cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:07:33 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -dif -name hb_64 -include rdft/scalar/hb.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 1038 FP additions, 644 FP multiplications,
|
cannam@167
|
32 * (or, 520 additions, 126 multiplications, 518 fused multiply/add),
|
cannam@167
|
33 * 192 stack variables, 15 constants, and 256 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hb.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hb_64(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
40 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@167
|
41 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
42 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@167
|
43 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
44 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@167
|
45 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
46 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
47 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@167
|
48 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
49 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
50 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
51 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
52 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
53 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
54 {
|
cannam@167
|
55 INT m;
|
cannam@167
|
56 for (m = mb, W = W + ((mb - 1) * 126); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 126, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
57 E Tv, Thy, T5B, T7n, Tey, TfP, TjB, Tkl, T2k, T6U, T2H, T7o, Tia, TiH, Tj8;
|
cannam@167
|
58 E Tk8, T5E, T6V, T9N, Tbz, T9Q, Tb7, Tev, Tgh, T8G, Tb6, T8N, TbA, TcU, TfO;
|
cannam@167
|
59 E Td5, Tgi, T10, Ti3, Tje, TjC, ThF, TiI, Tds, TeA, Tjb, TjD, Tdh, TeB, TfT;
|
cannam@167
|
60 E Tgl, TfW, Tgk, T39, T7r, T5H, T6Z, T8V, TbC, T9S, Tbb, T3A, T7q, T5G, T72;
|
cannam@167
|
61 E T92, TbD, T9T, Tbe, T1w, ThH, Tjq, Tke, Tjt, Tkf, ThO, TiK, Tec, TgT, Tfc;
|
cannam@167
|
62 E Tgb, Tel, TgU, Tfd, Tg8, T5a, T82, T83, T5n, T6i, T77, T7a, T6j, T9f, Tcb;
|
cannam@167
|
63 E Tcc, T9m, Tar, Tbj, Tbm, Tas, T21, ThQ, Tjj, Tkb, Tjm, Tkc, ThX, TiL, TdL;
|
cannam@167
|
64 E TgW, Tf9, Tg4, TdU, TgX, Tfa, Tg1, T4h, T7Z, T80, T4u, T6f, T7e, T7h, T6g;
|
cannam@167
|
65 E T9y, Tce, Tcf, T9F, Tau, Tbq, Tbt, Tav;
|
cannam@167
|
66 {
|
cannam@167
|
67 E T3, T6, T7, T5t, T24, Tes, Ter, T27, Ti4, T5w, Ta, TcR, Td, TcS, Te;
|
cannam@167
|
68 E T2d, Ti5, T5z, T5y, T2i, Tm, Td3, Ti7, T2p, T2u, T8I, Td0, T8H, Tt, TcY;
|
cannam@167
|
69 E Ti8, T2A, T2F, T8L, TcX, T8K;
|
cannam@167
|
70 {
|
cannam@167
|
71 E T1, T2, T4, T5;
|
cannam@167
|
72 T1 = cr[0];
|
cannam@167
|
73 T2 = ci[WS(rs, 31)];
|
cannam@167
|
74 T3 = T1 + T2;
|
cannam@167
|
75 T4 = cr[WS(rs, 16)];
|
cannam@167
|
76 T5 = ci[WS(rs, 15)];
|
cannam@167
|
77 T6 = T4 + T5;
|
cannam@167
|
78 T7 = T3 + T6;
|
cannam@167
|
79 T5t = T4 - T5;
|
cannam@167
|
80 T24 = T1 - T2;
|
cannam@167
|
81 }
|
cannam@167
|
82 {
|
cannam@167
|
83 E T25, T26, T5u, T5v;
|
cannam@167
|
84 T25 = ci[WS(rs, 47)];
|
cannam@167
|
85 T26 = cr[WS(rs, 48)];
|
cannam@167
|
86 Tes = T25 - T26;
|
cannam@167
|
87 T5u = ci[WS(rs, 63)];
|
cannam@167
|
88 T5v = cr[WS(rs, 32)];
|
cannam@167
|
89 Ter = T5u - T5v;
|
cannam@167
|
90 T27 = T25 + T26;
|
cannam@167
|
91 Ti4 = Ter + Tes;
|
cannam@167
|
92 T5w = T5u + T5v;
|
cannam@167
|
93 }
|
cannam@167
|
94 {
|
cannam@167
|
95 E T29, T2h, T2e, T2c;
|
cannam@167
|
96 {
|
cannam@167
|
97 E T8, T9, T2f, T2g;
|
cannam@167
|
98 T8 = cr[WS(rs, 8)];
|
cannam@167
|
99 T9 = ci[WS(rs, 23)];
|
cannam@167
|
100 Ta = T8 + T9;
|
cannam@167
|
101 T29 = T8 - T9;
|
cannam@167
|
102 T2f = ci[WS(rs, 39)];
|
cannam@167
|
103 T2g = cr[WS(rs, 56)];
|
cannam@167
|
104 T2h = T2f + T2g;
|
cannam@167
|
105 TcR = T2f - T2g;
|
cannam@167
|
106 }
|
cannam@167
|
107 {
|
cannam@167
|
108 E Tb, Tc, T2a, T2b;
|
cannam@167
|
109 Tb = ci[WS(rs, 7)];
|
cannam@167
|
110 Tc = cr[WS(rs, 24)];
|
cannam@167
|
111 Td = Tb + Tc;
|
cannam@167
|
112 T2e = Tb - Tc;
|
cannam@167
|
113 T2a = ci[WS(rs, 55)];
|
cannam@167
|
114 T2b = cr[WS(rs, 40)];
|
cannam@167
|
115 T2c = T2a + T2b;
|
cannam@167
|
116 TcS = T2a - T2b;
|
cannam@167
|
117 }
|
cannam@167
|
118 Te = Ta + Td;
|
cannam@167
|
119 T2d = T29 - T2c;
|
cannam@167
|
120 Ti5 = TcS + TcR;
|
cannam@167
|
121 T5z = T2e + T2h;
|
cannam@167
|
122 T5y = T29 + T2c;
|
cannam@167
|
123 T2i = T2e - T2h;
|
cannam@167
|
124 }
|
cannam@167
|
125 {
|
cannam@167
|
126 E Ti, T2l, T2t, Td1, Tl, T2q, T2o, Td2;
|
cannam@167
|
127 {
|
cannam@167
|
128 E Tg, Th, T2r, T2s;
|
cannam@167
|
129 Tg = cr[WS(rs, 4)];
|
cannam@167
|
130 Th = ci[WS(rs, 27)];
|
cannam@167
|
131 Ti = Tg + Th;
|
cannam@167
|
132 T2l = Tg - Th;
|
cannam@167
|
133 T2r = ci[WS(rs, 59)];
|
cannam@167
|
134 T2s = cr[WS(rs, 36)];
|
cannam@167
|
135 T2t = T2r + T2s;
|
cannam@167
|
136 Td1 = T2r - T2s;
|
cannam@167
|
137 }
|
cannam@167
|
138 {
|
cannam@167
|
139 E Tj, Tk, T2m, T2n;
|
cannam@167
|
140 Tj = cr[WS(rs, 20)];
|
cannam@167
|
141 Tk = ci[WS(rs, 11)];
|
cannam@167
|
142 Tl = Tj + Tk;
|
cannam@167
|
143 T2q = Tj - Tk;
|
cannam@167
|
144 T2m = ci[WS(rs, 43)];
|
cannam@167
|
145 T2n = cr[WS(rs, 52)];
|
cannam@167
|
146 T2o = T2m + T2n;
|
cannam@167
|
147 Td2 = T2m - T2n;
|
cannam@167
|
148 }
|
cannam@167
|
149 Tm = Ti + Tl;
|
cannam@167
|
150 Td3 = Td1 - Td2;
|
cannam@167
|
151 Ti7 = Td1 + Td2;
|
cannam@167
|
152 T2p = T2l - T2o;
|
cannam@167
|
153 T2u = T2q + T2t;
|
cannam@167
|
154 T8I = T2l + T2o;
|
cannam@167
|
155 Td0 = Ti - Tl;
|
cannam@167
|
156 T8H = T2t - T2q;
|
cannam@167
|
157 }
|
cannam@167
|
158 {
|
cannam@167
|
159 E Tp, T2w, T2E, TcV, Ts, T2B, T2z, TcW;
|
cannam@167
|
160 {
|
cannam@167
|
161 E Tn, To, T2C, T2D;
|
cannam@167
|
162 Tn = ci[WS(rs, 3)];
|
cannam@167
|
163 To = cr[WS(rs, 28)];
|
cannam@167
|
164 Tp = Tn + To;
|
cannam@167
|
165 T2w = Tn - To;
|
cannam@167
|
166 T2C = ci[WS(rs, 35)];
|
cannam@167
|
167 T2D = cr[WS(rs, 60)];
|
cannam@167
|
168 T2E = T2C + T2D;
|
cannam@167
|
169 TcV = T2C - T2D;
|
cannam@167
|
170 }
|
cannam@167
|
171 {
|
cannam@167
|
172 E Tq, Tr, T2x, T2y;
|
cannam@167
|
173 Tq = cr[WS(rs, 12)];
|
cannam@167
|
174 Tr = ci[WS(rs, 19)];
|
cannam@167
|
175 Ts = Tq + Tr;
|
cannam@167
|
176 T2B = Tq - Tr;
|
cannam@167
|
177 T2x = ci[WS(rs, 51)];
|
cannam@167
|
178 T2y = cr[WS(rs, 44)];
|
cannam@167
|
179 T2z = T2x + T2y;
|
cannam@167
|
180 TcW = T2x - T2y;
|
cannam@167
|
181 }
|
cannam@167
|
182 Tt = Tp + Ts;
|
cannam@167
|
183 TcY = Tp - Ts;
|
cannam@167
|
184 Ti8 = TcV + TcW;
|
cannam@167
|
185 T2A = T2w - T2z;
|
cannam@167
|
186 T2F = T2B - T2E;
|
cannam@167
|
187 T8L = T2w + T2z;
|
cannam@167
|
188 TcX = TcV - TcW;
|
cannam@167
|
189 T8K = T2B + T2E;
|
cannam@167
|
190 }
|
cannam@167
|
191 {
|
cannam@167
|
192 E Tf, Tu, T5x, T5A;
|
cannam@167
|
193 Tf = T7 + Te;
|
cannam@167
|
194 Tu = Tm + Tt;
|
cannam@167
|
195 Tv = Tf + Tu;
|
cannam@167
|
196 Thy = Tf - Tu;
|
cannam@167
|
197 T5x = T5t + T5w;
|
cannam@167
|
198 T5A = T5y - T5z;
|
cannam@167
|
199 T5B = FMA(KP707106781, T5A, T5x);
|
cannam@167
|
200 T7n = FNMS(KP707106781, T5A, T5x);
|
cannam@167
|
201 }
|
cannam@167
|
202 {
|
cannam@167
|
203 E Tew, Tex, Tjz, TjA;
|
cannam@167
|
204 Tew = Td0 - Td3;
|
cannam@167
|
205 Tex = TcY + TcX;
|
cannam@167
|
206 Tey = Tew - Tex;
|
cannam@167
|
207 TfP = Tew + Tex;
|
cannam@167
|
208 Tjz = Ti4 - Ti5;
|
cannam@167
|
209 TjA = Tm - Tt;
|
cannam@167
|
210 TjB = Tjz - TjA;
|
cannam@167
|
211 Tkl = TjA + Tjz;
|
cannam@167
|
212 }
|
cannam@167
|
213 {
|
cannam@167
|
214 E T28, T2j, T2v, T2G;
|
cannam@167
|
215 T28 = T24 - T27;
|
cannam@167
|
216 T2j = T2d + T2i;
|
cannam@167
|
217 T2k = FMA(KP707106781, T2j, T28);
|
cannam@167
|
218 T6U = FNMS(KP707106781, T2j, T28);
|
cannam@167
|
219 T2v = FNMS(KP414213562, T2u, T2p);
|
cannam@167
|
220 T2G = FMA(KP414213562, T2F, T2A);
|
cannam@167
|
221 T2H = T2v + T2G;
|
cannam@167
|
222 T7o = T2v - T2G;
|
cannam@167
|
223 }
|
cannam@167
|
224 {
|
cannam@167
|
225 E Ti6, Ti9, Tj6, Tj7;
|
cannam@167
|
226 Ti6 = Ti4 + Ti5;
|
cannam@167
|
227 Ti9 = Ti7 + Ti8;
|
cannam@167
|
228 Tia = Ti6 - Ti9;
|
cannam@167
|
229 TiH = Ti6 + Ti9;
|
cannam@167
|
230 Tj6 = T7 - Te;
|
cannam@167
|
231 Tj7 = Ti8 - Ti7;
|
cannam@167
|
232 Tj8 = Tj6 - Tj7;
|
cannam@167
|
233 Tk8 = Tj6 + Tj7;
|
cannam@167
|
234 }
|
cannam@167
|
235 {
|
cannam@167
|
236 E T5C, T5D, T9L, T9M;
|
cannam@167
|
237 T5C = FMA(KP414213562, T2p, T2u);
|
cannam@167
|
238 T5D = FNMS(KP414213562, T2A, T2F);
|
cannam@167
|
239 T5E = T5C + T5D;
|
cannam@167
|
240 T6V = T5D - T5C;
|
cannam@167
|
241 T9L = T5w - T5t;
|
cannam@167
|
242 T9M = T2d - T2i;
|
cannam@167
|
243 T9N = FMA(KP707106781, T9M, T9L);
|
cannam@167
|
244 Tbz = FNMS(KP707106781, T9M, T9L);
|
cannam@167
|
245 }
|
cannam@167
|
246 {
|
cannam@167
|
247 E T9O, T9P, Tet, Teu;
|
cannam@167
|
248 T9O = FMA(KP414213562, T8H, T8I);
|
cannam@167
|
249 T9P = FMA(KP414213562, T8K, T8L);
|
cannam@167
|
250 T9Q = T9O - T9P;
|
cannam@167
|
251 Tb7 = T9O + T9P;
|
cannam@167
|
252 Tet = Ter - Tes;
|
cannam@167
|
253 Teu = Ta - Td;
|
cannam@167
|
254 Tev = Tet - Teu;
|
cannam@167
|
255 Tgh = Teu + Tet;
|
cannam@167
|
256 }
|
cannam@167
|
257 {
|
cannam@167
|
258 E T8E, T8F, T8J, T8M;
|
cannam@167
|
259 T8E = T24 + T27;
|
cannam@167
|
260 T8F = T5y + T5z;
|
cannam@167
|
261 T8G = FNMS(KP707106781, T8F, T8E);
|
cannam@167
|
262 Tb6 = FMA(KP707106781, T8F, T8E);
|
cannam@167
|
263 T8J = FNMS(KP414213562, T8I, T8H);
|
cannam@167
|
264 T8M = FNMS(KP414213562, T8L, T8K);
|
cannam@167
|
265 T8N = T8J + T8M;
|
cannam@167
|
266 TbA = T8M - T8J;
|
cannam@167
|
267 }
|
cannam@167
|
268 {
|
cannam@167
|
269 E TcQ, TcT, TcZ, Td4;
|
cannam@167
|
270 TcQ = T3 - T6;
|
cannam@167
|
271 TcT = TcR - TcS;
|
cannam@167
|
272 TcU = TcQ - TcT;
|
cannam@167
|
273 TfO = TcQ + TcT;
|
cannam@167
|
274 TcZ = TcX - TcY;
|
cannam@167
|
275 Td4 = Td0 + Td3;
|
cannam@167
|
276 Td5 = TcZ - Td4;
|
cannam@167
|
277 Tgi = Td4 + TcZ;
|
cannam@167
|
278 }
|
cannam@167
|
279 }
|
cannam@167
|
280 {
|
cannam@167
|
281 E TC, Tdn, ThC, T3e, T3v, T8S, Tdk, T8P, TY, Tdf, ThA, T2S, T2X, T36, Tda;
|
cannam@167
|
282 E T35, TJ, Tdq, ThD, T3j, T3o, T3x, Tdl, T3w, TR, Tdc, Thz, T2N, T34, T8Z;
|
cannam@167
|
283 E Td9, T8W;
|
cannam@167
|
284 {
|
cannam@167
|
285 E Ty, T3r, T3u, Tdj, TB, T3a, T3d, Tdi;
|
cannam@167
|
286 {
|
cannam@167
|
287 E Tw, Tx, T3s, T3t;
|
cannam@167
|
288 Tw = cr[WS(rs, 2)];
|
cannam@167
|
289 Tx = ci[WS(rs, 29)];
|
cannam@167
|
290 Ty = Tw + Tx;
|
cannam@167
|
291 T3r = Tw - Tx;
|
cannam@167
|
292 T3s = ci[WS(rs, 45)];
|
cannam@167
|
293 T3t = cr[WS(rs, 50)];
|
cannam@167
|
294 T3u = T3s + T3t;
|
cannam@167
|
295 Tdj = T3s - T3t;
|
cannam@167
|
296 }
|
cannam@167
|
297 {
|
cannam@167
|
298 E Tz, TA, T3b, T3c;
|
cannam@167
|
299 Tz = cr[WS(rs, 18)];
|
cannam@167
|
300 TA = ci[WS(rs, 13)];
|
cannam@167
|
301 TB = Tz + TA;
|
cannam@167
|
302 T3a = Tz - TA;
|
cannam@167
|
303 T3b = ci[WS(rs, 61)];
|
cannam@167
|
304 T3c = cr[WS(rs, 34)];
|
cannam@167
|
305 T3d = T3b + T3c;
|
cannam@167
|
306 Tdi = T3b - T3c;
|
cannam@167
|
307 }
|
cannam@167
|
308 TC = Ty + TB;
|
cannam@167
|
309 Tdn = Ty - TB;
|
cannam@167
|
310 ThC = Tdi + Tdj;
|
cannam@167
|
311 T3e = T3a + T3d;
|
cannam@167
|
312 T3v = T3r - T3u;
|
cannam@167
|
313 T8S = T3r + T3u;
|
cannam@167
|
314 Tdk = Tdi - Tdj;
|
cannam@167
|
315 T8P = T3d - T3a;
|
cannam@167
|
316 }
|
cannam@167
|
317 {
|
cannam@167
|
318 E TU, T2O, T2W, Tdd, TX, T2T, T2R, Tde;
|
cannam@167
|
319 {
|
cannam@167
|
320 E TS, TT, T2U, T2V;
|
cannam@167
|
321 TS = cr[WS(rs, 6)];
|
cannam@167
|
322 TT = ci[WS(rs, 25)];
|
cannam@167
|
323 TU = TS + TT;
|
cannam@167
|
324 T2O = TS - TT;
|
cannam@167
|
325 T2U = ci[WS(rs, 41)];
|
cannam@167
|
326 T2V = cr[WS(rs, 54)];
|
cannam@167
|
327 T2W = T2U + T2V;
|
cannam@167
|
328 Tdd = T2U - T2V;
|
cannam@167
|
329 }
|
cannam@167
|
330 {
|
cannam@167
|
331 E TV, TW, T2P, T2Q;
|
cannam@167
|
332 TV = ci[WS(rs, 9)];
|
cannam@167
|
333 TW = cr[WS(rs, 22)];
|
cannam@167
|
334 TX = TV + TW;
|
cannam@167
|
335 T2T = TV - TW;
|
cannam@167
|
336 T2P = ci[WS(rs, 57)];
|
cannam@167
|
337 T2Q = cr[WS(rs, 38)];
|
cannam@167
|
338 T2R = T2P + T2Q;
|
cannam@167
|
339 Tde = T2P - T2Q;
|
cannam@167
|
340 }
|
cannam@167
|
341 TY = TU + TX;
|
cannam@167
|
342 Tdf = Tdd - Tde;
|
cannam@167
|
343 ThA = Tde + Tdd;
|
cannam@167
|
344 T2S = T2O + T2R;
|
cannam@167
|
345 T2X = T2T + T2W;
|
cannam@167
|
346 T36 = T2T - T2W;
|
cannam@167
|
347 Tda = TU - TX;
|
cannam@167
|
348 T35 = T2O - T2R;
|
cannam@167
|
349 }
|
cannam@167
|
350 {
|
cannam@167
|
351 E TF, T3f, T3n, Tdo, TI, T3k, T3i, Tdp;
|
cannam@167
|
352 {
|
cannam@167
|
353 E TD, TE, T3l, T3m;
|
cannam@167
|
354 TD = cr[WS(rs, 10)];
|
cannam@167
|
355 TE = ci[WS(rs, 21)];
|
cannam@167
|
356 TF = TD + TE;
|
cannam@167
|
357 T3f = TD - TE;
|
cannam@167
|
358 T3l = ci[WS(rs, 37)];
|
cannam@167
|
359 T3m = cr[WS(rs, 58)];
|
cannam@167
|
360 T3n = T3l + T3m;
|
cannam@167
|
361 Tdo = T3l - T3m;
|
cannam@167
|
362 }
|
cannam@167
|
363 {
|
cannam@167
|
364 E TG, TH, T3g, T3h;
|
cannam@167
|
365 TG = ci[WS(rs, 5)];
|
cannam@167
|
366 TH = cr[WS(rs, 26)];
|
cannam@167
|
367 TI = TG + TH;
|
cannam@167
|
368 T3k = TG - TH;
|
cannam@167
|
369 T3g = ci[WS(rs, 53)];
|
cannam@167
|
370 T3h = cr[WS(rs, 42)];
|
cannam@167
|
371 T3i = T3g + T3h;
|
cannam@167
|
372 Tdp = T3g - T3h;
|
cannam@167
|
373 }
|
cannam@167
|
374 TJ = TF + TI;
|
cannam@167
|
375 Tdq = Tdo - Tdp;
|
cannam@167
|
376 ThD = Tdp + Tdo;
|
cannam@167
|
377 T3j = T3f + T3i;
|
cannam@167
|
378 T3o = T3k + T3n;
|
cannam@167
|
379 T3x = T3k - T3n;
|
cannam@167
|
380 Tdl = TF - TI;
|
cannam@167
|
381 T3w = T3f - T3i;
|
cannam@167
|
382 }
|
cannam@167
|
383 {
|
cannam@167
|
384 E TN, T30, T33, Td8, TQ, T2J, T2M, Td7;
|
cannam@167
|
385 {
|
cannam@167
|
386 E TL, TM, T31, T32;
|
cannam@167
|
387 TL = ci[WS(rs, 1)];
|
cannam@167
|
388 TM = cr[WS(rs, 30)];
|
cannam@167
|
389 TN = TL + TM;
|
cannam@167
|
390 T30 = TL - TM;
|
cannam@167
|
391 T31 = ci[WS(rs, 49)];
|
cannam@167
|
392 T32 = cr[WS(rs, 46)];
|
cannam@167
|
393 T33 = T31 + T32;
|
cannam@167
|
394 Td8 = T31 - T32;
|
cannam@167
|
395 }
|
cannam@167
|
396 {
|
cannam@167
|
397 E TO, TP, T2K, T2L;
|
cannam@167
|
398 TO = cr[WS(rs, 14)];
|
cannam@167
|
399 TP = ci[WS(rs, 17)];
|
cannam@167
|
400 TQ = TO + TP;
|
cannam@167
|
401 T2J = TO - TP;
|
cannam@167
|
402 T2K = ci[WS(rs, 33)];
|
cannam@167
|
403 T2L = cr[WS(rs, 62)];
|
cannam@167
|
404 T2M = T2K + T2L;
|
cannam@167
|
405 Td7 = T2K - T2L;
|
cannam@167
|
406 }
|
cannam@167
|
407 TR = TN + TQ;
|
cannam@167
|
408 Tdc = TN - TQ;
|
cannam@167
|
409 Thz = Td7 + Td8;
|
cannam@167
|
410 T2N = T2J - T2M;
|
cannam@167
|
411 T34 = T30 - T33;
|
cannam@167
|
412 T8Z = T30 + T33;
|
cannam@167
|
413 Td9 = Td7 - Td8;
|
cannam@167
|
414 T8W = T2J + T2M;
|
cannam@167
|
415 }
|
cannam@167
|
416 {
|
cannam@167
|
417 E TK, TZ, Tdm, Tdr;
|
cannam@167
|
418 TK = TC + TJ;
|
cannam@167
|
419 TZ = TR + TY;
|
cannam@167
|
420 T10 = TK + TZ;
|
cannam@167
|
421 Ti3 = TK - TZ;
|
cannam@167
|
422 {
|
cannam@167
|
423 E Tjc, Tjd, ThB, ThE;
|
cannam@167
|
424 Tjc = TC - TJ;
|
cannam@167
|
425 Tjd = ThC - ThD;
|
cannam@167
|
426 Tje = Tjc + Tjd;
|
cannam@167
|
427 TjC = Tjc - Tjd;
|
cannam@167
|
428 ThB = Thz + ThA;
|
cannam@167
|
429 ThE = ThC + ThD;
|
cannam@167
|
430 ThF = ThB - ThE;
|
cannam@167
|
431 TiI = ThE + ThB;
|
cannam@167
|
432 }
|
cannam@167
|
433 Tdm = Tdk - Tdl;
|
cannam@167
|
434 Tdr = Tdn - Tdq;
|
cannam@167
|
435 Tds = FNMS(KP414213562, Tdr, Tdm);
|
cannam@167
|
436 TeA = FMA(KP414213562, Tdm, Tdr);
|
cannam@167
|
437 {
|
cannam@167
|
438 E Tj9, Tja, Tdb, Tdg;
|
cannam@167
|
439 Tj9 = Thz - ThA;
|
cannam@167
|
440 Tja = TR - TY;
|
cannam@167
|
441 Tjb = Tj9 - Tja;
|
cannam@167
|
442 TjD = Tja + Tj9;
|
cannam@167
|
443 Tdb = Td9 - Tda;
|
cannam@167
|
444 Tdg = Tdc - Tdf;
|
cannam@167
|
445 Tdh = FMA(KP414213562, Tdg, Tdb);
|
cannam@167
|
446 TeB = FNMS(KP414213562, Tdb, Tdg);
|
cannam@167
|
447 }
|
cannam@167
|
448 }
|
cannam@167
|
449 {
|
cannam@167
|
450 E TfR, TfS, TfU, TfV;
|
cannam@167
|
451 TfR = Tda + Td9;
|
cannam@167
|
452 TfS = Tdc + Tdf;
|
cannam@167
|
453 TfT = FNMS(KP414213562, TfS, TfR);
|
cannam@167
|
454 Tgl = FMA(KP414213562, TfR, TfS);
|
cannam@167
|
455 TfU = Tdl + Tdk;
|
cannam@167
|
456 TfV = Tdn + Tdq;
|
cannam@167
|
457 TfW = FMA(KP414213562, TfV, TfU);
|
cannam@167
|
458 Tgk = FNMS(KP414213562, TfU, TfV);
|
cannam@167
|
459 {
|
cannam@167
|
460 E T2Z, T6X, T38, T6Y, T2Y, T37;
|
cannam@167
|
461 T2Y = T2S - T2X;
|
cannam@167
|
462 T2Z = FMA(KP707106781, T2Y, T2N);
|
cannam@167
|
463 T6X = FNMS(KP707106781, T2Y, T2N);
|
cannam@167
|
464 T37 = T35 + T36;
|
cannam@167
|
465 T38 = FMA(KP707106781, T37, T34);
|
cannam@167
|
466 T6Y = FNMS(KP707106781, T37, T34);
|
cannam@167
|
467 T39 = FNMS(KP198912367, T38, T2Z);
|
cannam@167
|
468 T7r = FNMS(KP668178637, T6X, T6Y);
|
cannam@167
|
469 T5H = FMA(KP198912367, T2Z, T38);
|
cannam@167
|
470 T6Z = FMA(KP668178637, T6Y, T6X);
|
cannam@167
|
471 }
|
cannam@167
|
472 }
|
cannam@167
|
473 {
|
cannam@167
|
474 E T8R, Tb9, T8U, Tba, T8Q, T8T;
|
cannam@167
|
475 T8Q = T3x - T3w;
|
cannam@167
|
476 T8R = FNMS(KP707106781, T8Q, T8P);
|
cannam@167
|
477 Tb9 = FMA(KP707106781, T8Q, T8P);
|
cannam@167
|
478 T8T = T3j + T3o;
|
cannam@167
|
479 T8U = FNMS(KP707106781, T8T, T8S);
|
cannam@167
|
480 Tba = FMA(KP707106781, T8T, T8S);
|
cannam@167
|
481 T8V = FMA(KP668178637, T8U, T8R);
|
cannam@167
|
482 TbC = FMA(KP198912367, Tb9, Tba);
|
cannam@167
|
483 T9S = FNMS(KP668178637, T8R, T8U);
|
cannam@167
|
484 Tbb = FNMS(KP198912367, Tba, Tb9);
|
cannam@167
|
485 }
|
cannam@167
|
486 {
|
cannam@167
|
487 E T3q, T70, T3z, T71, T3p, T3y;
|
cannam@167
|
488 T3p = T3j - T3o;
|
cannam@167
|
489 T3q = FMA(KP707106781, T3p, T3e);
|
cannam@167
|
490 T70 = FNMS(KP707106781, T3p, T3e);
|
cannam@167
|
491 T3y = T3w + T3x;
|
cannam@167
|
492 T3z = FMA(KP707106781, T3y, T3v);
|
cannam@167
|
493 T71 = FNMS(KP707106781, T3y, T3v);
|
cannam@167
|
494 T3A = FMA(KP198912367, T3z, T3q);
|
cannam@167
|
495 T7q = FMA(KP668178637, T70, T71);
|
cannam@167
|
496 T5G = FNMS(KP198912367, T3q, T3z);
|
cannam@167
|
497 T72 = FNMS(KP668178637, T71, T70);
|
cannam@167
|
498 }
|
cannam@167
|
499 {
|
cannam@167
|
500 E T8Y, Tbc, T91, Tbd, T8X, T90;
|
cannam@167
|
501 T8X = T35 - T36;
|
cannam@167
|
502 T8Y = FNMS(KP707106781, T8X, T8W);
|
cannam@167
|
503 Tbc = FMA(KP707106781, T8X, T8W);
|
cannam@167
|
504 T90 = T2S + T2X;
|
cannam@167
|
505 T91 = FNMS(KP707106781, T90, T8Z);
|
cannam@167
|
506 Tbd = FMA(KP707106781, T90, T8Z);
|
cannam@167
|
507 T92 = FMA(KP668178637, T91, T8Y);
|
cannam@167
|
508 TbD = FMA(KP198912367, Tbc, Tbd);
|
cannam@167
|
509 T9T = FNMS(KP668178637, T8Y, T91);
|
cannam@167
|
510 Tbe = FNMS(KP198912367, Tbd, Tbc);
|
cannam@167
|
511 }
|
cannam@167
|
512 }
|
cannam@167
|
513 {
|
cannam@167
|
514 E T18, Ted, ThI, T4A, T5f, T9g, TdY, T95, T1u, Te4, ThM, T52, T57, T9c, Te1;
|
cannam@167
|
515 E T9b, T1f, Teg, ThJ, T4F, T4K, T5h, TdZ, T5g, T1n, Te9, ThL, T4R, T4W, T99;
|
cannam@167
|
516 E Te6, T98;
|
cannam@167
|
517 {
|
cannam@167
|
518 E T14, T5b, T5e, TdX, T17, T4w, T4z, TdW;
|
cannam@167
|
519 {
|
cannam@167
|
520 E T12, T13, T5c, T5d;
|
cannam@167
|
521 T12 = cr[WS(rs, 1)];
|
cannam@167
|
522 T13 = ci[WS(rs, 30)];
|
cannam@167
|
523 T14 = T12 + T13;
|
cannam@167
|
524 T5b = T12 - T13;
|
cannam@167
|
525 T5c = ci[WS(rs, 46)];
|
cannam@167
|
526 T5d = cr[WS(rs, 49)];
|
cannam@167
|
527 T5e = T5c + T5d;
|
cannam@167
|
528 TdX = T5c - T5d;
|
cannam@167
|
529 }
|
cannam@167
|
530 {
|
cannam@167
|
531 E T15, T16, T4x, T4y;
|
cannam@167
|
532 T15 = cr[WS(rs, 17)];
|
cannam@167
|
533 T16 = ci[WS(rs, 14)];
|
cannam@167
|
534 T17 = T15 + T16;
|
cannam@167
|
535 T4w = T15 - T16;
|
cannam@167
|
536 T4x = ci[WS(rs, 62)];
|
