comparison src/fftw-3.3.3/dft/simd/common/n1fv_25.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 07:36:55 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 25 -name n1fv_25 -include n1f.h */
29
30 /*
31 * This function contains 224 FP additions, 193 FP multiplications,
32 * (or, 43 additions, 12 multiplications, 181 fused multiply/add),
33 * 215 stack variables, 67 constants, and 50 memory accesses
34 */
35 #include "n1f.h"
36
37 static void n1fv_25(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38 {
39 DVK(KP792626838, +0.792626838241819413632131824093538848057784557);
40 DVK(KP876091699, +0.876091699473550838204498029706869638173524346);
41 DVK(KP803003575, +0.803003575438660414833440593570376004635464850);
42 DVK(KP617882369, +0.617882369114440893914546919006756321695042882);
43 DVK(KP242145790, +0.242145790282157779872542093866183953459003101);
44 DVK(KP968583161, +0.968583161128631119490168375464735813836012403);
45 DVK(KP999544308, +0.999544308746292983948881682379742149196758193);
46 DVK(KP683113946, +0.683113946453479238701949862233725244439656928);
47 DVK(KP559154169, +0.559154169276087864842202529084232643714075927);
48 DVK(KP904730450, +0.904730450839922351881287709692877908104763647);
49 DVK(KP829049696, +0.829049696159252993975487806364305442437946767);
50 DVK(KP831864738, +0.831864738706457140726048799369896829771167132);
51 DVK(KP916574801, +0.916574801383451584742370439148878693530976769);
52 DVK(KP894834959, +0.894834959464455102997960030820114611498661386);
53 DVK(KP809385824, +0.809385824416008241660603814668679683846476688);
54 DVK(KP447417479, +0.447417479732227551498980015410057305749330693);
55 DVK(KP860541664, +0.860541664367944677098261680920518816412804187);
56 DVK(KP897376177, +0.897376177523557693138608077137219684419427330);
57 DVK(KP876306680, +0.876306680043863587308115903922062583399064238);
58 DVK(KP681693190, +0.681693190061530575150324149145440022633095390);
59 DVK(KP560319534, +0.560319534973832390111614715371676131169633784);
60 DVK(KP855719849, +0.855719849902058969314654733608091555096772472);
61 DVK(KP237294955, +0.237294955877110315393888866460840817927895961);
62 DVK(KP949179823, +0.949179823508441261575555465843363271711583843);
63 DVK(KP904508497, +0.904508497187473712051146708591409529430077295);
64 DVK(KP997675361, +0.997675361079556513670859573984492383596555031);
65 DVK(KP262346850, +0.262346850930607871785420028382979691334784273);
66 DVK(KP763932022, +0.763932022500210303590826331268723764559381640);
67 DVK(KP992114701, +0.992114701314477831049793042785778521453036709);
68 DVK(KP690983005, +0.690983005625052575897706582817180941139845410);
69 DVK(KP952936919, +0.952936919628306576880750665357914584765951388);
70 DVK(KP998026728, +0.998026728428271561952336806863450553336905220);
71 DVK(KP570584518, +0.570584518783621657366766175430996792655723863);
72 DVK(KP669429328, +0.669429328479476605641803240971985825917022098);
73 DVK(KP923225144, +0.923225144846402650453449441572664695995209956);
74 DVK(KP906616052, +0.906616052148196230441134447086066874408359177);
75 DVK(KP956723877, +0.956723877038460305821989399535483155872969262);
76 DVK(KP522616830, +0.522616830205754336872861364785224694908468440);
77 DVK(KP945422727, +0.945422727388575946270360266328811958657216298);
78 DVK(KP912575812, +0.912575812670962425556968549836277086778922727);
79 DVK(KP982009705, +0.982009705009746369461829878184175962711969869);
80 DVK(KP921078979, +0.921078979742360627699756128143719920817673854);
81 DVK(KP734762448, +0.734762448793050413546343770063151342619912334);
82 DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
83 DVK(KP958953096, +0.958953096729998668045963838399037225970891871);
84 DVK(KP867381224, +0.867381224396525206773171885031575671309956167);
85 DVK(KP269969613, +0.269969613759572083574752974412347470060951301);
86 DVK(KP244189809, +0.244189809627953270309879511234821255780225091);
87 DVK(KP845997307, +0.845997307939530944175097360758058292389769300);
88 DVK(KP772036680, +0.772036680810363904029489473607579825330539880);
89 DVK(KP132830569, +0.132830569247582714407653942074819768844536507);
90 DVK(KP120146378, +0.120146378570687701782758537356596213647956445);
91 DVK(KP987388751, +0.987388751065621252324603216482382109400433949);
