2
* Copyright (c) 2003, 2006 Matteo Frigo
3
* Copyright (c) 2003, 2006 Massachusetts Institute of Technology
5
* This program is free software; you can redistribute it and/or modify
6
* it under the terms of the GNU General Public License as published by
7
* the Free Software Foundation; either version 2 of the License, or
8
* (at your option) any later version.
10
* This program is distributed in the hope that it will be useful,
11
* but WITHOUT ANY WARRANTY; without even the implied warranty of
12
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13
* GNU General Public License for more details.
15
* You should have received a copy of the GNU General Public License
16
* along with this program; if not, write to the Free Software
17
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21
/* This file was automatically generated --- DO NOT EDIT */
22
/* Generated on Sun Jul 2 15:39:31 EDT 2006 */
24
#include "codelet-rdft.h"
28
/* Generated by: ../../../genfft/gen_r2hc -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 128 -name r2hc_128 -include r2hc.h */
31
* This function contains 956 FP additions, 516 FP multiplications,
32
* (or, 440 additions, 0 multiplications, 516 fused multiply/add),
33
* 229 stack variables, and 256 memory accesses
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_r2hc.ml,v 1.18 2006-02-12 23:34:12 athena Exp $
44
static void r2hc_128(const R *I, R *ro, R *io, stride is, stride ros, stride ios, INT v, INT ivs, INT ovs)
46
DK(KP989176509, +0.989176509964780973451673738016243063983689533);
47
DK(KP803207531, +0.803207531480644909806676512963141923879569427);
48
DK(KP148335987, +0.148335987538347428753676511486911367000625355);
49
DK(KP741650546, +0.741650546272035369581266691172079863842265220);
50
DK(KP998795456, +0.998795456205172392714771604759100694443203615);
51
DK(KP740951125, +0.740951125354959091175616897495162729728955309);
52
DK(KP049126849, +0.049126849769467254105343321271313617079695752);
53
DK(KP906347169, +0.906347169019147157946142717268914412664134293);
54
DK(KP857728610, +0.857728610000272069902269984284770137042490799);
55
DK(KP970031253, +0.970031253194543992603984207286100251456865962);
56
DK(KP599376933, +0.599376933681923766271389869014404232837890546);
57
DK(KP250486960, +0.250486960191305461595702160124721208578685568);
58
DK(KP941544065, +0.941544065183020778412509402599502357185589796);
59
DK(KP903989293, +0.903989293123443331586200297230537048710132025);
60
DK(KP472964775, +0.472964775891319928124438237972992463904131113);
61
DK(KP357805721, +0.357805721314524104672487743774474392487532769);
62
DK(KP773010453, +0.773010453362736960810906609758469800971041293);
63
DK(KP995184726, +0.995184726672196886244836953109479921575474869);
64
DK(KP098491403, +0.098491403357164253077197521291327432293052451);
65
DK(KP820678790, +0.820678790828660330972281985331011598767386482);
66
DK(KP956940335, +0.956940335732208864935797886980269969482849206);
67
DK(KP881921264, +0.881921264348355029712756863660388349508442621);
68
DK(KP534511135, +0.534511135950791641089685961295362908582039528);
69
DK(KP303346683, +0.303346683607342391675883946941299872384187453);
70
DK(KP831469612, +0.831469612302545237078788377617905756738560812);
71
DK(KP980785280, +0.980785280403230449126182236134239036973933731);
72
DK(KP923879532, +0.923879532511286756128183189396788286822416626);
73
DK(KP198912367, +0.198912367379658006911597622644676228597850501);
74
DK(KP668178637, +0.668178637919298919997757686523080761552472251);
75
DK(KP414213562, +0.414213562373095048801688724209698078569671875);
76
DK(KP707106781, +0.707106781186547524400844362104849039284835938);
78
for (i = v; i > 0; i = i - 1, I = I + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(ros), MAKE_VOLATILE_STRIDE(ios)) {
81
E TcD, TdR, T5P, T8v, T27, T7r, Tf, Ta5, T7s, T5S, T8w, T2e, TdS, TcG, Tbn;
82
E Tu, TcK, TdU, TK, Ta6, T7w, T8y, T2o, T5U, TcN, TdV, TZ, Ta7, T7z, T8z;
83
E T2x, T5V, T1g, Taa, Tab, T1v, Tew, TcX, Tex, TcU, T6A, T2M, T9b, T7E, T9a;
84
E T7H, T6z, T2T, TeO, TdK, TeL, Tdz, T9p, T8d, T6O, T5G, T6L, T4X, Tc3, TaV;
85
E Tc4, Tbi, T9s, T8o, TeH, Tdp, TeE, Tde, T9i, T7U, T6H, T4r, T6E, T3I, TbW;
86
E Tao, TbX, TaL, T9l, T85, T1L, Tad, Tae, T20, Tez, Td6, TeA, Td3, T6x, T37;
87
E T9e, T7L, T9d, T7O, T6w, T3e, TbZ, T3Z, T4s, Tc0, TeF, Tds, T4t, T4g, T87;
88
E T80, TeI, Tdl, T86, T7X, TaM, TaD, Tb2, Tc6, T8e, T8f, T5e, T5H, Tb9, Tc7;
89
E TeM, TdN, T5I, T5v, T8q, T8j, TeP, TdG;
93
E T28, Ti, Tn, T2c, Ts, T29, Tl, To;
95
E T4, T23, T3, T25, Td, T5, T8, T9;
114
E Tg, T5N, T6, T24, Ta, Th;
129
T5P = FNMS(KP707106781, T5O, T5N);
130
T8v = FMA(KP707106781, T5O, T5N);
131
T27 = FMA(KP707106781, T26, T23);
132
T7r = FNMS(KP707106781, T26, T23);
152
E T2g, T2l, T2h, TF, TcI, TC, T2i, TI;
156
E Tw, T5Q, T2a, TcE, Tm, T2b, Tp, Tx;
158
T5Q = FMA(KP414213562, T28, T29);
159
T2a = FNMS(KP414213562, T29, T28);
170
T5R = FNMS(KP414213562, T2b, T2c);
171
T2d = FMA(KP414213562, T2c, T2b);
202
E T2p, T2u, T2q, TU, TcL, TR, T2r, TX;
206
E T2k, T7u, T2n, T7v, TL, TM;
215
TcK = FMA(KP414213562, TcJ, TcI);
216
TdU = FNMS(KP414213562, TcI, TcJ);
219
T2k = FMA(KP707106781, T2j, T2g);
220
T7u = FNMS(KP707106781, T2j, T2g);
221
T2n = FMA(KP707106781, T2m, T2l);
222
T7v = FNMS(KP707106781, T2m, T2l);
226
T7w = FMA(KP668178637, T7v, T7u);
227
T8y = FNMS(KP668178637, T7u, T7v);
228
T2o = FNMS(KP198912367, T2n, T2k);
229
T5U = FMA(KP198912367, T2k, T2n);
250
E T2A, T14, T2N, T17, T1b, T1e, T2D, T2O, T1r, T2I, T1q, T2Q, T2H, TcR, T1n;
253
E T2t, T7x, T2w, T7y, T12, T13;
262
TcN = FNMS(KP414213562, TcM, TcL);
263
TdV = FMA(KP414213562, TcL, TcM);
266
T2t = FMA(KP707106781, T2s, T2p);
267
T7x = FNMS(KP707106781, T2s, T2p);
268
T2w = FMA(KP707106781, T2v, T2u);
269
T7y = FNMS(KP707106781, T2v, T2u);
273
T7z = FNMS(KP668178637, T7y, T7x);
274
T8z = FMA(KP668178637, T7x, T7y);
275
T2x = FMA(KP198912367, T2w, T2t);
276
T5V = FNMS(KP198912367, T2t, T2w);
281
E T1c, T2B, T1d, T19, T1a;
284
T1c = I[WS(is, 114)];
291
E T1k, T2F, T1j, T1l, T1h, T1i, T2C;
301
T1l = I[WS(is, 106)];
303
E T1o, T1p, T2G, T1m;
304
T1o = I[WS(is, 122)];
311
T2Q = FMA(KP414213562, T2F, T2G);
312
T2H = FNMS(KP414213562, T2G, T2F);
320
E TcQ, TcV, T2K, T2R, T1u, TcT, TcW, TcS;
322
E T18, T1f, T1t, T2J;
331
T2K = FMA(KP414213562, T2J, T2I);
332
T2R = FNMS(KP414213562, T2I, T2J);
339
E T7C, T2E, T2L, T7D;
340
T7C = FNMS(KP707106781, T2D, T2A);
341
T2E = FMA(KP707106781, T2D, T2A);
344
Tew = FNMS(KP707106781, TcW, TcV);
345
TcX = FMA(KP707106781, TcW, TcV);
346
Tex = FNMS(KP707106781, TcT, TcQ);
347
TcU = FMA(KP707106781, TcT, TcQ);
352
T2P = FMA(KP707106781, T2O, T2N);
353
T7F = FNMS(KP707106781, T2O, T2N);
354
T6A = FNMS(KP923879532, T2L, T2E);
355
T2M = FMA(KP923879532, T2L, T2E);
356
T9b = FNMS(KP923879532, T7D, T7C);
357
T7E = FMA(KP923879532, T7D, T7C);
365
E T83, T84, T8m, T8n;
367
E TaP, T4z, TaQ, T5A, TaS, TaT, T4G, T5B, T4O, T5D, Tbh, Tdw, T4R, Tbc, T4S;
370
E T4x, T4y, T5y, T5z;
371
T4x = I[WS(is, 127)];
372
T9a = FNMS(KP923879532, T7G, T7F);
373
