comparison src/fftw-3.3.3/dft/simd/common/n1fv_32.c @ 10:37bf6b4a2645

Add FFTW3
author Chris Cannam
date Wed, 20 Mar 2013 15:35:50 +0000
parents
children
comparison
equal deleted inserted replaced
9:c0fb53affa76 10:37bf6b4a2645
1 /*
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 */
20
21 /* This file was automatically generated --- DO NOT EDIT */
22 /* Generated on Sun Nov 25 07:36:54 EST 2012 */
23
24 #include "codelet-dft.h"
25
26 #ifdef HAVE_FMA
27
28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n1fv_32 -include n1f.h */
29
30 /*
31 * This function contains 186 FP additions, 98 FP multiplications,
32 * (or, 88 additions, 0 multiplications, 98 fused multiply/add),
33 * 104 stack variables, 7 constants, and 64 memory accesses
34 */
35 #include "n1f.h"
36
37 static void n1fv_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(KP831469612, +0.831469612302545237078788377617905756738560812);
41 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
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 T1h, Tr, T1a, T1k, TI, T1b, T1L, T1P, T1I, T1G, T1O, T1Q, T1H, T1z, T1c;
54 V TZ;
55 {
56 V T2x, T1T, T2K, T1W, T1p, Tb, T1A, T16, Tu, TF, T2N, T2H, T2b, T2t, TY;
57 V T1w, TT, T1v, T20, T2C, Tj, Te, T2h, To, T2f, T23, T2D, TB, TG, Th;
58 V T2i, Tk;
59 {
60 V TL, TW, TP, TQ, T2F, T27, T28, TO;
61 {
62 V T1, T2, T12, T13, T4, T5, T7, T8;
63 T1 = LD(&(xi[0]), ivs, &(xi[0]));
64 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
65 T12 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
66 T13 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
67 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
68 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
69 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
70 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
71 {
72 V TM, T25, T26, TN;
73 {
74 V TJ, T3, T14, T1U, T6, T1V, T9, TK, TU, TV, T1R, T1S, Ta, T15;
75 TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
76 T1R = VADD(T1, T2);
77 T3 = VSUB(T1, T2);
78 T1S = VADD(T12, T13);
79 T14 = VSUB(T12, T13);
80 T1U = VADD(T4, T5);
81 T6 = VSUB(T4, T5);
82 T1V = VADD(T7, T8);
83 T9 = VSUB(T7, T8);
84 TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
85 TU = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
86 T2x = VSUB(T1R, T1S);
87 T1T = VADD(T1R, T1S);
88 TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
89 TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
90 T2K = VSUB(T1V, T1U);
91 T1W = VADD(T1U, T1V);
92 Ta = VADD(T6, T9);
93 T15 = VSUB(T9, T6);
94 T25 = VADD(TJ, TK);
95 TL = VSUB(TJ, TK);
96 T26 = VADD(TV, TU);
97 TW = VSUB(TU, TV);
98 TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
99 TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
100 T1p = VFNMS(LDK(KP707106781), Ta, T3);
101 Tb = VFMA(LDK(KP707106781), Ta, T3);
102 T1A = VFMA(LDK(KP707106781), T15, T14);
