comparison src/fftw-3.3.8/dft/simd/common/n2fv_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 08:05:08 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_notw_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n2fv_32 -with-ostride 2 -include dft/simd/n2f.h -store-multiple 2 */
29
30 /*
31 * This function contains 186 FP additions, 98 FP multiplications,
32 * (or, 88 additions, 0 multiplications, 98 fused multiply/add),
33 * 72 stack variables, 7 constants, and 80 memory accesses
34 */
35 #include "dft/simd/n2f.h"
36
37 static void n2fv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38 {
39 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
40 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
41 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
42 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
44 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
45 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
46 {
47 INT i;
48 const R *xi;
49 R *xo;
50 xi = ri;
51 xo = ro;
52 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(64, is), MAKE_VOLATILE_STRIDE(64, os)) {
53 V T1T, T1W, T2K, T2x, T16, T1A, Tb, T1p, TT, T1v, TY, T1w, T27, T2a, T2b;
54 V T2H, T2N, TC, T1s, TH, T1t, T20, T23, T24, T2E, T2O, T2g, T2j, Tq, T1B;
55 V T19, T1q, T2A, T2L;
56 {
57 V T3, T1R, T14, T1S, T6, T1U, T9, T1V, T15, Ta;
58 {
59 V T1, T2, T12, T13;
60 T1 = LD(&(xi[0]), ivs, &(xi[0]));
61 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
62 T3 = VSUB(T1, T2);
63 T1R = VADD(T1, T2);
64 T12 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
65 T13 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
66 T14 = VSUB(T12, T13);
67 T1S = VADD(T12, T13);
68 }
69 {
70 V T4, T5, T7, T8;
71 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
72 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
73 T6 = VSUB(T4, T5);
74 T1U = VADD(T4, T5);
75 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
76 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
77 T9 = VSUB(T7, T8);
78 T1V = VADD(T7, T8);
79 }
80 T1T = VADD(T1R, T1S);
81 T1W = VADD(T1U, T1V);
82 T2K = VSUB(T1V, T1U);
83 T2x = VSUB(T1R, T1S);
84 T15 = VSUB(T9, T6);
85 T16 = VFNMS(LDK(KP707106781), T15, T14);
86 T1A = VFMA(LDK(KP707106781), T15, T14);
87 Ta = VADD(T6, T9);
88 Tb = VFMA(LDK(KP707106781), Ta, T3);
89 T1p = VFNMS(LDK(KP707106781), Ta, T3);
90 }
91 {
92 V TL, T25, TW, T26, TO, T28, TR, T29;
93 {
94 V TJ, TK, TU, TV;
95 TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
96 TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
97 TL = VSUB(TJ, TK);
98 T25 = VADD(TJ, TK);
99 TU = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
100 TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
101 TW = VSUB(TU, TV);
102 T26 = VADD(TV, TU);
103 }
104 {
105 V TM, TN, TP, TQ;
106 TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
107 TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
108 TO = VSUB(TM, TN);
109 T28 = VADD(TM, TN);
110 TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
111 TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
112 TR = VSUB(TP, TQ);
113 T29 = VADD(TP, TQ);
114 }
115 {
116 V TS, TX, T2F, T2G;
