~diresu/blender/blender-command-port

« back to all changes in this revision

Viewing changes to extern/fftw/dft/simd/codelets/n1bv_64.c

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

Show diffs side-by-side

added added

removed removed

Lines of Context:
 
1
/*
 
2
 * Copyright (c) 2003, 2006 Matteo Frigo
 
3
 * Copyright (c) 2003, 2006 Massachusetts Institute of Technology
 
4
 *
 
5
 * This program is free software; you can redistribute it and/or modify
 
6
 * it under the terms of the GNU General Public License as published by
 
7
 * the Free Software Foundation; either version 2 of the License, or
 
8
 * (at your option) any later version.
 
9
 *
 
10
 * This program is distributed in the hope that it will be useful,
 
11
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 
12
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 
13
 * GNU General Public License for more details.
 
14
 *
 
15
 * You should have received a copy of the GNU General Public License
 
16
 * along with this program; if not, write to the Free Software
 
17
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 
18
 *
 
19
 */
 
20
 
 
21
/* This file was automatically generated --- DO NOT EDIT */
 
22
/* Generated on Sat Jul  1 14:30:46 EDT 2006 */
 
23
 
 
24
#include "codelet-dft.h"
 
25
 
 
26
#ifdef HAVE_FMA
 
27
 
 
28
/* Generated by: ../../../genfft/gen_notw_c -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 64 -name n1bv_64 -include n1b.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, and 128 memory accesses
 
34
 */
 
35
/*
 
36
 * Generator Id's : 
 
37
 * $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
 
38
 * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
 
39
 * $Id: gen_notw_c.ml,v 1.17 2006-02-12 23:34:12 athena Exp $
 
40
 */
 
41
 
 
42
#include "n1b.h"
 
43
 
 
44
static void n1bv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
 
45
{
 
46
     DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
 
47
     DVK(KP820678790, +0.820678790828660330972281985331011598767386482);
 
48
     DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
 
49
     DVK(KP534511135, +0.534511135950791641089685961295362908582039528);
 
50
     DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
 
51
     DVK(KP098491403, +0.098491403357164253077197521291327432293052451);
 
52
     DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
 
53
     DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
 
54
     DVK(KP303346683, +0.303346683607342391675883946941299872384187453);
 
55
     DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
 
56
     DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
57
     DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
 
58
     DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
 
59
     DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
 
60
     DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
61
     INT i;
 
62
     const R *xi;
 
63
     R *xo;
 
64
     xi = ii;
 
65
     xo = io;
 