103 T16 = VFNMS(LDK(KP707106781), T15, T14);
104 TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
105 }
106 T2F = VSUB(T25, T26);
107 T27 = VADD(T25, T26);
108 T28 = VADD(TM, TN);
109 TO = VSUB(TM, TN);
110 }
111 }
112 {
113 V Ty, T21, Tx, Tz, T1Y, T1Z;
114 {
115 V Ts, Tt, TD, T29, TR, TE, Tv, Tw;
116 Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
117 Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
118 TD = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
119 T29 = VADD(TP, TQ);
120 TR = VSUB(TP, TQ);
121 TE = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
122 Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
123 Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
124 Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
125 T1Y = VADD(Ts, Tt);
126 Tu = VSUB(Ts, Tt);
127 {
128 V T2G, T2a, TX, TS;
129 T2G = VSUB(T29, T28);
130 T2a = VADD(T28, T29);
131 TX = VSUB(TR, TO);
132 TS = VADD(TO, TR);
133 T1Z = VADD(TD, TE);
134 TF = VSUB(TD, TE);
135 T21 = VADD(Tv, Tw);
136 Tx = VSUB(Tv, Tw);
137 T2N = VFMA(LDK(KP414213562), T2F, T2G);
138 T2H = VFNMS(LDK(KP414213562), T2G, T2F);
139 T2b = VSUB(T27, T2a);
140 T2t = VADD(T27, T2a);
141 TY = VFMA(LDK(KP707106781), TX, TW);
142 T1w = VFNMS(LDK(KP707106781), TX, TW);
143 TT = VFMA(LDK(KP707106781), TS, TL);
144 T1v = VFNMS(LDK(KP707106781), TS, TL);
145 Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
146 }
147 }
148 T20 = VADD(T1Y, T1Z);
149 T2C = VSUB(T1Y, T1Z);
150 {
151 V Tc, Td, Tm, Tn;
152 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
153 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
154 Tm = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
155 Tn = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
156 {
157 V Tf, TA, T22, Tg;
158 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
159 TA = VSUB(Ty, Tz);
160 T22 = VADD(Ty, Tz);
161 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
162 Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
163 Te = VSUB(Tc, Td);
164 T2h = VADD(Tc, Td);
165 To = VSUB(Tm, Tn);
166 T2f = VADD(Tn, Tm);
167 T23 = VADD(T21, T22);
168 T2D = VSUB(T21, T22);
169 TB = VADD(Tx, TA);
170 TG = VSUB(Tx, TA);
171 Th = VSUB(Tf, Tg);
172 T2i = VADD(Tf, Tg);
173 Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
174 }
175 }
176 }
177 }
178 {
179 V T1t, TH, T1s, TC, T2P, T2U, T2n, T2d, T2w, T2u, T1q, T19, T1B, Tq, T2W;
180 V T2M, T2B, T2T, T2v, T2r, T2o, T2m, T2X, T2I;
181 {
182 V T1X, T2p, T2E, T2O, T2s, T2y, T2j, T17, Ti, T2e, Tl, T2c, T2l, T24;
183 T1X = VSUB(T1T, T1W);
184 T2p = VADD(T1T, T1W);
185 T2E = VFNMS(LDK(KP414213562), T2D, T2C);
186 T2O = VFMA(LDK(KP414213562), T2C, T2D);
187 T2s = VADD(T20, T23);
188 T24 = VSUB(T20, T23);
189 T1t = VFNMS(LDK(KP707106781), TG, TF);
190 TH = VFMA(LDK(KP707106781), TG, TF);
191 T1s = VFNMS(LDK(KP707106781), TB, Tu);
