← Back to branch summary

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

« back to all changes in this revision

Viewing changes to dft/simd/codelets/n1bv_13.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/n1bv_13.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 06:39:23 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 -sign 1 -n 13 -name n1bv_13 -include n1b.h */
29
 
 
30
 
/*
31
 
 * This function contains 88 FP additions, 63 FP multiplications,
32
 
 * (or, 31 additions, 6 multiplications, 57 fused multiply/add),
33
 
 * 96 stack variables, 23 constants, and 26 memory accesses
34
 
 */
35
 
#include "n1b.h"
36
 
 
37
 
static void n1bv_13(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
38
 
{
39
 
     DVK(KP904176221, +0.904176221990848204433795481776887926501523162);
40
 
     DVK(KP575140729, +0.575140729474003121368385547455453388461001608);
41
 
     DVK(KP300462606, +0.300462606288665774426601772289207995520941381);
42
 
     DVK(KP516520780, +0.516520780623489722840901288569017135705033622);
43
 
     DVK(KP522026385, +0.522026385161275033714027226654165028300441940);
44
 
     DVK(KP957805992, +0.957805992594665126462521754605754580515587217);
45
 
     DVK(KP600477271, +0.600477271932665282925769253334763009352012849);
46
 
     DVK(KP251768516, +0.251768516431883313623436926934233488546674281);
47
 
     DVK(KP503537032, +0.503537032863766627246873853868466977093348562);
48
 
     DVK(KP769338817, +0.769338817572980603471413688209101117038278899);
49
 
     DVK(KP859542535, +0.859542535098774820163672132761689612766401925);
50
 
     DVK(KP581704778, +0.581704778510515730456870384989698884939833902);
51
 
     DVK(KP853480001, +0.853480001859823990758994934970528322872359049);
52
 
     DVK(KP083333333, +0.083333333333333333333333333333333333333333333);
53
 
     DVK(KP226109445, +0.226109445035782405468510155372505010481906348);
54
 
     DVK(KP301479260, +0.301479260047709873958013540496673347309208464);
55
 
     DVK(KP686558370, +0.686558370781754340655719594850823015421401653);
56
 
     DVK(KP514918778, +0.514918778086315755491789696138117261566051239);
57
 
     DVK(KP038632954, +0.038632954644348171955506895830342264440241080);
58
 
     DVK(KP612264650, +0.612264650376756543746494474777125408779395514);
59
 
     DVK(KP302775637, +0.302775637731994646559610633735247973125648287);
60
 
     DVK(KP866025403, +0.866025403784438646763723170752936183471402627);
61
 
     DVK(KP500000000, +0.500000000000000000000000000000000000000000000);
62
 
     INT i;
63
 
     const R *xi;
64
 
     R *xo;
65
 
     xi = ii;
66
 
     xo = io;
67
 
     for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
68
 
          V T1, T7, T2, Tg, Tf, TN, Th, Tq, Ta, Tj, T5, Tr, Tk;
69
 
          T1 = LD(&(xi[0]), ivs, &(xi[0]));
70
 
          {
71
 
               V Td, Te, T8, T9, T3, T4;
72
 
               Td = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
73
 
               Te = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
74
 
               T7 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
75
 
               T8 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
76
 
               T9 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
77
 
               T2 = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
78
 
               T3 = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
79
 
               T4 = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
80
 
               Tg = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
81
 
               Tf = VADD(Td, Te);
82
 
               TN = VSUB(Td, Te);
83
 
               Th = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
84
 
               Tq = VSUB(T8, T9);
85
 
               Ta = VADD(T8, T9);
86
 
               Tj = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
87
 
               T5 = VADD(T3, T4);
88
 
               Tr = VSUB(T4, T3);
89
 
               Tk = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