cannam@167
|
537 T4y = cr[WS(rs, 33)];
|
cannam@167
|
538 T4z = T4x + T4y;
|
cannam@167
|
539 TdW = T4x - T4y;
|
cannam@167
|
540 }
|
cannam@167
|
541 T18 = T14 + T17;
|
cannam@167
|
542 Ted = T14 - T17;
|
cannam@167
|
543 ThI = TdW + TdX;
|
cannam@167
|
544 T4A = T4w + T4z;
|
cannam@167
|
545 T5f = T5b - T5e;
|
cannam@167
|
546 T9g = T5b + T5e;
|
cannam@167
|
547 TdY = TdW - TdX;
|
cannam@167
|
548 T95 = T4z - T4w;
|
cannam@167
|
549 }
|
cannam@167
|
550 {
|
cannam@167
|
551 E T1q, T53, T56, Te3, T1t, T4Y, T51, Te2;
|
cannam@167
|
552 {
|
cannam@167
|
553 E T1o, T1p, T54, T55;
|
cannam@167
|
554 T1o = ci[WS(rs, 2)];
|
cannam@167
|
555 T1p = cr[WS(rs, 29)];
|
cannam@167
|
556 T1q = T1o + T1p;
|
cannam@167
|
557 T53 = T1o - T1p;
|
cannam@167
|
558 T54 = ci[WS(rs, 50)];
|
cannam@167
|
559 T55 = cr[WS(rs, 45)];
|
cannam@167
|
560 T56 = T54 + T55;
|
cannam@167
|
561 Te3 = T54 - T55;
|
cannam@167
|
562 }
|
cannam@167
|
563 {
|
cannam@167
|
564 E T1r, T1s, T4Z, T50;
|
cannam@167
|
565 T1r = cr[WS(rs, 13)];
|
cannam@167
|
566 T1s = ci[WS(rs, 18)];
|
cannam@167
|
567 T1t = T1r + T1s;
|
cannam@167
|
568 T4Y = T1r - T1s;
|
cannam@167
|
569 T4Z = ci[WS(rs, 34)];
|
cannam@167
|
570 T50 = cr[WS(rs, 61)];
|
cannam@167
|
571 T51 = T4Z + T50;
|
cannam@167
|
572 Te2 = T4Z - T50;
|
cannam@167
|
573 }
|
cannam@167
|
574 T1u = T1q + T1t;
|
cannam@167
|
575 Te4 = Te2 - Te3;
|
cannam@167
|
576 ThM = Te2 + Te3;
|
cannam@167
|
577 T52 = T4Y - T51;
|
cannam@167
|
578 T57 = T53 - T56;
|
cannam@167
|
579 T9c = T4Y + T51;
|
cannam@167
|
580 Te1 = T1q - T1t;
|
cannam@167
|
581 T9b = T53 + T56;
|
cannam@167
|
582 }
|
cannam@167
|
583 {
|
cannam@167
|
584 E T1b, T4B, T4J, Tee, T1e, T4G, T4E, Tef;
|
cannam@167
|
585 {
|
cannam@167
|
586 E T19, T1a, T4H, T4I;
|
cannam@167
|
587 T19 = cr[WS(rs, 9)];
|
cannam@167
|
588 T1a = ci[WS(rs, 22)];
|
cannam@167
|
589 T1b = T19 + T1a;
|
cannam@167
|
590 T4B = T19 - T1a;
|
cannam@167
|
591 T4H = ci[WS(rs, 38)];
|
cannam@167
|
592 T4I = cr[WS(rs, 57)];
|
cannam@167
|
593 T4J = T4H + T4I;
|
cannam@167
|
594 Tee = T4H - T4I;
|
cannam@167
|
595 }
|
cannam@167
|
596 {
|
cannam@167
|
597 E T1c, T1d, T4C, T4D;
|
cannam@167
|
598 T1c = ci[WS(rs, 6)];
|
cannam@167
|
599 T1d = cr[WS(rs, 25)];
|
cannam@167
|
600 T1e = T1c + T1d;
|
cannam@167
|
601 T4G = T1c - T1d;
|
cannam@167
|
602 T4C = ci[WS(rs, 54)];
|
cannam@167
|
603 T4D = cr[WS(rs, 41)];
|
cannam@167
|
604 T4E = T4C + T4D;
|
cannam@167
|
605 Tef = T4C - T4D;
|
cannam@167
|
606 }
|
cannam@167
|
607 T1f = T1b + T1e;
|
cannam@167
|
608 Teg = Tee - Tef;
|
cannam@167
|
609 ThJ = Tef + Tee;
|
cannam@167
|
610 T4F = T4B + T4E;
|
cannam@167
|
611 T4K = T4G + T4J;
|
cannam@167
|
612 T5h = T4G - T4J;
|
cannam@167
|
613 TdZ = T1b - T1e;
|
cannam@167
|
614 T5g = T4B - T4E;
|
cannam@167
|
615 }
|
cannam@167
|
616 {
|
cannam@167
|
617 E T1j, T4S, T4V, Te8, T1m, T4N, T4Q, Te7;
|
cannam@167
|
618 {
|
cannam@167
|
619 E T1h, T1i, T4T, T4U;
|
cannam@167
|
620 T1h = cr[WS(rs, 5)];
|
cannam@167
|
621 T1i = ci[WS(rs, 26)];
|
cannam@167
|
622 T1j = T1h + T1i;
|
cannam@167
|
623 T4S = T1h - T1i;
|
cannam@167
|
624 T4T = ci[WS(rs, 42)];
|
cannam@167
|
625 T4U = cr[WS(rs, 53)];
|
cannam@167
|
626 T4V = T4T + T4U;
|
cannam@167
|
627 Te8 = T4T - T4U;
|
cannam@167
|
628 }
|
cannam@167
|
629 {
|
cannam@167
|
630 E T1k, T1l, T4O, T4P;
|
cannam@167
|
631 T1k = cr[WS(rs, 21)];
|
cannam@167
|
632 T1l = ci[WS(rs, 10)];
|
cannam@167
|
633 T1m = T1k + T1l;
|
cannam@167
|
634 T4N = T1k - T1l;
|
cannam@167
|
635 T4O = ci[WS(rs, 58)];
|
cannam@167
|
636 T4P = cr[WS(rs, 37)];
|
cannam@167
|
637 T4Q = T4O + T4P;
|
cannam@167
|
638 Te7 = T4O - T4P;
|
cannam@167
|
639 }
|
cannam@167
|
640 T1n = T1j + T1m;
|
cannam@167
|
641 Te9 = Te7 - Te8;
|
cannam@167
|
642 ThL = Te7 + Te8;
|
cannam@167
|
643 T4R = T4N + T4Q;
|
cannam@167
|
644 T4W = T4S - T4V;
|
cannam@167
|
645 T99 = T4Q - T4N;
|
cannam@167
|
646 Te6 = T1j - T1m;
|
cannam@167
|
647 T98 = T4S + T4V;
|
cannam@167
|
648 }
|
cannam@167
|
649 {
|
cannam@167
|
650 E T1g, T1v, Tjo, Tjp;
|
cannam@167
|
651 T1g = T18 + T1f;
|
cannam@167
|
652 T1v = T1n + T1u;
|
cannam@167
|
653 T1w = T1g + T1v;
|
cannam@167
|
654 ThH = T1g - T1v;
|
cannam@167
|
655 Tjo = ThI - ThJ;
|
cannam@167
|
656 Tjp = T1n - T1u;
|
cannam@167
|
657 Tjq = Tjo - Tjp;
|
cannam@167
|
658 Tke = Tjp + Tjo;
|
cannam@167
|
659 }
|
cannam@167
|
660 {
|
cannam@167
|
661 E Tjr, Tjs, ThK, ThN;
|
cannam@167
|
662 Tjr = T18 - T1f;
|
cannam@167
|
663 Tjs = ThM - ThL;
|
cannam@167
|
664 Tjt = Tjr - Tjs;
|
cannam@167
|
665 Tkf = Tjr + Tjs;
|
cannam@167
|
666 ThK = ThI + ThJ;
|
cannam@167
|
667 ThN = ThL + ThM;
|
cannam@167
|
668 ThO = ThK - ThN;
|
cannam@167
|
669 TiK = ThK + ThN;
|
cannam@167
|
670 }
|
cannam@167
|
671 {
|
cannam@167
|
672 E Te0, Tg9, Teb, Tga, Te5, Tea;
|
cannam@167
|
673 Te0 = TdY - TdZ;
|
cannam@167
|
674 Tg9 = Ted + Teg;
|
cannam@167
|
675 Te5 = Te1 + Te4;
|
cannam@167
|
676 Tea = Te6 - Te9;
|
cannam@167
|
677 Teb = Te5 - Tea;
|
cannam@167
|
678 Tga = Tea + Te5;
|
cannam@167
|
679 Tec = FNMS(KP707106781, Teb, Te0);
|
cannam@167
|
680 TgT = FMA(KP707106781, Tga, Tg9);
|
cannam@167
|
681 Tfc = FMA(KP707106781, Teb, Te0);
|
cannam@167
|
682 Tgb = FNMS(KP707106781, Tga, Tg9);
|
cannam@167
|
683 }
|
cannam@167
|
684 {
|
cannam@167
|
685 E Teh, Tg6, Tek, Tg7, Tei, Tej;
|
cannam@167
|
686 Teh = Ted - Teg;
|
cannam@167
|
687 Tg6 = TdZ + TdY;
|
cannam@167
|
688 Tei = Te6 + Te9;
|
cannam@167
|
689 Tej = Te4 - Te1;
|
cannam@167
|
690 Tek = Tei - Tej;
|
cannam@167
|
691 Tg7 = Tei + Tej;
|
cannam@167
|
692 Tel = FNMS(KP707106781, Tek, Teh);
|
cannam@167
|
693 TgU = FMA(KP707106781, Tg7, Tg6);
|
cannam@167
|
694 Tfd = FMA(KP707106781, Tek, Teh);
|
cannam@167
|
695 Tg8 = FNMS(KP707106781, Tg7, Tg6);
|
cannam@167
|
696 }
|
cannam@167
|
697 {
|
cannam@167
|
698 E T4M, T78, T5j, T75, T59, T76, T5m, T79, T4L, T5i;
|
cannam@167
|
699 T4L = T4F - T4K;
|
cannam@167
|
700 T4M = FMA(KP707106781, T4L, T4A);
|
cannam@167
|
701 T78 = FNMS(KP707106781, T4L, T4A);
|
cannam@167
|
702 T5i = T5g + T5h;
|
cannam@167
|
703 T5j = FMA(KP707106781, T5i, T5f);
|
cannam@167
|
704 T75 = FNMS(KP707106781, T5i, T5f);
|
cannam@167
|
705 {
|
cannam@167
|
706 E T4X, T58, T5k, T5l;
|
cannam@167
|
707 T4X = FMA(KP414213562, T4W, T4R);
|
cannam@167
|
708 T58 = FNMS(KP414213562, T57, T52);
|
cannam@167
|
709 T59 = T4X + T58;
|
cannam@167
|
710 T76 = T4X - T58;
|
cannam@167
|
711 T5k = FNMS(KP414213562, T4R, T4W);
|
cannam@167
|
712 T5l = FMA(KP414213562, T52, T57);
|
cannam@167
|
713 T5m = T5k + T5l;
|
cannam@167
|
714 T79 = T5l - T5k;
|
cannam@167
|
715 }
|
cannam@167
|
716 T5a = FNMS(KP923879532, T59, T4M);
|
cannam@167
|
717 T82 = FMA(KP923879532, T79, T78);
|
cannam@167
|
718 T83 = FMA(KP923879532, T76, T75);
|
cannam@167
|
719 T5n = FNMS(KP923879532, T5m, T5j);
|
cannam@167
|
720 T6i = FMA(KP923879532, T59, T4M);
|
cannam@167
|
721 T77 = FNMS(KP923879532, T76, T75);
|
cannam@167
|
722 T7a = FNMS(KP923879532, T79, T78);
|
cannam@167
|
723 T6j = FMA(KP923879532, T5m, T5j);
|
cannam@167
|
724 }
|
cannam@167
|
725 {
|
cannam@167
|
726 E T97, Tbk, T9i, Tbh, T9e, Tbi, T9l, Tbl, T96, T9h;
|
cannam@167
|
727 T96 = T5h - T5g;
|
cannam@167
|
728 T97 = FNMS(KP707106781, T96, T95);
|
cannam@167
|
729 Tbk = FMA(KP707106781, T96, T95);
|
cannam@167
|
730 T9h = T4F + T4K;
|
cannam@167
|
731 T9i = FNMS(KP707106781, T9h, T9g);
|
cannam@167
|
732 Tbh = FMA(KP707106781, T9h, T9g);
|
cannam@167
|
733 {
|
cannam@167
|
734 E T9a, T9d, T9j, T9k;
|
cannam@167
|
735 T9a = FMA(KP414213562, T99, T98);
|
cannam@167
|
736 T9d = FMA(KP414213562, T9c, T9b);
|
cannam@167
|
737 T9e = T9a - T9d;
|
cannam@167
|
738 Tbi = T9a + T9d;
|
cannam@167
|
739 T9j = FNMS(KP414213562, T98, T99);
|
cannam@167
|
740 T9k = FNMS(KP414213562, T9b, T9c);
|
cannam@167
|
741 T9l = T9j + T9k;
|
cannam@167
|
742 Tbl = T9j - T9k;
|
cannam@167
|
743 }
|
cannam@167
|
744 T9f = FNMS(KP923879532, T9e, T97);
|
cannam@167
|
745 Tcb = FMA(KP923879532, Tbl, Tbk);
|
cannam@167
|
746 Tcc = FMA(KP923879532, Tbi, Tbh);
|
cannam@167
|
747 T9m = FMA(KP923879532, T9l, T9i);
|
cannam@167
|
748 Tar = FNMS(KP923879532, T9l, T9i);
|
cannam@167
|
749 Tbj = FNMS(KP923879532, Tbi, Tbh);
|
cannam@167
|
750 Tbm = FNMS(KP923879532, Tbl, Tbk);
|
cannam@167
|
751 Tas = FMA(KP923879532, T9e, T97);
|
cannam@167
|
752 }
|
cannam@167
|
753 }
|
cannam@167
|
754 {
|
cannam@167
|
755 E T1D, TdM, ThR, T3H, T4m, T9z, Tdx, T9o, T1Z, TdD, ThV, T49, T4e, T9s, TdA;
|
cannam@167
|
756 E T9r, T1K, TdP, ThS, T3M, T3R, T4o, Tdy, T4n, T1S, TdI, ThU, T3Y, T43, T9v;
|
cannam@167
|
757 E TdF, T9u;
|
cannam@167
|
758 {
|
cannam@167
|
759 E T1z, T4i, T4l, Tdw, T1C, T3D, T3G, Tdv;
|
cannam@167
|
760 {
|
cannam@167
|
761 E T1x, T1y, T4j, T4k;
|
cannam@167
|
762 T1x = ci[0];
|
cannam@167
|
763 T1y = cr[WS(rs, 31)];
|
cannam@167
|
764 T1z = T1x + T1y;
|
cannam@167
|
765 T4i = T1x - T1y;
|
cannam@167
|
766 T4j = ci[WS(rs, 48)];
|
cannam@167
|
767 T4k = cr[WS(rs, 47)];
|
cannam@167
|
768 T4l = T4j + T4k;
|
cannam@167
|
769 Tdw = T4j - T4k;
|
cannam@167
|
770 }
|
cannam@167
|
771 {
|
cannam@167
|
772 E T1A, T1B, T3E, T3F;
|
cannam@167
|
773 T1A = cr[WS(rs, 15)];
|
cannam@167
|
774 T1B = ci[WS(rs, 16)];
|
cannam@167
|
775 T1C = T1A + T1B;
|
cannam@167
|
776 T3D = T1A - T1B;
|
cannam@167
|
777 T3E = ci[WS(rs, 32)];
|
cannam@167
|
778 T3F = cr[WS(rs, 63)];
|
cannam@167
|
779 T3G = T3E + T3F;
|
cannam@167
|
780 Tdv = T3E - T3F;
|
cannam@167
|
781 }
|
cannam@167
|
782 T1D = T1z + T1C;
|
cannam@167
|
783 TdM = T1z - T1C;
|
cannam@167
|
784 ThR = Tdv + Tdw;
|
cannam@167
|
785 T3H = T3D - T3G;
|
cannam@167
|
786 T4m = T4i - T4l;
|
cannam@167
|
787 T9z = T4i + T4l;
|
cannam@167
|
788 Tdx = Tdv - Tdw;
|
cannam@167
|
789 T9o = T3D + T3G;
|
cannam@167
|
790 }
|
cannam@167
|
791 {
|
cannam@167
|
792 E T1V, T4a, T4d, TdC, T1Y, T45, T48, TdB;
|
cannam@167
|
793 {
|
cannam@167
|
794 E T1T, T1U, T4b, T4c;
|
cannam@167
|
795 T1T = ci[WS(rs, 4)];
|
cannam@167
|
796 T1U = cr[WS(rs, 27)];
|
cannam@167
|
797 T1V = T1T + T1U;
|
cannam@167
|
798 T4a = T1T - T1U;
|
cannam@167
|
799 T4b = ci[WS(rs, 52)];
|
cannam@167
|
800 T4c = cr[WS(rs, 43)];
|
cannam@167
|
801 T4d = T4b + T4c;
|
cannam@167
|
802 TdC = T4b - T4c;
|
cannam@167
|
803 }
|
cannam@167
|
804 {
|
cannam@167
|
805 E T1W, T1X, T46, T47;
|
cannam@167
|
806 T1W = cr[WS(rs, 11)];
|
cannam@167
|
807 T1X = ci[WS(rs, 20)];
|
cannam@167
|
808 T1Y = T1W + T1X;
|
cannam@167
|
809 T45 = T1W - T1X;
|
cannam@167
|
810 T46 = ci[WS(rs, 36)];
|
cannam@167
|
811 T47 = cr[WS(rs, 59)];
|
cannam@167
|
812 T48 = T46 + T47;
|
cannam@167
|
813 TdB = T46 - T47;
|
cannam@167
|
814 }
|
cannam@167
|
815 T1Z = T1V + T1Y;
|
cannam@167
|
816 TdD = TdB - TdC;
|
cannam@167
|
817 ThV = TdB + TdC;
|
cannam@167
|
818 T49 = T45 - T48;
|
cannam@167
|
819 T4e = T4a - T4d;
|
cannam@167
|
820 T9s = T45 + T48;
|
cannam@167
|
821 TdA = T1V - T1Y;
|
cannam@167
|
822 T9r = T4a + T4d;
|
cannam@167
|
823 }
|
cannam@167
|
824 {
|
cannam@167
|
825 E T1G, T3I, T3Q, TdN, T1J, T3N, T3L, TdO;
|
cannam@167
|
826 {
|
cannam@167
|
827 E T1E, T1F, T3O, T3P;
|
cannam@167
|
828 T1E = cr[WS(rs, 7)];
|
cannam@167
|
829 T1F = ci[WS(rs, 24)];
|
cannam@167
|
830 T1G = T1E + T1F;
|
cannam@167
|
831 T3I = T1E - T1F;
|
cannam@167
|
832 T3O = ci[WS(rs, 40)];
|
cannam@167
|
833 T3P = cr[WS(rs, 55)];
|
cannam@167
|
834 T3Q = T3O + T3P;
|
cannam@167
|
835 TdN = T3O - T3P;
|
cannam@167
|
836 }
|
cannam@167
|
837 {
|
cannam@167
|
838 E T1H, T1I, T3J, T3K;
|
cannam@167
|
839 T1H = ci[WS(rs, 8)];
|
cannam@167
|
840 T1I = cr[WS(rs, 23)];
|
cannam@167
|
841 T1J = T1H + T1I;
|
cannam@167
|
842 T3N = T1H - T1I;
|
cannam@167
|
843 T3J = ci[WS(rs, 56)];
|
cannam@167
|
844 T3K = cr[WS(rs, 39)];
|
cannam@167
|
845 T3L = T3J + T3K;
|
cannam@167
|
846 TdO = T3J - T3K;
|
cannam@167
|
847 }
|
cannam@167
|
848 T1K = T1G + T1J;
|
cannam@167
|
849 TdP = TdN - TdO;
|
cannam@167
|
850 ThS = TdO + TdN;
|
cannam@167
|
851 T3M = T3I + T3L;
|
cannam@167
|
852 T3R = T3N + T3Q;
|
cannam@167
|
853 T4o = T3N - T3Q;
|
cannam@167
|
854 Tdy = T1G - T1J;
|
cannam@167
|
855 T4n = T3I - T3L;
|
cannam@167
|
856 }
|
cannam@167
|
857 {
|
cannam@167
|
858 E T1O, T3Z, T42, TdH, T1R, T3U, T3X, TdG;
|
cannam@167
|
859 {
|
cannam@167
|
860 E T1M, T1N, T40, T41;
|
cannam@167
|
861 T1M = cr[WS(rs, 3)];
|
cannam@167
|
862 T1N = ci[WS(rs, 28)];
|
cannam@167
|
863 T1O = T1M + T1N;
|
cannam@167
|
864 T3Z = T1M - T1N;
|
cannam@167
|
865 T40 = ci[WS(rs, 44)];
|
cannam@167
|
866 T41 = cr[WS(rs, 51)];
|
cannam@167
|
867 T42 = T40 + T41;
|
cannam@167
|
868 TdH = T40 - T41;
|
cannam@167
|
869 }
|
cannam@167
|
870 {
|
cannam@167
|
871 E T1P, T1Q, T3V, T3W;
|
cannam@167
|
872 T1P = cr[WS(rs, 19)];
|
cannam@167
|
873 T1Q = ci[WS(rs, 12)];
|
cannam@167
|
874 T1R = T1P + T1Q;
|
cannam@167
|
875 T3U = T1P - T1Q;
|
cannam@167
|
876 T3V = ci[WS(rs, 60)];
|
cannam@167
|
877 T3W = cr[WS(rs, 35)];
|
cannam@167
|
878 T3X = T3V + T3W;
|
cannam@167
|
879 TdG = T3V - T3W;
|
cannam@167
|
880 }
|
cannam@167
|
881 T1S = T1O + T1R;
|
cannam@167
|
882 TdI = TdG - TdH;
|
cannam@167
|
883 ThU = TdG + TdH;
|
cannam@167
|
884 T3Y = T3U + T3X;
|
cannam@167
|
885 T43 = T3Z - T42;
|
cannam@167
|
886 T9v = T3U - T3X;
|
cannam@167
|
887 TdF = T1O - T1R;
|
cannam@167
|
888 T9u = T3Z + T42;
|
cannam@167
|
889 }
|
cannam@167
|
890 {
|
cannam@167
|
891 E T1L, T20, Tjh, Tji;
|
cannam@167
|
892 T1L = T1D + T1K;
|
cannam@167
|
893 T20 = T1S + T1Z;
|
cannam@167
|
894 T21 = T1L + T20;
|
cannam@167
|
895 ThQ = T1L - T20;
|
cannam@167
|
896 Tjh = ThR - ThS;
|
cannam@167
|
897 Tji = T1S - T1Z;
|
cannam@167
|
898 Tjj = Tjh - Tji;
|
cannam@167
|
899 Tkb = Tji + Tjh;
|
cannam@167
|
900 }
|
cannam@167
|
901 {
|
cannam@167
|
902 E Tjk, Tjl, ThT, ThW;
|
cannam@167
|
903 Tjk = T1D - T1K;
|
cannam@167
|
904 Tjl = ThV - ThU;
|
cannam@167
|
905 Tjm = Tjk - Tjl;
|
cannam@167
|
906 Tkc = Tjk + Tjl;
|
cannam@167
|
907 ThT = ThR + ThS;
|
cannam@167
|
908 ThW = ThU + ThV;
|
cannam@167
|
909 ThX = ThT - ThW;
|
cannam@167
|
910 TiL = ThT + ThW;
|
cannam@167
|
911 }
|
cannam@167
|
912 {
|
cannam@167
|
913 E Tdz, Tg2, TdK, Tg3, TdE, TdJ;
|
cannam@167
|
914 Tdz = Tdx - Tdy;
|
cannam@167
|
915 Tg2 = TdM + TdP;
|
cannam@167
|
916 TdE = TdA + TdD;
|
cannam@167
|
917 TdJ = TdF - TdI;
|
cannam@167
|
918 TdK = TdE - TdJ;
|
cannam@167
|
919 Tg3 = TdJ + TdE;
|
cannam@167
|
920 TdL = FNMS(KP707106781, TdK, Tdz);
|
cannam@167
|
921 TgW = FMA(KP707106781, Tg3, Tg2);
|
cannam@167
|
922 Tf9 = FMA(KP707106781, TdK, Tdz);
|
cannam@167
|
923 Tg4 = FNMS(KP707106781, Tg3, Tg2);
|
cannam@167
|
924 }
|
cannam@167
|
925 {
|
cannam@167
|
926 E TdQ, TfZ, TdT, Tg0, TdR, TdS;
|
cannam@167
|
927 TdQ = TdM - TdP;
|
cannam@167
|
928 TfZ = Tdy + Tdx;
|
cannam@167
|
929 TdR = TdF + TdI;
|
cannam@167
|
930 TdS = TdD - TdA;
|
cannam@167
|
931 TdT = TdR - TdS;
|
cannam@167
|
932 Tg0 = TdR + TdS;
|
cannam@167
|
933 TdU = FNMS(KP707106781, TdT, TdQ);
|
cannam@167
|
934 TgX = FMA(KP707106781, Tg0, TfZ);
|
cannam@167
|
935 Tfa = FMA(KP707106781, TdT, TdQ);
|
cannam@167
|
936 Tg1 = FNMS(KP707106781, Tg0, TfZ);
|
cannam@167
|
937 }
|
cannam@167
|
938 {
|
cannam@167
|
939 E T3T, T7f, T4q, T7c, T4g, T7d, T4t, T7g, T3S, T4p;
|
cannam@167
|
940 T3S = T3M - T3R;
|
cannam@167
|
941 T3T = FMA(KP707106781, T3S, T3H);
|
cannam@167
|
942 T7f = FNMS(KP707106781, T3S, T3H);
|
cannam@167
|
943 T4p = T4n + T4o;
|
cannam@167
|
944 T4q = FMA(KP707106781, T4p, T4m);
|
cannam@167
|
945 T7c = FNMS(KP707106781, T4p, T4m);
|
cannam@167
|
946 {
|
cannam@167
|
947 E T44, T4f, T4r, T4s;
|
cannam@167
|
948 T44 = FMA(KP414213562, T43, T3Y);
|
cannam@167
|
949 T4f = FNMS(KP414213562, T4e, T49);
|
cannam@167
|
950 T4g = T44 + T4f;
|
cannam@167
|
951 T7d = T44 - T4f;
|
cannam@167
|
952 T4r = FNMS(KP414213562, T3Y, T43);
|
cannam@167
|
953 T4s = FMA(KP414213562, T49, T4e);
|
cannam@167
|
954 T4t = T4r + T4s;
|
cannam@167
|
955 T7g = T4s - T4r;
|
cannam@167
|
956 }
|
cannam@167
|
957 T4h = FNMS(KP923879532, T4g, T3T);
|
cannam@167
|
958 T7Z = FMA(KP923879532, T7g, T7f);
|
cannam@167
|
959 T80 = FMA(KP923879532, T7d, T7c);
|
cannam@167
|
960 T4u = FNMS(KP923879532, T4t, T4q);
|
cannam@167
|
961 T6f = FMA(KP923879532, T4g, T3T);
|
cannam@167
|
962 T7e = FNMS(KP923879532, T7d, T7c);
|
cannam@167
|
963 T7h = FNMS(KP923879532, T7g, T7f);
|
cannam@167
|
964 T6g = FMA(KP923879532, T4t, T4q);
|
cannam@167
|
965 }
|
cannam@167
|
966 {
|
cannam@167
|
967 E T9q, Tbr, T9B, Tbo, T9x, Tbp, T9E, Tbs, T9p, T9A;
|
cannam@167
|
968 T9p = T4n - T4o;
|
cannam@167
|
969 T9q = FNMS(KP707106781, T9p, T9o);
|
cannam@167
|
970 Tbr = FMA(KP707106781, T9p, T9o);
|
cannam@167
|
971 T9A = T3M + T3R;
|
cannam@167
|
972 T9B = FNMS(KP707106781, T9A, T9z);
|
cannam@167
|
973 Tbo = FMA(KP707106781, T9A, T9z);
|
cannam@167
|
974 {
|
cannam@167
|
975 E T9t, T9w, T9C, T9D;
|
cannam@167
|
976 T9t = FMA(KP414213562, T9s, T9r);
|
cannam@167
|
977 T9w = FNMS(KP414213562, T9v, T9u);
|
cannam@167
|
978 T9x = T9t - T9w;
|
cannam@167
|
979 Tbp = T9w + T9t;
|
cannam@167
|
980 T9C = FMA(KP414213562, T9u, T9v);
|
cannam@167
|
981 T9D = FNMS(KP414213562, T9r, T9s);
|
cannam@167
|
982 T9E = T9C - T9D;
|
cannam@167
|
983 Tbs = T9C + T9D;
|
cannam@167
|
984 }
|
cannam@167
|
985 T9y = FNMS(KP923879532, T9x, T9q);
|
cannam@167
|
986 Tce = FMA(KP923879532, Tbs, Tbr);
|
cannam@167
|
987 Tcf = FMA(KP923879532, Tbp, Tbo);
|
cannam@167
|
988 T9F = FNMS(KP923879532, T9E, T9B);
|
cannam@167
|
989 Tau = FMA(KP923879532, T9E, T9B);
|
cannam@167
|
990 Tbq = FNMS(KP923879532, Tbp, Tbo);
|
cannam@167
|
991 Tbt = FNMS(KP923879532, Tbs, Tbr);
|
cannam@167
|
992 Tav = FMA(KP923879532, T9x, T9q);
|
cannam@167
|
993 }
|
cannam@167
|
994 }
|
cannam@167
|
995 {
|
cannam@167
|
996 E T11, T22, TiE, TiJ, TiM, TiN;
|
cannam@167
|
997 T11 = Tv + T10;
|
cannam@167
|
998 T22 = T1w + T21;
|
cannam@167
|
999 TiE = T11 - T22;
|
cannam@167
|
1000 TiJ = TiH + TiI;
|
cannam@167
|
1001 TiM = TiK + TiL;
|
cannam@167
|
1002 TiN = TiJ - TiM;
|
cannam@167
|
1003 cr[0] = T11 + T22;
|
cannam@167
|
1004 ci[0] = TiJ + TiM;
|
cannam@167
|
1005 {
|
cannam@167
|
1006 E TiD, TiF, TiG, TiO;
|
cannam@167
|
1007 TiD = W[62];
|
cannam@167
|
1008 TiF = TiD * TiE;
|
cannam@167
|
1009 TiG = W[63];
|
cannam@167
|
1010 TiO = TiG * TiE;
|
cannam@167
|
1011 cr[WS(rs, 32)] = FNMS(TiG, TiN, TiF);
|
cannam@167
|
1012 ci[WS(rs, 32)] = FMA(TiD, TiN, TiO);
|
cannam@167
|
1013 }
|
cannam@167
|
1014 }
|
cannam@167
|
1015 {
|
cannam@167
|
1016 E TiS, Tj0, TiX, Tj3;
|
cannam@167
|
1017 {
|
cannam@167
|
1018 E TiQ, TiR, TiV, TiW;
|
cannam@167
|
1019 TiQ = Tv - T10;
|
cannam@167
|
1020 TiR = TiL - TiK;
|
cannam@167
|
1021 TiS = TiQ - TiR;
|
cannam@167
|
1022 Tj0 = TiQ + TiR;
|
cannam@167
|
1023 TiV = TiH - TiI;
|
cannam@167
|
1024 TiW = T1w - T21;
|
cannam@167
|
1025 TiX = TiV - TiW;
|
cannam@167
|
1026 Tj3 = TiW + TiV;
|
cannam@167
|
1027 }
|
cannam@167
|
1028 {
|
cannam@167
|
1029 E TiT, TiY, TiP, TiU;
|
cannam@167
|
1030 TiP = W[94];
|
cannam@167
|
1031 TiT = TiP * TiS;
|
cannam@167
|
1032 TiY = TiP * TiX;
|
cannam@167
|
1033 TiU = W[95];
|
cannam@167
|
1034 cr[WS(rs, 48)] = FNMS(TiU, TiX, TiT);
|
cannam@167
|
1035 ci[WS(rs, 48)] = FMA(TiU, TiS, TiY);
|
cannam@167
|
1036 }
|
cannam@167
|
1037 {
|
cannam@167
|
1038 E Tj1, Tj4, TiZ, Tj2;
|
cannam@167
|
1039 TiZ = W[30];
|
cannam@167
|
1040 Tj1 = TiZ * Tj0;
|
cannam@167
|
1041 Tj4 = TiZ * Tj3;
|
cannam@167
|
1042 Tj2 = W[31];
|
cannam@167
|
1043 cr[WS(rs, 16)] = FNMS(Tj2, Tj3, Tj1);
|
cannam@167
|
1044 ci[WS(rs, 16)] = FMA(Tj2, Tj0, Tj4);
|
cannam@167
|
1045 }
|
cannam@167
|
1046 }
|
cannam@167
|
1047 {
|
cannam@167
|
1048 E Tib, Tie, Tiy, Tiq, Ti0, TiB, Tii, Tiv;
|
cannam@167
|
1049 Tib = Ti3 + Tia;
|
cannam@167
|
1050 {
|
cannam@167
|
1051 E Tio, Tic, Tid, Tip;
|
cannam@167
|
1052 Tio = Thy - ThF;
|
cannam@167
|
1053 Tic = ThH + ThO;
|
cannam@167
|
1054 Tid = ThX - ThQ;
|
cannam@167
|
1055 Tip = Tid - Tic;
|
cannam@167
|
1056 Tie = Tic + Tid;
|
cannam@167
|
1057 Tiy = FMA(KP707106781, Tip, Tio);