117 TS = VADD(TO, TR);
118 TT = VFMA(LDK(KP707106781), TS, TL);
119 T1v = VFNMS(LDK(KP707106781), TS, TL);
120 TX = VSUB(TR, TO);
121 TY = VFMA(LDK(KP707106781), TX, TW);
122 T1w = VFNMS(LDK(KP707106781), TX, TW);
123 T27 = VADD(T25, T26);
124 T2a = VADD(T28, T29);
125 T2b = VSUB(T27, T2a);
126 T2F = VSUB(T25, T26);
127 T2G = VSUB(T29, T28);
128 T2H = VFNMS(LDK(KP414213562), T2G, T2F);
129 T2N = VFMA(LDK(KP414213562), T2F, T2G);
130 }
131 }
132 {
133 V Tu, T1Y, TF, T1Z, Tx, T21, TA, T22;
134 {
135 V Ts, Tt, TD, TE;
136 Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
137 Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
138 Tu = VSUB(Ts, Tt);
139 T1Y = VADD(Ts, Tt);
140 TD = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
141 TE = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
142 TF = VSUB(TD, TE);
143 T1Z = VADD(TD, TE);
144 }
145 {
146 V Tv, Tw, Ty, Tz;
147 Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
148 Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
149 Tx = VSUB(Tv, Tw);
150 T21 = VADD(Tv, Tw);
151 Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
152 Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
153 TA = VSUB(Ty, Tz);
154 T22 = VADD(Ty, Tz);
155 }
156 {
157 V TB, TG, T2C, T2D;
158 TB = VADD(Tx, TA);
159 TC = VFMA(LDK(KP707106781), TB, Tu);
160 T1s = VFNMS(LDK(KP707106781), TB, Tu);
161 TG = VSUB(Tx, TA);
162 TH = VFMA(LDK(KP707106781), TG, TF);
163 T1t = VFNMS(LDK(KP707106781), TG, TF);
164 T20 = VADD(T1Y, T1Z);
165 T23 = VADD(T21, T22);
166 T24 = VSUB(T20, T23);
167 T2C = VSUB(T1Y, T1Z);
168 T2D = VSUB(T21, T22);
169 T2E = VFNMS(LDK(KP414213562), T2D, T2C);
170 T2O = VFMA(LDK(KP414213562), T2C, T2D);
171 }
172 }
173 {
174 V Te, T2h, To, T2f, Th, T2i, Tl, T2e, Ti, Tp;
175 {
176 V Tc, Td, Tm, Tn;
177 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
178 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
179 Te = VSUB(Tc, Td);
180 T2h = VADD(Tc, Td);
181 Tm = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
182 Tn = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
183 To = VSUB(Tm, Tn);
184 T2f = VADD(Tn, Tm);
185 }
186 {
187 V Tf, Tg, Tj, Tk;
188 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
189 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
190 Th = VSUB(Tf, Tg);
191 T2i = VADD(Tf, Tg);
192 Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
193 Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
194 Tl = VSUB(Tj, Tk);
195 T2e = VADD(Tj, Tk);
196 }
197 T2g = VADD(T2e, T2f);
198 T2j = VADD(T2h, T2i);
199 Ti = VFNMS(LDK(KP414213562), Th, Te);
200 Tp = VFNMS(LDK(KP414213562), To, Tl);
201 Tq = VADD(Ti, Tp);
202 T1B = VSUB(Tp, Ti);
203 {
204 V T17, T18, T2y, T2z;
205 T17 = VFMA(LDK(KP414213562), Te, Th);
206 T18 = VFMA(LDK(KP414213562), Tl, To);
207 T19 = VSUB(T17, T18);
208 T1q = VADD(T17, T18);
209 T2y = VSUB(T2h, T2i);
210 T2z = VSUB(T2e, T2f);
211 T2A = VADD(T2y, T2z);
212 T2L = VSUB(T2z, T2y);
213 }
214 }
215 {
216 V T31, T32, T33, T34, T35, T36, T37, T38, T39, T3a, T3b, T3c;
217 {
218 V T2d, T2n, T2m, T2o;
219 {
220 V T1X, T2c, T2k, T2l;
221 T1X = VSUB(T1T, T1W);
222 T2c = VADD(T24, T2b);
