~diresu/blender/blender-command-port

« back to all changes in this revision

Viewing changes to extern/fftw/dft/codelets/standard/t1_16.c

  • Committer: theeth
  • Date: 2008-10-14 16:52:04 UTC
  • Revision ID: vcs-imports@canonical.com-20081014165204-r32w2gm6s0osvdhn
copy back trunk

Show diffs side-by-side

added added

removed removed

Lines of Context:
 
1
/*
 
2
 * Copyright (c) 2003, 2006 Matteo Frigo
 
3
 * Copyright (c) 2003, 2006 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 Sat Jul  1 13:59:13 EDT 2006 */
 
23
 
 
24
#include "codelet-dft.h"
 
25
 
 
26
#ifdef HAVE_FMA
 
27
 
 
28
/* Generated by: ../../../genfft/gen_twiddle -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 16 -name t1_16 -include t.h */
 
29
 
 
30
/*
 
31
 * This function contains 174 FP additions, 100 FP multiplications,
 
32
 * (or, 104 additions, 30 multiplications, 70 fused multiply/add),
 
33
 * 97 stack variables, and 64 memory accesses
 
34
 */
 
35
/*
 
36
 * Generator Id's : 
 
37
 * $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
 
38
 * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
 
39
 * $Id: gen_twiddle.ml,v 1.24 2006-02-12 23:34:12 athena Exp $
 
40
 */
 
41
 
 
42
#include "t.h"
 
43
 
 
44
static const R *t1_16(R *ri, R *ii, const R *W, stride ios, INT m, INT dist)
 
45
{
 
46
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
47
     DK(KP414213562, +0.414213562373095048801688724209698078569671875);
 
48
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
49
     INT i;
 
50
     for (i = m; i > 0; i = i - 1, ri = ri + dist, ii = ii + dist, W = W + 30, MAKE_VOLATILE_STRIDE(ios)) {
 
