← Back to branch summary

~diresu/blender/blender-command-port

« back to all changes in this revision

Viewing changes to extern/fftw/rdft/codelets/r2hc/hf_9.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:
extern/fftw/rdft/codelets/r2hc/hf_9.c
 
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 Sun Jul  2 15:50:22 EDT 2006 */
 
23
 
 
24
#include "codelet-rdft.h"
 
25
 
 
26
#ifdef HAVE_FMA
 
27
 
 
28
/* Generated by: ../../../genfft/gen_hc2hc -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 9 -dit -name hf_9 -include hf.h */
 
29
 
 
30
/*
 
31
 * This function contains 96 FP additions, 88 FP multiplications,
 
32
 * (or, 24 additions, 16 multiplications, 72 fused multiply/add),
 
33
 * 75 stack variables, and 36 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_hc2hc.ml,v 1.16 2006-02-12 23:34:12 athena Exp $
 
40
 */
 
41
 
 
42
#include "hf.h"
 
43
 
 
44
static const R *hf_9(R *rio, R *iio, const R *W, stride ios, INT m, INT dist)
 
45
{
 
46
     DK(KP777861913, +0.777861913430206160028177977318626690410586096);
 
47
     DK(KP852868531, +0.852868531952443209628250963940074071936020296);
 
48
     DK(KP839099631, +0.839099631177280011763127298123181364687434283);
 
49
     DK(KP492403876, +0.492403876506104029683371512294761506835321626);
 
50
     DK(KP984807753, +0.984807753012208059366743024589523013670643252);
 
51
     DK(KP954188894, +0.954188894138671133499268364187245676532219158);
 
52
     DK(KP363970234, +0.363970234266202361351047882776834043890471784);
 
53
     DK(KP176326980, +0.176326980708464973471090386868618986121633062);
 
54
     DK(KP866025403, +0.866025403784438646763723170752936183471402627);
 
55
     DK(KP500000000, +0.500000000000000000000000000000000000000000000);
 
56
     INT i;
 
57
     for (i = m - 2; i > 0; i = i - 2, rio = rio + dist, iio = iio - dist, W = W + 16, MAKE_VOLATILE_STRIDE(ios)) {
 