192 TC = VFMA(LDK(KP707106781), TB, Tu);
193 T2y = VSUB(T2h, T2i);
194 T2j = VADD(T2h, T2i);
195 T17 = VFMA(LDK(KP414213562), Te, Th);
196 Ti = VFNMS(LDK(KP414213562), Th, Te);
197 T2e = VADD(Tj, Tk);
198 Tl = VSUB(Tj, Tk);
199 T2c = VADD(T24, T2b);
200 T2l = VSUB(T2b, T24);
201 {
202 V T2L, T2A, T2q, T2k;
203 T2P = VSUB(T2N, T2O);
204 T2U = VADD(T2O, T2N);
205 {
206 V T2z, T2g, T18, Tp;
207 T2z = VSUB(T2e, T2f);
208 T2g = VADD(T2e, T2f);
209 T18 = VFMA(LDK(KP414213562), Tl, To);
210 Tp = VFNMS(LDK(KP414213562), To, Tl);
211 T2n = VFMA(LDK(KP707106781), T2c, T1X);
212 T2d = VFNMS(LDK(KP707106781), T2c, T1X);
213 T2w = VSUB(T2t, T2s);
214 T2u = VADD(T2s, T2t);
215 T2L = VSUB(T2z, T2y);
216 T2A = VADD(T2y, T2z);
217 T2q = VADD(T2j, T2g);
218 T2k = VSUB(T2g, T2j);
219 T1q = VADD(T17, T18);
220 T19 = VSUB(T17, T18);
221 T1B = VSUB(Tp, Ti);
222 Tq = VADD(Ti, Tp);
223 }
224 T2W = VFNMS(LDK(KP707106781), T2L, T2K);
225 T2M = VFMA(LDK(KP707106781), T2L, T2K);
226 T2B = VFMA(LDK(KP707106781), T2A, T2x);
227 T2T = VFNMS(LDK(KP707106781), T2A, T2x);
228 T2v = VSUB(T2p, T2q);
229 T2r = VADD(T2p, T2q);
230 T2o = VFMA(LDK(KP707106781), T2l, T2k);
231 T2m = VFNMS(LDK(KP707106781), T2l, T2k);
232 T2X = VSUB(T2H, T2E);
233 T2I = VADD(T2E, T2H);
234 }
235 }
236 {
237 V T2V, T2Z, T2Y, T30, T2R, T2J;
238 T2V = VFNMS(LDK(KP923879532), T2U, T2T);
239 T2Z = VFMA(LDK(KP923879532), T2U, T2T);
240 ST(&(xo[WS(os, 24)]), VFNMSI(T2w, T2v), ovs, &(xo[0]));
241 ST(&(xo[WS(os, 8)]), VFMAI(T2w, T2v), ovs, &(xo[0]));
242 ST(&(xo[0]), VADD(T2r, T2u), ovs, &(xo[0]));
243 ST(&(xo[WS(os, 16)]), VSUB(T2r, T2u), ovs, &(xo[0]));
244 ST(&(xo[WS(os, 28)]), VFNMSI(T2o, T2n), ovs, &(xo[0]));
245 ST(&(xo[WS(os, 4)]), VFMAI(T2o, T2n), ovs, &(xo[0]));
246 ST(&(xo[WS(os, 20)]), VFMAI(T2m, T2d), ovs, &(xo[0]));
247 ST(&(xo[WS(os, 12)]), VFNMSI(T2m, T2d), ovs, &(xo[0]));
248 T2Y = VFMA(LDK(KP923879532), T2X, T2W);
249 T30 = VFNMS(LDK(KP923879532), T2X, T2W);
250 T2R = VFMA(LDK(KP923879532), T2I, T2B);
251 T2J = VFNMS(LDK(KP923879532), T2I, T2B);
252 {
253 V T1J, T1r, T1C, T1M, T2S, T2Q, T1u, T1D, T1E, T1x;
254 T1J = VFNMS(LDK(KP923879532), T1q, T1p);
255 T1r = VFMA(LDK(KP923879532), T1q, T1p);
256 T1C = VFMA(LDK(KP923879532), T1B, T1A);
257 T1M = VFNMS(LDK(KP923879532), T1B, T1A);
258 ST(&(xo[WS(os, 6)]), VFNMSI(T30, T2Z), ovs, &(xo[0]));
259 ST(&(xo[WS(os, 26)]), VFMAI(T30, T2Z), ovs, &(xo[0]));
260 ST(&(xo[WS(os, 22)]), VFNMSI(T2Y, T2V), ovs, &(xo[0]));
261 ST(&(xo[WS(os, 10)]), VFMAI(T2Y, T2V), ovs, &(xo[0]));
262 T2S = VFMA(LDK(KP923879532), T2P, T2M);
263 T2Q = VFNMS(LDK(KP923879532), T2P, T2M);
264 T1u = VFMA(LDK(KP668178637), T1t, T1s);
265 T1D = VFNMS(LDK(KP668178637), T1s, T1t);
266 T1E = VFNMS(LDK(KP668178637), T1v, T1w);
267 T1x = VFMA(LDK(KP668178637), T1w, T1v);
268 {
269 V T1K, T1F, T1N, T1y;
270 T1h = VFNMS(LDK(KP923879532), Tq, Tb);