66
     for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
 
67
          V T5T, T5S, T5X, T65, T5Z, T5R, T67, T63, T5U, T64;
 
68
          {
 
69
               V T7, T26, T5k, T6A, T47, T69, T2V, T3z, T6B, T4e, T6a, T5n, T3M, T2Y, T27;
 
70
               V Tm, T3A, T3i, T29, TC, T5p, T4o, T6D, T6e, T3l, T3B, TR, T2a, T4x, T5q;
 
71
               V T6h, T6E, T39, T3H, T3I, T3c, T5N, T57, T72, T6w, T5O, T5e, T71, T6t, T2y;
 
72
               V T1W, T2x, T1N, T33, T34, T3E, T32, T1p, T2v, T1g, T2u, T4M, T5K, T6p, T6Z;
 
73
               V T6m, T6Y, T5L, T4T;
 
74
               {
 
75
                    V T4g, T4l, T3g, Tu, Tx, T4h, TA, T4i;
 
76
                    {
 
77
                         V T1, T2, T23, T24, T4, T5, T20, T21;
 
78
                         T1 = LD(&(xi[0]), ivs, &(xi[0]));
 
79
                         T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
 
80
                         T23 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
 
81
                         T24 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
 
82
                         T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
 
83
                         T5 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
 
84
                         T20 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
 
85
                         T21 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
 
86
                         {
 
87
                              V Ta, T48, Tk, T4c, T49, Td, Tf, Tg;
 
88
                              {
 
89
                                   V T8, T43, T3, T45, T25, T5i, T6, T44, T22, T9, Ti, Tj, Tb, Tc;
 
90
                                   T8 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
 
91
                                   T43 = VSUB(T1, T2);
 
92
                                   T3 = VADD(T1, T2);
 
93
                                   T45 = VSUB(T23, T24);
 
94
                                   T25 = VADD(T23, T24);
 
95
                                   T5i = VSUB(T4, T5);
 
96
                                   T6 = VADD(T4, T5);
 
97
                                   T44 = VSUB(T20, T21);
 
98
                                   T22 = VADD(T20, T21);
 
99
                                   T9 = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
 
100
                                   Ti = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
 
101
                                   Tj = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
 
102
                                   Tb = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
 
103
                                   Tc = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
 
104
                                   {
 
105
                                        V T2T, T46, T5j, T2U;
 
106
                                        T7 = VSUB(T3, T6);
 
107
                                        T2T = VADD(T3, T6);
 
108
                                        T46 = VADD(T44, T45);
 
109
                                        T5j = VSUB(T44, T45);
 
110
                                        T26 = VSUB(T22, T25);
 
111
                                        T2U = VADD(T22, T25);
 
112
                                        Ta = VADD(T8, T9);
 
113
                                        T48 = VSUB(T8, T9);
 
114
                                        Tk = VADD(Ti, Tj);
 
115
                                        T4c = VSUB(Tj, Ti);
 
116
                                        T5k = VFMA(LDK(KP707106781), T5j, T5i);
 
117
                                        T6A = VFNMS(LDK(KP707106781), T5j, T5i);
 
118
                                        T47 = VFMA(LDK(KP707106781), T46, T43);
 
119
                                        T69 = VFNMS(LDK(KP707106781), T46, T43);
 
120
                                        T2V = VADD(T2T, T2U);
 
121
                                        T3z = VSUB(T2T, T2U);
 
122
                                        T49 = VSUB(Tb, Tc);
 
123
                                        Td = VADD(Tb, Tc);
 
124
                                   }
 
125
                                   Tf = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
 
126
                                   Tg = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
 
127
                              }
 
128
                              {
 
129
                                   V Te, T2W, T5l, T4a, Tq, Tt, Tv, Tw, T5m, T4d, Tl, T2X, Ty, Tz, To;
 
130
                                   V Tp;
 
131
                                   To = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
 
132
                                   Tp = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
 
133
                                   {
 
134
                                        V Th, T4b, Tr, Ts;
 
135
                                        Tr = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
 
136
                                        Ts = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
 
137
                                        Te = VSUB(Ta, Td);
 
138
                                        T2W = VADD(Ta, Td);
 
139
                                        T5l = VFMA(LDK(KP414213562), T48, T49);
 
140
                                        T4a = VFNMS(LDK(KP414213562), T49, T48);
 
141
                                        Th = VADD(Tf, Tg);
 
142
                                        T4b = VSUB(Tf, Tg);
 
143
                                        Tq = VADD(To, Tp);
 
144
                                        T4g = VSUB(To, Tp);
 
145
                                        T4l = VSUB(Tr, Ts);
 
146
                                        Tt = VADD(Tr, Ts);
 
147
                                        Tv = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
 
148
                                        Tw = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
 
149
                                        T5m = VFMA(LDK(KP414213562), T4b, T4c);
 
150
                                        T4d = VFNMS(LDK(KP414213562), T4c, T4b);
 
151
                                        Tl = VSUB(Th, Tk);
 
152
                                        T2X = VADD(Th, Tk);
 
153
                                        Ty = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
 
154
                                        Tz = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
 
155
                                   }
 
156
                                   T3g = VADD(Tq, Tt);
 
157
                                   Tu = VSUB(Tq, Tt);
 
158
                                   Tx = VADD(Tv, Tw);
 
159
                                   T4h = VSUB(Tv, Tw);
 
160
                                   T6B = VSUB(T4a, T4d);
 
161
                                   T4e = VADD(T4a, T4d);
 
162
                                   T6a = VADD(T5l, T5m);
 
163
                                   T5n = VSUB(T5l, T5m);
 
164
                                   T3M = VSUB(T2W, T2X);
 
165
                                   T2Y = VADD(T2W, T2X);
 
166
                                   T27 = VSUB(Te, Tl);
 
167
                                   Tm = VADD(Te, Tl);
 
168
                                   TA = VADD(Ty, Tz);
 
169
                                   T4i = VSUB(Ty, Tz);
 
170
                              }
 
171
                         }
 
172
                    }
 
173
                    {
 
174
                         V TK, T4p, T4u, T4k, T6d, T4n, T6c, TL, TN, TO, T3j, TJ, TF, TI;
 
175
                         {
 
176
                              V TD, TE, TG, TH;
 
177
                              TD = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
 
178
                              TE = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
 
179
                              TG = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
 
180
                              TH = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
 
181
                              TK = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
 
182
                              {
 
183
                                   V T3h, TB, T4j, T4m;
 
184
                                   T3h = VADD(Tx, TA);
 
185
                                   TB = VSUB(Tx, TA);
 
186
                                   T4j = VADD(T4h, T4i);
 
187
                                   T4m = VSUB(T4h, T4i);
 
188
                                   T4p = VSUB(TD, TE);
 
189
                                   TF = VADD(TD, TE);
 
190
                                   T4u = VSUB(TH, TG);
 
191
                                   TI = VADD(TG, TH);
 
192
                                   T3A = VSUB(T3g, T3h);
 
193
                                   T3i = VADD(T3g, T3h);
 
194
                                   T29 = VFMA(LDK(KP414213562), Tu, TB);
 
195
                                   TC = VFNMS(LDK(KP414213562), TB, Tu);
 
196
                                   T4k = VFMA(LDK(KP707106781), T4j, T4g);
 
197
                                   T6d = VFNMS(LDK(KP707106781), T4j, T4g);
 
198
                                   T4n = VFMA(LDK(KP707106781), T4m, T4l);
 
199
                                   T6c = VFNMS(LDK(KP707106781), T4m, T4l);
 
200
                                   TL = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
 
201
                              }
 
202
                              TN = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
 
203
                              TO = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
 
204
                         }
 
205
                         T3j = VADD(TF, TI);
 
206
                         TJ = VSUB(TF, TI);
 
207
                         {
 
208
                              V T3a, T1E, T52, T5b, T1x, T4Z, T6r, T6u, T5a, T1U, T55, T5c, T1L, T3b;
 
209
                              {
 
210
                                   V T4V, T1t, T58, T1w, T1Q, T1T, T1I, T4Y, T59, T1J, T53, T1H;
 
211
                                   {
 
212
                                        V T1r, TM, T4r, TP, T4q, T1s, T1u, T1v;
 
213
                                        T1r = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
 
214
                                        T5p = VFMA(LDK(KP198912367), T4k, T4n);
 
215
                                        T4o = VFNMS(LDK(KP198912367), T4n, T4k);
 
216
                                        T6D = VFMA(LDK(KP668178637), T6c, T6d);
 
217
                                        T6e = VFNMS(LDK(KP668178637), T6d, T6c);
 
218
                                        TM = VADD(TK, TL);
 
219
                                        T4r = VSUB(TK, TL);
 
220
                                        TP = VADD(TN, TO);
 
221
                                        T4q = VSUB(TN, TO);
 
222
                                        T1s = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
 
223
                                        T1u = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
 
224
                                        T1v = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
 
225
                                        {
 
226
                                             V T1R, T4X, T6g, T4t, T6f, T4w, T1S, T1O, T1P;
 
227
                                             T1O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
 
228
                                             T1P = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
 
229
                                             T1R = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
 
230
                                             {
 
231
                                                  V T3k, TQ, T4s, T4v;
 
232
                                                  T3k = VADD(TP, TM);
 
233
                                                  TQ = VSUB(TM, TP);
 
234
                                                  T4s = VADD(T4q, T4r);
 
235
                                                  T4v = VSUB(T4r, T4q);
 
236
                                                  T4V = VSUB(T1r, T1s);
 
237
                                                  T1t = VADD(T1r, T1s);
 
238
                                                  T58 = VSUB(T1v, T1u);
 
239
                                                  T1w = VADD(T1u, T1v);
 
240
                                                  T4X = VSUB(T1O, T1P);
 
241
                                                  T1Q = VADD(T1O, T1P);
 
242
                                                  T3l = VADD(T3j, T3k);
 
243
                                                  T3B = VSUB(T3j, T3k);
 
244
                                                  TR = VFNMS(LDK(KP414213562), TQ, TJ);
 
245
                                                  T2a = VFMA(LDK(KP414213562), TJ, TQ);
 
246
                                                  T6g = VFNMS(LDK(KP707106781), T4s, T4p);
 
247
                                                  T4t = VFMA(LDK(KP707106781), T4s, T4p);
 
248
                                                  T6f = VFNMS(LDK(KP707106781), T4v, T4u);
 
249
                                                  T4w = VFMA(LDK(KP707106781), T4v, T4u);
 
250
                                                  T1S = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
 
251
                                             }
 
252
                                             {
 
253
                                                  V T4W, T1A, T50, T51, T1D, T1F, T1G;
 
254
                                                  {
 
255
                                                       V T1y, T1z, T1B, T1C;
 
256
                                                       T1y = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
 
257
                                                       T1z = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
 
258
                                                       T1B = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
 
259
                                                       T1C = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
 
260
                                                       T4x = VFNMS(LDK(KP198912367), T4w, T4t);
 
261
                                                       T5q = VFMA(LDK(KP198912367), T4t, T4w);
 
262
                                                       T6h = VFNMS(LDK(KP668178637), T6g, T6f);
 
263
                                                       T6E = VFMA(LDK(KP668178637), T6f, T6g);
 
264
                                                       T4W = VSUB(T1R, T1S);
 
265
                                                       T1T = VADD(T1R, T1S);
 
266
                                                       T1A = VADD(T1y, T1z);
 
267
                                                       T50 = VSUB(T1y, T1z);
 
268
                                                       T51 = VSUB(T1C, T1B);
 
269
                                                       T1D = VADD(T1B, T1C);
 
270
                                                  }
 
271
                                                  T1F = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
 
272
                                                  T1G = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
 
273
                                                  T1I = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
 
274
                                                  T4Y = VADD(T4W, T4X);
 
275
                                                  T59 = VSUB(T4X, T4W);
 
276
                                                  T1J = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
 
277
                                                  T3a = VADD(T1A, T1D);
 
278
                                                  T1E = VSUB(T1A, T1D);
 
279
                                                  T52 = VFMA(LDK(KP414213562), T51, T50);
 
280
                                                  T5b = VFNMS(LDK(KP414213562), T50, T51);
 
281
                                                  T53 = VSUB(T1F, T1G);
 
282
                                                  T1H = VADD(T1F, T1G);
 
283
                                             }
 
284
                                        }
 
285
                                   }
 
286
                                   {
 
287
                                        V T37, T54, T1K, T38;
 
288
                                        T1x = VSUB(T1t, T1w);
 
289
                                        T37 = VADD(T1t, T1w);
 
290
                                        T4Z = VFMA(LDK(KP707106781), T4Y, T4V);
 
291
                                        T6r = VFNMS(LDK(KP707106781), T4Y, T4V);
 
292
                                        T54 = VSUB(T1J, T1I);
 
293
                                        T1K = VADD(T1I, T1J);
 
294
                                        T6u = VFNMS(LDK(KP707106781), T59, T58);
 
295
                                        T5a = VFMA(LDK(KP707106781), T59, T58);
 
296
                                        T38 = VADD(T1T, T1Q);
 
297
                                        T1U = VSUB(T1Q, T1T);
 
298
                                        T55 = VFNMS(LDK(KP414213562), T54, T53);
 
299
                                        T5c = VFMA(LDK(KP414213562), T53, T54);
 
300
                                        T1L = VSUB(T1H, T1K);
 
301
                                        T3b = VADD(T1H, T1K);
 
302
                                        T39 = VADD(T37, T38);
 
303
                                        T3H = VSUB(T37, T38);
 
304
                                   }
 