58
          E T1K, T20, T1H, T1Z;
 
59
          {
 
60
               E T1, T1R, T10, T1Q, Te, T1W, Th, Tk, T1l, T1r, T1q, T1N, Ti, T1g, TT;
 
61
               E T12, Tn, Tq, Tp, T17, Tx, T14, To, Tj;
 
62
               {
 
63
                    E T9, Tc, TY, Ta, Tb, TX, T7, TZ, Td;
 
64
                    T1 = rio[0];
 
65
                    {
 
66
                         E T3, T6, T8, TW, T4, T2, T5;
 
67
                         T3 = rio[WS(ios, 3)];
 
68
                         T6 = iio[-WS(ios, 5)];
 
69
                         T2 = W[4];
 
70
                         T9 = rio[WS(ios, 6)];
 
71
                         Tc = iio[-WS(ios, 2)];
 
72
                         T8 = W[10];
 
73
                         TW = T2 * T6;
 
74
                         T4 = T2 * T3;
 
75
                         T5 = W[5];
 
76
                         TY = T8 * Tc;
 
77
                         Ta = T8 * T9;
 
78
                         Tb = W[11];
 
79
                         TX = FNMS(T5, T3, TW);
 
80
                         T7 = FMA(T5, T6, T4);
 
81
                    }
 
82
                    T1R = iio[-WS(ios, 8)];
 
83
                    TZ = FNMS(Tb, T9, TY);
 
84
                    Td = FMA(Tb, Tc, Ta);
 
85
                    {
 
86
                         E TB, TE, T1n, TC, TH, TK, T1k, TR, TG, TJ, TD, TA;
 
87
                         TB = rio[WS(ios, 2)];
 
88
                         TE = iio[-WS(ios, 6)];
 
89
                         T10 = TX - TZ;
 
90
                         T1Q = TX + TZ;
 
91
                         Te = T7 + Td;
 
92
                         T1W = Td - T7;
 
93
                         TA = W[2];
 
94
                         {
 
95
                              E TN, TQ, TP, T1j, TO, TM;
 
96
                              TN = rio[WS(ios, 8)];
 
97
                              TQ = iio[0];
 
98
                              TM = W[14];
 
99
                              T1n = TA * TE;
 
100
                              TC = TA * TB;
 
101
                              TP = W[15];
 
102
                              T1j = TM * TQ;
 
103
                              TO = TM * TN;
 
104
                              TH = rio[WS(ios, 5)];
 
105
                              TK = iio[-WS(ios, 3)];
 
106
                              T1k = FNMS(TP, TN, T1j);
 
107
                              TR = FMA(TP, TQ, TO);
 
108
                              TG = W[8];
 
109
                              TJ = W[9];
 
110
                         }
 
111
                         TD = W[3];
 
112
                         {
 
113
                              E T1o, TF, T1i, TL, T1h, TI, Tg, T1p, TS;
 
114
                              Th = rio[WS(ios, 1)];
 
115
                              T1h = TG * TK;
 
116
                              TI = TG * TH;
 
117
                              T1o = FNMS(TD, TB, T1n);
 
118
                              TF = FMA(TD, TE, TC);
 
119
                              T1i = FNMS(TJ, TH, T1h);
 
120
                              TL = FMA(TJ, TK, TI);
 
121
                              Tk = iio[-WS(ios, 7)];
 
122
                              Tg = W[0];
 
123
                              T1p = T1i + T1k;
 
124
                              T1l = T1i - T1k;
 
125
                              TS = TL + TR;
 
126
                              T1r = TR - TL;
 
127
                              T1q = FNMS(KP500000000, T1p, T1o);
 
128
                              T1N = T1o + T1p;
 
129
                              Ti = Tg * Th;
 
130
                              T1g = FNMS(KP500000000, TS, TF);
 
131
                              TT = TF + TS;
 
132
                              T12 = Tg * Tk;
 
133
                         }
 
134
                         {
 
135
                              E Tt, Tw, Ts, Tv, T16, Tu, Tm;
 
136
                              Tt = rio[WS(ios, 7)];
 
137
                              Tw = iio[-WS(ios, 1)];
 
138
                              Ts = W[12];
 
139
                              Tv = W[13];
 
140
                              Tn = rio[WS(ios, 4)];
 
141
                              Tq = iio[-WS(ios, 4)];
 
142
                              T16 = Ts * Tw;
 
143
                              Tu = Ts * Tt;
 
144
                              Tm = W[6];
 
145
                              Tp = W[7];
 
146
                              T17 = FNMS(Tv, Tt, T16);
 
147
                              Tx = FMA(Tv, Tw, Tu);
 
148
                              T14 = Tm * Tq;
 
149
                              To = Tm * Tn;
 
150
                         }
 
151
                         Tj = W[1];
 
152
                    }
 
153
               }
 
154
               {
 
155
                    E TV, Tf, T1d, T1a, T19, T1M, T1V, T1S, T1c, Tz, Tl, Ty;
 
156
                    TV = FNMS(KP500000000, Te, T1);
 
157
                    Tf = T1 + Te;
 
158
                    {
 
159
                         E T15, Tr, T13, T18;
 
160
                         T15 = FNMS(Tp, Tn, T14);
 
161
                         Tr = FMA(Tp, Tq, To);
 
162
                         T13 = FNMS(Tj, Th, T12);
 
163
                         Tl = FMA(Tj, Tk, Ti);
 
164
                         T18 = T15 + T17;
 
165
                         T1d = T15 - T17;
 
166
                         Ty = Tr + Tx;
 
167
                         T1a = Tr - Tx;
 
168
                         T19 = FNMS(KP500000000, T18, T13);
 
169
                         T1M = T13 + T18;
 
170
                         T1V = FNMS(KP500000000, T1Q, T1R);
 
171
                         T1S = T1Q + T1R;
 
172
                    }
 
173
                    T1c = FNMS(KP500000000, Ty, Tl);
 
174
                    Tz = Tl + Ty;
 
175
                    {
 
176
                         E T11, T1z, T1E, T1D, T21, T1X, T1I, T1C, T22, T1y, T24, T1u, T1U, TU;
 
177
                         T1U = TT - Tz;
 
178
                         TU = Tz + TT;
 
179
                         {
 
180
                              E T1P, T1O, T1L, T1T;
 
181
                              T1P = T1M + T1N;
 
182
                              T1O = T1M - T1N;
 
183
                              T11 = FNMS(KP866025403, T10, TV);
 
184
                              T1z = FMA(KP866025403, T10, TV);
 
185
                              T1L = FNMS(KP500000000, TU, Tf);
 
186
                              rio[0] = Tf + TU;
 