271 Tr = VFMA(LDK(KP923879532), Tq, Tb);
272 ST(&(xo[WS(os, 30)]), VFNMSI(T2S, T2R), ovs, &(xo[0]));
273 ST(&(xo[WS(os, 2)]), VFMAI(T2S, T2R), ovs, &(xo[0]));
274 ST(&(xo[WS(os, 18)]), VFMAI(T2Q, T2J), ovs, &(xo[0]));
275 ST(&(xo[WS(os, 14)]), VFNMSI(T2Q, T2J), ovs, &(xo[0]));
276 T1K = VADD(T1D, T1E);
277 T1F = VSUB(T1D, T1E);
278 T1N = VSUB(T1x, T1u);
279 T1y = VADD(T1u, T1x);
280 T1a = VFMA(LDK(KP923879532), T19, T16);
281 T1k = VFNMS(LDK(KP923879532), T19, T16);
282 TI = VFNMS(LDK(KP198912367), TH, TC);
283 T1b = VFMA(LDK(KP198912367), TC, TH);
284 T1L = VFMA(LDK(KP831469612), T1K, T1J);
285 T1P = VFNMS(LDK(KP831469612), T1K, T1J);
286 T1I = VFMA(LDK(KP831469612), T1F, T1C);
287 T1G = VFNMS(LDK(KP831469612), T1F, T1C);
288 T1O = VFMA(LDK(KP831469612), T1N, T1M);
289 T1Q = VFNMS(LDK(KP831469612), T1N, T1M);
290 T1H = VFMA(LDK(KP831469612), T1y, T1r);
291 T1z = VFNMS(LDK(KP831469612), T1y, T1r);
292 T1c = VFMA(LDK(KP198912367), TT, TY);
293 TZ = VFNMS(LDK(KP198912367), TY, TT);
294 }
295 }
296 }
297 }
298 }
299 {
300 V T1d, T1i, T10, T1l;
301 ST(&(xo[WS(os, 21)]), VFNMSI(T1O, T1L), ovs, &(xo[WS(os, 1)]));
302 ST(&(xo[WS(os, 11)]), VFMAI(T1O, T1L), ovs, &(xo[WS(os, 1)]));
303 ST(&(xo[WS(os, 27)]), VFMAI(T1Q, T1P), ovs, &(xo[WS(os, 1)]));
304 ST(&(xo[WS(os, 5)]), VFNMSI(T1Q, T1P), ovs, &(xo[WS(os, 1)]));
305 ST(&(xo[WS(os, 3)]), VFMAI(T1I, T1H), ovs, &(xo[WS(os, 1)]));
306 ST(&(xo[WS(os, 29)]), VFNMSI(T1I, T1H), ovs, &(xo[WS(os, 1)]));
307 ST(&(xo[WS(os, 19)]), VFMAI(T1G, T1z), ovs, &(xo[WS(os, 1)]));
308 ST(&(xo[WS(os, 13)]), VFNMSI(T1G, T1z), ovs, &(xo[WS(os, 1)]));
309 T1d = VSUB(T1b, T1c);
310 T1i = VADD(T1b, T1c);
311 T10 = VADD(TI, TZ);
312 T1l = VSUB(TZ, TI);
313 {
314 V T1n, T1j, T1e, T1g, T1o, T1m, T11, T1f;
315 T1n = VFMA(LDK(KP980785280), T1i, T1h);
316 T1j = VFNMS(LDK(KP980785280), T1i, T1h);
317 T1e = VFNMS(LDK(KP980785280), T1d, T1a);
318 T1g = VFMA(LDK(KP980785280), T1d, T1a);
319 T1o = VFMA(LDK(KP980785280), T1l, T1k);
320 T1m = VFNMS(LDK(KP980785280), T1l, T1k);
321 T11 = VFNMS(LDK(KP980785280), T10, Tr);
322 T1f = VFMA(LDK(KP980785280), T10, Tr);
323 ST(&(xo[WS(os, 23)]), VFMAI(T1m, T1j), ovs, &(xo[WS(os, 1)]));
324 ST(&(xo[WS(os, 9)]), VFNMSI(T1m, T1j), ovs, &(xo[WS(os, 1)]));
325 ST(&(xo[WS(os, 25)]), VFNMSI(T1o, T1n), ovs, &(xo[WS(os, 1)]));
326 ST(&(xo[WS(os, 7)]), VFMAI(T1o, T1n), ovs, &(xo[WS(os, 1)]));
327 ST(&(xo[WS(os, 31)]), VFMAI(T1g, T1f), ovs, &(xo[WS(os, 1)]));
328 ST(&(xo[WS(os, 1)]), VFNMSI(T1g, T1f), ovs, &(xo[WS(os, 1)]));
329 ST(&(xo[WS(os, 15)]), VFMAI(T1e, T11), ovs, &(xo[WS(os, 1)]));
330 ST(&(xo[WS(os, 17)]), VFNMSI(T1e, T11), ovs, &(xo[WS(os, 1)]));
331 }
332 }
333 }
334 }
335 VLEAVE();
336 }
337
338 static const kdft_desc desc = { 32, XSIMD_STRING("n1fv_32"), {88, 0, 98, 0}, &GENUS, 0, 0, 0, 0 };
339
340 void XSIMD(codelet_n1fv_32) (planner *p) {
341 X(kdft_register) (p, n1fv_32, &desc);