305
                              }
 
306
                              {
 
307
                                   V T4A, TW, T4N, TZ, T1j, T1m, T4O, T4D, T13, T4F, T16, T4G, T1a, T4I, T4J;
 
308
                                   V T1d;
 
309
                                   {
 
310
                                        V TU, TV, TX, TY, T56, T6v;
 
311
                                        TU = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
 
312
                                        T56 = VADD(T52, T55);
 
313
                                        T6v = VSUB(T55, T52);
 
314
                                        {
 
315
                                             V T5d, T6s, T1V, T1M;
 
316
                                             T5d = VADD(T5b, T5c);
 
317
                                             T6s = VSUB(T5c, T5b);
 
318
                                             T1V = VSUB(T1L, T1E);
 
319
                                             T1M = VADD(T1E, T1L);
 
320
                                             T3I = VSUB(T3b, T3a);
 
321
                                             T3c = VADD(T3a, T3b);
 
322
                                             T5N = VFNMS(LDK(KP923879532), T56, T4Z);
 
323
                                             T57 = VFMA(LDK(KP923879532), T56, T4Z);
 
324
                                             T72 = VFNMS(LDK(KP923879532), T6v, T6u);
 
325
                                             T6w = VFMA(LDK(KP923879532), T6v, T6u);
 
326
                                             T5O = VFNMS(LDK(KP923879532), T5d, T5a);
 
327
                                             T5e = VFMA(LDK(KP923879532), T5d, T5a);
 
328
                                             T71 = VFMA(LDK(KP923879532), T6s, T6r);
 
329
                                             T6t = VFNMS(LDK(KP923879532), T6s, T6r);
 
330
                                             T2y = VFNMS(LDK(KP707106781), T1V, T1U);
 
331
                                             T1W = VFMA(LDK(KP707106781), T1V, T1U);
 
332
                                             T2x = VFNMS(LDK(KP707106781), T1M, T1x);
 
333
                                             T1N = VFMA(LDK(KP707106781), T1M, T1x);
 
334
                                             TV = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
 
335
                                        }
 
336
                                        TX = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
 
337
                                        TY = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
 
338
                                        {
 
339
                                             V T1h, T1i, T1k, T1l;
 
340
                                             T1h = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
 
341
                                             T1i = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
 
342
                                             T1k = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
 
343
                                             T1l = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
 
344
                                             {
 
345
                                                  V T11, T4B, T4C, T12, T14, T15;
 
346
                                                  T11 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
 
347
                                                  T4A = VSUB(TU, TV);
 
348
                                                  TW = VADD(TU, TV);
 
349
                                                  T4N = VSUB(TX, TY);
 
350
                                                  TZ = VADD(TX, TY);
 
351
                                                  T1j = VADD(T1h, T1i);
 
352
                                                  T4B = VSUB(T1h, T1i);
 
353
                                                  T1m = VADD(T1k, T1l);
 
354
                                                  T4C = VSUB(T1k, T1l);
 
355
                                                  T12 = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
 
356
                                                  T14 = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
 
357
                                                  T15 = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
 
358
                                                  {
 
359
                                                       V T18, T19, T1b, T1c;
 
360
                                                       T18 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
 
361
                                                       T19 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
 
362
                                                       T1b = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
 
363
                                                       T1c = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
 
364
                                                       T4O = VSUB(T4B, T4C);
 
365
                                                       T4D = VADD(T4B, T4C);
 
366
                                                       T13 = VADD(T11, T12);
 
367
                                                       T4F = VSUB(T11, T12);
 
368
                                                       T16 = VADD(T14, T15);
 
369
                                                       T4G = VSUB(T14, T15);
 
370
                                                       T1a = VADD(T18, T19);
 
371
                                                       T4I = VSUB(T18, T19);
 
372
                                                       T4J = VSUB(T1b, T1c);
 
373
                                                       T1d = VADD(T1b, T1c);
 
374
                                                  }
 
375
                                             }
 
376
                                        }
 
377
                                   }
 
378
                                   {
 
379
                                        V T30, T10, T6k, T4E, T4Q, T4H, T17, T6n, T4P, T1e, T4K, T4R, T1n, T31;
 
380
                                        T30 = VADD(TW, TZ);
 
381
                                        T10 = VSUB(TW, TZ);
 
382
                                        T6k = VFNMS(LDK(KP707106781), T4D, T4A);
 
383
                                        T4E = VFMA(LDK(KP707106781), T4D, T4A);
 
384
                                        T4Q = VFMA(LDK(KP414213562), T4F, T4G);
 
385
                                        T4H = VFNMS(LDK(KP414213562), T4G, T4F);
 
386
                                        T33 = VADD(T13, T16);
 
387
                                        T17 = VSUB(T13, T16);
 
388
                                        T6n = VFNMS(LDK(KP707106781), T4O, T4N);
 
389
                                        T4P = VFMA(LDK(KP707106781), T4O, T4N);
 
390
                                        T34 = VADD(T1a, T1d);
 
391
                                        T1e = VSUB(T1a, T1d);
 
392
                                        T4K = VFMA(LDK(KP414213562), T4J, T4I);
 
393
                                        T4R = VFNMS(LDK(KP414213562), T4I, T4J);
 
394
                                        T1n = VSUB(T1j, T1m);
 
395
                                        T31 = VADD(T1j, T1m);
 
396
                                        {
 
397
                                             V T1f, T1o, T6o, T4L, T4S, T6l;
 
398
                                             T1f = VADD(T17, T1e);
 
399
                                             T1o = VSUB(T17, T1e);
 
400
                                             T6o = VSUB(T4H, T4K);
 
401
                                             T4L = VADD(T4H, T4K);
 
402
                                             T4S = VADD(T4Q, T4R);
 
403
                                             T6l = VSUB(T4Q, T4R);
 
404
                                             T3E = VSUB(T30, T31);
 
405
                                             T32 = VADD(T30, T31);
 
406
                                             T1p = VFMA(LDK(KP707106781), T1o, T1n);
 
407
                                             T2v = VFNMS(LDK(KP707106781), T1o, T1n);
 
408
                                             T1g = VFMA(LDK(KP707106781), T1f, T10);
 
409
                                             T2u = VFNMS(LDK(KP707106781), T1f, T10);
 
410
                                             T4M = VFMA(LDK(KP923879532), T4L, T4E);
 
411
                                             T5K = VFNMS(LDK(KP923879532), T4L, T4E);
 
412
                                             T6p = VFMA(LDK(KP923879532), T6o, T6n);
 
413
                                             T6Z = VFNMS(LDK(KP923879532), T6o, T6n);
 
414
                                             T6m = VFNMS(LDK(KP923879532), T6l, T6k);
 
415
                                             T6Y = VFMA(LDK(KP923879532), T6l, T6k);
 
416
                                             T5L = VFNMS(LDK(KP923879532), T4S, T4P);
 
417
                                             T4T = VFMA(LDK(KP923879532), T4S, T4P);
 
418
                                        }
 
419
                                   }
 
420
                              }
 
421
                         }
 
422
                    }
 
423
               }
 
424
               {
 
425
                    V T6b, T6F, T7f, T6X, T70, T79, T7a, T73, T6C, T76, T77, T6i;
 
426
                    {
 
427
                         V T2Z, T3r, T3s, T3m, T3d, T3v;
 
428
                         T2Z = VSUB(T2V, T2Y);
 
429
                         T3r = VADD(T2V, T2Y);
 
430
                         T3s = VADD(T3i, T3l);
 
431
                         T3m = VSUB(T3i, T3l);
 
432
                         T3d = VSUB(T39, T3c);
 
433
                         T3v = VADD(T39, T3c);
 
434
                         {
 
435
                              V T3x, T3t, T3Q, T3J, T3D, T3V, T3G, T3P, T3u, T36, T3O, T3Y, T6V, T6W;
 
436
                              {
 
437
                                   V T3N, T3C, T3F, T35;
 
438
                                   T3N = VSUB(T3A, T3B);
 
439
                                   T3C = VADD(T3A, T3B);
 
440
                                   T3F = VSUB(T33, T34);
 
441
                                   T35 = VADD(T33, T34);
 
442
                                   T3x = VADD(T3r, T3s);
 
443
                                   T3t = VSUB(T3r, T3s);
 
444
                                   T3Q = VFMA(LDK(KP414213562), T3H, T3I);
 
445
                                   T3J = VFNMS(LDK(KP414213562), T3I, T3H);
 
446
                                   T3D = VFMA(LDK(KP707106781), T3C, T3z);
 
447
                                   T3V = VFNMS(LDK(KP707106781), T3C, T3z);
 
448
                                   T3G = VFNMS(LDK(KP414213562), T3F, T3E);
 
449
                                   T3P = VFMA(LDK(KP414213562), T3E, T3F);
 
450
                                   T3u = VADD(T32, T35);
 
451
                                   T36 = VSUB(T32, T35);
 
452
                                   T3O = VFMA(LDK(KP707106781), T3N, T3M);
 
453
                                   T3Y = VFNMS(LDK(KP707106781), T3N, T3M);
 
454
                              }
 
455
                              T6b = VFNMS(LDK(KP923879532), T6a, T69);
 
456
                              T6V = VFMA(LDK(KP923879532), T6a, T69);
 
457
                              T6W = VADD(T6D, T6E);
 
458
                              T6F = VSUB(T6D, T6E);
 
459
                              {
 
460
                                   V T3R, T3W, T3K, T3Z;
 
461
                                   T3R = VSUB(T3P, T3Q);
 
462
                                   T3W = VADD(T3P, T3Q);
 
463
                                   T3K = VADD(T3G, T3J);
 
464
                                   T3Z = VSUB(T3G, T3J);
 
465
                                   {
 
466
                                        V T3e, T3n, T3w, T3y;
 
467
                                        T3e = VADD(T36, T3d);
 
468
                                        T3n = VSUB(T36, T3d);
 
469
                                        T3w = VSUB(T3u, T3v);
 
470
                                        T3y = VADD(T3u, T3v);
 
471
                                        {
 
472
                                             V T41, T3X, T3S, T3U;
 
473
                                             T41 = VFMA(LDK(KP923879532), T3W, T3V);
 
474
                                             T3X = VFNMS(LDK(KP923879532), T3W, T3V);
 
475
                                             T3S = VFNMS(LDK(KP923879532), T3R, T3O);
 
476
                                             T3U = VFMA(LDK(KP923879532), T3R, T3O);
 
477
                                             {
 
478
                                                  V T42, T40, T3L, T3T;
 
479
                                                  T42 = VFNMS(LDK(KP923879532), T3Z, T3Y);
 
480
                                                  T40 = VFMA(LDK(KP923879532), T3Z, T3Y);
 
481
                                                  T3L = VFNMS(LDK(KP923879532), T3K, T3D);
 
482
                                                  T3T = VFMA(LDK(KP923879532), T3K, T3D);
 
483
                                                  {
 
484
                                                       V T3o, T3q, T3f, T3p;
 
485
                                                       T3o = VFNMS(LDK(KP707106781), T3n, T3m);
 
486
                                                       T3q = VFMA(LDK(KP707106781), T3n, T3m);
 
487
                                                       T3f = VFNMS(LDK(KP707106781), T3e, T2Z);
 
488
                                                       T3p = VFMA(LDK(KP707106781), T3e, T2Z);
 
489
                                                       ST(&(xo[WS(os, 32)]), VSUB(T3x, T3y), ovs, &(xo[0]));
 
490
                                                       ST(&(xo[0]), VADD(T3x, T3y), ovs, &(xo[0]));
 
491
                                                       ST(&(xo[WS(os, 16)]), VFMAI(T3w, T3t), ovs, &(xo[0]));
 
492
                                                       ST(&(xo[WS(os, 48)]), VFNMSI(T3w, T3t), ovs, &(xo[0]));
 
493
                                                       ST(&(xo[WS(os, 44)]), VFNMSI(T40, T3X), ovs, &(xo[0]));
 
494
                                                       ST(&(xo[WS(os, 20)]), VFMAI(T40, T3X), ovs, &(xo[0]));
 
495
                                                       ST(&(xo[WS(os, 52)]), VFMAI(T42, T41), ovs, &(xo[0]));
 
496
                                                       ST(&(xo[WS(os, 12)]), VFNMSI(T42, T41), ovs, &(xo[0]));
 
497
                                                       ST(&(xo[WS(os, 4)]), VFMAI(T3U, T3T), ovs, &(xo[0]));
 
498
                                                       ST(&(xo[WS(os, 60)]), VFNMSI(T3U, T3T), ovs, &(xo[0]));
 
499
                                                       ST(&(xo[WS(os, 36)]), VFMAI(T3S, T3L), ovs, &(xo[0]));
 
500
                                                       ST(&(xo[WS(os, 28)]), VFNMSI(T3S, T3L), ovs, &(xo[0]));
 
501
                                                       ST(&(xo[WS(os, 56)]), VFNMSI(T3q, T3p), ovs, &(xo[0]));
 
502
                                                       ST(&(xo[WS(os, 8)]), VFMAI(T3q, T3p), ovs, &(xo[0]));
 
503
                                                       ST(&(xo[WS(os, 40)]), VFMAI(T3o, T3f), ovs, &(xo[0]));
 
504
                                                       ST(&(xo[WS(os, 24)]), VFNMSI(T3o, T3f), ovs, &(xo[0]));
 
505
                                                       T7f = VFNMS(LDK(KP831469612), T6W, T6V);
 
506
                                                       T6X = VFMA(LDK(KP831469612), T6W, T6V);
 
507
                                                  }
 
508
                                             }
 
509
                                        }
 
510
                                   }
 
511
                              }
 
512
                              T70 = VFMA(LDK(KP303346683), T6Z, T6Y);
 
513
                              T79 = VFNMS(LDK(KP303346683), T6Y, T6Z);
 
514
                              T7a = VFNMS(LDK(KP303346683), T71, T72);
 
515
                              T73 = VFMA(LDK(KP303346683), T72, T71);
 
516
                              T6C = VFMA(LDK(KP923879532), T6B, T6A);
 
517
                              T76 = VFNMS(LDK(KP923879532), T6B, T6A);
 
518
                              T77 = VSUB(T6e, T6h);
 
519
                              T6i = VADD(T6e, T6h);
 
520
                         }
 
521
                    }
 
522
                    {
 
523
                         V T2r, T2D, T2C, T2s, T5H, T5o, T5v, T5D, T5r, T5I, T5x, T5h, T5F, T5B;
 
524
                         {
 
525
                              V TT, T2f, T2n, T1Y, T28, T2b, T2l, T2p, T2j, T2k;
 
526
                              {
 
527
                                   V T1q, T2d, T7h, T7l, T2e, T1X, T75, T7d, T7m, T7k, T7c, T7e, Tn, TS;
 
528
                                   T2r = VFNMS(LDK(KP707106781), Tm, T7);
 
529
                                   Tn = VFMA(LDK(KP707106781), Tm, T7);
 
530
                                   TS = VADD(TC, TR);
 
531
                                   T2D = VSUB(TC, TR);
 
532
                                   {
 
533
                                        V T7b, T7j, T74, T7i, T78, T7g;
 
534
                                        T1q = VFNMS(LDK(KP198912367), T1p, T1g);
 
535
                                        T2d = VFMA(LDK(KP198912367), T1g, T1p);
 
536
                                        T7g = VADD(T79, T7a);
 