187
                              T1T = FNMS(KP500000000, T1P, T1S);
 
188
                              iio[0] = T1P + T1S;
 
189
                              rio[WS(ios, 3)] = FMA(KP866025403, T1O, T1L);
 
190
                              iio[-WS(ios, 6)] = FNMS(KP866025403, T1O, T1L);
 
191
                              iio[-WS(ios, 3)] = FMA(KP866025403, T1U, T1T);
 
192
                              rio[WS(ios, 6)] = FMS(KP866025403, T1U, T1T);
 
193
                         }
 
194
                         {
 
195
                              E T1B, T1m, T1w, T1f, T1s, T1A, T1b, T1e, T1x, T1t;
 
196
                              T1E = FNMS(KP866025403, T1a, T19);
 
197
                              T1b = FMA(KP866025403, T1a, T19);
 
198
                              T1e = FNMS(KP866025403, T1d, T1c);
 
199
                              T1D = FMA(KP866025403, T1d, T1c);
 
200
                              T1B = FMA(KP866025403, T1l, T1g);
 
201
                              T1m = FNMS(KP866025403, T1l, T1g);
 
202
                              T21 = FNMS(KP866025403, T1W, T1V);
 
203
                              T1X = FMA(KP866025403, T1W, T1V);
 
204
                              T1w = FNMS(KP176326980, T1b, T1e);
 
205
                              T1f = FMA(KP176326980, T1e, T1b);
 
206
                              T1s = FNMS(KP866025403, T1r, T1q);
 
207
                              T1A = FMA(KP866025403, T1r, T1q);
 
208
                              T1x = FMA(KP363970234, T1m, T1s);
 
209
                              T1t = FNMS(KP363970234, T1s, T1m);
 
210
                              T1I = FNMS(KP176326980, T1A, T1B);
 
211
                              T1C = FMA(KP176326980, T1B, T1A);
 
212
                              T22 = FMA(KP954188894, T1x, T1w);
 
213
                              T1y = FNMS(KP954188894, T1x, T1w);
 
214
                              T24 = FMA(KP954188894, T1t, T1f);
 
215
                              T1u = FNMS(KP954188894, T1t, T1f);
 
216
                         }
 
217
                         {
 
218
                              E T1Y, T1G, T23, T1F, T1J, T1v;
 
219
                              iio[-WS(ios, 2)] = FNMS(KP984807753, T22, T21);
 
220
                              T1v = FNMS(KP492403876, T1u, T11);
 
221
                              rio[WS(ios, 2)] = FMA(KP984807753, T1u, T11);
 
222
                              T1F = FMA(KP839099631, T1E, T1D);
 
223
                              T1J = FNMS(KP839099631, T1D, T1E);
 
224
                              iio[-WS(ios, 5)] = FNMS(KP852868531, T1y, T1v);
 
225
                              iio[-WS(ios, 8)] = FMA(KP852868531, T1y, T1v);
 
226
                              T1K = FMA(KP777861913, T1J, T1I);
 
227
                              T1Y = FNMS(KP777861913, T1J, T1I);
 
228
                              T1G = FMA(KP777861913, T1F, T1C);
 
229
                              T20 = FNMS(KP777861913, T1F, T1C);
 
230
                              T23 = FMA(KP492403876, T22, T21);
 
231
                              iio[-WS(ios, 1)] = FNMS(KP984807753, T1Y, T1X);
 
232
                              rio[WS(ios, 1)] = FMA(KP984807753, T1G, T1z);
 
233
                              T1H = FNMS(KP492403876, T1G, T1z);
 
234
                              rio[WS(ios, 8)] = -(FMA(KP852868531, T24, T23));
 
235
                              rio[WS(ios, 5)] = FMS(KP852868531, T24, T23);
 
236
                              T1Z = FMA(KP492403876, T1Y, T1X);
 
237
                         }
 
238
                    }
 
239
               }
 
240
          }
 
241
          iio[-WS(ios, 7)] = FNMS(KP852868531, T1K, T1H);
 
242
          rio[WS(ios, 4)] = FMA(KP852868531, T1K, T1H);
 