223 T2d = VFNMS(LDK(KP707106781), T2c, T1X);
224 T2n = VFMA(LDK(KP707106781), T2c, T1X);
225 T2k = VSUB(T2g, T2j);
226 T2l = VSUB(T2b, T24);
227 T2m = VFNMS(LDK(KP707106781), T2l, T2k);
228 T2o = VFMA(LDK(KP707106781), T2l, T2k);
229 }
230 T31 = VFNMSI(T2m, T2d);
231 STM2(&(xo[24]), T31, ovs, &(xo[0]));
232 T32 = VFMAI(T2o, T2n);
233 STM2(&(xo[8]), T32, ovs, &(xo[0]));
234 T33 = VFMAI(T2m, T2d);
235 STM2(&(xo[40]), T33, ovs, &(xo[0]));
236 T34 = VFNMSI(T2o, T2n);
237 STM2(&(xo[56]), T34, ovs, &(xo[0]));
238 }
239 {
240 V T2r, T2v, T2u, T2w;
241 {
242 V T2p, T2q, T2s, T2t;
243 T2p = VADD(T1T, T1W);
244 T2q = VADD(T2j, T2g);
245 T2r = VADD(T2p, T2q);
246 T2v = VSUB(T2p, T2q);
247 T2s = VADD(T20, T23);
248 T2t = VADD(T27, T2a);
249 T2u = VADD(T2s, T2t);
250 T2w = VSUB(T2t, T2s);
251 }
252 T35 = VSUB(T2r, T2u);
253 STM2(&(xo[32]), T35, ovs, &(xo[0]));
254 T36 = VFMAI(T2w, T2v);
255 STM2(&(xo[16]), T36, ovs, &(xo[0]));
256 T37 = VADD(T2r, T2u);
257 STM2(&(xo[0]), T37, ovs, &(xo[0]));
258 T38 = VFNMSI(T2w, T2v);
259 STM2(&(xo[48]), T38, ovs, &(xo[0]));
260 }
261 {
262 V T2V, T2Z, T2Y, T30;
263 {
264 V T2T, T2U, T2W, T2X;
265 T2T = VFNMS(LDK(KP707106781), T2A, T2x);
266 T2U = VADD(T2O, T2N);
267 T2V = VFNMS(LDK(KP923879532), T2U, T2T);
268 T2Z = VFMA(LDK(KP923879532), T2U, T2T);
269 T2W = VFNMS(LDK(KP707106781), T2L, T2K);
270 T2X = VSUB(T2H, T2E);
271 T2Y = VFMA(LDK(KP923879532), T2X, T2W);
272 T30 = VFNMS(LDK(KP923879532), T2X, T2W);
273 }
274 T39 = VFMAI(T2Y, T2V);
275 STM2(&(xo[20]), T39, ovs, &(xo[0]));
276 T3a = VFMAI(T30, T2Z);
277 STM2(&(xo[52]), T3a, ovs, &(xo[0]));
278 T3b = VFNMSI(T2Y, T2V);
279 STM2(&(xo[44]), T3b, ovs, &(xo[0]));
280 T3c = VFNMSI(T30, T2Z);
281 STM2(&(xo[12]), T3c, ovs, &(xo[0]));
282 }
283 {
284 V T3d, T3e, T3f, T3g;
285 {
286 V T2J, T2R, T2Q, T2S;
287 {
288 V T2B, T2I, T2M, T2P;
289 T2B = VFMA(LDK(KP707106781), T2A, T2x);
290 T2I = VADD(T2E, T2H);
291 T2J = VFNMS(LDK(KP923879532), T2I, T2B);
292 T2R = VFMA(LDK(KP923879532), T2I, T2B);
293 T2M = VFMA(LDK(KP707106781), T2L, T2K);
294 T2P = VSUB(T2N, T2O);
295 T2Q = VFNMS(LDK(KP923879532), T2P, T2M);
296 T2S = VFMA(LDK(KP923879532), T2P, T2M);
297 }
298 T3d = VFNMSI(T2Q, T2J);
299 STM2(&(xo[28]), T3d, ovs, &(xo[0]));
300 T3e = VFMAI(T2S, T2R);
301 STM2(&(xo[4]), T3e, ovs, &(xo[0]));
302 T3f = VFMAI(T2Q, T2J);
303 STM2(&(xo[36]), T3f, ovs, &(xo[0]));
304 T3g = VFNMSI(T2S, T2R);
305 STM2(&(xo[60]), T3g, ovs, &(xo[0]));
306 }
307 {
308 V T1r, T1C, T1M, T1J, T1F, T1K, T1y, T1N;
309 T1r = VFMA(LDK(KP923879532), T1q, T1p);
310 T1C = VFMA(LDK(KP923879532), T1B, T1A);
311 T1M = VFNMS(LDK(KP923879532), T1B, T1A);
312 T1J = VFNMS(LDK(KP923879532), T1q, T1p);
313 {
314 V T1D, T1E, T1u, T1x;
315 T1D = VFNMS(LDK(KP668178637), T1s, T1t);
316 T1E = VFNMS(LDK(KP668178637), T1v, T1w);
317 T1F = VSUB(T1D, T1E);
318 T1K = VADD(T1D, T1E);
319 T1u = VFMA(LDK(KP668178637), T1t, T1s);
320 T1x = VFMA(LDK(KP668178637), T1w, T1v);
321 T1y = VADD(T1u, T1x);
322 T1N = VSUB(T1x, T1u);
323 }
324 {
325 V T1z, T1G, T3h, T3i;
326 T1z = VFNMS(LDK(KP831469612), T1y, T1r);
327 T1G = VFNMS(LDK(KP831469612), T1F, T1C);
328 T3h = VFNMSI(T1G, T1z);
329 STM2(&(xo[26]), T3h, ovs, &(xo[2]));
