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

« back to all changes in this revision

Viewing changes to dft/simd/common/n1fv_64.c

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

Show diffs side-by-side

added added

removed removed

Lines of Context:
 
1
/*
 
2
 * Copyright (c) 2003, 2007-11 Matteo Frigo
 
3
 * Copyright (c) 2003, 2007-11 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 Wed Jul 27 06:13:33 EDT 2011 */
 
23
 
 
24
#include "codelet-dft.h"
 
25
 
 
26
#ifdef HAVE_FMA
 
27
 
 
28
/* Generated by: ../../../genfft/gen_notw_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n1fv_64 -include n1f.h */
 
29
 
 
30
/*
 
31
 * This function contains 456 FP additions, 258 FP multiplications,
 
32
 * (or, 198 additions, 0 multiplications, 258 fused multiply/add),
 
33
 * 168 stack variables, 15 constants, and 128 memory accesses
 
34
 */
 
35
#include "n1f.h"
 
36
 
 
37
static void n1fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
 
38
{
 
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);
 
54
     {
 
55
          INT i;
 
56
          const R *xi;
 
57
          R *xo;
 
58
          xi = ri;
 
59
          xo = ro;
 
60
          for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
 
61
               V T5T, T5S, T5X, T65, T5Z, T5R, T67, T63, T5U, T64;
 
62
               {
 
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;
 
67
                    V T6m, T6Y, T5L, T4T;
 
68
                    {
 
69
                         V T4g, T4l, T3j, Tu, Tx, T4h, TA, T4i;
 
70
                         {
 
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]));
 
80
                              {
 
81
                                   V Ta, T48, Tk, T4c, T49, Td, Tf, Tg;
 
82
                                   {
 
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]));
 
85
                                        T43 = VSUB(T1, T2);
 
86
                                        T3 = VADD(T1, T2);
 
87
                                        T44 = VSUB(T23, T24);
 
88
                                        T25 = VADD(T23, T24);
 
89
                                        T5i = VSUB(T4, T5);
 
90
                                        T6 = VADD(T4, T5);
 
91
                                        T45 = VSUB(T20, T21);
 
92
                                        T22 = VADD(T20, T21);
 
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]));
 
98
                                        {
 
99
                                             V T2T, T46, T5j, T2U;
 
100
                                             T7 = VSUB(T3, T6);
 
101
                                             T2T = VADD(T3, T6);
 
102
                                             T46 = VADD(T44, T45);
 
103
                                             T5j = VSUB(T45, T44);
 
104
                                             T26 = VSUB(T22, T25);
 
105
                                             T2U = VADD(T25, T22);
 
106
                                             Ta = VADD(T8, T9);
 
107
                                             T48 = VSUB(T8, T9);
 
108
                                             Tk = VADD(Ti, Tj);
 
109
                                             T4c = VSUB(Tj, Ti);
 
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);
 
116
                                             T49 = VSUB(Tb, Tc);
 
117
                                             Td = VADD(Tb, Tc);
 
118
                                        }
 
119
                                        Tf = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
 
120
                                        Tg = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
 
121
                                   }
 
122
                                   {
 
123
                                        V Te, T2W, T5l, T4a, Tq, Tt, Tv, Tw, T5m, T4d, Tl, T2X, Ty, Tz, To;
 
124
                                        V Tp;
 
125
                                        To = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
 
126
                                        Tp = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
 
127
                                        {
 
128
                                             V Th, T4b, Tr, Ts;
 
129
                                             Tr = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
 
130
                                             Ts = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
 
131
                                             Te = VSUB(Ta, Td);
 
132
                                             T2W = VADD(Ta, Td);
 
133
                                             T5l = VFMA(LDK(KP414213562), T48, T49);
 
134
                                             T4a = VFNMS(LDK(KP414213562), T49, T48);
 
135
                                             Th = VADD(Tf, Tg);
 
136
                                             T4b = VSUB(Tf, Tg);
 
137
                                             Tq = VADD(To, Tp);
 
138
                                             T4g = VSUB(To, Tp);
 
139
                                             T4l = VSUB(Tr, Ts);
 
140
                                             Tt = VADD(Tr, Ts);
 
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);
 
145
                                             Tl = VSUB(Th, Tk);
 
146
                                             T2X = VADD(Th, Tk);
 
147
                                             Ty = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
 
148
                                             Tz = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
 
149
                                        }
 
150
                                        T3j = VADD(Tq, Tt);
 
151
                                        Tu = VSUB(Tq, Tt);
 
152
                                        Tx = VADD(Tv, Tw);
 
153
                                        T4h = VSUB(Tv, Tw);
 
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);
 
160
                                        T27 = VSUB(Tl, Te);
 
161
                                        Tm = VADD(Te, Tl);
 
162
                                        TA = VADD(Ty, Tz);
 
163
                                        T4i = VSUB(Ty, Tz);
 
164
                                   }
 
165
                              }
 
166
                         }
 
167
                         {
 
168
                              V TK, T4p, T4u, T4k, T6d, T4n, T6c, TL, TN, TO, T3g, TJ, TF, TI;
 
169
                              {
 
170
                                   V TD, TE, TG, TH;
 
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]));
 
176
                                   {
 
177
                                        V T3k, TB, T4j, T4m;
 
178
                                        T3k = VADD(Tx, TA);
 
179
                                        TB = VSUB(Tx, TA);
 
180
                                        T4j = VADD(T4h, T4i);
 
181
                                        T4m = VSUB(T4h, T4i);
 
182
                                        T4p = VSUB(TD, TE);
 
183
                                        TF = VADD(TD, TE);
 
184
                                        T4u = VSUB(TH, TG);
 
185
                                        TI = VADD(TG, TH);
 
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]));
 
195
                                   }
 
196
                                   TN = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
 
197
                                   TO = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
 
198
                              }
 
199
                              T3g = VADD(TF, TI);
 
200
                              TJ = VSUB(TF, TI);
 
201
                              {
 
202
                                   V T3a, T1E, T52, T5b, T1x, T4Z, T6r, T6u, T5a, T1U, T55, T5c, T1L, T3b;
 
203
                                   {
 
204
                                        V T4V, T1t, T58, T1w, T1Q, T1T, T1I, T4Y, T59, T1J, T53, T1H;
 
205
                                        {
 
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);
 
212
                                             TM = VADD(TK, TL);
 
213
                                             T4r = VSUB(TK, TL);
 
214
                                             TP = VADD(TN, TO);
 
215
                                             T4q = VSUB(TN, TO);
 
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)]));
 
219
                                             {
 
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)]));
 
224
                                                  {
 
225
                                                       V T3h, TQ, T4s, T4v;
 
226
                                                       T3h = VADD(TP, TM);
 
227
                                                       TQ = VSUB(TM, TP);
 
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)]));
 
245
                                                  }
 
246
                                                  {
 
247
                                                       V T4W, T1A, T50, T51, T1D, T1F, T1G;
 
248
                                                       {
 
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);
 
264
                                                       }
 
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);
 
277
                                                  }
 
278
                                             }
 
279
                                        }
 
280
                                        {
 
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);
 
298
                                        }
 
299
                                   }
 
300
                                   {
 
301
                                        V T4A, TW, T4N, TZ, T1j, T1m, T4O, T4D, T13, T4F, T16, T4G, T1a, T4I, T4J;
 
302
                                        V T1d;
 
303
                                        {
 
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);
 
308
                                             {
 
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)]));
 
329
                                             }
 
330
                                             TX = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
 
331
                                             TY = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
 
332
                                             {
 
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)]));
 
338
                                                  {
 
339
                                                       V T11, T4B, T4C, T12, T14, T15;
 
340
                                                       T11 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
 
341
                                                       T4A = VSUB(TU, TV);
 
342
                                                       TW = VADD(TU, TV);
 
343
                                                       T4N = VSUB(TX, TY);
 
344
                                                       TZ = VADD(TX, TY);
 
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)]));
 
352
                                                       {
 
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);
 
368
                                                       }
 
369
                                                  }
 
370
                                             }
 
371
                                        }
 
372
                                        {
 
373
                                             V T30, T10, T6k, T4E, T4Q, T4H, T17, T6n, T4P, T1e, T4K, T4R, T1n, T31;
 
374
                                             T30 = VADD(TW, TZ);
 
375
                                             T10 = VSUB(TW, TZ);
 
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);
 
390
                                             {
 
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);
 
