← Back to branch summary

~ubuntu-branches/ubuntu/utopic/fftw3/utopic

« back to all changes in this revision

Viewing changes to dft/simd/codelets/n1fv_20.c

  • Committer: Package Import Robot
  • Author(s): Matthias Klose
  • Date: 2011-12-14 13:21:22 UTC
  • mfrom: (3.1.5 sid)
  • Revision ID: package-import@ubuntu.com-20111214132122-l4avyl2kkr7vq5aj
Tags: 3.3-1ubuntu1
* Merge with Debian; remaining changes:
  - Revert the ARM workaround.

Show diffs side-by-side

added added

removed removed

Lines of Context:
dft/simd/codelets/n1fv_20.c
1
 
/*
2
 
 * Copyright (c) 2003, 2007-8 Matteo Frigo
3
 
 * Copyright (c) 2003, 2007-8 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., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
18
 
 *
19
 
 */
20
 
 
21
 
/* This file was automatically generated --- DO NOT EDIT */
22
 
/* Generated on Sun Jul 12 06:39:13 EDT 2009 */
23
 
 
24
 
#include "codelet-dft.h"
25
 
 
26
 
#ifdef HAVE_FMA
27
 
 
28
 
/* Generated by: ../../../genfft/gen_notw_c -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 20 -name n1fv_20 -include n1f.h */
29
 
