2
* Copyright (c) 2003, 2007-8 Matteo Frigo
3
* Copyright (c) 2003, 2007-8 Massachusetts Institute of Technology
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.
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.
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
21
/* This file was automatically generated --- DO NOT EDIT */
22
/* Generated on Sun Jul 12 06:39:55 EDT 2009 */
24
#include "codelet-dft.h"
28
/* Generated by: ../../../genfft/gen_notw_c -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n2fv_64 -with-ostride 2 -include n2f.h -store-multiple 2 */
31
* This function contains 456 FP additions, 258 FP multiplications,
32
* (or, 198 additions, 0 multiplications, 258 fused multiply/add),
33
* 178 stack variables, 15 constants, and 160 memory accesses
37
static void n2fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
39
DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
40
DVK(KP820678790, +0.820678790828660330972281985331011598767386482);
41
DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
42
DVK(KP534511135, +0.534511135950791641089685961295362908582039528);
43
DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
44
DVK(KP098491403, +0.098491403357164253077197521291327432293052451);
45
DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
46
DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
47
DVK(KP303346683, +0.303346683607342391675883946941299872384187453);
48
DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
49
DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
50
DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
51
DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
52
DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
53
DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
59
for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
60
V T7r, T7s, T7t, T7u, T5T, T5S, T5X, T65, T8a, T8b, T8e, T8g, T5Z, T5R, T67;
63
V T7, T26, T5k, T6A, T47, T69, T2V, T3z, T6B, T4e, T6a, T5n, T3M, T2Y, T27;
64
V Tm, T3A, T3l, T2a, TC, T5p, T4o, T6E, T6e, T3i, T3B, TR, T29, T4x, T5q;
65
V T6h, T6D, T39, T3H, T3I, T3c, T5N, T57, T72, T6w, T5O, T5e, T71, T6t, T2y;
66
V T1W, T2x, T1N, T33, T34, T3E, T32, T1p, T2v, T1g, T2u, T4M, T5K, T6p, T6Z;
69
V T4g, T4l, T3j, Tu, Tx, T4h, TA, T4i;
71
V T1, T2, T23, T24, T4, T5, T20, T21;
72
T1 = LD(&(xi[0]), ivs, &(xi[0]));
73
T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
74
T23 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
75
T24 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
76
T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
77
T5 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
78
T20 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
79
T21 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
81
V Ta, T48, Tk, T4c, T49, Td, Tf, Tg;
83
V T8, T43, T3, T44, T25, T5i, T6, T45, T22, T9, Ti, Tj, Tb, Tc;
84
T8 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
93
T9 = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
94
Ti = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
95
Tj = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
96
Tb = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
97
Tc = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
102
T46 = VADD(T44, T45);
103
T5j = VSUB(T45, T44);
104
T26 = VSUB(T22, T25);
105
T2U = VADD(T25, T22);
110
T5k = VFNMS(LDK(KP707106781), T5j, T5i);
111
T6A = VFMA(LDK(KP707106781), T5j, T5i);
112
T47 = VFMA(LDK(KP707106781), T46, T43);
113
T69 = VFNMS(LDK(KP707106781), T46, T43);
114
T2V = VADD(T2T, T2U);
115
T3z = VSUB(T2T, T2U);
119
Tf = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
120
Tg = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
123
V Te, T2W, T5l, T4a, Tq, Tt, Tv, Tw, T5m, T4d, Tl, T2X, Ty, Tz, To;
125
To = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
126
Tp = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
129
Tr = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
130
Ts = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
133
T5l = VFMA(LDK(KP414213562), T48, T49);
134
T4a = VFNMS(LDK(KP414213562), T49, T48);
141
Tv = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
142
Tw = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
143
T5m = VFMA(LDK(KP414213562), T4b, T4c);
144
T4d = VFNMS(LDK(KP414213562), T4c, T4b);
147
Ty = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
148
Tz = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
154
T6B = VSUB(T4d, T4a);
155
T4e = VADD(T4a, T4d);
156
T6a = VADD(T5l, T5m);
157
T5n = VSUB(T5l, T5m);
158
T3M = VSUB(T2X, T2W);
159
T2Y = VADD(T2W, T2X);
168
V TK, T4p, T4u, T4k, T6d, T4n, T6c, TL, TN, TO, T3g, TJ, TF, TI;
171
TD = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
172
TE = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
173
TG = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
174
TH = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
175
TK = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
180
T4j = VADD(T4h, T4i);
181
T4m = VSUB(T4h, T4i);
186
T3A = VSUB(T3j, T3k);
187
T3l = VADD(T3j, T3k);
188
T2a = VFMA(LDK(KP414213562), Tu, TB);
189
TC = VFNMS(LDK(KP414213562), TB, Tu);
190
T4k = VFMA(LDK(KP707106781), T4j, T4g);
191
T6d = VFNMS(LDK(KP707106781), T4j, T4g);
192
T4n = VFMA(LDK(KP707106781), T4m, T4l);
193
T6c = VFNMS(LDK(KP707106781), T4m, T4l);
194
TL = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
196
TN = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
197
TO = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
202
V T3a, T1E, T52, T5b, T1x, T4Z, T6r, T6u, T5a, T1U, T55, T5c, T1L, T3b;
204
V T4V, T1t, T58, T1w, T1Q, T1T, T1I, T4Y, T59, T1J, T53, T1H;
206
V T1r, TM, T4r, TP, T4q, T1s, T1u, T1v;
207
T1r = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
208
T5p = VFMA(LDK(KP198912367), T4k, T4n);
209
T4o = VFNMS(LDK(KP198912367), T4n, T4k);
210
T6E = VFMA(LDK(KP668178637), T6c, T6d);
211
T6e = VFNMS(LDK(KP668178637), T6d, T6c);
216
T1s = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
217
T1u = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
218
T1v = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
220
V T1R, T4X, T6g, T4t, T6f, T4w, T1S, T1O, T1P;
221
T1O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
222
T1P = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
223
T1R = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
228
T4s = VADD(T4q, T4r);
229
T4v = VSUB(T4r, T4q);
230
T4V = VSUB(T1r, T1s);
231
T1t = VADD(T1r, T1s);
232
T58 = VSUB(T1v, T1u);
233
T1w = VADD(T1u, T1v);
234
T4X = VSUB(T1O, T1P);
235
T1Q = VADD(T1O, T1P);
236