330 STN2(&(xo[24]), T31, T3h, ovs);
331 T3i = VFMAI(T1G, T1z);
332 STM2(&(xo[38]), T3i, ovs, &(xo[2]));
333 STN2(&(xo[36]), T3f, T3i, ovs);
334 }
335 {
336 V T1P, T1Q, T3j, T3k;
337 T1P = VFNMS(LDK(KP831469612), T1K, T1J);
338 T1Q = VFNMS(LDK(KP831469612), T1N, T1M);
339 T3j = VFNMSI(T1Q, T1P);
340 STM2(&(xo[10]), T3j, ovs, &(xo[2]));
341 STN2(&(xo[8]), T32, T3j, ovs);
342 T3k = VFMAI(T1Q, T1P);
343 STM2(&(xo[54]), T3k, ovs, &(xo[2]));
344 STN2(&(xo[52]), T3a, T3k, ovs);
345 }
346 {
347 V T1H, T1I, T3l, T3m;
348 T1H = VFMA(LDK(KP831469612), T1y, T1r);
349 T1I = VFMA(LDK(KP831469612), T1F, T1C);
350 T3l = VFNMSI(T1I, T1H);
351 STM2(&(xo[58]), T3l, ovs, &(xo[2]));
352 STN2(&(xo[56]), T34, T3l, ovs);
353 T3m = VFMAI(T1I, T1H);
354 STM2(&(xo[6]), T3m, ovs, &(xo[2]));
355 STN2(&(xo[4]), T3e, T3m, ovs);
356 }
357 {
358 V T1L, T1O, T3n, T3o;
359 T1L = VFMA(LDK(KP831469612), T1K, T1J);
360 T1O = VFMA(LDK(KP831469612), T1N, T1M);
361 T3n = VFMAI(T1O, T1L);
362 STM2(&(xo[22]), T3n, ovs, &(xo[2]));
363 STN2(&(xo[20]), T39, T3n, ovs);
364 T3o = VFNMSI(T1O, T1L);
365 STM2(&(xo[42]), T3o, ovs, &(xo[2]));
366 STN2(&(xo[40]), T33, T3o, ovs);
367 }
368 }
369 {
370 V Tr, T1a, T1k, T1h, T1d, T1i, T10, T1l;
371 Tr = VFMA(LDK(KP923879532), Tq, Tb);
372 T1a = VFMA(LDK(KP923879532), T19, T16);
373 T1k = VFNMS(LDK(KP923879532), T19, T16);
374 T1h = VFNMS(LDK(KP923879532), Tq, Tb);
375 {
376 V T1b, T1c, TI, TZ;
377 T1b = VFMA(LDK(KP198912367), TC, TH);
378 T1c = VFMA(LDK(KP198912367), TT, TY);
379 T1d = VSUB(T1b, T1c);
380 T1i = VADD(T1b, T1c);
381 TI = VFNMS(LDK(KP198912367), TH, TC);
382 TZ = VFNMS(LDK(KP198912367), TY, TT);
383 T10 = VADD(TI, TZ);
384 T1l = VSUB(TZ, TI);
385 }
386 {
387 V T11, T1e, T3p, T3q;
388 T11 = VFNMS(LDK(KP980785280), T10, Tr);
389 T1e = VFNMS(LDK(KP980785280), T1d, T1a);
390 T3p = VFNMSI(T1e, T11);
391 STM2(&(xo[34]), T3p, ovs, &(xo[2]));
392 STN2(&(xo[32]), T35, T3p, ovs);
393 T3q = VFMAI(T1e, T11);
394 STM2(&(xo[30]), T3q, ovs, &(xo[2]));
395 STN2(&(xo[28]), T3d, T3q, ovs);
396 }
397 {
398 V T1n, T1o, T3r, T3s;
399 T1n = VFMA(LDK(KP980785280), T1i, T1h);
400 T1o = VFMA(LDK(KP980785280), T1l, T1k);
401 T3r = VFMAI(T1o, T1n);
402 STM2(&(xo[14]), T3r, ovs, &(xo[2]));
403 STN2(&(xo[12]), T3c, T3r, ovs);
404 T3s = VFNMSI(T1o, T1n);
405 STM2(&(xo[50]), T3s, ovs, &(xo[2]));
406 STN2(&(xo[48]), T38, T3s, ovs);
407 }
408 {
409 V T1f, T1g, T3t, T3u;
410 T1f = VFMA(LDK(KP980785280), T10, Tr);
411 T1g = VFMA(LDK(KP980785280), T1d, T1a);
412 T3t = VFNMSI(T1g, T1f);
413 STM2(&(xo[2]), T3t, ovs, &(xo[2]));
414 STN2(&(xo[0]), T37, T3t, ovs);
415 T3u = VFMAI(T1g, T1f);
416 STM2(&(xo[62]), T3u, ovs, &(xo[2]));
417 STN2(&(xo[60]), T3g, T3u, ovs);
418 }
419 {
420 V T1j, T1m, T3v, T3w;
421 T1j = VFNMS(LDK(KP980785280), T1i, T1h);
422 T1m = VFNMS(LDK(KP980785280), T1l, T1k);
423 T3v = VFNMSI(T1m, T1j);
424 STM2(&(xo[18]), T3v, ovs, &(xo[2]));
425 STN2(&(xo[16]), T36, T3v, ovs);
426 T3w = VFMAI(T1m, T1j);
427 STM2(&(xo[46]), T3w, ovs, &(xo[2]));
428 STN2(&(xo[44]), T3b, T3w, ovs);
429 }
430 }
431 }
432 }
433 }
434 }
435 VLEAVE();
436 }
437
438 static const kdft_desc desc = { 32, XSIMD_STRING("n2fv_32"), {88, 0, 98, 0}, &GENUS, 0, 2, 0, 0 };