51
          E T3G, T3F;
 
52
          {
 
53
               E T3z, T3o, T8, T1I, T2o, T35, T2r, T1s, T2w, T36, T2p, T1F, T3k, T1N, T3A;
 
54
               E Tl, T1T, T2V, T1U, Tz, T29, T30, T2c, T11, TB, TE, T2h, T31, T2a, T1e;
 
55
               E TC, T1X, TH, TK, TG, TD, TJ;
 
56
               {
 
57
                    E Ta, Td, Tb, T1J, Tg, Tj, Tf, Tc, Ti;
 
58
                    {
 
59
                         E T1h, T1k, T1n, T2k, T1i, T1q, T1m, T1j, T1p;
 
60
                         {
 
61
                              E T1, T3n, T3, T6, T2, T5;
 
62
                              T1 = ri[0];
 
63
                              T3n = ii[0];
 
64
                              T3 = ri[WS(ios, 8)];
 
65
                              T6 = ii[WS(ios, 8)];
 
66
                              T2 = W[14];
 
67
                              T5 = W[15];
 
68
                              {
 
69
                                   E T3l, T4, T1g, T3m, T7;
 
70
                                   T1h = ri[WS(ios, 15)];
 
71
                                   T1k = ii[WS(ios, 15)];
 
72
                                   T3l = T2 * T6;
 
73
                                   T4 = T2 * T3;
 
74
                                   T1g = W[28];
 
75
                                   T1n = ri[WS(ios, 7)];
 
76
                                   T3m = FNMS(T5, T3, T3l);
 
77
                                   T7 = FMA(T5, T6, T4);
 
78
                                   T2k = T1g * T1k;
 
79
                                   T1i = T1g * T1h;
 
80
                                   T3z = T3n - T3m;
 
81
                                   T3o = T3m + T3n;
 
82
                                   T8 = T1 + T7;
 
83
                                   T1I = T1 - T7;
 
84
                                   T1q = ii[WS(ios, 7)];
 
85
                                   T1m = W[12];
 
86
                              }
 
87
                              T1j = W[29];
 
88
                              T1p = W[13];
 
89
                         }
 
90
                         {
 
91
                              E T1u, T1x, T1v, T2s, T1A, T1D, T1z, T1w, T1C;
 
92
                              {
 
93
                                   E T2l, T1l, T2n, T1r, T2m, T1o, T1t;
 
94
                                   T1u = ri[WS(ios, 3)];
 
95
                                   T2m = T1m * T1q;
 
96
                                   T1o = T1m * T1n;
 
97
                                   T2l = FNMS(T1j, T1h, T2k);
 
98
                                   T1l = FMA(T1j, T1k, T1i);
 
99
                                   T2n = FNMS(T1p, T1n, T2m);
 
100
                                   T1r = FMA(T1p, T1q, T1o);
 
101
                                   T1x = ii[WS(ios, 3)];
 
102
                                   T1t = W[4];
 
103
                                   T2o = T2l - T2n;
 
104
                                   T35 = T2l + T2n;
 
105
                                   T2r = T1l - T1r;
 
106
                                   T1s = T1l + T1r;
 
107
                                   T1v = T1t * T1u;
 
108
                                   T2s = T1t * T1x;
 
109
                              }
 
110
                              T1A = ri[WS(ios, 11)];
 
111
                              T1D = ii[WS(ios, 11)];
 
112
                              T1z = W[20];
 
113
                              T1w = W[5];
 
114
                              T1C = W[21];
 
115
                              {
 
116
                                   E T2t, T1y, T2v, T1E, T2u, T1B, T9;
 
117
                                   Ta = ri[WS(ios, 4)];
 
118
                                   T2u = T1z * T1D;
 
119
                                   T1B = T1z * T1A;
 
120
                                   T2t = FNMS(T1w, T1u, T2s);
 
121
                                   T1y = FMA(T1w, T1x, T1v);
 
122
                                   T2v = FNMS(T1C, T1A, T2u);
 
123
                                   T1E = FMA(T1C, T1D, T1B);
 
124
                                   Td = ii[WS(ios, 4)];
 
125
                                   T9 = W[6];
 
126
                                   T2w = T2t - T2v;
 
127
                                   T36 = T2t + T2v;
 
128
                                   T2p = T1y - T1E;
 
129
                                   T1F = T1y + T1E;
 
130
                                   Tb = T9 * Ta;
 
131
                                   T1J = T9 * Td;
 
132
                              }
 
133
                              Tg = ri[WS(ios, 12)];
 
134
                              Tj = ii[WS(ios, 12)];
 
135
                              Tf = W[22];
 
136
                              Tc = W[7];
 
137
                              Ti = W[23];
 
138
                         }
 
139
                    }
 
140
                    {
 
141
                         E TQ, TT, TR, T25, TW, TZ, TV, TS, TY;
 
142
                         {
 
143
                              E To, Tr, Tp, T1P, Tu, Tx, Tt, Tq, Tw;
 
144
                              {
 
145
                                   E T1K, Te, T1M, Tk, T1L, Th, Tn;
 
146
                                   To = ri[WS(ios, 2)];
 
147
                                   T1L = Tf * Tj;
 
148
                                   Th = Tf * Tg;
 
149
                                   T1K = FNMS(Tc, Ta, T1J);
 
150
                                   Te = FMA(Tc, Td, Tb);
 
151
                                   T1M = FNMS(Ti, Tg, T1L);
 
152
                                   Tk = FMA(Ti, Tj, Th);
 
153
                                   Tr = ii[WS(ios, 2)];
 
154
                                   Tn = W[2];
 
155
                                   T3k = T1K + T1M;
 
156
                                   T1N = T1K - T1M;
 
157
                                   T3A = Te - Tk;
 
158
                                   Tl = Te + Tk;
 
159
                                   Tp = Tn * To;
 
160
                                   T1P = Tn * Tr;
 
161
                              }
 
162
                              Tu = ri[WS(ios, 10)];
 
163
                              Tx = ii[WS(ios, 10)];
 
164
                              Tt = W[18];
 
165
                              Tq = W[3];
 
166
                              Tw = W[19];
 
167
                              {
 
168
                                   E T1Q, Ts, T1S, Ty, T1R, Tv, TP;
 
169
                                   TQ = ri[WS(ios, 1)];
 
170
                                   T1R = Tt * Tx;
 
171
