comparison src/fftw-3.3.8/dft/simd/common/n1bv_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 08:04:58 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 -sign 1 -n 32 -name n1bv_32 -include dft/simd/n1b.h */
29
30 /*
31 * This function contains 186 FP additions, 98 FP multiplications,
32 * (or, 88 additions, 0 multiplications, 98 fused multiply/add),
33 * 58 stack variables, 7 constants, and 64 memory accesses
34 */
35 #include "dft/simd/n1b.h"
36
37 static void n1bv_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 = ii;
51 xo = io;
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, T2O, TC, T1s, TH, T1t, T20, T23, T24, T2E, T2N, 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(T1U, T1V);
83 T2x = VSUB(T1R, T1S);
84 T15 = VSUB(T6, T9);
85 T16 = VFMA(LDK(KP707106781), T15, T14);
86 T1A = VFNMS(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 T2O = 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 T2N = VFMA(LDK(KP414213562), T2C, T2D);
171 }
172 }
173 {
174 V Te, T2e, To, T2i, Th, T2f, Tl, T2h, 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 T2e = 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 T2i = 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 T2f = 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 T2h = 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(Ti, Tp);
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(T2e, T2f);
210 T2z = VSUB(T2h, T2i);
211 T2A = VADD(T2y, T2z);
212 T2L = VSUB(T2y, T2z);
213 }
214 }
215 {
216 V T2d, T2n, T2m, T2o;
217 {
218 V T1X, T2c, T2k, T2l;
219 T1X = VSUB(T1T, T1W);
220 T2c = VADD(T24, T2b);
221 T2d = VFNMS(LDK(KP707106781), T2c, T1X);
222 T2n = VFMA(LDK(KP707106781), T2c, T1X);
223 T2k = VSUB(T2g, T2j);
224 T2l = VSUB(T24, T2b);
225 T2m = VFNMS(LDK(KP707106781), T2l, T2k);
226 T2o = VFMA(LDK(KP707106781), T2l, T2k);
227 }
228 ST(&(xo[WS(os, 12)]), VFNMSI(T2m, T2d), ovs, &(xo[0]));
229 ST(&(xo[WS(os, 28)]), VFNMSI(T2o, T2n), ovs, &(xo[0]));
230 ST(&(xo[WS(os, 20)]), VFMAI(T2m, T2d), ovs, &(xo[0]));
231 ST(&(xo[WS(os, 4)]), VFMAI(T2o, T2n), ovs, &(xo[0]));
232 }
233 {
234 V T2r, T2v, T2u, T2w;
235 {
236 V T2p, T2q, T2s, T2t;
237 T2p = VADD(T1T, T1W);
238 T2q = VADD(T2g, T2j);
239 T2r = VSUB(T2p, T2q);
240 T2v = VADD(T2p, T2q);
241 T2s = VADD(T20, T23);
242 T2t = VADD(T27, T2a);
243 T2u = VSUB(T2s, T2t);
244 T2w = VADD(T2s, T2t);
245 }
246 ST(&(xo[WS(os, 24)]), VFNMSI(T2u, T2r), ovs, &(xo[0]));
247 ST(&(xo[0]), VADD(T2v, T2w), ovs, &(xo[0]));
248 ST(&(xo[WS(os, 8)]), VFMAI(T2u, T2r), ovs, &(xo[0]));
249 ST(&(xo[WS(os, 16)]), VSUB(T2v, T2w), ovs, &(xo[0]));
250 }
251 {
252 V T2V, T2Z, T2Y, T30;
253 {
254 V T2T, T2U, T2W, T2X;