92 DVK(KP893101515, +0.893101515366181661711202267938416198338079437);
93 DVK(KP786782374, +0.786782374965295178365099601674911834788448471);
94 DVK(KP869845200, +0.869845200362138853122720822420327157933056305);
95 DVK(KP447533225, +0.447533225982656890041886979663652563063114397);
96 DVK(KP494780565, +0.494780565770515410344588413655324772219443730);
97 DVK(KP578046249, +0.578046249379945007321754579646815604023525655);
98 DVK(KP522847744, +0.522847744331509716623755382187077770911012542);
99 DVK(KP059835404, +0.059835404262124915169548397419498386427871950);
100 DVK(KP066152395, +0.066152395967733048213034281011006031460903353);
101 DVK(KP603558818, +0.603558818296015001454675132653458027918768137);
102 DVK(KP667278218, +0.667278218140296670899089292254759909713898805);
103 DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
104 DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
105 DVK(KP618033988, +0.618033988749894848204586834365638117720309180);
106 {
107 INT i;
108 const R *xi;
109 R *xo;
110 xi = ri;
111 xo = ro;
112 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(50, is), MAKE_VOLATILE_STRIDE(50, os)) {
113 V T1g, T1k, T1I, T24, T2a, T1G, T1A, T1l, T1B, T1H, T1d;
114 {
115 V T2z, T1q, Ta, T9, T3n, Ty, Tl, T2O, T2W, T2l, T2s, TV, T1i, T1K, T1S;
116 V T3z, T3t, Tk, T3o, Tp, T2g, T2N, T2V, T2o, T2t, T1a, T1j, T1J, T1R, Tz;
117 V Tt, TA, Tw;
118 {
119 V T1, T5, T6, T2, T3;
120 T1 = LD(&(xi[0]), ivs, &(xi[0]));
121 T5 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
122 T6 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
123 T2 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
124 T3 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
125 {
126 V TH, TW, TK, TS, T10, T8, TN, TT, T17, TZ, T11;
127 TH = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
128 TW = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
129 {
130 V TI, TJ, TL, T7, T1p, T4, T1o, TM, TX, TY;
131 TI = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
132 TJ = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
133 TL = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
134 T7 = VADD(T5, T6);
135 T1p = VSUB(T5, T6);
136 T4 = VADD(T2, T3);
137 T1o = VSUB(T2, T3);
138 TM = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
139 TX = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
140 TK = VADD(TI, TJ);
141 TS = VSUB(TI, TJ);
142 TY = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
143 T10 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
144 T2z = VFNMS(LDK(KP618033988), T1o, T1p);
145 T1q = VFMA(LDK(KP618033988), T1p, T1o);
146 Ta = VSUB(T4, T7);
147 T8 = VADD(T4, T7);
148 TN = VADD(TL, TM);
149 TT = VSUB(TM, TL);
150 T17 = VSUB(TX, TY);
151 TZ = VADD(TX, TY);
152 T11 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
153 }
154 {
155 V Tc, T2m, T19, Tn, To, Tr, Tj, T16, T2n, Ts, Tu, Tv;
156 {
157 V TU, T2j, TO, TQ, T12, T18;
158 Tc = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
159 T9 = VFNMS(LDK(KP250000000), T8, T1);
160 T3n = VADD(T1, T8);
161 TU = VFNMS(LDK(KP618033988), TT, TS);
162 T2j = VFMA(LDK(KP618033988), TS, TT);
163 TO = VADD(TK, TN);
164 TQ = VSUB(TN, TK);
165 T12 = VADD(T10, T11);
166 T18 = VSUB(T10, T11);
167 Ty = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
168 {
169 V T3r, T15, T13, Tf, Ti, T2k, TR, TP, T3s, T14;
170 {
171 V Td, Te, Tg, Th;
172 Td = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
173 Te = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
174 Tg = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
175 Th = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
176 TP = VFNMS(LDK(KP250000000), TO, TH);
177 T3r = VADD(TH, TO);
178 T2m = VFNMS(LDK(KP618033988), T17, T18);
179 T19 = VFMA(LDK(KP618033988), T18, T17);
180 T15 = VSUB(T12, TZ);
181 T13 = VADD(TZ, T12);
182 Tf = VADD(Td, Te);
183 Tn = VSUB(Td, Te);
184 To = VSUB(Th, Tg);
185 Ti = VADD(Tg, Th);
186 }
187 T2k = VFMA(LDK(KP559016994), TQ, TP);
188 TR = VFNMS(LDK(KP559016994), TQ, TP);
189 Tr = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
190 T3s = VADD(TW, T13);
191 T14 = VFNMS(LDK(KP250000000), T13, TW);
192 Tj = VADD(Tf, Ti);
193 Tl = VSUB(Tf, Ti);
194 T2O = VFNMS(LDK(KP667278218), T2k, T2j);
