Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:39:55 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-rdft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 64 -dit -name hf_64 -include hf.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 1038 FP additions, 644 FP multiplications,
|
Chris@10
|
32 * (or, 520 additions, 126 multiplications, 518 fused multiply/add),
|
Chris@10
|
33 * 246 stack variables, 15 constants, and 256 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hf.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hf_64(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@10
|
40 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@10
|
41 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@10
|
42 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@10
|
43 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
44 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
45 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@10
|
46 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
Chris@10
|
47 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
Chris@10
|
48 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
Chris@10
|
49 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
Chris@10
|
50 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
51 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@10
|
52 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
53 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@10
|
54 {
|
Chris@10
|
55 INT m;
|
Chris@10
|
56 for (m = mb, W = W + ((mb - 1) * 126); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 126, MAKE_VOLATILE_STRIDE(128, rs)) {
|
Chris@10
|
57 E Tku, Tky, Tkt, Tkx;
|
Chris@10
|
58 {
|
Chris@10
|
59 E TiV, Tjm, T7e, TcA, TjR, Tkl, Tm, TeM, T7Q, TcI, TeZ, Thr, T1G, TeW, TcJ;
|
Chris@10
|
60 E T7X, T87, TcN, Tf5, Thw, T29, Tf8, TcQ, T8u, Taq, Tdm, Tgc, ThX, T5K, TfS;
|
Chris@10
|
61 E Tdx, Tbj, TcB, T7l, TiP, TeP, Tjl, TN, TcC, T7s, T7B, TcF, TeU, Ths, T7I;
|
Chris@10
|
62 E TcG, T1f, TeR, T8G, TcU, Tfg, ThB, T32, Tfj, TcX, T93, T9h, Td3, TfK, ThM;
|
Chris@10
|
63 E T3X, Tfr, Tde, Taa, Thx, Tfb, Tf6, T2A, T8x, TcO, T8m, TcR, Tfm, ThC, T3t;
|
Chris@10
|
64 E Tfh, T96, TcV, T8V, TcY, ThN, Tfu, TfL, T4o, Tad, Td4, T9w, Tdf, TfV, ThY;
|
Chris@10
|
65 E T6b, Tg9, Tbm, Tdn, TaF, Tdy, ThJ, T4Q, TfN, TfA, Taf, T9M, Td8, Tdh, ThI;
|
Chris@10
|
66 E T5h, TfO, TfF, Tag, Ta1, Tdb, Tdi, ThU, T6D, Tgf, Tg1, Tbo, TaV, Tdr, TdA;
|
Chris@10
|
67 E Tb2, Tds, Tg5, ThT, Tg2, T74, Tdt, Tb9;
|
Chris@10
|
68 {
|
Chris@10
|
69 E T7a, Te, T78, T8, TjQ, TiU, T7c, Tk;
|
Chris@10
|
70 {
|
Chris@10
|
71 E T1, TiT, TiS, T7, Tg, Tj, Tf, Ti, T7b, Th;
|
Chris@10
|
72 T1 = cr[0];
|
Chris@10
|
73 TiT = ci[0];
|
Chris@10
|
74 {
|
Chris@10
|
75 E T3, T6, T2, T5;
|
Chris@10
|
76 T3 = cr[WS(rs, 32)];
|
Chris@10
|
77 T6 = ci[WS(rs, 32)];
|
Chris@10
|
78 T2 = W[62];
|
Chris@10
|
79 T5 = W[63];
|
Chris@10
|
80 {
|
Chris@10
|
81 E Ta, Td, Tc, T79, Tb, TiR, T4, T9;
|
Chris@10
|
82 Ta = cr[WS(rs, 16)];
|
Chris@10
|
83 Td = ci[WS(rs, 16)];
|
Chris@10
|
84 TiR = T2 * T6;
|
Chris@10
|
85 T4 = T2 * T3;
|
Chris@10
|
86 T9 = W[30];
|
Chris@10
|
87 Tc = W[31];
|
Chris@10
|
88 TiS = FNMS(T5, T3, TiR);
|
Chris@10
|
89 T7 = FMA(T5, T6, T4);
|
Chris@10
|
90 T79 = T9 * Td;
|
Chris@10
|
91 Tb = T9 * Ta;
|
Chris@10
|
92 Tg = cr[WS(rs, 48)];
|
Chris@10
|
93 Tj = ci[WS(rs, 48)];
|
Chris@10
|
94 T7a = FNMS(Tc, Ta, T79);
|
Chris@10
|
95 Te = FMA(Tc, Td, Tb);
|
Chris@10
|
96 Tf = W[94];
|
Chris@10
|
97 Ti = W[95];
|
Chris@10
|
98 }
|
Chris@10
|
99 }
|
Chris@10
|
100 T78 = T1 - T7;
|
Chris@10
|
101 T8 = T1 + T7;
|
Chris@10
|
102 TjQ = TiT - TiS;
|
Chris@10
|
103 TiU = TiS + TiT;
|
Chris@10
|
104 T7b = Tf * Tj;
|
Chris@10
|
105 Th = Tf * Tg;
|
Chris@10
|
106 T7c = FNMS(Ti, Tg, T7b);
|
Chris@10
|
107 Tk = FMA(Ti, Tj, Th);
|
Chris@10
|
108 }
|
Chris@10
|
109 {
|
Chris@10
|
110 E T7S, T1l, T7O, T1E, T1u, T1x, T1w, T7U, T1r, T7L, T1v;
|
Chris@10
|
111 {
|
Chris@10
|
112 E T1A, T1D, T1C, T7N, T1B;
|
Chris@10
|
113 {
|
Chris@10
|
114 E T1h, T1k, T1g, T1j, T7R, T1i, T1z;
|
Chris@10
|
115 T1h = cr[WS(rs, 60)];
|
Chris@10
|
116 T1k = ci[WS(rs, 60)];
|
Chris@10
|
117 {
|
Chris@10
|
118 E T7d, TiQ, Tl, TjP;
|
Chris@10
|
119 T7d = T7a - T7c;
|
Chris@10
|
120 TiQ = T7a + T7c;
|
Chris@10
|
121 Tl = Te + Tk;
|
Chris@10
|
122 TjP = Te - Tk;
|
Chris@10
|
123 TiV = TiQ + TiU;
|
Chris@10
|
124 Tjm = TiU - TiQ;
|
Chris@10
|
125 T7e = T78 - T7d;
|
Chris@10
|
126 TcA = T78 + T7d;
|
Chris@10
|
127 TjR = TjP + TjQ;
|
Chris@10
|
128 Tkl = TjQ - TjP;
|
Chris@10
|
129 Tm = T8 + Tl;
|
Chris@10
|
130 TeM = T8 - Tl;
|
Chris@10
|
131 T1g = W[118];
|
Chris@10
|
132 }
|
Chris@10
|
133 T1j = W[119];
|
Chris@10
|
134 T1A = cr[WS(rs, 44)];
|
Chris@10
|
135 T1D = ci[WS(rs, 44)];
|
Chris@10
|
136 T7R = T1g * T1k;
|
Chris@10
|
137 T1i = T1g * T1h;
|
Chris@10
|
138 T1z = W[86];
|
Chris@10
|
139 T1C = W[87];
|
Chris@10
|
140 T7S = FNMS(T1j, T1h, T7R);
|
Chris@10
|
141 T1l = FMA(T1j, T1k, T1i);
|
Chris@10
|
142 T7N = T1z * T1D;
|
Chris@10
|
143 T1B = T1z * T1A;
|
Chris@10
|
144 }
|
Chris@10
|
145 {
|
Chris@10
|
146 E T1n, T1q, T1m, T1p, T7T, T1o, T1t;
|
Chris@10
|
147 T1n = cr[WS(rs, 28)];
|
Chris@10
|
148 T1q = ci[WS(rs, 28)];
|
Chris@10
|
149 T7O = FNMS(T1C, T1A, T7N);
|
Chris@10
|
150 T1E = FMA(T1C, T1D, T1B);
|
Chris@10
|
151 T1m = W[54];
|
Chris@10
|
152 T1p = W[55];
|
Chris@10
|
153 T1u = cr[WS(rs, 12)];
|
Chris@10
|
154 T1x = ci[WS(rs, 12)];
|
Chris@10
|
155 T7T = T1m * T1q;
|
Chris@10
|
156 T1o = T1m * T1n;
|
Chris@10
|
157 T1t = W[22];
|
Chris@10
|
158 T1w = W[23];
|
Chris@10
|
159 T7U = FNMS(T1p, T1n, T7T);
|
Chris@10
|
160 T1r = FMA(T1p, T1q, T1o);
|
Chris@10
|
161 T7L = T1t * T1x;
|
Chris@10
|
162 T1v = T1t * T1u;
|
Chris@10
|
163 }
|
Chris@10
|
164 }
|
Chris@10
|
165 {
|
Chris@10
|
166 E T7V, TeX, T1s, T7K, T7M, T1y;
|
Chris@10
|
167 T7V = T7S - T7U;
|
Chris@10
|
168 TeX = T7S + T7U;
|
Chris@10
|
169 T1s = T1l + T1r;
|
Chris@10
|
170 T7K = T1l - T1r;
|
Chris@10
|
171 T7M = FNMS(T1w, T1u, T7L);
|
Chris@10
|
172 T1y = FMA(T1w, T1x, T1v);
|
Chris@10
|
173 {
|
Chris@10
|
174 E TeY, T7P, T7W, T1F;
|
Chris@10
|
175 TeY = T7M + T7O;
|
Chris@10
|
176 T7P = T7M - T7O;
|
Chris@10
|
177 T7W = T1y - T1E;
|
Chris@10
|
178 T1F = T1y + T1E;
|
Chris@10
|
179 T7Q = T7K - T7P;
|
Chris@10
|
180 TcI = T7K + T7P;
|
Chris@10
|
181 TeZ = TeX - TeY;
|
Chris@10
|
182 Thr = TeX + TeY;
|
Chris@10
|
183 T1G = T1s + T1F;
|
Chris@10
|
184 TeW = T1s - T1F;
|
Chris@10
|
185 TcJ = T7V - T7W;
|
Chris@10
|
186 T7X = T7V + T7W;
|
Chris@10
|
187 }
|
Chris@10
|
188 }
|
Chris@10
|
189 }
|
Chris@10
|
190 }
|
Chris@10
|
191 {
|
Chris@10
|
192 E T8p, T1O, T85, T27, T1X, T20, T1Z, T8r, T1U, T82, T1Y;
|
Chris@10
|
193 {
|
Chris@10
|
194 E T23, T26, T25, T84, T24;
|
Chris@10
|
195 {
|
Chris@10
|
196 E T1K, T1N, T1J, T1M, T8o, T1L, T22;
|
Chris@10
|
197 T1K = cr[WS(rs, 2)];
|
Chris@10
|
198 T1N = ci[WS(rs, 2)];
|
Chris@10
|
199 T1J = W[2];
|
Chris@10
|
200 T1M = W[3];
|
Chris@10
|
201 T23 = cr[WS(rs, 50)];
|
Chris@10
|
202 T26 = ci[WS(rs, 50)];
|
Chris@10
|
203 T8o = T1J * T1N;
|
Chris@10
|
204 T1L = T1J * T1K;
|
Chris@10
|
205 T22 = W[98];
|
Chris@10
|
206 T25 = W[99];
|
Chris@10
|
207 T8p = FNMS(T1M, T1K, T8o);
|
Chris@10
|
208 T1O = FMA(T1M, T1N, T1L);
|
Chris@10
|
209 T84 = T22 * T26;
|
Chris@10
|
210 T24 = T22 * T23;
|
Chris@10
|
211 }
|
Chris@10
|
212 {
|
Chris@10
|
213 E T1Q, T1T, T1P, T1S, T8q, T1R, T1W;
|
Chris@10
|
214 T1Q = cr[WS(rs, 34)];
|
Chris@10
|
215 T1T = ci[WS(rs, 34)];
|
Chris@10
|
216 T85 = FNMS(T25, T23, T84);
|
Chris@10
|
217 T27 = FMA(T25, T26, T24);
|
Chris@10
|
218 T1P = W[66];
|
Chris@10
|
219 T1S = W[67];
|
Chris@10
|
220 T1X = cr[WS(rs, 18)];
|
Chris@10
|
221 T20 = ci[WS(rs, 18)];
|
Chris@10
|
222 T8q = T1P * T1T;
|
Chris@10
|
223 T1R = T1P * T1Q;
|
Chris@10
|
224 T1W = W[34];
|
Chris@10
|
225 T1Z = W[35];
|
Chris@10
|
226 T8r = FNMS(T1S, T1Q, T8q);
|
Chris@10
|
227 T1U = FMA(T1S, T1T, T1R);
|
Chris@10
|
228 T82 = T1W * T20;
|
Chris@10
|
229 T1Y = T1W * T1X;
|
Chris@10
|
230 }
|
Chris@10
|
231 }
|
Chris@10
|
232 {
|
Chris@10
|
233 E T8s, Tf3, T1V, T81, T83, T21;
|
Chris@10
|
234 T8s = T8p - T8r;
|
Chris@10
|
235 Tf3 = T8p + T8r;
|
Chris@10
|
236 T1V = T1O + T1U;
|
Chris@10
|
237 T81 = T1O - T1U;
|
Chris@10
|
238 T83 = FNMS(T1Z, T1X, T82);
|
Chris@10
|
239 T21 = FMA(T1Z, T20, T1Y);
|
Chris@10
|
240 {
|
Chris@10
|
241 E Tf4, T86, T8t, T28;
|
Chris@10
|
242 Tf4 = T83 + T85;
|
Chris@10
|
243 T86 = T83 - T85;
|
Chris@10
|
244 T8t = T21 - T27;
|
Chris@10
|
245 T28 = T21 + T27;
|
Chris@10
|
246 T87 = T81 - T86;
|
Chris@10
|
247 TcN = T81 + T86;
|
Chris@10
|
248 Tf5 = Tf3 - Tf4;
|
Chris@10
|
249 Thw = Tf3 + Tf4;
|
Chris@10
|
250 T29 = T1V + T28;
|
Chris@10
|
251 Tf8 = T1V - T28;
|
Chris@10
|
252 TcQ = T8s - T8t;
|
Chris@10
|
253 T8u = T8s + T8t;
|
Chris@10
|
254 }
|
Chris@10
|
255 }
|
Chris@10
|
256 }
|
Chris@10
|
257 {
|
Chris@10
|
258 E Tbf, T5p, Tao, T5I, T5y, T5B, T5A, Tbh, T5v, Tal, T5z;
|
Chris@10
|
259 {
|
Chris@10
|
260 E T5E, T5H, T5G, Tan, T5F;
|
Chris@10
|
261 {
|
Chris@10
|
262 E T5l, T5o, T5k, T5n, Tbe, T5m, T5D;
|
Chris@10
|
263 T5l = cr[WS(rs, 63)];
|
Chris@10
|
264 T5o = ci[WS(rs, 63)];
|
Chris@10
|
265 T5k = W[124];
|
Chris@10
|
266 T5n = W[125];
|
Chris@10
|
267 T5E = cr[WS(rs, 47)];
|
Chris@10
|
268 T5H = ci[WS(rs, 47)];
|
Chris@10
|
269 Tbe = T5k * T5o;
|
Chris@10
|
270 T5m = T5k * T5l;
|
Chris@10
|
271 T5D = W[92];
|
Chris@10
|
272 T5G = W[93];
|
Chris@10
|
273 Tbf = FNMS(T5n, T5l, Tbe);
|
Chris@10
|
274 T5p = FMA(T5n, T5o, T5m);
|
Chris@10
|
275 Tan = T5D * T5H;
|
Chris@10
|
276 T5F = T5D * T5E;
|
Chris@10
|
277 }
|
Chris@10
|
278 {
|
Chris@10
|
279 E T5r, T5u, T5q, T5t, Tbg, T5s, T5x;
|
Chris@10
|
280 T5r = cr[WS(rs, 31)];
|
Chris@10
|
281 T5u = ci[WS(rs, 31)];
|
Chris@10
|
282 Tao = FNMS(T5G, T5E, Tan);
|
Chris@10
|
283 T5I = FMA(T5G, T5H, T5F);
|
Chris@10
|
284 T5q = W[60];
|
Chris@10
|
285 T5t = W[61];
|
Chris@10
|
286 T5y = cr[WS(rs, 15)];
|
Chris@10
|
287 T5B = ci[WS(rs, 15)];
|
Chris@10
|
288 Tbg = T5q * T5u;
|
Chris@10
|
289 T5s = T5q * T5r;
|
Chris@10
|
290 T5x = W[28];
|
Chris@10
|
291 T5A = W[29];
|
Chris@10
|
292 Tbh = FNMS(T5t, T5r, Tbg);
|
Chris@10
|
293 T5v = FMA(T5t, T5u, T5s);
|
Chris@10
|
294 Tal = T5x * T5B;
|
Chris@10
|
295 T5z = T5x * T5y;
|
Chris@10
|
296 }
|
Chris@10
|
297 }
|
Chris@10
|
298 {
|
Chris@10
|
299 E Tbi, Tga, T5w, Tak, Tam, T5C;
|
Chris@10
|
300 Tbi = Tbf - Tbh;
|
Chris@10
|
301 Tga = Tbf + Tbh;
|
Chris@10
|
302 T5w = T5p + T5v;
|
Chris@10
|
303 Tak = T5p - T5v;
|
Chris@10
|
304 Tam = FNMS(T5A, T5y, Tal);
|
Chris@10
|
305 T5C = FMA(T5A, T5B, T5z);
|
Chris@10
|
306 {
|
Chris@10
|
307 E Tgb, Tap, T5J, Tbd;
|
Chris@10
|
308 Tgb = Tam + Tao;
|
Chris@10
|
309 Tap = Tam - Tao;
|
Chris@10
|
310 T5J = T5C + T5I;
|
Chris@10
|
311 Tbd = T5I - T5C;
|
Chris@10
|
312 Taq = Tak - Tap;
|
Chris@10
|
313 Tdm = Tak + Tap;
|
Chris@10
|
314 Tgc = Tga - Tgb;
|
Chris@10
|
315 ThX = Tga + Tgb;
|
Chris@10
|
316 T5K = T5w + T5J;
|
Chris@10
|
317 TfS = T5w - T5J;
|
Chris@10
|
318 Tdx = Tbi + Tbd;
|
Chris@10
|
319 Tbj = Tbd - Tbi;
|
Chris@10
|
320 }
|
Chris@10
|
321 }
|
Chris@10
|
322 }
|
Chris@10
|
323 {
|
Chris@10
|
324 E T7z, T1d, T7G, TeS, T11, T7v, T7x, T17, T7r, T7m;
|
Chris@10
|
325 {
|
Chris@10
|
326 E T7h, Ts, T7q, TL, TB, TE, TD, T7j, Ty, T7n, TC;
|
Chris@10
|
327 {
|
Chris@10
|
328 E TH, TK, TJ, T7p, TI;
|
Chris@10
|
329 {
|
Chris@10
|
330 E To, Tr, Tn, Tq, T7g, Tp, TG;
|
Chris@10
|
331 To = cr[WS(rs, 8)];
|
Chris@10
|
332 Tr = ci[WS(rs, 8)];
|
Chris@10
|
333 Tn = W[14];
|
Chris@10
|
334 Tq = W[15];
|
Chris@10
|
335 TH = cr[WS(rs, 24)];
|
Chris@10
|
336 TK = ci[WS(rs, 24)];
|
Chris@10
|
337 T7g = Tn * Tr;
|
Chris@10
|
338 Tp = Tn * To;
|
Chris@10
|
339 TG = W[46];
|
Chris@10
|
340 TJ = W[47];
|
Chris@10
|
341 T7h = FNMS(Tq, To, T7g);
|
Chris@10
|
342 Ts = FMA(Tq, Tr, Tp);
|
Chris@10
|
343 T7p = TG * TK;
|
Chris@10
|
344 TI = TG * TH;
|
Chris@10
|
345 }
|
Chris@10
|
346 {
|
Chris@10
|
347 E Tu, Tx, Tt, Tw, T7i, Tv, TA;
|
Chris@10
|
348 Tu = cr[WS(rs, 40)];
|
Chris@10
|
349 Tx = ci[WS(rs, 40)];
|
Chris@10
|