412
                                             }
 
413
                                        }
 
414
                                   }
 
415
                              }
 
416
                         }
 
417
                    }
 
418
                    {
 
419
                         V T6b, T6F, T7f, T6X, T70, T79, T7a, T73, T6C, T76, T77, T6i;
 
420
                         {
 
421
                              V T2Z, T3r, T3s, T3m, T3d, T3v;
 
422
                              T2Z = VSUB(T2V, T2Y);
 
423
                              T3r = VADD(T2V, T2Y);
 
424
                              T3s = VADD(T3l, T3i);
 
425
                              T3m = VSUB(T3i, T3l);
 
426
                              T3d = VSUB(T39, T3c);
 
427
                              T3v = VADD(T39, T3c);
 
428
                              {
 
429
                                   V T3x, T3t, T3P, T3J, T3D, T3V, T3Q, T3G, T36, T3u, T3Y, T3O, T6V, T6W;
 
430
                                   {
 
431
                                        V T3N, T3C, T3F, T35;
 
432
                                        T3N = VSUB(T3B, T3A);
 
433
                                        T3C = VADD(T3A, T3B);
 
434
                                        T3F = VSUB(T33, T34);
 
435
                                        T35 = VADD(T33, T34);
 
436
                                        T3x = VSUB(T3r, T3s);
 
437
                                        T3t = VADD(T3r, T3s);
 
438
                                        T3P = VFMA(LDK(KP414213562), T3H, T3I);
 
439
                                        T3J = VFNMS(LDK(KP414213562), T3I, T3H);
 
440
                                        T3D = VFMA(LDK(KP707106781), T3C, T3z);
 
441
                                        T3V = VFNMS(LDK(KP707106781), T3C, T3z);
 
442
                                        T3Q = VFMA(LDK(KP414213562), T3E, T3F);
 
443
                                        T3G = VFNMS(LDK(KP414213562), T3F, T3E);
 
444
                                        T36 = VSUB(T32, T35);
 
445
                                        T3u = VADD(T32, T35);
 
446
                                        T3Y = VFNMS(LDK(KP707106781), T3N, T3M);
 
447
                                        T3O = VFMA(LDK(KP707106781), T3N, T3M);
 
448
                                   }
 
449
                                   T6b = VFNMS(LDK(KP923879532), T6a, T69);
 
450
                                   T6V = VFMA(LDK(KP923879532), T6a, T69);
 
451
                                   T6W = VADD(T6E, T6D);
 
452
                                   T6F = VSUB(T6D, T6E);
 
453
                                   {
 
454
                                        V T3K, T3Z, T3e, T3n;
 
455
                                        T3K = VADD(T3G, T3J);
 
456
                                        T3Z = VSUB(T3J, T3G);
 
457
                                        T3e = VADD(T36, T3d);
 
458
                                        T3n = VSUB(T3d, T36);
 
459
                                        {
 
460
                                             V T3w, T3y, T3R, T3W;
 
461
                                             T3w = VADD(T3u, T3v);
 
462
                                             T3y = VSUB(T3v, T3u);
 
463
                                             T3R = VSUB(T3P, T3Q);
 
464
                                             T3W = VADD(T3Q, T3P);
 
465
                                             {
 
466
                                                  V T42, T40, T3L, T3T;
 
467
                                                  T42 = VFNMS(LDK(KP923879532), T3Z, T3Y);
 
468
                                                  T40 = VFMA(LDK(KP923879532), T3Z, T3Y);
 
469
                                                  T3L = VFNMS(LDK(KP923879532), T3K, T3D);
 
470
                                                  T3T = VFMA(LDK(KP923879532), T3K, T3D);
 
471
                                                  {
 
472
                                                       V T3o, T3q, T3f, T3p;
 
473
                                                       T3o = VFNMS(LDK(KP707106781), T3n, T3m);
 
474
                                                       T3q = VFMA(LDK(KP707106781), T3n, T3m);
 
475
                                                       T3f = VFNMS(LDK(KP707106781), T3e, T2Z);
 
476
                                                       T3p = VFMA(LDK(KP707106781), T3e, T2Z);
 
477
                                                       ST(&(xo[WS(os, 48)]), VFNMSI(T3y, T3x), ovs, &(xo[0]));
 
478
                                                       ST(&(xo[WS(os, 16)]), VFMAI(T3y, T3x), ovs, &(xo[0]));
 
479
                                                       ST(&(xo[0]), VADD(T3t, T3w), ovs, &(xo[0]));
 
480
                                                       ST(&(xo[WS(os, 32)]), VSUB(T3t, T3w), ovs, &(xo[0]));
 
481
                                                       {
 
482
                                                            V T41, T3X, T3S, T3U;
 
483
                                                            T41 = VFMA(LDK(KP923879532), T3W, T3V);
 
484
                                                            T3X = VFNMS(LDK(KP923879532), T3W, T3V);
 
485
                                                            T3S = VFNMS(LDK(KP923879532), T3R, T3O);
 
486
                                                            T3U = VFMA(LDK(KP923879532), T3R, T3O);
 
487
                                                            ST(&(xo[WS(os, 8)]), VFMAI(T3q, T3p), ovs, &(xo[0]));
 
488
                                                            ST(&(xo[WS(os, 56)]), VFNMSI(T3q, T3p), ovs, &(xo[0]));
 
489
                                                            ST(&(xo[WS(os, 40)]), VFMAI(T3o, T3f), ovs, &(xo[0]));
 
490
                                                            ST(&(xo[WS(os, 24)]), VFNMSI(T3o, T3f), ovs, &(xo[0]));
 
491
                                                            ST(&(xo[WS(os, 44)]), VFNMSI(T40, T3X), ovs, &(xo[0]));
 
492
                                                            ST(&(xo[WS(os, 20)]), VFMAI(T40, T3X), ovs, &(xo[0]));
 
493
                                                            ST(&(xo[WS(os, 52)]), VFMAI(T42, T41), ovs, &(xo[0]));
 
494
                                                            ST(&(xo[WS(os, 12)]), VFNMSI(T42, T41), ovs, &(xo[0]));
 
495
                                                            ST(&(xo[WS(os, 4)]), VFMAI(T3U, T3T), ovs, &(xo[0]));
 
496
                                                            ST(&(xo[WS(os, 60)]), VFNMSI(T3U, T3T), ovs, &(xo[0]));
 
497
                                                            ST(&(xo[WS(os, 36)]), VFMAI(T3S, T3L), ovs, &(xo[0]));
 
498
                                                            ST(&(xo[WS(os, 28)]), VFNMSI(T3S, T3L), ovs, &(xo[0]));
 
499
                                                            T7f = VFNMS(LDK(KP831469612), T6W, T6V);
 
500
                                                            T6X = VFMA(LDK(KP831469612), T6W, T6V);
 
501
                                                       }
 
502
                                                  }
 
503
                                             }
 
504
                                        }
 
505
                                   }
 
506
                                   T70 = VFMA(LDK(KP303346683), T6Z, T6Y);
 
507
                                   T79 = VFNMS(LDK(KP303346683), T6Y, T6Z);
 
508
                                   T7a = VFNMS(LDK(KP303346683), T71, T72);
 
509
                                   T73 = VFMA(LDK(KP303346683), T72, T71);
 
510
                                   T6C = VFNMS(LDK(KP923879532), T6B, T6A);
 
511
                                   T76 = VFMA(LDK(KP923879532), T6B, T6A);
 
512
                                   T77 = VSUB(T6e, T6h);
 
513
                                   T6i = VADD(T6e, T6h);
 
514
                              }
 
515
                         }
 
516
                         {
 
517
                              V T2r, T2D, T2C, T2s, T5H, T5o, T5v, T5D, T5r, T5I, T5x, T5h, T5F, T5B;
 
518
                              {
 
519
                                   V TT, T2f, T2n, T1Y, T28, T2b, T2l, T2p, T2j, T2k;
 
520
                                   {
 
521
                                        V T1X, T2d, T7h, T7l, T2e, T1q, T75, T7d, T7m, T7k, T7c, T7e, Tn, TS;
 
522
                                        T2r = VFNMS(LDK(KP707106781), Tm, T7);
 
523
                                        Tn = VFMA(LDK(KP707106781), Tm, T7);
 
524
                                        TS = VADD(TC, TR);
 
525
                                        T2D = VSUB(TR, TC);
 
526
                                        {
 
527
                                             V T7b, T7j, T74, T7i, T78, T7g;
 
528
                                             T1X = VFNMS(LDK(KP198912367), T1W, T1N);
 
529
                                             T2d = VFMA(LDK(KP198912367), T1N, T1W);
 
530
                                             T7g = VADD(T79, T7a);
 
531
                                             T7b = VSUB(T79, T7a);
 
532
                                             T7j = VSUB(T73, T70);
 
533
                                             T74 = VADD(T70, T73);
 
534
                                             T7i = VFNMS(LDK(KP831469612), T77, T76);
 
535
                                             T78 = VFMA(LDK(KP831469612), T77, T76);
 
536
                                             T2j = VFNMS(LDK(KP923879532), TS, Tn);
 
537
                                             TT = VFMA(LDK(KP923879532), TS, Tn);
 
538
                                             T7h = VFMA(LDK(KP956940335), T7g, T7f);
 
539
                                             T7l = VFNMS(LDK(KP956940335), T7g, T7f);
 
540
                                             T2e = VFMA(LDK(KP198912367), T1g, T1p);
 
541
                                             T1q = VFNMS(LDK(KP198912367), T1p, T1g);
 
542
                                             T75 = VFNMS(LDK(KP956940335), T74, T6X);
 
543
                                             T7d = VFMA(LDK(KP956940335), T74, T6X);
 
544
                                             T7m = VFNMS(LDK(KP956940335), T7j, T7i);
 
545
                                             T7k = VFMA(LDK(KP956940335), T7j, T7i);
 
546
                                             T7c = VFNMS(LDK(KP956940335), T7b, T78);
 
547
                                             T7e = VFMA(LDK(KP956940335), T7b, T78);
 
548
                                        }
 
549
                                        T2k = VADD(T2e, T2d);
 
550
                                        T2f = VSUB(T2d, T2e);
 
551
                                        ST(&(xo[WS(os, 45)]), VFNMSI(T7k, T7h), ovs, &(xo[WS(os, 1)]));
 
552
                                        ST(&(xo[WS(os, 19)]), VFMAI(T7k, T7h), ovs, &(xo[WS(os, 1)]));
 
553
                                        ST(&(xo[WS(os, 51)]), VFMAI(T7m, T7l), ovs, &(xo[WS(os, 1)]));
 
554
                                        ST(&(xo[WS(os, 13)]), VFNMSI(T7m, T7l), ovs, &(xo[WS(os, 1)]));
 
555
                                        ST(&(xo[WS(os, 3)]), VFMAI(T7e, T7d), ovs, &(xo[WS(os, 1)]));
 