255 T2T = VFNMS(LDK(KP707106781), T2A, T2x);
256 T2U = VADD(T2N, T2O);
257 T2V = VFNMS(LDK(KP923879532), T2U, T2T);
258 T2Z = VFMA(LDK(KP923879532), T2U, T2T);
259 T2W = VFNMS(LDK(KP707106781), T2L, T2K);
260 T2X = VSUB(T2E, T2H);
261 T2Y = VFMA(LDK(KP923879532), T2X, T2W);
262 T30 = VFNMS(LDK(KP923879532), T2X, T2W);
263 }
264 ST(&(xo[WS(os, 10)]), VFMAI(T2Y, T2V), ovs, &(xo[0]));
265 ST(&(xo[WS(os, 26)]), VFMAI(T30, T2Z), ovs, &(xo[0]));
266 ST(&(xo[WS(os, 22)]), VFNMSI(T2Y, T2V), ovs, &(xo[0]));
267 ST(&(xo[WS(os, 6)]), VFNMSI(T30, T2Z), ovs, &(xo[0]));
268 }
269 {
270 V T2J, T2R, T2Q, T2S;
271 {
272 V T2B, T2I, T2M, T2P;
273 T2B = VFMA(LDK(KP707106781), T2A, T2x);
274 T2I = VADD(T2E, T2H);
275 T2J = VFNMS(LDK(KP923879532), T2I, T2B);
276 T2R = VFMA(LDK(KP923879532), T2I, T2B);
277 T2M = VFMA(LDK(KP707106781), T2L, T2K);
278 T2P = VSUB(T2N, T2O);
279 T2Q = VFNMS(LDK(KP923879532), T2P, T2M);
280 T2S = VFMA(LDK(KP923879532), T2P, T2M);
281 }
282 ST(&(xo[WS(os, 14)]), VFNMSI(T2Q, T2J), ovs, &(xo[0]));
283 ST(&(xo[WS(os, 2)]), VFMAI(T2S, T2R), ovs, &(xo[0]));
284 ST(&(xo[WS(os, 18)]), VFMAI(T2Q, T2J), ovs, &(xo[0]));
285 ST(&(xo[WS(os, 30)]), VFNMSI(T2S, T2R), ovs, &(xo[0]));
286 }
287 {
288 V T1r, T1C, T1M, T1J, T1F, T1K, T1y, T1N;
289 T1r = VFMA(LDK(KP923879532), T1q, T1p);
290 T1C = VFNMS(LDK(KP923879532), T1B, T1A);
291 T1M = VFMA(LDK(KP923879532), T1B, T1A);
292 T1J = VFNMS(LDK(KP923879532), T1q, T1p);
293 {
294 V T1D, T1E, T1u, T1x;
295 T1D = VFNMS(LDK(KP668178637), T1s, T1t);
296 T1E = VFNMS(LDK(KP668178637), T1v, T1w);
297 T1F = VSUB(T1D, T1E);
298 T1K = VADD(T1D, T1E);
299 T1u = VFMA(LDK(KP668178637), T1t, T1s);
300 T1x = VFMA(LDK(KP668178637), T1w, T1v);
301 T1y = VADD(T1u, T1x);
302 T1N = VSUB(T1u, T1x);
303 }
304 {
305 V T1z, T1G, T1P, T1Q;
306 T1z = VFNMS(LDK(KP831469612), T1y, T1r);
307 T1G = VFNMS(LDK(KP831469612), T1F, T1C);
308 ST(&(xo[WS(os, 19)]), VFNMSI(T1G, T1z), ovs, &(xo[WS(os, 1)]));
309 ST(&(xo[WS(os, 13)]), VFMAI(T1G, T1z), ovs, &(xo[WS(os, 1)]));
310 T1P = VFNMS(LDK(KP831469612), T1K, T1J);
311 T1Q = VFMA(LDK(KP831469612), T1N, T1M);
312 ST(&(xo[WS(os, 5)]), VFMAI(T1Q, T1P), ovs, &(xo[WS(os, 1)]));
313 ST(&(xo[WS(os, 27)]), VFNMSI(T1Q, T1P), ovs, &(xo[WS(os, 1)]));
314 }
315 {
316 V T1H, T1I, T1L, T1O;
317 T1H = VFMA(LDK(KP831469612), T1y, T1r);
318 T1I = VFMA(LDK(KP831469612), T1F, T1C);
319 ST(&(xo[WS(os, 3)]), VFNMSI(T1I, T1H), ovs, &(xo[WS(os, 1)]));
320 ST(&(xo[WS(os, 29)]), VFMAI(T1I, T1H), ovs, &(xo[WS(os, 1)]));
321 T1L = VFMA(LDK(KP831469612), T1K, T1J);
322 T1O = VFNMS(LDK(KP831469612), T1N, T1M);
323 ST(&(xo[WS(os, 11)]), VFNMSI(T1O, T1L), ovs, &(xo[WS(os, 1)]));