T7H = FMA(KP923879532, T7G, T7F);
374
T6z = FNMS(KP923879532, T2S, T2P);
375
T2T = FMA(KP923879532, T2S, T2P);
380
E T4A, T4B, T4D, T4E;
387
T4D = I[WS(is, 111)];
390
E T4K, Tbf, Tbg, T4N, T4P, T4Q;
392
E T4I, T4C, T4F, T4J, T4L, T4M;
399
T4L = I[WS(is, 103)];
408
T4P = I[WS(is, 119)];
410
T4O = FMA(KP414213562, T4N, T4K);
411
T5D = FNMS(KP414213562, T4K, T4N);
422
E T4H, T8b, TaR, Tdv, TdI, TaU, T4U, Tbd, T5C;
423
T4H = FMA(KP707106781, T4G, T4z);
424
T8b = FNMS(KP707106781, T4G, T4z);
431
T8m = FNMS(KP707106781, T5B, T5A);
432
T5C = FMA(KP707106781, T5B, T5A);
434
E Tbe, Tdx, T5E, T4V;
437
T5E = FMA(KP414213562, T4R, T4U);
438
T4V = FNMS(KP414213562, T4U, T4R);
440
E Tdy, TdJ, T5F, T8c, T4W;
447
TeO = FNMS(KP707106781, TdJ, TdI);
448
TdK = FMA(KP707106781, TdJ, TdI);
449
TeL = FNMS(KP707106781, Tdy, Tdv);
450
Tdz = FMA(KP707106781, Tdy, Tdv);
451
T9p = FNMS(KP923879532, T8c, T8b);
452
T8d = FMA(KP923879532, T8c, T8b);
453
T6O = FNMS(KP923879532, T5F, T5C);
454
T5G = FMA(KP923879532, T5F, T5C);
455
T6L = FNMS(KP923879532, T4W, T4H);
456
T4X = FMA(KP923879532, T4W, T4H);
466
E Tai, T3k, Taj, T4l, Tal, Tam, T4m, T3r, T3D, TaF, T3C, Tdb, TaK, T3z, T4o;
469
E T4j, T4k, T3i, T3j;
473
T9s = FMA(KP923879532, T8n, T8m);
474
T8o = FNMS(KP923879532, T8n, T8m);
479
E T3o, T3n, T3p, T3l, T3m;
482
T3o = I[WS(is, 113)];
489
E T3w, TaI, T3v, T3x, T3t, T3u, T3q;
499
T3x = I[WS(is, 105)];
501
E T3A, T3B, TaJ, T3y;
502
T3A = I[WS(is, 121)];
511
T3z = FNMS(KP414213562, T3y, T3v);
512
T4o = FMA(KP414213562, T3v, T3y);
519
E T3s, T7S, Tak, Tda, Tdn, Tan, T3F, TaG, T4n;
520
T3s = FMA(KP707106781, T3r, T3k);
521
T7S = FNMS(KP707106781, T3r, T3k);
528
T83 = FNMS(KP707106781, T4m, T4l);
529
T4n = FMA(KP707106781, T4m, T4l);
531
E TaH, Tdc, T4p, T3G;
534
T4p = FNMS(KP414213562, T3C, T3F);
535
T3G = FMA(KP414213562, T3F, T3C);
537
E Tdd, Tdo, T4q, T7T, T3H;
544
TeH = FNMS(KP707106781, Tdo, Tdn);
545
Tdp = FMA(KP707106781, Tdo, Tdn);
546
TeE = FNMS(KP707106781, Tdd, Tda);
547
Tde = FMA(KP707106781, Tdd, Tda);
548
T9i = FNMS(KP923879532, T7T, T7S);
549
T7U = FMA(KP923879532, T7T, T7S);
550
T6H = FNMS(KP923879532, T4q, T4n);
551
T4r = FMA(KP923879532, T4q, T4n);
552
T6E = FNMS(KP923879532, T3H, T3s);
553
T3I = FMA(KP923879532, T3H, T3s);
563
E T7N, T3d, T3a, T7M;
565
E T2V, T1z, T38, T1C, T1G, T1J, T2Y, T39, T1W, T33, T1V, T3b, T32, Td0, T1S;
568
E T1A, T1B, T1x, T1y;
569
T1x = I[WS(is, 126)];
572
T9l = FNMS(KP923879532, T84, T83);
573
T85 = FMA(KP923879532, T84, T83);
578
E T1H, T2W, T1I, T1E, T1F;
581
T1H = I[WS(is, 110)];
588
E T1P, T30, T1O, T1Q, T1M, T1N, T2X;
598
T1Q = I[WS(is, 102)];
600
E T1T, T1U, T31, T1R;
601
T1T = I[WS(is, 118)];
608
T3b = FMA(KP414213562, T30, T31);
609
T32 = FNMS(KP414213562, T31, T30);
618
E TcZ, Td4, T35, T3c, T1Z, Td2, Td5, Td1;
620
E T1D, T1K, T1Y, T34;
629
T35 = FMA(KP414213562, T34, T33);
630
T3c = FNMS(KP414213562, T33, T34);
637
E T7J, T2Z, T36, T7K;
638
T7J = FNMS(KP707106781, T2Y, T2V);
639
T2Z = FMA(KP707106781, T2Y, T2V);
642
Tez = FNMS(KP707106781, Td5, Td4);
643
Td6 = FMA(KP707106781, Td5, Td4);
644
TeA = FNMS(KP707106781, Td2, TcZ);
645
Td3 = FMA(KP707106781, Td2, TcZ);
650
T3a = FMA(KP707106781, T39, T38);
651
T7M = FNMS(KP707106781, T39, T38);
652
T6x = FNMS(KP923879532, T36, T2Z);
653
T37 = FMA(KP923879532, T36, T2Z);
654
T9e = FNMS(KP923879532, T7K, T7J);
655
T7L = FMA(KP923879532, T7K, T7J);
660
E Tav, T7V, T7W, TaC;
662
E T3L, T3W, Tdf, Tar, T42, T4d, Tay, Tdi, T46, Tau, Tdg, T3X, T3S, Taz, T45;
665
E T3J, T3K, T3U, T3V;
667
T9d = FNMS(KP923879532, T7N, T7M);
668
T7O = FMA(KP923879532, T7N, T7M);
669
T6w = FNMS(KP923879532, T3d, T3a);
670
T3e = FMA(KP923879532, T3d, T3a);
673
T3V = I[WS(is, 101)];
675
E T40, Tap, Taq, T41, T4b, T4c;
676
T40 = I[WS(is, 125)];
693
E T3M, T3N, T3P, T3Q;
698
T3P = I[WS(is, 117)];
701
E T43, Tas, T3O, Tat, T3R, T44;
708
T46 = I[WS(is, 109)];
719
E Tdq, Tdh, T49, T4e, Tdr, Tdk;
723
E T3T, T3Y, TaA, T48, Tdj, TaB;
724
T3T = FMA(KP707106781, T3S, T3L);
725
T7V = FNMS(KP707106781, T3S, T3L);
726
T7W = FNMS(KP707106781, T3X, T3W);
727
T3Y = FMA(KP707106781, T3X, T3W);
730
Tdq = FNMS(KP414213562, Tdf, Tdg);
731
Tdh = FMA(KP414213562, Tdg, Tdf);
732
T3Z = FNMS(KP198912367, T3Y, T3T);
733
T4s = FMA(KP198912367, T3T, T3Y);
740
Tdr = FMA(KP414213562, Tdi, Tdj);
741
Tdk = FNMS(KP414213562, Tdj, Tdi);
744
E T7Z, T7Y, T4f, T4a;
745
T7Z = FNMS(KP707106781, T4e, T4d);
746
T4f = FMA(KP707106781, T4e, T4d);
747
T4a = FMA(KP707106781, T49, T42);
748
T7Y = FNMS(KP707106781, T49, T42);
751
T4t = FNMS(KP198912367, T4a, T4f);
752
T4g = FMA(KP198912367, T4f, T4a);
753
T87 = FMA(KP668178637, T7Y, T7Z);
754
T80 = FNMS(KP668178637, T7Z, T7Y);
761
E T50, T5b, TdA, TaY, T5h, T5s, Tb5, TdD, T5l, Tb1, TdB, T5c, T57, Tb6, T5k;
764
E T4Y, T4Z, T59, T5a;
766
T86 = FNMS(KP668178637, T7V, T7W);
767
T7X = FMA(KP668178637, T7W, T7V);
774
E T5f, TaW, TaX, T5g, T5q, T5r;
775
T5f = I[WS(is, 123)];
792
E T51, T52, T54, T55;
797
T54 = I[WS(is, 115)];
800
E T5i, TaZ, T53, Tb0, T56, T5j;
807
T5l = I[WS(is, 107)];
818
E TdL, TdC, T5o, T5t, TdM, TdF;
822
E T58, T5d, Tb7, T5n, TdE, Tb8;
823
T58 = FMA(KP707106781, T57, T50);
824
T8e = FNMS(KP707106781, T57, T50);
825
T8f = FNMS(KP707106781, T5c, T5b);
826
T5d = FMA(KP707106781, T5c, T5b);
829
TdL = FNMS(KP414213562, TdA, TdB);
830
TdC = FMA(KP414213562, TdB, TdA);
831
T5e = FMA(KP198912367, T5d, T58);
832
T5H = FNMS(KP198912367, T58, T5d);
839
TdM = FMA(KP414213562, TdD, TdE);
840
TdF = FNMS(KP414213562, TdE, TdD);
843
E T8i, T8h, T5u, T5p;
844
T8i = FNMS(KP707106781, T5t, T5s);
845
T5u = FMA(KP707106781, T5t, T5s);
846
T5p = FMA(KP707106781, T5o, T5h);
847
T8h = FNMS(KP707106781, T5o, T5h);
850
T5I = FMA(KP198912367, T5p, T5u);
851
T5v = FNMS(KP198912367, T5u, T5p);
852
T8q = FNMS(KP668178637, T8h, T8i);
853
T8j = FMA(KP668178637, T8i, T8h);
864
E T8p, T8g, TcH, TdW, TdT, TcO, Tfp, Tfk, Tfj, Tfq;
866
E Tbj, Tba, Tcy, Tco, TcB, Tcl, Tcx, Tcv, Tcz, Tcr;
868
E Tch, Tct, Tcp, Tcq, Tci, T1w, TbV, T11, Tcf, Tc9, T21, Tcj, Tcm, TbY, Tc1;
874
T8p = FMA(KP668178637, T8e, T8f);
875
T8g = FNMS(KP668178637, T8f, T8e);
898
E Tcb, T22, Tce, Tc2;
904
E Tcd, Tcg, Tca, Tcc;
906
ro[WS(ros, 32)] = T11 - T22;
908
io[WS(ios, 32)] = Tcf - Tce;
912
ro[WS(ros, 64)] = Tcd - Tcg;
913
io[WS(ios, 48)] = FMS(KP707106781, Tcc, Tcb);
914
io[WS(ios, 16)] = FMA(KP707106781, Tcc, Tcb);
915
ro[WS(ros, 16)] = FMA(KP707106781, Tca, TbV);
916
ro[WS(ros, 48)] = FNMS(KP707106781, Tca, TbV);
919
Tcy = FNMS(KP414213562, Tcm, Tcn);
920
Tco = FMA(KP414213562, Tcn, Tcm);
923
TcB = FNMS(KP707106781, Tck, Tch);
924
Tcl = FMA(KP707106781, Tck, Tch);
925
Tcx = FMA(KP707106781, Tcu, Tct);
926
Tcv = FNMS(KP707106781, Tcu, Tct);
927
Tcz = FMA(KP414213562, Tcp, Tcq);
928
Tcr = FNMS(KP414213562, Tcq, Tcp);
931
E TbT, TbO, TbN, TbU;
933
E Ta9, TbB, Tbb, TbL, Tbp, TbM, Tag, Tbk, TbR, TbJ, Tbw, TaO, TbC, Tbs, TbQ;
936
E Tbq, Tbr, TbH, TbI;
938
E Tbo, Ta8, Tac, Taf;
942
E TcC, TcA, Tcw, Tcs;
947
ro[WS(ros, 24)] = FMA(KP923879532, TcC, TcB);
948
ro[WS(ros, 40)] = FNMS(KP923879532, TcC, TcB);
949