350 T7q = FNMS(TJ, TH, T7p);
|
Chris@10
|
351 TL = FMA(TJ, TK, TI);
|
Chris@10
|
352 Tt = W[78];
|
Chris@10
|
353 Tw = W[79];
|
Chris@10
|
354 TB = cr[WS(rs, 56)];
|
Chris@10
|
355 TE = ci[WS(rs, 56)];
|
Chris@10
|
356 T7i = Tt * Tx;
|
Chris@10
|
357 Tv = Tt * Tu;
|
Chris@10
|
358 TA = W[110];
|
Chris@10
|
359 TD = W[111];
|
Chris@10
|
360 T7j = FNMS(Tw, Tu, T7i);
|
Chris@10
|
361 Ty = FMA(Tw, Tx, Tv);
|
Chris@10
|
362 T7n = TA * TE;
|
Chris@10
|
363 TC = TA * TB;
|
Chris@10
|
364 }
|
Chris@10
|
365 }
|
Chris@10
|
366 {
|
Chris@10
|
367 E T7k, TeO, Tz, T7f, T7o, TF, TeN, TM;
|
Chris@10
|
368 T7k = T7h - T7j;
|
Chris@10
|
369 TeO = T7h + T7j;
|
Chris@10
|
370 Tz = Ts + Ty;
|
Chris@10
|
371 T7f = Ts - Ty;
|
Chris@10
|
372 T7o = FNMS(TD, TB, T7n);
|
Chris@10
|
373 TF = FMA(TD, TE, TC);
|
Chris@10
|
374 T7r = T7o - T7q;
|
Chris@10
|
375 TeN = T7o + T7q;
|
Chris@10
|
376 TM = TF + TL;
|
Chris@10
|
377 T7m = TF - TL;
|
Chris@10
|
378 TcB = T7f + T7k;
|
Chris@10
|
379 T7l = T7f - T7k;
|
Chris@10
|
380 TiP = TeO + TeN;
|
Chris@10
|
381 TeP = TeN - TeO;
|
Chris@10
|
382 Tjl = Tz - TM;
|
Chris@10
|
383 TN = Tz + TM;
|
Chris@10
|
384 }
|
Chris@10
|
385 }
|
Chris@10
|
386 {
|
Chris@10
|
387 E T7D, TU, T13, T16, T7F, T10, T12, T15, T7w, T14;
|
Chris@10
|
388 {
|
Chris@10
|
389 E T19, T1c, T18, T1b;
|
Chris@10
|
390 {
|
Chris@10
|
391 E TQ, TT, TS, T7C, TR, TP;
|
Chris@10
|
392 TQ = cr[WS(rs, 4)];
|
Chris@10
|
393 TT = ci[WS(rs, 4)];
|
Chris@10
|
394 TP = W[6];
|
Chris@10
|
395 TcC = T7m - T7r;
|
Chris@10
|
396 T7s = T7m + T7r;
|
Chris@10
|
397 TS = W[7];
|
Chris@10
|
398 T7C = TP * TT;
|
Chris@10
|
399 TR = TP * TQ;
|
Chris@10
|
400 T19 = cr[WS(rs, 52)];
|
Chris@10
|
401 T1c = ci[WS(rs, 52)];
|
Chris@10
|
402 T7D = FNMS(TS, TQ, T7C);
|
Chris@10
|
403 TU = FMA(TS, TT, TR);
|
Chris@10
|
404 T18 = W[102];
|
Chris@10
|
405 T1b = W[103];
|
Chris@10
|
406 }
|
Chris@10
|
407 {
|
Chris@10
|
408 E TW, TZ, TY, T7E, TX, T7y, T1a, TV;
|
Chris@10
|
409 TW = cr[WS(rs, 36)];
|
Chris@10
|
410 TZ = ci[WS(rs, 36)];
|
Chris@10
|
411 T7y = T18 * T1c;
|
Chris@10
|
412 T1a = T18 * T19;
|
Chris@10
|
413 TV = W[70];
|
Chris@10
|
414 TY = W[71];
|
Chris@10
|
415 T7z = FNMS(T1b, T19, T7y);
|
Chris@10
|
416 T1d = FMA(T1b, T1c, T1a);
|
Chris@10
|
417 T7E = TV * TZ;
|
Chris@10
|
418 TX = TV * TW;
|
Chris@10
|
419 T13 = cr[WS(rs, 20)];
|
Chris@10
|
420 T16 = ci[WS(rs, 20)];
|
Chris@10
|
421 T7F = FNMS(TY, TW, T7E);
|
Chris@10
|
422 T10 = FMA(TY, TZ, TX);
|
Chris@10
|
423 T12 = W[38];
|
Chris@10
|
424 T15 = W[39];
|
Chris@10
|
425 }
|
Chris@10
|
426 }
|
Chris@10
|
427 T7G = T7D - T7F;
|
Chris@10
|
428 TeS = T7D + T7F;
|
Chris@10
|
429 T11 = TU + T10;
|
Chris@10
|
430 T7v = TU - T10;
|
Chris@10
|
431 T7w = T12 * T16;
|
Chris@10
|
432 T14 = T12 * T13;
|
Chris@10
|
433 T7x = FNMS(T15, T13, T7w);
|
Chris@10
|
434 T17 = FMA(T15, T16, T14);
|
Chris@10
|
435 }
|
Chris@10
|
436 {
|
Chris@10
|
437 E T8Y, T2H, T8E, T30, T2Q, T2T, T2S, T90, T2N, T8B, T2R;
|
Chris@10
|
438 {
|
Chris@10
|
439 E T2W, T2Z, T2Y, T8D, T2X;
|
Chris@10
|
440 {
|
Chris@10
|
441 E T2D, T2G, T2C, T2F, T8X, T2E, T2V;
|
Chris@10
|
442 T2D = cr[WS(rs, 62)];
|
Chris@10
|
443 T2G = ci[WS(rs, 62)];
|
Chris@10
|
444 {
|
Chris@10
|
445 E TeT, T7A, T1e, T7H;
|
Chris@10
|
446 TeT = T7x + T7z;
|
Chris@10
|
447 T7A = T7x - T7z;
|
Chris@10
|
448 T1e = T17 + T1d;
|
Chris@10
|
449 T7H = T17 - T1d;
|
Chris@10
|
450 T7B = T7v - T7A;
|
Chris@10
|
451 TcF = T7v + T7A;
|
Chris@10
|
452 TeU = TeS - TeT;
|
Chris@10
|
453 Ths = TeS + TeT;
|
Chris@10
|
454 T7I = T7G + T7H;
|
Chris@10
|
455 TcG = T7G - T7H;
|
Chris@10
|
456 T1f = T11 + T1e;
|
Chris@10
|
457 TeR = T11 - T1e;
|
Chris@10
|
458 T2C = W[122];
|
Chris@10
|
459 }
|
Chris@10
|
460 T2F = W[123];
|
Chris@10
|
461 T2W = cr[WS(rs, 46)];
|
Chris@10
|
462 T2Z = ci[WS(rs, 46)];
|
Chris@10
|
463 T8X = T2C * T2G;
|
Chris@10
|
464 T2E = T2C * T2D;
|
Chris@10
|
465 T2V = W[90];
|
Chris@10
|
466 T2Y = W[91];
|
Chris@10
|
467 T8Y = FNMS(T2F, T2D, T8X);
|
Chris@10
|
468 T2H = FMA(T2F, T2G, T2E);
|
Chris@10
|
469 T8D = T2V * T2Z;
|
Chris@10
|
470 T2X = T2V * T2W;
|
Chris@10
|
471 }
|
Chris@10
|
472 {
|
Chris@10
|
473 E T2J, T2M, T2I, T2L, T8Z, T2K, T2P;
|
Chris@10
|
474 T2J = cr[WS(rs, 30)];
|
Chris@10
|
475 T2M = ci[WS(rs, 30)];
|
Chris@10
|
476 T8E = FNMS(T2Y, T2W, T8D);
|
Chris@10
|
477 T30 = FMA(T2Y, T2Z, T2X);
|
Chris@10
|
478 T2I = W[58];
|
Chris@10
|
479 T2L = W[59];
|
Chris@10
|
480 T2Q = cr[WS(rs, 14)];
|
Chris@10
|
481 T2T = ci[WS(rs, 14)];
|
Chris@10
|
482 T8Z = T2I * T2M;
|
Chris@10
|
483 T2K = T2I * T2J;
|
Chris@10
|
484 T2P = W[26];
|
Chris@10
|
485 T2S = W[27];
|
Chris@10
|
486 T90 = FNMS(T2L, T2J, T8Z);
|
Chris@10
|
487 T2N = FMA(T2L, T2M, T2K);
|
Chris@10
|
488 T8B = T2P * T2T;
|
Chris@10
|
489 T2R = T2P * T2Q;
|
Chris@10
|
490 }
|
Chris@10
|
491 }
|
Chris@10
|
492 {
|
Chris@10
|
493 E T91, Tfe, T2O, T8A, T8C, T2U;
|
Chris@10
|
494 T91 = T8Y - T90;
|
Chris@10
|
495 Tfe = T8Y + T90;
|
Chris@10
|
496 T2O = T2H + T2N;
|
Chris@10
|
497 T8A = T2H - T2N;
|
Chris@10
|
498 T8C = FNMS(T2S, T2Q, T8B);
|
Chris@10
|
499 T2U = FMA(T2S, T2T, T2R);
|
Chris@10
|
500 {
|
Chris@10
|
501 E Tff, T8F, T92, T31;
|
Chris@10
|
502 Tff = T8C + T8E;
|
Chris@10
|
503 T8F = T8C - T8E;
|
Chris@10
|
504 T92 = T2U - T30;
|
Chris@10
|
505 T31 = T2U + T30;
|
Chris@10
|
506 T8G = T8A - T8F;
|
Chris@10
|
507 TcU = T8A + T8F;
|
Chris@10
|
508 Tfg = Tfe - Tff;
|
Chris@10
|
509 ThB = Tfe + Tff;
|
Chris@10
|
510 T32 = T2O + T31;
|
Chris@10
|
511 Tfj = T2O - T31;
|
Chris@10
|
512 TcX = T91 - T92;
|
Chris@10
|
513 T93 = T91 + T92;
|
Chris@10
|
514 }
|
Chris@10
|
515 }
|
Chris@10
|
516 }
|
Chris@10
|
517 {
|
Chris@10
|
518 E Ta5, T3C, T9f, T3V, T3L, T3O, T3N, Ta7, T3I, T9c, T3M;
|
Chris@10
|
519 {
|
Chris@10
|
520 E T3R, T3U, T3T, T9e, T3S;
|
Chris@10
|
521 {
|
Chris@10
|
522 E T3y, T3B, T3x, T3A, Ta4, T3z, T3Q;
|
Chris@10
|
523 T3y = cr[WS(rs, 1)];
|
Chris@10
|
524 T3B = ci[WS(rs, 1)];
|
Chris@10
|
525 T3x = W[0];
|
Chris@10
|
526 T3A = W[1];
|
Chris@10
|
527 T3R = cr[WS(rs, 49)];
|
Chris@10
|
528 T3U = ci[WS(rs, 49)];
|
Chris@10
|
529 Ta4 = T3x * T3B;
|
Chris@10
|
530 T3z = T3x * T3y;
|
Chris@10
|
531 T3Q = W[96];
|
Chris@10
|
532 T3T = W[97];
|
Chris@10
|
533 Ta5 = FNMS(T3A, T3y, Ta4);
|
Chris@10
|
534 T3C = FMA(T3A, T3B, T3z);
|
Chris@10
|
535 T9e = T3Q * T3U;
|
Chris@10
|
536 T3S = T3Q * T3R;
|
Chris@10
|
537 }
|
Chris@10
|
538 {
|
Chris@10
|
539 E T3E, T3H, T3D, T3G, Ta6, T3F, T3K;
|
Chris@10
|
540 T3E = cr[WS(rs, 33)];
|
Chris@10
|
541 T3H = ci[WS(rs, 33)];
|
Chris@10
|
542 T9f = FNMS(T3T, T3R, T9e);
|
Chris@10
|
543 T3V = FMA(T3T, T3U, T3S);
|
Chris@10
|
544 T3D = W[64];
|
Chris@10
|
545 T3G = W[65];
|
Chris@10
|
546 T3L = cr[WS(rs, 17)];
|
Chris@10
|
547 T3O = ci[WS(rs, 17)];
|
Chris@10
|
548 Ta6 = T3D * T3H;
|
Chris@10
|
549 T3F = T3D * T3E;
|
Chris@10
|
550 T3K = W[32];
|
Chris@10
|
551 T3N = W[33];
|
Chris@10
|
552 Ta7 = FNMS(T3G, T3E, Ta6);
|
Chris@10
|
553 T3I = FMA(T3G, T3H, T3F);
|
Chris@10
|
554 T9c = T3K * T3O;
|
Chris@10
|
555 T3M = T3K * T3L;
|
Chris@10
|
556 }
|
Chris@10
|
557 }
|
Chris@10
|
558 {
|
Chris@10
|
559 E Ta8, TfI, T3J, T9b, T9d, T3P;
|
Chris@10
|
560 Ta8 = Ta5 - Ta7;
|
Chris@10
|
561 TfI = Ta5 + Ta7;
|
Chris@10
|
562 T3J = T3C + T3I;
|
Chris@10
|
563 T9b = T3C - T3I;
|
Chris@10
|
564 T9d = FNMS(T3N, T3L, T9c);
|
Chris@10
|
565 T3P = FMA(T3N, T3O, T3M);
|
Chris@10
|
566 {
|
Chris@10
|
567 E TfJ, T9g, Ta9, T3W;
|
Chris@10
|
568 TfJ = T9d + T9f;
|
Chris@10
|
569 T9g = T9d - T9f;
|
Chris@10
|
570 Ta9 = T3P - T3V;
|
Chris@10
|
571 T3W = T3P + T3V;
|
Chris@10
|
572 T9h = T9b - T9g;
|
Chris@10
|
573 Td3 = T9b + T9g;
|
Chris@10
|
574 TfK = TfI - TfJ;
|
Chris@10
|
575 ThM = TfI + TfJ;
|
Chris@10
|
576 T3X = T3J + T3W;
|
Chris@10
|
577 Tfr = T3J - T3W;
|
Chris@10
|
578 Tde = Ta8 - Ta9;
|
Chris@10
|
579 Taa = Ta8 + Ta9;
|
Chris@10
|
580 }
|
Chris@10
|
581 }
|
Chris@10
|
582 }
|
Chris@10
|
583 }
|
Chris@10
|
584 {
|
Chris@10
|
585 E TaC, T69, Taw, TfU, T5X, Tar, TaA, T63;
|
Chris@10
|
586 {
|
Chris@10
|
587 E T8S, T3r, T8M, Tfl, T3f, T8H, T8Q, T3l;
|
Chris@10
|
588 {
|
Chris@10
|
589 E T8k, T8f, T8v, T8e;
|
Chris@10
|
590 {
|
Chris@10
|
591 E T8a, T2f, T8j, T2y, T2o, T2r, T2q, T8c, T2l, T8g, T2p;
|
Chris@10
|
592 {
|
Chris@10
|
593 E T2u, T2x, T2w, T8i, T2v;
|
Chris@10
|
594 {
|
Chris@10
|
595 E T2b, T2e, T2a, T2d, T89, T2c, T2t;
|
Chris@10
|
596 T2b = cr[WS(rs, 10)];
|
Chris@10
|
597 T2e = ci[WS(rs, 10)];
|
Chris@10
|
598 T2a = W[18];
|
Chris@10
|
599 T2d = W[19];
|
Chris@10
|
600 T2u = cr[WS(rs, 26)];
|
Chris@10
|
601 T2x = ci[WS(rs, 26)];
|
Chris@10
|
602 T89 = T2a * T2e;
|
Chris@10
|
603 T2c = T2a * T2b;
|
Chris@10
|
604 T2t = W[50];
|
Chris@10
|
605 T2w = W[51];
|
Chris@10
|
606 T8a = FNMS(T2d, T2b, T89);
|
Chris@10
|
607 T2f = FMA(T2d, T2e, T2c);
|
Chris@10
|
608 T8i = T2t * T2x;
|
Chris@10
|
609 T2v = T2t * T2u;
|
Chris@10
|
610 }
|
Chris@10
|
611 {
|
Chris@10
|
612 E T2h, T2k, T2g, T2j, T8b, T2i, T2n;
|
Chris@10
|
613 T2h = cr[WS(rs, 42)];
|
Chris@10
|
614 T2k = ci[WS(rs, 42)];
|
Chris@10
|
615 T8j = FNMS(T2w, T2u, T8i);
|
Chris@10
|
616 T2y = FMA(T2w, T2x, T2v);
|
Chris@10
|
617 T2g = W[82];
|
Chris@10
|
618 T2j = W[83];
|
Chris@10
|
619 T2o = cr[WS(rs, 58)];
|
Chris@10
|
620 T2r = ci[WS(rs, 58)];
|
Chris@10
|
621 T8b = T2g * T2k;
|
Chris@10
|
622 T2i = T2g * T2h;
|
Chris@10
|
623 T2n = W[114];
|
Chris@10
|
624 T2q = W[115];
|
Chris@10
|
625 T8c = FNMS(T2j, T2h, T8b);
|
Chris@10
|
626 T2l = FMA(T2j, T2k, T2i);
|
Chris@10
|
627 T8g = T2n * T2r;
|
Chris@10
|
628 T2p = T2n * T2o;
|
Chris@10
|
629 }
|
Chris@10
|
630 }
|
Chris@10
|
631 {
|
Chris@10
|
632 E T8d, Tfa, T2m, T88, T8h, T2s, Tf9, T2z;
|
Chris@10
|
633 T8d = T8a - T8c;
|
Chris@10
|
634 Tfa = T8a + T8c;
|
Chris@10
|
635 T2m = T2f + T2l;
|
Chris@10
|
636 T88 = T2f - T2l;
|
Chris@10
|
637 T8h = FNMS(T2q, T2o, T8g);
|
Chris@10
|
638 T2s = FMA(T2q, T2r, T2p);
|
Chris@10
|
639 T8k = T8h - T8j;
|
Chris@10
|
640 Tf9 = T8h + T8j;
|
Chris@10
|
641 T2z = T2s + T2y;
|
Chris@10
|
642 T8f = T2s - T2y;
|
Chris@10
|
643 T8v = T88 + T8d;
|
Chris@10
|
644 T8e = T88 - T8d;
|
Chris@10
|
645 Thx = Tfa + Tf9;
|
Chris@10
|
646 Tfb = Tf9 - Tfa;
|
Chris@10
|
647 Tf6 = T2m - T2z;
|
Chris@10
|
648 T2A = T2m + T2z;
|
Chris@10
|
649 }
|
Chris@10
|
650 }
|
Chris@10
|
651 {
|
Chris@10
|
652 E T38, T8J, T3h, T3k, T8L, T3e, T3g, T3j, T8P, T3i;
|
Chris@10
|
653 {
|
Chris@10
|
654 E T3n, T3q, T3m, T3p;
|
Chris@10
|
655 {
|
Chris@10
|
656 E T34, T37, T33, T8w, T8l, T36, T8I, T35;
|
Chris@10
|
657 T34 = cr[WS(rs, 6)];
|
Chris@10
|
658 T37 = ci[WS(rs, 6)];
|
Chris@10
|
659 T33 = W[10];
|
Chris@10
|
660 T8w = T8k - T8f;
|
Chris@10
|
661 T8l = T8f + T8k;
|
Chris@10
|
662 T36 = W[11];
|
Chris@10
|
663 T8I = T33 * T37;
|
Chris@10
|
664 T35 = T33 * T34;
|
Chris@10
|
665 T8x = T8v + T8w;
|
Chris@10
|
666 TcO = T8v - T8w;
|
Chris@10
|
667 T8m = T8e + T8l;
|
Chris@10
|
668 TcR = T8l - T8e;
|
Chris@10
|
669 T38 = FMA(T36, T37, T35);
|
Chris@10
|
670 T8J = FNMS(T36, T34, T8I);
|
Chris@10
|
671 }
|
Chris@10
|
672 T3n = cr[WS(rs, 22)];
|
Chris@10
|
673 T3q = ci[WS(rs, 22)];
|
Chris@10
|
674 T3m = W[42];
|
Chris@10
|
675 T3p = W[43];
|
Chris@10
|
676 {
|
Chris@10
|
677 E T3a, T3d, T3c, T8K, T3b, T8R, T3o, T39;
|
Chris@10
|
678 T3a = cr[WS(rs, 38)];
|
Chris@10
|
679 T3d = ci[WS(rs, 38)];
|
Chris@10
|
680 T8R = T3m * T3q;
|
Chris@10
|
681 T3o = T3m * T3n;
|
Chris@10
|
682 T39 = W[74];
|
Chris@10
|
683 T3c = W[75];
|
Chris@10
|
684 T8S = FNMS(T3p, T3n, T8R);
|
Chris@10
|
685 T3r = FMA(T3p, T3q, T3o);
|
Chris@10
|
686 T8K = T39 * T3d;
|
Chris@10
|
687 T3b = T39 * T3a;
|
Chris@10
|
688 T3h = cr[WS(rs, 54)];
|
Chris@10
|
689 T3k = ci[WS(rs, 54)];
|
Chris@10
|
690 T8L = FNMS(T3c, T3a, T8K);
|
Chris@10
|
691 T3e = FMA(T3c, T3d, T3b);
|
Chris@10
|
692 T3g = W[106];
|
Chris@10
|
693 T3j = W[107];
|
Chris@10
|
694 }
|
Chris@10
|
695 }
|
Chris@10
|
696 T8M = T8J - T8L;
|
Chris@10
|
697 Tfl = T8J + T8L;
|
Chris@10
|
698 T3f = T38 + T3e;
|
Chris@10
|
699 T8H = T38 - T3e;
|
Chris@10
|
700 T8P = T3g * T3k;
|
Chris@10
|
701 T3i = T3g * T3h;
|
Chris@10
|
702 T8Q = FNMS(T3j, T3h, T8P);
|
Chris@10
|
703 T3l = FMA(T3j, T3k, T3i);
|