195 T2W = VFMA(LDK(KP603558818), T2j, T2k);
196 T2l = VFMA(LDK(KP066152395), T2k, T2j);
197 T2s = VFNMS(LDK(KP059835404), T2j, T2k);
198 TV = VFNMS(LDK(KP522847744), TU, TR);
199 T1i = VFMA(LDK(KP578046249), TR, TU);
200 T1K = VFNMS(LDK(KP494780565), TR, TU);
201 T1S = VFMA(LDK(KP447533225), TU, TR);
202 T16 = VFNMS(LDK(KP559016994), T15, T14);
203 T2n = VFMA(LDK(KP559016994), T15, T14);
204 T3z = VSUB(T3r, T3s);
205 T3t = VADD(T3r, T3s);
206 Ts = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
207 Tu = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
208 Tv = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
209 }
210 }
211 Tk = VFNMS(LDK(KP250000000), Tj, Tc);
212 T3o = VADD(Tc, Tj);
213 Tp = VFNMS(LDK(KP618033988), To, Tn);
214 T2g = VFMA(LDK(KP618033988), Tn, To);
215 T2N = VFMA(LDK(KP066152395), T2n, T2m);
216 T2V = VFNMS(LDK(KP059835404), T2m, T2n);
217 T2o = VFMA(LDK(KP869845200), T2n, T2m);
218 T2t = VFNMS(LDK(KP786782374), T2m, T2n);
219 T1a = VFNMS(LDK(KP893101515), T19, T16);
220 T1j = VFMA(LDK(KP987388751), T16, T19);
221 T1J = VFNMS(LDK(KP120146378), T19, T16);
222 T1R = VFMA(LDK(KP132830569), T16, T19);
223 Tz = VADD(Ts, Tr);
224 Tt = VSUB(Tr, Ts);
225 TA = VADD(Tv, Tu);
226 Tw = VSUB(Tu, Tv);
227 }
228 }
229 }
230 {
231 V T2p, T2I, T2u, T2C, Tx, T2d, T2X, T34, T2P, T3b, T2b, Tb, T2Q, T2Z, T2h;
232 V T2w, Tq, T1e, T1M, T1U, TE, T2c, T3q, T3y;
233 T2p = VFNMS(LDK(KP772036680), T2o, T2l);
234 T2I = VFMA(LDK(KP772036680), T2o, T2l);
235 T2u = VFMA(LDK(KP772036680), T2t, T2s);
236 T2C = VFNMS(LDK(KP772036680), T2t, T2s);
237 {
238 V TD, TB, Tm, T2f, T3p, TC;
239 Tx = VFMA(LDK(KP618033988), Tw, Tt);
240 T2d = VFNMS(LDK(KP618033988), Tt, Tw);
241 TD = VSUB(Tz, TA);
242 TB = VADD(Tz, TA);
243 Tm = VFMA(LDK(KP559016994), Tl, Tk);
244 T2f = VFNMS(LDK(KP559016994), Tl, Tk);
245 T2X = VFMA(LDK(KP845997307), T2W, T2V);
246 T34 = VFNMS(LDK(KP845997307), T2W, T2V);
247 T2P = VFNMS(LDK(KP845997307), T2O, T2N);
248 T3b = VFMA(LDK(KP845997307), T2O, T2N);
249 T2b = VFNMS(LDK(KP559016994), Ta, T9);
250 Tb = VFMA(LDK(KP559016994), Ta, T9);
251 T3p = VADD(Ty, TB);
252 TC = VFMS(LDK(KP250000000), TB, Ty);
253 T2Q = VFNMS(LDK(KP522847744), T2g, T2f);
254 T2Z = VFMA(LDK(KP578046249), T2f, T2g);
255 T2h = VFMA(LDK(KP893101515), T2g, T2f);
256 T2w = VFNMS(LDK(KP987388751), T2f, T2g);
257 Tq = VFNMS(LDK(KP244189809), Tp, Tm);
258 T1e = VFMA(LDK(KP269969613), Tm, Tp);
259 T1M = VFMA(LDK(KP667278218), Tm, Tp);
260 T1U = VFNMS(LDK(KP603558818), Tp, Tm);
261 TE = VFNMS(LDK(KP559016994), TD, TC);
262 T2c = VFMA(LDK(KP559016994), TD, TC);
263 T3q = VADD(T3o, T3p);
264 T3y = VSUB(T3o, T3p);
265 }
266 {
267 V T1Z, T25, T1P, T22, T1X, TG, T1b, T28, T1t, T1y, T1x, T1E, T1Q, T1Y;
268 {
269 V T26, T1L, T1T, TF, T1f, T1W, T3m, T3g, T2M, T2G, T39, T3j, T21, T1O, T20;
270 V T27;
271 T26 = VFMA(LDK(KP867381224), T1K, T1J);
272 T1L = VFNMS(LDK(KP867381224), T1K, T1J);
273 T20 = VFNMS(LDK(KP958953096), T1S, T1R);
274 T1T = VFMA(LDK(KP958953096), T1S, T1R);
275 {
276 V T2R, T2Y, T2e, T2v, T1N, T1V;
277 T2R = VFNMS(LDK(KP494780565), T2c, T2d);
278 T2Y = VFMA(LDK(KP447533225), T2d, T2c);
279 T2e = VFMA(LDK(KP120146378), T2d, T2c);
280 T2v = VFNMS(LDK(KP132830569), T2c, T2d);
281 TF = VFNMS(LDK(KP667278218), TE, Tx);
282 T1f = VFMA(LDK(KP603558818), Tx, TE);
283 T1N = VFMA(LDK(KP869845200), TE, Tx);
284 T1V = VFNMS(LDK(KP786782374), Tx, TE);
285 {
286 V T3A, T3C, T3w, T3u;
287 T3A = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T3z, T3y));
288 T3C = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T3y, T3z));
289 T3w = VSUB(T3q, T3t);
290 T3u = VADD(T3q, T3t);
291 {
292 V T2B, T2x, T2H, T2i;
293 T2B = VFMA(LDK(KP734762448), T2w, T2v);
294 T2x = VFNMS(LDK(KP734762448), T2w, T2v);
295 T2H = VFNMS(LDK(KP734762448), T2h, T2e);
296 T2i = VFMA(LDK(KP734762448), T2h, T2e);
297 {
298 V T30, T35, T3c, T2S, T3v;
299 T30 = VFNMS(LDK(KP921078979), T2Z, T2Y);
300 T35 = VFMA(LDK(KP921078979), T2Z, T2Y);
301 T3c = VFMA(LDK(KP982009705), T2R, T2Q);
302 T2S = VFNMS(LDK(KP982009705), T2R, T2Q);
303 T1W = VFMA(LDK(KP912575812), T1V, T1U);
304 T1Z = VFNMS(LDK(KP912575812), T1V, T1U);
305 T1O = VFMA(LDK(KP912575812), T1N, T1M);
306 T25 = VFNMS(LDK(KP912575812), T1N, T1M);
307 ST(&(xo[0]), VADD(T3u, T3n), ovs, &(xo[0]));