324 ST(&(xo[WS(os, 21)]), VFMAI(T1O, T1L), ovs, &(xo[WS(os, 1)]));
325 }
326 }
327 {
328 V Tr, T1a, T1k, T1h, T1d, T1i, T10, T1l;
329 Tr = VFMA(LDK(KP923879532), Tq, Tb);
330 T1a = VFMA(LDK(KP923879532), T19, T16);
331 T1k = VFNMS(LDK(KP923879532), T19, T16);
332 T1h = VFNMS(LDK(KP923879532), Tq, Tb);
333 {
334 V T1b, T1c, TI, TZ;
335 T1b = VFMA(LDK(KP198912367), TC, TH);
336 T1c = VFMA(LDK(KP198912367), TT, TY);
337 T1d = VSUB(T1b, T1c);
338 T1i = VADD(T1b, T1c);
339 TI = VFNMS(LDK(KP198912367), TH, TC);
340 TZ = VFNMS(LDK(KP198912367), TY, TT);
341 T10 = VADD(TI, TZ);
342 T1l = VSUB(TI, TZ);
343 }
344 {
345 V T11, T1e, T1n, T1o;
346 T11 = VFNMS(LDK(KP980785280), T10, Tr);
347 T1e = VFNMS(LDK(KP980785280), T1d, T1a);
348 ST(&(xo[WS(os, 15)]), VFNMSI(T1e, T11), ovs, &(xo[WS(os, 1)]));
349 ST(&(xo[WS(os, 17)]), VFMAI(T1e, T11), ovs, &(xo[WS(os, 1)]));
350 T1n = VFMA(LDK(KP980785280), T1i, T1h);
351 T1o = VFNMS(LDK(KP980785280), T1l, T1k);
352 ST(&(xo[WS(os, 7)]), VFNMSI(T1o, T1n), ovs, &(xo[WS(os, 1)]));
353 ST(&(xo[WS(os, 25)]), VFMAI(T1o, T1n), ovs, &(xo[WS(os, 1)]));
354 }
355 {
356 V T1f, T1g, T1j, T1m;
357 T1f = VFMA(LDK(KP980785280), T10, Tr);
358 T1g = VFMA(LDK(KP980785280), T1d, T1a);
359 ST(&(xo[WS(os, 31)]), VFNMSI(T1g, T1f), ovs, &(xo[WS(os, 1)]));
360 ST(&(xo[WS(os, 1)]), VFMAI(T1g, T1f), ovs, &(xo[WS(os, 1)]));
361 T1j = VFNMS(LDK(KP980785280), T1i, T1h);
362 T1m = VFMA(LDK(KP980785280), T1l, T1k);
363 ST(&(xo[WS(os, 9)]), VFMAI(T1m, T1j), ovs, &(xo[WS(os, 1)]));
364 ST(&(xo[WS(os, 23)]), VFNMSI(T1m, T1j), ovs, &(xo[WS(os, 1)]));
365 }
366 }
367 }
368 }
369 VLEAVE();
370 }
371
372 static const kdft_desc desc = { 32, XSIMD_STRING("n1bv_32"), {88, 0, 98, 0}, &GENUS, 0, 0, 0, 0 };
373
374 void XSIMD(codelet_n1bv_32) (planner *p) {
375 X(kdft_register) (p, n1bv_32, &desc);
376 }
377
378 #else
379
380 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 32 -name n1bv_32 -include dft/simd/n1b.h */
381
382 /*
383 * This function contains 186 FP additions, 42 FP multiplications,
384 * (or, 170 additions, 26 multiplications, 16 fused multiply/add),
385 * 58 stack variables, 7 constants, and 64 memory accesses
386 */
387 #include "dft/simd/n1b.h"
388
389 static void n1bv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
390 {
391 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
392 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
393 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
394 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
395 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
396 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
397 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
