171 TF = VSUB(TB, TE);
172 TG = VFNMS(LDK(KP414213562), TF, Ty);
173 T27 = VFMA(LDK(KP414213562), Ty, TF);
174 T3e = VADD(Tu, Tx);
175 T3f = VADD(TB, TE);
176 T3g = VADD(T3e, T3f);
177 T3y = VSUB(T3e, T3f);
178 }
179 }
180 {
181 V T1v, T1y, T1S, T1Q, T1I, T1K, T1L, T1B, T1F, T1G;
182 {
183 V T1u, T1x, T1R, T1P;
184 T1u = LD(&(x[WS(rs, 31)]), ms, &(x[WS(rs, 1)]));
185 T1v = VZMUL(T1t, T1u);
186 T1x = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
187 T1y = VZMUL(T1w, T1x);
188 T1R = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
189 T1S = VZMUL(Tf, T1R);
190 T1P = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
191 T1Q = VZMUL(T1O, T1P);
192 {
193 V T1H, T1J, T1A, T1E;
194 T1H = LD(&(x[WS(rs, 27)]), ms, &(x[WS(rs, 1)]));
195 T1I = VZMUL(T7, T1H);
196 T1J = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
197 T1K = VZMUL(T6, T1J);
198 T1L = VSUB(T1I, T1K);
199 T1A = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
200 T1B = VZMUL(T3, T1A);
201 T1E = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
202 T1F = VZMUL(T1D, T1E);
203 T1G = VSUB(T1B, T1F);
204 }
205 }
206 {
207 V T1z, T1M, T38, T39;
208 T1z = VSUB(T1v, T1y);
209 T1M = VADD(T1G, T1L);
210 T1N = VFMA(LDK(KP707106781), T1M, T1z);
211 T2v = VFNMS(LDK(KP707106781), T1M, T1z);
212 T38 = VADD(T1B, T1F);
213 T39 = VADD(T1I, T1K);
214 T3a = VADD(T38, T39);
215 T3G = VSUB(T39, T38);
216 }
217 {
218 V T1T, T1U, T35, T36;
219 T1T = VSUB(T1Q, T1S);
220 T1U = VSUB(T1L, T1G);
221 T1V = VFMA(LDK(KP707106781), T1U, T1T);
222 T2w = VFNMS(LDK(KP707106781), T1U, T1T);
223 T35 = VADD(T1v, T1y);
224 T36 = VADD(T1S, T1Q);
225 T37 = VADD(T35, T36);
226 T3F = VSUB(T35, T36);
227 }
228 }
229 {
230 V T11, T14, T1o, T1l, T1e, T1g, T1h, T17, T1a, T1b;
231 {
232 V T10, T13, T1n, T1k;
233 T10 = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
234 T11 = VZMUL(T2, T10);
235 T13 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
236 T14 = VZMUL(T12, T13);
237 T1n = LD(&(x[WS(rs, 25)]), ms, &(x[WS(rs, 1)]));
238 T1o = VZMUL(T1m, T1n);
239 T1k = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
240 T1l = VZMUL(T5, T1k);
241 {
242 V T1d, T1f, T16, T19;
243 T1d = LD(&(x[WS(rs, 29)]), ms, &(x[WS(rs, 1)]));
244 T1e = VZMUL(T1c, T1d);
245 T1f = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
246 T1g = VZMUL(TK, T1f);
247 T1h = VSUB(T1e, T1g);
248 T16 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
249 T17 = VZMUL(TP, T16);
250 T19 = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
251 T1a = VZMUL(T18, T19);
252 T1b = VSUB(T17, T1a);
253 }
254 }
255 {
256 V T15, T1i, T31, T32;
257 T15 = VSUB(T11, T14);
258 T1i = VADD(T1b, T1h);
259 T1j = VFMA(LDK(KP707106781), T1i, T15);
260 T2s = VFNMS(LDK(KP707106781), T1i, T15);
261 T31 = VADD(T17, T1a);
262 T32 = VADD(T1e, T1g);
263 T33 = VADD(T31, T32);
264 T3D = VSUB(T31, T32);
265 }
266 {
267 V T1p, T1q, T2Y, T2Z;
268 T1p = VSUB(T1l, T1o);