 
30
 
/*
31
 
 * This function contains 104 FP additions, 50 FP multiplications,
32
 
 * (or, 58 additions, 4 multiplications, 46 fused multiply/add),
33
 
 * 71 stack variables, 4 constants, and 40 memory accesses
34
 
 */
35
 
#include "n1f.h"
36
 
 
37
 
static void n1fv_20(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38
 
{
39
 
     DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
40
 
     DVK(KP618033988, +0.618033988749894848204586834365638117720309180);
41
 
     DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
42
 
     DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
43
 
     INT i;
44
 
     const R *xi;
45
 
     R *xo;
46
 
     xi = ri;
47
 
     xo = ro;
48
 
     for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
49
 
          V TU, TI, TP, TX, TM, TW, TT, TF;
50
 
          {
51
 
               V T3, Tm, T1r, T13, Ta, TN, TH, TA, TG, Tt, Th, TO, T1u, T1C, T1n;
52
 
               V T1a, T1m, T1h, T1x, T1D, TE, Ti;
53
 
               {
54
 
                    V T1, T2, Tk, Tl;
55
 
                    T1 = LD(&(xi[0]), ivs, &(xi[0]));
56
 
                    T2 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
57
 
                    Tk = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
58
 
                    Tl = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
59
 
                    {
60
 
                         V T14, T6, T1c, Tw, Tn, T1f, Tz, T17, T9, To, Tq, T1b, Td, Tr, Te;
61
 
                         V Tf, T15, Tp;
62
 
                         {
63
 
                              V Tx, Ty, T7, T8, Tb, Tc;
64
 
                              {
65
 
                                   V T4, T5, Tu, Tv, T11, T12;
66
 
                                   T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
67
 
                                   T5 = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
68
 
                                   Tu = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
69
 
                                   Tv = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
70
 
                                   Tx = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
71
 
                                   T3 = VSUB(T1, T2);
72
 
                                   T11 = VADD(T1, T2);
73
 
                                   Tm = VSUB(Tk, Tl);
74
 
                                   T12 = VADD(Tk, Tl);
75
 
                                   T14 = VADD(T4, T5);
76
 
                                   T6 = VSUB(T4, T5);
77
 
                                   T1c = VADD(Tu, Tv);
78
 
                                   Tw = VSUB(Tu, Tv);
79
 
                                   Ty = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
80
 
                                   T7 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
81
 
                                   T8 = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
82
 
                                   T1r = VADD(T11, T12);
83
 
                                   T13 = VSUB(T11, T12);
84
 
                              }
85
 
                              Tb = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
86
 
                              Tc = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
87
 
                              Tn = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
88
 
                              T1f = VADD(Tx, Ty);
89
 
                              Tz = VSUB(Tx, Ty);
90
 
                              T17 = VADD(T7, T8);
91
 
                              T9 = VSUB(T7, T8);
92
 
                              To = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
93
 
                              Tq = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
94
 
                              T1b = VADD(Tb, Tc);
95
 
                              Td = VSUB(Tb, Tc);
96
 
                              Tr = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
97
 
                              Te = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
98
 
                              Tf = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
99
 
                         }
100
 
                         Ta = VADD(T6, T9);
101
 
                         TN = VSUB(T6, T9);
102
 
                         T15 = VADD(Tn, To);
103
 
                         Tp = VSUB(Tn, To);
104
 
                         TH = VSUB(Tz, Tw);
105
 
                         TA = VADD(Tw, Tz);
106
 
                         {
107
 
                              V T1d, T1v, T18, Ts, T1e, Tg, T16, T1s;
108
 
                              T1d = VSUB(T1b, T1c);
109
 
                              T1v = VADD(T1b, T1c);
110
 
                              T18 = VADD(Tq, Tr);
111
 
                              Ts = VSUB(Tq, Tr);
112
 
                              T1e = VADD(Te, Tf);
113
 
                              Tg = VSUB(Te, Tf);
114
 
                              T16 = VSUB(T14, T15);
115
 
                              T1s = VADD(T14, T15);
116
 
                              {
117
 
                                   V T1t, T19, T1w, T1g;
118
 
                                   T1t = VADD(T17, T18);
119
 
                                   T19 = VSUB(T17, T18);
120
 
                                   TG = VSUB(Ts, Tp);
121
 
                                   Tt = VADD(Tp, Ts);