Chris@10
|
704 }
|
Chris@10
|
705 }
|
Chris@10
|
706 {
|
Chris@10
|
707 E T9u, T9p, Tab, T9o;
|
Chris@10
|
708 {
|
Chris@10
|
709 E T9k, T43, T9t, T4m, T4c, T4f, T4e, T9m, T49, T9q, T4d;
|
Chris@10
|
710 {
|
Chris@10
|
711 E T4i, T4l, T4k, T9s, T4j;
|
Chris@10
|
712 {
|
Chris@10
|
713 E T3Z, T42, T3Y, T41, T9j, T40, T4h;
|
Chris@10
|
714 {
|
Chris@10
|
715 E T94, T8N, T8T, Tfk, T8O, T3s, T8U, T95;
|
Chris@10
|
716 T3Z = cr[WS(rs, 9)];
|
Chris@10
|
717 T94 = T8H + T8M;
|
Chris@10
|
718 T8N = T8H - T8M;
|
Chris@10
|
719 T8T = T8Q - T8S;
|
Chris@10
|
720 Tfk = T8Q + T8S;
|
Chris@10
|
721 T8O = T3l - T3r;
|
Chris@10
|
722 T3s = T3l + T3r;
|
Chris@10
|
723 T42 = ci[WS(rs, 9)];
|
Chris@10
|
724 Tfm = Tfk - Tfl;
|
Chris@10
|
725 ThC = Tfl + Tfk;
|
Chris@10
|
726 T8U = T8O + T8T;
|
Chris@10
|
727 T95 = T8T - T8O;
|
Chris@10
|
728 T3t = T3f + T3s;
|
Chris@10
|
729 Tfh = T3f - T3s;
|
Chris@10
|
730 T96 = T94 + T95;
|
Chris@10
|
731 TcV = T94 - T95;
|
Chris@10
|
732 T8V = T8N + T8U;
|
Chris@10
|
733 TcY = T8U - T8N;
|
Chris@10
|
734 T3Y = W[16];
|
Chris@10
|
735 }
|
Chris@10
|
736 T41 = W[17];
|
Chris@10
|
737 T4i = cr[WS(rs, 25)];
|
Chris@10
|
738 T4l = ci[WS(rs, 25)];
|
Chris@10
|
739 T9j = T3Y * T42;
|
Chris@10
|
740 T40 = T3Y * T3Z;
|
Chris@10
|
741 T4h = W[48];
|
Chris@10
|
742 T4k = W[49];
|
Chris@10
|
743 T9k = FNMS(T41, T3Z, T9j);
|
Chris@10
|
744 T43 = FMA(T41, T42, T40);
|
Chris@10
|
745 T9s = T4h * T4l;
|
Chris@10
|
746 T4j = T4h * T4i;
|
Chris@10
|
747 }
|
Chris@10
|
748 {
|
Chris@10
|
749 E T45, T48, T44, T47, T9l, T46, T4b;
|
Chris@10
|
750 T45 = cr[WS(rs, 41)];
|
Chris@10
|
751 T48 = ci[WS(rs, 41)];
|
Chris@10
|
752 T9t = FNMS(T4k, T4i, T9s);
|
Chris@10
|
753 T4m = FMA(T4k, T4l, T4j);
|
Chris@10
|
754 T44 = W[80];
|
Chris@10
|
755 T47 = W[81];
|
Chris@10
|
756 T4c = cr[WS(rs, 57)];
|
Chris@10
|
757 T4f = ci[WS(rs, 57)];
|
Chris@10
|
758 T9l = T44 * T48;
|
Chris@10
|
759 T46 = T44 * T45;
|
Chris@10
|
760 T4b = W[112];
|
Chris@10
|
761 T4e = W[113];
|
Chris@10
|
762 T9m = FNMS(T47, T45, T9l);
|
Chris@10
|
763 T49 = FMA(T47, T48, T46);
|
Chris@10
|
764 T9q = T4b * T4f;
|
Chris@10
|
765 T4d = T4b * T4c;
|
Chris@10
|
766 }
|
Chris@10
|
767 }
|
Chris@10
|
768 {
|
Chris@10
|
769 E T9n, Tft, T4a, T9i, T9r, T4g, Tfs, T4n;
|
Chris@10
|
770 T9n = T9k - T9m;
|
Chris@10
|
771 Tft = T9k + T9m;
|
Chris@10
|
772 T4a = T43 + T49;
|
Chris@10
|
773 T9i = T43 - T49;
|
Chris@10
|
774 T9r = FNMS(T4e, T4c, T9q);
|
Chris@10
|
775 T4g = FMA(T4e, T4f, T4d);
|
Chris@10
|
776 T9u = T9r - T9t;
|
Chris@10
|
777 Tfs = T9r + T9t;
|
Chris@10
|
778 T4n = T4g + T4m;
|
Chris@10
|
779 T9p = T4g - T4m;
|
Chris@10
|
780 Tab = T9i + T9n;
|
Chris@10
|
781 T9o = T9i - T9n;
|
Chris@10
|
782 ThN = Tft + Tfs;
|
Chris@10
|
783 Tfu = Tfs - Tft;
|
Chris@10
|
784 TfL = T4a - T4n;
|
Chris@10
|
785 T4o = T4a + T4n;
|
Chris@10
|
786 }
|
Chris@10
|
787 }
|
Chris@10
|
788 {
|
Chris@10
|
789 E T5Q, Tat, T5Z, T62, Tav, T5W, T5Y, T61, Taz, T60;
|
Chris@10
|
790 {
|
Chris@10
|
791 E T65, T68, T64, T67;
|
Chris@10
|
792 {
|
Chris@10
|
793 E T5M, T5P, T5L, Tac, T9v, T5O, Tas, T5N;
|
Chris@10
|
794 T5M = cr[WS(rs, 7)];
|
Chris@10
|
795 T5P = ci[WS(rs, 7)];
|
Chris@10
|
796 T5L = W[12];
|
Chris@10
|
797 Tac = T9u - T9p;
|
Chris@10
|
798 T9v = T9p + T9u;
|
Chris@10
|
799 T5O = W[13];
|
Chris@10
|
800 Tas = T5L * T5P;
|
Chris@10
|
801 T5N = T5L * T5M;
|
Chris@10
|
802 Tad = Tab + Tac;
|
Chris@10
|
803 Td4 = Tab - Tac;
|
Chris@10
|
804 T9w = T9o + T9v;
|
Chris@10
|
805 Tdf = T9v - T9o;
|
Chris@10
|
806 T5Q = FMA(T5O, T5P, T5N);
|
Chris@10
|
807 Tat = FNMS(T5O, T5M, Tas);
|
Chris@10
|
808 }
|
Chris@10
|
809 T65 = cr[WS(rs, 23)];
|
Chris@10
|
810 T68 = ci[WS(rs, 23)];
|
Chris@10
|
811 T64 = W[44];
|
Chris@10
|
812 T67 = W[45];
|
Chris@10
|
813 {
|
Chris@10
|
814 E T5S, T5V, T5U, Tau, T5T, TaB, T66, T5R;
|
Chris@10
|
815 T5S = cr[WS(rs, 39)];
|
Chris@10
|
816 T5V = ci[WS(rs, 39)];
|
Chris@10
|
817 TaB = T64 * T68;
|
Chris@10
|
818 T66 = T64 * T65;
|
Chris@10
|
819 T5R = W[76];
|
Chris@10
|
820 T5U = W[77];
|
Chris@10
|
821 TaC = FNMS(T67, T65, TaB);
|
Chris@10
|
822 T69 = FMA(T67, T68, T66);
|
Chris@10
|
823 Tau = T5R * T5V;
|
Chris@10
|
824 T5T = T5R * T5S;
|
Chris@10
|
825 T5Z = cr[WS(rs, 55)];
|
Chris@10
|
826 T62 = ci[WS(rs, 55)];
|
Chris@10
|
827 Tav = FNMS(T5U, T5S, Tau);
|
Chris@10
|
828 T5W = FMA(T5U, T5V, T5T);
|
Chris@10
|
829 T5Y = W[108];
|
Chris@10
|
830 T61 = W[109];
|
Chris@10
|
831 }
|
Chris@10
|
832 }
|
Chris@10
|
833 Taw = Tat - Tav;
|
Chris@10
|
834 TfU = Tat + Tav;
|
Chris@10
|
835 T5X = T5Q + T5W;
|
Chris@10
|
836 Tar = T5Q - T5W;
|
Chris@10
|
837 Taz = T5Y * T62;
|
Chris@10
|
838 T60 = T5Y * T5Z;
|
Chris@10
|
839 TaA = FNMS(T61, T5Z, Taz);
|
Chris@10
|
840 T63 = FMA(T61, T62, T60);
|
Chris@10
|
841 }
|
Chris@10
|
842 }
|
Chris@10
|
843 }
|
Chris@10
|
844 {
|
Chris@10
|
845 E T9T, Td9, TfE, TfB, Tda, Ta0;
|
Chris@10
|
846 {
|
Chris@10
|
847 E T9E, Td6, Tfz, Tfw, Td7, T9L;
|
Chris@10
|
848 {
|
Chris@10
|
849 E T9G, T4v, T9C, T4O, T4E, T4H, T4G, T9I, T4B, T9z, T4F;
|
Chris@10
|
850 {
|
Chris@10
|
851 E T4K, T4N, T4M, T9B, T4L;
|
Chris@10
|
852 {
|
Chris@10
|
853 E T4r, T4u, T4q, T4t, T9F, T4s, T4J;
|
Chris@10
|
854 {
|
Chris@10
|
855 E Tbl, Tax, TaD, TfT, Tay, T6a, TaE, Tbk;
|
Chris@10
|
856 T4r = cr[WS(rs, 5)];
|
Chris@10
|
857 Tbl = Tar + Taw;
|
Chris@10
|
858 Tax = Tar - Taw;
|
Chris@10
|
859 TaD = TaA - TaC;
|
Chris@10
|
860 TfT = TaA + TaC;
|
Chris@10
|
861 Tay = T63 - T69;
|
Chris@10
|
862 T6a = T63 + T69;
|
Chris@10
|
863 T4u = ci[WS(rs, 5)];
|
Chris@10
|
864 TfV = TfT - TfU;
|
Chris@10
|
865 ThY = TfU + TfT;
|
Chris@10
|
866 TaE = Tay + TaD;
|
Chris@10
|
867 Tbk = Tay - TaD;
|
Chris@10
|
868 T6b = T5X + T6a;
|
Chris@10
|
869 Tg9 = T6a - T5X;
|
Chris@10
|
870 Tbm = Tbk - Tbl;
|
Chris@10
|
871 Tdn = Tbl + Tbk;
|
Chris@10
|
872 TaF = Tax + TaE;
|
Chris@10
|
873 Tdy = TaE - Tax;
|
Chris@10
|
874 T4q = W[8];
|
Chris@10
|
875 }
|
Chris@10
|
876 T4t = W[9];
|
Chris@10
|
877 T4K = cr[WS(rs, 53)];
|
Chris@10
|
878 T4N = ci[WS(rs, 53)];
|
Chris@10
|
879 T9F = T4q * T4u;
|
Chris@10
|
880 T4s = T4q * T4r;
|
Chris@10
|
881 T4J = W[104];
|
Chris@10
|
882 T4M = W[105];
|
Chris@10
|
883 T9G = FNMS(T4t, T4r, T9F);
|
Chris@10
|
884 T4v = FMA(T4t, T4u, T4s);
|
Chris@10
|
885 T9B = T4J * T4N;
|
Chris@10
|
886 T4L = T4J * T4K;
|
Chris@10
|
887 }
|
Chris@10
|
888 {
|
Chris@10
|
889 E T4x, T4A, T4w, T4z, T9H, T4y, T4D;
|
Chris@10
|
890 T4x = cr[WS(rs, 37)];
|
Chris@10
|
891 T4A = ci[WS(rs, 37)];
|
Chris@10
|
892 T9C = FNMS(T4M, T4K, T9B);
|
Chris@10
|
893 T4O = FMA(T4M, T4N, T4L);
|
Chris@10
|
894 T4w = W[72];
|
Chris@10
|
895 T4z = W[73];
|
Chris@10
|
896 T4E = cr[WS(rs, 21)];
|
Chris@10
|
897 T4H = ci[WS(rs, 21)];
|
Chris@10
|
898 T9H = T4w * T4A;
|
Chris@10
|
899 T4y = T4w * T4x;
|
Chris@10
|
900 T4D = W[40];
|
Chris@10
|
901 T4G = W[41];
|
Chris@10
|
902 T9I = FNMS(T4z, T4x, T9H);
|
Chris@10
|
903 T4B = FMA(T4z, T4A, T4y);
|
Chris@10
|
904 T9z = T4D * T4H;
|
Chris@10
|
905 T4F = T4D * T4E;
|
Chris@10
|
906 }
|
Chris@10
|
907 }
|
Chris@10
|
908 {
|
Chris@10
|
909 E T9J, Tfx, T4C, T9y, T9A, T4I;
|
Chris@10
|
910 T9J = T9G - T9I;
|
Chris@10
|
911 Tfx = T9G + T9I;
|
Chris@10
|
912 T4C = T4v + T4B;
|
Chris@10
|
913 T9y = T4v - T4B;
|
Chris@10
|
914 T9A = FNMS(T4G, T4E, T9z);
|
Chris@10
|
915 T4I = FMA(T4G, T4H, T4F);
|
Chris@10
|
916 {
|
Chris@10
|
917 E Tfy, T9D, T9K, T4P;
|
Chris@10
|
918 Tfy = T9A + T9C;
|
Chris@10
|
919 T9D = T9A - T9C;
|
Chris@10
|
920 T9K = T4I - T4O;
|
Chris@10
|
921 T4P = T4I + T4O;
|
Chris@10
|
922 T9E = T9y - T9D;
|
Chris@10
|
923 Td6 = T9y + T9D;
|
Chris@10
|
924 Tfz = Tfx - Tfy;
|
Chris@10
|
925 ThJ = Tfx + Tfy;
|
Chris@10
|
926 Tfw = T4C - T4P;
|
Chris@10
|
927 T4Q = T4C + T4P;
|
Chris@10
|
928 Td7 = T9J - T9K;
|
Chris@10
|
929 T9L = T9J + T9K;
|
Chris@10
|
930 }
|
Chris@10
|
931 }
|
Chris@10
|
932 }
|
Chris@10
|
933 {
|
Chris@10
|
934 E T9V, T4W, T9R, T5f, T55, T58, T57, T9X, T52, T9O, T56;
|
Chris@10
|
935 {
|
Chris@10
|
936 E T5b, T5e, T5d, T9Q, T5c;
|
Chris@10
|
937 {
|
Chris@10
|
938 E T4S, T4V, T4R, T4U, T9U, T4T, T5a;
|
Chris@10
|
939 T4S = cr[WS(rs, 61)];
|
Chris@10
|
940 TfN = Tfw + Tfz;
|
Chris@10
|
941 TfA = Tfw - Tfz;
|
Chris@10
|
942 Taf = FMA(KP414213562, T9E, T9L);
|
Chris@10
|
943 T9M = FNMS(KP414213562, T9L, T9E);
|
Chris@10
|
944 Td8 = FMA(KP414213562, Td7, Td6);
|
Chris@10
|
945 Tdh = FNMS(KP414213562, Td6, Td7);
|
Chris@10
|
946 T4V = ci[WS(rs, 61)];
|
Chris@10
|
947 T4R = W[120];
|
Chris@10
|
948 T4U = W[121];
|
Chris@10
|
949 T5b = cr[WS(rs, 45)];
|
Chris@10
|
950 T5e = ci[WS(rs, 45)];
|
Chris@10
|
951 T9U = T4R * T4V;
|
Chris@10
|
952 T4T = T4R * T4S;
|
Chris@10
|
953 T5a = W[88];
|
Chris@10
|
954 T5d = W[89];
|
Chris@10
|
955 T9V = FNMS(T4U, T4S, T9U);
|
Chris@10
|
956 T4W = FMA(T4U, T4V, T4T);
|
Chris@10
|
957 T9Q = T5a * T5e;
|
Chris@10
|
958 T5c = T5a * T5b;
|
Chris@10
|
959 }
|
Chris@10
|
960 {
|
Chris@10
|
961 E T4Y, T51, T4X, T50, T9W, T4Z, T54;
|
Chris@10
|
962 T4Y = cr[WS(rs, 29)];
|
Chris@10
|
963 T51 = ci[WS(rs, 29)];
|
Chris@10
|
964 T9R = FNMS(T5d, T5b, T9Q);
|
Chris@10
|
965 T5f = FMA(T5d, T5e, T5c);
|
Chris@10
|
966 T4X = W[56];
|
Chris@10
|
967 T50 = W[57];
|
Chris@10
|
968 T55 = cr[WS(rs, 13)];
|
Chris@10
|
969 T58 = ci[WS(rs, 13)];
|
Chris@10
|
970 T9W = T4X * T51;
|
Chris@10
|
971 T4Z = T4X * T4Y;
|
Chris@10
|
972 T54 = W[24];
|
Chris@10
|
973 T57 = W[25];
|
Chris@10
|
974 T9X = FNMS(T50, T4Y, T9W);
|
Chris@10
|
975 T52 = FMA(T50, T51, T4Z);
|
Chris@10
|
976 T9O = T54 * T58;
|
Chris@10
|
977 T56 = T54 * T55;
|
Chris@10
|
978 }
|
Chris@10
|
979 }
|
Chris@10
|
980 {
|
Chris@10
|
981 E T9Y, TfC, T53, T9N, T9P, T59;
|
Chris@10
|
982 T9Y = T9V - T9X;
|
Chris@10
|
983 TfC = T9V + T9X;
|
Chris@10
|
984 T53 = T4W + T52;
|
Chris@10
|
985 T9N = T4W - T52;
|
Chris@10
|
986 T9P = FNMS(T57, T55, T9O);
|
Chris@10
|
987 T59 = FMA(T57, T58, T56);
|
Chris@10
|
988 {
|
Chris@10
|
989 E TfD, T9S, T9Z, T5g;
|
Chris@10
|
990 TfD = T9P + T9R;
|
Chris@10
|
991 T9S = T9P - T9R;
|
Chris@10
|
992 T9Z = T59 - T5f;
|
Chris@10
|
993 T5g = T59 + T5f;
|
Chris@10
|
994 T9T = T9N - T9S;
|
Chris@10
|
995 Td9 = T9N + T9S;
|
Chris@10
|
996 TfE = TfC - TfD;
|
Chris@10
|
997 ThI = TfC + TfD;
|
Chris@10
|
998 TfB = T53 - T5g;
|
Chris@10
|
999 T5h = T53 + T5g;
|
Chris@10
|
1000 Tda = T9Y - T9Z;
|
Chris@10
|
1001 Ta0 = T9Y + T9Z;
|
Chris@10
|
1002 }
|
Chris@10
|
1003 }
|
Chris@10
|
1004 }
|
Chris@10
|
1005 }
|
Chris@10
|
1006 {
|
Chris@10
|
1007 E TaN, Tdp, Tg0, TfX, Tdq, TaU;
|
Chris@10
|
1008 {
|
Chris@10
|
1009 E TaQ, T6i, TaL, T6B, T6r, T6u, T6t, TaS, T6o, TaI, T6s;
|
Chris@10
|
1010 {
|
Chris@10
|
1011 E T6x, T6A, T6z, TaK, T6y;
|
Chris@10
|
1012 {
|
Chris@10
|
1013 E T6e, T6h, T6d, T6g, TaP, T6f, T6w;
|
Chris@10
|
1014 T6e = cr[WS(rs, 3)];
|
Chris@10
|
1015 TfO = TfE - TfB;
|
Chris@10
|
1016 TfF = TfB + TfE;
|
Chris@10
|
1017 Tag = FNMS(KP414213562, T9T, Ta0);
|
Chris@10
|
1018 Ta1 = FMA(KP414213562, Ta0, T9T);
|
Chris@10
|
1019 Tdb = FNMS(KP414213562, Tda, Td9);
|
Chris@10
|
1020 Tdi = FMA(KP414213562, Td9, Tda);
|
Chris@10
|
1021 T6h = ci[WS(rs, 3)];
|
Chris@10
|
1022 T6d = W[4];
|
Chris@10
|
1023 T6g = W[5];
|
Chris@10
|
1024 T6x = cr[WS(rs, 51)];
|
Chris@10
|
1025 T6A = ci[WS(rs, 51)];
|
Chris@10
|
1026 TaP = T6d * T6h;
|
Chris@10
|
1027 T6f = T6d * T6e;
|
Chris@10
|
1028 T6w = W[100];
|
Chris@10
|
1029 T6z = W[101];
|
Chris@10
|
1030 TaQ = FNMS(T6g, T6e, TaP);
|
Chris@10
|
1031 T6i = FMA(T6g, T6h, T6f);
|
Chris@10
|
1032 TaK = T6w * T6A;
|
Chris@10
|
1033 T6y = T6w * T6x;
|
Chris@10
|
1034 }
|
Chris@10
|
1035 {
|
Chris@10
|
1036 E T6k, T6n, T6j, T6m, TaR, T6l, T6q;
|
Chris@10
|
1037 T6k = cr[WS(rs, 35)];
|
Chris@10
|
1038 T6n = ci[WS(rs, 35)];
|
Chris@10
|
1039 TaL = FNMS(T6z, T6x, TaK);
|
Chris@10
|
1040 T6B = FMA(T6z, T6A, T6y);
|
Chris@10
|
1041 T6j = W[68];
|
Chris@10
|
1042 T6m = W[69];
|
Chris@10
|
1043 T6r = cr[WS(rs, 19)];
|
Chris@10
|
1044 T6u = ci[WS(rs, 19)];
|
Chris@10
|
1045 TaR = T6j * T6n;
|
Chris@10
|
1046 T6l = T6j * T6k;
|
Chris@10
|
1047 T6q = W[36];
|
Chris@10
|
1048 T6t = W[37];
|
Chris@10
|
1049 TaS = FNMS(T6m, T6k, TaR);
|
Chris@10
|