308 T3v = VFNMS(LDK(KP250000000), T3u, T3n);
309 {
310 V T2y, T2J, T2q, T2D;
311 T2y = VFMA(LDK(KP945422727), T2x, T2u);
312 T2J = VFMA(LDK(KP522616830), T2x, T2I);
313 T2q = VFMA(LDK(KP956723877), T2p, T2i);
314 T2D = VFNMS(LDK(KP522616830), T2i, T2C);
315 {
316 V T3e, T31, T36, T2T;
317 T3e = VFMA(LDK(KP906616052), T30, T2X);
318 T31 = VFNMS(LDK(KP906616052), T30, T2X);
319 T36 = VFNMS(LDK(KP923225144), T2S, T2P);
320 T2T = VFMA(LDK(KP923225144), T2S, T2P);
321 {
322 V T3k, T3d, T3x, T3B;
323 T3k = VFNMS(LDK(KP669429328), T3b, T3c);
324 T3d = VFMA(LDK(KP570584518), T3c, T3b);
325 T3x = VFMA(LDK(KP559016994), T3w, T3v);
326 T3B = VFNMS(LDK(KP559016994), T3w, T3v);
327 {
328 V T2A, T2K, T2r, T2E;
329 T2A = VMUL(LDK(KP998026728), VFMA(LDK(KP952936919), T2z, T2y));
330 T2K = VFNMS(LDK(KP690983005), T2J, T2u);
331 T2r = VFMA(LDK(KP992114701), T2q, T2b);
332 T2E = VFMA(LDK(KP763932022), T2D, T2p);
333 {
334 V T32, T3a, T37, T3h;
335 T32 = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T2z, T31));
336 T3a = VFMA(LDK(KP262346850), T31, T2z);
337 T37 = VFNMS(LDK(KP997675361), T36, T35);
338 T3h = VFNMS(LDK(KP904508497), T36, T34);
339 {
340 V T2U, T33, T3l, T3f;
341 T2U = VFMA(LDK(KP949179823), T2T, T2b);
342 T33 = VFNMS(LDK(KP237294955), T2T, T2b);
343 T3l = VFNMS(LDK(KP669429328), T3e, T3k);
344 T3f = VFMA(LDK(KP618033988), T3e, T3d);
345 ST(&(xo[WS(os, 20)]), VFMAI(T3A, T3x), ovs, &(xo[0]));
346 ST(&(xo[WS(os, 5)]), VFNMSI(T3A, T3x), ovs, &(xo[WS(os, 1)]));
347 ST(&(xo[WS(os, 15)]), VFNMSI(T3C, T3B), ovs, &(xo[WS(os, 1)]));
348 ST(&(xo[WS(os, 10)]), VFMAI(T3C, T3B), ovs, &(xo[0]));
349 {
350 V T2L, T2F, T38, T3i;
351 T2L = VFMA(LDK(KP855719849), T2K, T2H);
352 ST(&(xo[WS(os, 22)]), VFMAI(T2A, T2r), ovs, &(xo[0]));
353 ST(&(xo[WS(os, 3)]), VFNMSI(T2A, T2r), ovs, &(xo[WS(os, 1)]));
354 T2F = VFNMS(LDK(KP855719849), T2E, T2B);
355 T38 = VFMA(LDK(KP560319534), T37, T34);
356 T3i = VFNMS(LDK(KP681693190), T3h, T35);
357 ST(&(xo[WS(os, 23)]), VFMAI(T32, T2U), ovs, &(xo[WS(os, 1)]));
358 ST(&(xo[WS(os, 2)]), VFNMSI(T32, T2U), ovs, &(xo[0]));
359 T3m = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T3l, T3a));
360 T3g = VMUL(LDK(KP951056516), VFNMS(LDK(KP949179823), T3f, T3a));
361 T2M = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T2L, T2z));
362 T2G = VFMA(LDK(KP897376177), T2F, T2b);
363 T39 = VFNMS(LDK(KP949179823), T38, T33);
364 T3j = VFNMS(LDK(KP860541664), T3i, T33);
365 T21 = VFMA(LDK(KP447417479), T1O, T20);
366 }
367 }
368 }
369 }
370 }
371 }
372 }
373 }
374 }
375 }
376 }
377 T1P = VFNMS(LDK(KP809385824), T1O, T1L);
378 ST(&(xo[WS(os, 17)]), VFMAI(T2M, T2G), ovs, &(xo[WS(os, 1)]));
379 ST(&(xo[WS(os, 8)]), VFNMSI(T2M, T2G), ovs, &(xo[0]));
380 ST(&(xo[WS(os, 12)]), VFMAI(T3g, T39), ovs, &(xo[0]));
381 ST(&(xo[WS(os, 13)]), VFNMSI(T3g, T39), ovs, &(xo[WS(os, 1)]));
382 ST(&(xo[WS(os, 7)]), VFMAI(T3m, T3j), ovs, &(xo[WS(os, 1)]));
383 ST(&(xo[WS(os, 18)]), VFNMSI(T3m, T3j), ovs, &(xo[0]));
384 T22 = VFMA(LDK(KP690983005), T21, T1L);
385 T27 = VFMA(LDK(KP447417479), T1W, T26);
386 T1X = VFMA(LDK(KP894834959), T1W, T1T);
387 {
388 V T1r, T1s, T1v, T1w;
389 T1r = VFNMS(LDK(KP916574801), T1f, T1e);
390 T1g = VFMA(LDK(KP916574801), T1f, T1e);
391 T1k = VFNMS(LDK(KP831864738), T1j, T1i);
392 T1s = VFMA(LDK(KP831864738), T1j, T1i);
393 T1v = VFNMS(LDK(KP829049696), TF, Tq);
394 TG = VFMA(LDK(KP829049696), TF, Tq);
395 T1b = VFMA(LDK(KP831864738), T1a, TV);
396 T1w = VFNMS(LDK(KP831864738), T1a, TV);
397 T28 = VFNMS(LDK(KP763932022), T27, T1T);
398 T1t = VFMA(LDK(KP904730450), T1s, T1r);
399 T1y = VFNMS(LDK(KP904730450), T1s, T1r);
400 T1x = VFMA(LDK(KP559154169), T1w, T1v);
401 T1E = VFNMS(LDK(KP683113946), T1v, T1w);
402 }
403 }
404 T1Q = VFNMS(LDK(KP992114701), T1P, Tb);
405 T1Y = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T1X, T1q));
406 {
407 V T1u, T1F, T1z, T1h, T1c, T23, T29;
408 T23 = VFNMS(LDK(KP999544308), T22, T1Z);
409 T29 = VFNMS(LDK(KP999544308), T28, T25);
410 T1I = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1t, T1q));
411 T1u = VFNMS(LDK(KP242145790), T1t, T1q);
412 T1F = VFMA(LDK(KP617882369), T1y, T1E);
413 T1z = VFMA(LDK(KP559016994), T1y, T1x);
414 T1h = VFNMS(LDK(KP904730450), T1b, TG);
415 T1c = VFMA(LDK(KP904730450), T1b, TG);