556
                                        ST(&(xo[WS(os, 61)]), VFNMSI(T7e, T7d), ovs, &(xo[WS(os, 1)]));
 
557
                                        ST(&(xo[WS(os, 35)]), VFMAI(T7c, T75), ovs, &(xo[WS(os, 1)]));
 
558
                                        ST(&(xo[WS(os, 29)]), VFNMSI(T7c, T75), ovs, &(xo[WS(os, 1)]));
 
559
                                        T2n = VSUB(T1X, T1q);
 
560
                                        T1Y = VADD(T1q, T1X);
 
561
                                        T2C = VFNMS(LDK(KP707106781), T27, T26);
 
562
                                        T28 = VFMA(LDK(KP707106781), T27, T26);
 
563
                                        T2b = VSUB(T29, T2a);
 
564
                                        T2s = VADD(T2a, T29);
 
565
                                   }
 
566
                                   T2l = VFNMS(LDK(KP980785280), T2k, T2j);
 
567
                                   T2p = VFMA(LDK(KP980785280), T2k, T2j);
 
568
                                   {
 
569
                                        V T5z, T4z, T5A, T5g;
 
570
                                        {
 
571
                                             V T4f, T4y, T1Z, T2h, T4U, T5t, T2m, T2c, T5u, T5f;
 
572
                                             T5H = VFNMS(LDK(KP923879532), T4e, T47);
 
573
                                             T4f = VFMA(LDK(KP923879532), T4e, T47);
 
574
                                             T4y = VADD(T4o, T4x);
 
575
                                             T5T = VSUB(T4x, T4o);
 
576
                                             T1Z = VFNMS(LDK(KP980785280), T1Y, TT);
 
577
                                             T2h = VFMA(LDK(KP980785280), T1Y, TT);
 
578
                                             T4U = VFNMS(LDK(KP098491403), T4T, T4M);
 
579
                                             T5t = VFMA(LDK(KP098491403), T4M, T4T);
 
580
                                             T2m = VFNMS(LDK(KP923879532), T2b, T28);
 
581
                                             T2c = VFMA(LDK(KP923879532), T2b, T28);
 
582
                                             T5u = VFMA(LDK(KP098491403), T57, T5e);
 
583
                                             T5f = VFNMS(LDK(KP098491403), T5e, T57);
 
584
                                             T5z = VFNMS(LDK(KP980785280), T4y, T4f);
 
585
                                             T4z = VFMA(LDK(KP980785280), T4y, T4f);
 
586
                                             T5S = VFNMS(LDK(KP923879532), T5n, T5k);
 
587
                                             T5o = VFMA(LDK(KP923879532), T5n, T5k);
 
588
                                             {
 
589
                                                  V T2o, T2q, T2i, T2g;
 
590
                                                  T2o = VFMA(LDK(KP980785280), T2n, T2m);
 
591
                                                  T2q = VFNMS(LDK(KP980785280), T2n, T2m);
 
592
                                                  T2i = VFMA(LDK(KP980785280), T2f, T2c);
 
593
                                                  T2g = VFNMS(LDK(KP980785280), T2f, T2c);
 
594
                                                  T5A = VADD(T5t, T5u);
 
595
                                                  T5v = VSUB(T5t, T5u);
 
596
                                                  T5D = VSUB(T5f, T4U);
 
597
                                                  T5g = VADD(T4U, T5f);
 
598
                                                  ST(&(xo[WS(os, 46)]), VFNMSI(T2o, T2l), ovs, &(xo[0]));
 
599
                                                  ST(&(xo[WS(os, 18)]), VFMAI(T2o, T2l), ovs, &(xo[0]));
 
600
                                                  ST(&(xo[WS(os, 50)]), VFMAI(T2q, T2p), ovs, &(xo[0]));
 
601
                                                  ST(&(xo[WS(os, 14)]), VFNMSI(T2q, T2p), ovs, &(xo[0]));
 
602
                                                  ST(&(xo[WS(os, 2)]), VFMAI(T2i, T2h), ovs, &(xo[0]));
 
603
                                                  ST(&(xo[WS(os, 62)]), VFNMSI(T2i, T2h), ovs, &(xo[0]));
 
604
                                                  ST(&(xo[WS(os, 34)]), VFMAI(T2g, T1Z), ovs, &(xo[0]));
 
605
                                                  ST(&(xo[WS(os, 30)]), VFNMSI(T2g, T1Z), ovs, &(xo[0]));
 
606
                                                  T5r = VSUB(T5p, T5q);
 
607
                                                  T5I = VADD(T5p, T5q);
 
608
                                             }
 
609
                                        }
 
610
                                        T5x = VFMA(LDK(KP995184726), T5g, T4z);
 
611
                                        T5h = VFNMS(LDK(KP995184726), T5g, T4z);
 
612
                                        T5F = VFMA(LDK(KP995184726), T5A, T5z);
 
613
                                        T5B = VFNMS(LDK(KP995184726), T5A, T5z);
 
614
                                   }
 
615
                              }
 
616
                              {
 
617
                                   V T6J, T6R, T6L, T6z, T6T, T6P;
 
618
                                   {
 
619
                                        V T6N, T6j, T6O, T6y;
 
620
                                        {
 
621
                                             V T6q, T6H, T5C, T5s, T6I, T6x;
 
622
                                             T6q = VFNMS(LDK(KP534511135), T6p, T6m);
 
623
                                             T6H = VFMA(LDK(KP534511135), T6m, T6p);
 
624
                                             T5C = VFNMS(LDK(KP980785280), T5r, T5o);
 
625
                                             T5s = VFMA(LDK(KP980785280), T5r, T5o);
 
626
                                             T6I = VFMA(LDK(KP534511135), T6t, T6w);
 
627
                                             T6x = VFNMS(LDK(KP534511135), T6w, T6t);
 
628
                                             T6N = VFMA(LDK(KP831469612), T6i, T6b);
 
629
                                             T6j = VFNMS(LDK(KP831469612), T6i, T6b);
 
630
                                             {
 
631
                                                  V T5E, T5G, T5y, T5w;
 
632
                                                  T5E = VFNMS(LDK(KP995184726), T5D, T5C);
 
633
                                                  T5G = VFMA(LDK(KP995184726), T5D, T5C);
 
634
                                                  T5y = VFMA(LDK(KP995184726), T5v, T5s);
 
635
                                                  T5w = VFNMS(LDK(KP995184726), T5v, T5s);
 
636
                                                  T6O = VADD(T6H, T6I);
 
637
                                                  T6J = VSUB(T6H, T6I);
 
638
                                                  T6R = VSUB(T6x, T6q);
 
639
                                                  T6y = VADD(T6q, T6x);
 
640
                                                  ST(&(xo[WS(os, 47)]), VFMAI(T5E, T5B), ovs, &(xo[WS(os, 1)]));
 
641
                                                  ST(&(xo[WS(os, 17)]), VFNMSI(T5E, T5B), ovs, &(xo[WS(os, 1)]));
 
642
                                                  ST(&(xo[WS(os, 49)]), VFNMSI(T5G, T5F), ovs, &(xo[WS(os, 1)]));
 
643
                                                  ST(&(xo[WS(os, 15)]), VFMAI(T5G, T5F), ovs, &(xo[WS(os, 1)]));
 
644
                                                  ST(&(xo[WS(os, 63)]), VFMAI(T5y, T5x), ovs, &(xo[WS(os, 1)]));
 
645
                                                  ST(&(xo[WS(os, 1)]), VFNMSI(T5y, T5x), ovs, &(xo[WS(os, 1)]));
 
646
                                                  ST(&(xo[WS(os, 31)]), VFMAI(T5w, T5h), ovs, &(xo[WS(os, 1)]));
 
647
                                                  ST(&(xo[WS(os, 33)]), VFNMSI(T5w, T5h), ovs, &(xo[WS(os, 1)]));
 
648
                                             }
 
649
                                        }
 
650
                                        T6L = VFMA(LDK(KP881921264), T6y, T6j);
 
651
                                        T6z = VFNMS(LDK(KP881921264), T6y, T6j);
 
652
                                        T6T = VFMA(LDK(KP881921264), T6O, T6N);
 
653
                                        T6P = VFNMS(LDK(KP881921264), T6O, T6N);
 
654
                                   }
 
655
                                   {
 
656
                                        V T2H, T2P, T2J, T2B, T2R, T2N;
 
657
                                        {
 
658
                                             V T2L, T2t, T2M, T2A;
 
659
                                             {
 
660
                                                  V T2z, T2F, T6Q, T6G, T2G, T2w;
 
661
                                                  T2z = VFMA(LDK(KP668178637), T2y, T2x);
 
662
                                                  T2F = VFNMS(LDK(KP668178637), T2x, T2y);
 
663
                                                  T6Q = VFMA(LDK(KP831469612), T6F, T6C);
 
664
                                                  T6G = VFNMS(LDK(KP831469612), T6F, T6C);
 
665
                                                  T2G = VFNMS(LDK(KP668178637), T2u, T2v);
 
666
                                                  T2w = VFMA(LDK(KP668178637), T2v, T2u);
 
667
                                                  T2L = VFNMS(LDK(KP923879532), T2s, T2r);
 
668
                                                  T2t = VFMA(LDK(KP923879532), T2s, T2r);
 
669
                                                  {
 
670
                                                       V T6S, T6U, T6M, T6K;
 
671
                                                       T6S = VFNMS(LDK(KP881921264), T6R, T6Q);
 
672
                                                       T6U = VFMA(LDK(KP881921264), T6R, T6Q);
 
673
                                                       T6M = VFMA(LDK(KP881921264), T6J, T6G);
 
674
                                                       T6K = VFNMS(LDK(KP881921264), T6J, T6G);
 
675
                                                       T2M = VADD(T2G, T2F);
 
676
                                                       T2H = VSUB(T2F, T2G);
 
677
                                                       T2P = VSUB(T2z, T2w);
 
678
                                                       T2A = VADD(T2w, T2z);
 
679
                                                       ST(&(xo[WS(os, 43)]), VFMAI(T6S, T6P), ovs, &(xo[WS(os, 1)]));
 
680
                                                       ST(&(xo[WS(os, 21)]), VFNMSI(T6S, T6P), ovs, &(xo[WS(os, 1)]));
 
681
                                                       ST(&(xo[WS(os, 53)]), VFNMSI(T6U, T6T), ovs, &(xo[WS(os, 1)]));
 
682
                                                       ST(&(xo[WS(os, 11)]), VFMAI(T6U, T6T), ovs, &(xo[WS(os, 1)]));
 
683
                                                       ST(&(xo[WS(os, 59)]), VFMAI(T6M, T6L), ovs, &(xo[WS(os, 1)]));
 
684
                                                       ST(&(xo[WS(os, 5)]), VFNMSI(T6M, T6L), ovs, &(xo[WS(os, 1)]));
 
685
                                                       ST(&(xo[WS(os, 27)]), VFMAI(T6K, T6z), ovs, &(xo[WS(os, 1)]));
 
686
                                                       ST(&(xo[WS(os, 37)]), VFNMSI(T6K, T6z), ovs, &(xo[WS(os, 1)]));
 