                                   Tv = Tt * Tu;
 
172
                                   T1Q = FNMS(Tq, To, T1P);
 
173
                                   Ts = FMA(Tq, Tr, Tp);
 
174
                                   T1S = FNMS(Tw, Tu, T1R);
 
175
                                   Ty = FMA(Tw, Tx, Tv);
 
176
                                   TT = ii[WS(ios, 1)];
 
177
                                   TP = W[0];
 
178
                                   T1T = T1Q - T1S;
 
179
                                   T2V = T1Q + T1S;
 
180
                                   T1U = Ts - Ty;
 
181
                                   Tz = Ts + Ty;
 
182
                                   TR = TP * TQ;
 
183
                                   T25 = TP * TT;
 
184
                              }
 
185
                              TW = ri[WS(ios, 9)];
 
186
                              TZ = ii[WS(ios, 9)];
 
187
                              TV = W[16];
 
188
                              TS = W[1];
 
189
                              TY = W[17];
 
190
                         }
 
191
                         {
 
192
                              E T13, T16, T14, T2d, T19, T1c, T18, T15, T1b;
 
193
                              {
 
194
                                   E T26, TU, T28, T10, T27, TX, T12;
 
195
                                   T13 = ri[WS(ios, 5)];
 
196
                                   T27 = TV * TZ;
 
197
                                   TX = TV * TW;
 
198
                                   T26 = FNMS(TS, TQ, T25);
 
199
                                   TU = FMA(TS, TT, TR);
 
200
                                   T28 = FNMS(TY, TW, T27);
 
201
                                   T10 = FMA(TY, TZ, TX);
 
202
                                   T16 = ii[WS(ios, 5)];
 
203
                                   T12 = W[8];
 
204
                                   T29 = T26 - T28;
 
205
                                   T30 = T26 + T28;
 
206
                                   T2c = TU - T10;
 
207
                                   T11 = TU + T10;
 
208
                                   T14 = T12 * T13;
 
209
                                   T2d = T12 * T16;
 
210
                              }
 
211
                              T19 = ri[WS(ios, 13)];
 
212
                              T1c = ii[WS(ios, 13)];
 
213
                              T18 = W[24];
 
214
                              T15 = W[9];
 
215
                              T1b = W[25];
 
216
                              {
 
217
                                   E T2e, T17, T2g, T1d, T2f, T1a, TA;
 
218
                                   TB = ri[WS(ios, 14)];
 
219
                                   T2f = T18 * T1c;
 
220
                                   T1a = T18 * T19;
 
221
                                   T2e = FNMS(T15, T13, T2d);
 
222
                                   T17 = FMA(T15, T16, T14);
 
223
                                   T2g = FNMS(T1b, T19, T2f);
 
224
                                   T1d = FMA(T1b, T1c, T1a);
 
225
                                   TE = ii[WS(ios, 14)];
 
226
                                   TA = W[26];
 
227
                                   T2h = T2e - T2g;
 
228
                                   T31 = T2e + T2g;
 
229
                                   T2a = T17 - T1d;
 
230
                                   T1e = T17 + T1d;
 
231
                                   TC = TA * TB;
 
232
                                   T1X = TA * TE;
 
233
                              }
 
234
                              TH = ri[WS(ios, 6)];
 
235
                              TK = ii[WS(ios, 6)];
 
236
                              TG = W[10];
 
237
                              TD = W[27];
 
238
                              TJ = W[11];
 
239
                         }
 
240
                    }
 
241
               }
 
242
               {
 
243
                    E T2U, T3u, T2Z, T21, T1W, T34, T2X, T3f, T32, T3t, T1H, T3q, T3e, TO, T3g;
 
244
                    E T37, T3r, T3s, T3h, T3i;
 
245
                    {
 
246
                         E Tm, T1Y, TF, T20, TL, T3p, T1Z, TI;
 
247
                         T2U = T8 - Tl;
 
248
                         Tm = T8 + Tl;
 
249
                         T1Z = TG * TK;
 
250
                         TI = TG * TH;
 
251
                         T1Y = FNMS(TD, TB, T1X);
 
252
                         TF = FMA(TD, TE, TC);
 
253
                         T20 = FNMS(TJ, TH, T1Z);
 
254
                         TL = FMA(TJ, TK, TI);
 
255
                         T3p = T3k + T3o;
 
256
                         T3u = T3o - T3k;
 
257
                         {
 
258
                              E T1f, TM, T1G, T3j, T2W, TN;
 
259
                              T2Z = T11 - T1e;
 
260
                              T1f = T11 + T1e;
 
261
                              T21 = T1Y - T20;
 
262
                              T2W = T1Y + T20;
 
263
                              T1W = TF - TL;
 
264
                              TM = TF + TL;
 
265
                              T1G = T1s + T1F;
 
266
                              T34 = T1s - T1F;
 
267
                              T2X = T2V - T2W;
 
268
                              T3j = T2V + T2W;
 
269
                              T3f = T30 + T31;
 
270
                              T32 = T30 - T31;
 
271
                              T3t = TM - Tz;
 
272
                              TN = Tz + TM;
 
273
                              T3r = T1G - T1f;
 
274
                              T1H = T1f + T1G;
 
275
                              T3s = T3p - T3j;
 
276
                              T3q = T3j + T3p;
 
277
                              T3e = Tm - TN;
 
278
                              TO = Tm + TN;
 
279
                              T3g = T35 + T36;
 
280
                              T37 = T35 - T36;
 
281
                         }
 
282
                    }
 
283
                    ii[WS(ios, 12)] = T3s - T3r;
 
284
                    ii[WS(ios, 4)] = T3r + T3s;
 
285
                    ri[0] = TO + T1H;
 
286
                    ri[WS(ios, 8)] = TO - T1H;
 
287
                    T3h = T3f - T3g;
 
288
                    T3i = T3f + T3g;
 
289
                    {
 
290
                         E T3a, T2Y, T3x, T3v, T3b, T33;
 
291
                         ii[0] = T3i + T3q;
 
292
                         ii[WS(ios, 8)] = T3q - T3i;
 
293
                         ri[WS(ios, 4)] = T3e + T3h;
 
294
                         ri[WS(ios, 12)] = T3e - T3h;
 