342 }
343
344 #else /* HAVE_FMA */
345
346 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name n1fv_32 -include n1f.h */
347
348 /*
349 * This function contains 186 FP additions, 42 FP multiplications,
350 * (or, 170 additions, 26 multiplications, 16 fused multiply/add),
351 * 58 stack variables, 7 constants, and 64 memory accesses
352 */
353 #include "n1f.h"
354
355 static void n1fv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
356 {
357 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
358 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
359 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
360 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
361 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
362 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
363 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
364 {
365 INT i;
366 const R *xi;
367 R *xo;
368 xi = ri;
369 xo = ro;
370 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)) {
371 V T1T, T1W, T2K, T2x, T16, T1A, Tb, T1p, TT, T1v, TY, T1w, T27, T2a, T2b;
372 V T2H, T2O, TC, T1s, TH, T1t, T20, T23, T24, T2E, T2N, T2g, T2j, Tq, T1B;
373 V T19, T1q, T2A, T2L;
374 {
375 V T3, T1R, T15, T1S, T6, T1U, T9, T1V, T12, Ta;
376 {
377 V T1, T2, T13, T14;
378 T1 = LD(&(xi[0]), ivs, &(xi[0]));
379 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
380 T3 = VSUB(T1, T2);
381 T1R = VADD(T1, T2);
382 T13 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
383 T14 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
384 T15 = VSUB(T13, T14);
385 T1S = VADD(T13, T14);
386 }
387 {
388 V T4, T5, T7, T8;
389 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
390 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
391 T6 = VSUB(T4, T5);
392 T1U = VADD(T4, T5);
393 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
394 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
395 T9 = VSUB(T7, T8);
396 T1V = VADD(T7, T8);
397 }
398 T1T = VADD(T1R, T1S);
399 T1W = VADD(T1U, T1V);
400 T2K = VSUB(T1V, T1U);
401 T2x = VSUB(T1R, T1S);
402 T12 = VMUL(LDK(KP707106781), VSUB(T9, T6));
403 T16 = VSUB(T12, T15);
404 T1A = VADD(T15, T12);
405 Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
406 Tb = VADD(T3, Ta);
407 T1p = VSUB(T3, Ta);
408 }
409 {
410 V TL, T25, TX, T26, TO, T28, TR, T29;
411 {
412 V TJ, TK, TV, TW;
413 TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
414 TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
415 TL = VSUB(TJ, TK);
416 T25 = VADD(TJ, TK);
417 TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
418 TW = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
419 TX = VSUB(TV, TW);
420 T26 = VADD(TV, TW);
421 }
422 {
423 V TM, TN, TP, TQ;
424 TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
425 TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
426 TO = VSUB(TM, TN);