269 T1q = VSUB(T1b, T1h);
270 T1r = VFMA(LDK(KP707106781), T1q, T1p);
271 T2t = VFNMS(LDK(KP707106781), T1q, T1p);
272 T2Y = VADD(T11, T14);
273 T2Z = VADD(T1l, T1o);
274 T30 = VADD(T2Y, T2Z);
275 T3C = VSUB(T2Y, T2Z);
276 }
277 }
278 {
279 V T3r, T3v, T3u, T3w;
280 {
281 V T3p, T3q, T3s, T3t;
282 T3p = VADD(T2T, T2W);
283 T3q = VADD(T3g, T3j);
284 T3r = VSUB(T3p, T3q);
285 T3v = VADD(T3p, T3q);
286 T3s = VADD(T30, T33);
287 T3t = VADD(T37, T3a);
288 T3u = VSUB(T3s, T3t);
289 T3w = VADD(T3s, T3t);
290 }
291 ST(&(x[WS(rs, 24)]), VFNMSI(T3u, T3r), ms, &(x[0]));
292 ST(&(x[0]), VADD(T3v, T3w), ms, &(x[0]));
293 ST(&(x[WS(rs, 8)]), VFMAI(T3u, T3r), ms, &(x[0]));
294 ST(&(x[WS(rs, 16)]), VSUB(T3v, T3w), ms, &(x[0]));
295 }
296 {
297 V T2X, T3k, T3c, T3l, T34, T3b;
298 T2X = VSUB(T2T, T2W);
299 T3k = VSUB(T3g, T3j);
300 T34 = VSUB(T30, T33);
301 T3b = VSUB(T37, T3a);
302 T3c = VADD(T34, T3b);
303 T3l = VSUB(T34, T3b);
304 {
305 V T3d, T3m, T3n, T3o;
306 T3d = VFNMS(LDK(KP707106781), T3c, T2X);
307 T3m = VFNMS(LDK(KP707106781), T3l, T3k);
308 ST(&(x[WS(rs, 12)]), VFNMSI(T3m, T3d), ms, &(x[0]));
309 ST(&(x[WS(rs, 20)]), VFMAI(T3m, T3d), ms, &(x[0]));
310 T3n = VFMA(LDK(KP707106781), T3c, T2X);
311 T3o = VFMA(LDK(KP707106781), T3l, T3k);
312 ST(&(x[WS(rs, 4)]), VFMAI(T3o, T3n), ms, &(x[0]));
313 ST(&(x[WS(rs, 28)]), VFNMSI(T3o, T3n), ms, &(x[0]));
314 }
315 }
316 {
317 V T3B, T3T, T3M, T3W, T3I, T3X, T3P, T3U, T3A, T3L;
318 T3A = VADD(T3y, T3z);
319 T3B = VFMA(LDK(KP707106781), T3A, T3x);
320 T3T = VFNMS(LDK(KP707106781), T3A, T3x);
321 T3L = VSUB(T3y, T3z);
322 T3M = VFMA(LDK(KP707106781), T3L, T3K);
323 T3W = VFNMS(LDK(KP707106781), T3L, T3K);
324 {
325 V T3E, T3H, T3N, T3O;
326 T3E = VFNMS(LDK(KP414213562), T3D, T3C);
327 T3H = VFNMS(LDK(KP414213562), T3G, T3F);
328 T3I = VADD(T3E, T3H);
329 T3X = VSUB(T3E, T3H);
330 T3N = VFMA(LDK(KP414213562), T3C, T3D);
331 T3O = VFMA(LDK(KP414213562), T3F, T3G);
332 T3P = VSUB(T3N, T3O);
333 T3U = VADD(T3N, T3O);
334 }
335 {
336 V T3J, T3Q, T3Z, T40;
337 T3J = VFNMS(LDK(KP923879532), T3I, T3B);
338 T3Q = VFNMS(LDK(KP923879532), T3P, T3M);
339 ST(&(x[WS(rs, 14)]), VFNMSI(T3Q, T3J), ms, &(x[0]));
340 ST(&(x[WS(rs, 18)]), VFMAI(T3Q, T3J), ms, &(x[0]));
341 T3Z = VFMA(LDK(KP923879532), T3U, T3T);
342 T40 = VFNMS(LDK(KP923879532), T3X, T3W);
343 ST(&(x[WS(rs, 6)]), VFNMSI(T40, T3Z), ms, &(x[0]));
344 ST(&(x[WS(rs, 26)]), VFMAI(T40, T3Z), ms, &(x[0]));
345 }
346 {
347 V T3R, T3S, T3V, T3Y;
348 T3R = VFMA(LDK(KP923879532), T3I, T3B);
349 T3S = VFMA(LDK(KP923879532), T3P, T3M);
350 ST(&(x[WS(rs, 30)]), VFNMSI(T3S, T3R), ms, &(x[0]));
351 ST(&(x[WS(rs, 2)]), VFMAI(T3S, T3R), ms, &(x[0]));
352 T3V = VFNMS(LDK(KP923879532), T3U, T3T);
353 T3Y = VFMA(LDK(KP923879532), T3X, T3W);
354 ST(&(x[WS(rs, 10)]), VFMAI(T3Y, T3V), ms, &(x[0]));
355 ST(&(x[WS(rs, 22)]), VFNMSI(T3Y, T3V), ms, &(x[0]));
356 }
357 }
358 {