439
440 void XSIMD(codelet_n2fv_32) (planner *p) {
441 X(kdft_register) (p, n2fv_32, &desc);
442 }
443
444 #else
445
446 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n2fv_32 -with-ostride 2 -include dft/simd/n2f.h -store-multiple 2 */
447
448 /*
449 * This function contains 186 FP additions, 42 FP multiplications,
450 * (or, 170 additions, 26 multiplications, 16 fused multiply/add),
451 * 72 stack variables, 7 constants, and 80 memory accesses
452 */
453 #include "dft/simd/n2f.h"
454
455 static void n2fv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
456 {
457 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
458 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
459 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
460 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
461 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
462 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
463 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
464 {
465 INT i;
466 const R *xi;
467 R *xo;
468 xi = ri;
469 xo = ro;
470 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(64, is), MAKE_VOLATILE_STRIDE(64, os)) {
471 V T1T, T1W, T2K, T2x, T16, T1A, Tb, T1p, TT, T1v, TY, T1w, T27, T2a, T2b;
472 V T2H, T2O, TC, T1s, TH, T1t, T20, T23, T24, T2E, T2N, T2g, T2j, Tq, T1B;
473 V T19, T1q, T2A, T2L;
474 {
475 V T3, T1R, T15, T1S, T6, T1U, T9, T1V, T12, Ta;
476 {
477 V T1, T2, T13, T14;
478 T1 = LD(&(xi[0]), ivs, &(xi[0]));
479 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
480 T3 = VSUB(T1, T2);
481 T1R = VADD(T1, T2);
482 T13 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
483 T14 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
484 T15 = VSUB(T13, T14);
485 T1S = VADD(T13, T14);
486 }
487 {
488 V T4, T5, T7, T8;
489 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
490 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
491 T6 = VSUB(T4, T5);
492 T1U = VADD(T4, T5);
493 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
494 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
495 T9 = VSUB(T7, T8);
496 T1V = VADD(T7, T8);
497 }
498 T1T = VADD(T1R, T1S);
499 T1W = VADD(T1U, T1V);
500 T2K = VSUB(T1V, T1U);
501 T2x = VSUB(T1R, T1S);
502 T12 = VMUL(LDK(KP707106781), VSUB(T9, T6));
503 T16 = VSUB(T12, T15);
504 T1A = VADD(T15, T12);
505 Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
506 Tb = VADD(T3, Ta);
507 T1p = VSUB(T3, Ta);
508 }
509 {
510 V TL, T25, TX, T26, TO, T28, TR, T29;
511 {
512 V TJ, TK, TV, TW;
513 TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
514 TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
515 TL = VSUB(TJ, TK);
516 T25 = VADD(TJ, TK);
517 TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
518 TW = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
519 TX = VSUB(TV, TW);
520 T26 = VADD(TV, TW);
521 }
522 {
523 V TM, TN, TP, TQ;
524 TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
525 TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
526 TO = VSUB(TM, TN);
527 T28 = VADD(TM, TN);
528 TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