|
cannam@167
|
1058 Tiq = FNMS(KP707106781, Tip, Tio);
|
cannam@167
|
1059 }
|
cannam@167
|
1060 {
|
cannam@167
|
1061 E ThG, Tit, ThZ, Tiu, ThP, ThY;
|
cannam@167
|
1062 ThG = Thy + ThF;
|
cannam@167
|
1063 Tit = Tia - Ti3;
|
cannam@167
|
1064 ThP = ThH - ThO;
|
cannam@167
|
1065 ThY = ThQ + ThX;
|
cannam@167
|
1066 ThZ = ThP + ThY;
|
cannam@167
|
1067 Tiu = ThP - ThY;
|
cannam@167
|
1068 Ti0 = FNMS(KP707106781, ThZ, ThG);
|
cannam@167
|
1069 TiB = FMA(KP707106781, Tiu, Tit);
|
cannam@167
|
1070 Tii = FMA(KP707106781, ThZ, ThG);
|
cannam@167
|
1071 Tiv = FNMS(KP707106781, Tiu, Tit);
|
cannam@167
|
1072 }
|
cannam@167
|
1073 {
|
cannam@167
|
1074 E Tir, Tiw, Tin, Tis;
|
cannam@167
|
1075 Tin = W[110];
|
cannam@167
|
1076 Tir = Tin * Tiq;
|
cannam@167
|
1077 Tiw = Tin * Tiv;
|
cannam@167
|
1078 Tis = W[111];
|
cannam@167
|
1079 cr[WS(rs, 56)] = FNMS(Tis, Tiv, Tir);
|
cannam@167
|
1080 ci[WS(rs, 56)] = FMA(Tis, Tiq, Tiw);
|
cannam@167
|
1081 }
|
cannam@167
|
1082 {
|
cannam@167
|
1083 E Tiz, TiC, Tix, TiA;
|
cannam@167
|
1084 Tix = W[46];
|
cannam@167
|
1085 Tiz = Tix * Tiy;
|
cannam@167
|
1086 TiC = Tix * TiB;
|
cannam@167
|
1087 TiA = W[47];
|
cannam@167
|
1088 cr[WS(rs, 24)] = FNMS(TiA, TiB, Tiz);
|
cannam@167
|
1089 ci[WS(rs, 24)] = FMA(TiA, Tiy, TiC);
|
cannam@167
|
1090 }
|
cannam@167
|
1091 {
|
cannam@167
|
1092 E Tif, Ti2, Tig, Thx, Ti1;
|
cannam@167
|
1093 Tif = FNMS(KP707106781, Tie, Tib);
|
cannam@167
|
1094 Ti2 = W[79];
|
cannam@167
|
1095 Tig = Ti2 * Ti0;
|
cannam@167
|
1096 Thx = W[78];
|
cannam@167
|
1097 Ti1 = Thx * Ti0;
|
cannam@167
|
1098 cr[WS(rs, 40)] = FNMS(Ti2, Tif, Ti1);
|
cannam@167
|
1099 ci[WS(rs, 40)] = FMA(Thx, Tif, Tig);
|
cannam@167
|
1100 }
|
cannam@167
|
1101 {
|
cannam@167
|
1102 E Til, Tik, Tim, Tih, Tij;
|
cannam@167
|
1103 Til = FMA(KP707106781, Tie, Tib);
|
cannam@167
|
1104 Tik = W[15];
|
cannam@167
|
1105 Tim = Tik * Tii;
|
cannam@167
|
1106 Tih = W[14];
|
cannam@167
|
1107 Tij = Tih * Tii;
|
cannam@167
|
1108 cr[WS(rs, 8)] = FNMS(Tik, Til, Tij);
|
cannam@167
|
1109 ci[WS(rs, 8)] = FMA(Tih, Til, Tim);
|
cannam@167
|
1110 }
|
cannam@167
|
1111 }
|
cannam@167
|
1112 {
|
cannam@167
|
1113 E Tjw, Tk2, Tk5, TjF, TjI, TjU, TjZ, TjM;
|
cannam@167
|
1114 {
|
cannam@167
|
1115 E TjE, TjX, Tjg, TjS, TjG, TjH, TjT, Tjv, TjY, Tjf, Tjn, Tju;
|
cannam@167
|
1116 TjE = TjC - TjD;
|
cannam@167
|
1117 TjX = FNMS(KP707106781, TjE, TjB);
|
cannam@167
|
1118 Tjf = Tjb - Tje;
|
cannam@167
|
1119 Tjg = FMA(KP707106781, Tjf, Tj8);
|
cannam@167
|
1120 TjS = FNMS(KP707106781, Tjf, Tj8);
|
cannam@167
|
1121 TjG = FMA(KP414213562, Tjq, Tjt);
|
cannam@167
|
1122 TjH = FNMS(KP414213562, Tjj, Tjm);
|
cannam@167
|
1123 TjT = TjG + TjH;
|
cannam@167
|
1124 Tjn = FMA(KP414213562, Tjm, Tjj);
|
cannam@167
|
1125 Tju = FNMS(KP414213562, Tjt, Tjq);
|
cannam@167
|
1126 Tjv = Tjn - Tju;
|
cannam@167
|
1127 TjY = Tju + Tjn;
|
cannam@167
|
1128 Tjw = FNMS(KP923879532, Tjv, Tjg);
|
cannam@167
|
1129 Tk2 = FMA(KP923879532, TjT, TjS);
|
cannam@167
|
1130 Tk5 = FMA(KP923879532, TjY, TjX);
|
cannam@167
|
1131 TjF = FMA(KP707106781, TjE, TjB);
|
cannam@167
|
1132 TjI = TjG - TjH;
|
cannam@167
|
1133 TjU = FNMS(KP923879532, TjT, TjS);
|
cannam@167
|
1134 TjZ = FNMS(KP923879532, TjY, TjX);
|
cannam@167
|
1135 TjM = FMA(KP923879532, Tjv, Tjg);
|
cannam@167
|
1136 }
|
cannam@167
|
1137 {
|
cannam@167
|
1138 E TjV, Tk0, TjR, TjW;
|
cannam@167
|
1139 TjR = W[54];
|
cannam@167
|
1140 TjV = TjR * TjU;
|
cannam@167
|
1141 Tk0 = TjR * TjZ;
|
cannam@167
|
1142 TjW = W[55];
|
cannam@167
|
1143 cr[WS(rs, 28)] = FNMS(TjW, TjZ, TjV);
|
cannam@167
|
1144 ci[WS(rs, 28)] = FMA(TjW, TjU, Tk0);
|
cannam@167
|
1145 }
|
cannam@167
|
1146 {
|
cannam@167
|
1147 E Tk3, Tk6, Tk1, Tk4;
|
cannam@167
|
1148 Tk1 = W[118];
|
cannam@167
|
1149 Tk3 = Tk1 * Tk2;
|
cannam@167
|
1150 Tk6 = Tk1 * Tk5;
|
cannam@167
|
1151 Tk4 = W[119];
|
cannam@167
|
1152 cr[WS(rs, 60)] = FNMS(Tk4, Tk5, Tk3);
|
cannam@167
|
1153 ci[WS(rs, 60)] = FMA(Tk4, Tk2, Tk6);
|
cannam@167
|
1154 }
|
cannam@167
|
1155 {
|
cannam@167
|
1156 E TjJ, Tjy, TjK, Tj5, Tjx;
|
cannam@167
|
1157 TjJ = FNMS(KP923879532, TjI, TjF);
|
cannam@167
|
1158 Tjy = W[87];
|
cannam@167
|
1159 TjK = Tjy * Tjw;
|
cannam@167
|
1160 Tj5 = W[86];
|
cannam@167
|
1161 Tjx = Tj5 * Tjw;
|
cannam@167
|
1162 cr[WS(rs, 44)] = FNMS(Tjy, TjJ, Tjx);
|
cannam@167
|
1163 ci[WS(rs, 44)] = FMA(Tj5, TjJ, TjK);
|
cannam@167
|
1164 }
|
cannam@167
|
1165 {
|
cannam@167
|
1166 E TjP, TjO, TjQ, TjL, TjN;
|
cannam@167
|
1167 TjP = FMA(KP923879532, TjI, TjF);
|
cannam@167
|
1168 TjO = W[23];
|
cannam@167
|
1169 TjQ = TjO * TjM;
|
cannam@167
|
1170 TjL = W[22];
|
cannam@167
|
1171 TjN = TjL * TjM;
|
cannam@167
|
1172 cr[WS(rs, 12)] = FNMS(TjO, TjP, TjN);
|
cannam@167
|
1173 ci[WS(rs, 12)] = FMA(TjL, TjP, TjQ);
|
cannam@167
|
1174 }
|
cannam@167
|
1175 }
|
cannam@167
|
1176 {
|
cannam@167
|
1177 E Tki, TkK, TkN, Tkn, Tkq, TkC, TkH, Tku;
|
cannam@167
|
1178 {
|
cannam@167
|
1179 E Tkm, TkF, Tka, TkA, Tko, Tkp, TkB, Tkh, TkG, Tk9, Tkd, Tkg;
|
cannam@167
|
1180 Tkm = Tje + Tjb;
|
cannam@167
|
1181 TkF = FMA(KP707106781, Tkm, Tkl);
|
cannam@167
|
1182 Tk9 = TjC + TjD;
|
cannam@167
|
1183 Tka = FNMS(KP707106781, Tk9, Tk8);
|
cannam@167
|
1184 TkA = FMA(KP707106781, Tk9, Tk8);
|
cannam@167
|
1185 Tko = FNMS(KP414213562, Tke, Tkf);
|
cannam@167
|
1186 Tkp = FMA(KP414213562, Tkb, Tkc);
|
cannam@167
|
1187 TkB = Tko + Tkp;
|
cannam@167
|
1188 Tkd = FNMS(KP414213562, Tkc, Tkb);
|
cannam@167
|
1189 Tkg = FMA(KP414213562, Tkf, Tke);
|
cannam@167
|
1190 Tkh = Tkd - Tkg;
|
cannam@167
|
1191 TkG = Tkg + Tkd;
|
cannam@167
|
1192 Tki = FNMS(KP923879532, Tkh, Tka);
|
cannam@167
|
1193 TkK = FMA(KP923879532, TkB, TkA);
|
cannam@167
|
1194 TkN = FMA(KP923879532, TkG, TkF);
|
cannam@167
|
1195 Tkn = FNMS(KP707106781, Tkm, Tkl);
|
cannam@167
|
1196 Tkq = Tko - Tkp;
|
cannam@167
|
1197 TkC = FNMS(KP923879532, TkB, TkA);
|
cannam@167
|
1198 TkH = FNMS(KP923879532, TkG, TkF);
|
cannam@167
|
1199 Tku = FMA(KP923879532, Tkh, Tka);
|
cannam@167
|
1200 }
|
cannam@167
|
1201 {
|
cannam@167
|
1202 E TkD, TkI, Tkz, TkE;
|
cannam@167
|
1203 Tkz = W[70];
|
cannam@167
|
1204 TkD = Tkz * TkC;
|
cannam@167
|
1205 TkI = Tkz * TkH;
|
cannam@167
|
1206 TkE = W[71];
|
cannam@167
|
1207 cr[WS(rs, 36)] = FNMS(TkE, TkH, TkD);
|
cannam@167
|
1208 ci[WS(rs, 36)] = FMA(TkE, TkC, TkI);
|
cannam@167
|
1209 }
|
cannam@167
|
1210 {
|
cannam@167
|
1211 E TkL, TkO, TkJ, TkM;
|
cannam@167
|
1212 TkJ = W[6];
|
cannam@167
|
1213 TkL = TkJ * TkK;
|
cannam@167
|
1214 TkO = TkJ * TkN;
|
cannam@167
|
1215 TkM = W[7];
|
cannam@167
|
1216 cr[WS(rs, 4)] = FNMS(TkM, TkN, TkL);
|
cannam@167
|
1217 ci[WS(rs, 4)] = FMA(TkM, TkK, TkO);
|
cannam@167
|
1218 }
|
cannam@167
|
1219 {
|
cannam@167
|
1220 E Tkr, Tkk, Tks, Tk7, Tkj;
|
cannam@167
|
1221 Tkr = FNMS(KP923879532, Tkq, Tkn);
|
cannam@167
|
1222 Tkk = W[103];
|
cannam@167
|
1223 Tks = Tkk * Tki;
|
cannam@167
|
1224 Tk7 = W[102];
|
cannam@167
|
1225 Tkj = Tk7 * Tki;
|
cannam@167
|
1226 cr[WS(rs, 52)] = FNMS(Tkk, Tkr, Tkj);
|
cannam@167
|
1227 ci[WS(rs, 52)] = FMA(Tk7, Tkr, Tks);
|
cannam@167
|
1228 }
|
cannam@167
|
1229 {
|
cannam@167
|
1230 E Tkx, Tkw, Tky, Tkt, Tkv;
|
cannam@167
|
1231 Tkx = FMA(KP923879532, Tkq, Tkn);
|
cannam@167
|
1232 Tkw = W[39];
|
cannam@167
|
1233 Tky = Tkw * Tku;
|
cannam@167
|
1234 Tkt = W[38];
|
cannam@167
|
1235 Tkv = Tkt * Tku;
|
cannam@167
|
1236 cr[WS(rs, 20)] = FNMS(Tkw, Tkx, Tkv);
|
cannam@167
|
1237 ci[WS(rs, 20)] = FMA(Tkt, Tkx, Tky);
|
cannam@167
|
1238 }
|
cannam@167
|
1239 }
|
cannam@167
|
1240 {
|
cannam@167
|
1241 E T5q, T66, T69, T5J, T5M, T5Y, T63, T5Q;
|
cannam@167
|
1242 {
|
cannam@167
|
1243 E T5F, T5I, T61, T5K, T5L, T5X, T3C, T5W, T5p, T62;
|
cannam@167
|
1244 T5F = FNMS(KP923879532, T5E, T5B);
|
cannam@167
|
1245 T5I = T5G - T5H;
|
cannam@167
|
1246 T61 = FNMS(KP980785280, T5I, T5F);
|
cannam@167
|
1247 T5K = FMA(KP820678790, T5a, T5n);
|
cannam@167
|
1248 T5L = FNMS(KP820678790, T4h, T4u);
|
cannam@167
|
1249 T5X = T5K + T5L;
|
cannam@167
|
1250 {
|
cannam@167
|
1251 E T2I, T3B, T4v, T5o;
|
cannam@167
|
1252 T2I = FNMS(KP923879532, T2H, T2k);
|
cannam@167
|
1253 T3B = T39 - T3A;
|
cannam@167
|
1254 T3C = FMA(KP980785280, T3B, T2I);
|
cannam@167
|
1255 T5W = FNMS(KP980785280, T3B, T2I);
|
cannam@167
|
1256 T4v = FMA(KP820678790, T4u, T4h);
|
cannam@167
|
1257 T5o = FNMS(KP820678790, T5n, T5a);
|
cannam@167
|
1258 T5p = T4v - T5o;
|
cannam@167
|
1259 T62 = T5o + T4v;
|
cannam@167
|
1260 }
|
cannam@167
|
1261 T5q = FNMS(KP773010453, T5p, T3C);
|
cannam@167
|
1262 T66 = FMA(KP773010453, T5X, T5W);
|
cannam@167
|
1263 T69 = FMA(KP773010453, T62, T61);
|
cannam@167
|
1264 T5J = FMA(KP980785280, T5I, T5F);
|
cannam@167
|
1265 T5M = T5K - T5L;
|
cannam@167
|
1266 T5Y = FNMS(KP773010453, T5X, T5W);
|
cannam@167
|
1267 T63 = FNMS(KP773010453, T62, T61);
|
cannam@167
|
1268 T5Q = FMA(KP773010453, T5p, T3C);
|
cannam@167
|
1269 }
|
cannam@167
|
1270 {
|
cannam@167
|
1271 E T5Z, T64, T5V, T60;
|
cannam@167
|
1272 T5V = W[48];
|
cannam@167
|
1273 T5Z = T5V * T5Y;
|
cannam@167
|
1274 T64 = T5V * T63;
|
cannam@167
|
1275 T60 = W[49];
|
cannam@167
|
1276 cr[WS(rs, 25)] = FNMS(T60, T63, T5Z);
|
cannam@167
|
1277 ci[WS(rs, 25)] = FMA(T60, T5Y, T64);
|
cannam@167
|
1278 }
|
cannam@167
|
1279 {
|
cannam@167
|
1280 E T67, T6a, T65, T68;
|
cannam@167
|
1281 T65 = W[112];
|
cannam@167
|
1282 T67 = T65 * T66;
|
cannam@167
|
1283 T6a = T65 * T69;
|
cannam@167
|
1284 T68 = W[113];
|
cannam@167
|
1285 cr[WS(rs, 57)] = FNMS(T68, T69, T67);
|
cannam@167
|
1286 ci[WS(rs, 57)] = FMA(T68, T66, T6a);
|
cannam@167
|
1287 }
|
cannam@167
|
1288 {
|
cannam@167
|
1289 E T5N, T5s, T5O, T23, T5r;
|
cannam@167
|
1290 T5N = FNMS(KP773010453, T5M, T5J);
|
cannam@167
|
1291 T5s = W[81];
|
cannam@167
|
1292 T5O = T5s * T5q;
|
cannam@167
|
1293 T23 = W[80];
|
cannam@167
|
1294 T5r = T23 * T5q;
|
cannam@167
|
1295 cr[WS(rs, 41)] = FNMS(T5s, T5N, T5r);
|
cannam@167
|
1296 ci[WS(rs, 41)] = FMA(T23, T5N, T5O);
|
cannam@167
|
1297 }
|
cannam@167
|
1298 {
|
cannam@167
|
1299 E T5T, T5S, T5U, T5P, T5R;
|
cannam@167
|
1300 T5T = FMA(KP773010453, T5M, T5J);
|
cannam@167
|
1301 T5S = W[17];
|
cannam@167
|
1302 T5U = T5S * T5Q;
|
cannam@167
|
1303 T5P = W[16];
|
cannam@167
|
1304 T5R = T5P * T5Q;
|
cannam@167
|
1305 cr[WS(rs, 9)] = FNMS(T5S, T5T, T5R);
|
cannam@167
|
1306 ci[WS(rs, 9)] = FMA(T5P, T5T, T5U);
|
cannam@167
|
1307 }
|
cannam@167
|
1308 }
|
cannam@167
|
1309 {
|
cannam@167
|
1310 E Tge, TgG, TgK, Tgr, Tgu, TgC, TgF, Tgx;
|
cannam@167
|
1311 {
|
cannam@167
|
1312 E Tg5, Tgc, Tgd, Tgj, Tgm, Tgn, TfY, TgA, Tgq, TgB;
|
cannam@167
|
1313 Tg5 = FMA(KP668178637, Tg4, Tg1);
|
cannam@167
|
1314 Tgc = FNMS(KP668178637, Tgb, Tg8);
|
cannam@167
|
1315 Tgd = Tg5 - Tgc;
|
cannam@167
|
1316 Tgj = FNMS(KP707106781, Tgi, Tgh);
|
cannam@167
|
1317 Tgm = Tgk - Tgl;
|
cannam@167
|
1318 Tgn = FMA(KP923879532, Tgm, Tgj);
|
cannam@167
|
1319 {
|
cannam@167
|
1320 E TfQ, TfX, Tgo, Tgp;
|
cannam@167
|
1321 TfQ = FNMS(KP707106781, TfP, TfO);
|
cannam@167
|
1322 TfX = TfT - TfW;
|
cannam@167
|
1323 TfY = FMA(KP923879532, TfX, TfQ);
|
cannam@167
|
1324 TgA = FNMS(KP923879532, TfX, TfQ);
|
cannam@167
|
1325 Tgo = FMA(KP668178637, Tg8, Tgb);
|
cannam@167
|
1326 Tgp = FNMS(KP668178637, Tg1, Tg4);
|
cannam@167
|
1327 Tgq = Tgo - Tgp;
|
cannam@167
|
1328 TgB = Tgo + Tgp;
|
cannam@167
|
1329 }
|
cannam@167
|
1330 Tge = FNMS(KP831469612, Tgd, TfY);
|
cannam@167
|
1331 TgG = Tgc + Tg5;
|
cannam@167
|
1332 TgK = FMA(KP831469612, TgB, TgA);
|
cannam@167
|
1333 Tgr = FNMS(KP831469612, Tgq, Tgn);
|
cannam@167
|
1334 Tgu = FMA(KP831469612, Tgd, TfY);
|
cannam@167
|
1335 TgC = FNMS(KP831469612, TgB, TgA);
|
cannam@167
|
1336 TgF = FNMS(KP923879532, Tgm, Tgj);
|
cannam@167
|
1337 Tgx = FMA(KP831469612, Tgq, Tgn);
|
cannam@167
|
1338 }
|
cannam@167
|
1339 {
|
cannam@167
|
1340 E Tgf, Tgs, TfN, Tgg;
|
cannam@167
|
1341 TfN = W[82];
|
cannam@167
|
1342 Tgf = TfN * Tge;
|
cannam@167
|
1343 Tgs = TfN * Tgr;
|
cannam@167
|
1344 Tgg = W[83];
|
cannam@167
|
1345 cr[WS(rs, 42)] = FNMS(Tgg, Tgr, Tgf);
|
cannam@167
|
1346 ci[WS(rs, 42)] = FMA(Tgg, Tge, Tgs);
|
cannam@167
|
1347 }
|
cannam@167
|
1348 {
|
cannam@167
|
1349 E Tgv, Tgy, Tgt, Tgw;
|
cannam@167
|
1350 Tgt = W[18];
|
cannam@167
|
1351 Tgv = Tgt * Tgu;
|
cannam@167
|
1352 Tgy = Tgt * Tgx;
|
cannam@167
|
1353 Tgw = W[19];
|
cannam@167
|
1354 cr[WS(rs, 10)] = FNMS(Tgw, Tgx, Tgv);
|
cannam@167
|
1355 ci[WS(rs, 10)] = FMA(Tgw, Tgu, Tgy);
|
cannam@167
|
1356 }
|
cannam@167
|
1357 {
|
cannam@167
|
1358 E TgH, TgE, TgI, Tgz, TgD;
|
cannam@167
|
1359 TgH = FNMS(KP831469612, TgG, TgF);
|
cannam@167
|
1360 TgE = W[51];
|
cannam@167
|
1361 TgI = TgE * TgC;
|
cannam@167
|
1362 Tgz = W[50];
|
cannam@167
|
1363 TgD = Tgz * TgC;
|
cannam@167
|
1364 cr[WS(rs, 26)] = FNMS(TgE, TgH, TgD);
|
cannam@167
|
1365 ci[WS(rs, 26)] = FMA(Tgz, TgH, TgI);
|
cannam@167
|
1366 }
|
cannam@167
|
1367 {
|
cannam@167
|
1368 E TgN, TgM, TgO, TgJ, TgL;
|
cannam@167
|
1369 TgN = FMA(KP831469612, TgG, TgF);
|
cannam@167
|
1370 TgM = W[115];
|
cannam@167
|
1371 TgO = TgM * TgK;
|
cannam@167
|
1372 TgJ = W[114];
|
cannam@167
|
1373 TgL = TgJ * TgK;
|
cannam@167
|
1374 cr[WS(rs, 58)] = FNMS(TgM, TgN, TgL);
|
cannam@167
|
1375 ci[WS(rs, 58)] = FMA(TgJ, TgN, TgO);
|
cannam@167
|
1376 }
|
cannam@167
|
1377 }
|
cannam@167
|
1378 {
|
cannam@167
|
1379 E Th0, Ths, Thv, Th5, Th8, Thk, Thp, Thc;
|
cannam@167
|
1380 {
|
cannam@167
|
1381 E Th3, Th4, Thn, Th6, Th7, Thj, TgS, Thi, TgZ, Tho;
|
cannam@167
|
1382 Th3 = FMA(KP707106781, Tgi, Tgh);
|
cannam@167
|
1383 Th4 = TfW + TfT;
|
cannam@167
|
1384 Thn = FNMS(KP923879532, Th4, Th3);
|
cannam@167
|
1385 Th6 = FMA(KP198912367, TgT, TgU);
|
cannam@167
|
1386 Th7 = FNMS(KP198912367, TgW, TgX);
|
cannam@167
|
1387 Thj = Th7 - Th6;
|
cannam@167
|
1388 {
|
cannam@167
|
1389 E TgQ, TgR, TgV, TgY;
|
cannam@167
|
1390 TgQ = FMA(KP707106781, TfP, TfO);
|
cannam@167
|
1391 TgR = Tgk + Tgl;
|
cannam@167
|
1392 TgS = FMA(KP923879532, TgR, TgQ);
|
cannam@167
|
1393 Thi = FNMS(KP923879532, TgR, TgQ);
|
cannam@167
|
1394 TgV = FNMS(KP198912367, TgU, TgT);
|
cannam@167
|
1395 TgY = FMA(KP198912367, TgX, TgW);
|
cannam@167
|
1396 TgZ = TgV + TgY;
|
cannam@167
|
1397 Tho = TgV - TgY;
|
cannam@167
|
1398 }
|
cannam@167
|
1399 Th0 = FNMS(KP980785280, TgZ, TgS);
|
cannam@167
|
1400 Ths = FMA(KP980785280, Thj, Thi);
|
cannam@167
|
1401 Thv = FMA(KP980785280, Tho, Thn);
|
cannam@167
|
1402 Th5 = FMA(KP923879532, Th4, Th3);
|
cannam@167
|
1403 Th8 = Th6 + Th7;
|
cannam@167
|
1404 Thk = FNMS(KP980785280, Thj, Thi);
|
cannam@167
|
1405 Thp = FNMS(KP980785280, Tho, Thn);
|
cannam@167
|
1406 Thc = FMA(KP980785280, TgZ, TgS);
|
cannam@167
|
1407 }
|
cannam@167
|
1408 {
|
cannam@167
|
1409 E Thl, Thq, Thh, Thm;
|
cannam@167
|
1410 Thh = W[98];
|
cannam@167
|
1411 Thl = Thh * Thk;
|
cannam@167
|
1412 Thq = Thh * Thp;
|
cannam@167
|
1413 Thm = W[99];
|
cannam@167
|
1414 cr[WS(rs, 50)] = FNMS(Thm, Thp, Thl);
|
cannam@167
|
1415 ci[WS(rs, 50)] = FMA(Thm, Thk, Thq);
|
cannam@167
|
1416 }
|
cannam@167
|
1417 {
|
cannam@167
|
1418 E Tht, Thw, Thr, Thu;
|
cannam@167
|
1419 Thr = W[34];
|
cannam@167
|
1420 Tht = Thr * Ths;
|
cannam@167
|
1421 Thw = Thr * Thv;
|
cannam@167
|
1422 Thu = W[35];
|
cannam@167
|
1423 cr[WS(rs, 18)] = FNMS(Thu, Thv, Tht);
|
cannam@167
|
1424 ci[WS(rs, 18)] = FMA(Thu, Ths, Thw);
|
cannam@167
|
1425 }
|
cannam@167
|
1426 {
|
cannam@167
|
1427 E Th9, Th2, Tha, TgP, Th1;
|
cannam@167
|
1428 Th9 = FNMS(KP980785280, Th8, Th5);
|
cannam@167
|
1429 Th2 = W[67];
|
cannam@167
|
1430 Tha = Th2 * Th0;
|
cannam@167
|
1431 TgP = W[66];
|
cannam@167
|
1432 Th1 = TgP * Th0;
|
cannam@167
|
1433 cr[WS(rs, 34)] = FNMS(Th2, Th9, Th1);
|
cannam@167
|
1434 ci[WS(rs, 34)] = FMA(TgP, Th9, Tha);
|
cannam@167
|
1435 }
|
cannam@167
|
1436 {
|
cannam@167
|
1437 E Thf, The, Thg, Thb, Thd;
|
cannam@167
|
1438 Thf = FMA(KP980785280, Th8, Th5);
|
cannam@167
|
1439 The = W[3];
|
cannam@167
|
1440 Thg = The * Thc;
|
cannam@167
|
1441 Thb = W[2];
|
cannam@167
|
1442 Thd = Thb * Thc;
|
cannam@167
|
1443 cr[WS(rs, 2)] = FNMS(The, Thf, Thd);
|
cannam@167
|
1444 ci[WS(rs, 2)] = FMA(Thb, Thf, Thg);
|
cannam@167
|
1445 }
|
cannam@167
|
1446 }
|
cannam@167
|
1447 {
|
cannam@167
|
1448 E T6m, T6O, T6R, T6r, T6u, T6G, T6L, T6y;
|
cannam@167
|
1449 {
|
cannam@167
|
1450 E T6p, T6q, T6J, T6s, T6t, T6F, T6e, T6E, T6l, T6K;
|
cannam@167
|
1451 T6p = FMA(KP923879532, T5E, T5B);
|
cannam@167
|
1452 T6q = T3A + T39;
|
cannam@167
|
1453 T6J = FMA(KP980785280, T6q, T6p);
|
cannam@167
|
1454 T6s = FNMS(KP098491403, T6i, T6j);
|
cannam@167
|
1455 T6t = FMA(KP098491403, T6f, T6g);
|
cannam@167
|
1456 T6F = T6s + T6t;
|
cannam@167
|
1457 {
|
cannam@167
|
1458 E T6c, T6d, T6h, T6k;
|
cannam@167
|
1459 T6c = FMA(KP923879532, T2H, T2k);
|
cannam@167
|
1460 T6d = T5G + T5H;
|
cannam@167
|
1461 T6e = FNMS(KP980785280, T6d, T6c);
|
cannam@167
|
1462 T6E = FMA(KP980785280, T6d, T6c);
|
cannam@167
|
1463 T6h = FNMS(KP098491403, T6g, T6f);
|
cannam@167
|
1464 T6k = FMA(KP098491403, T6j, T6i);
|
cannam@167
|
1465 T6l = T6h - T6k;
|
cannam@167
|
1466 T6K = T6k + T6h;
|
cannam@167
|
1467 }
|
cannam@167
|
1468 T6m = FNMS(KP995184726, T6l, T6e);
|
cannam@167
|
1469 T6O = FMA(KP995184726, T6F, T6E);
|
cannam@167
|
1470 T6R = FMA(KP995184726, T6K, T6J);
|
cannam@167
|
1471 T6r = FNMS(KP980785280, T6q, T6p);
|
cannam@167
|
1472 T6u = T6s - T6t;
|
cannam@167
|
1473 T6G = FNMS(KP995184726, T6F, T6E);
|
cannam@167
|
1474 T6L = FNMS(KP995184726, T6K, T6J);
|
cannam@167
|
1475 T6y = FMA(KP995184726, T6l, T6e);
|
cannam@167
|
1476 }
|
cannam@167
|
1477 {
|
cannam@167
|
1478 E T6H, T6M, T6D, T6I;
|
cannam@167
|
1479 T6D = W[64];
|
cannam@167
|
1480 T6H = T6D * T6G;
|
cannam@167
|
1481 T6M = T6D * T6L;
|
cannam@167
|
1482 T6I = W[65];
|
cannam@167
|
1483 cr[WS(rs, 33)] = FNMS(T6I, T6L, T6H);
|
cannam@167
|
1484 ci[WS(rs, 33)] = FMA(T6I, T6G, T6M);
|
cannam@167
|
1485 }
|
cannam@167
|
1486 {
|
cannam@167
|
1487 E T6P, T6S, T6N, T6Q;
|
cannam@167
|
1488 T6N = W[0];
|
cannam@167
|
1489 T6P = T6N * T6O;
|
cannam@167
|
1490 T6S = T6N * T6R;
|
cannam@167
|
1491 T6Q = W[1];
|
cannam@167
|
1492 cr[WS(rs, 1)] = FNMS(T6Q, T6R, T6P);
|
cannam@167
|
1493 ci[WS(rs, 1)] = FMA(T6Q, T6O, T6S);
|
cannam@167
|
1494 }
|
cannam@167
|
1495 {
|
cannam@167
|
1496 E T6v, T6o, T6w, T6b, T6n;
|
cannam@167
|
1497 T6v = FNMS(KP995184726, T6u, T6r);
|
cannam@167
|
1498 T6o = W[97];
|
cannam@167
|
1499 T6w = T6o * T6m;
|
cannam@167
|
1500 T6b = W[96];
|
cannam@167
|
1501 T6n = T6b * T6m;
|
cannam@167
|
1502 cr[WS(rs, 49)] = FNMS(T6o, T6v, T6n);
|
cannam@167
|
1503 ci[WS(rs, 49)] = FMA(T6b, T6v, T6w);
|
cannam@167
|
1504 }
|
cannam@167
|
1505 {
|
cannam@167
|
1506 E T6B, T6A, T6C, T6x, T6z;
|
cannam@167
|
1507 T6B = FMA(KP995184726, T6u, T6r);
|
cannam@167
|
1508 T6A = W[33];
|
cannam@167
|
1509 T6C = T6A * T6y;
|
cannam@167
|
1510 T6x = W[32];
|
cannam@167
|
1511 T6z = T6x * T6y;
|
cannam@167
|
1512 cr[WS(rs, 17)] = FNMS(T6A, T6B, T6z);
|
cannam@167
|
1513 ci[WS(rs, 17)] = FMA(T6x, T6B, T6C);
|
cannam@167
|
1514 }
|
cannam@167
|
1515 }
|
cannam@167
|
1516 {
|
cannam@167
|
1517 E Tbw, Tc2, Tc5, TbF, TbI, TbU, TbZ, TbM;
|
cannam@167
|
1518 {
|
cannam@167
|
1519 E TbB, TbE, TbX, TbG, TbH, TbT, Tbg, TbS, Tbv, TbY;
|
cannam@167
|
1520 TbB = FMA(KP923879532, TbA, Tbz);
|
cannam@167
|
1521 TbE = TbC - TbD;
|
cannam@167
|
1522 TbX = FNMS(KP980785280, TbE, TbB);
|
cannam@167
|
1523 TbG = FMA(KP820678790, Tbj, Tbm);
|
cannam@167
|
1524 TbH = FMA(KP820678790, Tbq, Tbt);
|
cannam@167
|
1525 TbT = TbG + TbH;
|
cannam@167
|
1526 {
|
cannam@167
|
1527 E Tb8, Tbf, Tbn, Tbu;
|
cannam@167
|
1528 Tb8 = FNMS(KP923879532, Tb7, Tb6);
|
cannam@167
|
1529 Tbf = Tbb + Tbe;
|
cannam@167
|
1530 Tbg = FNMS(KP980785280, Tbf, Tb8);
|
cannam@167
|
1531 TbS = FMA(KP980785280, Tbf, Tb8);
|
cannam@167
|
1532 Tbn = FNMS(KP820678790, Tbm, Tbj);