398 {
399 INT i;
400 const R *xi;
401 R *xo;
402 xi = ii;
403 xo = io;
404 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)) {
405 V T2f, T2k, T2N, T2M, T19, T1B, Tb, T1p, TT, T1v, TY, T1w, T2E, T2F, T2G;
406 V T24, T2o, TC, T1s, TH, T1t, T2B, T2C, T2D, T1X, T2n, T2I, T2J, Tq, T1A;
407 V T14, T1q, T2c, T2l;
408 {
409 V T3, T2i, T18, T2j, T6, T2d, T9, T2e, T15, Ta;
410 {
411 V T1, T2, T16, T17;
412 T1 = LD(&(xi[0]), ivs, &(xi[0]));
413 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
414 T3 = VSUB(T1, T2);
415 T2i = VADD(T1, T2);
416 T16 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
417 T17 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
418 T18 = VSUB(T16, T17);
419 T2j = VADD(T16, T17);
420 }
421 {
422 V T4, T5, T7, T8;
423 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
424 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
425 T6 = VSUB(T4, T5);
426 T2d = VADD(T4, T5);
427 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
428 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
429 T9 = VSUB(T7, T8);
430 T2e = VADD(T7, T8);
431 }
432 T2f = VSUB(T2d, T2e);
433 T2k = VSUB(T2i, T2j);
434 T2N = VADD(T2d, T2e);
435 T2M = VADD(T2i, T2j);
436 T15 = VMUL(LDK(KP707106781), VSUB(T6, T9));
437 T19 = VSUB(T15, T18);
438 T1B = VADD(T18, T15);
439 Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
440 Tb = VSUB(T3, Ta);
441 T1p = VADD(T3, Ta);
442 }
443 {
444 V TL, T21, TW, T1Y, TO, T22, TS, T1Z;
445 {
446 V TJ, TK, TU, TV;
447 TJ = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
448 TK = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
449 TL = VSUB(TJ, TK);
450 T21 = VADD(TJ, TK);
451 TU = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
452 TV = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
453 TW = VSUB(TU, TV);
454 T1Y = VADD(TU, TV);
455 }
456 {
457 V TM, TN, TQ, TR;
458 TM = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
459 TN = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
460 TO = VSUB(TM, TN);
461 T22 = VADD(TM, TN);
462 TQ = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
463 TR = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
464 TS = VSUB(TQ, TR);
465 T1Z = VADD(TQ, TR);
466 }
467 {
468 V TP, TX, T20, T23;
469 TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
470 TT = VSUB(TP, TS);
471 T1v = VADD(TS, TP);
472 TX = VMUL(LDK(KP707106781), VADD(TL, TO));
473 TY = VSUB(TW, TX);
474 T1w = VADD(TW, TX);
475 T2E = VADD(T1Y, T1Z);
476 T2F = VADD(T21, T22);
477 T2G = VSUB(T2E, T2F);
478 T20 = VSUB(T1Y, T1Z);
479 T23 = VSUB(T21, T22);
480 T24 = VFMA(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T23));