T3i = VADD(T3g, T3h);
237
T3B = VSUB(T3g, T3h);
238
TR = VFNMS(LDK(KP414213562), TQ, TJ);
239
T29 = VFMA(LDK(KP414213562), TJ, TQ);
240
T6g = VFNMS(LDK(KP707106781), T4s, T4p);
241
T4t = VFMA(LDK(KP707106781), T4s, T4p);
242
T6f = VFNMS(LDK(KP707106781), T4v, T4u);
243
T4w = VFMA(LDK(KP707106781), T4v, T4u);
244
T1S = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
247
V T4W, T1A, T50, T51, T1D, T1F, T1G;
249
V T1y, T1z, T1B, T1C;
250
T1y = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
251
T1z = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
252
T1B = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
253
T1C = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
254
T4x = VFNMS(LDK(KP198912367), T4w, T4t);
255
T5q = VFMA(LDK(KP198912367), T4t, T4w);
256
T6h = VFNMS(LDK(KP668178637), T6g, T6f);
257
T6D = VFMA(LDK(KP668178637), T6f, T6g);
258
T4W = VSUB(T1R, T1S);
259
T1T = VADD(T1R, T1S);
260
T1A = VADD(T1y, T1z);
261
T50 = VSUB(T1y, T1z);
262
T51 = VSUB(T1C, T1B);
263
T1D = VADD(T1B, T1C);
265
T1F = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
266
T1G = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
267
T1I = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
268
T4Y = VADD(T4W, T4X);
269
T59 = VSUB(T4X, T4W);
270
T1J = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
271
T3a = VADD(T1A, T1D);
272
T1E = VSUB(T1A, T1D);
273
T52 = VFMA(LDK(KP414213562), T51, T50);
274
T5b = VFNMS(LDK(KP414213562), T50, T51);
275
T53 = VSUB(T1F, T1G);
276
T1H = VADD(T1F, T1G);
281
V T37, T54, T1K, T38;
282
T1x = VSUB(T1t, T1w);
283
T37 = VADD(T1t, T1w);
284
T4Z = VFMA(LDK(KP707106781), T4Y, T4V);
285
T6r = VFNMS(LDK(KP707106781), T4Y, T4V);
286
T54 = VSUB(T1J, T1I);
287
T1K = VADD(T1I, T1J);
288
T6u = VFNMS(LDK(KP707106781), T59, T58);
289
T5a = VFMA(LDK(KP707106781), T59, T58);
290
T38 = VADD(T1T, T1Q);
291
T1U = VSUB(T1Q, T1T);
292
T55 = VFNMS(LDK(KP414213562), T54, T53);
293
T5c = VFMA(LDK(KP414213562), T53, T54);
294
T1L = VSUB(T1H, T1K);
295
T3b = VADD(T1H, T1K);
296
T39 = VADD(T37, T38);
297
T3H = VSUB(T37, T38);
301
V T4A, TW, T4N, TZ, T1j, T1m, T4O, T4D, T13, T4F, T16, T4G, T1a, T4I, T4J;
304
V TU, TV, TX, TY, T56, T6v;
305
TU = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
306
T56 = VADD(T52, T55);
307
T6v = VSUB(T55, T52);
309
V T5d, T6s, T1V, T1M;
310
T5d = VADD(T5b, T5c);
311
T6s = VSUB(T5c, T5b);
312
T1V = VSUB(T1L, T1E);
313
T1M = VADD(T1E, T1L);
314
T3I = VSUB(T3b, T3a);
315
T3c = VADD(T3a, T3b);
316
T5N = VFNMS(LDK(KP923879532), T56, T4Z);
317
T57 = VFMA(LDK(KP923879532), T56, T4Z);
318
T72 = VFNMS(LDK(KP923879532), T6v, T6u);
319
T6w = VFMA(LDK(KP923879532), T6v, T6u);
320
T5O = VFNMS(LDK(KP923879532), T5d, T5a);
321
T5e = VFMA(LDK(KP923879532), T5d, T5a);
322
T71 = VFMA(LDK(KP923879532), T6s, T6r);
323
T6t = VFNMS(LDK(KP923879532), T6s, T6r);
324
T2y = VFNMS(LDK(KP707106781), T1V, T1U);
325
T1W = VFMA(LDK(KP707106781), T1V, T1U);
326
T2x = VFNMS(LDK(KP707106781), T1M, T1x);
327
T1N = VFMA(LDK(KP707106781), T1M, T1x);
328
TV = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
330
TX = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
331
TY = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
333
V T1h, T1i, T1k, T1l;
334
T1h = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
335
T1i = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
336
T1k = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
337
T1l = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
339
V T11, T4B, T4C, T12, T14, T15;
340
T11 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
345
T1j = VADD(T1h, T1i);
346
T4B = VSUB(T1h, T1i);
347
T1m = VADD(T1k, T1l);
348
T4C = VSUB(T1k, T1l);
349
T12 = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
350
T14 = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
351
T15 = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
353
V T18, T19, T1b, T1c;
354
T18 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
355
T19 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
356
T1b = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
357
T1c = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
358
T4O = VSUB(T4B, T4C);
359
T4D = VADD(T4B, T4C);
360
T13 = VADD(T11, T12);
361
T4F = VSUB(T11, T12);
362
T16 = VADD(T14, T15);
363
T4G = VSUB(T14, T15);
364
T1a = VADD(T18, T19);
365
T4I = VSUB(T18, T19);
366
T4J = VSUB(T1b, T1c);
367
T1d = VADD(T1b, T1c);
373
V T30, T10, T6k, T4E, T4Q, T4H, T17, T6n, T4P, T1e, T4K, T4R, T1n, T31;
376
T6k = VFNMS(LDK(KP707106781), T4D, T4A);
377
T4E = VFMA(LDK(KP707106781), T4D, T4A);
378
T4Q = VFMA(LDK(KP414213562), T4F, T4G);
379
T4H = VFNMS(LDK(KP414213562), T4G, T4F);
380
T33 = VADD(T13, T16);
381
T17 = VSUB(T13, T16);
382
T6n = VFNMS(LDK(KP707106781), T4O, T4N);
383
T4P = VFMA(LDK(KP707106781), T4O, T4N);
384
T34 = VADD(T1a, T1d);
385
T1e = VSUB(T1a, T1d);
386
T4K = VFMA(LDK(KP414213562), T4J, T4I);
387
T4R = VFNMS(LDK(KP414213562), T4I, T4J);
388
T1n = VSUB(T1j, T1m);
389
T31 = VADD(T1j, T1m);
391
V T1f, T1o, T6o, T4L, T4S, T6l;
392
T1f = VADD(T17, T1e);
393
T1o = VSUB(T17, T1e);
394
T6o = VSUB(T4H, T4K);
395
T4L = VADD(T4H, T4K);
396
T4S = VADD(T4Q, T4R);
397
T6l = VSUB(T4Q, T4R);
398
T3E = VSUB(T30, T31);
399
T32 = VADD(T30, T31);
400
T1p = VFMA(LDK(KP707106781), T1o, T1n);
401
T2v = VFNMS(LDK(KP707106781), T1o, T1n);
402
T1g = VFMA(LDK(KP707106781), T1f, T10);
403
T2u = VFNMS(LDK(KP707106781), T1f, T10);
404
T4M = VFMA(LDK(KP923879532), T4L, T4E);
405
T5K = VFNMS(LDK(KP923879532), T4L, T4E);
406
T6p = VFMA(LDK(KP923879532), T6o, T6n);
407
T6Z = VFNMS(LDK(KP923879532), T6o, T6n);
408
T6m = VFNMS(LDK(KP923879532), T6l, T6k);
409
T6Y = VFMA(LDK(KP923879532), T6l, T6k);
410
T5L = VFNMS(LDK(KP923879532), T4S, T4P);
411
T4T = VFMA(LDK(KP923879532), T4S, T4P);
419
V T6b, T6F, T7n, T7o, T7p, T7q, T7v, T7w, T7x, T7y, T7z, T7A, T7B, T7C, T7f;
420
V T6X, T70, T79, T7a, T73, T6C, T76, T77, T6i;
422
V T2Z, T3r, T3s, T3m, T3d, T3v;
423
T2Z = VSUB(T2V, T2Y);
424
T3r = VADD(T2V, T2Y);
425
T3s = VADD(T3l, T3i);
426
T3m = VSUB(T3i, T3l);
427
T3d = VSUB(T39, T3c);
428
T3v = VADD(T39, T3c);
430
V T3x, T3t, T3P, T3J, T3D, T3V, T3Q, T3G, T36, T3u, T3Y, T3O, T6V, T6W;
432
V T3N, T3C, T3F, T35;
433
T3N = VSUB(T3B, T3A);
434
T3C = VADD(T3A, T3B);
435
T3F = VSUB(T33, T34);
436
T35 = VADD(T33, T34);
437
T3x = VSUB(T3r, T3s);
438
T3t = VADD(T3r, T3s);
439
T3P = VFMA(LDK(KP414213562), T3H, T3I);
440
T3J = VFNMS(LDK(KP414213562), T3I, T3H);
441
T3D = VFMA(LDK(KP707106781), T3C, T3z);
442
T3V = VFNMS(LDK(KP707106781), T3C, T3z);
443
T3Q = VFMA(LDK(KP414213562), T3E, T3F);
444
T3G = VFNMS(LDK(KP414213562), T3F, T3E);