1050 T6o = FMA(T6m, T6n, T6l);
|
Chris@10
|
1051 TaI = T6q * T6u;
|
Chris@10
|
1052 T6s = T6q * T6r;
|
Chris@10
|
1053 }
|
Chris@10
|
1054 }
|
Chris@10
|
1055 {
|
Chris@10
|
1056 E TaT, TfY, T6p, TaH, TaJ, T6v;
|
Chris@10
|
1057 TaT = TaQ - TaS;
|
Chris@10
|
1058 TfY = TaQ + TaS;
|
Chris@10
|
1059 T6p = T6i + T6o;
|
Chris@10
|
1060 TaH = T6i - T6o;
|
Chris@10
|
1061 TaJ = FNMS(T6t, T6r, TaI);
|
Chris@10
|
1062 T6v = FMA(T6t, T6u, T6s);
|
Chris@10
|
1063 {
|
Chris@10
|
1064 E TfZ, TaM, T6C, TaO;
|
Chris@10
|
1065 TfZ = TaJ + TaL;
|
Chris@10
|
1066 TaM = TaJ - TaL;
|
Chris@10
|
1067 T6C = T6v + T6B;
|
Chris@10
|
1068 TaO = T6B - T6v;
|
Chris@10
|
1069 TaN = TaH - TaM;
|
Chris@10
|
1070 Tdp = TaH + TaM;
|
Chris@10
|
1071 Tg0 = TfY - TfZ;
|
Chris@10
|
1072 ThU = TfY + TfZ;
|
Chris@10
|
1073 TfX = T6p - T6C;
|
Chris@10
|
1074 T6D = T6p + T6C;
|
Chris@10
|
1075 Tdq = TaT + TaO;
|
Chris@10
|
1076 TaU = TaO - TaT;
|
Chris@10
|
1077 }
|
Chris@10
|
1078 }
|
Chris@10
|
1079 }
|
Chris@10
|
1080 {
|
Chris@10
|
1081 E Tb5, T6J, Tb0, T72, T6S, T6V, T6U, Tb7, T6P, TaX, T6T;
|
Chris@10
|
1082 {
|
Chris@10
|
1083 E T6Y, T71, T70, TaZ, T6Z;
|
Chris@10
|
1084 {
|
Chris@10
|
1085 E T6F, T6I, T6E, T6H, Tb4, T6G, T6X;
|
Chris@10
|
1086 T6F = cr[WS(rs, 59)];
|
Chris@10
|
1087 Tgf = TfX + Tg0;
|
Chris@10
|
1088 Tg1 = TfX - Tg0;
|
Chris@10
|
1089 Tbo = FNMS(KP414213562, TaN, TaU);
|
Chris@10
|
1090 TaV = FMA(KP414213562, TaU, TaN);
|
Chris@10
|
1091 Tdr = FMA(KP414213562, Tdq, Tdp);
|
Chris@10
|
1092 TdA = FNMS(KP414213562, Tdp, Tdq);
|
Chris@10
|
1093 T6I = ci[WS(rs, 59)];
|
Chris@10
|
1094 T6E = W[116];
|
Chris@10
|
1095 T6H = W[117];
|
Chris@10
|
1096 T6Y = cr[WS(rs, 43)];
|
Chris@10
|
1097 T71 = ci[WS(rs, 43)];
|
Chris@10
|
1098 Tb4 = T6E * T6I;
|
Chris@10
|
1099 T6G = T6E * T6F;
|
Chris@10
|
1100 T6X = W[84];
|
Chris@10
|
1101 T70 = W[85];
|
Chris@10
|
1102 Tb5 = FNMS(T6H, T6F, Tb4);
|
Chris@10
|
1103 T6J = FMA(T6H, T6I, T6G);
|
Chris@10
|
1104 TaZ = T6X * T71;
|
Chris@10
|
1105 T6Z = T6X * T6Y;
|
Chris@10
|
1106 }
|
Chris@10
|
1107 {
|
Chris@10
|
1108 E T6L, T6O, T6K, T6N, Tb6, T6M, T6R;
|
Chris@10
|
1109 T6L = cr[WS(rs, 27)];
|
Chris@10
|
1110 T6O = ci[WS(rs, 27)];
|
Chris@10
|
1111 Tb0 = FNMS(T70, T6Y, TaZ);
|
Chris@10
|
1112 T72 = FMA(T70, T71, T6Z);
|
Chris@10
|
1113 T6K = W[52];
|
Chris@10
|
1114 T6N = W[53];
|
Chris@10
|
1115 T6S = cr[WS(rs, 11)];
|
Chris@10
|
1116 T6V = ci[WS(rs, 11)];
|
Chris@10
|
1117 Tb6 = T6K * T6O;
|
Chris@10
|
1118 T6M = T6K * T6L;
|
Chris@10
|
1119 T6R = W[20];
|
Chris@10
|
1120 T6U = W[21];
|
Chris@10
|
1121 Tb7 = FNMS(T6N, T6L, Tb6);
|
Chris@10
|
1122 T6P = FMA(T6N, T6O, T6M);
|
Chris@10
|
1123 TaX = T6R * T6V;
|
Chris@10
|
1124 T6T = T6R * T6S;
|
Chris@10
|
1125 }
|
Chris@10
|
1126 }
|
Chris@10
|
1127 {
|
Chris@10
|
1128 E Tb8, Tg3, T6Q, TaW, TaY, T6W;
|
Chris@10
|
1129 Tb8 = Tb5 - Tb7;
|
Chris@10
|
1130 Tg3 = Tb5 + Tb7;
|
Chris@10
|
1131 T6Q = T6J + T6P;
|
Chris@10
|
1132 TaW = T6J - T6P;
|
Chris@10
|
1133 TaY = FNMS(T6U, T6S, TaX);
|
Chris@10
|
1134 T6W = FMA(T6U, T6V, T6T);
|
Chris@10
|
1135 {
|
Chris@10
|
1136 E Tg4, Tb1, T73, Tb3;
|
Chris@10
|
1137 Tg4 = TaY + Tb0;
|
Chris@10
|
1138 Tb1 = TaY - Tb0;
|
Chris@10
|
1139 T73 = T6W + T72;
|
Chris@10
|
1140 Tb3 = T72 - T6W;
|
Chris@10
|
1141 Tb2 = TaW - Tb1;
|
Chris@10
|
1142 Tds = TaW + Tb1;
|
Chris@10
|
1143 Tg5 = Tg3 - Tg4;
|
Chris@10
|
1144 ThT = Tg3 + Tg4;
|
Chris@10
|
1145 Tg2 = T6Q - T73;
|
Chris@10
|
1146 T74 = T6Q + T73;
|
Chris@10
|
1147 Tdt = Tb8 + Tb3;
|
Chris@10
|
1148 Tb9 = Tb3 - Tb8;
|
Chris@10
|
1149 }
|
Chris@10
|
1150 }
|
Chris@10
|
1151 }
|
Chris@10
|
1152 }
|
Chris@10
|
1153 }
|
Chris@10
|
1154 }
|
Chris@10
|
1155 {
|
Chris@10
|
1156 E Thq, Tge, Tg6, Tdu, TdB, Tj7, Thv, ThA, Tht, Tj8, ThD, Thy, ThS, Ti0, ThZ;
|
Chris@10
|
1157 E ThV, ThH, ThP, ThO, ThK, Tkm, TcD, Tk0, Tk4, TjZ, Tk3, Tik, Tin;
|
Chris@10
|
1158 {
|
Chris@10
|
1159 E Tbp, Tba, TiI, TiL;
|
Chris@10
|
1160 {
|
Chris@10
|
1161 E Tio, T1I, Tj1, T3v, Tj2, TiX, TiN, Tir, T76, TiJ, TiC, TiG, T5j, Tit, Tiw;
|
Chris@10
|
1162 E TiK;
|
Chris@10
|
1163 {
|
Chris@10
|
1164 E TiO, TiW, Tip, Tiq;
|
Chris@10
|
1165 {
|
Chris@10
|
1166 E TO, T1H, T2B, T3u;
|
Chris@10
|
1167 Thq = Tm - TN;
|
Chris@10
|
1168 TO = Tm + TN;
|
Chris@10
|
1169 Tge = Tg2 - Tg5;
|
Chris@10
|
1170 Tg6 = Tg2 + Tg5;
|
Chris@10
|
1171 Tbp = FMA(KP414213562, Tb2, Tb9);
|
Chris@10
|
1172 Tba = FNMS(KP414213562, Tb9, Tb2);
|
Chris@10
|
1173 Tdu = FNMS(KP414213562, Tdt, Tds);
|
Chris@10
|
1174 TdB = FMA(KP414213562, Tds, Tdt);
|
Chris@10
|
1175 T1H = T1f + T1G;
|
Chris@10
|
1176 Tj7 = T1f - T1G;
|
Chris@10
|
1177 Thv = T29 - T2A;
|
Chris@10
|
1178 T2B = T29 + T2A;
|
Chris@10
|
1179 T3u = T32 + T3t;
|
Chris@10
|
1180 ThA = T32 - T3t;
|
Chris@10
|
1181 Tht = Thr - Ths;
|
Chris@10
|
1182 TiO = Ths + Thr;
|
Chris@10
|
1183 Tio = TO - T1H;
|
Chris@10
|
1184 T1I = TO + T1H;
|
Chris@10
|
1185 Tj1 = T2B - T3u;
|
Chris@10
|
1186 T3v = T2B + T3u;
|
Chris@10
|
1187 TiW = TiP + TiV;
|
Chris@10
|
1188 Tj8 = TiV - TiP;
|
Chris@10
|
1189 }
|
Chris@10
|
1190 ThD = ThB - ThC;
|
Chris@10
|
1191 Tip = ThB + ThC;
|
Chris@10
|
1192 Tiq = Thw + Thx;
|
Chris@10
|
1193 Thy = Thw - Thx;
|
Chris@10
|
1194 {
|
Chris@10
|
1195 E T6c, T75, Tiz, TiA;
|
Chris@10
|
1196 ThS = T5K - T6b;
|
Chris@10
|
1197 T6c = T5K + T6b;
|
Chris@10
|
1198 Tj2 = TiW - TiO;
|
Chris@10
|
1199 TiX = TiO + TiW;
|
Chris@10
|
1200 TiN = Tiq + Tip;
|
Chris@10
|
1201 Tir = Tip - Tiq;
|
Chris@10
|
1202 T75 = T6D + T74;
|
Chris@10
|
1203 Ti0 = T74 - T6D;
|
Chris@10
|
1204 ThZ = ThX - ThY;
|
Chris@10
|
1205 Tiz = ThX + ThY;
|
Chris@10
|
1206 TiA = ThU + ThT;
|
Chris@10
|
1207 ThV = ThT - ThU;
|
Chris@10
|
1208 {
|
Chris@10
|
1209 E T4p, Tiy, TiB, T5i, Tiu, Tiv;
|
Chris@10
|
1210 ThH = T3X - T4o;
|
Chris@10
|
1211 T4p = T3X + T4o;
|
Chris@10
|
1212 T76 = T6c + T75;
|
Chris@10
|
1213 Tiy = T6c - T75;
|
Chris@10
|
1214 TiJ = Tiz + TiA;
|
Chris@10
|
1215 TiB = Tiz - TiA;
|
Chris@10
|
1216 T5i = T4Q + T5h;
|
Chris@10
|
1217 ThP = T4Q - T5h;
|
Chris@10
|
1218 ThO = ThM - ThN;
|
Chris@10
|
1219 Tiu = ThM + ThN;
|
Chris@10
|
1220 Tiv = ThJ + ThI;
|
Chris@10
|
1221 ThK = ThI - ThJ;
|
Chris@10
|
1222 TiC = Tiy - TiB;
|
Chris@10
|
1223 TiG = Tiy + TiB;
|
Chris@10
|
1224 T5j = T4p + T5i;
|
Chris@10
|
1225 Tit = T4p - T5i;
|
Chris@10
|
1226 Tiw = Tiu - Tiv;
|
Chris@10
|
1227 TiK = Tiu + Tiv;
|
Chris@10
|
1228 }
|
Chris@10
|
1229 }
|
Chris@10
|
1230 }
|
Chris@10
|
1231 {
|
Chris@10
|
1232 E TiZ, TiD, TiH, TiE, Tis, TiM, TiY, Tj0;
|
Chris@10
|
1233 {
|
Chris@10
|
1234 E T3w, TiF, Tix, T77, Tj5, Tj3, Tj6, Tj4;
|
Chris@10
|
1235 TiI = T1I - T3v;
|
Chris@10
|
1236 T3w = T1I + T3v;
|
Chris@10
|
1237 TiF = Tit - Tiw;
|
Chris@10
|
1238 Tix = Tit + Tiw;
|
Chris@10
|
1239 T77 = T5j + T76;
|
Chris@10
|
1240 TiZ = T76 - T5j;
|
Chris@10
|
1241 Tj5 = Tj2 - Tj1;
|
Chris@10
|
1242 Tj3 = Tj1 + Tj2;
|
Chris@10
|
1243 TiD = Tix + TiC;
|
Chris@10
|
1244 Tj4 = TiC - Tix;
|
Chris@10
|
1245 cr[0] = T3w + T77;
|
Chris@10
|
1246 ci[WS(rs, 31)] = T3w - T77;
|
Chris@10
|
1247 Tj6 = TiG - TiF;
|
Chris@10
|
1248 TiH = TiF + TiG;
|
Chris@10
|
1249 ci[WS(rs, 39)] = FMA(KP707106781, Tj4, Tj3);
|
Chris@10
|
1250 cr[WS(rs, 56)] = FMS(KP707106781, Tj4, Tj3);
|
Chris@10
|
1251 TiE = Tio + Tir;
|
Chris@10
|
1252 Tis = Tio - Tir;
|
Chris@10
|
1253 ci[WS(rs, 55)] = FMA(KP707106781, Tj6, Tj5);
|
Chris@10
|
1254 cr[WS(rs, 40)] = FMS(KP707106781, Tj6, Tj5);
|
Chris@10
|
1255 }
|
Chris@10
|
1256 TiL = TiJ - TiK;
|
Chris@10
|
1257 TiM = TiK + TiJ;
|
Chris@10
|
1258 cr[WS(rs, 8)] = FMA(KP707106781, TiD, Tis);
|
Chris@10
|
1259 ci[WS(rs, 23)] = FNMS(KP707106781, TiD, Tis);
|
Chris@10
|
1260 ci[WS(rs, 7)] = FMA(KP707106781, TiH, TiE);
|
Chris@10
|
1261 cr[WS(rs, 24)] = FNMS(KP707106781, TiH, TiE);
|
Chris@10
|
1262 TiY = TiN + TiX;
|
Chris@10
|
1263 Tj0 = TiX - TiN;
|
Chris@10
|
1264 ci[WS(rs, 63)] = TiM + TiY;
|
Chris@10
|
1265 cr[WS(rs, 32)] = TiM - TiY;
|
Chris@10
|
1266 ci[WS(rs, 47)] = TiZ + Tj0;
|
Chris@10
|
1267 cr[WS(rs, 48)] = TiZ - Tj0;
|
Chris@10
|
1268 }
|
Chris@10
|
1269 }
|
Chris@10
|
1270 {
|
Chris@10
|
1271 E TjW, TbB, Tk2, T99, TbF, TbL, Tbv, Taj, Tcu, Tcy, Tci, Tce, Tcr, Tcx, Tch;
|
Chris@10
|
1272 E Tc7, Tcn, Tkg, Tka, TbZ, TbP, T7J, TbO, T7u, Tk7, TjT, TbI, TbM, Tbw, Tbs;
|
Chris@10
|
1273 E T7Y, TbQ;
|
Chris@10
|
1274 {
|
Chris@10
|
1275 E TbX, TbW, TbU, TbT, Tc1, Tc5, Tc4, Tc2, TaG, Tbq, Tbn, Tcb, Tcs, Tca, Tcc;
|
Chris@10
|
1276 E Tbb, Tcm, TbV;
|
Chris@10
|
1277 {
|
Chris@10
|
1278 E T8W, Tbz, T8z, T97, T8n, T8y;
|
Chris@10
|
1279 TbX = FNMS(KP707106781, T8m, T87);
|
Chris@10
|
1280 T8n = FMA(KP707106781, T8m, T87);
|
Chris@10
|
1281 T8y = FMA(KP707106781, T8x, T8u);
|
Chris@10
|
1282 TbW = FNMS(KP707106781, T8x, T8u);
|
Chris@10
|
1283 TbU = FNMS(KP707106781, T8V, T8G);
|
Chris@10
|
1284 T8W = FMA(KP707106781, T8V, T8G);
|
Chris@10
|
1285 ci[WS(rs, 15)] = TiI + TiL;
|
Chris@10
|
1286 cr[WS(rs, 16)] = TiI - TiL;
|
Chris@10
|
1287 Tbz = FMA(KP198912367, T8n, T8y);
|
Chris@10
|
1288 T8z = FNMS(KP198912367, T8y, T8n);
|
Chris@10
|
1289 T97 = FMA(KP707106781, T96, T93);
|
Chris@10
|
1290 TbT = FNMS(KP707106781, T96, T93);
|
Chris@10
|
1291 {
|
Chris@10
|
1292 E Tae, TbD, Ta3, Tah;
|
Chris@10
|
1293 {
|
Chris@10
|
1294 E T9x, Ta2, TbA, T98;
|
Chris@10
|
1295 Tc1 = FNMS(KP707106781, T9w, T9h);
|
Chris@10
|
1296 T9x = FMA(KP707106781, T9w, T9h);
|
Chris@10
|
1297 Ta2 = T9M + Ta1;
|
Chris@10
|
1298 Tc5 = Ta1 - T9M;
|
Chris@10
|
1299 Tc4 = FNMS(KP707106781, Tad, Taa);
|
Chris@10
|
1300 Tae = FMA(KP707106781, Tad, Taa);
|
Chris@10
|
1301 TbA = FNMS(KP198912367, T8W, T97);
|
Chris@10
|
1302 T98 = FMA(KP198912367, T97, T8W);
|
Chris@10
|
1303 TbD = FNMS(KP923879532, Ta2, T9x);
|
Chris@10
|
1304 Ta3 = FMA(KP923879532, Ta2, T9x);
|
Chris@10
|
1305 TjW = Tbz + TbA;
|
Chris@10
|
1306 TbB = Tbz - TbA;
|
Chris@10
|
1307 Tk2 = T98 - T8z;
|
Chris@10
|
1308 T99 = T8z + T98;
|
Chris@10
|
1309 Tah = Taf + Tag;
|
Chris@10
|
1310 Tc2 = Taf - Tag;
|
Chris@10
|
1311 }
|
Chris@10
|
1312 {
|
Chris@10
|
1313 E Tc8, Tc9, TbE, Tai;
|
Chris@10
|
1314 TaG = FMA(KP707106781, TaF, Taq);
|
Chris@10
|
1315 Tc8 = FNMS(KP707106781, TaF, Taq);
|
Chris@10
|
1316 Tc9 = Tbp - Tbo;
|
Chris@10
|
1317 Tbq = Tbo + Tbp;
|
Chris@10
|
1318 Tbn = FMA(KP707106781, Tbm, Tbj);
|
Chris@10
|
1319 Tcb = FNMS(KP707106781, Tbm, Tbj);
|
Chris@10
|
1320 TbE = FNMS(KP923879532, Tah, Tae);
|
Chris@10
|
1321 Tai = FMA(KP923879532, Tah, Tae);
|
Chris@10
|
1322 Tcs = FMA(KP923879532, Tc9, Tc8);
|
Chris@10
|
1323 Tca = FNMS(KP923879532, Tc9, Tc8);
|
Chris@10
|
1324 TbF = FMA(KP820678790, TbE, TbD);
|
Chris@10
|
1325 TbL = FNMS(KP820678790, TbD, TbE);
|
Chris@10
|
1326 Tbv = FMA(KP098491403, Ta3, Tai);
|
Chris@10
|
1327 Taj = FNMS(KP098491403, Tai, Ta3);
|
Chris@10
|
1328 Tcc = Tba - TaV;
|
Chris@10
|
1329 Tbb = TaV + Tba;
|
Chris@10
|
1330 }
|
Chris@10
|
1331 }
|
Chris@10
|
1332 }
|
Chris@10
|
1333 {
|
Chris@10
|
1334 E Tcp, Tc3, Tct, Tcd, Tcq, Tc6;
|
Chris@10
|
1335 Tct = FNMS(KP923879532, Tcc, Tcb);
|
Chris@10
|
1336 Tcd = FMA(KP923879532, Tcc, Tcb);
|
Chris@10
|
1337 Tcp = FMA(KP923879532, Tc2, Tc1);
|
Chris@10
|
1338 Tc3 = FNMS(KP923879532, Tc2, Tc1);
|
Chris@10
|
1339 Tcu = FMA(KP303346683, Tct, Tcs);
|
Chris@10
|
1340 Tcy = FNMS(KP303346683, Tcs, Tct);
|
Chris@10
|
1341 Tci = FMA(KP534511135, Tca, Tcd);
|
Chris@10
|
1342 Tce = FNMS(KP534511135, Tcd, Tca);
|
Chris@10
|
1343 Tcq = FMA(KP923879532, Tc5, Tc4);
|
Chris@10
|
1344 Tc6 = FNMS(KP923879532, Tc5, Tc4);
|
Chris@10
|
1345 Tcm = FNMS(KP668178637, TbT, TbU);
|
Chris@10
|
1346 TbV = FMA(KP668178637, TbU, TbT);
|
Chris@10
|
1347 Tcr = FMA(KP303346683, Tcq, Tcp);
|
Chris@10
|
1348 Tcx = FNMS(KP303346683, Tcp, Tcq);
|
Chris@10
|
1349 Tch = FMA(KP534511135, Tc3, Tc6);
|
Chris@10
|
1350 Tc7 = FNMS(KP534511135, Tc6, Tc3);
|
Chris@10
|
1351 }
|
Chris@10
|
1352 {
|
Chris@10
|
1353 E TbG, Tbc, Tcl, TbY;
|
Chris@10
|
1354 Tcl = FMA(KP668178637, TbW, TbX);
|
Chris@10
|
1355 TbY = FNMS(KP668178637, TbX, TbW);
|
Chris@10
|
1356 TbG = FNMS(KP923879532, Tbb, TaG);
|
Chris@10
|
1357 Tbc = FMA(KP923879532, Tbb, TaG);
|
Chris@10
|
1358 Tcn = Tcl + Tcm;
|
Chris@10
|
1359 Tkg = Tcl - Tcm;