90
 
          }
91
 
          {
92
 
               V Tt, Ti, Ty, Tb, Ts, TQ, Tx, T6, Tu, Tl;
93
 
               Tt = VSUB(Tg, Th);
94
 
               Ti = VADD(Tg, Th);
95
 
               Ty = VFMS(LDK(KP500000000), Ta, T7);
96
 
               Tb = VADD(T7, Ta);
97
 
               Ts = VSUB(Tq, Tr);
98
 
               TQ = VADD(Tr, Tq);
99
 
               Tx = VFNMS(LDK(KP500000000), T5, T2);
100
 
               T6 = VADD(T2, T5);
101
 
               Tu = VSUB(Tj, Tk);
102
 
               Tl = VADD(Tj, Tk);
103
 
               {
104
 
                    V TK, Tz, Tc, TX, Tv, TO, TL, Tm;
105
 
                    TK = VADD(Tx, Ty);
106
 
                    Tz = VSUB(Tx, Ty);
107
 
                    Tc = VADD(T6, Tb);
108
 
                    TX = VSUB(T6, Tb);
109
 
                    Tv = VSUB(Tt, Tu);
110
 
                    TO = VADD(Tt, Tu);
111
 
                    TL = VSUB(Ti, Tl);
112
 
                    Tm = VADD(Ti, Tl);
113
 
                    {
114
 
                         V TF, Tw, TP, TY, TT, TM, TA, Tn;
115
 
                         TF = VSUB(Ts, Tv);
116
 
                         Tw = VADD(Ts, Tv);
117
 
                         TP = VFNMS(LDK(KP500000000), TO, TN);
118
 
                         TY = VADD(TN, TO);
119
 
                         TT = VFNMS(LDK(KP866025403), TL, TK);
120
 
                         TM = VFMA(LDK(KP866025403), TL, TK);
121
 
                         TA = VFNMS(LDK(KP500000000), Tm, Tf);
122
 
                         Tn = VADD(Tf, Tm);
123
 
                         {
124
 
                              V T1f, T1n, TI, T18, T1k, T1c, TD, T17, T10, T1m, T16, T1e, TU, TR;
125
 
                              TU = VFNMS(LDK(KP866025403), TQ, TP);
126
 
                              TR = VFMA(LDK(KP866025403), TQ, TP);
127
 
                              {
128
 
                                   V TZ, T15, TE, TB;
129
 
                                   TZ = VFMA(LDK(KP302775637), TY, TX);
130
 
                                   T15 = VFNMS(LDK(KP302775637), TX, TY);
131
 
                                   TE = VSUB(Tz, TA);
132
 
                                   TB = VADD(Tz, TA);
133
 
                                   {
134
 
                                        V TH, To, TV, T13;
135
 
                                        TH = VSUB(Tc, Tn);
136
 
                                        To = VADD(Tc, Tn);
137
 
                                        TV = VFNMS(LDK(KP612264650), TU, TT);
138
 
                                        T13 = VFMA(LDK(KP612264650), TT, TU);
139
 
                                        {
140
 
                                             V TS, T12, TG, T1b;
141
 
                                             TS = VFNMS(LDK(KP038632954), TR, TM);
142
 
                                             T12 = VFMA(LDK(KP038632954), TM, TR);
143
 
                                             TG = VFNMS(LDK(KP514918778), TF, TE);
144
 
                                             T1b = VFMA(LDK(KP686558370), TE, TF);
145
 
                                             {
146
 
                                                  V TC, T1a, Tp, TW, T14;
147
 
                                                  TC = VFMA(LDK(KP301479260), TB, Tw);
148
 
                                                  T1a = VFNMS(LDK(KP226109445), Tw, TB);
149
 
                                                  Tp = VFNMS(LDK(KP083333333), To, T1);
150
 
                                                  ST(&(xo[0]), VADD(T1, To), ovs, &(xo[0]));
151
 
                                                  T1f = VFMA(LDK(KP853480001), TV, TS);
152
 
                                                  TW = VFNMS(LDK(KP853480001), TV, TS);
153
 
                                                  T1n = VFMA(LDK(KP853480001), T13, T12);
154
 
                                                  T14 = VFNMS(LDK(KP853480001), T13, T12);
155
 