295
                         T3a = T2U - T2X;
 
296
                         T2Y = T2U + T2X;
 
297
                         T3x = T3u - T3t;
 
298
                         T3v = T3t + T3u;
 
299
                         T3b = T32 - T2Z;
 
300
                         T33 = T2Z + T32;
 
301
                         {
 
302
                              E T2E, T1O, T3B, T3H, T2x, T2q, T3C, T23, T2S, T2O, T2K, T2J, T3I, T2H, T2B;
 
303
                              E T2j;
 
304
                              {
 
305
                                   E T2F, T1V, T22, T2G, T3c, T38;
 
306
                                   T2E = T1I + T1N;
 
307
                                   T1O = T1I - T1N;
 
308
                                   T3B = T3z - T3A;
 
309
                                   T3H = T3A + T3z;
 
310
                                   T3c = T34 + T37;
 
311
                                   T38 = T34 - T37;
 
312
                                   T2F = T1U + T1T;
 
313
                                   T1V = T1T - T1U;
 
314
                                   {
 
315
                                        E T3d, T3w, T3y, T39;
 
316
                                        T3d = T3b - T3c;
 
317
                                        T3w = T3b + T3c;
 
318
                                        T3y = T38 - T33;
 
319
                                        T39 = T33 + T38;
 
320
                                        ri[WS(ios, 6)] = FMA(KP707106781, T3d, T3a);
 
321
                                        ri[WS(ios, 14)] = FNMS(KP707106781, T3d, T3a);
 
322
                                        ii[WS(ios, 10)] = FNMS(KP707106781, T3w, T3v);
 
323
                                        ii[WS(ios, 2)] = FMA(KP707106781, T3w, T3v);
 
324
                                        ii[WS(ios, 14)] = FNMS(KP707106781, T3y, T3x);
 
325
                                        ii[WS(ios, 6)] = FMA(KP707106781, T3y, T3x);
 
326
                                        ri[WS(ios, 2)] = FMA(KP707106781, T39, T2Y);
 
327
                                        ri[WS(ios, 10)] = FNMS(KP707106781, T39, T2Y);
 
328
                                        T22 = T1W + T21;
 
329
                                        T2G = T1W - T21;
 
330
                                   }
 
331
                                   {
 
332
                                        E T2M, T2N, T2b, T2i;
 
333
                                        T2x = T2r - T2w;
 
334
                                        T2M = T2r + T2w;
 
335
                                        T2N = T2o - T2p;
 
336
                                        T2q = T2o + T2p;
 
337
                                        T3C = T1V + T22;
 
338
                                        T23 = T1V - T22;
 
339
                                        T2S = FMA(KP414213562, T2M, T2N);
 
340
                                        T2O = FNMS(KP414213562, T2N, T2M);
 
341
                                        T2K = T29 - T2a;
 
342
                                        T2b = T29 + T2a;
 
343
                                        T2i = T2c - T2h;
 
344
                                        T2J = T2c + T2h;
 
345
                                        T3I = T2G - T2F;
 
346
                                        T2H = T2F + T2G;
 
347
                                        T2B = FNMS(KP414213562, T2b, T2i);
 
348
                                        T2j = FMA(KP414213562, T2i, T2b);
 
349
                                   }
 
350
                              }
 
351
                              {
 
352
                                   E T2R, T2L, T3L, T3M;
 
353
                                   {
 
354
                                        E T2A, T24, T2C, T2y, T3J, T3K, T2D, T2z;
 
355
                                        T2A = FNMS(KP707106781, T23, T1O);
 
356
                                        T24 = FMA(KP707106781, T23, T1O);
 
357
                                        T2R = FNMS(KP414213562, T2J, T2K);
 
358
                                        T2L = FMA(KP414213562, T2K, T2J);
 
359
                                        T2C = FMA(KP414213562, T2q, T2x);
 
360
                                        T2y = FNMS(KP414213562, T2x, T2q);
 
361
                                        T3J = FMA(KP707106781, T3I, T3H);
 
362
                                        T3L = FNMS(KP707106781, T3I, T3H);
 
363
                                        T3K = T2C - T2B;
 
364
                                        T2D = T2B + T2C;
 
365
                                        T3M = T2j + T2y;
 
366
                                        T2z = T2j - T2y;
 
367
                                        ii[WS(ios, 11)] = FNMS(KP923879532, T3K, T3J);
 
368
                                        ii[WS(ios, 3)] = FMA(KP923879532, T3K, T3J);
 
369
                                        ri[WS(ios, 3)] = FMA(KP923879532, T2z, T24);
 
370
                                        ri[WS(ios, 11)] = FNMS(KP923879532, T2z, T24);
 
371
                                        ri[WS(ios, 15)] = FMA(KP923879532, T2D, T2A);
 
372
                                        ri[WS(ios, 7)] = FNMS(KP923879532, T2D, T2A);
 
373
                                   }
 
374
                                   {
 
375
                                        E T2Q, T3D, T3E, T2T, T2I, T2P;
 
376
                                        T2Q = FNMS(KP707106781, T2H, T2E);
 
377
                                        T2I = FMA(KP707106781, T2H, T2E);
 
378
                                        T2P = T2L + T2O;
 
379
                                        T3G = T2O - T2L;
 
380
                                        T3F = FNMS(KP707106781, T3C, T3B);
 
381
                                        T3D = FMA(KP707106781, T3C, T3B);
 
382
                                        ii[WS(ios, 15)] = FMA(KP923879532, T3M, T3L);
 
383
                                        ii[WS(ios, 7)] = FNMS(KP923879532, T3M, T3L);
 
384
                                        ri[WS(ios, 1)] = FMA(KP923879532, T2P, T2I);
 