|
Chris@10
|
1360 Tka = TbY + TbV;
|
Chris@10
|
1361 TbZ = TbV - TbY;
|
Chris@10
|
1362 {
|
Chris@10
|
1363 E T7t, TjS, TbH, Tbr;
|
Chris@10
|
1364 Tkm = T7s - T7l;
|
Chris@10
|
1365 T7t = T7l + T7s;
|
Chris@10
|
1366 TjS = TcB - TcC;
|
Chris@10
|
1367 TcD = TcB + TcC;
|
Chris@10
|
1368 TbP = FMA(KP414213562, T7B, T7I);
|
Chris@10
|
1369 T7J = FNMS(KP414213562, T7I, T7B);
|
Chris@10
|
1370 TbH = FNMS(KP923879532, Tbq, Tbn);
|
Chris@10
|
1371 Tbr = FMA(KP923879532, Tbq, Tbn);
|
Chris@10
|
1372 TbO = FNMS(KP707106781, T7t, T7e);
|
Chris@10
|
1373 T7u = FMA(KP707106781, T7t, T7e);
|
Chris@10
|
1374 Tk7 = FNMS(KP707106781, TjS, TjR);
|
Chris@10
|
1375 TjT = FMA(KP707106781, TjS, TjR);
|
Chris@10
|
1376 TbI = FMA(KP820678790, TbH, TbG);
|
Chris@10
|
1377 TbM = FNMS(KP820678790, TbG, TbH);
|
Chris@10
|
1378 Tbw = FMA(KP098491403, Tbc, Tbr);
|
Chris@10
|
1379 Tbs = FNMS(KP098491403, Tbr, Tbc);
|
Chris@10
|
1380 T7Y = FMA(KP414213562, T7X, T7Q);
|
Chris@10
|
1381 TbQ = FNMS(KP414213562, T7Q, T7X);
|
Chris@10
|
1382 }
|
Chris@10
|
1383 }
|
Chris@10
|
1384 }
|
Chris@10
|
1385 {
|
Chris@10
|
1386 E Tk1, TjV, Tck, TbS, Tkd, Tcz, Tkh, Tcf, TjY, Tk6, Tke, Tcv, Tki, Tcj;
|
Chris@10
|
1387 {
|
Chris@10
|
1388 E Tbu, TbC, Tkb, Tkc, Tkj, Tkk, Tbx, TbJ;
|
Chris@10
|
1389 {
|
Chris@10
|
1390 E Tbt, Tkf, Tk9, T9a, TbK, TbN, Tby;
|
Chris@10
|
1391 Tk0 = Tbs - Taj;
|
Chris@10
|
1392 Tbt = Taj + Tbs;
|
Chris@10
|
1393 {
|
Chris@10
|
1394 E Tk8, T7Z, TjU, TbR, T80;
|
Chris@10
|
1395 Tk8 = T7Y - T7J;
|
Chris@10
|
1396 T7Z = T7J + T7Y;
|
Chris@10
|
1397 TjU = TbP + TbQ;
|
Chris@10
|
1398 TbR = TbP - TbQ;
|
Chris@10
|
1399 Tkf = FNMS(KP923879532, Tk8, Tk7);
|
Chris@10
|
1400 Tk9 = FMA(KP923879532, Tk8, Tk7);
|
Chris@10
|
1401 Tby = FNMS(KP923879532, T7Z, T7u);
|
Chris@10
|
1402 T80 = FMA(KP923879532, T7Z, T7u);
|
Chris@10
|
1403 Tk1 = FNMS(KP923879532, TjU, TjT);
|
Chris@10
|
1404 TjV = FMA(KP923879532, TjU, TjT);
|
Chris@10
|
1405 Tck = FMA(KP923879532, TbR, TbO);
|
Chris@10
|
1406 TbS = FNMS(KP923879532, TbR, TbO);
|
Chris@10
|
1407 T9a = FMA(KP980785280, T99, T80);
|
Chris@10
|
1408 Tbu = FNMS(KP980785280, T99, T80);
|
Chris@10
|
1409 }
|
Chris@10
|
1410 TbC = FMA(KP980785280, TbB, Tby);
|
Chris@10
|
1411 TbK = FNMS(KP980785280, TbB, Tby);
|
Chris@10
|
1412 TbN = TbL + TbM;
|
Chris@10
|
1413 Tk4 = TbL - TbM;
|
Chris@10
|
1414 Tkd = FNMS(KP831469612, Tka, Tk9);
|
Chris@10
|
1415 Tkb = FMA(KP831469612, Tka, Tk9);
|
Chris@10
|
1416 ci[0] = FMA(KP995184726, Tbt, T9a);
|
Chris@10
|
1417 cr[WS(rs, 31)] = FNMS(KP995184726, Tbt, T9a);
|
Chris@10
|
1418 ci[WS(rs, 8)] = FNMS(KP773010453, TbN, TbK);
|
Chris@10
|
1419 cr[WS(rs, 23)] = FMA(KP773010453, TbN, TbK);
|
Chris@10
|
1420 Tkc = Tcx - Tcy;
|
Chris@10
|
1421 Tcz = Tcx + Tcy;
|
Chris@10
|
1422 Tkh = FMA(KP831469612, Tkg, Tkf);
|
Chris@10
|
1423 Tkj = FNMS(KP831469612, Tkg, Tkf);
|
Chris@10
|
1424 Tkk = Tce - Tc7;
|
Chris@10
|
1425 Tcf = Tc7 + Tce;
|
Chris@10
|
1426 }
|
Chris@10
|
1427 ci[WS(rs, 60)] = FMA(KP956940335, Tkc, Tkb);
|
Chris@10
|
1428 cr[WS(rs, 35)] = FMS(KP956940335, Tkc, Tkb);
|
Chris@10
|
1429 ci[WS(rs, 52)] = FMA(KP881921264, Tkk, Tkj);
|
Chris@10
|
1430 cr[WS(rs, 43)] = FMS(KP881921264, Tkk, Tkj);
|
Chris@10
|
1431 Tbx = Tbv + Tbw;
|
Chris@10
|
1432 TjY = Tbw - Tbv;
|
Chris@10
|
1433 TbJ = TbF + TbI;
|
Chris@10
|
1434 Tk6 = TbI - TbF;
|
Chris@10
|
1435 cr[WS(rs, 15)] = FMA(KP995184726, Tbx, Tbu);
|
Chris@10
|
1436 ci[WS(rs, 16)] = FNMS(KP995184726, Tbx, Tbu);
|
Chris@10
|
1437 cr[WS(rs, 7)] = FMA(KP773010453, TbJ, TbC);
|
Chris@10
|
1438 ci[WS(rs, 24)] = FNMS(KP773010453, TbJ, TbC);
|
Chris@10
|
1439 Tke = Tcu - Tcr;
|
Chris@10
|
1440 Tcv = Tcr + Tcu;
|
Chris@10
|
1441 Tki = Tci - Tch;
|
Chris@10
|
1442 Tcj = Tch + Tci;
|
Chris@10
|
1443 }
|
Chris@10
|
1444 {
|
Chris@10
|
1445 E Tcg, Tco, TjX, Tk5, Tc0, Tcw;
|
Chris@10
|
1446 Tcg = FNMS(KP831469612, TbZ, TbS);
|
Chris@10
|
1447 Tc0 = FMA(KP831469612, TbZ, TbS);
|
Chris@10
|
1448 ci[WS(rs, 44)] = FMA(KP956940335, Tke, Tkd);
|
Chris@10
|
1449 cr[WS(rs, 51)] = FMS(KP956940335, Tke, Tkd);
|
Chris@10
|
1450 ci[WS(rs, 36)] = FMA(KP881921264, Tki, Tkh);
|
Chris@10
|
1451 cr[WS(rs, 59)] = FMS(KP881921264, Tki, Tkh);
|
Chris@10
|
1452 Tco = FMA(KP831469612, Tcn, Tck);
|
Chris@10
|
1453 Tcw = FNMS(KP831469612, Tcn, Tck);
|
Chris@10
|
1454 TjZ = FNMS(KP980785280, TjW, TjV);
|
Chris@10
|
1455 TjX = FMA(KP980785280, TjW, TjV);
|
Chris@10
|
1456 ci[WS(rs, 4)] = FMA(KP881921264, Tcf, Tc0);
|
Chris@10
|
1457 cr[WS(rs, 27)] = FNMS(KP881921264, Tcf, Tc0);
|
Chris@10
|
1458 ci[WS(rs, 12)] = FNMS(KP956940335, Tcz, Tcw);
|
Chris@10
|
1459 cr[WS(rs, 19)] = FMA(KP956940335, Tcz, Tcw);
|
Chris@10
|
1460 Tk3 = FMA(KP980785280, Tk2, Tk1);
|
Chris@10
|
1461 Tk5 = FNMS(KP980785280, Tk2, Tk1);
|
Chris@10
|
1462 ci[WS(rs, 32)] = FMA(KP995184726, TjY, TjX);
|
Chris@10
|
1463 cr[WS(rs, 63)] = FMS(KP995184726, TjY, TjX);
|
Chris@10
|
1464 ci[WS(rs, 40)] = FMA(KP773010453, Tk6, Tk5);
|
Chris@10
|
1465 cr[WS(rs, 55)] = FMS(KP773010453, Tk6, Tk5);
|
Chris@10
|
1466 cr[WS(rs, 11)] = FMA(KP881921264, Tcj, Tcg);
|
Chris@10
|
1467 ci[WS(rs, 20)] = FNMS(KP881921264, Tcj, Tcg);
|
Chris@10
|
1468 cr[WS(rs, 3)] = FMA(KP956940335, Tcv, Tco);
|
Chris@10
|
1469 ci[WS(rs, 28)] = FNMS(KP956940335, Tcv, Tco);
|
Chris@10
|
1470 }
|
Chris@10
|
1471 }
|
Chris@10
|
1472 }
|
Chris@10
|
1473 }
|
Chris@10
|
1474 {
|
Chris@10
|
1475 E Ti8, Thu, Tjf, Tj9, Tib, Tjg, Tja, ThF, Tig, ThW, Tif, Til, Ti6, ThR;
|
Chris@10
|
1476 ci[WS(rs, 48)] = FMA(KP995184726, Tk0, TjZ);
|
Chris@10
|
1477 cr[WS(rs, 47)] = FMS(KP995184726, Tk0, TjZ);
|
Chris@10
|
1478 ci[WS(rs, 56)] = FMA(KP773010453, Tk4, Tk3);
|
Chris@10
|
1479 cr[WS(rs, 39)] = FMS(KP773010453, Tk4, Tk3);
|
Chris@10
|
1480 Ti8 = Thq + Tht;
|
Chris@10
|
1481 Thu = Thq - Tht;
|
Chris@10
|
1482 Tjf = Tj8 - Tj7;
|
Chris@10
|
1483 Tj9 = Tj7 + Tj8;
|
Chris@10
|
1484 {
|
Chris@10
|
1485 E Tid, ThL, Tie, ThQ;
|
Chris@10
|
1486 {
|
Chris@10
|
1487 E Ti9, Thz, Tia, ThE;
|
Chris@10
|
1488 Ti9 = Thv - Thy;
|
Chris@10
|
1489 Thz = Thv + Thy;
|
Chris@10
|
1490 Tia = ThA + ThD;
|
Chris@10
|
1491 ThE = ThA - ThD;
|
Chris@10
|
1492 Tib = Ti9 + Tia;
|
Chris@10
|
1493 Tjg = Tia - Ti9;
|
Chris@10
|
1494 Tja = Thz - ThE;
|
Chris@10
|
1495 ThF = Thz + ThE;
|
Chris@10
|
1496 Tid = ThH + ThK;
|
Chris@10
|
1497 ThL = ThH - ThK;
|
Chris@10
|
1498 }
|
Chris@10
|
1499 Tie = ThO + ThP;
|
Chris@10
|
1500 ThQ = ThO - ThP;
|
Chris@10
|
1501 Tig = ThS + ThV;
|
Chris@10
|
1502 ThW = ThS - ThV;
|
Chris@10
|
1503 Tif = FNMS(KP414213562, Tie, Tid);
|
Chris@10
|
1504 Til = FMA(KP414213562, Tid, Tie);
|
Chris@10
|
1505 Ti6 = FNMS(KP414213562, ThL, ThQ);
|
Chris@10
|
1506 ThR = FMA(KP414213562, ThQ, ThL);
|
Chris@10
|
1507 }
|
Chris@10
|
1508 {
|
Chris@10
|
1509 E Ti4, ThG, Tjh, Tjj, Tih, Ti1;
|
Chris@10
|
1510 Ti4 = FNMS(KP707106781, ThF, Thu);
|
Chris@10
|
1511 ThG = FMA(KP707106781, ThF, Thu);
|
Chris@10
|
1512 Tjh = FMA(KP707106781, Tjg, Tjf);
|
Chris@10
|
1513 Tjj = FNMS(KP707106781, Tjg, Tjf);
|
Chris@10
|
1514 Tih = Ti0 - ThZ;
|
Chris@10
|
1515 Ti1 = ThZ + Ti0;
|
Chris@10
|
1516 {
|
Chris@10
|
1517 E Tje, Tjd, Tjb, Tjc;
|
Chris@10
|
1518 {
|
Chris@10
|
1519 E Tic, Tim, Ti5, Ti2, Tij, Tii;
|
Chris@10
|
1520 Tik = FNMS(KP707106781, Tib, Ti8);
|
Chris@10
|
1521 Tic = FMA(KP707106781, Tib, Ti8);
|
Chris@10
|
1522 Tii = FNMS(KP414213562, Tih, Tig);
|
Chris@10
|
1523 Tim = FMA(KP414213562, Tig, Tih);
|
Chris@10
|
1524 Ti5 = FMA(KP414213562, ThW, Ti1);
|
Chris@10
|
1525 Ti2 = FNMS(KP414213562, Ti1, ThW);
|
Chris@10
|
1526 Tij = Tif + Tii;
|
Chris@10
|
1527 Tje = Tii - Tif;
|
Chris@10
|
1528 Tjd = FNMS(KP707106781, Tja, Tj9);
|
Chris@10
|
1529 Tjb = FMA(KP707106781, Tja, Tj9);
|
Chris@10
|
1530 {
|
Chris@10
|
1531 E Ti7, Tji, Tjk, Ti3;
|
Chris@10
|
1532 Ti7 = Ti5 - Ti6;
|
Chris@10
|
1533 Tji = Ti6 + Ti5;
|
Chris@10
|
1534 Tjk = Ti2 - ThR;
|
Chris@10
|
1535 Ti3 = ThR + Ti2;
|
Chris@10
|
1536 ci[WS(rs, 3)] = FMA(KP923879532, Tij, Tic);
|
Chris@10
|
1537 cr[WS(rs, 28)] = FNMS(KP923879532, Tij, Tic);
|
Chris@10
|
1538 ci[WS(rs, 11)] = FMA(KP923879532, Ti7, Ti4);
|
Chris@10
|
1539 cr[WS(rs, 20)] = FNMS(KP923879532, Ti7, Ti4);
|
Chris@10
|
1540 ci[WS(rs, 59)] = FMA(KP923879532, Tji, Tjh);
|
Chris@10
|
1541 cr[WS(rs, 36)] = FMS(KP923879532, Tji, Tjh);
|
Chris@10
|
1542 ci[WS(rs, 43)] = FMA(KP923879532, Tjk, Tjj);
|
Chris@10
|
1543 cr[WS(rs, 52)] = FMS(KP923879532, Tjk, Tjj);
|
Chris@10
|
1544 cr[WS(rs, 4)] = FMA(KP923879532, Ti3, ThG);
|
Chris@10
|
1545 ci[WS(rs, 27)] = FNMS(KP923879532, Ti3, ThG);
|
Chris@10
|
1546 Tjc = Tim - Til;
|
Chris@10
|
1547 Tin = Til + Tim;
|
Chris@10
|
1548 }
|
Chris@10
|
1549 }
|
Chris@10
|
1550 ci[WS(rs, 35)] = FMA(KP923879532, Tjc, Tjb);
|
Chris@10
|
1551 cr[WS(rs, 60)] = FMS(KP923879532, Tjc, Tjb);
|
Chris@10
|
1552 ci[WS(rs, 51)] = FMA(KP923879532, Tje, Tjd);
|
Chris@10
|
1553 cr[WS(rs, 44)] = FMS(KP923879532, Tje, Tjd);
|
Chris@10
|
1554 }
|
Chris@10
|
1555 }
|
Chris@10
|
1556 }
|
Chris@10
|
1557 {
|
Chris@10
|
1558 E Tjy, Tju, Tjt, Tjx;
|
Chris@10
|
1559 {
|
Chris@10
|
1560 E TjD, TjJ, Tgo, Tf2, Tjp, Tjv, Tha, TgI, Tgg, Tgd, Tgr, Tjw, Tjq, Tfp, Thk;
|
Chris@10
|
1561 E Tho, Th7, Th4, Tgv, TgB, Tgl, TfR, TjE, Thd, TjK, TgP, Tgw, Tg8, Thh, Thn;
|
Chris@10
|
1562 E Th8, TgX;
|
Chris@10
|
1563 {
|
Chris@10
|
1564 E TgK, TgJ, TgN, TgM, TfW, Th1, Thi, Th0, Th2, Tg7;
|
Chris@10
|
1565 {
|
Chris@10
|
1566 E TgE, TeQ, TjB, Tjn, TgF, TgG, TjC, Tf1, TeV, Tf0;
|
Chris@10
|
1567 TgE = TeM - TeP;
|
Chris@10
|
1568 TeQ = TeM + TeP;
|
Chris@10
|
1569 TjB = Tjm - Tjl;
|
Chris@10
|
1570 Tjn = Tjl + Tjm;
|
Chris@10
|
1571 TgF = TeR + TeU;
|
Chris@10
|
1572 TeV = TeR - TeU;
|
Chris@10
|
1573 cr[WS(rs, 12)] = FMA(KP923879532, Tin, Tik);
|
Chris@10
|
1574 ci[WS(rs, 19)] = FNMS(KP923879532, Tin, Tik);
|
Chris@10
|
1575 Tf0 = TeW + TeZ;
|
Chris@10
|
1576 TgG = TeW - TeZ;
|
Chris@10
|
1577 TjC = Tf0 - TeV;
|
Chris@10
|
1578 Tf1 = TeV + Tf0;
|
Chris@10
|
1579 {
|
Chris@10
|
1580 E Tfi, Tgp, Tfd, Tfn;
|
Chris@10
|
1581 {
|
Chris@10
|
1582 E Tf7, Tjo, TgH, Tfc;
|
Chris@10
|
1583 TgK = Tf5 - Tf6;
|
Chris@10
|
1584 Tf7 = Tf5 + Tf6;
|
Chris@10
|
1585 TjD = FMA(KP707106781, TjC, TjB);
|
Chris@10
|
1586 TjJ = FNMS(KP707106781, TjC, TjB);
|
Chris@10
|
1587 Tgo = FMA(KP707106781, Tf1, TeQ);
|
Chris@10
|
1588 Tf2 = FNMS(KP707106781, Tf1, TeQ);
|
Chris@10
|
1589 Tjo = TgF - TgG;
|
Chris@10
|
1590 TgH = TgF + TgG;
|
Chris@10
|
1591 Tfc = Tf8 + Tfb;
|
Chris@10
|
1592 TgJ = Tf8 - Tfb;
|
Chris@10
|
1593 TgN = Tfg - Tfh;
|
Chris@10
|
1594 Tfi = Tfg + Tfh;
|
Chris@10
|
1595 Tjp = FMA(KP707106781, Tjo, Tjn);
|
Chris@10
|
1596 Tjv = FNMS(KP707106781, Tjo, Tjn);
|
Chris@10
|
1597 Tha = FNMS(KP707106781, TgH, TgE);
|
Chris@10
|
1598 TgI = FMA(KP707106781, TgH, TgE);
|
Chris@10
|
1599 Tgp = FNMS(KP414213562, Tf7, Tfc);
|
Chris@10
|
1600 Tfd = FMA(KP414213562, Tfc, Tf7);
|
Chris@10
|
1601 Tfn = Tfj + Tfm;
|
Chris@10
|
1602 TgM = Tfj - Tfm;
|
Chris@10
|
1603 }
|
Chris@10
|
1604 {
|
Chris@10
|
1605 E TgY, TgZ, Tgq, Tfo;
|
Chris@10
|
1606 TfW = TfS + TfV;
|
Chris@10
|
1607 TgY = TfS - TfV;
|
Chris@10
|
1608 TgZ = Tgf + Tge;
|
Chris@10
|
1609 Tgg = Tge - Tgf;
|
Chris@10
|
1610 Tgd = Tg9 - Tgc;
|
Chris@10
|
1611 Th1 = Tgc + Tg9;
|
Chris@10
|
1612 Tgq = FMA(KP414213562, Tfi, Tfn);
|
Chris@10
|
1613 Tfo = FNMS(KP414213562, Tfn, Tfi);
|
Chris@10
|
1614 Thi = FNMS(KP707106781, TgZ, TgY);
|
Chris@10
|
1615 Th0 = FMA(KP707106781, TgZ, TgY);
|
Chris@10
|
1616 Tgr = Tgp + Tgq;
|
Chris@10
|
1617 Tjw = Tgq - Tgp;
|
Chris@10
|
1618 Tjq = Tfd + Tfo;
|
Chris@10
|
1619 Tfp = Tfd - Tfo;
|
Chris@10
|
1620 Th2 = Tg6 - Tg1;
|
Chris@10
|
1621 Tg7 = Tg1 + Tg6;
|
Chris@10
|
1622 }
|
Chris@10
|
1623 }
|
Chris@10
|
1624 }
|
Chris@10
|
1625 {
|
Chris@10
|
1626 E TgR, TgV, TgU, TgS, Thc, TgL;
|
Chris@10
|
1627 {
|
Chris@10
|
1628 E TfM, Tgt, TfH, TfP, Tgu, TfQ;
|
Chris@10
|
1629 {
|
Chris@10
|
1630 E Tfv, TfG, Thj, Th3;
|
Chris@10
|
1631 TgR = Tfr - Tfu;
|
Chris@10
|
1632 Tfv = Tfr + Tfu;
|
Chris@10
|
1633 TfG = TfA + TfF;
|
Chris@10
|
1634 TgV = TfF - TfA;
|
Chris@10
|
1635 TgU = TfK - TfL;
|
Chris@10
|
1636 TfM = TfK + TfL;
|
Chris@10
|
1637 Thj = FNMS(KP707106781, Th2, Th1);