529 TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
530 TR = VSUB(TP, TQ);
531 T29 = VADD(TP, TQ);
532 }
533 {
534 V TS, TU, T2F, T2G;
535 TS = VMUL(LDK(KP707106781), VADD(TO, TR));
536 TT = VADD(TL, TS);
537 T1v = VSUB(TL, TS);
538 TU = VMUL(LDK(KP707106781), VSUB(TR, TO));
539 TY = VSUB(TU, TX);
540 T1w = VADD(TX, TU);
541 T27 = VADD(T25, T26);
542 T2a = VADD(T28, T29);
543 T2b = VSUB(T27, T2a);
544 T2F = VSUB(T25, T26);
545 T2G = VSUB(T29, T28);
546 T2H = VFNMS(LDK(KP382683432), T2G, VMUL(LDK(KP923879532), T2F));
547 T2O = VFMA(LDK(KP382683432), T2F, VMUL(LDK(KP923879532), T2G));
548 }
549 }
550 {
551 V Tu, T1Y, TG, T1Z, Tx, T21, TA, T22;
552 {
553 V Ts, Tt, TE, TF;
554 Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
555 Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
556 Tu = VSUB(Ts, Tt);
557 T1Y = VADD(Ts, Tt);
558 TE = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
559 TF = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
560 TG = VSUB(TE, TF);
561 T1Z = VADD(TE, TF);
562 }
563 {
564 V Tv, Tw, Ty, Tz;
565 Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
566 Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
567 Tx = VSUB(Tv, Tw);
568 T21 = VADD(Tv, Tw);
569 Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
570 Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
571 TA = VSUB(Ty, Tz);
572 T22 = VADD(Ty, Tz);
573 }
574 {
575 V TB, TD, T2C, T2D;
576 TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
577 TC = VADD(Tu, TB);
578 T1s = VSUB(Tu, TB);
579 TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
580 TH = VSUB(TD, TG);
581 T1t = VADD(TG, TD);
582 T20 = VADD(T1Y, T1Z);
583 T23 = VADD(T21, T22);
584 T24 = VSUB(T20, T23);
585 T2C = VSUB(T1Y, T1Z);
586 T2D = VSUB(T22, T21);
587 T2E = VFMA(LDK(KP923879532), T2C, VMUL(LDK(KP382683432), T2D));
588 T2N = VFNMS(LDK(KP382683432), T2C, VMUL(LDK(KP923879532), T2D));
589 }
590 }
591 {
592 V Te, T2h, To, T2f, Th, T2i, Tl, T2e, Ti, Tp;
593 {
594 V Tc, Td, Tm, Tn;
595 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
596 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
597 Te = VSUB(Tc, Td);
598 T2h = VADD(Tc, Td);
599 Tm = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
600 Tn = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
601 To = VSUB(Tm, Tn);
602 T2f = VADD(Tm, Tn);
603 }
604 {
605 V Tf, Tg, Tj, Tk;
606 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
607 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
608 Th = VSUB(Tf, Tg);
609 T2i = VADD(Tf, Tg);
610 Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
611 Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
612 Tl = VSUB(Tj, Tk);
613 T2e = VADD(Tj, Tk);
614 }
615 T2g = VADD(T2e, T2f);
616 T2j = VADD(T2h, T2i);
617 Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
618 Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
619 Tq = VADD(Ti, Tp);
620 T1B = VSUB(Tp, Ti);
621 {
622 V T17, T18, T2y, T2z;
623 T17 = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
624 T18 = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
625 T19 = VSUB(T17, T18);
626 T1q = VADD(T18, T17);
627 T2y = VSUB(T2h, T2i);
628 T2z = VSUB(T2e, T2f);