io[WS(ios, 56)] = FMS(KP923879532, TcA, Tcx);
950
io[WS(ios, 8)] = FMA(KP923879532, TcA, Tcx);
951
io[WS(ios, 40)] = FMA(KP923879532, Tcw, Tcv);
952
io[WS(ios, 24)] = FMS(KP923879532, Tcw, Tcv);
953
ro[WS(ros, 8)] = FMA(KP923879532, Tcs, Tcl);
954
ro[WS(ros, 56)] = FNMS(KP923879532, Tcs, Tcl);
955
Ta9 = FMA(KP707106781, Ta8, Ta5);
956
TbB = FNMS(KP707106781, Ta8, Ta5);
958
Tbq = FNMS(KP414213562, Taa, Tab);
959
Tac = FMA(KP414213562, Tab, Taa);
960
Taf = FNMS(KP414213562, Tae, Tad);
961
Tbr = FMA(KP414213562, Tad, Tae);
962
Tbb = FMA(KP707106781, Tba, TaV);
963
TbH = FNMS(KP707106781, Tba, TaV);
964
TbL = FNMS(KP707106781, Tbo, Tbn);
965
Tbp = FMA(KP707106781, Tbo, Tbn);
968
TbI = FNMS(KP707106781, Tbj, Tbi);
969
Tbk = FMA(KP707106781, Tbj, Tbi);
972
E TbE, TbF, TaE, TaN;
973
TbE = FNMS(KP707106781, TaD, Tao);
974
TaE = FMA(KP707106781, TaD, Tao);
975
TaN = FMA(KP707106781, TaM, TaL);
976
TbF = FNMS(KP707106781, TaM, TaL);
977
TbR = FNMS(KP668178637, TbH, TbI);
978
TbJ = FMA(KP668178637, TbI, TbH);
979
Tbw = FNMS(KP198912367, TaE, TaN);
980
TaO = FMA(KP198912367, TaN, TaE);
983
TbQ = FMA(KP668178637, TbE, TbF);
984
TbG = FNMS(KP668178637, TbF, TbE);
988
E Tbz, Tah, Tbv, Tbt, Tbx, Tbl;
989
Tbz = FNMS(KP923879532, Tag, Ta9);
990
Tah = FMA(KP923879532, Tag, Ta9);
991
Tbv = FMA(KP923879532, Tbs, Tbp);
992
Tbt = FNMS(KP923879532, Tbs, Tbp);
993
Tbx = FMA(KP198912367, Tbb, Tbk);
994
Tbl = FNMS(KP198912367, Tbk, Tbb);
996
E TbD, TbK, TbP, TbS;
997
TbT = FNMS(KP923879532, TbC, TbB);
998
TbD = FMA(KP923879532, TbC, TbB);
1000
E TbA, Tby, Tbu, Tbm;
1005
ro[WS(ros, 28)] = FMA(KP980785280, TbA, Tbz);
1006
ro[WS(ros, 36)] = FNMS(KP980785280, TbA, Tbz);
1007
io[WS(ios, 60)] = FMS(KP980785280, Tby, Tbv);
1008
io[WS(ios, 4)] = FMA(KP980785280, Tby, Tbv);
1009
io[WS(ios, 36)] = FMA(KP980785280, Tbu, Tbt);
1010
io[WS(ios, 28)] = FMS(KP980785280, Tbu, Tbt);
1011
ro[WS(ros, 4)] = FMA(KP980785280, Tbm, Tah);
1012
ro[WS(ros, 60)] = FNMS(KP980785280, Tbm, Tah);
1016
TbN = FMA(KP923879532, TbM, TbL);
1017
TbP = FNMS(KP923879532, TbM, TbL);
1020
ro[WS(ros, 12)] = FMA(KP831469612, TbK, TbD);
1021
ro[WS(ros, 52)] = FNMS(KP831469612, TbK, TbD);
1022
io[WS(ios, 52)] = FNMS(KP831469612, TbS, TbP);
1023
io[WS(ios, 12)] = -(FMA(KP831469612, TbS, TbP));
1028
E TeN, Tf7, Tev, Tfm, Tfc, TeQ, TeX, TeW, Tfn, Tff, Tfi, TeC, Tf2, TeK, Tfh;
1031
E TeG, TeJ, Tfd, Tfe, Tey, TeB, TeT, TeU;
1033
E Tet, Teu, Tfa, Tfb;
1034
TcH = FMA(KP707106781, TcG, TcD);
1035
Tet = FNMS(KP707106781, TcG, TcD);
1036
io[WS(ios, 44)] = FMS(KP831469612, TbO, TbN);
1037
io[WS(ios, 20)] = FMA(KP831469612, TbO, TbN);
1038
ro[WS(ros, 20)] = FMA(KP831469612, TbU, TbT);
1039
ro[WS(ros, 44)] = FNMS(KP831469612, TbU, TbT);
1042
TeG = FNMS(KP923879532, TeF, TeE);
1043
Tfa = FMA(KP923879532, TeF, TeE);
1044
Tfb = FMA(KP923879532, TeI, TeH);
1045
TeJ = FNMS(KP923879532, TeI, TeH);
1046
TeN = FNMS(KP923879532, TeM, TeL);
1047
Tfd = FMA(KP923879532, TeM, TeL);
1048
Tf7 = FMA(KP923879532, Teu, Tet);
1049
Tev = FNMS(KP923879532, Teu, Tet);
1050
Tfm = FMA(KP303346683, Tfa, Tfb);
1051
Tfc = FNMS(KP303346683, Tfb, Tfa);
1052
Tfe = FNMS(KP923879532, TeP, TeO);
1053
TeQ = FMA(KP923879532, TeP, TeO);
1054
TeX = FNMS(KP668178637, Tew, Tex);
1055
Tey = FMA(KP668178637, Tex, Tew);
1056
TeB = FNMS(KP668178637, TeA, Tez);
1057
TeW = FMA(KP668178637, Tez, TeA);
1059
Tfn = FNMS(KP303346683, Tfd, Tfe);
1060
Tff = FMA(KP303346683, Tfe, Tfd);
1063
TdT = FMA(KP707106781, TdS, TdR);
1064
TeT = FNMS(KP707106781, TdS, TdR);
1067
Tf2 = FNMS(KP534511135, TeG, TeJ);
1068
TeK = FMA(KP534511135, TeJ, TeG);
1069
Tfh = FNMS(KP923879532, TeU, TeT);
1070
TeV = FMA(KP923879532, TeU, TeT);
1073
E Tf5, TeD, TeY, Tf3, TeR;
1074
Tf5 = FNMS(KP831469612, TeC, Tev);
1075
TeD = FMA(KP831469612, TeC, Tev);
1078
Tf3 = FMA(KP534511135, TeN, TeQ);
1079
TeR = FNMS(KP534511135, TeQ, TeN);
1081
E Tf1, TeZ, Tf6, Tf4, Tf0, TeS;
1082
Tf1 = FMA(KP831469612, TeY, TeV);
1083
TeZ = FNMS(KP831469612, TeY, TeV);
1088
io[WS(ios, 54)] = FMS(KP881921264, Tf4, Tf1);
1089
io[WS(ios, 10)] = FMA(KP881921264, Tf4, Tf1);
1090
io[WS(ios, 42)] = FMA(KP881921264, Tf0, TeZ);
1091
io[WS(ios, 22)] = FMS(KP881921264, Tf0, TeZ);
1092
ro[WS(ros, 10)] = FMA(KP881921264, TeS, TeD);
1093
ro[WS(ros, 54)] = FNMS(KP881921264, TeS, TeD);
1094
ro[WS(ros, 42)] = FNMS(KP881921264, Tf6, Tf5);
1095
ro[WS(ros, 22)] = FMA(KP881921264, Tf6, Tf5);
1099
E Tf9, Tfg, Tfl, Tfo;
1100
Tfp = FNMS(KP831469612, Tf8, Tf7);
1101
Tf9 = FMA(KP831469612, Tf8, Tf7);
1104
Tfj = FNMS(KP831469612, Tfi, Tfh);
1105
Tfl = FMA(KP831469612, Tfi, Tfh);
1108
ro[WS(ros, 6)] = FMA(KP956940335, Tfg, Tf9);
1109
ro[WS(ros, 58)] = FNMS(KP956940335, Tfg, Tf9);
1110
io[WS(ios, 58)] = FNMS(KP956940335, Tfo, Tfl);
1111
io[WS(ios, 6)] = -(FMA(KP956940335, Tfo, Tfl));
1117
E T2f, T5W, T5T, T2y, T5J, T5w, T4u, T4h, T7p, T7q;
1119
E Ter, Tem, Tel, Tes;
1121
E TdH, Te9, TcP, Teo, Tee, TdO, TdY, TdZ, Tep, Teh, Tek, Td8, Te4, Tdu, Tej;
1124
E Tdm, Tdt, Tef, Teg, TcY, Td7, Tec, Ted;
1125
io[WS(ios, 38)] = FMS(KP956940335, Tfk, Tfj);
1126
io[WS(ios, 26)] = FMA(KP956940335, Tfk, Tfj);
1127
ro[WS(ros, 26)] = FMA(KP956940335, Tfq, Tfp);
1128
ro[WS(ros, 38)] = FNMS(KP956940335, Tfq, Tfp);
1129
Tdm = FMA(KP923879532, Tdl, Tde);
1130
Tec = FNMS(KP923879532, Tdl, Tde);
1131
Ted = FNMS(KP923879532, Tds, Tdp);
1132
Tdt = FMA(KP923879532, Tds, Tdp);
1133
TdH = FMA(KP923879532, TdG, Tdz);
1134
Tef = FNMS(KP923879532, TdG, Tdz);
1135
Te9 = FNMS(KP923879532, TcO, TcH);
1136
TcP = FMA(KP923879532, TcO, TcH);
1137
Teo = FMA(KP820678790, Tec, Ted);
1138
Tee = FNMS(KP820678790, Ted, Tec);
1139
Teg = FNMS(KP923879532, TdN, TdK);
1140
TdO = FMA(KP923879532, TdN, TdK);
1141
TdY = FNMS(KP198912367, TcU, TcX);
1142
TcY = FMA(KP198912367, TcX, TcU);
1143
Td7 = FNMS(KP198912367, Td6, Td3);
1144
TdZ = FMA(KP198912367, Td3, Td6);
1145
Tep = FNMS(KP820678790, Tef, Teg);
1146
Teh = FMA(KP820678790, Teg, Tef);
1149
Te4 = FNMS(KP098491403, Tdm, Tdt);
1150
Tdu = FMA(KP098491403, Tdt, Tdm);
1151
Tej = FNMS(KP923879532, TdW, TdT);
1152
TdX = FMA(KP923879532, TdW, TdT);
1155
E Te7, Td9, Te0, Te5, TdP;
1156
Te7 = FNMS(KP980785280, Td8, TcP);
1157
Td9 = FMA(KP980785280, Td8, TcP);
1160
Te5 = FMA(KP098491403, TdH, TdO);
1161
TdP = FNMS(KP098491403, TdO, TdH);
1163
E Te3, Te1, Te8, Te6, Te2, TdQ;
1164
Te3 = FMA(KP980785280, Te0, TdX);
1165
Te1 = FNMS(KP980785280, Te0, TdX);
1170
io[WS(ios, 62)] = FMS(KP995184726, Te6, Te3);
1171
io[WS(ios, 2)] = FMA(KP995184726, Te6, Te3);
1172
io[WS(ios, 34)] = FMA(KP995184726, Te2, Te1);
1173
io[WS(ios, 30)] = FMS(KP995184726, Te2, Te1);
1174
ro[WS(ros, 2)] = FMA(KP995184726, TdQ, Td9);
1175
ro[WS(ros, 62)] = FNMS(KP995184726, TdQ, Td9);
1176
ro[WS(ros, 34)] = FNMS(KP995184726, Te8, Te7);
1177
ro[WS(ros, 30)] = FMA(KP995184726, Te8, Te7);
1181
E Teb, Tei, Ten, Teq;
1182
Ter = FNMS(KP980785280, Tea, Te9);
1183
Teb = FMA(KP980785280, Tea, Te9);
1186
Tel = FMA(KP980785280, Tek, Tej);
1187
Ten = FNMS(KP980785280, Tek, Tej);
1190
ro[WS(ros, 14)] = FMA(KP773010453, Tei, Teb);
1191
ro[WS(ros, 50)] = FNMS(KP773010453, Tei, Teb);