537
                                        T7b = VSUB(T79, T7a);
 
538
                                        T7j = VSUB(T70, T73);
 
539
                                        T74 = VADD(T70, T73);
 
540
                                        T7i = VFNMS(LDK(KP831469612), T77, T76);
 
541
                                        T78 = VFMA(LDK(KP831469612), T77, T76);
 
542
                                        T2j = VFNMS(LDK(KP923879532), TS, Tn);
 
543
                                        TT = VFMA(LDK(KP923879532), TS, Tn);
 
544
                                        T7h = VFMA(LDK(KP956940335), T7g, T7f);
 
545
                                        T7l = VFNMS(LDK(KP956940335), T7g, T7f);
 
546
                                        T2e = VFMA(LDK(KP198912367), T1N, T1W);
 
547
                                        T1X = VFNMS(LDK(KP198912367), T1W, T1N);
 
548
                                        T75 = VFNMS(LDK(KP956940335), T74, T6X);
 
549
                                        T7d = VFMA(LDK(KP956940335), T74, T6X);
 
550
                                        T7m = VFMA(LDK(KP956940335), T7j, T7i);
 
551
                                        T7k = VFNMS(LDK(KP956940335), T7j, T7i);
 
552
                                        T7c = VFNMS(LDK(KP956940335), T7b, T78);
 
553
                                        T7e = VFMA(LDK(KP956940335), T7b, T78);
 
554
                                   }
 
555
                                   T2k = VADD(T2d, T2e);
 
556
                                   T2f = VSUB(T2d, T2e);
 
557
                                   ST(&(xo[WS(os, 45)]), VFMAI(T7k, T7h), ovs, &(xo[WS(os, 1)]));
 
558
                                   ST(&(xo[WS(os, 19)]), VFNMSI(T7k, T7h), ovs, &(xo[WS(os, 1)]));
 
559
                                   ST(&(xo[WS(os, 51)]), VFNMSI(T7m, T7l), ovs, &(xo[WS(os, 1)]));
 
560
                                   ST(&(xo[WS(os, 13)]), VFMAI(T7m, T7l), ovs, &(xo[WS(os, 1)]));
 
561
                                   ST(&(xo[WS(os, 61)]), VFMAI(T7e, T7d), ovs, &(xo[WS(os, 1)]));
 
562
                                   ST(&(xo[WS(os, 3)]), VFNMSI(T7e, T7d), ovs, &(xo[WS(os, 1)]));
 
563
                                   ST(&(xo[WS(os, 29)]), VFMAI(T7c, T75), ovs, &(xo[WS(os, 1)]));
 
564
                                   ST(&(xo[WS(os, 35)]), VFNMSI(T7c, T75), ovs, &(xo[WS(os, 1)]));
 
565
                                   T2n = VSUB(T1q, T1X);
 
566
                                   T1Y = VADD(T1q, T1X);
 
567
                                   T2C = VFNMS(LDK(KP707106781), T27, T26);
 
568
                                   T28 = VFMA(LDK(KP707106781), T27, T26);
 
569
                                   T2b = VSUB(T29, T2a);
 
570
                                   T2s = VADD(T29, T2a);
 
571
                              }
 
572
                              T2l = VFNMS(LDK(KP980785280), T2k, T2j);
 
573
                              T2p = VFMA(LDK(KP980785280), T2k, T2j);
 
574
                              {
 
575
                                   V T5z, T4z, T5A, T5g;
 
576
                                   {
 
577
                                        V T4f, T4y, T1Z, T2h, T4U, T5t, T2m, T2c, T5u, T5f;
 
578
                                        T5H = VFNMS(LDK(KP923879532), T4e, T47);
 
579
                                        T4f = VFMA(LDK(KP923879532), T4e, T47);
 
580
                                        T4y = VADD(T4o, T4x);
 
581
                                        T5T = VSUB(T4o, T4x);
 
582
                                        T1Z = VFNMS(LDK(KP980785280), T1Y, TT);
 
583
                                        T2h = VFMA(LDK(KP980785280), T1Y, TT);
 
584
                                        T4U = VFNMS(LDK(KP098491403), T4T, T4M);
 
585
                                        T5t = VFMA(LDK(KP098491403), T4M, T4T);
 
586
                                        T2m = VFNMS(LDK(KP923879532), T2b, T28);
 
587
                                        T2c = VFMA(LDK(KP923879532), T2b, T28);
 
588
                                        T5u = VFMA(LDK(KP098491403), T57, T5e);
 
589
                                        T5f = VFNMS(LDK(KP098491403), T5e, T57);
 
590
                                        T5z = VFNMS(LDK(KP980785280), T4y, T4f);
 
591
                                        T4z = VFMA(LDK(KP980785280), T4y, T4f);
 
592
                                        T5S = VFNMS(LDK(KP923879532), T5n, T5k);
 
593
                                        T5o = VFMA(LDK(KP923879532), T5n, T5k);
 
594
                                        {
 
595
                                             V T2o, T2q, T2i, T2g;
 
596
                                             T2o = VFMA(LDK(KP980785280), T2n, T2m);
 
597
                                             T2q = VFNMS(LDK(KP980785280), T2n, T2m);
 
598
                                             T2i = VFMA(LDK(KP980785280), T2f, T2c);
 
599
                                             T2g = VFNMS(LDK(KP980785280), T2f, T2c);
 
600
                                             T5A = VADD(T5t, T5u);
 
601
                                             T5v = VSUB(T5t, T5u);
 
602
                                             T5D = VSUB(T4U, T5f);
 
603
                                             T5g = VADD(T4U, T5f);
 
604
                                             ST(&(xo[WS(os, 46)]), VFNMSI(T2o, T2l), ovs, &(xo[0]));
 
605
                                             ST(&(xo[WS(os, 18)]), VFMAI(T2o, T2l), ovs, &(xo[0]));
 
606
                                             ST(&(xo[WS(os, 50)]), VFMAI(T2q, T2p), ovs, &(xo[0]));
 
607
                                             ST(&(xo[WS(os, 14)]), VFNMSI(T2q, T2p), ovs, &(xo[0]));
 
608
                                             ST(&(xo[WS(os, 2)]), VFMAI(T2i, T2h), ovs, &(xo[0]));
 
609
                                             ST(&(xo[WS(os, 62)]), VFNMSI(T2i, T2h), ovs, &(xo[0]));
 
610
                                             ST(&(xo[WS(os, 34)]), VFMAI(T2g, T1Z), ovs, &(xo[0]));
 
611
                                             ST(&(xo[WS(os, 30)]), VFNMSI(T2g, T1Z), ovs, &(xo[0]));
 
612
                                             T5r = VSUB(T5p, T5q);
 
613
                                             T5I = VADD(T5p, T5q);
 
614
                                        }
 
615
                                   }
 
616
                                   T5x = VFMA(LDK(KP995184726), T5g, T4z);
 
617
                                   T5h = VFNMS(LDK(KP995184726), T5g, T4z);
 
618
                                   T5F = VFMA(LDK(KP995184726), T5A, T5z);
 
619
                                   T5B = VFNMS(LDK(KP995184726), T5A, T5z);
 
620
                              }
 
621
                         }
 
622
                         {
 
623
                              V T6J, T6R, T6L, T6z, T6T, T6P;
 
624
                              {
 
625
                                   V T6N, T6j, T6O, T6y;
 
626
                                   {
 
627
                                        V T6q, T6H, T5C, T5s, T6I, T6x;
 
628
                                        T6q = VFNMS(LDK(KP534511135), T6p, T6m);
 
629
                                        T6H = VFMA(LDK(KP534511135), T6m, T6p);
 
630
                                        T5C = VFNMS(LDK(KP980785280), T5r, T5o);
 
631
                                        T5s = VFMA(LDK(KP980785280), T5r, T5o);
 
632
                                        T6I = VFMA(LDK(KP534511135), T6t, T6w);
 
633
                                        T6x = VFNMS(LDK(KP534511135), T6w, T6t);
 
634
                                        T6N = VFMA(LDK(KP831469612), T6i, T6b);
 
635
                                        T6j = VFNMS(LDK(KP831469612), T6i, T6b);
 
636
                                        {
 
637
                                             V T5E, T5G, T5y, T5w;
 
638
                                             T5E = VFMA(LDK(KP995184726), T5D, T5C);
 
639
                                             T5G = VFNMS(LDK(KP995184726), T5D, T5C);
 
640
                                             T5y = VFMA(LDK(KP995184726), T5v, T5s);
 
641
                                             T5w = VFNMS(LDK(KP995184726), T5v, T5s);
 
642
                                             T6O = VADD(T6H, T6I);
 
643
                                             T6J = VSUB(T6H, T6I);
 
644
                                             T6R = VSUB(T6q, T6x);
 
645
                                             T6y = VADD(T6q, T6x);
 
646
                                             ST(&(xo[WS(os, 47)]), VFNMSI(T5E, T5B), ovs, &(xo[WS(os, 1)]));
 
647
                                             ST(&(xo[WS(os, 17)]), VFMAI(T5E, T5B), ovs, &(xo[WS(os, 1)]));
 
648
                                             ST(&(xo[WS(os, 49)]), VFMAI(T5G, T5F), ovs, &(xo[WS(os, 1)]));
 
649
                                             ST(&(xo[WS(os, 15)]), VFNMSI(T5G, T5F), ovs, &(xo[WS(os, 1)]));
 
650
                                             ST(&(xo[WS(os, 1)]), VFMAI(T5y, T5x), ovs, &(xo[WS(os, 1)]));
 
651
                                             ST(&(xo[WS(os, 63)]), VFNMSI(T5y, T5x), ovs, &(xo[WS(os, 1)]));
 
652
                                             ST(&(xo[WS(os, 33)]), VFMAI(T5w, T5h), ovs, &(xo[WS(os, 1)]));
 
653
                                             ST(&(xo[WS(os, 31)]), VFNMSI(T5w, T5h), ovs, &(xo[WS(os, 1)]));
 
654
                                        }
 
655
                                   }
 
656
                                   T6L = VFMA(LDK(KP881921264), T6y, T6j);
 
657
                                   T6z = VFNMS(LDK(KP881921264), T6y, T6j);
 
658
                                   T6T = VFMA(LDK(KP881921264), T6O, T6N);
 
659
                                   T6P = VFNMS(LDK(KP881921264), T6O, T6N);
 
660
                              }
 
661
                              {
 
662
                                   V T2H, T2P, T2J, T2B, T2R, T2N;
 
663
                                   {
 
664
                                        V T2L, T2t, T2M, T2A;
 
665
                                        {
 
666
                                             V T2w, T2F, T6Q, T6G, T2G, T2z;
 
667
                                             T2w = VFMA(LDK(KP668178637), T2v, T2u);
 
668
                                             T2F = VFNMS(LDK(KP668178637), T2u, T2v);
 
669
                                             T6Q = VFNMS(LDK(KP831469612), T6F, T6C);
 
670
                                             T6G = VFMA(LDK(KP831469612), T6F, T6C);
 
671
                                             T2G = VFNMS(LDK(KP668178637), T2x, T2y);
 
672
                                             T2z = VFMA(LDK(KP668178637), T2y, T2x);
 
673
                                             T2L = VFNMS(LDK(KP923879532), T2s, T2r);
 
674
                                             T2t = VFMA(LDK(KP923879532), T2s, T2r);
 
675
                                             {
 
676
                                                  V T6S, T6U, T6M, T6K;
 
677
                                                  T6S = VFMA(LDK(KP881921264), T6R, T6Q);
 
678
                                                  T6U = VFNMS(LDK(KP881921264), T6R, T6Q);
 
679
                                                  T6M = VFMA(LDK(KP881921264), T6J, T6G);
 
680
                                                  T6K = VFNMS(LDK(KP881921264), T6J, T6G);
 
681
                                                  T2M = VADD(T2F, T2G);
 
682
                                                  T2H = VSUB(T2F, T2G);
 
683
                                                  T2P = VSUB(T2w, T2z);
 
684
                                                  T2A = VADD(T2w, T2z);
 
685
                                                  ST(&(xo[WS(os, 43)]), VFNMSI(T6S, T6P), ovs, &(xo[WS(os, 1)]));
 
686
                                                  ST(&(xo[WS(os, 21)]), VFMAI(T6S, T6P), ovs, &(xo[WS(os, 1)]));
 
687
                                                  ST(&(xo[WS(os, 53)]), VFMAI(T6U, T6T), ovs, &(xo[WS(os, 1)]));
 
688
                                                  ST(&(xo[WS(os, 11)]), VFNMSI(T6U, T6T), ovs, &(xo[WS(os, 1)]));
 
689
                                                  ST(&(xo[WS(os, 5)]), VFMAI(T6M, T6L), ovs, &(xo[WS(os, 1)]));
 
690
                                                  ST(&(xo[WS(os, 59)]), VFNMSI(T6M, T6L), ovs, &(xo[WS(os, 1)]));
 
691
                                                  ST(&(xo[WS(os, 37)]), VFMAI(T6K, T6z), ovs, &(xo[WS(os, 1)]));
 
692
                                                  ST(&(xo[WS(os, 27)]), VFNMSI(T6K, T6z), ovs, &(xo[WS(os, 1)]));
 
693
                                             }
 
694
                                        }
 
695
                                        T2J = VFMA(LDK(KP831469612), T2A, T2t);
 
696
                                        T2B = VFNMS(LDK(KP831469612), T2A, T2t);
 
697
                                        T2R = VFNMS(LDK(KP831469612), T2M, T2L);
 
698
                                        T2N = VFMA(LDK(KP831469612), T2M, T2L);
 
699
                                   }
 
700
                                   {
 
701
                                        V T61, T5J, T62, T5Q;
 
702
                                        {
 
703
                                             V T5M, T5V, T2O, T2E, T5W, T5P;
 
704
                                             T5M = VFMA(LDK(KP820678790), T5L, T5K);
 
705
                                             T5V = VFNMS(LDK(KP820678790), T5K, T5L);
 
706
                                             T2O = VFMA(LDK(KP923879532), T2D, T2C);
 
707
                                             T2E = VFNMS(LDK(KP923879532), T2D, T2C);
 
708
                                             T5W = VFNMS(LDK(KP820678790), T5N, T5O);
 
709
                                             T5P = VFMA(LDK(KP820678790), T5O, T5N);
 
710
                                             T61 = VFNMS(LDK(KP980785280), T5I, T5H);
 
711
                                             T5J = VFMA(LDK(KP980785280), T5I, T5H);
 
712
                                             {
 
713
                                                  V T2Q, T2S, T2K, T2I;
 
714
                                                  T2Q = VFNMS(LDK(KP831469612), T2P, T2O);
 
715
                                                  T2S = VFMA(LDK(KP831469612), T2P, T2O);
 
716
                                                  T2K = VFMA(LDK(KP831469612), T2H, T2E);
 
717
                                                  T2I = VFNMS(LDK(KP831469612), T2H, T2E);
 
718
                                                  T62 = VADD(T5V, T5W);
 
719
                                                  T5X = VSUB(T5V, T5W);
 
720
                                                  T65 = VSUB(T5M, T5P);
 
721
                                                  T5Q = VADD(T5M, T5P);
 
722
                                                  ST(&(xo[WS(os, 42)]), VFMAI(T2Q, T2N), ovs, &(xo[0]));
 
723
                                                  ST(&(xo[WS(os, 22)]), VFNMSI(T2Q, T2N), ovs, &(xo[0]));
 
724
                                                  ST(&(xo[WS(os, 54)]), VFNMSI(T2S, T2R), ovs, &(xo[0]));
 
725
                                                  ST(&(xo[WS(os, 10)]), VFMAI(T2S, T2R), ovs, &(xo[0]));
 
726
                                                  ST(&(xo[WS(os, 58)]), VFMAI(T2K, T2J), ovs, &(xo[0]));
 
727
                                                  ST(&(xo[WS(os, 6)]), VFNMSI(T2K, T2J), ovs, &(xo[0]));
 
728
                                                  ST(&(xo[WS(os, 26)]), VFMAI(T2I, T2B), ovs, &(xo[0]));
 
729
                                                  ST(&(xo[WS(os, 38)]), VFNMSI(T2I, T2B), ovs, &(xo[0]));
 
730
                                             }
 
731
                                        }
 
732
                                        T5Z = VFMA(LDK(KP773010453), T5Q, T5J);
 
733
                                        T5R = VFNMS(LDK(KP773010453), T5Q, T5J);
 
734
                                        T67 = VFNMS(LDK(KP773010453), T62, T61);
 
735
                                        T63 = VFMA(LDK(KP773010453), T62, T61);
 
736
                                   }
 
737
                              }
 
738
                         }
 
739
                    }
 
740
               }
 
741
          }
 
742
          T5U = VFNMS(LDK(KP980785280), T5T, T5S);
 
743
          T64 = VFMA(LDK(KP980785280), T5T, T5S);
 
744
          {
 
745
               V T68, T66, T5Y, T60;
 
746
               T68 = VFMA(LDK(KP773010453), T65, T64);
 
747
               T66 = VFNMS(LDK(KP773010453), T65, T64);
 
748
               T5Y = VFNMS(LDK(KP773010453), T5X, T5U);
 
749
               T60 = VFMA(LDK(KP773010453), T5X, T5U);
 
750
               ST(&(xo[WS(os, 41)]), VFMAI(T66, T63), ovs, &(xo[WS(os, 1)]));
 
751
               ST(&(xo[WS(os, 23)]), VFNMSI(T66, T63), ovs, &(xo[WS(os, 1)]));
 
752
               ST(&(xo[WS(os, 55)]), VFNMSI(T68, T67), ovs, &(xo[WS(os, 1)]));
 
753
               ST(&(xo[WS(os, 9)]), VFMAI(T68, T67), ovs, &(xo[WS(os, 1)]));
 
754
               ST(&(xo[WS(os, 57)]), VFMAI(T60, T5Z), ovs, &(xo[WS(os, 1)]));
 
755
               ST(&(xo[WS(os, 7)]), VFNMSI(T60, T5Z), ovs, &(xo[WS(os, 1)]));
 
756
               ST(&(xo[WS(os, 25)]), VFMAI(T5Y, T5R), ovs, &(xo[WS(os, 1)]));
 
757
               ST(&(xo[WS(os, 39)]), VFNMSI(T5Y, T5R), ovs, &(xo[WS(os, 1)]));
 
758
          }
 
759
     }
 
760
}
 
761
 
 
762
static const kdft_desc desc = { 64, "n1bv_64", {198, 0, 258, 0}, &GENUS, 0, 0, 0, 0 };
 