416 ST(&(xo[WS(os, 21)]), VFNMSI(T1Y, T1Q), ovs, &(xo[WS(os, 1)]));
417 ST(&(xo[WS(os, 4)]), VFMAI(T1Y, T1Q), ovs, &(xo[0]));
418 T24 = VFNMS(LDK(KP803003575), T23, Tb);
419 T2a = VMUL(LDK(KP951056516), VFNMS(LDK(KP803003575), T29, T1q));
420 T1G = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T1F, T1u));
421 T1A = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1z, T1u));
422 T1l = VFNMS(LDK(KP904730450), T1k, T1h);
423 T1B = VADD(T1g, T1h);
424 T1H = VFMA(LDK(KP968583161), T1c, Tb);
425 T1d = VFNMS(LDK(KP242145790), T1c, Tb);
426 }
427 }
428 }
429 }
430 ST(&(xo[WS(os, 9)]), VFMAI(T2a, T24), ovs, &(xo[WS(os, 1)]));
431 ST(&(xo[WS(os, 16)]), VFNMSI(T2a, T24), ovs, &(xo[0]));
432 {
433 V T1m, T1C, T1n, T1D;
434 T1m = VFNMS(LDK(KP618033988), T1l, T1g);
435 T1C = VFNMS(LDK(KP683113946), T1B, T1k);
436 ST(&(xo[WS(os, 24)]), VFMAI(T1I, T1H), ovs, &(xo[0]));
437 ST(&(xo[WS(os, 1)]), VFNMSI(T1I, T1H), ovs, &(xo[WS(os, 1)]));
438 T1n = VFNMS(LDK(KP876091699), T1m, T1d);
439 T1D = VFMA(LDK(KP792626838), T1C, T1d);
440 ST(&(xo[WS(os, 19)]), VFMAI(T1A, T1n), ovs, &(xo[WS(os, 1)]));
441 ST(&(xo[WS(os, 6)]), VFNMSI(T1A, T1n), ovs, &(xo[0]));
442 ST(&(xo[WS(os, 14)]), VFMAI(T1G, T1D), ovs, &(xo[0]));
443 ST(&(xo[WS(os, 11)]), VFNMSI(T1G, T1D), ovs, &(xo[WS(os, 1)]));
444 }
445 }
446 }
447 VLEAVE();
448 }
449
450 static const kdft_desc desc = { 25, XSIMD_STRING("n1fv_25"), {43, 12, 181, 0}, &GENUS, 0, 0, 0, 0 };
451
452 void XSIMD(codelet_n1fv_25) (planner *p) {
453 X(kdft_register) (p, n1fv_25, &desc);
454 }
455
456 #else /* HAVE_FMA */
457
458 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 25 -name n1fv_25 -include n1f.h */
459
460 /*
461 * This function contains 224 FP additions, 140 FP multiplications,
462 * (or, 146 additions, 62 multiplications, 78 fused multiply/add),
463 * 115 stack variables, 40 constants, and 50 memory accesses
464 */
465 #include "n1f.h"
466
467 static void n1fv_25(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
468 {
469 DVK(KP809016994, +0.809016994374947424102293417182819058860154590);
470 DVK(KP309016994, +0.309016994374947424102293417182819058860154590);
471 DVK(KP770513242, +0.770513242775789230803009636396177847271667672);
472 DVK(KP1_274847979, +1.274847979497379420353425623352032390869834596);
473 DVK(KP992114701, +0.992114701314477831049793042785778521453036709);
474 DVK(KP250666467, +0.250666467128608490746237519633017587885836494);
475 DVK(KP637423989, +0.637423989748689710176712811676016195434917298);
476 DVK(KP1_541026485, +1.541026485551578461606019272792355694543335344);
477 DVK(KP125333233, +0.125333233564304245373118759816508793942918247);
478 DVK(KP1_984229402, +1.984229402628955662099586085571557042906073418);
479 DVK(KP248689887, +0.248689887164854788242283746006447968417567406);
480 DVK(KP1_937166322, +1.937166322257262238980336750929471627672024806);
481 DVK(KP497379774, +0.497379774329709576484567492012895936835134813);
482 DVK(KP968583161, +0.968583161128631119490168375464735813836012403);
483 DVK(KP904827052, +0.904827052466019527713668647932697593970413911);
484 DVK(KP851558583, +0.851558583130145297725004891488503407959946084);
485 DVK(KP425779291, +0.425779291565072648862502445744251703979973042);
486 DVK(KP1_809654104, +1.809654104932039055427337295865395187940827822);
487 DVK(KP844327925, +0.844327925502015078548558063966681505381659241);
488 DVK(KP1_071653589, +1.071653589957993236542617535735279956127150691);
489 DVK(KP481753674, +0.481753674101715274987191502872129653528542010);
490 DVK(KP1_752613360, +1.752613360087727174616231807844125166798128477);
491 DVK(KP535826794, +0.535826794978996618271308767867639978063575346);
492 DVK(KP1_688655851, +1.688655851004030157097116127933363010763318483);
493 DVK(KP963507348, +0.963507348203430549974383005744259307057084020);
494 DVK(KP876306680, +0.876306680043863587308115903922062583399064238);
495 DVK(KP998026728, +0.998026728428271561952336806863450553336905220);
496 DVK(KP125581039, +0.125581039058626752152356449131262266244969664);
497 DVK(KP684547105, +0.684547105928688673732283357621209269889519233);
498 DVK(KP1_457937254, +1.457937254842823046293460638110518222745143328);
499 DVK(KP062790519, +0.062790519529313376076178224565631133122484832);