427 T28 = VADD(TM, TN);
428 TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
429 TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
430 TR = VSUB(TP, TQ);
431 T29 = VADD(TP, TQ);
432 }
433 {
434 V TS, TU, T2F, T2G;
435 TS = VMUL(LDK(KP707106781), VADD(TO, TR));
436 TT = VADD(TL, TS);
437 T1v = VSUB(TL, TS);
438 TU = VMUL(LDK(KP707106781), VSUB(TR, TO));
439 TY = VSUB(TU, TX);
440 T1w = VADD(TX, TU);
441 T27 = VADD(T25, T26);
442 T2a = VADD(T28, T29);
443 T2b = VSUB(T27, T2a);
444 T2F = VSUB(T25, T26);
445 T2G = VSUB(T29, T28);
446 T2H = VFNMS(LDK(KP382683432), T2G, VMUL(LDK(KP923879532), T2F));
447 T2O = VFMA(LDK(KP382683432), T2F, VMUL(LDK(KP923879532), T2G));
448 }
449 }
450 {
451 V Tu, T1Y, TG, T1Z, Tx, T21, TA, T22;
452 {
453 V Ts, Tt, TE, TF;
454 Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
455 Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
456 Tu = VSUB(Ts, Tt);
457 T1Y = VADD(Ts, Tt);
458 TE = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
459 TF = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
460 TG = VSUB(TE, TF);
461 T1Z = VADD(TE, TF);
462 }
463 {
464 V Tv, Tw, Ty, Tz;
465 Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
466 Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
467 Tx = VSUB(Tv, Tw);
468 T21 = VADD(Tv, Tw);
469 Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
470 Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
471 TA = VSUB(Ty, Tz);
472 T22 = VADD(Ty, Tz);
473 }
474 {
475 V TB, TD, T2C, T2D;
476 TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
477 TC = VADD(Tu, TB);
478 T1s = VSUB(Tu, TB);
479 TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
480 TH = VSUB(TD, TG);
481 T1t = VADD(TG, TD);
482 T20 = VADD(T1Y, T1Z);
483 T23 = VADD(T21, T22);
484 T24 = VSUB(T20, T23);
485 T2C = VSUB(T1Y, T1Z);
486 T2D = VSUB(T22, T21);
487 T2E = VFMA(LDK(KP923879532), T2C, VMUL(LDK(KP382683432), T2D));
488 T2N = VFNMS(LDK(KP382683432), T2C, VMUL(LDK(KP923879532), T2D));
489 }
490 }
491 {
492 V Te, T2h, To, T2f, Th, T2i, Tl, T2e, Ti, Tp;
493 {
494 V Tc, Td, Tm, Tn;
495 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
496 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
497 Te = VSUB(Tc, Td);
498 T2h = VADD(Tc, Td);
499 Tm = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
500 Tn = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
501 To = VSUB(Tm, Tn);
502 T2f = VADD(Tm, Tn);
503 }
504 {
505 V Tf, Tg, Tj, Tk;
506 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
507 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
508 Th = VSUB(Tf, Tg);
509 T2i = VADD(Tf, Tg);
510 Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
511 Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
512 Tl = VSUB(Tj, Tk);
513 T2e = VADD(Tj, Tk);
514 }
515 T2g = VADD(T2e, T2f);
516 T2j = VADD(T2h, T2i);