                                                  TI = VFMA(LDK(KP581704778), TH, TG);
156
 
                                                  T18 = VFNMS(LDK(KP859542535), TG, TH);
157
 
                                                  T1k = VFMA(LDK(KP769338817), T1b, T1a);
158
 
                                                  T1c = VFNMS(LDK(KP769338817), T1b, T1a);
159
 
                                                  TD = VFMA(LDK(KP503537032), TC, Tp);
160
 
                                                  T17 = VFNMS(LDK(KP251768516), TC, Tp);
161
 
                                                  T10 = VMUL(LDK(KP600477271), VFMA(LDK(KP957805992), TZ, TW));
162
 
                                                  T1m = VFNMS(LDK(KP522026385), TW, TZ);
163
 
                                                  T16 = VMUL(LDK(KP600477271), VFMA(LDK(KP957805992), T15, T14));
164
 
                                                  T1e = VFNMS(LDK(KP522026385), T14, T15);
165
 
                                             }
166
 
                                        }
167
 
                                   }
168
 
                              }
169
 
                              {
170
 
                                   V T1o, T1q, T1g, T1i, T1d, T1h, T1l, T1p;
171
 
                                   {
172
 
                                        V T11, TJ, T19, T1j;
173
 
                                        T11 = VFMA(LDK(KP516520780), TI, TD);
174
 
                                        TJ = VFNMS(LDK(KP516520780), TI, TD);
175
 
                                        T19 = VFMA(LDK(KP300462606), T18, T17);
176
 
                                        T1j = VFNMS(LDK(KP300462606), T18, T17);
177
 
                                        T1o = VMUL(LDK(KP575140729), VFNMS(LDK(KP904176221), T1n, T1m));
178
 
                                        T1q = VMUL(LDK(KP575140729), VFMA(LDK(KP904176221), T1n, T1m));
179
 
                                        T1g = VMUL(LDK(KP575140729), VFMA(LDK(KP904176221), T1f, T1e));
180
 
                                        T1i = VMUL(LDK(KP575140729), VFNMS(LDK(KP904176221), T1f, T1e));
181
 
                                        ST(&(xo[WS(os, 12)]), VFMAI(T16, T11), ovs, &(xo[0]));
182
 
                                        ST(&(xo[WS(os, 1)]), VFNMSI(T16, T11), ovs, &(xo[WS(os, 1)]));
183
 
                                        ST(&(xo[WS(os, 8)]), VFNMSI(T10, TJ), ovs, &(xo[0]));
184
 
                                        ST(&(xo[WS(os, 5)]), VFMAI(T10, TJ), ovs, &(xo[WS(os, 1)]));
185
 
                                        T1d = VFNMS(LDK(KP503537032), T1c, T19);
186
 
                                        T1h = VFMA(LDK(KP503537032), T1c, T19);
187
 
                                        T1l = VFNMS(LDK(KP503537032), T1k, T1j);
188
 
                                        T1p = VFMA(LDK(KP503537032), T1k, T1j);
189
 
                                   }
190
 
                                   ST(&(xo[WS(os, 9)]), VFNMSI(T1g, T1d), ovs, &(xo[WS(os, 1)]));
191
 
                                   ST(&(xo[WS(os, 4)]), VFMAI(T1g, T1d), ovs, &(xo[0]));
192
 
                                   ST(&(xo[WS(os, 10)]), VFMAI(T1i, T1h), ovs, &(xo[0]));
193
 
                                   ST(&(xo[WS(os, 3)]), VFNMSI(T1i, T1h), ovs, &(xo[WS(os, 1)]));
194
 
                                   ST(&(xo[WS(os, 7)]), VFNMSI(T1o, T1l), ovs, &(xo[WS(os, 1)]));
195
 
                                   ST(&(xo[WS(os, 6)]), VFMAI(T1o, T1l), ovs, &(xo[0]));
196
 
                                   ST(&(xo[WS(os, 11)]), VFNMSI(T1q, T1p), ovs, &(xo[WS(os, 1)]));
197
 
                                   ST(&(xo[WS(os, 2)]), VFMAI(T1q, T1p), ovs, &(xo[0]));
198
 
                              }
199
 
                         }
200
 
                    }
201
 
               }
202
 
          }
203
 
     }
204
 
}
205
 
 
206
 
static const kdft_desc desc = { 13, "n1bv_13", {31, 6, 57, 0}, &GENUS, 0, 0, 0, 0 };
207
 