|
cannam@167
|
1533 Tbu = FNMS(KP820678790, Tbt, Tbq);
|
cannam@167
|
1534 Tbv = Tbn + Tbu;
|
cannam@167
|
1535 TbY = Tbn - Tbu;
|
cannam@167
|
1536 }
|
cannam@167
|
1537 Tbw = FNMS(KP773010453, Tbv, Tbg);
|
cannam@167
|
1538 Tc2 = FMA(KP773010453, TbT, TbS);
|
cannam@167
|
1539 Tc5 = FNMS(KP773010453, TbY, TbX);
|
cannam@167
|
1540 TbF = FMA(KP980785280, TbE, TbB);
|
cannam@167
|
1541 TbI = TbG - TbH;
|
cannam@167
|
1542 TbU = FNMS(KP773010453, TbT, TbS);
|
cannam@167
|
1543 TbZ = FMA(KP773010453, TbY, TbX);
|
cannam@167
|
1544 TbM = FMA(KP773010453, Tbv, Tbg);
|
cannam@167
|
1545 }
|
cannam@167
|
1546 {
|
cannam@167
|
1547 E TbV, Tc0, TbR, TbW;
|
cannam@167
|
1548 TbR = W[44];
|
cannam@167
|
1549 TbV = TbR * TbU;
|
cannam@167
|
1550 Tc0 = TbR * TbZ;
|
cannam@167
|
1551 TbW = W[45];
|
cannam@167
|
1552 cr[WS(rs, 23)] = FNMS(TbW, TbZ, TbV);
|
cannam@167
|
1553 ci[WS(rs, 23)] = FMA(TbW, TbU, Tc0);
|
cannam@167
|
1554 }
|
cannam@167
|
1555 {
|
cannam@167
|
1556 E Tc3, Tc6, Tc1, Tc4;
|
cannam@167
|
1557 Tc1 = W[108];
|
cannam@167
|
1558 Tc3 = Tc1 * Tc2;
|
cannam@167
|
1559 Tc6 = Tc1 * Tc5;
|
cannam@167
|
1560 Tc4 = W[109];
|
cannam@167
|
1561 cr[WS(rs, 55)] = FNMS(Tc4, Tc5, Tc3);
|
cannam@167
|
1562 ci[WS(rs, 55)] = FMA(Tc4, Tc2, Tc6);
|
cannam@167
|
1563 }
|
cannam@167
|
1564 {
|
cannam@167
|
1565 E TbJ, Tby, TbK, Tb5, Tbx;
|
cannam@167
|
1566 TbJ = FNMS(KP773010453, TbI, TbF);
|
cannam@167
|
1567 Tby = W[77];
|
cannam@167
|
1568 TbK = Tby * Tbw;
|
cannam@167
|
1569 Tb5 = W[76];
|
cannam@167
|
1570 Tbx = Tb5 * Tbw;
|
cannam@167
|
1571 cr[WS(rs, 39)] = FNMS(Tby, TbJ, Tbx);
|
cannam@167
|
1572 ci[WS(rs, 39)] = FMA(Tb5, TbJ, TbK);
|
cannam@167
|
1573 }
|
cannam@167
|
1574 {
|
cannam@167
|
1575 E TbP, TbO, TbQ, TbL, TbN;
|
cannam@167
|
1576 TbP = FMA(KP773010453, TbI, TbF);
|
cannam@167
|
1577 TbO = W[13];
|
cannam@167
|
1578 TbQ = TbO * TbM;
|
cannam@167
|
1579 TbL = W[12];
|
cannam@167
|
1580 TbN = TbL * TbM;
|
cannam@167
|
1581 cr[WS(rs, 7)] = FNMS(TbO, TbP, TbN);
|
cannam@167
|
1582 ci[WS(rs, 7)] = FMA(TbL, TbP, TbQ);
|
cannam@167
|
1583 }
|
cannam@167
|
1584 }
|
cannam@167
|
1585 {
|
cannam@167
|
1586 E Tay, Tb0, Tb3, TaD, TaG, TaS, TaX, TaK;
|
cannam@167
|
1587 {
|
cannam@167
|
1588 E TaB, TaC, TaV, TaE, TaF, TaR, Taq, TaQ, Tax, TaW;
|
cannam@167
|
1589 TaB = FMA(KP923879532, T9Q, T9N);
|
cannam@167
|
1590 TaC = T8V - T92;
|
cannam@167
|
1591 TaV = FNMS(KP831469612, TaC, TaB);
|
cannam@167
|
1592 TaE = FMA(KP303346683, Tar, Tas);
|
cannam@167
|
1593 TaF = FMA(KP303346683, Tau, Tav);
|
cannam@167
|
1594 TaR = TaE + TaF;
|
cannam@167
|
1595 {
|
cannam@167
|
1596 E Tao, Tap, Tat, Taw;
|
cannam@167
|
1597 Tao = FNMS(KP923879532, T8N, T8G);
|
cannam@167
|
1598 Tap = T9S + T9T;
|
cannam@167
|
1599 Taq = FMA(KP831469612, Tap, Tao);
|
cannam@167
|
1600 TaQ = FNMS(KP831469612, Tap, Tao);
|
cannam@167
|
1601 Tat = FNMS(KP303346683, Tas, Tar);
|
cannam@167
|
1602 Taw = FNMS(KP303346683, Tav, Tau);
|
cannam@167
|
1603 Tax = Tat + Taw;
|
cannam@167
|
1604 TaW = Tat - Taw;
|
cannam@167
|
1605 }
|
cannam@167
|
1606 Tay = FNMS(KP956940335, Tax, Taq);
|
cannam@167
|
1607 Tb0 = FMA(KP956940335, TaR, TaQ);
|
cannam@167
|
1608 Tb3 = FNMS(KP956940335, TaW, TaV);
|
cannam@167
|
1609 TaD = FMA(KP831469612, TaC, TaB);
|
cannam@167
|
1610 TaG = TaE - TaF;
|
cannam@167
|
1611 TaS = FNMS(KP956940335, TaR, TaQ);
|
cannam@167
|
1612 TaX = FMA(KP956940335, TaW, TaV);
|
cannam@167
|
1613 TaK = FMA(KP956940335, Tax, Taq);
|
cannam@167
|
1614 }
|
cannam@167
|
1615 {
|
cannam@167
|
1616 E TaT, TaY, TaP, TaU;
|
cannam@167
|
1617 TaP = W[36];
|
cannam@167
|
1618 TaT = TaP * TaS;
|
cannam@167
|
1619 TaY = TaP * TaX;
|
cannam@167
|
1620 TaU = W[37];
|
cannam@167
|
1621 cr[WS(rs, 19)] = FNMS(TaU, TaX, TaT);
|
cannam@167
|
1622 ci[WS(rs, 19)] = FMA(TaU, TaS, TaY);
|
cannam@167
|
1623 }
|
cannam@167
|
1624 {
|
cannam@167
|
1625 E Tb1, Tb4, TaZ, Tb2;
|
cannam@167
|
1626 TaZ = W[100];
|
cannam@167
|
1627 Tb1 = TaZ * Tb0;
|
cannam@167
|
1628 Tb4 = TaZ * Tb3;
|
cannam@167
|
1629 Tb2 = W[101];
|
cannam@167
|
1630 cr[WS(rs, 51)] = FNMS(Tb2, Tb3, Tb1);
|
cannam@167
|
1631 ci[WS(rs, 51)] = FMA(Tb2, Tb0, Tb4);
|
cannam@167
|
1632 }
|
cannam@167
|
1633 {
|
cannam@167
|
1634 E TaH, TaA, TaI, Tan, Taz;
|
cannam@167
|
1635 TaH = FNMS(KP956940335, TaG, TaD);
|
cannam@167
|
1636 TaA = W[69];
|
cannam@167
|
1637 TaI = TaA * Tay;
|
cannam@167
|
1638 Tan = W[68];
|
cannam@167
|
1639 Taz = Tan * Tay;
|
cannam@167
|
1640 cr[WS(rs, 35)] = FNMS(TaA, TaH, Taz);
|
cannam@167
|
1641 ci[WS(rs, 35)] = FMA(Tan, TaH, TaI);
|
cannam@167
|
1642 }
|
cannam@167
|
1643 {
|
cannam@167
|
1644 E TaN, TaM, TaO, TaJ, TaL;
|
cannam@167
|
1645 TaN = FMA(KP956940335, TaG, TaD);
|
cannam@167
|
1646 TaM = W[5];
|
cannam@167
|
1647 TaO = TaM * TaK;
|
cannam@167
|
1648 TaJ = W[4];
|
cannam@167
|
1649 TaL = TaJ * TaK;
|
cannam@167
|
1650 cr[WS(rs, 3)] = FNMS(TaM, TaN, TaL);
|
cannam@167
|
1651 ci[WS(rs, 3)] = FMA(TaJ, TaN, TaO);
|
cannam@167
|
1652 }
|
cannam@167
|
1653 }
|
cannam@167
|
1654 {
|
cannam@167
|
1655 E Tfg, TfI, TfL, Tfl, Tfo, TfA, TfF, Tfs;
|
cannam@167
|
1656 {
|
cannam@167
|
1657 E Tfj, Tfk, TfD, Tfm, Tfn, Tfz, Tf8, Tfy, Tff, TfE;
|
cannam@167
|
1658 Tfj = FNMS(KP707106781, Tey, Tev);
|
cannam@167
|
1659 Tfk = Tds + Tdh;
|
cannam@167
|
1660 TfD = FMA(KP923879532, Tfk, Tfj);
|
cannam@167
|
1661 Tfm = FMA(KP198912367, Tfc, Tfd);
|
cannam@167
|
1662 Tfn = FNMS(KP198912367, Tf9, Tfa);
|
cannam@167
|
1663 Tfz = Tfm + Tfn;
|
cannam@167
|
1664 {
|
cannam@167
|
1665 E Tf6, Tf7, Tfb, Tfe;
|
cannam@167
|
1666 Tf6 = FNMS(KP707106781, Td5, TcU);
|
cannam@167
|
1667 Tf7 = TeA + TeB;
|
cannam@167
|
1668 Tf8 = FNMS(KP923879532, Tf7, Tf6);
|
cannam@167
|
1669 Tfy = FMA(KP923879532, Tf7, Tf6);
|
cannam@167
|
1670 Tfb = FMA(KP198912367, Tfa, Tf9);
|
cannam@167
|
1671 Tfe = FNMS(KP198912367, Tfd, Tfc);
|
cannam@167
|
1672 Tff = Tfb - Tfe;
|
cannam@167
|
1673 TfE = Tfe + Tfb;
|
cannam@167
|
1674 }
|
cannam@167
|
1675 Tfg = FNMS(KP980785280, Tff, Tf8);
|
cannam@167
|
1676 TfI = FMA(KP980785280, Tfz, Tfy);
|
cannam@167
|
1677 TfL = FMA(KP980785280, TfE, TfD);
|
cannam@167
|
1678 Tfl = FNMS(KP923879532, Tfk, Tfj);
|
cannam@167
|
1679 Tfo = Tfm - Tfn;
|
cannam@167
|
1680 TfA = FNMS(KP980785280, Tfz, Tfy);
|
cannam@167
|
1681 TfF = FNMS(KP980785280, TfE, TfD);
|
cannam@167
|
1682 Tfs = FMA(KP980785280, Tff, Tf8);
|
cannam@167
|
1683 }
|
cannam@167
|
1684 {
|
cannam@167
|
1685 E TfB, TfG, Tfx, TfC;
|
cannam@167
|
1686 Tfx = W[58];
|
cannam@167
|
1687 TfB = Tfx * TfA;
|
cannam@167
|
1688 TfG = Tfx * TfF;
|
cannam@167
|
1689 TfC = W[59];
|
cannam@167
|
1690 cr[WS(rs, 30)] = FNMS(TfC, TfF, TfB);
|
cannam@167
|
1691 ci[WS(rs, 30)] = FMA(TfC, TfA, TfG);
|
cannam@167
|
1692 }
|
cannam@167
|
1693 {
|
cannam@167
|
1694 E TfJ, TfM, TfH, TfK;
|
cannam@167
|
1695 TfH = W[122];
|
cannam@167
|
1696 TfJ = TfH * TfI;
|
cannam@167
|
1697 TfM = TfH * TfL;
|
cannam@167
|
1698 TfK = W[123];
|
cannam@167
|
1699 cr[WS(rs, 62)] = FNMS(TfK, TfL, TfJ);
|
cannam@167
|
1700 ci[WS(rs, 62)] = FMA(TfK, TfI, TfM);
|
cannam@167
|
1701 }
|
cannam@167
|
1702 {
|
cannam@167
|
1703 E Tfp, Tfi, Tfq, Tf5, Tfh;
|
cannam@167
|
1704 Tfp = FNMS(KP980785280, Tfo, Tfl);
|
cannam@167
|
1705 Tfi = W[91];
|
cannam@167
|
1706 Tfq = Tfi * Tfg;
|
cannam@167
|
1707 Tf5 = W[90];
|
cannam@167
|
1708 Tfh = Tf5 * Tfg;
|
cannam@167
|
1709 cr[WS(rs, 46)] = FNMS(Tfi, Tfp, Tfh);
|
cannam@167
|
1710 ci[WS(rs, 46)] = FMA(Tf5, Tfp, Tfq);
|
cannam@167
|
1711 }
|
cannam@167
|
1712 {
|
cannam@167
|
1713 E Tfv, Tfu, Tfw, Tfr, Tft;
|
cannam@167
|
1714 Tfv = FMA(KP980785280, Tfo, Tfl);
|
cannam@167
|
1715 Tfu = W[27];
|
cannam@167
|
1716 Tfw = Tfu * Tfs;
|
cannam@167
|
1717 Tfr = W[26];
|
cannam@167
|
1718 Tft = Tfr * Tfs;
|
cannam@167
|
1719 cr[WS(rs, 14)] = FNMS(Tfu, Tfv, Tft);
|
cannam@167
|
1720 ci[WS(rs, 14)] = FMA(Tfr, Tfv, Tfw);
|
cannam@167
|
1721 }
|
cannam@167
|
1722 }
|
cannam@167
|
1723 {
|
cannam@167
|
1724 E T7k, T7Q, T7T, T7t, T7w, T7I, T7N, T7A;
|
cannam@167
|
1725 {
|
cannam@167
|
1726 E T7p, T7s, T7L, T7u, T7v, T7H, T74, T7G, T7j, T7M;
|
cannam@167
|
1727 T7p = FMA(KP923879532, T7o, T7n);
|
cannam@167
|
1728 T7s = T7q - T7r;
|
cannam@167
|
1729 T7L = FNMS(KP831469612, T7s, T7p);
|
cannam@167
|
1730 T7u = FMA(KP534511135, T77, T7a);
|
cannam@167
|
1731 T7v = FNMS(KP534511135, T7e, T7h);
|
cannam@167
|
1732 T7H = T7v - T7u;
|
cannam@167
|
1733 {
|
cannam@167
|
1734 E T6W, T73, T7b, T7i;
|
cannam@167
|
1735 T6W = FMA(KP923879532, T6V, T6U);
|
cannam@167
|
1736 T73 = T6Z - T72;
|
cannam@167
|
1737 T74 = FMA(KP831469612, T73, T6W);
|
cannam@167
|
1738 T7G = FNMS(KP831469612, T73, T6W);
|
cannam@167
|
1739 T7b = FNMS(KP534511135, T7a, T77);
|
cannam@167
|
1740 T7i = FMA(KP534511135, T7h, T7e);
|
cannam@167
|
1741 T7j = T7b + T7i;
|
cannam@167
|
1742 T7M = T7b - T7i;
|
cannam@167
|
1743 }
|
cannam@167
|
1744 T7k = FNMS(KP881921264, T7j, T74);
|
cannam@167
|
1745 T7Q = FMA(KP881921264, T7H, T7G);
|
cannam@167
|
1746 T7T = FMA(KP881921264, T7M, T7L);
|
cannam@167
|
1747 T7t = FMA(KP831469612, T7s, T7p);
|
cannam@167
|
1748 T7w = T7u + T7v;
|
cannam@167
|
1749 T7I = FNMS(KP881921264, T7H, T7G);
|
cannam@167
|
1750 T7N = FNMS(KP881921264, T7M, T7L);
|
cannam@167
|
1751 T7A = FMA(KP881921264, T7j, T74);
|
cannam@167
|
1752 }
|
cannam@167
|
1753 {
|
cannam@167
|
1754 E T7J, T7O, T7F, T7K;
|
cannam@167
|
1755 T7F = W[104];
|
cannam@167
|
1756 T7J = T7F * T7I;
|
cannam@167
|
1757 T7O = T7F * T7N;
|
cannam@167
|
1758 T7K = W[105];
|
cannam@167
|
1759 cr[WS(rs, 53)] = FNMS(T7K, T7N, T7J);
|
cannam@167
|
1760 ci[WS(rs, 53)] = FMA(T7K, T7I, T7O);
|
cannam@167
|
1761 }
|
cannam@167
|
1762 {
|
cannam@167
|
1763 E T7R, T7U, T7P, T7S;
|
cannam@167
|
1764 T7P = W[40];
|
cannam@167
|
1765 T7R = T7P * T7Q;
|
cannam@167
|
1766 T7U = T7P * T7T;
|
cannam@167
|
1767 T7S = W[41];
|
cannam@167
|
1768 cr[WS(rs, 21)] = FNMS(T7S, T7T, T7R);
|
cannam@167
|
1769 ci[WS(rs, 21)] = FMA(T7S, T7Q, T7U);
|
cannam@167
|
1770 }
|
cannam@167
|
1771 {
|
cannam@167
|
1772 E T7x, T7m, T7y, T6T, T7l;
|
cannam@167
|
1773 T7x = FNMS(KP881921264, T7w, T7t);
|
cannam@167
|
1774 T7m = W[73];
|
cannam@167
|
1775 T7y = T7m * T7k;
|
cannam@167
|
1776 T6T = W[72];
|
cannam@167
|
1777 T7l = T6T * T7k;
|
cannam@167
|
1778 cr[WS(rs, 37)] = FNMS(T7m, T7x, T7l);
|
cannam@167
|
1779 ci[WS(rs, 37)] = FMA(T6T, T7x, T7y);
|
cannam@167
|
1780 }
|
cannam@167
|
1781 {
|
cannam@167
|
1782 E T7D, T7C, T7E, T7z, T7B;
|
cannam@167
|
1783 T7D = FMA(KP881921264, T7w, T7t);
|
cannam@167
|
1784 T7C = W[9];
|
cannam@167
|
1785 T7E = T7C * T7A;
|
cannam@167
|
1786 T7z = W[8];
|
cannam@167
|
1787 T7B = T7z * T7A;
|
cannam@167
|
1788 cr[WS(rs, 5)] = FNMS(T7C, T7D, T7B);
|
cannam@167
|
1789 ci[WS(rs, 5)] = FMA(T7z, T7D, T7E);
|
cannam@167
|
1790 }
|
cannam@167
|
1791 }
|
cannam@167
|
1792 {
|
cannam@167
|
1793 E T86, T8u, T8y, T8f, T8i, T8q, T8t, T8l;
|
cannam@167
|
1794 {
|
cannam@167
|
1795 E T81, T84, T85, T89, T8a, T8b, T7Y, T8o, T8e, T8p;
|
cannam@167
|
1796 T81 = FMA(KP303346683, T80, T7Z);
|
cannam@167
|
1797 T84 = FNMS(KP303346683, T83, T82);
|
cannam@167
|
1798 T85 = T81 - T84;
|
cannam@167
|
1799 T89 = FNMS(KP923879532, T7o, T7n);
|
cannam@167
|
1800 T8a = T72 + T6Z;
|
cannam@167
|
1801 T8b = FNMS(KP831469612, T8a, T89);
|
cannam@167
|
1802 {
|
cannam@167
|
1803 E T7W, T7X, T8c, T8d;
|
cannam@167
|
1804 T7W = FNMS(KP923879532, T6V, T6U);
|
cannam@167
|
1805 T7X = T7q + T7r;
|
cannam@167
|
1806 T7Y = FNMS(KP831469612, T7X, T7W);
|
cannam@167
|
1807 T8o = FMA(KP831469612, T7X, T7W);
|
cannam@167
|
1808 T8c = FMA(KP303346683, T82, T83);
|
cannam@167
|
1809 T8d = FNMS(KP303346683, T7Z, T80);
|
cannam@167
|
1810 T8e = T8c - T8d;
|
cannam@167
|
1811 T8p = T8c + T8d;
|
cannam@167
|
1812 }
|
cannam@167
|
1813 T86 = FNMS(KP956940335, T85, T7Y);
|
cannam@167
|
1814 T8u = T84 + T81;
|
cannam@167
|
1815 T8y = FMA(KP956940335, T8p, T8o);
|
cannam@167
|
1816 T8f = FNMS(KP956940335, T8e, T8b);
|
cannam@167
|
1817 T8i = FMA(KP956940335, T85, T7Y);
|
cannam@167
|
1818 T8q = FNMS(KP956940335, T8p, T8o);
|
cannam@167
|
1819 T8t = FMA(KP831469612, T8a, T89);
|
cannam@167
|
1820 T8l = FMA(KP956940335, T8e, T8b);
|
cannam@167
|
1821 }
|
cannam@167
|
1822 {
|
cannam@167
|
1823 E T87, T8g, T7V, T88;
|
cannam@167
|
1824 T7V = W[88];
|
cannam@167
|
1825 T87 = T7V * T86;
|
cannam@167
|
1826 T8g = T7V * T8f;
|
cannam@167
|
1827 T88 = W[89];
|
cannam@167
|
1828 cr[WS(rs, 45)] = FNMS(T88, T8f, T87);
|
cannam@167
|
1829 ci[WS(rs, 45)] = FMA(T88, T86, T8g);
|
cannam@167
|
1830 }
|
cannam@167
|
1831 {
|
cannam@167
|
1832 E T8j, T8m, T8h, T8k;
|
cannam@167
|
1833 T8h = W[24];
|
cannam@167
|
1834 T8j = T8h * T8i;
|
cannam@167
|
1835 T8m = T8h * T8l;
|
cannam@167
|
1836 T8k = W[25];
|
cannam@167
|
1837 cr[WS(rs, 13)] = FNMS(T8k, T8l, T8j);
|
cannam@167
|
1838 ci[WS(rs, 13)] = FMA(T8k, T8i, T8m);
|
cannam@167
|
1839 }
|
cannam@167
|
1840 {
|
cannam@167
|
1841 E T8v, T8s, T8w, T8n, T8r;
|
cannam@167
|
1842 T8v = FNMS(KP956940335, T8u, T8t);
|
cannam@167
|
1843 T8s = W[57];
|
cannam@167
|
1844 T8w = T8s * T8q;
|
cannam@167
|
1845 T8n = W[56];
|
cannam@167
|
1846 T8r = T8n * T8q;
|
cannam@167
|
1847 cr[WS(rs, 29)] = FNMS(T8s, T8v, T8r);
|
cannam@167
|
1848 ci[WS(rs, 29)] = FMA(T8n, T8v, T8w);
|
cannam@167
|
1849 }
|
cannam@167
|
1850 {
|
cannam@167
|
1851 E T8B, T8A, T8C, T8x, T8z;
|
cannam@167
|
1852 T8B = FMA(KP956940335, T8u, T8t);
|
cannam@167
|
1853 T8A = W[121];
|
cannam@167
|
1854 T8C = T8A * T8y;
|
cannam@167
|
1855 T8x = W[120];
|
cannam@167
|
1856 T8z = T8x * T8y;
|
cannam@167
|
1857 cr[WS(rs, 61)] = FNMS(T8A, T8B, T8z);
|
cannam@167
|
1858 ci[WS(rs, 61)] = FMA(T8x, T8B, T8C);
|
cannam@167
|
1859 }
|
cannam@167
|
1860 }
|
cannam@167
|
1861 {
|
cannam@167
|
1862 E T9I, Tai, Tal, T9V, T9Y, Taa, Taf, Ta2;
|
cannam@167
|
1863 {
|
cannam@167
|
1864 E T9R, T9U, Tad, T9W, T9X, Ta9, T94, Ta8, T9H, Tae;
|
cannam@167
|
1865 T9R = FNMS(KP923879532, T9Q, T9N);
|
cannam@167
|
1866 T9U = T9S - T9T;
|
cannam@167
|
1867 Tad = FNMS(KP831469612, T9U, T9R);
|
cannam@167
|
1868 T9W = FMA(KP534511135, T9f, T9m);
|
cannam@167
|
1869 T9X = FMA(KP534511135, T9y, T9F);
|
cannam@167
|
1870 Ta9 = T9W + T9X;
|
cannam@167
|
1871 {
|
cannam@167
|
1872 E T8O, T93, T9n, T9G;
|
cannam@167
|
1873 T8O = FMA(KP923879532, T8N, T8G);
|
cannam@167
|
1874 T93 = T8V + T92;
|
cannam@167
|
1875 T94 = FNMS(KP831469612, T93, T8O);
|
cannam@167
|
1876 Ta8 = FMA(KP831469612, T93, T8O);
|
cannam@167
|
1877 T9n = FNMS(KP534511135, T9m, T9f);
|
cannam@167
|
1878 T9G = FNMS(KP534511135, T9F, T9y);
|
cannam@167
|
1879 T9H = T9n + T9G;
|
cannam@167
|
1880 Tae = T9G - T9n;
|
cannam@167
|
1881 }
|
cannam@167
|
1882 T9I = FMA(KP881921264, T9H, T94);
|
cannam@167
|
1883 Tai = FMA(KP881921264, Ta9, Ta8);
|
cannam@167
|
1884 Tal = FNMS(KP881921264, Tae, Tad);
|
cannam@167
|
1885 T9V = FMA(KP831469612, T9U, T9R);
|
cannam@167
|
1886 T9Y = T9W - T9X;
|
cannam@167
|
1887 Taa = FNMS(KP881921264, Ta9, Ta8);
|
cannam@167
|
1888 Taf = FMA(KP881921264, Tae, Tad);
|
cannam@167
|
1889 Ta2 = FNMS(KP881921264, T9H, T94);
|
cannam@167
|
1890 }
|
cannam@167
|
1891 {
|
cannam@167
|
1892 E Tab, Tag, Ta7, Tac;
|
cannam@167
|
1893 Ta7 = W[52];
|
cannam@167
|
1894 Tab = Ta7 * Taa;
|
cannam@167
|
1895 Tag = Ta7 * Taf;
|
cannam@167
|
1896 Tac = W[53];
|
cannam@167
|
1897 cr[WS(rs, 27)] = FNMS(Tac, Taf, Tab);
|
cannam@167
|
1898 ci[WS(rs, 27)] = FMA(Tac, Taa, Tag);
|
cannam@167
|
1899 }
|
cannam@167
|
1900 {
|
cannam@167
|
1901 E Taj, Tam, Tah, Tak;
|
cannam@167
|
1902 Tah = W[116];
|
cannam@167
|
1903 Taj = Tah * Tai;
|
cannam@167
|
1904 Tam = Tah * Tal;
|
cannam@167
|
1905 Tak = W[117];
|
cannam@167
|
1906 cr[WS(rs, 59)] = FNMS(Tak, Tal, Taj);
|
cannam@167
|
1907 ci[WS(rs, 59)] = FMA(Tak, Tai, Tam);
|
cannam@167
|
1908 }
|
cannam@167
|
1909 {
|
cannam@167
|
1910 E T9Z, T9K, Ta0, T8D, T9J;
|
cannam@167
|
1911 T9Z = FNMS(KP881921264, T9Y, T9V);
|
cannam@167
|
1912 T9K = W[85];
|
cannam@167
|
1913 Ta0 = T9K * T9I;
|
cannam@167
|
1914 T8D = W[84];
|
cannam@167
|
1915 T9J = T8D * T9I;
|
cannam@167
|
1916 cr[WS(rs, 43)] = FNMS(T9K, T9Z, T9J);
|
cannam@167
|
1917 ci[WS(rs, 43)] = FMA(T8D, T9Z, Ta0);
|
cannam@167
|
1918 }
|
cannam@167
|
1919 {
|
cannam@167
|
1920 E Ta5, Ta4, Ta6, Ta1, Ta3;
|
cannam@167
|
1921 Ta5 = FMA(KP881921264, T9Y, T9V);
|
cannam@167
|
1922 Ta4 = W[21];
|
cannam@167
|
1923 Ta6 = Ta4 * Ta2;
|
cannam@167
|
1924 Ta1 = W[20];
|
cannam@167
|
1925 Ta3 = Ta1 * Ta2;
|
cannam@167
|
1926 cr[WS(rs, 11)] = FNMS(Ta4, Ta5, Ta3);
|
cannam@167
|
1927 ci[WS(rs, 11)] = FMA(Ta1, Ta5, Ta6);
|
cannam@167
|
1928 }
|
cannam@167
|
1929 }
|
cannam@167
|
1930 {
|
cannam@167
|
1931 E Teo, Tf0, Tf3, TeD, TeG, TeS, TeX, TeK;
|
cannam@167
|
1932 {
|
cannam@167
|
1933 E Tez, TeC, TeV, TeE, TeF, TeR, Tdu, TeQ, Ten, TeW;
|
cannam@167
|
1934 Tez = FMA(KP707106781, Tey, Tev);
|
cannam@167
|
1935 TeC = TeA - TeB;
|
cannam@167
|
1936 TeV = FMA(KP923879532, TeC, Tez);
|
cannam@167
|
1937 TeE = FNMS(KP668178637, Tec, Tel);
|
cannam@167
|
1938 TeF = FMA(KP668178637, TdL, TdU);
|
cannam@167
|
1939 TeR = TeE + TeF;
|
cannam@167
|
1940 {
|
cannam@167
|
1941 E Td6, Tdt, TdV, Tem;
|
cannam@167
|
1942 Td6 = FMA(KP707106781, Td5, TcU);
|
cannam@167
|
1943 Tdt = Tdh - Tds;
|
cannam@167
|
1944 Tdu = FNMS(KP923879532, Tdt, Td6);
|
cannam@167
|
1945 TeQ = FMA(KP923879532, Tdt, Td6);
|
cannam@167
|
1946 TdV = FNMS(KP668178637, TdU, TdL);
|
cannam@167
|
1947 Tem = FMA(KP668178637, Tel, Tec);
|
cannam@167
|
1948 Ten = TdV - Tem;
|
cannam@167
|
1949 TeW = Tem + TdV;
|
cannam@167
|
1950 }
|
cannam@167
|
1951 Teo = FNMS(KP831469612, Ten, Tdu);
|
cannam@167
|
1952 Tf0 = FMA(KP831469612, TeR, TeQ);
|
cannam@167
|
1953 Tf3 = FMA(KP831469612, TeW, TeV);
|
cannam@167
|
1954 TeD = FNMS(KP923879532, TeC, Tez);
|
cannam@167
|
1955 TeG = TeE - TeF;
|
cannam@167
|
1956 TeS = FNMS(KP831469612, TeR, TeQ);
|
cannam@167
|
1957 TeX = FNMS(KP831469612, TeW, TeV);
|
cannam@167
|
1958 TeK = FMA(KP831469612, Ten, Tdu);
|
cannam@167
|
1959 }
|
cannam@167
|
1960 {
|
cannam@167
|
1961 E TeT, TeY, TeP, TeU;
|
cannam@167
|
1962 TeP = W[74];
|
cannam@167
|
1963 TeT = TeP * TeS;
|
cannam@167
|
1964 TeY = TeP * TeX;
|
cannam@167
|
1965 TeU = W[75];
|
cannam@167
|
1966 cr[WS(rs, 38)] = FNMS(TeU, TeX, TeT);
|
cannam@167
|
1967 ci[WS(rs, 38)] = FMA(TeU, TeS, TeY);
|
cannam@167
|
1968 }
|
cannam@167
|
1969 {
|
cannam@167
|
1970 E Tf1, Tf4, TeZ, Tf2;
|
cannam@167
|
1971 TeZ = W[10];
|
cannam@167
|
1972 Tf1 = TeZ * Tf0;
|
cannam@167
|
1973 Tf4 = TeZ * Tf3;
|
cannam@167
|
1974 Tf2 = W[11];
|
cannam@167
|
1975 cr[WS(rs, 6)] = FNMS(Tf2, Tf3, Tf1);
|
cannam@167
|
1976 ci[WS(rs, 6)] = FMA(Tf2, Tf0, Tf4);
|
cannam@167
|
1977 }
|
cannam@167
|
1978 {
|
cannam@167
|
1979 E TeH, Teq, TeI, TcP, Tep;
|
cannam@167
|
1980 TeH = FNMS(KP831469612, TeG, TeD);
|
cannam@167
|
1981 Teq = W[107];
|
cannam@167
|
1982 TeI = Teq * Teo;
|
cannam@167
|
1983 TcP = W[106];
|
cannam@167
|
1984 Tep = TcP * Teo;
|
cannam@167
|
1985 cr[WS(rs, 54)] = FNMS(Teq, TeH, Tep);
|
cannam@167
|
1986 ci[WS(rs, 54)] = FMA(TcP, TeH, TeI);
|
cannam@167
|
1987 }
|
cannam@167
|
1988 {
|
cannam@167
|
1989 E TeN, TeM, TeO, TeJ, TeL;
|
cannam@167
|
1990 TeN = FMA(KP831469612, TeG, TeD);
|
cannam@167
|
1991 TeM = W[43];
|
cannam@167
|
1992 TeO = TeM * TeK;
|
cannam@167
|
1993 TeJ = W[42];
|
cannam@167
|
1994 TeL = TeJ * TeK;
|
cannam@167
|
1995 cr[WS(rs, 22)] = FNMS(TeM, TeN, TeL);
|
cannam@167
|
1996 ci[WS(rs, 22)] = FMA(TeJ, TeN, TeO);
|
cannam@167
|
1997 }
|
cannam@167
|
1998 }
|
cannam@167
|
1999 {
|
cannam@167
|
2000 E Tci, TcK, TcN, Tcn, Tcq, TcC, TcH, Tcu;
|
cannam@167
|
2001 {
|
cannam@167
|
2002 E Tcl, Tcm, TcF, Tco, Tcp, TcB, Tca, TcA, Tch, TcG;
|
cannam@167
|
2003 Tcl = FNMS(KP923879532, TbA, Tbz);
|
cannam@167
|
2004 Tcm = Tbe - Tbb;
|
cannam@167
|
2005 TcF = FNMS(KP980785280, Tcm, Tcl);
|
cannam@167
|
2006 Tco = FMA(KP098491403, Tcb, Tcc);
|
cannam@167
|
2007 Tcp = FMA(KP098491403, Tce, Tcf);