1192
io[WS(ios, 50)] = FNMS(KP773010453, Teq, Ten);
1193
io[WS(ios, 14)] = -(FMA(KP773010453, Teq, Ten));
1197
E T77, T6v, T7i, T6C, T78, T6Y, T7h, T6V, T6N, T7d, T6P, T6F, T6I;
1199
E T6W, T6X, T6T, T6U, T6M;
1201
E T6t, T6u, T6y, T6B;
1202
T2f = FMA(KP923879532, T2e, T27);
1203
T6t = FNMS(KP923879532, T2e, T27);
1204
io[WS(ios, 46)] = FMS(KP773010453, Tem, Tel);
1205
io[WS(ios, 18)] = FMA(KP773010453, Tem, Tel);
1206
ro[WS(ros, 18)] = FMA(KP773010453, Tes, Ter);
1207
ro[WS(ros, 46)] = FNMS(KP773010453, Tes, Ter);
1210
T6W = FNMS(KP820678790, T6w, T6x);
1211
T6y = FMA(KP820678790, T6x, T6w);
1212
T6B = FNMS(KP820678790, T6A, T6z);
1213
T6X = FMA(KP820678790, T6z, T6A);
1214
T77 = FMA(KP980785280, T6u, T6t);
1215
T6v = FNMS(KP980785280, T6u, T6t);
1219
T5T = FMA(KP923879532, T5S, T5P);
1220
T6T = FNMS(KP923879532, T5S, T5P);
1227
T7h = FMA(KP980785280, T6U, T6T);
1228
T6V = FNMS(KP980785280, T6U, T6T);
1229
T6N = FNMS(KP980785280, T6M, T6L);
1230
T7d = FMA(KP980785280, T6M, T6L);
1239
E T75, T7f, T7n, T7c, T7m, T76;
1241
E T6D, T72, T6R, T73, T6K, T71, T6Z, T7e, T6Q, T74, T70, T6S;
1242
T75 = FNMS(KP773010453, T6C, T6v);
1243
T6D = FMA(KP773010453, T6C, T6v);
1244
T7e = FNMS(KP980785280, T6P, T6O);
1245
T6Q = FMA(KP980785280, T6P, T6O);
1247
E T7a, T6G, T7b, T6J;
1248
T7a = FMA(KP980785280, T6F, T6E);
1249
T6G = FNMS(KP980785280, T6F, T6E);
1250
T7b = FMA(KP980785280, T6I, T6H);
1251
T6J = FNMS(KP980785280, T6I, T6H);
1252
T7f = FMA(KP357805721, T7e, T7d);
1253
T7n = FNMS(KP357805721, T7d, T7e);
1254
T72 = FMA(KP472964775, T6N, T6Q);
1255
T6R = FNMS(KP472964775, T6Q, T6N);
1256
T7c = FMA(KP357805721, T7b, T7a);
1257
T7m = FNMS(KP357805721, T7a, T7b);
1258
T73 = FMA(KP472964775, T6G, T6J);
1259
T6K = FNMS(KP472964775, T6J, T6G);
1261
T71 = FNMS(KP773010453, T6Y, T6V);
1262
T6Z = FMA(KP773010453, T6Y, T6V);
1267
io[WS(ios, 55)] = FMA(KP903989293, T74, T71);
1268
io[WS(ios, 9)] = FMS(KP903989293, T74, T71);
1269
ro[WS(ros, 9)] = FMA(KP903989293, T6S, T6D);
1270
ro[WS(ros, 55)] = FNMS(KP903989293, T6S, T6D);
1271
io[WS(ios, 41)] = FMS(KP903989293, T70, T6Z);
1272
io[WS(ios, 23)] = FMA(KP903989293, T70, T6Z);
1275
E T7k, T7j, T7l, T7o, T79, T7g;
1276
T7p = FNMS(KP773010453, T78, T77);
1277
T79 = FMA(KP773010453, T78, T77);
1280
T7j = FNMS(KP773010453, T7i, T7h);
1281
T7l = FMA(KP773010453, T7i, T7h);
1282
ro[WS(ros, 23)] = FMA(KP903989293, T76, T75);
1283
ro[WS(ros, 41)] = FNMS(KP903989293, T76, T75);
1284
ro[WS(ros, 7)] = FMA(KP941544065, T7g, T79);
1285
ro[WS(ros, 57)] = FNMS(KP941544065, T7g, T79);
1288
io[WS(ios, 57)] = FMS(KP941544065, T7o, T7l);
1289
io[WS(ios, 7)] = FMA(KP941544065, T7o, T7l);
1290
io[WS(ios, 39)] = FMA(KP941544065, T7k, T7j);
1291
io[WS(ios, 25)] = FMS(KP941544065, T7k, T7j);
1297
E T7t, T8A, T8x, T7A, T8r, T8k, T88, T81, Ta3, Ta4, T6r, T6s;
1299
E T9L, T99, T9W, T9g, T9M, T9C, T9V, T9z, T9k, T9O, T9T, Ta0, T9H, T9v, T9m;
1301
E T9B, T9c, T9f, T9A, T97, T98;
1302
T7t = FMA(KP923879532, T7s, T7r);
1303
T97 = FNMS(KP923879532, T7s, T7r);
1306
T9B = FNMS(KP534511135, T9a, T9b);
1307
T9c = FMA(KP534511135, T9b, T9a);
1308
ro[WS(ros, 25)] = FNMS(KP941544065, T7q, T7p);
1309
ro[WS(ros, 39)] = FMA(KP941544065, T7q, T7p);
1310
T9L = FMA(KP831469612, T98, T97);
1311
T99 = FNMS(KP831469612, T98, T97);
1312
T9f = FNMS(KP534511135, T9e, T9d);
1313
T9A = FMA(KP534511135, T9d, T9e);
1315
E T9x, T9y, T9q, T9t;
1316
T8x = FMA(KP923879532, T8w, T8v);
1317
T9x = FNMS(KP923879532, T8w, T8v);
1329
E T9R, T9r, T9S, T9u, T9j;
1332
T9V = FNMS(KP831469612, T9y, T9x);
1333
T9z = FMA(KP831469612, T9y, T9x);
1334
T9R = FMA(KP831469612, T9q, T9p);
1335
T9r = FNMS(KP831469612, T9q, T9p);
1336
T9S = FMA(KP831469612, T9t, T9s);
1337
T9u = FNMS(KP831469612, T9t, T9s);
1338
T9k = FNMS(KP831469612, T9j, T9i);
1339
T9O = FMA(KP831469612, T9j, T9i);
1340
T9T = FNMS(KP250486960, T9S, T9R);
1341
Ta0 = FMA(KP250486960, T9R, T9S);
1342
T9H = FNMS(KP599376933, T9r, T9u);
1343
T9v = FMA(KP599376933, T9u, T9r);
1350
E T9J, T9h, T9F, T9D, T9P, T9n;
1351
T9J = FNMS(KP881921264, T9g, T99);
1352
T9h = FMA(KP881921264, T9g, T99);
1353
T9F = FMA(KP881921264, T9C, T9z);
1354
T9D = FNMS(KP881921264, T9C, T9z);
1355
T9P = FMA(KP831469612, T9m, T9l);
1356
T9n = FNMS(KP831469612, T9m, T9l);
1358
E T9Y, T9X, T9Z, Ta2;
1360
E T9N, Ta1, T9G, T9o, T9U, T9Q;
1361
Ta3 = FNMS(KP881921264, T9M, T9L);
1362
T9N = FMA(KP881921264, T9M, T9L);
1363
T9Q = FNMS(KP250486960, T9P, T9O);
1364
Ta1 = FMA(KP250486960, T9O, T9P);
1365
T9G = FNMS(KP599376933, T9k, T9n);
1366
T9o = FMA(KP599376933, T9n, T9k);
1369
T9X = FNMS(KP881921264, T9W, T9V);
1370
T9Z = FMA(KP881921264, T9W, T9V);
1372
E T9K, T9I, T9E, T9w;
1377
ro[WS(ros, 5)] = FMA(KP970031253, T9U, T9N);
1378
ro[WS(ros, 59)] = FNMS(KP970031253, T9U, T9N);
1379
ro[WS(ros, 21)] = FNMS(KP857728610, T9K, T9J);
1380
ro[WS(ros, 43)] = FMA(KP857728610, T9K, T9J);
1381
io[WS(ios, 53)] = FMS(KP857728610, T9I, T9F);
1382
io[WS(ios, 11)] = FMA(KP857728610, T9I, T9F);
1383
io[WS(ios, 43)] = FMA(KP857728610, T9E, T9D);
1384
io[WS(ios, 21)] = FMS(KP857728610, T9E, T9D);
1385
ro[WS(ros, 11)] = FMA(KP857728610, T9w, T9h);
1386
ro[WS(ros, 53)] = FNMS(KP857728610, T9w, T9h);
1391
io[WS(ios, 59)] = FMA(KP970031253, Ta2, T9Z);
1392
io[WS(ios, 5)] = FMS(KP970031253, Ta2, T9Z);
1393
io[WS(ios, 37)] = FMS(KP970031253, T9Y, T9X);
1394
io[WS(ios, 27)] = FMA(KP970031253, T9Y, T9X);
1399
E T69, T2z, T6k, T3g, T6a, T60, T6j, T5X, T4i, T6c, T6h, T6p, T64, T5L;
1401
E T5Y, T2U, T3f, T5Z;
1402
T5Y = FMA(KP098491403, T2M, T2T);
1403
T2U = FNMS(KP098491403, T2T, T2M);
1404
ro[WS(ros, 27)] = FMA(KP970031253, Ta4, Ta3);
1405
ro[WS(ros, 37)] = FNMS(KP970031253, Ta4, Ta3);
1406
T69 = FNMS(KP980785280, T2y, T2f);
1407
T2z = FMA(KP980785280, T2y, T2f);
1408
T3f = FMA(KP098491403, T3e, T37);
1409
T5Z = FNMS(KP098491403, T37, T3e);
1415
E T6f, T5x, T6g, T5K;
1416
T6j = FNMS(KP980785280, T5W, T5T);
1417
T5X = FMA(KP980785280, T5W, T5T);
1418
T6f = FNMS(KP980785280, T5w, T4X);
1419
T5x = FMA(KP980785280, T5w, T4X);
1420
T6g = FNMS(KP980785280, T5J, T5G);
1421
T5K = FMA(KP980785280, T5J, T5G);
1422
T4i = FMA(KP980785280, T4h, T3I);
1423
T6c = FNMS(KP980785280, T4h, T3I);
1424
T6h = FMA(KP906347169, T6g, T6f);
1425
T6p = FNMS(KP906347169, T6f, T6g);
1426
T64 = FMA(KP049126849, T5x, T5K);
1427
T5L = FNMS(KP049126849, T5K, T5x);
1431
E T67, T3h, T63, T61, T6d, T4v;
1432
T67 = FNMS(KP995184726, T3g, T2z);
1433
T3h = FMA(KP995184726, T3g, T2z);
1434
T63 = FMA(KP995184726, T60, T5X);
1435
T61 = FNMS(KP995184726, T60, T5X);
1436
T6d = FNMS(KP980785280, T4u, T4r);
1437
T4v = FMA(KP980785280, T4u, T4r);
1439
E T6m, T6l, T6n, T6q;
1441
E T6b, T6o, T65, T4w, T6i, T6e;
1442
T6r = FNMS(KP995184726, T6a, T69);
1443
T6b = FMA(KP995184726, T6a, T69);
1444
T6e = FMA(KP906347169, T6d, T6c);
1445
T6o = FNMS(KP906347169, T6c, T6d);
1446
T65 = FMA(KP049126849, T4i, T4v);
1447
T4w = FNMS(KP049126849, T4v, T4i);
1450
T6l = FNMS(KP995184726, T6k, T6j);
1451
T6n = FMA(KP995184726, T6k, T6j);
1453
E T68, T66, T62, T5M;
1458
ro[WS(ros, 15)] = FMA(KP740951125, T6i, T6b);