122
 
                                   T1w = VADD(T1e, T1f);
123
 
                                   T1g = VSUB(T1e, T1f);
124
 
                                   Th = VADD(Td, Tg);
125
 
                                   TO = VSUB(Td, Tg);
126
 
                                   T1u = VADD(T1s, T1t);
127
 
                                   T1C = VSUB(T1s, T1t);
128
 
                                   T1n = VSUB(T16, T19);
129
 
                                   T1a = VADD(T16, T19);
130
 
                                   T1m = VSUB(T1d, T1g);
131
 
                                   T1h = VADD(T1d, T1g);
132
 
                                   T1x = VADD(T1v, T1w);
133
 
                                   T1D = VSUB(T1v, T1w);
134
 
                              }
135
 
                         }
136
 
                    }
137
 
               }
138
 
               TE = VSUB(Ta, Th);
139
 
               Ti = VADD(Ta, Th);
140
 
               {
141
 
                    V TL, T1k, T1A, Tj, TD, T1E, T1G, TK, TC, T1j, T1z, T1i, T1y, TB;
142
 
                    TL = VSUB(TA, Tt);
143
 
                    TB = VADD(Tt, TA);
144
 
                    T1i = VADD(T1a, T1h);
145
 
                    T1k = VSUB(T1a, T1h);
146
 
                    T1y = VADD(T1u, T1x);
147
 
                    T1A = VSUB(T1u, T1x);
148
 
                    Tj = VADD(T3, Ti);
149
 
                    TD = VFNMS(LDK(KP250000000), Ti, T3);
150
 
                    T1E = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T1D, T1C));
151
 
                    T1G = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T1C, T1D));
152
 
                    TK = VFNMS(LDK(KP250000000), TB, Tm);
153
 
                    TC = VADD(Tm, TB);
154
 
                    T1j = VFNMS(LDK(KP250000000), T1i, T13);
155
 
                    ST(&(xo[0]), VADD(T1r, T1y), ovs, &(xo[0]));
156
 
                    T1z = VFNMS(LDK(KP250000000), T1y, T1r);
157
 
                    ST(&(xo[WS(os, 10)]), VADD(T13, T1i), ovs, &(xo[0]));
158
 
                    {
159
 
                         V T1p, T1l, T1o, T1q, T1F, T1B;
160
 
                         TU = VFNMS(LDK(KP618033988), TG, TH);
161
 
                         TI = VFMA(LDK(KP618033988), TH, TG);
162
 
                         TP = VFMA(LDK(KP618033988), TO, TN);
163
 
                         TX = VFNMS(LDK(KP618033988), TN, TO);
164
 
                         ST(&(xo[WS(os, 15)]), VFMAI(TC, Tj), ovs, &(xo[WS(os, 1)]));
165
 
                         ST(&(xo[WS(os, 5)]), VFNMSI(TC, Tj), ovs, &(xo[WS(os, 1)]));
166
 
                         T1p = VFMA(LDK(KP559016994), T1k, T1j);
167
 
                         T1l = VFNMS(LDK(KP559016994), T1k, T1j);
168
 
                         T1o = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T1n, T1m));
169
 
                         T1q = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T1m, T1n));
170
 
                         T1F = VFNMS(LDK(KP559016994), T1A, T1z);
171
 
                         T1B = VFMA(LDK(KP559016994), T1A, T1z);
172
 
                         ST(&(xo[WS(os, 14)]), VFMAI(T1q, T1p), ovs, &(xo[0]));
173
 
                         ST(&(xo[WS(os, 6)]), VFNMSI(T1q, T1p), ovs, &(xo[0]));
174
 
                         ST(&(xo[WS(os, 18)]), VFNMSI(T1o, T1l), ovs, &(xo[0]));
175
 
                         ST(&(xo[WS(os, 2)]), VFMAI(T1o, T1l), ovs, &(xo[0]));
176
 
                         ST(&(xo[WS(os, 16)]), VFNMSI(T1E, T1B), ovs, &(xo[0]));
177
 
                         ST(&(xo[WS(os, 4)]), VFMAI(T1E, T1B), ovs, &(xo[0]));
178
 
                         ST(&(xo[WS(os, 12)]), VFMAI(T1G, T1F), ovs, &(xo[0]));
179
 
                         ST(&(xo[WS(os, 8)]), VFNMSI(T1G, T1F), ovs, &(xo[0]));
180
 
                         TM = VFNMS(LDK(KP559016994), TL, TK);
181
 
                         TW = VFMA(LDK(KP559016994), TL, TK);
182
 
                         TT = VFNMS(LDK(KP559016994), TE, TD);
183
 
                         TF = VFMA(LDK(KP559016994), TE, TD);
184
 
                    }
185
 
               }
186
 
          }
187
 
          {
188
 
               V T10, TY, TQ, TS, TJ, TR, TZ, TV;
189
 
               T10 = VFMA(LDK(KP951056516), TX, TW);
190
 
               TY = VFNMS(LDK(KP951056516), TX, TW);
191
 
               TQ = VFMA(LDK(KP951056516), TP, TM);
192
 
               TS = VFNMS(LDK(KP951056516), TP, TM);
193
 
               TJ = VFMA(LDK(KP951056516), TI, TF);
194
 
               TR = VFNMS(LDK(KP951056516), TI, TF);
195
 
               TZ = VFMA(LDK(KP951056516), TU, TT);
196
 
               TV = VFNMS(LDK(KP951056516), TU, TT);
197
 
               ST(&(xo[WS(os, 11)]), VFMAI(TS, TR), ovs, &(xo[WS(os, 1)]));
198
 
               ST(&(xo[WS(os, 9)]), VFNMSI(TS, TR), ovs, &(xo[WS(os, 1)]));
199
 
               ST(&(xo[WS(os, 19)]), VFMAI(TQ, TJ), ovs, &(xo[WS(os, 1)]));
200
 
               ST(&(xo[WS(os, 1)]), VFNMSI(TQ, TJ), ovs, &(xo[WS(os, 1)]));
201
 
               ST(&(xo[WS(os, 3)]), VFMAI(TY, TV), ovs, &(xo[WS(os, 1)]));
202
 
               ST(&(xo[WS(os, 17)]), VFNMSI(TY, TV), ovs, &(xo[WS(os, 1)]));
203
 
               ST(&(xo[WS(os, 7)]), VFMAI(T10, TZ), ovs, &(xo[WS(os, 1)]));
204
 
               ST(&(xo[WS(os, 13)]), VFNMSI(T10, TZ), ovs, &(xo[WS(os, 1)]));
205
 
          }
206
 
     }
207
 
}
208
 
 
209
 
static const kdft_desc desc = { 20, "n1fv_20", {58, 4, 46, 0}, &GENUS, 0, 0, 0, 0 };
210
 