|
cannam@167
|
2008 TcB = Tco + Tcp;
|
cannam@167
|
2009 {
|
cannam@167
|
2010 E Tc8, Tc9, Tcd, Tcg;
|
cannam@167
|
2011 Tc8 = FMA(KP923879532, Tb7, Tb6);
|
cannam@167
|
2012 Tc9 = TbC + TbD;
|
cannam@167
|
2013 Tca = FNMS(KP980785280, Tc9, Tc8);
|
cannam@167
|
2014 TcA = FMA(KP980785280, Tc9, Tc8);
|
cannam@167
|
2015 Tcd = FNMS(KP098491403, Tcc, Tcb);
|
cannam@167
|
2016 Tcg = FNMS(KP098491403, Tcf, Tce);
|
cannam@167
|
2017 Tch = Tcd + Tcg;
|
cannam@167
|
2018 TcG = Tcg - Tcd;
|
cannam@167
|
2019 }
|
cannam@167
|
2020 Tci = FMA(KP995184726, Tch, Tca);
|
cannam@167
|
2021 TcK = FMA(KP995184726, TcB, TcA);
|
cannam@167
|
2022 TcN = FNMS(KP995184726, TcG, TcF);
|
cannam@167
|
2023 Tcn = FMA(KP980785280, Tcm, Tcl);
|
cannam@167
|
2024 Tcq = Tco - Tcp;
|
cannam@167
|
2025 TcC = FNMS(KP995184726, TcB, TcA);
|
cannam@167
|
2026 TcH = FMA(KP995184726, TcG, TcF);
|
cannam@167
|
2027 Tcu = FNMS(KP995184726, Tch, Tca);
|
cannam@167
|
2028 }
|
cannam@167
|
2029 {
|
cannam@167
|
2030 E TcD, TcI, Tcz, TcE;
|
cannam@167
|
2031 Tcz = W[60];
|
cannam@167
|
2032 TcD = Tcz * TcC;
|
cannam@167
|
2033 TcI = Tcz * TcH;
|
cannam@167
|
2034 TcE = W[61];
|
cannam@167
|
2035 cr[WS(rs, 31)] = FNMS(TcE, TcH, TcD);
|
cannam@167
|
2036 ci[WS(rs, 31)] = FMA(TcE, TcC, TcI);
|
cannam@167
|
2037 }
|
cannam@167
|
2038 {
|
cannam@167
|
2039 E TcL, TcO, TcJ, TcM;
|
cannam@167
|
2040 TcJ = W[124];
|
cannam@167
|
2041 TcL = TcJ * TcK;
|
cannam@167
|
2042 TcO = TcJ * TcN;
|
cannam@167
|
2043 TcM = W[125];
|
cannam@167
|
2044 cr[WS(rs, 63)] = FNMS(TcM, TcN, TcL);
|
cannam@167
|
2045 ci[WS(rs, 63)] = FMA(TcM, TcK, TcO);
|
cannam@167
|
2046 }
|
cannam@167
|
2047 {
|
cannam@167
|
2048 E Tcr, Tck, Tcs, Tc7, Tcj;
|
cannam@167
|
2049 Tcr = FNMS(KP995184726, Tcq, Tcn);
|
cannam@167
|
2050 Tck = W[93];
|
cannam@167
|
2051 Tcs = Tck * Tci;
|
cannam@167
|
2052 Tc7 = W[92];
|
cannam@167
|
2053 Tcj = Tc7 * Tci;
|
cannam@167
|
2054 cr[WS(rs, 47)] = FNMS(Tck, Tcr, Tcj);
|
cannam@167
|
2055 ci[WS(rs, 47)] = FMA(Tc7, Tcr, Tcs);
|
cannam@167
|
2056 }
|
cannam@167
|
2057 {
|
cannam@167
|
2058 E Tcx, Tcw, Tcy, Tct, Tcv;
|
cannam@167
|
2059 Tcx = FMA(KP995184726, Tcq, Tcn);
|
cannam@167
|
2060 Tcw = W[29];
|
cannam@167
|
2061 Tcy = Tcw * Tcu;
|
cannam@167
|
2062 Tct = W[28];
|
cannam@167
|
2063 Tcv = Tct * Tcu;
|
cannam@167
|
2064 cr[WS(rs, 15)] = FNMS(Tcw, Tcx, Tcv);
|
cannam@167
|
2065 ci[WS(rs, 15)] = FMA(Tct, Tcx, Tcy);
|
cannam@167
|
2066 }
|
cannam@167
|
2067 }
|
cannam@167
|
2068 }
|
cannam@167
|
2069 }
|
cannam@167
|
2070 }
|
cannam@167
|
2071
|
cannam@167
|
2072 static const tw_instr twinstr[] = {
|
cannam@167
|
2073 {TW_FULL, 1, 64},
|
cannam@167
|
2074 {TW_NEXT, 1, 0}
|
cannam@167
|
2075 };
|
cannam@167
|
2076
|
cannam@167
|
2077 static const hc2hc_desc desc = { 64, "hb_64", twinstr, &GENUS, {520, 126, 518, 0} };
|
cannam@167
|
2078
|
cannam@167
|
2079 void X(codelet_hb_64) (planner *p) {
|
cannam@167
|
2080 X(khc2hc_register) (p, hb_64, &desc);
|
cannam@167
|
2081 }
|
cannam@167
|
2082 #else
|
cannam@167
|
2083
|
cannam@167
|
2084 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -dif -name hb_64 -include rdft/scalar/hb.h */
|
cannam@167
|
2085
|
cannam@167
|
2086 /*
|
cannam@167
|
2087 * This function contains 1038 FP additions, 500 FP multiplications,
|
cannam@167
|
2088 * (or, 808 additions, 270 multiplications, 230 fused multiply/add),
|
cannam@167
|
2089 * 196 stack variables, 15 constants, and 256 memory accesses
|
cannam@167
|
2090 */
|
cannam@167
|
2091 #include "rdft/scalar/hb.h"
|
cannam@167
|
2092
|
cannam@167
|
2093 static void hb_64(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
2094 {
|
cannam@167
|
2095 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
cannam@167
|
2096 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
2097 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
2098 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
cannam@167
|
2099 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
cannam@167
|
2100 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
2101 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
2102 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
cannam@167
|
2103 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
2104 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
2105 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
2106 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
2107 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
2108 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
2109 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
2110 {
|
cannam@167
|
2111 INT m;
|
cannam@167
|
2112 for (m = mb, W = W + ((mb - 1) * 126); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 126, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
2113 E Tf, T8C, Tfa, Thk, Tgg, ThM, T2c, T5O, T4K, T6g, Tag, TdE, TcA, Te6, T7P;
|
cannam@167
|
2114 E T94, TK, T7o, T38, T4P, Tfv, Thn, T5W, T6j, Tb0, TdK, Tfs, Tho, T8K, T97;
|
cannam@167
|
2115 E Tb7, TdL, TZ, T7l, T2P, T4Q, Tfo, Thq, T5T, T6k, TaH, TdH, Tfl, Thr, T8H;
|
cannam@167
|
2116 E T98, TaO, TdI, Tu, T95, Tfh, ThN, Tgj, Thl, T2v, T6h, T4N, T5P, Tav, Te7;
|
cannam@167
|
2117 E TcD, TdF, T7S, T8D, T1L, T20, T7A, T7D, T7G, T7H, T40, T62, Tg1, Thv, Tg8;
|
cannam@167
|
2118 E Thz, Tg5, Thw, T4t, T5Z, T4j, T60, T4w, T63, TbY, TdS, Tcd, TdQ, TfU, Thy;
|
cannam@167
|
2119 E T8P, T9z, T8S, T9A, Tcl, TdP, Tco, TdT, T1g, T1v, T7r, T7u, T7x, T7y, T3j;
|
cannam@167
|
2120 E T69, TfI, ThD, TfP, ThG, TfM, ThC, T3M, T66, T3C, T67, T3P, T6a, Tbl, TdZ;
|
cannam@167
|
2121 E TbA, TdX, TfB, ThF, T8W, T9C, T8Z, T9D, TbI, TdW, TbL, Te0;
|
cannam@167
|
2122 {
|
cannam@167
|
2123 E T3, Ta6, T6, Tcu, T4I, Ta7, T4F, Tcv, Td, Tcy, T27, Tae, Ta, Tcx, T2a;
|
cannam@167
|
2124 E Tab;
|
cannam@167
|
2125 {
|
cannam@167
|
2126 E T1, T2, T4D, T4E;
|
cannam@167
|
2127 T1 = cr[0];
|
cannam@167
|
2128 T2 = ci[WS(rs, 31)];
|
cannam@167
|
2129 T3 = T1 + T2;
|
cannam@167
|
2130 Ta6 = T1 - T2;
|
cannam@167
|
2131 {
|
cannam@167
|
2132 E T4, T5, T4G, T4H;
|
cannam@167
|
2133 T4 = cr[WS(rs, 16)];
|
cannam@167
|
2134 T5 = ci[WS(rs, 15)];
|
cannam@167
|
2135 T6 = T4 + T5;
|
cannam@167
|
2136 Tcu = T4 - T5;
|
cannam@167
|
2137 T4G = ci[WS(rs, 47)];
|
cannam@167
|
2138 T4H = cr[WS(rs, 48)];
|
cannam@167
|
2139 T4I = T4G - T4H;
|
cannam@167
|
2140 Ta7 = T4G + T4H;
|
cannam@167
|
2141 }
|
cannam@167
|
2142 T4D = ci[WS(rs, 63)];
|
cannam@167
|
2143 T4E = cr[WS(rs, 32)];
|
cannam@167
|
2144 T4F = T4D - T4E;
|
cannam@167
|
2145 Tcv = T4D + T4E;
|
cannam@167
|
2146 {
|
cannam@167
|
2147 E Tb, Tc, Tac, T25, T26, Tad;
|
cannam@167
|
2148 Tb = ci[WS(rs, 7)];
|
cannam@167
|
2149 Tc = cr[WS(rs, 24)];
|
cannam@167
|
2150 Tac = Tb - Tc;
|
cannam@167
|
2151 T25 = ci[WS(rs, 39)];
|
cannam@167
|
2152 T26 = cr[WS(rs, 56)];
|
cannam@167
|
2153 Tad = T25 + T26;
|
cannam@167
|
2154 Td = Tb + Tc;
|
cannam@167
|
2155 Tcy = Tac + Tad;
|
cannam@167
|
2156 T27 = T25 - T26;
|
cannam@167
|
2157 Tae = Tac - Tad;
|
cannam@167
|
2158 }
|
cannam@167
|
2159 {
|
cannam@167
|
2160 E T8, T9, Ta9, T28, T29, Taa;
|
cannam@167
|
2161 T8 = cr[WS(rs, 8)];
|
cannam@167
|
2162 T9 = ci[WS(rs, 23)];
|
cannam@167
|
2163 Ta9 = T8 - T9;
|
cannam@167
|
2164 T28 = ci[WS(rs, 55)];
|
cannam@167
|
2165 T29 = cr[WS(rs, 40)];
|
cannam@167
|
2166 Taa = T28 + T29;
|
cannam@167
|
2167 Ta = T8 + T9;
|
cannam@167
|
2168 Tcx = Ta9 + Taa;
|
cannam@167
|
2169 T2a = T28 - T29;
|
cannam@167
|
2170 Tab = Ta9 - Taa;
|
cannam@167
|
2171 }
|
cannam@167
|
2172 }
|
cannam@167
|
2173 {
|
cannam@167
|
2174 E T7, Te, Tf8, Tf9;
|
cannam@167
|
2175 T7 = T3 + T6;
|
cannam@167
|
2176 Te = Ta + Td;
|
cannam@167
|
2177 Tf = T7 + Te;
|
cannam@167
|
2178 T8C = T7 - Te;
|
cannam@167
|
2179 Tf8 = Ta6 + Ta7;
|
cannam@167
|
2180 Tf9 = KP707106781 * (Tcx + Tcy);
|
cannam@167
|
2181 Tfa = Tf8 - Tf9;
|
cannam@167
|
2182 Thk = Tf8 + Tf9;
|
cannam@167
|
2183 }
|
cannam@167
|
2184 {
|
cannam@167
|
2185 E Tge, Tgf, T24, T2b;
|
cannam@167
|
2186 Tge = Tcv - Tcu;
|
cannam@167
|
2187 Tgf = KP707106781 * (Tab - Tae);
|
cannam@167
|
2188 Tgg = Tge + Tgf;
|
cannam@167
|
2189 ThM = Tge - Tgf;
|
cannam@167
|
2190 T24 = T3 - T6;
|
cannam@167
|
2191 T2b = T27 - T2a;
|
cannam@167
|
2192 T2c = T24 + T2b;
|
cannam@167
|
2193 T5O = T24 - T2b;
|
cannam@167
|
2194 }
|
cannam@167
|
2195 {
|
cannam@167
|
2196 E T4C, T4J, Ta8, Taf;
|
cannam@167
|
2197 T4C = Ta - Td;
|
cannam@167
|
2198 T4J = T4F - T4I;
|
cannam@167
|
2199 T4K = T4C + T4J;
|
cannam@167
|
2200 T6g = T4J - T4C;
|
cannam@167
|
2201 Ta8 = Ta6 - Ta7;
|
cannam@167
|
2202 Taf = KP707106781 * (Tab + Tae);
|
cannam@167
|
2203 Tag = Ta8 - Taf;
|
cannam@167
|
2204 TdE = Ta8 + Taf;
|
cannam@167
|
2205 }
|
cannam@167
|
2206 {
|
cannam@167
|
2207 E Tcw, Tcz, T7N, T7O;
|
cannam@167
|
2208 Tcw = Tcu + Tcv;
|
cannam@167
|
2209 Tcz = KP707106781 * (Tcx - Tcy);
|
cannam@167
|
2210 TcA = Tcw - Tcz;
|
cannam@167
|
2211 Te6 = Tcw + Tcz;
|
cannam@167
|
2212 T7N = T4F + T4I;
|
cannam@167
|
2213 T7O = T2a + T27;
|
cannam@167
|
2214 T7P = T7N + T7O;
|
cannam@167
|
2215 T94 = T7N - T7O;
|
cannam@167
|
2216 }
|
cannam@167
|
2217 }
|
cannam@167
|
2218 {
|
cannam@167
|
2219 E TC, Tb1, T2Z, TaQ, T2X, Tb2, T7m, TaR, TJ, Tb4, Tb5, T2Q, T36, TaV, TaY;
|
cannam@167
|
2220 E T7n, Tfq, Tfr;
|
cannam@167
|
2221 {
|
cannam@167
|
2222 E Tw, Tx, Ty, Tz, TA, TB;
|
cannam@167
|
2223 Tw = cr[WS(rs, 2)];
|
cannam@167
|
2224 Tx = ci[WS(rs, 29)];
|
cannam@167
|
2225 Ty = Tw + Tx;
|
cannam@167
|
2226 Tz = cr[WS(rs, 18)];
|
cannam@167
|
2227 TA = ci[WS(rs, 13)];
|
cannam@167
|
2228 TB = Tz + TA;
|
cannam@167
|
2229 TC = Ty + TB;
|
cannam@167
|
2230 Tb1 = Tz - TA;
|
cannam@167
|
2231 T2Z = Ty - TB;
|
cannam@167
|
2232 TaQ = Tw - Tx;
|
cannam@167
|
2233 }
|
cannam@167
|
2234 {
|
cannam@167
|
2235 E T2R, T2S, T2T, T2U, T2V, T2W;
|
cannam@167
|
2236 T2R = ci[WS(rs, 61)];
|
cannam@167
|
2237 T2S = cr[WS(rs, 34)];
|
cannam@167
|
2238 T2T = T2R - T2S;
|
cannam@167
|
2239 T2U = ci[WS(rs, 45)];
|
cannam@167
|
2240 T2V = cr[WS(rs, 50)];
|
cannam@167
|
2241 T2W = T2U - T2V;
|
cannam@167
|
2242 T2X = T2T - T2W;
|
cannam@167
|
2243 Tb2 = T2R + T2S;
|
cannam@167
|
2244 T7m = T2T + T2W;
|
cannam@167
|
2245 TaR = T2U + T2V;
|
cannam@167
|
2246 }
|
cannam@167
|
2247 {
|
cannam@167
|
2248 E TF, TaT, T35, TaU, TI, TaW, T32, TaX;
|
cannam@167
|
2249 {
|
cannam@167
|
2250 E TD, TE, T33, T34;
|
cannam@167
|
2251 TD = cr[WS(rs, 10)];
|
cannam@167
|
2252 TE = ci[WS(rs, 21)];
|
cannam@167
|
2253 TF = TD + TE;
|
cannam@167
|
2254 TaT = TD - TE;
|
cannam@167
|
2255 T33 = ci[WS(rs, 53)];
|
cannam@167
|
2256 T34 = cr[WS(rs, 42)];
|
cannam@167
|
2257 T35 = T33 - T34;
|
cannam@167
|
2258 TaU = T33 + T34;
|
cannam@167
|
2259 }
|
cannam@167
|
2260 {
|
cannam@167
|
2261 E TG, TH, T30, T31;
|
cannam@167
|
2262 TG = ci[WS(rs, 5)];
|
cannam@167
|
2263 TH = cr[WS(rs, 26)];
|
cannam@167
|
2264 TI = TG + TH;
|
cannam@167
|
2265 TaW = TG - TH;
|
cannam@167
|
2266 T30 = ci[WS(rs, 37)];
|
cannam@167
|
2267 T31 = cr[WS(rs, 58)];
|
cannam@167
|
2268 T32 = T30 - T31;
|
cannam@167
|
2269 TaX = T30 + T31;
|
cannam@167
|
2270 }
|
cannam@167
|
2271 TJ = TF + TI;
|
cannam@167
|
2272 Tb4 = TaT + TaU;
|
cannam@167
|
2273 Tb5 = TaW + TaX;
|
cannam@167
|
2274 T2Q = TF - TI;
|
cannam@167
|
2275 T36 = T32 - T35;
|
cannam@167
|
2276 TaV = TaT - TaU;
|
cannam@167
|
2277 TaY = TaW - TaX;
|
cannam@167
|
2278 T7n = T35 + T32;
|
cannam@167
|
2279 }
|
cannam@167
|
2280 TK = TC + TJ;
|
cannam@167
|
2281 T7o = T7m + T7n;
|
cannam@167
|
2282 {
|
cannam@167
|
2283 E T2Y, T37, Tft, Tfu;
|
cannam@167
|
2284 T2Y = T2Q + T2X;
|
cannam@167
|
2285 T37 = T2Z + T36;
|
cannam@167
|
2286 T38 = FMA(KP923879532, T2Y, KP382683432 * T37);
|
cannam@167
|
2287 T4P = FNMS(KP382683432, T2Y, KP923879532 * T37);
|
cannam@167
|
2288 Tft = TaQ + TaR;
|
cannam@167
|
2289 Tfu = KP707106781 * (Tb4 + Tb5);
|
cannam@167
|
2290 Tfv = Tft - Tfu;
|
cannam@167
|
2291 Thn = Tft + Tfu;
|
cannam@167
|
2292 }
|
cannam@167
|
2293 {
|
cannam@167
|
2294 E T5U, T5V, TaS, TaZ;
|
cannam@167
|
2295 T5U = T2X - T2Q;
|
cannam@167
|
2296 T5V = T2Z - T36;
|
cannam@167
|
2297 T5W = FMA(KP382683432, T5U, KP923879532 * T5V);
|
cannam@167
|
2298 T6j = FNMS(KP923879532, T5U, KP382683432 * T5V);
|
cannam@167
|
2299 TaS = TaQ - TaR;
|
cannam@167
|
2300 TaZ = KP707106781 * (TaV + TaY);
|
cannam@167
|
2301 Tb0 = TaS - TaZ;
|
cannam@167
|
2302 TdK = TaS + TaZ;
|
cannam@167
|
2303 }
|
cannam@167
|
2304 Tfq = Tb2 - Tb1;
|
cannam@167
|
2305 Tfr = KP707106781 * (TaV - TaY);
|
cannam@167
|
2306 Tfs = Tfq + Tfr;
|
cannam@167
|
2307 Tho = Tfq - Tfr;
|
cannam@167
|
2308 {
|
cannam@167
|
2309 E T8I, T8J, Tb3, Tb6;
|
cannam@167
|
2310 T8I = TC - TJ;
|
cannam@167
|
2311 T8J = T7m - T7n;
|
cannam@167
|
2312 T8K = T8I + T8J;
|
cannam@167
|
2313 T97 = T8I - T8J;
|
cannam@167
|
2314 Tb3 = Tb1 + Tb2;
|
cannam@167
|
2315 Tb6 = KP707106781 * (Tb4 - Tb5);
|
cannam@167
|
2316 Tb7 = Tb3 - Tb6;
|
cannam@167
|
2317 TdL = Tb3 + Tb6;
|
cannam@167
|
2318 }
|
cannam@167
|
2319 }
|
cannam@167
|
2320 {
|
cannam@167
|
2321 E TR, TaI, T2G, Tax, T2E, TaJ, T7j, Tay, TY, TaL, TaM, T2x, T2N, TaC, TaF;
|
cannam@167
|
2322 E T7k, Tfj, Tfk;
|
cannam@167
|
2323 {
|
cannam@167
|
2324 E TL, TM, TN, TO, TP, TQ;
|
cannam@167
|
2325 TL = ci[WS(rs, 1)];
|
cannam@167
|
2326 TM = cr[WS(rs, 30)];
|
cannam@167
|
2327 TN = TL + TM;
|
cannam@167
|
2328 TO = cr[WS(rs, 14)];
|
cannam@167
|
2329 TP = ci[WS(rs, 17)];
|
cannam@167
|
2330 TQ = TO + TP;
|
cannam@167
|
2331 TR = TN + TQ;
|
cannam@167
|
2332 TaI = TL - TM;
|
cannam@167
|
2333 T2G = TN - TQ;
|
cannam@167
|
2334 Tax = TO - TP;
|
cannam@167
|
2335 }
|
cannam@167
|
2336 {
|
cannam@167
|
2337 E T2y, T2z, T2A, T2B, T2C, T2D;
|
cannam@167
|
2338 T2y = ci[WS(rs, 33)];
|
cannam@167
|
2339 T2z = cr[WS(rs, 62)];
|
cannam@167
|
2340 T2A = T2y - T2z;
|
cannam@167
|
2341 T2B = ci[WS(rs, 49)];
|
cannam@167
|
2342 T2C = cr[WS(rs, 46)];
|
cannam@167
|
2343 T2D = T2B - T2C;
|
cannam@167
|
2344 T2E = T2A - T2D;
|
cannam@167
|
2345 TaJ = T2B + T2C;
|
cannam@167
|
2346 T7j = T2A + T2D;
|
cannam@167
|
2347 Tay = T2y + T2z;
|
cannam@167
|
2348 }
|
cannam@167
|
2349 {
|
cannam@167
|
2350 E TU, TaA, T2M, TaB, TX, TaD, T2J, TaE;
|
cannam@167
|
2351 {
|
cannam@167
|
2352 E TS, TT, T2K, T2L;
|
cannam@167
|
2353 TS = cr[WS(rs, 6)];
|
cannam@167
|
2354 TT = ci[WS(rs, 25)];
|
cannam@167
|
2355 TU = TS + TT;
|
cannam@167
|
2356 TaA = TS - TT;
|
cannam@167
|
2357 T2K = ci[WS(rs, 57)];
|
cannam@167
|
2358 T2L = cr[WS(rs, 38)];
|
cannam@167
|
2359 T2M = T2K - T2L;
|
cannam@167
|
2360 TaB = T2K + T2L;
|
cannam@167
|
2361 }
|
cannam@167
|
2362 {
|
cannam@167
|
2363 E TV, TW, T2H, T2I;
|
cannam@167
|
2364 TV = ci[WS(rs, 9)];
|
cannam@167
|
2365 TW = cr[WS(rs, 22)];
|
cannam@167
|
2366 TX = TV + TW;
|
cannam@167
|
2367 TaD = TV - TW;
|
cannam@167
|
2368 T2H = ci[WS(rs, 41)];
|
cannam@167
|
2369 T2I = cr[WS(rs, 54)];
|
cannam@167
|
2370 T2J = T2H - T2I;
|
cannam@167
|
2371 TaE = T2H + T2I;
|
cannam@167
|
2372 }
|
cannam@167
|
2373 TY = TU + TX;
|
cannam@167
|
2374 TaL = TaA - TaB;
|
cannam@167
|
2375 TaM = TaD - TaE;
|
cannam@167
|
2376 T2x = TU - TX;
|
cannam@167
|
2377 T2N = T2J - T2M;
|
cannam@167
|
2378 TaC = TaA + TaB;
|
cannam@167
|
2379 TaF = TaD + TaE;
|
cannam@167
|
2380 T7k = T2M + T2J;
|
cannam@167
|
2381 }
|
cannam@167
|
2382 TZ = TR + TY;
|
cannam@167
|
2383 T7l = T7j + T7k;
|
cannam@167
|
2384 {
|
cannam@167
|
2385 E T2F, T2O, Tfm, Tfn;
|
cannam@167
|
2386 T2F = T2x + T2E;
|
cannam@167
|
2387 T2O = T2G + T2N;
|
cannam@167
|
2388 T2P = FNMS(KP382683432, T2O, KP923879532 * T2F);
|
cannam@167
|
2389 T4Q = FMA(KP382683432, T2F, KP923879532 * T2O);
|
cannam@167
|
2390 Tfm = TaI + TaJ;
|
cannam@167
|
2391 Tfn = KP707106781 * (TaC + TaF);
|
cannam@167
|
2392 Tfo = Tfm - Tfn;
|
cannam@167
|
2393 Thq = Tfm + Tfn;
|
cannam@167
|
2394 }
|
cannam@167
|
2395 {
|
cannam@167
|
2396 E T5R, T5S, Taz, TaG;
|
cannam@167
|
2397 T5R = T2E - T2x;
|
cannam@167
|
2398 T5S = T2G - T2N;
|
cannam@167
|
2399 T5T = FNMS(KP923879532, T5S, KP382683432 * T5R);
|
cannam@167
|
2400 T6k = FMA(KP923879532, T5R, KP382683432 * T5S);
|
cannam@167
|
2401 Taz = Tax - Tay;
|
cannam@167
|
2402 TaG = KP707106781 * (TaC - TaF);
|
cannam@167
|
2403 TaH = Taz - TaG;
|
cannam@167
|
2404 TdH = Taz + TaG;
|
cannam@167
|
2405 }
|
cannam@167
|
2406 Tfj = KP707106781 * (TaL - TaM);
|
cannam@167
|
2407 Tfk = Tax + Tay;
|
cannam@167
|
2408 Tfl = Tfj - Tfk;
|
cannam@167
|
2409 Thr = Tfk + Tfj;
|
cannam@167
|
2410 {
|
cannam@167
|
2411 E T8F, T8G, TaK, TaN;
|
cannam@167
|
2412 T8F = T7j - T7k;
|
cannam@167
|
2413 T8G = TR - TY;
|
cannam@167
|
2414 T8H = T8F - T8G;
|
cannam@167
|
2415 T98 = T8G + T8F;
|
cannam@167
|
2416 TaK = TaI - TaJ;
|
cannam@167
|
2417 TaN = KP707106781 * (TaL + TaM);
|
cannam@167
|
2418 TaO = TaK - TaN;
|
cannam@167
|
2419 TdI = TaK + TaN;
|
cannam@167
|
2420 }
|
cannam@167
|
2421 }
|
cannam@167
|
2422 {
|
cannam@167
|
2423 E Ti, T2j, Tl, T2g, T2d, T2k, Tfc, Tfb, Tat, Taq, Tp, T2s, Ts, T2p, T2m;
|
cannam@167
|
2424 E T2t, Tff, Tfe, Tam, Taj;
|
cannam@167
|
2425 {
|
cannam@167
|
2426 E Tar, Tas, Tao, Tap;
|
cannam@167
|
2427 {
|
cannam@167
|
2428 E Tg, Th, T2h, T2i;
|
cannam@167
|
2429 Tg = cr[WS(rs, 4)];
|
cannam@167
|
2430 Th = ci[WS(rs, 27)];
|
cannam@167
|
2431 Ti = Tg + Th;
|
cannam@167
|
2432 Tar = Tg - Th;
|
cannam@167
|
2433 T2h = ci[WS(rs, 43)];
|
cannam@167
|
2434 T2i = cr[WS(rs, 52)];
|
cannam@167
|
2435 T2j = T2h - T2i;
|
cannam@167
|
2436 Tas = T2h + T2i;
|
cannam@167
|
2437 }
|
cannam@167
|
2438 {
|
cannam@167
|
2439 E Tj, Tk, T2e, T2f;
|
cannam@167
|
2440 Tj = cr[WS(rs, 20)];
|
cannam@167
|
2441 Tk = ci[WS(rs, 11)];
|
cannam@167
|
2442 Tl = Tj + Tk;
|
cannam@167
|
2443 Tao = Tj - Tk;
|
cannam@167
|
2444 T2e = ci[WS(rs, 59)];
|
cannam@167
|
2445 T2f = cr[WS(rs, 36)];
|
cannam@167
|
2446 T2g = T2e - T2f;
|
cannam@167
|
2447 Tap = T2e + T2f;
|
cannam@167
|
2448 }
|
cannam@167
|
2449 T2d = Ti - Tl;
|
cannam@167
|
2450 T2k = T2g - T2j;
|
cannam@167
|
2451 Tfc = Tap - Tao;
|
cannam@167
|
2452 Tfb = Tar + Tas;
|
cannam@167
|
2453 Tat = Tar - Tas;
|
cannam@167
|
2454 Taq = Tao + Tap;
|
cannam@167
|
2455 }
|
cannam@167
|
2456 {
|
cannam@167
|
2457 E Tak, Tal, Tah, Tai;
|
cannam@167
|
2458 {
|
cannam@167
|
2459 E Tn, To, T2q, T2r;
|
cannam@167
|
2460 Tn = ci[WS(rs, 3)];
|
cannam@167
|
2461 To = cr[WS(rs, 28)];
|
cannam@167
|
2462 Tp = Tn + To;
|
cannam@167
|
2463 Tak = Tn - To;
|
cannam@167
|
2464 T2q = ci[WS(rs, 51)];
|
cannam@167
|
2465 T2r = cr[WS(rs, 44)];
|
cannam@167
|
2466 T2s = T2q - T2r;
|
cannam@167
|
2467 Tal = T2q + T2r;
|
cannam@167
|
2468 }
|
cannam@167
|
2469 {
|
cannam@167
|
2470 E Tq, Tr, T2n, T2o;
|
cannam@167
|
2471 Tq = cr[WS(rs, 12)];
|
cannam@167
|
2472 Tr = ci[WS(rs, 19)];
|
cannam@167
|
2473 Ts = Tq + Tr;
|
cannam@167
|
2474 Tah = Tq - Tr;
|
cannam@167
|
2475 T2n = ci[WS(rs, 35)];
|
cannam@167
|
2476 T2o = cr[WS(rs, 60)];
|
cannam@167
|
2477 T2p = T2n - T2o;
|
cannam@167
|
2478 Tai = T2n + T2o;
|
cannam@167
|
2479 }
|
cannam@167
|
2480 T2m = Tp - Ts;
|
cannam@167
|
2481 T2t = T2p - T2s;
|
cannam@167
|
2482 Tff = Tah + Tai;
|
cannam@167
|
2483 Tfe = Tak + Tal;
|
cannam@167
|
2484 Tam = Tak - Tal;
|
cannam@167
|
2485 Taj = Tah - Tai;
|
cannam@167
|
2486 }
|
cannam@167
|
2487 {
|
cannam@167
|
2488 E Tm, Tt, Tfd, Tfg;
|
cannam@167
|
2489 Tm = Ti + Tl;
|
cannam@167
|
2490 Tt = Tp + Ts;
|
cannam@167
|
2491 Tu = Tm + Tt;
|
cannam@167
|
2492 T95 = Tm - Tt;
|
cannam@167
|
2493 Tfd = FNMS(KP923879532, Tfc, KP382683432 * Tfb);
|
cannam@167
|
2494 Tfg = FNMS(KP923879532, Tff, KP382683432 * Tfe);
|
cannam@167
|
2495 Tfh = Tfd + Tfg;
|
cannam@167
|
2496 ThN = Tfd - Tfg;
|
cannam@167
|
2497 }
|
cannam@167
|
2498 {
|
cannam@167
|
2499 E Tgh, Tgi, T2l, T2u;
|
cannam@167
|
2500 Tgh = FMA(KP382683432, Tfc, KP923879532 * Tfb);
|
cannam@167
|
2501 Tgi = FMA(KP382683432, Tff, KP923879532 * Tfe);
|
cannam@167
|
2502 Tgj = Tgh - Tgi;
|
cannam@167
|
2503 Thl = Tgh + Tgi;
