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 08:04:21 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "dft/codelet-dft.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_twiddle.native -fma -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 64 -name t2_64 -include dft/scalar/t.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 1154 FP additions, 840 FP multiplications,
|
cannam@167
|
32 * (or, 520 additions, 206 multiplications, 634 fused multiply/add),
|
cannam@167
|
33 * 316 stack variables, 15 constants, and 256 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "dft/scalar/t.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void t2_64(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
40 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
41 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
42 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
43 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@167
|
44 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@167
|
45 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@167
|
46 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@167
|
47 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
48 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
49 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
50 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
51 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
52 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
53 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
54 {
|
cannam@167
|
55 INT m;
|
cannam@167
|
56 for (m = mb, W = W + (mb * 10); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 10, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
57 E T2, T3, Tc, T8, Te, T5, T6, Tr, T7, TJ, T14, T3d, T3i, TG, T10;
|
cannam@167
|
58 E T3a, T3g, TL, TP, Tb, Td, T17, Tt, Tu, T1i, Ti, T2U, T1t, T7B, T5O;
|
cannam@167
|
59 E T3N, T3U, T1I, T3G, T3R, T79, T1x, T3D, T2l, T3X, T2d, T1M, T4B, T4x, T4T;
|
cannam@167
|
60 E T2h, T29, T5s, T81, T5w, T7X, T7N, T7h, T64, T6a, T6e, T7l, T60, T7R, T5A;
|
cannam@167
|
61 E T6h, T6J, T7o, T5E, T6k, T6N, T7r, T2X, T6t, T6x, TO, TK, TQ, T7c, TU;
|
cannam@167
|
62 E T2x, T2u, T2y, T7E, T2C, T4b, T48, T4c, T5R, T4g, T3m, T3j, T3n, T4W, T3r;
|
cannam@167
|
63 E Tx, Ty, TC, T1Z, T23, T4s, T4p, T70, T6W, T19, T41, T44, T1a, T1e, T35;
|
cannam@167
|
64 E T31, T59, T55, T1k, T1R, T1V, T1l, T1p, T2Q, T2N, T8i, T8e, Th, T4E, T4H;
|
cannam@167
|
65 E Tj, Tn, T3A, T3w, T5n, T5j;
|
cannam@167
|
66 {
|
cannam@167
|
67 E T1H, Tg, Tw, T1s, T2g, TH, T2t, T47, T3h, T28, T4w, T3M, T2c, T4A, T3Q;
|
cannam@167
|
68 E T1w, T2k, T1L, T5r, T80;
|
cannam@167
|
69 {
|
cannam@167
|
70 E TI, T13, TF, TZ, Ta, T4, T9, Ts;
|
cannam@167
|
71 T2 = W[0];
|
cannam@167
|
72 T3 = W[2];
|
cannam@167
|
73 T4 = T2 * T3;
|
cannam@167
|
74 Tc = W[5];
|
cannam@167
|
75 TI = T3 * Tc;
|
cannam@167
|
76 T13 = T2 * Tc;
|
cannam@167
|
77 T8 = W[4];
|
cannam@167
|
78 Te = W[6];
|
cannam@167
|
79 TF = T3 * T8;
|
cannam@167
|
80 T1H = T8 * Te;
|
cannam@167
|
81 TZ = T2 * T8;
|
cannam@167
|
82 T5 = W[1];
|
cannam@167
|
83 T6 = W[3];
|
cannam@167
|
84 Ta = T2 * T6;
|
cannam@167
|
85 Tr = FMA(T5, T6, T4);
|
cannam@167
|
86 T7 = FNMS(T5, T6, T4);
|
cannam@167
|
87 Tg = T7 * Tc;
|
cannam@167
|
88 Tw = Tr * Tc;
|
cannam@167
|
89 T1s = T3 * Te;
|
cannam@167
|
90 T2g = T2 * Te;
|
cannam@167
|
91 TJ = FMA(T6, T8, TI);
|
cannam@167
|
92 T14 = FNMS(T5, T8, T13);
|
cannam@167
|
93 T3d = FMA(T5, T8, T13);
|
cannam@167
|
94 T3i = FNMS(T6, T8, TI);
|
cannam@167
|
95 TG = FNMS(T6, Tc, TF);
|
cannam@167
|
96 TH = TG * Te;
|
cannam@167
|
97 T10 = FMA(T5, Tc, TZ);
|
cannam@167
|
98 T2t = T10 * Te;
|
cannam@167
|
99 T3a = FNMS(T5, Tc, TZ);
|
cannam@167
|
100 T47 = T3a * Te;
|
cannam@167
|
101 T3g = FMA(T6, Tc, TF);
|
cannam@167
|
102 T3h = T3g * Te;
|
cannam@167
|
103 TL = W[8];
|
cannam@167
|
104 T28 = T3 * TL;
|
cannam@167
|
105 T4w = T8 * TL;
|
cannam@167
|
106 T3M = T2 * TL;
|
cannam@167
|
107 TP = W[9];
|
cannam@167
|
108 T2c = T3 * TP;
|
cannam@167
|
109 T4A = T8 * TP;
|
cannam@167
|
110 T3Q = T2 * TP;
|
cannam@167
|
111 T9 = T7 * T8;
|
cannam@167
|
112 Tb = FMA(T5, T3, Ta);
|
cannam@167
|
113 Td = FMA(Tb, Tc, T9);
|
cannam@167
|
114 T17 = FNMS(Tb, Tc, T9);
|
cannam@167
|
115 Ts = Tr * T8;
|
cannam@167
|
116 Tt = FNMS(T5, T3, Ta);
|
cannam@167
|
117 Tu = FNMS(Tt, Tc, Ts);
|
cannam@167
|
118 T1i = FMA(Tt, Tc, Ts);
|
cannam@167
|
119 Ti = W[7];
|
cannam@167
|
120 T1w = T3 * Ti;
|
cannam@167
|
121 T2k = T2 * Ti;
|
cannam@167
|
122 T1L = T8 * Ti;
|
cannam@167
|
123 T2U = FMA(Tc, Ti, T1H);
|
cannam@167
|
124 }
|
cannam@167
|
125 T1t = FMA(T6, Ti, T1s);
|
cannam@167
|
126 T7B = FNMS(T14, Ti, T2t);
|
cannam@167
|
127 T5O = FNMS(T3d, Ti, T47);
|
cannam@167
|
128 T3N = FMA(T5, TP, T3M);
|
cannam@167
|
129 T3U = FNMS(T6, Ti, T1s);
|
cannam@167
|
130 T1I = FNMS(Tc, Ti, T1H);
|
cannam@167
|
131 T3G = FNMS(T5, Te, T2k);
|
cannam@167
|
132 T3R = FNMS(T5, TL, T3Q);
|
cannam@167
|
133 T79 = FNMS(TJ, Ti, TH);
|
cannam@167
|
134 T1x = FNMS(T6, Te, T1w);
|
cannam@167
|
135 T3D = FMA(T5, Ti, T2g);
|
cannam@167
|
136 T2l = FMA(T5, Te, T2k);
|
cannam@167
|
137 T3X = FMA(T6, Te, T1w);
|
cannam@167
|
138 T2d = FNMS(T6, TL, T2c);
|
cannam@167
|
139 T1M = FMA(Tc, Te, T1L);
|
cannam@167
|
140 T4B = FNMS(Tc, TL, T4A);
|
cannam@167
|
141 T4x = FMA(Tc, TP, T4w);
|
cannam@167
|
142 T4T = FNMS(T3i, Ti, T3h);
|
cannam@167
|
143 T2h = FNMS(T5, Ti, T2g);
|
cannam@167
|
144 T29 = FMA(T6, TP, T28);
|
cannam@167
|
145 T5r = T3g * TL;
|
cannam@167
|
146 T5s = FMA(T3i, TP, T5r);
|
cannam@167
|
147 T80 = T7 * TP;
|
cannam@167
|
148 T81 = FNMS(Tb, TL, T80);
|
cannam@167
|
149 {
|
cannam@167
|
150 E T5v, T7W, T7M, T7g, T63;
|
cannam@167
|
151 T5v = T3g * TP;
|
cannam@167
|
152 T5w = FNMS(T3i, TL, T5v);
|
cannam@167
|
153 T7W = T7 * TL;
|
cannam@167
|
154 T7X = FMA(Tb, TP, T7W);
|
cannam@167
|
155 T7M = TG * TL;
|
cannam@167
|
156 T7N = FMA(TJ, TP, T7M);
|
cannam@167
|
157 T7g = T10 * TL;
|
cannam@167
|
158 T7h = FMA(T14, TP, T7g);
|
cannam@167
|
159 T63 = T3a * TP;
|
cannam@167
|
160 T64 = FNMS(T3d, TL, T63);
|
cannam@167
|
161 }
|
cannam@167
|
162 {
|
cannam@167
|
163 E T69, T6d, T7k, T5Z, T7Q, T5z;
|
cannam@167
|
164 T69 = Tr * TL;
|
cannam@167
|
165 T6a = FMA(Tt, TP, T69);
|
cannam@167
|
166 T6d = Tr * TP;
|
cannam@167
|
167 T6e = FNMS(Tt, TL, T6d);
|
cannam@167
|
168 T7k = T10 * TP;
|
cannam@167
|
169 T7l = FNMS(T14, TL, T7k);
|
cannam@167
|
170 T5Z = T3a * TL;
|
cannam@167
|
171 T60 = FMA(T3d, TP, T5Z);
|
cannam@167
|
172 T7Q = TG * TP;
|
cannam@167
|
173 T7R = FNMS(TJ, TL, T7Q);
|
cannam@167
|
174 T5z = Tr * Te;
|
cannam@167
|
175 T5A = FMA(Tt, Ti, T5z);
|
cannam@167
|
176 T6h = FNMS(Tt, Ti, T5z);
|
cannam@167
|
177 }
|
cannam@167
|
178 {
|
cannam@167
|
179 E T6I, T5D, T6M, T6s, T6w;
|
cannam@167
|
180 T6I = T7 * Te;
|
cannam@167
|
181 T6J = FNMS(Tb, Ti, T6I);
|
cannam@167
|
182 T7o = FMA(Tb, Ti, T6I);
|
cannam@167
|
183 T5D = Tr * Ti;
|
cannam@167
|
184 T5E = FNMS(Tt, Te, T5D);
|
cannam@167
|
185 T6k = FMA(Tt, Te, T5D);
|
cannam@167
|
186 T6M = T7 * Ti;
|
cannam@167
|
187 T6N = FMA(Tb, Te, T6M);
|
cannam@167
|
188 T7r = FNMS(Tb, Te, T6M);
|
cannam@167
|
189 T6s = T2U * TL;
|
cannam@167
|
190 T6w = T2U * TP;
|
cannam@167
|
191 T2X = FNMS(Tc, Te, T1L);
|
cannam@167
|
192 T6t = FMA(T2X, TP, T6s);
|
cannam@167
|
193 T6x = FNMS(T2X, TL, T6w);
|
cannam@167
|
194 {
|
cannam@167
|
195 E TN, TM, TT, T2w, T2v, T2B;
|
cannam@167
|
196 TN = TG * Ti;
|
cannam@167
|
197 TO = FNMS(TJ, Te, TN);
|
cannam@167
|
198 TK = FMA(TJ, Ti, TH);
|
cannam@167
|
199 TM = TK * TL;
|
cannam@167
|
200 TT = TK * TP;
|
cannam@167
|
201 TQ = FMA(TO, TP, TM);
|
cannam@167
|
202 T7c = FMA(TJ, Te, TN);
|
cannam@167
|
203 TU = FNMS(TO, TL, TT);
|
cannam@167
|
204 T2w = T10 * Ti;
|
cannam@167
|
205 T2x = FNMS(T14, Te, T2w);
|
cannam@167
|
206 T2u = FMA(T14, Ti, T2t);
|
cannam@167
|
207 T2v = T2u * TL;
|
cannam@167
|
208 T2B = T2u * TP;
|
cannam@167
|
209 T2y = FMA(T2x, TP, T2v);
|
cannam@167
|
210 T7E = FMA(T14, Te, T2w);
|
cannam@167
|
211 T2C = FNMS(T2x, TL, T2B);
|
cannam@167
|
212 }
|
cannam@167
|
213 }
|
cannam@167
|
214 {
|
cannam@167
|
215 E T4a, T49, T4f, T3l, T3k, T3q;
|
cannam@167
|
216 T4a = T3a * Ti;
|
cannam@167
|
217 T4b = FNMS(T3d, Te, T4a);
|
cannam@167
|
218 T48 = FMA(T3d, Ti, T47);
|
cannam@167
|
219 T49 = T48 * TL;
|
cannam@167
|
220 T4f = T48 * TP;
|
cannam@167
|
221 T4c = FMA(T4b, TP, T49);
|
cannam@167
|
222 T5R = FMA(T3d, Te, T4a);
|
cannam@167
|
223 T4g = FNMS(T4b, TL, T4f);
|
cannam@167
|
224 T3l = T3g * Ti;
|
cannam@167
|
225 T3m = FNMS(T3i, Te, T3l);
|
cannam@167
|
226 T3j = FMA(T3i, Ti, T3h);
|
cannam@167
|
227 T3k = T3j * TL;
|
cannam@167
|
228 T3q = T3j * TP;
|
cannam@167
|
229 T3n = FMA(T3m, TP, T3k);
|
cannam@167
|
230 T4W = FMA(T3i, Te, T3l);
|
cannam@167
|
231 T3r = FNMS(T3m, TL, T3q);
|
cannam@167
|
232 {
|
cannam@167
|
233 E T1Y, T22, Tv, TB, T6Z, T6V;
|
cannam@167
|
234 T1Y = Tu * TL;
|
cannam@167
|
235 T22 = Tu * TP;
|
cannam@167
|
236 Tv = Tu * Te;
|
cannam@167
|
237 TB = Tu * Ti;
|
cannam@167
|
238 Tx = FMA(Tt, T8, Tw);
|
cannam@167
|
239 Ty = FMA(Tx, Ti, Tv);
|
cannam@167
|
240 TC = FNMS(Tx, Te, TB);
|
cannam@167
|
241 T1Z = FMA(Tx, TP, T1Y);
|
cannam@167
|
242 T23 = FNMS(Tx, TL, T22);
|
cannam@167
|
243 T4s = FMA(Tx, Te, TB);
|
cannam@167
|
244 T4p = FNMS(Tx, Ti, Tv);
|
cannam@167
|
245 T6Z = Ty * TP;
|
cannam@167
|
246 T70 = FNMS(TC, TL, T6Z);
|
cannam@167
|
247 T6V = Ty * TL;
|
cannam@167
|
248 T6W = FMA(TC, TP, T6V);
|
cannam@167
|
249 }
|
cannam@167
|
250 }
|
cannam@167
|
251 {
|
cannam@167
|
252 E T30, T34, T18, T1d, T58, T54;
|
cannam@167
|
253 T30 = T17 * TL;
|
cannam@167
|
254 T34 = T17 * TP;
|
cannam@167
|
255 T18 = T17 * Te;
|
cannam@167
|
256 T1d = T17 * Ti;
|
cannam@167
|
257 T19 = FMA(Tb, T8, Tg);
|
cannam@167
|
258 T41 = FMA(T19, Ti, T18);
|
cannam@167
|
259 T44 = FNMS(T19, Te, T1d);
|
cannam@167
|
260 T1a = FNMS(T19, Ti, T18);
|
cannam@167
|
261 T1e = FMA(T19, Te, T1d);
|
cannam@167
|
262 T35 = FNMS(T19, TL, T34);
|
cannam@167
|
263 T31 = FMA(T19, TP, T30);
|
cannam@167
|
264 T58 = T41 * TP;
|
cannam@167
|
265 T59 = FNMS(T44, TL, T58);
|
cannam@167
|
266 T54 = T41 * TL;
|
cannam@167
|
267 T55 = FMA(T44, TP, T54);
|
cannam@167
|
268 }
|
cannam@167
|
269 {
|
cannam@167
|
270 E T1j, T1o, T1Q, T1U, T8h, T8d;
|
cannam@167
|
271 T1j = T1i * TL;
|
cannam@167
|
272 T1o = T1i * TP;
|
cannam@167
|
273 T1Q = T1i * Te;
|
cannam@167
|
274 T1U = T1i * Ti;
|
cannam@167
|
275 T1k = FNMS(Tt, T8, Tw);
|
cannam@167
|
276 T1R = FMA(T1k, Ti, T1Q);
|
cannam@167
|
277 T1V = FNMS(T1k, Te, T1U);
|
cannam@167
|
278 T1l = FMA(T1k, TP, T1j);
|
cannam@167
|
279 T1p = FNMS(T1k, TL, T1o);
|
cannam@167
|
280 T2Q = FMA(T1k, Te, T1U);
|
cannam@167
|
281 T2N = FNMS(T1k, Ti, T1Q);
|
cannam@167
|
282 T8h = T1R * TP;
|
cannam@167
|
283 T8i = FNMS(T1V, TL, T8h);
|
cannam@167
|
284 T8d = T1R * TL;
|
cannam@167
|
285 T8e = FMA(T1V, TP, T8d);
|
cannam@167
|
286 }
|
cannam@167
|
287 {
|
cannam@167
|
288 E T3v, T3z, Tf, Tm, T5m, T5i;
|
cannam@167
|
289 T3v = Td * TL;
|
cannam@167
|
290 T3z = Td * TP;
|
cannam@167
|
291 Tf = Td * Te;
|
cannam@167
|
292 Tm = Td * Ti;
|
cannam@167
|
293 Th = FNMS(Tb, T8, Tg);
|
cannam@167
|
294 T4E = FMA(Th, Ti, Tf);
|
cannam@167
|
295 T4H = FNMS(Th, Te, Tm);
|
cannam@167
|
296 Tj = FNMS(Th, Ti, Tf);
|
cannam@167
|
297 Tn = FMA(Th, Te, Tm);
|
cannam@167
|
298 T3A = FNMS(Th, TL, T3z);
|
cannam@167
|
299 T3w = FMA(Th, TP, T3v);
|
cannam@167
|
300 T5m = T4E * TP;
|
cannam@167
|
301 T5n = FNMS(T4H, TL, T5m);
|
cannam@167
|
302 T5i = T4E * TL;
|
cannam@167
|
303 T5j = FMA(T4H, TP, T5i);
|
cannam@167
|
304 }
|
cannam@167
|
305 }
|
cannam@167
|
306 {
|
cannam@167
|
307 E TY, Tg4, Tl9, TlD, T8w, TdS, Tkd, TkE, T2G, Tge, Tgh, TiK, T98, Te1, T9f;
|
cannam@167
|
308 E Te0, T39, Tgq, Tgn, TiN, T9p, Te5, T9M, Te8, T74, Thr, Thc, Tja, TbI, TeE;
|
cannam@167
|
309 E TcB, TeP, T1B, TkD, Tg7, Tk7, T8D, TdT, T8K, TdU, T27, Tg9, Tgc, TiJ, T8T;
|
cannam@167
|
310 E TdY, T90, TdX, T4k, TgB, Tgy, TiT, T9Y, Tec, Tal, Tef, T5d, Th0, TgL, TiZ;
|
cannam@167
|
311 E Taz, Tel, Tbs, Tew, T3K, Tgo, Tgt, TiO, T9E, Te9, T9P, Te6, T4L, Tgz, TgE;
|
cannam@167
|
312 E TiU, Tad, Teg, Tao, Ted, T5I, TgM, Th3, Tj0, TaO, Tex, Tbv, Tem, T7v, Thd;
|
cannam@167
|
313 E Thu, Tjb, TbX, TeQ, TcE, TeF, T68, Tj5, TgS, Th5, Tbj, Tez, Tbx, Teq, T6B;
|
cannam@167
|
314 E Tj6, TgX, Th6, Tb4, TeA, Tby, Tet, T7V, Tjg, Thj, Thw, Tcs, TeS, TcG, TeJ;
|
cannam@167
|
315 E T8m, Tjh, Tho, Thx, Tcd, TeT, TcH, TeM;
|
cannam@167
|
316 {
|
cannam@167
|
317 E T1, Tkb, Tp, Tka, TE, T8s, TW, T8u;
|
cannam@167
|
318 T1 = ri[0];
|
cannam@167
|
319 Tkb = ii[0];
|
cannam@167
|
320 {
|
cannam@167
|
321 E Tk, Tl, To, Tk9;
|
cannam@167
|
322 Tk = ri[WS(rs, 32)];
|
cannam@167
|
323 Tl = Tj * Tk;
|
cannam@167
|
324 To = ii[WS(rs, 32)];
|
cannam@167
|
325 Tk9 = Tj * To;
|
cannam@167
|
326 Tp = FMA(Tn, To, Tl);
|
cannam@167
|
327 Tka = FNMS(Tn, Tk, Tk9);
|
cannam@167
|
328 }
|
cannam@167
|
329 {
|
cannam@167
|
330 E Tz, TA, TD, T8r;
|
cannam@167
|
331 Tz = ri[WS(rs, 16)];
|
cannam@167
|
332 TA = Ty * Tz;
|
cannam@167
|
333 TD = ii[WS(rs, 16)];
|
cannam@167
|
334 T8r = Ty * TD;
|
cannam@167
|
335 TE = FMA(TC, TD, TA);
|
cannam@167
|
336 T8s = FNMS(TC, Tz, T8r);
|
cannam@167
|
337 }
|
cannam@167
|
338 {
|
cannam@167
|
339 E TR, TS, TV, T8t;
|
cannam@167
|
340 TR = ri[WS(rs, 48)];
|
cannam@167
|
341 TS = TQ * TR;
|
cannam@167
|
342 TV = ii[WS(rs, 48)];
|
cannam@167
|
343 T8t = TQ * TV;
|
cannam@167
|
344 TW = FMA(TU, TV, TS);
|
cannam@167
|
345 T8u = FNMS(TU, TR, T8t);
|
cannam@167
|
346 }
|
cannam@167
|
347 {
|
cannam@167
|
348 E Tq, TX, Tl7, Tl8;
|
cannam@167
|
349 Tq = T1 + Tp;
|
cannam@167
|
350 TX = TE + TW;
|
cannam@167
|
351 TY = Tq + TX;
|
cannam@167
|
352 Tg4 = Tq - TX;
|
cannam@167
|
353 Tl7 = Tkb - Tka;
|
cannam@167
|
354 Tl8 = TE - TW;
|
cannam@167
|
355 Tl9 = Tl7 - Tl8;
|
cannam@167
|
356 TlD = Tl8 + Tl7;
|
cannam@167
|
357 }
|
cannam@167
|
358 {
|
cannam@167
|
359 E T8q, T8v, Tk8, Tkc;
|
cannam@167
|
360 T8q = T1 - Tp;
|
cannam@167
|
361 T8v = T8s - T8u;
|
cannam@167
|
362 T8w = T8q - T8v;
|
cannam@167
|
363 TdS = T8q + T8v;
|
cannam@167
|
364 Tk8 = T8s + T8u;
|
cannam@167
|
365 Tkc = Tka + Tkb;
|
cannam@167
|
366 Tkd = Tk8 + Tkc;
|
cannam@167
|
367 TkE = Tkc - Tk8;
|
cannam@167
|
368 }
|
cannam@167
|
369 }
|
cannam@167
|
370 {
|
cannam@167
|
371 E T2f, T93, T2E, T9d, T2n, T95, T2s, T9b;
|
cannam@167
|
372 {
|
cannam@167
|
373 E T2a, T2b, T2e, T92;
|
cannam@167
|
374 T2a = ri[WS(rs, 60)];
|
cannam@167
|
375 T2b = T29 * T2a;
|
cannam@167
|
376 T2e = ii[WS(rs, 60)];
|
cannam@167
|
377 T92 = T29 * T2e;
|
cannam@167
|
378 T2f = FMA(T2d, T2e, T2b);
|
cannam@167
|
379 T93 = FNMS(T2d, T2a, T92);
|
cannam@167
|
380 }
|
cannam@167
|
381 {
|
cannam@167
|
382 E T2z, T2A, T2D, T9c;
|
cannam@167
|
383 T2z = ri[WS(rs, 44)];
|
cannam@167
|
384 T2A = T2y * T2z;
|
cannam@167
|
385 T2D = ii[WS(rs, 44)];
|
cannam@167
|
386 T9c = T2y * T2D;
|
cannam@167
|
387 T2E = FMA(T2C, T2D, T2A);
|
cannam@167
|
388 T9d = FNMS(T2C, T2z, T9c);
|
cannam@167
|
389 }
|
cannam@167
|
390 {
|
cannam@167
|
391 E T2i, T2j, T2m, T94;
|
cannam@167
|
392 T2i = ri[WS(rs, 28)];
|
cannam@167
|
393 T2j = T2h * T2i;
|
cannam@167
|
394 T2m = ii[WS(rs, 28)];
|
cannam@167
|
395 T94 = T2h * T2m;
|
cannam@167
|
396 T2n = FMA(T2l, T2m, T2j);
|
cannam@167
|
397 T95 = FNMS(T2l, T2i, T94);
|
cannam@167
|
398 }
|
cannam@167
|
399 {
|
cannam@167
|
400 E T2p, T2q, T2r, T9a;
|
cannam@167
|
401 T2p = ri[WS(rs, 12)];
|
cannam@167
|
402 T2q = TG * T2p;
|
cannam@167
|
403 T2r = ii[WS(rs, 12)];
|
cannam@167
|
404 T9a = TG * T2r;
|
cannam@167
|
405 T2s = FMA(TJ, T2r, T2q);
|
cannam@167
|
406 T9b = FNMS(TJ, T2p, T9a);
|
cannam@167
|
407 }
|
cannam@167
|
408 {
|
cannam@167
|
409 E T2o, T2F, Tgf, Tgg;
|
cannam@167
|
410 T2o = T2f + T2n;
|
cannam@167
|
411 T2F = T2s + T2E;
|
cannam@167
|
412 T2G = T2o + T2F;
|
cannam@167
|
413 Tge = T2o - T2F;
|
cannam@167
|
414 Tgf = T93 + T95;
|
cannam@167
|
415 Tgg = T9b + T9d;
|
cannam@167
|
416 Tgh = Tgf - Tgg;
|
cannam@167
|
417 TiK = Tgf + Tgg;
|
cannam@167
|
418 }
|
cannam@167
|
419 {
|
cannam@167
|
420 E T96, T97, T99, T9e;
|
cannam@167
|
421 T96 = T93 - T95;
|
cannam@167
|
422 T97 = T2s - T2E;
|
cannam@167
|
423 T98 = T96 + T97;
|
cannam@167
|
424 Te1 = T96 - T97;
|
cannam@167
|
425 T99 = T2f - T2n;
|
cannam@167
|
426 T9e = T9b - T9d;
|
cannam@167
|
427 T9f = T99 - T9e;
|
cannam@167
|
428 Te0 = T99 + T9e;
|
cannam@167
|
429 }
|
cannam@167
|
430 }
|
cannam@167
|
431 {
|
cannam@167
|
432 E T2M, T9k, T37, T9K, T2S, T9m, T2Z, T9I;
|
cannam@167
|
433 {
|
cannam@167
|
434 E T2J, T2K, T2L, T9j;
|
cannam@167
|
435 T2J = ri[WS(rs, 2)];
|
cannam@167
|
436 T2K = Tr * T2J;
|
cannam@167
|
437 T2L = ii[WS(rs, 2)];
|
cannam@167
|
438 T9j = Tr * T2L;
|
cannam@167
|
439 T2M = FMA(Tt, T2L, T2K);
|
cannam@167
|
440 T9k = FNMS(Tt, T2J, T9j);
|
cannam@167
|
441 }
|
cannam@167
|
442 {
|
cannam@167
|
443 E T32, T33, T36, T9J;
|
cannam@167
|
444 T32 = ri[WS(rs, 50)];
|
cannam@167
|
445 T33 = T31 * T32;
|
cannam@167
|
446 T36 = ii[WS(rs, 50)];
|
cannam@167
|
447 T9J = T31 * T36;
|
cannam@167
|
448 T37 = FMA(T35, T36, T33);
|
cannam@167
|
449 T9K = FNMS(T35, T32, T9J);
|
cannam@167
|
450 }
|
cannam@167
|
451 {
|
cannam@167
|
452 E T2O, T2P, T2R, T9l;
|
cannam@167
|
453 T2O = ri[WS(rs, 34)];
|
cannam@167
|
454 T2P = T2N * T2O;
|
cannam@167
|
455 T2R = ii[WS(rs, 34)];
|
cannam@167
|
456 T9l = T2N * T2R;
|
cannam@167
|
457 T2S = FMA(T2Q, T2R, T2P);
|
cannam@167
|
458 T9m = FNMS(T2Q, T2O, T9l);
|
cannam@167
|
459 }
|
cannam@167
|
460 {
|
cannam@167
|
461 E T2V, T2W, T2Y, T9H;
|
cannam@167
|
462 T2V = ri[WS(rs, 18)];
|
cannam@167
|
463 T2W = T2U * T2V;
|
cannam@167
|
464 T2Y = ii[WS(rs, 18)];
|
cannam@167
|
465 T9H = T2U * T2Y;
|
cannam@167
|
466 T2Z = FMA(T2X, T2Y, T2W);
|
cannam@167
|
467 T9I = FNMS(T2X, T2V, T9H);
|
cannam@167
|
468 }
|
cannam@167
|
469 {
|
cannam@167
|
470 E T2T, T38, Tgl, Tgm;
|
cannam@167
|
471 T2T = T2M + T2S;
|
cannam@167
|
472 T38 = T2Z + T37;
|
cannam@167
|
473 T39 = T2T + T38;
|
cannam@167
|
474 Tgq = T2T - T38;
|
cannam@167
|
475 Tgl = T9k + T9m;
|
cannam@167
|
476 Tgm = T9I + T9K;
|
cannam@167
|
477 Tgn = Tgl - Tgm;
|
cannam@167
|
478 TiN = Tgl + Tgm;
|
cannam@167
|
479 }
|
cannam@167
|
480 {
|
cannam@167
|
481 E T9n, T9o, T9G, T9L;
|
cannam@167
|
482 T9n = T9k - T9m;
|
cannam@167
|
483 T9o = T2Z - T37;
|
cannam@167
|
484 T9p = T9n + T9o;
|
cannam@167
|
485 Te5 = T9n - T9o;
|
cannam@167
|
486 T9G = T2M - T2S;
|
cannam@167
|
487 T9L = T9I - T9K;
|
cannam@167
|
488 T9M = T9G - T9L;
|
cannam@167
|
489 Te8 = T9G + T9L;
|
cannam@167
|
490 }
|
cannam@167
|
491 }
|
cannam@167
|
492 {
|
cannam@167
|
493 E T6H, TbD, T72, Tcz, T6P, TbF, T6U, Tcx;
|
cannam@167
|
494 {
|
cannam@167
|
495 E T6E, T6F, T6G, TbC;
|
cannam@167
|
496 T6E = ri[WS(rs, 63)];
|
cannam@167
|
497 T6F = TL * T6E;
|
cannam@167
|
498 T6G = ii[WS(rs, 63)];
|
cannam@167
|
499 TbC = TL * T6G;
|
cannam@167
|
500 T6H = FMA(TP, T6G, T6F);
|
cannam@167
|
501 TbD = FNMS(TP, T6E, TbC);
|
cannam@167
|
502 }
|
cannam@167
|
503 {
|
cannam@167
|
504 E T6X, T6Y, T71, Tcy;
|
cannam@167
|
505 T6X = ri[WS(rs, 47)];
|
cannam@167
|
506 T6Y = T6W * T6X;
|
cannam@167
|
507 T71 = ii[WS(rs, 47)];
|
cannam@167
|
508 Tcy = T6W * T71;
|
cannam@167
|
509 T72 = FMA(T70, T71, T6Y);
|
cannam@167
|
510 Tcz = FNMS(T70, T6X, Tcy);
|
cannam@167
|
511 }
|
cannam@167
|
512 {
|
cannam@167
|
513 E T6K, T6L, T6O, TbE;
|
cannam@167
|
514 T6K = ri[WS(rs, 31)];
|
cannam@167
|
515 T6L = T6J * T6K;
|
cannam@167
|
516 T6O = ii[WS(rs, 31)];
|
cannam@167
|
517 TbE = T6J * T6O;
|
cannam@167
|
518 T6P = FMA(T6N, T6O, T6L);
|
cannam@167
|
519 TbF = FNMS(T6N, T6K, TbE);
|
cannam@167
|
520 }
|
cannam@167
|
521 {
|
cannam@167
|
522 E T6R, T6S, T6T, Tcw;
|
cannam@167
|
523 T6R = ri[WS(rs, 15)];
|
cannam@167
|
524 T6S = TK * T6R;
|
cannam@167
|
525 T6T = ii[WS(rs, 15)];
|
cannam@167
|
526 Tcw = TK * T6T;
|
cannam@167
|
527 T6U = FMA(TO, T6T, T6S);
|
cannam@167
|
528 Tcx = FNMS(TO, T6R, Tcw);
|
cannam@167
|
529 }
|
cannam@167
|
530 {
|
cannam@167
|
531 E T6Q, T73, Tha, Thb;
|
cannam@167
|
532 T6Q = T6H + T6P;
|
cannam@167
|
533 T73 = T6U + T72;
|
cannam@167
|
534 T74 = T6Q + T73;
|
cannam@167
|
535 Thr = T6Q - T73;
|
cannam@167
|
536 Tha = TbD + TbF;
|
cannam@167
|
537 Thb = Tcx + Tcz;
|
cannam@167
|
538 Thc = Tha - Thb;
|
cannam@167
|
539 Tja = Tha + Thb;
|
cannam@167
|
540 }
|
cannam@167
|
541 {
|
cannam@167
|
542 E TbG, TbH, Tcv, TcA;
|
cannam@167
|
543 TbG = TbD - TbF;
|
cannam@167
|
544 TbH = T6U - T72;
|
cannam@167
|
545 TbI = TbG + TbH;
|
cannam@167
|
546 TeE = TbG - TbH;
|
cannam@167
|
547 Tcv = T6H - T6P;
|
cannam@167
|
548 TcA = Tcx - Tcz;
|
cannam@167
|
549 TcB = Tcv - TcA;
|
cannam@167
|
550 TeP = Tcv + TcA;
|
cannam@167
|
551 }
|
cannam@167
|
552 }
|
cannam@167
|
553 {
|
cannam@167
|
554 E T16, T8y, T1z, T8I, T1g, T8A, T1r, T8G;
|
cannam@167
|
555 {
|
cannam@167
|
556 E T11, T12, T15, T8x;
|
cannam@167
|
557 T11 = ri[WS(rs, 8)];
|
cannam@167
|
558 T12 = T10 * T11;
|
cannam@167
|
559 T15 = ii[WS(rs, 8)];
|
cannam@167
|
560 T8x = T10 * T15;
|
cannam@167
|
561 T16 = FMA(T14, T15, T12);
|
cannam@167
|