359 V TZ, T2h, T2d, T2i, T1X, T2l, T2a, T2k;
360 {
361 V Ts, TY, T2b, T2c;
362 Ts = VFMA(LDK(KP707106781), Tr, Tb);
363 TY = VADD(TG, TX);
364 TZ = VFMA(LDK(KP923879532), TY, Ts);
365 T2h = VFNMS(LDK(KP923879532), TY, Ts);
366 T2b = VFMA(LDK(KP198912367), T1j, T1r);
367 T2c = VFMA(LDK(KP198912367), T1N, T1V);
368 T2d = VSUB(T2b, T2c);
369 T2i = VADD(T2b, T2c);
370 }
371 {
372 V T1s, T1W, T26, T29;
373 T1s = VFNMS(LDK(KP198912367), T1r, T1j);
374 T1W = VFNMS(LDK(KP198912367), T1V, T1N);
375 T1X = VADD(T1s, T1W);
376 T2l = VSUB(T1s, T1W);
377 T26 = VFMA(LDK(KP707106781), T25, T24);
378 T29 = VSUB(T27, T28);
379 T2a = VFMA(LDK(KP923879532), T29, T26);
380 T2k = VFNMS(LDK(KP923879532), T29, T26);
381 }
382 {
383 V T1Y, T2e, T2n, T2o;
384 T1Y = VFNMS(LDK(KP980785280), T1X, TZ);
385 T2e = VFNMS(LDK(KP980785280), T2d, T2a);
386 ST(&(x[WS(rs, 15)]), VFNMSI(T2e, T1Y), ms, &(x[WS(rs, 1)]));
387 ST(&(x[WS(rs, 17)]), VFMAI(T2e, T1Y), ms, &(x[WS(rs, 1)]));
388 T2n = VFMA(LDK(KP980785280), T2i, T2h);
389 T2o = VFNMS(LDK(KP980785280), T2l, T2k);
390 ST(&(x[WS(rs, 7)]), VFNMSI(T2o, T2n), ms, &(x[WS(rs, 1)]));
391 ST(&(x[WS(rs, 25)]), VFMAI(T2o, T2n), ms, &(x[WS(rs, 1)]));
392 }
393 {
394 V T2f, T2g, T2j, T2m;
395 T2f = VFMA(LDK(KP980785280), T1X, TZ);
396 T2g = VFMA(LDK(KP980785280), T2d, T2a);
397 ST(&(x[WS(rs, 31)]), VFNMSI(T2g, T2f), ms, &(x[WS(rs, 1)]));
398 ST(&(x[WS(rs, 1)]), VFMAI(T2g, T2f), ms, &(x[WS(rs, 1)]));
399 T2j = VFNMS(LDK(KP980785280), T2i, T2h);
400 T2m = VFMA(LDK(KP980785280), T2l, T2k);
401 ST(&(x[WS(rs, 9)]), VFMAI(T2m, T2j), ms, &(x[WS(rs, 1)]));
402 ST(&(x[WS(rs, 23)]), VFNMSI(T2m, T2j), ms, &(x[WS(rs, 1)]));
403 }
404 }
405 {
406 V T2r, T2J, T2F, T2K, T2y, T2N, T2C, T2M;
407 {
408 V T2p, T2q, T2D, T2E;
409 T2p = VFNMS(LDK(KP707106781), Tr, Tb);
410 T2q = VADD(T27, T28);
411 T2r = VFMA(LDK(KP923879532), T2q, T2p);
412 T2J = VFNMS(LDK(KP923879532), T2q, T2p);
413 T2D = VFNMS(LDK(KP668178637), T2s, T2t);
414 T2E = VFNMS(LDK(KP668178637), T2v, T2w);
415 T2F = VSUB(T2D, T2E);
416 T2K = VADD(T2D, T2E);
417 }
418 {
419 V T2u, T2x, T2A, T2B;
420 T2u = VFMA(LDK(KP668178637), T2t, T2s);
421 T2x = VFMA(LDK(KP668178637), T2w, T2v);
422 T2y = VADD(T2u, T2x);
423 T2N = VSUB(T2u, T2x);
424 T2A = VFNMS(LDK(KP707106781), T25, T24);
425 T2B = VSUB(TG, TX);
426 T2C = VFNMS(LDK(KP923879532), T2B, T2A);
427 T2M = VFMA(LDK(KP923879532), T2B, T2A);
428 }
429 {
430 V T2z, T2G, T2P, T2Q;
431 T2z = VFNMS(LDK(KP831469612), T2y, T2r);
432 T2G = VFNMS(LDK(KP831469612), T2F, T2C);
433 ST(&(x[WS(rs, 19)]), VFNMSI(T2G, T2z), ms, &(x[WS(rs, 1)]));
434 ST(&(x[WS(rs, 13)]), VFMAI(T2G, T2z), ms, &(x[WS(rs, 1)]));
435 T2P = VFNMS(LDK(KP831469612), T2K, T2J);
436 T2Q = VFMA(LDK(KP831469612), T2N, T2M);
437 ST(&(x[WS(rs, 5)]), VFMAI(T2Q, T2P), ms, &(x[WS(rs, 1)]));
438 ST(&(x[WS(rs, 27)]), VFNMSI(T2Q, T2P), ms, &(x[WS(rs, 1)]));
439 }
440 {
441 V T2H, T2I, T2L, T2O;
442 T2H = VFMA(LDK(KP831469612), T2y, T2r);