|
cannam@167
|
2504 T2l = T2d - T2k;
|
cannam@167
|
2505 T2u = T2m + T2t;
|
cannam@167
|
2506 T2v = KP707106781 * (T2l + T2u);
|
cannam@167
|
2507 T6h = KP707106781 * (T2l - T2u);
|
cannam@167
|
2508 }
|
cannam@167
|
2509 {
|
cannam@167
|
2510 E T4L, T4M, Tan, Tau;
|
cannam@167
|
2511 T4L = T2d + T2k;
|
cannam@167
|
2512 T4M = T2t - T2m;
|
cannam@167
|
2513 T4N = KP707106781 * (T4L + T4M);
|
cannam@167
|
2514 T5P = KP707106781 * (T4M - T4L);
|
cannam@167
|
2515 Tan = FNMS(KP382683432, Tam, KP923879532 * Taj);
|
cannam@167
|
2516 Tau = FMA(KP923879532, Taq, KP382683432 * Tat);
|
cannam@167
|
2517 Tav = Tan - Tau;
|
cannam@167
|
2518 Te7 = Tau + Tan;
|
cannam@167
|
2519 }
|
cannam@167
|
2520 {
|
cannam@167
|
2521 E TcB, TcC, T7Q, T7R;
|
cannam@167
|
2522 TcB = FNMS(KP382683432, Taq, KP923879532 * Tat);
|
cannam@167
|
2523 TcC = FMA(KP382683432, Taj, KP923879532 * Tam);
|
cannam@167
|
2524 TcD = TcB - TcC;
|
cannam@167
|
2525 TdF = TcB + TcC;
|
cannam@167
|
2526 T7Q = T2g + T2j;
|
cannam@167
|
2527 T7R = T2p + T2s;
|
cannam@167
|
2528 T7S = T7Q + T7R;
|
cannam@167
|
2529 T8D = T7R - T7Q;
|
cannam@167
|
2530 }
|
cannam@167
|
2531 }
|
cannam@167
|
2532 {
|
cannam@167
|
2533 E T1z, T1C, T1D, Tcf, TbO, T4o, T4r, T7B, Tcg, TbP, T1G, T3Y, T1J, T3V, T1K;
|
cannam@167
|
2534 E T7C, Tcj, Tci, TbW, TbT, T1S, TfV, TfW, T41, T48, Tc8, Tcb, T7E, T1Z, TfY;
|
cannam@167
|
2535 E TfZ, T4a, T4h, Tc1, Tc4, T7F;
|
cannam@167
|
2536 {
|
cannam@167
|
2537 E T1x, T1y, T1A, T1B;
|
cannam@167
|
2538 T1x = ci[0];
|
cannam@167
|
2539 T1y = cr[WS(rs, 31)];
|
cannam@167
|
2540 T1z = T1x + T1y;
|
cannam@167
|
2541 T1A = cr[WS(rs, 15)];
|
cannam@167
|
2542 T1B = ci[WS(rs, 16)];
|
cannam@167
|
2543 T1C = T1A + T1B;
|
cannam@167
|
2544 T1D = T1z + T1C;
|
cannam@167
|
2545 Tcf = T1A - T1B;
|
cannam@167
|
2546 TbO = T1x - T1y;
|
cannam@167
|
2547 }
|
cannam@167
|
2548 {
|
cannam@167
|
2549 E T4m, T4n, T4p, T4q;
|
cannam@167
|
2550 T4m = ci[WS(rs, 32)];
|
cannam@167
|
2551 T4n = cr[WS(rs, 63)];
|
cannam@167
|
2552 T4o = T4m - T4n;
|
cannam@167
|
2553 T4p = ci[WS(rs, 48)];
|
cannam@167
|
2554 T4q = cr[WS(rs, 47)];
|
cannam@167
|
2555 T4r = T4p - T4q;
|
cannam@167
|
2556 T7B = T4o + T4r;
|
cannam@167
|
2557 Tcg = T4m + T4n;
|
cannam@167
|
2558 TbP = T4p + T4q;
|
cannam@167
|
2559 }
|
cannam@167
|
2560 {
|
cannam@167
|
2561 E TbR, TbS, TbU, TbV;
|
cannam@167
|
2562 {
|
cannam@167
|
2563 E T1E, T1F, T3W, T3X;
|
cannam@167
|
2564 T1E = cr[WS(rs, 7)];
|
cannam@167
|
2565 T1F = ci[WS(rs, 24)];
|
cannam@167
|
2566 T1G = T1E + T1F;
|
cannam@167
|
2567 TbR = T1E - T1F;
|
cannam@167
|
2568 T3W = ci[WS(rs, 56)];
|
cannam@167
|
2569 T3X = cr[WS(rs, 39)];
|
cannam@167
|
2570 T3Y = T3W - T3X;
|
cannam@167
|
2571 TbS = T3W + T3X;
|
cannam@167
|
2572 }
|
cannam@167
|
2573 {
|
cannam@167
|
2574 E T1H, T1I, T3T, T3U;
|
cannam@167
|
2575 T1H = ci[WS(rs, 8)];
|
cannam@167
|
2576 T1I = cr[WS(rs, 23)];
|
cannam@167
|
2577 T1J = T1H + T1I;
|
cannam@167
|
2578 TbU = T1H - T1I;
|
cannam@167
|
2579 T3T = ci[WS(rs, 40)];
|
cannam@167
|
2580 T3U = cr[WS(rs, 55)];
|
cannam@167
|
2581 T3V = T3T - T3U;
|
cannam@167
|
2582 TbV = T3T + T3U;
|
cannam@167
|
2583 }
|
cannam@167
|
2584 T1K = T1G + T1J;
|
cannam@167
|
2585 T7C = T3Y + T3V;
|
cannam@167
|
2586 Tcj = TbU + TbV;
|
cannam@167
|
2587 Tci = TbR + TbS;
|
cannam@167
|
2588 TbW = TbU - TbV;
|
cannam@167
|
2589 TbT = TbR - TbS;
|
cannam@167
|
2590 }
|
cannam@167
|
2591 {
|
cannam@167
|
2592 E T1O, Tc9, T47, Tca, T1R, Tc6, T44, Tc7;
|
cannam@167
|
2593 {
|
cannam@167
|
2594 E T1M, T1N, T45, T46;
|
cannam@167
|
2595 T1M = cr[WS(rs, 3)];
|
cannam@167
|
2596 T1N = ci[WS(rs, 28)];
|
cannam@167
|
2597 T1O = T1M + T1N;
|
cannam@167
|
2598 Tc9 = T1M - T1N;
|
cannam@167
|
2599 T45 = ci[WS(rs, 44)];
|
cannam@167
|
2600 T46 = cr[WS(rs, 51)];
|
cannam@167
|
2601 T47 = T45 - T46;
|
cannam@167
|
2602 Tca = T45 + T46;
|
cannam@167
|
2603 }
|
cannam@167
|
2604 {
|
cannam@167
|
2605 E T1P, T1Q, T42, T43;
|
cannam@167
|
2606 T1P = cr[WS(rs, 19)];
|
cannam@167
|
2607 T1Q = ci[WS(rs, 12)];
|
cannam@167
|
2608 T1R = T1P + T1Q;
|
cannam@167
|
2609 Tc6 = T1P - T1Q;
|
cannam@167
|
2610 T42 = ci[WS(rs, 60)];
|
cannam@167
|
2611 T43 = cr[WS(rs, 35)];
|
cannam@167
|
2612 T44 = T42 - T43;
|
cannam@167
|
2613 Tc7 = T42 + T43;
|
cannam@167
|
2614 }
|
cannam@167
|
2615 T1S = T1O + T1R;
|
cannam@167
|
2616 TfV = Tc9 + Tca;
|
cannam@167
|
2617 TfW = Tc7 - Tc6;
|
cannam@167
|
2618 T41 = T1O - T1R;
|
cannam@167
|
2619 T48 = T44 - T47;
|
cannam@167
|
2620 Tc8 = Tc6 + Tc7;
|
cannam@167
|
2621 Tcb = Tc9 - Tca;
|
cannam@167
|
2622 T7E = T44 + T47;
|
cannam@167
|
2623 }
|
cannam@167
|
2624 {
|
cannam@167
|
2625 E T1V, Tc2, T4g, Tc3, T1Y, TbZ, T4d, Tc0;
|
cannam@167
|
2626 {
|
cannam@167
|
2627 E T1T, T1U, T4e, T4f;
|
cannam@167
|
2628 T1T = ci[WS(rs, 4)];
|
cannam@167
|
2629 T1U = cr[WS(rs, 27)];
|
cannam@167
|
2630 T1V = T1T + T1U;
|
cannam@167
|
2631 Tc2 = T1T - T1U;
|
cannam@167
|
2632 T4e = ci[WS(rs, 52)];
|
cannam@167
|
2633 T4f = cr[WS(rs, 43)];
|
cannam@167
|
2634 T4g = T4e - T4f;
|
cannam@167
|
2635 Tc3 = T4e + T4f;
|
cannam@167
|
2636 }
|
cannam@167
|
2637 {
|
cannam@167
|
2638 E T1W, T1X, T4b, T4c;
|
cannam@167
|
2639 T1W = cr[WS(rs, 11)];
|
cannam@167
|
2640 T1X = ci[WS(rs, 20)];
|
cannam@167
|
2641 T1Y = T1W + T1X;
|
cannam@167
|
2642 TbZ = T1W - T1X;
|
cannam@167
|
2643 T4b = ci[WS(rs, 36)];
|
cannam@167
|
2644 T4c = cr[WS(rs, 59)];
|
cannam@167
|
2645 T4d = T4b - T4c;
|
cannam@167
|
2646 Tc0 = T4b + T4c;
|
cannam@167
|
2647 }
|
cannam@167
|
2648 T1Z = T1V + T1Y;
|
cannam@167
|
2649 TfY = Tc2 + Tc3;
|
cannam@167
|
2650 TfZ = TbZ + Tc0;
|
cannam@167
|
2651 T4a = T1V - T1Y;
|
cannam@167
|
2652 T4h = T4d - T4g;
|
cannam@167
|
2653 Tc1 = TbZ - Tc0;
|
cannam@167
|
2654 Tc4 = Tc2 - Tc3;
|
cannam@167
|
2655 T7F = T4d + T4g;
|
cannam@167
|
2656 }
|
cannam@167
|
2657 T1L = T1D + T1K;
|
cannam@167
|
2658 T20 = T1S + T1Z;
|
cannam@167
|
2659 T7A = T1L - T20;
|
cannam@167
|
2660 T7D = T7B + T7C;
|
cannam@167
|
2661 T7G = T7E + T7F;
|
cannam@167
|
2662 T7H = T7D - T7G;
|
cannam@167
|
2663 {
|
cannam@167
|
2664 E T3S, T3Z, TfX, Tg0;
|
cannam@167
|
2665 T3S = T1z - T1C;
|
cannam@167
|
2666 T3Z = T3V - T3Y;
|
cannam@167
|
2667 T40 = T3S + T3Z;
|
cannam@167
|
2668 T62 = T3S - T3Z;
|
cannam@167
|
2669 TfX = FNMS(KP923879532, TfW, KP382683432 * TfV);
|
cannam@167
|
2670 Tg0 = FNMS(KP923879532, TfZ, KP382683432 * TfY);
|
cannam@167
|
2671 Tg1 = TfX + Tg0;
|
cannam@167
|
2672 Thv = TfX - Tg0;
|
cannam@167
|
2673 }
|
cannam@167
|
2674 {
|
cannam@167
|
2675 E Tg6, Tg7, Tg3, Tg4;
|
cannam@167
|
2676 Tg6 = FMA(KP382683432, TfW, KP923879532 * TfV);
|
cannam@167
|
2677 Tg7 = FMA(KP382683432, TfZ, KP923879532 * TfY);
|
cannam@167
|
2678 Tg8 = Tg6 - Tg7;
|
cannam@167
|
2679 Thz = Tg6 + Tg7;
|
cannam@167
|
2680 Tg3 = KP707106781 * (TbT - TbW);
|
cannam@167
|
2681 Tg4 = Tcf + Tcg;
|
cannam@167
|
2682 Tg5 = Tg3 - Tg4;
|
cannam@167
|
2683 Thw = Tg4 + Tg3;
|
cannam@167
|
2684 }
|
cannam@167
|
2685 {
|
cannam@167
|
2686 E T4l, T4s, T49, T4i;
|
cannam@167
|
2687 T4l = T1G - T1J;
|
cannam@167
|
2688 T4s = T4o - T4r;
|
cannam@167
|
2689 T4t = T4l + T4s;
|
cannam@167
|
2690 T5Z = T4s - T4l;
|
cannam@167
|
2691 T49 = T41 - T48;
|
cannam@167
|
2692 T4i = T4a + T4h;
|
cannam@167
|
2693 T4j = KP707106781 * (T49 + T4i);
|
cannam@167
|
2694 T60 = KP707106781 * (T49 - T4i);
|
cannam@167
|
2695 }
|
cannam@167
|
2696 {
|
cannam@167
|
2697 E T4u, T4v, TbQ, TbX;
|
cannam@167
|
2698 T4u = T41 + T48;
|
cannam@167
|
2699 T4v = T4h - T4a;
|
cannam@167
|
2700 T4w = KP707106781 * (T4u + T4v);
|
cannam@167
|
2701 T63 = KP707106781 * (T4v - T4u);
|
cannam@167
|
2702 TbQ = TbO - TbP;
|
cannam@167
|
2703 TbX = KP707106781 * (TbT + TbW);
|
cannam@167
|
2704 TbY = TbQ - TbX;
|
cannam@167
|
2705 TdS = TbQ + TbX;
|
cannam@167
|
2706 }
|
cannam@167
|
2707 {
|
cannam@167
|
2708 E Tc5, Tcc, TfS, TfT;
|
cannam@167
|
2709 Tc5 = FNMS(KP382683432, Tc4, KP923879532 * Tc1);
|
cannam@167
|
2710 Tcc = FMA(KP923879532, Tc8, KP382683432 * Tcb);
|
cannam@167
|
2711 Tcd = Tc5 - Tcc;
|
cannam@167
|
2712 TdQ = Tcc + Tc5;
|
cannam@167
|
2713 TfS = TbO + TbP;
|
cannam@167
|
2714 TfT = KP707106781 * (Tci + Tcj);
|
cannam@167
|
2715 TfU = TfS - TfT;
|
cannam@167
|
2716 Thy = TfS + TfT;
|
cannam@167
|
2717 }
|
cannam@167
|
2718 {
|
cannam@167
|
2719 E T8N, T8O, T8Q, T8R;
|
cannam@167
|
2720 T8N = T7B - T7C;
|
cannam@167
|
2721 T8O = T1S - T1Z;
|
cannam@167
|
2722 T8P = T8N - T8O;
|
cannam@167
|
2723 T9z = T8O + T8N;
|
cannam@167
|
2724 T8Q = T1D - T1K;
|
cannam@167
|
2725 T8R = T7F - T7E;
|
cannam@167
|
2726 T8S = T8Q - T8R;
|
cannam@167
|
2727 T9A = T8Q + T8R;
|
cannam@167
|
2728 }
|
cannam@167
|
2729 {
|
cannam@167
|
2730 E Tch, Tck, Tcm, Tcn;
|
cannam@167
|
2731 Tch = Tcf - Tcg;
|
cannam@167
|
2732 Tck = KP707106781 * (Tci - Tcj);
|
cannam@167
|
2733 Tcl = Tch - Tck;
|
cannam@167
|
2734 TdP = Tch + Tck;
|
cannam@167
|
2735 Tcm = FNMS(KP382683432, Tc8, KP923879532 * Tcb);
|
cannam@167
|
2736 Tcn = FMA(KP382683432, Tc1, KP923879532 * Tc4);
|
cannam@167
|
2737 Tco = Tcm - Tcn;
|
cannam@167
|
2738 TdT = Tcm + Tcn;
|
cannam@167
|
2739 }
|
cannam@167
|
2740 }
|
cannam@167
|
2741 {
|
cannam@167
|
2742 E T14, T17, T18, TbC, Tbb, T3H, T3K, T7s, TbD, Tbc, T1b, T3h, T1e, T3e, T1f;
|
cannam@167
|
2743 E T7t, TbG, TbF, Tbj, Tbg, T1n, TfC, TfD, T3k, T3r, Tbv, Tby, T7v, T1u, TfF;
|
cannam@167
|
2744 E TfG, T3t, T3A, Tbo, Tbr, T7w;
|
cannam@167
|
2745 {
|
cannam@167
|
2746 E T12, T13, T15, T16;
|
cannam@167
|
2747 T12 = cr[WS(rs, 1)];
|
cannam@167
|
2748 T13 = ci[WS(rs, 30)];
|
cannam@167
|
2749 T14 = T12 + T13;
|
cannam@167
|
2750 T15 = cr[WS(rs, 17)];
|
cannam@167
|
2751 T16 = ci[WS(rs, 14)];
|
cannam@167
|
2752 T17 = T15 + T16;
|
cannam@167
|
2753 T18 = T14 + T17;
|
cannam@167
|
2754 TbC = T15 - T16;
|
cannam@167
|
2755 Tbb = T12 - T13;
|
cannam@167
|
2756 }
|
cannam@167
|
2757 {
|
cannam@167
|
2758 E T3F, T3G, T3I, T3J;
|
cannam@167
|
2759 T3F = ci[WS(rs, 62)];
|
cannam@167
|
2760 T3G = cr[WS(rs, 33)];
|
cannam@167
|
2761 T3H = T3F - T3G;
|
cannam@167
|
2762 T3I = ci[WS(rs, 46)];
|
cannam@167
|
2763 T3J = cr[WS(rs, 49)];
|
cannam@167
|
2764 T3K = T3I - T3J;
|
cannam@167
|
2765 T7s = T3H + T3K;
|
cannam@167
|
2766 TbD = T3F + T3G;
|
cannam@167
|
2767 Tbc = T3I + T3J;
|
cannam@167
|
2768 }
|
cannam@167
|
2769 {
|
cannam@167
|
2770 E Tbe, Tbf, Tbh, Tbi;
|
cannam@167
|
2771 {
|
cannam@167
|
2772 E T19, T1a, T3f, T3g;
|
cannam@167
|
2773 T19 = cr[WS(rs, 9)];
|
cannam@167
|
2774 T1a = ci[WS(rs, 22)];
|
cannam@167
|
2775 T1b = T19 + T1a;
|
cannam@167
|
2776 Tbe = T19 - T1a;
|
cannam@167
|
2777 T3f = ci[WS(rs, 54)];
|
cannam@167
|
2778 T3g = cr[WS(rs, 41)];
|
cannam@167
|
2779 T3h = T3f - T3g;
|
cannam@167
|
2780 Tbf = T3f + T3g;
|
cannam@167
|
2781 }
|
cannam@167
|
2782 {
|
cannam@167
|
2783 E T1c, T1d, T3c, T3d;
|
cannam@167
|
2784 T1c = ci[WS(rs, 6)];
|
cannam@167
|
2785 T1d = cr[WS(rs, 25)];
|
cannam@167
|
2786 T1e = T1c + T1d;
|
cannam@167
|
2787 Tbh = T1c - T1d;
|
cannam@167
|
2788 T3c = ci[WS(rs, 38)];
|
cannam@167
|
2789 T3d = cr[WS(rs, 57)];
|
cannam@167
|
2790 T3e = T3c - T3d;
|
cannam@167
|
2791 Tbi = T3c + T3d;
|
cannam@167
|
2792 }
|
cannam@167
|
2793 T1f = T1b + T1e;
|
cannam@167
|
2794 T7t = T3h + T3e;
|
cannam@167
|
2795 TbG = Tbh + Tbi;
|
cannam@167
|
2796 TbF = Tbe + Tbf;
|
cannam@167
|
2797 Tbj = Tbh - Tbi;
|
cannam@167
|
2798 Tbg = Tbe - Tbf;
|
cannam@167
|
2799 }
|
cannam@167
|
2800 {
|
cannam@167
|
2801 E T1j, Tbw, T3q, Tbx, T1m, Tbt, T3n, Tbu;
|
cannam@167
|
2802 {
|
cannam@167
|
2803 E T1h, T1i, T3o, T3p;
|
cannam@167
|
2804 T1h = cr[WS(rs, 5)];
|
cannam@167
|
2805 T1i = ci[WS(rs, 26)];
|
cannam@167
|
2806 T1j = T1h + T1i;
|
cannam@167
|
2807 Tbw = T1h - T1i;
|
cannam@167
|
2808 T3o = ci[WS(rs, 42)];
|
cannam@167
|
2809 T3p = cr[WS(rs, 53)];
|
cannam@167
|
2810 T3q = T3o - T3p;
|
cannam@167
|
2811 Tbx = T3o + T3p;
|
cannam@167
|
2812 }
|
cannam@167
|
2813 {
|
cannam@167
|
2814 E T1k, T1l, T3l, T3m;
|
cannam@167
|
2815 T1k = cr[WS(rs, 21)];
|
cannam@167
|
2816 T1l = ci[WS(rs, 10)];
|
cannam@167
|
2817 T1m = T1k + T1l;
|
cannam@167
|
2818 Tbt = T1k - T1l;
|
cannam@167
|
2819 T3l = ci[WS(rs, 58)];
|
cannam@167
|
2820 T3m = cr[WS(rs, 37)];
|
cannam@167
|
2821 T3n = T3l - T3m;
|
cannam@167
|
2822 Tbu = T3l + T3m;
|
cannam@167
|
2823 }
|
cannam@167
|
2824 T1n = T1j + T1m;
|
cannam@167
|
2825 TfC = Tbw + Tbx;
|
cannam@167
|
2826 TfD = Tbu - Tbt;
|
cannam@167
|
2827 T3k = T1j - T1m;
|
cannam@167
|
2828 T3r = T3n - T3q;
|
cannam@167
|
2829 Tbv = Tbt + Tbu;
|
cannam@167
|
2830 Tby = Tbw - Tbx;
|
cannam@167
|
2831 T7v = T3n + T3q;
|
cannam@167
|
2832 }
|
cannam@167
|
2833 {
|
cannam@167
|
2834 E T1q, Tbp, T3z, Tbq, T1t, Tbm, T3w, Tbn;
|
cannam@167
|
2835 {
|
cannam@167
|
2836 E T1o, T1p, T3x, T3y;
|
cannam@167
|
2837 T1o = ci[WS(rs, 2)];
|
cannam@167
|
2838 T1p = cr[WS(rs, 29)];
|
cannam@167
|
2839 T1q = T1o + T1p;
|
cannam@167
|
2840 Tbp = T1o - T1p;
|
cannam@167
|
2841 T3x = ci[WS(rs, 50)];
|
cannam@167
|
2842 T3y = cr[WS(rs, 45)];
|
cannam@167
|
2843 T3z = T3x - T3y;
|
cannam@167
|
2844 Tbq = T3x + T3y;
|
cannam@167
|
2845 }
|
cannam@167
|
2846 {
|
cannam@167
|
2847 E T1r, T1s, T3u, T3v;
|
cannam@167
|
2848 T1r = cr[WS(rs, 13)];
|
cannam@167
|
2849 T1s = ci[WS(rs, 18)];
|
cannam@167
|
2850 T1t = T1r + T1s;
|
cannam@167
|
2851 Tbm = T1r - T1s;
|
cannam@167
|
2852 T3u = ci[WS(rs, 34)];
|
cannam@167
|
2853 T3v = cr[WS(rs, 61)];
|
cannam@167
|
2854 T3w = T3u - T3v;
|
cannam@167
|
2855 Tbn = T3u + T3v;
|
cannam@167
|
2856 }
|
cannam@167
|
2857 T1u = T1q + T1t;
|
cannam@167
|
2858 TfF = Tbp + Tbq;
|
cannam@167
|
2859 TfG = Tbm + Tbn;
|
cannam@167
|
2860 T3t = T1q - T1t;
|
cannam@167
|
2861 T3A = T3w - T3z;
|
cannam@167
|
2862 Tbo = Tbm - Tbn;
|
cannam@167
|
2863 Tbr = Tbp - Tbq;
|
cannam@167
|
2864 T7w = T3w + T3z;
|
cannam@167
|
2865 }
|
cannam@167
|
2866 T1g = T18 + T1f;
|
cannam@167
|
2867 T1v = T1n + T1u;
|
cannam@167
|
2868 T7r = T1g - T1v;
|
cannam@167
|
2869 T7u = T7s + T7t;
|
cannam@167
|
2870 T7x = T7v + T7w;
|
cannam@167
|
2871 T7y = T7u - T7x;
|
cannam@167
|
2872 {
|
cannam@167
|
2873 E T3b, T3i, TfE, TfH;
|
cannam@167
|
2874 T3b = T14 - T17;
|
cannam@167
|
2875 T3i = T3e - T3h;
|
cannam@167
|
2876 T3j = T3b + T3i;
|
cannam@167
|
2877 T69 = T3b - T3i;
|
cannam@167
|
2878 TfE = FNMS(KP923879532, TfD, KP382683432 * TfC);
|
cannam@167
|
2879 TfH = FNMS(KP923879532, TfG, KP382683432 * TfF);
|
cannam@167
|
2880 TfI = TfE + TfH;
|
cannam@167
|
2881 ThD = TfE - TfH;
|
cannam@167
|
2882 }
|
cannam@167
|
2883 {
|
cannam@167
|
2884 E TfN, TfO, TfK, TfL;
|
cannam@167
|
2885 TfN = FMA(KP382683432, TfD, KP923879532 * TfC);
|
cannam@167
|
2886 TfO = FMA(KP382683432, TfG, KP923879532 * TfF);
|
cannam@167
|
2887 TfP = TfN - TfO;
|
cannam@167
|
2888 ThG = TfN + TfO;
|
cannam@167
|
2889 TfK = TbD - TbC;
|
cannam@167
|
2890 TfL = KP707106781 * (Tbg - Tbj);
|
cannam@167
|
2891 TfM = TfK + TfL;
|
cannam@167
|
2892 ThC = TfK - TfL;
|
cannam@167
|
2893 }
|
cannam@167
|
2894 {
|
cannam@167
|
2895 E T3E, T3L, T3s, T3B;
|
cannam@167
|
2896 T3E = T1b - T1e;
|
cannam@167
|
2897 T3L = T3H - T3K;
|
cannam@167
|
2898 T3M = T3E + T3L;
|
cannam@167
|
2899 T66 = T3L - T3E;
|
cannam@167
|
2900 T3s = T3k - T3r;
|
cannam@167
|
2901 T3B = T3t + T3A;
|
cannam@167
|
2902 T3C = KP707106781 * (T3s + T3B);
|
cannam@167
|
2903 T67 = KP707106781 * (T3s - T3B);
|
cannam@167
|
2904 }
|
cannam@167
|
2905 {
|
cannam@167
|
2906 E T3N, T3O, Tbd, Tbk;
|
cannam@167
|
2907 T3N = T3k + T3r;
|
cannam@167
|
2908 T3O = T3A - T3t;
|
cannam@167
|
2909 T3P = KP707106781 * (T3N + T3O);
|
cannam@167
|
2910 T6a = KP707106781 * (T3O - T3N);
|
cannam@167
|
2911 Tbd = Tbb - Tbc;
|
cannam@167
|
2912 Tbk = KP707106781 * (Tbg + Tbj);
|
cannam@167
|
2913 Tbl = Tbd - Tbk;
|
cannam@167
|
2914 TdZ = Tbd + Tbk;
|
cannam@167
|
2915 }
|
cannam@167
|
2916 {
|
cannam@167
|
2917 E Tbs, Tbz, Tfz, TfA;
|
cannam@167
|
2918 Tbs = FNMS(KP382683432, Tbr, KP923879532 * Tbo);
|
cannam@167
|
2919 Tbz = FMA(KP923879532, Tbv, KP382683432 * Tby);
|
cannam@167
|
2920 TbA = Tbs - Tbz;
|
cannam@167
|
2921 TdX = Tbz + Tbs;
|
cannam@167
|
2922 Tfz = Tbb + Tbc;
|
cannam@167
|
2923 TfA = KP707106781 * (TbF + TbG);
|
cannam@167
|
2924 TfB = Tfz - TfA;
|
cannam@167
|
2925 ThF = Tfz + TfA;
|
cannam@167
|
2926 }
|
cannam@167
|
2927 {
|
cannam@167
|
2928 E T8U, T8V, T8X, T8Y;
|
cannam@167
|
2929 T8U = T7s - T7t;
|
cannam@167
|
2930 T8V = T1n - T1u;
|
cannam@167
|
2931 T8W = T8U - T8V;
|
cannam@167
|
2932 T9C = T8V + T8U;
|
cannam@167
|
2933 T8X = T18 - T1f;
|
cannam@167
|
2934 T8Y = T7w - T7v;
|
cannam@167
|
2935 T8Z = T8X - T8Y;
|
cannam@167
|
2936 T9D = T8X + T8Y;
|
cannam@167
|
2937 }
|
cannam@167
|
2938 {
|
cannam@167
|
2939 E TbE, TbH, TbJ, TbK;
|
cannam@167
|
2940 TbE = TbC + TbD;
|
cannam@167
|
2941 TbH = KP707106781 * (TbF - TbG);
|
cannam@167
|
2942 TbI = TbE - TbH;
|
cannam@167
|
2943 TdW = TbE + TbH;
|
cannam@167
|
2944 TbJ = FNMS(KP382683432, Tbv, KP923879532 * Tby);
|
cannam@167
|
2945 TbK = FMA(KP382683432, Tbo, KP923879532 * Tbr);
|
cannam@167
|
2946 TbL = TbJ - TbK;
|
cannam@167
|
2947 Te0 = TbJ + TbK;
|
cannam@167
|
2948 }
|
cannam@167
|
2949 }
|
cannam@167
|
2950 {
|
cannam@167
|
2951 E T11, T8q, T8n, T8r, T22, T8v, T8k, T8u;
|
cannam@167
|
2952 {
|
cannam@167
|
2953 E Tv, T10, T8l, T8m;
|
cannam@167
|
2954 Tv = Tf + Tu;
|
cannam@167
|
2955 T10 = TK + TZ;
|
cannam@167
|
2956 T11 = Tv + T10;
|
cannam@167
|
2957 T8q = Tv - T10;
|
cannam@167
|
2958 T8l = T7u + T7x;
|
cannam@167
|
2959 T8m = T7D + T7G;
|
cannam@167
|
2960 T8n = T8l + T8m;
|
cannam@167
|
2961 T8r = T8m - T8l;
|
cannam@167
|
2962 }
|
cannam@167
|
2963 {
|
cannam@167
|
2964 E T1w, T21, T8i, T8j;
|
cannam@167
|
2965 T1w = T1g + T1v;
|
cannam@167
|
2966 T21 = T1L + T20;
|
cannam@167
|
2967 T22 = T1w + T21;
|
cannam@167
|
2968 T8v = T1w - T21;
|
cannam@167
|
2969 T8i = T7P + T7S;
|
cannam@167
|
2970 T8j = T7o + T7l;
|
cannam@167
|
2971 T8k = T8i + T8j;
|
cannam@167
|
2972 T8u = T8i - T8j;
|
cannam@167
|
2973 }
|
cannam@167
|
2974 cr[0] = T11 + T22;
|
cannam@167
|
2975 ci[0] = T8k + T8n;
|
cannam@167
|
2976 {
|
cannam@167
|
2977 E T8g, T8o, T8f, T8h;
|
cannam@167
|
2978 T8g = T11 - T22;
|
cannam@167
|
2979 T8o = T8k - T8n;
|
cannam@167
|
2980 T8f = W[62];
|
cannam@167
|
2981 T8h = W[63];
|
cannam@167
|
2982 cr[WS(rs, 32)] = FNMS(T8h, T8o, T8f * T8g);
|
cannam@167
|
2983 ci[WS(rs, 32)] = FMA(T8h, T8g, T8f * T8o);
|
cannam@167
|
2984 }
|
cannam@167
|
2985 {
|
cannam@167
|
2986 E T8s, T8w, T8p, T8t;
|
cannam@167
|
2987 T8s = T8q - T8r;
|
cannam@167
|
2988 T8w = T8u - T8v;
|
cannam@167
|
2989 T8p = W[94];
|
cannam@167
|
2990 T8t = W[95];
|
cannam@167
|
2991 cr[WS(rs, 48)] = FNMS(T8t, T8w, T8p * T8s);
|
cannam@167
|
2992 ci[WS(rs, 48)] = FMA(T8p, T8w, T8t * T8s);
|
cannam@167
|
2993 }
|
cannam@167
|
2994 {
|
cannam@167
|
2995 E T8y, T8A, T8x, T8z;
|
cannam@167
|
2996 T8y = T8q + T8r;
|
cannam@167
|
2997 T8A = T8v + T8u;
|
cannam@167
|
2998 T8x = W[30];
|
cannam@167
|
2999 T8z = W[31];
|
cannam@167
|
3000 cr[WS(rs, 16)] = FNMS(T8z, T8A, T8x * T8y);
|
cannam@167
|
3001 ci[WS(rs, 16)] = FMA(T8x, T8A, T8z * T8y);
|
cannam@167
|
3002 }
|
cannam@167
|
3003 }
|
cannam@167
|
3004 {
|
cannam@167
|
3005 E T9y, T9U, T9N, T9V, T9F, T9Z, T9K, T9Y;
|
cannam@167
|
3006 {
|
cannam@167
|
3007 E T9w, T9x, T9L, T9M;
|
cannam@167
|
3008 T9w = T8C + T8D;
|
cannam@167
|
3009 T9x = KP707106781 * (T97 + T98);
|
cannam@167
|
3010 T9y = T9w - T9x;
|
cannam@167
|
3011 T9U = T9w + T9x;
|
cannam@167
|
3012 T9L = FNMS(KP382683432, T9C, KP923879532 * T9D);
|
cannam@167
|
3013 T9M = FMA(KP382683432, T9z, KP923879532 * T9A);
|
cannam@167
|
3014 T9N = T9L - T9M;
|
cannam@167
|
3015 T9V = T9L + T9M;
|
cannam@167
|
3016 }
|
cannam@167
|
3017 {
|
cannam@167
|
3018 E T9B, T9E, T9I, T9J;
|
cannam@167
|
3019 T9B = FNMS(KP382683432, T9A, KP923879532 * T9z);
|
cannam@167
|
3020 T9E = FMA(KP923879532, T9C, KP382683432 * T9D);
|
cannam@167
|
3021 T9F = T9B - T9E;