1459
ro[WS(ros, 49)] = FNMS(KP740951125, T6i, T6b);
1460
ro[WS(ros, 31)] = FMA(KP998795456, T68, T67);
1461
ro[WS(ros, 33)] = FNMS(KP998795456, T68, T67);
1462
io[WS(ios, 63)] = FMA(KP998795456, T66, T63);
1463
io[WS(ios, 1)] = FMS(KP998795456, T66, T63);
1464
io[WS(ios, 33)] = FMS(KP998795456, T62, T61);
1465
io[WS(ios, 31)] = FMA(KP998795456, T62, T61);
1466
ro[WS(ros, 1)] = FMA(KP998795456, T5M, T3h);
1467
ro[WS(ros, 63)] = FNMS(KP998795456, T5M, T3h);
1472
io[WS(ios, 49)] = FMS(KP740951125, T6q, T6n);
1473
io[WS(ios, 15)] = FMA(KP740951125, T6q, T6n);
1474
io[WS(ios, 47)] = FMA(KP740951125, T6m, T6l);
1475
io[WS(ios, 17)] = FMS(KP740951125, T6m, T6l);
1480
E T8N, T7B, T8Y, T7Q, T8O, T8E, T8X, T8B, T82, T8Q, T8V, T92, T8J, T8t;
1482
E T8C, T7I, T7P, T8D;
1483
T8C = FNMS(KP303346683, T7E, T7H);
1484
T7I = FMA(KP303346683, T7H, T7E);
1485
ro[WS(ros, 17)] = FNMS(KP740951125, T6s, T6r);
1486
ro[WS(ros, 47)] = FMA(KP740951125, T6s, T6r);
1487
T8N = FNMS(KP831469612, T7A, T7t);
1488
T7B = FMA(KP831469612, T7A, T7t);
1489
T7P = FNMS(KP303346683, T7O, T7L);
1490
T8D = FMA(KP303346683, T7L, T7O);
1496
E T8T, T8l, T8U, T8s;
1497
T8X = FNMS(KP831469612, T8A, T8x);
1498
T8B = FMA(KP831469612, T8A, T8x);
1499
T8T = FNMS(KP831469612, T8k, T8d);
1500
T8l = FMA(KP831469612, T8k, T8d);
1501
T8U = FNMS(KP831469612, T8r, T8o);
1502
T8s = FMA(KP831469612, T8r, T8o);
1503
T82 = FMA(KP831469612, T81, T7U);
1504
T8Q = FNMS(KP831469612, T81, T7U);
1505
T8V = FNMS(KP741650546, T8U, T8T);
1506
T92 = FMA(KP741650546, T8T, T8U);
1507
T8J = FNMS(KP148335987, T8l, T8s);
1508
T8t = FMA(KP148335987, T8s, T8l);
1512
E T8L, T7R, T8H, T8F, T8R, T89;
1513
T8L = FNMS(KP956940335, T7Q, T7B);
1514
T7R = FMA(KP956940335, T7Q, T7B);
1515
T8H = FMA(KP956940335, T8E, T8B);
1516
T8F = FNMS(KP956940335, T8E, T8B);
1517
T8R = FNMS(KP831469612, T88, T85);
1518
T89 = FMA(KP831469612, T88, T85);
1520
E T90, T8Z, T91, T94;
1522
E T8P, T93, T8I, T8a, T8W, T8S;
1523
T95 = FNMS(KP956940335, T8O, T8N);
1524
T8P = FMA(KP956940335, T8O, T8N);
1525
T8S = FNMS(KP741650546, T8R, T8Q);
1526
T93 = FMA(KP741650546, T8Q, T8R);
1527
T8I = FNMS(KP148335987, T82, T89);
1528
T8a = FMA(KP148335987, T89, T82);
1531
T8Z = FMA(KP956940335, T8Y, T8X);
1532
T91 = FNMS(KP956940335, T8Y, T8X);
1534
E T8M, T8K, T8G, T8u;
1539
ro[WS(ros, 13)] = FMA(KP803207531, T8W, T8P);
1540
ro[WS(ros, 51)] = FNMS(KP803207531, T8W, T8P);
1541
ro[WS(ros, 29)] = FNMS(KP989176509, T8M, T8L);
1542
ro[WS(ros, 35)] = FMA(KP989176509, T8M, T8L);
1543
io[WS(ios, 61)] = FMS(KP989176509, T8K, T8H);
1544
io[WS(ios, 3)] = FMA(KP989176509, T8K, T8H);
1545
io[WS(ios, 35)] = FMA(KP989176509, T8G, T8F);
1546
io[WS(ios, 29)] = FMS(KP989176509, T8G, T8F);
1547
ro[WS(ros, 3)] = FMA(KP989176509, T8u, T7R);
1548
ro[WS(ros, 61)] = FNMS(KP989176509, T8u, T7R);
1553
io[WS(ios, 51)] = FMA(KP803207531, T94, T91);
1554
io[WS(ios, 13)] = FMS(KP803207531, T94, T91);
1555
io[WS(ios, 45)] = FMS(KP803207531, T90, T8Z);
1556
io[WS(ios, 19)] = FMA(KP803207531, T90, T8Z);
1564
ro[WS(ros, 19)] = FMA(KP803207531, T96, T95);
1565
ro[WS(ros, 45)] = FNMS(KP803207531, T96, T95);
1569
static const kr2hc_desc desc = { 128, "r2hc_128", {440, 0, 516, 0}, &GENUS, 0, 0, 0, 0, 0 };
1571
void X(codelet_r2hc_128) (planner *p) {
1572
X(kr2hc_register) (p, r2hc_128, &desc);
1575
#else /* HAVE_FMA */
1577
/* Generated by: ../../../genfft/gen_r2hc -compact -variables 4 -pipeline-latency 4 -n 128 -name r2hc_128 -include r2hc.h */
1580
* This function contains 956 FP additions, 330 FP multiplications,
1581
* (or, 812 additions, 186 multiplications, 144 fused multiply/add),
1582
* 186 stack variables, and 256 memory accesses
1586
* $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
1587
* $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
1588
* $Id: gen_r2hc.ml,v 1.18 2006-02-12 23:34:12 athena Exp $
1593
static void r2hc_128(const R *I, R *ro, R *io, stride is, stride ros, stride ios, INT v, INT ivs, INT ovs)
1595
DK(KP803207531, +0.803207531480644909806676512963141923879569427);
1596
DK(KP595699304, +0.595699304492433343467036528829969889511926338);
1597
DK(KP146730474, +0.146730474455361751658850129646717819706215317);
1598
DK(KP989176509, +0.989176509964780973451673738016243063983689533);
1599
DK(KP740951125, +0.740951125354959091175616897495162729728955309);
1600
DK(KP671558954, +0.671558954847018400625376850427421803228750632);
1601
DK(KP049067674, +0.049067674327418014254954976942682658314745363);
1602
DK(KP998795456, +0.998795456205172392714771604759100694443203615);
1603
DK(KP242980179, +0.242980179903263889948274162077471118320990783);
1604
DK(KP970031253, +0.970031253194543992603984207286100251456865962);
1605
DK(KP514102744, +0.514102744193221726593693838968815772608049120);
1606
DK(KP857728610, +0.857728610000272069902269984284770137042490799);
1607
DK(KP336889853, +0.336889853392220050689253212619147570477766780);
1608
DK(KP941544065, +0.941544065183020778412509402599502357185589796);
1609
DK(KP427555093, +0.427555093430282094320966856888798534304578629);
1610
DK(KP903989293, +0.903989293123443331586200297230537048710132025);
1611
DK(KP098017140, +0.098017140329560601994195563888641845861136673);
1612
DK(KP995184726, +0.995184726672196886244836953109479921575474869);
1613
DK(KP634393284, +0.634393284163645498215171613225493370675687095);
1614
DK(KP773010453, +0.773010453362736960810906609758469800971041293);
1615
DK(KP881921264, +0.881921264348355029712756863660388349508442621);
1616
DK(KP471396736, +0.471396736825997648556387625905254377657460319);
1617
DK(KP956940335, +0.956940335732208864935797886980269969482849206);
1618
DK(KP290284677, +0.290284677254462367636192375817395274691476278);
1619
DK(KP555570233, +0.555570233019602224742830813948532874374937191);
1620
DK(KP831469612, +0.831469612302545237078788377617905756738560812);
1621
DK(KP195090322, +0.195090322016128267848284868477022240927691618);
1622
DK(KP980785280, +0.980785280403230449126182236134239036973933731);
1623
DK(KP382683432, +0.382683432365089771728459984030398866761344562);
1624
DK(KP923879532, +0.923879532511286756128183189396788286822416626);
1625
DK(KP707106781, +0.707106781186547524400844362104849039284835938);
1627
for (i = v; i > 0; i = i - 1, I = I + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(ros), MAKE_VOLATILE_STRIDE(ios)) {
1628
E TcD, TdU, T27, T7r, T5S, T8y, Tf, Ta5, Tu, Tbq, TcG, TdV, T2e, T8z, T5V;
1629
E T7s, TK, Ta6, TcK, TdX, T2o, T5X, T7w, T8B, TZ, Ta7, TcN, TdY, T2x, T5Y;
1630
E T7z, T8C, T1g, Taa, TcU, TeA, TcX, Tez, T1v, Tab, T2M, T6z, T7E, T9e, T7H;
1631
E T9d, T2T, T6A, T4X, T6L, Tdz, TeL, TdK, TeP, T5G, T6P, T8d, T9p, TaV, Tc3;
1632
E Tbi, Tc4, T8o, T9t, T3I, T6H, Tde, TeH, Tdp, TeF, T4r, T6F, T7U, T9l, Tao;
1633
E TbW, TaL, TbX, T85, T9j, T1L, Tad, Td3, Tew, Td6, Tex, T20, Tae, T37, T6x;
1634
E T7L, T9a, T7O, T9b, T3e, T6w, TbZ, Tc0, T3Z, T4s, Tds, TeI, T4g, T4t, T80;
1635
E T87, Tdl, TeE, T7X, T86, TaD, TaM, Tc6, Tc7, T5e, T5H, TdN, TeM, T5v, T5I;
1636
E T8j, T8q, TdG, TeO, T8g, T8p, Tba, Tbj;