562 T8y = FNMS(T14, T11, T8x);
|
cannam@167
|
563 }
|
cannam@167
|
564 {
|
cannam@167
|
565 E T1u, T1v, T1y, T8H;
|
cannam@167
|
566 T1u = ri[WS(rs, 24)];
|
cannam@167
|
567 T1v = T1t * T1u;
|
cannam@167
|
568 T1y = ii[WS(rs, 24)];
|
cannam@167
|
569 T8H = T1t * T1y;
|
cannam@167
|
570 T1z = FMA(T1x, T1y, T1v);
|
cannam@167
|
571 T8I = FNMS(T1x, T1u, T8H);
|
cannam@167
|
572 }
|
cannam@167
|
573 {
|
cannam@167
|
574 E T1b, T1c, T1f, T8z;
|
cannam@167
|
575 T1b = ri[WS(rs, 40)];
|
cannam@167
|
576 T1c = T1a * T1b;
|
cannam@167
|
577 T1f = ii[WS(rs, 40)];
|
cannam@167
|
578 T8z = T1a * T1f;
|
cannam@167
|
579 T1g = FMA(T1e, T1f, T1c);
|
cannam@167
|
580 T8A = FNMS(T1e, T1b, T8z);
|
cannam@167
|
581 }
|
cannam@167
|
582 {
|
cannam@167
|
583 E T1m, T1n, T1q, T8F;
|
cannam@167
|
584 T1m = ri[WS(rs, 56)];
|
cannam@167
|
585 T1n = T1l * T1m;
|
cannam@167
|
586 T1q = ii[WS(rs, 56)];
|
cannam@167
|
587 T8F = T1l * T1q;
|
cannam@167
|
588 T1r = FMA(T1p, T1q, T1n);
|
cannam@167
|
589 T8G = FNMS(T1p, T1m, T8F);
|
cannam@167
|
590 }
|
cannam@167
|
591 {
|
cannam@167
|
592 E T1h, T1A, Tg5, Tg6;
|
cannam@167
|
593 T1h = T16 + T1g;
|
cannam@167
|
594 T1A = T1r + T1z;
|
cannam@167
|
595 T1B = T1h + T1A;
|
cannam@167
|
596 TkD = T1A - T1h;
|
cannam@167
|
597 Tg5 = T8y + T8A;
|
cannam@167
|
598 Tg6 = T8G + T8I;
|
cannam@167
|
599 Tg7 = Tg5 - Tg6;
|
cannam@167
|
600 Tk7 = Tg5 + Tg6;
|
cannam@167
|
601 }
|
cannam@167
|
602 {
|
cannam@167
|
603 E T8B, T8C, T8E, T8J;
|
cannam@167
|
604 T8B = T8y - T8A;
|
cannam@167
|
605 T8C = T16 - T1g;
|
cannam@167
|
606 T8D = T8B - T8C;
|
cannam@167
|
607 TdT = T8C + T8B;
|
cannam@167
|
608 T8E = T1r - T1z;
|
cannam@167
|
609 T8J = T8G - T8I;
|
cannam@167
|
610 T8K = T8E + T8J;
|
cannam@167
|
611 TdU = T8E - T8J;
|
cannam@167
|
612 }
|
cannam@167
|
613 }
|
cannam@167
|
614 {
|
cannam@167
|
615 E T1G, T8O, T25, T8Y, T1O, T8Q, T1X, T8W;
|
cannam@167
|
616 {
|
cannam@167
|
617 E T1D, T1E, T1F, T8N;
|
cannam@167
|
618 T1D = ri[WS(rs, 4)];
|
cannam@167
|
619 T1E = T7 * T1D;
|
cannam@167
|
620 T1F = ii[WS(rs, 4)];
|
cannam@167
|
621 T8N = T7 * T1F;
|
cannam@167
|
622 T1G = FMA(Tb, T1F, T1E);
|
cannam@167
|
623 T8O = FNMS(Tb, T1D, T8N);
|
cannam@167
|
624 }
|
cannam@167
|
625 {
|
cannam@167
|
626 E T20, T21, T24, T8X;
|
cannam@167
|
627 T20 = ri[WS(rs, 52)];
|
cannam@167
|
628 T21 = T1Z * T20;
|
cannam@167
|
629 T24 = ii[WS(rs, 52)];
|
cannam@167
|
630 T8X = T1Z * T24;
|
cannam@167
|
631 T25 = FMA(T23, T24, T21);
|
cannam@167
|
632 T8Y = FNMS(T23, T20, T8X);
|
cannam@167
|
633 }
|
cannam@167
|
634 {
|
cannam@167
|
635 E T1J, T1K, T1N, T8P;
|
cannam@167
|
636 T1J = ri[WS(rs, 36)];
|
cannam@167
|
637 T1K = T1I * T1J;
|
cannam@167
|
638 T1N = ii[WS(rs, 36)];
|
cannam@167
|
639 T8P = T1I * T1N;
|
cannam@167
|
640 T1O = FMA(T1M, T1N, T1K);
|
cannam@167
|
641 T8Q = FNMS(T1M, T1J, T8P);
|
cannam@167
|
642 }
|
cannam@167
|
643 {
|
cannam@167
|
644 E T1S, T1T, T1W, T8V;
|
cannam@167
|
645 T1S = ri[WS(rs, 20)];
|
cannam@167
|
646 T1T = T1R * T1S;
|
cannam@167
|
647 T1W = ii[WS(rs, 20)];
|
cannam@167
|
648 T8V = T1R * T1W;
|
cannam@167
|
649 T1X = FMA(T1V, T1W, T1T);
|
cannam@167
|
650 T8W = FNMS(T1V, T1S, T8V);
|
cannam@167
|
651 }
|
cannam@167
|
652 {
|
cannam@167
|
653 E T1P, T26, Tga, Tgb;
|
cannam@167
|
654 T1P = T1G + T1O;
|
cannam@167
|
655 T26 = T1X + T25;
|
cannam@167
|
656 T27 = T1P + T26;
|
cannam@167
|
657 Tg9 = T1P - T26;
|
cannam@167
|
658 Tga = T8O + T8Q;
|
cannam@167
|
659 Tgb = T8W + T8Y;
|
cannam@167
|
660 Tgc = Tga - Tgb;
|
cannam@167
|
661 TiJ = Tga + Tgb;
|
cannam@167
|
662 }
|
cannam@167
|
663 {
|
cannam@167
|
664 E T8R, T8S, T8U, T8Z;
|
cannam@167
|
665 T8R = T8O - T8Q;
|
cannam@167
|
666 T8S = T1X - T25;
|
cannam@167
|
667 T8T = T8R + T8S;
|
cannam@167
|
668 TdY = T8R - T8S;
|
cannam@167
|
669 T8U = T1G - T1O;
|
cannam@167
|
670 T8Z = T8W - T8Y;
|
cannam@167
|
671 T90 = T8U - T8Z;
|
cannam@167
|
672 TdX = T8U + T8Z;
|
cannam@167
|
673 }
|
cannam@167
|
674 }
|
cannam@167
|
675 {
|
cannam@167
|
676 E T3T, T9T, T4i, Taj, T3Z, T9V, T46, Tah;
|
cannam@167
|
677 {
|
cannam@167
|
678 E T3O, T3P, T3S, T9S;
|
cannam@167
|
679 T3O = ri[WS(rs, 62)];
|
cannam@167
|
680 T3P = T3N * T3O;
|
cannam@167
|
681 T3S = ii[WS(rs, 62)];
|
cannam@167
|
682 T9S = T3N * T3S;
|
cannam@167
|
683 T3T = FMA(T3R, T3S, T3P);
|
cannam@167
|
684 T9T = FNMS(T3R, T3O, T9S);
|
cannam@167
|
685 }
|
cannam@167
|
686 {
|
cannam@167
|
687 E T4d, T4e, T4h, Tai;
|
cannam@167
|
688 T4d = ri[WS(rs, 46)];
|
cannam@167
|
689 T4e = T4c * T4d;
|
cannam@167
|
690 T4h = ii[WS(rs, 46)];
|
cannam@167
|
691 Tai = T4c * T4h;
|
cannam@167
|
692 T4i = FMA(T4g, T4h, T4e);
|
cannam@167
|
693 Taj = FNMS(T4g, T4d, Tai);
|
cannam@167
|
694 }
|
cannam@167
|
695 {
|
cannam@167
|
696 E T3V, T3W, T3Y, T9U;
|
cannam@167
|
697 T3V = ri[WS(rs, 30)];
|
cannam@167
|
698 T3W = T3U * T3V;
|
cannam@167
|
699 T3Y = ii[WS(rs, 30)];
|
cannam@167
|
700 T9U = T3U * T3Y;
|
cannam@167
|
701 T3Z = FMA(T3X, T3Y, T3W);
|
cannam@167
|
702 T9V = FNMS(T3X, T3V, T9U);
|
cannam@167
|
703 }
|
cannam@167
|
704 {
|
cannam@167
|
705 E T42, T43, T45, Tag;
|
cannam@167
|
706 T42 = ri[WS(rs, 14)];
|
cannam@167
|
707 T43 = T41 * T42;
|
cannam@167
|
708 T45 = ii[WS(rs, 14)];
|
cannam@167
|
709 Tag = T41 * T45;
|
cannam@167
|
710 T46 = FMA(T44, T45, T43);
|
cannam@167
|
711 Tah = FNMS(T44, T42, Tag);
|
cannam@167
|
712 }
|
cannam@167
|
713 {
|
cannam@167
|
714 E T40, T4j, Tgw, Tgx;
|
cannam@167
|
715 T40 = T3T + T3Z;
|
cannam@167
|
716 T4j = T46 + T4i;
|
cannam@167
|
717 T4k = T40 + T4j;
|
cannam@167
|
718 TgB = T40 - T4j;
|
cannam@167
|
719 Tgw = T9T + T9V;
|
cannam@167
|
720 Tgx = Tah + Taj;
|
cannam@167
|
721 Tgy = Tgw - Tgx;
|
cannam@167
|
722 TiT = Tgw + Tgx;
|
cannam@167
|
723 }
|
cannam@167
|
724 {
|
cannam@167
|
725 E T9W, T9X, Taf, Tak;
|
cannam@167
|
726 T9W = T9T - T9V;
|
cannam@167
|
727 T9X = T46 - T4i;
|
cannam@167
|
728 T9Y = T9W + T9X;
|
cannam@167
|
729 Tec = T9W - T9X;
|
cannam@167
|
730 Taf = T3T - T3Z;
|
cannam@167
|
731 Tak = Tah - Taj;
|
cannam@167
|
732 Tal = Taf - Tak;
|
cannam@167
|
733 Tef = Taf + Tak;
|
cannam@167
|
734 }
|
cannam@167
|
735 }
|
cannam@167
|
736 {
|
cannam@167
|
737 E T4S, Tau, T5b, Tbq, T4Y, Taw, T53, Tbo;
|
cannam@167
|
738 {
|
cannam@167
|
739 E T4P, T4Q, T4R, Tat;
|
cannam@167
|
740 T4P = ri[WS(rs, 1)];
|
cannam@167
|
741 T4Q = T2 * T4P;
|
cannam@167
|
742 T4R = ii[WS(rs, 1)];
|
cannam@167
|
743 Tat = T2 * T4R;
|
cannam@167
|
744 T4S = FMA(T5, T4R, T4Q);
|
cannam@167
|
745 Tau = FNMS(T5, T4P, Tat);
|
cannam@167
|
746 }
|
cannam@167
|
747 {
|
cannam@167
|
748 E T56, T57, T5a, Tbp;
|
cannam@167
|
749 T56 = ri[WS(rs, 49)];
|
cannam@167
|
750 T57 = T55 * T56;
|
cannam@167
|
751 T5a = ii[WS(rs, 49)];
|
cannam@167
|
752 Tbp = T55 * T5a;
|
cannam@167
|
753 T5b = FMA(T59, T5a, T57);
|
cannam@167
|
754 Tbq = FNMS(T59, T56, Tbp);
|
cannam@167
|
755 }
|
cannam@167
|
756 {
|
cannam@167
|
757 E T4U, T4V, T4X, Tav;
|
cannam@167
|
758 T4U = ri[WS(rs, 33)];
|
cannam@167
|
759 T4V = T4T * T4U;
|
cannam@167
|
760 T4X = ii[WS(rs, 33)];
|
cannam@167
|
761 Tav = T4T * T4X;
|
cannam@167
|
762 T4Y = FMA(T4W, T4X, T4V);
|
cannam@167
|
763 Taw = FNMS(T4W, T4U, Tav);
|
cannam@167
|
764 }
|
cannam@167
|
765 {
|
cannam@167
|
766 E T50, T51, T52, Tbn;
|
cannam@167
|
767 T50 = ri[WS(rs, 17)];
|
cannam@167
|
768 T51 = T48 * T50;
|
cannam@167
|
769 T52 = ii[WS(rs, 17)];
|
cannam@167
|
770 Tbn = T48 * T52;
|
cannam@167
|
771 T53 = FMA(T4b, T52, T51);
|
cannam@167
|
772 Tbo = FNMS(T4b, T50, Tbn);
|
cannam@167
|
773 }
|
cannam@167
|
774 {
|
cannam@167
|
775 E T4Z, T5c, TgJ, TgK;
|
cannam@167
|
776 T4Z = T4S + T4Y;
|
cannam@167
|
777 T5c = T53 + T5b;
|
cannam@167
|
778 T5d = T4Z + T5c;
|
cannam@167
|
779 Th0 = T4Z - T5c;
|
cannam@167
|
780 TgJ = Tau + Taw;
|
cannam@167
|
781 TgK = Tbo + Tbq;
|
cannam@167
|
782 TgL = TgJ - TgK;
|
cannam@167
|
783 TiZ = TgJ + TgK;
|
cannam@167
|
784 }
|
cannam@167
|
785 {
|
cannam@167
|
786 E Tax, Tay, Tbm, Tbr;
|
cannam@167
|
787 Tax = Tau - Taw;
|
cannam@167
|
788 Tay = T53 - T5b;
|
cannam@167
|
789 Taz = Tax + Tay;
|
cannam@167
|
790 Tel = Tax - Tay;
|
cannam@167
|
791 Tbm = T4S - T4Y;
|
cannam@167
|
792 Tbr = Tbo - Tbq;
|
cannam@167
|
793 Tbs = Tbm - Tbr;
|
cannam@167
|
794 Tew = Tbm + Tbr;
|
cannam@167
|
795 }
|
cannam@167
|
796 }
|
cannam@167
|
797 {
|
cannam@167
|
798 E T3f, T9s, T3I, T9B, T3t, T9u, T3C, T9z;
|
cannam@167
|
799 {
|
cannam@167
|
800 E T3b, T3c, T3e, T9r;
|
cannam@167
|
801 T3b = ri[WS(rs, 10)];
|
cannam@167
|
802 T3c = T3a * T3b;
|
cannam@167
|
803 T3e = ii[WS(rs, 10)];
|
cannam@167
|
804 T9r = T3a * T3e;
|
cannam@167
|
805 T3f = FMA(T3d, T3e, T3c);
|
cannam@167
|
806 T9s = FNMS(T3d, T3b, T9r);
|
cannam@167
|
807 }
|
cannam@167
|
808 {
|
cannam@167
|
809 E T3E, T3F, T3H, T9A;
|
cannam@167
|
810 T3E = ri[WS(rs, 26)];
|
cannam@167
|
811 T3F = T3D * T3E;
|
cannam@167
|
812 T3H = ii[WS(rs, 26)];
|
cannam@167
|
813 T9A = T3D * T3H;
|
cannam@167
|
814 T3I = FMA(T3G, T3H, T3F);
|
cannam@167
|
815 T9B = FNMS(T3G, T3E, T9A);
|
cannam@167
|
816 }
|
cannam@167
|
817 {
|
cannam@167
|
818 E T3o, T3p, T3s, T9t;
|
cannam@167
|
819 T3o = ri[WS(rs, 42)];
|
cannam@167
|
820 T3p = T3n * T3o;
|
cannam@167
|
821 T3s = ii[WS(rs, 42)];
|
cannam@167
|
822 T9t = T3n * T3s;
|
cannam@167
|
823 T3t = FMA(T3r, T3s, T3p);
|
cannam@167
|
824 T9u = FNMS(T3r, T3o, T9t);
|
cannam@167
|
825 }
|
cannam@167
|
826 {
|
cannam@167
|
827 E T3x, T3y, T3B, T9y;
|
cannam@167
|
828 T3x = ri[WS(rs, 58)];
|
cannam@167
|
829 T3y = T3w * T3x;
|
cannam@167
|
830 T3B = ii[WS(rs, 58)];
|
cannam@167
|
831 T9y = T3w * T3B;
|
cannam@167
|
832 T3C = FMA(T3A, T3B, T3y);
|
cannam@167
|
833 T9z = FNMS(T3A, T3x, T9y);
|
cannam@167
|
834 }
|
cannam@167
|
835 {
|
cannam@167
|
836 E T3u, T3J, Tgr, Tgs;
|
cannam@167
|
837 T3u = T3f + T3t;
|
cannam@167
|
838 T3J = T3C + T3I;
|
cannam@167
|
839 T3K = T3u + T3J;
|
cannam@167
|
840 Tgo = T3J - T3u;
|
cannam@167
|
841 Tgr = T9s + T9u;
|
cannam@167
|
842 Tgs = T9z + T9B;
|
cannam@167
|
843 Tgt = Tgr - Tgs;
|
cannam@167
|
844 TiO = Tgr + Tgs;
|
cannam@167
|
845 {
|
cannam@167
|
846 E T9w, T9O, T9D, T9N;
|
cannam@167
|
847 {
|
cannam@167
|
848 E T9q, T9v, T9x, T9C;
|
cannam@167
|
849 T9q = T3f - T3t;
|
cannam@167
|
850 T9v = T9s - T9u;
|
cannam@167
|
851 T9w = T9q + T9v;
|
cannam@167
|
852 T9O = T9v - T9q;
|
cannam@167
|
853 T9x = T3C - T3I;
|
cannam@167
|
854 T9C = T9z - T9B;
|
cannam@167
|
855 T9D = T9x - T9C;
|
cannam@167
|
856 T9N = T9x + T9C;
|
cannam@167
|
857 }
|
cannam@167
|
858 T9E = T9w - T9D;
|
cannam@167
|
859 Te9 = T9w + T9D;
|
cannam@167
|
860 T9P = T9N - T9O;
|
cannam@167
|
861 Te6 = T9O + T9N;
|
cannam@167
|
862 }
|
cannam@167
|
863 }
|
cannam@167
|
864 }
|
cannam@167
|
865 {
|
cannam@167
|
866 E T4o, Ta1, T4J, Taa, T4u, Ta3, T4D, Ta8;
|
cannam@167
|
867 {
|
cannam@167
|
868 E T4l, T4m, T4n, Ta0;
|
cannam@167
|
869 T4l = ri[WS(rs, 6)];
|
cannam@167
|
870 T4m = T3g * T4l;
|
cannam@167
|
871 T4n = ii[WS(rs, 6)];
|
cannam@167
|
872 Ta0 = T3g * T4n;
|
cannam@167
|
873 T4o = FMA(T3i, T4n, T4m);
|
cannam@167
|
874 Ta1 = FNMS(T3i, T4l, Ta0);
|
cannam@167
|
875 }
|
cannam@167
|
876 {
|
cannam@167
|
877 E T4F, T4G, T4I, Ta9;
|
cannam@167
|
878 T4F = ri[WS(rs, 22)];
|
cannam@167
|
879 T4G = T4E * T4F;
|
cannam@167
|
880 T4I = ii[WS(rs, 22)];
|
cannam@167
|
881 Ta9 = T4E * T4I;
|
cannam@167
|
882 T4J = FMA(T4H, T4I, T4G);
|
cannam@167
|
883 Taa = FNMS(T4H, T4F, Ta9);
|
cannam@167
|
884 }
|
cannam@167
|
885 {
|
cannam@167
|
886 E T4q, T4r, T4t, Ta2;
|
cannam@167
|
887 T4q = ri[WS(rs, 38)];
|
cannam@167
|
888 T4r = T4p * T4q;
|
cannam@167
|
889 T4t = ii[WS(rs, 38)];
|
cannam@167
|
890 Ta2 = T4p * T4t;
|
cannam@167
|
891 T4u = FMA(T4s, T4t, T4r);
|
cannam@167
|
892 Ta3 = FNMS(T4s, T4q, Ta2);
|
cannam@167
|
893 }
|
cannam@167
|
894 {
|
cannam@167
|
895 E T4y, T4z, T4C, Ta7;
|
cannam@167
|
896 T4y = ri[WS(rs, 54)];
|
cannam@167
|
897 T4z = T4x * T4y;
|
cannam@167
|
898 T4C = ii[WS(rs, 54)];
|
cannam@167
|
899 Ta7 = T4x * T4C;
|
cannam@167
|
900 T4D = FMA(T4B, T4C, T4z);
|
cannam@167
|
901 Ta8 = FNMS(T4B, T4y, Ta7);
|
cannam@167
|
902 }
|
cannam@167
|
903 {
|
cannam@167
|
904 E T4v, T4K, TgC, TgD;
|
cannam@167
|
905 T4v = T4o + T4u;
|
cannam@167
|
906 T4K = T4D + T4J;
|
cannam@167
|
907 T4L = T4v + T4K;
|
cannam@167
|
908 Tgz = T4K - T4v;
|
cannam@167
|
909 TgC = Ta1 + Ta3;
|
cannam@167
|
910 TgD = Ta8 + Taa;
|
cannam@167
|
911 TgE = TgC - TgD;
|
cannam@167
|
912 TiU = TgC + TgD;
|
cannam@167
|
913 {
|
cannam@167
|
914 E Ta5, Tan, Tac, Tam;
|
cannam@167
|
915 {
|
cannam@167
|
916 E T9Z, Ta4, Ta6, Tab;
|
cannam@167
|
917 T9Z = T4o - T4u;
|
cannam@167
|
918 Ta4 = Ta1 - Ta3;
|
cannam@167
|
919 Ta5 = T9Z + Ta4;
|
cannam@167
|
920 Tan = Ta4 - T9Z;
|
cannam@167
|
921 Ta6 = T4D - T4J;
|
cannam@167
|
922 Tab = Ta8 - Taa;
|
cannam@167
|
923 Tac = Ta6 - Tab;
|
cannam@167
|
924 Tam = Ta6 + Tab;
|
cannam@167
|
925 }
|
cannam@167
|
926 Tad = Ta5 - Tac;
|
cannam@167
|
927 Teg = Ta5 + Tac;
|
cannam@167
|
928 Tao = Tam - Tan;
|
cannam@167
|
929 Ted = Tan + Tam;
|
cannam@167
|
930 }
|
cannam@167
|
931 }
|
cannam@167
|
932 }
|
cannam@167
|
933 {
|
cannam@167
|
934 E T5h, TaC, T5G, TaL, T5p, TaE, T5y, TaJ;
|
cannam@167
|
935 {
|
cannam@167
|
936 E T5e, T5f, T5g, TaB;
|
cannam@167
|
937 T5e = ri[WS(rs, 9)];
|
cannam@167
|
938 T5f = T8 * T5e;
|
cannam@167
|
939 T5g = ii[WS(rs, 9)];
|
cannam@167
|
940 TaB = T8 * T5g;
|
cannam@167
|
941 T5h = FMA(Tc, T5g, T5f);
|
cannam@167
|
942 TaC = FNMS(Tc, T5e, TaB);
|
cannam@167
|
943 }
|
cannam@167
|
944 {
|
cannam@167
|
945 E T5B, T5C, T5F, TaK;
|
cannam@167
|
946 T5B = ri[WS(rs, 25)];
|
cannam@167
|
947 T5C = T5A * T5B;
|
cannam@167
|
948 T5F = ii[WS(rs, 25)];
|
cannam@167
|
949 TaK = T5A * T5F;
|
cannam@167
|
950 T5G = FMA(T5E, T5F, T5C);
|
cannam@167
|
951 TaL = FNMS(T5E, T5B, TaK);
|
cannam@167
|
952 }
|
cannam@167
|
953 {
|
cannam@167
|
954 E T5k, T5l, T5o, TaD;
|
cannam@167
|
955 T5k = ri[WS(rs, 41)];
|
cannam@167
|
956 T5l = T5j * T5k;
|
cannam@167
|
957 T5o = ii[WS(rs, 41)];
|
cannam@167
|
958 TaD = T5j * T5o;
|
cannam@167
|
959 T5p = FMA(T5n, T5o, T5l);
|
cannam@167
|
960 TaE = FNMS(T5n, T5k, TaD);
|
cannam@167
|
961 }
|
cannam@167
|
962 {
|
cannam@167
|
963 E T5t, T5u, T5x, TaI;
|
cannam@167
|
964 T5t = ri[WS(rs, 57)];
|
cannam@167
|
965 T5u = T5s * T5t;
|
cannam@167
|
966 T5x = ii[WS(rs, 57)];
|
cannam@167
|
967 TaI = T5s * T5x;
|
cannam@167
|
968 T5y = FMA(T5w, T5x, T5u);
|
cannam@167
|
969 TaJ = FNMS(T5w, T5t, TaI);
|
cannam@167
|
970 }
|
cannam@167
|
971 {
|
cannam@167
|
972 E T5q, T5H, Th1, Th2;
|
cannam@167
|
973 T5q = T5h + T5p;
|
cannam@167
|
974 T5H = T5y + T5G;
|
cannam@167
|
975 T5I = T5q + T5H;
|
cannam@167
|
976 TgM = T5H - T5q;
|
cannam@167
|
977 Th1 = TaC + TaE;
|
cannam@167
|
978 Th2 = TaJ + TaL;
|
cannam@167
|
979 Th3 = Th1 - Th2;
|
cannam@167
|
980 Tj0 = Th1 + Th2;
|
cannam@167
|
981 {
|
cannam@167
|
982 E TaG, Tbu, TaN, Tbt;
|
cannam@167
|
983 {
|
cannam@167
|
984 E TaA, TaF, TaH, TaM;
|
cannam@167
|
985 TaA = T5h - T5p;
|
cannam@167
|
986 TaF = TaC - TaE;
|
cannam@167
|
987 TaG = TaA + TaF;
|
cannam@167
|
988 Tbu = TaF - TaA;
|
cannam@167
|
989 TaH = T5y - T5G;
|
cannam@167
|
990 TaM = TaJ - TaL;
|
cannam@167
|
991 TaN = TaH - TaM;
|
cannam@167
|
992 Tbt = TaH + TaM;
|
cannam@167
|
993 }
|
cannam@167
|
994 TaO = TaG - TaN;
|
cannam@167
|
995 Tex = TaG + TaN;
|
cannam@167
|
996 Tbv = Tbt - Tbu;
|
cannam@167
|
997 Tem = Tbu + Tbt;
|
cannam@167
|
998 }
|
cannam@167
|
999 }
|
cannam@167
|
1000 }
|
cannam@167
|
1001 {
|
cannam@167
|
1002 E T78, TbL, T7t, TbU, T7e, TbN, T7n, TbS;
|
cannam@167
|
1003 {
|
cannam@167
|
1004 E T75, T76, T77, TbK;
|
cannam@167
|
1005 T75 = ri[WS(rs, 7)];
|
cannam@167
|
1006 T76 = T1i * T75;
|
cannam@167
|
1007 T77 = ii[WS(rs, 7)];
|
cannam@167
|
1008 TbK = T1i * T77;
|
cannam@167
|
1009 T78 = FMA(T1k, T77, T76);
|
cannam@167
|
1010 TbL = FNMS(T1k, T75, TbK);
|
cannam@167
|
1011 }
|
cannam@167
|
1012 {
|
cannam@167
|
1013 E T7p, T7q, T7s, TbT;
|
cannam@167
|
1014 T7p = ri[WS(rs, 23)];
|
cannam@167
|
1015 T7q = T7o * T7p;
|
cannam@167
|
1016 T7s = ii[WS(rs, 23)];
|
cannam@167
|
1017 TbT = T7o * T7s;
|
cannam@167
|
1018 T7t = FMA(T7r, T7s, T7q);
|
cannam@167
|
1019 TbU = FNMS(T7r, T7p, TbT);
|
cannam@167
|
1020 }
|
cannam@167
|
1021 {
|
cannam@167
|
1022 E T7a, T7b, T7d, TbM;
|
cannam@167
|
1023 T7a = ri[WS(rs, 39)];
|
cannam@167
|
1024 T7b = T79 * T7a;
|
cannam@167
|
1025 T7d = ii[WS(rs, 39)];
|
cannam@167
|
1026 TbM = T79 * T7d;
|
cannam@167
|
1027 T7e = FMA(T7c, T7d, T7b);
|
cannam@167
|
1028 TbN = FNMS(T7c, T7a, TbM);
|
cannam@167
|
1029 }
|
cannam@167
|
1030 {
|
cannam@167
|
1031 E T7i, T7j, T7m, TbR;
|
cannam@167
|
1032 T7i = ri[WS(rs, 55)];
|
cannam@167
|
1033 T7j = T7h * T7i;
|
cannam@167
|
1034 T7m = ii[WS(rs, 55)];
|
cannam@167
|
1035 TbR = T7h * T7m;
|
cannam@167
|
1036 T7n = FMA(T7l, T7m, T7j);
|
cannam@167
|
1037 TbS = FNMS(T7l, T7i, TbR);
|
cannam@167
|
1038 }
|
cannam@167
|
1039 {
|
cannam@167
|
1040 E T7f, T7u, Ths, Tht;
|
cannam@167
|
1041 T7f = T78 + T7e;
|
cannam@167
|
1042 T7u = T7n + T7t;
|
cannam@167
|
1043 T7v = T7f + T7u;
|
cannam@167
|
1044 Thd = T7u - T7f;
|
cannam@167
|
1045 Ths = TbL + TbN;
|
cannam@167
|
1046 Tht = TbS + TbU;
|
cannam@167
|
1047 Thu = Ths - Tht;
|
cannam@167
|
1048 Tjb = Ths + Tht;
|
cannam@167
|
1049 {
|
cannam@167
|
1050 E TbP, TcD, TbW, TcC;
|
cannam@167
|
1051 {
|
cannam@167
|
1052 E TbJ, TbO, TbQ, TbV;
|
cannam@167
|
1053 TbJ = T78 - T7e;
|
cannam@167
|
1054 TbO = TbL - TbN;
|
cannam@167
|
1055 TbP = TbJ + TbO;
|
cannam@167
|
1056 TcD = TbO - TbJ;
|
cannam@167
|
1057 TbQ = T7n - T7t;
|
cannam@167
|
1058 TbV = TbS - TbU;
|
cannam@167
|
1059 TbW = TbQ - TbV;
|
cannam@167
|
1060 TcC = TbQ + TbV;
|
cannam@167
|
1061 }
|
cannam@167
|
1062 TbX = TbP - TbW;
|
cannam@167
|
1063 TeQ = TbP + TbW;
|
cannam@167
|
1064 TcE = TcC - TcD;
|
cannam@167
|
1065 TeF = TcD + TcC;
|
cannam@167
|
1066 }
|
cannam@167
|
1067 }
|
cannam@167
|
1068 }
|
cannam@167
|
1069 {
|
cannam@167
|
1070 E T5N, Tbd, T66, Tb9, T5T, Tbf, T5Y, Tb7;
|
cannam@167
|
1071 {
|
cannam@167
|
1072 E T5K, T5L, T5M, Tbc;
|
cannam@167
|
1073 T5K = ri[WS(rs, 5)];
|
cannam@167
|
1074 T5L = Td * T5K;
|
cannam@167
|
1075 T5M = ii[WS(rs, 5)];
|
cannam@167
|
1076 Tbc = Td * T5M;
|
cannam@167
|
1077 T5N = FMA(Th, T5M, T5L);
|
cannam@167
|
1078 Tbd = FNMS(Th, T5K, Tbc);
|
cannam@167
|
1079 }
|
cannam@167
|
1080 {
|
cannam@167
|
1081 E T61, T62, T65, Tb8;
|
cannam@167
|
1082 T61 = ri[WS(rs, 53)];
|
cannam@167
|
1083 T62 = T60 * T61;
|
cannam@167
|
1084 T65 = ii[WS(rs, 53)];
|
cannam@167
|
1085 Tb8 = T60 * T65;
|
cannam@167
|
1086 T66 = FMA(T64, T65, T62);
|
cannam@167
|
1087 Tb9 = FNMS(T64, T61, Tb8);
|
cannam@167
|
1088 }
|
cannam@167
|
1089 {
|
cannam@167
|
1090 E T5P, T5Q, T5S, Tbe;
|
cannam@167
|
1091 T5P = ri[WS(rs, 37)];
|
cannam@167
|
1092 T5Q = T5O * T5P;
|
cannam@167
|
1093 T5S = ii[WS(rs, 37)];
|
cannam@167
|
1094 Tbe = T5O * T5S;
|
cannam@167
|
1095 T5T = FMA(T5R, T5S, T5Q);
|
cannam@167
|
1096 Tbf = FNMS(T5R, T5P, Tbe);
|
cannam@167
|
1097 }
|
cannam@167
|
1098 {
|
cannam@167
|
1099 E T5V, T5W, T5X, Tb6;
|
cannam@167
|
1100 T5V = ri[WS(rs, 21)];
|
cannam@167
|
1101 T5W = T3j * T5V;
|
cannam@167
|
1102 T5X = ii[WS(rs, 21)];
|
cannam@167
|
1103 Tb6 = T3j * T5X;
|
cannam@167
|
1104 T5Y = FMA(T3m, T5X, T5W);
|
cannam@167
|
1105 Tb7 = FNMS(T3m, T5V, Tb6);
|
cannam@167
|
1106 }
|
cannam@167
|
1107 {
|
cannam@167
|
1108 E T5U, T67, TgR, TgO, TgP, TgQ;
|
cannam@167
|
1109 T5U = T5N + T5T;
|
cannam@167
|
1110 T67 = T5Y + T66;
|
cannam@167
|
1111 TgR = T5U - T67;
|
cannam@167
|
1112 TgO = Tbd + Tbf;
|
cannam@167
|
1113 TgP = Tb7 + Tb9;
|
cannam@167
|
1114 TgQ = TgO - TgP;
|
cannam@167
|
1115 T68 = T5U + T67;
|
cannam@167
|
1116 Tj5 = TgO + TgP;
|
cannam@167
|
1117 TgS = TgQ - TgR;
|
cannam@167
|
1118 Th5 = TgR + TgQ;
|
cannam@167
|
1119 }
|
cannam@167
|
1120 {
|
cannam@167
|
1121 E Tbb, Tep, Tbi, Teo;
|
cannam@167
|
1122 {
|
cannam@167
|
1123 E Tb5, Tba, Tbg, Tbh;
|
cannam@167
|
1124 Tb5 = T5N - T5T;
|
cannam@167
|
1125 Tba = Tb7 - Tb9;
|
cannam@167
|
1126 Tbb = Tb5 - Tba;
|
cannam@167
|
1127 Tep = Tb5 + Tba;
|
cannam@167
|
1128 Tbg = Tbd - Tbf;
|
cannam@167
|
1129 Tbh = T5Y - T66;
|
cannam@167
|
1130 Tbi = Tbg + Tbh;
|
cannam@167
|
1131 Teo = Tbg - Tbh;
|
cannam@167
|
1132 }
|
cannam@167
|
1133 Tbj = FNMS(KP414213562, Tbi, Tbb);
|
cannam@167
|
1134 Tez = FMA(KP414213562, Teo, Tep);
|
cannam@167
|
1135 Tbx = FMA(KP414213562, Tbb, Tbi);
|
cannam@167
|
1136 Teq = FNMS(KP414213562, Tep, Teo);
|
cannam@167
|
1137 }
|
cannam@167
|
1138 }
|
cannam@167
|
1139 {
|
cannam@167
|
1140 E T6g, TaY, T6z, TaU, T6m, Tb0, T6r, TaS;
|
cannam@167
|
1141 {
|
cannam@167
|
1142 E T6b, T6c, T6f, TaX;
|
cannam@167
|
1143 T6b = ri[WS(rs, 61)];
|
cannam@167
|
1144 T6c = T6a * T6b;
|
cannam@167
|
1145 T6f = ii[WS(rs, 61)];
|
cannam@167
|
1146 TaX = T6a * T6f;
|
cannam@167
|
1147 T6g = FMA(T6e, T6f, T6c);
|
cannam@167
|
1148 TaY = FNMS(T6e, T6b, TaX);
|
cannam@167
|
1149 }
|
cannam@167
|
1150 {
|
cannam@167
|
1151 E T6u, T6v, T6y, TaT;
|
cannam@167
|
1152 T6u = ri[WS(rs, 45)];
|
cannam@167
|
1153 T6v = T6t * T6u;
|
cannam@167
|
1154 T6y = ii[WS(rs, 45)];
|
cannam@167
|
1155 TaT = T6t * T6y;
|
cannam@167
|
1156 T6z = FMA(T6x, T6y, T6v);