445
T36 = VSUB(T32, T35);
446
T3u = VADD(T32, T35);
447
T3Y = VFNMS(LDK(KP707106781), T3N, T3M);
448
T3O = VFMA(LDK(KP707106781), T3N, T3M);
450
T6b = VFNMS(LDK(KP923879532), T6a, T69);
451
T6V = VFMA(LDK(KP923879532), T6a, T69);
452
T6W = VADD(T6E, T6D);
453
T6F = VSUB(T6D, T6E);
455
V T3K, T3Z, T3e, T3n;
456
T3K = VADD(T3G, T3J);
457
T3Z = VSUB(T3J, T3G);
458
T3e = VADD(T36, T3d);
459
T3n = VSUB(T3d, T36);
461
V T3w, T3y, T3R, T3W;
462
T3w = VADD(T3u, T3v);
463
T3y = VSUB(T3v, T3u);
464
T3R = VSUB(T3P, T3Q);
465
T3W = VADD(T3Q, T3P);
467
V T42, T40, T3L, T3T;
468
T42 = VFNMS(LDK(KP923879532), T3Z, T3Y);
469
T40 = VFMA(LDK(KP923879532), T3Z, T3Y);
470
T3L = VFNMS(LDK(KP923879532), T3K, T3D);
471
T3T = VFMA(LDK(KP923879532), T3K, T3D);
473
V T3o, T3q, T3f, T3p;
474
T3o = VFNMS(LDK(KP707106781), T3n, T3m);
475
T3q = VFMA(LDK(KP707106781), T3n, T3m);
476
T3f = VFNMS(LDK(KP707106781), T3e, T2Z);
477
T3p = VFMA(LDK(KP707106781), T3e, T2Z);
478
T7n = VFNMSI(T3y, T3x);
479
STM2(&(xo[96]), T7n, ovs, &(xo[0]));
480
T7o = VFMAI(T3y, T3x);
481
STM2(&(xo[32]), T7o, ovs, &(xo[0]));
482
T7p = VADD(T3t, T3w);
483
STM2(&(xo[0]), T7p, ovs, &(xo[0]));
484
T7q = VSUB(T3t, T3w);
485
STM2(&(xo[64]), T7q, ovs, &(xo[0]));
487
V T41, T3X, T3S, T3U;
488
T41 = VFMA(LDK(KP923879532), T3W, T3V);
489
T3X = VFNMS(LDK(KP923879532), T3W, T3V);
490
T3S = VFNMS(LDK(KP923879532), T3R, T3O);
491
T3U = VFMA(LDK(KP923879532), T3R, T3O);
492
T7r = VFMAI(T3q, T3p);
493
STM2(&(xo[16]), T7r, ovs, &(xo[0]));
494
T7s = VFNMSI(T3q, T3p);
495
STM2(&(xo[112]), T7s, ovs, &(xo[0]));
496
T7t = VFMAI(T3o, T3f);
497
STM2(&(xo[80]), T7t, ovs, &(xo[0]));
498
T7u = VFNMSI(T3o, T3f);
499
STM2(&(xo[48]), T7u, ovs, &(xo[0]));
500
T7v = VFNMSI(T40, T3X);
501
STM2(&(xo[88]), T7v, ovs, &(xo[0]));
502
T7w = VFMAI(T40, T3X);
503
STM2(&(xo[40]), T7w, ovs, &(xo[0]));
504
T7x = VFMAI(T42, T41);
505
STM2(&(xo[104]), T7x, ovs, &(xo[0]));
506
T7y = VFNMSI(T42, T41);
507
STM2(&(xo[24]), T7y, ovs, &(xo[0]));
508
T7z = VFMAI(T3U, T3T);
509
STM2(&(xo[8]), T7z, ovs, &(xo[0]));
510
T7A = VFNMSI(T3U, T3T);
511
STM2(&(xo[120]), T7A, ovs, &(xo[0]));
512
T7B = VFMAI(T3S, T3L);
513
STM2(&(xo[72]), T7B, ovs, &(xo[0]));
514
T7C = VFNMSI(T3S, T3L);
515
STM2(&(xo[56]), T7C, ovs, &(xo[0]));
516
T7f = VFNMS(LDK(KP831469612), T6W, T6V);
517
T6X = VFMA(LDK(KP831469612), T6W, T6V);
523
T70 = VFMA(LDK(KP303346683), T6Z, T6Y);
524
T79 = VFNMS(LDK(KP303346683), T6Y, T6Z);
525
T7a = VFNMS(LDK(KP303346683), T71, T72);
526
T73 = VFMA(LDK(KP303346683), T72, T71);
527
T6C = VFNMS(LDK(KP923879532), T6B, T6A);
528
T76 = VFMA(LDK(KP923879532), T6B, T6A);
529
T77 = VSUB(T6e, T6h);
530
T6i = VADD(T6e, T6h);
534
V T2r, T2D, T2C, T2s, T5H, T5o, T5v, T5D, T7L, T7O, T7Q, T7S, T5r, T5I, T5x;
537
V TT, T2f, T7E, T7F, T7H, T7J, T2n, T1Y, T28, T2b, T2l, T2p, T2j, T2k;
539
V T1X, T2d, T7h, T7l, T2e, T1q, T75, T7d, T7m, T7k, T7c, T7e, Tn, TS;
540
T2r = VFNMS(LDK(KP707106781), Tm, T7);
541
Tn = VFMA(LDK(KP707106781), Tm, T7);
545
V T7b, T7j, T74, T7i, T78, T7g;
546
T1X = VFNMS(LDK(KP198912367), T1W, T1N);
547
T2d = VFMA(LDK(KP198912367), T1N, T1W);
548
T7g = VADD(T79, T7a);
549
T7b = VSUB(T79, T7a);
550
T7j = VSUB(T73, T70);
551
T74 = VADD(T70, T73);
552
T7i = VFNMS(LDK(KP831469612), T77, T76);
553
T78 = VFMA(LDK(KP831469612), T77, T76);
554
T2j = VFNMS(LDK(KP923879532), TS, Tn);
555
TT = VFMA(LDK(KP923879532), TS, Tn);
556
T7h = VFMA(LDK(KP956940335), T7g, T7f);
557
T7l = VFNMS(LDK(KP956940335), T7g, T7f);
558
T2e = VFMA(LDK(KP198912367), T1g, T1p);
559
T1q = VFNMS(LDK(KP198912367), T1p, T1g);
560
T75 = VFNMS(LDK(KP956940335), T74, T6X);
561
T7d = VFMA(LDK(KP956940335), T74, T6X);
562
T7m = VFNMS(LDK(KP956940335), T7j, T7i);
563
T7k = VFMA(LDK(KP956940335), T7j, T7i);
564
T7c = VFNMS(LDK(KP956940335), T7b, T78);
565
T7e = VFMA(LDK(KP956940335), T7b, T78);
567
T2k = VADD(T2e, T2d);
568
T2f = VSUB(T2d, T2e);
570
V T7D, T7G, T7I, T7K;
571
T7D = VFNMSI(T7k, T7h);
572
STM2(&(xo[90]), T7D, ovs, &(xo[2]));
573
STN2(&(xo[88]), T7v, T7D, ovs);
574
T7E = VFMAI(T7k, T7h);
575
STM2(&(xo[38]), T7E, ovs, &(xo[2]));
576
T7F = VFMAI(T7m, T7l);
577
STM2(&(xo[102]), T7F, ovs, &(xo[2]));
578
T7G = VFNMSI(T7m, T7l);
579
STM2(&(xo[26]), T7G, ovs, &(xo[2]));
580
STN2(&(xo[24]), T7y, T7G, ovs);
581
T7H = VFMAI(T7e, T7d);
582
STM2(&(xo[6]), T7H, ovs, &(xo[2]));
583
T7I = VFNMSI(T7e, T7d);
584
STM2(&(xo[122]), T7I, ovs, &(xo[2]));
585
STN2(&(xo[120]), T7A, T7I, ovs);
586
T7J = VFMAI(T7c, T75);
587
STM2(&(xo[70]), T7J, ovs, &(xo[2]));
588
T7K = VFNMSI(T7c, T75);
589
STM2(&(xo[58]), T7K, ovs, &(xo[2]));
590
STN2(&(xo[56]), T7C, T7K, ovs);
591
T2n = VSUB(T1X, T1q);
592
T1Y = VADD(T1q, T1X);
594
T2C = VFNMS(LDK(KP707106781), T27, T26);
595
T28 = VFMA(LDK(KP707106781), T27, T26);
596
T2b = VSUB(T29, T2a);
597
T2s = VADD(T2a, T29);
599
T2l = VFNMS(LDK(KP980785280), T2k, T2j);
600
T2p = VFMA(LDK(KP980785280), T2k, T2j);
602
V T5z, T4z, T5A, T5g;
604
V T4f, T4y, T1Z, T2h, T4U, T5t, T2m, T2c, T5u, T5f;
605
T5H = VFNMS(LDK(KP923879532), T4e, T47);
606
T4f = VFMA(LDK(KP923879532), T4e, T47);
607
T4y = VADD(T4o, T4x);
608
T5T = VSUB(T4x, T4o);
609
T1Z = VFNMS(LDK(KP980785280), T1Y, TT);
610
T2h = VFMA(LDK(KP980785280), T1Y, TT);
611
T4U = VFNMS(LDK(KP098491403), T4T, T4M);
612
T5t = VFMA(LDK(KP098491403), T4M, T4T);
613
T2m = VFNMS(LDK(KP923879532), T2b, T28);
614
T2c = VFMA(LDK(KP923879532), T2b, T28);
615
T5u = VFMA(LDK(KP098491403), T57, T5e);
616
T5f = VFNMS(LDK(KP098491403), T5e, T57);
617
T5z = VFNMS(LDK(KP980785280), T4y, T4f);
618
T4z = VFMA(LDK(KP980785280), T4y, T4f);
619
T5S = VFNMS(LDK(KP923879532), T5n, T5k);
620
T5o = VFMA(LDK(KP923879532), T5n, T5k);
622
V T2o, T2q, T2i, T2g;
623
T2o = VFMA(LDK(KP980785280), T2n, T2m);
624
T2q = VFNMS(LDK(KP980785280), T2n, T2m);
625
T2i = VFMA(LDK(KP980785280), T2f, T2c);
626
T2g = VFNMS(LDK(KP980785280), T2f, T2c);
627
T5A = VADD(T5t, T5u);
628
T5v = VSUB(T5t, T5u);
629
T5D = VSUB(T5f, T4U);
630
T5g = VADD(T4U, T5f);
631
T7L = VFNMSI(T2o, T2l);
632
STM2(&(xo[92]), T7L, ovs, &(xo[0]));
634
V T7M, T7N, T7P, T7R;
635
T7M = VFMAI(T2o, T2l);
636
STM2(&(xo[36]), T7M, ovs, &(xo[0]));
637
STN2(&(xo[36]), T7M, T7E, ovs);
638
T7N = VFMAI(T2q, T2p);
639
STM2(&(xo[100]), T7N, ovs, &(xo[0]));
640
STN2(&(xo[100]), T7N, T7F, ovs);
641
T7O = VFNMSI(T2q, T2p);
642
STM2(&(xo[28]), T7O, ovs, &(xo[0]));
643
T7P = VFMAI(T2i, T2h);
644
STM2(&(xo[4]), T7P, ovs, &(xo[0]));
645
STN2(&(xo[4]), T7P, T7H, ovs);
646
T7Q = VFNMSI(T2i, T2h);
647
STM2(&(xo[124]), T7Q, ovs, &(xo[0]));
648
T7R = VFMAI(T2g, T1Z);
649
STM2(&(xo[68]), T7R, ovs, &(xo[0]));
650
STN2(&(xo[68]), T7R, T7J, ovs);
651
T7S = VFNMSI(T2g, T1Z);