1638
E T3, T23, Td, T25, T6, T5R, Ta, T24;
1645
Tb = I[WS(is, 112)];
1665
T26 = KP707106781 * (T24 + T25);
1668
T5Q = KP707106781 * (T25 - T24);
1678
E Ti, T28, Ts, T2c, Tl, T29, Tp, T2b;
1693
Tk = I[WS(is, 104)];
1696
Tn = I[WS(is, 120)];
1709
TcG = KP707106781 * (TcE + TcF);
1710
TdV = KP707106781 * (TcF - TcE);
1713
E T2a, T2d, T5T, T5U;
1714
T2a = FNMS(KP382683432, T29, KP923879532 * T28);
1715
T2d = FMA(KP923879532, T2b, KP382683432 * T2c);
1718
T5T = FNMS(KP923879532, T2c, KP382683432 * T2b);
1719
T5U = FMA(KP382683432, T28, KP923879532 * T29);
1725
E Ty, T2g, TB, T2m, TF, T2l, TI, T2j;
1733
TA = I[WS(is, 100)];
1737
E TD, TE, T2h, TG, TH, T2i;
1741
TG = I[WS(is, 116)];
1745
T2l = KP707106781 * (T2i - T2h);
1747
T2j = KP707106781 * (T2h + T2i);
1758
TcK = FMA(KP923879532, TcI, KP382683432 * TcJ);
1759
TdX = FNMS(KP382683432, TcI, KP923879532 * TcJ);
1762
E T2k, T2n, T7u, T7v;
1765
T2o = FMA(KP980785280, T2k, KP195090322 * T2n);
1766
T5X = FNMS(KP195090322, T2k, KP980785280 * T2n);
1769
T7w = FMA(KP831469612, T7u, KP555570233 * T7v);
1770
T8B = FNMS(KP555570233, T7u, KP831469612 * T7v);
1774
E TN, T2p, TQ, T2v, TU, T2u, TX, T2s;
1777
TL = I[WS(is, 124)];
1786
E TS, TT, T2q, TV, TW, T2r;
1790
TV = I[WS(is, 108)];
1794
T2u = KP707106781 * (T2r - T2q);
1796
T2s = KP707106781 * (T2q + T2r);
1807
TcN = FNMS(KP382683432, TcM, KP923879532 * TcL);
1808
TdY = FMA(KP382683432, TcL, KP923879532 * TcM);
1811
E T2t, T2w, T7x, T7y;
1814
T2x = FNMS(KP195090322, T2w, KP980785280 * T2t);
1815
T5Y = FMA(KP195090322, T2t, KP980785280 * T2w);
1818
T7z = FNMS(KP555570233, T7y, KP831469612 * T7x);
1819
T8C = FMA(KP555570233, T7x, KP831469612 * T7y);
1823
E T14, T2N, T17, T2D, T1b, T2O, T1e, T2C, T1j, T1m, T2K, TcR, T2Q, T1q, T1t;
1826
E T12, T13, T15, T16;
1828
T13 = I[WS(is, 66)];
1831
T15 = I[WS(is, 34)];
1832
T16 = I[WS(is, 98)];
1837
E T19, T1a, T2B, T1c, T1d, T2A;
1838
T19 = I[WS(is, 18)];
1839
T1a = I[WS(is, 82)];
1841
T1c = I[WS(is, 114)];
1842
T1d = I[WS(is, 50)];
1845
T2O = KP707106781 * (T2B + T2A);
1847
T2C = KP707106781 * (T2A - T2B);
1850
E T2I, T2J, T2F, T2G;
1852
E T1h, T1i, T1k, T1l;
1853
T1h = I[WS(is, 10)];
1854
T1i = I[WS(is, 74)];
1857
T1k = I[WS(is, 42)];
1858
T1l = I[WS(is, 106)];
1862
T2K = FMA(KP382683432, T2I, KP923879532 * T2J);
1864
T2Q = FNMS(KP382683432, T2J, KP923879532 * T2I);
1866
E T1o, T1p, T1r, T1s;
1867
T1o = I[WS(is, 122)];
1868
T1p = I[WS(is, 58)];
1871
T1r = I[WS(is, 26)];
1872
T1s = I[WS(is, 90)];
1876
T2H = FNMS(KP923879532, T2G, KP382683432 * T2F);
1878
T2R = FMA(KP923879532, T2F, KP382683432 * T2G);
1881
E T18, T1f, TcQ, TcT;
1887
TcT = KP707106781 * (TcR + TcS);
1892
E TcV, TcW, T1n, T1u;
1894
TcW = KP707106781 * (TcS - TcR);
1903
E T2E, T2L, T7C, T7D;
1914
E T7F, T7G, T2P, T2S;
1926
E T4z, TaP, T5B, TaQ, T4G, TaT, T5y, TaS, Tbf, Tbg, T4O, Tdw, T5E, Tbc, Tbd;
1929
E T4x, T4y, T5z, T5A;
1930
T4x = I[WS(is, 127)];
1931
T4y = I[WS(is, 63)];
1934
T5z = I[WS(is, 31)];
1935
T5A = I[WS(is, 95)];
1940
E T4A, T4B, T4C, T4D, T4E, T4F;
1941
T4A = I[WS(is, 15)];
1942
T4B = I[WS(is, 79)];
1944
T4D = I[WS(is, 111)];
1945
T4E = I[WS(is, 47)];
1947
T4G = KP707106781 * (T4C + T4F);
1949
T5y = KP707106781 * (T4F - T4C);
1953
E T4K, T4N, T4R, T4U;
1955
E T4I, T4J, T4L, T4M;
1957
T4J = I[WS(is, 71)];
1960
T4L = I[WS(is, 39)];
1961
T4M = I[WS(is, 103)];
1965
T4O = FNMS(KP382683432, T4N, KP923879532 * T4K);
1967
T5E = FMA(KP382683432, T4K, KP923879532 * T4N);
1969
E T4P, T4Q, T4S, T4T;
1970
T4P = I[WS(is, 119)];
1971
T4Q = I[WS(is, 55)];
1974
T4S = I[WS(is, 23)];
1975
T4T = I[WS(is, 87)];
1979
T4V = FMA(KP923879532, T4R, KP382683432 * T4U);
1981
T5D = FNMS(KP923879532, T4U, KP382683432 * T4R);
1984
E T4H, T4W, Tdv, Tdy;
1990
Tdy = KP707106781 * (Tdw + Tdx);
1995
E TdI, TdJ, T5C, T5F;
1997
TdJ = KP707106781 * (Tdx - Tdw);
2006
E T8b, T8c, TaR, TaU;
2017
E Tbe, Tbh, T8m, T8n;
2029
E T3k, Tai, T4m, Taj, T3r, Tam, T4j, Tal, TaI, TaJ, T3z, Tdb, T4p, TaF, TaG;
2032
E T3i, T3j, T4k, T4l;
2034
T3j = I[WS(is, 65)];
2037
T4k = I[WS(is, 33)];
2038
T4l = I[WS(is, 97)];
2043
E T3l, T3m, T3n, T3o, T3p, T3q;
2044
T3l = I[WS(is, 17)];
2045
T3m = I[WS(is, 81)];
2047
T3o = I[WS(is, 113)];
2048
T3p = I[WS(is, 49)];
2050
T3r = KP707106781 * (T3n + T3q);
2052
T4j = KP707106781 * (T3q - T3n);
2056
E T3v, T3y, T3C, T3F;
2058
E T3t, T3u, T3w, T3x;
2060
T3u = I[WS(is, 73)];
2063
T3w = I[WS(is, 41)];
2064
T3x = I[WS(is, 105)];
2068
T3z = FNMS(KP382683432, T3y, KP923879532 * T3v);
2070
T4p = FMA(KP382683432, T3v, KP923879532 * T3y);
2072
E T3A, T3B, T3D, T3E;
2073
T3A = I[WS(is, 121)];
2074
T3B = I[WS(is, 57)];
2077
T3D = I[WS(is, 25)];
2078
T3E = I[WS(is, 89)];
2082
T3G = FMA(KP923879532, T3C, KP382683432 * T3F);
2084
T4o = FNMS(KP923879532, T3F, KP382683432 * T3C);
2087
E T3s, T3H, Tda, Tdd;
2093
Tdd = KP707106781 * (Tdb + Tdc);
2098
E Tdn, Tdo, T4n, T4q;
2100
Tdo = KP707106781 * (Tdc - Tdb);
2109
E T7S, T7T, Tak, Tan;
2120
E TaH, TaK, T83, T84;
2132
E T1z, T2V, T1C, T39, T1G, T38, T1J, T2Y, T1O, T1R, T32, Td0, T3c, T1V, T1Y;
2135
E T1x, T1y, T1A, T1B;
2136
T1x = I[WS(is, 126)];
2137
T1y = I[WS(is, 62)];
2140
T1A = I[WS(is, 30)];
2141
T1B = I[WS(is, 94)];
2146
E T1E, T1F, T2W, T1H, T1I, T2X;
2147
T1E = I[WS(is, 14)];
2148
T1F = I[WS(is, 78)];
2150
T1H = I[WS(is, 110)];
2151
T1I = I[WS(is, 46)];
2154
T38 = KP707106781 * (T2X - T2W);
2156
T2Y = KP707106781 * (T2W + T2X);
2159
E T30, T31, T33, T34;
2161
E T1M, T1N, T1P, T1Q;
2163
T1N = I[WS(is, 70)];
2166
T1P = I[WS(is, 38)];
2167
T1Q = I[WS(is, 102)];
2171
T32 = FNMS(KP382683432, T31, KP923879532 * T30);
2173
T3c = FMA(KP382683432, T30, KP923879532 * T31);
2175
E T1T, T1U, T1W, T1X;
2176
T1T = I[WS(is, 118)];
2177
T1U = I[WS(is, 54)];
2180
T1W = I[WS(is, 22)];
2181
T1X = I[WS(is, 86)];
2185
T35 = FMA(KP923879532, T33, KP382683432 * T34);
2187
T3b = FNMS(KP923879532, T34, KP382683432 * T33);
2190
E T1D, T1K, TcZ, Td2;
2196
Td2 = KP707106781 * (Td0 + Td1);
2201
E Td4, Td5, T1S, T1Z;
2203
Td5 = KP707106781 * (Td1 - Td0);
2212
E T2Z, T36, T7J, T7K;
2223
E T7M, T7N, T3a, T3d;
2235
E T3L, Tdf, T3X, Tar, T42, Tdi, T4e, Tay, T3S, Tdg, T3U, Tau, T49, Tdj, T4b;
2238
E T3J, T3K, Tap, T3V, T3W, Taq;
2240
T3K = I[WS(is, 69)];
2242
T3V = I[WS(is, 37)];
2243
T3W = I[WS(is, 101)];
2251
E T40, T41, Taw, T4c, T4d, Tax;
2252
T40 = I[WS(is, 125)];
2253
T41 = I[WS(is, 61)];
2255
T4c = I[WS(is, 29)];
2256
T4d = I[WS(is, 93)];
2264
E T3O, Tas, T3R, Tat;
2266
E T3M, T3N, T3P, T3Q;
2267
T3M = I[WS(is, 21)];
2268
T3N = I[WS(is, 85)];
2271
T3P = I[WS(is, 117)];
2272
T3Q = I[WS(is, 53)];
2276
T3S = KP707106781 * (T3O + T3R);
2278
T3U = KP707106781 * (T3R - T3O);
2282
E T45, Taz, T48, TaA;
2284
E T43, T44, T46, T47;
2285
T43 = I[WS(is, 13)];
2286
T44 = I[WS(is, 77)];
2289
T46 = I[WS(is, 109)];
2290
T47 = I[WS(is, 45)];
2294
T49 = KP707106781 * (T45 + T48);
2296
T4b = KP707106781 * (T48 - T45);
2302
E T3T, T3Y, Tdq, Tdr;
2305
T3Z = FMA(KP980785280, T3T, KP195090322 * T3Y);
2306
T4s = FNMS(KP195090322, T3T, KP980785280 * T3Y);
2307
Tdq = FNMS(KP382683432, Tdf, KP923879532 * Tdg);
2308