763
void X(codelet_n1bv_64) (planner *p) {
 
764
     X(kdft_register) (p, n1bv_64, &desc);
 
765
}
 
766
 
 
767
#else                           /* HAVE_FMA */
 
768
 
 
769
/* Generated by: ../../../genfft/gen_notw_c -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 64 -name n1bv_64 -include n1b.h */
 
770
 
 
771
/*
 
772
 * This function contains 456 FP additions, 124 FP multiplications,
 
773
 * (or, 404 additions, 72 multiplications, 52 fused multiply/add),
 
774
 * 108 stack variables, and 128 memory accesses
 
775
 */
 
776
/*
 
777
 * Generator Id's : 
 
778
 * $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
 
779
 * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
 
780
 * $Id: gen_notw_c.ml,v 1.17 2006-02-12 23:34:12 athena Exp $
 
781
 */
 
782
 
 
783
#include "n1b.h"
 
784
 
 
785
static void n1bv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
 
786
{
 
787
     DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
 
788
     DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
 
789
     DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
 
790
     DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
 
791
     DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
 
792
     DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
 
793
     DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
 
794
     DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
 
795
     DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
 
796
     DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
 
797
     DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
 
798
     DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
 
799
     DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
 
800
     DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
 
801
     DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
 
802
     INT i;
 
803
     const R *xi;
 
804
     R *xo;
 
805
     xi = ii;
 
806
     xo = io;
 