|
cannam@167
|
1157 TaU = FNMS(T6x, T6u, TaT);
|
cannam@167
|
1158 }
|
cannam@167
|
1159 {
|
cannam@167
|
1160 E T6i, T6j, T6l, TaZ;
|
cannam@167
|
1161 T6i = ri[WS(rs, 29)];
|
cannam@167
|
1162 T6j = T6h * T6i;
|
cannam@167
|
1163 T6l = ii[WS(rs, 29)];
|
cannam@167
|
1164 TaZ = T6h * T6l;
|
cannam@167
|
1165 T6m = FMA(T6k, T6l, T6j);
|
cannam@167
|
1166 Tb0 = FNMS(T6k, T6i, TaZ);
|
cannam@167
|
1167 }
|
cannam@167
|
1168 {
|
cannam@167
|
1169 E T6o, T6p, T6q, TaR;
|
cannam@167
|
1170 T6o = ri[WS(rs, 13)];
|
cannam@167
|
1171 T6p = T17 * T6o;
|
cannam@167
|
1172 T6q = ii[WS(rs, 13)];
|
cannam@167
|
1173 TaR = T17 * T6q;
|
cannam@167
|
1174 T6r = FMA(T19, T6q, T6p);
|
cannam@167
|
1175 TaS = FNMS(T19, T6o, TaR);
|
cannam@167
|
1176 }
|
cannam@167
|
1177 {
|
cannam@167
|
1178 E T6n, T6A, TgT, TgU, TgV, TgW;
|
cannam@167
|
1179 T6n = T6g + T6m;
|
cannam@167
|
1180 T6A = T6r + T6z;
|
cannam@167
|
1181 TgT = T6n - T6A;
|
cannam@167
|
1182 TgU = TaY + Tb0;
|
cannam@167
|
1183 TgV = TaS + TaU;
|
cannam@167
|
1184 TgW = TgU - TgV;
|
cannam@167
|
1185 T6B = T6n + T6A;
|
cannam@167
|
1186 Tj6 = TgU + TgV;
|
cannam@167
|
1187 TgX = TgT + TgW;
|
cannam@167
|
1188 Th6 = TgT - TgW;
|
cannam@167
|
1189 }
|
cannam@167
|
1190 {
|
cannam@167
|
1191 E TaW, Tes, Tb3, Ter;
|
cannam@167
|
1192 {
|
cannam@167
|
1193 E TaQ, TaV, Tb1, Tb2;
|
cannam@167
|
1194 TaQ = T6g - T6m;
|
cannam@167
|
1195 TaV = TaS - TaU;
|
cannam@167
|
1196 TaW = TaQ - TaV;
|
cannam@167
|
1197 Tes = TaQ + TaV;
|
cannam@167
|
1198 Tb1 = TaY - Tb0;
|
cannam@167
|
1199 Tb2 = T6r - T6z;
|
cannam@167
|
1200 Tb3 = Tb1 + Tb2;
|
cannam@167
|
1201 Ter = Tb1 - Tb2;
|
cannam@167
|
1202 }
|
cannam@167
|
1203 Tb4 = FMA(KP414213562, Tb3, TaW);
|
cannam@167
|
1204 TeA = FNMS(KP414213562, Ter, Tes);
|
cannam@167
|
1205 Tby = FNMS(KP414213562, TaW, Tb3);
|
cannam@167
|
1206 Tet = FMA(KP414213562, Tes, Ter);
|
cannam@167
|
1207 }
|
cannam@167
|
1208 }
|
cannam@167
|
1209 {
|
cannam@167
|
1210 E T7A, Tcm, T7T, Tci, T7G, Tco, T7L, Tcg;
|
cannam@167
|
1211 {
|
cannam@167
|
1212 E T7x, T7y, T7z, Tcl;
|
cannam@167
|
1213 T7x = ri[WS(rs, 3)];
|
cannam@167
|
1214 T7y = T3 * T7x;
|
cannam@167
|
1215 T7z = ii[WS(rs, 3)];
|
cannam@167
|
1216 Tcl = T3 * T7z;
|
cannam@167
|
1217 T7A = FMA(T6, T7z, T7y);
|
cannam@167
|
1218 Tcm = FNMS(T6, T7x, Tcl);
|
cannam@167
|
1219 }
|
cannam@167
|
1220 {
|
cannam@167
|
1221 E T7O, T7P, T7S, Tch;
|
cannam@167
|
1222 T7O = ri[WS(rs, 51)];
|
cannam@167
|
1223 T7P = T7N * T7O;
|
cannam@167
|
1224 T7S = ii[WS(rs, 51)];
|
cannam@167
|
1225 Tch = T7N * T7S;
|
cannam@167
|
1226 T7T = FMA(T7R, T7S, T7P);
|
cannam@167
|
1227 Tci = FNMS(T7R, T7O, Tch);
|
cannam@167
|
1228 }
|
cannam@167
|
1229 {
|
cannam@167
|
1230 E T7C, T7D, T7F, Tcn;
|
cannam@167
|
1231 T7C = ri[WS(rs, 35)];
|
cannam@167
|
1232 T7D = T7B * T7C;
|
cannam@167
|
1233 T7F = ii[WS(rs, 35)];
|
cannam@167
|
1234 Tcn = T7B * T7F;
|
cannam@167
|
1235 T7G = FMA(T7E, T7F, T7D);
|
cannam@167
|
1236 Tco = FNMS(T7E, T7C, Tcn);
|
cannam@167
|
1237 }
|
cannam@167
|
1238 {
|
cannam@167
|
1239 E T7I, T7J, T7K, Tcf;
|
cannam@167
|
1240 T7I = ri[WS(rs, 19)];
|
cannam@167
|
1241 T7J = T2u * T7I;
|
cannam@167
|
1242 T7K = ii[WS(rs, 19)];
|
cannam@167
|
1243 Tcf = T2u * T7K;
|
cannam@167
|
1244 T7L = FMA(T2x, T7K, T7J);
|
cannam@167
|
1245 Tcg = FNMS(T2x, T7I, Tcf);
|
cannam@167
|
1246 }
|
cannam@167
|
1247 {
|
cannam@167
|
1248 E T7H, T7U, Thi, Thf, Thg, Thh;
|
cannam@167
|
1249 T7H = T7A + T7G;
|
cannam@167
|
1250 T7U = T7L + T7T;
|
cannam@167
|
1251 Thi = T7H - T7U;
|
cannam@167
|
1252 Thf = Tcm + Tco;
|
cannam@167
|
1253 Thg = Tcg + Tci;
|
cannam@167
|
1254 Thh = Thf - Thg;
|
cannam@167
|
1255 T7V = T7H + T7U;
|
cannam@167
|
1256 Tjg = Thf + Thg;
|
cannam@167
|
1257 Thj = Thh - Thi;
|
cannam@167
|
1258 Thw = Thi + Thh;
|
cannam@167
|
1259 }
|
cannam@167
|
1260 {
|
cannam@167
|
1261 E Tck, TeI, Tcr, TeH;
|
cannam@167
|
1262 {
|
cannam@167
|
1263 E Tce, Tcj, Tcp, Tcq;
|
cannam@167
|
1264 Tce = T7A - T7G;
|
cannam@167
|
1265 Tcj = Tcg - Tci;
|
cannam@167
|
1266 Tck = Tce - Tcj;
|
cannam@167
|
1267 TeI = Tce + Tcj;
|
cannam@167
|
1268 Tcp = Tcm - Tco;
|
cannam@167
|
1269 Tcq = T7L - T7T;
|
cannam@167
|
1270 Tcr = Tcp + Tcq;
|
cannam@167
|
1271 TeH = Tcp - Tcq;
|
cannam@167
|
1272 }
|
cannam@167
|
1273 Tcs = FNMS(KP414213562, Tcr, Tck);
|
cannam@167
|
1274 TeS = FMA(KP414213562, TeH, TeI);
|
cannam@167
|
1275 TcG = FMA(KP414213562, Tck, Tcr);
|
cannam@167
|
1276 TeJ = FNMS(KP414213562, TeI, TeH);
|
cannam@167
|
1277 }
|
cannam@167
|
1278 }
|
cannam@167
|
1279 {
|
cannam@167
|
1280 E T83, Tc7, T8k, Tc3, T87, Tc9, T8c, Tc1;
|
cannam@167
|
1281 {
|
cannam@167
|
1282 E T7Y, T7Z, T82, Tc6;
|
cannam@167
|
1283 T7Y = ri[WS(rs, 59)];
|
cannam@167
|
1284 T7Z = T7X * T7Y;
|
cannam@167
|
1285 T82 = ii[WS(rs, 59)];
|
cannam@167
|
1286 Tc6 = T7X * T82;
|
cannam@167
|
1287 T83 = FMA(T81, T82, T7Z);
|
cannam@167
|
1288 Tc7 = FNMS(T81, T7Y, Tc6);
|
cannam@167
|
1289 }
|
cannam@167
|
1290 {
|
cannam@167
|
1291 E T8f, T8g, T8j, Tc2;
|
cannam@167
|
1292 T8f = ri[WS(rs, 43)];
|
cannam@167
|
1293 T8g = T8e * T8f;
|
cannam@167
|
1294 T8j = ii[WS(rs, 43)];
|
cannam@167
|
1295 Tc2 = T8e * T8j;
|
cannam@167
|
1296 T8k = FMA(T8i, T8j, T8g);
|
cannam@167
|
1297 Tc3 = FNMS(T8i, T8f, Tc2);
|
cannam@167
|
1298 }
|
cannam@167
|
1299 {
|
cannam@167
|
1300 E T84, T85, T86, Tc8;
|
cannam@167
|
1301 T84 = ri[WS(rs, 27)];
|
cannam@167
|
1302 T85 = Te * T84;
|
cannam@167
|
1303 T86 = ii[WS(rs, 27)];
|
cannam@167
|
1304 Tc8 = Te * T86;
|
cannam@167
|
1305 T87 = FMA(Ti, T86, T85);
|
cannam@167
|
1306 Tc9 = FNMS(Ti, T84, Tc8);
|
cannam@167
|
1307 }
|
cannam@167
|
1308 {
|
cannam@167
|
1309 E T89, T8a, T8b, Tc0;
|
cannam@167
|
1310 T89 = ri[WS(rs, 11)];
|
cannam@167
|
1311 T8a = Tu * T89;
|
cannam@167
|
1312 T8b = ii[WS(rs, 11)];
|
cannam@167
|
1313 Tc0 = Tu * T8b;
|
cannam@167
|
1314 T8c = FMA(Tx, T8b, T8a);
|
cannam@167
|
1315 Tc1 = FNMS(Tx, T89, Tc0);
|
cannam@167
|
1316 }
|
cannam@167
|
1317 {
|
cannam@167
|
1318 E T88, T8l, Thk, Thl, Thm, Thn;
|
cannam@167
|
1319 T88 = T83 + T87;
|
cannam@167
|
1320 T8l = T8c + T8k;
|
cannam@167
|
1321 Thk = T88 - T8l;
|
cannam@167
|
1322 Thl = Tc7 + Tc9;
|
cannam@167
|
1323 Thm = Tc1 + Tc3;
|
cannam@167
|
1324 Thn = Thl - Thm;
|
cannam@167
|
1325 T8m = T88 + T8l;
|
cannam@167
|
1326 Tjh = Thl + Thm;
|
cannam@167
|
1327 Tho = Thk + Thn;
|
cannam@167
|
1328 Thx = Thk - Thn;
|
cannam@167
|
1329 }
|
cannam@167
|
1330 {
|
cannam@167
|
1331 E Tc5, TeL, Tcc, TeK;
|
cannam@167
|
1332 {
|
cannam@167
|
1333 E TbZ, Tc4, Tca, Tcb;
|
cannam@167
|
1334 TbZ = T83 - T87;
|
cannam@167
|
1335 Tc4 = Tc1 - Tc3;
|
cannam@167
|
1336 Tc5 = TbZ - Tc4;
|
cannam@167
|
1337 TeL = TbZ + Tc4;
|
cannam@167
|
1338 Tca = Tc7 - Tc9;
|
cannam@167
|
1339 Tcb = T8c - T8k;
|
cannam@167
|
1340 Tcc = Tca + Tcb;
|
cannam@167
|
1341 TeK = Tca - Tcb;
|
cannam@167
|
1342 }
|
cannam@167
|
1343 Tcd = FMA(KP414213562, Tcc, Tc5);
|
cannam@167
|
1344 TeT = FNMS(KP414213562, TeK, TeL);
|
cannam@167
|
1345 TcH = FNMS(KP414213562, Tc5, Tcc);
|
cannam@167
|
1346 TeM = FMA(KP414213562, TeL, TeK);
|
cannam@167
|
1347 }
|
cannam@167
|
1348 }
|
cannam@167
|
1349 {
|
cannam@167
|
1350 E T2I, TjG, T4N, Tkj, Tkf, Tkk, TjJ, Tk5, T8o, Tk2, TjU, TjY, T6D, Tk1, TjP;
|
cannam@167
|
1351 E TjX;
|
cannam@167
|
1352 {
|
cannam@167
|
1353 E T1C, T2H, TjH, TjI;
|
cannam@167
|
1354 T1C = TY + T1B;
|
cannam@167
|
1355 T2H = T27 + T2G;
|
cannam@167
|
1356 T2I = T1C + T2H;
|
cannam@167
|
1357 TjG = T1C - T2H;
|
cannam@167
|
1358 {
|
cannam@167
|
1359 E T3L, T4M, Tk6, Tke;
|
cannam@167
|
1360 T3L = T39 + T3K;
|
cannam@167
|
1361 T4M = T4k + T4L;
|
cannam@167
|
1362 T4N = T3L + T4M;
|
cannam@167
|
1363 Tkj = T4M - T3L;
|
cannam@167
|
1364 Tk6 = TiJ + TiK;
|
cannam@167
|
1365 Tke = Tk7 + Tkd;
|
cannam@167
|
1366 Tkf = Tk6 + Tke;
|
cannam@167
|
1367 Tkk = Tke - Tk6;
|
cannam@167
|
1368 }
|
cannam@167
|
1369 TjH = TiN + TiO;
|
cannam@167
|
1370 TjI = TiT + TiU;
|
cannam@167
|
1371 TjJ = TjH - TjI;
|
cannam@167
|
1372 Tk5 = TjH + TjI;
|
cannam@167
|
1373 {
|
cannam@167
|
1374 E T7w, T8n, TjQ, TjR, TjS, TjT;
|
cannam@167
|
1375 T7w = T74 + T7v;
|
cannam@167
|
1376 T8n = T7V + T8m;
|
cannam@167
|
1377 TjQ = T7w - T8n;
|
cannam@167
|
1378 TjR = Tja + Tjb;
|
cannam@167
|
1379 TjS = Tjg + Tjh;
|
cannam@167
|
1380 TjT = TjR - TjS;
|
cannam@167
|
1381 T8o = T7w + T8n;
|
cannam@167
|
1382 Tk2 = TjR + TjS;
|
cannam@167
|
1383 TjU = TjQ - TjT;
|
cannam@167
|
1384 TjY = TjQ + TjT;
|
cannam@167
|
1385 }
|
cannam@167
|
1386 {
|
cannam@167
|
1387 E T5J, T6C, TjL, TjM, TjN, TjO;
|
cannam@167
|
1388 T5J = T5d + T5I;
|
cannam@167
|
1389 T6C = T68 + T6B;
|
cannam@167
|
1390 TjL = T5J - T6C;
|
cannam@167
|
1391 TjM = TiZ + Tj0;
|
cannam@167
|
1392 TjN = Tj5 + Tj6;
|
cannam@167
|
1393 TjO = TjM - TjN;
|
cannam@167
|
1394 T6D = T5J + T6C;
|
cannam@167
|
1395 Tk1 = TjM + TjN;
|
cannam@167
|
1396 TjP = TjL + TjO;
|
cannam@167
|
1397 TjX = TjO - TjL;
|
cannam@167
|
1398 }
|
cannam@167
|
1399 }
|
cannam@167
|
1400 {
|
cannam@167
|
1401 E T4O, T8p, Tk4, Tkg;
|
cannam@167
|
1402 T4O = T2I + T4N;
|
cannam@167
|
1403 T8p = T6D + T8o;
|
cannam@167
|
1404 ri[WS(rs, 32)] = T4O - T8p;
|
cannam@167
|
1405 ri[0] = T4O + T8p;
|
cannam@167
|
1406 Tk4 = Tk1 + Tk2;
|
cannam@167
|
1407 Tkg = Tk5 + Tkf;
|
cannam@167
|
1408 ii[0] = Tk4 + Tkg;
|
cannam@167
|
1409 ii[WS(rs, 32)] = Tkg - Tk4;
|
cannam@167
|
1410 }
|
cannam@167
|
1411 {
|
cannam@167
|
1412 E TjK, TjV, Tkl, Tkm;
|
cannam@167
|
1413 TjK = TjG + TjJ;
|
cannam@167
|
1414 TjV = TjP + TjU;
|
cannam@167
|
1415 ri[WS(rs, 40)] = FNMS(KP707106781, TjV, TjK);
|
cannam@167
|
1416 ri[WS(rs, 8)] = FMA(KP707106781, TjV, TjK);
|
cannam@167
|
1417 Tkl = Tkj + Tkk;
|
cannam@167
|
1418 Tkm = TjX + TjY;
|
cannam@167
|
1419 ii[WS(rs, 8)] = FMA(KP707106781, Tkm, Tkl);
|
cannam@167
|
1420 ii[WS(rs, 40)] = FNMS(KP707106781, Tkm, Tkl);
|
cannam@167
|
1421 }
|
cannam@167
|
1422 {
|
cannam@167
|
1423 E TjW, TjZ, Tkn, Tko;
|
cannam@167
|
1424 TjW = TjG - TjJ;
|
cannam@167
|
1425 TjZ = TjX - TjY;
|
cannam@167
|
1426 ri[WS(rs, 56)] = FNMS(KP707106781, TjZ, TjW);
|
cannam@167
|
1427 ri[WS(rs, 24)] = FMA(KP707106781, TjZ, TjW);
|
cannam@167
|
1428 Tkn = Tkk - Tkj;
|
cannam@167
|
1429 Tko = TjU - TjP;
|
cannam@167
|
1430 ii[WS(rs, 24)] = FMA(KP707106781, Tko, Tkn);
|
cannam@167
|
1431 ii[WS(rs, 56)] = FNMS(KP707106781, Tko, Tkn);
|
cannam@167
|
1432 }
|
cannam@167
|
1433 {
|
cannam@167
|
1434 E Tk0, Tk3, Tkh, Tki;
|
cannam@167
|
1435 Tk0 = T2I - T4N;
|
cannam@167
|
1436 Tk3 = Tk1 - Tk2;
|
cannam@167
|
1437 ri[WS(rs, 48)] = Tk0 - Tk3;
|
cannam@167
|
1438 ri[WS(rs, 16)] = Tk0 + Tk3;
|
cannam@167
|
1439 Tkh = T8o - T6D;
|
cannam@167
|
1440 Tki = Tkf - Tk5;
|
cannam@167
|
1441 ii[WS(rs, 16)] = Tkh + Tki;
|
cannam@167
|
1442 ii[WS(rs, 48)] = Tki - Tkh;
|
cannam@167
|
1443 }
|
cannam@167
|
1444 }
|
cannam@167
|
1445 {
|
cannam@167
|
1446 E TiM, Tjq, Tkr, Tkx, TiX, Tky, Tjt, Tks, Tj9, TjD, Tjn, Tjx, Tjk, TjE, Tjo;
|
cannam@167
|
1447 E TjA;
|
cannam@167
|
1448 {
|
cannam@167
|
1449 E TiI, TiL, Tkp, Tkq;
|
cannam@167
|
1450 TiI = TY - T1B;
|
cannam@167
|
1451 TiL = TiJ - TiK;
|
cannam@167
|
1452 TiM = TiI - TiL;
|
cannam@167
|
1453 Tjq = TiI + TiL;
|
cannam@167
|
1454 Tkp = T2G - T27;
|
cannam@167
|
1455 Tkq = Tkd - Tk7;
|
cannam@167
|
1456 Tkr = Tkp + Tkq;
|
cannam@167
|
1457 Tkx = Tkq - Tkp;
|
cannam@167
|
1458 }
|
cannam@167
|
1459 {
|
cannam@167
|
1460 E TiR, Tjr, TiW, Tjs;
|
cannam@167
|
1461 {
|
cannam@167
|
1462 E TiP, TiQ, TiS, TiV;
|
cannam@167
|
1463 TiP = TiN - TiO;
|
cannam@167
|
1464 TiQ = T39 - T3K;
|
cannam@167
|
1465 TiR = TiP - TiQ;
|
cannam@167
|
1466 Tjr = TiQ + TiP;
|
cannam@167
|
1467 TiS = T4k - T4L;
|
cannam@167
|
1468 TiV = TiT - TiU;
|
cannam@167
|
1469 TiW = TiS + TiV;
|
cannam@167
|
1470 Tjs = TiS - TiV;
|
cannam@167
|
1471 }
|
cannam@167
|
1472 TiX = TiR - TiW;
|
cannam@167
|
1473 Tky = Tjs - Tjr;
|
cannam@167
|
1474 Tjt = Tjr + Tjs;
|
cannam@167
|
1475 Tks = TiR + TiW;
|
cannam@167
|
1476 }
|
cannam@167
|
1477 {
|
cannam@167
|
1478 E Tj3, Tjw, Tj8, Tjv;
|
cannam@167
|
1479 {
|
cannam@167
|
1480 E Tj1, Tj2, Tj4, Tj7;
|
cannam@167
|
1481 Tj1 = TiZ - Tj0;
|
cannam@167
|
1482 Tj2 = T6B - T68;
|
cannam@167
|
1483 Tj3 = Tj1 - Tj2;
|
cannam@167
|
1484 Tjw = Tj1 + Tj2;
|
cannam@167
|
1485 Tj4 = T5d - T5I;
|
cannam@167
|
1486 Tj7 = Tj5 - Tj6;
|
cannam@167
|
1487 Tj8 = Tj4 - Tj7;
|
cannam@167
|
1488 Tjv = Tj4 + Tj7;
|
cannam@167
|
1489 }
|
cannam@167
|
1490 Tj9 = FMA(KP414213562, Tj8, Tj3);
|
cannam@167
|
1491 TjD = FNMS(KP414213562, Tjv, Tjw);
|
cannam@167
|
1492 Tjn = FNMS(KP414213562, Tj3, Tj8);
|
cannam@167
|
1493 Tjx = FMA(KP414213562, Tjw, Tjv);
|
cannam@167
|
1494 }
|
cannam@167
|
1495 {
|
cannam@167
|
1496 E Tje, Tjz, Tjj, Tjy;
|
cannam@167
|
1497 {
|
cannam@167
|
1498 E Tjc, Tjd, Tjf, Tji;
|
cannam@167
|
1499 Tjc = Tja - Tjb;
|
cannam@167
|
1500 Tjd = T8m - T7V;
|
cannam@167
|
1501 Tje = Tjc - Tjd;
|
cannam@167
|
1502 Tjz = Tjc + Tjd;
|
cannam@167
|
1503 Tjf = T74 - T7v;
|
cannam@167
|
1504 Tji = Tjg - Tjh;
|
cannam@167
|
1505 Tjj = Tjf - Tji;
|
cannam@167
|
1506 Tjy = Tjf + Tji;
|
cannam@167
|
1507 }
|
cannam@167
|
1508 Tjk = FNMS(KP414213562, Tjj, Tje);
|
cannam@167
|
1509 TjE = FMA(KP414213562, Tjy, Tjz);
|
cannam@167
|
1510 Tjo = FMA(KP414213562, Tje, Tjj);
|
cannam@167
|
1511 TjA = FNMS(KP414213562, Tjz, Tjy);
|
cannam@167
|
1512 }
|
cannam@167
|
1513 {
|
cannam@167
|
1514 E TiY, Tjl, Tkz, TkA;
|
cannam@167
|
1515 TiY = FMA(KP707106781, TiX, TiM);
|
cannam@167
|
1516 Tjl = Tj9 - Tjk;
|
cannam@167
|
1517 ri[WS(rs, 44)] = FNMS(KP923879532, Tjl, TiY);
|
cannam@167
|
1518 ri[WS(rs, 12)] = FMA(KP923879532, Tjl, TiY);
|
cannam@167
|
1519 Tkz = FMA(KP707106781, Tky, Tkx);
|
cannam@167
|
1520 TkA = Tjo - Tjn;
|
cannam@167
|
1521 ii[WS(rs, 12)] = FMA(KP923879532, TkA, Tkz);
|
cannam@167
|
1522 ii[WS(rs, 44)] = FNMS(KP923879532, TkA, Tkz);
|
cannam@167
|
1523 }
|
cannam@167
|
1524 {
|
cannam@167
|
1525 E Tjm, Tjp, TkB, TkC;
|
cannam@167
|
1526 Tjm = FNMS(KP707106781, TiX, TiM);
|
cannam@167
|
1527 Tjp = Tjn + Tjo;
|
cannam@167
|
1528 ri[WS(rs, 28)] = FNMS(KP923879532, Tjp, Tjm);
|
cannam@167
|
1529 ri[WS(rs, 60)] = FMA(KP923879532, Tjp, Tjm);
|
cannam@167
|
1530 TkB = FNMS(KP707106781, Tky, Tkx);
|
cannam@167
|
1531 TkC = Tj9 + Tjk;
|
cannam@167
|
1532 ii[WS(rs, 28)] = FNMS(KP923879532, TkC, TkB);
|
cannam@167
|
1533 ii[WS(rs, 60)] = FMA(KP923879532, TkC, TkB);
|
cannam@167
|
1534 }
|
cannam@167
|
1535 {
|
cannam@167
|
1536 E Tju, TjB, Tkt, Tku;
|
cannam@167
|
1537 Tju = FMA(KP707106781, Tjt, Tjq);
|
cannam@167
|
1538 TjB = Tjx + TjA;
|
cannam@167
|
1539 ri[WS(rs, 36)] = FNMS(KP923879532, TjB, Tju);
|
cannam@167
|
1540 ri[WS(rs, 4)] = FMA(KP923879532, TjB, Tju);
|
cannam@167
|
1541 Tkt = FMA(KP707106781, Tks, Tkr);
|
cannam@167
|
1542 Tku = TjD + TjE;
|
cannam@167
|
1543 ii[WS(rs, 4)] = FMA(KP923879532, Tku, Tkt);
|
cannam@167
|
1544 ii[WS(rs, 36)] = FNMS(KP923879532, Tku, Tkt);
|
cannam@167
|
1545 }
|
cannam@167
|
1546 {
|
cannam@167
|
1547 E TjC, TjF, Tkv, Tkw;
|
cannam@167
|
1548 TjC = FNMS(KP707106781, Tjt, Tjq);
|
cannam@167
|
1549 TjF = TjD - TjE;
|
cannam@167
|
1550 ri[WS(rs, 52)] = FNMS(KP923879532, TjF, TjC);
|
cannam@167
|
1551 ri[WS(rs, 20)] = FMA(KP923879532, TjF, TjC);
|
cannam@167
|
1552 Tkv = FNMS(KP707106781, Tks, Tkr);
|
cannam@167
|
1553 Tkw = TjA - Tjx;
|
cannam@167
|
1554 ii[WS(rs, 20)] = FMA(KP923879532, Tkw, Tkv);
|
cannam@167
|
1555 ii[WS(rs, 52)] = FNMS(KP923879532, Tkw, Tkv);
|
cannam@167
|
1556 }
|
cannam@167
|
1557 }
|
cannam@167
|
1558 {
|
cannam@167
|
1559 E Tgk, Tl1, ThG, TkV, Ti0, TkN, Tis, TkH, TgH, TkO, ThJ, TkI, Tim, TiG, Tiq;
|
cannam@167
|
1560 E TiC, Th9, ThT, ThD, ThN, Ti7, Tl2, Tiv, TkW, Tif, TiF, Tip, Tiz, ThA, ThU;
|
cannam@167
|
1561 E ThE, ThQ;
|
cannam@167
|
1562 {
|
cannam@167
|
1563 E Tg8, TkT, Tgj, TkU, Tgd, Tgi;
|
cannam@167
|
1564 Tg8 = Tg4 + Tg7;
|
cannam@167
|
1565 TkT = TkE - TkD;
|
cannam@167
|
1566 Tgd = Tg9 + Tgc;
|
cannam@167
|
1567 Tgi = Tge - Tgh;
|
cannam@167
|
1568 Tgj = Tgd + Tgi;
|
cannam@167
|
1569 TkU = Tgi - Tgd;
|
cannam@167
|
1570 Tgk = FNMS(KP707106781, Tgj, Tg8);
|
cannam@167
|
1571 Tl1 = FNMS(KP707106781, TkU, TkT);
|
cannam@167
|
1572 ThG = FMA(KP707106781, Tgj, Tg8);
|
cannam@167
|
1573 TkV = FMA(KP707106781, TkU, TkT);
|
cannam@167
|
1574 }
|
cannam@167
|
1575 {
|
cannam@167
|
1576 E ThW, TkF, ThZ, TkG, ThX, ThY;
|
cannam@167
|
1577 ThW = Tg4 - Tg7;
|
cannam@167
|
1578 TkF = TkD + TkE;
|
cannam@167
|
1579 ThX = Tgc - Tg9;
|
cannam@167
|
1580 ThY = Tge + Tgh;
|
cannam@167
|
1581 ThZ = ThX - ThY;
|
cannam@167
|
1582 TkG = ThX + ThY;
|
cannam@167
|
1583 Ti0 = FMA(KP707106781, ThZ, ThW);
|
cannam@167
|
1584 TkN = FNMS(KP707106781, TkG, TkF);
|
cannam@167
|
1585 Tis = FNMS(KP707106781, ThZ, ThW);
|
cannam@167
|
1586 TkH = FMA(KP707106781, TkG, TkF);
|
cannam@167
|
1587 }
|
cannam@167
|
1588 {
|
cannam@167
|
1589 E Tgv, ThH, TgG, ThI;
|
cannam@167
|
1590 {
|
cannam@167
|
1591 E Tgp, Tgu, TgA, TgF;
|
cannam@167
|
1592 Tgp = Tgn + Tgo;
|
cannam@167
|
1593 Tgu = Tgq + Tgt;
|
cannam@167
|
1594 Tgv = FNMS(KP414213562, Tgu, Tgp);
|
cannam@167
|
1595 ThH = FMA(KP414213562, Tgp, Tgu);
|
cannam@167
|
1596 TgA = Tgy + Tgz;
|
cannam@167
|
1597 TgF = TgB + TgE;
|
cannam@167
|
1598 TgG = FMA(KP414213562, TgF, TgA);
|
cannam@167
|
1599 ThI = FNMS(KP414213562, TgA, TgF);
|
cannam@167
|
1600 }
|
cannam@167
|
1601 TgH = Tgv - TgG;
|
cannam@167
|
1602 TkO = ThI - ThH;
|
cannam@167
|
1603 ThJ = ThH + ThI;
|
cannam@167
|
1604 TkI = Tgv + TgG;
|
cannam@167
|
1605 }
|
cannam@167
|
1606 {
|
cannam@167
|
1607 E Tii, TiB, Til, TiA;
|
cannam@167
|
1608 {
|
cannam@167
|
1609 E Tig, Tih, Tij, Tik;
|
cannam@167
|
1610 Tig = Thr - Thu;
|
cannam@167
|
1611 Tih = Tho - Thj;
|
cannam@167
|
1612 Tii = FNMS(KP707106781, Tih, Tig);
|
cannam@167
|
1613 TiB = FMA(KP707106781, Tih, Tig);
|
cannam@167
|
1614 Tij = Thc - Thd;
|
cannam@167
|
1615 Tik = Thw - Thx;
|
cannam@167
|
1616 Til = FNMS(KP707106781, Tik, Tij);
|
cannam@167
|
1617 TiA = FMA(KP707106781, Tik, Tij);
|
cannam@167
|
1618 }
|
cannam@167
|
1619 Tim = FNMS(KP668178637, Til, Tii);
|
cannam@167
|
1620 TiG = FMA(KP198912367, TiA, TiB);
|
cannam@167
|
1621 Tiq = FMA(KP668178637, Tii, Til);
|
cannam@167
|
1622 TiC = FNMS(KP198912367, TiB, TiA);
|
cannam@167
|
1623 }
|
cannam@167
|
1624 {
|
cannam@167
|
1625 E TgZ, ThM, Th8, ThL;
|
cannam@167
|
1626 {
|
cannam@167
|
1627 E TgN, TgY, Th4, Th7;
|
cannam@167
|
1628 TgN = TgL + TgM;
|
cannam@167
|
1629 TgY = TgS + TgX;
|
cannam@167
|
1630 TgZ = FNMS(KP707106781, TgY, TgN);
|
cannam@167
|
1631 ThM = FMA(KP707106781, TgY, TgN);
|
cannam@167
|
1632 Th4 = Th0 + Th3;
|
cannam@167
|
1633 Th7 = Th5 + Th6;
|
cannam@167
|
1634 Th8 = FNMS(KP707106781, Th7, Th4);
|
cannam@167
|
1635 ThL = FMA(KP707106781, Th7, Th4);
|
cannam@167
|
1636 }
|
cannam@167
|
1637 Th9 = FMA(KP668178637, Th8, TgZ);
|
cannam@167
|
1638 ThT = FNMS(KP198912367, ThL, ThM);
|
cannam@167
|
1639 ThD = FNMS(KP668178637, TgZ, Th8);
|
cannam@167
|
1640 ThN = FMA(KP198912367, ThM, ThL);
|
cannam@167
|
1641 }
|
cannam@167
|
1642 {
|
cannam@167
|
1643 E Ti3, Tit, Ti6, Tiu;
|
cannam@167
|
1644 {
|
cannam@167
|
1645 E Ti1, Ti2, Ti4, Ti5;
|
cannam@167
|
1646 Ti1 = Tgn - Tgo;
|
cannam@167
|
1647 Ti2 = Tgq - Tgt;
|
cannam@167
|
1648 Ti3 = FMA(KP414213562, Ti2, Ti1);
|
cannam@167
|
1649 Tit = FNMS(KP414213562, Ti1, Ti2);
|
cannam@167
|
1650 Ti4 = Tgy - Tgz;
|
cannam@167
|
1651 Ti5 = TgB - TgE;
|
cannam@167
|
1652 Ti6 = FNMS(KP414213562, Ti5, Ti4);
|
cannam@167
|
1653 Tiu = FMA(KP414213562, Ti4, Ti5);
|
cannam@167
|
1654 }
|
cannam@167
|
1655 Ti7 = Ti3 - Ti6;
|
cannam@167
|
1656 Tl2 = Ti3 + Ti6;
|
cannam@167
|
1657 Tiv = Tit + Tiu;
|
cannam@167
|
1658 TkW = Tiu - Tit;
|
cannam@167
|
1659 }
|
cannam@167
|
1660 {
|
cannam@167
|
1661 E Tib, Tiy, Tie, Tix;
|
cannam@167
|
1662 {
|
cannam@167
|
1663 E Ti9, Tia, Tic, Tid;
|
cannam@167
|
1664 Ti9 = Th0 - Th3;
|
cannam@167
|
1665 Tia = TgX - TgS;
|
cannam@167
|
1666 Tib = FNMS(KP707106781, Tia, Ti9);
|
cannam@167
|
1667 Tiy = FMA(KP707106781, Tia, Ti9);
|
cannam@167
|
1668 Tic = TgL - TgM;
|
cannam@167
|
1669 Tid = Th5 - Th6;
|
cannam@167
|
1670 Tie = FNMS(KP707106781, Tid, Tic);
|
cannam@167
|
1671 Tix = FMA(KP707106781, Tid, Tic);
|
cannam@167
|
1672 }
|
cannam@167
|
1673 Tif = FMA(KP668178637, Tie, Tib);
|
cannam@167
|
1674 TiF = FNMS(KP198912367, Tix, Tiy);
|
cannam@167
|
1675 Tip = FNMS(KP668178637, Tib, Tie);
|
cannam@167
|
1676 Tiz = FMA(KP198912367, Tiy, Tix);
|
cannam@167
|
1677 }
|
cannam@167
|
1678 {
|
cannam@167
|
1679 E Thq, ThP, Thz, ThO;
|
cannam@167
|
1680 {
|
cannam@167
|
1681 E The, Thp, Thv, Thy;
|
cannam@167
|
1682 The = Thc + Thd;
|
cannam@167
|
1683 Thp = Thj + Tho;
|
cannam@167
|
1684 Thq = FNMS(KP707106781, Thp, The);
|
cannam@167
|
1685 ThP = FMA(KP707106781, Thp, The);
|
cannam@167
|
1686 Thv = Thr + Thu;
|
cannam@167
|
1687 Thy = Thw + Thx;
|
cannam@167
|
1688 Thz = FNMS(KP707106781, Thy, Thv);
|
cannam@167
|
1689 ThO = FMA(KP707106781, Thy, Thv);
|
cannam@167
|
1690 }
|
cannam@167
|
1691 ThA = FNMS(KP668178637, Thz, Thq);
|
cannam@167
|
1692 ThU = FMA(KP198912367, ThO, ThP);
|
cannam@167
|
1693 ThE = FMA(KP668178637, Thq, Thz);
|
cannam@167
|