 
211
 
void X(codelet_n1fv_20) (planner *p) {
212
 
     X(kdft_register) (p, n1fv_20, &desc);
213
 
}
214
 
 
215
 
#else                           /* HAVE_FMA */
216
 
 
217
 
/* Generated by: ../../../genfft/gen_notw_c -simd -compact -variables 4 -pipeline-latency 8 -n 20 -name n1fv_20 -include n1f.h */
218
 
 
219
 
/*
220
 
 * This function contains 104 FP additions, 24 FP multiplications,
221
 
 * (or, 92 additions, 12 multiplications, 12 fused multiply/add),
222
 
 * 53 stack variables, 4 constants, and 40 memory accesses
223
 
 */
224
 
#include "n1f.h"
225
 
 
226
 
static void n1fv_20(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
227
 
{
228
 
     DVK(KP587785252, +0.587785252292473129168705954639072768597652438);
229
 
     DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
230
 
     DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
231
 
     DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
232
 
     INT i;
233
 
     const R *xi;
234
 
     R *xo;
235
 
     xi = ri;
236
 
     xo = ro;
237
 
     for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
238
 
          V T3, T1B, Tm, T1i, TG, TN, TO, TH, T13, T16, T1k, T1u, T1v, T1z, T1r;
239
 
          V T1s, T1y, T1a, T1d, T1j, Ti, TD, TB, TL, Tj, TC;
240
 
          {
241
 
               V T1, T2, T1g, Tk, Tl, T1h;
242
 
               T1 = LD(&(xi[0]), ivs, &(xi[0]));
243
 
               T2 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
244
 
               T1g = VADD(T1, T2);
245
 
               Tk = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
246
 
               Tl = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
247
 
               T1h = VADD(Tk, Tl);
248
 
               T3 = VSUB(T1, T2);
249
 
               T1B = VADD(T1g, T1h);
250
 
               Tm = VSUB(Tk, Tl);
251
 
               T1i = VSUB(T1g, T1h);
252
 
          }
253
 
          {
254
 
               V T6, T18, Tw, T12, Tz, T15, T9, T1b, Td, T11, Tp, T19, Ts, T1c, Tg;
255
 
               V T14;
256
 
               {
257
 
                    V T4, T5, Tu, Tv;
258
 
                    T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
259
 
                    T5 = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
260
 
                    T6 = VSUB(T4, T5);
261
 
                    T18 = VADD(T4, T5);
262
 
                    Tu = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
263
 
                    Tv = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
264
 
                    Tw = VSUB(Tu, Tv);
265
 
                    T12 = VADD(Tu, Tv);
266
 
               }
267
 
               {
268
 
                    V Tx, Ty, T7, T8;
269
 
                    Tx = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
270
 
                    Ty = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
271
 
                    Tz = VSUB(Tx, Ty);
272
 
                    T15 = VADD(Tx, Ty);
273
 
                    T7 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
274
 
                    T8 = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
275
 
                    T9 = VSUB(T7, T8);
276
 
                    T1b = VADD(T7, T8);
277
 
               }
278
 
               {
279
 
                    V Tb, Tc, Tn, To;
280
 
                    Tb = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