652
STM2(&(xo[60]), T7S, ovs, &(xo[0]));
653
T5r = VSUB(T5p, T5q);
654
T5I = VADD(T5p, T5q);
658
T5x = VFMA(LDK(KP995184726), T5g, T4z);
659
T5h = VFNMS(LDK(KP995184726), T5g, T4z);
660
T5F = VFMA(LDK(KP995184726), T5A, T5z);
661
T5B = VFNMS(LDK(KP995184726), T5A, T5z);
665
V T6J, T6R, T6L, T6z, T6T, T6P;
667
V T6N, T6j, T6O, T6y;
669
V T6q, T6H, T5C, T5s, T6I, T6x;
670
T6q = VFNMS(LDK(KP534511135), T6p, T6m);
671
T6H = VFMA(LDK(KP534511135), T6m, T6p);
672
T5C = VFNMS(LDK(KP980785280), T5r, T5o);
673
T5s = VFMA(LDK(KP980785280), T5r, T5o);
674
T6I = VFMA(LDK(KP534511135), T6t, T6w);
675
T6x = VFNMS(LDK(KP534511135), T6w, T6t);
676
T6N = VFMA(LDK(KP831469612), T6i, T6b);
677
T6j = VFNMS(LDK(KP831469612), T6i, T6b);
679
V T5E, T5G, T5y, T5w;
680
T5E = VFNMS(LDK(KP995184726), T5D, T5C);
681
T5G = VFMA(LDK(KP995184726), T5D, T5C);
682
T5y = VFMA(LDK(KP995184726), T5v, T5s);
683
T5w = VFNMS(LDK(KP995184726), T5v, T5s);
684
T6O = VADD(T6H, T6I);
685
T6J = VSUB(T6H, T6I);
686
T6R = VSUB(T6x, T6q);
687
T6y = VADD(T6q, T6x);
689
V T7T, T7U, T7V, T7W;
690
T7T = VFMAI(T5E, T5B);
691
STM2(&(xo[94]), T7T, ovs, &(xo[2]));
692
STN2(&(xo[92]), T7L, T7T, ovs);
693
T7U = VFNMSI(T5E, T5B);
694
STM2(&(xo[34]), T7U, ovs, &(xo[2]));
695
STN2(&(xo[32]), T7o, T7U, ovs);
696
T7V = VFNMSI(T5G, T5F);
697
STM2(&(xo[98]), T7V, ovs, &(xo[2]));
698
STN2(&(xo[96]), T7n, T7V, ovs);
699
T7W = VFMAI(T5G, T5F);
700
STM2(&(xo[30]), T7W, ovs, &(xo[2]));
701
STN2(&(xo[28]), T7O, T7W, ovs);
703
V T7X, T7Y, T7Z, T80;
704
T7X = VFMAI(T5y, T5x);
705
STM2(&(xo[126]), T7X, ovs, &(xo[2]));
706
STN2(&(xo[124]), T7Q, T7X, ovs);
707
T7Y = VFNMSI(T5y, T5x);
708
STM2(&(xo[2]), T7Y, ovs, &(xo[2]));
709
STN2(&(xo[0]), T7p, T7Y, ovs);
710
T7Z = VFMAI(T5w, T5h);
711
STM2(&(xo[62]), T7Z, ovs, &(xo[2]));
712
STN2(&(xo[60]), T7S, T7Z, ovs);
713
T80 = VFNMSI(T5w, T5h);
714
STM2(&(xo[66]), T80, ovs, &(xo[2]));
715
STN2(&(xo[64]), T7q, T80, ovs);
720
T6L = VFMA(LDK(KP881921264), T6y, T6j);
721
T6z = VFNMS(LDK(KP881921264), T6y, T6j);
722
T6T = VFMA(LDK(KP881921264), T6O, T6N);
723
T6P = VFNMS(LDK(KP881921264), T6O, T6N);
726
V T2H, T2P, T81, T84, T85, T87, T2J, T2B, T2R, T2N;
728
V T2L, T2t, T2M, T2A;
730
V T2z, T2F, T6Q, T6G, T2G, T2w;
731
T2z = VFMA(LDK(KP668178637), T2y, T2x);
732
T2F = VFNMS(LDK(KP668178637), T2x, T2y);
733
T6Q = VFMA(LDK(KP831469612), T6F, T6C);
734
T6G = VFNMS(LDK(KP831469612), T6F, T6C);
735
T2G = VFNMS(LDK(KP668178637), T2u, T2v);
736
T2w = VFMA(LDK(KP668178637), T2v, T2u);
737
T2L = VFNMS(LDK(KP923879532), T2s, T2r);
738
T2t = VFMA(LDK(KP923879532), T2s, T2r);
740
V T6S, T6U, T6M, T6K;
741
T6S = VFNMS(LDK(KP881921264), T6R, T6Q);
742
T6U = VFMA(LDK(KP881921264), T6R, T6Q);
743
T6M = VFMA(LDK(KP881921264), T6J, T6G);
744
T6K = VFNMS(LDK(KP881921264), T6J, T6G);
745
T2M = VADD(T2G, T2F);
746
T2H = VSUB(T2F, T2G);
747
T2P = VSUB(T2z, T2w);
748
T2A = VADD(T2w, T2z);
749
T81 = VFMAI(T6S, T6P);
750
STM2(&(xo[86]), T81, ovs, &(xo[2]));
752
V T82, T83, T86, T88;
753
T82 = VFNMSI(T6S, T6P);
754
STM2(&(xo[42]), T82, ovs, &(xo[2]));
755
STN2(&(xo[40]), T7w, T82, ovs);
756
T83 = VFNMSI(T6U, T6T);
757
STM2(&(xo[106]), T83, ovs, &(xo[2]));
758
STN2(&(xo[104]), T7x, T83, ovs);
759
T84 = VFMAI(T6U, T6T);
760
STM2(&(xo[22]), T84, ovs, &(xo[2]));
761
T85 = VFMAI(T6M, T6L);
762
STM2(&(xo[118]), T85, ovs, &(xo[2]));
763
T86 = VFNMSI(T6M, T6L);
764
STM2(&(xo[10]), T86, ovs, &(xo[2]));
765
STN2(&(xo[8]), T7z, T86, ovs);
766
T87 = VFMAI(T6K, T6z);
767
STM2(&(xo[54]), T87, ovs, &(xo[2]));
768
T88 = VFNMSI(T6K, T6z);
769
STM2(&(xo[74]), T88, ovs, &(xo[2]));
770
STN2(&(xo[72]), T7B, T88, ovs);
774
T2J = VFMA(LDK(KP831469612), T2A, T2t);
775
T2B = VFNMS(LDK(KP831469612), T2A, T2t);
776
T2R = VFNMS(LDK(KP831469612), T2M, T2L);
777
T2N = VFMA(LDK(KP831469612), T2M, T2L);
780
V T61, T5J, T62, T5Q;
782
V T5M, T5V, T2O, T2E, T5W, T5P;
783
T5M = VFMA(LDK(KP820678790), T5L, T5K);
784
T5V = VFNMS(LDK(KP820678790), T5K, T5L);
785
T2O = VFMA(LDK(KP923879532), T2D, T2C);
786
T2E = VFNMS(LDK(KP923879532), T2D, T2C);
787
T5W = VFNMS(LDK(KP820678790), T5N, T5O);
788
T5P = VFMA(LDK(KP820678790), T5O, T5N);
789
T61 = VFNMS(LDK(KP980785280), T5I, T5H);
790
T5J = VFMA(LDK(KP980785280), T5I, T5H);
792
V T2Q, T2S, T2K, T2I;
793
T2Q = VFNMS(LDK(KP831469612), T2P, T2O);
794
T2S = VFMA(LDK(KP831469612), T2P, T2O);
795
T2K = VFMA(LDK(KP831469612), T2H, T2E);
796
T2I = VFNMS(LDK(KP831469612), T2H, T2E);
797
T62 = VADD(T5V, T5W);
798
T5X = VSUB(T5V, T5W);
799
T65 = VSUB(T5P, T5M);
800
T5Q = VADD(T5M, T5P);
802
V T89, T8c, T8d, T8f;
803
T89 = VFMAI(T2Q, T2N);
804
STM2(&(xo[84]), T89, ovs, &(xo[0]));
805
STN2(&(xo[84]), T89, T81, ovs);
806
T8a = VFNMSI(T2Q, T2N);
807
STM2(&(xo[44]), T8a, ovs, &(xo[0]));
808
T8b = VFNMSI(T2S, T2R);
809
STM2(&(xo[108]), T8b, ovs, &(xo[0]));
810
T8c = VFMAI(T2S, T2R);
811
STM2(&(xo[20]), T8c, ovs, &(xo[0]));
812
STN2(&(xo[20]), T8c, T84, ovs);
813
T8d = VFMAI(T2K, T2J);
814
STM2(&(xo[116]), T8d, ovs, &(xo[0]));
815
STN2(&(xo[116]), T8d, T85, ovs);
816
T8e = VFNMSI(T2K, T2J);
817
STM2(&(xo[12]), T8e, ovs, &(xo[0]));
818
T8f = VFMAI(T2I, T2B);
819
STM2(&(xo[52]), T8f, ovs, &(xo[0]));
820
STN2(&(xo[52]), T8f, T87, ovs);
821
T8g = VFNMSI(T2I, T2B);
822
STM2(&(xo[76]), T8g, ovs, &(xo[0]));
826
T5Z = VFMA(LDK(KP773010453), T5Q, T5J);
827
T5R = VFNMS(LDK(KP773010453), T5Q, T5J);
828
T67 = VFNMS(LDK(KP773010453), T62, T61);
829
T63 = VFMA(LDK(KP773010453), T62, T61);
836
T5U = VFMA(LDK(KP980785280), T5T, T5S);
837
T64 = VFNMS(LDK(KP980785280), T5T, T5S);
839
V T68, T66, T5Y, T60;
840
T68 = VFNMS(LDK(KP773010453), T65, T64);
841
T66 = VFMA(LDK(KP773010453), T65, T64);
842
T5Y = VFNMS(LDK(KP773010453), T5X, T5U);
843
T60 = VFMA(LDK(KP773010453), T5X, T5U);
845
V T8h, T8i, T8j, T8k;
846
T8h = VFNMSI(T66, T63);
847
STM2(&(xo[82]), T8h, ovs, &(xo[2]));
848
STN2(&(xo[80]), T7t, T8h, ovs);
849
T8i = VFMAI(T66, T63);
850
STM2(&(xo[46]), T8i, ovs, &(xo[2]));
851
STN2(&(xo[44]), T8a, T8i, ovs);
852
T8j = VFMAI(T68, T67);
853
STM2(&(xo[110]), T8j, ovs, &(xo[2]));
854
STN2(&(xo[108]), T8b, T8j, ovs);
855
T8k = VFNMSI(T68, T67);
856
STM2(&(xo[18]), T8k, ovs, &(xo[2]));
857
STN2(&(xo[16]), T7r, T8k, ovs);
859
V T8l, T8m, T8n, T8o;
860
T8l = VFMAI(T60, T5Z);
861
STM2(&(xo[14]), T8l, ovs, &(xo[2]));
862
STN2(&(xo[12]), T8e, T8l, ovs);
863
T8m = VFNMSI(T60, T5Z);
864
STM2(&(xo[114]), T8m, ovs, &(xo[2]));
865
STN2(&(xo[112]), T7s, T8m, ovs);
866
T8n = VFMAI(T5Y, T5R);
867
STM2(&(xo[78]), T8n, ovs, &(xo[2]));
868
STN2(&(xo[76]), T8g, T8n, ovs);
869
T8o = VFNMSI(T5Y, T5R);
870
STM2(&(xo[50]), T8o, ovs, &(xo[2]));
871
STN2(&(xo[48]), T7u, T8o, ovs);
878