1694 ThQ = FNMS(KP198912367, ThP, ThO);
|
cannam@167
|
1695 }
|
cannam@167
|
1696 {
|
cannam@167
|
1697 E TgI, ThB, TkP, TkQ;
|
cannam@167
|
1698 TgI = FMA(KP923879532, TgH, Tgk);
|
cannam@167
|
1699 ThB = Th9 - ThA;
|
cannam@167
|
1700 ri[WS(rs, 42)] = FNMS(KP831469612, ThB, TgI);
|
cannam@167
|
1701 ri[WS(rs, 10)] = FMA(KP831469612, ThB, TgI);
|
cannam@167
|
1702 TkP = FMA(KP923879532, TkO, TkN);
|
cannam@167
|
1703 TkQ = ThE - ThD;
|
cannam@167
|
1704 ii[WS(rs, 10)] = FMA(KP831469612, TkQ, TkP);
|
cannam@167
|
1705 ii[WS(rs, 42)] = FNMS(KP831469612, TkQ, TkP);
|
cannam@167
|
1706 }
|
cannam@167
|
1707 {
|
cannam@167
|
1708 E ThC, ThF, TkR, TkS;
|
cannam@167
|
1709 ThC = FNMS(KP923879532, TgH, Tgk);
|
cannam@167
|
1710 ThF = ThD + ThE;
|
cannam@167
|
1711 ri[WS(rs, 26)] = FNMS(KP831469612, ThF, ThC);
|
cannam@167
|
1712 ri[WS(rs, 58)] = FMA(KP831469612, ThF, ThC);
|
cannam@167
|
1713 TkR = FNMS(KP923879532, TkO, TkN);
|
cannam@167
|
1714 TkS = Th9 + ThA;
|
cannam@167
|
1715 ii[WS(rs, 26)] = FNMS(KP831469612, TkS, TkR);
|
cannam@167
|
1716 ii[WS(rs, 58)] = FMA(KP831469612, TkS, TkR);
|
cannam@167
|
1717 }
|
cannam@167
|
1718 {
|
cannam@167
|
1719 E ThK, ThR, TkJ, TkK;
|
cannam@167
|
1720 ThK = FMA(KP923879532, ThJ, ThG);
|
cannam@167
|
1721 ThR = ThN + ThQ;
|
cannam@167
|
1722 ri[WS(rs, 34)] = FNMS(KP980785280, ThR, ThK);
|
cannam@167
|
1723 ri[WS(rs, 2)] = FMA(KP980785280, ThR, ThK);
|
cannam@167
|
1724 TkJ = FMA(KP923879532, TkI, TkH);
|
cannam@167
|
1725 TkK = ThT + ThU;
|
cannam@167
|
1726 ii[WS(rs, 2)] = FMA(KP980785280, TkK, TkJ);
|
cannam@167
|
1727 ii[WS(rs, 34)] = FNMS(KP980785280, TkK, TkJ);
|
cannam@167
|
1728 }
|
cannam@167
|
1729 {
|
cannam@167
|
1730 E ThS, ThV, TkL, TkM;
|
cannam@167
|
1731 ThS = FNMS(KP923879532, ThJ, ThG);
|
cannam@167
|
1732 ThV = ThT - ThU;
|
cannam@167
|
1733 ri[WS(rs, 50)] = FNMS(KP980785280, ThV, ThS);
|
cannam@167
|
1734 ri[WS(rs, 18)] = FMA(KP980785280, ThV, ThS);
|
cannam@167
|
1735 TkL = FNMS(KP923879532, TkI, TkH);
|
cannam@167
|
1736 TkM = ThQ - ThN;
|
cannam@167
|
1737 ii[WS(rs, 18)] = FMA(KP980785280, TkM, TkL);
|
cannam@167
|
1738 ii[WS(rs, 50)] = FNMS(KP980785280, TkM, TkL);
|
cannam@167
|
1739 }
|
cannam@167
|
1740 {
|
cannam@167
|
1741 E Ti8, Tin, TkX, TkY;
|
cannam@167
|
1742 Ti8 = FMA(KP923879532, Ti7, Ti0);
|
cannam@167
|
1743 Tin = Tif + Tim;
|
cannam@167
|
1744 ri[WS(rs, 38)] = FNMS(KP831469612, Tin, Ti8);
|
cannam@167
|
1745 ri[WS(rs, 6)] = FMA(KP831469612, Tin, Ti8);
|
cannam@167
|
1746 TkX = FMA(KP923879532, TkW, TkV);
|
cannam@167
|
1747 TkY = Tip + Tiq;
|
cannam@167
|
1748 ii[WS(rs, 6)] = FMA(KP831469612, TkY, TkX);
|
cannam@167
|
1749 ii[WS(rs, 38)] = FNMS(KP831469612, TkY, TkX);
|
cannam@167
|
1750 }
|
cannam@167
|
1751 {
|
cannam@167
|
1752 E Tio, Tir, TkZ, Tl0;
|
cannam@167
|
1753 Tio = FNMS(KP923879532, Ti7, Ti0);
|
cannam@167
|
1754 Tir = Tip - Tiq;
|
cannam@167
|
1755 ri[WS(rs, 54)] = FNMS(KP831469612, Tir, Tio);
|
cannam@167
|
1756 ri[WS(rs, 22)] = FMA(KP831469612, Tir, Tio);
|
cannam@167
|
1757 TkZ = FNMS(KP923879532, TkW, TkV);
|
cannam@167
|
1758 Tl0 = Tim - Tif;
|
cannam@167
|
1759 ii[WS(rs, 22)] = FMA(KP831469612, Tl0, TkZ);
|
cannam@167
|
1760 ii[WS(rs, 54)] = FNMS(KP831469612, Tl0, TkZ);
|
cannam@167
|
1761 }
|
cannam@167
|
1762 {
|
cannam@167
|
1763 E Tiw, TiD, Tl3, Tl4;
|
cannam@167
|
1764 Tiw = FNMS(KP923879532, Tiv, Tis);
|
cannam@167
|
1765 TiD = Tiz - TiC;
|
cannam@167
|
1766 ri[WS(rs, 46)] = FNMS(KP980785280, TiD, Tiw);
|
cannam@167
|
1767 ri[WS(rs, 14)] = FMA(KP980785280, TiD, Tiw);
|
cannam@167
|
1768 Tl3 = FNMS(KP923879532, Tl2, Tl1);
|
cannam@167
|
1769 Tl4 = TiG - TiF;
|
cannam@167
|
1770 ii[WS(rs, 14)] = FMA(KP980785280, Tl4, Tl3);
|
cannam@167
|
1771 ii[WS(rs, 46)] = FNMS(KP980785280, Tl4, Tl3);
|
cannam@167
|
1772 }
|
cannam@167
|
1773 {
|
cannam@167
|
1774 E TiE, TiH, Tl5, Tl6;
|
cannam@167
|
1775 TiE = FMA(KP923879532, Tiv, Tis);
|
cannam@167
|
1776 TiH = TiF + TiG;
|
cannam@167
|
1777 ri[WS(rs, 30)] = FNMS(KP980785280, TiH, TiE);
|
cannam@167
|
1778 ri[WS(rs, 62)] = FMA(KP980785280, TiH, TiE);
|
cannam@167
|
1779 Tl5 = FMA(KP923879532, Tl2, Tl1);
|
cannam@167
|
1780 Tl6 = Tiz + TiC;
|
cannam@167
|
1781 ii[WS(rs, 30)] = FNMS(KP980785280, Tl6, Tl5);
|
cannam@167
|
1782 ii[WS(rs, 62)] = FMA(KP980785280, Tl6, Tl5);
|
cannam@167
|
1783 }
|
cannam@167
|
1784 }
|
cannam@167
|
1785 {
|
cannam@167
|
1786 E Tar, TlO, TcT, TlI, TbB, Td3, TcN, TcX, Tdw, TdQ, TdA, TdM, Tdp, TdP, Tdz;
|
cannam@167
|
1787 E TdJ, Tdh, Tm2, TdF, TlW, TcK, Td4, TcO, Td0, T9i, TlV, Tm1, TcQ, Tda, TlH;
|
cannam@167
|
1788 E TlN, TdC;
|
cannam@167
|
1789 {
|
cannam@167
|
1790 E T9R, TcR, Taq, TcS;
|
cannam@167
|
1791 {
|
cannam@167
|
1792 E T9F, T9Q, Tae, Tap;
|
cannam@167
|
1793 T9F = FNMS(KP707106781, T9E, T9p);
|
cannam@167
|
1794 T9Q = FNMS(KP707106781, T9P, T9M);
|
cannam@167
|
1795 T9R = FNMS(KP668178637, T9Q, T9F);
|
cannam@167
|
1796 TcR = FMA(KP668178637, T9F, T9Q);
|
cannam@167
|
1797 Tae = FNMS(KP707106781, Tad, T9Y);
|
cannam@167
|
1798 Tap = FNMS(KP707106781, Tao, Tal);
|
cannam@167
|
1799 Taq = FMA(KP668178637, Tap, Tae);
|
cannam@167
|
1800 TcS = FNMS(KP668178637, Tae, Tap);
|
cannam@167
|
1801 }
|
cannam@167
|
1802 Tar = T9R - Taq;
|
cannam@167
|
1803 TlO = TcS - TcR;
|
cannam@167
|
1804 TcT = TcR + TcS;
|
cannam@167
|
1805 TlI = T9R + Taq;
|
cannam@167
|
1806 }
|
cannam@167
|
1807 {
|
cannam@167
|
1808 E Tbl, TcW, TbA, TcV;
|
cannam@167
|
1809 {
|
cannam@167
|
1810 E TaP, Tbk, Tbw, Tbz;
|
cannam@167
|
1811 TaP = FNMS(KP707106781, TaO, Taz);
|
cannam@167
|
1812 Tbk = Tb4 - Tbj;
|
cannam@167
|
1813 Tbl = FNMS(KP923879532, Tbk, TaP);
|
cannam@167
|
1814 TcW = FMA(KP923879532, Tbk, TaP);
|
cannam@167
|
1815 Tbw = FNMS(KP707106781, Tbv, Tbs);
|
cannam@167
|
1816 Tbz = Tbx - Tby;
|
cannam@167
|
1817 TbA = FNMS(KP923879532, Tbz, Tbw);
|
cannam@167
|
1818 TcV = FMA(KP923879532, Tbz, Tbw);
|
cannam@167
|
1819 }
|
cannam@167
|
1820 TbB = FMA(KP534511135, TbA, Tbl);
|
cannam@167
|
1821 Td3 = FNMS(KP303346683, TcV, TcW);
|
cannam@167
|
1822 TcN = FNMS(KP534511135, Tbl, TbA);
|
cannam@167
|
1823 TcX = FMA(KP303346683, TcW, TcV);
|
cannam@167
|
1824 }
|
cannam@167
|
1825 {
|
cannam@167
|
1826 E Tds, TdL, Tdv, TdK;
|
cannam@167
|
1827 {
|
cannam@167
|
1828 E Tdq, Tdr, Tdt, Tdu;
|
cannam@167
|
1829 Tdq = FMA(KP707106781, TcE, TcB);
|
cannam@167
|
1830 Tdr = Tcs + Tcd;
|
cannam@167
|
1831 Tds = FNMS(KP923879532, Tdr, Tdq);
|
cannam@167
|
1832 TdL = FMA(KP923879532, Tdr, Tdq);
|
cannam@167
|
1833 Tdt = FMA(KP707106781, TbX, TbI);
|
cannam@167
|
1834 Tdu = TcG + TcH;
|
cannam@167
|
1835 Tdv = FNMS(KP923879532, Tdu, Tdt);
|
cannam@167
|
1836 TdK = FMA(KP923879532, Tdu, Tdt);
|
cannam@167
|
1837 }
|
cannam@167
|
1838 Tdw = FNMS(KP820678790, Tdv, Tds);
|
cannam@167
|
1839 TdQ = FMA(KP098491403, TdK, TdL);
|
cannam@167
|
1840 TdA = FMA(KP820678790, Tds, Tdv);
|
cannam@167
|
1841 TdM = FNMS(KP098491403, TdL, TdK);
|
cannam@167
|
1842 }
|
cannam@167
|
1843 {
|
cannam@167
|
1844 E Tdl, TdI, Tdo, TdH;
|
cannam@167
|
1845 {
|
cannam@167
|
1846 E Tdj, Tdk, Tdm, Tdn;
|
cannam@167
|
1847 Tdj = FMA(KP707106781, Tbv, Tbs);
|
cannam@167
|
1848 Tdk = Tbj + Tb4;
|
cannam@167
|
1849 Tdl = FNMS(KP923879532, Tdk, Tdj);
|
cannam@167
|
1850 TdI = FMA(KP923879532, Tdk, Tdj);
|
cannam@167
|
1851 Tdm = FMA(KP707106781, TaO, Taz);
|
cannam@167
|
1852 Tdn = Tbx + Tby;
|
cannam@167
|
1853 Tdo = FNMS(KP923879532, Tdn, Tdm);
|
cannam@167
|
1854 TdH = FMA(KP923879532, Tdn, Tdm);
|
cannam@167
|
1855 }
|
cannam@167
|
1856 Tdp = FMA(KP820678790, Tdo, Tdl);
|
cannam@167
|
1857 TdP = FNMS(KP098491403, TdH, TdI);
|
cannam@167
|
1858 Tdz = FNMS(KP820678790, Tdl, Tdo);
|
cannam@167
|
1859 TdJ = FMA(KP098491403, TdI, TdH);
|
cannam@167
|
1860 }
|
cannam@167
|
1861 {
|
cannam@167
|
1862 E Tdd, TdD, Tdg, TdE;
|
cannam@167
|
1863 {
|
cannam@167
|
1864 E Tdb, Tdc, Tde, Tdf;
|
cannam@167
|
1865 Tdb = FMA(KP707106781, T9E, T9p);
|
cannam@167
|
1866 Tdc = FMA(KP707106781, T9P, T9M);
|
cannam@167
|
1867 Tdd = FMA(KP198912367, Tdc, Tdb);
|
cannam@167
|
1868 TdD = FNMS(KP198912367, Tdb, Tdc);
|
cannam@167
|
1869 Tde = FMA(KP707106781, Tad, T9Y);
|
cannam@167
|
1870 Tdf = FMA(KP707106781, Tao, Tal);
|
cannam@167
|
1871 Tdg = FNMS(KP198912367, Tdf, Tde);
|
cannam@167
|
1872 TdE = FMA(KP198912367, Tde, Tdf);
|
cannam@167
|
1873 }
|
cannam@167
|
1874 Tdh = Tdd - Tdg;
|
cannam@167
|
1875 Tm2 = Tdd + Tdg;
|
cannam@167
|
1876 TdF = TdD + TdE;
|
cannam@167
|
1877 TlW = TdE - TdD;
|
cannam@167
|
1878 }
|
cannam@167
|
1879 {
|
cannam@167
|
1880 E Tcu, TcZ, TcJ, TcY;
|
cannam@167
|
1881 {
|
cannam@167
|
1882 E TbY, Tct, TcF, TcI;
|
cannam@167
|
1883 TbY = FNMS(KP707106781, TbX, TbI);
|
cannam@167
|
1884 Tct = Tcd - Tcs;
|
cannam@167
|
1885 Tcu = FNMS(KP923879532, Tct, TbY);
|
cannam@167
|
1886 TcZ = FMA(KP923879532, Tct, TbY);
|
cannam@167
|
1887 TcF = FNMS(KP707106781, TcE, TcB);
|
cannam@167
|
1888 TcI = TcG - TcH;
|
cannam@167
|
1889 TcJ = FNMS(KP923879532, TcI, TcF);
|
cannam@167
|
1890 TcY = FMA(KP923879532, TcI, TcF);
|
cannam@167
|
1891 }
|
cannam@167
|
1892 TcK = FNMS(KP534511135, TcJ, Tcu);
|
cannam@167
|
1893 Td4 = FMA(KP303346683, TcY, TcZ);
|
cannam@167
|
1894 TcO = FMA(KP534511135, Tcu, TcJ);
|
cannam@167
|
1895 Td0 = FNMS(KP303346683, TcZ, TcY);
|
cannam@167
|
1896 }
|
cannam@167
|
1897 {
|
cannam@167
|
1898 E T8M, Td6, TlF, TlT, T9h, TlU, Td9, TlG, T8L, TlE;
|
cannam@167
|
1899 T8L = T8D - T8K;
|
cannam@167
|
1900 T8M = FMA(KP707106781, T8L, T8w);
|
cannam@167
|
1901 Td6 = FNMS(KP707106781, T8L, T8w);
|
cannam@167
|
1902 TlE = TdU - TdT;
|
cannam@167
|
1903 TlF = FMA(KP707106781, TlE, TlD);
|
cannam@167
|
1904 TlT = FNMS(KP707106781, TlE, TlD);
|
cannam@167
|
1905 {
|
cannam@167
|
1906 E T91, T9g, Td7, Td8;
|
cannam@167
|
1907 T91 = FMA(KP414213562, T90, T8T);
|
cannam@167
|
1908 T9g = FNMS(KP414213562, T9f, T98);
|
cannam@167
|
1909 T9h = T91 - T9g;
|
cannam@167
|
1910 TlU = T91 + T9g;
|
cannam@167
|
1911 Td7 = FNMS(KP414213562, T8T, T90);
|
cannam@167
|
1912 Td8 = FMA(KP414213562, T98, T9f);
|
cannam@167
|
1913 Td9 = Td7 + Td8;
|
cannam@167
|
1914 TlG = Td8 - Td7;
|
cannam@167
|
1915 }
|
cannam@167
|
1916 T9i = FNMS(KP923879532, T9h, T8M);
|
cannam@167
|
1917 TlV = FNMS(KP923879532, TlU, TlT);
|
cannam@167
|
1918 Tm1 = FMA(KP923879532, TlU, TlT);
|
cannam@167
|
1919 TcQ = FMA(KP923879532, T9h, T8M);
|
cannam@167
|
1920 Tda = FNMS(KP923879532, Td9, Td6);
|
cannam@167
|
1921 TlH = FMA(KP923879532, TlG, TlF);
|
cannam@167
|
1922 TlN = FNMS(KP923879532, TlG, TlF);
|
cannam@167
|
1923 TdC = FMA(KP923879532, Td9, Td6);
|
cannam@167
|
1924 }
|
cannam@167
|
1925 {
|
cannam@167
|
1926 E Tas, TcL, TlP, TlQ;
|
cannam@167
|
1927 Tas = FMA(KP831469612, Tar, T9i);
|
cannam@167
|
1928 TcL = TbB - TcK;
|
cannam@167
|
1929 ri[WS(rs, 43)] = FNMS(KP881921264, TcL, Tas);
|
cannam@167
|
1930 ri[WS(rs, 11)] = FMA(KP881921264, TcL, Tas);
|
cannam@167
|
1931 TlP = FMA(KP831469612, TlO, TlN);
|
cannam@167
|
1932 TlQ = TcO - TcN;
|
cannam@167
|
1933 ii[WS(rs, 11)] = FMA(KP881921264, TlQ, TlP);
|
cannam@167
|
1934 ii[WS(rs, 43)] = FNMS(KP881921264, TlQ, TlP);
|
cannam@167
|
1935 }
|
cannam@167
|
1936 {
|
cannam@167
|
1937 E TcM, TcP, TlR, TlS;
|
cannam@167
|
1938 TcM = FNMS(KP831469612, Tar, T9i);
|
cannam@167
|
1939 TcP = TcN + TcO;
|
cannam@167
|
1940 ri[WS(rs, 27)] = FNMS(KP881921264, TcP, TcM);
|
cannam@167
|
1941 ri[WS(rs, 59)] = FMA(KP881921264, TcP, TcM);
|
cannam@167
|
1942 TlR = FNMS(KP831469612, TlO, TlN);
|
cannam@167
|
1943 TlS = TbB + TcK;
|
cannam@167
|
1944 ii[WS(rs, 27)] = FNMS(KP881921264, TlS, TlR);
|
cannam@167
|
1945 ii[WS(rs, 59)] = FMA(KP881921264, TlS, TlR);
|
cannam@167
|
1946 }
|
cannam@167
|
1947 {
|
cannam@167
|
1948 E TcU, Td1, TlJ, TlK;
|
cannam@167
|
1949 TcU = FMA(KP831469612, TcT, TcQ);
|
cannam@167
|
1950 Td1 = TcX + Td0;
|
cannam@167
|
1951 ri[WS(rs, 35)] = FNMS(KP956940335, Td1, TcU);
|
cannam@167
|
1952 ri[WS(rs, 3)] = FMA(KP956940335, Td1, TcU);
|
cannam@167
|
1953 TlJ = FMA(KP831469612, TlI, TlH);
|
cannam@167
|
1954 TlK = Td3 + Td4;
|
cannam@167
|
1955 ii[WS(rs, 3)] = FMA(KP956940335, TlK, TlJ);
|
cannam@167
|
1956 ii[WS(rs, 35)] = FNMS(KP956940335, TlK, TlJ);
|
cannam@167
|
1957 }
|
cannam@167
|
1958 {
|
cannam@167
|
1959 E Td2, Td5, TlL, TlM;
|
cannam@167
|
1960 Td2 = FNMS(KP831469612, TcT, TcQ);
|
cannam@167
|
1961 Td5 = Td3 - Td4;
|
cannam@167
|
1962 ri[WS(rs, 51)] = FNMS(KP956940335, Td5, Td2);
|
cannam@167
|
1963 ri[WS(rs, 19)] = FMA(KP956940335, Td5, Td2);
|
cannam@167
|
1964 TlL = FNMS(KP831469612, TlI, TlH);
|
cannam@167
|
1965 TlM = Td0 - TcX;
|
cannam@167
|
1966 ii[WS(rs, 19)] = FMA(KP956940335, TlM, TlL);
|
cannam@167
|
1967 ii[WS(rs, 51)] = FNMS(KP956940335, TlM, TlL);
|
cannam@167
|
1968 }
|
cannam@167
|
1969 {
|
cannam@167
|
1970 E Tdi, Tdx, TlX, TlY;
|
cannam@167
|
1971 Tdi = FMA(KP980785280, Tdh, Tda);
|
cannam@167
|
1972 Tdx = Tdp + Tdw;
|
cannam@167
|
1973 ri[WS(rs, 39)] = FNMS(KP773010453, Tdx, Tdi);
|
cannam@167
|
1974 ri[WS(rs, 7)] = FMA(KP773010453, Tdx, Tdi);
|
cannam@167
|
1975 TlX = FMA(KP980785280, TlW, TlV);
|
cannam@167
|
1976 TlY = Tdz + TdA;
|
cannam@167
|
1977 ii[WS(rs, 7)] = FMA(KP773010453, TlY, TlX);
|
cannam@167
|
1978 ii[WS(rs, 39)] = FNMS(KP773010453, TlY, TlX);
|
cannam@167
|
1979 }
|
cannam@167
|
1980 {
|
cannam@167
|
1981 E Tdy, TdB, TlZ, Tm0;
|
cannam@167
|
1982 Tdy = FNMS(KP980785280, Tdh, Tda);
|
cannam@167
|
1983 TdB = Tdz - TdA;
|
cannam@167
|
1984 ri[WS(rs, 55)] = FNMS(KP773010453, TdB, Tdy);
|
cannam@167
|
1985 ri[WS(rs, 23)] = FMA(KP773010453, TdB, Tdy);
|
cannam@167
|
1986 TlZ = FNMS(KP980785280, TlW, TlV);
|
cannam@167
|
1987 Tm0 = Tdw - Tdp;
|
cannam@167
|
1988 ii[WS(rs, 23)] = FMA(KP773010453, Tm0, TlZ);
|
cannam@167
|
1989 ii[WS(rs, 55)] = FNMS(KP773010453, Tm0, TlZ);
|
cannam@167
|
1990 }
|
cannam@167
|
1991 {
|
cannam@167
|
1992 E TdG, TdN, Tm3, Tm4;
|
cannam@167
|
1993 TdG = FNMS(KP980785280, TdF, TdC);
|
cannam@167
|
1994 TdN = TdJ - TdM;
|
cannam@167
|
1995 ri[WS(rs, 47)] = FNMS(KP995184726, TdN, TdG);
|
cannam@167
|
1996 ri[WS(rs, 15)] = FMA(KP995184726, TdN, TdG);
|
cannam@167
|
1997 Tm3 = FNMS(KP980785280, Tm2, Tm1);
|
cannam@167
|
1998 Tm4 = TdQ - TdP;
|
cannam@167
|
1999 ii[WS(rs, 15)] = FMA(KP995184726, Tm4, Tm3);
|
cannam@167
|
2000 ii[WS(rs, 47)] = FNMS(KP995184726, Tm4, Tm3);
|
cannam@167
|
2001 }
|
cannam@167
|
2002 {
|
cannam@167
|
2003 E TdO, TdR, Tm5, Tm6;
|
cannam@167
|
2004 TdO = FMA(KP980785280, TdF, TdC);
|
cannam@167
|
2005 TdR = TdP + TdQ;
|
cannam@167
|
2006 ri[WS(rs, 31)] = FNMS(KP995184726, TdR, TdO);
|
cannam@167
|
2007 ri[WS(rs, 63)] = FMA(KP995184726, TdR, TdO);
|
cannam@167
|
2008 Tm5 = FMA(KP980785280, Tm2, Tm1);
|
cannam@167
|
2009 Tm6 = TdJ + TdM;
|
cannam@167
|
2010 ii[WS(rs, 31)] = FNMS(KP995184726, Tm6, Tm5);
|
cannam@167
|
2011 ii[WS(rs, 63)] = FMA(KP995184726, Tm6, Tm5);
|
cannam@167
|
2012 }
|
cannam@167
|
2013 }
|
cannam@167
|
2014 {
|
cannam@167
|
2015 E Tej, Tlk, Tf5, Tle, TeD, Tff, TeZ, Tf9, TfI, Tg2, TfM, TfY, TfB, Tg1, TfL;
|
cannam@167
|
2016 E TfV, Tft, Tly, TfR, Tls, TeW, Tfg, Tf0, Tfc, Te4, Tlr, Tlx, Tf2, Tfm, Tld;
|
cannam@167
|
2017 E Tlj, TfO;
|
cannam@167
|
2018 {
|
cannam@167
|
2019 E Teb, Tf3, Tei, Tf4;
|
cannam@167
|
2020 {
|
cannam@167
|
2021 E Te7, Tea, Tee, Teh;
|
cannam@167
|
2022 Te7 = FMA(KP707106781, Te6, Te5);
|
cannam@167
|
2023 Tea = FMA(KP707106781, Te9, Te8);
|
cannam@167
|
2024 Teb = FNMS(KP198912367, Tea, Te7);
|
cannam@167
|
2025 Tf3 = FMA(KP198912367, Te7, Tea);
|
cannam@167
|
2026 Tee = FMA(KP707106781, Ted, Tec);
|
cannam@167
|
2027 Teh = FMA(KP707106781, Teg, Tef);
|
cannam@167
|
2028 Tei = FMA(KP198912367, Teh, Tee);
|
cannam@167
|
2029 Tf4 = FNMS(KP198912367, Tee, Teh);
|
cannam@167
|
2030 }
|
cannam@167
|
2031 Tej = Teb - Tei;
|
cannam@167
|
2032 Tlk = Tf4 - Tf3;
|
cannam@167
|
2033 Tf5 = Tf3 + Tf4;
|
cannam@167
|
2034 Tle = Teb + Tei;
|
cannam@167
|
2035 }
|
cannam@167
|
2036 {
|
cannam@167
|
2037 E Tev, Tf8, TeC, Tf7;
|
cannam@167
|
2038 {
|
cannam@167
|
2039 E Ten, Teu, Tey, TeB;
|
cannam@167
|
2040 Ten = FMA(KP707106781, Tem, Tel);
|
cannam@167
|
2041 Teu = Teq + Tet;
|
cannam@167
|
2042 Tev = FNMS(KP923879532, Teu, Ten);
|
cannam@167
|
2043 Tf8 = FMA(KP923879532, Teu, Ten);
|
cannam@167
|
2044 Tey = FMA(KP707106781, Tex, Tew);
|
cannam@167
|
2045 TeB = Tez + TeA;
|
cannam@167
|
2046 TeC = FNMS(KP923879532, TeB, Tey);
|
cannam@167
|
2047 Tf7 = FMA(KP923879532, TeB, Tey);
|
cannam@167
|
2048 }
|
cannam@167
|
2049 TeD = FMA(KP820678790, TeC, Tev);
|
cannam@167
|
2050 Tff = FNMS(KP098491403, Tf7, Tf8);
|
cannam@167
|
2051 TeZ = FNMS(KP820678790, Tev, TeC);
|
cannam@167
|
2052 Tf9 = FMA(KP098491403, Tf8, Tf7);
|
cannam@167
|
2053 }
|
cannam@167
|
2054 {
|
cannam@167
|
2055 E TfE, TfX, TfH, TfW;
|
cannam@167
|
2056 {
|
cannam@167
|
2057 E TfC, TfD, TfF, TfG;
|
cannam@167
|
2058 TfC = FNMS(KP707106781, TeQ, TeP);
|
cannam@167
|
2059 TfD = TeM - TeJ;
|
cannam@167
|
2060 TfE = FNMS(KP923879532, TfD, TfC);
|
cannam@167
|
2061 TfX = FMA(KP923879532, TfD, TfC);
|
cannam@167
|
2062 TfF = FNMS(KP707106781, TeF, TeE);
|
cannam@167
|
2063 TfG = TeS - TeT;
|
cannam@167
|
2064 TfH = FNMS(KP923879532, TfG, TfF);
|
cannam@167
|
2065 TfW = FMA(KP923879532, TfG, TfF);
|
cannam@167
|
2066 }
|
cannam@167
|
2067 TfI = FNMS(KP534511135, TfH, TfE);
|
cannam@167
|
2068 Tg2 = FMA(KP303346683, TfW, TfX);
|
cannam@167
|
2069 TfM = FMA(KP534511135, TfE, TfH);
|
cannam@167
|
2070 TfY = FNMS(KP303346683, TfX, TfW);
|
cannam@167
|
2071 }
|
cannam@167
|
2072 {
|
cannam@167
|
2073 E Tfx, TfU, TfA, TfT;
|
cannam@167
|
2074 {
|
cannam@167
|
2075 E Tfv, Tfw, Tfy, Tfz;
|
cannam@167
|
2076 Tfv = FNMS(KP707106781, Tex, Tew);
|
cannam@167
|
2077 Tfw = Tet - Teq;
|
cannam@167
|
2078 Tfx = FNMS(KP923879532, Tfw, Tfv);
|
cannam@167
|
2079 TfU = FMA(KP923879532, Tfw, Tfv);
|
cannam@167
|
2080 Tfy = FNMS(KP707106781, Tem, Tel);
|
cannam@167
|
2081 Tfz = Tez - TeA;
|
cannam@167
|
2082 TfA = FNMS(KP923879532, Tfz, Tfy);
|
cannam@167
|
2083 TfT = FMA(KP923879532, Tfz, Tfy);
|
cannam@167
|
2084 }
|
cannam@167
|
2085 TfB = FMA(KP534511135, TfA, Tfx);
|
cannam@167
|
2086 Tg1 = FNMS(KP303346683, TfT, TfU);
|
cannam@167
|
2087 TfL = FNMS(KP534511135, Tfx, TfA);
|
cannam@167
|
2088 TfV = FMA(KP303346683, TfU, TfT);
|
cannam@167
|
2089 }
|
cannam@167
|
2090 {
|
cannam@167
|
2091 E Tfp, TfP, Tfs, TfQ;
|
cannam@167
|
2092 {
|
cannam@167
|
2093 E Tfn, Tfo, Tfq, Tfr;
|
cannam@167
|
2094 Tfn = FNMS(KP707106781, Te6, Te5);
|
cannam@167
|
2095 Tfo = FNMS(KP707106781, Te9, Te8);
|
cannam@167
|
2096 Tfp = FMA(KP668178637, Tfo, Tfn);
|
cannam@167
|
2097 TfP = FNMS(KP668178637, Tfn, Tfo);
|
cannam@167
|
2098 Tfq = FNMS(KP707106781, Ted, Tec);
|
cannam@167
|
2099 Tfr = FNMS(KP707106781, Teg, Tef);
|
cannam@167
|
2100 Tfs = FNMS(KP668178637, Tfr, Tfq);
|
cannam@167
|
2101 TfQ = FMA(KP668178637, Tfq, Tfr);
|
cannam@167
|
2102 }
|
cannam@167
|
2103 Tft = Tfp - Tfs;
|
cannam@167
|
2104 Tly = Tfp + Tfs;
|
cannam@167
|
2105 TfR = TfP + TfQ;
|
cannam@167
|
2106 Tls = TfQ - TfP;
|
cannam@167
|
2107 }
|
cannam@167
|
2108 {
|
cannam@167
|
2109 E TeO, Tfb, TeV, Tfa;
|
cannam@167
|
2110 {
|
cannam@167
|
2111 E TeG, TeN, TeR, TeU;
|
cannam@167
|
2112 TeG = FMA(KP707106781, TeF, TeE);
|
cannam@167
|
2113 TeN = TeJ + TeM;
|
cannam@167
|
2114 TeO = FNMS(KP923879532, TeN, TeG);
|
cannam@167
|
2115 Tfb = FMA(KP923879532, TeN, TeG);
|
cannam@167
|
2116 TeR = FMA(KP707106781, TeQ, TeP);
|
cannam@167
|
2117 TeU = TeS + TeT;
|
cannam@167
|
2118 TeV = FNMS(KP923879532, TeU, TeR);
|
cannam@167
|
2119 Tfa = FMA(KP923879532, TeU, TeR);
|
cannam@167
|
2120 }
|
cannam@167
|
2121 TeW = FNMS(KP820678790, TeV, TeO);
|
cannam@167
|
2122 Tfg = FMA(KP098491403, Tfa, Tfb);
|
cannam@167
|
2123 Tf0 = FMA(KP820678790, TeO, TeV);
|
cannam@167
|
2124 Tfc = FNMS(KP098491403, Tfb, Tfa);
|
cannam@167
|
2125 }
|
cannam@167
|
2126 {
|
cannam@167
|
2127 E TdW, Tfi, Tlb, Tlp, Te3, Tlq, Tfl, Tlc, TdV, Tla;
|
cannam@167
|
2128 TdV = TdT + TdU;
|
cannam@167
|
2129 TdW = FMA(KP707106781, TdV, TdS);
|
cannam@167
|
2130 Tfi = FNMS(KP707106781, TdV, TdS);
|
cannam@167
|
2131 Tla = T8D + T8K;
|
cannam@167
|
2132 Tlb = FMA(KP707106781, Tla, Tl9);
|
cannam@167
|
2133 Tlp = FNMS(KP707106781, Tla, Tl9);
|
cannam@167
|
2134 {
|
cannam@167
|
2135 E TdZ, Te2, Tfj, Tfk;
|
cannam@167
|
2136 TdZ = FMA(KP414213562, TdY, TdX);
|
cannam@167
|
2137 Te2 = FNMS(KP414213562, Te1, Te0);
|
cannam@167
|
2138 Te3 = TdZ + Te2;
|
cannam@167
|
2139 Tlq = Te2 - TdZ;
|
cannam@167
|
2140 Tfj = FNMS(KP414213562, TdX, TdY);
|
cannam@167
|
2141 Tfk = FMA(KP414213562, Te0, Te1);
|
cannam@167
|
2142 Tfl = Tfj - Tfk;
|
cannam@167
|
2143 Tlc = Tfj + Tfk;
|
cannam@167
|
2144 }
|
cannam@167
|
2145 Te4 = FNMS(KP923879532, Te3, TdW);
|
cannam@167
|
2146 Tlr = FMA(KP923879532, Tlq, Tlp);
|
cannam@167
|
2147 Tlx = FNMS(KP923879532, Tlq, Tlp);
|
cannam@167
|
2148 Tf2 = FMA(KP923879532, Te3, TdW);
|
cannam@167
|
2149 Tfm = FMA(KP923879532, Tfl, Tfi);
|
cannam@167
|
2150 Tld = FMA(KP923879532, Tlc, Tlb);
|
cannam@167
|
2151 Tlj = FNMS(KP923879532, Tlc, Tlb);
|
cannam@167
|
2152 TfO = FNMS(KP923879532, Tfl, Tfi);
|
cannam@167
|
2153 }
|
cannam@167
|
2154 {
|
cannam@167
|
2155 E Tek, TeX, Tll, Tlm;
|
cannam@167
|
2156 Tek = FMA(KP980785280, Tej, Te4);