281
 
                    Tc = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
282
 
                    Td = VSUB(Tb, Tc);
283
 
                    T11 = VADD(Tb, Tc);
284
 
                    Tn = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
285
 
                    To = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
286
 
                    Tp = VSUB(Tn, To);
287
 
                    T19 = VADD(Tn, To);
288
 
               }
289
 
               {
290
 
                    V Tq, Tr, Te, Tf;
291
 
                    Tq = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
292
 
                    Tr = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
293
 
                    Ts = VSUB(Tq, Tr);
294
 
                    T1c = VADD(Tq, Tr);
295
 
                    Te = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
296
 
                    Tf = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
297
 
                    Tg = VSUB(Te, Tf);
298
 
                    T14 = VADD(Te, Tf);
299
 
               }
300
 
               TG = VSUB(Ts, Tp);
301
 
               TN = VSUB(T6, T9);
302
 
               TO = VSUB(Td, Tg);
303
 
               TH = VSUB(Tz, Tw);
304
 
               T13 = VSUB(T11, T12);
305
 
               T16 = VSUB(T14, T15);
306
 
               T1k = VADD(T13, T16);
307
 
               T1u = VADD(T11, T12);
308
 
               T1v = VADD(T14, T15);
309
 
               T1z = VADD(T1u, T1v);
310
 
               T1r = VADD(T18, T19);
311
 
               T1s = VADD(T1b, T1c);
312
 
               T1y = VADD(T1r, T1s);
313
 
               T1a = VSUB(T18, T19);
314
 
               T1d = VSUB(T1b, T1c);
315
 
               T1j = VADD(T1a, T1d);
316
 
               {
317
 
                    V Ta, Th, Tt, TA;
318
 
                    Ta = VADD(T6, T9);
319
 
                    Th = VADD(Td, Tg);
320
 
                    Ti = VADD(Ta, Th);
321
 
                    TD = VMUL(LDK(KP559016994), VSUB(Ta, Th));
322
 
                    Tt = VADD(Tp, Ts);
323
 
                    TA = VADD(Tw, Tz);
324
 
                    TB = VADD(Tt, TA);
325
 
                    TL = VMUL(LDK(KP559016994), VSUB(TA, Tt));
326
 
               }
327
 
          }
328
 
          Tj = VADD(T3, Ti);
329
 
          TC = VBYI(VADD(Tm, TB));
330
 
          ST(&(xo[WS(os, 5)]), VSUB(Tj, TC), ovs, &(xo[WS(os, 1)]));
331
 
          ST(&(xo[WS(os, 15)]), VADD(Tj, TC), ovs, &(xo[WS(os, 1)]));
332
 
          {
333
 
               V T1A, T1C, T1D, T1x, T1G, T1t, T1w, T1F, T1E;
334
 
               T1A = VMUL(LDK(KP559016994), VSUB(T1y, T1z));
335
 
               T1C = VADD(T1y, T1z);
336
 
               T1D = VFNMS(LDK(KP250000000), T1C, T1B);
337
 
               T1t = VSUB(T1r, T1s);
338
 
               T1w = VSUB(T1u, T1v);
339
 
               T1x = VBYI(VFMA(LDK(KP951056516), T1t, VMUL(LDK(KP587785252), T1w)));
340
 
               T1G = VBYI(VFNMS(LDK(KP587785252), T1t, VMUL(LDK(KP951056516), T1w)));
341
 
               ST(&(xo[0]), VADD(T1B, T1C), ovs, &(xo[0]));
342
 
               T1F = VSUB(T1D, T1A);
343
 
               ST(&(xo[WS(os, 8)]), VSUB(T1F, T1G), ovs, &(xo[0]));
344
 
               ST(&(xo[WS(os, 12)]), VADD(T1G, T1F), ovs, &(xo[0]));
345
 
               T1E = VADD(T1A, T1D);
346
 
               ST(&(xo[WS(os, 4)]), VADD(T1x, T1E), ovs, &(xo[0]));