481 T2o = VFNMS(LDK(KP382683432), T20, VMUL(LDK(KP923879532), T23));
482 }
483 }
484 {
485 V Tu, T1U, TF, T1R, Tx, T1V, TB, T1S;
486 {
487 V Ts, Tt, TD, TE;
488 Ts = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
489 Tt = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
490 Tu = VSUB(Ts, Tt);
491 T1U = VADD(Ts, Tt);
492 TD = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
493 TE = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
494 TF = VSUB(TD, TE);
495 T1R = VADD(TD, TE);
496 }
497 {
498 V Tv, Tw, Tz, TA;
499 Tv = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
500 Tw = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
501 Tx = VSUB(Tv, Tw);
502 T1V = VADD(Tv, Tw);
503 Tz = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
504 TA = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
505 TB = VSUB(Tz, TA);
506 T1S = VADD(Tz, TA);
507 }
508 {
509 V Ty, TG, T1T, T1W;
510 Ty = VMUL(LDK(KP707106781), VSUB(Tu, Tx));
511 TC = VSUB(Ty, TB);
512 T1s = VADD(TB, Ty);
513 TG = VMUL(LDK(KP707106781), VADD(Tu, Tx));
514 TH = VSUB(TF, TG);
515 T1t = VADD(TF, TG);
516 T2B = VADD(T1R, T1S);
517 T2C = VADD(T1U, T1V);
518 T2D = VSUB(T2B, T2C);
519 T1T = VSUB(T1R, T1S);
520 T1W = VSUB(T1U, T1V);
521 T1X = VFNMS(LDK(KP382683432), T1W, VMUL(LDK(KP923879532), T1T));
522 T2n = VFMA(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1W));
523 }
524 }
525 {
526 V Te, T26, To, T29, Th, T27, Tl, T2a, Ti, Tp;
527 {
528 V Tc, Td, Tm, Tn;
529 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
530 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
531 Te = VSUB(Tc, Td);
532 T26 = VADD(Tc, Td);
533 Tm = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
534 Tn = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
535 To = VSUB(Tm, Tn);
536 T29 = VADD(Tm, Tn);
537 }
538 {
539 V Tf, Tg, Tj, Tk;
540 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
541 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
542 Th = VSUB(Tf, Tg);
543 T27 = VADD(Tf, Tg);
544 Tj = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
545 Tk = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
546 Tl = VSUB(Tj, Tk);
547 T2a = VADD(Tj, Tk);
548 }
549 T2I = VADD(T26, T27);
550 T2J = VADD(T29, T2a);
551 Ti = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
552 Tp = VFNMS(LDK(KP382683432), To, VMUL(LDK(KP923879532), Tl));
553 Tq = VSUB(Ti, Tp);
554 T1A = VADD(Ti, Tp);
555 {
556 V T12, T13, T28, T2b;
557 T12 = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
558 T13 = VFMA(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
559 T14 = VSUB(T12, T13);
560 T1q = VADD(T12, T13);
561 T28 = VSUB(T26, T27);
562 T2b = VSUB(T29, T2a);