static const kdft_desc desc = { 64, "n2fv_64", {198, 0, 258, 0}, &GENUS, 0, 2, 0, 0 };
880
void X(codelet_n2fv_64) (planner *p) {
881
X(kdft_register) (p, n2fv_64, &desc);
886
/* Generated by: ../../../genfft/gen_notw_c -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n2fv_64 -with-ostride 2 -include n2f.h -store-multiple 2 */
889
* This function contains 456 FP additions, 124 FP multiplications,
890
* (or, 404 additions, 72 multiplications, 52 fused multiply/add),
891
* 128 stack variables, 15 constants, and 160 memory accesses
895
static void n2fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
897
DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
898
DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
899
DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
900
DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
901
DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
902
DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
903
DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
904
DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
905
DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
906
DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
907
DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
908
DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
909
DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
910
DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
911
DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
917
for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
918
V T4p, T5q, Tb, T39, T2n, T3A, T6f, T6T, Tq, T3B, T6i, T76, T2i, T3a, T4w;
919
V T5r, TI, T2p, T6C, T6V, T3h, T3E, T4L, T5u, TZ, T2q, T6F, T6U, T3e, T3D;
920
V T4E, T5t, T23, T2N, T6t, T71, T6w, T72, T2c, T2O, T3t, T41, T5f, T5R, T5k;
921
V T5S, T3w, T42, T1s, T2K, T6m, T6Y, T6p, T6Z, T1B, T2L, T3m, T3Y, T4Y, T5O;
922
V T53, T5P, T3p, T3Z;
924
V T3, T4n, T2m, T4o, T6, T5p, T9, T5o;
927
T1 = LD(&(xi[0]), ivs, &(xi[0]));
928
T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
931
T2k = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
932
T2l = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
933
T2m = VSUB(T2k, T2l);
934
T4o = VADD(T2k, T2l);
938
T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
939
T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
942
T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
943
T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
947
T4p = VSUB(T4n, T4o);
948
T5q = VSUB(T5o, T5p);
951
Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
954
T2j = VMUL(LDK(KP707106781), VSUB(T9, T6));
955
T2n = VSUB(T2j, T2m);
956
T3A = VADD(T2m, T2j);
957
T6d = VADD(T4n, T4o);
958
T6e = VADD(T5p, T5o);
959
T6f = VADD(T6d, T6e);
960
T6T = VSUB(T6d, T6e);
964
V Te, T4q, To, T4u, Th, T4r, Tl, T4t;
967
Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
968
Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
971
Tm = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
972
Tn = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
978
Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
979
Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
982
Tj = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
983
Tk = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
989
Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
990
Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
993
T6g = VADD(T4q, T4r);
994
T6h = VADD(T4t, T4u);
995
T6i = VADD(T6g, T6h);
996
T76 = VSUB(T6h, T6g);
999
V T2g, T2h, T4s, T4v;
1000
T2g = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
1001
T2h = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
1002
T2i = VSUB(T2g, T2h);
1003
T3a = VADD(T2h, T2g);
1004
T4s = VSUB(T4q, T4r);
1005
T4v = VSUB(T4t, T4u);
1006
T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v));
1007
T5r = VMUL(LDK(KP707106781), VSUB(T4v, T4s));
1011
V Tu, T4F, TG, T4G, TB, T4J, TD, T4I;
1014
Ts = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
1015
Tt = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
1018
TE = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
1019
TF = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
1023
V Tv, Tw, Tx, Ty, Tz, TA;
1024
Tv = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
1025
Tw = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
1027
Ty = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
1028
Tz = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
1030
TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
1032
TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
1040
TI = VFMA(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
1041
T2p = VFNMS(LDK(KP195090322), TH, VMUL(LDK(KP980785280), TC));
1042
T6A = VADD(T4F, T4G);
1043
T6B = VADD(T4J, T4I);
1044
T6C = VADD(T6A, T6B);
1045
T6V = VSUB(T6A, T6B);
1048
V T3f, T3g, T4H, T4K;
1051
T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f));
1052
T3E = VFMA(LDK(KP555570233), T3f, VMUL(LDK(KP831469612), T3g));
1053
T4H = VSUB(T4F, T4G);
1054
T4K = VSUB(T4I, T4J);
1055
T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H));
1056
T5u = VFMA(LDK(KP382683432), T4H, VMUL(LDK(KP923879532), T4K));
1060
V TS, T4z, TW, T4y, TP, T4C, TX, T4B;
1063
TQ = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
1064
TR = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
1067
TU = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
1068
TV = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
1072
V TJ, TK, TL, TM, TN, TO;
1073
TJ = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
1074
TK = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
1076
TM = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
1077
TN = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
1079
TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
1081
TX = VMUL(LDK(KP707106781), VADD(TO, TL));
1089
TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
1090
T2q = VFMA(LDK(KP980785280), TY, VMUL(LDK(KP195090322), TT));
1091
T6D = VADD(T4y, T4z);
1092
T6E = VADD(T4C, T4B);
1093
T6F = VADD(T6D, T6E);
1094
T6U = VSUB(T6D, T6E);
1097
V T3c, T3d, T4A, T4D;
1100
T3e = VFMA(LDK(KP831469612), T3c, VMUL(LDK(KP555570233), T3d));
1101
T3D = VFNMS(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d));
1102
T4A = VSUB(T4y, T4z);
1103
T4D = VSUB(T4B, T4C);
1104
T4E = VFMA(LDK(KP923879532), T4A, VMUL(LDK(KP382683432), T4D));
1105
T5t = VFNMS(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D));
1109
V T1F, T55, T2a, T56, T1M, T5h, T27, T5g, T58, T59, T1U, T5a, T25, T5b, T5c;
1112
V T1D, T1E, T28, T29;
1113
T1D = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