500 DVK(KP1_996053456, +1.996053456856543123904673613726901106673810439);
501 DVK(KP1_369094211, +1.369094211857377347464566715242418539779038465);
502 DVK(KP728968627, +0.728968627421411523146730319055259111372571664);
503 DVK(KP293892626, +0.293892626146236564584352977319536384298826219);
504 DVK(KP475528258, +0.475528258147576786058219666689691071702849317);
505 DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
506 DVK(KP587785252, +0.587785252292473129168705954639072768597652438);
507 DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
508 DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
509 {
510 INT i;
511 const R *xi;
512 R *xo;
513 xi = ri;
514 xo = ro;
515 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(50, is), MAKE_VOLATILE_STRIDE(50, os)) {
516 V T7, T1g, T26, Ta, T2R, T2N, T2O, T2P, T19, T1Y, T16, T1Z, T1a, T2v, T1l;
517 V T2m, TU, T21, TR, T22, TV, T2u, T1k, T2l, T2K, T2L, T2M, TE, T1R, TB;
518 V T1S, TF, T2r, T1i, T2j, Tp, T1U, Tm, T1V, Tq, T2s, T1h, T2i;
519 {
520 V T8, T6, T1f, T3, T1e, T25, T9;
521 T8 = LD(&(xi[0]), ivs, &(xi[0]));
522 {
523 V T4, T5, T1, T2;
524 T4 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
525 T5 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
526 T6 = VADD(T4, T5);
527 T1f = VSUB(T4, T5);
528 T1 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
529 T2 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
530 T3 = VADD(T1, T2);
531 T1e = VSUB(T1, T2);
532 }
533 T7 = VMUL(LDK(KP559016994), VSUB(T3, T6));
534 T1g = VFMA(LDK(KP951056516), T1e, VMUL(LDK(KP587785252), T1f));
535 T25 = VMUL(LDK(KP951056516), T1f);
536 T26 = VFNMS(LDK(KP587785252), T1e, T25);
537 T9 = VADD(T3, T6);
538 Ta = VFNMS(LDK(KP250000000), T9, T8);
539 T2R = VADD(T8, T9);
540 }
541 {
542 V TO, T13, TN, TT, TP, TS, T12, T18, T14, T17, T15, TQ;
543 TO = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
544 T13 = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
545 {
546 V TH, TI, TJ, TK, TL, TM;
547 TH = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
548 TI = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
549 TJ = VADD(TH, TI);
550 TK = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
551 TL = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
552 TM = VADD(TK, TL);
553 TN = VMUL(LDK(KP559016994), VSUB(TJ, TM));
554 TT = VSUB(TK, TL);
555 TP = VADD(TJ, TM);
556 TS = VSUB(TH, TI);
557 }
558 {
559 V TW, TX, TY, TZ, T10, T11;
560 TW = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
561 TX = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
562 TY = VADD(TW, TX);
563 TZ = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
564 T10 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
565 T11 = VADD(TZ, T10);
566 T12 = VMUL(LDK(KP559016994), VSUB(TY, T11));
567 T18 = VSUB(TZ, T10);
568 T14 = VADD(TY, T11);
569 T17 = VSUB(TW, TX);
570 }
571 T2N = VADD(TO, TP);
572 T2O = VADD(T13, T14);
573 T2P = VADD(T2N, T2O);
574 T19 = VFMA(LDK(KP475528258), T17, VMUL(LDK(KP293892626), T18));
575 T1Y = VFNMS(LDK(KP293892626), T17, VMUL(LDK(KP475528258), T18));
576 T15 = VFNMS(LDK(KP250000000), T14, T13);
577 T16 = VADD(T12, T15);
578 T1Z = VSUB(T15, T12);
579 T1a = VFNMS(LDK(KP1_369094211), T19, VMUL(LDK(KP728968627), T16));
580 T2v = VFMA(LDK(KP1_996053456), T1Y, VMUL(LDK(KP062790519), T1Z));
581 T1l = VFMA(LDK(KP1_457937254), T19, VMUL(LDK(KP684547105), T16));
582 T2m = VFNMS(LDK(KP998026728), T1Z, VMUL(LDK(KP125581039), T1Y));
583 TU = VFMA(LDK(KP475528258), TS, VMUL(LDK(KP293892626), TT));
584 T21 = VFNMS(LDK(KP293892626), TS, VMUL(LDK(KP475528258), TT));
585 TQ = VFNMS(LDK(KP250000000), TP, TO);
586 TR = VADD(TN, TQ);
587 T22 = VSUB(TQ, TN);
588 TV = VFNMS(LDK(KP963507348), TU, VMUL(LDK(KP876306680), TR));
589 T2u = VFMA(LDK(KP1_688655851), T21, VMUL(LDK(KP535826794), T22));
590 T1k = VFMA(LDK(KP1_752613360), TU, VMUL(LDK(KP481753674), TR));
591 T2l = VFNMS(LDK(KP844327925), T22, VMUL(LDK(KP1_071653589), T21));
592 }
593 {
594 V Tj, Ty, Ti, To, Tk, Tn, Tx, TD, Tz, TC, TA, Tl;
595 Tj = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
596 Ty = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
597 {
598 V Tc, Td, Te, Tf, Tg, Th;
599 Tc = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