807
     for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
 
808
          V T4p, T5u, Tb, T3A, T2q, T3v, T6G, T78, Tq, T3w, T6B, T79, T2l, T3B, T4w;
 
809
          V T5r, TI, T2g, T6u, T74, T3q, T3D, T4E, T5o, TZ, T2h, T6x, T75, T3t, T3E;
 
810
          V T4L, T5p, T23, T2N, T6m, T70, T6p, T71, T2c, T2O, T3i, T3Y, T5f, T5R, T5k;
 
811
          V T5S, T3l, T3Z, T1s, T2K, T6f, T6X, T6i, T6Y, T1B, T2L, T3b, T3V, T4Y, T5O;
 
812
          V T53, T5P, T3e, T3W;
 
813
          {
 
814
               V T3, T4n, T2p, T4o, T6, T5s, T9, T5t;
 
815
               {
 
816
                    V T1, T2, T2n, T2o;
 
817
                    T1 = LD(&(xi[0]), ivs, &(xi[0]));
 
818
                    T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
 
819
                    T3 = VSUB(T1, T2);
 
820
                    T4n = VADD(T1, T2);
 
821
                    T2n = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
 
822
                    T2o = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
 
823
                    T2p = VSUB(T2n, T2o);
 
824
                    T4o = VADD(T2n, T2o);
 
825
               }
 
826
               {
 
827
                    V T4, T5, T7, T8;
 
828
                    T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
 
829
                    T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
 
830
                    T6 = VSUB(T4, T5);
 
831
                    T5s = VADD(T4, T5);
 
832
                    T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
 
833
                    T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
 
834
                    T9 = VSUB(T7, T8);
 
835
                    T5t = VADD(T7, T8);
 
836
               }
 
837
               T4p = VSUB(T4n, T4o);
 
838
               T5u = VSUB(T5s, T5t);
 
839
               {
 
840
                    V Ta, T2m, T6E, T6F;
 
841
                    Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
 
842
                    Tb = VSUB(T3, Ta);
 
843
                    T3A = VADD(T3, Ta);
 
844
                    T2m = VMUL(LDK(KP707106781), VSUB(T6, T9));
 
845
                    T2q = VSUB(T2m, T2p);
 
846
                    T3v = VADD(T2p, T2m);
 
847
                    T6E = VADD(T4n, T4o);
 
848
                    T6F = VADD(T5s, T5t);
 
849
                    T6G = VSUB(T6E, T6F);
 
850
                    T78 = VADD(T6E, T6F);
 
851
               }
 
852
          }
 
853
          {
 
854
               V Te, T4q, To, T4t, Th, T4r, Tl, T4u;
 
855
               {
 
856
                    V Tc, Td, Tm, Tn;
 
857
                    Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
 
858
                    Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
 
859
                    Te = VSUB(Tc, Td);
 
860
                    T4q = VADD(Tc, Td);
 
861
                    Tm = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
 
862
                    Tn = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
 
863
                    To = VSUB(Tm, Tn);
 
864
                    T4t = VADD(Tm, Tn);
 
865
               }
 
866
               {
 
867
                    V Tf, Tg, Tj, Tk;
 
868
                    Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
 
869
                    Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
 
870
                    Th = VSUB(Tf, Tg);
 
871
                    T4r = VADD(Tf, Tg);
 
872
                    Tj = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
 
873
                    Tk = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
 
874
                    Tl = VSUB(Tj, Tk);
 
875
                    T4u = VADD(Tj, Tk);
 
876
               }
 
877
               {
 
878
                    V Ti, Tp, T6z, T6A;
 
879
                    Ti = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
 
880
                    Tp = VFNMS(LDK(KP382683432), To, VMUL(LDK(KP923879532), Tl));
 
881
                    Tq = VSUB(Ti, Tp);
 
882
                    T3w = VADD(Ti, Tp);
 
883
                    T6z = VADD(T4q, T4r);
 
884
                    T6A = VADD(T4t, T4u);
 
885
                    T6B = VSUB(T6z, T6A);
 
886
                    T79 = VADD(T6z, T6A);
 
887
               }
 
888
               {
 
889
                    V T2j, T2k, T4s, T4v;
 
890
                    T2j = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
 
891
                    T2k = VFMA(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
 
892
                    T2l = VSUB(T2j, T2k);
 
893
                    T3B = VADD(T2j, T2k);
 
894
                    T4s = VSUB(T4q, T4r);
 
895
                    T4v = VSUB(T4t, T4u);
 
896
                    T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v));
 
897
                    T5r = VMUL(LDK(KP707106781), VSUB(T4s, T4v));
 
898
               }
 
899
          }
 
900
          {
 
901
               V TB, T4z, TF, T4y, Ty, T4C, TG, T4B;
 
902
               {
 
903
                    V Tz, TA, TD, TE;
 
904
                    Tz = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
 
905
                    TA = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
 
906
                    TB = VSUB(Tz, TA);
 
907
                    T4z = VADD(Tz, TA);
 
908
                    TD = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
 
909
                    TE = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
 
910
                    TF = VSUB(TD, TE);
 
911
                    T4y = VADD(TD, TE);
 
912
                    {
 
913
                         V Ts, Tt, Tu, Tv, Tw, Tx;
 
914
                         Ts = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
 
915
                         Tt = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
 
916
                         Tu = VSUB(Ts, Tt);
 
917
                         Tv = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
 
918
                         Tw = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
 
919
                         Tx = VSUB(Tv, Tw);
 
920
                         Ty = VMUL(LDK(KP707106781), VSUB(Tu, Tx));
 
921
                         T4C = VADD(Tv, Tw);
 
922
                         TG = VMUL(LDK(KP707106781), VADD(Tu, Tx));
 
923
                         T4B = VADD(Ts, Tt);
 
924
                    }
 
925
               }
 
926
               {
 
927
                    V TC, TH, T6s, T6t;
 
928
                    TC = VSUB(Ty, TB);
 
929
                    TH = VSUB(TF, TG);
 
930
                    TI = VFMA(LDK(KP831469612), TC, VMUL(LDK(KP555570233), TH));
 
931
                    T2g = VFNMS(LDK(KP555570233), TC, VMUL(LDK(KP831469612), TH));
 
932
                    T6s = VADD(T4y, T4z);
 
933
                    T6t = VADD(T4B, T4C);
 
934
                    T6u = VSUB(T6s, T6t);
 
935
                    T74 = VADD(T6s, T6t);
 
936
               }
 
937
               {
 
938
                    V T3o, T3p, T4A, T4D;
 
939
                    T3o = VADD(TB, Ty);
 
940
                    T3p = VADD(TF, TG);
 
941
                    T3q = VFMA(LDK(KP980785280), T3o, VMUL(LDK(KP195090322), T3p));
 
942
                    T3D = VFNMS(LDK(KP195090322), T3o, VMUL(LDK(KP980785280), T3p));
 
943
                    T4A = VSUB(T4y, T4z);
 
944
                    T4D = VSUB(T4B, T4C);
 
945
                    T4E = VFMA(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D));
 
946
                    T5o = VFNMS(LDK(KP382683432), T4D, VMUL(LDK(KP923879532), T4A));
 
947
               }
 
948
          }
 
949
          {
 
950
               V TS, T4J, TW, T4I, TP, T4G, TX, T4F;
 
951
               {
 
952
                    V TQ, TR, TU, TV;
 
953
                    TQ = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
 
954
                    TR = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
 
955
                    TS = VSUB(TQ, TR);
 
956
                    T4J = VADD(TQ, TR);
 
957
                    TU = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
 
958
                    TV = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
 
959
                    TW = VSUB(TU, TV);
 
960
                    T4I = VADD(TU, TV);
 
961
                    {
 
962
                         V TJ, TK, TL, TM, TN, TO;
 
963
                         TJ = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
 
964
                         TK = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
 
965
                         TL = VSUB(TJ, TK);
 
966
                         TM = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
 
967
                         TN = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
 
968
                         TO = VSUB(TM, TN);
 
969
                         TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
 
970
                         T4G = VADD(TM, TN);
 
971
                         TX = VMUL(LDK(KP707106781), VADD(TL, TO));
 
972
                         T4F = VADD(TJ, TK);
 
973
                    }
 
974
               }
 
975
               {
 
976
                    V TT, TY, T6v, T6w;
 
977
                    TT = VSUB(TP, TS);
 
978
                    TY = VSUB(TW, TX);
 
979
                    TZ = VFNMS(LDK(KP555570233), TY, VMUL(LDK(KP831469612), TT));
 
980
                    T2h = VFMA(LDK(KP555570233), TT, VMUL(LDK(KP831469612), TY));
 
981
                    T6v = VADD(T4I, T4J);
 
982
                    T6w = VADD(T4F, T4G);
 
983
                    T6x = VSUB(T6v, T6w);
 
984
                    T75 = VADD(T6v, T6w);
 
985
               }
 
986
               {
 
987
                    V T3r, T3s, T4H, T4K;
 
988
                    T3r = VADD(TS, TP);
 
989
                    T3s = VADD(TW, TX);
 
990
                    T3t = VFNMS(LDK(KP195090322), T3s, VMUL(LDK(KP980785280), T3r));
 
991
                    T3E = VFMA(LDK(KP195090322), T3r, VMUL(LDK(KP980785280), T3s));
 
992
                    T4H = VSUB(T4F, T4G);
 
993
                    T4K = VSUB(T4I, T4J);
 
994
                    T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H));
 
995
                    T5p = VFMA(LDK(KP923879532), T4K, VMUL(LDK(KP382683432), T4H));
 
996
               }
 
997
          }
 
998
          {
 
999
               V T21, T5h, T26, T5g, T1Y, T5d, T27, T5c, T55, T56, T1J, T57, T29, T58, T59;
 
1000
               V T1Q, T5a, T2a;
 
1001
               {
 
1002
                    V T1Z, T20, T24, T25;
 
1003
                    T1Z = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
 
1004
                    T20 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
 
1005
                    T21 = VSUB(T1Z, T20);
 
1006
                    T5h = VADD(T1Z, T20);
 
1007
                    T24 = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
 
1008
                    T25 = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
 
1009
                    T26 = VSUB(T24, T25);
 
1010
                    T5g = VADD(T24, T25);
 
1011
               }
 
1012
               {
 
1013
                    V T1S, T1T, T1U, T1V, T1W, T1X;
 
1014
                    T1S = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
 
1015
                    T1T = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
 
1016
                    T1U = VSUB(T1S, T1T);
 
1017
                    T1V = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
 
1018
                    T1W = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
 
1019
                    T1X = VSUB(T1V, T1W);
 
1020
                    T1Y = VMUL(LDK(KP707106781), VSUB(T1U, T1X));
 
1021
                    T5d = VADD(T1V, T1W);
 
1022
                    T27 = VMUL(LDK(KP707106781), VADD(T1U, T1X));
 
1023
                    T5c = VADD(T1S, T1T);
 
1024
               }
 
1025
               {
 
1026
                    V T1F, T1I, T1M, T1P;
 
1027
                    {
 
1028
                         V T1D, T1E, T1G, T1H;
 
1029
                         T1D = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
 
1030
                         T1E = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
 
1031
                         T1F = VSUB(T1D, T1E);
 
1032
                         T55 = VADD(T1D, T1E);
 
1033
                         T1G = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
 
1034
                         T1H = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
 
1035
                         T1I = VSUB(T1G, T1H);
 
1036
                         T56 = VADD(T1G, T1H);
 
1037
                    }
 
1038
                    T1J = VFNMS(LDK(KP382683432), T1I, VMUL(LDK(KP923879532), T1F));
 
1039
                    T57 = VSUB(T55, T56);
 
1040
                    T29 = VFMA(LDK(KP382683432), T1F, VMUL(LDK(KP923879532), T1I));
 
1041
                    {
 
1042
                         V T1K, T1L, T1N, T1O;
 
1043
                         T1K = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
 
1044
                         T1L = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
 
1045
                         T1M = VSUB(T1K, T1L);
 
1046
                         T58 = VADD(T1K, T1L);
 
1047
                         T1N = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
 
1048
                         T1O = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
 
1049
                         T1P = VSUB(T1N, T1O);
 
1050
                         T59 = VADD(T1N, T1O);
 
1051
                    }
 
1052
                    T1Q = VFMA(LDK(KP923879532), T1M, VMUL(LDK(KP382683432), T1P));
 
1053
                    T5a = VSUB(T58, T59);
 