|
cannam@167
|
3022 T9Z = T9E + T9B;
|
cannam@167
|
3023 T9I = T95 + T94;
|
cannam@167
|
3024 T9J = KP707106781 * (T8K + T8H);
|
cannam@167
|
3025 T9K = T9I - T9J;
|
cannam@167
|
3026 T9Y = T9I + T9J;
|
cannam@167
|
3027 }
|
cannam@167
|
3028 {
|
cannam@167
|
3029 E T9G, T9O, T9v, T9H;
|
cannam@167
|
3030 T9G = T9y - T9F;
|
cannam@167
|
3031 T9O = T9K - T9N;
|
cannam@167
|
3032 T9v = W[102];
|
cannam@167
|
3033 T9H = W[103];
|
cannam@167
|
3034 cr[WS(rs, 52)] = FNMS(T9H, T9O, T9v * T9G);
|
cannam@167
|
3035 ci[WS(rs, 52)] = FMA(T9H, T9G, T9v * T9O);
|
cannam@167
|
3036 }
|
cannam@167
|
3037 {
|
cannam@167
|
3038 E Ta2, Ta4, Ta1, Ta3;
|
cannam@167
|
3039 Ta2 = T9U + T9V;
|
cannam@167
|
3040 Ta4 = T9Y + T9Z;
|
cannam@167
|
3041 Ta1 = W[6];
|
cannam@167
|
3042 Ta3 = W[7];
|
cannam@167
|
3043 cr[WS(rs, 4)] = FNMS(Ta3, Ta4, Ta1 * Ta2);
|
cannam@167
|
3044 ci[WS(rs, 4)] = FMA(Ta1, Ta4, Ta3 * Ta2);
|
cannam@167
|
3045 }
|
cannam@167
|
3046 {
|
cannam@167
|
3047 E T9Q, T9S, T9P, T9R;
|
cannam@167
|
3048 T9Q = T9y + T9F;
|
cannam@167
|
3049 T9S = T9K + T9N;
|
cannam@167
|
3050 T9P = W[38];
|
cannam@167
|
3051 T9R = W[39];
|
cannam@167
|
3052 cr[WS(rs, 20)] = FNMS(T9R, T9S, T9P * T9Q);
|
cannam@167
|
3053 ci[WS(rs, 20)] = FMA(T9R, T9Q, T9P * T9S);
|
cannam@167
|
3054 }
|
cannam@167
|
3055 {
|
cannam@167
|
3056 E T9W, Ta0, T9T, T9X;
|
cannam@167
|
3057 T9W = T9U - T9V;
|
cannam@167
|
3058 Ta0 = T9Y - T9Z;
|
cannam@167
|
3059 T9T = W[70];
|
cannam@167
|
3060 T9X = W[71];
|
cannam@167
|
3061 cr[WS(rs, 36)] = FNMS(T9X, Ta0, T9T * T9W);
|
cannam@167
|
3062 ci[WS(rs, 36)] = FMA(T9T, Ta0, T9X * T9W);
|
cannam@167
|
3063 }
|
cannam@167
|
3064 }
|
cannam@167
|
3065 {
|
cannam@167
|
3066 E T8M, T9k, T9d, T9l, T91, T9p, T9a, T9o;
|
cannam@167
|
3067 {
|
cannam@167
|
3068 E T8E, T8L, T9b, T9c;
|
cannam@167
|
3069 T8E = T8C - T8D;
|
cannam@167
|
3070 T8L = KP707106781 * (T8H - T8K);
|
cannam@167
|
3071 T8M = T8E - T8L;
|
cannam@167
|
3072 T9k = T8E + T8L;
|
cannam@167
|
3073 T9b = FNMS(KP923879532, T8W, KP382683432 * T8Z);
|
cannam@167
|
3074 T9c = FMA(KP923879532, T8P, KP382683432 * T8S);
|
cannam@167
|
3075 T9d = T9b - T9c;
|
cannam@167
|
3076 T9l = T9b + T9c;
|
cannam@167
|
3077 }
|
cannam@167
|
3078 {
|
cannam@167
|
3079 E T8T, T90, T96, T99;
|
cannam@167
|
3080 T8T = FNMS(KP923879532, T8S, KP382683432 * T8P);
|
cannam@167
|
3081 T90 = FMA(KP382683432, T8W, KP923879532 * T8Z);
|
cannam@167
|
3082 T91 = T8T - T90;
|
cannam@167
|
3083 T9p = T90 + T8T;
|
cannam@167
|
3084 T96 = T94 - T95;
|
cannam@167
|
3085 T99 = KP707106781 * (T97 - T98);
|
cannam@167
|
3086 T9a = T96 - T99;
|
cannam@167
|
3087 T9o = T96 + T99;
|
cannam@167
|
3088 }
|
cannam@167
|
3089 {
|
cannam@167
|
3090 E T92, T9e, T8B, T93;
|
cannam@167
|
3091 T92 = T8M - T91;
|
cannam@167
|
3092 T9e = T9a - T9d;
|
cannam@167
|
3093 T8B = W[118];
|
cannam@167
|
3094 T93 = W[119];
|
cannam@167
|
3095 cr[WS(rs, 60)] = FNMS(T93, T9e, T8B * T92);
|
cannam@167
|
3096 ci[WS(rs, 60)] = FMA(T93, T92, T8B * T9e);
|
cannam@167
|
3097 }
|
cannam@167
|
3098 {
|
cannam@167
|
3099 E T9s, T9u, T9r, T9t;
|
cannam@167
|
3100 T9s = T9k + T9l;
|
cannam@167
|
3101 T9u = T9o + T9p;
|
cannam@167
|
3102 T9r = W[22];
|
cannam@167
|
3103 T9t = W[23];
|
cannam@167
|
3104 cr[WS(rs, 12)] = FNMS(T9t, T9u, T9r * T9s);
|
cannam@167
|
3105 ci[WS(rs, 12)] = FMA(T9r, T9u, T9t * T9s);
|
cannam@167
|
3106 }
|
cannam@167
|
3107 {
|
cannam@167
|
3108 E T9g, T9i, T9f, T9h;
|
cannam@167
|
3109 T9g = T8M + T91;
|
cannam@167
|
3110 T9i = T9a + T9d;
|
cannam@167
|
3111 T9f = W[54];
|
cannam@167
|
3112 T9h = W[55];
|
cannam@167
|
3113 cr[WS(rs, 28)] = FNMS(T9h, T9i, T9f * T9g);
|
cannam@167
|
3114 ci[WS(rs, 28)] = FMA(T9h, T9g, T9f * T9i);
|
cannam@167
|
3115 }
|
cannam@167
|
3116 {
|
cannam@167
|
3117 E T9m, T9q, T9j, T9n;
|
cannam@167
|
3118 T9m = T9k - T9l;
|
cannam@167
|
3119 T9q = T9o - T9p;
|
cannam@167
|
3120 T9j = W[86];
|
cannam@167
|
3121 T9n = W[87];
|
cannam@167
|
3122 cr[WS(rs, 44)] = FNMS(T9n, T9q, T9j * T9m);
|
cannam@167
|
3123 ci[WS(rs, 44)] = FMA(T9j, T9q, T9n * T9m);
|
cannam@167
|
3124 }
|
cannam@167
|
3125 }
|
cannam@167
|
3126 {
|
cannam@167
|
3127 E T7q, T84, T7X, T85, T7J, T89, T7U, T88;
|
cannam@167
|
3128 {
|
cannam@167
|
3129 E T7i, T7p, T7V, T7W;
|
cannam@167
|
3130 T7i = Tf - Tu;
|
cannam@167
|
3131 T7p = T7l - T7o;
|
cannam@167
|
3132 T7q = T7i + T7p;
|
cannam@167
|
3133 T84 = T7i - T7p;
|
cannam@167
|
3134 T7V = T7r + T7y;
|
cannam@167
|
3135 T7W = T7H - T7A;
|
cannam@167
|
3136 T7X = KP707106781 * (T7V + T7W);
|
cannam@167
|
3137 T85 = KP707106781 * (T7W - T7V);
|
cannam@167
|
3138 }
|
cannam@167
|
3139 {
|
cannam@167
|
3140 E T7z, T7I, T7M, T7T;
|
cannam@167
|
3141 T7z = T7r - T7y;
|
cannam@167
|
3142 T7I = T7A + T7H;
|
cannam@167
|
3143 T7J = KP707106781 * (T7z + T7I);
|
cannam@167
|
3144 T89 = KP707106781 * (T7z - T7I);
|
cannam@167
|
3145 T7M = TK - TZ;
|
cannam@167
|
3146 T7T = T7P - T7S;
|
cannam@167
|
3147 T7U = T7M + T7T;
|
cannam@167
|
3148 T88 = T7T - T7M;
|
cannam@167
|
3149 }
|
cannam@167
|
3150 {
|
cannam@167
|
3151 E T7K, T7Y, T7h, T7L;
|
cannam@167
|
3152 T7K = T7q - T7J;
|
cannam@167
|
3153 T7Y = T7U - T7X;
|
cannam@167
|
3154 T7h = W[78];
|
cannam@167
|
3155 T7L = W[79];
|
cannam@167
|
3156 cr[WS(rs, 40)] = FNMS(T7L, T7Y, T7h * T7K);
|
cannam@167
|
3157 ci[WS(rs, 40)] = FMA(T7L, T7K, T7h * T7Y);
|
cannam@167
|
3158 }
|
cannam@167
|
3159 {
|
cannam@167
|
3160 E T8c, T8e, T8b, T8d;
|
cannam@167
|
3161 T8c = T84 + T85;
|
cannam@167
|
3162 T8e = T88 + T89;
|
cannam@167
|
3163 T8b = W[46];
|
cannam@167
|
3164 T8d = W[47];
|
cannam@167
|
3165 cr[WS(rs, 24)] = FNMS(T8d, T8e, T8b * T8c);
|
cannam@167
|
3166 ci[WS(rs, 24)] = FMA(T8b, T8e, T8d * T8c);
|
cannam@167
|
3167 }
|
cannam@167
|
3168 {
|
cannam@167
|
3169 E T80, T82, T7Z, T81;
|
cannam@167
|
3170 T80 = T7q + T7J;
|
cannam@167
|
3171 T82 = T7U + T7X;
|
cannam@167
|
3172 T7Z = W[14];
|
cannam@167
|
3173 T81 = W[15];
|
cannam@167
|
3174 cr[WS(rs, 8)] = FNMS(T81, T82, T7Z * T80);
|
cannam@167
|
3175 ci[WS(rs, 8)] = FMA(T81, T80, T7Z * T82);
|
cannam@167
|
3176 }
|
cannam@167
|
3177 {
|
cannam@167
|
3178 E T86, T8a, T83, T87;
|
cannam@167
|
3179 T86 = T84 - T85;
|
cannam@167
|
3180 T8a = T88 - T89;
|
cannam@167
|
3181 T83 = W[110];
|
cannam@167
|
3182 T87 = W[111];
|
cannam@167
|
3183 cr[WS(rs, 56)] = FNMS(T87, T8a, T83 * T86);
|
cannam@167
|
3184 ci[WS(rs, 56)] = FMA(T83, T8a, T87 * T86);
|
cannam@167
|
3185 }
|
cannam@167
|
3186 }
|
cannam@167
|
3187 {
|
cannam@167
|
3188 E T6K, T76, T6W, T7a, T6R, T7b, T6Z, T77;
|
cannam@167
|
3189 {
|
cannam@167
|
3190 E T6I, T6J, T6U, T6V;
|
cannam@167
|
3191 T6I = T5O + T5P;
|
cannam@167
|
3192 T6J = T6j + T6k;
|
cannam@167
|
3193 T6K = T6I - T6J;
|
cannam@167
|
3194 T76 = T6I + T6J;
|
cannam@167
|
3195 T6U = T6g + T6h;
|
cannam@167
|
3196 T6V = T5W + T5T;
|
cannam@167
|
3197 T6W = T6U - T6V;
|
cannam@167
|
3198 T7a = T6U + T6V;
|
cannam@167
|
3199 {
|
cannam@167
|
3200 E T6N, T6Y, T6Q, T6X;
|
cannam@167
|
3201 {
|
cannam@167
|
3202 E T6L, T6M, T6O, T6P;
|
cannam@167
|
3203 T6L = T5Z + T60;
|
cannam@167
|
3204 T6M = T62 + T63;
|
cannam@167
|
3205 T6N = FNMS(KP555570233, T6M, KP831469612 * T6L);
|
cannam@167
|
3206 T6Y = FMA(KP555570233, T6L, KP831469612 * T6M);
|
cannam@167
|
3207 T6O = T66 + T67;
|
cannam@167
|
3208 T6P = T69 + T6a;
|
cannam@167
|
3209 T6Q = FMA(KP831469612, T6O, KP555570233 * T6P);
|
cannam@167
|
3210 T6X = FNMS(KP555570233, T6O, KP831469612 * T6P);
|
cannam@167
|
3211 }
|
cannam@167
|
3212 T6R = T6N - T6Q;
|
cannam@167
|
3213 T7b = T6Q + T6N;
|
cannam@167
|
3214 T6Z = T6X - T6Y;
|
cannam@167
|
3215 T77 = T6X + T6Y;
|
cannam@167
|
3216 }
|
cannam@167
|
3217 }
|
cannam@167
|
3218 {
|
cannam@167
|
3219 E T6S, T70, T6H, T6T;
|
cannam@167
|
3220 T6S = T6K - T6R;
|
cannam@167
|
3221 T70 = T6W - T6Z;
|
cannam@167
|
3222 T6H = W[106];
|
cannam@167
|
3223 T6T = W[107];
|
cannam@167
|
3224 cr[WS(rs, 54)] = FNMS(T6T, T70, T6H * T6S);
|
cannam@167
|
3225 ci[WS(rs, 54)] = FMA(T6T, T6S, T6H * T70);
|
cannam@167
|
3226 }
|
cannam@167
|
3227 {
|
cannam@167
|
3228 E T7e, T7g, T7d, T7f;
|
cannam@167
|
3229 T7e = T76 + T77;
|
cannam@167
|
3230 T7g = T7a + T7b;
|
cannam@167
|
3231 T7d = W[10];
|
cannam@167
|
3232 T7f = W[11];
|
cannam@167
|
3233 cr[WS(rs, 6)] = FNMS(T7f, T7g, T7d * T7e);
|
cannam@167
|
3234 ci[WS(rs, 6)] = FMA(T7d, T7g, T7f * T7e);
|
cannam@167
|
3235 }
|
cannam@167
|
3236 {
|
cannam@167
|
3237 E T72, T74, T71, T73;
|
cannam@167
|
3238 T72 = T6K + T6R;
|
cannam@167
|
3239 T74 = T6W + T6Z;
|
cannam@167
|
3240 T71 = W[42];
|
cannam@167
|
3241 T73 = W[43];
|
cannam@167
|
3242 cr[WS(rs, 22)] = FNMS(T73, T74, T71 * T72);
|
cannam@167
|
3243 ci[WS(rs, 22)] = FMA(T73, T72, T71 * T74);
|
cannam@167
|
3244 }
|
cannam@167
|
3245 {
|
cannam@167
|
3246 E T78, T7c, T75, T79;
|
cannam@167
|
3247 T78 = T76 - T77;
|
cannam@167
|
3248 T7c = T7a - T7b;
|
cannam@167
|
3249 T75 = W[74];
|
cannam@167
|
3250 T79 = W[75];
|
cannam@167
|
3251 cr[WS(rs, 38)] = FNMS(T79, T7c, T75 * T78);
|
cannam@167
|
3252 ci[WS(rs, 38)] = FMA(T75, T7c, T79 * T78);
|
cannam@167
|
3253 }
|
cannam@167
|
3254 }
|
cannam@167
|
3255 {
|
cannam@167
|
3256 E T3a, T52, T4S, T56, T4z, T57, T4V, T53;
|
cannam@167
|
3257 {
|
cannam@167
|
3258 E T2w, T39, T4O, T4R;
|
cannam@167
|
3259 T2w = T2c - T2v;
|
cannam@167
|
3260 T39 = T2P - T38;
|
cannam@167
|
3261 T3a = T2w + T39;
|
cannam@167
|
3262 T52 = T2w - T39;
|
cannam@167
|
3263 T4O = T4K - T4N;
|
cannam@167
|
3264 T4R = T4P - T4Q;
|
cannam@167
|
3265 T4S = T4O + T4R;
|
cannam@167
|
3266 T56 = T4O - T4R;
|
cannam@167
|
3267 {
|
cannam@167
|
3268 E T3R, T4T, T4y, T4U;
|
cannam@167
|
3269 {
|
cannam@167
|
3270 E T3D, T3Q, T4k, T4x;
|
cannam@167
|
3271 T3D = T3j - T3C;
|
cannam@167
|
3272 T3Q = T3M - T3P;
|
cannam@167
|
3273 T3R = FNMS(KP831469612, T3Q, KP555570233 * T3D);
|
cannam@167
|
3274 T4T = FMA(KP831469612, T3D, KP555570233 * T3Q);
|
cannam@167
|
3275 T4k = T40 - T4j;
|
cannam@167
|
3276 T4x = T4t - T4w;
|
cannam@167
|
3277 T4y = FMA(KP555570233, T4k, KP831469612 * T4x);
|
cannam@167
|
3278 T4U = FNMS(KP831469612, T4k, KP555570233 * T4x);
|
cannam@167
|
3279 }
|
cannam@167
|
3280 T4z = T3R + T4y;
|
cannam@167
|
3281 T57 = T3R - T4y;
|
cannam@167
|
3282 T4V = T4T + T4U;
|
cannam@167
|
3283 T53 = T4U - T4T;
|
cannam@167
|
3284 }
|
cannam@167
|
3285 }
|
cannam@167
|
3286 {
|
cannam@167
|
3287 E T4A, T4W, T23, T4B;
|
cannam@167
|
3288 T4A = T3a - T4z;
|
cannam@167
|
3289 T4W = T4S - T4V;
|
cannam@167
|
3290 T23 = W[82];
|
cannam@167
|
3291 T4B = W[83];
|
cannam@167
|
3292 cr[WS(rs, 42)] = FNMS(T4B, T4W, T23 * T4A);
|
cannam@167
|
3293 ci[WS(rs, 42)] = FMA(T4B, T4A, T23 * T4W);
|
cannam@167
|
3294 }
|
cannam@167
|
3295 {
|
cannam@167
|
3296 E T5a, T5c, T59, T5b;
|
cannam@167
|
3297 T5a = T52 + T53;
|
cannam@167
|
3298 T5c = T56 + T57;
|
cannam@167
|
3299 T59 = W[50];
|
cannam@167
|
3300 T5b = W[51];
|
cannam@167
|
3301 cr[WS(rs, 26)] = FNMS(T5b, T5c, T59 * T5a);
|
cannam@167
|
3302 ci[WS(rs, 26)] = FMA(T59, T5c, T5b * T5a);
|
cannam@167
|
3303 }
|
cannam@167
|
3304 {
|
cannam@167
|
3305 E T4Y, T50, T4X, T4Z;
|
cannam@167
|
3306 T4Y = T3a + T4z;
|
cannam@167
|
3307 T50 = T4S + T4V;
|
cannam@167
|
3308 T4X = W[18];
|
cannam@167
|
3309 T4Z = W[19];
|
cannam@167
|
3310 cr[WS(rs, 10)] = FNMS(T4Z, T50, T4X * T4Y);
|
cannam@167
|
3311 ci[WS(rs, 10)] = FMA(T4Z, T4Y, T4X * T50);
|
cannam@167
|
3312 }
|
cannam@167
|
3313 {
|
cannam@167
|
3314 E T54, T58, T51, T55;
|
cannam@167
|
3315 T54 = T52 - T53;
|
cannam@167
|
3316 T58 = T56 - T57;
|
cannam@167
|
3317 T51 = W[114];
|
cannam@167
|
3318 T55 = W[115];
|
cannam@167
|
3319 cr[WS(rs, 58)] = FNMS(T55, T58, T51 * T54);
|
cannam@167
|
3320 ci[WS(rs, 58)] = FMA(T51, T58, T55 * T54);
|
cannam@167
|
3321 }
|
cannam@167
|
3322 }
|
cannam@167
|
3323 {
|
cannam@167
|
3324 E T5g, T5C, T5s, T5G, T5n, T5H, T5v, T5D;
|
cannam@167
|
3325 {
|
cannam@167
|
3326 E T5e, T5f, T5q, T5r;
|
cannam@167
|
3327 T5e = T2c + T2v;
|
cannam@167
|
3328 T5f = T4P + T4Q;
|
cannam@167
|
3329 T5g = T5e + T5f;
|
cannam@167
|
3330 T5C = T5e - T5f;
|
cannam@167
|
3331 T5q = T4K + T4N;
|
cannam@167
|
3332 T5r = T38 + T2P;
|
cannam@167
|
3333 T5s = T5q + T5r;
|
cannam@167
|
3334 T5G = T5q - T5r;
|
cannam@167
|
3335 {
|
cannam@167
|
3336 E T5j, T5t, T5m, T5u;
|
cannam@167
|
3337 {
|
cannam@167
|
3338 E T5h, T5i, T5k, T5l;
|
cannam@167
|
3339 T5h = T3j + T3C;
|
cannam@167
|
3340 T5i = T3M + T3P;
|
cannam@167
|
3341 T5j = FNMS(KP195090322, T5i, KP980785280 * T5h);
|
cannam@167
|
3342 T5t = FMA(KP195090322, T5h, KP980785280 * T5i);
|
cannam@167
|
3343 T5k = T40 + T4j;
|
cannam@167
|
3344 T5l = T4t + T4w;
|
cannam@167
|
3345 T5m = FMA(KP980785280, T5k, KP195090322 * T5l);
|
cannam@167
|
3346 T5u = FNMS(KP195090322, T5k, KP980785280 * T5l);
|
cannam@167
|
3347 }
|
cannam@167
|
3348 T5n = T5j + T5m;
|
cannam@167
|
3349 T5H = T5j - T5m;
|
cannam@167
|
3350 T5v = T5t + T5u;
|
cannam@167
|
3351 T5D = T5u - T5t;
|
cannam@167
|
3352 }
|
cannam@167
|
3353 }
|
cannam@167
|
3354 {
|
cannam@167
|
3355 E T5o, T5w, T5d, T5p;
|
cannam@167
|
3356 T5o = T5g - T5n;
|
cannam@167
|
3357 T5w = T5s - T5v;
|
cannam@167
|
3358 T5d = W[66];
|
cannam@167
|
3359 T5p = W[67];
|
cannam@167
|
3360 cr[WS(rs, 34)] = FNMS(T5p, T5w, T5d * T5o);
|
cannam@167
|
3361 ci[WS(rs, 34)] = FMA(T5p, T5o, T5d * T5w);
|
cannam@167
|
3362 }
|
cannam@167
|
3363 {
|
cannam@167
|
3364 E T5K, T5M, T5J, T5L;
|
cannam@167
|
3365 T5K = T5C + T5D;
|
cannam@167
|
3366 T5M = T5G + T5H;
|
cannam@167
|
3367 T5J = W[34];
|
cannam@167
|
3368 T5L = W[35];
|
cannam@167
|
3369 cr[WS(rs, 18)] = FNMS(T5L, T5M, T5J * T5K);
|
cannam@167
|
3370 ci[WS(rs, 18)] = FMA(T5J, T5M, T5L * T5K);
|
cannam@167
|
3371 }
|
cannam@167
|
3372 {
|
cannam@167
|
3373 E T5y, T5A, T5x, T5z;
|
cannam@167
|
3374 T5y = T5g + T5n;
|
cannam@167
|
3375 T5A = T5s + T5v;
|
cannam@167
|
3376 T5x = W[2];
|
cannam@167
|
3377 T5z = W[3];
|
cannam@167
|
3378 cr[WS(rs, 2)] = FNMS(T5z, T5A, T5x * T5y);
|
cannam@167
|
3379 ci[WS(rs, 2)] = FMA(T5z, T5y, T5x * T5A);
|
cannam@167
|
3380 }
|
cannam@167
|
3381 {
|
cannam@167
|
3382 E T5E, T5I, T5B, T5F;
|
cannam@167
|
3383 T5E = T5C - T5D;
|
cannam@167
|
3384 T5I = T5G - T5H;
|
cannam@167
|
3385 T5B = W[98];
|
cannam@167
|
3386 T5F = W[99];
|
cannam@167
|
3387 cr[WS(rs, 50)] = FNMS(T5F, T5I, T5B * T5E);
|
cannam@167
|
3388 ci[WS(rs, 50)] = FMA(T5B, T5I, T5F * T5E);
|
cannam@167
|
3389 }
|
cannam@167
|
3390 }
|
cannam@167
|
3391 {
|
cannam@167
|
3392 E T5Y, T6w, T6m, T6A, T6d, T6B, T6p, T6x;
|
cannam@167
|
3393 {
|
cannam@167
|
3394 E T5Q, T5X, T6i, T6l;
|
cannam@167
|
3395 T5Q = T5O - T5P;
|
cannam@167
|
3396 T5X = T5T - T5W;
|
cannam@167
|
3397 T5Y = T5Q - T5X;
|
cannam@167
|
3398 T6w = T5Q + T5X;
|
cannam@167
|
3399 T6i = T6g - T6h;
|
cannam@167
|
3400 T6l = T6j - T6k;
|
cannam@167
|
3401 T6m = T6i - T6l;
|
cannam@167
|
3402 T6A = T6i + T6l;
|
cannam@167
|
3403 {
|
cannam@167
|
3404 E T65, T6o, T6c, T6n;
|
cannam@167
|
3405 {
|
cannam@167
|
3406 E T61, T64, T68, T6b;
|
cannam@167
|
3407 T61 = T5Z - T60;
|
cannam@167
|
3408 T64 = T62 - T63;
|
cannam@167
|
3409 T65 = FNMS(KP980785280, T64, KP195090322 * T61);
|
cannam@167
|
3410 T6o = FMA(KP980785280, T61, KP195090322 * T64);
|
cannam@167
|
3411 T68 = T66 - T67;
|
cannam@167
|
3412 T6b = T69 - T6a;
|
cannam@167
|
3413 T6c = FMA(KP195090322, T68, KP980785280 * T6b);
|
cannam@167
|
3414 T6n = FNMS(KP980785280, T68, KP195090322 * T6b);
|
cannam@167
|
3415 }
|
cannam@167
|
3416 T6d = T65 - T6c;
|
cannam@167
|
3417 T6B = T6c + T65;
|
cannam@167
|
3418 T6p = T6n - T6o;
|
cannam@167
|
3419 T6x = T6n + T6o;
|
cannam@167
|
3420 }
|
cannam@167
|
3421 }
|
cannam@167
|
3422 {
|
cannam@167
|
3423 E T6e, T6q, T5N, T6f;
|
cannam@167
|
3424 T6e = T5Y - T6d;
|
cannam@167
|
3425 T6q = T6m - T6p;
|
cannam@167
|
3426 T5N = W[122];
|
cannam@167
|
3427 T6f = W[123];
|
cannam@167
|
3428 cr[WS(rs, 62)] = FNMS(T6f, T6q, T5N * T6e);
|
cannam@167
|
3429 ci[WS(rs, 62)] = FMA(T6f, T6e, T5N * T6q);
|
cannam@167
|
3430 }
|
cannam@167
|
3431 {
|
cannam@167
|
3432 E T6E, T6G, T6D, T6F;
|
cannam@167
|
3433 T6E = T6w + T6x;
|
cannam@167
|
3434 T6G = T6A + T6B;
|
cannam@167
|
3435 T6D = W[26];
|
cannam@167
|
3436 T6F = W[27];
|
cannam@167
|
3437 cr[WS(rs, 14)] = FNMS(T6F, T6G, T6D * T6E);
|
cannam@167
|
3438 ci[WS(rs, 14)] = FMA(T6D, T6G, T6F * T6E);
|
cannam@167
|
3439 }
|
cannam@167
|
3440 {
|
cannam@167
|
3441 E T6s, T6u, T6r, T6t;
|
cannam@167
|
3442 T6s = T5Y + T6d;
|
cannam@167
|
3443 T6u = T6m + T6p;
|
cannam@167
|
3444 T6r = W[58];
|
cannam@167
|
3445 T6t = W[59];
|
cannam@167
|
3446 cr[WS(rs, 30)] = FNMS(T6t, T6u, T6r * T6s);
|
cannam@167
|
3447 ci[WS(rs, 30)] = FMA(T6t, T6s, T6r * T6u);
|
cannam@167
|
3448 }
|
cannam@167
|
3449 {
|
cannam@167
|
3450 E T6y, T6C, T6v, T6z;
|
cannam@167
|
3451 T6y = T6w - T6x;
|
cannam@167
|
3452 T6C = T6A - T6B;
|
cannam@167
|
3453 T6v = W[90];
|
cannam@167
|
3454 T6z = W[91];
|
cannam@167
|
3455 cr[WS(rs, 46)] = FNMS(T6z, T6C, T6v * T6y);
|
cannam@167
|
3456 ci[WS(rs, 46)] = FMA(T6v, T6C, T6z * T6y);
|
cannam@167
|
3457 }
|
cannam@167
|
3458 }
|
cannam@167
|
3459 {
|
cannam@167
|
3460 E Tba, Tdw, TcS, Tdi, TcI, Tds, TcW, Td6, Tcr, TcX, TcL, TcT, Tdd, Tdx, Tdl;
|
cannam@167
|
3461 E Tdt;
|
cannam@167
|
3462 {
|
cannam@167
|
3463 E Taw, Tdg, Tb9, Tdh, TaP, Tb8;
|
cannam@167
|
3464 Taw = Tag - Tav;
|
cannam@167
|
3465 Tdg = TcA + TcD;
|
cannam@167
|
3466 TaP = FNMS(KP831469612, TaO, KP555570233 * TaH);
|
cannam@167
|
3467 Tb8 = FMA(KP831469612, Tb0, KP555570233 * Tb7);
|
cannam@167
|
3468 Tb9 = TaP - Tb8;
|
cannam@167
|
3469 Tdh = Tb8 + TaP;
|
cannam@167
|
3470 Tba = Taw + Tb9;
|
cannam@167
|
3471 Tdw = Tdg - Tdh;
|
cannam@167
|
3472 TcS = Taw - Tb9;
|
cannam@167
|
3473 Tdi = Tdg + Tdh;
|
cannam@167
|
3474 }
|
cannam@167
|
3475 {
|
cannam@167
|
3476 E TcE, Td4, TcH, Td5, TcF, TcG;
|
cannam@167
|
3477 TcE = TcA - TcD;
|
cannam@167
|
3478 Td4 = Tag + Tav;
|
cannam@167
|
3479 TcF = FNMS(KP831469612, Tb7, KP555570233 * Tb0);
|
cannam@167
|
3480 TcG = FMA(KP555570233, TaO, KP831469612 * TaH);
|
cannam@167
|
3481 TcH = TcF - TcG;
|
cannam@167
|
3482 Td5 = TcF + TcG;
|
cannam@167
|
3483 TcI = TcE + TcH;
|
cannam@167
|
3484 Tds = Td4 - Td5;
|
cannam@167
|
3485 TcW = TcE - TcH;
|
cannam@167
|
3486 Td6 = Td4 + Td5;
|
cannam@167
|
3487 }
|
cannam@167
|
3488 {
|
cannam@167
|
3489 E TbN, TcJ, Tcq, TcK;
|
cannam@167
|
3490 {
|
cannam@167
|
3491 E TbB, TbM, Tce, Tcp;
|
cannam@167
|
3492 TbB = Tbl - TbA;
|
cannam@167
|
3493 TbM = TbI - TbL;
|
cannam@167
|
3494 TbN = FNMS(KP956940335, TbM, KP290284677 * TbB);
|
cannam@167
|
3495 TcJ = FMA(KP956940335, TbB, KP290284677 * TbM);
|
cannam@167
|
3496 Tce = TbY - Tcd;
|
cannam@167
|
3497 Tcp = Tcl - Tco;
|
cannam@167
|
3498 Tcq = FMA(KP290284677, Tce, KP956940335 * Tcp);
|
cannam@167
|
3499 TcK = FNMS(KP956940335, Tce, KP290284677 * Tcp);
|
cannam@167
|
3500 }
|
cannam@167
|
3501 Tcr = TbN + Tcq;
|
cannam@167
|
3502 TcX = TbN - Tcq;
|
cannam@167
|
3503 TcL = TcJ + TcK;
|
cannam@167
|
3504 TcT = TcK - TcJ;
|
cannam@167
|
3505 }
|
cannam@167
|
3506 {
|
cannam@167
|
3507 E Td9, Tdj, Tdc, Tdk;
|
cannam@167
|
3508 {
|
cannam@167
|
3509 E Td7, Td8, Tda, Tdb;
|
cannam@167
|
3510 Td7 = Tbl + TbA;
|
cannam@167
|
3511 Td8 = TbI + TbL;
|
cannam@167
|
3512 Td9 = FNMS(KP471396736, Td8, KP881921264 * Td7);
|
cannam@167
|
3513 Tdj = FMA(KP471396736, Td7, KP881921264 * Td8);
|
cannam@167
|
3514 Tda = TbY + Tcd;
|
cannam@167
|
3515 Tdb = Tcl + Tco;
|
cannam@167
|
3516 Tdc = FMA(KP881921264, Tda, KP471396736 * Tdb);
|
cannam@167
|
3517 Tdk = FNMS(KP471396736, Tda, KP881921264 * Tdb);
|
cannam@167
|
3518 }
|
cannam@167
|
3519 Tdd = Td9 + Tdc;
|
cannam@167
|
3520 Tdx = Td9 - Tdc;
|
cannam@167
|
3521 Tdl = Tdj + Tdk;
|
cannam@167
|
3522 Tdt = Tdk - Tdj;
|
cannam@167
|
3523 }
|
cannam@167
|
3524 {
|
cannam@167
|
3525 E Tcs, TcM, Ta5, Tct;
|
cannam@167
|
3526 Tcs = Tba - Tcr;
|
cannam@167
|
3527 TcM = TcI - TcL;
|
cannam@167