|
Chris@10
|
1638 Th3 = FMA(KP707106781, Th2, Th1);
|
Chris@10
|
1639 Tgt = FMA(KP707106781, TfG, Tfv);
|
Chris@10
|
1640 TfH = FNMS(KP707106781, TfG, Tfv);
|
Chris@10
|
1641 Thk = FMA(KP668178637, Thj, Thi);
|
Chris@10
|
1642 Tho = FNMS(KP668178637, Thi, Thj);
|
Chris@10
|
1643 Th7 = FMA(KP198912367, Th0, Th3);
|
Chris@10
|
1644 Th4 = FNMS(KP198912367, Th3, Th0);
|
Chris@10
|
1645 TfP = TfN + TfO;
|
Chris@10
|
1646 TgS = TfN - TfO;
|
Chris@10
|
1647 }
|
Chris@10
|
1648 Tgu = FMA(KP707106781, TfP, TfM);
|
Chris@10
|
1649 TfQ = FNMS(KP707106781, TfP, TfM);
|
Chris@10
|
1650 Thc = FNMS(KP414213562, TgJ, TgK);
|
Chris@10
|
1651 TgL = FMA(KP414213562, TgK, TgJ);
|
Chris@10
|
1652 Tgv = FNMS(KP198912367, Tgu, Tgt);
|
Chris@10
|
1653 TgB = FMA(KP198912367, Tgt, Tgu);
|
Chris@10
|
1654 Tgl = FNMS(KP668178637, TfH, TfQ);
|
Chris@10
|
1655 TfR = FMA(KP668178637, TfQ, TfH);
|
Chris@10
|
1656 }
|
Chris@10
|
1657 {
|
Chris@10
|
1658 E Thf, TgT, Thb, TgO, Thg, TgW;
|
Chris@10
|
1659 Thb = FMA(KP414213562, TgM, TgN);
|
Chris@10
|
1660 TgO = FNMS(KP414213562, TgN, TgM);
|
Chris@10
|
1661 Thf = FNMS(KP707106781, TgS, TgR);
|
Chris@10
|
1662 TgT = FMA(KP707106781, TgS, TgR);
|
Chris@10
|
1663 TjE = Thc + Thb;
|
Chris@10
|
1664 Thd = Thb - Thc;
|
Chris@10
|
1665 TjK = TgL - TgO;
|
Chris@10
|
1666 TgP = TgL + TgO;
|
Chris@10
|
1667 Thg = FNMS(KP707106781, TgV, TgU);
|
Chris@10
|
1668 TgW = FMA(KP707106781, TgV, TgU);
|
Chris@10
|
1669 Tgw = FMA(KP707106781, Tg7, TfW);
|
Chris@10
|
1670 Tg8 = FNMS(KP707106781, Tg7, TfW);
|
Chris@10
|
1671 Thh = FNMS(KP668178637, Thg, Thf);
|
Chris@10
|
1672 Thn = FMA(KP668178637, Thf, Thg);
|
Chris@10
|
1673 Th8 = FNMS(KP198912367, TgT, TgW);
|
Chris@10
|
1674 TgX = FMA(KP198912367, TgW, TgT);
|
Chris@10
|
1675 }
|
Chris@10
|
1676 }
|
Chris@10
|
1677 }
|
Chris@10
|
1678 {
|
Chris@10
|
1679 E TjH, Th9, TjL, Tjs, TjA, Thl, TjI, Th5, TjM, Thp;
|
Chris@10
|
1680 {
|
Chris@10
|
1681 E Tgk, Tfq, TgA, Tgs, TjN, Tgy, Tgm, TgD, Tgj, TjO, Tgn, Tgz;
|
Chris@10
|
1682 Tgk = FNMS(KP923879532, Tfp, Tf2);
|
Chris@10
|
1683 Tfq = FMA(KP923879532, Tfp, Tf2);
|
Chris@10
|
1684 TgA = FNMS(KP923879532, Tgr, Tgo);
|
Chris@10
|
1685 Tgs = FMA(KP923879532, Tgr, Tgo);
|
Chris@10
|
1686 {
|
Chris@10
|
1687 E TjF, Tgx, Tgh, TjG, TgC, Tgi;
|
Chris@10
|
1688 TjH = FNMS(KP923879532, TjE, TjD);
|
Chris@10
|
1689 TjF = FMA(KP923879532, TjE, TjD);
|
Chris@10
|
1690 Tgx = FMA(KP707106781, Tgg, Tgd);
|
Chris@10
|
1691 Tgh = FNMS(KP707106781, Tgg, Tgd);
|
Chris@10
|
1692 TjG = Th8 + Th7;
|
Chris@10
|
1693 Th9 = Th7 - Th8;
|
Chris@10
|
1694 TjL = FMA(KP923879532, TjK, TjJ);
|
Chris@10
|
1695 TjN = FNMS(KP923879532, TjK, TjJ);
|
Chris@10
|
1696 Tgy = FNMS(KP198912367, Tgx, Tgw);
|
Chris@10
|
1697 TgC = FMA(KP198912367, Tgw, Tgx);
|
Chris@10
|
1698 Tgm = FNMS(KP668178637, Tg8, Tgh);
|
Chris@10
|
1699 Tgi = FMA(KP668178637, Tgh, Tg8);
|
Chris@10
|
1700 ci[WS(rs, 61)] = FMA(KP980785280, TjG, TjF);
|
Chris@10
|
1701 cr[WS(rs, 34)] = FMS(KP980785280, TjG, TjF);
|
Chris@10
|
1702 TgD = TgB + TgC;
|
Chris@10
|
1703 Tjs = TgC - TgB;
|
Chris@10
|
1704 TjA = Tgi - TfR;
|
Chris@10
|
1705 Tgj = TfR + Tgi;
|
Chris@10
|
1706 TjO = Thk - Thh;
|
Chris@10
|
1707 Thl = Thh + Thk;
|
Chris@10
|
1708 }
|
Chris@10
|
1709 cr[WS(rs, 14)] = FMA(KP980785280, TgD, TgA);
|
Chris@10
|
1710 ci[WS(rs, 17)] = FNMS(KP980785280, TgD, TgA);
|
Chris@10
|
1711 cr[WS(rs, 6)] = FMA(KP831469612, Tgj, Tfq);
|
Chris@10
|
1712 ci[WS(rs, 25)] = FNMS(KP831469612, Tgj, Tfq);
|
Chris@10
|
1713 ci[WS(rs, 53)] = FMA(KP831469612, TjO, TjN);
|
Chris@10
|
1714 cr[WS(rs, 42)] = FMS(KP831469612, TjO, TjN);
|
Chris@10
|
1715 Tgn = Tgl + Tgm;
|
Chris@10
|
1716 Tjy = Tgl - Tgm;
|
Chris@10
|
1717 Tgz = Tgv + Tgy;
|
Chris@10
|
1718 Tju = Tgy - Tgv;
|
Chris@10
|
1719 ci[WS(rs, 9)] = FNMS(KP831469612, Tgn, Tgk);
|
Chris@10
|
1720 cr[WS(rs, 22)] = FMA(KP831469612, Tgn, Tgk);
|
Chris@10
|
1721 ci[WS(rs, 1)] = FMA(KP980785280, Tgz, Tgs);
|
Chris@10
|
1722 cr[WS(rs, 30)] = FNMS(KP980785280, Tgz, Tgs);
|
Chris@10
|
1723 TjI = Th4 - TgX;
|
Chris@10
|
1724 Th5 = TgX + Th4;
|
Chris@10
|
1725 TjM = Thn + Tho;
|
Chris@10
|
1726 Thp = Thn - Tho;
|
Chris@10
|
1727 }
|
Chris@10
|
1728 {
|
Chris@10
|
1729 E Th6, The, Tjr, Tjz, TgQ, Thm;
|
Chris@10
|
1730 Th6 = FNMS(KP923879532, TgP, TgI);
|
Chris@10
|
1731 TgQ = FMA(KP923879532, TgP, TgI);
|
Chris@10
|
1732 ci[WS(rs, 45)] = FMA(KP980785280, TjI, TjH);
|
Chris@10
|
1733 cr[WS(rs, 50)] = FMS(KP980785280, TjI, TjH);
|
Chris@10
|
1734 ci[WS(rs, 37)] = FNMS(KP831469612, TjM, TjL);
|
Chris@10
|
1735 cr[WS(rs, 58)] = -(FMA(KP831469612, TjM, TjL));
|
Chris@10
|
1736 The = FMA(KP923879532, Thd, Tha);
|
Chris@10
|
1737 Thm = FNMS(KP923879532, Thd, Tha);
|
Chris@10
|
1738 Tjt = FNMS(KP923879532, Tjq, Tjp);
|
Chris@10
|
1739 Tjr = FMA(KP923879532, Tjq, Tjp);
|
Chris@10
|
1740 cr[WS(rs, 2)] = FMA(KP980785280, Th5, TgQ);
|
Chris@10
|
1741 ci[WS(rs, 29)] = FNMS(KP980785280, Th5, TgQ);
|
Chris@10
|
1742 cr[WS(rs, 10)] = FMA(KP831469612, Thp, Thm);
|
Chris@10
|
1743 ci[WS(rs, 21)] = FNMS(KP831469612, Thp, Thm);
|
Chris@10
|
1744 Tjx = FMA(KP923879532, Tjw, Tjv);
|
Chris@10
|
1745 Tjz = FNMS(KP923879532, Tjw, Tjv);
|
Chris@10
|
1746 ci[WS(rs, 33)] = FMA(KP980785280, Tjs, Tjr);
|
Chris@10
|
1747 cr[WS(rs, 62)] = FMS(KP980785280, Tjs, Tjr);
|
Chris@10
|
1748 ci[WS(rs, 41)] = FMA(KP831469612, TjA, Tjz);
|
Chris@10
|
1749 cr[WS(rs, 54)] = FMS(KP831469612, TjA, Tjz);
|
Chris@10
|
1750 ci[WS(rs, 13)] = FMA(KP980785280, Th9, Th6);
|
Chris@10
|
1751 cr[WS(rs, 18)] = FNMS(KP980785280, Th9, Th6);
|
Chris@10
|
1752 ci[WS(rs, 5)] = FMA(KP831469612, Thl, The);
|
Chris@10
|
1753 cr[WS(rs, 26)] = FNMS(KP831469612, Thl, The);
|
Chris@10
|
1754 }
|
Chris@10
|
1755 }
|
Chris@10
|
1756 }
|
Chris@10
|
1757 {
|
Chris@10
|
1758 E Tkq, TdN, Tkw, Td1, TdR, TdX, TdI, Tdl, TeG, TeK, Tet, Teq, TeD, TeJ, Teu;
|
Chris@10
|
1759 E Tej, Tez, TkK, TkE, Teb, Te2, TcH, Te0, TcE, TkB, Tkn, TdU, TdY, TdH, TdE;
|
Chris@10
|
1760 E TcK, Te1;
|
Chris@10
|
1761 {
|
Chris@10
|
1762 E Te6, Te5, Te9, Te8, Ted, Teh, Teg, Tee, Tdo, TdC, Tdz, Ten, TeE, Tem, Teo;
|
Chris@10
|
1763 E Tdv, Tex, Te7;
|
Chris@10
|
1764 {
|
Chris@10
|
1765 E TcP, TcS, TcW, TcZ;
|
Chris@10
|
1766 Te6 = FNMS(KP707106781, TcO, TcN);
|
Chris@10
|
1767 TcP = FMA(KP707106781, TcO, TcN);
|
Chris@10
|
1768 ci[WS(rs, 49)] = FMA(KP980785280, Tju, Tjt);
|
Chris@10
|
1769 cr[WS(rs, 46)] = FMS(KP980785280, Tju, Tjt);
|
Chris@10
|
1770 ci[WS(rs, 57)] = FMA(KP831469612, Tjy, Tjx);
|
Chris@10
|
1771 cr[WS(rs, 38)] = FMS(KP831469612, Tjy, Tjx);
|
Chris@10
|
1772 TcS = FMA(KP707106781, TcR, TcQ);
|
Chris@10
|
1773 Te5 = FNMS(KP707106781, TcR, TcQ);
|
Chris@10
|
1774 Te9 = FNMS(KP707106781, TcV, TcU);
|
Chris@10
|
1775 TcW = FMA(KP707106781, TcV, TcU);
|
Chris@10
|
1776 TcZ = FMA(KP707106781, TcY, TcX);
|
Chris@10
|
1777 Te8 = FNMS(KP707106781, TcY, TcX);
|
Chris@10
|
1778 {
|
Chris@10
|
1779 E Tdg, TdP, Tdd, Tdj;
|
Chris@10
|
1780 {
|
Chris@10
|
1781 E Td5, TdM, TcT, TdL, Td0, Tdc;
|
Chris@10
|
1782 Ted = FNMS(KP707106781, Td4, Td3);
|
Chris@10
|
1783 Td5 = FMA(KP707106781, Td4, Td3);
|
Chris@10
|
1784 TdM = FNMS(KP198912367, TcP, TcS);
|
Chris@10
|
1785 TcT = FMA(KP198912367, TcS, TcP);
|
Chris@10
|
1786 TdL = FMA(KP198912367, TcW, TcZ);
|
Chris@10
|
1787 Td0 = FNMS(KP198912367, TcZ, TcW);
|
Chris@10
|
1788 Tdc = Td8 + Tdb;
|
Chris@10
|
1789 Teh = Td8 - Tdb;
|
Chris@10
|
1790 Teg = FNMS(KP707106781, Tdf, Tde);
|
Chris@10
|
1791 Tdg = FMA(KP707106781, Tdf, Tde);
|
Chris@10
|
1792 Tkq = TdM + TdL;
|
Chris@10
|
1793 TdN = TdL - TdM;
|
Chris@10
|
1794 Tkw = TcT - Td0;
|
Chris@10
|
1795 Td1 = TcT + Td0;
|
Chris@10
|
1796 TdP = FNMS(KP923879532, Tdc, Td5);
|
Chris@10
|
1797 Tdd = FMA(KP923879532, Tdc, Td5);
|
Chris@10
|
1798 Tdj = Tdh + Tdi;
|
Chris@10
|
1799 Tee = Tdi - Tdh;
|
Chris@10
|
1800 }
|
Chris@10
|
1801 {
|
Chris@10
|
1802 E Tek, Tel, TdQ, Tdk;
|
Chris@10
|
1803 Tdo = FMA(KP707106781, Tdn, Tdm);
|
Chris@10
|
1804 Tek = FNMS(KP707106781, Tdn, Tdm);
|
Chris@10
|
1805 Tel = TdB - TdA;
|
Chris@10
|
1806 TdC = TdA + TdB;
|
Chris@10
|
1807 Tdz = FMA(KP707106781, Tdy, Tdx);
|
Chris@10
|
1808 Ten = FNMS(KP707106781, Tdy, Tdx);
|
Chris@10
|
1809 TdQ = FNMS(KP923879532, Tdj, Tdg);
|
Chris@10
|
1810 Tdk = FMA(KP923879532, Tdj, Tdg);
|
Chris@10
|
1811 TeE = FMA(KP923879532, Tel, Tek);
|
Chris@10
|
1812 Tem = FNMS(KP923879532, Tel, Tek);
|
Chris@10
|
1813 TdR = FNMS(KP820678790, TdQ, TdP);
|
Chris@10
|
1814 TdX = FMA(KP820678790, TdP, TdQ);
|
Chris@10
|
1815 TdI = FNMS(KP098491403, Tdd, Tdk);
|
Chris@10
|
1816 Tdl = FMA(KP098491403, Tdk, Tdd);
|
Chris@10
|
1817 Teo = Tdu - Tdr;
|
Chris@10
|
1818 Tdv = Tdr + Tdu;
|
Chris@10
|
1819 }
|
Chris@10
|
1820 }
|
Chris@10
|
1821 }
|
Chris@10
|
1822 {
|
Chris@10
|
1823 E TeB, Tef, TeF, Tep, TeC, Tei;
|
Chris@10
|
1824 TeF = FNMS(KP923879532, Teo, Ten);
|
Chris@10
|
1825 Tep = FMA(KP923879532, Teo, Ten);
|
Chris@10
|
1826 TeB = FMA(KP923879532, Tee, Ted);
|
Chris@10
|
1827 Tef = FNMS(KP923879532, Tee, Ted);
|
Chris@10
|
1828 TeG = FMA(KP303346683, TeF, TeE);
|
Chris@10
|
1829 TeK = FNMS(KP303346683, TeE, TeF);
|
Chris@10
|
1830 Tet = FMA(KP534511135, Tem, Tep);
|
Chris@10
|
1831 Teq = FNMS(KP534511135, Tep, Tem);
|
Chris@10
|
1832 TeC = FMA(KP923879532, Teh, Teg);
|
Chris@10
|
1833 Tei = FNMS(KP923879532, Teh, Teg);
|
Chris@10
|
1834 Tex = FNMS(KP668178637, Te5, Te6);
|
Chris@10
|
1835 Te7 = FMA(KP668178637, Te6, Te5);
|
Chris@10
|
1836 TeD = FNMS(KP303346683, TeC, TeB);
|
Chris@10
|
1837 TeJ = FMA(KP303346683, TeB, TeC);
|
Chris@10
|
1838 Teu = FNMS(KP534511135, Tef, Tei);
|
Chris@10
|
1839 Tej = FMA(KP534511135, Tei, Tef);
|
Chris@10
|
1840 }
|
Chris@10
|
1841 {
|
Chris@10
|
1842 E TdS, Tdw, Tey, Tea, TdT, TdD;
|
Chris@10
|
1843 Tey = FMA(KP668178637, Te8, Te9);
|
Chris@10
|
1844 Tea = FNMS(KP668178637, Te9, Te8);
|
Chris@10
|
1845 TdS = FNMS(KP923879532, Tdv, Tdo);
|
Chris@10
|
1846 Tdw = FMA(KP923879532, Tdv, Tdo);
|
Chris@10
|
1847 Tez = Tex + Tey;
|
Chris@10
|
1848 TkK = Tey - Tex;
|
Chris@10
|
1849 TkE = Te7 + Tea;
|
Chris@10
|
1850 Teb = Te7 - Tea;
|
Chris@10
|
1851 Te2 = FNMS(KP414213562, TcF, TcG);
|
Chris@10
|
1852 TcH = FMA(KP414213562, TcG, TcF);
|
Chris@10
|
1853 TdT = FNMS(KP923879532, TdC, Tdz);
|
Chris@10
|
1854 TdD = FMA(KP923879532, TdC, Tdz);
|
Chris@10
|
1855 Te0 = FNMS(KP707106781, TcD, TcA);
|
Chris@10
|
1856 TcE = FMA(KP707106781, TcD, TcA);
|
Chris@10
|
1857 TkB = FNMS(KP707106781, Tkm, Tkl);
|
Chris@10
|
1858 Tkn = FMA(KP707106781, Tkm, Tkl);
|
Chris@10
|
1859 TdU = FMA(KP820678790, TdT, TdS);
|
Chris@10
|
1860 TdY = FNMS(KP820678790, TdS, TdT);
|
Chris@10
|
1861 TdH = FMA(KP098491403, Tdw, TdD);
|
Chris@10
|
1862 TdE = FNMS(KP098491403, TdD, Tdw);
|
Chris@10
|
1863 TcK = FNMS(KP414213562, TcJ, TcI);
|
Chris@10
|
1864 Te1 = FMA(KP414213562, TcI, TcJ);
|
Chris@10
|
1865 }
|
Chris@10
|
1866 }
|
Chris@10
|
1867 {
|
Chris@10
|
1868 E Tkv, Tkp, Tew, Te4, TkH, TeL, TkL, Ter, Tks, TkA, TkI, TeH, TkM, Tev;
|
Chris@10
|
1869 {
|
Chris@10
|
1870 E TdG, TdO, TkF, TkG, TkN, TkO, TdJ, TdV;
|
Chris@10
|
1871 {
|
Chris@10
|
1872 E TdF, TkJ, TkD, Td2, TdW, TdZ, TdK;
|
Chris@10
|
1873 Tku = TdE - Tdl;
|
Chris@10
|
1874 TdF = Tdl + TdE;
|
Chris@10
|
1875 {
|
Chris@10
|
1876 E TkC, TcL, Tko, Te3, TcM;
|
Chris@10
|
1877 TkC = TcH - TcK;
|
Chris@10
|
1878 TcL = TcH + TcK;
|
Chris@10
|
1879 Tko = Te2 + Te1;
|
Chris@10
|
1880 Te3 = Te1 - Te2;
|
Chris@10
|
1881 TkJ = FNMS(KP923879532, TkC, TkB);
|
Chris@10
|
1882 TkD = FMA(KP923879532, TkC, TkB);
|
Chris@10
|
1883 TdK = FNMS(KP923879532, TcL, TcE);
|
Chris@10
|
1884 TcM = FMA(KP923879532, TcL, TcE);
|
Chris@10
|
1885 Tkv = FNMS(KP923879532, Tko, Tkn);
|
Chris@10
|
1886 Tkp = FMA(KP923879532, Tko, Tkn);
|
Chris@10
|
1887 Tew = FMA(KP923879532, Te3, Te0);
|
Chris@10
|
1888 Te4 = FNMS(KP923879532, Te3, Te0);
|
Chris@10
|
1889 Td2 = FMA(KP980785280, Td1, TcM);
|
Chris@10
|
1890 TdG = FNMS(KP980785280, Td1, TcM);
|
Chris@10
|
1891 }
|
Chris@10
|
1892 TdO = FMA(KP980785280, TdN, TdK);
|
Chris@10
|
1893 TdW = FNMS(KP980785280, TdN, TdK);
|
Chris@10
|
1894 TdZ = TdX - TdY;
|
Chris@10
|
1895 Tky = TdX + TdY;
|