385
                                        ri[WS(ios, 9)] = FNMS(KP923879532, T2P, T2I);
 
386
                                        T3E = T2R + T2S;
 
387
                                        T2T = T2R - T2S;
 
388
                                        ii[WS(ios, 9)] = FNMS(KP923879532, T3E, T3D);
 
389
                                        ii[WS(ios, 1)] = FMA(KP923879532, T3E, T3D);
 
390
                                        ri[WS(ios, 5)] = FMA(KP923879532, T2T, T2Q);
 
391
                                        ri[WS(ios, 13)] = FNMS(KP923879532, T2T, T2Q);
 
392
                                   }
 
393
                              }
 
394
                         }
 
395
                    }
 
396
               }
 
397
          }
 
398
          ii[WS(ios, 13)] = FNMS(KP923879532, T3G, T3F);
 
399
          ii[WS(ios, 5)] = FMA(KP923879532, T3G, T3F);
 
400
     }
 
401
     return W;
 
402
}
 
403
 
 
404
static const tw_instr twinstr[] = {
 
405
     {TW_FULL, 0, 16},
 
406
     {TW_NEXT, 1, 0}
 
407
};
 
408
 
 
409
static const ct_desc desc = { 16, "t1_16", twinstr, &GENUS, {104, 30, 70, 0}, 0, 0, 0 };
 
410
 
 
411
void X(codelet_t1_16) (planner *p) {
 
412
     X(kdft_dit_register) (p, t1_16, &desc);
 
413
}
 
414
#else                           /* HAVE_FMA */
 
415
 
 
416
/* Generated by: ../../../genfft/gen_twiddle -compact -variables 4 -pipeline-latency 4 -n 16 -name t1_16 -include t.h */
 
417
 
 
418
/*
 
419
 * This function contains 174 FP additions, 84 FP multiplications,
 
420
 * (or, 136 additions, 46 multiplications, 38 fused multiply/add),
 
421
 * 52 stack variables, and 64 memory accesses
 
422
 */
 
423
/*
 
424
 * Generator Id's : 
 
425
 * $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
 
426
 * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
 
427
 * $Id: gen_twiddle.ml,v 1.24 2006-02-12 23:34:12 athena Exp $
 
428
 */
 
429
 
 
430
#include "t.h"
 
431
 
 
432
static const R *t1_16(R *ri, R *ii, const R *W, stride ios, INT m, INT dist)
 
433
{
 
434
     DK(KP382683432, +0.382683432365089771728459984030398866761344562);
 
435
     DK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
436
     DK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
437
     INT i;
 
438
     for (i = m; i > 0; i = i - 1, ri = ri + dist, ii = ii + dist, W = W + 30, MAKE_VOLATILE_STRIDE(ios)) {
 