443 T2I = VFMA(LDK(KP831469612), T2F, T2C);
444 ST(&(x[WS(rs, 3)]), VFNMSI(T2I, T2H), ms, &(x[WS(rs, 1)]));
445 ST(&(x[WS(rs, 29)]), VFMAI(T2I, T2H), ms, &(x[WS(rs, 1)]));
446 T2L = VFMA(LDK(KP831469612), T2K, T2J);
447 T2O = VFNMS(LDK(KP831469612), T2N, T2M);
448 ST(&(x[WS(rs, 11)]), VFNMSI(T2O, T2L), ms, &(x[WS(rs, 1)]));
449 ST(&(x[WS(rs, 21)]), VFMAI(T2O, T2L), ms, &(x[WS(rs, 1)]));
450 }
451 }
452 }
453 }
454 }
455 VLEAVE();
456 }
457
458 static const tw_instr twinstr[] = {
459 VTW(0, 1),
460 VTW(0, 3),
461 VTW(0, 9),
462 VTW(0, 27),
463 {TW_NEXT, VL, 0}
464 };
465
466 static const ct_desc desc = { 32, XSIMD_STRING("t3bv_32"), twinstr, &GENUS, {146, 116, 98, 0}, 0, 0, 0 };
467
468 void XSIMD(codelet_t3bv_32) (planner *p) {
469 X(kdft_dit_register) (p, t3bv_32, &desc);
470 }
471 #else
472
473 /* Generated by: ../../../genfft/gen_twiddle_c.native -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -no-generate-bytw -n 32 -name t3bv_32 -include dft/simd/t3b.h -sign 1 */
474
475 /*
476 * This function contains 244 FP additions, 158 FP multiplications,
477 * (or, 228 additions, 142 multiplications, 16 fused multiply/add),
478 * 90 stack variables, 7 constants, and 64 memory accesses
479 */
480 #include "dft/simd/t3b.h"
481
482 static void t3bv_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
483 {
484 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
485 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
486 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
487 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
488 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
489 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
490 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
491 {
492 INT m;
493 R *x;
494 x = ii;
495 for (m = mb, W = W + (mb * ((TWVL / VL) * 8)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 8), MAKE_VOLATILE_STRIDE(32, rs)) {
496 V T2, T5, T3, T4, Tc, T1v, TH, Tz, Tn, T6, TS, Tf, TK, T7, T8;
497 V Tv, T1I, T25, Tg, Tk, T1N, T1Q, TC, T16, T12, T1w, TL, TP, TT, T1m;
498 V T1f;
499 T2 = LDW(&(W[0]));
500 T5 = LDW(&(W[TWVL * 4]));
501 T3 = LDW(&(W[TWVL * 2]));
502 T4 = VZMULJ(T2, T3);
503 Tc = VZMUL(T2, T3);
504 T1v = VZMULJ(T2, T5);
505 TH = VZMULJ(T3, T5);
506 Tz = VZMUL(T2, T5);
507 Tn = VZMUL(T3, T5);
508 T6 = VZMUL(T4, T5);
509 TS = VZMUL(Tc, T5);
510 Tf = VZMULJ(T4, T5);
511 TK = VZMULJ(Tc, T5);
512 T7 = LDW(&(W[TWVL * 6]));
513 T8 = VZMULJ(T6, T7);
514 Tv = VZMULJ(T5, T7);
515 T1I = VZMULJ(Tc, T7);
516 T25 = VZMULJ(T3, T7);
517 Tg = VZMULJ(Tf, T7);
518 Tk = VZMUL(T2, T7);
519 T1N = VZMUL(Tc, T7);
520 T1Q = VZMULJ(Tn, T7);
521 TC = VZMULJ(T2, T7);
522 T16 = VZMUL(T4, T7);
523 T12 = VZMULJ(TH, T7);
524 T1w = VZMULJ(T1v, T7);
525 TL = VZMULJ(TK, T7);
526 TP = VZMUL(T3, T7);
527 TT = VZMULJ(TS, T7);
528 T1m = VZMULJ(Tz, T7);
529 T1f = VZMULJ(T4, T7);
530 {