|
3528 Ta5 = W[88];
|
cannam@167
|
3529 Tct = W[89];
|
cannam@167
|
3530 cr[WS(rs, 45)] = FNMS(Tct, TcM, Ta5 * Tcs);
|
cannam@167
|
3531 ci[WS(rs, 45)] = FMA(Tct, Tcs, Ta5 * TcM);
|
cannam@167
|
3532 }
|
cannam@167
|
3533 {
|
cannam@167
|
3534 E Tdu, Tdy, Tdr, Tdv;
|
cannam@167
|
3535 Tdu = Tds - Tdt;
|
cannam@167
|
3536 Tdy = Tdw - Tdx;
|
cannam@167
|
3537 Tdr = W[104];
|
cannam@167
|
3538 Tdv = W[105];
|
cannam@167
|
3539 cr[WS(rs, 53)] = FNMS(Tdv, Tdy, Tdr * Tdu);
|
cannam@167
|
3540 ci[WS(rs, 53)] = FMA(Tdr, Tdy, Tdv * Tdu);
|
cannam@167
|
3541 }
|
cannam@167
|
3542 {
|
cannam@167
|
3543 E TdA, TdC, Tdz, TdB;
|
cannam@167
|
3544 TdA = Tds + Tdt;
|
cannam@167
|
3545 TdC = Tdw + Tdx;
|
cannam@167
|
3546 Tdz = W[40];
|
cannam@167
|
3547 TdB = W[41];
|
cannam@167
|
3548 cr[WS(rs, 21)] = FNMS(TdB, TdC, Tdz * TdA);
|
cannam@167
|
3549 ci[WS(rs, 21)] = FMA(Tdz, TdC, TdB * TdA);
|
cannam@167
|
3550 }
|
cannam@167
|
3551 {
|
cannam@167
|
3552 E TcO, TcQ, TcN, TcP;
|
cannam@167
|
3553 TcO = Tba + Tcr;
|
cannam@167
|
3554 TcQ = TcI + TcL;
|
cannam@167
|
3555 TcN = W[24];
|
cannam@167
|
3556 TcP = W[25];
|
cannam@167
|
3557 cr[WS(rs, 13)] = FNMS(TcP, TcQ, TcN * TcO);
|
cannam@167
|
3558 ci[WS(rs, 13)] = FMA(TcP, TcO, TcN * TcQ);
|
cannam@167
|
3559 }
|
cannam@167
|
3560 {
|
cannam@167
|
3561 E TcU, TcY, TcR, TcV;
|
cannam@167
|
3562 TcU = TcS - TcT;
|
cannam@167
|
3563 TcY = TcW - TcX;
|
cannam@167
|
3564 TcR = W[120];
|
cannam@167
|
3565 TcV = W[121];
|
cannam@167
|
3566 cr[WS(rs, 61)] = FNMS(TcV, TcY, TcR * TcU);
|
cannam@167
|
3567 ci[WS(rs, 61)] = FMA(TcR, TcY, TcV * TcU);
|
cannam@167
|
3568 }
|
cannam@167
|
3569 {
|
cannam@167
|
3570 E Tde, Tdm, Td3, Tdf;
|
cannam@167
|
3571 Tde = Td6 - Tdd;
|
cannam@167
|
3572 Tdm = Tdi - Tdl;
|
cannam@167
|
3573 Td3 = W[72];
|
cannam@167
|
3574 Tdf = W[73];
|
cannam@167
|
3575 cr[WS(rs, 37)] = FNMS(Tdf, Tdm, Td3 * Tde);
|
cannam@167
|
3576 ci[WS(rs, 37)] = FMA(Tdf, Tde, Td3 * Tdm);
|
cannam@167
|
3577 }
|
cannam@167
|
3578 {
|
cannam@167
|
3579 E Tdo, Tdq, Tdn, Tdp;
|
cannam@167
|
3580 Tdo = Td6 + Tdd;
|
cannam@167
|
3581 Tdq = Tdi + Tdl;
|
cannam@167
|
3582 Tdn = W[8];
|
cannam@167
|
3583 Tdp = W[9];
|
cannam@167
|
3584 cr[WS(rs, 5)] = FNMS(Tdp, Tdq, Tdn * Tdo);
|
cannam@167
|
3585 ci[WS(rs, 5)] = FMA(Tdp, Tdo, Tdn * Tdq);
|
cannam@167
|
3586 }
|
cannam@167
|
3587 {
|
cannam@167
|
3588 E Td0, Td2, TcZ, Td1;
|
cannam@167
|
3589 Td0 = TcS + TcT;
|
cannam@167
|
3590 Td2 = TcW + TcX;
|
cannam@167
|
3591 TcZ = W[56];
|
cannam@167
|
3592 Td1 = W[57];
|
cannam@167
|
3593 cr[WS(rs, 29)] = FNMS(Td1, Td2, TcZ * Td0);
|
cannam@167
|
3594 ci[WS(rs, 29)] = FMA(TcZ, Td2, Td1 * Td0);
|
cannam@167
|
3595 }
|
cannam@167
|
3596 }
|
cannam@167
|
3597 {
|
cannam@167
|
3598 E Tfy, Thc, Tgy, TgY, Tgo, Th8, TgC, TgM, Tgb, TgD, Tgr, Tgz, TgT, Thd, Th1;
|
cannam@167
|
3599 E Th9;
|
cannam@167
|
3600 {
|
cannam@167
|
3601 E Tfi, TgW, Tfx, TgX, Tfp, Tfw;
|
cannam@167
|
3602 Tfi = Tfa - Tfh;
|
cannam@167
|
3603 TgW = Tgg + Tgj;
|
cannam@167
|
3604 Tfp = FNMS(KP555570233, Tfo, KP831469612 * Tfl);
|
cannam@167
|
3605 Tfw = FMA(KP831469612, Tfs, KP555570233 * Tfv);
|
cannam@167
|
3606 Tfx = Tfp - Tfw;
|
cannam@167
|
3607 TgX = Tfw + Tfp;
|
cannam@167
|
3608 Tfy = Tfi + Tfx;
|
cannam@167
|
3609 Thc = TgW - TgX;
|
cannam@167
|
3610 Tgy = Tfi - Tfx;
|
cannam@167
|
3611 TgY = TgW + TgX;
|
cannam@167
|
3612 }
|
cannam@167
|
3613 {
|
cannam@167
|
3614 E Tgk, TgK, Tgn, TgL, Tgl, Tgm;
|
cannam@167
|
3615 Tgk = Tgg - Tgj;
|
cannam@167
|
3616 TgK = Tfa + Tfh;
|
cannam@167
|
3617 Tgl = FNMS(KP555570233, Tfs, KP831469612 * Tfv);
|
cannam@167
|
3618 Tgm = FMA(KP555570233, Tfl, KP831469612 * Tfo);
|
cannam@167
|
3619 Tgn = Tgl - Tgm;
|
cannam@167
|
3620 TgL = Tgl + Tgm;
|
cannam@167
|
3621 Tgo = Tgk + Tgn;
|
cannam@167
|
3622 Th8 = TgK - TgL;
|
cannam@167
|
3623 TgC = Tgk - Tgn;
|
cannam@167
|
3624 TgM = TgK + TgL;
|
cannam@167
|
3625 }
|
cannam@167
|
3626 {
|
cannam@167
|
3627 E TfR, Tgp, Tga, Tgq;
|
cannam@167
|
3628 {
|
cannam@167
|
3629 E TfJ, TfQ, Tg2, Tg9;
|
cannam@167
|
3630 TfJ = TfB - TfI;
|
cannam@167
|
3631 TfQ = TfM - TfP;
|
cannam@167
|
3632 TfR = FNMS(KP881921264, TfQ, KP471396736 * TfJ);
|
cannam@167
|
3633 Tgp = FMA(KP881921264, TfJ, KP471396736 * TfQ);
|
cannam@167
|
3634 Tg2 = TfU - Tg1;
|
cannam@167
|
3635 Tg9 = Tg5 - Tg8;
|
cannam@167
|
3636 Tga = FMA(KP471396736, Tg2, KP881921264 * Tg9);
|
cannam@167
|
3637 Tgq = FNMS(KP881921264, Tg2, KP471396736 * Tg9);
|
cannam@167
|
3638 }
|
cannam@167
|
3639 Tgb = TfR + Tga;
|
cannam@167
|
3640 TgD = TfR - Tga;
|
cannam@167
|
3641 Tgr = Tgp + Tgq;
|
cannam@167
|
3642 Tgz = Tgq - Tgp;
|
cannam@167
|
3643 }
|
cannam@167
|
3644 {
|
cannam@167
|
3645 E TgP, TgZ, TgS, Th0;
|
cannam@167
|
3646 {
|
cannam@167
|
3647 E TgN, TgO, TgQ, TgR;
|
cannam@167
|
3648 TgN = TfB + TfI;
|
cannam@167
|
3649 TgO = TfM + TfP;
|
cannam@167
|
3650 TgP = FNMS(KP290284677, TgO, KP956940335 * TgN);
|
cannam@167
|
3651 TgZ = FMA(KP290284677, TgN, KP956940335 * TgO);
|
cannam@167
|
3652 TgQ = TfU + Tg1;
|
cannam@167
|
3653 TgR = Tg5 + Tg8;
|
cannam@167
|
3654 TgS = FMA(KP956940335, TgQ, KP290284677 * TgR);
|
cannam@167
|
3655 Th0 = FNMS(KP290284677, TgQ, KP956940335 * TgR);
|
cannam@167
|
3656 }
|
cannam@167
|
3657 TgT = TgP + TgS;
|
cannam@167
|
3658 Thd = TgP - TgS;
|
cannam@167
|
3659 Th1 = TgZ + Th0;
|
cannam@167
|
3660 Th9 = Th0 - TgZ;
|
cannam@167
|
3661 }
|
cannam@167
|
3662 {
|
cannam@167
|
3663 E Tgc, Tgs, Tf7, Tgd;
|
cannam@167
|
3664 Tgc = Tfy - Tgb;
|
cannam@167
|
3665 Tgs = Tgo - Tgr;
|
cannam@167
|
3666 Tf7 = W[84];
|
cannam@167
|
3667 Tgd = W[85];
|
cannam@167
|
3668 cr[WS(rs, 43)] = FNMS(Tgd, Tgs, Tf7 * Tgc);
|
cannam@167
|
3669 ci[WS(rs, 43)] = FMA(Tgd, Tgc, Tf7 * Tgs);
|
cannam@167
|
3670 }
|
cannam@167
|
3671 {
|
cannam@167
|
3672 E Tha, The, Th7, Thb;
|
cannam@167
|
3673 Tha = Th8 - Th9;
|
cannam@167
|
3674 The = Thc - Thd;
|
cannam@167
|
3675 Th7 = W[100];
|
cannam@167
|
3676 Thb = W[101];
|
cannam@167
|
3677 cr[WS(rs, 51)] = FNMS(Thb, The, Th7 * Tha);
|
cannam@167
|
3678 ci[WS(rs, 51)] = FMA(Th7, The, Thb * Tha);
|
cannam@167
|
3679 }
|
cannam@167
|
3680 {
|
cannam@167
|
3681 E Thg, Thi, Thf, Thh;
|
cannam@167
|
3682 Thg = Th8 + Th9;
|
cannam@167
|
3683 Thi = Thc + Thd;
|
cannam@167
|
3684 Thf = W[36];
|
cannam@167
|
3685 Thh = W[37];
|
cannam@167
|
3686 cr[WS(rs, 19)] = FNMS(Thh, Thi, Thf * Thg);
|
cannam@167
|
3687 ci[WS(rs, 19)] = FMA(Thf, Thi, Thh * Thg);
|
cannam@167
|
3688 }
|
cannam@167
|
3689 {
|
cannam@167
|
3690 E Tgu, Tgw, Tgt, Tgv;
|
cannam@167
|
3691 Tgu = Tfy + Tgb;
|
cannam@167
|
3692 Tgw = Tgo + Tgr;
|
cannam@167
|
3693 Tgt = W[20];
|
cannam@167
|
3694 Tgv = W[21];
|
cannam@167
|
3695 cr[WS(rs, 11)] = FNMS(Tgv, Tgw, Tgt * Tgu);
|
cannam@167
|
3696 ci[WS(rs, 11)] = FMA(Tgv, Tgu, Tgt * Tgw);
|
cannam@167
|
3697 }
|
cannam@167
|
3698 {
|
cannam@167
|
3699 E TgA, TgE, Tgx, TgB;
|
cannam@167
|
3700 TgA = Tgy - Tgz;
|
cannam@167
|
3701 TgE = TgC - TgD;
|
cannam@167
|
3702 Tgx = W[116];
|
cannam@167
|
3703 TgB = W[117];
|
cannam@167
|
3704 cr[WS(rs, 59)] = FNMS(TgB, TgE, Tgx * TgA);
|
cannam@167
|
3705 ci[WS(rs, 59)] = FMA(Tgx, TgE, TgB * TgA);
|
cannam@167
|
3706 }
|
cannam@167
|
3707 {
|
cannam@167
|
3708 E TgU, Th2, TgJ, TgV;
|
cannam@167
|
3709 TgU = TgM - TgT;
|
cannam@167
|
3710 Th2 = TgY - Th1;
|
cannam@167
|
3711 TgJ = W[68];
|
cannam@167
|
3712 TgV = W[69];
|
cannam@167
|
3713 cr[WS(rs, 35)] = FNMS(TgV, Th2, TgJ * TgU);
|
cannam@167
|
3714 ci[WS(rs, 35)] = FMA(TgV, TgU, TgJ * Th2);
|
cannam@167
|
3715 }
|
cannam@167
|
3716 {
|
cannam@167
|
3717 E Th4, Th6, Th3, Th5;
|
cannam@167
|
3718 Th4 = TgM + TgT;
|
cannam@167
|
3719 Th6 = TgY + Th1;
|
cannam@167
|
3720 Th3 = W[4];
|
cannam@167
|
3721 Th5 = W[5];
|
cannam@167
|
3722 cr[WS(rs, 3)] = FNMS(Th5, Th6, Th3 * Th4);
|
cannam@167
|
3723 ci[WS(rs, 3)] = FMA(Th5, Th4, Th3 * Th6);
|
cannam@167
|
3724 }
|
cannam@167
|
3725 {
|
cannam@167
|
3726 E TgG, TgI, TgF, TgH;
|
cannam@167
|
3727 TgG = Tgy + Tgz;
|
cannam@167
|
3728 TgI = TgC + TgD;
|
cannam@167
|
3729 TgF = W[52];
|
cannam@167
|
3730 TgH = W[53];
|
cannam@167
|
3731 cr[WS(rs, 27)] = FNMS(TgH, TgI, TgF * TgG);
|
cannam@167
|
3732 ci[WS(rs, 27)] = FMA(TgF, TgI, TgH * TgG);
|
cannam@167
|
3733 }
|
cannam@167
|
3734 }
|
cannam@167
|
3735 {
|
cannam@167
|
3736 E TdO, Tf0, Tem, TeM, Tec, TeW, Teq, TeA, Te3, Ter, Tef, Ten, TeH, Tf1, TeP;
|
cannam@167
|
3737 E TeX;
|
cannam@167
|
3738 {
|
cannam@167
|
3739 E TdG, TeK, TdN, TeL, TdJ, TdM;
|
cannam@167
|
3740 TdG = TdE - TdF;
|
cannam@167
|
3741 TeK = Te6 + Te7;
|
cannam@167
|
3742 TdJ = FNMS(KP195090322, TdI, KP980785280 * TdH);
|
cannam@167
|
3743 TdM = FMA(KP195090322, TdK, KP980785280 * TdL);
|
cannam@167
|
3744 TdN = TdJ - TdM;
|
cannam@167
|
3745 TeL = TdM + TdJ;
|
cannam@167
|
3746 TdO = TdG - TdN;
|
cannam@167
|
3747 Tf0 = TeK + TeL;
|
cannam@167
|
3748 Tem = TdG + TdN;
|
cannam@167
|
3749 TeM = TeK - TeL;
|
cannam@167
|
3750 }
|
cannam@167
|
3751 {
|
cannam@167
|
3752 E Te8, Tey, Teb, Tez, Te9, Tea;
|
cannam@167
|
3753 Te8 = Te6 - Te7;
|
cannam@167
|
3754 Tey = TdE + TdF;
|
cannam@167
|
3755 Te9 = FNMS(KP195090322, TdL, KP980785280 * TdK);
|
cannam@167
|
3756 Tea = FMA(KP980785280, TdI, KP195090322 * TdH);
|
cannam@167
|
3757 Teb = Te9 - Tea;
|
cannam@167
|
3758 Tez = Te9 + Tea;
|
cannam@167
|
3759 Tec = Te8 - Teb;
|
cannam@167
|
3760 TeW = Tey + Tez;
|
cannam@167
|
3761 Teq = Te8 + Teb;
|
cannam@167
|
3762 TeA = Tey - Tez;
|
cannam@167
|
3763 }
|
cannam@167
|
3764 {
|
cannam@167
|
3765 E TdV, Tee, Te2, Ted;
|
cannam@167
|
3766 {
|
cannam@167
|
3767 E TdR, TdU, TdY, Te1;
|
cannam@167
|
3768 TdR = TdP - TdQ;
|
cannam@167
|
3769 TdU = TdS - TdT;
|
cannam@167
|
3770 TdV = FNMS(KP773010453, TdU, KP634393284 * TdR);
|
cannam@167
|
3771 Tee = FMA(KP773010453, TdR, KP634393284 * TdU);
|
cannam@167
|
3772 TdY = TdW - TdX;
|
cannam@167
|
3773 Te1 = TdZ - Te0;
|
cannam@167
|
3774 Te2 = FMA(KP634393284, TdY, KP773010453 * Te1);
|
cannam@167
|
3775 Ted = FNMS(KP773010453, TdY, KP634393284 * Te1);
|
cannam@167
|
3776 }
|
cannam@167
|
3777 Te3 = TdV - Te2;
|
cannam@167
|
3778 Ter = Te2 + TdV;
|
cannam@167
|
3779 Tef = Ted - Tee;
|
cannam@167
|
3780 Ten = Ted + Tee;
|
cannam@167
|
3781 }
|
cannam@167
|
3782 {
|
cannam@167
|
3783 E TeD, TeO, TeG, TeN;
|
cannam@167
|
3784 {
|
cannam@167
|
3785 E TeB, TeC, TeE, TeF;
|
cannam@167
|
3786 TeB = TdP + TdQ;
|
cannam@167
|
3787 TeC = TdS + TdT;
|
cannam@167
|
3788 TeD = FNMS(KP098017140, TeC, KP995184726 * TeB);
|
cannam@167
|
3789 TeO = FMA(KP098017140, TeB, KP995184726 * TeC);
|
cannam@167
|
3790 TeE = TdW + TdX;
|
cannam@167
|
3791 TeF = TdZ + Te0;
|
cannam@167
|
3792 TeG = FMA(KP995184726, TeE, KP098017140 * TeF);
|
cannam@167
|
3793 TeN = FNMS(KP098017140, TeE, KP995184726 * TeF);
|
cannam@167
|
3794 }
|
cannam@167
|
3795 TeH = TeD - TeG;
|
cannam@167
|
3796 Tf1 = TeG + TeD;
|
cannam@167
|
3797 TeP = TeN - TeO;
|
cannam@167
|
3798 TeX = TeN + TeO;
|
cannam@167
|
3799 }
|
cannam@167
|
3800 {
|
cannam@167
|
3801 E Te4, Teg, TdD, Te5;
|
cannam@167
|
3802 Te4 = TdO - Te3;
|
cannam@167
|
3803 Teg = Tec - Tef;
|
cannam@167
|
3804 TdD = W[112];
|
cannam@167
|
3805 Te5 = W[113];
|
cannam@167
|
3806 cr[WS(rs, 57)] = FNMS(Te5, Teg, TdD * Te4);
|
cannam@167
|
3807 ci[WS(rs, 57)] = FMA(Te5, Te4, TdD * Teg);
|
cannam@167
|
3808 }
|
cannam@167
|
3809 {
|
cannam@167
|
3810 E TeY, Tf2, TeV, TeZ;
|
cannam@167
|
3811 TeY = TeW - TeX;
|
cannam@167
|
3812 Tf2 = Tf0 - Tf1;
|
cannam@167
|
3813 TeV = W[64];
|
cannam@167
|
3814 TeZ = W[65];
|
cannam@167
|
3815 cr[WS(rs, 33)] = FNMS(TeZ, Tf2, TeV * TeY);
|
cannam@167
|
3816 ci[WS(rs, 33)] = FMA(TeV, Tf2, TeZ * TeY);
|
cannam@167
|
3817 }
|
cannam@167
|
3818 {
|
cannam@167
|
3819 E Tf4, Tf6, Tf3, Tf5;
|
cannam@167
|
3820 Tf4 = TeW + TeX;
|
cannam@167
|
3821 Tf6 = Tf0 + Tf1;
|
cannam@167
|
3822 Tf3 = W[0];
|
cannam@167
|
3823 Tf5 = W[1];
|
cannam@167
|
3824 cr[WS(rs, 1)] = FNMS(Tf5, Tf6, Tf3 * Tf4);
|
cannam@167
|
3825 ci[WS(rs, 1)] = FMA(Tf3, Tf6, Tf5 * Tf4);
|
cannam@167
|
3826 }
|
cannam@167
|
3827 {
|
cannam@167
|
3828 E Tei, Tek, Teh, Tej;
|
cannam@167
|
3829 Tei = TdO + Te3;
|
cannam@167
|
3830 Tek = Tec + Tef;
|
cannam@167
|
3831 Teh = W[48];
|
cannam@167
|
3832 Tej = W[49];
|
cannam@167
|
3833 cr[WS(rs, 25)] = FNMS(Tej, Tek, Teh * Tei);
|
cannam@167
|
3834 ci[WS(rs, 25)] = FMA(Tej, Tei, Teh * Tek);
|
cannam@167
|
3835 }
|
cannam@167
|
3836 {
|
cannam@167
|
3837 E Teo, Tes, Tel, Tep;
|
cannam@167
|
3838 Teo = Tem - Ten;
|
cannam@167
|
3839 Tes = Teq - Ter;
|
cannam@167
|
3840 Tel = W[80];
|
cannam@167
|
3841 Tep = W[81];
|
cannam@167
|
3842 cr[WS(rs, 41)] = FNMS(Tep, Tes, Tel * Teo);
|
cannam@167
|
3843 ci[WS(rs, 41)] = FMA(Tel, Tes, Tep * Teo);
|
cannam@167
|
3844 }
|
cannam@167
|
3845 {
|
cannam@167
|
3846 E TeI, TeQ, Tex, TeJ;
|
cannam@167
|
3847 TeI = TeA - TeH;
|
cannam@167
|
3848 TeQ = TeM - TeP;
|
cannam@167
|
3849 Tex = W[96];
|
cannam@167
|
3850 TeJ = W[97];
|
cannam@167
|
3851 cr[WS(rs, 49)] = FNMS(TeJ, TeQ, Tex * TeI);
|
cannam@167
|
3852 ci[WS(rs, 49)] = FMA(TeJ, TeI, Tex * TeQ);
|
cannam@167
|
3853 }
|
cannam@167
|
3854 {
|
cannam@167
|
3855 E TeS, TeU, TeR, TeT;
|
cannam@167
|
3856 TeS = TeA + TeH;
|
cannam@167
|
3857 TeU = TeM + TeP;
|
cannam@167
|
3858 TeR = W[32];
|
cannam@167
|
3859 TeT = W[33];
|
cannam@167
|
3860 cr[WS(rs, 17)] = FNMS(TeT, TeU, TeR * TeS);
|
cannam@167
|
3861 ci[WS(rs, 17)] = FMA(TeT, TeS, TeR * TeU);
|
cannam@167
|
3862 }
|
cannam@167
|
3863 {
|
cannam@167
|
3864 E Teu, Tew, Tet, Tev;
|
cannam@167
|
3865 Teu = Tem + Ten;
|
cannam@167
|
3866 Tew = Teq + Ter;
|
cannam@167
|
3867 Tet = W[16];
|
cannam@167
|
3868 Tev = W[17];
|
cannam@167
|
3869 cr[WS(rs, 9)] = FNMS(Tev, Tew, Tet * Teu);
|
cannam@167
|
3870 ci[WS(rs, 9)] = FMA(Tet, Tew, Tev * Teu);
|
cannam@167
|
3871 }
|
cannam@167
|
3872 }
|
cannam@167
|
3873 {
|
cannam@167
|
3874 E Thu, TiG, Ti2, Tis, ThS, TiC, Ti6, Tig, ThJ, Ti7, ThV, Ti3, Tin, TiH, Tiv;
|
cannam@167
|
3875 E TiD;
|
cannam@167
|
3876 {
|
cannam@167
|
3877 E Thm, Tiq, Tht, Tir, Thp, Ths;
|
cannam@167
|
3878 Thm = Thk - Thl;
|
cannam@167
|
3879 Tiq = ThM - ThN;
|
cannam@167
|
3880 Thp = FNMS(KP980785280, Tho, KP195090322 * Thn);
|
cannam@167
|
3881 Ths = FNMS(KP980785280, Thr, KP195090322 * Thq);
|
cannam@167
|
3882 Tht = Thp + Ths;
|
cannam@167
|
3883 Tir = Thp - Ths;
|
cannam@167
|
3884 Thu = Thm - Tht;
|
cannam@167
|
3885 TiG = Tiq - Tir;
|
cannam@167
|
3886 Ti2 = Thm + Tht;
|
cannam@167
|
3887 Tis = Tiq + Tir;
|
cannam@167
|
3888 }
|
cannam@167
|
3889 {
|
cannam@167
|
3890 E ThO, Tie, ThR, Tif, ThP, ThQ;
|
cannam@167
|
3891 ThO = ThM + ThN;
|
cannam@167
|
3892 Tie = Thk + Thl;
|
cannam@167
|
3893 ThP = FMA(KP195090322, Tho, KP980785280 * Thn);
|
cannam@167
|
3894 ThQ = FMA(KP195090322, Thr, KP980785280 * Thq);
|
cannam@167
|
3895 ThR = ThP - ThQ;
|
cannam@167
|
3896 Tif = ThP + ThQ;
|
cannam@167
|
3897 ThS = ThO - ThR;
|
cannam@167
|
3898 TiC = Tie + Tif;
|
cannam@167
|
3899 Ti6 = ThO + ThR;
|
cannam@167
|
3900 Tig = Tie - Tif;
|
cannam@167
|
3901 }
|
cannam@167
|
3902 {
|
cannam@167
|
3903 E ThB, ThU, ThI, ThT;
|
cannam@167
|
3904 {
|
cannam@167
|
3905 E Thx, ThA, ThE, ThH;
|
cannam@167
|
3906 Thx = Thv - Thw;
|
cannam@167
|
3907 ThA = Thy - Thz;
|
cannam@167
|
3908 ThB = FNMS(KP634393284, ThA, KP773010453 * Thx);
|
cannam@167
|
3909 ThU = FMA(KP634393284, Thx, KP773010453 * ThA);
|
cannam@167
|
3910 ThE = ThC + ThD;
|
cannam@167
|
3911 ThH = ThF - ThG;
|
cannam@167
|
3912 ThI = FMA(KP773010453, ThE, KP634393284 * ThH);
|
cannam@167
|
3913 ThT = FNMS(KP634393284, ThE, KP773010453 * ThH);
|
cannam@167
|
3914 }
|
cannam@167
|
3915 ThJ = ThB - ThI;
|
cannam@167
|
3916 Ti7 = ThI + ThB;
|
cannam@167
|
3917 ThV = ThT - ThU;
|
cannam@167
|
3918 Ti3 = ThT + ThU;
|
cannam@167
|
3919 }
|
cannam@167
|
3920 {
|
cannam@167
|
3921 E Tij, Tit, Tim, Tiu;
|
cannam@167
|
3922 {
|
cannam@167
|
3923 E Tih, Tii, Tik, Til;
|
cannam@167
|
3924 Tih = ThF + ThG;
|
cannam@167
|
3925 Tii = ThC - ThD;
|
cannam@167
|
3926 Tij = FNMS(KP995184726, Tii, KP098017140 * Tih);
|
cannam@167
|
3927 Tit = FMA(KP098017140, Tii, KP995184726 * Tih);
|
cannam@167
|
3928 Tik = Thy + Thz;
|
cannam@167
|
3929 Til = Thw + Thv;
|
cannam@167
|
3930 Tim = FNMS(KP995184726, Til, KP098017140 * Tik);
|
cannam@167
|
3931 Tiu = FMA(KP098017140, Til, KP995184726 * Tik);
|
cannam@167
|
3932 }
|
cannam@167
|
3933 Tin = Tij + Tim;
|
cannam@167
|
3934 TiH = Tij - Tim;
|
cannam@167
|
3935 Tiv = Tit - Tiu;
|
cannam@167
|
3936 TiD = Tit + Tiu;
|
cannam@167
|
3937 }
|
cannam@167
|
3938 {
|
cannam@167
|
3939 E ThK, ThW, Thj, ThL;
|
cannam@167
|
3940 ThK = Thu - ThJ;
|
cannam@167
|
3941 ThW = ThS - ThV;
|
cannam@167
|
3942 Thj = W[108];
|
cannam@167
|
3943 ThL = W[109];
|
cannam@167
|
3944 cr[WS(rs, 55)] = FNMS(ThL, ThW, Thj * ThK);
|
cannam@167
|
3945 ci[WS(rs, 55)] = FMA(ThL, ThK, Thj * ThW);
|
cannam@167
|
3946 }
|
cannam@167
|
3947 {
|
cannam@167
|
3948 E TiE, TiI, TiB, TiF;
|
cannam@167
|
3949 TiE = TiC - TiD;
|
cannam@167
|
3950 TiI = TiG + TiH;
|
cannam@167
|
3951 TiB = W[60];
|
cannam@167
|
3952 TiF = W[61];
|
cannam@167
|
3953 cr[WS(rs, 31)] = FNMS(TiF, TiI, TiB * TiE);
|
cannam@167
|
3954 ci[WS(rs, 31)] = FMA(TiB, TiI, TiF * TiE);
|
cannam@167
|
3955 }
|
cannam@167
|
3956 {
|
cannam@167
|
3957 E TiK, TiM, TiJ, TiL;
|
cannam@167
|
3958 TiK = TiC + TiD;
|
cannam@167
|
3959 TiM = TiG - TiH;
|
cannam@167
|
3960 TiJ = W[124];
|
cannam@167
|
3961 TiL = W[125];
|
cannam@167
|
3962 cr[WS(rs, 63)] = FNMS(TiL, TiM, TiJ * TiK);
|
cannam@167
|
3963 ci[WS(rs, 63)] = FMA(TiJ, TiM, TiL * TiK);
|
cannam@167
|
3964 }
|
cannam@167
|
3965 {
|
cannam@167
|
3966 E ThY, Ti0, ThX, ThZ;
|
cannam@167
|
3967 ThY = Thu + ThJ;
|
cannam@167
|
3968 Ti0 = ThS + ThV;
|
cannam@167
|
3969 ThX = W[44];
|
cannam@167
|
3970 ThZ = W[45];
|
cannam@167
|
3971 cr[WS(rs, 23)] = FNMS(ThZ, Ti0, ThX * ThY);
|
cannam@167
|
3972 ci[WS(rs, 23)] = FMA(ThZ, ThY, ThX * Ti0);
|
cannam@167
|
3973 }
|
cannam@167
|
3974 {
|
cannam@167
|
3975 E Ti4, Ti8, Ti1, Ti5;
|
cannam@167
|
3976 Ti4 = Ti2 - Ti3;
|
cannam@167
|
3977 Ti8 = Ti6 - Ti7;
|
cannam@167
|
3978 Ti1 = W[76];
|
cannam@167
|
3979 Ti5 = W[77];
|
cannam@167
|
3980 cr[WS(rs, 39)] = FNMS(Ti5, Ti8, Ti1 * Ti4);
|
cannam@167
|
3981 ci[WS(rs, 39)] = FMA(Ti1, Ti8, Ti5 * Ti4);
|
cannam@167
|
3982 }
|
cannam@167
|
3983 {
|
cannam@167
|
3984 E Tio, Tiw, Tid, Tip;
|
cannam@167
|
3985 Tio = Tig - Tin;
|
cannam@167
|
3986 Tiw = Tis - Tiv;
|
cannam@167
|
3987 Tid = W[92];
|
cannam@167
|
3988 Tip = W[93];
|
cannam@167
|
3989 cr[WS(rs, 47)] = FNMS(Tip, Tiw, Tid * Tio);
|
cannam@167
|
3990 ci[WS(rs, 47)] = FMA(Tip, Tio, Tid * Tiw);
|
cannam@167
|
3991 }
|
cannam@167
|
3992 {
|
cannam@167
|
3993 E Tiy, TiA, Tix, Tiz;
|
cannam@167
|
3994 Tiy = Tig + Tin;
|
cannam@167
|
3995 TiA = Tis + Tiv;
|
cannam@167
|
3996 Tix = W[28];
|
cannam@167
|
3997 Tiz = W[29];
|
cannam@167
|
3998 cr[WS(rs, 15)] = FNMS(Tiz, TiA, Tix * Tiy);
|
cannam@167
|
3999 ci[WS(rs, 15)] = FMA(Tiz, Tiy, Tix * TiA);
|
cannam@167
|
4000 }
|
cannam@167
|
4001 {
|
cannam@167
|
4002 E Tia, Tic, Ti9, Tib;
|
cannam@167
|
4003 Tia = Ti2 + Ti3;
|
cannam@167
|
4004 Tic = Ti6 + Ti7;
|
cannam@167
|
4005 Ti9 = W[12];
|
cannam@167
|
4006 Tib = W[13];
|
cannam@167
|
4007 cr[WS(rs, 7)] = FNMS(Tib, Tic, Ti9 * Tia);
|
cannam@167
|
4008 ci[WS(rs, 7)] = FMA(Ti9, Tic, Tib * Tia);
|
cannam@167
|
4009 }
|
cannam@167
|
4010 }
|
cannam@167
|
4011 }
|
cannam@167
|
4012 }
|
cannam@167
|
4013 }
|
cannam@167
|
4014
|
cannam@167
|
4015 static const tw_instr twinstr[] = {
|
cannam@167
|
4016 {TW_FULL, 1, 64},
|
cannam@167
|
4017 {TW_NEXT, 1, 0}
|
cannam@167
|
4018 };
|
cannam@167
|
4019
|
cannam@167
|
4020 static const hc2hc_desc desc = { 64, "hb_64", twinstr, &GENUS, {808, 270, 230, 0} };
|
cannam@167
|
4021
|
cannam@167
|
4022 void X(codelet_hb_64) (planner *p) {
|
cannam@167
|
4023 X(khc2hc_register) (p, hb_64, &desc);
|
cannam@167
|
4024 }
|
cannam@167
|
4025 #endif
|