629 T2A = VMUL(LDK(KP707106781), VADD(T2y, T2z));
630 T2L = VMUL(LDK(KP707106781), VSUB(T2z, T2y));
631 }
632 }
633 {
634 V T31, T32, T33, T34, T35, T36, T37, T38, T39, T3a, T3b, T3c;
635 {
636 V T2d, T2n, T2m, T2o;
637 {
638 V T1X, T2c, T2k, T2l;
639 T1X = VSUB(T1T, T1W);
640 T2c = VMUL(LDK(KP707106781), VADD(T24, T2b));
641 T2d = VADD(T1X, T2c);
642 T2n = VSUB(T1X, T2c);
643 T2k = VSUB(T2g, T2j);
644 T2l = VMUL(LDK(KP707106781), VSUB(T2b, T24));
645 T2m = VBYI(VADD(T2k, T2l));
646 T2o = VBYI(VSUB(T2l, T2k));
647 }
648 T31 = VSUB(T2d, T2m);
649 STM2(&(xo[56]), T31, ovs, &(xo[0]));
650 T32 = VADD(T2n, T2o);
651 STM2(&(xo[24]), T32, ovs, &(xo[0]));
652 T33 = VADD(T2d, T2m);
653 STM2(&(xo[8]), T33, ovs, &(xo[0]));
654 T34 = VSUB(T2n, T2o);
655 STM2(&(xo[40]), T34, ovs, &(xo[0]));
656 }
657 {
658 V T2r, T2v, T2u, T2w;
659 {
660 V T2p, T2q, T2s, T2t;
661 T2p = VADD(T1T, T1W);
662 T2q = VADD(T2j, T2g);
663 T2r = VADD(T2p, T2q);
664 T2v = VSUB(T2p, T2q);
665 T2s = VADD(T20, T23);
666 T2t = VADD(T27, T2a);
667 T2u = VADD(T2s, T2t);
668 T2w = VBYI(VSUB(T2t, T2s));
669 }
670 T35 = VSUB(T2r, T2u);
671 STM2(&(xo[32]), T35, ovs, &(xo[0]));
672 T36 = VADD(T2v, T2w);
673 STM2(&(xo[16]), T36, ovs, &(xo[0]));
674 T37 = VADD(T2r, T2u);
675 STM2(&(xo[0]), T37, ovs, &(xo[0]));
676 T38 = VSUB(T2v, T2w);
677 STM2(&(xo[48]), T38, ovs, &(xo[0]));
678 }
679 {
680 V T2V, T2Z, T2Y, T30;
681 {
682 V T2T, T2U, T2W, T2X;
683 T2T = VSUB(T2H, T2E);
684 T2U = VSUB(T2L, T2K);
685 T2V = VBYI(VSUB(T2T, T2U));
686 T2Z = VBYI(VADD(T2U, T2T));
687 T2W = VSUB(T2x, T2A);
688 T2X = VSUB(T2O, T2N);
689 T2Y = VSUB(T2W, T2X);
690 T30 = VADD(T2W, T2X);
691 }
692 T39 = VADD(T2V, T2Y);
693 STM2(&(xo[20]), T39, ovs, &(xo[0]));
694 T3a = VSUB(T30, T2Z);
695 STM2(&(xo[52]), T3a, ovs, &(xo[0]));
696 T3b = VSUB(T2Y, T2V);
697 STM2(&(xo[44]), T3b, ovs, &(xo[0]));
698 T3c = VADD(T2Z, T30);
699 STM2(&(xo[12]), T3c, ovs, &(xo[0]));
700 }
701 {
702 V T3d, T3e, T3f, T3g;
703 {
704 V T2J, T2R, T2Q, T2S;
705 {
706 V T2B, T2I, T2M, T2P;
707 T2B = VADD(T2x, T2A);
708 T2I = VADD(T2E, T2H);
709 T2J = VADD(T2B, T2I);
710 T2R = VSUB(T2B, T2I);
711 T2M = VADD(T2K, T2L);
712 T2P = VADD(T2N, T2O);
713 T2Q = VBYI(VADD(T2M, T2P));
714 T2S = VBYI(VSUB(T2P, T2M));
715 }
716 T3d = VSUB(T2J, T2Q);
717 STM2(&(xo[60]), T3d, ovs, &(xo[0]));
718 T3e = VADD(T2R, T2S);
719 STM2(&(xo[28]), T3e, ovs, &(xo[0]));
720 T3f = VADD(T2J, T2Q);
721 STM2(&(xo[4]), T3f, ovs, &(xo[0]));
722 T3g = VSUB(T2R, T2S);
723 STM2(&(xo[36]), T3g, ovs, &(xo[0]));
724 }
725 {
726 V T1r, T1C, T1M, T1K, T1F, T1N, T1y, T1J;
727 T1r = VADD(T1p, T1q);
728 T1C = VADD(T1A, T1B);
729 T1M = VSUB(T1p, T1q);
730 T1K = VSUB(T1B, T1A);
731 {
732 V T1D, T1E, T1u, T1x;
733 T1D = VFNMS(LDK(KP555570233), T1s, VMUL(LDK(KP831469612), T1t));
734 T1E = VFMA(LDK(KP555570233), T1v, VMUL(LDK(KP831469612), T1w));
735 T1F = VADD(T1D, T1E);
736 T1N = VSUB(T1E, T1D);
737 T1u = VFMA(LDK(KP831469612), T1s, VMUL(LDK(KP555570233), T1t));
738 T1x = VFNMS(LDK(KP555570233), T1w, VMUL(LDK(KP831469612), T1v));
739 T1y = VADD(T1u, T1x);
740 T1J = VSUB(T1x, T1u);
741 }
742 {
743 V T1z, T1G, T3h, T3i;
744 T1z = VADD(T1r, T1y);
745 T1G = VBYI(VADD(T1C, T1F));