687
                                                  }
 
688
                                             }
 
689
                                             T2J = VFMA(LDK(KP831469612), T2A, T2t);
 
690
                                             T2B = VFNMS(LDK(KP831469612), T2A, T2t);
 
691
                                             T2R = VFNMS(LDK(KP831469612), T2M, T2L);
 
692
                                             T2N = VFMA(LDK(KP831469612), T2M, T2L);
 
693
                                        }
 
694
                                        {
 
695
                                             V T61, T5J, T62, T5Q;
 
696
                                             {
 
697
                                                  V T5M, T5V, T2O, T2E, T5W, T5P;
 
698
                                                  T5M = VFMA(LDK(KP820678790), T5L, T5K);
 
699
                                                  T5V = VFNMS(LDK(KP820678790), T5K, T5L);
 
700
                                                  T2O = VFMA(LDK(KP923879532), T2D, T2C);
 
701
                                                  T2E = VFNMS(LDK(KP923879532), T2D, T2C);
 
702
                                                  T5W = VFNMS(LDK(KP820678790), T5N, T5O);
 
703
                                                  T5P = VFMA(LDK(KP820678790), T5O, T5N);
 
704
                                                  T61 = VFNMS(LDK(KP980785280), T5I, T5H);
 
705
                                                  T5J = VFMA(LDK(KP980785280), T5I, T5H);
 
706
                                                  {
 
707
                                                       V T2Q, T2S, T2K, T2I;
 
708
                                                       T2Q = VFNMS(LDK(KP831469612), T2P, T2O);
 
709
                                                       T2S = VFMA(LDK(KP831469612), T2P, T2O);
 
710
                                                       T2K = VFMA(LDK(KP831469612), T2H, T2E);
 
711
                                                       T2I = VFNMS(LDK(KP831469612), T2H, T2E);
 
712
                                                       T62 = VADD(T5V, T5W);
 
713
                                                       T5X = VSUB(T5V, T5W);
 
714
                                                       T65 = VSUB(T5P, T5M);
 
715
                                                       T5Q = VADD(T5M, T5P);
 
716
                                                       ST(&(xo[WS(os, 42)]), VFMAI(T2Q, T2N), ovs, &(xo[0]));
 
717
                                                       ST(&(xo[WS(os, 22)]), VFNMSI(T2Q, T2N), ovs, &(xo[0]));
 
718
                                                       ST(&(xo[WS(os, 54)]), VFNMSI(T2S, T2R), ovs, &(xo[0]));
 
719
                                                       ST(&(xo[WS(os, 10)]), VFMAI(T2S, T2R), ovs, &(xo[0]));
 
720
                                                       ST(&(xo[WS(os, 58)]), VFMAI(T2K, T2J), ovs, &(xo[0]));
 
721
                                                       ST(&(xo[WS(os, 6)]), VFNMSI(T2K, T2J), ovs, &(xo[0]));
 
722
                                                       ST(&(xo[WS(os, 26)]), VFMAI(T2I, T2B), ovs, &(xo[0]));
 
723
                                                       ST(&(xo[WS(os, 38)]), VFNMSI(T2I, T2B), ovs, &(xo[0]));
 
724
                                                  }
 
725
                                             }
 
726
                                             T5Z = VFMA(LDK(KP773010453), T5Q, T5J);
 
727
                                             T5R = VFNMS(LDK(KP773010453), T5Q, T5J);
 
728
                                             T67 = VFNMS(LDK(KP773010453), T62, T61);
 
729
                                             T63 = VFMA(LDK(KP773010453), T62, T61);
 
730
                                        }
 
731
                                   }
 
732
                              }
 
733
                         }
 
734
                    }
 
735
               }
 
736
               T5U = VFMA(LDK(KP980785280), T5T, T5S);
 
737
               T64 = VFNMS(LDK(KP980785280), T5T, T5S);
 
738
               {
 
739
                    V T68, T66, T5Y, T60;
 
740
                    T68 = VFNMS(LDK(KP773010453), T65, T64);
 
741
                    T66 = VFMA(LDK(KP773010453), T65, T64);
 
742
                    T5Y = VFNMS(LDK(KP773010453), T5X, T5U);
 
743
                    T60 = VFMA(LDK(KP773010453), T5X, T5U);
 
744
                    ST(&(xo[WS(os, 41)]), VFNMSI(T66, T63), ovs, &(xo[WS(os, 1)]));
 
745
                    ST(&(xo[WS(os, 23)]), VFMAI(T66, T63), ovs, &(xo[WS(os, 1)]));
 
746
                    ST(&(xo[WS(os, 55)]), VFMAI(T68, T67), ovs, &(xo[WS(os, 1)]));
 
747
                    ST(&(xo[WS(os, 9)]), VFNMSI(T68, T67), ovs, &(xo[WS(os, 1)]));
 
748
                    ST(&(xo[WS(os, 7)]), VFMAI(T60, T5Z), ovs, &(xo[WS(os, 1)]));
 
749
                    ST(&(xo[WS(os, 57)]), VFNMSI(T60, T5Z), ovs, &(xo[WS(os, 1)]));
 
750
                    ST(&(xo[WS(os, 39)]), VFMAI(T5Y, T5R), ovs, &(xo[WS(os, 1)]));
 
751
                    ST(&(xo[WS(os, 25)]), VFNMSI(T5Y, T5R), ovs, &(xo[WS(os, 1)]));
 
752
               }
 
753
          }
 
754
     }
 
755
     VLEAVE();
 
756
}
 
757
 
 
758
static const kdft_desc desc = { 64, XSIMD_STRING("n1fv_64"), {198, 0, 258, 0}, &GENUS, 0, 0, 0, 0 };
 
759
 
 
760
void XSIMD(codelet_n1fv_64) (planner *p) {
 
761
     X(kdft_register) (p, n1fv_64, &desc);
 
762
}
 
763
 
 
764
#else                           /* HAVE_FMA */
 
765
 
 
766
/* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n1fv_64 -include n1f.h */
 
767
 
 
768
/*
 
769
 * This function contains 456 FP additions, 124 FP multiplications,
 
770
 * (or, 404 additions, 72 multiplications, 52 fused multiply/add),
 
771
 * 108 stack variables, 15 constants, and 128 memory accesses
 
772
 */
 
773
#include "n1f.h"
 
774
 
 
775
static void n1fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
 
776
{
 
777
     DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
 
778
     DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
 
779
     DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
 
780
     DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
 
781
     DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
 
782
     DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
 
783
     DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
 
784
     DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
 
785
     DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
 
786
     DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
 
787
     DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
 
788
     DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
 
789
     DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
 
790
     DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
791
     DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
792
     {
 
793
          INT i;
 
794
          const R *xi;
 
795
          R *xo;
 
796
          xi = ri;
 
797
          xo = ro;
 
798
          for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
 
799
               V T4p, T5q, Tb, T39, T2n, T3A, T6f, T6T, Tq, T3B, T6i, T76, T2i, T3a, T4w;
 
800
               V T5r, TI, T2p, T6C, T6V, T3h, T3E, T4L, T5u, TZ, T2q, T6F, T6U, T3e, T3D;
 
801
               V T4E, T5t, T23, T2N, T6t, T71, T6w, T72, T2c, T2O, T3t, T41, T5f, T5R, T5k;
 
802
               V T5S, T3w, T42, T1s, T2K, T6m, T6Y, T6p, T6Z, T1B, T2L, T3m, T3Y, T4Y, T5O;
 
803
               V T53, T5P, T3p, T3Z;
 
804
               {
 
805
                    V T3, T4n, T2m, T4o, T6, T5p, T9, T5o;
 
806
                    {
 
807
                         V T1, T2, T2k, T2l;
 
808
                         T1 = LD(&(xi[0]), ivs, &(xi[0]));
 
809
                         T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
 
810
                         T3 = VSUB(T1, T2);
 
811
                         T4n = VADD(T1, T2);
 
812
                         T2k = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
 
813
                         T2l = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
 
814
                         T2m = VSUB(T2k, T2l);
 
815
                         T4o = VADD(T2k, T2l);
 
816
                    }
 
817
                    {
 
818
                         V T4, T5, T7, T8;
 
819
                         T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
 
820
                         T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
 
821
                         T6 = VSUB(T4, T5);
 
822
                         T5p = VADD(T4, T5);
 
823
                         T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
 
824
                         T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
 
825
                         T9 = VSUB(T7, T8);
 
826
                         T5o = VADD(T7, T8);
 
827
                    }
 
828
                    T4p = VSUB(T4n, T4o);
 
829
                    T5q = VSUB(T5o, T5p);
 
830
                    {
 
831
                         V Ta, T2j, T6d, T6e;
 
832
                         Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
 
833
                         Tb = VADD(T3, Ta);
 
834
                         T39 = VSUB(T3, Ta);
 
835
                         T2j = VMUL(LDK(KP707106781), VSUB(T9, T6));
 
836
                         T2n = VSUB(T2j, T2m);
 
837
                         T3A = VADD(T2m, T2j);
 
838
                         T6d = VADD(T4n, T4o);
 
839
                         T6e = VADD(T5p, T5o);
 
840
                         T6f = VADD(T6d, T6e);
 
841
                         T6T = VSUB(T6d, T6e);
 
842
                    }
 
843
               }
 
844
               {
 
845
                    V Te, T4q, To, T4u, Th, T4r, Tl, T4t;
 
846
                    {
 
847
                         V Tc, Td, Tm, Tn;
 
848
                         Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
 
849
                         Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
 
850
                         Te = VSUB(Tc, Td);
 
851
                         T4q = VADD(Tc, Td);
 
852
                         Tm = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
 
853
                         Tn = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
 
854
                         To = VSUB(Tm, Tn);
 
855
                         T4u = VADD(Tm, Tn);
 
856
                    }
 
857
                    {
 
858
                         V Tf, Tg, Tj, Tk;
 
859
                         Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
 
860
                         Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
 
861
                         Th = VSUB(Tf, Tg);
 
862
                         T4r = VADD(Tf, Tg);
 
863
                         Tj = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
 
864
                         Tk = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
 
865
                         Tl = VSUB(Tj, Tk);
 
866
                         T4t = VADD(Tj, Tk);
 
867
                    }
 
868
                    {
 
869
                         V Ti, Tp, T6g, T6h;
 
870
                         Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
 
871
                         Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
 