|
cannam@167
|
2157 TeX = TeD - TeW;
|
cannam@167
|
2158 ri[WS(rs, 41)] = FNMS(KP773010453, TeX, Tek);
|
cannam@167
|
2159 ri[WS(rs, 9)] = FMA(KP773010453, TeX, Tek);
|
cannam@167
|
2160 Tll = FMA(KP980785280, Tlk, Tlj);
|
cannam@167
|
2161 Tlm = Tf0 - TeZ;
|
cannam@167
|
2162 ii[WS(rs, 9)] = FMA(KP773010453, Tlm, Tll);
|
cannam@167
|
2163 ii[WS(rs, 41)] = FNMS(KP773010453, Tlm, Tll);
|
cannam@167
|
2164 }
|
cannam@167
|
2165 {
|
cannam@167
|
2166 E TeY, Tf1, Tln, Tlo;
|
cannam@167
|
2167 TeY = FNMS(KP980785280, Tej, Te4);
|
cannam@167
|
2168 Tf1 = TeZ + Tf0;
|
cannam@167
|
2169 ri[WS(rs, 25)] = FNMS(KP773010453, Tf1, TeY);
|
cannam@167
|
2170 ri[WS(rs, 57)] = FMA(KP773010453, Tf1, TeY);
|
cannam@167
|
2171 Tln = FNMS(KP980785280, Tlk, Tlj);
|
cannam@167
|
2172 Tlo = TeD + TeW;
|
cannam@167
|
2173 ii[WS(rs, 25)] = FNMS(KP773010453, Tlo, Tln);
|
cannam@167
|
2174 ii[WS(rs, 57)] = FMA(KP773010453, Tlo, Tln);
|
cannam@167
|
2175 }
|
cannam@167
|
2176 {
|
cannam@167
|
2177 E Tf6, Tfd, Tlf, Tlg;
|
cannam@167
|
2178 Tf6 = FMA(KP980785280, Tf5, Tf2);
|
cannam@167
|
2179 Tfd = Tf9 + Tfc;
|
cannam@167
|
2180 ri[WS(rs, 33)] = FNMS(KP995184726, Tfd, Tf6);
|
cannam@167
|
2181 ri[WS(rs, 1)] = FMA(KP995184726, Tfd, Tf6);
|
cannam@167
|
2182 Tlf = FMA(KP980785280, Tle, Tld);
|
cannam@167
|
2183 Tlg = Tff + Tfg;
|
cannam@167
|
2184 ii[WS(rs, 1)] = FMA(KP995184726, Tlg, Tlf);
|
cannam@167
|
2185 ii[WS(rs, 33)] = FNMS(KP995184726, Tlg, Tlf);
|
cannam@167
|
2186 }
|
cannam@167
|
2187 {
|
cannam@167
|
2188 E Tfe, Tfh, Tlh, Tli;
|
cannam@167
|
2189 Tfe = FNMS(KP980785280, Tf5, Tf2);
|
cannam@167
|
2190 Tfh = Tff - Tfg;
|
cannam@167
|
2191 ri[WS(rs, 49)] = FNMS(KP995184726, Tfh, Tfe);
|
cannam@167
|
2192 ri[WS(rs, 17)] = FMA(KP995184726, Tfh, Tfe);
|
cannam@167
|
2193 Tlh = FNMS(KP980785280, Tle, Tld);
|
cannam@167
|
2194 Tli = Tfc - Tf9;
|
cannam@167
|
2195 ii[WS(rs, 17)] = FMA(KP995184726, Tli, Tlh);
|
cannam@167
|
2196 ii[WS(rs, 49)] = FNMS(KP995184726, Tli, Tlh);
|
cannam@167
|
2197 }
|
cannam@167
|
2198 {
|
cannam@167
|
2199 E Tfu, TfJ, Tlt, Tlu;
|
cannam@167
|
2200 Tfu = FMA(KP831469612, Tft, Tfm);
|
cannam@167
|
2201 TfJ = TfB + TfI;
|
cannam@167
|
2202 ri[WS(rs, 37)] = FNMS(KP881921264, TfJ, Tfu);
|
cannam@167
|
2203 ri[WS(rs, 5)] = FMA(KP881921264, TfJ, Tfu);
|
cannam@167
|
2204 Tlt = FMA(KP831469612, Tls, Tlr);
|
cannam@167
|
2205 Tlu = TfL + TfM;
|
cannam@167
|
2206 ii[WS(rs, 5)] = FMA(KP881921264, Tlu, Tlt);
|
cannam@167
|
2207 ii[WS(rs, 37)] = FNMS(KP881921264, Tlu, Tlt);
|
cannam@167
|
2208 }
|
cannam@167
|
2209 {
|
cannam@167
|
2210 E TfK, TfN, Tlv, Tlw;
|
cannam@167
|
2211 TfK = FNMS(KP831469612, Tft, Tfm);
|
cannam@167
|
2212 TfN = TfL - TfM;
|
cannam@167
|
2213 ri[WS(rs, 53)] = FNMS(KP881921264, TfN, TfK);
|
cannam@167
|
2214 ri[WS(rs, 21)] = FMA(KP881921264, TfN, TfK);
|
cannam@167
|
2215 Tlv = FNMS(KP831469612, Tls, Tlr);
|
cannam@167
|
2216 Tlw = TfI - TfB;
|
cannam@167
|
2217 ii[WS(rs, 21)] = FMA(KP881921264, Tlw, Tlv);
|
cannam@167
|
2218 ii[WS(rs, 53)] = FNMS(KP881921264, Tlw, Tlv);
|
cannam@167
|
2219 }
|
cannam@167
|
2220 {
|
cannam@167
|
2221 E TfS, TfZ, Tlz, TlA;
|
cannam@167
|
2222 TfS = FNMS(KP831469612, TfR, TfO);
|
cannam@167
|
2223 TfZ = TfV - TfY;
|
cannam@167
|
2224 ri[WS(rs, 45)] = FNMS(KP956940335, TfZ, TfS);
|
cannam@167
|
2225 ri[WS(rs, 13)] = FMA(KP956940335, TfZ, TfS);
|
cannam@167
|
2226 Tlz = FNMS(KP831469612, Tly, Tlx);
|
cannam@167
|
2227 TlA = Tg2 - Tg1;
|
cannam@167
|
2228 ii[WS(rs, 13)] = FMA(KP956940335, TlA, Tlz);
|
cannam@167
|
2229 ii[WS(rs, 45)] = FNMS(KP956940335, TlA, Tlz);
|
cannam@167
|
2230 }
|
cannam@167
|
2231 {
|
cannam@167
|
2232 E Tg0, Tg3, TlB, TlC;
|
cannam@167
|
2233 Tg0 = FMA(KP831469612, TfR, TfO);
|
cannam@167
|
2234 Tg3 = Tg1 + Tg2;
|
cannam@167
|
2235 ri[WS(rs, 29)] = FNMS(KP956940335, Tg3, Tg0);
|
cannam@167
|
2236 ri[WS(rs, 61)] = FMA(KP956940335, Tg3, Tg0);
|
cannam@167
|
2237 TlB = FMA(KP831469612, Tly, Tlx);
|
cannam@167
|
2238 TlC = TfV + TfY;
|
cannam@167
|
2239 ii[WS(rs, 29)] = FNMS(KP956940335, TlC, TlB);
|
cannam@167
|
2240 ii[WS(rs, 61)] = FMA(KP956940335, TlC, TlB);
|
cannam@167
|
2241 }
|
cannam@167
|
2242 }
|
cannam@167
|
2243 }
|
cannam@167
|
2244 }
|
cannam@167
|
2245 }
|
cannam@167
|
2246 }
|
cannam@167
|
2247
|
cannam@167
|
2248 static const tw_instr twinstr[] = {
|
cannam@167
|
2249 {TW_CEXP, 0, 1},
|
cannam@167
|
2250 {TW_CEXP, 0, 3},
|
cannam@167
|
2251 {TW_CEXP, 0, 9},
|
cannam@167
|
2252 {TW_CEXP, 0, 27},
|
cannam@167
|
2253 {TW_CEXP, 0, 63},
|
cannam@167
|
2254 {TW_NEXT, 1, 0}
|
cannam@167
|
2255 };
|
cannam@167
|
2256
|
cannam@167
|
2257 static const ct_desc desc = { 64, "t2_64", twinstr, &GENUS, {520, 206, 634, 0}, 0, 0, 0 };
|
cannam@167
|
2258
|
cannam@167
|
2259 void X(codelet_t2_64) (planner *p) {
|
cannam@167
|
2260 X(kdft_dit_register) (p, t2_64, &desc);
|
cannam@167
|
2261 }
|
cannam@167
|
2262 #else
|
cannam@167
|
2263
|
cannam@167
|
2264 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 64 -name t2_64 -include dft/scalar/t.h */
|
cannam@167
|
2265
|
cannam@167
|
2266 /*
|
cannam@167
|
2267 * This function contains 1154 FP additions, 660 FP multiplications,
|
cannam@167
|
2268 * (or, 880 additions, 386 multiplications, 274 fused multiply/add),
|
cannam@167
|
2269 * 302 stack variables, 15 constants, and 256 memory accesses
|
cannam@167
|
2270 */
|
cannam@167
|
2271 #include "dft/scalar/t.h"
|
cannam@167
|
2272
|
cannam@167
|
2273 static void t2_64(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
2274 {
|
cannam@167
|
2275 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
cannam@167
|
2276 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
2277 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
cannam@167
|
2278 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
2279 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
cannam@167
|
2280 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
2281 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
cannam@167
|
2282 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
2283 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
2284 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
2285 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
2286 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
2287 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
2288 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
2289 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
2290 {
|
cannam@167
|
2291 INT m;
|
cannam@167
|
2292 for (m = mb, W = W + (mb * 10); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 10, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
2293 E T2, T5, T3, T6, Te, T9, TP, T3e, T1e, T39, T3c, TT, T1a, T37, T8;
|
cannam@167
|
2294 E Tw, Td, Ty, Tm, Th, T1C, T3K, T1V, T3x, T3I, T1G, T1R, T3v, T2m, T2q;
|
cannam@167
|
2295 E T5Y, T6u, T53, T5B, T62, T6w, T57, T5D, T2V, T2X, Tg, TE, T3Y, T3V, T3j;
|
cannam@167
|
2296 E Tl, TA, T3g, T1j, T1t, TV, T2C, T2z, T1u, TZ, T1h, To, T1p, T6j, T6H;
|
cannam@167
|
2297 E Ts, T1l, T6l, T6F, T2P, T4b, T4x, T5i, T2R, T49, T4z, T5g, TG, T4k, T4m;
|
cannam@167
|
2298 E TK, T21, T3O, T3Q, T25, TW, T10, T11, T79, T6X, T5M, T6b, T1v, T30, T69;
|
cannam@167
|
2299 E T77, T13, T2F, T2D, T6p, T6O, T1x, T2a, T2f, T6V, T28, T6r, T2h, T6Q, T32;
|
cannam@167
|
2300 E T5K, T5w, T4G, T4Q, T3m, T4h, T4I, T5y, T3k, T4f, T41, T4S, T4Y, T3q, T3D;
|
cannam@167
|
2301 E T3F, T5r, T3s, T4W, T3Z, T5p;
|
cannam@167
|
2302 {
|
cannam@167
|
2303 E Ta, Tj, Tx, TC, Tf, Tk, Tz, TD, T1B, T1E, T2o, T2l, T1T, T1Q, T1A;
|
cannam@167
|
2304 E T1F, T2p, T2k, T1U, T1P;
|
cannam@167
|
2305 {
|
cannam@167
|
2306 E T4, T1d, T19, Tb, T1c, T7, Tc, T18, TR, TO, TS, TN;
|
cannam@167
|
2307 T2 = W[0];
|
cannam@167
|
2308 T5 = W[1];
|
cannam@167
|
2309 T3 = W[2];
|
cannam@167
|
2310 T6 = W[3];
|
cannam@167
|
2311 Te = W[5];
|
cannam@167
|
2312 T9 = W[4];
|
cannam@167
|
2313 T4 = T2 * T3;
|
cannam@167
|
2314 T1d = T5 * T9;
|
cannam@167
|
2315 T19 = T5 * Te;
|
cannam@167
|
2316 Tb = T2 * T6;
|
cannam@167
|
2317 T1c = T2 * Te;
|
cannam@167
|
2318 T7 = T5 * T6;
|
cannam@167
|
2319 Tc = T5 * T3;
|
cannam@167
|
2320 T18 = T2 * T9;
|
cannam@167
|
2321 TR = T3 * Te;
|
cannam@167
|
2322 TO = T6 * Te;
|
cannam@167
|
2323 TS = T6 * T9;
|
cannam@167
|
2324 TN = T3 * T9;
|
cannam@167
|
2325 TP = TN - TO;
|
cannam@167
|
2326 T3e = TR - TS;
|
cannam@167
|
2327 T1e = T1c - T1d;
|
cannam@167
|
2328 T39 = T1c + T1d;
|
cannam@167
|
2329 T3c = TN + TO;
|
cannam@167
|
2330 TT = TR + TS;
|
cannam@167
|
2331 T1a = T18 + T19;
|
cannam@167
|
2332 T37 = T18 - T19;
|
cannam@167
|
2333 T8 = T4 - T7;
|
cannam@167
|
2334 Ta = T8 * T9;
|
cannam@167
|
2335 Tj = T8 * Te;
|
cannam@167
|
2336 Tw = T4 + T7;
|
cannam@167
|
2337 Tx = Tw * T9;
|
cannam@167
|
2338 TC = Tw * Te;
|
cannam@167
|
2339 Td = Tb + Tc;
|
cannam@167
|
2340 Tf = Td * Te;
|
cannam@167
|
2341 Tk = Td * T9;
|
cannam@167
|
2342 Ty = Tb - Tc;
|
cannam@167
|
2343 Tz = Ty * Te;
|
cannam@167
|
2344 TD = Ty * T9;
|
cannam@167
|
2345 Tm = W[7];
|
cannam@167
|
2346 T1B = T6 * Tm;
|
cannam@167
|
2347 T1E = T3 * Tm;
|
cannam@167
|
2348 T2o = T2 * Tm;
|
cannam@167
|
2349 T2l = T5 * Tm;
|
cannam@167
|
2350 T1T = T9 * Tm;
|
cannam@167
|
2351 T1Q = Te * Tm;
|
cannam@167
|
2352 Th = W[6];
|
cannam@167
|
2353 T1A = T3 * Th;
|
cannam@167
|
2354 T1F = T6 * Th;
|
cannam@167
|
2355 T2p = T5 * Th;
|
cannam@167
|
2356 T2k = T2 * Th;
|
cannam@167
|
2357 T1U = Te * Th;
|
cannam@167
|
2358 T1P = T9 * Th;
|
cannam@167
|
2359 }
|
cannam@167
|
2360 T1C = T1A + T1B;
|
cannam@167
|
2361 T3K = T1E + T1F;
|
cannam@167
|
2362 T1V = T1T + T1U;
|
cannam@167
|
2363 T3x = T2o - T2p;
|
cannam@167
|
2364 T3I = T1A - T1B;
|
cannam@167
|
2365 T1G = T1E - T1F;
|
cannam@167
|
2366 T1R = T1P - T1Q;
|
cannam@167
|
2367 {
|
cannam@167
|
2368 E T5W, T5X, T55, T56;
|
cannam@167
|
2369 T3v = T2k + T2l;
|
cannam@167
|
2370 T2m = T2k - T2l;
|
cannam@167
|
2371 T2q = T2o + T2p;
|
cannam@167
|
2372 T5W = T8 * Th;
|
cannam@167
|
2373 T5X = Td * Tm;
|
cannam@167
|
2374 T5Y = T5W - T5X;
|
cannam@167
|
2375 T6u = T5W + T5X;
|
cannam@167
|
2376 {
|
cannam@167
|
2377 E T51, T52, T60, T61;
|
cannam@167
|
2378 T51 = Tw * Th;
|
cannam@167
|
2379 T52 = Ty * Tm;
|
cannam@167
|
2380 T53 = T51 + T52;
|
cannam@167
|
2381 T5B = T51 - T52;
|
cannam@167
|
2382 T60 = T8 * Tm;
|
cannam@167
|
2383 T61 = Td * Th;
|
cannam@167
|
2384 T62 = T60 + T61;
|
cannam@167
|
2385 T6w = T60 - T61;
|
cannam@167
|
2386 }
|
cannam@167
|
2387 T55 = Tw * Tm;
|
cannam@167
|
2388 T56 = Ty * Th;
|
cannam@167
|
2389 T57 = T55 - T56;
|
cannam@167
|
2390 T5D = T55 + T56;
|
cannam@167
|
2391 {
|
cannam@167
|
2392 E Ti, Tq, TF, TJ, T3W, T3X, T3T, T3U, T3h, T3i, Tn, Tr, TB, TI, T3d;
|
cannam@167
|
2393 E T3f, T1k, T1o, T1Z, T23, TQ, TU, T2A, T2B, T2x, T2y, T20, T24, TX, TY;
|
cannam@167
|
2394 E T1i, T1n;
|
cannam@167
|
2395 T2V = T1P + T1Q;
|
cannam@167
|
2396 T2X = T1T - T1U;
|
cannam@167
|
2397 Tg = Ta + Tf;
|
cannam@167
|
2398 Ti = Tg * Th;
|
cannam@167
|
2399 Tq = Tg * Tm;
|
cannam@167
|
2400 TE = TC + TD;
|
cannam@167
|
2401 TF = TE * Tm;
|
cannam@167
|
2402 TJ = TE * Th;
|
cannam@167
|
2403 T3W = T37 * Tm;
|
cannam@167
|
2404 T3X = T39 * Th;
|
cannam@167
|
2405 T3Y = T3W - T3X;
|
cannam@167
|
2406 T3T = T37 * Th;
|
cannam@167
|
2407 T3U = T39 * Tm;
|
cannam@167
|
2408 T3V = T3T + T3U;
|
cannam@167
|
2409 T3h = T3c * Tm;
|
cannam@167
|
2410 T3i = T3e * Th;
|
cannam@167
|
2411 T3j = T3h - T3i;
|
cannam@167
|
2412 Tl = Tj - Tk;
|
cannam@167
|
2413 Tn = Tl * Tm;
|
cannam@167
|
2414 Tr = Tl * Th;
|
cannam@167
|
2415 TA = Tx - Tz;
|
cannam@167
|
2416 TB = TA * Th;
|
cannam@167
|
2417 TI = TA * Tm;
|
cannam@167
|
2418 T3d = T3c * Th;
|
cannam@167
|
2419 T3f = T3e * Tm;
|
cannam@167
|
2420 T3g = T3d + T3f;
|
cannam@167
|
2421 T1j = Tj + Tk;
|
cannam@167
|
2422 T1k = T1j * Tm;
|
cannam@167
|
2423 T1o = T1j * Th;
|
cannam@167
|
2424 T1t = Tx + Tz;
|
cannam@167
|
2425 T1Z = T1t * Th;
|
cannam@167
|
2426 T23 = T1t * Tm;
|
cannam@167
|
2427 TQ = TP * Th;
|
cannam@167
|
2428 TU = TT * Tm;
|
cannam@167
|
2429 TV = TQ + TU;
|
cannam@167
|
2430 T2A = T1a * Tm;
|
cannam@167
|
2431 T2B = T1e * Th;
|
cannam@167
|
2432 T2C = T2A - T2B;
|
cannam@167
|
2433 T2x = T1a * Th;
|
cannam@167
|
2434 T2y = T1e * Tm;
|
cannam@167
|
2435 T2z = T2x + T2y;
|
cannam@167
|
2436 T1u = TC - TD;
|
cannam@167
|
2437 T20 = T1u * Tm;
|
cannam@167
|
2438 T24 = T1u * Th;
|
cannam@167
|
2439 TX = TP * Tm;
|
cannam@167
|
2440 TY = TT * Th;
|
cannam@167
|
2441 TZ = TX - TY;
|
cannam@167
|
2442 T1h = Ta - Tf;
|
cannam@167
|
2443 T1i = T1h * Th;
|
cannam@167
|
2444 T1n = T1h * Tm;
|
cannam@167
|
2445 To = Ti - Tn;
|
cannam@167
|
2446 T1p = T1n + T1o;
|
cannam@167
|
2447 T6j = TQ - TU;
|
cannam@167
|
2448 T6H = T2A + T2B;
|
cannam@167
|
2449 Ts = Tq + Tr;
|
cannam@167
|
2450 T1l = T1i - T1k;
|
cannam@167
|
2451 T6l = TX + TY;
|
cannam@167
|
2452 T6F = T2x - T2y;
|
cannam@167
|
2453 T2P = T1Z - T20;
|
cannam@167
|
2454 T4b = TI + TJ;
|
cannam@167
|
2455 T4x = T3d - T3f;
|
cannam@167
|
2456 T5i = T3W + T3X;
|
cannam@167
|
2457 T2R = T23 + T24;
|
cannam@167
|
2458 T49 = TB - TF;
|
cannam@167
|
2459 T4z = T3h + T3i;
|
cannam@167
|
2460 T5g = T3T - T3U;
|
cannam@167
|
2461 TG = TB + TF;
|
cannam@167
|
2462 T4k = Ti + Tn;
|
cannam@167
|
2463 T4m = Tq - Tr;
|
cannam@167
|
2464 TK = TI - TJ;
|
cannam@167
|
2465 T21 = T1Z + T20;
|
cannam@167
|
2466 T3O = T1i + T1k;
|
cannam@167
|
2467 T3Q = T1n - T1o;
|
cannam@167
|
2468 T25 = T23 - T24;
|
cannam@167
|
2469 TW = W[8];
|
cannam@167
|
2470 T10 = W[9];
|
cannam@167
|
2471 T11 = FMA(TV, TW, TZ * T10);
|
cannam@167
|
2472 T79 = FNMS(T25, TW, T21 * T10);
|
cannam@167
|
2473 T6X = FNMS(Td, TW, T8 * T10);
|
cannam@167
|
2474 T5M = FNMS(T2X, TW, T2V * T10);
|
cannam@167
|
2475 T6b = FNMS(TK, TW, TG * T10);
|
cannam@167
|
2476 T1v = FMA(T1t, TW, T1u * T10);
|
cannam@167
|
2477 T30 = FMA(T1h, TW, T1j * T10);
|
cannam@167
|
2478 T69 = FMA(TG, TW, TK * T10);
|
cannam@167
|
2479 T77 = FMA(T21, TW, T25 * T10);
|
cannam@167
|
2480 T13 = FNMS(TZ, TW, TV * T10);
|
cannam@167
|
2481 T2F = FNMS(T2C, TW, T2z * T10);
|
cannam@167
|
2482 T2D = FMA(T2z, TW, T2C * T10);
|
cannam@167
|
2483 T6p = FMA(T1a, TW, T1e * T10);
|
cannam@167
|
2484 T6O = FMA(TP, TW, TT * T10);
|
cannam@167
|
2485 T1x = FNMS(T1u, TW, T1t * T10);
|
cannam@167
|
2486 T2a = FNMS(TE, TW, TA * T10);
|
cannam@167
|
2487 T2f = FMA(T3, TW, T6 * T10);
|
cannam@167
|
2488 T6V = FMA(T8, TW, Td * T10);
|
cannam@167
|
2489 T28 = FMA(TA, TW, TE * T10);
|
cannam@167
|
2490 T6r = FNMS(T1e, TW, T1a * T10);
|
cannam@167
|
2491 T2h = FNMS(T6, TW, T3 * T10);
|
cannam@167
|
2492 T6Q = FNMS(TT, TW, TP * T10);
|
cannam@167
|
2493 T32 = FNMS(T1j, TW, T1h * T10);
|
cannam@167
|
2494 T5K = FMA(T2V, TW, T2X * T10);
|
cannam@167
|
2495 T5w = FMA(Tw, TW, Ty * T10);
|
cannam@167
|
2496 T4G = FMA(T3O, TW, T3Q * T10);
|
cannam@167
|
2497 T4Q = FMA(T4k, TW, T4m * T10);
|
cannam@167
|
2498 T3m = FNMS(T3j, TW, T3g * T10);
|
cannam@167
|
2499 T4h = FNMS(Te, TW, T9 * T10);
|
cannam@167
|
2500 T4I = FNMS(T3Q, TW, T3O * T10);
|
cannam@167
|
2501 T5y = FNMS(Ty, TW, Tw * T10);
|
cannam@167
|
2502 T3k = FMA(T3g, TW, T3j * T10);
|
cannam@167
|
2503 T4f = FMA(T9, TW, Te * T10);
|
cannam@167
|
2504 T41 = FNMS(T3Y, TW, T3V * T10);
|
cannam@167
|
2505 T4S = FNMS(T4m, TW, T4k * T10);
|
cannam@167
|
2506 T4Y = FNMS(T3e, TW, T3c * T10);
|
cannam@167
|
2507 T3q = FMA(Tg, TW, Tl * T10);
|
cannam@167
|
2508 T3D = FMA(T2, TW, T5 * T10);
|
cannam@167
|
2509 T3F = FNMS(T5, TW, T2 * T10);
|
cannam@167
|
2510 T5r = FNMS(T39, TW, T37 * T10);
|
cannam@167
|
2511 T3s = FNMS(Tl, TW, Tg * T10);
|
cannam@167
|
2512 T4W = FMA(T3c, TW, T3e * T10);
|
cannam@167
|
2513 T3Z = FMA(T3V, TW, T3Y * T10);
|
cannam@167
|
2514 T5p = FMA(T37, TW, T39 * T10);
|
cannam@167
|
2515 }
|
cannam@167
|
2516 }
|
cannam@167
|
2517 }
|
cannam@167
|
2518 {
|
cannam@167
|
2519 E T17, TdV, Tj3, Tjx, T7l, TbJ, Ti3, Tix, T1K, Tiw, TdY, ThY, T7w, Tj0, TbM;
|
cannam@167
|
2520 E Tjw, T2e, TgA, T7I, TaY, TbQ, Tda, Te4, TfO, T2J, TgB, T7T, TaZ, TbT, Tdb;
|
cannam@167
|
2521 E Te9, TfP, T36, T3B, TgH, TgE, TgF, TgG, T80, TbW, Tel, TfT, T8b, Tc0, T8k;
|
cannam@167
|
2522 E TbX, Teg, TfS, T8h, TbZ, T45, T4q, TgJ, TgK, TgL, TgM, T8r, Tc6, Tew, TfW;
|
cannam@167
|
2523 E T8C, Tc4, T8L, Tc7, Ter, TfV, T8I, Tc3, T6B, Th1, Tfm, Tga, Th8, ThI, T9N;
|
cannam@167
|
2524 E Tcv, T9Y, TcH, Tav, Tcw, Tf5, Tg7, Tas, TcG, T5c, TgV, TeV, Tg0, TgS, ThD;
|
cannam@167
|
2525 E T8U, Tcc, T95, Tco, T9C, Tcd, TeE, Tg3, T9z, Tcn, T5R, TgT, TeO, TeW, TgY;
|
cannam@167
|
2526 E ThE, T9h, T9F, T9s, T9E, Tck, Tcq, TeJ, TeX, Tch, Tcr, T7e, Th9, Tff, Tfn;
|
cannam@167
|
2527 E Th4, ThJ, Taa, Tay, Tal, Tax, TcD, TcJ, Tfa, Tfo, TcA, TcK;
|
cannam@167
|
2528 {
|
cannam@167
|
2529 E T1, Ti1, Tu, Ti0, TM, T7i, T15, T7j, Tp, Tt;
|
cannam@167
|
2530 T1 = ri[0];
|
cannam@167
|
2531 Ti1 = ii[0];
|
cannam@167
|
2532 Tp = ri[WS(rs, 32)];
|
cannam@167
|
2533 Tt = ii[WS(rs, 32)];
|
cannam@167
|
2534 Tu = FMA(To, Tp, Ts * Tt);
|
cannam@167
|
2535 Ti0 = FNMS(Ts, Tp, To * Tt);
|
cannam@167
|
2536 {
|
cannam@167
|
2537 E TH, TL, T12, T14;
|
cannam@167
|
2538 TH = ri[WS(rs, 16)];
|
cannam@167
|
2539 TL = ii[WS(rs, 16)];
|
cannam@167
|
2540 TM = FMA(TG, TH, TK * TL);
|
cannam@167
|
2541 T7i = FNMS(TK, TH, TG * TL);
|
cannam@167
|
2542 T12 = ri[WS(rs, 48)];
|
cannam@167
|
2543 T14 = ii[WS(rs, 48)];
|
cannam@167
|
2544 T15 = FMA(T11, T12, T13 * T14);
|
cannam@167
|
2545 T7j = FNMS(T13, T12, T11 * T14);
|
cannam@167
|
2546 }
|
cannam@167
|
2547 {
|
cannam@167
|
2548 E Tv, T16, Tj1, Tj2;
|
cannam@167
|
2549 Tv = T1 + Tu;
|
cannam@167
|
2550 T16 = TM + T15;
|
cannam@167
|
2551 T17 = Tv + T16;
|
cannam@167
|
2552 TdV = Tv - T16;
|
cannam@167
|
2553 Tj1 = Ti1 - Ti0;
|
cannam@167
|
2554 Tj2 = TM - T15;
|
cannam@167
|
2555 Tj3 = Tj1 - Tj2;
|
cannam@167
|
2556 Tjx = Tj2 + Tj1;
|
cannam@167
|
2557 }
|
cannam@167
|
2558 {
|
cannam@167
|
2559 E T7h, T7k, ThZ, Ti2;
|
cannam@167
|
2560 T7h = T1 - Tu;
|
cannam@167
|
2561 T7k = T7i - T7j;
|
cannam@167
|
2562 T7l = T7h - T7k;
|
cannam@167
|
2563 TbJ = T7h + T7k;
|
cannam@167
|
2564 ThZ = T7i + T7j;
|
cannam@167
|
2565 Ti2 = Ti0 + Ti1;
|
cannam@167
|
2566 Ti3 = ThZ + Ti2;
|
cannam@167
|
2567 Tix = Ti2 - ThZ;
|
cannam@167
|
2568 }
|
cannam@167
|
2569 }
|
cannam@167
|
2570 {
|
cannam@167
|
2571 E T1g, T7m, T1r, T7n, T7o, T7p, T1z, T7s, T1I, T7t, T7r, T7u;
|
cannam@167
|
2572 {
|
cannam@167
|
2573 E T1b, T1f, T1m, T1q;
|
cannam@167
|
2574 T1b = ri[WS(rs, 8)];
|
cannam@167
|
2575 T1f = ii[WS(rs, 8)];
|
cannam@167
|
2576 T1g = FMA(T1a, T1b, T1e * T1f);
|
cannam@167
|
2577 T7m = FNMS(T1e, T1b, T1a * T1f);
|
cannam@167
|
2578 T1m = ri[WS(rs, 40)];
|
cannam@167
|
2579 T1q = ii[WS(rs, 40)];
|
cannam@167
|
2580 T1r = FMA(T1l, T1m, T1p * T1q);
|
cannam@167
|
2581 T7n = FNMS(T1p, T1m, T1l * T1q);
|
cannam@167
|
2582 }
|
cannam@167
|
2583 T7o = T7m - T7n;
|
cannam@167
|
2584 T7p = T1g - T1r;
|
cannam@167
|
2585 {
|
cannam@167
|
2586 E T1w, T1y, T1D, T1H;
|
cannam@167
|
2587 T1w = ri[WS(rs, 56)];
|
cannam@167
|
2588 T1y = ii[WS(rs, 56)];
|
cannam@167
|
2589 T1z = FMA(T1v, T1w, T1x * T1y);
|
cannam@167
|
2590 T7s = FNMS(T1x, T1w, T1v * T1y);
|
cannam@167
|
2591 T1D = ri[WS(rs, 24)];
|
cannam@167
|
2592 T1H = ii[WS(rs, 24)];
|
cannam@167
|
2593 T1I = FMA(T1C, T1D, T1G * T1H);
|
cannam@167
|
2594 T7t = FNMS(T1G, T1D, T1C * T1H);
|
cannam@167
|
2595 }
|
cannam@167
|
2596 T7r = T1z - T1I;
|
cannam@167
|
2597 T7u = T7s - T7t;
|
cannam@167
|
2598 {
|
cannam@167
|
2599 E T1s, T1J, TdW, TdX;
|
cannam@167
|
2600 T1s = T1g + T1r;
|
cannam@167
|
2601 T1J = T1z + T1I;
|
cannam@167
|
2602 T1K = T1s + T1J;
|
cannam@167
|
2603 Tiw = T1J - T1s;
|
cannam@167
|
2604 TdW = T7m + T7n;
|
cannam@167
|
2605 TdX = T7s + T7t;
|
cannam@167
|
2606 TdY = TdW - TdX;
|
cannam@167
|
2607 ThY = TdW + TdX;
|
cannam@167
|
2608 }
|
cannam@167
|
2609 {
|
cannam@167
|
2610 E T7q, T7v, TbK, TbL;
|
cannam@167
|
2611 T7q = T7o - T7p;
|
cannam@167
|
2612 T7v = T7r + T7u;
|
cannam@167
|
2613 T7w = KP707106781 * (T7q - T7v);
|
cannam@167
|
2614 Tj0 = KP707106781 * (T7q + T7v);
|
cannam@167
|
2615 TbK = T7p + T7o;
|
cannam@167
|
2616 TbL = T7r - T7u;
|
cannam@167
|
2617 TbM = KP707106781 * (TbK + TbL);
|
cannam@167
|
2618 Tjw = KP707106781 * (TbL - TbK);
|
cannam@167
|
2619 }
|
cannam@167
|
2620 }
|
cannam@167
|
2621 {
|
cannam@167
|
2622 E T1Y, Te0, T7A, T7D, T2d, Te1, T7B, T7G, T7C, T7H;
|
cannam@167
|
2623 {
|
cannam@167
|
2624 E T1O, T7y, T1X, T7z;
|
cannam@167
|
2625 {
|
cannam@167
|
2626 E T1M, T1N, T1S, T1W;
|
cannam@167
|
2627 T1M = ri[WS(rs, 4)];
|
cannam@167
|
2628 T1N = ii[WS(rs, 4)];
|
cannam@167
|
2629 T1O = FMA(T8, T1M, Td * T1N);
|
cannam@167
|
2630 T7y = FNMS(Td, T1M, T8 * T1N);
|
cannam@167
|
2631 T1S = ri[WS(rs, 36)];
|
cannam@167
|
2632 T1W = ii[WS(rs, 36)];
|
cannam@167
|
2633 T1X = FMA(T1R, T1S, T1V * T1W);
|
cannam@167
|
2634 T7z = FNMS(T1V, T1S, T1R * T1W);
|
cannam@167
|
2635 }
|
cannam@167
|
2636 T1Y = T1O + T1X;
|
cannam@167
|
2637 Te0 = T7y + T7z;
|
cannam@167
|
2638 T7A = T7y - T7z;
|
cannam@167
|
2639 T7D = T1O - T1X;
|
cannam@167
|
2640 }
|
cannam@167
|
2641 {
|
cannam@167
|
2642 E T27, T7E, T2c, T7F;