517 Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
518 Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
519 Tq = VADD(Ti, Tp);
520 T1B = VSUB(Tp, Ti);
521 {
522 V T17, T18, T2y, T2z;
523 T17 = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
524 T18 = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
525 T19 = VSUB(T17, T18);
526 T1q = VADD(T18, T17);
527 T2y = VSUB(T2h, T2i);
528 T2z = VSUB(T2e, T2f);
529 T2A = VMUL(LDK(KP707106781), VADD(T2y, T2z));
530 T2L = VMUL(LDK(KP707106781), VSUB(T2z, T2y));
531 }
532 }
533 {
534 V T2d, T2n, T2m, T2o;
535 {
536 V T1X, T2c, T2k, T2l;
537 T1X = VSUB(T1T, T1W);
538 T2c = VMUL(LDK(KP707106781), VADD(T24, T2b));
539 T2d = VADD(T1X, T2c);
540 T2n = VSUB(T1X, T2c);
541 T2k = VSUB(T2g, T2j);
542 T2l = VMUL(LDK(KP707106781), VSUB(T2b, T24));
543 T2m = VBYI(VADD(T2k, T2l));
544 T2o = VBYI(VSUB(T2l, T2k));
545 }
546 ST(&(xo[WS(os, 28)]), VSUB(T2d, T2m), ovs, &(xo[0]));
547 ST(&(xo[WS(os, 12)]), VADD(T2n, T2o), ovs, &(xo[0]));
548 ST(&(xo[WS(os, 4)]), VADD(T2d, T2m), ovs, &(xo[0]));
549 ST(&(xo[WS(os, 20)]), VSUB(T2n, T2o), ovs, &(xo[0]));
550 }
551 {
552 V T2r, T2v, T2u, T2w;
553 {
554 V T2p, T2q, T2s, T2t;
555 T2p = VADD(T1T, T1W);
556 T2q = VADD(T2j, T2g);
557 T2r = VADD(T2p, T2q);
558 T2v = VSUB(T2p, T2q);
559 T2s = VADD(T20, T23);
560 T2t = VADD(T27, T2a);
561 T2u = VADD(T2s, T2t);
562 T2w = VBYI(VSUB(T2t, T2s));
563 }
564 ST(&(xo[WS(os, 16)]), VSUB(T2r, T2u), ovs, &(xo[0]));
565 ST(&(xo[WS(os, 8)]), VADD(T2v, T2w), ovs, &(xo[0]));
566 ST(&(xo[0]), VADD(T2r, T2u), ovs, &(xo[0]));
567 ST(&(xo[WS(os, 24)]), VSUB(T2v, T2w), ovs, &(xo[0]));
568 }
569 {
570 V T2V, T2Z, T2Y, T30;
571 {
572 V T2T, T2U, T2W, T2X;
573 T2T = VSUB(T2H, T2E);
574 T2U = VSUB(T2L, T2K);
575 T2V = VBYI(VSUB(T2T, T2U));
576 T2Z = VBYI(VADD(T2U, T2T));
577 T2W = VSUB(T2x, T2A);
578 T2X = VSUB(T2O, T2N);
579 T2Y = VSUB(T2W, T2X);
580 T30 = VADD(T2W, T2X);
581 }
582 ST(&(xo[WS(os, 10)]), VADD(T2V, T2Y), ovs, &(xo[0]));
583 ST(&(xo[WS(os, 26)]), VSUB(T30, T2Z), ovs, &(xo[0]));
584 ST(&(xo[WS(os, 22)]), VSUB(T2Y, T2V), ovs, &(xo[0]));
585 ST(&(xo[WS(os, 6)]), VADD(T2Z, T30), ovs, &(xo[0]));
586 }
587 {
588 V T2J, T2R, T2Q, T2S;
589 {
590 V T2B, T2I, T2M, T2P;
591 T2B = VADD(T2x, T2A);
592 T2I = VADD(T2E, T2H);
593 T2J = VADD(T2B, T2I);
594 T2R = VSUB(T2B, T2I);
595 T2M = VADD(T2K, T2L);
596 T2P = VADD(T2N, T2O);
597 T2Q = VBYI(VADD(T2M, T2P));
598 T2S = VBYI(VSUB(T2P, T2M));
599 }
600 ST(&(xo[WS(os, 30)]), VSUB(T2J, T2Q), ovs, &(xo[0]));
601 ST(&(xo[WS(os, 14)]), VADD(T2R, T2S), ovs, &(xo[0]));
602 ST(&(xo[WS(os, 2)]), VADD(T2J, T2Q), ovs, &(xo[0]));
603 ST(&(xo[WS(os, 18)]), VSUB(T2R, T2S), ovs, &(xo[0]));
604 }
605 {
606 V T1r, T1C, T1M, T1K, T1F, T1N, T1y, T1J;
607 T1r = VADD(T1p, T1q);
608 T1C = VADD(T1A, T1B);
609 T1M = VSUB(T1p, T1q);
610 T1K = VSUB(T1B, T1A);
611 {
612 V T1D, T1E, T1u, T1x;