Tdr = FMA(KP382683432, Tdi, KP923879532 * Tdj);
2313
E T4a, T4f, T7Y, T7Z;
2316
T4g = FNMS(KP195090322, T4f, KP980785280 * T4a);
2317
T4t = FMA(KP195090322, T4a, KP980785280 * T4f);
2320
T80 = FNMS(KP555570233, T7Z, KP831469612 * T7Y);
2321
T87 = FMA(KP555570233, T7Y, KP831469612 * T7Z);
2323
Tdh = FMA(KP923879532, Tdf, KP382683432 * Tdg);
2324
Tdk = FNMS(KP382683432, Tdj, KP923879532 * Tdi);
2328
E T7V, T7W, Tav, TaC;
2331
T7X = FMA(KP831469612, T7V, KP555570233 * T7W);
2332
T86 = FNMS(KP555570233, T7V, KP831469612 * T7W);
2335
TaD = KP707106781 * (Tav + TaC);
2336
TaM = KP707106781 * (TaC - Tav);
2340
E T50, TdA, T5c, TaY, T5h, TdD, T5t, Tb5, T57, TdB, T59, Tb1, T5o, TdE, T5q;
2343
E T4Y, T4Z, TaW, T5a, T5b, TaX;
2345
T4Z = I[WS(is, 67)];
2347
T5a = I[WS(is, 35)];
2348
T5b = I[WS(is, 99)];
2356
E T5f, T5g, Tb3, T5r, T5s, Tb4;
2357
T5f = I[WS(is, 123)];
2358
T5g = I[WS(is, 59)];
2360
T5r = I[WS(is, 27)];
2361
T5s = I[WS(is, 91)];
2369
E T53, TaZ, T56, Tb0;
2371
E T51, T52, T54, T55;
2372
T51 = I[WS(is, 19)];
2373
T52 = I[WS(is, 83)];
2376
T54 = I[WS(is, 115)];
2377
T55 = I[WS(is, 51)];
2381
T57 = KP707106781 * (T53 + T56);
2383
T59 = KP707106781 * (T56 - T53);
2387
E T5k, Tb6, T5n, Tb7;
2389
E T5i, T5j, T5l, T5m;
2390
T5i = I[WS(is, 11)];
2391
T5j = I[WS(is, 75)];
2394
T5l = I[WS(is, 107)];
2395
T5m = I[WS(is, 43)];
2399
T5o = KP707106781 * (T5k + T5n);
2401
T5q = KP707106781 * (T5n - T5k);
2407
E T58, T5d, TdL, TdM;
2410
T5e = FMA(KP980785280, T58, KP195090322 * T5d);
2411
T5H = FNMS(KP195090322, T58, KP980785280 * T5d);
2412
TdL = FNMS(KP382683432, TdA, KP923879532 * TdB);
2413
TdM = FMA(KP382683432, TdD, KP923879532 * TdE);
2418
E T5p, T5u, T8h, T8i;
2421
T5v = FNMS(KP195090322, T5u, KP980785280 * T5p);
2422
T5I = FMA(KP195090322, T5p, KP980785280 * T5u);
2425
T8j = FNMS(KP555570233, T8i, KP831469612 * T8h);
2426
T8q = FMA(KP555570233, T8h, KP831469612 * T8i);
2428
TdC = FMA(KP923879532, TdA, KP382683432 * TdB);
2429
TdF = FNMS(KP382683432, TdE, KP923879532 * TdD);
2433
E T8e, T8f, Tb2, Tb9;
2436
T8g = FMA(KP831469612, T8e, KP555570233 * T8f);
2437
T8p = FNMS(KP555570233, T8e, KP831469612 * T8f);
2440
Tba = KP707106781 * (Tb2 + Tb9);
2441
Tbj = KP707106781 * (Tb9 - Tb2);
2445
E T11, TbV, Tc9, Tcf, T22, Tcb, Tc2, Tce;
2447
E Tv, T10, Tc5, Tc8;
2458
E T1w, T21, TbY, Tc1;
2468
ro[WS(ros, 32)] = T11 - T22;
2469
io[WS(ios, 32)] = Tcf - Tce;
2471
E Tca, Tcc, Tcd, Tcg;
2472
Tca = KP707106781 * (Tc2 + Tc9);
2473
ro[WS(ros, 48)] = TbV - Tca;
2474
ro[WS(ros, 16)] = TbV + Tca;
2475
Tcc = KP707106781 * (Tc9 - Tc2);
2476
io[WS(ios, 16)] = Tcb + Tcc;
2477
io[WS(ios, 48)] = Tcc - Tcb;
2480
ro[WS(ros, 64)] = Tcd - Tcg;
2485
E Tch, Tcu, Tck, Tct, Tco, Tcy, Tcr, Tcz, Tci, Tcj;
2490
Tck = KP707106781 * (Tci + Tcj);
2491
Tct = KP707106781 * (Tcj - Tci);
2493
E Tcm, Tcn, Tcp, Tcq;
2496
Tco = FMA(KP923879532, Tcm, KP382683432 * Tcn);
2497
Tcy = FNMS(KP382683432, Tcm, KP923879532 * Tcn);
2500
Tcr = FNMS(KP382683432, Tcq, KP923879532 * Tcp);
2501
Tcz = FMA(KP382683432, Tcp, KP923879532 * Tcq);
2504
E Tcl, Tcs, Tcx, TcA;
2507
ro[WS(ros, 56)] = Tcl - Tcs;
2508
ro[WS(ros, 8)] = Tcl + Tcs;
2511
io[WS(ios, 8)] = Tcx + TcA;
2512
io[WS(ios, 56)] = TcA - Tcx;
2515
E Tcv, Tcw, TcB, TcC;
2518
io[WS(ios, 24)] = Tcv + Tcw;
2519
io[WS(ios, 40)] = Tcw - Tcv;
2522
ro[WS(ros, 40)] = TcB - TcC;
2523
ro[WS(ros, 24)] = TcB + TcC;
2527
E Ta9, TbB, Tbs, TbM, Tag, TbL, TbJ, TbR, TaO, Tbw, Tbp, TbC, TbG, TbQ, Tbl;
2529
Ta8 = KP707106781 * (Ta6 + Ta7);
2532
Tbr = KP707106781 * (Ta7 - Ta6);
2536
E Tac, Taf, TbH, TbI;
2537
Tac = FMA(KP923879532, Taa, KP382683432 * Tab);
2538
Taf = FNMS(KP382683432, Tae, KP923879532 * Tad);
2543
TbJ = FNMS(KP555570233, TbI, KP831469612 * TbH);
2544
TbR = FMA(KP555570233, TbH, KP831469612 * TbI);
2547
E TaE, TaN, Tbn, Tbo;
2550
TaO = FMA(KP980785280, TaE, KP195090322 * TaN);
2551
Tbw = FNMS(KP195090322, TaE, KP980785280 * TaN);
2552
Tbn = FNMS(KP382683432, Taa, KP923879532 * Tab);
2553
Tbo = FMA(KP382683432, Tad, KP923879532 * Tae);
2558
E TbE, TbF, Tbb, Tbk;
2561
TbG = FMA(KP831469612, TbE, KP555570233 * TbF);
2562
TbQ = FNMS(KP555570233, TbE, KP831469612 * TbF);
2565
Tbl = FNMS(KP195090322, Tbk, KP980785280 * Tbb);
2566
Tbx = FMA(KP195090322, Tbb, KP980785280 * Tbk);
2569
E Tah, Tbm, Tbv, Tby;
2572
ro[WS(ros, 60)] = Tah - Tbm;
2573
ro[WS(ros, 4)] = Tah + Tbm;
2576
io[WS(ios, 4)] = Tbv + Tby;
2577
io[WS(ios, 60)] = Tby - Tbv;
2580
E Tbt, Tbu, Tbz, TbA;
2583
io[WS(ios, 28)] = Tbt + Tbu;
2584
io[WS(ios, 36)] = Tbu - Tbt;
2587
ro[WS(ros, 36)] = Tbz - TbA;
2588
ro[WS(ros, 28)] = Tbz + TbA;
2591
E TbD, TbK, TbP, TbS;
2594
ro[WS(ros, 52)] = TbD - TbK;
2595
ro[WS(ros, 12)] = TbD + TbK;
2598
io[WS(ios, 12)] = TbP + TbS;
2599
io[WS(ios, 52)] = TbS - TbP;
2602
E TbN, TbO, TbT, TbU;
2605
io[WS(ios, 20)] = TbN + TbO;
2606
io[WS(ios, 44)] = TbO - TbN;
2609
ro[WS(ros, 44)] = TbT - TbU;
2610
ro[WS(ros, 20)] = TbT + TbU;
2614
E Tev, Tf7, Tfc, Tfm, Tff, Tfn, TeC, Tfh, TeK, Tf2, TeV, Tf8, TeY, Tfi, TeR;
2617
E Tet, Teu, Tfa, Tfb;
2624
Tfc = FMA(KP290284677, Tfa, KP956940335 * Tfb);
2625
Tfm = FNMS(KP290284677, Tfb, KP956940335 * Tfa);
2628
E Tfd, Tfe, Tey, TeB;
2631
Tff = FNMS(KP290284677, Tfe, KP956940335 * Tfd);
2632
Tfn = FMA(KP956940335, Tfe, KP290284677 * Tfd);
2633
Tey = FMA(KP555570233, Tew, KP831469612 * Tex);
2634
TeB = FNMS(KP555570233, TeA, KP831469612 * Tez);
2639
E TeG, TeJ, TeT, TeU;
2642
TeK = FMA(KP471396736, TeG, KP881921264 * TeJ);
2643
Tf2 = FNMS(KP471396736, TeJ, KP881921264 * TeG);
2644
TeT = FNMS(KP555570233, Tex, KP831469612 * Tew);
2645
TeU = FMA(KP831469612, TeA, KP555570233 * Tez);
2650
E TeW, TeX, TeN, TeQ;
2657
TeR = FNMS(KP471396736, TeQ, KP881921264 * TeN);
2658
Tf3 = FMA(KP881921264, TeQ, KP471396736 * TeN);
2661
E TeD, TeS, Tf1, Tf4;
2664
ro[WS(ros, 54)] = TeD - TeS;
2665
ro[WS(ros, 10)] = TeD + TeS;
2668
io[WS(ios, 10)] = Tf1 + Tf4;
2669
io[WS(ios, 54)] = Tf4 - Tf1;
2672
E TeZ, Tf0, Tf5, Tf6;
2675
io[WS(ios, 22)] = TeZ + Tf0;
2676
io[WS(ios, 42)] = Tf0 - TeZ;
2679
ro[WS(ros, 42)] = Tf5 - Tf6;
2680
ro[WS(ros, 22)] = Tf5 + Tf6;
2683
E Tf9, Tfg, Tfl, Tfo;
2686
ro[WS(ros, 58)] = Tf9 - Tfg;
2687
ro[WS(ros, 6)] = Tf9 + Tfg;
2690
io[WS(ios, 6)] = Tfl + Tfo;
2691
io[WS(ios, 58)] = Tfo - Tfl;
2694
E Tfj, Tfk, Tfp, Tfq;
2697
io[WS(ios, 26)] = Tfj + Tfk;
2698
io[WS(ios, 38)] = Tfk - Tfj;
2701
ro[WS(ros, 38)] = Tfp - Tfq;
2702
ro[WS(ros, 26)] = Tfp + Tfq;
2706
E TcP, Te9, Tee, Teo, Teh, Tep, Td8, Tej, Tdu, Te4, TdT, Tea, Te0, Tek, TdP;
2709
E TcH, TcO, Tec, Ted;
2716
Tee = FMA(KP773010453, Tec, KP634393284 * Ted);
2717
Teo = FNMS(KP634393284, Tec, KP773010453 * Ted);
2720
E Tef, Teg, TcY, Td7;
2723
Teh = FNMS(KP634393284, Teg, KP773010453 * Tef);
2724
Tep = FMA(KP634393284, Tef, KP773010453 * Teg);
2725
TcY = FMA(KP980785280, TcU, KP195090322 * TcX);
2726
Td7 = FNMS(KP195090322, Td6, KP980785280 * Td3);
2731
E Tdm, Tdt, TdR, TdS;
2734
Tdu = FMA(KP995184726, Tdm, KP098017140 * Tdt);
2735
Te4 = FNMS(KP098017140, Tdm, KP995184726 * Tdt);
2736
TdR = FNMS(KP195090322, TcU, KP980785280 * TcX);
2737
TdS = FMA(KP195090322, Td3, KP980785280 * Td6);