243
          iio[-WS(ios, 4)] = FMA(KP852868531, T20, T1Z);
 
244
          rio[WS(ios, 7)] = FMS(KP852868531, T20, T1Z);
 
245
     }
 
246
     return W;
 
247
}
 
248
 
 
249
static const tw_instr twinstr[] = {
 
250
     {TW_FULL, 0, 9},
 
251
     {TW_NEXT, 1, 0}
 
252
};
 
253
 
 
254
static const hc2hc_desc desc = { 9, "hf_9", twinstr, &GENUS, {24, 16, 72, 0}, 0, 0, 0 };
 
255
 
 
256
void X(codelet_hf_9) (planner *p) {
 
257
     X(khc2hc_register) (p, hf_9, &desc);
 
258
}
 
259
#else                           /* HAVE_FMA */
 
260
 
 
261
/* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -n 9 -dit -name hf_9 -include hf.h */
 
262
 
 
263
/*
 
264
 * This function contains 96 FP additions, 72 FP multiplications,
 
265
 * (or, 60 additions, 36 multiplications, 36 fused multiply/add),
 
266
 * 41 stack variables, and 36 memory accesses
 
267
 */
 
268
/*
 
269
 * Generator Id's : 
 
270
 * $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
 
271
 * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
 
272
 * $Id: gen_hc2hc.ml,v 1.16 2006-02-12 23:34:12 athena Exp $
 
273
 */
 
274
 
 
275
#include "hf.h"
 
276
 
 
277
static const R *hf_9(R *rio, R *iio, const R *W, stride ios, INT m, INT dist)
 
278
{
 
279
     DK(KP939692620, +0.939692620785908384054109277324731469936208134);
 
280
     DK(KP342020143, +0.342020143325668733044099614682259580763083368);
 
281
     DK(KP984807753, +0.984807753012208059366743024589523013670643252);
 
282
     DK(KP173648177, +0.173648177666930348851716626769314796000375677);
 
283
     DK(KP642787609, +0.642787609686539326322643409907263432907559884);
 
284
     DK(KP766044443, +0.766044443118978035202392650555416673935832457);
 
285
     DK(KP500000000, +0.500000000000000000000000000000000000000000000);
 
286
     DK(KP866025403, +0.866025403784438646763723170752936183471402627);
 
287
     INT i;
 
288
     for (i = m - 2; i > 0; i = i - 2, rio = rio + dist, iio = iio - dist, W = W + 16, MAKE_VOLATILE_STRIDE(ios)) {
 