872
                         Tq = VADD(Ti, Tp);
 
873
                         T3B = VSUB(Tp, Ti);
 
874
                         T6g = VADD(T4q, T4r);
 
875
                         T6h = VADD(T4t, T4u);
 
876
                         T6i = VADD(T6g, T6h);
 
877
                         T76 = VSUB(T6h, T6g);
 
878
                    }
 
879
                    {
 
880
                         V T2g, T2h, T4s, T4v;
 
881
                         T2g = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
 
882
                         T2h = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
 
883
                         T2i = VSUB(T2g, T2h);
 
884
                         T3a = VADD(T2h, T2g);
 
885
                         T4s = VSUB(T4q, T4r);
 
886
                         T4v = VSUB(T4t, T4u);
 
887
                         T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v));
 
888
                         T5r = VMUL(LDK(KP707106781), VSUB(T4v, T4s));
 
889
                    }
 
890
               }
 
891
               {
 
892
                    V Tu, T4F, TG, T4G, TB, T4J, TD, T4I;
 
893
                    {
 
894
                         V Ts, Tt, TE, TF;
 
895
                         Ts = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
 
896
                         Tt = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
 
897
                         Tu = VSUB(Ts, Tt);
 
898
                         T4F = VADD(Ts, Tt);
 
899
                         TE = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
 
900
                         TF = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
 
901
                         TG = VSUB(TE, TF);
 
902
                         T4G = VADD(TE, TF);
 
903
                         {
 
904
                              V Tv, Tw, Tx, Ty, Tz, TA;
 
905
                              Tv = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
 
906
                              Tw = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
 
907
                              Tx = VSUB(Tv, Tw);
 
908
                              Ty = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
 
909
                              Tz = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
 
910
                              TA = VSUB(Ty, Tz);
 
911
                              TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
 
912
                              T4J = VADD(Tv, Tw);
 
913
                              TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
 
914
                              T4I = VADD(Ty, Tz);
 
915
                         }
 
916
                    }
 
917
                    {
 
918
                         V TC, TH, T6A, T6B;
 
919
                         TC = VADD(Tu, TB);
 
920
                         TH = VSUB(TD, TG);
 
921
                         TI = VFMA(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
 
922
                         T2p = VFNMS(LDK(KP195090322), TH, VMUL(LDK(KP980785280), TC));
 
923
                         T6A = VADD(T4F, T4G);
 
924
                         T6B = VADD(T4J, T4I);
 
925
                         T6C = VADD(T6A, T6B);
 
926
                         T6V = VSUB(T6A, T6B);
 
927
                    }
 
928
                    {
 
929
                         V T3f, T3g, T4H, T4K;
 
930
                         T3f = VSUB(Tu, TB);
 
931
                         T3g = VADD(TG, TD);
 
932
                         T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f));
 
933
                         T3E = VFMA(LDK(KP555570233), T3f, VMUL(LDK(KP831469612), T3g));
 
934
                         T4H = VSUB(T4F, T4G);
 
935
                         T4K = VSUB(T4I, T4J);
 
936
                         T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H));
 
937
                         T5u = VFMA(LDK(KP382683432), T4H, VMUL(LDK(KP923879532), T4K));
 
938
                    }
 
939
               }
 
940
               {
 
941
                    V TS, T4z, TW, T4y, TP, T4C, TX, T4B;
 
942
                    {
 
943
                         V TQ, TR, TU, TV;
 
944
                         TQ = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
 
945
                         TR = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
 
946
                         TS = VSUB(TQ, TR);
 
947
                         T4z = VADD(TQ, TR);
 
948
                         TU = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
 
949
                         TV = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
 
950
                         TW = VSUB(TU, TV);
 
951
                         T4y = VADD(TU, TV);
 
952
                         {
 
953
                              V TJ, TK, TL, TM, TN, TO;
 
954
                              TJ = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
 
955
                              TK = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
 
956
                              TL = VSUB(TJ, TK);
 
957
                              TM = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
 
958
                              TN = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
 
959
                              TO = VSUB(TM, TN);
 
960
                              TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
 
961
                              T4C = VADD(TM, TN);
 
962
                              TX = VMUL(LDK(KP707106781), VADD(TO, TL));
 
963
                              T4B = VADD(TJ, TK);
 
964
                         }
 
965
                    }
 
966
                    {
 
967
                         V TT, TY, T6D, T6E;
 
968
                         TT = VSUB(TP, TS);
 
969
                         TY = VADD(TW, TX);
 
970
                         TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
 
971
                         T2q = VFMA(LDK(KP980785280), TY, VMUL(LDK(KP195090322), TT));
 
972
                         T6D = VADD(T4y, T4z);
 
973
                         T6E = VADD(T4C, T4B);
 
974
                         T6F = VADD(T6D, T6E);
 
975
                         T6U = VSUB(T6D, T6E);
 
976
                    }
 
977
                    {
 
978
                         V T3c, T3d, T4A, T4D;
 
979
                         T3c = VSUB(TW, TX);
 
980
                         T3d = VADD(TS, TP);
 
981
                         T3e = VFMA(LDK(KP831469612), T3c, VMUL(LDK(KP555570233), T3d));
 
982
                         T3D = VFNMS(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d));
 
983
                         T4A = VSUB(T4y, T4z);
 
984
                         T4D = VSUB(T4B, T4C);
 
985
                         T4E = VFMA(LDK(KP923879532), T4A, VMUL(LDK(KP382683432), T4D));
 
986
                         T5t = VFNMS(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D));
 
987
                    }
 
988
               }
 
989
               {
 
990
                    V T1F, T55, T2a, T56, T1M, T5h, T27, T5g, T58, T59, T1U, T5a, T25, T5b, T5c;
 
991
                    V T21, T5d, T24;
 
992
                    {
 
993
                         V T1D, T1E, T28, T29;
 
994
                         T1D = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
 
995
                         T1E = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
 
996
                         T1F = VSUB(T1D, T1E);
 
997
                         T55 = VADD(T1D, T1E);
 
998
                         T28 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
 
999
                         T29 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
 
1000
                         T2a = VSUB(T28, T29);
 
1001
                         T56 = VADD(T28, T29);
 
1002
                    }
 
1003
                    {
 
1004
                         V T1G, T1H, T1I, T1J, T1K, T1L;
 
1005
                         T1G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
 
1006
                         T1H = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
 
1007
                         T1I = VSUB(T1G, T1H);
 
1008
                         T1J = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
 
1009
                         T1K = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
 
1010
                         T1L = VSUB(T1J, T1K);
 
1011
                         T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L));
 
1012
                         T5h = VADD(T1G, T1H);
 
1013
                         T27 = VMUL(LDK(KP707106781), VSUB(T1L, T1I));
 
1014
                         T5g = VADD(T1J, T1K);
 
1015
                    }
 
1016
                    {
 
1017
                         V T1Q, T1T, T1X, T20;
 
1018
                         {
 
1019
                              V T1O, T1P, T1R, T1S;
 
1020
                              T1O = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
 
1021
                              T1P = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
 
1022
                              T1Q = VSUB(T1O, T1P);
 
1023
                              T58 = VADD(T1O, T1P);
 
1024
                              T1R = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
 
1025
                              T1S = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
 
1026
                              T1T = VSUB(T1R, T1S);
 
1027
                              T59 = VADD(T1R, T1S);
 
1028
                         }
 
1029
                         T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q));
 
1030
                         T5a = VSUB(T58, T59);
 
1031
                         T25 = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T));
 
1032
                         {
 
1033
                              V T1V, T1W, T1Y, T1Z;
 
1034
                              T1V = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
 
1035
                              T1W = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
 
1036
                              T1X = VSUB(T1V, T1W);
 
1037
                              T5b = VADD(T1V, T1W);
 
1038