2742
E TdW, TdZ, TdH, TdO;
2749
TdP = FNMS(KP098017140, TdO, KP995184726 * TdH);
2750
Te5 = FMA(KP098017140, TdH, KP995184726 * TdO);
2753
E Td9, TdQ, Te3, Te6;
2756
ro[WS(ros, 62)] = Td9 - TdQ;
2757
ro[WS(ros, 2)] = Td9 + TdQ;
2760
io[WS(ios, 2)] = Te3 + Te6;
2761
io[WS(ios, 62)] = Te6 - Te3;
2764
E Te1, Te2, Te7, Te8;
2767
io[WS(ios, 30)] = Te1 + Te2;
2768
io[WS(ios, 34)] = Te2 - Te1;
2771
ro[WS(ros, 34)] = Te7 - Te8;
2772
ro[WS(ros, 30)] = Te7 + Te8;
2775
E Teb, Tei, Ten, Teq;
2778
ro[WS(ros, 50)] = Teb - Tei;
2779
ro[WS(ros, 14)] = Teb + Tei;
2782
io[WS(ios, 14)] = Ten + Teq;
2783
io[WS(ios, 50)] = Teq - Ten;
2786
E Tel, Tem, Ter, Tes;
2789
io[WS(ios, 18)] = Tel + Tem;
2790
io[WS(ios, 46)] = Tem - Tel;
2793
ro[WS(ros, 46)] = Ter - Tes;
2794
ro[WS(ros, 18)] = Ter + Tes;
2798
E T6v, T77, T6C, T7h, T6Y, T7i, T6V, T78, T6R, T7n, T73, T7f, T6K, T7m, T72;
2801
E T6t, T6u, T6T, T6U;
2807
E T6y, T6B, T6W, T6X;
2808
T6y = FMA(KP773010453, T6w, KP634393284 * T6x);
2809
T6B = FNMS(KP634393284, T6A, KP773010453 * T6z);
2817
T6T = FNMS(KP634393284, T6w, KP773010453 * T6x);
2818
T6U = FMA(KP634393284, T6z, KP773010453 * T6A);
2822
E T6N, T7d, T6Q, T7e, T6M, T6O;
2829
T6R = FNMS(KP427555093, T6Q, KP903989293 * T6N);
2830
T7n = FMA(KP941544065, T7e, KP336889853 * T7d);
2831
T73 = FMA(KP903989293, T6Q, KP427555093 * T6N);
2832
T7f = FNMS(KP336889853, T7e, KP941544065 * T7d);
2835
E T6G, T7a, T6J, T7b, T6E, T6I;
2842
T6K = FMA(KP427555093, T6G, KP903989293 * T6J);
2843
T7m = FNMS(KP336889853, T7b, KP941544065 * T7a);
2844
T72 = FNMS(KP427555093, T6J, KP903989293 * T6G);
2845
T7c = FMA(KP336889853, T7a, KP941544065 * T7b);
2849
E T6D, T6S, T71, T74;
2852
ro[WS(ros, 55)] = T6D - T6S;
2853
ro[WS(ros, 9)] = T6D + T6S;
2856
io[WS(ios, 9)] = T71 + T74;
2857
io[WS(ios, 55)] = T74 - T71;
2860
E T6Z, T70, T75, T76;
2863
io[WS(ios, 23)] = T6Z + T70;
2864
io[WS(ios, 41)] = T70 - T6Z;
2867
ro[WS(ros, 41)] = T75 - T76;
2868
ro[WS(ros, 23)] = T75 + T76;
2871
E T79, T7g, T7l, T7o;
2874
ro[WS(ros, 57)] = T79 - T7g;
2875
ro[WS(ros, 7)] = T79 + T7g;
2878
io[WS(ios, 7)] = T7l + T7o;
2879
io[WS(ios, 57)] = T7o - T7l;
2882
E T7j, T7k, T7p, T7q;
2885
io[WS(ios, 25)] = T7j + T7k;
2886
io[WS(ios, 39)] = T7k - T7j;
2889
ro[WS(ros, 39)] = T7p - T7q;
2890
ro[WS(ros, 25)] = T7p + T7q;
2894
E T99, T9L, T9g, T9V, T9C, T9W, T9z, T9M, T9v, Ta1, T9H, T9T, T9o, Ta0, T9G;
2897
E T97, T98, T9x, T9y;
2903
E T9c, T9f, T9A, T9B;
2904
T9c = FMA(KP471396736, T9a, KP881921264 * T9b);
2905
T9f = FNMS(KP471396736, T9e, KP881921264 * T9d);
2913
T9x = FNMS(KP471396736, T9b, KP881921264 * T9a);
2914
T9y = FMA(KP881921264, T9e, KP471396736 * T9d);
2918
E T9r, T9R, T9u, T9S, T9q, T9s;
2925
T9v = FNMS(KP514102744, T9u, KP857728610 * T9r);
2926
Ta1 = FMA(KP970031253, T9S, KP242980179 * T9R);
2927
T9H = FMA(KP857728610, T9u, KP514102744 * T9r);
2928
T9T = FNMS(KP242980179, T9S, KP970031253 * T9R);
2931
E T9k, T9O, T9n, T9P, T9i, T9m;
2938
T9o = FMA(KP514102744, T9k, KP857728610 * T9n);
2939
Ta0 = FNMS(KP242980179, T9P, KP970031253 * T9O);
2940
T9G = FNMS(KP514102744, T9n, KP857728610 * T9k);
2941
T9Q = FMA(KP242980179, T9O, KP970031253 * T9P);
2945
E T9h, T9w, T9F, T9I;
2948
ro[WS(ros, 53)] = T9h - T9w;
2949
ro[WS(ros, 11)] = T9h + T9w;
2952
io[WS(ios, 11)] = T9F + T9I;
2953
io[WS(ios, 53)] = T9I - T9F;
2956
E T9D, T9E, T9J, T9K;
2959
io[WS(ios, 21)] = T9D + T9E;
2960
io[WS(ios, 43)] = T9E - T9D;
2963
ro[WS(ros, 43)] = T9J - T9K;
2964
ro[WS(ros, 21)] = T9J + T9K;
2967
E T9N, T9U, T9Z, Ta2;
2970
ro[WS(ros, 59)] = T9N - T9U;
2971
ro[WS(ros, 5)] = T9N + T9U;
2974
io[WS(ios, 5)] = T9Z + Ta2;
2975
io[WS(ios, 59)] = Ta2 - T9Z;
2978
E T9X, T9Y, Ta3, Ta4;
2981
io[WS(ios, 27)] = T9X + T9Y;
2982
io[WS(ios, 37)] = T9Y - T9X;
2985
ro[WS(ros, 37)] = Ta3 - Ta4;
2986
ro[WS(ros, 27)] = Ta3 + Ta4;
2990
E T2z, T69, T3g, T6j, T60, T6k, T5P, T6a, T5L, T6p, T65, T6h, T4w, T6o, T64;
2993
E T2f, T2y, T5N, T5O;
2999
E T2U, T3f, T5W, T5Z;
3000
T2U = FMA(KP098017140, T2M, KP995184726 * T2T);
3001
T3f = FNMS(KP098017140, T3e, KP995184726 * T37);
3009
T5N = FNMS(KP098017140, T2T, KP995184726 * T2M);
3010
T5O = FMA(KP995184726, T3e, KP098017140 * T37);
3014
E T5x, T6f, T5K, T6g, T5w, T5J;
3021
T5L = FNMS(KP049067674, T5K, KP998795456 * T5x);
3022
T6p = FMA(KP671558954, T6f, KP740951125 * T6g);
3023
T65 = FMA(KP049067674, T5x, KP998795456 * T5K);
3024
T6h = FNMS(KP671558954, T6g, KP740951125 * T6f);
3027
E T4i, T6c, T4v, T6d, T4h, T4u;
3034
T4w = FMA(KP998795456, T4i, KP049067674 * T4v);
3035
T6o = FNMS(KP671558954, T6c, KP740951125 * T6d);
3036
T64 = FNMS(KP049067674, T4i, KP998795456 * T4v);
3037
T6e = FMA(KP740951125, T6c, KP671558954 * T6d);
3041
E T3h, T5M, T63, T66;
3044
ro[WS(ros, 63)] = T3h - T5M;
3045
ro[WS(ros, 1)] = T3h + T5M;
3048
io[WS(ios, 1)] = T63 + T66;
3049
io[WS(ios, 63)] = T66 - T63;
3052
E T61, T62, T67, T68;
3055
io[WS(ios, 31)] = T61 + T62;
3056
io[WS(ios, 33)] = T62 - T61;
3059
ro[WS(ros, 33)] = T67 - T68;
3060
ro[WS(ros, 31)] = T67 + T68;
3063
E T6b, T6i, T6n, T6q;
3066
ro[WS(ros, 49)] = T6b - T6i;
3067
ro[WS(ros, 15)] = T6b + T6i;
3070
io[WS(ios, 15)] = T6n + T6q;
3071
io[WS(ios, 49)] = T6q - T6n;
3074
E T6l, T6m, T6r, T6s;
3077
io[WS(ios, 17)] = T6l + T6m;
3078
io[WS(ios, 47)] = T6m - T6l;
3081
ro[WS(ros, 47)] = T6r - T6s;
3082
ro[WS(ros, 17)] = T6r + T6s;
3086
E T7B, T8N, T7Q, T8X, T8E, T8Y, T8x, T8O, T8t, T93, T8J, T8V, T8a, T92, T8I;
3089
E T7t, T7A, T8v, T8w;
3095
E T7I, T7P, T8A, T8D;
3096
T7I = FMA(KP956940335, T7E, KP290284677 * T7H);
3097
T7P = FNMS(KP290284677, T7O, KP956940335 * T7L);
3105
T8v = FNMS(KP290284677, T7E, KP956940335 * T7H);
3106
T8w = FMA(KP290284677, T7L, KP956940335 * T7O);
3110
E T8l, T8T, T8s, T8U, T8k, T8r;
3117
T8t = FNMS(KP146730474, T8s, KP989176509 * T8l);
3118
T93 = FMA(KP595699304, T8T, KP803207531 * T8U);
3119
T8J = FMA(KP146730474, T8l, KP989176509 * T8s);
3120
T8V = FNMS(KP595699304, T8U, KP803207531 * T8T);
3123
E T82, T8Q, T89, T8R, T81, T88;
3130
T8a = FMA(KP989176509, T82, KP146730474 * T89);
3131
T92 = FNMS(KP595699304, T8Q, KP803207531 * T8R);
3132
T8I = FNMS(KP146730474, T82, KP989176509 * T89);
3133
T8S = FMA(KP803207531, T8Q, KP595699304 * T8R);
3137
E T7R, T8u, T8H, T8K;
3140
ro[WS(ros, 61)] = T7R - T8u;
3141
ro[WS(ros, 3)] = T7R + T8u;
3144
io[WS(ios, 3)] = T8H + T8K;
3145
io[WS(ios, 61)] = T8K - T8H;
3148
E T8F, T8G, T8L, T8M;
3151
io[WS(ios, 29)] = T8F + T8G;
3152
io[WS(ios, 35)] = T8G - T8F;
3155
ro[WS(ros, 35)] = T8L - T8M;
3156
ro[WS(ros, 29)] = T8L + T8M;
3159
E T8P, T8W, T91, T94;
3162
ro[WS(ros, 51)] = T8P - T8W;
3163
ro[WS(ros, 13)] = T8P + T8W;
3166
io[WS(ios, 13)] = T91 + T94;
3167
io[WS(ios, 51)] = T94 - T91;
3170
E T8Z, T90, T95, T96;
3173
io[WS(ios, 19)] = T8Z + T90;
3174
io[WS(ios, 45)] = T90 - T8Z;
3177
ro[WS(ros, 45)] = T95 - T96;
3178
ro[WS(ros, 19)] = T95 + T96;
3184
static const kr2hc_desc desc = { 128, "r2hc_128", {812, 186, 144, 0}, &GENUS, 0, 0, 0, 0, 0 };
3186
void X(codelet_r2hc_128) (planner *p) {
3187
X(kr2hc_register) (p, r2hc_128, &desc);
3190
#endif /* HAVE_FMA */