289
          E T1, T1B, TQ, T1G, Tc, TN, T1A, T1H, TL, T1x, T17, T1o, T1c, T1n, Tu;
 
290
          E T1w, TW, T1k, T11, T1l;
 
291
          {
 
292
               E T6, TO, Tb, TP;
 
293
               T1 = rio[0];
 
294
               T1B = iio[-WS(ios, 8)];
 
295
               {
 
296
                    E T3, T5, T2, T4;
 
297
                    T3 = rio[WS(ios, 3)];
 
298
                    T5 = iio[-WS(ios, 5)];
 
299
                    T2 = W[4];
 
300
                    T4 = W[5];
 
301
                    T6 = FMA(T2, T3, T4 * T5);
 
302
                    TO = FNMS(T4, T3, T2 * T5);
 
303
               }
 
304
               {
 
305
                    E T8, Ta, T7, T9;
 
306
                    T8 = rio[WS(ios, 6)];
 
307
                    Ta = iio[-WS(ios, 2)];
 
308
                    T7 = W[10];
 
309
                    T9 = W[11];
 
310
                    Tb = FMA(T7, T8, T9 * Ta);
 
311
                    TP = FNMS(T9, T8, T7 * Ta);
 
312
               }
 
313
               TQ = KP866025403 * (TO - TP);
 
314
               T1G = KP866025403 * (Tb - T6);
 
315
               Tc = T6 + Tb;
 
316
               TN = FNMS(KP500000000, Tc, T1);
 
317
               T1A = TO + TP;
 
318
               T1H = FNMS(KP500000000, T1A, T1B);
 
319
          }
 
320
          {
 
321
               E Tz, T19, TE, T14, TJ, T15, TK, T1a;
 
322
               {
 
323
                    E Tw, Ty, Tv, Tx;
 
324
                    Tw = rio[WS(ios, 2)];
 
325
                    Ty = iio[-WS(ios, 6)];
 
326
                    Tv = W[2];
 
327
                    Tx = W[3];
 
328
                    Tz = FMA(Tv, Tw, Tx * Ty);
 
329
                    T19 = FNMS(Tx, Tw, Tv * Ty);
 
330
               }
 
331
               {
 
332
                    E TB, TD, TA, TC;
 
333
                    TB = rio[WS(ios, 5)];
 
334
                    TD = iio[-WS(ios, 3)];
 
335
                    TA = W[8];
 
336
                    TC = W[9];
 
337
                    TE = FMA(TA, TB, TC * TD);
 
338
                    T14 = FNMS(TC, TB, TA * TD);
 
339
               }
 
340
               {
 
341
                    E TG, TI, TF, TH;
 
342
                    TG = rio[WS(ios, 8)];
 
343
                    TI = iio[0];
 
344
                    TF = W[14];
 
345
                    TH = W[15];
 
346
                    TJ = FMA(TF, TG, TH * TI);
 
347
                    T15 = FNMS(TH, TG, TF * TI);
 
348
               }
 
349
               TK = TE + TJ;
 
350
               T1a = T14 + T15;
 
351
               TL = Tz + TK;
 
352
               T1x = T19 + T1a;
 
353
               {
 
354
                    E T13, T16, T18, T1b;
 
355
                    T13 = FNMS(KP500000000, TK, Tz);
 
356
                    T16 = KP866025403 * (T14 - T15);
 
357
                    T17 = T13 + T16;
 
358
                    T1o = T13 - T16;
 
359
                    T18 = KP866025403 * (TJ - TE);
 
360
                    T1b = FNMS(KP500000000, T1a, T19);
 
361
                    T1c = T18 + T1b;
 
362
                    T1n = T1b - T18;
 
363
               }
 
364
          }
 
365
          {
 
366
               E Ti, TY, Tn, TT, Ts, TU, Tt, TZ;
 
367
               {
 
368
                    E Tf, Th, Te, Tg;
 
369
                    Tf = rio[WS(ios, 1)];
 
370
                    Th = iio[-WS(ios, 7)];
 
371
                    Te = W[0];
 
372
                    Tg = W[1];
 
373
                    Ti = FMA(Te, Tf, Tg * Th);
 
374
                    TY = FNMS(Tg, Tf, Te * Th);
 
375
               }
 
376
               {
 
377
                    E Tk, Tm, Tj, Tl;
 
378
                    Tk = rio[WS(ios, 4)];
 
379
                    Tm = iio[-WS(ios, 4)];
 
380
                    Tj = W[6];
 
381
                    Tl = W[7];
 
382
                    Tn = FMA(Tj, Tk, Tl * Tm);
 
383
                    TT = FNMS(Tl, Tk, Tj * Tm);
 
384
               }
 
385
               {
 
386
                    E Tp, Tr, To, Tq;
 
387
                    Tp = rio[WS(ios, 7)];
 
388
                    Tr = iio[-WS(ios, 1)];
 
389
                    To = W[12];
 
390
                    Tq = W[13];
 
391
                    Ts = FMA(To, Tp, Tq * Tr);
 
392
                    TU = FNMS(Tq, Tp, To * Tr);
 
393
               }
 
394
               Tt = Tn + Ts;
 
395
               TZ = TT + TU;
 
396
               Tu = Ti + Tt;
 
397
               T1w = TY + TZ;
 
398
               {
 
399
                    E TS, TV, TX, T10;
 
400
                    TS = FNMS(KP500000000, Tt, Ti);
 
401
                    TV = KP866025403 * (TT - TU);
 
402
                    TW = TS + TV;
 
403
                    T1k = TS - TV;
 
404
                    TX = KP866025403 * (Ts - Tn);
 
405
                    T10 = FNMS(KP500000000, TZ, TY);
 
406
                    T11 = TX + T10;
 
407
                    T1l = T10 - TX;
 
408
               }
 
409
          }
 
410
          {
 
411
               E T1y, Td, TM, T1v;
 
412
               T1y = KP866025403 * (T1w - T1x);
 
413
               Td = T1 + Tc;
 
414
               TM = Tu + TL;
 
415
               T1v = FNMS(KP500000000, TM, Td);
 
416
               rio[0] = Td + TM;
 