531 V Tb, T28, T3k, T3M, Tr, T22, T3f, T3N, TX, T20, T3b, T3J, TG, T1Z, T38;
532 V T3I, T1M, T2v, T33, T3F, T1V, T2w, T30, T3E, T1j, T2s, T2W, T3C, T1r, T2t;
533 V T2T, T3B;
534 {
535 V T1, T27, Ta, T24, T26, T9, T23, T3i, T3j;
536 T1 = LD(&(x[0]), ms, &(x[0]));
537 T26 = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
538 T27 = VZMUL(T25, T26);
539 T9 = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
540 Ta = VZMUL(T8, T9);
541 T23 = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
542 T24 = VZMUL(T1v, T23);
543 Tb = VSUB(T1, Ta);
544 T28 = VSUB(T24, T27);
545 T3i = VADD(T1, Ta);
546 T3j = VADD(T24, T27);
547 T3k = VSUB(T3i, T3j);
548 T3M = VADD(T3i, T3j);
549 }
550 {
551 V Te, Tp, Ti, Tm;
552 {
553 V Td, To, Th, Tl;
554 Td = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
555 Te = VZMUL(Tc, Td);
556 To = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
557 Tp = VZMUL(Tn, To);
558 Th = LD(&(x[WS(rs, 20)]), ms, &(x[0]));
559 Ti = VZMUL(Tg, Th);
560 Tl = LD(&(x[WS(rs, 28)]), ms, &(x[0]));
561 Tm = VZMUL(Tk, Tl);
562 }
563 {
564 V Tj, Tq, T3d, T3e;
565 Tj = VSUB(Te, Ti);
566 Tq = VSUB(Tm, Tp);
567 Tr = VMUL(LDK(KP707106781), VADD(Tj, Tq));
568 T22 = VMUL(LDK(KP707106781), VSUB(Tj, Tq));
569 T3d = VADD(Te, Ti);
570 T3e = VADD(Tm, Tp);
571 T3f = VSUB(T3d, T3e);
572 T3N = VADD(T3d, T3e);
573 }
574 }
575 {
576 V TJ, TV, TN, TR;
577 {
578 V TI, TU, TM, TQ;
579 TI = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
580 TJ = VZMUL(TH, TI);
581 TU = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
582 TV = VZMUL(TT, TU);
583 TM = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
584 TN = VZMUL(TL, TM);
585 TQ = LD(&(x[WS(rs, 30)]), ms, &(x[0]));
586 TR = VZMUL(TP, TQ);
587 }
588 {
589 V TO, TW, T39, T3a;
590 TO = VSUB(TJ, TN);
591 TW = VSUB(TR, TV);
592 TX = VFNMS(LDK(KP382683432), TW, VMUL(LDK(KP923879532), TO));
593 T20 = VFMA(LDK(KP923879532), TW, VMUL(LDK(KP382683432), TO));
594 T39 = VADD(TR, TV);
595 T3a = VADD(TJ, TN);
596 T3b = VSUB(T39, T3a);
597 T3J = VADD(T39, T3a);
598 }
599 }
600 {
601 V Tu, TE, Tx, TB;
602 {
603 V Tt, TD, Tw, TA;
604 Tt = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
605 Tu = VZMUL(T4, Tt);
606 TD = LD(&(x[WS(rs, 26)]), ms, &(x[0]));
607 TE = VZMUL(TC, TD);
608 Tw = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
609 Tx = VZMUL(Tv, Tw);
610 TA = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
611 TB = VZMUL(Tz, TA);
612 }
613 {
614 V Ty, TF, T36, T37;
615 Ty = VSUB(Tu, Tx);
616 TF = VSUB(TB, TE);
617 TG = VFMA(LDK(KP382683432), Ty, VMUL(LDK(KP923879532), TF));
618 T1Z = VFNMS(LDK(KP382683432), TF, VMUL(LDK(KP923879532), Ty));
619 T36 = VADD(Tu, Tx);
620 T37 = VADD(TB, TE);
621 T38 = VSUB(T36, T37);
622 T3I = VADD(T36, T37);
623 }
624 }
625 {
626 V T1H, T1K, T1S, T1P, T1B, T1D, T1E, T1u, T1y, T1z;
627 {
628 V T1G, T1J, T1R, T1O;
629 T1G = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