600 Td = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
601 Te = VADD(Tc, Td);
602 Tf = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
603 Tg = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
604 Th = VADD(Tf, Tg);
605 Ti = VMUL(LDK(KP559016994), VSUB(Te, Th));
606 To = VSUB(Tf, Tg);
607 Tk = VADD(Te, Th);
608 Tn = VSUB(Tc, Td);
609 }
610 {
611 V Tr, Ts, Tt, Tu, Tv, Tw;
612 Tr = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
613 Ts = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
614 Tt = VADD(Tr, Ts);
615 Tu = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
616 Tv = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
617 Tw = VADD(Tu, Tv);
618 Tx = VMUL(LDK(KP559016994), VSUB(Tt, Tw));
619 TD = VSUB(Tu, Tv);
620 Tz = VADD(Tt, Tw);
621 TC = VSUB(Tr, Ts);
622 }
623 T2K = VADD(Tj, Tk);
624 T2L = VADD(Ty, Tz);
625 T2M = VADD(T2K, T2L);
626 TE = VFMA(LDK(KP475528258), TC, VMUL(LDK(KP293892626), TD));
627 T1R = VFNMS(LDK(KP293892626), TC, VMUL(LDK(KP475528258), TD));
628 TA = VFNMS(LDK(KP250000000), Tz, Ty);
629 TB = VADD(Tx, TA);
630 T1S = VSUB(TA, Tx);
631 TF = VFNMS(LDK(KP1_688655851), TE, VMUL(LDK(KP535826794), TB));
632 T2r = VFNMS(LDK(KP425779291), T1S, VMUL(LDK(KP1_809654104), T1R));
633 T1i = VFMA(LDK(KP1_071653589), TE, VMUL(LDK(KP844327925), TB));
634 T2j = VFMA(LDK(KP851558583), T1R, VMUL(LDK(KP904827052), T1S));
635 Tp = VFMA(LDK(KP475528258), Tn, VMUL(LDK(KP293892626), To));
636 T1U = VFNMS(LDK(KP293892626), Tn, VMUL(LDK(KP475528258), To));
637 Tl = VFNMS(LDK(KP250000000), Tk, Tj);
638 Tm = VADD(Ti, Tl);
639 T1V = VSUB(Tl, Ti);
640 Tq = VFNMS(LDK(KP497379774), Tp, VMUL(LDK(KP968583161), Tm));
641 T2s = VFMA(LDK(KP963507348), T1U, VMUL(LDK(KP876306680), T1V));
642 T1h = VFMA(LDK(KP1_937166322), Tp, VMUL(LDK(KP248689887), Tm));
643 T2i = VFNMS(LDK(KP481753674), T1V, VMUL(LDK(KP1_752613360), T1U));
644 }
645 {
646 V T2Q, T2S, T2T, T2X, T2Y, T2V, T2W, T2Z, T2U;
647 T2Q = VMUL(LDK(KP559016994), VSUB(T2M, T2P));
648 T2S = VADD(T2M, T2P);
649 T2T = VFNMS(LDK(KP250000000), T2S, T2R);
650 T2V = VSUB(T2K, T2L);
651 T2W = VSUB(T2N, T2O);
652 T2X = VBYI(VFMA(LDK(KP951056516), T2V, VMUL(LDK(KP587785252), T2W)));
653 T2Y = VBYI(VFNMS(LDK(KP587785252), T2V, VMUL(LDK(KP951056516), T2W)));
654 ST(&(xo[0]), VADD(T2R, T2S), ovs, &(xo[0]));
655 T2Z = VSUB(T2T, T2Q);
656 ST(&(xo[WS(os, 10)]), VADD(T2Y, T2Z), ovs, &(xo[0]));
657 ST(&(xo[WS(os, 15)]), VSUB(T2Z, T2Y), ovs, &(xo[WS(os, 1)]));
658 T2U = VADD(T2Q, T2T);
659 ST(&(xo[WS(os, 5)]), VSUB(T2U, T2X), ovs, &(xo[WS(os, 1)]));
660 ST(&(xo[WS(os, 20)]), VADD(T2X, T2U), ovs, &(xo[0]));
661 }
662 {
663 V T2t, T2y, T2z, T2w, T1T, T1W, T1X, T2c, T2d, T2e, T29, T2a, T2b, T20, T23;
664 V T24, T2p, T2o, T2q, T28, T2D, T2C, T2E, T2x, T2F;
665 T2t = VSUB(T2r, T2s);
666 T2y = VADD(T2i, T2j);
667 T2z = VSUB(T2l, T2m);
668 T2w = VSUB(T2u, T2v);
669 T1T = VFNMS(LDK(KP125333233), T1S, VMUL(LDK(KP1_984229402), T1R));
670 T1W = VFMA(LDK(KP1_457937254), T1U, VMUL(LDK(KP684547105), T1V));
671 T1X = VSUB(T1T, T1W);
672 T2c = VFNMS(LDK(KP1_996053456), T21, VMUL(LDK(KP062790519), T22));
673 T2d = VFMA(LDK(KP1_541026485), T1Y, VMUL(LDK(KP637423989), T1Z));
674 T2e = VSUB(T2c, T2d);
675 T29 = VFNMS(LDK(KP1_369094211), T1U, VMUL(LDK(KP728968627), T1V));
676 T2a = VFMA(LDK(KP250666467), T1R, VMUL(LDK(KP992114701), T1S));
677 T2b = VSUB(T29, T2a);
678 T20 = VFNMS(LDK(KP770513242), T1Z, VMUL(LDK(KP1_274847979), T1Y));
679 T23 = VFMA(LDK(KP125581039), T21, VMUL(LDK(KP998026728), T22));
680 T24 = VSUB(T20, T23);
681 {
682 V T2k, T2n, T2A, T2B;
683 T2k = VSUB(T2i, T2j);
684 T2n = VADD(T2l, T2m);
685 T2p = VADD(T2k, T2n);
686 T2o = VMUL(LDK(KP559016994), VSUB(T2k, T2n));
687 T2q = VFNMS(LDK(KP250000000), T2p, T26);
688 T28 = VSUB(Ta, T7);
689 T2A = VADD(T2s, T2r);
690 T2B = VADD(T2u, T2v);
691 T2D = VADD(T2A, T2B);
692 T2C = VMUL(LDK(KP559016994), VSUB(T2A, T2B));
693 T2E = VFNMS(LDK(KP250000000), T2D, T28);
694 }
695 {
696 V T2I, T2J, T27, T2f;
697 T2I = VBYI(VADD(T26, T2p));
698 T2J = VADD(T28, T2D);
699 ST(&(xo[WS(os, 2)]), VADD(T2I, T2J), ovs, &(xo[0]));
700 ST(&(xo[WS(os, 23)]), VSUB(T2J, T2I), ovs, &(xo[WS(os, 1)]));
701 T27 = VBYI(VSUB(VADD(T1X, T24), T26));
702 T2f = VADD(T28, VADD(T2b, T2e));
703 ST(&(xo[WS(os, 3)]), VADD(T27, T2f), ovs, &(xo[WS(os, 1)]));