563 T2c = VMUL(LDK(KP707106781), VSUB(T28, T2b));
564 T2l = VMUL(LDK(KP707106781), VADD(T28, T2b));
565 }
566 }
567 {
568 V T2L, T2R, T2Q, T2S;
569 {
570 V T2H, T2K, T2O, T2P;
571 T2H = VMUL(LDK(KP707106781), VSUB(T2D, T2G));
572 T2K = VSUB(T2I, T2J);
573 T2L = VBYI(VSUB(T2H, T2K));
574 T2R = VBYI(VADD(T2K, T2H));
575 T2O = VSUB(T2M, T2N);
576 T2P = VMUL(LDK(KP707106781), VADD(T2D, T2G));
577 T2Q = VSUB(T2O, T2P);
578 T2S = VADD(T2O, T2P);
579 }
580 ST(&(xo[WS(os, 12)]), VADD(T2L, T2Q), ovs, &(xo[0]));
581 ST(&(xo[WS(os, 28)]), VSUB(T2S, T2R), ovs, &(xo[0]));
582 ST(&(xo[WS(os, 20)]), VSUB(T2Q, T2L), ovs, &(xo[0]));
583 ST(&(xo[WS(os, 4)]), VADD(T2R, T2S), ovs, &(xo[0]));
584 }
585 {
586 V T2h, T2r, T2q, T2s;
587 {
588 V T25, T2g, T2m, T2p;
589 T25 = VSUB(T1X, T24);
590 T2g = VSUB(T2c, T2f);
591 T2h = VBYI(VSUB(T25, T2g));
592 T2r = VBYI(VADD(T2g, T25));
593 T2m = VSUB(T2k, T2l);
594 T2p = VSUB(T2n, T2o);
595 T2q = VSUB(T2m, T2p);
596 T2s = VADD(T2m, T2p);
597 }
598 ST(&(xo[WS(os, 10)]), VADD(T2h, T2q), ovs, &(xo[0]));
599 ST(&(xo[WS(os, 26)]), VSUB(T2s, T2r), ovs, &(xo[0]));
600 ST(&(xo[WS(os, 22)]), VSUB(T2q, T2h), ovs, &(xo[0]));
601 ST(&(xo[WS(os, 6)]), VADD(T2r, T2s), ovs, &(xo[0]));
602 }
603 {
604 V T2V, T2Z, T2Y, T30;
605 {
606 V T2T, T2U, T2W, T2X;
607 T2T = VADD(T2M, T2N);
608 T2U = VADD(T2I, T2J);
609 T2V = VSUB(T2T, T2U);
610 T2Z = VADD(T2T, T2U);
611 T2W = VADD(T2B, T2C);
612 T2X = VADD(T2E, T2F);
613 T2Y = VBYI(VSUB(T2W, T2X));
614 T30 = VADD(T2W, T2X);
615 }
616 ST(&(xo[WS(os, 24)]), VSUB(T2V, T2Y), ovs, &(xo[0]));
617 ST(&(xo[0]), VADD(T2Z, T30), ovs, &(xo[0]));
618 ST(&(xo[WS(os, 8)]), VADD(T2V, T2Y), ovs, &(xo[0]));
619 ST(&(xo[WS(os, 16)]), VSUB(T2Z, T30), ovs, &(xo[0]));
620 }
621 {
622 V T2v, T2z, T2y, T2A;
623 {
624 V T2t, T2u, T2w, T2x;
625 T2t = VADD(T2k, T2l);
626 T2u = VADD(T1X, T24);
627 T2v = VADD(T2t, T2u);
628 T2z = VSUB(T2t, T2u);
629 T2w = VADD(T2f, T2c);
630 T2x = VADD(T2n, T2o);
631 T2y = VBYI(VADD(T2w, T2x));
632 T2A = VBYI(VSUB(T2x, T2w));
633 }
634 ST(&(xo[WS(os, 30)]), VSUB(T2v, T2y), ovs, &(xo[0]));
635 ST(&(xo[WS(os, 14)]), VADD(T2z, T2A), ovs, &(xo[0]));
636 ST(&(xo[WS(os, 2)]), VADD(T2v, T2y), ovs, &(xo[0]));
637 ST(&(xo[WS(os, 18)]), VSUB(T2z, T2A), ovs, &(xo[0]));
638 }
639 {
640 V T1r, T1C, T1M, T1K, T1F, T1N, T1y, T1J;
641 T1r = VSUB(T1p, T1q);
642 T1C = VSUB(T1A, T1B);
643 T1M = VADD(T1p, T1q);
644 T1K = VADD(T1B, T1A);
645 {
646 V T1D, T1E, T1u, T1x;
647 T1D = VFNMS(LDK(KP195090322), T1s, VMUL(LDK(KP980785280), T1t));
648 T1E = VFMA(LDK(KP195090322), T1v, VMUL(LDK(KP980785280), T1w));
649 T1F = VSUB(T1D, T1E);
650 T1N = VADD(T1D, T1E);