417
               rio[WS(ios, 3)] = T1v + T1y;
 
418
               iio[-WS(ios, 6)] = T1v - T1y;
 
419
          }
 
420
          {
 
421
               E T1D, T1z, T1C, T1E;
 
422
               T1D = KP866025403 * (TL - Tu);
 
423
               T1z = T1w + T1x;
 
424
               T1C = T1A + T1B;
 
425
               T1E = FNMS(KP500000000, T1z, T1C);
 
426
               iio[0] = T1z + T1C;
 
427
               iio[-WS(ios, 3)] = T1D + T1E;
 
428
               rio[WS(ios, 6)] = T1D - T1E;
 
429
          }
 
430
          {
 
431
               E TR, T1I, T1e, T1J, T1i, T1F, T1f, T1K;
 
432
               TR = TN + TQ;
 
433
               T1I = T1G + T1H;
 
434
               {
 
435
                    E T12, T1d, T1g, T1h;
 
436
                    T12 = FMA(KP766044443, TW, KP642787609 * T11);
 
437
                    T1d = FMA(KP173648177, T17, KP984807753 * T1c);
 
438
                    T1e = T12 + T1d;
 
439
                    T1J = KP866025403 * (T1d - T12);
 
440
                    T1g = FNMS(KP642787609, TW, KP766044443 * T11);
 
441
                    T1h = FNMS(KP984807753, T17, KP173648177 * T1c);
 
442
                    T1i = KP866025403 * (T1g - T1h);
 
443
                    T1F = T1g + T1h;
 
444
               }
 
445
               rio[WS(ios, 1)] = TR + T1e;
 
446
               iio[-WS(ios, 1)] = T1F + T1I;
 
447
               T1f = FNMS(KP500000000, T1e, TR);
 
448
               iio[-WS(ios, 7)] = T1f - T1i;
 
449
               rio[WS(ios, 4)] = T1f + T1i;
 
450
               T1K = FNMS(KP500000000, T1F, T1I);
 
451
               rio[WS(ios, 7)] = T1J - T1K;
 
452
               iio[-WS(ios, 4)] = T1J + T1K;
 
453
          }
 
454
          {
 
455
               E T1j, T1M, T1q, T1O, T1u, T1L, T1r, T1N;
 
456
               T1j = TN - TQ;
 
457
               T1M = T1H - T1G;
 
458
               {
 
459
                    E T1m, T1p, T1s, T1t;
 
460
                    T1m = FMA(KP173648177, T1k, KP984807753 * T1l);
 
461
                    T1p = FNMS(KP939692620, T1o, KP342020143 * T1n);
 
462
                    T1q = T1m + T1p;
 
463
                    T1O = KP866025403 * (T1p - T1m);
 
464
                    T1s = FNMS(KP984807753, T1k, KP173648177 * T1l);
 
465
                    T1t = FMA(KP342020143, T1o, KP939692620 * T1n);
 
466
                    T1u = KP866025403 * (T1s + T1t);
 
467
                    T1L = T1s - T1t;
 
468
               }
 
469
               rio[WS(ios, 2)] = T1j + T1q;
 
470
               iio[-WS(ios, 2)] = T1L + T1M;
 
471
               T1r = FNMS(KP500000000, T1q, T1j);
 
472
               iio[-WS(ios, 8)] = T1r - T1u;
 
473
               iio[-WS(ios, 5)] = T1r + T1u;
 
474
               T1N = FMS(KP500000000, T1L, T1M);
 
475
               rio[WS(ios, 5)] = T1N - T1O;
 
476
               rio[WS(ios, 8)] = T1O + T1N;
 
477
          }
 
478
     }
 
479
     return W;
 
480
}
 
481
 
 
482
static const tw_instr twinstr[] = {
 
483
     {TW_FULL, 0, 9},
 
484
     {TW_NEXT, 1, 0}
 
485
};
 
486
 
 
487
static const hc2hc_desc desc = { 9, "hf_9", twinstr, &GENUS, {60, 36, 36, 0}, 0, 0, 0 };
 
488
 
 
489
void X(codelet_hf_9) (planner *p) {
 
490
     X(khc2hc_register) (p, hf_9, &desc);
 
491
}
 
492
#endif                          /* HAVE_FMA */