439
          E T7, T37, T1t, T2U, Ti, T38, T1w, T2R, Tu, T2s, T1C, T2c, TF, T2t, T1H;
 
440
          E T2d, T1f, T1q, T2B, T2C, T2D, T2E, T1Z, T2j, T24, T2k, TS, T13, T2w, T2x;
 
441
          E T2y, T2z, T1O, T2g, T1T, T2h;
 
442
          {
 
443
               E T1, T2T, T6, T2S;
 
444
               T1 = ri[0];
 
445
               T2T = ii[0];
 
446
               {
 
447
                    E T3, T5, T2, T4;
 
448
                    T3 = ri[WS(ios, 8)];
 
449
                    T5 = ii[WS(ios, 8)];
 
450
                    T2 = W[14];
 
451
                    T4 = W[15];
 
452
                    T6 = FMA(T2, T3, T4 * T5);
 
453
                    T2S = FNMS(T4, T3, T2 * T5);
 
454
               }
 
455
               T7 = T1 + T6;
 
456
               T37 = T2T - T2S;
 
457
               T1t = T1 - T6;
 
458
               T2U = T2S + T2T;
 
459
          }
 
460
          {
 
461
               E Tc, T1u, Th, T1v;
 
462
               {
 
463
                    E T9, Tb, T8, Ta;
 
464
                    T9 = ri[WS(ios, 4)];
 
465
                    Tb = ii[WS(ios, 4)];
 
466
                    T8 = W[6];
 
467
                    Ta = W[7];
 
468
                    Tc = FMA(T8, T9, Ta * Tb);
 
469
                    T1u = FNMS(Ta, T9, T8 * Tb);
 
470
               }
 
471
               {
 
472
                    E Te, Tg, Td, Tf;
 
473
                    Te = ri[WS(ios, 12)];
 
474
                    Tg = ii[WS(ios, 12)];
 
475
                    Td = W[22];
 
476
                    Tf = W[23];
 
477
                    Th = FMA(Td, Te, Tf * Tg);
 
478
                    T1v = FNMS(Tf, Te, Td * Tg);
 
479
               }
 
480
               Ti = Tc + Th;
 
481
               T38 = Tc - Th;
 
482
               T1w = T1u - T1v;
 
483
               T2R = T1u + T1v;
 
484
          }
 
485
          {
 
486
               E To, T1y, Tt, T1z, T1A, T1B;
 
487
               {
 
488
                    E Tl, Tn, Tk, Tm;
 
489
                    Tl = ri[WS(ios, 2)];
 
490
                    Tn = ii[WS(ios, 2)];
 
491
                    Tk = W[2];
 
492
                    Tm = W[3];
 
493
                    To = FMA(Tk, Tl, Tm * Tn);
 
494
                    T1y = FNMS(Tm, Tl, Tk * Tn);
 
495
               }
 
496
               {
 
497
                    E Tq, Ts, Tp, Tr;
 
498
                    Tq = ri[WS(ios, 10)];
 
499
                    Ts = ii[WS(ios, 10)];
 
500
                    Tp = W[18];
 
501
                    Tr = W[19];
 
502
                    Tt = FMA(Tp, Tq, Tr * Ts);
 
503
                    T1z = FNMS(Tr, Tq, Tp * Ts);
 
504
               }
 
505
               Tu = To + Tt;
 
506
               T2s = T1y + T1z;
 
507
               T1A = T1y - T1z;
 
508
               T1B = To - Tt;
 
509
               T1C = T1A - T1B;
 
510
               T2c = T1B + T1A;
 
511
          }
 
512
          {
 
513
               E Tz, T1E, TE, T1F, T1D, T1G;
 
514
               {
 
515
                    E Tw, Ty, Tv, Tx;
 
516
                    Tw = ri[WS(ios, 14)];
 
517
                    Ty = ii[WS(ios, 14)];
 
518
                    Tv = W[26];
 
519
                    Tx = W[27];
 
520
                    Tz = FMA(Tv, Tw, Tx * Ty);
 
521
                    T1E = FNMS(Tx, Tw, Tv * Ty);
 
522
               }
 
523
               {
 
524
                    E TB, TD, TA, TC;
 
525
                    TB = ri[WS(ios, 6)];
 
526
                    TD = ii[WS(ios, 6)];
 
527
                    TA = W[10];
 
528
                    TC = W[11];
 
529
                    TE = FMA(TA, TB, TC * TD);
 
530
                    T1F = FNMS(TC, TB, TA * TD);
 
531
               }
 
532
               TF = Tz + TE;
 
533
               T2t = T1E + T1F;
 
534
               T1D = Tz - TE;
 
535
               T1G = T1E - T1F;
 
536
               T1H = T1D + T1G;
 
537
               T2d = T1D - T1G;
 
538
          }
 
539
          {
 
540
               E T19, T20, T1p, T1X, T1e, T21, T1k, T1W;
 
541
               {
 
542
                    E T16, T18, T15, T17;
 
543
                    T16 = ri[WS(ios, 15)];
 
544
                    T18 = ii[WS(ios, 15)];
 
545
                    T15 = W[28];
 
546
                    T17 = W[29];
 
547
                    T19 = FMA(T15, T16, T17 * T18);
 
548
                    T20 = FNMS(T17, T16, T15 * T18);
 
549
               }
 
550
               {
 
551
                    E T1m, T1o, T1l, T1n;
 
552
                    T1m = ri[WS(ios, 11)];
 
553
                    T1o = ii[WS(ios, 11)];
 
554
                    T1l = W[20];
 
555
                    T1n = W[21];
 
556
                    T1p = FMA(T1l, T1m, T1n * T1o);
 
557
                    T1X = FNMS(T1n, T1m, T1l * T1o);
 
558
               }
 
559
               {
 
560
                    E T1b, T1d, T1a, T1c;
 
561
                    T1b = ri[WS(ios, 7)];
 
562
                    T1d = ii[WS(ios, 7)];
 
563
                    T1a = W[12];
 
564
                    T1c = W[13];
 
565
                    T1e = FMA(T1a, T1b, T1c * T1d);
 
566
                    T21 = FNMS(T1c, T1b, T1a * T1d);
 
567
               }
 
568
               {
 
569
                    E T1h, T1j, T1g, T1i;
 
570
                    T1h = ri[WS(ios, 3)];
 
571
                    T1j = ii[WS(ios, 3)];
 
572
                    T1g = W[4];
 
573
                    T1i = W[5];
 
574
                    T1k = FMA(T1g, T1h, T1i * T1j);
 
575
                    T1W = FNMS(T1i, T1h, T1g * T1j);
 