1054
                    T2a = VFNMS(LDK(KP382683432), T1M, VMUL(LDK(KP923879532), T1P));
 
1055
               }
 
1056
               {
 
1057
                    V T1R, T22, T6k, T6l;
 
1058
                    T1R = VSUB(T1J, T1Q);
 
1059
                    T22 = VSUB(T1Y, T21);
 
1060
                    T23 = VSUB(T1R, T22);
 
1061
                    T2N = VADD(T22, T1R);
 
1062
                    T6k = VADD(T5g, T5h);
 
1063
                    T6l = VADD(T5c, T5d);
 
1064
                    T6m = VSUB(T6k, T6l);
 
1065
                    T70 = VADD(T6k, T6l);
 
1066
               }
 
1067
               {
 
1068
                    V T6n, T6o, T28, T2b;
 
1069
                    T6n = VADD(T55, T56);
 
1070
                    T6o = VADD(T58, T59);
 
1071
                    T6p = VSUB(T6n, T6o);
 
1072
                    T71 = VADD(T6n, T6o);
 
1073
                    T28 = VSUB(T26, T27);
 
1074
                    T2b = VSUB(T29, T2a);
 
1075
                    T2c = VSUB(T28, T2b);
 
1076
                    T2O = VADD(T28, T2b);
 
1077
               }
 
1078
               {
 
1079
                    V T3g, T3h, T5b, T5e;
 
1080
                    T3g = VADD(T26, T27);
 
1081
                    T3h = VADD(T1J, T1Q);
 
1082
                    T3i = VADD(T3g, T3h);
 
1083
                    T3Y = VSUB(T3g, T3h);
 
1084
                    T5b = VMUL(LDK(KP707106781), VSUB(T57, T5a));
 
1085
                    T5e = VSUB(T5c, T5d);
 
1086
                    T5f = VSUB(T5b, T5e);
 
1087
                    T5R = VADD(T5e, T5b);
 
1088
               }
 
1089
               {
 
1090
                    V T5i, T5j, T3j, T3k;
 
1091
                    T5i = VSUB(T5g, T5h);
 
1092
                    T5j = VMUL(LDK(KP707106781), VADD(T57, T5a));
 
1093
                    T5k = VSUB(T5i, T5j);
 
1094
                    T5S = VADD(T5i, T5j);
 
1095
                    T3j = VADD(T21, T1Y);
 
1096
                    T3k = VADD(T29, T2a);
 
1097
                    T3l = VADD(T3j, T3k);
 
1098
                    T3Z = VSUB(T3k, T3j);
 
1099
               }
 
1100
          }
 
1101
          {
 
1102
               V T1q, T50, T1v, T4Z, T1n, T4W, T1w, T4V, T4O, T4P, T18, T4Q, T1y, T4R, T4S;
 
1103
               V T1f, T4T, T1z;
 
1104
               {
 
1105
                    V T1o, T1p, T1t, T1u;
 
1106
                    T1o = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
 
1107
                    T1p = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
 
1108
                    T1q = VSUB(T1o, T1p);
 
1109
                    T50 = VADD(T1o, T1p);
 
1110
                    T1t = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
 
1111
                    T1u = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
 
1112
                    T1v = VSUB(T1t, T1u);
 
1113
                    T4Z = VADD(T1t, T1u);
 
1114
               }
 
1115
               {
 
1116
                    V T1h, T1i, T1j, T1k, T1l, T1m;
 
1117
                    T1h = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
 
1118
                    T1i = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
 
1119
                    T1j = VSUB(T1h, T1i);
 
1120
                    T1k = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
 
1121
                    T1l = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
 
1122
                    T1m = VSUB(T1k, T1l);
 
1123
                    T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m));
 
1124
                    T4W = VADD(T1k, T1l);
 
1125
                    T1w = VMUL(LDK(KP707106781), VADD(T1j, T1m));
 
1126
                    T4V = VADD(T1h, T1i);
 
1127
               }
 
1128
               {
 
1129
                    V T14, T17, T1b, T1e;
 
1130
                    {
 
1131
                         V T12, T13, T15, T16;
 
1132
                         T12 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
 
1133
                         T13 = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
 
1134
                         T14 = VSUB(T12, T13);
 
1135
                         T4O = VADD(T12, T13);
 
1136
                         T15 = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
 
1137
                         T16 = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
 
1138
                         T17 = VSUB(T15, T16);
 
1139
                         T4P = VADD(T15, T16);
 
1140
                    }
 
1141
                    T18 = VFNMS(LDK(KP382683432), T17, VMUL(LDK(KP923879532), T14));
 
1142
                    T4Q = VSUB(T4O, T4P);
 
1143
                    T1y = VFMA(LDK(KP382683432), T14, VMUL(LDK(KP923879532), T17));
 
1144
                    {
 
1145
                         V T19, T1a, T1c, T1d;
 
1146
                         T19 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
 
1147
                         T1a = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
 
1148
                         T1b = VSUB(T19, T1a);
 
1149
                         T4R = VADD(T19, T1a);
 
1150
                         T1c = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
 
1151
                         T1d = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
 
1152
                         T1e = VSUB(T1c, T1d);
 
1153
                         T4S = VADD(T1c, T1d);
 
1154
                    }
 
1155
                    T1f = VFMA(LDK(KP923879532), T1b, VMUL(LDK(KP382683432), T1e));
 
1156
                    T4T = VSUB(T4R, T4S);
 
1157
                    T1z = VFNMS(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e));
 
1158
               }
 
1159
               {
 
1160
                    V T1g, T1r, T6d, T6e;
 
1161
                    T1g = VSUB(T18, T1f);
 
1162
                    T1r = VSUB(T1n, T1q);
 
1163
                    T1s = VSUB(T1g, T1r);
 
1164
                    T2K = VADD(T1r, T1g);
 
1165
                    T6d = VADD(T4Z, T50);
 
1166
                    T6e = VADD(T4V, T4W);
 
1167
                    T6f = VSUB(T6d, T6e);
 
1168
                    T6X = VADD(T6d, T6e);
 
1169
               }
 
1170
               {
 
1171
                    V T6g, T6h, T1x, T1A;
 
1172
                    T6g = VADD(T4O, T4P);
 
1173
                    T6h = VADD(T4R, T4S);
 
1174
                    T6i = VSUB(T6g, T6h);
 
1175
                    T6Y = VADD(T6g, T6h);
 
1176
                    T1x = VSUB(T1v, T1w);
 
1177
                    T1A = VSUB(T1y, T1z);
 
1178
                    T1B = VSUB(T1x, T1A);
 
1179
                    T2L = VADD(T1x, T1A);
 
1180
               }
 
1181
               {
 
1182
                    V T39, T3a, T4U, T4X;
 
1183
                    T39 = VADD(T1v, T1w);
 
1184
                    T3a = VADD(T18, T1f);
 
1185
                    T3b = VADD(T39, T3a);
 
1186
                    T3V = VSUB(T39, T3a);
 
1187
                    T4U = VMUL(LDK(KP707106781), VSUB(T4Q, T4T));
 
1188
                    T4X = VSUB(T4V, T4W);
 
1189
                    T4Y = VSUB(T4U, T4X);
 
1190
                    T5O = VADD(T4X, T4U);
 
1191
               }
 
1192
               {
 
1193
                    V T51, T52, T3c, T3d;
 
1194
                    T51 = VSUB(T4Z, T50);
 
1195
                    T52 = VMUL(LDK(KP707106781), VADD(T4Q, T4T));
 
1196
                    T53 = VSUB(T51, T52);
 
1197
                    T5P = VADD(T51, T52);
 
1198
                    T3c = VADD(T1q, T1n);
 
1199
                    T3d = VADD(T1y, T1z);
 
1200
                    T3e = VADD(T3c, T3d);
 
1201
                    T3W = VSUB(T3d, T3c);
 
1202
               }
 
1203
          }
 
1204
          {
 
1205
               V T7h, T7l, T7k, T7m;
 
1206
               {
 
1207
                    V T7f, T7g, T7i, T7j;
 
1208
                    T7f = VADD(T78, T79);
 
1209
                    T7g = VADD(T74, T75);
 
1210
                    T7h = VSUB(T7f, T7g);
 
1211
                    T7l = VADD(T7f, T7g);
 
1212
                    T7i = VADD(T6X, T6Y);
 
1213
                    T7j = VADD(T70, T71);
 
1214
                    T7k = VBYI(VSUB(T7i, T7j));
 
1215
                    T7m = VADD(T7i, T7j);
 
1216
               }
 
1217
               ST(&(xo[WS(os, 48)]), VSUB(T7h, T7k), ovs, &(xo[0]));
 
1218
               ST(&(xo[0]), VADD(T7l, T7m), ovs, &(xo[0]));
 
1219
               ST(&(xo[WS(os, 16)]), VADD(T7h, T7k), ovs, &(xo[0]));
 
1220
               ST(&(xo[WS(os, 32)]), VSUB(T7l, T7m), ovs, &(xo[0]));
 
1221
          }
 
1222
          {
 
1223
               V T76, T7a, T73, T7b, T6Z, T72;
 
1224
               T76 = VSUB(T74, T75);
 
1225
               T7a = VSUB(T78, T79);
 
1226
               T6Z = VSUB(T6X, T6Y);
 
1227
               T72 = VSUB(T70, T71);
 
1228
               T73 = VMUL(LDK(KP707106781), VSUB(T6Z, T72));
 
1229
               T7b = VMUL(LDK(KP707106781), VADD(T6Z, T72));
 
1230
               {
 
1231
                    V T77, T7c, T7d, T7e;
 
1232
                    T77 = VBYI(VSUB(T73, T76));
 
1233
                    T7c = VSUB(T7a, T7b);
 
1234
                    ST(&(xo[WS(os, 24)]), VADD(T77, T7c), ovs, &(xo[0]));
 
1235
                    ST(&(xo[WS(os, 40)]), VSUB(T7c, T77), ovs, &(xo[0]));
 
1236
                    T7d = VBYI(VADD(T76, T73));
 
1237
                    T7e = VADD(T7a, T7b);
 
1238
                    ST(&(xo[WS(os, 8)]), VADD(T7d, T7e), ovs, &(xo[0]));
 
1239
                    ST(&(xo[WS(os, 56)]), VSUB(T7e, T7d), ovs, &(xo[0]));
 
1240
               }
 
1241
          }
 
1242
          {
 
1243
               V T6C, T6S, T6I, T6P, T6r, T6Q, T6L, T6T, T6y, T6H;
 
1244
               T6y = VMUL(LDK(KP707106781), VSUB(T6u, T6x));
 
1245
               T6C = VSUB(T6y, T6B);
 
1246
               T6S = VADD(T6B, T6y);
 
1247
               T6H = VMUL(LDK(KP707106781), VADD(T6u, T6x));
 
1248
               T6I = VSUB(T6G, T6H);
 
1249
               T6P = VADD(T6G, T6H);
 
1250
               {
 
1251
                    V T6j, T6q, T6J, T6K;
 
1252
                    T6j = VFNMS(LDK(KP382683432), T6i, VMUL(LDK(KP923879532), T6f));
 
1253
                    T6q = VFMA(LDK(KP923879532), T6m, VMUL(LDK(KP382683432), T6p));
 
1254
                    T6r = VSUB(T6j, T6q);
 
1255
                    T6Q = VADD(T6j, T6q);
 
1256
                    T6J = VFMA(LDK(KP382683432), T6f, VMUL(LDK(KP923879532), T6i));
 
1257
                    T6K = VFNMS(LDK(KP382683432), T6m, VMUL(LDK(KP923879532), T6p));
 
1258
                    T6L = VSUB(T6J, T6K);
 
1259
                    T6T = VADD(T6J, T6K);
 
1260
               }
 
1261
               {
 
1262
                    V T6D, T6M, T6V, T6W;
 
1263
                    T6D = VBYI(VSUB(T6r, T6C));
 
1264
                    T6M = VSUB(T6I, T6L);
 
1265
                    ST(&(xo[WS(os, 20)]), VADD(T6D, T6M), ovs, &(xo[0]));
 
1266
                    ST(&(xo[WS(os, 44)]), VSUB(T6M, T6D), ovs, &(xo[0]));
 
1267
                    T6V = VSUB(T6P, T6Q);
 
1268
                    T6W = VBYI(VSUB(T6T, T6S));
 
1269
                    ST(&(xo[WS(os, 36)]), VSUB(T6V, T6W), ovs, &(xo[0]));
 
1270
                    ST(&(xo[WS(os, 28)]), VADD(T6V, T6W), ovs, &(xo[0]));
 
1271
               }
 
1272
               {
 
1273
                    V T6N, T6O, T6R, T6U;
 
1274
                    T6N = VBYI(VADD(T6C, T6r));
 
1275
                    T6O = VADD(T6I, T6L);
 
1276
                    ST(&(xo[WS(os, 12)]), VADD(T6N, T6O), ovs, &(xo[0]));
 
1277
                    ST(&(xo[WS(os, 52)]), VSUB(T6O, T6N), ovs, &(xo[0]));
 
1278
                    T6R = VADD(T6P, T6Q);
 
1279
                    T6U = VBYI(VADD(T6S, T6T));
 
1280
                    ST(&(xo[WS(os, 60)]), VSUB(T6R, T6U), ovs, &(xo[0]));
 
1281
                    ST(&(xo[WS(os, 4)]), VADD(T6R, T6U), ovs, &(xo[0]));
 
1282
               }
 
1283
          }
 
1284
          {
 
1285
               V T5N, T68, T61, T69, T5U, T65, T5Y, T66;
 
1286
               {
 
1287
                    V T5L, T5M, T5Z, T60;
 
1288
                    T5L = VADD(T4p, T4w);
 
1289
                    T5M = VADD(T5o, T5p);
 
1290
                    T5N = VSUB(T5L, T5M);
 
1291
                    T68 = VADD(T5L, T5M);
 
1292
                    T5Z = VFNMS(LDK(KP195090322), T5O, VMUL(LDK(KP980785280), T5P));
 
1293
                    T60 = VFMA(LDK(KP195090322), T5R, VMUL(LDK(KP980785280), T5S));
 
1294
                    T61 = VSUB(T5Z, T60);
 
1295
                    T69 = VADD(T5Z, T60);
 
1296
               }
 
1297
               {
 
1298
                    V T5Q, T5T, T5W, T5X;
 
1299
                    T5Q = VFMA(LDK(KP980785280), T5O, VMUL(LDK(KP195090322), T5P));
 
1300
                    T5T = VFNMS(LDK(KP195090322), T5S, VMUL(LDK(KP980785280), T5R));
 
1301
                    T5U = VSUB(T5Q, T5T);
 
1302
                    T65 = VADD(T5Q, T5T);
 
1303
                    T5W = VADD(T4E, T4L);
 
1304
                    T5X = VADD(T5u, T5r);
 
1305
                    T5Y = VSUB(T5W, T5X);
 
1306
                    T66 = VADD(T5X, T5W);
 
1307
               }
 
1308
               {
 
1309
                    V T5V, T62, T6b, T6c;
 
1310
                    T5V = VADD(T5N, T5U);
 
1311
                    T62 = VBYI(VADD(T5Y, T61));
 
1312
                    ST(&(xo[WS(os, 50)]), VSUB(T5V, T62), ovs, &(xo[0]));
 
1313
                    ST(&(xo[WS(os, 14)]), VADD(T5V, T62), ovs, &(xo[0]));
 
1314
                    T6b = VBYI(VADD(T66, T65));
 
1315
                    T6c = VADD(T68, T69);
 
1316
                    ST(&(xo[WS(os, 2)]), VADD(T6b, T6c), ovs, &(xo[0]));
 
1317
                    ST(&(xo[WS(os, 62)]), VSUB(T6c, T6b), ovs, &(xo[0]));
 