 
208
 
void X(codelet_n1bv_13) (planner *p) {
209
 
     X(kdft_register) (p, n1bv_13, &desc);
210
 
}
211
 
 
212
 
#else                           /* HAVE_FMA */
213
 
 
214
 
/* Generated by: ../../../genfft/gen_notw_c -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 13 -name n1bv_13 -include n1b.h */
215
 
 
216
 
/*
217
 
 * This function contains 88 FP additions, 34 FP multiplications,
218
 
 * (or, 69 additions, 15 multiplications, 19 fused multiply/add),
219
 
 * 60 stack variables, 20 constants, and 26 memory accesses
220
 
 */
221
 
#include "n1b.h"
222
 
 
223
 
static void n1bv_13(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
224
 
{
225
 
     DVK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
226
 
     DVK(KP083333333, +0.083333333333333333333333333333333333333333333);
227
 
     DVK(KP075902986, +0.075902986037193865983102897245103540356428373);
228
 
     DVK(KP251768516, +0.251768516431883313623436926934233488546674281);
229
 
     DVK(KP132983124, +0.132983124607418643793760531921092974399165133);
230
 
     DVK(KP258260390, +0.258260390311744861420450644284508567852516811);
231
 
     DVK(KP1_732050807, +1.732050807568877293527446341505872366942805254);
232
 
     DVK(KP300238635, +0.300238635966332641462884626667381504676006424);
233
 
     DVK(KP011599105, +0.011599105605768290721655456654083252189827041);
234
 
     DVK(KP256247671, +0.256247671582936600958684654061725059144125175);
235
 
     DVK(KP156891391, +0.156891391051584611046832726756003269660212636);
236
 
     DVK(KP174138601, +0.174138601152135905005660794929264742616964676);
237
 
     DVK(KP575140729, +0.575140729474003121368385547455453388461001608);
238
 
     DVK(KP503537032, +0.503537032863766627246873853868466977093348562);
239
 
     DVK(KP113854479, +0.113854479055790798974654345867655310534642560);
240
 
     DVK(KP265966249, +0.265966249214837287587521063842185948798330267);
241
 
     DVK(KP387390585, +0.387390585467617292130675966426762851778775217);
242
 
     DVK(KP300462606, +0.300462606288665774426601772289207995520941381);
243
 
     DVK(KP866025403, +0.866025403784438646763723170752936183471402627);
244
 
     DVK(KP500000000, +0.500000000000000000000000000000000000000000000);
245
 
     INT i;
246
 
     const R *xi;
247
 
     R *xo;
248
 
     xi = ii;
249
 
     xo = io;
250
 
     for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
251
 
          V TW, Tb, Tm, Ts, TB, TR, TX, TK, TU, Tz, TC, TN, TT;
252
 
          TW = LD(&(xi[0]), ivs, &(xi[0]));
253
 
          {
254
 
               V Te, TH, Ta, Tu, Tp, T5, Tt, To, Th, Tw, Tk, Tx, Tl, TI, Tc;
255
 
               V Td, Tq, Tr;
256
 
               Tc = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
257
 
               Td = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
258
 
               Te = VSUB(Tc, Td);
259
 
               TH = VADD(Tc, Td);
260
 
               {
261
 
                    V T6, T7, T8, T9;
262
 
                    T6 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
263
 
                    T7 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
264
 
                    T8 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
265
 
                    T9 = VADD(T7, T8);
266
 
                    Ta = VADD(T6, T9);
267
 
                    Tu = VFNMS(LDK(KP500000000), T9, T6);
268
 
                    Tp = VSUB(T7, T8);
269
 
               }