Chris@10
|
1896 TkH = FNMS(KP831469612, TkE, TkD);
|
Chris@10
|
1897 TkF = FMA(KP831469612, TkE, TkD);
|
Chris@10
|
1898 cr[WS(rs, 1)] = FMA(KP995184726, TdF, Td2);
|
Chris@10
|
1899 ci[WS(rs, 30)] = FNMS(KP995184726, TdF, Td2);
|
Chris@10
|
1900 cr[WS(rs, 9)] = FMA(KP773010453, TdZ, TdW);
|
Chris@10
|
1901 ci[WS(rs, 22)] = FNMS(KP773010453, TdZ, TdW);
|
Chris@10
|
1902 TkG = TeJ + TeK;
|
Chris@10
|
1903 TeL = TeJ - TeK;
|
Chris@10
|
1904 TkL = FMA(KP831469612, TkK, TkJ);
|
Chris@10
|
1905 TkN = FNMS(KP831469612, TkK, TkJ);
|
Chris@10
|
1906 TkO = Teq - Tej;
|
Chris@10
|
1907 Ter = Tej + Teq;
|
Chris@10
|
1908 }
|
Chris@10
|
1909 ci[WS(rs, 34)] = FNMS(KP956940335, TkG, TkF);
|
Chris@10
|
1910 cr[WS(rs, 61)] = -(FMA(KP956940335, TkG, TkF));
|
Chris@10
|
1911 ci[WS(rs, 42)] = FMA(KP881921264, TkO, TkN);
|
Chris@10
|
1912 cr[WS(rs, 53)] = FMS(KP881921264, TkO, TkN);
|
Chris@10
|
1913 TdJ = TdH - TdI;
|
Chris@10
|
1914 Tks = TdI + TdH;
|
Chris@10
|
1915 TdV = TdR + TdU;
|
Chris@10
|
1916 TkA = TdU - TdR;
|
Chris@10
|
1917 ci[WS(rs, 14)] = FMA(KP995184726, TdJ, TdG);
|
Chris@10
|
1918 cr[WS(rs, 17)] = FNMS(KP995184726, TdJ, TdG);
|
Chris@10
|
1919 ci[WS(rs, 6)] = FMA(KP773010453, TdV, TdO);
|
Chris@10
|
1920 cr[WS(rs, 25)] = FNMS(KP773010453, TdV, TdO);
|
Chris@10
|
1921 TkI = TeG - TeD;
|
Chris@10
|
1922 TeH = TeD + TeG;
|
Chris@10
|
1923 TkM = Teu + Tet;
|
Chris@10
|
1924 Tev = Tet - Teu;
|
Chris@10
|
1925 }
|
Chris@10
|
1926 {
|
Chris@10
|
1927 E Tes, TeA, Tkr, Tkz, Tec, TeI;
|
Chris@10
|
1928 Tes = FNMS(KP831469612, Teb, Te4);
|
Chris@10
|
1929 Tec = FMA(KP831469612, Teb, Te4);
|
Chris@10
|
1930 ci[WS(rs, 50)] = FMA(KP956940335, TkI, TkH);
|
Chris@10
|
1931 cr[WS(rs, 45)] = FMS(KP956940335, TkI, TkH);
|
Chris@10
|
1932 ci[WS(rs, 58)] = FMA(KP881921264, TkM, TkL);
|
Chris@10
|
1933 cr[WS(rs, 37)] = FMS(KP881921264, TkM, TkL);
|
Chris@10
|
1934 TeA = FMA(KP831469612, Tez, Tew);
|
Chris@10
|
1935 TeI = FNMS(KP831469612, Tez, Tew);
|
Chris@10
|
1936 Tkt = FNMS(KP980785280, Tkq, Tkp);
|
Chris@10
|
1937 Tkr = FMA(KP980785280, Tkq, Tkp);
|
Chris@10
|
1938 cr[WS(rs, 5)] = FMA(KP881921264, Ter, Tec);
|
Chris@10
|
1939 ci[WS(rs, 26)] = FNMS(KP881921264, Ter, Tec);
|
Chris@10
|
1940 cr[WS(rs, 13)] = FMA(KP956940335, TeL, TeI);
|
Chris@10
|
1941 ci[WS(rs, 18)] = FNMS(KP956940335, TeL, TeI);
|
Chris@10
|
1942 Tkx = FMA(KP980785280, Tkw, Tkv);
|
Chris@10
|
1943 Tkz = FNMS(KP980785280, Tkw, Tkv);
|
Chris@10
|
1944 ci[WS(rs, 62)] = FMA(KP995184726, Tks, Tkr);
|
Chris@10
|
1945 cr[WS(rs, 33)] = FMS(KP995184726, Tks, Tkr);
|
Chris@10
|
1946 ci[WS(rs, 54)] = FMA(KP773010453, TkA, Tkz);
|
Chris@10
|
1947 cr[WS(rs, 41)] = FMS(KP773010453, TkA, Tkz);
|
Chris@10
|
1948 ci[WS(rs, 10)] = FMA(KP881921264, Tev, Tes);
|
Chris@10
|
1949 cr[WS(rs, 21)] = FNMS(KP881921264, Tev, Tes);
|
Chris@10
|
1950 ci[WS(rs, 2)] = FMA(KP956940335, TeH, TeA);
|
Chris@10
|
1951 cr[WS(rs, 29)] = FNMS(KP956940335, TeH, TeA);
|
Chris@10
|
1952 }
|
Chris@10
|
1953 }
|
Chris@10
|
1954 }
|
Chris@10
|
1955 }
|
Chris@10
|
1956 }
|
Chris@10
|
1957 }
|
Chris@10
|
1958 ci[WS(rs, 46)] = FMA(KP995184726, Tku, Tkt);
|
Chris@10
|
1959 cr[WS(rs, 49)] = FMS(KP995184726, Tku, Tkt);
|
Chris@10
|
1960 ci[WS(rs, 38)] = FNMS(KP773010453, Tky, Tkx);
|
Chris@10
|
1961 cr[WS(rs, 57)] = -(FMA(KP773010453, Tky, Tkx));
|
Chris@10
|
1962 }
|
Chris@10
|
1963 }
|
Chris@10
|
1964 }
|
Chris@10
|
1965
|
Chris@10
|
1966 static const tw_instr twinstr[] = {
|
Chris@10
|
1967 {TW_FULL, 1, 64},
|
Chris@10
|
1968 {TW_NEXT, 1, 0}
|
Chris@10
|
1969 };
|
Chris@10
|
1970
|
Chris@10
|
1971 static const hc2hc_desc desc = { 64, "hf_64", twinstr, &GENUS, {520, 126, 518, 0} };
|
Chris@10
|
1972
|
Chris@10
|
1973 void X(codelet_hf_64) (planner *p) {
|
Chris@10
|
1974 X(khc2hc_register) (p, hf_64, &desc);
|
Chris@10
|
1975 }
|
Chris@10
|
1976 #else /* HAVE_FMA */
|
Chris@10
|
1977
|
Chris@10
|
1978 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 64 -dit -name hf_64 -include hf.h */
|
Chris@10
|
1979
|
Chris@10
|
1980 /*
|
Chris@10
|
1981 * This function contains 1038 FP additions, 500 FP multiplications,
|
Chris@10
|
1982 * (or, 808 additions, 270 multiplications, 230 fused multiply/add),
|
Chris@10
|
1983 * 176 stack variables, 15 constants, and 256 memory accesses
|
Chris@10
|
1984 */
|
Chris@10
|
1985 #include "hf.h"
|
Chris@10
|
1986
|
Chris@10
|
1987 static void hf_64(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
1988 {
|
Chris@10
|
1989 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
Chris@10
|
1990 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@10
|
1991 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@10
|
1992 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
Chris@10
|
1993 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@10
|
1994 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
1995 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
Chris@10
|
1996 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@10
|
1997 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@10
|
1998 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
Chris@10
|
1999 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
2000 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@10
|
2001 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
2002 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
2003 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
2004 {
|
Chris@10
|
2005 INT m;
|
Chris@10
|
2006 for (m = mb, W = W + ((mb - 1) * 126); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 126, MAKE_VOLATILE_STRIDE(128, rs)) {
|
Chris@10
|
2007 E Tj, TcL, ThT, Tin, T6b, Taz, TgT, Thn, TG, Thm, TcO, TgO, T6m, Tim, TaC;
|
Chris@10
|
2008 E ThQ, T14, Tfr, T6y, T9O, TaG, Tc0, TcU, TeE, T1r, Tfq, T6J, T9P, TaJ, Tc1;
|
Chris@10
|
2009 E TcZ, TeF, T1Q, T2d, Tfu, Tfv, Tfw, Tfx, T6Q, TaM, Tdb, TeI, T71, TaQ, T7a;
|
Chris@10
|
2010 E TaN, Td6, TeJ, T77, TaP, T2B, T2Y, Tfz, TfA, TfB, TfC, T7h, TaW, Tdm, TeL;
|
Chris@10
|
2011 E T7s, TaU, T7B, TaX, Tdh, TeM, T7y, TaT, T5j, TfR, Tec, TeX, TfY, Tgy, T8D;
|
Chris@10
|
2012 E Tbl, T8O, Tbx, T9l, Tbm, TdV, Tf0, T9i, Tbw, T3M, TfL, TdL, TeT, TfI, Tgt;
|
Chris@10
|
2013 E T7K, Tbd, T7V, Tb3, T8s, Tbe, Tdu, TeQ, T8p, Tb2, T4x, TfJ, TdE, TdM, TfO;
|
Chris@10
|
2014 E Tgu, T87, T8u, T8i, T8v, Tba, Tbh, Tdz, TdN, Tb7, Tbg, T64, TfZ, Te5, Ted;
|
Chris@10
|
2015 E TfU, Tgz, T90, T9n, T9b, T9o, Tbt, TbA, Te0, Tee, Tbq, Tbz;
|
Chris@10
|
2016 {
|
Chris@10
|
2017 E T1, TgR, T6, TgQ, Tc, T68, Th, T69;
|
Chris@10
|
2018 T1 = cr[0];
|
Chris@10
|
2019 TgR = ci[0];
|
Chris@10
|
2020 {
|
Chris@10
|
2021 E T3, T5, T2, T4;
|
Chris@10
|
2022 T3 = cr[WS(rs, 32)];
|
Chris@10
|
2023 T5 = ci[WS(rs, 32)];
|
Chris@10
|
2024 T2 = W[62];
|
Chris@10
|
2025 T4 = W[63];
|
Chris@10
|
2026 T6 = FMA(T2, T3, T4 * T5);
|
Chris@10
|
2027 TgQ = FNMS(T4, T3, T2 * T5);
|
Chris@10
|
2028 }
|
Chris@10
|
2029 {
|
Chris@10
|
2030 E T9, Tb, T8, Ta;
|
Chris@10
|
2031 T9 = cr[WS(rs, 16)];
|
Chris@10
|
2032 Tb = ci[WS(rs, 16)];
|
Chris@10
|
2033 T8 = W[30];
|
Chris@10
|
2034 Ta = W[31];
|
Chris@10
|
2035 Tc = FMA(T8, T9, Ta * Tb);
|
Chris@10
|
2036 T68 = FNMS(Ta, T9, T8 * Tb);
|
Chris@10
|
2037 }
|
Chris@10
|
2038 {
|
Chris@10
|
2039 E Te, Tg, Td, Tf;
|
Chris@10
|
2040 Te = cr[WS(rs, 48)];
|
Chris@10
|
2041 Tg = ci[WS(rs, 48)];
|
Chris@10
|
2042 Td = W[94];
|
Chris@10
|
2043 Tf = W[95];
|
Chris@10
|
2044 Th = FMA(Td, Te, Tf * Tg);
|
Chris@10
|
2045 T69 = FNMS(Tf, Te, Td * Tg);
|
Chris@10
|
2046 }
|
Chris@10
|
2047 {
|
Chris@10
|
2048 E T7, Ti, ThR, ThS;
|
Chris@10
|
2049 T7 = T1 + T6;
|
Chris@10
|
2050 Ti = Tc + Th;
|
Chris@10
|
2051 Tj = T7 + Ti;
|
Chris@10
|
2052 TcL = T7 - Ti;
|
Chris@10
|
2053 ThR = Tc - Th;
|
Chris@10
|
2054 ThS = TgR - TgQ;
|
Chris@10
|
2055 ThT = ThR + ThS;
|
Chris@10
|
2056 Tin = ThS - ThR;
|
Chris@10
|
2057 }
|
Chris@10
|
2058 {
|
Chris@10
|
2059 E T67, T6a, TgP, TgS;
|
Chris@10
|
2060 T67 = T1 - T6;
|
Chris@10
|
2061 T6a = T68 - T69;
|
Chris@10
|
2062 T6b = T67 - T6a;
|
Chris@10
|
2063 Taz = T67 + T6a;
|
Chris@10
|
2064 TgP = T68 + T69;
|
Chris@10
|
2065 TgS = TgQ + TgR;
|
Chris@10
|
2066 TgT = TgP + TgS;
|
Chris@10
|
2067 Thn = TgS - TgP;
|
Chris@10
|
2068 }
|
Chris@10
|
2069 }
|
Chris@10
|
2070 {
|
Chris@10
|
2071 E To, T6d, Tt, T6e, T6c, T6f, Tz, T6i, TE, T6j, T6h, T6k;
|
Chris@10
|
2072 {
|
Chris@10
|
2073 E Tl, Tn, Tk, Tm;
|
Chris@10
|
2074 Tl = cr[WS(rs, 8)];
|
Chris@10
|
2075 Tn = ci[WS(rs, 8)];
|
Chris@10
|
2076 Tk = W[14];
|
Chris@10
|
2077 Tm = W[15];
|
Chris@10
|
2078 To = FMA(Tk, Tl, Tm * Tn);
|
Chris@10
|
2079 T6d = FNMS(Tm, Tl, Tk * Tn);
|
Chris@10
|
2080 }
|
Chris@10
|
2081 {
|
Chris@10
|
2082 E Tq, Ts, Tp, Tr;
|
Chris@10
|
2083 Tq = cr[WS(rs, 40)];
|
Chris@10
|
2084 Ts = ci[WS(rs, 40)];
|
Chris@10
|
2085 Tp = W[78];
|
Chris@10
|
2086 Tr = W[79];
|
Chris@10
|
2087 Tt = FMA(Tp, Tq, Tr * Ts);
|
Chris@10
|
2088 T6e = FNMS(Tr, Tq, Tp * Ts);
|
Chris@10
|
2089 }
|
Chris@10
|
2090 T6c = To - Tt;
|
Chris@10
|
2091 T6f = T6d - T6e;
|
Chris@10
|
2092 {
|
Chris@10
|
2093 E Tw, Ty, Tv, Tx;
|
Chris@10
|
2094 Tw = cr[WS(rs, 56)];
|
Chris@10
|
2095 Ty = ci[WS(rs, 56)];
|
Chris@10
|
2096 Tv = W[110];
|
Chris@10
|
2097 Tx = W[111];
|
Chris@10
|
2098 Tz = FMA(Tv, Tw, Tx * Ty);
|
Chris@10
|
2099 T6i = FNMS(Tx, Tw, Tv * Ty);
|
Chris@10
|
2100 }
|
Chris@10
|
2101 {
|
Chris@10
|
2102 E TB, TD, TA, TC;
|
Chris@10
|
2103 TB = cr[WS(rs, 24)];
|
Chris@10
|
2104 TD = ci[WS(rs, 24)];
|
Chris@10
|
2105 TA = W[46];
|
Chris@10
|
2106 TC = W[47];
|
Chris@10
|
2107 TE = FMA(TA, TB, TC * TD);
|
Chris@10
|
2108 T6j = FNMS(TC, TB, TA * TD);
|
Chris@10
|
2109 }
|
Chris@10
|
2110 T6h = Tz - TE;
|
Chris@10
|
2111 T6k = T6i - T6j;
|
Chris@10
|
2112 {
|
Chris@10
|
2113 E Tu, TF, TcM, TcN;
|
Chris@10
|
2114 Tu = To + Tt;
|
Chris@10
|
2115 TF = Tz + TE;
|
Chris@10
|
2116 TG = Tu + TF;
|
Chris@10
|
2117 Thm = Tu - TF;
|
Chris@10
|
2118 TcM = T6i + T6j;
|
Chris@10
|
2119 TcN = T6d + T6e;
|
Chris@10
|
2120 TcO = TcM - TcN;
|
Chris@10
|
2121 TgO = TcN + TcM;
|
Chris@10
|
2122 }
|
Chris@10
|
2123 {
|
Chris@10
|
2124 E T6g, T6l, TaA, TaB;
|
Chris@10
|
2125 T6g = T6c - T6f;
|
Chris@10
|
2126 T6l = T6h + T6k;
|
Chris@10
|
2127 T6m = KP707106781 * (T6g + T6l);
|
Chris@10
|
2128 Tim = KP707106781 * (T6l - T6g);
|
Chris@10
|
2129 TaA = T6c + T6f;
|
Chris@10
|
2130 TaB = T6h - T6k;
|
Chris@10
|
2131 TaC = KP707106781 * (TaA + TaB);
|
Chris@10
|
2132 ThQ = KP707106781 * (TaA - TaB);
|
Chris@10
|
2133 }
|
Chris@10
|
2134 }
|
Chris@10
|
2135 {
|
Chris@10
|
2136 E TS, TcR, T6o, T6v, T13, TcS, T6r, T6w, T6s, T6x;
|
Chris@10
|
2137 {
|
Chris@10
|
2138 E TM, T6t, TR, T6u;
|
Chris@10
|
2139 {
|
Chris@10
|
2140 E TJ, TL, TI, TK;
|
Chris@10
|
2141 TJ = cr[WS(rs, 4)];
|
Chris@10
|
2142 TL = ci[WS(rs, 4)];
|
Chris@10
|
2143 TI = W[6];
|
Chris@10
|
2144 TK = W[7];
|
Chris@10
|
2145 TM = FMA(TI, TJ, TK * TL);
|
Chris@10
|
2146 T6t = FNMS(TK, TJ, TI * TL);
|
Chris@10
|
2147 }
|
Chris@10
|
2148 {
|
Chris@10
|
2149 E TO, TQ, TN, TP;
|
Chris@10
|
2150 TO = cr[WS(rs, 36)];
|
Chris@10
|
2151 TQ = ci[WS(rs, 36)];
|
Chris@10
|
2152 TN = W[70];
|
Chris@10
|
2153 TP = W[71];
|
Chris@10
|
2154 TR = FMA(TN, TO, TP * TQ);
|
Chris@10
|
2155 T6u = FNMS(TP, TO, TN * TQ);
|
Chris@10
|
2156 }
|
Chris@10
|
2157 TS = TM + TR;
|
Chris@10
|
2158 TcR = T6t + T6u;
|
Chris@10
|
2159 T6o = TM - TR;
|
Chris@10
|
2160 T6v = T6t - T6u;
|
Chris@10
|
2161 }
|
Chris@10
|
2162 {
|
Chris@10
|
2163 E TX, T6p, T12, T6q;
|
Chris@10
|
2164 {
|
Chris@10
|
2165 E TU, TW, TT, TV;
|
Chris@10
|
2166 TU = cr[WS(rs, 20)];
|
Chris@10
|
2167 TW = ci[WS(rs, 20)];
|
Chris@10
|
2168 TT = W[38];
|
Chris@10
|
2169 TV = W[39];
|
Chris@10
|
2170 TX = FMA(TT, TU, TV * TW);
|
Chris@10
|
2171 T6p = FNMS(TV, TU, TT * TW);
|
Chris@10
|
2172 }
|
Chris@10
|
2173 {
|
Chris@10
|
2174 E TZ, T11, TY, T10;
|
Chris@10
|
2175 TZ = cr[WS(rs, 52)];
|
Chris@10
|
2176 T11 = ci[WS(rs, 52)];
|
Chris@10
|
2177 TY = W[102];
|
Chris@10
|
2178 T10 = W[103];
|
Chris@10
|
2179 T12 = FMA(TY, TZ, T10 * T11);
|
Chris@10
|
2180 T6q = FNMS(T10, TZ, TY * T11);
|
Chris@10
|
2181 }
|
Chris@10
|
2182 T13 = TX + T12;
|
Chris@10
|