630 T1H = VZMUL(Tf, T1G);
631 T1J = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
632 T1K = VZMUL(T1I, T1J);
633 T1R = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
634 T1S = VZMUL(T1Q, T1R);
635 T1O = LD(&(x[WS(rs, 31)]), ms, &(x[WS(rs, 1)]));
636 T1P = VZMUL(T1N, T1O);
637 {
638 V T1A, T1C, T1t, T1x;
639 T1A = LD(&(x[WS(rs, 27)]), ms, &(x[WS(rs, 1)]));
640 T1B = VZMUL(T7, T1A);
641 T1C = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
642 T1D = VZMUL(T6, T1C);
643 T1E = VSUB(T1B, T1D);
644 T1t = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
645 T1u = VZMUL(T3, T1t);
646 T1x = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
647 T1y = VZMUL(T1w, T1x);
648 T1z = VSUB(T1u, T1y);
649 }
650 }
651 {
652 V T1F, T1L, T31, T32;
653 T1F = VMUL(LDK(KP707106781), VSUB(T1z, T1E));
654 T1L = VSUB(T1H, T1K);
655 T1M = VSUB(T1F, T1L);
656 T2v = VADD(T1L, T1F);
657 T31 = VADD(T1u, T1y);
658 T32 = VADD(T1B, T1D);
659 T33 = VSUB(T31, T32);
660 T3F = VADD(T31, T32);
661 }
662 {
663 V T1T, T1U, T2Y, T2Z;
664 T1T = VSUB(T1P, T1S);
665 T1U = VMUL(LDK(KP707106781), VADD(T1z, T1E));
666 T1V = VSUB(T1T, T1U);
667 T2w = VADD(T1T, T1U);
668 T2Y = VADD(T1P, T1S);
669 T2Z = VADD(T1H, T1K);
670 T30 = VSUB(T2Y, T2Z);
671 T3E = VADD(T2Y, T2Z);
672 }
673 }
674 {
675 V T1e, T1h, T1o, T1l, T18, T1a, T1b, T11, T14, T15;
676 {
677 V T1d, T1g, T1n, T1k;
678 T1d = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
679 T1e = VZMUL(T5, T1d);
680 T1g = LD(&(x[WS(rs, 25)]), ms, &(x[WS(rs, 1)]));
681 T1h = VZMUL(T1f, T1g);
682 T1n = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
683 T1o = VZMUL(T1m, T1n);
684 T1k = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
685 T1l = VZMUL(T2, T1k);
686 {
687 V T17, T19, T10, T13;
688 T17 = LD(&(x[WS(rs, 29)]), ms, &(x[WS(rs, 1)]));
689 T18 = VZMUL(T16, T17);
690 T19 = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
691 T1a = VZMUL(TS, T19);
692 T1b = VSUB(T18, T1a);
693 T10 = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
694 T11 = VZMUL(TK, T10);
695 T13 = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
696 T14 = VZMUL(T12, T13);
697 T15 = VSUB(T11, T14);
698 }
699 }
700 {
701 V T1c, T1i, T2U, T2V;
702 T1c = VMUL(LDK(KP707106781), VSUB(T15, T1b));
703 T1i = VSUB(T1e, T1h);
704 T1j = VSUB(T1c, T1i);
705 T2s = VADD(T1i, T1c);
706 T2U = VADD(T11, T14);
707 T2V = VADD(T18, T1a);
708 T2W = VSUB(T2U, T2V);
709 T3C = VADD(T2U, T2V);
710 }
711 {
712 V T1p, T1q, T2R, T2S;
713 T1p = VSUB(T1l, T1o);
714 T1q = VMUL(LDK(KP707106781), VADD(T15, T1b));
715 T1r = VSUB(T1p, T1q);
716 T2t = VADD(T1p, T1q);
717 T2R = VADD(T1l, T1o);
718 T2S = VADD(T1e, T1h);
719 T2T = VSUB(T2R, T2S);
720 T3B = VADD(T2R, T2S);
721 }
722 }
723 {
724 V T3V, T3Z, T3Y, T40;
725 {
726 V T3T, T3U, T3W, T3X;
727 T3T = VADD(T3M, T3N);
728 T3U = VADD(T3I, T3J);
729 T3V = VSUB(T3T, T3U);
730 T3Z = VADD(T3T, T3U);
731 T3W = VADD(T3B, T3C);