|
cannam@167
|
2643 {
|
cannam@167
|
2644 E T22, T26, T29, T2b;
|
cannam@167
|
2645 T22 = ri[WS(rs, 20)];
|
cannam@167
|
2646 T26 = ii[WS(rs, 20)];
|
cannam@167
|
2647 T27 = FMA(T21, T22, T25 * T26);
|
cannam@167
|
2648 T7E = FNMS(T25, T22, T21 * T26);
|
cannam@167
|
2649 T29 = ri[WS(rs, 52)];
|
cannam@167
|
2650 T2b = ii[WS(rs, 52)];
|
cannam@167
|
2651 T2c = FMA(T28, T29, T2a * T2b);
|
cannam@167
|
2652 T7F = FNMS(T2a, T29, T28 * T2b);
|
cannam@167
|
2653 }
|
cannam@167
|
2654 T2d = T27 + T2c;
|
cannam@167
|
2655 Te1 = T7E + T7F;
|
cannam@167
|
2656 T7B = T27 - T2c;
|
cannam@167
|
2657 T7G = T7E - T7F;
|
cannam@167
|
2658 }
|
cannam@167
|
2659 T2e = T1Y + T2d;
|
cannam@167
|
2660 TgA = Te0 + Te1;
|
cannam@167
|
2661 T7C = T7A + T7B;
|
cannam@167
|
2662 T7H = T7D - T7G;
|
cannam@167
|
2663 T7I = FNMS(KP923879532, T7H, KP382683432 * T7C);
|
cannam@167
|
2664 TaY = FMA(KP923879532, T7C, KP382683432 * T7H);
|
cannam@167
|
2665 {
|
cannam@167
|
2666 E TbO, TbP, Te2, Te3;
|
cannam@167
|
2667 TbO = T7A - T7B;
|
cannam@167
|
2668 TbP = T7D + T7G;
|
cannam@167
|
2669 TbQ = FNMS(KP382683432, TbP, KP923879532 * TbO);
|
cannam@167
|
2670 Tda = FMA(KP382683432, TbO, KP923879532 * TbP);
|
cannam@167
|
2671 Te2 = Te0 - Te1;
|
cannam@167
|
2672 Te3 = T1Y - T2d;
|
cannam@167
|
2673 Te4 = Te2 - Te3;
|
cannam@167
|
2674 TfO = Te3 + Te2;
|
cannam@167
|
2675 }
|
cannam@167
|
2676 }
|
cannam@167
|
2677 {
|
cannam@167
|
2678 E T2t, Te6, T7L, T7O, T2I, Te7, T7M, T7R, T7N, T7S;
|
cannam@167
|
2679 {
|
cannam@167
|
2680 E T2j, T7J, T2s, T7K;
|
cannam@167
|
2681 {
|
cannam@167
|
2682 E T2g, T2i, T2n, T2r;
|
cannam@167
|
2683 T2g = ri[WS(rs, 60)];
|
cannam@167
|
2684 T2i = ii[WS(rs, 60)];
|
cannam@167
|
2685 T2j = FMA(T2f, T2g, T2h * T2i);
|
cannam@167
|
2686 T7J = FNMS(T2h, T2g, T2f * T2i);
|
cannam@167
|
2687 T2n = ri[WS(rs, 28)];
|
cannam@167
|
2688 T2r = ii[WS(rs, 28)];
|
cannam@167
|
2689 T2s = FMA(T2m, T2n, T2q * T2r);
|
cannam@167
|
2690 T7K = FNMS(T2q, T2n, T2m * T2r);
|
cannam@167
|
2691 }
|
cannam@167
|
2692 T2t = T2j + T2s;
|
cannam@167
|
2693 Te6 = T7J + T7K;
|
cannam@167
|
2694 T7L = T7J - T7K;
|
cannam@167
|
2695 T7O = T2j - T2s;
|
cannam@167
|
2696 }
|
cannam@167
|
2697 {
|
cannam@167
|
2698 E T2w, T7P, T2H, T7Q;
|
cannam@167
|
2699 {
|
cannam@167
|
2700 E T2u, T2v, T2E, T2G;
|
cannam@167
|
2701 T2u = ri[WS(rs, 12)];
|
cannam@167
|
2702 T2v = ii[WS(rs, 12)];
|
cannam@167
|
2703 T2w = FMA(TP, T2u, TT * T2v);
|
cannam@167
|
2704 T7P = FNMS(TT, T2u, TP * T2v);
|
cannam@167
|
2705 T2E = ri[WS(rs, 44)];
|
cannam@167
|
2706 T2G = ii[WS(rs, 44)];
|
cannam@167
|
2707 T2H = FMA(T2D, T2E, T2F * T2G);
|
cannam@167
|
2708 T7Q = FNMS(T2F, T2E, T2D * T2G);
|
cannam@167
|
2709 }
|
cannam@167
|
2710 T2I = T2w + T2H;
|
cannam@167
|
2711 Te7 = T7P + T7Q;
|
cannam@167
|
2712 T7M = T2w - T2H;
|
cannam@167
|
2713 T7R = T7P - T7Q;
|
cannam@167
|
2714 }
|
cannam@167
|
2715 T2J = T2t + T2I;
|
cannam@167
|
2716 TgB = Te6 + Te7;
|
cannam@167
|
2717 T7N = T7L + T7M;
|
cannam@167
|
2718 T7S = T7O - T7R;
|
cannam@167
|
2719 T7T = FMA(KP382683432, T7N, KP923879532 * T7S);
|
cannam@167
|
2720 TaZ = FNMS(KP923879532, T7N, KP382683432 * T7S);
|
cannam@167
|
2721 {
|
cannam@167
|
2722 E TbR, TbS, Te5, Te8;
|
cannam@167
|
2723 TbR = T7L - T7M;
|
cannam@167
|
2724 TbS = T7O + T7R;
|
cannam@167
|
2725 TbT = FMA(KP923879532, TbR, KP382683432 * TbS);
|
cannam@167
|
2726 Tdb = FNMS(KP382683432, TbR, KP923879532 * TbS);
|
cannam@167
|
2727 Te5 = T2t - T2I;
|
cannam@167
|
2728 Te8 = Te6 - Te7;
|
cannam@167
|
2729 Te9 = Te5 + Te8;
|
cannam@167
|
2730 TfP = Te5 - Te8;
|
cannam@167
|
2731 }
|
cannam@167
|
2732 }
|
cannam@167
|
2733 {
|
cannam@167
|
2734 E T2O, T7W, T2T, T7X, T2U, Tec, T2Z, T8e, T34, T8f, T35, Ted, T3p, Tei, T86;
|
cannam@167
|
2735 E T89, T3A, Tej, T81, T84;
|
cannam@167
|
2736 {
|
cannam@167
|
2737 E T2M, T2N, T2Q, T2S;
|
cannam@167
|
2738 T2M = ri[WS(rs, 2)];
|
cannam@167
|
2739 T2N = ii[WS(rs, 2)];
|
cannam@167
|
2740 T2O = FMA(Tw, T2M, Ty * T2N);
|
cannam@167
|
2741 T7W = FNMS(Ty, T2M, Tw * T2N);
|
cannam@167
|
2742 T2Q = ri[WS(rs, 34)];
|
cannam@167
|
2743 T2S = ii[WS(rs, 34)];
|
cannam@167
|
2744 T2T = FMA(T2P, T2Q, T2R * T2S);
|
cannam@167
|
2745 T7X = FNMS(T2R, T2Q, T2P * T2S);
|
cannam@167
|
2746 }
|
cannam@167
|
2747 T2U = T2O + T2T;
|
cannam@167
|
2748 Tec = T7W + T7X;
|
cannam@167
|
2749 {
|
cannam@167
|
2750 E T2W, T2Y, T31, T33;
|
cannam@167
|
2751 T2W = ri[WS(rs, 18)];
|
cannam@167
|
2752 T2Y = ii[WS(rs, 18)];
|
cannam@167
|
2753 T2Z = FMA(T2V, T2W, T2X * T2Y);
|
cannam@167
|
2754 T8e = FNMS(T2X, T2W, T2V * T2Y);
|
cannam@167
|
2755 T31 = ri[WS(rs, 50)];
|
cannam@167
|
2756 T33 = ii[WS(rs, 50)];
|
cannam@167
|
2757 T34 = FMA(T30, T31, T32 * T33);
|
cannam@167
|
2758 T8f = FNMS(T32, T31, T30 * T33);
|
cannam@167
|
2759 }
|
cannam@167
|
2760 T35 = T2Z + T34;
|
cannam@167
|
2761 Ted = T8e + T8f;
|
cannam@167
|
2762 {
|
cannam@167
|
2763 E T3b, T87, T3o, T88;
|
cannam@167
|
2764 {
|
cannam@167
|
2765 E T38, T3a, T3l, T3n;
|
cannam@167
|
2766 T38 = ri[WS(rs, 10)];
|
cannam@167
|
2767 T3a = ii[WS(rs, 10)];
|
cannam@167
|
2768 T3b = FMA(T37, T38, T39 * T3a);
|
cannam@167
|
2769 T87 = FNMS(T39, T38, T37 * T3a);
|
cannam@167
|
2770 T3l = ri[WS(rs, 42)];
|
cannam@167
|
2771 T3n = ii[WS(rs, 42)];
|
cannam@167
|
2772 T3o = FMA(T3k, T3l, T3m * T3n);
|
cannam@167
|
2773 T88 = FNMS(T3m, T3l, T3k * T3n);
|
cannam@167
|
2774 }
|
cannam@167
|
2775 T3p = T3b + T3o;
|
cannam@167
|
2776 Tei = T87 + T88;
|
cannam@167
|
2777 T86 = T3b - T3o;
|
cannam@167
|
2778 T89 = T87 - T88;
|
cannam@167
|
2779 }
|
cannam@167
|
2780 {
|
cannam@167
|
2781 E T3u, T82, T3z, T83;
|
cannam@167
|
2782 {
|
cannam@167
|
2783 E T3r, T3t, T3w, T3y;
|
cannam@167
|
2784 T3r = ri[WS(rs, 58)];
|
cannam@167
|
2785 T3t = ii[WS(rs, 58)];
|
cannam@167
|
2786 T3u = FMA(T3q, T3r, T3s * T3t);
|
cannam@167
|
2787 T82 = FNMS(T3s, T3r, T3q * T3t);
|
cannam@167
|
2788 T3w = ri[WS(rs, 26)];
|
cannam@167
|
2789 T3y = ii[WS(rs, 26)];
|
cannam@167
|
2790 T3z = FMA(T3v, T3w, T3x * T3y);
|
cannam@167
|
2791 T83 = FNMS(T3x, T3w, T3v * T3y);
|
cannam@167
|
2792 }
|
cannam@167
|
2793 T3A = T3u + T3z;
|
cannam@167
|
2794 Tej = T82 + T83;
|
cannam@167
|
2795 T81 = T3u - T3z;
|
cannam@167
|
2796 T84 = T82 - T83;
|
cannam@167
|
2797 }
|
cannam@167
|
2798 T36 = T2U + T35;
|
cannam@167
|
2799 T3B = T3p + T3A;
|
cannam@167
|
2800 TgH = T36 - T3B;
|
cannam@167
|
2801 TgE = Tec + Ted;
|
cannam@167
|
2802 TgF = Tei + Tej;
|
cannam@167
|
2803 TgG = TgE - TgF;
|
cannam@167
|
2804 {
|
cannam@167
|
2805 E T7Y, T7Z, Teh, Tek;
|
cannam@167
|
2806 T7Y = T7W - T7X;
|
cannam@167
|
2807 T7Z = T2Z - T34;
|
cannam@167
|
2808 T80 = T7Y + T7Z;
|
cannam@167
|
2809 TbW = T7Y - T7Z;
|
cannam@167
|
2810 Teh = T2U - T35;
|
cannam@167
|
2811 Tek = Tei - Tej;
|
cannam@167
|
2812 Tel = Teh - Tek;
|
cannam@167
|
2813 TfT = Teh + Tek;
|
cannam@167
|
2814 }
|
cannam@167
|
2815 {
|
cannam@167
|
2816 E T85, T8a, T8i, T8j;
|
cannam@167
|
2817 T85 = T81 - T84;
|
cannam@167
|
2818 T8a = T86 + T89;
|
cannam@167
|
2819 T8b = KP707106781 * (T85 - T8a);
|
cannam@167
|
2820 Tc0 = KP707106781 * (T8a + T85);
|
cannam@167
|
2821 T8i = T89 - T86;
|
cannam@167
|
2822 T8j = T81 + T84;
|
cannam@167
|
2823 T8k = KP707106781 * (T8i - T8j);
|
cannam@167
|
2824 TbX = KP707106781 * (T8i + T8j);
|
cannam@167
|
2825 }
|
cannam@167
|
2826 {
|
cannam@167
|
2827 E Tee, Tef, T8d, T8g;
|
cannam@167
|
2828 Tee = Tec - Ted;
|
cannam@167
|
2829 Tef = T3A - T3p;
|
cannam@167
|
2830 Teg = Tee - Tef;
|
cannam@167
|
2831 TfS = Tee + Tef;
|
cannam@167
|
2832 T8d = T2O - T2T;
|
cannam@167
|
2833 T8g = T8e - T8f;
|
cannam@167
|
2834 T8h = T8d - T8g;
|
cannam@167
|
2835 TbZ = T8d + T8g;
|
cannam@167
|
2836 }
|
cannam@167
|
2837 }
|
cannam@167
|
2838 {
|
cannam@167
|
2839 E T3H, T8n, T3M, T8o, T3N, Ten, T3S, T8F, T43, T8G, T44, Teo, T4e, Tet, T8x;
|
cannam@167
|
2840 E T8A, T4p, Teu, T8s, T8v;
|
cannam@167
|
2841 {
|
cannam@167
|
2842 E T3E, T3G, T3J, T3L;
|
cannam@167
|
2843 T3E = ri[WS(rs, 62)];
|
cannam@167
|
2844 T3G = ii[WS(rs, 62)];
|
cannam@167
|
2845 T3H = FMA(T3D, T3E, T3F * T3G);
|
cannam@167
|
2846 T8n = FNMS(T3F, T3E, T3D * T3G);
|
cannam@167
|
2847 T3J = ri[WS(rs, 30)];
|
cannam@167
|
2848 T3L = ii[WS(rs, 30)];
|
cannam@167
|
2849 T3M = FMA(T3I, T3J, T3K * T3L);
|
cannam@167
|
2850 T8o = FNMS(T3K, T3J, T3I * T3L);
|
cannam@167
|
2851 }
|
cannam@167
|
2852 T3N = T3H + T3M;
|
cannam@167
|
2853 Ten = T8n + T8o;
|
cannam@167
|
2854 {
|
cannam@167
|
2855 E T3P, T3R, T40, T42;
|
cannam@167
|
2856 T3P = ri[WS(rs, 14)];
|
cannam@167
|
2857 T3R = ii[WS(rs, 14)];
|
cannam@167
|
2858 T3S = FMA(T3O, T3P, T3Q * T3R);
|
cannam@167
|
2859 T8F = FNMS(T3Q, T3P, T3O * T3R);
|
cannam@167
|
2860 T40 = ri[WS(rs, 46)];
|
cannam@167
|
2861 T42 = ii[WS(rs, 46)];
|
cannam@167
|
2862 T43 = FMA(T3Z, T40, T41 * T42);
|
cannam@167
|
2863 T8G = FNMS(T41, T40, T3Z * T42);
|
cannam@167
|
2864 }
|
cannam@167
|
2865 T44 = T3S + T43;
|
cannam@167
|
2866 Teo = T8F + T8G;
|
cannam@167
|
2867 {
|
cannam@167
|
2868 E T48, T8y, T4d, T8z;
|
cannam@167
|
2869 {
|
cannam@167
|
2870 E T46, T47, T4a, T4c;
|
cannam@167
|
2871 T46 = ri[WS(rs, 6)];
|
cannam@167
|
2872 T47 = ii[WS(rs, 6)];
|
cannam@167
|
2873 T48 = FMA(T3c, T46, T3e * T47);
|
cannam@167
|
2874 T8y = FNMS(T3e, T46, T3c * T47);
|
cannam@167
|
2875 T4a = ri[WS(rs, 38)];
|
cannam@167
|
2876 T4c = ii[WS(rs, 38)];
|
cannam@167
|
2877 T4d = FMA(T49, T4a, T4b * T4c);
|
cannam@167
|
2878 T8z = FNMS(T4b, T4a, T49 * T4c);
|
cannam@167
|
2879 }
|
cannam@167
|
2880 T4e = T48 + T4d;
|
cannam@167
|
2881 Tet = T8y + T8z;
|
cannam@167
|
2882 T8x = T48 - T4d;
|
cannam@167
|
2883 T8A = T8y - T8z;
|
cannam@167
|
2884 }
|
cannam@167
|
2885 {
|
cannam@167
|
2886 E T4j, T8t, T4o, T8u;
|
cannam@167
|
2887 {
|
cannam@167
|
2888 E T4g, T4i, T4l, T4n;
|
cannam@167
|
2889 T4g = ri[WS(rs, 54)];
|
cannam@167
|
2890 T4i = ii[WS(rs, 54)];
|
cannam@167
|
2891 T4j = FMA(T4f, T4g, T4h * T4i);
|
cannam@167
|
2892 T8t = FNMS(T4h, T4g, T4f * T4i);
|
cannam@167
|
2893 T4l = ri[WS(rs, 22)];
|
cannam@167
|
2894 T4n = ii[WS(rs, 22)];
|
cannam@167
|
2895 T4o = FMA(T4k, T4l, T4m * T4n);
|
cannam@167
|
2896 T8u = FNMS(T4m, T4l, T4k * T4n);
|
cannam@167
|
2897 }
|
cannam@167
|
2898 T4p = T4j + T4o;
|
cannam@167
|
2899 Teu = T8t + T8u;
|
cannam@167
|
2900 T8s = T4j - T4o;
|
cannam@167
|
2901 T8v = T8t - T8u;
|
cannam@167
|
2902 }
|
cannam@167
|
2903 T45 = T3N + T44;
|
cannam@167
|
2904 T4q = T4e + T4p;
|
cannam@167
|
2905 TgJ = T45 - T4q;
|
cannam@167
|
2906 TgK = Ten + Teo;
|
cannam@167
|
2907 TgL = Tet + Teu;
|
cannam@167
|
2908 TgM = TgK - TgL;
|
cannam@167
|
2909 {
|
cannam@167
|
2910 E T8p, T8q, Tes, Tev;
|
cannam@167
|
2911 T8p = T8n - T8o;
|
cannam@167
|
2912 T8q = T3S - T43;
|
cannam@167
|
2913 T8r = T8p + T8q;
|
cannam@167
|
2914 Tc6 = T8p - T8q;
|
cannam@167
|
2915 Tes = T3N - T44;
|
cannam@167
|
2916 Tev = Tet - Teu;
|
cannam@167
|
2917 Tew = Tes - Tev;
|
cannam@167
|
2918 TfW = Tes + Tev;
|
cannam@167
|
2919 }
|
cannam@167
|
2920 {
|
cannam@167
|
2921 E T8w, T8B, T8J, T8K;
|
cannam@167
|
2922 T8w = T8s - T8v;
|
cannam@167
|
2923 T8B = T8x + T8A;
|
cannam@167
|
2924 T8C = KP707106781 * (T8w - T8B);
|
cannam@167
|
2925 Tc4 = KP707106781 * (T8B + T8w);
|
cannam@167
|
2926 T8J = T8A - T8x;
|
cannam@167
|
2927 T8K = T8s + T8v;
|
cannam@167
|
2928 T8L = KP707106781 * (T8J - T8K);
|
cannam@167
|
2929 Tc7 = KP707106781 * (T8J + T8K);
|
cannam@167
|
2930 }
|
cannam@167
|
2931 {
|
cannam@167
|
2932 E Tep, Teq, T8E, T8H;
|
cannam@167
|
2933 Tep = Ten - Teo;
|
cannam@167
|
2934 Teq = T4p - T4e;
|
cannam@167
|
2935 Ter = Tep - Teq;
|
cannam@167
|
2936 TfV = Tep + Teq;
|
cannam@167
|
2937 T8E = T3H - T3M;
|
cannam@167
|
2938 T8H = T8F - T8G;
|
cannam@167
|
2939 T8I = T8E - T8H;
|
cannam@167
|
2940 Tc3 = T8E + T8H;
|
cannam@167
|
2941 }
|
cannam@167
|
2942 }
|
cannam@167
|
2943 {
|
cannam@167
|
2944 E T5V, Tao, T64, Tap, T65, Tfi, T68, T9K, T6d, T9L, T6e, Tfj, T6o, Tf2, T9Q;
|
cannam@167
|
2945 E T9R, T6z, Tf3, T9T, T9W;
|
cannam@167
|
2946 {
|
cannam@167
|
2947 E T5T, T5U, T5Z, T63;
|
cannam@167
|
2948 T5T = ri[WS(rs, 63)];
|
cannam@167
|
2949 T5U = ii[WS(rs, 63)];
|
cannam@167
|
2950 T5V = FMA(TW, T5T, T10 * T5U);
|
cannam@167
|
2951 Tao = FNMS(T10, T5T, TW * T5U);
|
cannam@167
|
2952 T5Z = ri[WS(rs, 31)];
|
cannam@167
|
2953 T63 = ii[WS(rs, 31)];
|
cannam@167
|
2954 T64 = FMA(T5Y, T5Z, T62 * T63);
|
cannam@167
|
2955 Tap = FNMS(T62, T5Z, T5Y * T63);
|
cannam@167
|
2956 }
|
cannam@167
|
2957 T65 = T5V + T64;
|
cannam@167
|
2958 Tfi = Tao + Tap;
|
cannam@167
|
2959 {
|
cannam@167
|
2960 E T66, T67, T6a, T6c;
|
cannam@167
|
2961 T66 = ri[WS(rs, 15)];
|
cannam@167
|
2962 T67 = ii[WS(rs, 15)];
|
cannam@167
|
2963 T68 = FMA(TV, T66, TZ * T67);
|
cannam@167
|
2964 T9K = FNMS(TZ, T66, TV * T67);
|
cannam@167
|
2965 T6a = ri[WS(rs, 47)];
|
cannam@167
|
2966 T6c = ii[WS(rs, 47)];
|
cannam@167
|
2967 T6d = FMA(T69, T6a, T6b * T6c);
|
cannam@167
|
2968 T9L = FNMS(T6b, T6a, T69 * T6c);
|
cannam@167
|
2969 }
|
cannam@167
|
2970 T6e = T68 + T6d;
|
cannam@167
|
2971 Tfj = T9K + T9L;
|
cannam@167
|
2972 {
|
cannam@167
|
2973 E T6i, T9O, T6n, T9P;
|
cannam@167
|
2974 {
|
cannam@167
|
2975 E T6g, T6h, T6k, T6m;
|
cannam@167
|
2976 T6g = ri[WS(rs, 7)];
|
cannam@167
|
2977 T6h = ii[WS(rs, 7)];
|
cannam@167
|
2978 T6i = FMA(T1t, T6g, T1u * T6h);
|
cannam@167
|
2979 T9O = FNMS(T1u, T6g, T1t * T6h);
|
cannam@167
|
2980 T6k = ri[WS(rs, 39)];
|
cannam@167
|
2981 T6m = ii[WS(rs, 39)];
|
cannam@167
|
2982 T6n = FMA(T6j, T6k, T6l * T6m);
|
cannam@167
|
2983 T9P = FNMS(T6l, T6k, T6j * T6m);
|
cannam@167
|
2984 }
|
cannam@167
|
2985 T6o = T6i + T6n;
|
cannam@167
|
2986 Tf2 = T9O + T9P;
|
cannam@167
|
2987 T9Q = T9O - T9P;
|
cannam@167
|
2988 T9R = T6i - T6n;
|
cannam@167
|
2989 }
|
cannam@167
|
2990 {
|
cannam@167
|
2991 E T6t, T9U, T6y, T9V;
|
cannam@167
|
2992 {
|
cannam@167
|
2993 E T6q, T6s, T6v, T6x;
|
cannam@167
|
2994 T6q = ri[WS(rs, 55)];
|
cannam@167
|
2995 T6s = ii[WS(rs, 55)];
|
cannam@167
|
2996 T6t = FMA(T6p, T6q, T6r * T6s);
|
cannam@167
|
2997 T9U = FNMS(T6r, T6q, T6p * T6s);
|
cannam@167
|
2998 T6v = ri[WS(rs, 23)];
|
cannam@167
|
2999 T6x = ii[WS(rs, 23)];
|
cannam@167
|
3000 T6y = FMA(T6u, T6v, T6w * T6x);
|
cannam@167
|
3001 T9V = FNMS(T6w, T6v, T6u * T6x);
|
cannam@167
|
3002 }
|
cannam@167
|
3003 T6z = T6t + T6y;
|
cannam@167
|
3004 Tf3 = T9U + T9V;
|
cannam@167
|
3005 T9T = T6t - T6y;
|
cannam@167
|
3006 T9W = T9U - T9V;
|
cannam@167
|
3007 }
|
cannam@167
|
3008 {
|
cannam@167
|
3009 E T6f, T6A, Tfk, Tfl;
|
cannam@167
|
3010 T6f = T65 + T6e;
|
cannam@167
|
3011 T6A = T6o + T6z;
|
cannam@167
|
3012 T6B = T6f + T6A;
|
cannam@167
|
3013 Th1 = T6f - T6A;
|
cannam@167
|
3014 Tfk = Tfi - Tfj;
|
cannam@167
|
3015 Tfl = T6z - T6o;
|
cannam@167
|
3016 Tfm = Tfk - Tfl;
|
cannam@167
|
3017 Tga = Tfk + Tfl;
|
cannam@167
|
3018 }
|
cannam@167
|
3019 {
|
cannam@167
|
3020 E Th6, Th7, T9J, T9M;
|
cannam@167
|
3021 Th6 = Tfi + Tfj;
|
cannam@167
|
3022 Th7 = Tf2 + Tf3;
|
cannam@167
|
3023 Th8 = Th6 - Th7;
|
cannam@167
|
3024 ThI = Th6 + Th7;
|
cannam@167
|
3025 T9J = T5V - T64;
|
cannam@167
|
3026 T9M = T9K - T9L;
|
cannam@167
|
3027 T9N = T9J - T9M;
|
cannam@167
|
3028 Tcv = T9J + T9M;
|
cannam@167
|
3029 }
|
cannam@167
|
3030 {
|
cannam@167
|
3031 E T9S, T9X, Tat, Tau;
|
cannam@167
|
3032 T9S = T9Q - T9R;
|
cannam@167
|
3033 T9X = T9T + T9W;
|
cannam@167
|
3034 T9Y = KP707106781 * (T9S - T9X);
|
cannam@167
|
3035 TcH = KP707106781 * (T9S + T9X);
|
cannam@167
|
3036 Tat = T9T - T9W;
|
cannam@167
|
3037 Tau = T9R + T9Q;
|
cannam@167
|
3038 Tav = KP707106781 * (Tat - Tau);
|
cannam@167
|
3039 Tcw = KP707106781 * (Tau + Tat);
|
cannam@167
|
3040 }
|
cannam@167
|
3041 {
|
cannam@167
|
3042 E Tf1, Tf4, Taq, Tar;
|
cannam@167
|
3043 Tf1 = T65 - T6e;
|
cannam@167
|
3044 Tf4 = Tf2 - Tf3;
|
cannam@167
|
3045 Tf5 = Tf1 - Tf4;
|
cannam@167
|
3046 Tg7 = Tf1 + Tf4;
|
cannam@167
|
3047 Taq = Tao - Tap;
|
cannam@167
|
3048 Tar = T68 - T6d;
|
cannam@167
|
3049 Tas = Taq + Tar;
|
cannam@167
|
3050 TcG = Taq - Tar;
|
cannam@167
|
3051 }
|
cannam@167
|
3052 }
|
cannam@167
|
3053 {
|
cannam@167
|
3054 E T4w, T8Q, T4B, T8R, T4C, TeA, T4F, T9w, T4K, T9x, T4L, TeB, T4V, TeS, T90;
|
cannam@167
|
3055 E T93, T5a, TeT, T8V, T8Y;
|
cannam@167
|
3056 {
|
cannam@167
|
3057 E T4u, T4v, T4y, T4A;
|
cannam@167
|
3058 T4u = ri[WS(rs, 1)];
|
cannam@167
|
3059 T4v = ii[WS(rs, 1)];
|
cannam@167
|
3060 T4w = FMA(T2, T4u, T5 * T4v);
|
cannam@167
|
3061 T8Q = FNMS(T5, T4u, T2 * T4v);
|
cannam@167
|
3062 T4y = ri[WS(rs, 33)];
|
cannam@167
|
3063 T4A = ii[WS(rs, 33)];
|
cannam@167
|
3064 T4B = FMA(T4x, T4y, T4z * T4A);
|
cannam@167
|
3065 T8R = FNMS(T4z, T4y, T4x * T4A);
|
cannam@167
|
3066 }
|
cannam@167
|
3067 T4C = T4w + T4B;
|
cannam@167
|
3068 TeA = T8Q + T8R;
|
cannam@167
|
3069 {
|
cannam@167
|
3070 E T4D, T4E, T4H, T4J;
|
cannam@167
|
3071 T4D = ri[WS(rs, 17)];
|
cannam@167
|
3072 T4E = ii[WS(rs, 17)];
|
cannam@167
|
3073 T4F = FMA(T3V, T4D, T3Y * T4E);
|
cannam@167
|
3074 T9w = FNMS(T3Y, T4D, T3V * T4E);
|
cannam@167
|
3075 T4H = ri[WS(rs, 49)];
|
cannam@167
|
3076 T4J = ii[WS(rs, 49)];
|
cannam@167
|
3077 T4K = FMA(T4G, T4H, T4I * T4J);
|
cannam@167
|
3078 T9x = FNMS(T4I, T4H, T4G * T4J);
|
cannam@167
|
3079 }
|
cannam@167
|
3080 T4L = T4F + T4K;
|
cannam@167
|
3081 TeB = T9w + T9x;
|
cannam@167
|
3082 {
|
cannam@167
|
3083 E T4P, T91, T4U, T92;
|
cannam@167
|
3084 {
|
cannam@167
|
3085 E T4N, T4O, T4R, T4T;
|
cannam@167
|
3086 T4N = ri[WS(rs, 9)];
|
cannam@167
|
3087 T4O = ii[WS(rs, 9)];
|
cannam@167
|
3088 T4P = FMA(T9, T4N, Te * T4O);
|
cannam@167
|
3089 T91 = FNMS(Te, T4N, T9 * T4O);
|
cannam@167
|
3090 T4R = ri[WS(rs, 41)];
|
cannam@167
|
3091 T4T = ii[WS(rs, 41)];
|
cannam@167
|
3092 T4U = FMA(T4Q, T4R, T4S * T4T);
|
cannam@167
|
3093 T92 = FNMS(T4S, T4R, T4Q * T4T);
|
cannam@167
|
3094 }
|
cannam@167
|
3095 T4V = T4P + T4U;
|
cannam@167
|
3096 TeS = T91 + T92;
|
cannam@167
|
3097 T90 = T4P - T4U;
|
cannam@167
|
3098 T93 = T91 - T92;
|
cannam@167
|
3099 }
|
cannam@167
|
3100 {
|
cannam@167
|
3101 E T50, T8W, T59, T8X;
|
cannam@167
|
3102 {
|
cannam@167
|
3103 E T4X, T4Z, T54, T58;
|
cannam@167
|
3104 T4X = ri[WS(rs, 57)];
|
cannam@167
|
3105 T4Z = ii[WS(rs, 57)];
|
cannam@167
|
3106 T50 = FMA(T4W, T4X, T4Y * T4Z);
|
cannam@167
|
3107 T8W = FNMS(T4Y, T4X, T4W * T4Z);
|
cannam@167
|
3108 T54 = ri[WS(rs, 25)];
|
cannam@167
|
3109 T58 = ii[WS(rs, 25)];
|
cannam@167
|
3110 T59 = FMA(T53, T54, T57 * T58);
|
cannam@167
|
3111 T8X = FNMS(T57, T54, T53 * T58);
|
cannam@167
|
3112 }
|
cannam@167
|
3113 T5a = T50 + T59;
|
cannam@167
|
3114 TeT = T8W + T8X;
|
cannam@167
|
3115 T8V = T50 - T59;
|
cannam@167
|
3116 T8Y = T8W - T8X;
|
cannam@167
|
3117 }
|
cannam@167
|
3118 {
|
cannam@167
|
3119 E T4M, T5b, TeR, TeU;
|
cannam@167
|
3120 T4M = T4C + T4L;
|
cannam@167
|
3121 T5b = T4V + T5a;
|
cannam@167
|
3122 T5c = T4M + T5b;
|
cannam@167
|
3123 TgV = T4M - T5b;
|
cannam@167
|
3124 TeR = T4C - T4L;
|
cannam@167
|
3125 TeU = TeS - TeT;
|
cannam@167
|
3126 TeV = TeR - TeU;
|
cannam@167
|
3127 Tg0 = TeR + TeU;
|
cannam@167
|
3128 }
|
cannam@167
|
3129 {
|
cannam@167
|
3130 E TgQ, TgR, T8S, T8T;
|
cannam@167
|
3131 TgQ = TeA + TeB;
|
cannam@167
|
3132 TgR = TeS + TeT;
|
cannam@167
|
3133 TgS = TgQ - TgR;
|
cannam@167
|
3134 ThD = TgQ + TgR;
|
cannam@167
|
3135 T8S = T8Q - T8R;
|
cannam@167
|
3136 T8T = T4F - T4K;
|
cannam@167
|
3137 T8U = T8S + T8T;
|
cannam@167
|
3138 Tcc = T8S - T8T;
|
cannam@167
|
3139 }
|
cannam@167
|
3140 {
|
cannam@167
|
3141 E T8Z, T94, T9A, T9B;
|
cannam@167
|
3142 T8Z = T8V - T8Y;
|
cannam@167
|
3143 T94 = T90 + T93;
|
cannam@167
|
3144 T95 = KP707106781 * (T8Z - T94);
|
cannam@167
|
3145 Tco = KP707106781 * (T94 + T8Z);
|
cannam@167
|
3146 T9A = T93 - T90;
|
cannam@167
|
3147 T9B = T8V + T8Y;
|
cannam@167
|
3148 T9C = KP707106781 * (T9A - T9B);
|
cannam@167
|
3149 Tcd = KP707106781 * (T9A + T9B);
|
cannam@167
|
3150 }
|
cannam@167
|
3151 {
|
cannam@167
|
3152 E TeC, TeD, T9v, T9y;
|
cannam@167
|
3153 TeC = TeA - TeB;
|
cannam@167
|
3154 TeD = T5a - T4V;
|
cannam@167
|
3155 TeE = TeC - TeD;
|
cannam@167
|
3156 Tg3 = TeC + TeD;
|
cannam@167
|
3157 T9v = T4w - T4B;
|
cannam@167
|
3158 T9y = T9w - T9x;
|
cannam@167
|
3159 T9z = T9v - T9y;
|
cannam@167
|
3160 Tcn = T9v + T9y;
|
cannam@167
|
3161 }
|
cannam@167
|
3162 }
|
cannam@167
|
3163 {
|
cannam@167
|
3164 E T5l, TeL, T9k, T9n, T5P, TeH, T9a, T9f, T5u, TeM, T9l, T9q, T5G, TeG, T97;
|
cannam@167
|
3165 E T9e;
|
cannam@167
|
3166 {
|
cannam@167
|
3167 E T5f, T9i, T5k, T9j;
|
cannam@167
|
3168 {
|
cannam@167
|
3169 E T5d, T5e, T5h, T5j;
|
cannam@167
|
3170 T5d = ri[WS(rs, 5)];
|
cannam@167
|
3171 T5e = ii[WS(rs, 5)];
|
cannam@167
|
3172 T5f = FMA(Tg, T5d, Tl * T5e);
|
cannam@167
|
3173 T9i = FNMS(Tl, T5d, Tg * T5e);
|
cannam@167
|
3174 T5h = ri[WS(rs, 37)];
|
cannam@167
|
3175 T5j = ii[WS(rs, 37)];
|
cannam@167
|
3176 T5k = FMA(T5g, T5h, T5i * T5j);
|
cannam@167
|
3177 T9j = FNMS(T5i, T5h, T5g * T5j);
|
cannam@167
|
3178 }
|
cannam@167
|
3179 T5l = T5f + T5k;
|
cannam@167
|
3180 TeL = T9i + T9j;
|
cannam@167
|
3181 T9k = T9i - T9j;
|
cannam@167
|
3182 T9n = T5f - T5k;
|
cannam@167
|
3183 }
|