1114
T1E = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
1115
T1F = VSUB(T1D, T1E);
1116
T55 = VADD(T1D, T1E);
1117
T28 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
1118
T29 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
1119
T2a = VSUB(T28, T29);
1120
T56 = VADD(T28, T29);
1123
V T1G, T1H, T1I, T1J, T1K, T1L;
1124
T1G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
1125
T1H = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
1126
T1I = VSUB(T1G, T1H);
1127
T1J = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
1128
T1K = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
1129
T1L = VSUB(T1J, T1K);
1130
T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L));
1131
T5h = VADD(T1G, T1H);
1132
T27 = VMUL(LDK(KP707106781), VSUB(T1L, T1I));
1133
T5g = VADD(T1J, T1K);
1136
V T1Q, T1T, T1X, T20;
1138
V T1O, T1P, T1R, T1S;
1139
T1O = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
1140
T1P = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
1141
T1Q = VSUB(T1O, T1P);
1142
T58 = VADD(T1O, T1P);
1143
T1R = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
1144
T1S = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
1145
T1T = VSUB(T1R, T1S);
1146
T59 = VADD(T1R, T1S);
1148
T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q));
1149
T5a = VSUB(T58, T59);
1150
T25 = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T));
1152
V T1V, T1W, T1Y, T1Z;
1153
T1V = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
1154
T1W = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
1155
T1X = VSUB(T1V, T1W);
1156
T5b = VADD(T1V, T1W);
1157
T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
1158
T1Z = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
1159
T20 = VSUB(T1Y, T1Z);
1160
T5c = VADD(T1Y, T1Z);
1162
T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20));
1163
T5d = VSUB(T5b, T5c);
1164
T24 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X));
1167
V T1N, T22, T6r, T6s;
1168
T1N = VADD(T1F, T1M);
1169
T22 = VADD(T1U, T21);
1170
T23 = VSUB(T1N, T22);
1171
T2N = VADD(T1N, T22);
1172
T6r = VADD(T55, T56);
1173
T6s = VADD(T5h, T5g);
1174
T6t = VADD(T6r, T6s);
1175
T71 = VSUB(T6r, T6s);
1178
V T6u, T6v, T26, T2b;
1179
T6u = VADD(T58, T59);
1180
T6v = VADD(T5b, T5c);
1181
T6w = VADD(T6u, T6v);
1182
T72 = VSUB(T6v, T6u);
1183
T26 = VSUB(T24, T25);
1184
T2b = VSUB(T27, T2a);
1185
T2c = VSUB(T26, T2b);
1186
T2O = VADD(T2b, T26);
1189
V T3r, T3s, T57, T5e;
1190
T3r = VSUB(T1F, T1M);
1191
T3s = VADD(T25, T24);
1192
T3t = VADD(T3r, T3s);
1193
T41 = VSUB(T3r, T3s);
1194
T57 = VSUB(T55, T56);
1195
T5e = VMUL(LDK(KP707106781), VADD(T5a, T5d));
1196
T5f = VADD(T57, T5e);
1197
T5R = VSUB(T57, T5e);
1200
V T5i, T5j, T3u, T3v;
1201
T5i = VSUB(T5g, T5h);
1202
T5j = VMUL(LDK(KP707106781), VSUB(T5d, T5a));
1203
T5k = VADD(T5i, T5j);
1204
T5S = VSUB(T5j, T5i);
1205
T3u = VADD(T2a, T27);
1206
T3v = VSUB(T21, T1U);
1207
T3w = VADD(T3u, T3v);
1208
T42 = VSUB(T3v, T3u);
1212
V T1q, T4P, T1v, T4O, T1n, T50, T1w, T4Z, T4U, T4V, T18, T4W, T1z, T4R, T4S;
1215
V T1o, T1p, T1t, T1u;
1216
T1o = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
1217
T1p = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
1218
T1q = VSUB(T1o, T1p);
1219
T4P = VADD(T1o, T1p);
1220
T1t = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
1221
T1u = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
1222
T1v = VSUB(T1t, T1u);
1223
T4O = VADD(T1t, T1u);
1226
V T1h, T1i, T1j, T1k, T1l, T1m;
1227
T1h = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
1228
T1i = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
1229
T1j = VSUB(T1h, T1i);
1230
T1k = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
1231
T1l = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
1232
T1m = VSUB(T1k, T1l);
1233
T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m));
1234
T50 = VADD(T1k, T1l);
1235
T1w = VMUL(LDK(KP707106781), VADD(T1m, T1j));
1236
T4Z = VADD(T1h, T1i);
1239
V T14, T17, T1b, T1e;
1241
V T12, T13, T15, T16;
1242
T12 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
1243
T13 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
1244
T14 = VSUB(T12, T13);
1245
T4U = VADD(T12, T13);
1246
T15 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
1247
T16 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
1248
T17 = VSUB(T15, T16);
1249
T4V = VADD(T15, T16);
1251
T18 = VFNMS(LDK(KP923879532), T17, VMUL(LDK(KP382683432), T14));
1252
T4W = VSUB(T4U, T4V);
1253
T1z = VFMA(LDK(KP923879532), T14, VMUL(LDK(KP382683432), T17));
1255
V T19, T1a, T1c, T1d;
1256
T19 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
1257
T1a = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
1258
T1b = VSUB(T19, T1a);
1259
T4R = VADD(T19, T1a);
1260
T1c = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
1261
T1d = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
1262
T1e = VSUB(T1c, T1d);
1263
T4S = VADD(T1c, T1d);
1265
T1f = VFMA(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e));
1266
T4T = VSUB(T4R, T4S);
1267
T1y = VFNMS(LDK(KP382683432), T1e, VMUL(LDK(KP923879532), T1b));
1270
V T1g, T1r, T6k, T6l;
1271
T1g = VSUB(T18, T1f);
1272
T1r = VSUB(T1n, T1q);
1273
T1s = VSUB(T1g, T1r);
1274
T2K = VADD(T1r, T1g);
1275
T6k = VADD(T4O, T4P);
1276
T6l = VADD(T50, T4Z);
1277
T6m = VADD(T6k, T6l);
1278
T6Y = VSUB(T6k, T6l);
1281
V T6n, T6o, T1x, T1A;
1282
T6n = VADD(T4R, T4S);
1283
T6o = VADD(T4U, T4V);
1284
T6p = VADD(T6n, T6o);
1285
T6Z = VSUB(T6o, T6n);
1286
T1x = VADD(T1v, T1w);
1287
T1A = VADD(T1y, T1z);
1288
T1B = VSUB(T1x, T1A);
1289
T2L = VADD(T1x, T1A);
1292
V T3k, T3l, T4Q, T4X;
1293
T3k = VSUB(T1v, T1w);
1294
T3l = VADD(T1f, T18);
1295
T3m = VADD(T3k, T3l);
1296
T3Y = VSUB(T3k, T3l);
1297
T4Q = VSUB(T4O, T4P);
1298
T4X = VMUL(LDK(KP707106781), VADD(T4T, T4W));
1299
T4Y = VADD(T4Q, T4X);
1300
T5O = VSUB(T4Q, T4X);
1303
V T51, T52, T3n, T3o;
1304
T51 = VSUB(T4Z, T50);
1305
T52 = VMUL(LDK(KP707106781), VSUB(T4W, T4T));
1306
T53 = VADD(T51, T52);
1307
T5P = VSUB(T52, T51);
1308
T3n = VADD(T1q, T1n);
1309
T3o = VSUB(T1z, T1y);
1310
T3p = VADD(T3n, T3o);
1311
T3Z = VSUB(T3o, T3n);
1315
V T7n, T7o, T7p, T7q, T7r, T7s, T7t, T7u, T7v, T7w, T7x, T7y, T7z, T7A, T7B;
1316
V T7C, T7D, T7E, T7F, T7G, T7H, T7I, T7J, T7K;
1318
V T6N, T6R, T6Q, T6S;
1320
V T6L, T6M, T6O, T6P;
1321
T6L = VADD(T6f, T6i);
1322
T6M = VADD(T6F, T6C);
1323
T6N = VADD(T6L, T6M);
1324
T6R = VSUB(T6L, T6M);
1325
T6O = VADD(T6m, T6p);
1326
T6P = VADD(T6t, T6w);
1327
T6Q = VADD(T6O, T6P);
1328
T6S = VBYI(VSUB(T6P, T6O));
1330
T7n = VSUB(T6N, T6Q);
1331
STM2(&(xo[64]), T7n, ovs, &(xo[0]));
1332
T7o = VADD(T6R, T6S);