651 T1u = VFMA(LDK(KP980785280), T1s, VMUL(LDK(KP195090322), T1t));
652 T1x = VFNMS(LDK(KP195090322), T1w, VMUL(LDK(KP980785280), T1v));
653 T1y = VSUB(T1u, T1x);
654 T1J = VADD(T1u, T1x);
655 }
656 {
657 V T1z, T1G, T1P, T1Q;
658 T1z = VADD(T1r, T1y);
659 T1G = VBYI(VADD(T1C, T1F));
660 ST(&(xo[WS(os, 25)]), VSUB(T1z, T1G), ovs, &(xo[WS(os, 1)]));
661 ST(&(xo[WS(os, 7)]), VADD(T1z, T1G), ovs, &(xo[WS(os, 1)]));
662 T1P = VBYI(VADD(T1K, T1J));
663 T1Q = VADD(T1M, T1N);
664 ST(&(xo[WS(os, 1)]), VADD(T1P, T1Q), ovs, &(xo[WS(os, 1)]));
665 ST(&(xo[WS(os, 31)]), VSUB(T1Q, T1P), ovs, &(xo[WS(os, 1)]));
666 }
667 {
668 V T1H, T1I, T1L, T1O;
669 T1H = VSUB(T1r, T1y);
670 T1I = VBYI(VSUB(T1F, T1C));
671 ST(&(xo[WS(os, 23)]), VSUB(T1H, T1I), ovs, &(xo[WS(os, 1)]));
672 ST(&(xo[WS(os, 9)]), VADD(T1H, T1I), ovs, &(xo[WS(os, 1)]));
673 T1L = VBYI(VSUB(T1J, T1K));
674 T1O = VSUB(T1M, T1N);
675 ST(&(xo[WS(os, 15)]), VADD(T1L, T1O), ovs, &(xo[WS(os, 1)]));
676 ST(&(xo[WS(os, 17)]), VSUB(T1O, T1L), ovs, &(xo[WS(os, 1)]));
677 }
678 }
679 {
680 V Tr, T1a, T1k, T1i, T1d, T1l, T10, T1h;
681 Tr = VSUB(Tb, Tq);
682 T1a = VSUB(T14, T19);
683 T1k = VADD(Tb, Tq);
684 T1i = VADD(T19, T14);
685 {
686 V T1b, T1c, TI, TZ;
687 T1b = VFNMS(LDK(KP555570233), TC, VMUL(LDK(KP831469612), TH));
688 T1c = VFMA(LDK(KP555570233), TT, VMUL(LDK(KP831469612), TY));
689 T1d = VSUB(T1b, T1c);
690 T1l = VADD(T1b, T1c);
691 TI = VFMA(LDK(KP831469612), TC, VMUL(LDK(KP555570233), TH));
692 TZ = VFNMS(LDK(KP555570233), TY, VMUL(LDK(KP831469612), TT));
693 T10 = VSUB(TI, TZ);
694 T1h = VADD(TI, TZ);
695 }
696 {
697 V T11, T1e, T1n, T1o;
698 T11 = VADD(Tr, T10);
699 T1e = VBYI(VADD(T1a, T1d));
700 ST(&(xo[WS(os, 27)]), VSUB(T11, T1e), ovs, &(xo[WS(os, 1)]));
701 ST(&(xo[WS(os, 5)]), VADD(T11, T1e), ovs, &(xo[WS(os, 1)]));
702 T1n = VBYI(VADD(T1i, T1h));
703 T1o = VADD(T1k, T1l);
704 ST(&(xo[WS(os, 3)]), VADD(T1n, T1o), ovs, &(xo[WS(os, 1)]));
705 ST(&(xo[WS(os, 29)]), VSUB(T1o, T1n), ovs, &(xo[WS(os, 1)]));
706 }
707 {
708 V T1f, T1g, T1j, T1m;
709 T1f = VSUB(Tr, T10);
710 T1g = VBYI(VSUB(T1d, T1a));
711 ST(&(xo[WS(os, 21)]), VSUB(T1f, T1g), ovs, &(xo[WS(os, 1)]));
712 ST(&(xo[WS(os, 11)]), VADD(T1f, T1g), ovs, &(xo[WS(os, 1)]));
713 T1j = VBYI(VSUB(T1h, T1i));
714 T1m = VSUB(T1k, T1l);
715 ST(&(xo[WS(os, 13)]), VADD(T1j, T1m), ovs, &(xo[WS(os, 1)]));
716 ST(&(xo[WS(os, 19)]), VSUB(T1m, T1j), ovs, &(xo[WS(os, 1)]));
717 }
718 }
719 }
720 }
721 VLEAVE();
722 }
723
724 static const kdft_desc desc = { 32, XSIMD_STRING("n1bv_32"), {170, 26, 16, 0}, &GENUS, 0, 0, 0, 0 };
725
726 void XSIMD(codelet_n1bv_32) (planner *p) {
727 X(kdft_register) (p, n1bv_32, &desc);
728 }
729
730 #endif