732 T3X = VADD(T3E, T3F);
733 T3Y = VBYI(VSUB(T3W, T3X));
734 T40 = VADD(T3W, T3X);
735 }
736 ST(&(x[WS(rs, 24)]), VSUB(T3V, T3Y), ms, &(x[0]));
737 ST(&(x[0]), VADD(T3Z, T40), ms, &(x[0]));
738 ST(&(x[WS(rs, 8)]), VADD(T3V, T3Y), ms, &(x[0]));
739 ST(&(x[WS(rs, 16)]), VSUB(T3Z, T40), ms, &(x[0]));
740 }
741 {
742 V T3K, T3O, T3H, T3P, T3D, T3G;
743 T3K = VSUB(T3I, T3J);
744 T3O = VSUB(T3M, T3N);
745 T3D = VSUB(T3B, T3C);
746 T3G = VSUB(T3E, T3F);
747 T3H = VMUL(LDK(KP707106781), VSUB(T3D, T3G));
748 T3P = VMUL(LDK(KP707106781), VADD(T3D, T3G));
749 {
750 V T3L, T3Q, T3R, T3S;
751 T3L = VBYI(VSUB(T3H, T3K));
752 T3Q = VSUB(T3O, T3P);
753 ST(&(x[WS(rs, 12)]), VADD(T3L, T3Q), ms, &(x[0]));
754 ST(&(x[WS(rs, 20)]), VSUB(T3Q, T3L), ms, &(x[0]));
755 T3R = VBYI(VADD(T3K, T3H));
756 T3S = VADD(T3O, T3P);
757 ST(&(x[WS(rs, 4)]), VADD(T3R, T3S), ms, &(x[0]));
758 ST(&(x[WS(rs, 28)]), VSUB(T3S, T3R), ms, &(x[0]));
759 }
760 }
761 {
762 V T3g, T3w, T3m, T3t, T35, T3u, T3p, T3x, T3c, T3l;
763 T3c = VMUL(LDK(KP707106781), VSUB(T38, T3b));
764 T3g = VSUB(T3c, T3f);
765 T3w = VADD(T3f, T3c);
766 T3l = VMUL(LDK(KP707106781), VADD(T38, T3b));
767 T3m = VSUB(T3k, T3l);
768 T3t = VADD(T3k, T3l);
769 {
770 V T2X, T34, T3n, T3o;
771 T2X = VFNMS(LDK(KP382683432), T2W, VMUL(LDK(KP923879532), T2T));
772 T34 = VFMA(LDK(KP923879532), T30, VMUL(LDK(KP382683432), T33));
773 T35 = VSUB(T2X, T34);
774 T3u = VADD(T2X, T34);
775 T3n = VFMA(LDK(KP382683432), T2T, VMUL(LDK(KP923879532), T2W));
776 T3o = VFNMS(LDK(KP382683432), T30, VMUL(LDK(KP923879532), T33));
777 T3p = VSUB(T3n, T3o);
778 T3x = VADD(T3n, T3o);
779 }
780 {
781 V T3h, T3q, T3z, T3A;
782 T3h = VBYI(VSUB(T35, T3g));
783 T3q = VSUB(T3m, T3p);
784 ST(&(x[WS(rs, 10)]), VADD(T3h, T3q), ms, &(x[0]));
785 ST(&(x[WS(rs, 22)]), VSUB(T3q, T3h), ms, &(x[0]));
786 T3z = VSUB(T3t, T3u);
787 T3A = VBYI(VSUB(T3x, T3w));
788 ST(&(x[WS(rs, 18)]), VSUB(T3z, T3A), ms, &(x[0]));
789 ST(&(x[WS(rs, 14)]), VADD(T3z, T3A), ms, &(x[0]));
790 }
791 {
792 V T3r, T3s, T3v, T3y;
793 T3r = VBYI(VADD(T3g, T35));
794 T3s = VADD(T3m, T3p);
795 ST(&(x[WS(rs, 6)]), VADD(T3r, T3s), ms, &(x[0]));
796 ST(&(x[WS(rs, 26)]), VSUB(T3s, T3r), ms, &(x[0]));
797 T3v = VADD(T3t, T3u);
798 T3y = VBYI(VADD(T3w, T3x));
799 ST(&(x[WS(rs, 30)]), VSUB(T3v, T3y), ms, &(x[0]));
800 ST(&(x[WS(rs, 2)]), VADD(T3v, T3y), ms, &(x[0]));
801 }
802 }
803 {
804 V TZ, T2k, T2d, T2l, T1X, T2h, T2a, T2i;
805 {
806 V Ts, TY, T2b, T2c;
807 Ts = VSUB(Tb, Tr);
808 TY = VSUB(TG, TX);
809 TZ = VSUB(Ts, TY);
810 T2k = VADD(Ts, TY);
811 T2b = VFNMS(LDK(KP555570233), T1j, VMUL(LDK(KP831469612), T1r));
812 T2c = VFMA(LDK(KP555570233), T1M, VMUL(LDK(KP831469612), T1V));
813 T2d = VSUB(T2b, T2c);
814 T2l = VADD(T2b, T2c);
815 }
816 {
817 V T1s, T1W, T21, T29;
818 T1s = VFMA(LDK(KP831469612), T1j, VMUL(LDK(KP555570233), T1r));
819 T1W = VFNMS(LDK(KP555570233), T1V, VMUL(LDK(KP831469612), T1M));