                              T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
 
1039
                              T1Z = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
 
1040
                              T20 = VSUB(T1Y, T1Z);
 
1041
                              T5c = VADD(T1Y, T1Z);
 
1042
                         }
 
1043
                         T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20));
 
1044
                         T5d = VSUB(T5b, T5c);
 
1045
                         T24 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X));
 
1046
                    }
 
1047
                    {
 
1048
                         V T1N, T22, T6r, T6s;
 
1049
                         T1N = VADD(T1F, T1M);
 
1050
                         T22 = VADD(T1U, T21);
 
1051
                         T23 = VSUB(T1N, T22);
 
1052
                         T2N = VADD(T1N, T22);
 
1053
                         T6r = VADD(T55, T56);
 
1054
                         T6s = VADD(T5h, T5g);
 
1055
                         T6t = VADD(T6r, T6s);
 
1056
                         T71 = VSUB(T6r, T6s);
 
1057
                    }
 
1058
                    {
 
1059
                         V T6u, T6v, T26, T2b;
 
1060
                         T6u = VADD(T58, T59);
 
1061
                         T6v = VADD(T5b, T5c);
 
1062
                         T6w = VADD(T6u, T6v);
 
1063
                         T72 = VSUB(T6v, T6u);
 
1064
                         T26 = VSUB(T24, T25);
 
1065
                         T2b = VSUB(T27, T2a);
 
1066
                         T2c = VSUB(T26, T2b);
 
1067
                         T2O = VADD(T2b, T26);
 
1068
                    }
 
1069
                    {
 
1070
                         V T3r, T3s, T57, T5e;
 
1071
                         T3r = VSUB(T1F, T1M);
 
1072
                         T3s = VADD(T25, T24);
 
1073
                         T3t = VADD(T3r, T3s);
 
1074
                         T41 = VSUB(T3r, T3s);
 
1075
                         T57 = VSUB(T55, T56);
 
1076
                         T5e = VMUL(LDK(KP707106781), VADD(T5a, T5d));
 
1077
                         T5f = VADD(T57, T5e);
 
1078
                         T5R = VSUB(T57, T5e);
 
1079
                    }
 
1080
                    {
 
1081
                         V T5i, T5j, T3u, T3v;
 
1082
                         T5i = VSUB(T5g, T5h);
 
1083
                         T5j = VMUL(LDK(KP707106781), VSUB(T5d, T5a));
 
1084
                         T5k = VADD(T5i, T5j);
 
1085
                         T5S = VSUB(T5j, T5i);
 
1086
                         T3u = VADD(T2a, T27);
 
1087
                         T3v = VSUB(T21, T1U);
 
1088
                         T3w = VADD(T3u, T3v);
 
1089
                         T42 = VSUB(T3v, T3u);
 
1090
                    }
 
1091
               }
 
1092
               {
 
1093
                    V T1q, T4P, T1v, T4O, T1n, T50, T1w, T4Z, T4U, T4V, T18, T4W, T1z, T4R, T4S;
 
1094
                    V T1f, T4T, T1y;
 
1095
                    {
 
1096
                         V T1o, T1p, T1t, T1u;
 
1097
                         T1o = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
 
1098
                         T1p = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
 
1099
                         T1q = VSUB(T1o, T1p);
 
1100
                         T4P = VADD(T1o, T1p);
 
1101
                         T1t = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
 
1102
                         T1u = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
 
1103
                         T1v = VSUB(T1t, T1u);
 
1104
                         T4O = VADD(T1t, T1u);
 
1105
                    }
 
1106
                    {
 
1107
                         V T1h, T1i, T1j, T1k, T1l, T1m;
 
1108
                         T1h = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
 
1109
                         T1i = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
 
1110
                         T1j = VSUB(T1h, T1i);
 
1111
                         T1k = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
 
1112
                         T1l = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
 
1113
                         T1m = VSUB(T1k, T1l);
 
1114
                         T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m));
 
1115
                         T50 = VADD(T1k, T1l);
 
1116
                         T1w = VMUL(LDK(KP707106781), VADD(T1m, T1j));
 
1117
                         T4Z = VADD(T1h, T1i);
 
1118
                    }
 
1119
                    {
 
1120
                         V T14, T17, T1b, T1e;
 
1121
                         {
 
1122
                              V T12, T13, T15, T16;
 
1123
                              T12 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
 
1124
                              T13 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
 
1125
                              T14 = VSUB(T12, T13);
 
1126
                              T4U = VADD(T12, T13);
 
1127
                              T15 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
 
1128
                              T16 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
 
1129
                              T17 = VSUB(T15, T16);
 
1130
                              T4V = VADD(T15, T16);
 
1131
                         }
 
1132
                         T18 = VFNMS(LDK(KP923879532), T17, VMUL(LDK(KP382683432), T14));
 
1133
                         T4W = VSUB(T4U, T4V);
 
1134
                         T1z = VFMA(LDK(KP923879532), T14, VMUL(LDK(KP382683432), T17));
 
1135
                         {
 
1136
                              V T19, T1a, T1c, T1d;
 
1137
                              T19 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
 
1138
                              T1a = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
 
1139
                              T1b = VSUB(T19, T1a);
 
1140
                              T4R = VADD(T19, T1a);
 
1141
                              T1c = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
 
1142
                              T1d = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
 
1143
                              T1e = VSUB(T1c, T1d);
 
1144
                              T4S = VADD(T1c, T1d);
 
1145
                         }
 
1146
                         T1f = VFMA(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e));
 
1147
                         T4T = VSUB(T4R, T4S);
 
1148
                         T1y = VFNMS(LDK(KP382683432), T1e, VMUL(LDK(KP923879532), T1b));
 
1149
                    }
 
1150
                    {
 
1151
                         V T1g, T1r, T6k, T6l;
 
1152
                         T1g = VSUB(T18, T1f);
 
1153
                         T1r = VSUB(T1n, T1q);
 
1154
                         T1s = VSUB(T1g, T1r);
 
1155
                         T2K = VADD(T1r, T1g);
 
1156
                         T6k = VADD(T4O, T4P);
 
1157
                         T6l = VADD(T50, T4Z);
 
1158
                         T6m = VADD(T6k, T6l);
 
1159
                         T6Y = VSUB(T6k, T6l);
 
1160
                    }
 
1161
                    {
 
1162
                         V T6n, T6o, T1x, T1A;
 
1163
                         T6n = VADD(T4R, T4S);
 
1164
                         T6o = VADD(T4U, T4V);
 
1165
                         T6p = VADD(T6n, T6o);
 
1166
                         T6Z = VSUB(T6o, T6n);
 
1167
                         T1x = VADD(T1v, T1w);
 
1168
                         T1A = VADD(T1y, T1z);
 
1169
                         T1B = VSUB(T1x, T1A);
 
1170
                         T2L = VADD(T1x, T1A);
 
1171
                    }
 
1172
                    {
 
1173
                         V T3k, T3l, T4Q, T4X;
 
1174
                         T3k = VSUB(T1v, T1w);
 
1175
                         T3l = VADD(T1f, T18);
 
1176
                         T3m = VADD(T3k, T3l);
 
1177
                         T3Y = VSUB(T3k, T3l);
 
1178
                         T4Q = VSUB(T4O, T4P);
 
1179
                         T4X = VMUL(LDK(KP707106781), VADD(T4T, T4W));
 
1180
                         T4Y = VADD(T4Q, T4X);
 
1181
                         T5O = VSUB(T4Q, T4X);
 
1182
                    }
 
1183
                    {
 
1184
                         V T51, T52, T3n, T3o;
 
1185
                         T51 = VSUB(T4Z, T50);
 
1186
                         T52 = VMUL(LDK(KP707106781), VSUB(T4W, T4T));
 
1187
                         T53 = VADD(T51, T52);
 
1188
                         T5P = VSUB(T52, T51);
 
1189
                         T3n = VADD(T1q, T1n);
 
1190
                         T3o = VSUB(T1z, T1y);
 
1191
                         T3p = VADD(T3n, T3o);
 
1192
                         T3Z = VSUB(T3o, T3n);
 
1193
                    }
 
1194
               }
 
1195
               {
 
1196
                    V T6N, T6R, T6Q, T6S;
 
1197
                    {
 
1198
                         V T6L, T6M, T6O, T6P;
 
1199
                         T6L = VADD(T6f, T6i);
 
1200
                         T6M = VADD(T6F, T6C);
 
1201
                         T6N = VADD(T6L, T6M);
 
1202
                         T6R = VSUB(T6L, T6M);
 
1203
                         T6O = VADD(T6m, T6p);
 
1204
                         T6P = VADD(T6t, T6w);
 
1205
                         T6Q = VADD(T6O, T6P);
 
1206
                         T6S = VBYI(VSUB(T6P, T6O));
 
1207
                    }
 
1208
                    ST(&(xo[WS(os, 32)]), VSUB(T6N, T6Q), ovs, &(xo[0]));
 
1209
                    ST(&(xo[WS(os, 16)]), VADD(T6R, T6S), ovs, &(xo[0]));
 
1210
                    ST(&(xo[0]), VADD(T6N, T6Q), ovs, &(xo[0]));
 
1211
                    ST(&(xo[WS(os, 48)]), VSUB(T6R, T6S), ovs, &(xo[0]));
 
1212
               }
 
1213
               {
 
1214
                    V T6j, T6G, T6y, T6H, T6q, T6x;
 
1215
                    T6j = VSUB(T6f, T6i);
 
1216
                    T6G = VSUB(T6C, T6F);
 
1217
                    T6q = VSUB(T6m, T6p);
 
1218
                    T6x = VSUB(T6t, T6w);
 
1219
                    T6y = VMUL(LDK(KP707106781), VADD(T6q, T6x));
 