576
               }
 
577
               T1f = T19 + T1e;
 
578
               T1q = T1k + T1p;
 
579
               T2B = T1f - T1q;
 
580
               T2C = T20 + T21;
 
581
               T2D = T1W + T1X;
 
582
               T2E = T2C - T2D;
 
583
               {
 
584
                    E T1V, T1Y, T22, T23;
 
585
                    T1V = T19 - T1e;
 
586
                    T1Y = T1W - T1X;
 
587
                    T1Z = T1V - T1Y;
 
588
                    T2j = T1V + T1Y;
 
589
                    T22 = T20 - T21;
 
590
                    T23 = T1k - T1p;
 
591
                    T24 = T22 + T23;
 
592
                    T2k = T22 - T23;
 
593
               }
 
594
          }
 
595
          {
 
596
               E TM, T1K, T12, T1R, TR, T1L, TX, T1Q;
 
597
               {
 
598
                    E TJ, TL, TI, TK;
 
599
                    TJ = ri[WS(ios, 1)];
 
600
                    TL = ii[WS(ios, 1)];
 
601
                    TI = W[0];
 
602
                    TK = W[1];
 
603
                    TM = FMA(TI, TJ, TK * TL);
 
604
                    T1K = FNMS(TK, TJ, TI * TL);
 
605
               }
 
606
               {
 
607
                    E TZ, T11, TY, T10;
 
608
                    TZ = ri[WS(ios, 13)];
 
609
                    T11 = ii[WS(ios, 13)];
 
610
                    TY = W[24];
 
611
                    T10 = W[25];
 
612
                    T12 = FMA(TY, TZ, T10 * T11);
 
613
                    T1R = FNMS(T10, TZ, TY * T11);
 
614
               }
 
615
               {
 
616
                    E TO, TQ, TN, TP;
 
617
                    TO = ri[WS(ios, 9)];
 
618
                    TQ = ii[WS(ios, 9)];
 
619
                    TN = W[16];
 
620
                    TP = W[17];
 
621
                    TR = FMA(TN, TO, TP * TQ);
 
622
                    T1L = FNMS(TP, TO, TN * TQ);
 
623
               }
 
624
               {
 
625
                    E TU, TW, TT, TV;
 
626
                    TU = ri[WS(ios, 5)];
 
627
                    TW = ii[WS(ios, 5)];
 
628
                    TT = W[8];
 
629
                    TV = W[9];
 
630
                    TX = FMA(TT, TU, TV * TW);
 
631
                    T1Q = FNMS(TV, TU, TT * TW);
 
632
               }
 
633
               TS = TM + TR;
 
634
               T13 = TX + T12;
 
635
               T2w = TS - T13;
 
636
               T2x = T1K + T1L;
 
637
               T2y = T1Q + T1R;
 
638
               T2z = T2x - T2y;
 
639
               {
 
640
                    E T1M, T1N, T1P, T1S;
 
641
                    T1M = T1K - T1L;
 
642
                    T1N = TX - T12;
 
643
                    T1O = T1M + T1N;
 
644
                    T2g = T1M - T1N;
 
645
                    T1P = TM - TR;
 
646
                    T1S = T1Q - T1R;
 
647
                    T1T = T1P - T1S;
 
648
                    T2h = T1P + T1S;
 
649
               }
 
650
          }
 
651
          {
 
652
               E T1J, T27, T3g, T3i, T26, T3h, T2a, T3d;
 
653
               {
 
654
                    E T1x, T1I, T3e, T3f;
 
655
                    T1x = T1t - T1w;
 
656
                    T1I = KP707106781 * (T1C - T1H);
 
657
                    T1J = T1x + T1I;
 
658
                    T27 = T1x - T1I;
 
659
                    T3e = KP707106781 * (T2d - T2c);
 
660
                    T3f = T38 + T37;
 
661
                    T3g = T3e + T3f;
 
662
                    T3i = T3f - T3e;
 
663
               }
 
664
               {
 
665
                    E T1U, T25, T28, T29;
 
666
                    T1U = FMA(KP923879532, T1O, KP382683432 * T1T);
 
667
                    T25 = FNMS(KP923879532, T24, KP382683432 * T1Z);
 
668
                    T26 = T1U + T25;
 
669
                    T3h = T25 - T1U;
 
670
                    T28 = FNMS(KP923879532, T1T, KP382683432 * T1O);
 
671
                    T29 = FMA(KP382683432, T24, KP923879532 * T1Z);
 
672
                    T2a = T28 - T29;
 
673
                    T3d = T28 + T29;
 
674
               }
 
675
               ri[WS(ios, 11)] = T1J - T26;
 
676
               ii[WS(ios, 11)] = T3g - T3d;
 
677
               ri[WS(ios, 3)] = T1J + T26;
 
678
               ii[WS(ios, 3)] = T3d + T3g;
 
679
               ri[WS(ios, 15)] = T27 - T2a;
 
680
               ii[WS(ios, 15)] = T3i - T3h;
 
681
               ri[WS(ios, 7)] = T27 + T2a;
 
682
               ii[WS(ios, 7)] = T3h + T3i;
 
683
          }
 
684
          {
 
685
               E T2v, T2H, T32, T34, T2G, T33, T2K, T2Z;
 
686
               {
 
687
                    E T2r, T2u, T30, T31;
 
688
                    T2r = T7 - Ti;
 
689
                    T2u = T2s - T2t;
 
690
                    T2v = T2r + T2u;
 
691
                    T2H = T2r - T2u;
 
692
                    T30 = TF - Tu;
 
693
                    T31 = T2U - T2R;
 
694
                    T32 = T30 + T31;
 
695
                    T34 = T31 - T30;
 
696
               }
 
697
               {
 
698
                    E T2A, T2F, T2I, T2J;
 
699
                    T2A = T2w + T2z;
 
700
                    T2F = T2B - T2E;
 
701
                    T2G = KP707106781 * (T2A + T2F);
 
702
                    T33 = KP707106781 * (T2F - T2A);
 
703
                    T2I = T2z - T2w;
 
704
                    T2J = T2B + T2E;
 
705
                    T2K = KP707106781 * (T2I - T2J);
 
706
                    T2Z = KP707106781 * (T2I + T2J);
 
707
               }
 
708
               ri[WS(ios, 10)] = T2v - T2G;
 