270
 
               {
271
 
                    V T1, T2, T3, T4;
272
 
                    T1 = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
273
 
                    T2 = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
274
 
                    T3 = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
275
 
                    T4 = VADD(T2, T3);
276
 
                    T5 = VADD(T1, T4);
277
 
                    Tt = VFNMS(LDK(KP500000000), T4, T1);
278
 
                    To = VSUB(T2, T3);
279
 
               }
280
 
               {
281
 
                    V Tf, Tg, Ti, Tj;
282
 
                    Tf = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
283
 
                    Tg = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
284
 
                    Th = VSUB(Tf, Tg);
285
 
                    Tw = VADD(Tf, Tg);
286
 
                    Ti = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
287
 
                    Tj = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
288
 
                    Tk = VSUB(Ti, Tj);
289
 
                    Tx = VADD(Ti, Tj);
290
 
               }
291
 
               Tl = VADD(Th, Tk);
292
 
               TI = VADD(Tw, Tx);
293
 
               Tb = VSUB(T5, Ta);
294
 
               Tm = VADD(Te, Tl);
295
 
               Tq = VMUL(LDK(KP866025403), VSUB(To, Tp));
296
 
               Tr = VFNMS(LDK(KP500000000), Tl, Te);
297
 
               Ts = VADD(Tq, Tr);
298
 
               TB = VSUB(Tq, Tr);
299
 
               {
300
 
                    V TP, TQ, TG, TJ;
301
 
                    TP = VADD(T5, Ta);
302
 
                    TQ = VADD(TH, TI);
303
 
                    TR = VMUL(LDK(KP300462606), VSUB(TP, TQ));
304
 
                    TX = VADD(TP, TQ);
305
 
                    TG = VADD(Tt, Tu);
306
 
                    TJ = VFNMS(LDK(KP500000000), TI, TH);
307
 
                    TK = VSUB(TG, TJ);
308
 
                    TU = VADD(TG, TJ);
309
 
               }
310
 
               {
311
 
                    V Tv, Ty, TL, TM;
312
 
                    Tv = VSUB(Tt, Tu);
313
 
                    Ty = VMUL(LDK(KP866025403), VSUB(Tw, Tx));
314
 
                    Tz = VSUB(Tv, Ty);
315
 
                    TC = VADD(Tv, Ty);
316
 
                    TL = VADD(To, Tp);
317
 
                    TM = VSUB(Th, Tk);
318
 
                    TN = VSUB(TL, TM);
319
 
                    TT = VADD(TL, TM);
320
 
               }
321
 
          }
322
 
          ST(&(xo[0]), VADD(TW, TX), ovs, &(xo[0]));
323
 
          {
324
 
               V T1c, T1n, T11, T14, T17, T1k, Tn, TE, T18, T1j, TS, T1m, TZ, T1f, TA;
325
 
               V TD;
326
 
               {
327
 
                    V T1a, T1b, T12, T13;
328
 
                    T1a = VFMA(LDK(KP387390585), TN, VMUL(LDK(KP265966249), TK));
329
 
                    T1b = VFNMS(LDK(KP503537032), TU, VMUL(LDK(KP113854479), TT));
330
 
                    T1c = VSUB(T1a, T1b);
331
 
                    T1n = VADD(T1a, T1b);
332
 
                    T11 = VFMA(LDK(KP575140729), Tb, VMUL(LDK(KP174138601), Tm));
333
 
                    T12 = VFNMS(LDK(KP256247671), Tz, VMUL(LDK(KP156891391), Ts));
334
 
                    T13 = VFMA(LDK(KP011599105), TB, VMUL(LDK(KP300238635), TC));
335
 
                    T14 = VADD(T12, T13);
336
 
                    T17 = VSUB(T11, T14);
337
 
                    T1k = VMUL(LDK(KP1_732050807), VSUB(T12, T13));