746 T3h = VSUB(T1z, T1G);
747 STM2(&(xo[58]), T3h, ovs, &(xo[2]));
748 STN2(&(xo[56]), T31, T3h, ovs);
749 T3i = VADD(T1z, T1G);
750 STM2(&(xo[6]), T3i, ovs, &(xo[2]));
751 STN2(&(xo[4]), T3f, T3i, ovs);
752 }
753 {
754 V T1P, T1Q, T3j, T3k;
755 T1P = VBYI(VADD(T1K, T1J));
756 T1Q = VADD(T1M, T1N);
757 T3j = VADD(T1P, T1Q);
758 STM2(&(xo[10]), T3j, ovs, &(xo[2]));
759 STN2(&(xo[8]), T33, T3j, ovs);
760 T3k = VSUB(T1Q, T1P);
761 STM2(&(xo[54]), T3k, ovs, &(xo[2]));
762 STN2(&(xo[52]), T3a, T3k, ovs);
763 }
764 {
765 V T1H, T1I, T3l, T3m;
766 T1H = VSUB(T1r, T1y);
767 T1I = VBYI(VSUB(T1F, T1C));
768 T3l = VSUB(T1H, T1I);
769 STM2(&(xo[38]), T3l, ovs, &(xo[2]));
770 STN2(&(xo[36]), T3g, T3l, ovs);
771 T3m = VADD(T1H, T1I);
772 STM2(&(xo[26]), T3m, ovs, &(xo[2]));
773 STN2(&(xo[24]), T32, T3m, ovs);
774 }
775 {
776 V T1L, T1O, T3n, T3o;
777 T1L = VBYI(VSUB(T1J, T1K));
778 T1O = VSUB(T1M, T1N);
779 T3n = VADD(T1L, T1O);
780 STM2(&(xo[22]), T3n, ovs, &(xo[2]));
781 STN2(&(xo[20]), T39, T3n, ovs);
782 T3o = VSUB(T1O, T1L);
783 STM2(&(xo[42]), T3o, ovs, &(xo[2]));
784 STN2(&(xo[40]), T34, T3o, ovs);
785 }
786 }
787 {
788 V Tr, T1a, T1k, T1i, T1d, T1l, T10, T1h;
789 Tr = VADD(Tb, Tq);
790 T1a = VADD(T16, T19);
791 T1k = VSUB(Tb, Tq);
792 T1i = VSUB(T19, T16);
793 {
794 V T1b, T1c, TI, TZ;
795 T1b = VFNMS(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
796 T1c = VFMA(LDK(KP195090322), TT, VMUL(LDK(KP980785280), TY));
797 T1d = VADD(T1b, T1c);
798 T1l = VSUB(T1c, T1b);
799 TI = VFMA(LDK(KP980785280), TC, VMUL(LDK(KP195090322), TH));
800 TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
801 T10 = VADD(TI, TZ);
802 T1h = VSUB(TZ, TI);
803 }
804 {
805 V T11, T1e, T3p, T3q;
806 T11 = VADD(Tr, T10);
807 T1e = VBYI(VADD(T1a, T1d));
808 T3p = VSUB(T11, T1e);
809 STM2(&(xo[62]), T3p, ovs, &(xo[2]));
810 STN2(&(xo[60]), T3d, T3p, ovs);
811 T3q = VADD(T11, T1e);
812 STM2(&(xo[2]), T3q, ovs, &(xo[2]));
813 STN2(&(xo[0]), T37, T3q, ovs);
814 }
815 {
816 V T1n, T1o, T3r, T3s;
817 T1n = VBYI(VADD(T1i, T1h));
818 T1o = VADD(T1k, T1l);
819 T3r = VADD(T1n, T1o);
820 STM2(&(xo[14]), T3r, ovs, &(xo[2]));
821 STN2(&(xo[12]), T3c, T3r, ovs);
822 T3s = VSUB(T1o, T1n);
823 STM2(&(xo[50]), T3s, ovs, &(xo[2]));
824 STN2(&(xo[48]), T38, T3s, ovs);
825 }
826 {
827 V T1f, T1g, T3t, T3u;
828 T1f = VSUB(Tr, T10);
829 T1g = VBYI(VSUB(T1d, T1a));
830 T3t = VSUB(T1f, T1g);
831 STM2(&(xo[34]), T3t, ovs, &(xo[2]));
832 STN2(&(xo[32]), T35, T3t, ovs);
833 T3u = VADD(T1f, T1g);
834 STM2(&(xo[30]), T3u, ovs, &(xo[2]));
835 STN2(&(xo[28]), T3e, T3u, ovs);
836 }
837 {
838 V T1j, T1m, T3v, T3w;
839 T1j = VBYI(VSUB(T1h, T1i));
840 T1m = VSUB(T1k, T1l);
841 T3v = VADD(T1j, T1m);
842 STM2(&(xo[18]), T3v, ovs, &(xo[2]));
843 STN2(&(xo[16]), T36, T3v, ovs);
844 T3w = VSUB(T1m, T1j);
845 STM2(&(xo[46]), T3w, ovs, &(xo[2]));
846 STN2(&(xo[44]), T3b, T3w, ovs);
847 }
848 }
849 }
850 }
851 }
852 }
853 VLEAVE();
854 }
855
856 static const kdft_desc desc = { 32, XSIMD_STRING("n2fv_32"), {170, 26, 16, 0}, &GENUS, 0, 2, 0, 0 };
857
858 void XSIMD(codelet_n2fv_32) (planner *p) {
859 X(kdft_register) (p, n2fv_32, &desc);
860 }
861
862 #endif