cannam@167
|
3184 {
|
cannam@167
|
3185 E T5J, T98, T5O, T99;
|
cannam@167
|
3186 {
|
cannam@167
|
3187 E T5H, T5I, T5L, T5N;
|
cannam@167
|
3188 T5H = ri[WS(rs, 13)];
|
cannam@167
|
3189 T5I = ii[WS(rs, 13)];
|
cannam@167
|
3190 T5J = FMA(T1h, T5H, T1j * T5I);
|
cannam@167
|
3191 T98 = FNMS(T1j, T5H, T1h * T5I);
|
cannam@167
|
3192 T5L = ri[WS(rs, 45)];
|
cannam@167
|
3193 T5N = ii[WS(rs, 45)];
|
cannam@167
|
3194 T5O = FMA(T5K, T5L, T5M * T5N);
|
cannam@167
|
3195 T99 = FNMS(T5M, T5L, T5K * T5N);
|
cannam@167
|
3196 }
|
cannam@167
|
3197 T5P = T5J + T5O;
|
cannam@167
|
3198 TeH = T98 + T99;
|
cannam@167
|
3199 T9a = T98 - T99;
|
cannam@167
|
3200 T9f = T5J - T5O;
|
cannam@167
|
3201 }
|
cannam@167
|
3202 {
|
cannam@167
|
3203 E T5o, T9o, T5t, T9p;
|
cannam@167
|
3204 {
|
cannam@167
|
3205 E T5m, T5n, T5q, T5s;
|
cannam@167
|
3206 T5m = ri[WS(rs, 21)];
|
cannam@167
|
3207 T5n = ii[WS(rs, 21)];
|
cannam@167
|
3208 T5o = FMA(T3g, T5m, T3j * T5n);
|
cannam@167
|
3209 T9o = FNMS(T3j, T5m, T3g * T5n);
|
cannam@167
|
3210 T5q = ri[WS(rs, 53)];
|
cannam@167
|
3211 T5s = ii[WS(rs, 53)];
|
cannam@167
|
3212 T5t = FMA(T5p, T5q, T5r * T5s);
|
cannam@167
|
3213 T9p = FNMS(T5r, T5q, T5p * T5s);
|
cannam@167
|
3214 }
|
cannam@167
|
3215 T5u = T5o + T5t;
|
cannam@167
|
3216 TeM = T9o + T9p;
|
cannam@167
|
3217 T9l = T5o - T5t;
|
cannam@167
|
3218 T9q = T9o - T9p;
|
cannam@167
|
3219 }
|
cannam@167
|
3220 {
|
cannam@167
|
3221 E T5A, T9c, T5F, T9d;
|
cannam@167
|
3222 {
|
cannam@167
|
3223 E T5x, T5z, T5C, T5E;
|
cannam@167
|
3224 T5x = ri[WS(rs, 61)];
|
cannam@167
|
3225 T5z = ii[WS(rs, 61)];
|
cannam@167
|
3226 T5A = FMA(T5w, T5x, T5y * T5z);
|
cannam@167
|
3227 T9c = FNMS(T5y, T5x, T5w * T5z);
|
cannam@167
|
3228 T5C = ri[WS(rs, 29)];
|
cannam@167
|
3229 T5E = ii[WS(rs, 29)];
|
cannam@167
|
3230 T5F = FMA(T5B, T5C, T5D * T5E);
|
cannam@167
|
3231 T9d = FNMS(T5D, T5C, T5B * T5E);
|
cannam@167
|
3232 }
|
cannam@167
|
3233 T5G = T5A + T5F;
|
cannam@167
|
3234 TeG = T9c + T9d;
|
cannam@167
|
3235 T97 = T5A - T5F;
|
cannam@167
|
3236 T9e = T9c - T9d;
|
cannam@167
|
3237 }
|
cannam@167
|
3238 {
|
cannam@167
|
3239 E T5v, T5Q, TeK, TeN;
|
cannam@167
|
3240 T5v = T5l + T5u;
|
cannam@167
|
3241 T5Q = T5G + T5P;
|
cannam@167
|
3242 T5R = T5v + T5Q;
|
cannam@167
|
3243 TgT = T5Q - T5v;
|
cannam@167
|
3244 TeK = T5l - T5u;
|
cannam@167
|
3245 TeN = TeL - TeM;
|
cannam@167
|
3246 TeO = TeK + TeN;
|
cannam@167
|
3247 TeW = TeN - TeK;
|
cannam@167
|
3248 }
|
cannam@167
|
3249 {
|
cannam@167
|
3250 E TgW, TgX, T9b, T9g;
|
cannam@167
|
3251 TgW = TeL + TeM;
|
cannam@167
|
3252 TgX = TeG + TeH;
|
cannam@167
|
3253 TgY = TgW - TgX;
|
cannam@167
|
3254 ThE = TgW + TgX;
|
cannam@167
|
3255 T9b = T97 - T9a;
|
cannam@167
|
3256 T9g = T9e + T9f;
|
cannam@167
|
3257 T9h = FNMS(KP923879532, T9g, KP382683432 * T9b);
|
cannam@167
|
3258 T9F = FMA(KP382683432, T9g, KP923879532 * T9b);
|
cannam@167
|
3259 }
|
cannam@167
|
3260 {
|
cannam@167
|
3261 E T9m, T9r, Tci, Tcj;
|
cannam@167
|
3262 T9m = T9k + T9l;
|
cannam@167
|
3263 T9r = T9n - T9q;
|
cannam@167
|
3264 T9s = FMA(KP923879532, T9m, KP382683432 * T9r);
|
cannam@167
|
3265 T9E = FNMS(KP923879532, T9r, KP382683432 * T9m);
|
cannam@167
|
3266 Tci = T9k - T9l;
|
cannam@167
|
3267 Tcj = T9n + T9q;
|
cannam@167
|
3268 Tck = FMA(KP382683432, Tci, KP923879532 * Tcj);
|
cannam@167
|
3269 Tcq = FNMS(KP382683432, Tcj, KP923879532 * Tci);
|
cannam@167
|
3270 }
|
cannam@167
|
3271 {
|
cannam@167
|
3272 E TeF, TeI, Tcf, Tcg;
|
cannam@167
|
3273 TeF = T5G - T5P;
|
cannam@167
|
3274 TeI = TeG - TeH;
|
cannam@167
|
3275 TeJ = TeF - TeI;
|
cannam@167
|
3276 TeX = TeF + TeI;
|
cannam@167
|
3277 Tcf = T97 + T9a;
|
cannam@167
|
3278 Tcg = T9e - T9f;
|
cannam@167
|
3279 Tch = FNMS(KP382683432, Tcg, KP923879532 * Tcf);
|
cannam@167
|
3280 Tcr = FMA(KP923879532, Tcg, KP382683432 * Tcf);
|
cannam@167
|
3281 }
|
cannam@167
|
3282 }
|
cannam@167
|
3283 {
|
cannam@167
|
3284 E T6K, Tf6, Ta2, Ta5, T7c, Tfd, Tae, Taj, T6T, Tf7, Ta3, Ta8, T73, Tfc, Tad;
|
cannam@167
|
3285 E Tag;
|
cannam@167
|
3286 {
|
cannam@167
|
3287 E T6E, Ta0, T6J, Ta1;
|
cannam@167
|
3288 {
|
cannam@167
|
3289 E T6C, T6D, T6G, T6I;
|
cannam@167
|
3290 T6C = ri[WS(rs, 3)];
|
cannam@167
|
3291 T6D = ii[WS(rs, 3)];
|
cannam@167
|
3292 T6E = FMA(T3, T6C, T6 * T6D);
|
cannam@167
|
3293 Ta0 = FNMS(T6, T6C, T3 * T6D);
|
cannam@167
|
3294 T6G = ri[WS(rs, 35)];
|
cannam@167
|
3295 T6I = ii[WS(rs, 35)];
|
cannam@167
|
3296 T6J = FMA(T6F, T6G, T6H * T6I);
|
cannam@167
|
3297 Ta1 = FNMS(T6H, T6G, T6F * T6I);
|
cannam@167
|
3298 }
|
cannam@167
|
3299 T6K = T6E + T6J;
|
cannam@167
|
3300 Tf6 = Ta0 + Ta1;
|
cannam@167
|
3301 Ta2 = Ta0 - Ta1;
|
cannam@167
|
3302 Ta5 = T6E - T6J;
|
cannam@167
|
3303 }
|
cannam@167
|
3304 {
|
cannam@167
|
3305 E T76, Tah, T7b, Tai;
|
cannam@167
|
3306 {
|
cannam@167
|
3307 E T74, T75, T78, T7a;
|
cannam@167
|
3308 T74 = ri[WS(rs, 11)];
|
cannam@167
|
3309 T75 = ii[WS(rs, 11)];
|
cannam@167
|
3310 T76 = FMA(TA, T74, TE * T75);
|
cannam@167
|
3311 Tah = FNMS(TE, T74, TA * T75);
|
cannam@167
|
3312 T78 = ri[WS(rs, 43)];
|
cannam@167
|
3313 T7a = ii[WS(rs, 43)];
|
cannam@167
|
3314 T7b = FMA(T77, T78, T79 * T7a);
|
cannam@167
|
3315 Tai = FNMS(T79, T78, T77 * T7a);
|
cannam@167
|
3316 }
|
cannam@167
|
3317 T7c = T76 + T7b;
|
cannam@167
|
3318 Tfd = Tah + Tai;
|
cannam@167
|
3319 Tae = T76 - T7b;
|
cannam@167
|
3320 Taj = Tah - Tai;
|
cannam@167
|
3321 }
|
cannam@167
|
3322 {
|
cannam@167
|
3323 E T6N, Ta6, T6S, Ta7;
|
cannam@167
|
3324 {
|
cannam@167
|
3325 E T6L, T6M, T6P, T6R;
|
cannam@167
|
3326 T6L = ri[WS(rs, 19)];
|
cannam@167
|
3327 T6M = ii[WS(rs, 19)];
|
cannam@167
|
3328 T6N = FMA(T2z, T6L, T2C * T6M);
|
cannam@167
|
3329 Ta6 = FNMS(T2C, T6L, T2z * T6M);
|
cannam@167
|
3330 T6P = ri[WS(rs, 51)];
|
cannam@167
|
3331 T6R = ii[WS(rs, 51)];
|
cannam@167
|
3332 T6S = FMA(T6O, T6P, T6Q * T6R);
|
cannam@167
|
3333 Ta7 = FNMS(T6Q, T6P, T6O * T6R);
|
cannam@167
|
3334 }
|
cannam@167
|
3335 T6T = T6N + T6S;
|
cannam@167
|
3336 Tf7 = Ta6 + Ta7;
|
cannam@167
|
3337 Ta3 = T6N - T6S;
|
cannam@167
|
3338 Ta8 = Ta6 - Ta7;
|
cannam@167
|
3339 }
|
cannam@167
|
3340 {
|
cannam@167
|
3341 E T6Z, Tab, T72, Tac;
|
cannam@167
|
3342 {
|
cannam@167
|
3343 E T6W, T6Y, T70, T71;
|
cannam@167
|
3344 T6W = ri[WS(rs, 59)];
|
cannam@167
|
3345 T6Y = ii[WS(rs, 59)];
|
cannam@167
|
3346 T6Z = FMA(T6V, T6W, T6X * T6Y);
|
cannam@167
|
3347 Tab = FNMS(T6X, T6W, T6V * T6Y);
|
cannam@167
|
3348 T70 = ri[WS(rs, 27)];
|
cannam@167
|
3349 T71 = ii[WS(rs, 27)];
|
cannam@167
|
3350 T72 = FMA(Th, T70, Tm * T71);
|
cannam@167
|
3351 Tac = FNMS(Tm, T70, Th * T71);
|
cannam@167
|
3352 }
|
cannam@167
|
3353 T73 = T6Z + T72;
|
cannam@167
|
3354 Tfc = Tab + Tac;
|
cannam@167
|
3355 Tad = Tab - Tac;
|
cannam@167
|
3356 Tag = T6Z - T72;
|
cannam@167
|
3357 }
|
cannam@167
|
3358 {
|
cannam@167
|
3359 E T6U, T7d, Tfb, Tfe;
|
cannam@167
|
3360 T6U = T6K + T6T;
|
cannam@167
|
3361 T7d = T73 + T7c;
|
cannam@167
|
3362 T7e = T6U + T7d;
|
cannam@167
|
3363 Th9 = T7d - T6U;
|
cannam@167
|
3364 Tfb = T73 - T7c;
|
cannam@167
|
3365 Tfe = Tfc - Tfd;
|
cannam@167
|
3366 Tff = Tfb + Tfe;
|
cannam@167
|
3367 Tfn = Tfb - Tfe;
|
cannam@167
|
3368 }
|
cannam@167
|
3369 {
|
cannam@167
|
3370 E Th2, Th3, Ta4, Ta9;
|
cannam@167
|
3371 Th2 = Tf6 + Tf7;
|
cannam@167
|
3372 Th3 = Tfc + Tfd;
|
cannam@167
|
3373 Th4 = Th2 - Th3;
|
cannam@167
|
3374 ThJ = Th2 + Th3;
|
cannam@167
|
3375 Ta4 = Ta2 + Ta3;
|
cannam@167
|
3376 Ta9 = Ta5 - Ta8;
|
cannam@167
|
3377 Taa = FNMS(KP923879532, Ta9, KP382683432 * Ta4);
|
cannam@167
|
3378 Tay = FMA(KP923879532, Ta4, KP382683432 * Ta9);
|
cannam@167
|
3379 }
|
cannam@167
|
3380 {
|
cannam@167
|
3381 E Taf, Tak, TcB, TcC;
|
cannam@167
|
3382 Taf = Tad + Tae;
|
cannam@167
|
3383 Tak = Tag - Taj;
|
cannam@167
|
3384 Tal = FMA(KP382683432, Taf, KP923879532 * Tak);
|
cannam@167
|
3385 Tax = FNMS(KP923879532, Taf, KP382683432 * Tak);
|
cannam@167
|
3386 TcB = Tad - Tae;
|
cannam@167
|
3387 TcC = Tag + Taj;
|
cannam@167
|
3388 TcD = FMA(KP923879532, TcB, KP382683432 * TcC);
|
cannam@167
|
3389 TcJ = FNMS(KP382683432, TcB, KP923879532 * TcC);
|
cannam@167
|
3390 }
|
cannam@167
|
3391 {
|
cannam@167
|
3392 E Tf8, Tf9, Tcy, Tcz;
|
cannam@167
|
3393 Tf8 = Tf6 - Tf7;
|
cannam@167
|
3394 Tf9 = T6K - T6T;
|
cannam@167
|
3395 Tfa = Tf8 - Tf9;
|
cannam@167
|
3396 Tfo = Tf9 + Tf8;
|
cannam@167
|
3397 Tcy = Ta2 - Ta3;
|
cannam@167
|
3398 Tcz = Ta5 + Ta8;
|
cannam@167
|
3399 TcA = FNMS(KP382683432, Tcz, KP923879532 * Tcy);
|
cannam@167
|
3400 TcK = FMA(KP382683432, Tcy, KP923879532 * Tcz);
|
cannam@167
|
3401 }
|
cannam@167
|
3402 }
|
cannam@167
|
3403 {
|
cannam@167
|
3404 E T2L, Thx, ThU, ThV, Ti5, Tib, T4s, Tia, T7g, Ti7, ThG, ThO, ThL, ThP, ThA;
|
cannam@167
|
3405 E ThW;
|
cannam@167
|
3406 {
|
cannam@167
|
3407 E T1L, T2K, ThS, ThT;
|
cannam@167
|
3408 T1L = T17 + T1K;
|
cannam@167
|
3409 T2K = T2e + T2J;
|
cannam@167
|
3410 T2L = T1L + T2K;
|
cannam@167
|
3411 Thx = T1L - T2K;
|
cannam@167
|
3412 ThS = ThD + ThE;
|
cannam@167
|
3413 ThT = ThI + ThJ;
|
cannam@167
|
3414 ThU = ThS - ThT;
|
cannam@167
|
3415 ThV = ThS + ThT;
|
cannam@167
|
3416 }
|
cannam@167
|
3417 {
|
cannam@167
|
3418 E ThX, Ti4, T3C, T4r;
|
cannam@167
|
3419 ThX = TgA + TgB;
|
cannam@167
|
3420 Ti4 = ThY + Ti3;
|
cannam@167
|
3421 Ti5 = ThX + Ti4;
|
cannam@167
|
3422 Tib = Ti4 - ThX;
|
cannam@167
|
3423 T3C = T36 + T3B;
|
cannam@167
|
3424 T4r = T45 + T4q;
|
cannam@167
|
3425 T4s = T3C + T4r;
|
cannam@167
|
3426 Tia = T4r - T3C;
|
cannam@167
|
3427 }
|
cannam@167
|
3428 {
|
cannam@167
|
3429 E T5S, T7f, ThC, ThF;
|
cannam@167
|
3430 T5S = T5c + T5R;
|
cannam@167
|
3431 T7f = T6B + T7e;
|
cannam@167
|
3432 T7g = T5S + T7f;
|
cannam@167
|
3433 Ti7 = T7f - T5S;
|
cannam@167
|
3434 ThC = T5c - T5R;
|
cannam@167
|
3435 ThF = ThD - ThE;
|
cannam@167
|
3436 ThG = ThC + ThF;
|
cannam@167
|
3437 ThO = ThF - ThC;
|
cannam@167
|
3438 }
|
cannam@167
|
3439 {
|
cannam@167
|
3440 E ThH, ThK, Thy, Thz;
|
cannam@167
|
3441 ThH = T6B - T7e;
|
cannam@167
|
3442 ThK = ThI - ThJ;
|
cannam@167
|
3443 ThL = ThH - ThK;
|
cannam@167
|
3444 ThP = ThH + ThK;
|
cannam@167
|
3445 Thy = TgE + TgF;
|
cannam@167
|
3446 Thz = TgK + TgL;
|
cannam@167
|
3447 ThA = Thy - Thz;
|
cannam@167
|
3448 ThW = Thy + Thz;
|
cannam@167
|
3449 }
|
cannam@167
|
3450 {
|
cannam@167
|
3451 E T4t, Ti6, ThR, Ti8;
|
cannam@167
|
3452 T4t = T2L + T4s;
|
cannam@167
|
3453 ri[WS(rs, 32)] = T4t - T7g;
|
cannam@167
|
3454 ri[0] = T4t + T7g;
|
cannam@167
|
3455 Ti6 = ThW + Ti5;
|
cannam@167
|
3456 ii[0] = ThV + Ti6;
|
cannam@167
|
3457 ii[WS(rs, 32)] = Ti6 - ThV;
|
cannam@167
|
3458 ThR = T2L - T4s;
|
cannam@167
|
3459 ri[WS(rs, 48)] = ThR - ThU;
|
cannam@167
|
3460 ri[WS(rs, 16)] = ThR + ThU;
|
cannam@167
|
3461 Ti8 = Ti5 - ThW;
|
cannam@167
|
3462 ii[WS(rs, 16)] = Ti7 + Ti8;
|
cannam@167
|
3463 ii[WS(rs, 48)] = Ti8 - Ti7;
|
cannam@167
|
3464 }
|
cannam@167
|
3465 {
|
cannam@167
|
3466 E ThB, ThM, Ti9, Tic;
|
cannam@167
|
3467 ThB = Thx + ThA;
|
cannam@167
|
3468 ThM = KP707106781 * (ThG + ThL);
|
cannam@167
|
3469 ri[WS(rs, 40)] = ThB - ThM;
|
cannam@167
|
3470 ri[WS(rs, 8)] = ThB + ThM;
|
cannam@167
|
3471 Ti9 = KP707106781 * (ThO + ThP);
|
cannam@167
|
3472 Tic = Tia + Tib;
|
cannam@167
|
3473 ii[WS(rs, 8)] = Ti9 + Tic;
|
cannam@167
|
3474 ii[WS(rs, 40)] = Tic - Ti9;
|
cannam@167
|
3475 }
|
cannam@167
|
3476 {
|
cannam@167
|
3477 E ThN, ThQ, Tid, Tie;
|
cannam@167
|
3478 ThN = Thx - ThA;
|
cannam@167
|
3479 ThQ = KP707106781 * (ThO - ThP);
|
cannam@167
|
3480 ri[WS(rs, 56)] = ThN - ThQ;
|
cannam@167
|
3481 ri[WS(rs, 24)] = ThN + ThQ;
|
cannam@167
|
3482 Tid = KP707106781 * (ThL - ThG);
|
cannam@167
|
3483 Tie = Tib - Tia;
|
cannam@167
|
3484 ii[WS(rs, 24)] = Tid + Tie;
|
cannam@167
|
3485 ii[WS(rs, 56)] = Tie - Tid;
|
cannam@167
|
3486 }
|
cannam@167
|
3487 }
|
cannam@167
|
3488 {
|
cannam@167
|
3489 E TgD, Thh, Thr, Thv, Tij, Tip, TgO, Tig, Th0, The, Thk, Tio, Tho, Thu, Thb;
|
cannam@167
|
3490 E Thf;
|
cannam@167
|
3491 {
|
cannam@167
|
3492 E Tgz, TgC, Thp, Thq;
|
cannam@167
|
3493 Tgz = T17 - T1K;
|
cannam@167
|
3494 TgC = TgA - TgB;
|
cannam@167
|
3495 TgD = Tgz - TgC;
|
cannam@167
|
3496 Thh = Tgz + TgC;
|
cannam@167
|
3497 Thp = Th1 + Th4;
|
cannam@167
|
3498 Thq = Th8 + Th9;
|
cannam@167
|
3499 Thr = FNMS(KP382683432, Thq, KP923879532 * Thp);
|
cannam@167
|
3500 Thv = FMA(KP923879532, Thq, KP382683432 * Thp);
|
cannam@167
|
3501 }
|
cannam@167
|
3502 {
|
cannam@167
|
3503 E Tih, Tii, TgI, TgN;
|
cannam@167
|
3504 Tih = T2J - T2e;
|
cannam@167
|
3505 Tii = Ti3 - ThY;
|
cannam@167
|
3506 Tij = Tih + Tii;
|
cannam@167
|
3507 Tip = Tii - Tih;
|
cannam@167
|
3508 TgI = TgG - TgH;
|
cannam@167
|
3509 TgN = TgJ + TgM;
|
cannam@167
|
3510 TgO = KP707106781 * (TgI - TgN);
|
cannam@167
|
3511 Tig = KP707106781 * (TgI + TgN);
|
cannam@167
|
3512 }
|
cannam@167
|
3513 {
|
cannam@167
|
3514 E TgU, TgZ, Thi, Thj;
|
cannam@167
|
3515 TgU = TgS - TgT;
|
cannam@167
|
3516 TgZ = TgV - TgY;
|
cannam@167
|
3517 Th0 = FMA(KP923879532, TgU, KP382683432 * TgZ);
|
cannam@167
|
3518 The = FNMS(KP923879532, TgZ, KP382683432 * TgU);
|
cannam@167
|
3519 Thi = TgH + TgG;
|
cannam@167
|
3520 Thj = TgJ - TgM;
|
cannam@167
|
3521 Thk = KP707106781 * (Thi + Thj);
|
cannam@167
|
3522 Tio = KP707106781 * (Thj - Thi);
|
cannam@167
|
3523 }
|
cannam@167
|
3524 {
|
cannam@167
|
3525 E Thm, Thn, Th5, Tha;
|
cannam@167
|
3526 Thm = TgS + TgT;
|
cannam@167
|
3527 Thn = TgV + TgY;
|
cannam@167
|
3528 Tho = FMA(KP382683432, Thm, KP923879532 * Thn);
|
cannam@167
|
3529 Thu = FNMS(KP382683432, Thn, KP923879532 * Thm);
|
cannam@167
|
3530 Th5 = Th1 - Th4;
|
cannam@167
|
3531 Tha = Th8 - Th9;
|
cannam@167
|
3532 Thb = FNMS(KP923879532, Tha, KP382683432 * Th5);
|
cannam@167
|
3533 Thf = FMA(KP382683432, Tha, KP923879532 * Th5);
|
cannam@167
|
3534 }
|
cannam@167
|
3535 {
|
cannam@167
|
3536 E TgP, Thc, Tin, Tiq;
|
cannam@167
|
3537 TgP = TgD + TgO;
|
cannam@167
|
3538 Thc = Th0 + Thb;
|
cannam@167
|
3539 ri[WS(rs, 44)] = TgP - Thc;
|
cannam@167
|
3540 ri[WS(rs, 12)] = TgP + Thc;
|
cannam@167
|
3541 Tin = The + Thf;
|
cannam@167
|
3542 Tiq = Tio + Tip;
|
cannam@167
|
3543 ii[WS(rs, 12)] = Tin + Tiq;
|
cannam@167
|
3544 ii[WS(rs, 44)] = Tiq - Tin;
|
cannam@167
|
3545 }
|
cannam@167
|
3546 {
|
cannam@167
|
3547 E Thd, Thg, Tir, Tis;
|
cannam@167
|
3548 Thd = TgD - TgO;
|
cannam@167
|
3549 Thg = The - Thf;
|
cannam@167
|
3550 ri[WS(rs, 60)] = Thd - Thg;
|
cannam@167
|
3551 ri[WS(rs, 28)] = Thd + Thg;
|
cannam@167
|
3552 Tir = Thb - Th0;
|
cannam@167
|
3553 Tis = Tip - Tio;
|
cannam@167
|
3554 ii[WS(rs, 28)] = Tir + Tis;
|
cannam@167
|
3555 ii[WS(rs, 60)] = Tis - Tir;
|
cannam@167
|
3556 }
|
cannam@167
|
3557 {
|
cannam@167
|
3558 E Thl, Ths, Tif, Tik;
|
cannam@167
|
3559 Thl = Thh + Thk;
|
cannam@167
|
3560 Ths = Tho + Thr;
|
cannam@167
|
3561 ri[WS(rs, 36)] = Thl - Ths;
|
cannam@167
|
3562 ri[WS(rs, 4)] = Thl + Ths;
|
cannam@167
|
3563 Tif = Thu + Thv;
|
cannam@167
|
3564 Tik = Tig + Tij;
|
cannam@167
|
3565 ii[WS(rs, 4)] = Tif + Tik;
|
cannam@167
|
3566 ii[WS(rs, 36)] = Tik - Tif;
|
cannam@167
|
3567 }
|
cannam@167
|
3568 {
|
cannam@167
|
3569 E Tht, Thw, Til, Tim;
|
cannam@167
|
3570 Tht = Thh - Thk;
|
cannam@167
|
3571 Thw = Thu - Thv;
|
cannam@167
|
3572 ri[WS(rs, 52)] = Tht - Thw;
|
cannam@167
|
3573 ri[WS(rs, 20)] = Tht + Thw;
|
cannam@167
|
3574 Til = Thr - Tho;
|
cannam@167
|
3575 Tim = Tij - Tig;
|
cannam@167
|
3576 ii[WS(rs, 20)] = Til + Tim;
|
cannam@167
|
3577 ii[WS(rs, 52)] = Tim - Til;
|
cannam@167
|
3578 }
|
cannam@167
|
3579 }
|
cannam@167
|
3580 {
|
cannam@167
|
3581 E Teb, Tfx, Tey, TiK, TiN, TiT, TfA, TiS, Tfr, TfL, Tfv, TfH, Tf0, TfK, Tfu;
|
cannam@167
|
3582 E TfE;
|
cannam@167
|
3583 {
|
cannam@167
|
3584 E TdZ, Tea, Tfy, Tfz;
|
cannam@167
|
3585 TdZ = TdV - TdY;
|
cannam@167
|
3586 Tea = KP707106781 * (Te4 - Te9);
|
cannam@167
|
3587 Teb = TdZ - Tea;
|
cannam@167
|
3588 Tfx = TdZ + Tea;
|
cannam@167
|
3589 {
|
cannam@167
|
3590 E Tem, Tex, TiL, TiM;
|
cannam@167
|
3591 Tem = FNMS(KP923879532, Tel, KP382683432 * Teg);
|
cannam@167
|
3592 Tex = FMA(KP382683432, Ter, KP923879532 * Tew);
|
cannam@167
|
3593 Tey = Tem - Tex;
|
cannam@167
|
3594 TiK = Tem + Tex;
|
cannam@167
|
3595 TiL = KP707106781 * (TfP - TfO);
|
cannam@167
|
3596 TiM = Tix - Tiw;
|
cannam@167
|
3597 TiN = TiL + TiM;
|
cannam@167
|
3598 TiT = TiM - TiL;
|
cannam@167
|
3599 }
|
cannam@167
|
3600 Tfy = FMA(KP923879532, Teg, KP382683432 * Tel);
|
cannam@167
|
3601 Tfz = FNMS(KP923879532, Ter, KP382683432 * Tew);
|
cannam@167
|
3602 TfA = Tfy + Tfz;
|
cannam@167
|
3603 TiS = Tfz - Tfy;
|
cannam@167
|
3604 {
|
cannam@167
|
3605 E Tfh, TfF, Tfq, TfG, Tfg, Tfp;
|
cannam@167
|
3606 Tfg = KP707106781 * (Tfa - Tff);
|
cannam@167
|
3607 Tfh = Tf5 - Tfg;
|
cannam@167
|
3608 TfF = Tf5 + Tfg;
|
cannam@167
|
3609 Tfp = KP707106781 * (Tfn - Tfo);
|
cannam@167
|
3610 Tfq = Tfm - Tfp;
|
cannam@167
|
3611 TfG = Tfm + Tfp;
|
cannam@167
|
3612 Tfr = FNMS(KP980785280, Tfq, KP195090322 * Tfh);
|
cannam@167
|
3613 TfL = FMA(KP831469612, TfG, KP555570233 * TfF);
|
cannam@167
|
3614 Tfv = FMA(KP195090322, Tfq, KP980785280 * Tfh);
|
cannam@167
|
3615 TfH = FNMS(KP555570233, TfG, KP831469612 * TfF);
|
cannam@167
|
3616 }
|
cannam@167
|
3617 {
|
cannam@167
|
3618 E TeQ, TfC, TeZ, TfD, TeP, TeY;
|
cannam@167
|
3619 TeP = KP707106781 * (TeJ - TeO);
|
cannam@167
|
3620 TeQ = TeE - TeP;
|
cannam@167
|
3621 TfC = TeE + TeP;
|
cannam@167
|
3622 TeY = KP707106781 * (TeW - TeX);
|
cannam@167
|
3623 TeZ = TeV - TeY;
|
cannam@167
|
3624 TfD = TeV + TeY;
|
cannam@167
|
3625 Tf0 = FMA(KP980785280, TeQ, KP195090322 * TeZ);
|
cannam@167
|
3626 TfK = FNMS(KP555570233, TfD, KP831469612 * TfC);
|
cannam@167
|
3627 Tfu = FNMS(KP980785280, TeZ, KP195090322 * TeQ);
|
cannam@167
|
3628 TfE = FMA(KP555570233, TfC, KP831469612 * TfD);
|
cannam@167
|
3629 }
|
cannam@167
|
3630 }
|
cannam@167
|
3631 {
|
cannam@167
|
3632 E Tez, Tfs, TiR, TiU;
|
cannam@167
|
3633 Tez = Teb + Tey;
|
cannam@167
|
3634 Tfs = Tf0 + Tfr;
|
cannam@167
|
3635 ri[WS(rs, 46)] = Tez - Tfs;
|
cannam@167
|
3636 ri[WS(rs, 14)] = Tez + Tfs;
|
cannam@167
|
3637 TiR = Tfu + Tfv;
|
cannam@167
|
3638 TiU = TiS + TiT;
|
cannam@167
|
3639 ii[WS(rs, 14)] = TiR + TiU;
|
cannam@167
|
3640 ii[WS(rs, 46)] = TiU - TiR;
|
cannam@167
|
3641 }
|
cannam@167
|
3642 {
|
cannam@167
|
3643 E Tft, Tfw, TiV, TiW;
|
cannam@167
|
3644 Tft = Teb - Tey;
|
cannam@167
|
3645 Tfw = Tfu - Tfv;
|
cannam@167
|
3646 ri[WS(rs, 62)] = Tft - Tfw;
|
cannam@167
|
3647 ri[WS(rs, 30)] = Tft + Tfw;
|
cannam@167
|
3648 TiV = Tfr - Tf0;
|
cannam@167
|
3649 TiW = TiT - TiS;
|
cannam@167
|
3650 ii[WS(rs, 30)] = TiV + TiW;
|
cannam@167
|
3651 ii[WS(rs, 62)] = TiW - TiV;
|
cannam@167
|
3652 }
|
cannam@167
|
3653 {
|
cannam@167
|
3654 E TfB, TfI, TiJ, TiO;
|
cannam@167
|
3655 TfB = Tfx + TfA;
|
cannam@167
|
3656 TfI = TfE + TfH;
|
cannam@167
|
3657 ri[WS(rs, 38)] = TfB - TfI;
|
cannam@167
|
3658 ri[WS(rs, 6)] = TfB + TfI;
|
cannam@167
|
3659 TiJ = TfK + TfL;
|
cannam@167
|
3660 TiO = TiK + TiN;
|
cannam@167
|
3661 ii[WS(rs, 6)] = TiJ + TiO;
|
cannam@167
|
3662 ii[WS(rs, 38)] = TiO - TiJ;
|
cannam@167
|
3663 }
|
cannam@167
|
3664 {
|
cannam@167
|
3665 E TfJ, TfM, TiP, TiQ;
|
cannam@167
|
3666 TfJ = Tfx - TfA;
|
cannam@167
|
3667 TfM = TfK - TfL;
|
cannam@167
|
3668 ri[WS(rs, 54)] = TfJ - TfM;
|
cannam@167
|
3669 ri[WS(rs, 22)] = TfJ + TfM;
|
cannam@167
|
3670 TiP = TfH - TfE;
|
cannam@167
|
3671 TiQ = TiN - TiK;
|
cannam@167
|
3672 ii[WS(rs, 22)] = TiP + TiQ;
|
cannam@167
|
3673 ii[WS(rs, 54)] = TiQ - TiP;
|
cannam@167
|
3674 }
|
cannam@167
|
3675 }
|
cannam@167
|
3676 {
|
cannam@167
|
3677 E TfR, Tgj, TfY, Tiu, Tiz, TiF, Tgm, TiE, Tgd, Tgx, Tgh, Tgt, Tg6, Tgw, Tgg;
|
cannam@167
|
3678 E Tgq;
|
cannam@167
|
3679 {
|
cannam@167
|
3680 E TfN, TfQ, Tgk, Tgl;
|
cannam@167
|
3681 TfN = TdV + TdY;
|
cannam@167
|
3682 TfQ = KP707106781 * (TfO + TfP);
|
cannam@167
|
3683 TfR = TfN - TfQ;
|
cannam@167
|
3684 Tgj = TfN + TfQ;
|
cannam@167
|
3685 {
|
cannam@167
|
3686 E TfU, TfX, Tiv, Tiy;
|
cannam@167
|
3687 TfU = FNMS(KP382683432, TfT, KP923879532 * TfS);
|
cannam@167
|
3688 TfX = FMA(KP923879532, TfV, KP382683432 * TfW);
|
cannam@167
|
3689 TfY = TfU - TfX;
|
cannam@167
|
3690 Tiu = TfU + TfX;
|
cannam@167
|
3691 Tiv = KP707106781 * (Te4 + Te9);
|
cannam@167
|
3692 Tiy = Tiw + Tix;
|
cannam@167
|
3693 Tiz = Tiv + Tiy;
|
cannam@167
|
3694 TiF = Tiy - Tiv;
|
cannam@167
|
3695 }
|
cannam@167
|
3696 Tgk = FMA(KP382683432, TfS, KP923879532 * TfT);
|
cannam@167
|
3697 Tgl = FNMS(KP382683432, TfV, KP923879532 * TfW);
|
cannam@167
|
3698 Tgm = Tgk + Tgl;
|
cannam@167
|
3699 TiE = Tgl - Tgk;
|
cannam@167
|
3700 {
|
cannam@167
|
3701 E Tg9, Tgr, Tgc, Tgs, Tg8, Tgb;
|
cannam@167
|
3702 Tg8 = KP707106781 * (Tfo + Tfn);
|
cannam@167
|
3703 Tg9 = Tg7 - Tg8;
|
cannam@167
|
3704 Tgr = Tg7 + Tg8;
|
cannam@167
|
3705 Tgb = KP707106781 * (Tfa + Tff);
|
cannam@167
|
3706 Tgc = Tga - Tgb;
|
cannam@167
|
3707 Tgs = Tga + Tgb;
|
cannam@167
|
3708 Tgd = FNMS(KP831469612, Tgc, KP555570233 * Tg9);
|
cannam@167
|
3709 Tgx = FMA(KP195090322, Tgr, KP980785280 * Tgs);
|
cannam@167
|