1220
                    T6H = VMUL(LDK(KP707106781), VSUB(T6x, T6q));
 
1221
                    {
 
1222
                         V T6z, T6I, T6J, T6K;
 
1223
                         T6z = VADD(T6j, T6y);
 
1224
                         T6I = VBYI(VADD(T6G, T6H));
 
1225
                         ST(&(xo[WS(os, 56)]), VSUB(T6z, T6I), ovs, &(xo[0]));
 
1226
                         ST(&(xo[WS(os, 8)]), VADD(T6z, T6I), ovs, &(xo[0]));
 
1227
                         T6J = VSUB(T6j, T6y);
 
1228
                         T6K = VBYI(VSUB(T6H, T6G));
 
1229
                         ST(&(xo[WS(os, 40)]), VSUB(T6J, T6K), ovs, &(xo[0]));
 
1230
                         ST(&(xo[WS(os, 24)]), VADD(T6J, T6K), ovs, &(xo[0]));
 
1231
                    }
 
1232
               }
 
1233
               {
 
1234
                    V T6X, T7i, T78, T7g, T74, T7f, T7b, T7j, T6W, T77;
 
1235
                    T6W = VMUL(LDK(KP707106781), VADD(T6U, T6V));
 
1236
                    T6X = VADD(T6T, T6W);
 
1237
                    T7i = VSUB(T6T, T6W);
 
1238
                    T77 = VMUL(LDK(KP707106781), VSUB(T6V, T6U));
 
1239
                    T78 = VADD(T76, T77);
 
1240
                    T7g = VSUB(T77, T76);
 
1241
                    {
 
1242
                         V T70, T73, T79, T7a;
 
1243
                         T70 = VFMA(LDK(KP923879532), T6Y, VMUL(LDK(KP382683432), T6Z));
 
1244
                         T73 = VFNMS(LDK(KP382683432), T72, VMUL(LDK(KP923879532), T71));
 
1245
                         T74 = VADD(T70, T73);
 
1246
                         T7f = VSUB(T73, T70);
 
1247
                         T79 = VFNMS(LDK(KP382683432), T6Y, VMUL(LDK(KP923879532), T6Z));
 
1248
                         T7a = VFMA(LDK(KP382683432), T71, VMUL(LDK(KP923879532), T72));
 
1249
                         T7b = VADD(T79, T7a);
 
1250
                         T7j = VSUB(T7a, T79);
 
1251
                    }
 
1252
                    {
 
1253
                         V T75, T7c, T7l, T7m;
 
1254
                         T75 = VADD(T6X, T74);
 
1255
                         T7c = VBYI(VADD(T78, T7b));
 
1256
                         ST(&(xo[WS(os, 60)]), VSUB(T75, T7c), ovs, &(xo[0]));
 
1257
                         ST(&(xo[WS(os, 4)]), VADD(T75, T7c), ovs, &(xo[0]));
 
1258
                         T7l = VBYI(VADD(T7g, T7f));
 
1259
                         T7m = VADD(T7i, T7j);
 
1260
                         ST(&(xo[WS(os, 12)]), VADD(T7l, T7m), ovs, &(xo[0]));
 
1261
                         ST(&(xo[WS(os, 52)]), VSUB(T7m, T7l), ovs, &(xo[0]));
 
1262
                    }
 
1263
                    {
 
1264
                         V T7d, T7e, T7h, T7k;
 
1265
                         T7d = VSUB(T6X, T74);
 
1266
                         T7e = VBYI(VSUB(T7b, T78));
 
1267
                         ST(&(xo[WS(os, 36)]), VSUB(T7d, T7e), ovs, &(xo[0]));
 
1268
                         ST(&(xo[WS(os, 28)]), VADD(T7d, T7e), ovs, &(xo[0]));
 
1269
                         T7h = VBYI(VSUB(T7f, T7g));
 
1270
                         T7k = VSUB(T7i, T7j);
 
1271
                         ST(&(xo[WS(os, 20)]), VADD(T7h, T7k), ovs, &(xo[0]));
 
1272
                         ST(&(xo[WS(os, 44)]), VSUB(T7k, T7h), ovs, &(xo[0]));
 
1273
                    }
 
1274
               }
 
1275
               {
 
1276
                    V T5N, T68, T61, T69, T5U, T65, T5Y, T66;
 
1277
                    {
 
1278
                         V T5L, T5M, T5Z, T60;
 
1279
                         T5L = VSUB(T4p, T4w);
 
1280
                         T5M = VSUB(T5u, T5t);
 
1281
                         T5N = VADD(T5L, T5M);
 
1282
                         T68 = VSUB(T5L, T5M);
 
1283
                         T5Z = VFNMS(LDK(KP555570233), T5O, VMUL(LDK(KP831469612), T5P));
 
1284
                         T60 = VFMA(LDK(KP555570233), T5R, VMUL(LDK(KP831469612), T5S));
 
1285
                         T61 = VADD(T5Z, T60);
 
1286
                         T69 = VSUB(T60, T5Z);
 
1287
                    }
 
1288
                    {
 
1289
                         V T5Q, T5T, T5W, T5X;
 
1290
                         T5Q = VFMA(LDK(KP831469612), T5O, VMUL(LDK(KP555570233), T5P));
 
1291
                         T5T = VFNMS(LDK(KP555570233), T5S, VMUL(LDK(KP831469612), T5R));
 
1292
                         T5U = VADD(T5Q, T5T);
 
1293
                         T65 = VSUB(T5T, T5Q);
 
1294
                         T5W = VSUB(T5r, T5q);
 
1295
                         T5X = VSUB(T4L, T4E);
 
1296
                         T5Y = VADD(T5W, T5X);
 
1297
                         T66 = VSUB(T5X, T5W);
 
1298
                    }
 
1299
                    {
 
1300
                         V T5V, T62, T6b, T6c;
 
1301
                         T5V = VADD(T5N, T5U);
 
1302
                         T62 = VBYI(VADD(T5Y, T61));
 
1303
                         ST(&(xo[WS(os, 58)]), VSUB(T5V, T62), ovs, &(xo[0]));
 
1304
                         ST(&(xo[WS(os, 6)]), VADD(T5V, T62), ovs, &(xo[0]));
 
1305
                         T6b = VBYI(VADD(T66, T65));
 
1306
                         T6c = VADD(T68, T69);
 
1307
                         ST(&(xo[WS(os, 10)]), VADD(T6b, T6c), ovs, &(xo[0]));
 
1308
                         ST(&(xo[WS(os, 54)]), VSUB(T6c, T6b), ovs, &(xo[0]));
 
1309
                    }
 
1310
                    {
 
1311
                         V T63, T64, T67, T6a;
 
1312
                         T63 = VSUB(T5N, T5U);
 
1313
                         T64 = VBYI(VSUB(T61, T5Y));
 
1314
                         ST(&(xo[WS(os, 38)]), VSUB(T63, T64), ovs, &(xo[0]));
 
1315
                         ST(&(xo[WS(os, 26)]), VADD(T63, T64), ovs, &(xo[0]));
 
1316
                         T67 = VBYI(VSUB(T65, T66));
 
1317
                         T6a = VSUB(T68, T69);
 
1318
                         ST(&(xo[WS(os, 22)]), VADD(T67, T6a), ovs, &(xo[0]));
 
1319
                         ST(&(xo[WS(os, 42)]), VSUB(T6a, T67), ovs, &(xo[0]));
 
1320
                    }
 
1321
               }
 
1322
               {
 
1323
                    V T11, T2C, T2v, T2D, T2e, T2z, T2s, T2A;
 
1324
                    {
 
1325
                         V Tr, T10, T2t, T2u;
 
1326
                         Tr = VSUB(Tb, Tq);
 
1327
                         T10 = VSUB(TI, TZ);
 
1328
                         T11 = VADD(Tr, T10);
 
1329
                         T2C = VSUB(Tr, T10);
 
1330
                         T2t = VFNMS(LDK(KP634393284), T1B, VMUL(LDK(KP773010453), T1s));
 
1331
                         T2u = VFMA(LDK(KP773010453), T2c, VMUL(LDK(KP634393284), T23));
 
1332
                         T2v = VADD(T2t, T2u);
 
1333
                         T2D = VSUB(T2u, T2t);
 
1334
                    }
 
1335
                    {
 
1336
                         V T1C, T2d, T2o, T2r;
 
1337
                         T1C = VFMA(LDK(KP634393284), T1s, VMUL(LDK(KP773010453), T1B));
 
1338
                         T2d = VFNMS(LDK(KP634393284), T2c, VMUL(LDK(KP773010453), T23));
 
1339
                         T2e = VADD(T1C, T2d);
 
1340
                         T2z = VSUB(T2d, T1C);
 
1341
                         T2o = VSUB(T2i, T2n);
 
1342
                         T2r = VSUB(T2p, T2q);
 
1343
                         T2s = VADD(T2o, T2r);
 
1344
                         T2A = VSUB(T2r, T2o);
 
1345
                    }
 
1346
                    {
 
1347
                         V T2f, T2w, T2F, T2G;
 
1348
                         T2f = VADD(T11, T2e);
 
1349
                         T2w = VBYI(VADD(T2s, T2v));
 
1350
                         ST(&(xo[WS(os, 57)]), VSUB(T2f, T2w), ovs, &(xo[WS(os, 1)]));
 
1351
                         ST(&(xo[WS(os, 7)]), VADD(T2f, T2w), ovs, &(xo[WS(os, 1)]));
 
1352
                         T2F = VBYI(VADD(T2A, T2z));
 
1353
                         T2G = VADD(T2C, T2D);
 
1354
                         ST(&(xo[WS(os, 9)]), VADD(T2F, T2G), ovs, &(xo[WS(os, 1)]));
 
1355
                         ST(&(xo[WS(os, 55)]), VSUB(T2G, T2F), ovs, &(xo[WS(os, 1)]));
 
1356
                    }
 
1357
                    {
 
1358
                         V T2x, T2y, T2B, T2E;
 
1359
                         T2x = VSUB(T11, T2e);
 
1360
                         T2y = VBYI(VSUB(T2v, T2s));
 
1361
                         ST(&(xo[WS(os, 39)]), VSUB(T2x, T2y), ovs, &(xo[WS(os, 1)]));
 
1362
                         ST(&(xo[WS(os, 25)]), VADD(T2x, T2y), ovs, &(xo[WS(os, 1)]));
 
1363
                         T2B = VBYI(VSUB(T2z, T2A));
 
1364
                         T2E = VSUB(T2C, T2D);
 
1365
                         ST(&(xo[WS(os, 23)]), VADD(T2B, T2E), ovs, &(xo[WS(os, 1)]));
 
1366
                         ST(&(xo[WS(os, 41)]), VSUB(T2E, T2B), ovs, &(xo[WS(os, 1)]));
 
1367
                    }
 
1368
               }
 
1369
               {
 
1370
                    V T3j, T3Q, T3J, T3R, T3y, T3N, T3G, T3O;
 
1371
                    {
 
1372
                         V T3b, T3i, T3H, T3I;
 
1373
                         T3b = VADD(T39, T3a);
 
1374
                         T3i = VADD(T3e, T3h);
 
1375
                         T3j = VADD(T3b, T3i);
 
1376
                         T3Q = VSUB(T3b, T3i);
 
1377
                         T3H = VFNMS(LDK(KP290284677), T3m, VMUL(LDK(KP956940335), T3p));
 
1378
                         T3I = VFMA(LDK(KP290284677), T3t, VMUL(LDK(KP956940335), T3w));
 
1379
                         T3J = VADD(T3H, T3I);
 
1380
                         T3R = VSUB(T3I, T3H);
 
1381
                    }
 
1382
                    {
 
1383
                         V T3q, T3x, T3C, T3F;
 
1384
                         T3q = VFMA(LDK(KP956940335), T3m, VMUL(LDK(KP290284677), T3p));
 
1385
                         T3x = VFNMS(LDK(KP290284677), T3w, VMUL(LDK(KP956940335), T3t));
 
1386
                         T3y = VADD(T3q, T3x);
 
1387
                         T3N = VSUB(T3x, T3q);
 
1388
                         T3C = VADD(T3A, T3B);
 
1389
                         T3F = VADD(T3D, T3E);
 
1390
                         T3G = VADD(T3C, T3F);
 
1391
                         T3O = VSUB(T3F, T3C);
 