347
 
               ST(&(xo[WS(os, 16)]), VSUB(T1E, T1x), ovs, &(xo[0]));
348
 
          }
349
 
          {
350
 
               V T1n, T1l, T1m, T1f, T1q, T17, T1e, T1p, T1o;
351
 
               T1n = VMUL(LDK(KP559016994), VSUB(T1j, T1k));
352
 
               T1l = VADD(T1j, T1k);
353
 
               T1m = VFNMS(LDK(KP250000000), T1l, T1i);
354
 
               T17 = VSUB(T13, T16);
355
 
               T1e = VSUB(T1a, T1d);
356
 
               T1f = VBYI(VFNMS(LDK(KP587785252), T1e, VMUL(LDK(KP951056516), T17)));
357
 
               T1q = VBYI(VFMA(LDK(KP951056516), T1e, VMUL(LDK(KP587785252), T17)));
358
 
               ST(&(xo[WS(os, 10)]), VADD(T1i, T1l), ovs, &(xo[0]));
359
 
               T1p = VADD(T1n, T1m);
360
 
               ST(&(xo[WS(os, 6)]), VSUB(T1p, T1q), ovs, &(xo[0]));
361
 
               ST(&(xo[WS(os, 14)]), VADD(T1q, T1p), ovs, &(xo[0]));
362
 
               T1o = VSUB(T1m, T1n);
363
 
               ST(&(xo[WS(os, 2)]), VADD(T1f, T1o), ovs, &(xo[0]));
364
 
               ST(&(xo[WS(os, 18)]), VSUB(T1o, T1f), ovs, &(xo[0]));
365
 
          }
366
 
          {
367
 
               V TI, TP, TX, TU, TM, TW, TF, TT, TK, TE;
368
 
               TI = VFMA(LDK(KP951056516), TG, VMUL(LDK(KP587785252), TH));
369
 
               TP = VFMA(LDK(KP951056516), TN, VMUL(LDK(KP587785252), TO));
370
 
               TX = VFNMS(LDK(KP587785252), TN, VMUL(LDK(KP951056516), TO));
371
 
               TU = VFNMS(LDK(KP587785252), TG, VMUL(LDK(KP951056516), TH));
372
 
               TK = VFMS(LDK(KP250000000), TB, Tm);
373
 
               TM = VADD(TK, TL);
374
 
               TW = VSUB(TL, TK);
375
 
               TE = VFNMS(LDK(KP250000000), Ti, T3);
376
 
               TF = VADD(TD, TE);
377
 
               TT = VSUB(TE, TD);
378
 
               {
379
 
                    V TJ, TQ, TZ, T10;
380
 
                    TJ = VADD(TF, TI);
381
 
                    TQ = VBYI(VSUB(TM, TP));
382
 
                    ST(&(xo[WS(os, 19)]), VSUB(TJ, TQ), ovs, &(xo[WS(os, 1)]));
383
 
                    ST(&(xo[WS(os, 1)]), VADD(TJ, TQ), ovs, &(xo[WS(os, 1)]));
384
 
                    TZ = VADD(TT, TU);
385
 
                    T10 = VBYI(VADD(TX, TW));
386
 
                    ST(&(xo[WS(os, 13)]), VSUB(TZ, T10), ovs, &(xo[WS(os, 1)]));
387
 
                    ST(&(xo[WS(os, 7)]), VADD(TZ, T10), ovs, &(xo[WS(os, 1)]));
388
 
               }
389
 
               {
390
 
                    V TR, TS, TV, TY;
391
 
                    TR = VSUB(TF, TI);
392
 
                    TS = VBYI(VADD(TP, TM));
393
 
                    ST(&(xo[WS(os, 11)]), VSUB(TR, TS), ovs, &(xo[WS(os, 1)]));
394
 
                    ST(&(xo[WS(os, 9)]), VADD(TR, TS), ovs, &(xo[WS(os, 1)]));
395
 
                    TV = VSUB(TT, TU);
396
 
                    TY = VBYI(VSUB(TW, TX));
397
 
                    ST(&(xo[WS(os, 17)]), VSUB(TV, TY), ovs, &(xo[WS(os, 1)]));
398
 
                    ST(&(xo[WS(os, 3)]), VADD(TV, TY), ovs, &(xo[WS(os, 1)]));
399
 
               }
400
 
          }
401
 
     }
402
 
}
403
 
 
404
 
static const kdft_desc desc = { 20, "n1fv_20", {92, 12, 12, 0}, &GENUS, 0, 0, 0, 0 };
405
 
 
406
 
void X(codelet_n1fv_20) (planner *p) {
407
 
     X(kdft_register) (p, n1fv_20, &desc);
408
 
}
409
 
 
410
 
#endif                          /* HAVE_FMA */