3710 Tgh = FMA(KP831469612, Tg9, KP555570233 * Tgc);
|
cannam@167
|
3711 Tgt = FNMS(KP195090322, Tgs, KP980785280 * Tgr);
|
cannam@167
|
3712 }
|
cannam@167
|
3713 {
|
cannam@167
|
3714 E Tg2, Tgo, Tg5, Tgp, Tg1, Tg4;
|
cannam@167
|
3715 Tg1 = KP707106781 * (TeO + TeJ);
|
cannam@167
|
3716 Tg2 = Tg0 - Tg1;
|
cannam@167
|
3717 Tgo = Tg0 + Tg1;
|
cannam@167
|
3718 Tg4 = KP707106781 * (TeW + TeX);
|
cannam@167
|
3719 Tg5 = Tg3 - Tg4;
|
cannam@167
|
3720 Tgp = Tg3 + Tg4;
|
cannam@167
|
3721 Tg6 = FMA(KP555570233, Tg2, KP831469612 * Tg5);
|
cannam@167
|
3722 Tgw = FNMS(KP195090322, Tgo, KP980785280 * Tgp);
|
cannam@167
|
3723 Tgg = FNMS(KP831469612, Tg2, KP555570233 * Tg5);
|
cannam@167
|
3724 Tgq = FMA(KP980785280, Tgo, KP195090322 * Tgp);
|
cannam@167
|
3725 }
|
cannam@167
|
3726 }
|
cannam@167
|
3727 {
|
cannam@167
|
3728 E TfZ, Tge, TiD, TiG;
|
cannam@167
|
3729 TfZ = TfR + TfY;
|
cannam@167
|
3730 Tge = Tg6 + Tgd;
|
cannam@167
|
3731 ri[WS(rs, 42)] = TfZ - Tge;
|
cannam@167
|
3732 ri[WS(rs, 10)] = TfZ + Tge;
|
cannam@167
|
3733 TiD = Tgg + Tgh;
|
cannam@167
|
3734 TiG = TiE + TiF;
|
cannam@167
|
3735 ii[WS(rs, 10)] = TiD + TiG;
|
cannam@167
|
3736 ii[WS(rs, 42)] = TiG - TiD;
|
cannam@167
|
3737 }
|
cannam@167
|
3738 {
|
cannam@167
|
3739 E Tgf, Tgi, TiH, TiI;
|
cannam@167
|
3740 Tgf = TfR - TfY;
|
cannam@167
|
3741 Tgi = Tgg - Tgh;
|
cannam@167
|
3742 ri[WS(rs, 58)] = Tgf - Tgi;
|
cannam@167
|
3743 ri[WS(rs, 26)] = Tgf + Tgi;
|
cannam@167
|
3744 TiH = Tgd - Tg6;
|
cannam@167
|
3745 TiI = TiF - TiE;
|
cannam@167
|
3746 ii[WS(rs, 26)] = TiH + TiI;
|
cannam@167
|
3747 ii[WS(rs, 58)] = TiI - TiH;
|
cannam@167
|
3748 }
|
cannam@167
|
3749 {
|
cannam@167
|
3750 E Tgn, Tgu, Tit, TiA;
|
cannam@167
|
3751 Tgn = Tgj + Tgm;
|
cannam@167
|
3752 Tgu = Tgq + Tgt;
|
cannam@167
|
3753 ri[WS(rs, 34)] = Tgn - Tgu;
|
cannam@167
|
3754 ri[WS(rs, 2)] = Tgn + Tgu;
|
cannam@167
|
3755 Tit = Tgw + Tgx;
|
cannam@167
|
3756 TiA = Tiu + Tiz;
|
cannam@167
|
3757 ii[WS(rs, 2)] = Tit + TiA;
|
cannam@167
|
3758 ii[WS(rs, 34)] = TiA - Tit;
|
cannam@167
|
3759 }
|
cannam@167
|
3760 {
|
cannam@167
|
3761 E Tgv, Tgy, TiB, TiC;
|
cannam@167
|
3762 Tgv = Tgj - Tgm;
|
cannam@167
|
3763 Tgy = Tgw - Tgx;
|
cannam@167
|
3764 ri[WS(rs, 50)] = Tgv - Tgy;
|
cannam@167
|
3765 ri[WS(rs, 18)] = Tgv + Tgy;
|
cannam@167
|
3766 TiB = Tgt - Tgq;
|
cannam@167
|
3767 TiC = Tiz - Tiu;
|
cannam@167
|
3768 ii[WS(rs, 18)] = TiB + TiC;
|
cannam@167
|
3769 ii[WS(rs, 50)] = TiC - TiB;
|
cannam@167
|
3770 }
|
cannam@167
|
3771 }
|
cannam@167
|
3772 {
|
cannam@167
|
3773 E T7V, TaH, TjN, TjT, T8O, TjS, TaK, TjK, T9I, TaU, TaE, TaO, TaB, TaV, TaF;
|
cannam@167
|
3774 E TaR;
|
cannam@167
|
3775 {
|
cannam@167
|
3776 E T7x, T7U, TjL, TjM;
|
cannam@167
|
3777 T7x = T7l - T7w;
|
cannam@167
|
3778 T7U = T7I - T7T;
|
cannam@167
|
3779 T7V = T7x - T7U;
|
cannam@167
|
3780 TaH = T7x + T7U;
|
cannam@167
|
3781 TjL = TaZ - TaY;
|
cannam@167
|
3782 TjM = Tjx - Tjw;
|
cannam@167
|
3783 TjN = TjL + TjM;
|
cannam@167
|
3784 TjT = TjM - TjL;
|
cannam@167
|
3785 }
|
cannam@167
|
3786 {
|
cannam@167
|
3787 E T8m, TaI, T8N, TaJ;
|
cannam@167
|
3788 {
|
cannam@167
|
3789 E T8c, T8l, T8D, T8M;
|
cannam@167
|
3790 T8c = T80 - T8b;
|
cannam@167
|
3791 T8l = T8h - T8k;
|
cannam@167
|
3792 T8m = FNMS(KP980785280, T8l, KP195090322 * T8c);
|
cannam@167
|
3793 TaI = FMA(KP980785280, T8c, KP195090322 * T8l);
|
cannam@167
|
3794 T8D = T8r - T8C;
|
cannam@167
|
3795 T8M = T8I - T8L;
|
cannam@167
|
3796 T8N = FMA(KP195090322, T8D, KP980785280 * T8M);
|
cannam@167
|
3797 TaJ = FNMS(KP980785280, T8D, KP195090322 * T8M);
|
cannam@167
|
3798 }
|
cannam@167
|
3799 T8O = T8m - T8N;
|
cannam@167
|
3800 TjS = TaJ - TaI;
|
cannam@167
|
3801 TaK = TaI + TaJ;
|
cannam@167
|
3802 TjK = T8m + T8N;
|
cannam@167
|
3803 }
|
cannam@167
|
3804 {
|
cannam@167
|
3805 E T9u, TaM, T9H, TaN;
|
cannam@167
|
3806 {
|
cannam@167
|
3807 E T96, T9t, T9D, T9G;
|
cannam@167
|
3808 T96 = T8U - T95;
|
cannam@167
|
3809 T9t = T9h - T9s;
|
cannam@167
|
3810 T9u = T96 - T9t;
|
cannam@167
|
3811 TaM = T96 + T9t;
|
cannam@167
|
3812 T9D = T9z - T9C;
|
cannam@167
|
3813 T9G = T9E - T9F;
|
cannam@167
|
3814 T9H = T9D - T9G;
|
cannam@167
|
3815 TaN = T9D + T9G;
|
cannam@167
|
3816 }
|
cannam@167
|
3817 T9I = FMA(KP995184726, T9u, KP098017140 * T9H);
|
cannam@167
|
3818 TaU = FNMS(KP634393284, TaN, KP773010453 * TaM);
|
cannam@167
|
3819 TaE = FNMS(KP995184726, T9H, KP098017140 * T9u);
|
cannam@167
|
3820 TaO = FMA(KP634393284, TaM, KP773010453 * TaN);
|
cannam@167
|
3821 }
|
cannam@167
|
3822 {
|
cannam@167
|
3823 E Tan, TaP, TaA, TaQ;
|
cannam@167
|
3824 {
|
cannam@167
|
3825 E T9Z, Tam, Taw, Taz;
|
cannam@167
|
3826 T9Z = T9N - T9Y;
|
cannam@167
|
3827 Tam = Taa - Tal;
|
cannam@167
|
3828 Tan = T9Z - Tam;
|
cannam@167
|
3829 TaP = T9Z + Tam;
|
cannam@167
|
3830 Taw = Tas - Tav;
|
cannam@167
|
3831 Taz = Tax - Tay;
|
cannam@167
|
3832 TaA = Taw - Taz;
|
cannam@167
|
3833 TaQ = Taw + Taz;
|
cannam@167
|
3834 }
|
cannam@167
|
3835 TaB = FNMS(KP995184726, TaA, KP098017140 * Tan);
|
cannam@167
|
3836 TaV = FMA(KP773010453, TaQ, KP634393284 * TaP);
|
cannam@167
|
3837 TaF = FMA(KP098017140, TaA, KP995184726 * Tan);
|
cannam@167
|
3838 TaR = FNMS(KP634393284, TaQ, KP773010453 * TaP);
|
cannam@167
|
3839 }
|
cannam@167
|
3840 {
|
cannam@167
|
3841 E T8P, TaC, TjR, TjU;
|
cannam@167
|
3842 T8P = T7V + T8O;
|
cannam@167
|
3843 TaC = T9I + TaB;
|
cannam@167
|
3844 ri[WS(rs, 47)] = T8P - TaC;
|
cannam@167
|
3845 ri[WS(rs, 15)] = T8P + TaC;
|
cannam@167
|
3846 TjR = TaE + TaF;
|
cannam@167
|
3847 TjU = TjS + TjT;
|
cannam@167
|
3848 ii[WS(rs, 15)] = TjR + TjU;
|
cannam@167
|
3849 ii[WS(rs, 47)] = TjU - TjR;
|
cannam@167
|
3850 }
|
cannam@167
|
3851 {
|
cannam@167
|
3852 E TaD, TaG, TjV, TjW;
|
cannam@167
|
3853 TaD = T7V - T8O;
|
cannam@167
|
3854 TaG = TaE - TaF;
|
cannam@167
|
3855 ri[WS(rs, 63)] = TaD - TaG;
|
cannam@167
|
3856 ri[WS(rs, 31)] = TaD + TaG;
|
cannam@167
|
3857 TjV = TaB - T9I;
|
cannam@167
|
3858 TjW = TjT - TjS;
|
cannam@167
|
3859 ii[WS(rs, 31)] = TjV + TjW;
|
cannam@167
|
3860 ii[WS(rs, 63)] = TjW - TjV;
|
cannam@167
|
3861 }
|
cannam@167
|
3862 {
|
cannam@167
|
3863 E TaL, TaS, TjJ, TjO;
|
cannam@167
|
3864 TaL = TaH + TaK;
|
cannam@167
|
3865 TaS = TaO + TaR;
|
cannam@167
|
3866 ri[WS(rs, 39)] = TaL - TaS;
|
cannam@167
|
3867 ri[WS(rs, 7)] = TaL + TaS;
|
cannam@167
|
3868 TjJ = TaU + TaV;
|
cannam@167
|
3869 TjO = TjK + TjN;
|
cannam@167
|
3870 ii[WS(rs, 7)] = TjJ + TjO;
|
cannam@167
|
3871 ii[WS(rs, 39)] = TjO - TjJ;
|
cannam@167
|
3872 }
|
cannam@167
|
3873 {
|
cannam@167
|
3874 E TaT, TaW, TjP, TjQ;
|
cannam@167
|
3875 TaT = TaH - TaK;
|
cannam@167
|
3876 TaW = TaU - TaV;
|
cannam@167
|
3877 ri[WS(rs, 55)] = TaT - TaW;
|
cannam@167
|
3878 ri[WS(rs, 23)] = TaT + TaW;
|
cannam@167
|
3879 TjP = TaR - TaO;
|
cannam@167
|
3880 TjQ = TjN - TjK;
|
cannam@167
|
3881 ii[WS(rs, 23)] = TjP + TjQ;
|
cannam@167
|
3882 ii[WS(rs, 55)] = TjQ - TjP;
|
cannam@167
|
3883 }
|
cannam@167
|
3884 }
|
cannam@167
|
3885 {
|
cannam@167
|
3886 E TbV, TcT, Tjj, Tjp, Tca, Tjo, TcW, Tjg, Tcu, Td6, TcQ, Td0, TcN, Td7, TcR;
|
cannam@167
|
3887 E Td3;
|
cannam@167
|
3888 {
|
cannam@167
|
3889 E TbN, TbU, Tjh, Tji;
|
cannam@167
|
3890 TbN = TbJ - TbM;
|
cannam@167
|
3891 TbU = TbQ - TbT;
|
cannam@167
|
3892 TbV = TbN - TbU;
|
cannam@167
|
3893 TcT = TbN + TbU;
|
cannam@167
|
3894 Tjh = Tdb - Tda;
|
cannam@167
|
3895 Tji = Tj3 - Tj0;
|
cannam@167
|
3896 Tjj = Tjh + Tji;
|
cannam@167
|
3897 Tjp = Tji - Tjh;
|
cannam@167
|
3898 }
|
cannam@167
|
3899 {
|
cannam@167
|
3900 E Tc2, TcU, Tc9, TcV;
|
cannam@167
|
3901 {
|
cannam@167
|
3902 E TbY, Tc1, Tc5, Tc8;
|
cannam@167
|
3903 TbY = TbW - TbX;
|
cannam@167
|
3904 Tc1 = TbZ - Tc0;
|
cannam@167
|
3905 Tc2 = FNMS(KP831469612, Tc1, KP555570233 * TbY);
|
cannam@167
|
3906 TcU = FMA(KP555570233, Tc1, KP831469612 * TbY);
|
cannam@167
|
3907 Tc5 = Tc3 - Tc4;
|
cannam@167
|
3908 Tc8 = Tc6 - Tc7;
|
cannam@167
|
3909 Tc9 = FMA(KP831469612, Tc5, KP555570233 * Tc8);
|
cannam@167
|
3910 TcV = FNMS(KP831469612, Tc8, KP555570233 * Tc5);
|
cannam@167
|
3911 }
|
cannam@167
|
3912 Tca = Tc2 - Tc9;
|
cannam@167
|
3913 Tjo = TcV - TcU;
|
cannam@167
|
3914 TcW = TcU + TcV;
|
cannam@167
|
3915 Tjg = Tc2 + Tc9;
|
cannam@167
|
3916 }
|
cannam@167
|
3917 {
|
cannam@167
|
3918 E Tcm, TcY, Tct, TcZ;
|
cannam@167
|
3919 {
|
cannam@167
|
3920 E Tce, Tcl, Tcp, Tcs;
|
cannam@167
|
3921 Tce = Tcc - Tcd;
|
cannam@167
|
3922 Tcl = Tch - Tck;
|
cannam@167
|
3923 Tcm = Tce - Tcl;
|
cannam@167
|
3924 TcY = Tce + Tcl;
|
cannam@167
|
3925 Tcp = Tcn - Tco;
|
cannam@167
|
3926 Tcs = Tcq - Tcr;
|
cannam@167
|
3927 Tct = Tcp - Tcs;
|
cannam@167
|
3928 TcZ = Tcp + Tcs;
|
cannam@167
|
3929 }
|
cannam@167
|
3930 Tcu = FMA(KP956940335, Tcm, KP290284677 * Tct);
|
cannam@167
|
3931 Td6 = FNMS(KP471396736, TcZ, KP881921264 * TcY);
|
cannam@167
|
3932 TcQ = FNMS(KP956940335, Tct, KP290284677 * Tcm);
|
cannam@167
|
3933 Td0 = FMA(KP471396736, TcY, KP881921264 * TcZ);
|
cannam@167
|
3934 }
|
cannam@167
|
3935 {
|
cannam@167
|
3936 E TcF, Td1, TcM, Td2;
|
cannam@167
|
3937 {
|
cannam@167
|
3938 E Tcx, TcE, TcI, TcL;
|
cannam@167
|
3939 Tcx = Tcv - Tcw;
|
cannam@167
|
3940 TcE = TcA - TcD;
|
cannam@167
|
3941 TcF = Tcx - TcE;
|
cannam@167
|
3942 Td1 = Tcx + TcE;
|
cannam@167
|
3943 TcI = TcG - TcH;
|
cannam@167
|
3944 TcL = TcJ - TcK;
|
cannam@167
|
3945 TcM = TcI - TcL;
|
cannam@167
|
3946 Td2 = TcI + TcL;
|
cannam@167
|
3947 }
|
cannam@167
|
3948 TcN = FNMS(KP956940335, TcM, KP290284677 * TcF);
|
cannam@167
|
3949 Td7 = FMA(KP881921264, Td2, KP471396736 * Td1);
|
cannam@167
|
3950 TcR = FMA(KP290284677, TcM, KP956940335 * TcF);
|
cannam@167
|
3951 Td3 = FNMS(KP471396736, Td2, KP881921264 * Td1);
|
cannam@167
|
3952 }
|
cannam@167
|
3953 {
|
cannam@167
|
3954 E Tcb, TcO, Tjn, Tjq;
|
cannam@167
|
3955 Tcb = TbV + Tca;
|
cannam@167
|
3956 TcO = Tcu + TcN;
|
cannam@167
|
3957 ri[WS(rs, 45)] = Tcb - TcO;
|
cannam@167
|
3958 ri[WS(rs, 13)] = Tcb + TcO;
|
cannam@167
|
3959 Tjn = TcQ + TcR;
|
cannam@167
|
3960 Tjq = Tjo + Tjp;
|
cannam@167
|
3961 ii[WS(rs, 13)] = Tjn + Tjq;
|
cannam@167
|
3962 ii[WS(rs, 45)] = Tjq - Tjn;
|
cannam@167
|
3963 }
|
cannam@167
|
3964 {
|
cannam@167
|
3965 E TcP, TcS, Tjr, Tjs;
|
cannam@167
|
3966 TcP = TbV - Tca;
|
cannam@167
|
3967 TcS = TcQ - TcR;
|
cannam@167
|
3968 ri[WS(rs, 61)] = TcP - TcS;
|
cannam@167
|
3969 ri[WS(rs, 29)] = TcP + TcS;
|
cannam@167
|
3970 Tjr = TcN - Tcu;
|
cannam@167
|
3971 Tjs = Tjp - Tjo;
|
cannam@167
|
3972 ii[WS(rs, 29)] = Tjr + Tjs;
|
cannam@167
|
3973 ii[WS(rs, 61)] = Tjs - Tjr;
|
cannam@167
|
3974 }
|
cannam@167
|
3975 {
|
cannam@167
|
3976 E TcX, Td4, Tjf, Tjk;
|
cannam@167
|
3977 TcX = TcT + TcW;
|
cannam@167
|
3978 Td4 = Td0 + Td3;
|
cannam@167
|
3979 ri[WS(rs, 37)] = TcX - Td4;
|
cannam@167
|
3980 ri[WS(rs, 5)] = TcX + Td4;
|
cannam@167
|
3981 Tjf = Td6 + Td7;
|
cannam@167
|
3982 Tjk = Tjg + Tjj;
|
cannam@167
|
3983 ii[WS(rs, 5)] = Tjf + Tjk;
|
cannam@167
|
3984 ii[WS(rs, 37)] = Tjk - Tjf;
|
cannam@167
|
3985 }
|
cannam@167
|
3986 {
|
cannam@167
|
3987 E Td5, Td8, Tjl, Tjm;
|
cannam@167
|
3988 Td5 = TcT - TcW;
|
cannam@167
|
3989 Td8 = Td6 - Td7;
|
cannam@167
|
3990 ri[WS(rs, 53)] = Td5 - Td8;
|
cannam@167
|
3991 ri[WS(rs, 21)] = Td5 + Td8;
|
cannam@167
|
3992 Tjl = Td3 - Td0;
|
cannam@167
|
3993 Tjm = Tjj - Tjg;
|
cannam@167
|
3994 ii[WS(rs, 21)] = Tjl + Tjm;
|
cannam@167
|
3995 ii[WS(rs, 53)] = Tjm - Tjl;
|
cannam@167
|
3996 }
|
cannam@167
|
3997 }
|
cannam@167
|
3998 {
|
cannam@167
|
3999 E Tdd, TdF, Tj5, Tjb, Tdk, Tja, TdI, TiY, Tds, TdS, TdC, TdM, Tdz, TdT, TdD;
|
cannam@167
|
4000 E TdP;
|
cannam@167
|
4001 {
|
cannam@167
|
4002 E Td9, Tdc, TiZ, Tj4;
|
cannam@167
|
4003 Td9 = TbJ + TbM;
|
cannam@167
|
4004 Tdc = Tda + Tdb;
|
cannam@167
|
4005 Tdd = Td9 - Tdc;
|
cannam@167
|
4006 TdF = Td9 + Tdc;
|
cannam@167
|
4007 TiZ = TbQ + TbT;
|
cannam@167
|
4008 Tj4 = Tj0 + Tj3;
|
cannam@167
|
4009 Tj5 = TiZ + Tj4;
|
cannam@167
|
4010 Tjb = Tj4 - TiZ;
|
cannam@167
|
4011 }
|
cannam@167
|
4012 {
|
cannam@167
|
4013 E Tdg, TdG, Tdj, TdH;
|
cannam@167
|
4014 {
|
cannam@167
|
4015 E Tde, Tdf, Tdh, Tdi;
|
cannam@167
|
4016 Tde = TbW + TbX;
|
cannam@167
|
4017 Tdf = TbZ + Tc0;
|
cannam@167
|
4018 Tdg = FNMS(KP195090322, Tdf, KP980785280 * Tde);
|
cannam@167
|
4019 TdG = FMA(KP980785280, Tdf, KP195090322 * Tde);
|
cannam@167
|
4020 Tdh = Tc3 + Tc4;
|
cannam@167
|
4021 Tdi = Tc6 + Tc7;
|
cannam@167
|
4022 Tdj = FMA(KP195090322, Tdh, KP980785280 * Tdi);
|
cannam@167
|
4023 TdH = FNMS(KP195090322, Tdi, KP980785280 * Tdh);
|
cannam@167
|
4024 }
|
cannam@167
|
4025 Tdk = Tdg - Tdj;
|
cannam@167
|
4026 Tja = TdH - TdG;
|
cannam@167
|
4027 TdI = TdG + TdH;
|
cannam@167
|
4028 TiY = Tdg + Tdj;
|
cannam@167
|
4029 }
|
cannam@167
|
4030 {
|
cannam@167
|
4031 E Tdo, TdK, Tdr, TdL;
|
cannam@167
|
4032 {
|
cannam@167
|
4033 E Tdm, Tdn, Tdp, Tdq;
|
cannam@167
|
4034 Tdm = Tcn + Tco;
|
cannam@167
|
4035 Tdn = Tck + Tch;
|
cannam@167
|
4036 Tdo = Tdm - Tdn;
|
cannam@167
|
4037 TdK = Tdm + Tdn;
|
cannam@167
|
4038 Tdp = Tcc + Tcd;
|
cannam@167
|
4039 Tdq = Tcq + Tcr;
|
cannam@167
|
4040 Tdr = Tdp - Tdq;
|
cannam@167
|
4041 TdL = Tdp + Tdq;
|
cannam@167
|
4042 }
|
cannam@167
|
4043 Tds = FMA(KP634393284, Tdo, KP773010453 * Tdr);
|
cannam@167
|
4044 TdS = FNMS(KP098017140, TdK, KP995184726 * TdL);
|
cannam@167
|
4045 TdC = FNMS(KP773010453, Tdo, KP634393284 * Tdr);
|
cannam@167
|
4046 TdM = FMA(KP995184726, TdK, KP098017140 * TdL);
|
cannam@167
|
4047 }
|
cannam@167
|
4048 {
|
cannam@167
|
4049 E Tdv, TdN, Tdy, TdO;
|
cannam@167
|
4050 {
|
cannam@167
|
4051 E Tdt, Tdu, Tdw, Tdx;
|
cannam@167
|
4052 Tdt = Tcv + Tcw;
|
cannam@167
|
4053 Tdu = TcK + TcJ;
|
cannam@167
|
4054 Tdv = Tdt - Tdu;
|
cannam@167
|
4055 TdN = Tdt + Tdu;
|
cannam@167
|
4056 Tdw = TcG + TcH;
|
cannam@167
|
4057 Tdx = TcA + TcD;
|
cannam@167
|
4058 Tdy = Tdw - Tdx;
|
cannam@167
|
4059 TdO = Tdw + Tdx;
|
cannam@167
|
4060 }
|
cannam@167
|
4061 Tdz = FNMS(KP773010453, Tdy, KP634393284 * Tdv);
|
cannam@167
|
4062 TdT = FMA(KP098017140, TdN, KP995184726 * TdO);
|
cannam@167
|
4063 TdD = FMA(KP773010453, Tdv, KP634393284 * Tdy);
|
cannam@167
|
4064 TdP = FNMS(KP098017140, TdO, KP995184726 * TdN);
|
cannam@167
|
4065 }
|
cannam@167
|
4066 {
|
cannam@167
|
4067 E Tdl, TdA, Tj9, Tjc;
|
cannam@167
|
4068 Tdl = Tdd + Tdk;
|
cannam@167
|
4069 TdA = Tds + Tdz;
|
cannam@167
|
4070 ri[WS(rs, 41)] = Tdl - TdA;
|
cannam@167
|
4071 ri[WS(rs, 9)] = Tdl + TdA;
|
cannam@167
|
4072 Tj9 = TdC + TdD;
|
cannam@167
|
4073 Tjc = Tja + Tjb;
|
cannam@167
|
4074 ii[WS(rs, 9)] = Tj9 + Tjc;
|
cannam@167
|
4075 ii[WS(rs, 41)] = Tjc - Tj9;
|
cannam@167
|
4076 }
|
cannam@167
|
4077 {
|
cannam@167
|
4078 E TdB, TdE, Tjd, Tje;
|
cannam@167
|
4079 TdB = Tdd - Tdk;
|
cannam@167
|
4080 TdE = TdC - TdD;
|
cannam@167
|
4081 ri[WS(rs, 57)] = TdB - TdE;
|
cannam@167
|
4082 ri[WS(rs, 25)] = TdB + TdE;
|
cannam@167
|
4083 Tjd = Tdz - Tds;
|
cannam@167
|
4084 Tje = Tjb - Tja;
|
cannam@167
|
4085 ii[WS(rs, 25)] = Tjd + Tje;
|
cannam@167
|
4086 ii[WS(rs, 57)] = Tje - Tjd;
|
cannam@167
|
4087 }
|
cannam@167
|
4088 {
|
cannam@167
|
4089 E TdJ, TdQ, TiX, Tj6;
|
cannam@167
|
4090 TdJ = TdF + TdI;
|
cannam@167
|
4091 TdQ = TdM + TdP;
|
cannam@167
|
4092 ri[WS(rs, 33)] = TdJ - TdQ;
|
cannam@167
|
4093 ri[WS(rs, 1)] = TdJ + TdQ;
|
cannam@167
|
4094 TiX = TdS + TdT;
|
cannam@167
|
4095 Tj6 = TiY + Tj5;
|
cannam@167
|
4096 ii[WS(rs, 1)] = TiX + Tj6;
|
cannam@167
|
4097 ii[WS(rs, 33)] = Tj6 - TiX;
|
cannam@167
|
4098 }
|
cannam@167
|
4099 {
|
cannam@167
|
4100 E TdR, TdU, Tj7, Tj8;
|
cannam@167
|
4101 TdR = TdF - TdI;
|
cannam@167
|
4102 TdU = TdS - TdT;
|
cannam@167
|
4103 ri[WS(rs, 49)] = TdR - TdU;
|
cannam@167
|
4104 ri[WS(rs, 17)] = TdR + TdU;
|
cannam@167
|
4105 Tj7 = TdP - TdM;
|
cannam@167
|
4106 Tj8 = Tj5 - TiY;
|
cannam@167
|
4107 ii[WS(rs, 17)] = Tj7 + Tj8;
|
cannam@167
|
4108 ii[WS(rs, 49)] = Tj8 - Tj7;
|
cannam@167
|
4109 }
|
cannam@167
|
4110 }
|
cannam@167
|
4111 {
|
cannam@167
|
4112 E Tb1, Tbt, Tjz, TjF, Tb8, TjE, Tbw, Tju, Tbg, TbG, Tbq, TbA, Tbn, TbH, Tbr;
|
cannam@167
|
4113 E TbD;
|
cannam@167
|
4114 {
|
cannam@167
|
4115 E TaX, Tb0, Tjv, Tjy;
|
cannam@167
|
4116 TaX = T7l + T7w;
|
cannam@167
|
4117 Tb0 = TaY + TaZ;
|
cannam@167
|
4118 Tb1 = TaX - Tb0;
|
cannam@167
|
4119 Tbt = TaX + Tb0;
|
cannam@167
|
4120 Tjv = T7I + T7T;
|
cannam@167
|
4121 Tjy = Tjw + Tjx;
|
cannam@167
|
4122 Tjz = Tjv + Tjy;
|
cannam@167
|
4123 TjF = Tjy - Tjv;
|
cannam@167
|
4124 }
|
cannam@167
|
4125 {
|
cannam@167
|
4126 E Tb4, Tbu, Tb7, Tbv;
|
cannam@167
|
4127 {
|
cannam@167
|
4128 E Tb2, Tb3, Tb5, Tb6;
|
cannam@167
|
4129 Tb2 = T80 + T8b;
|
cannam@167
|
4130 Tb3 = T8h + T8k;
|
cannam@167
|
4131 Tb4 = FNMS(KP555570233, Tb3, KP831469612 * Tb2);
|
cannam@167
|
4132 Tbu = FMA(KP555570233, Tb2, KP831469612 * Tb3);
|
cannam@167
|
4133 Tb5 = T8r + T8C;
|
cannam@167
|
4134 Tb6 = T8I + T8L;
|
cannam@167
|
4135 Tb7 = FMA(KP831469612, Tb5, KP555570233 * Tb6);
|
cannam@167
|
4136 Tbv = FNMS(KP555570233, Tb5, KP831469612 * Tb6);
|
cannam@167
|
4137 }
|
cannam@167
|
4138 Tb8 = Tb4 - Tb7;
|
cannam@167
|
4139 TjE = Tbv - Tbu;
|
cannam@167
|
4140 Tbw = Tbu + Tbv;
|
cannam@167
|
4141 Tju = Tb4 + Tb7;
|
cannam@167
|
4142 }
|
cannam@167
|
4143 {
|
cannam@167
|
4144 E Tbc, Tby, Tbf, Tbz;
|
cannam@167
|
4145 {
|
cannam@167
|
4146 E Tba, Tbb, Tbd, Tbe;
|
cannam@167
|
4147 Tba = T9z + T9C;
|
cannam@167
|
4148 Tbb = T9s + T9h;
|
cannam@167
|
4149 Tbc = Tba - Tbb;
|
cannam@167
|
4150 Tby = Tba + Tbb;
|
cannam@167
|
4151 Tbd = T8U + T95;
|
cannam@167
|
4152 Tbe = T9E + T9F;
|
cannam@167
|
4153 Tbf = Tbd - Tbe;
|
cannam@167
|
4154 Tbz = Tbd + Tbe;
|
cannam@167
|
4155 }
|
cannam@167
|
4156 Tbg = FMA(KP471396736, Tbc, KP881921264 * Tbf);
|
cannam@167
|
4157 TbG = FNMS(KP290284677, Tby, KP956940335 * Tbz);
|
cannam@167
|
4158 Tbq = FNMS(KP881921264, Tbc, KP471396736 * Tbf);
|
cannam@167
|
4159 TbA = FMA(KP956940335, Tby, KP290284677 * Tbz);
|
cannam@167
|
4160 }
|
cannam@167
|
4161 {
|
cannam@167
|
4162 E Tbj, TbB, Tbm, TbC;
|
cannam@167
|
4163 {
|
cannam@167
|
4164 E Tbh, Tbi, Tbk, Tbl;
|
cannam@167
|
4165 Tbh = T9N + T9Y;
|
cannam@167
|
4166 Tbi = Tay + Tax;
|
cannam@167
|
4167 Tbj = Tbh - Tbi;
|
cannam@167
|
4168 TbB = Tbh + Tbi;
|
cannam@167
|
4169 Tbk = Tas + Tav;
|
cannam@167
|
4170 Tbl = Taa + Tal;
|
cannam@167
|
4171 Tbm = Tbk - Tbl;
|
cannam@167
|
4172 TbC = Tbk + Tbl;
|
cannam@167
|
4173 }
|
cannam@167
|
4174 Tbn = FNMS(KP881921264, Tbm, KP471396736 * Tbj);
|
cannam@167
|
4175 TbH = FMA(KP290284677, TbB, KP956940335 * TbC);
|
cannam@167
|
4176 Tbr = FMA(KP881921264, Tbj, KP471396736 * Tbm);
|
cannam@167
|
4177 TbD = FNMS(KP290284677, TbC, KP956940335 * TbB);
|
cannam@167
|
4178 }
|
cannam@167
|
4179 {
|
cannam@167
|
4180 E Tb9, Tbo, TjD, TjG;
|
cannam@167
|
4181 Tb9 = Tb1 + Tb8;
|
cannam@167
|
4182 Tbo = Tbg + Tbn;
|
cannam@167
|
4183 ri[WS(rs, 43)] = Tb9 - Tbo;
|
cannam@167
|
4184 ri[WS(rs, 11)] = Tb9 + Tbo;
|
cannam@167
|
4185 TjD = Tbq + Tbr;
|
cannam@167
|
4186 TjG = TjE + TjF;
|
cannam@167
|
4187 ii[WS(rs, 11)] = TjD + TjG;
|
cannam@167
|
4188 ii[WS(rs, 43)] = TjG - TjD;
|
cannam@167
|
4189 }
|
cannam@167
|
4190 {
|
cannam@167
|
4191 E Tbp, Tbs, TjH, TjI;
|
cannam@167
|
4192 Tbp = Tb1 - Tb8;
|
cannam@167
|
4193 Tbs = Tbq - Tbr;
|
cannam@167
|
4194 ri[WS(rs, 59)] = Tbp - Tbs;
|
cannam@167
|
4195 ri[WS(rs, 27)] = Tbp + Tbs;
|
cannam@167
|
4196 TjH = Tbn - Tbg;
|
cannam@167
|
4197 TjI = TjF - TjE;
|
cannam@167
|
4198 ii[WS(rs, 27)] = TjH + TjI;
|
cannam@167
|
4199 ii[WS(rs, 59)] = TjI - TjH;
|
cannam@167
|
4200 }
|
cannam@167
|
4201 {
|
cannam@167
|
4202 E Tbx, TbE, Tjt, TjA;
|
cannam@167
|
4203 Tbx = Tbt + Tbw;
|
cannam@167
|
4204 TbE = TbA + TbD;
|
cannam@167
|
4205 ri[WS(rs, 35)] = Tbx - TbE;
|
cannam@167
|
4206 ri[WS(rs, 3)] = Tbx + TbE;
|
cannam@167
|
4207 Tjt = TbG + TbH;
|
cannam@167
|
4208 TjA = Tju + Tjz;
|
cannam@167
|
4209 ii[WS(rs, 3)] = Tjt + TjA;
|
cannam@167
|
4210 ii[WS(rs, 35)] = TjA - Tjt;
|
cannam@167
|
4211 }
|
cannam@167
|
4212 {
|
cannam@167
|
4213 E TbF, TbI, TjB, TjC;
|
cannam@167
|
4214 TbF = Tbt - Tbw;
|
cannam@167
|
4215 TbI = TbG - TbH;
|
cannam@167
|
4216 ri[WS(rs, 51)] = TbF - TbI;
|
cannam@167
|
4217 ri[WS(rs, 19)] = TbF + TbI;
|
cannam@167
|
4218 TjB = TbD - TbA;
|
cannam@167
|
4219 TjC = Tjz - Tju;
|
cannam@167
|
4220 ii[WS(rs, 19)] = TjB + TjC;
|
cannam@167
|
4221 ii[WS(rs, 51)] = TjC - TjB;
|
cannam@167
|
4222 }
|
cannam@167
|
4223 }
|
cannam@167
|
4224 }
|
cannam@167
|
4225 }
|
cannam@167
|
4226 }
|
cannam@167
|
4227 }
|
cannam@167
|
4228
|
cannam@167
|
4229 static const tw_instr twinstr[] = {
|
cannam@167
|
4230 {TW_CEXP, 0, 1},
|
cannam@167
|
4231 {TW_CEXP, 0, 3},
|
cannam@167
|
4232 {TW_CEXP, 0, 9},
|
cannam@167
|
4233 {TW_CEXP, 0, 27},
|
cannam@167
|
4234 {TW_CEXP, 0, 63},
|
cannam@167
|
4235 {TW_NEXT, 1, 0}
|
cannam@167
|
4236 };
|
cannam@167
|
4237
|
cannam@167
|
4238 static const ct_desc desc = { 64, "t2_64", twinstr, &GENUS, {880, 386, 274, 0}, 0, 0, 0 };
|
cannam@167
|
4239
|
cannam@167
|
4240 void X(codelet_t2_64) (planner *p) {
|
cannam@167
|
4241 X(kdft_dit_register) (p, t2_64, &desc);
|
cannam@167
|
4242 }
|
cannam@167
|
4243 #endif
|