1318
               }
 
1319
               {
 
1320
                    V T63, T64, T67, T6a;
 
1321
                    T63 = VSUB(T5N, T5U);
 
1322
                    T64 = VBYI(VSUB(T61, T5Y));
 
1323
                    ST(&(xo[WS(os, 46)]), VSUB(T63, T64), ovs, &(xo[0]));
 
1324
                    ST(&(xo[WS(os, 18)]), VADD(T63, T64), ovs, &(xo[0]));
 
1325
                    T67 = VBYI(VSUB(T65, T66));
 
1326
                    T6a = VSUB(T68, T69);
 
1327
                    ST(&(xo[WS(os, 30)]), VADD(T67, T6a), ovs, &(xo[0]));
 
1328
                    ST(&(xo[WS(os, 34)]), VSUB(T6a, T67), ovs, &(xo[0]));
 
1329
               }
 
1330
          }
 
1331
          {
 
1332
               V T11, T2C, T2v, T2D, T2e, T2z, T2s, T2A;
 
1333
               {
 
1334
                    V Tr, T10, T2t, T2u;
 
1335
                    Tr = VSUB(Tb, Tq);
 
1336
                    T10 = VSUB(TI, TZ);
 
1337
                    T11 = VSUB(Tr, T10);
 
1338
                    T2C = VADD(Tr, T10);
 
1339
                    T2t = VFNMS(LDK(KP471396736), T1s, VMUL(LDK(KP881921264), T1B));
 
1340
                    T2u = VFMA(LDK(KP471396736), T23, VMUL(LDK(KP881921264), T2c));
 
1341
                    T2v = VSUB(T2t, T2u);
 
1342
                    T2D = VADD(T2t, T2u);
 
1343
               }
 
1344
               {
 
1345
                    V T1C, T2d, T2i, T2r;
 
1346
                    T1C = VFMA(LDK(KP881921264), T1s, VMUL(LDK(KP471396736), T1B));
 
1347
                    T2d = VFNMS(LDK(KP471396736), T2c, VMUL(LDK(KP881921264), T23));
 
1348
                    T2e = VSUB(T1C, T2d);
 
1349
                    T2z = VADD(T1C, T2d);
 
1350
                    T2i = VSUB(T2g, T2h);
 
1351
                    T2r = VSUB(T2l, T2q);
 
1352
                    T2s = VSUB(T2i, T2r);
 
1353
                    T2A = VADD(T2r, T2i);
 
1354
               }
 
1355
               {
 
1356
                    V T2f, T2w, T2F, T2G;
 
1357
                    T2f = VADD(T11, T2e);
 
1358
                    T2w = VBYI(VADD(T2s, T2v));
 
1359
                    ST(&(xo[WS(os, 53)]), VSUB(T2f, T2w), ovs, &(xo[WS(os, 1)]));
 
1360
                    ST(&(xo[WS(os, 11)]), VADD(T2f, T2w), ovs, &(xo[WS(os, 1)]));
 
1361
                    T2F = VBYI(VADD(T2A, T2z));
 
1362
                    T2G = VADD(T2C, T2D);
 
1363
                    ST(&(xo[WS(os, 5)]), VADD(T2F, T2G), ovs, &(xo[WS(os, 1)]));
 
1364
                    ST(&(xo[WS(os, 59)]), VSUB(T2G, T2F), ovs, &(xo[WS(os, 1)]));
 
1365
               }
 
1366
               {
 
1367
                    V T2x, T2y, T2B, T2E;
 
1368
                    T2x = VSUB(T11, T2e);
 
1369
                    T2y = VBYI(VSUB(T2v, T2s));
 
1370
                    ST(&(xo[WS(os, 43)]), VSUB(T2x, T2y), ovs, &(xo[WS(os, 1)]));
 
1371
                    ST(&(xo[WS(os, 21)]), VADD(T2x, T2y), ovs, &(xo[WS(os, 1)]));
 
1372
                    T2B = VBYI(VSUB(T2z, T2A));
 
1373
                    T2E = VSUB(T2C, T2D);
 
1374
                    ST(&(xo[WS(os, 27)]), VADD(T2B, T2E), ovs, &(xo[WS(os, 1)]));
 
1375
                    ST(&(xo[WS(os, 37)]), VSUB(T2E, T2B), ovs, &(xo[WS(os, 1)]));
 
1376
               }
 
1377
          }
 
1378
          {
 
1379
               V T3n, T3O, T3J, T3R, T3y, T3Q, T3G, T3N;
 
1380
               {
 
1381
                    V T3f, T3m, T3H, T3I;
 
1382
                    T3f = VFNMS(LDK(KP098017140), T3e, VMUL(LDK(KP995184726), T3b));
 
1383
                    T3m = VFMA(LDK(KP995184726), T3i, VMUL(LDK(KP098017140), T3l));
 
1384
                    T3n = VSUB(T3f, T3m);
 
1385
                    T3O = VADD(T3f, T3m);
 
1386
                    T3H = VFMA(LDK(KP098017140), T3b, VMUL(LDK(KP995184726), T3e));
 
1387
                    T3I = VFNMS(LDK(KP098017140), T3i, VMUL(LDK(KP995184726), T3l));
 
1388
                    T3J = VSUB(T3H, T3I);
 
1389
                    T3R = VADD(T3H, T3I);
 
1390
               }
 
1391
               {
 
1392
                    V T3u, T3x, T3C, T3F;
 
1393
                    T3u = VADD(T3q, T3t);
 
1394
                    T3x = VADD(T3v, T3w);
 
1395
                    T3y = VSUB(T3u, T3x);
 
1396
                    T3Q = VADD(T3x, T3u);
 
1397
                    T3C = VADD(T3A, T3B);
 
1398
                    T3F = VADD(T3D, T3E);
 
1399
                    T3G = VSUB(T3C, T3F);
 
1400
                    T3N = VADD(T3C, T3F);
 
1401
               }
 
1402
               {
 
1403
                    V T3z, T3K, T3T, T3U;
 
1404
                    T3z = VBYI(VSUB(T3n, T3y));
 
1405
                    T3K = VSUB(T3G, T3J);
 
1406
                    ST(&(xo[WS(os, 17)]), VADD(T3z, T3K), ovs, &(xo[WS(os, 1)]));
 
1407
                    ST(&(xo[WS(os, 47)]), VSUB(T3K, T3z), ovs, &(xo[WS(os, 1)]));
 
1408
                    T3T = VSUB(T3N, T3O);
 
1409
                    T3U = VBYI(VSUB(T3R, T3Q));
 
1410
                    ST(&(xo[WS(os, 33)]), VSUB(T3T, T3U), ovs, &(xo[WS(os, 1)]));
 
1411
                    ST(&(xo[WS(os, 31)]), VADD(T3T, T3U), ovs, &(xo[WS(os, 1)]));
 
1412
               }
 
1413
               {
 
1414
                    V T3L, T3M, T3P, T3S;
 
1415
                    T3L = VBYI(VADD(T3y, T3n));
 
1416
                    T3M = VADD(T3G, T3J);
 
1417
                    ST(&(xo[WS(os, 15)]), VADD(T3L, T3M), ovs, &(xo[WS(os, 1)]));
 
1418
                    ST(&(xo[WS(os, 49)]), VSUB(T3M, T3L), ovs, &(xo[WS(os, 1)]));
 
1419
                    T3P = VADD(T3N, T3O);
 
1420
                    T3S = VBYI(VADD(T3Q, T3R));
 
1421
                    ST(&(xo[WS(os, 63)]), VSUB(T3P, T3S), ovs, &(xo[WS(os, 1)]));
 
1422
                    ST(&(xo[WS(os, 1)]), VADD(T3P, T3S), ovs, &(xo[WS(os, 1)]));
 
1423
               }
 
1424
          }
 
1425
          {
 
1426
               V T4N, T5G, T5z, T5H, T5m, T5D, T5w, T5E;
 
1427
               {
 
1428
                    V T4x, T4M, T5x, T5y;
 
1429
                    T4x = VSUB(T4p, T4w);
 
1430
                    T4M = VSUB(T4E, T4L);
 
1431
                    T4N = VSUB(T4x, T4M);
 
1432
                    T5G = VADD(T4x, T4M);
 
1433
                    T5x = VFNMS(LDK(KP555570233), T4Y, VMUL(LDK(KP831469612), T53));
 
1434
                    T5y = VFMA(LDK(KP555570233), T5f, VMUL(LDK(KP831469612), T5k));
 
1435
                    T5z = VSUB(T5x, T5y);
 
1436
                    T5H = VADD(T5x, T5y);
 
1437
               }
 
1438
               {
 
1439
                    V T54, T5l, T5q, T5v;
 
1440
                    T54 = VFMA(LDK(KP831469612), T4Y, VMUL(LDK(KP555570233), T53));
 
1441
                    T5l = VFNMS(LDK(KP555570233), T5k, VMUL(LDK(KP831469612), T5f));
 
1442
                    T5m = VSUB(T54, T5l);
 
1443
                    T5D = VADD(T54, T5l);
 
1444
                    T5q = VSUB(T5o, T5p);
 
1445
                    T5v = VSUB(T5r, T5u);
 
1446
                    T5w = VSUB(T5q, T5v);
 
1447
                    T5E = VADD(T5v, T5q);
 
1448
               }
 
1449
               {
 
1450
                    V T5n, T5A, T5J, T5K;
 
1451
                    T5n = VADD(T4N, T5m);
 
1452
                    T5A = VBYI(VADD(T5w, T5z));
 
1453
                    ST(&(xo[WS(os, 54)]), VSUB(T5n, T5A), ovs, &(xo[0]));
 
1454
                    ST(&(xo[WS(os, 10)]), VADD(T5n, T5A), ovs, &(xo[0]));
 
1455
                    T5J = VBYI(VADD(T5E, T5D));
 
1456
                    T5K = VADD(T5G, T5H);
 
1457
                    ST(&(xo[WS(os, 6)]), VADD(T5J, T5K), ovs, &(xo[0]));
 
1458
                    ST(&(xo[WS(os, 58)]), VSUB(T5K, T5J), ovs, &(xo[0]));
 
1459
               }
 
1460
               {
 
1461
                    V T5B, T5C, T5F, T5I;
 
1462
                    T5B = VSUB(T4N, T5m);
 
1463
                    T5C = VBYI(VSUB(T5z, T5w));
 
1464
                    ST(&(xo[WS(os, 42)]), VSUB(T5B, T5C), ovs, &(xo[0]));
 
1465
                    ST(&(xo[WS(os, 22)]), VADD(T5B, T5C), ovs, &(xo[0]));
 
1466
                    T5F = VBYI(VSUB(T5D, T5E));
 
1467
                    T5I = VSUB(T5G, T5H);
 
1468
                    ST(&(xo[WS(os, 26)]), VADD(T5F, T5I), ovs, &(xo[0]));
 
1469
                    ST(&(xo[WS(os, 38)]), VSUB(T5I, T5F), ovs, &(xo[0]));
 
1470
               }
 
1471
          }
 
1472
          {
 
1473
               V T2J, T34, T2X, T35, T2Q, T31, T2U, T32;
 
1474
               {
 
1475
                    V T2H, T2I, T2V, T2W;
 
1476
                    T2H = VADD(Tb, Tq);
 
1477
                    T2I = VADD(T2g, T2h);
 
1478
                    T2J = VSUB(T2H, T2I);
 
1479
                    T34 = VADD(T2H, T2I);
 
1480
                    T2V = VFNMS(LDK(KP290284677), T2K, VMUL(LDK(KP956940335), T2L));
 
1481
                    T2W = VFMA(LDK(KP290284677), T2N, VMUL(LDK(KP956940335), T2O));
 
1482
                    T2X = VSUB(T2V, T2W);
 
1483
                    T35 = VADD(T2V, T2W);
 
1484
               }
 
1485
               {
 
1486
                    V T2M, T2P, T2S, T2T;
 
1487
                    T2M = VFMA(LDK(KP956940335), T2K, VMUL(LDK(KP290284677), T2L));
 
1488
                    T2P = VFNMS(LDK(KP290284677), T2O, VMUL(LDK(KP956940335), T2N));
 
1489
                    T2Q = VSUB(T2M, T2P);
 
1490
                    T31 = VADD(T2M, T2P);
 
1491
                    T2S = VADD(TI, TZ);
 
1492
                    T2T = VADD(T2q, T2l);
 
1493
                    T2U = VSUB(T2S, T2T);
 
1494
                    T32 = VADD(T2T, T2S);
 
1495
               }
 
1496
               {
 
1497
                    V T2R, T2Y, T37, T38;
 
1498
                    T2R = VADD(T2J, T2Q);
 
1499
                    T2Y = VBYI(VADD(T2U, T2X));
 
1500
                    ST(&(xo[WS(os, 51)]), VSUB(T2R, T2Y), ovs, &(xo[WS(os, 1)]));
 
1501
                    ST(&(xo[WS(os, 13)]), VADD(T2R, T2Y), ovs, &(xo[WS(os, 1)]));
 
1502
                    T37 = VBYI(VADD(T32, T31));
 
1503
                    T38 = VADD(T34, T35);
 
1504
                    ST(&(xo[WS(os, 3)]), VADD(T37, T38), ovs, &(xo[WS(os, 1)]));
 
1505
                    ST(&(xo[WS(os, 61)]), VSUB(T38, T37), ovs, &(xo[WS(os, 1)]));
 
1506
               }
 
1507
               {
 
1508
                    V T2Z, T30, T33, T36;
 
1509
                    T2Z = VSUB(T2J, T2Q);
 
1510
                    T30 = VBYI(VSUB(T2X, T2U));
 
1511
                    ST(&(xo[WS(os, 45)]), VSUB(T2Z, T30), ovs, &(xo[WS(os, 1)]));
 
1512
                    ST(&(xo[WS(os, 19)]), VADD(T2Z, T30), ovs, &(xo[WS(os, 1)]));
 
1513
                    T33 = VBYI(VSUB(T31, T32));
 
1514
                    T36 = VSUB(T34, T35);
 
1515
                    ST(&(xo[WS(os, 29)]), VADD(T33, T36), ovs, &(xo[WS(os, 1)]));
 
1516
                    ST(&(xo[WS(os, 35)]), VSUB(T36, T33), ovs, &(xo[WS(os, 1)]));
 
1517
               }
 
1518
          }
 
1519
          {
 
1520
               V T41, T4g, T4b, T4j, T44, T4i, T48, T4f;
 
1521
               {
 
1522
                    V T3X, T40, T49, T4a;
 
1523
                    T3X = VFNMS(LDK(KP634393284), T3W, VMUL(LDK(KP773010453), T3V));
 
1524
                    T40 = VFMA(LDK(KP773010453), T3Y, VMUL(LDK(KP634393284), T3Z));
 
1525
                    T41 = VSUB(T3X, T40);
 
1526
                    T4g = VADD(T3X, T40);
 
1527
                    T49 = VFMA(LDK(KP634393284), T3V, VMUL(LDK(KP773010453), T3W));
 
1528
                    T4a = VFNMS(LDK(KP634393284), T3Y, VMUL(LDK(KP773010453), T3Z));
 
1529
                    T4b = VSUB(T49, T4a);
 
1530
                    T4j = VADD(T49, T4a);
 
1531
               }
 
1532
               {
 
1533
                    V T42, T43, T46, T47;
 
1534
                    T42 = VSUB(T3D, T3E);
 
1535
                    T43 = VSUB(T3w, T3v);
 
1536
                    T44 = VSUB(T42, T43);
 
1537
                    T4i = VADD(T43, T42);
 
1538
                    T46 = VSUB(T3A, T3B);
 
1539
                    T47 = VSUB(T3q, T3t);
 
1540
                    T48 = VSUB(T46, T47);
 
1541
                    T4f = VADD(T46, T47);
 
1542
               }
 
1543
               {
 
1544
                    V T45, T4c, T4l, T4m;
 
1545
                    T45 = VBYI(VSUB(T41, T44));
 
1546
                    T4c = VSUB(T48, T4b);
 
1547
                    ST(&(xo[WS(os, 23)]), VADD(T45, T4c), ovs, &(xo[WS(os, 1)]));
 
1548
                    ST(&(xo[WS(os, 41)]), VSUB(T4c, T45), ovs, &(xo[WS(os, 1)]));
 
1549
                    T4l = VSUB(T4f, T4g);
 
1550
                    T4m = VBYI(VSUB(T4j, T4i));
 
1551
                    ST(&(xo[WS(os, 39)]), VSUB(T4l, T4m), ovs, &(xo[WS(os, 1)]));
 
1552
                    ST(&(xo[WS(os, 25)]), VADD(T4l, T4m), ovs, &(xo[WS(os, 1)]));
 
1553
               }
 
1554
               {
 
1555
                    V T4d, T4e, T4h, T4k;
 
1556
                    T4d = VBYI(VADD(T44, T41));
 
1557
                    T4e = VADD(T48, T4b);
 
1558
                    ST(&(xo[WS(os, 9)]), VADD(T4d, T4e), ovs, &(xo[WS(os, 1)]));
 
1559
                    ST(&(xo[WS(os, 55)]), VSUB(T4e, T4d), ovs, &(xo[WS(os, 1)]));
 
1560
                    T4h = VADD(T4f, T4g);
 
1561
                    T4k = VBYI(VADD(T4i, T4j));
 
1562
                    ST(&(xo[WS(os, 57)]), VSUB(T4h, T4k), ovs, &(xo[WS(os, 1)]));
 
1563
                    ST(&(xo[WS(os, 7)]), VADD(T4h, T4k), ovs, &(xo[WS(os, 1)]));
 
1564
               }
 
1565
          }
 
1566
     }
 
1567
}
 
1568
 
 
1569
static const kdft_desc desc = { 64, "n1bv_64", {404, 72, 52, 0}, &GENUS, 0, 0, 0, 0 };
 
1570
void X(codelet_n1bv_64) (planner *p) {
 
1571
     X(kdft_register) (p, n1bv_64, &desc);
 
1572
}
 
1573
 
 
1574
#endif                          /* HAVE_FMA */