1333
STM2(&(xo[32]), T7o, ovs, &(xo[0]));
1334
T7p = VADD(T6N, T6Q);
1335
STM2(&(xo[0]), T7p, ovs, &(xo[0]));
1336
T7q = VSUB(T6R, T6S);
1337
STM2(&(xo[96]), T7q, ovs, &(xo[0]));
1340
V T6j, T6G, T6y, T6H, T6q, T6x;
1341
T6j = VSUB(T6f, T6i);
1342
T6G = VSUB(T6C, T6F);
1343
T6q = VSUB(T6m, T6p);
1344
T6x = VSUB(T6t, T6w);
1345
T6y = VMUL(LDK(KP707106781), VADD(T6q, T6x));
1346
T6H = VMUL(LDK(KP707106781), VSUB(T6x, T6q));
1348
V T6z, T6I, T6J, T6K;
1349
T6z = VADD(T6j, T6y);
1350
T6I = VBYI(VADD(T6G, T6H));
1351
T7r = VSUB(T6z, T6I);
1352
STM2(&(xo[112]), T7r, ovs, &(xo[0]));
1353
T7s = VADD(T6z, T6I);
1354
STM2(&(xo[16]), T7s, ovs, &(xo[0]));
1355
T6J = VSUB(T6j, T6y);
1356
T6K = VBYI(VSUB(T6H, T6G));
1357
T7t = VSUB(T6J, T6K);
1358
STM2(&(xo[80]), T7t, ovs, &(xo[0]));
1359
T7u = VADD(T6J, T6K);
1360
STM2(&(xo[48]), T7u, ovs, &(xo[0]));
1364
V T6X, T7i, T78, T7g, T74, T7f, T7b, T7j, T6W, T77;
1365
T6W = VMUL(LDK(KP707106781), VADD(T6U, T6V));
1366
T6X = VADD(T6T, T6W);
1367
T7i = VSUB(T6T, T6W);
1368
T77 = VMUL(LDK(KP707106781), VSUB(T6V, T6U));
1369
T78 = VADD(T76, T77);
1370
T7g = VSUB(T77, T76);
1372
V T70, T73, T79, T7a;
1373
T70 = VFMA(LDK(KP923879532), T6Y, VMUL(LDK(KP382683432), T6Z));
1374
T73 = VFNMS(LDK(KP382683432), T72, VMUL(LDK(KP923879532), T71));
1375
T74 = VADD(T70, T73);
1376
T7f = VSUB(T73, T70);
1377
T79 = VFNMS(LDK(KP382683432), T6Y, VMUL(LDK(KP923879532), T6Z));
1378
T7a = VFMA(LDK(KP382683432), T71, VMUL(LDK(KP923879532), T72));
1379
T7b = VADD(T79, T7a);
1380
T7j = VSUB(T7a, T79);
1383
V T75, T7c, T7l, T7m;
1384
T75 = VADD(T6X, T74);
1385
T7c = VBYI(VADD(T78, T7b));
1386
T7v = VSUB(T75, T7c);
1387
STM2(&(xo[120]), T7v, ovs, &(xo[0]));
1388
T7w = VADD(T75, T7c);
1389
STM2(&(xo[8]), T7w, ovs, &(xo[0]));
1390
T7l = VBYI(VADD(T7g, T7f));
1391
T7m = VADD(T7i, T7j);
1392
T7x = VADD(T7l, T7m);
1393
STM2(&(xo[24]), T7x, ovs, &(xo[0]));
1394
T7y = VSUB(T7m, T7l);
1395
STM2(&(xo[104]), T7y, ovs, &(xo[0]));
1398
V T7d, T7e, T7h, T7k;
1399
T7d = VSUB(T6X, T74);
1400
T7e = VBYI(VSUB(T7b, T78));
1401
T7z = VSUB(T7d, T7e);
1402
STM2(&(xo[72]), T7z, ovs, &(xo[0]));
1403
T7A = VADD(T7d, T7e);
1404
STM2(&(xo[56]), T7A, ovs, &(xo[0]));
1405
T7h = VBYI(VSUB(T7f, T7g));
1406
T7k = VSUB(T7i, T7j);
1407
T7B = VADD(T7h, T7k);
1408
STM2(&(xo[40]), T7B, ovs, &(xo[0]));
1409
T7C = VSUB(T7k, T7h);
1410
STM2(&(xo[88]), T7C, ovs, &(xo[0]));
1414
V T5N, T68, T61, T69, T5U, T65, T5Y, T66;
1416
V T5L, T5M, T5Z, T60;
1417
T5L = VSUB(T4p, T4w);
1418
T5M = VSUB(T5u, T5t);
1419
T5N = VADD(T5L, T5M);
1420
T68 = VSUB(T5L, T5M);
1421
T5Z = VFNMS(LDK(KP555570233), T5O, VMUL(LDK(KP831469612), T5P));
1422
T60 = VFMA(LDK(KP555570233), T5R, VMUL(LDK(KP831469612), T5S));
1423
T61 = VADD(T5Z, T60);
1424
T69 = VSUB(T60, T5Z);
1427
V T5Q, T5T, T5W, T5X;
1428
T5Q = VFMA(LDK(KP831469612), T5O, VMUL(LDK(KP555570233), T5P));
1429
T5T = VFNMS(LDK(KP555570233), T5S, VMUL(LDK(KP831469612), T5R));
1430
T5U = VADD(T5Q, T5T);
1431
T65 = VSUB(T5T, T5Q);
1432
T5W = VSUB(T5r, T5q);
1433
T5X = VSUB(T4L, T4E);
1434
T5Y = VADD(T5W, T5X);
1435
T66 = VSUB(T5X, T5W);
1438
V T5V, T62, T6b, T6c;
1439
T5V = VADD(T5N, T5U);
1440
T62 = VBYI(VADD(T5Y, T61));
1441
T7D = VSUB(T5V, T62);
1442
STM2(&(xo[116]), T7D, ovs, &(xo[0]));
1443
T7E = VADD(T5V, T62);
1444
STM2(&(xo[12]), T7E, ovs, &(xo[0]));
1445
T6b = VBYI(VADD(T66, T65));
1446
T6c = VADD(T68, T69);
1447
T7F = VADD(T6b, T6c);
1448
STM2(&(xo[20]), T7F, ovs, &(xo[0]));
1449
T7G = VSUB(T6c, T6b);
1450
STM2(&(xo[108]), T7G, ovs, &(xo[0]));
1453
V T63, T64, T67, T6a;
1454
T63 = VSUB(T5N, T5U);
1455
T64 = VBYI(VSUB(T61, T5Y));
1456
T7H = VSUB(T63, T64);
1457
STM2(&(xo[76]), T7H, ovs, &(xo[0]));
1458
T7I = VADD(T63, T64);
1459
STM2(&(xo[52]), T7I, ovs, &(xo[0]));
1460
T67 = VBYI(VSUB(T65, T66));
1461
T6a = VSUB(T68, T69);
1462
T7J = VADD(T67, T6a);
1463
STM2(&(xo[44]), T7J, ovs, &(xo[0]));
1464
T7K = VSUB(T6a, T67);
1465
STM2(&(xo[84]), T7K, ovs, &(xo[0]));
1469
V T7U, T7W, T7X, T7Z;
1471
V T11, T2C, T2v, T2D, T2e, T2z, T2s, T2A;
1473
V Tr, T10, T2t, T2u;
1476
T11 = VADD(Tr, T10);
1477
T2C = VSUB(Tr, T10);
1478
T2t = VFNMS(LDK(KP634393284), T1B, VMUL(LDK(KP773010453), T1s));
1479
T2u = VFMA(LDK(KP773010453), T2c, VMUL(LDK(KP634393284), T23));
1480
T2v = VADD(T2t, T2u);
1481
T2D = VSUB(T2u, T2t);
1484
V T1C, T2d, T2o, T2r;
1485
T1C = VFMA(LDK(KP634393284), T1s, VMUL(LDK(KP773010453), T1B));
1486
T2d = VFNMS(LDK(KP634393284), T2c, VMUL(LDK(KP773010453), T23));
1487
T2e = VADD(T1C, T2d);
1488
T2z = VSUB(T2d, T1C);
1489
T2o = VSUB(T2i, T2n);
1490
T2r = VSUB(T2p, T2q);
1491
T2s = VADD(T2o, T2r);
1492
T2A = VSUB(T2r, T2o);
1495
V T2f, T2w, T7L, T7M;
1496
T2f = VADD(T11, T2e);
1497
T2w = VBYI(VADD(T2s, T2v));
1498
T7L = VSUB(T2f, T2w);
1499
STM2(&(xo[114]), T7L, ovs, &(xo[2]));
1500
STN2(&(xo[112]), T7r, T7L, ovs);
1501
T7M = VADD(T2f, T2w);
1502
STM2(&(xo[14]), T7M, ovs, &(xo[2]));
1503
STN2(&(xo[12]), T7E, T7M, ovs);
1506
V T2F, T2G, T7N, T7O;
1507
T2F = VBYI(VADD(T2A, T2z));
1508
T2G = VADD(T2C, T2D);
1509
T7N = VADD(T2F, T2G);
1510
STM2(&(xo[18]), T7N, ovs, &(xo[2]));
1511
STN2(&(xo[16]), T7s, T7N, ovs);
1512
T7O = VSUB(T2G, T2F);
1513
STM2(&(xo[110]), T7O, ovs, &(xo[2]));
1514
STN2(&(xo[108]), T7G, T7O, ovs);
1517
V T2x, T2y, T7P, T7Q;
1518
T2x = VSUB(T11, T2e);
1519
T2y = VBYI(VSUB(T2v, T2s));
1520
T7P = VSUB(T2x, T2y);
1521
STM2(&(xo[78]), T7P, ovs, &(xo[2]));
1522
STN2(&(xo[76]), T7H, T7P, ovs);
1523
T7Q = VADD(T2x, T2y);
1524
STM2(&(xo[50]), T7Q, ovs, &(xo[2]));
1525
STN2(&(xo[48]), T7u, T7Q, ovs);
1528
V T2B, T2E, T7R, T7S;
1529
T2B = VBYI(VSUB(T2z, T2A));
1530
T2E = VSUB(T2C, T2D);
1531
T7R = VADD(T2B, T2E);
1532
STM2(&(xo[46]), T7R, ovs, &(xo[2]));
1533
STN2(&(xo[44]), T7J, T7R, ovs);
1534
T7S = VSUB(T2E, T2B);
1535
STM2(&(xo[82]), T7S, ovs, &(xo[2]));
1536
STN2(&(xo[80]), T7t, T7S, ovs);
1540
V T3j, T3Q, T3J, T3R, T3y, T3N, T3G, T3O;
1542
V T3b, T3i, T3H, T3I;
1543
T3b = VADD(T39, T3a);
1544
T3i = VADD(T3e, T3h);
1545
T3j = VADD(T3b, T3i);
1546
T3Q = VSUB(T3b, T3i);
1547
T3H = VFNMS(LDK(KP290284677), T3m, VMUL(LDK(KP956940335), T3p));
1548
T3I = VFMA(LDK(KP290284677), T3t, VMUL(LDK(KP956940335), T3w));
1549
T3J = VADD(T3H, T3I);
1550
T3R = VSUB(T3I, T3H);
1553
V T3q, T3x, T3C, T3F;
1554
T3q = VFMA(LDK(KP956940335), T3m, VMUL(LDK(KP290284677), T3p));
1555
T3x = VFNMS(LDK(KP290284677), T3w, VMUL(LDK(KP956940335), T3t));
1556
T3y = VADD(T3q, T3x);
1557
T3N = VSUB(T3x, T3q);
1558
T3C = VADD(T3A, T3B);
1559
T3F = VADD(T3D, T3E);
1560
T3G = VADD(T3C, T3F);
1561
T3O = VSUB(T3F, T3C);
1564
V T3z, T3K, T7T, T3T, T3U, T7V;
1565
T3z = VADD(T3j, T3y);
1566
T3K = VBYI(VADD(T3G, T3J));
1567