338
 
               }
339
 
               Tn = VFNMS(LDK(KP575140729), Tm, VMUL(LDK(KP174138601), Tb));
340
 
               TA = VFMA(LDK(KP256247671), Ts, VMUL(LDK(KP156891391), Tz));
341
 
               TD = VFNMS(LDK(KP011599105), TC, VMUL(LDK(KP300238635), TB));
342
 
               TE = VADD(TA, TD);
343
 
               T18 = VMUL(LDK(KP1_732050807), VSUB(TD, TA));
344
 
               T1j = VSUB(Tn, TE);
345
 
               {
346
 
                    V TO, T1e, TV, TY, T1d;
347
 
                    TO = VFNMS(LDK(KP132983124), TN, VMUL(LDK(KP258260390), TK));
348
 
                    T1e = VSUB(TR, TO);
349
 
                    TV = VFMA(LDK(KP251768516), TT, VMUL(LDK(KP075902986), TU));
350
 
                    TY = VFNMS(LDK(KP083333333), TX, TW);
351
 
                    T1d = VSUB(TY, TV);
352
 
                    TS = VFMA(LDK(KP2_000000000), TO, TR);
353
 
                    T1m = VADD(T1e, T1d);
354
 
                    TZ = VFMA(LDK(KP2_000000000), TV, TY);
355
 
                    T1f = VSUB(T1d, T1e);
356
 
               }
357
 
               {
358
 
                    V TF, T10, T1l, T1o;
359
 
                    TF = VBYI(VFMA(LDK(KP2_000000000), TE, Tn));
360
 
                    T10 = VADD(TS, TZ);
361
 
                    ST(&(xo[WS(os, 1)]), VADD(TF, T10), ovs, &(xo[WS(os, 1)]));
362
 
                    ST(&(xo[WS(os, 12)]), VSUB(T10, TF), ovs, &(xo[0]));
363
 
                    {
364
 
                         V T15, T16, T1p, T1q;
365
 
                         T15 = VBYI(VFMA(LDK(KP2_000000000), T14, T11));
366
 
                         T16 = VSUB(TZ, TS);
367
 
                         ST(&(xo[WS(os, 5)]), VADD(T15, T16), ovs, &(xo[WS(os, 1)]));
368
 
                         ST(&(xo[WS(os, 8)]), VSUB(T16, T15), ovs, &(xo[0]));
369
 
                         T1p = VADD(T1n, T1m);
370
 
                         T1q = VBYI(VADD(T1j, T1k));
371
 
                         ST(&(xo[WS(os, 4)]), VSUB(T1p, T1q), ovs, &(xo[0]));
372
 
                         ST(&(xo[WS(os, 9)]), VADD(T1q, T1p), ovs, &(xo[WS(os, 1)]));
373
 
                    }
374
 
                    T1l = VBYI(VSUB(T1j, T1k));
375
 
                    T1o = VSUB(T1m, T1n);
376
 
                    ST(&(xo[WS(os, 3)]), VADD(T1l, T1o), ovs, &(xo[WS(os, 1)]));
377
 
                    ST(&(xo[WS(os, 10)]), VSUB(T1o, T1l), ovs, &(xo[0]));
378
 
                    {
379
 
                         V T1h, T1i, T19, T1g;
380
 
                         T1h = VBYI(VADD(T18, T17));
381
 
                         T1i = VSUB(T1f, T1c);
382
 
                         ST(&(xo[WS(os, 6)]), VADD(T1h, T1i), ovs, &(xo[0]));
383
 
                         ST(&(xo[WS(os, 7)]), VSUB(T1i, T1h), ovs, &(xo[WS(os, 1)]));
384
 
                         T19 = VBYI(VSUB(T17, T18));
385
 
                         T1g = VADD(T1c, T1f);
386
 
                         ST(&(xo[WS(os, 2)]), VADD(T19, T1g), ovs, &(xo[0]));
387
 
                         ST(&(xo[WS(os, 11)]), VSUB(T1g, T19), ovs, &(xo[WS(os, 1)]));
388
 
                    }
389
 
               }
390
 
          }
391
 
     }
392
 
}
393
 
 
394
 
static const kdft_desc desc = { 13, "n1bv_13", {69, 15, 19, 0}, &GENUS, 0, 0, 0, 0 };
395
 
 
396
 
void X(codelet_n1bv_13) (planner *p) {
397
 
     X(kdft_register) (p, n1bv_13, &desc);
398
 
}
399
 
 
400
 
#endif                          /* HAVE_FMA */