820 T1X = VSUB(T1s, T1W);
821 T2h = VADD(T1s, T1W);
822 T21 = VSUB(T1Z, T20);
823 T29 = VSUB(T22, T28);
824 T2a = VSUB(T21, T29);
825 T2i = VADD(T29, T21);
826 }
827 {
828 V T1Y, T2e, T2n, T2o;
829 T1Y = VADD(TZ, T1X);
830 T2e = VBYI(VADD(T2a, T2d));
831 ST(&(x[WS(rs, 27)]), VSUB(T1Y, T2e), ms, &(x[WS(rs, 1)]));
832 ST(&(x[WS(rs, 5)]), VADD(T1Y, T2e), ms, &(x[WS(rs, 1)]));
833 T2n = VBYI(VADD(T2i, T2h));
834 T2o = VADD(T2k, T2l);
835 ST(&(x[WS(rs, 3)]), VADD(T2n, T2o), ms, &(x[WS(rs, 1)]));
836 ST(&(x[WS(rs, 29)]), VSUB(T2o, T2n), ms, &(x[WS(rs, 1)]));
837 }
838 {
839 V T2f, T2g, T2j, T2m;
840 T2f = VSUB(TZ, T1X);
841 T2g = VBYI(VSUB(T2d, T2a));
842 ST(&(x[WS(rs, 21)]), VSUB(T2f, T2g), ms, &(x[WS(rs, 1)]));
843 ST(&(x[WS(rs, 11)]), VADD(T2f, T2g), ms, &(x[WS(rs, 1)]));
844 T2j = VBYI(VSUB(T2h, T2i));
845 T2m = VSUB(T2k, T2l);
846 ST(&(x[WS(rs, 13)]), VADD(T2j, T2m), ms, &(x[WS(rs, 1)]));
847 ST(&(x[WS(rs, 19)]), VSUB(T2m, T2j), ms, &(x[WS(rs, 1)]));
848 }
849 }
850 {
851 V T2r, T2M, T2F, T2N, T2y, T2J, T2C, T2K;
852 {
853 V T2p, T2q, T2D, T2E;
854 T2p = VADD(Tb, Tr);
855 T2q = VADD(T1Z, T20);
856 T2r = VSUB(T2p, T2q);
857 T2M = VADD(T2p, T2q);
858 T2D = VFNMS(LDK(KP195090322), T2s, VMUL(LDK(KP980785280), T2t));
859 T2E = VFMA(LDK(KP195090322), T2v, VMUL(LDK(KP980785280), T2w));
860 T2F = VSUB(T2D, T2E);
861 T2N = VADD(T2D, T2E);
862 }
863 {
864 V T2u, T2x, T2A, T2B;
865 T2u = VFMA(LDK(KP980785280), T2s, VMUL(LDK(KP195090322), T2t));
866 T2x = VFNMS(LDK(KP195090322), T2w, VMUL(LDK(KP980785280), T2v));
867 T2y = VSUB(T2u, T2x);
868 T2J = VADD(T2u, T2x);
869 T2A = VADD(TG, TX);
870 T2B = VADD(T28, T22);
871 T2C = VSUB(T2A, T2B);
872 T2K = VADD(T2B, T2A);
873 }
874 {
875 V T2z, T2G, T2P, T2Q;
876 T2z = VADD(T2r, T2y);
877 T2G = VBYI(VADD(T2C, T2F));
878 ST(&(x[WS(rs, 25)]), VSUB(T2z, T2G), ms, &(x[WS(rs, 1)]));
879 ST(&(x[WS(rs, 7)]), VADD(T2z, T2G), ms, &(x[WS(rs, 1)]));
880 T2P = VBYI(VADD(T2K, T2J));
881 T2Q = VADD(T2M, T2N);
882 ST(&(x[WS(rs, 1)]), VADD(T2P, T2Q), ms, &(x[WS(rs, 1)]));
883 ST(&(x[WS(rs, 31)]), VSUB(T2Q, T2P), ms, &(x[WS(rs, 1)]));
884 }
885 {
886 V T2H, T2I, T2L, T2O;
887 T2H = VSUB(T2r, T2y);
888 T2I = VBYI(VSUB(T2F, T2C));
889 ST(&(x[WS(rs, 23)]), VSUB(T2H, T2I), ms, &(x[WS(rs, 1)]));
890 ST(&(x[WS(rs, 9)]), VADD(T2H, T2I), ms, &(x[WS(rs, 1)]));
891 T2L = VBYI(VSUB(T2J, T2K));
892 T2O = VSUB(T2M, T2N);
893 ST(&(x[WS(rs, 15)]), VADD(T2L, T2O), ms, &(x[WS(rs, 1)]));
894 ST(&(x[WS(rs, 17)]), VSUB(T2O, T2L), ms, &(x[WS(rs, 1)]));
895 }
896 }
897 }
898 }
899 }
900 VLEAVE();
901 }
902
903 static const tw_instr twinstr[] = {
904 VTW(0, 1),
905 VTW(0, 3),
906 VTW(0, 9),
907 VTW(0, 27),
908 {TW_NEXT, VL, 0}
909 };
910
911 static const ct_desc desc = { 32, XSIMD_STRING("t3bv_32"), twinstr, &GENUS, {228, 142, 16, 0}, 0, 0, 0 };
912
913 void XSIMD(codelet_t3bv_32) (planner *p) {
914 X(kdft_dit_register) (p, t3bv_32, &desc);
915 }
916 #endif