709
               ii[WS(ios, 10)] = T32 - T2Z;
 
710
               ri[WS(ios, 2)] = T2v + T2G;
 
711
               ii[WS(ios, 2)] = T2Z + T32;
 
712
               ri[WS(ios, 14)] = T2H - T2K;
 
713
               ii[WS(ios, 14)] = T34 - T33;
 
714
               ri[WS(ios, 6)] = T2H + T2K;
 
715
               ii[WS(ios, 6)] = T33 + T34;
 
716
          }
 
717
          {
 
718
               E T2f, T2n, T3a, T3c, T2m, T3b, T2q, T35;
 
719
               {
 
720
                    E T2b, T2e, T36, T39;
 
721
                    T2b = T1t + T1w;
 
722
                    T2e = KP707106781 * (T2c + T2d);
 
723
                    T2f = T2b + T2e;
 
724
                    T2n = T2b - T2e;
 
725
                    T36 = KP707106781 * (T1C + T1H);
 
726
                    T39 = T37 - T38;
 
727
                    T3a = T36 + T39;
 
728
                    T3c = T39 - T36;
 
729
               }
 
730
               {
 
731
                    E T2i, T2l, T2o, T2p;
 
732
                    T2i = FMA(KP382683432, T2g, KP923879532 * T2h);
 
733
                    T2l = FNMS(KP382683432, T2k, KP923879532 * T2j);
 
734
                    T2m = T2i + T2l;
 
735
                    T3b = T2l - T2i;
 
736
                    T2o = FNMS(KP382683432, T2h, KP923879532 * T2g);
 
737
                    T2p = FMA(KP923879532, T2k, KP382683432 * T2j);
 
738
                    T2q = T2o - T2p;
 
739
                    T35 = T2o + T2p;
 
740
               }
 
741
               ri[WS(ios, 9)] = T2f - T2m;
 
742
               ii[WS(ios, 9)] = T3a - T35;
 
743
               ri[WS(ios, 1)] = T2f + T2m;
 
744
               ii[WS(ios, 1)] = T35 + T3a;
 
745
               ri[WS(ios, 13)] = T2n - T2q;
 
746
               ii[WS(ios, 13)] = T3c - T3b;
 
747
               ri[WS(ios, 5)] = T2n + T2q;
 
748
               ii[WS(ios, 5)] = T3b + T3c;
 
749
          }
 
750
          {
 
751
               E TH, T2L, T2W, T2Y, T1s, T2X, T2O, T2P;
 
752
               {
 
753
                    E Tj, TG, T2Q, T2V;
 
754
                    Tj = T7 + Ti;
 
755
                    TG = Tu + TF;
 
756
                    TH = Tj + TG;
 
757
                    T2L = Tj - TG;
 
758
                    T2Q = T2s + T2t;
 
759
                    T2V = T2R + T2U;
 
760
                    T2W = T2Q + T2V;
 
761
                    T2Y = T2V - T2Q;
 
762
               }
 
763
               {
 
764
                    E T14, T1r, T2M, T2N;
 
765
                    T14 = TS + T13;
 
766
                    T1r = T1f + T1q;
 
767
                    T1s = T14 + T1r;
 
768
                    T2X = T1r - T14;
 
769
                    T2M = T2x + T2y;
 
770
                    T2N = T2C + T2D;
 
771
                    T2O = T2M - T2N;
 
772
                    T2P = T2M + T2N;
 
773
               }
 
774
               ri[WS(ios, 8)] = TH - T1s;
 
775
               ii[WS(ios, 8)] = T2W - T2P;
 
776
               ri[0] = TH + T1s;
 
777
               ii[0] = T2P + T2W;
 
778
               ri[WS(ios, 12)] = T2L - T2O;
 
779
               ii[WS(ios, 12)] = T2Y - T2X;
 
780
               ri[WS(ios, 4)] = T2L + T2O;
 
781
               ii[WS(ios, 4)] = T2X + T2Y;
 
782
          }
 
783
     }
 
784
     return W;
 
785
}
 
786
 
 
787
static const tw_instr twinstr[] = {
 
788
     {TW_FULL, 0, 16},
 
789
     {TW_NEXT, 1, 0}
 
790
};
 
791
 
 
792
static const ct_desc desc = { 16, "t1_16", twinstr, &GENUS, {136, 46, 38, 0}, 0, 0, 0 };
 
793
 
 
794
void X(codelet_t1_16) (planner *p) {
 
795
     X(kdft_dit_register) (p, t1_16, &desc);
 
796
}
 
797
#endif                          /* HAVE_FMA */