704 ST(&(xo[WS(os, 22)]), VSUB(T2f, T27), ovs, &(xo[0]));
705 }
706 T2x = VBYI(VADD(T2o, VADD(T2q, VFNMS(LDK(KP587785252), T2w, VMUL(LDK(KP951056516), T2t)))));
707 T2F = VFMA(LDK(KP951056516), T2y, VFMA(LDK(KP587785252), T2z, VADD(T2C, T2E)));
708 ST(&(xo[WS(os, 7)]), VADD(T2x, T2F), ovs, &(xo[WS(os, 1)]));
709 ST(&(xo[WS(os, 18)]), VSUB(T2F, T2x), ovs, &(xo[0]));
710 {
711 V T2G, T2H, T2g, T2h;
712 T2G = VBYI(VADD(T2q, VSUB(VFMA(LDK(KP587785252), T2t, VMUL(LDK(KP951056516), T2w)), T2o)));
713 T2H = VFMA(LDK(KP587785252), T2y, VSUB(VFNMS(LDK(KP951056516), T2z, T2E), T2C));
714 ST(&(xo[WS(os, 12)]), VADD(T2G, T2H), ovs, &(xo[0]));
715 ST(&(xo[WS(os, 13)]), VSUB(T2H, T2G), ovs, &(xo[WS(os, 1)]));
716 T2g = VFMA(LDK(KP309016994), T2b, VFNMS(LDK(KP809016994), T2e, VFNMS(LDK(KP587785252), VADD(T23, T20), VFNMS(LDK(KP951056516), VADD(T1W, T1T), T28))));
717 T2h = VBYI(VSUB(VFNMS(LDK(KP587785252), VADD(T2c, T2d), VFNMS(LDK(KP809016994), T24, VFNMS(LDK(KP951056516), VADD(T29, T2a), VMUL(LDK(KP309016994), T1X)))), T26));
718 ST(&(xo[WS(os, 17)]), VSUB(T2g, T2h), ovs, &(xo[WS(os, 1)]));
719 ST(&(xo[WS(os, 8)]), VADD(T2g, T2h), ovs, &(xo[0]));
720 }
721 }
722 {
723 V T1p, T1u, T1w, T1q, T1B, T1C, T1D, T1L, T1M, T1N, T1I, T1J, T1K, T1E, T1F;
724 V T1G, T1n, T1r, T1s, Tb, T1c, T1v, T1x, T1t, T1y;
725 T1p = VSUB(TF, Tq);
726 T1u = VSUB(T1i, T1h);
727 T1w = VSUB(T1l, T1k);
728 T1q = VSUB(TV, T1a);
729 T1B = VFMA(LDK(KP1_688655851), Tp, VMUL(LDK(KP535826794), Tm));
730 T1C = VFMA(LDK(KP1_541026485), TE, VMUL(LDK(KP637423989), TB));
731 T1D = VSUB(T1B, T1C);
732 T1L = VFMA(LDK(KP851558583), TU, VMUL(LDK(KP904827052), TR));
733 T1M = VFMA(LDK(KP1_984229402), T19, VMUL(LDK(KP125333233), T16));
734 T1N = VADD(T1L, T1M);
735 T1I = VFNMS(LDK(KP844327925), Tm, VMUL(LDK(KP1_071653589), Tp));
736 T1J = VFNMS(LDK(KP1_274847979), TE, VMUL(LDK(KP770513242), TB));
737 T1K = VADD(T1I, T1J);
738 T1E = VFNMS(LDK(KP425779291), TR, VMUL(LDK(KP1_809654104), TU));
739 T1F = VFNMS(LDK(KP992114701), T16, VMUL(LDK(KP250666467), T19));
740 T1G = VADD(T1E, T1F);
741 {
742 V T1j, T1m, TG, T1b;
743 T1j = VADD(T1h, T1i);
744 T1m = VADD(T1k, T1l);
745 T1n = VADD(T1j, T1m);
746 T1r = VFMS(LDK(KP250000000), T1n, T1g);
747 T1s = VMUL(LDK(KP559016994), VSUB(T1m, T1j));
748 Tb = VADD(T7, Ta);
749 TG = VADD(Tq, TF);
750 T1b = VADD(TV, T1a);
751 T1c = VADD(TG, T1b);
752 T1v = VFNMS(LDK(KP250000000), T1c, Tb);
753 T1x = VMUL(LDK(KP559016994), VSUB(TG, T1b));
754 }
755 {
756 V T1d, T1o, T1H, T1O;
757 T1d = VADD(Tb, T1c);
758 T1o = VBYI(VADD(T1g, T1n));
759 ST(&(xo[WS(os, 1)]), VSUB(T1d, T1o), ovs, &(xo[WS(os, 1)]));
760 ST(&(xo[WS(os, 24)]), VADD(T1d, T1o), ovs, &(xo[0]));
761 T1H = VADD(Tb, VADD(T1D, T1G));
762 T1O = VBYI(VADD(T1g, VSUB(T1K, T1N)));
763 ST(&(xo[WS(os, 21)]), VSUB(T1H, T1O), ovs, &(xo[WS(os, 1)]));
764 ST(&(xo[WS(os, 4)]), VADD(T1H, T1O), ovs, &(xo[0]));
765 }
766 T1t = VBYI(VADD(VFMA(LDK(KP587785252), T1p, VMUL(LDK(KP951056516), T1q)), VSUB(T1r, T1s)));
767 T1y = VFMA(LDK(KP587785252), T1u, VFNMS(LDK(KP951056516), T1w, VSUB(T1v, T1x)));
768 ST(&(xo[WS(os, 11)]), VADD(T1t, T1y), ovs, &(xo[WS(os, 1)]));
769 ST(&(xo[WS(os, 14)]), VSUB(T1y, T1t), ovs, &(xo[0]));
770 {
771 V T1z, T1A, T1P, T1Q;
772 T1z = VBYI(VADD(VFNMS(LDK(KP587785252), T1q, VMUL(LDK(KP951056516), T1p)), VADD(T1r, T1s)));
773 T1A = VFMA(LDK(KP951056516), T1u, VADD(T1x, VFMA(LDK(KP587785252), T1w, T1v)));
774 ST(&(xo[WS(os, 6)]), VADD(T1z, T1A), ovs, &(xo[0]));
775 ST(&(xo[WS(os, 19)]), VSUB(T1A, T1z), ovs, &(xo[WS(os, 1)]));
776 T1P = VBYI(VADD(T1g, VFMA(LDK(KP309016994), T1K, VFMA(LDK(KP587785252), VSUB(T1F, T1E), VFNMS(LDK(KP951056516), VADD(T1B, T1C), VMUL(LDK(KP809016994), T1N))))));
777 T1Q = VFMA(LDK(KP309016994), T1D, VFMA(LDK(KP951056516), VSUB(T1I, T1J), VFMA(LDK(KP587785252), VSUB(T1M, T1L), VFNMS(LDK(KP809016994), T1G, Tb))));
778 ST(&(xo[WS(os, 9)]), VADD(T1P, T1Q), ovs, &(xo[WS(os, 1)]));
779 ST(&(xo[WS(os, 16)]), VSUB(T1Q, T1P), ovs, &(xo[0]));
780 }
781 }
782 }
783 }
784 VLEAVE();
785 }
786
787 static const kdft_desc desc = { 25, XSIMD_STRING("n1fv_25"), {146, 62, 78, 0}, &GENUS, 0, 0, 0, 0 };
788
789 void XSIMD(codelet_n1fv_25) (planner *p) {
790 X(kdft_register) (p, n1fv_25, &desc);
791 }
792
793 #endif /* HAVE_FMA */