T7T = VSUB(T3z, T3K);
1568
STM2(&(xo[122]), T7T, ovs, &(xo[2]));
1569
STN2(&(xo[120]), T7v, T7T, ovs);
1570
T7U = VADD(T3z, T3K);
1571
STM2(&(xo[6]), T7U, ovs, &(xo[2]));
1572
T3T = VBYI(VADD(T3O, T3N));
1573
T3U = VADD(T3Q, T3R);
1574
T7V = VADD(T3T, T3U);
1575
STM2(&(xo[26]), T7V, ovs, &(xo[2]));
1576
STN2(&(xo[24]), T7x, T7V, ovs);
1577
T7W = VSUB(T3U, T3T);
1578
STM2(&(xo[102]), T7W, ovs, &(xo[2]));
1581
V T3L, T3M, T7Y, T3P, T3S, T80;
1582
T3L = VSUB(T3j, T3y);
1583
T3M = VBYI(VSUB(T3J, T3G));
1584
T7X = VSUB(T3L, T3M);
1585
STM2(&(xo[70]), T7X, ovs, &(xo[2]));
1586
T7Y = VADD(T3L, T3M);
1587
STM2(&(xo[58]), T7Y, ovs, &(xo[2]));
1588
STN2(&(xo[56]), T7A, T7Y, ovs);
1589
T3P = VBYI(VSUB(T3N, T3O));
1590
T3S = VSUB(T3Q, T3R);
1591
T7Z = VADD(T3P, T3S);
1592
STM2(&(xo[38]), T7Z, ovs, &(xo[2]));
1593
T80 = VSUB(T3S, T3P);
1594
STM2(&(xo[90]), T80, ovs, &(xo[2]));
1595
STN2(&(xo[88]), T7C, T80, ovs);
1599
V T81, T83, T86, T88;
1601
V T4N, T5G, T5z, T5H, T5m, T5D, T5w, T5E;
1603
V T4x, T4M, T5x, T5y;
1604
T4x = VADD(T4p, T4w);
1605
T4M = VADD(T4E, T4L);
1606
T4N = VADD(T4x, T4M);
1607
T5G = VSUB(T4x, T4M);
1608
T5x = VFNMS(LDK(KP195090322), T4Y, VMUL(LDK(KP980785280), T53));
1609
T5y = VFMA(LDK(KP195090322), T5f, VMUL(LDK(KP980785280), T5k));
1610
T5z = VADD(T5x, T5y);
1611
T5H = VSUB(T5y, T5x);
1614
V T54, T5l, T5s, T5v;
1615
T54 = VFMA(LDK(KP980785280), T4Y, VMUL(LDK(KP195090322), T53));
1616
T5l = VFNMS(LDK(KP195090322), T5k, VMUL(LDK(KP980785280), T5f));
1617
T5m = VADD(T54, T5l);
1618
T5D = VSUB(T5l, T54);
1619
T5s = VADD(T5q, T5r);
1620
T5v = VADD(T5t, T5u);
1621
T5w = VADD(T5s, T5v);
1622
T5E = VSUB(T5v, T5s);
1625
V T5n, T5A, T82, T5J, T5K, T84;
1626
T5n = VADD(T4N, T5m);
1627
T5A = VBYI(VADD(T5w, T5z));
1628
T81 = VSUB(T5n, T5A);
1629
STM2(&(xo[124]), T81, ovs, &(xo[0]));
1630
T82 = VADD(T5n, T5A);
1631
STM2(&(xo[4]), T82, ovs, &(xo[0]));
1632
STN2(&(xo[4]), T82, T7U, ovs);
1633
T5J = VBYI(VADD(T5E, T5D));
1634
T5K = VADD(T5G, T5H);
1635
T83 = VADD(T5J, T5K);
1636
STM2(&(xo[28]), T83, ovs, &(xo[0]));
1637
T84 = VSUB(T5K, T5J);
1638
STM2(&(xo[100]), T84, ovs, &(xo[0]));
1639
STN2(&(xo[100]), T84, T7W, ovs);
1642
V T5B, T5C, T85, T5F, T5I, T87;
1643
T5B = VSUB(T4N, T5m);
1644
T5C = VBYI(VSUB(T5z, T5w));
1645
T85 = VSUB(T5B, T5C);
1646
STM2(&(xo[68]), T85, ovs, &(xo[0]));
1647
STN2(&(xo[68]), T85, T7X, ovs);
1648
T86 = VADD(T5B, T5C);
1649
STM2(&(xo[60]), T86, ovs, &(xo[0]));
1650
T5F = VBYI(VSUB(T5D, T5E));
1651
T5I = VSUB(T5G, T5H);
1652
T87 = VADD(T5F, T5I);
1653
STM2(&(xo[36]), T87, ovs, &(xo[0]));
1654
STN2(&(xo[36]), T87, T7Z, ovs);
1655
T88 = VSUB(T5I, T5F);
1656
STM2(&(xo[92]), T88, ovs, &(xo[0]));
1660
V T2J, T34, T2X, T35, T2Q, T31, T2U, T32;
1662
V T2H, T2I, T2V, T2W;
1664
T2I = VADD(T2q, T2p);
1665
T2J = VADD(T2H, T2I);
1666
T34 = VSUB(T2H, T2I);
1667
T2V = VFNMS(LDK(KP098017140), T2L, VMUL(LDK(KP995184726), T2K));
1668
T2W = VFMA(LDK(KP995184726), T2O, VMUL(LDK(KP098017140), T2N));
1669
T2X = VADD(T2V, T2W);
1670
T35 = VSUB(T2W, T2V);
1673
V T2M, T2P, T2S, T2T;
1674
T2M = VFMA(LDK(KP098017140), T2K, VMUL(LDK(KP995184726), T2L));
1675
T2P = VFNMS(LDK(KP098017140), T2O, VMUL(LDK(KP995184726), T2N));
1676
T2Q = VADD(T2M, T2P);
1677
T31 = VSUB(T2P, T2M);
1678
T2S = VADD(T2n, T2i);
1680
T2U = VADD(T2S, T2T);
1681
T32 = VSUB(T2T, T2S);
1684
V T2R, T2Y, T89, T8a;
1685
T2R = VADD(T2J, T2Q);
1686
T2Y = VBYI(VADD(T2U, T2X));
1687
T89 = VSUB(T2R, T2Y);
1688
STM2(&(xo[126]), T89, ovs, &(xo[2]));
1689
STN2(&(xo[124]), T81, T89, ovs);
1690
T8a = VADD(T2R, T2Y);
1691
STM2(&(xo[2]), T8a, ovs, &(xo[2]));
1692
STN2(&(xo[0]), T7p, T8a, ovs);
1695
V T37, T38, T8b, T8c;
1696
T37 = VBYI(VADD(T32, T31));
1697
T38 = VADD(T34, T35);
1698
T8b = VADD(T37, T38);
1699
STM2(&(xo[30]), T8b, ovs, &(xo[2]));
1700
STN2(&(xo[28]), T83, T8b, ovs);
1701
T8c = VSUB(T38, T37);
1702
STM2(&(xo[98]), T8c, ovs, &(xo[2]));
1703
STN2(&(xo[96]), T7q, T8c, ovs);
1706
V T2Z, T30, T8d, T8e;
1707
T2Z = VSUB(T2J, T2Q);
1708
T30 = VBYI(VSUB(T2X, T2U));
1709
T8d = VSUB(T2Z, T30);
1710
STM2(&(xo[66]), T8d, ovs, &(xo[2]));
1711
STN2(&(xo[64]), T7n, T8d, ovs);
1712
T8e = VADD(T2Z, T30);
1713
STM2(&(xo[62]), T8e, ovs, &(xo[2]));
1714
STN2(&(xo[60]), T86, T8e, ovs);
1717
V T33, T36, T8f, T8g;
1718
T33 = VBYI(VSUB(T31, T32));
1719
T36 = VSUB(T34, T35);
1720
T8f = VADD(T33, T36);
1721
STM2(&(xo[34]), T8f, ovs, &(xo[2]));
1722
STN2(&(xo[32]), T7o, T8f, ovs);
1723
T8g = VSUB(T36, T33);
1724
STM2(&(xo[94]), T8g, ovs, &(xo[2]));
1725
STN2(&(xo[92]), T88, T8g, ovs);
1729
V T3X, T4i, T4b, T4j, T44, T4f, T48, T4g;
1731
V T3V, T3W, T49, T4a;
1732
T3V = VSUB(T39, T3a);
1733
T3W = VSUB(T3E, T3D);
1734
T3X = VADD(T3V, T3W);
1735
T4i = VSUB(T3V, T3W);
1736
T49 = VFNMS(LDK(KP471396736), T3Y, VMUL(LDK(KP881921264), T3Z));
1737
T4a = VFMA(LDK(KP471396736), T41, VMUL(LDK(KP881921264), T42));
1738
T4b = VADD(T49, T4a);
1739
T4j = VSUB(T4a, T49);
1742
V T40, T43, T46, T47;
1743
T40 = VFMA(LDK(KP881921264), T3Y, VMUL(LDK(KP471396736), T3Z));
1744
T43 = VFNMS(LDK(KP471396736), T42, VMUL(LDK(KP881921264), T41));
1745
T44 = VADD(T40, T43);
1746
T4f = VSUB(T43, T40);
1747
T46 = VSUB(T3B, T3A);
1748
T47 = VSUB(T3h, T3e);
1749
T48 = VADD(T46, T47);
1750
T4g = VSUB(T47, T46);
1753
V T45, T4c, T8h, T8i;
1754
T45 = VADD(T3X, T44);
1755
T4c = VBYI(VADD(T48, T4b));
1756
T8h = VSUB(T45, T4c);
1757
STM2(&(xo[118]), T8h, ovs, &(xo[2]));
1758
STN2(&(xo[116]), T7D, T8h, ovs);
1759
T8i = VADD(T45, T4c);
1760
STM2(&(xo[10]), T8i, ovs, &(xo[2]));
1761
STN2(&(xo[8]), T7w, T8i, ovs);
1764
V T4l, T4m, T8j, T8k;
1765
T4l = VBYI(VADD(T4g, T4f));
1766
T4m = VADD(T4i, T4j);
1767
T8j = VADD(T4l, T4m);
1768
STM2(&(xo[22]), T8j, ovs, &(xo[2]));
1769
STN2(&(xo[20]), T7F, T8j, ovs);
1770
T8k = VSUB(T4m, T4l);
1771
STM2(&(xo[106]), T8k, ovs, &(xo[2]));
1772
STN2(&(xo[104]), T7y, T8k, ovs);
1775
V T4d, T4e, T8l, T8m;
1776
T4d = VSUB(T3X, T44);
1777
T4e = VBYI(VSUB(T4b, T48));
1778
T8l = VSUB(T4d, T4e);
1779
STM2(&(xo[74]), T8l, ovs, &(xo[2]));
1780
STN2(&(xo[72]), T7z, T8l, ovs);
1781
T8m = VADD(T4d, T4e);
1782
STM2(&(xo[54]), T8m, ovs, &(xo[2]));
1783
STN2(&(xo[52]), T7I, T8m, ovs);
1786
V T4h, T4k, T8n, T8o;
1787
T4h = VBYI(VSUB(T4f, T4g));
1788
T4k = VSUB(T4i, T4j);
1789
T8n = VADD(T4h, T4k);
1790
STM2(&(xo[42]), T8n, ovs, &(xo[2]));
1791
STN2(&(xo[40]), T7B, T8n, ovs);
1792
T8o = VSUB(T4k, T4h);
1793
STM2(&(xo[86]), T8o, ovs, &(xo[2]));
1794
STN2(&(xo[84]), T7K, T8o, ovs);
1803
static const kdft_desc desc = { 64, "n2fv_64", {404, 72, 52, 0}, &GENUS, 0, 2, 0, 0 };
1805
void X(codelet_n2fv_64) (planner *p) {
1806
X(kdft_register) (p, n2fv_64, &desc);
1809
#endif /* HAVE_FMA */