1392
                    }
 
1393
                    {
 
1394
                         V T3z, T3K, T3T, T3U;
 
1395
                         T3z = VADD(T3j, T3y);
 
1396
                         T3K = VBYI(VADD(T3G, T3J));
 
1397
                         ST(&(xo[WS(os, 61)]), VSUB(T3z, T3K), ovs, &(xo[WS(os, 1)]));
 
1398
                         ST(&(xo[WS(os, 3)]), VADD(T3z, T3K), ovs, &(xo[WS(os, 1)]));
 
1399
                         T3T = VBYI(VADD(T3O, T3N));
 
1400
                         T3U = VADD(T3Q, T3R);
 
1401
                         ST(&(xo[WS(os, 13)]), VADD(T3T, T3U), ovs, &(xo[WS(os, 1)]));
 
1402
                         ST(&(xo[WS(os, 51)]), VSUB(T3U, T3T), ovs, &(xo[WS(os, 1)]));
 
1403
                    }
 
1404
                    {
 
1405
                         V T3L, T3M, T3P, T3S;
 
1406
                         T3L = VSUB(T3j, T3y);
 
1407
                         T3M = VBYI(VSUB(T3J, T3G));
 
1408
                         ST(&(xo[WS(os, 35)]), VSUB(T3L, T3M), ovs, &(xo[WS(os, 1)]));
 
1409
                         ST(&(xo[WS(os, 29)]), VADD(T3L, T3M), ovs, &(xo[WS(os, 1)]));
 
1410
                         T3P = VBYI(VSUB(T3N, T3O));
 
1411
                         T3S = VSUB(T3Q, T3R);
 
1412
                         ST(&(xo[WS(os, 19)]), VADD(T3P, T3S), ovs, &(xo[WS(os, 1)]));
 
1413
                         ST(&(xo[WS(os, 45)]), VSUB(T3S, T3P), ovs, &(xo[WS(os, 1)]));
 
1414
                    }
 
1415
               }
 
1416
               {
 
1417
                    V T4N, T5G, T5z, T5H, T5m, T5D, T5w, T5E;
 
1418
                    {
 
1419
                         V T4x, T4M, T5x, T5y;
 
1420
                         T4x = VADD(T4p, T4w);
 
1421
                         T4M = VADD(T4E, T4L);
 
1422
                         T4N = VADD(T4x, T4M);
 
1423
                         T5G = VSUB(T4x, T4M);
 
1424
                         T5x = VFNMS(LDK(KP195090322), T4Y, VMUL(LDK(KP980785280), T53));
 
1425
                         T5y = VFMA(LDK(KP195090322), T5f, VMUL(LDK(KP980785280), T5k));
 
1426
                         T5z = VADD(T5x, T5y);
 
1427
                         T5H = VSUB(T5y, T5x);
 
1428
                    }
 
1429
                    {
 
1430
                         V T54, T5l, T5s, T5v;
 
1431
                         T54 = VFMA(LDK(KP980785280), T4Y, VMUL(LDK(KP195090322), T53));
 
1432
                         T5l = VFNMS(LDK(KP195090322), T5k, VMUL(LDK(KP980785280), T5f));
 
1433
                         T5m = VADD(T54, T5l);
 
1434
                         T5D = VSUB(T5l, T54);
 
1435
                         T5s = VADD(T5q, T5r);
 
1436
                         T5v = VADD(T5t, T5u);
 
1437
                         T5w = VADD(T5s, T5v);
 
1438
                         T5E = VSUB(T5v, T5s);
 
1439
                    }
 
1440
                    {
 
1441
                         V T5n, T5A, T5J, T5K;
 
1442
                         T5n = VADD(T4N, T5m);
 
1443
                         T5A = VBYI(VADD(T5w, T5z));
 
1444
                         ST(&(xo[WS(os, 62)]), VSUB(T5n, T5A), ovs, &(xo[0]));
 
1445
                         ST(&(xo[WS(os, 2)]), VADD(T5n, T5A), ovs, &(xo[0]));
 
1446
                         T5J = VBYI(VADD(T5E, T5D));
 
1447
                         T5K = VADD(T5G, T5H);
 
1448
                         ST(&(xo[WS(os, 14)]), VADD(T5J, T5K), ovs, &(xo[0]));
 
1449
                         ST(&(xo[WS(os, 50)]), VSUB(T5K, T5J), ovs, &(xo[0]));
 
1450
                    }
 
1451
                    {
 
1452
                         V T5B, T5C, T5F, T5I;
 
1453
                         T5B = VSUB(T4N, T5m);
 
1454
                         T5C = VBYI(VSUB(T5z, T5w));
 
1455
                         ST(&(xo[WS(os, 34)]), VSUB(T5B, T5C), ovs, &(xo[0]));
 
1456
                         ST(&(xo[WS(os, 30)]), VADD(T5B, T5C), ovs, &(xo[0]));
 
1457
                         T5F = VBYI(VSUB(T5D, T5E));
 
1458
                         T5I = VSUB(T5G, T5H);
 
1459
                         ST(&(xo[WS(os, 18)]), VADD(T5F, T5I), ovs, &(xo[0]));
 
1460
                         ST(&(xo[WS(os, 46)]), VSUB(T5I, T5F), ovs, &(xo[0]));
 
1461
                    }
 
1462
               }
 
1463
               {
 
1464
                    V T2J, T34, T2X, T35, T2Q, T31, T2U, T32;
 
1465
                    {
 
1466
                         V T2H, T2I, T2V, T2W;
 
1467
                         T2H = VADD(Tb, Tq);
 
1468
                         T2I = VADD(T2q, T2p);
 
1469
                         T2J = VADD(T2H, T2I);
 
1470
                         T34 = VSUB(T2H, T2I);
 
1471
                         T2V = VFNMS(LDK(KP098017140), T2L, VMUL(LDK(KP995184726), T2K));
 
1472
                         T2W = VFMA(LDK(KP995184726), T2O, VMUL(LDK(KP098017140), T2N));
 
1473
                         T2X = VADD(T2V, T2W);
 
1474
                         T35 = VSUB(T2W, T2V);
 
1475
                    }
 
1476
                    {
 
1477
                         V T2M, T2P, T2S, T2T;
 
1478
                         T2M = VFMA(LDK(KP098017140), T2K, VMUL(LDK(KP995184726), T2L));
 
1479
                         T2P = VFNMS(LDK(KP098017140), T2O, VMUL(LDK(KP995184726), T2N));
 
1480
                         T2Q = VADD(T2M, T2P);
 
1481
                         T31 = VSUB(T2P, T2M);
 
1482
                         T2S = VADD(T2n, T2i);
 
1483
                         T2T = VADD(TZ, TI);
 
1484
                         T2U = VADD(T2S, T2T);
 
1485
                         T32 = VSUB(T2T, T2S);
 
1486
                    }
 
1487
                    {
 
1488
                         V T2R, T2Y, T37, T38;
 
1489
                         T2R = VADD(T2J, T2Q);
 
1490
                         T2Y = VBYI(VADD(T2U, T2X));
 
1491
                         ST(&(xo[WS(os, 63)]), VSUB(T2R, T2Y), ovs, &(xo[WS(os, 1)]));
 
1492
                         ST(&(xo[WS(os, 1)]), VADD(T2R, T2Y), ovs, &(xo[WS(os, 1)]));
 
1493
                         T37 = VBYI(VADD(T32, T31));
 
1494
                         T38 = VADD(T34, T35);
 
1495
                         ST(&(xo[WS(os, 15)]), VADD(T37, T38), ovs, &(xo[WS(os, 1)]));
 
1496
                         ST(&(xo[WS(os, 49)]), VSUB(T38, T37), ovs, &(xo[WS(os, 1)]));
 
1497
                    }
 
1498
                    {
 
1499
                         V T2Z, T30, T33, T36;
 
1500
                         T2Z = VSUB(T2J, T2Q);
 
1501
                         T30 = VBYI(VSUB(T2X, T2U));
 
1502
                         ST(&(xo[WS(os, 33)]), VSUB(T2Z, T30), ovs, &(xo[WS(os, 1)]));
 
1503
                         ST(&(xo[WS(os, 31)]), VADD(T2Z, T30), ovs, &(xo[WS(os, 1)]));
 
1504
                         T33 = VBYI(VSUB(T31, T32));
 
1505
                         T36 = VSUB(T34, T35);
 
1506
                         ST(&(xo[WS(os, 17)]), VADD(T33, T36), ovs, &(xo[WS(os, 1)]));
 
1507
                         ST(&(xo[WS(os, 47)]), VSUB(T36, T33), ovs, &(xo[WS(os, 1)]));
 
1508
                    }
 
1509
               }
 
1510
               {
 
1511
                    V T3X, T4i, T4b, T4j, T44, T4f, T48, T4g;
 
1512
                    {
 
1513
                         V T3V, T3W, T49, T4a;
 
1514
                         T3V = VSUB(T39, T3a);
 
1515
                         T3W = VSUB(T3E, T3D);
 
1516
                         T3X = VADD(T3V, T3W);
 
1517
                         T4i = VSUB(T3V, T3W);
 
1518
                         T49 = VFNMS(LDK(KP471396736), T3Y, VMUL(LDK(KP881921264), T3Z));
 
1519
                         T4a = VFMA(LDK(KP471396736), T41, VMUL(LDK(KP881921264), T42));
 
1520
                         T4b = VADD(T49, T4a);
 
1521
                         T4j = VSUB(T4a, T49);
 
1522
                    }
 
1523
                    {
 
1524
                         V T40, T43, T46, T47;
 
1525
                         T40 = VFMA(LDK(KP881921264), T3Y, VMUL(LDK(KP471396736), T3Z));
 
1526
                         T43 = VFNMS(LDK(KP471396736), T42, VMUL(LDK(KP881921264), T41));
 
1527
                         T44 = VADD(T40, T43);
 
1528
                         T4f = VSUB(T43, T40);
 
1529
                         T46 = VSUB(T3B, T3A);
 
1530
                         T47 = VSUB(T3h, T3e);
 
1531
                         T48 = VADD(T46, T47);
 
1532
                         T4g = VSUB(T47, T46);
 
1533
                    }
 
1534
                    {
 
1535
                         V T45, T4c, T4l, T4m;
 
1536
                         T45 = VADD(T3X, T44);
 
1537
                         T4c = VBYI(VADD(T48, T4b));
 
1538
                         ST(&(xo[WS(os, 59)]), VSUB(T45, T4c), ovs, &(xo[WS(os, 1)]));
 
1539
                         ST(&(xo[WS(os, 5)]), VADD(T45, T4c), ovs, &(xo[WS(os, 1)]));
 
1540
                         T4l = VBYI(VADD(T4g, T4f));
 
1541
                         T4m = VADD(T4i, T4j);
 
1542
                         ST(&(xo[WS(os, 11)]), VADD(T4l, T4m), ovs, &(xo[WS(os, 1)]));
 
1543
                         ST(&(xo[WS(os, 53)]), VSUB(T4m, T4l), ovs, &(xo[WS(os, 1)]));
 
1544
                    }
 
1545
                    {
 
1546
                         V T4d, T4e, T4h, T4k;
 
1547
                         T4d = VSUB(T3X, T44);
 
1548
                         T4e = VBYI(VSUB(T4b, T48));
 
1549
                         ST(&(xo[WS(os, 37)]), VSUB(T4d, T4e), ovs, &(xo[WS(os, 1)]));
 
1550
                         ST(&(xo[WS(os, 27)]), VADD(T4d, T4e), ovs, &(xo[WS(os, 1)]));
 
1551
                         T4h = VBYI(VSUB(T4f, T4g));
 
1552
                         T4k = VSUB(T4i, T4j);
 
1553
                         ST(&(xo[WS(os, 21)]), VADD(T4h, T4k), ovs, &(xo[WS(os, 1)]));
 
1554
                         ST(&(xo[WS(os, 43)]), VSUB(T4k, T4h), ovs, &(xo[WS(os, 1)]));
 
1555
                    }
 
1556
               }
 
1557
          }
 
1558
     }
 
1559
     VLEAVE();
 
1560
}
 
1561
 
 
1562
static const kdft_desc desc = { 64, XSIMD_STRING("n1fv_64"), {404, 72, 52, 0}, &GENUS, 0, 0, 0, 0 };
 
1563
 
 
1564
void XSIMD(codelet_n1fv_64) (planner *p) {
 
1565
     X(kdft_register) (p, n1fv_64, &desc);
 
1566
}
 
1567
 
 
1568
#endif                          /* HAVE_FMA */