613 T1D = VFNMS(LDK(KP555570233), T1s, VMUL(LDK(KP831469612), T1t));
614 T1E = VFMA(LDK(KP555570233), T1v, VMUL(LDK(KP831469612), T1w));
615 T1F = VADD(T1D, T1E);
616 T1N = VSUB(T1E, T1D);
617 T1u = VFMA(LDK(KP831469612), T1s, VMUL(LDK(KP555570233), T1t));
618 T1x = VFNMS(LDK(KP555570233), T1w, VMUL(LDK(KP831469612), T1v));
619 T1y = VADD(T1u, T1x);
620 T1J = VSUB(T1x, T1u);
621 }
622 {
623 V T1z, T1G, T1P, T1Q;
624 T1z = VADD(T1r, T1y);
625 T1G = VBYI(VADD(T1C, T1F));
626 ST(&(xo[WS(os, 29)]), VSUB(T1z, T1G), ovs, &(xo[WS(os, 1)]));
627 ST(&(xo[WS(os, 3)]), VADD(T1z, T1G), ovs, &(xo[WS(os, 1)]));
628 T1P = VBYI(VADD(T1K, T1J));
629 T1Q = VADD(T1M, T1N);
630 ST(&(xo[WS(os, 5)]), VADD(T1P, T1Q), ovs, &(xo[WS(os, 1)]));
631 ST(&(xo[WS(os, 27)]), VSUB(T1Q, T1P), ovs, &(xo[WS(os, 1)]));
632 }
633 {
634 V T1H, T1I, T1L, T1O;
635 T1H = VSUB(T1r, T1y);
636 T1I = VBYI(VSUB(T1F, T1C));
637 ST(&(xo[WS(os, 19)]), VSUB(T1H, T1I), ovs, &(xo[WS(os, 1)]));
638 ST(&(xo[WS(os, 13)]), VADD(T1H, T1I), ovs, &(xo[WS(os, 1)]));
639 T1L = VBYI(VSUB(T1J, T1K));
640 T1O = VSUB(T1M, T1N);
641 ST(&(xo[WS(os, 11)]), VADD(T1L, T1O), ovs, &(xo[WS(os, 1)]));
642 ST(&(xo[WS(os, 21)]), VSUB(T1O, T1L), ovs, &(xo[WS(os, 1)]));
643 }
644 }
645 {
646 V Tr, T1a, T1k, T1i, T1d, T1l, T10, T1h;
647 Tr = VADD(Tb, Tq);
648 T1a = VADD(T16, T19);
649 T1k = VSUB(Tb, Tq);
650 T1i = VSUB(T19, T16);
651 {
652 V T1b, T1c, TI, TZ;
653 T1b = VFNMS(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
654 T1c = VFMA(LDK(KP195090322), TT, VMUL(LDK(KP980785280), TY));
655 T1d = VADD(T1b, T1c);
656 T1l = VSUB(T1c, T1b);
657 TI = VFMA(LDK(KP980785280), TC, VMUL(LDK(KP195090322), TH));
658 TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
659 T10 = VADD(TI, TZ);
660 T1h = VSUB(TZ, TI);
661 }
662 {
663 V T11, T1e, T1n, T1o;
664 T11 = VADD(Tr, T10);
665 T1e = VBYI(VADD(T1a, T1d));
666 ST(&(xo[WS(os, 31)]), VSUB(T11, T1e), ovs, &(xo[WS(os, 1)]));
667 ST(&(xo[WS(os, 1)]), VADD(T11, T1e), ovs, &(xo[WS(os, 1)]));
668 T1n = VBYI(VADD(T1i, T1h));
669 T1o = VADD(T1k, T1l);
670 ST(&(xo[WS(os, 7)]), VADD(T1n, T1o), ovs, &(xo[WS(os, 1)]));
671 ST(&(xo[WS(os, 25)]), VSUB(T1o, T1n), ovs, &(xo[WS(os, 1)]));
672 }
673 {
674 V T1f, T1g, T1j, T1m;
675 T1f = VSUB(Tr, T10);
676 T1g = VBYI(VSUB(T1d, T1a));
677 ST(&(xo[WS(os, 17)]), VSUB(T1f, T1g), ovs, &(xo[WS(os, 1)]));
678 ST(&(xo[WS(os, 15)]), VADD(T1f, T1g), ovs, &(xo[WS(os, 1)]));
679 T1j = VBYI(VSUB(T1h, T1i));
680 T1m = VSUB(T1k, T1l);
681 ST(&(xo[WS(os, 9)]), VADD(T1j, T1m), ovs, &(xo[WS(os, 1)]));
682 ST(&(xo[WS(os, 23)]), VSUB(T1m, T1j), ovs, &(xo[WS(os, 1)]));
683 }
684 }
685 }
686 }
687 VLEAVE();
688 }
689
690 static const kdft_desc desc = { 32, XSIMD_STRING("n1fv_32"), {170, 26, 16, 0}, &GENUS, 0, 0, 0, 0 };
691
692 void XSIMD(codelet_n1fv_32) (planner *p) {
693 X(kdft_register) (p, n1fv_32, &desc);
694 }
695
696 #endif /* HAVE_FMA */