2183 TcS = T6p + T6q;
|
Chris@10
|
2184 T6r = T6p - T6q;
|
Chris@10
|
2185 T6w = TX - T12;
|
Chris@10
|
2186 }
|
Chris@10
|
2187 T14 = TS + T13;
|
Chris@10
|
2188 Tfr = TcR + TcS;
|
Chris@10
|
2189 T6s = T6o - T6r;
|
Chris@10
|
2190 T6x = T6v + T6w;
|
Chris@10
|
2191 T6y = FNMS(KP382683432, T6x, KP923879532 * T6s);
|
Chris@10
|
2192 T9O = FMA(KP923879532, T6x, KP382683432 * T6s);
|
Chris@10
|
2193 {
|
Chris@10
|
2194 E TaE, TaF, TcQ, TcT;
|
Chris@10
|
2195 TaE = T6v - T6w;
|
Chris@10
|
2196 TaF = T6o + T6r;
|
Chris@10
|
2197 TaG = FMA(KP382683432, TaE, KP923879532 * TaF);
|
Chris@10
|
2198 Tc0 = FNMS(KP923879532, TaE, KP382683432 * TaF);
|
Chris@10
|
2199 TcQ = TS - T13;
|
Chris@10
|
2200 TcT = TcR - TcS;
|
Chris@10
|
2201 TcU = TcQ + TcT;
|
Chris@10
|
2202 TeE = TcQ - TcT;
|
Chris@10
|
2203 }
|
Chris@10
|
2204 }
|
Chris@10
|
2205 {
|
Chris@10
|
2206 E T1f, TcW, T6B, T6E, T1q, TcX, T6C, T6H, T6D, T6I;
|
Chris@10
|
2207 {
|
Chris@10
|
2208 E T19, T6z, T1e, T6A;
|
Chris@10
|
2209 {
|
Chris@10
|
2210 E T16, T18, T15, T17;
|
Chris@10
|
2211 T16 = cr[WS(rs, 60)];
|
Chris@10
|
2212 T18 = ci[WS(rs, 60)];
|
Chris@10
|
2213 T15 = W[118];
|
Chris@10
|
2214 T17 = W[119];
|
Chris@10
|
2215 T19 = FMA(T15, T16, T17 * T18);
|
Chris@10
|
2216 T6z = FNMS(T17, T16, T15 * T18);
|
Chris@10
|
2217 }
|
Chris@10
|
2218 {
|
Chris@10
|
2219 E T1b, T1d, T1a, T1c;
|
Chris@10
|
2220 T1b = cr[WS(rs, 28)];
|
Chris@10
|
2221 T1d = ci[WS(rs, 28)];
|
Chris@10
|
2222 T1a = W[54];
|
Chris@10
|
2223 T1c = W[55];
|
Chris@10
|
2224 T1e = FMA(T1a, T1b, T1c * T1d);
|
Chris@10
|
2225 T6A = FNMS(T1c, T1b, T1a * T1d);
|
Chris@10
|
2226 }
|
Chris@10
|
2227 T1f = T19 + T1e;
|
Chris@10
|
2228 TcW = T6z + T6A;
|
Chris@10
|
2229 T6B = T6z - T6A;
|
Chris@10
|
2230 T6E = T19 - T1e;
|
Chris@10
|
2231 }
|
Chris@10
|
2232 {
|
Chris@10
|
2233 E T1k, T6F, T1p, T6G;
|
Chris@10
|
2234 {
|
Chris@10
|
2235 E T1h, T1j, T1g, T1i;
|
Chris@10
|
2236 T1h = cr[WS(rs, 12)];
|
Chris@10
|
2237 T1j = ci[WS(rs, 12)];
|
Chris@10
|
2238 T1g = W[22];
|
Chris@10
|
2239 T1i = W[23];
|
Chris@10
|
2240 T1k = FMA(T1g, T1h, T1i * T1j);
|
Chris@10
|
2241 T6F = FNMS(T1i, T1h, T1g * T1j);
|
Chris@10
|
2242 }
|
Chris@10
|
2243 {
|
Chris@10
|
2244 E T1m, T1o, T1l, T1n;
|
Chris@10
|
2245 T1m = cr[WS(rs, 44)];
|
Chris@10
|
2246 T1o = ci[WS(rs, 44)];
|
Chris@10
|
2247 T1l = W[86];
|
Chris@10
|
2248 T1n = W[87];
|
Chris@10
|
2249 T1p = FMA(T1l, T1m, T1n * T1o);
|
Chris@10
|
2250 T6G = FNMS(T1n, T1m, T1l * T1o);
|
Chris@10
|
2251 }
|
Chris@10
|
2252 T1q = T1k + T1p;
|
Chris@10
|
2253 TcX = T6F + T6G;
|
Chris@10
|
2254 T6C = T1k - T1p;
|
Chris@10
|
2255 T6H = T6F - T6G;
|
Chris@10
|
2256 }
|
Chris@10
|
2257 T1r = T1f + T1q;
|
Chris@10
|
2258 Tfq = TcW + TcX;
|
Chris@10
|
2259 T6D = T6B + T6C;
|
Chris@10
|
2260 T6I = T6E - T6H;
|
Chris@10
|
2261 T6J = FMA(KP382683432, T6D, KP923879532 * T6I);
|
Chris@10
|
2262 T9P = FNMS(KP923879532, T6D, KP382683432 * T6I);
|
Chris@10
|
2263 {
|
Chris@10
|
2264 E TaH, TaI, TcV, TcY;
|
Chris@10
|
2265 TaH = T6E + T6H;
|
Chris@10
|
2266 TaI = T6B - T6C;
|
Chris@10
|
2267 TaJ = FNMS(KP382683432, TaI, KP923879532 * TaH);
|
Chris@10
|
2268 Tc1 = FMA(KP923879532, TaI, KP382683432 * TaH);
|
Chris@10
|
2269 TcV = T1f - T1q;
|
Chris@10
|
2270 TcY = TcW - TcX;
|
Chris@10
|
2271 TcZ = TcV - TcY;
|
Chris@10
|
2272 TeF = TcV + TcY;
|
Chris@10
|
2273 }
|
Chris@10
|
2274 }
|
Chris@10
|
2275 {
|
Chris@10
|
2276 E T1y, T73, T1D, T74, T1E, Td7, T1J, T6N, T1O, T6O, T1P, Td8, T21, Td4, T6R;
|
Chris@10
|
2277 E T6U, T2c, Td3, T6W, T6Z;
|
Chris@10
|
2278 {
|
Chris@10
|
2279 E T1v, T1x, T1u, T1w;
|
Chris@10
|
2280 T1v = cr[WS(rs, 2)];
|
Chris@10
|
2281 T1x = ci[WS(rs, 2)];
|
Chris@10
|
2282 T1u = W[2];
|
Chris@10
|
2283 T1w = W[3];
|
Chris@10
|
2284 T1y = FMA(T1u, T1v, T1w * T1x);
|
Chris@10
|
2285 T73 = FNMS(T1w, T1v, T1u * T1x);
|
Chris@10
|
2286 }
|
Chris@10
|
2287 {
|
Chris@10
|
2288 E T1A, T1C, T1z, T1B;
|
Chris@10
|
2289 T1A = cr[WS(rs, 34)];
|
Chris@10
|
2290 T1C = ci[WS(rs, 34)];
|
Chris@10
|
2291 T1z = W[66];
|
Chris@10
|
2292 T1B = W[67];
|
Chris@10
|
2293 T1D = FMA(T1z, T1A, T1B * T1C);
|
Chris@10
|
2294 T74 = FNMS(T1B, T1A, T1z * T1C);
|
Chris@10
|
2295 }
|
Chris@10
|
2296 T1E = T1y + T1D;
|
Chris@10
|
2297 Td7 = T73 + T74;
|
Chris@10
|
2298 {
|
Chris@10
|
2299 E T1G, T1I, T1F, T1H;
|
Chris@10
|
2300 T1G = cr[WS(rs, 18)];
|
Chris@10
|
2301 T1I = ci[WS(rs, 18)];
|
Chris@10
|
2302 T1F = W[34];
|
Chris@10
|
2303 T1H = W[35];
|
Chris@10
|
2304 T1J = FMA(T1F, T1G, T1H * T1I);
|
Chris@10
|
2305 T6N = FNMS(T1H, T1G, T1F * T1I);
|
Chris@10
|
2306 }
|
Chris@10
|
2307 {
|
Chris@10
|
2308 E T1L, T1N, T1K, T1M;
|
Chris@10
|
2309 T1L = cr[WS(rs, 50)];
|
Chris@10
|
2310 T1N = ci[WS(rs, 50)];
|
Chris@10
|
2311 T1K = W[98];
|
Chris@10
|
2312 T1M = W[99];
|
Chris@10
|
2313 T1O = FMA(T1K, T1L, T1M * T1N);
|
Chris@10
|
2314 T6O = FNMS(T1M, T1L, T1K * T1N);
|
Chris@10
|
2315 }
|
Chris@10
|
2316 T1P = T1J + T1O;
|
Chris@10
|
2317 Td8 = T6N + T6O;
|
Chris@10
|
2318 {
|
Chris@10
|
2319 E T1V, T6S, T20, T6T;
|
Chris@10
|
2320 {
|
Chris@10
|
2321 E T1S, T1U, T1R, T1T;
|
Chris@10
|
2322 T1S = cr[WS(rs, 10)];
|
Chris@10
|
2323 T1U = ci[WS(rs, 10)];
|
Chris@10
|
2324 T1R = W[18];
|
Chris@10
|
2325 T1T = W[19];
|
Chris@10
|
2326 T1V = FMA(T1R, T1S, T1T * T1U);
|
Chris@10
|
2327 T6S = FNMS(T1T, T1S, T1R * T1U);
|
Chris@10
|
2328 }
|
Chris@10
|
2329 {
|
Chris@10
|
2330 E T1X, T1Z, T1W, T1Y;
|
Chris@10
|
2331 T1X = cr[WS(rs, 42)];
|
Chris@10
|
2332 T1Z = ci[WS(rs, 42)];
|
Chris@10
|
2333 T1W = W[82];
|
Chris@10
|
2334 T1Y = W[83];
|
Chris@10
|
2335 T20 = FMA(T1W, T1X, T1Y * T1Z);
|
Chris@10
|
2336 T6T = FNMS(T1Y, T1X, T1W * T1Z);
|
Chris@10
|
2337 }
|
Chris@10
|
2338 T21 = T1V + T20;
|
Chris@10
|
2339 Td4 = T6S + T6T;
|
Chris@10
|
2340 T6R = T1V - T20;
|
Chris@10
|
2341 T6U = T6S - T6T;
|
Chris@10
|
2342 }
|
Chris@10
|
2343 {
|
Chris@10
|
2344 E T26, T6X, T2b, T6Y;
|
Chris@10
|
2345 {
|
Chris@10
|
2346 E T23, T25, T22, T24;
|
Chris@10
|
2347 T23 = cr[WS(rs, 58)];
|
Chris@10
|
2348 T25 = ci[WS(rs, 58)];
|
Chris@10
|
2349 T22 = W[114];
|
Chris@10
|
2350 T24 = W[115];
|
Chris@10
|
2351 T26 = FMA(T22, T23, T24 * T25);
|
Chris@10
|
2352 T6X = FNMS(T24, T23, T22 * T25);
|
Chris@10
|
2353 }
|
Chris@10
|
2354 {
|
Chris@10
|
2355 E T28, T2a, T27, T29;
|
Chris@10
|
2356 T28 = cr[WS(rs, 26)];
|
Chris@10
|
2357 T2a = ci[WS(rs, 26)];
|
Chris@10
|
2358 T27 = W[50];
|
Chris@10
|
2359 T29 = W[51];
|
Chris@10
|
2360 T2b = FMA(T27, T28, T29 * T2a);
|
Chris@10
|
2361 T6Y = FNMS(T29, T28, T27 * T2a);
|
Chris@10
|
2362 }
|
Chris@10
|
2363 T2c = T26 + T2b;
|
Chris@10
|
2364 Td3 = T6X + T6Y;
|
Chris@10
|
2365 T6W = T26 - T2b;
|
Chris@10
|
2366 T6Z = T6X - T6Y;
|
Chris@10
|
2367 }
|
Chris@10
|
2368 T1Q = T1E + T1P;
|
Chris@10
|
2369 T2d = T21 + T2c;
|
Chris@10
|
2370 Tfu = T1Q - T2d;
|
Chris@10
|
2371 Tfv = Td7 + Td8;
|
Chris@10
|
2372 Tfw = Td4 + Td3;
|
Chris@10
|
2373 Tfx = Tfv - Tfw;
|
Chris@10
|
2374 {
|
Chris@10
|
2375 E T6M, T6P, Td9, Tda;
|
Chris@10
|
2376 T6M = T1y - T1D;
|
Chris@10
|
2377 T6P = T6N - T6O;
|
Chris@10
|
2378 T6Q = T6M - T6P;
|
Chris@10
|
2379 TaM = T6M + T6P;
|
Chris@10
|
2380 Td9 = Td7 - Td8;
|
Chris@10
|
2381 Tda = T21 - T2c;
|
Chris@10
|
2382 Tdb = Td9 - Tda;
|
Chris@10
|
2383 TeI = Td9 + Tda;
|
Chris@10
|
2384 }
|
Chris@10
|
2385 {
|
Chris@10
|
2386 E T6V, T70, T78, T79;
|
Chris@10
|
2387 T6V = T6R - T6U;
|
Chris@10
|
2388 T70 = T6W + T6Z;
|
Chris@10
|
2389 T71 = KP707106781 * (T6V + T70);
|
Chris@10
|
2390 TaQ = KP707106781 * (T70 - T6V);
|
Chris@10
|
2391 T78 = T6R + T6U;
|
Chris@10
|
2392 T79 = T6Z - T6W;
|
Chris@10
|
2393 T7a = KP707106781 * (T78 + T79);
|
Chris@10
|
2394 TaN = KP707106781 * (T78 - T79);
|
Chris@10
|
2395 }
|
Chris@10
|
2396 {
|
Chris@10
|
2397 E Td2, Td5, T75, T76;
|
Chris@10
|
2398 Td2 = T1E - T1P;
|
Chris@10
|
2399 Td5 = Td3 - Td4;
|
Chris@10
|
2400 Td6 = Td2 - Td5;
|
Chris@10
|
2401 TeJ = Td2 + Td5;
|
Chris@10
|
2402 T75 = T73 - T74;
|
Chris@10
|
2403 T76 = T1J - T1O;
|
Chris@10
|
2404 T77 = T75 + T76;
|
Chris@10
|
2405 TaP = T75 - T76;
|
Chris@10
|
2406 }
|
Chris@10
|
2407 }
|
Chris@10
|
2408 {
|
Chris@10
|
2409 E T2j, T7u, T2o, T7v, T2p, Tdd, T2u, T7e, T2z, T7f, T2A, Tde, T2M, Tdk, T7i;
|
Chris@10
|
2410 E T7l, T2X, Tdj, T7n, T7q;
|
Chris@10
|
2411 {
|
Chris@10
|
2412 E T2g, T2i, T2f, T2h;
|
Chris@10
|
2413 T2g = cr[WS(rs, 62)];
|
Chris@10
|
2414 T2i = ci[WS(rs, 62)];
|
Chris@10
|
2415 T2f = W[122];
|
Chris@10
|
2416 T2h = W[123];
|
Chris@10
|
2417 T2j = FMA(T2f, T2g, T2h * T2i);
|
Chris@10
|
2418 T7u = FNMS(T2h, T2g, T2f * T2i);
|
Chris@10
|
2419 }
|
Chris@10
|
2420 {
|
Chris@10
|
2421 E T2l, T2n, T2k, T2m;
|
Chris@10
|
2422 T2l = cr[WS(rs, 30)];
|
Chris@10
|
2423 T2n = ci[WS(rs, 30)];
|
Chris@10
|
2424 T2k = W[58];
|
Chris@10
|
2425 T2m = W[59];
|
Chris@10
|
2426 T2o = FMA(T2k, T2l, T2m * T2n);
|
Chris@10
|
2427 T7v = FNMS(T2m, T2l, T2k * T2n);
|
Chris@10
|
2428 }
|
Chris@10
|
2429 T2p = T2j + T2o;
|
Chris@10
|
2430 Tdd = T7u + T7v;
|
Chris@10
|
2431 {
|
Chris@10
|
2432 E T2r, T2t, T2q, T2s;
|
Chris@10
|
2433 T2r = cr[WS(rs, 14)];
|
Chris@10
|
2434 T2t = ci[WS(rs, 14)];
|
Chris@10
|
2435 T2q = W[26];
|
Chris@10
|
2436 T2s = W[27];
|
Chris@10
|
2437 T2u = FMA(T2q, T2r, T2s * T2t);
|
Chris@10
|
2438 T7e = FNMS(T2s, T2r, T2q * T2t);
|
Chris@10
|
2439 }
|
Chris@10
|
2440 {
|
Chris@10
|
2441 E T2w, T2y, T2v, T2x;
|
Chris@10
|
2442 T2w = cr[WS(rs, 46)];
|
Chris@10
|
2443 T2y = ci[WS(rs, 46)];
|
Chris@10
|
2444 T2v = W[90];
|
Chris@10
|
2445 T2x = W[91];
|
Chris@10
|
2446 T2z = FMA(T2v, T2w, T2x * T2y);
|
Chris@10
|
2447 T7f = FNMS(T2x, T2w, T2v * T2y);
|
Chris@10
|
2448 }
|
Chris@10
|
2449 T2A = T2u + T2z;
|
Chris@10
|
2450 Tde = T7e + T7f;
|
Chris@10
|
2451 {
|
Chris@10
|
2452 E T2G, T7j, T2L, T7k;
|
Chris@10
|
2453 {
|
Chris@10
|
2454 E T2D, T2F, T2C, T2E;
|
Chris@10
|
2455 T2D = cr[WS(rs, 6)];
|
Chris@10
|
2456 T2F = ci[WS(rs, 6)];
|
Chris@10
|
2457 T2C = W[10];
|
Chris@10
|
2458 T2E = W[11];
|
Chris@10
|
2459 T2G = FMA(T2C, T2D, T2E * T2F);
|
Chris@10
|
2460 T7j = FNMS(T2E, T2D, T2C * T2F);
|
Chris@10
|
2461 }
|
Chris@10
|
2462 {
|
Chris@10
|
2463 E T2I, T2K, T2H, T2J;
|
Chris@10
|
2464 T2I = cr[WS(rs, 38)];
|
Chris@10
|
2465 T2K = ci[WS(rs, 38)];
|
Chris@10
|
2466 T2H = W[74];
|
Chris@10
|
2467 T2J = W[75];
|
Chris@10
|
2468 T2L = FMA(T2H, T2I, T2J * T2K);
|
Chris@10
|
2469 T7k = FNMS(T2J, T2I, T2H * T2K);
|
Chris@10
|
2470 }
|
Chris@10
|
2471 T2M = T2G + T2L;
|
Chris@10
|
2472 Tdk = T7j + T7k;
|
Chris@10
|
2473 T7i = T2G - T2L;
|
Chris@10
|
2474 T7l = T7j - T7k;
|
Chris@10
|
2475 }
|
Chris@10
|
2476 {
|
Chris@10
|
2477 E T2R, T7o, T2W, T7p;
|
Chris@10
|
2478 {
|
Chris@10
|
2479 E T2O, T2Q, T2N, T2P;
|
Chris@10
|
2480 T2O = cr[WS(rs, 54)];
|
Chris@10
|
2481 T2Q = ci[WS(rs, 54)];
|
Chris@10
|
2482 T2N = W[106];
|
Chris@10
|
2483 T2P = W[107];
|
Chris@10
|
2484 T2R = FMA(T2N, T2O, T2P * T2Q);
|
Chris@10
|
2485 T7o = FNMS(T2P, T2O, T2N * T2Q);
|
Chris@10
|
2486 }
|
Chris@10
|
2487 {
|
Chris@10
|
2488 E T2T, T2V, T2S, T2U;
|
Chris@10
|
2489 T2T = cr[WS(rs, 22)];
|
Chris@10
|
2490 T2V = ci[WS(rs, 22)];
|
Chris@10
|
2491 T2S = W[42];
|
Chris@10
|
2492 T2U = W[43];
|
Chris@10
|
2493 T2W = FMA(T2S, T2T, T2U * T2V);
|
Chris@10
|
2494 T7p = FNMS(T2U, T2T, T2S * T2V);
|
Chris@10
|
2495 }
|
Chris@10
|
2496 T2X = T2R + T2W;
|
Chris@10
|
2497 Tdj = T7o + T7p;
|
Chris@10
|
2498 T7n = T2R - T2W;
|
Chris@10
|
2499 T7q = T7o - T7p;
|
Chris@10
|
2500 }
|
Chris@10
|
2501 T2B = T2p + T2A;
|
Chris@10
|
2502 T2Y = T2M + T2X;
|
Chris@10
|
2503 Tfz = T2B - T2Y;
|
Chris@10
|
2504 TfA = Tdd + Tde;
|
Chris@10
|
2505 TfB = Tdk + Tdj;
|
Chris@10
|
2506 TfC = TfA - TfB;
|
Chris@10
|
2507 {
|
Chris@10
|
2508 E T7d, T7g, Tdi, Tdl;
|
Chris@10
|
2509 T7d = T2j - T2o;
|
Chris@10
|
2510 T7g = T7e - T7f;
|
Chris@10
|
2511 T7h = T7d - T7g;
|
Chris@10
|
2512 TaW = T7d + T7g;
|
Chris@10
|
2513 Tdi = T2p - T2A;
|
Chris@10
|
2514 Tdl = Tdj - Tdk;
|
Chris@10
|
2515 Tdm = Tdi - Tdl;
|
Chris@10
|
2516 TeL = Tdi + Tdl;
|
Chris@10
|
2517 }
|
Chris@10
|
2518 {
|
Chris@10
|
2519 E T7m, T7r, T7z, T7A;
|
Chris@10
|
2520 T7m = T7i - T7l;
|
Chris@10
|
2521 T7r = T7n + T7q;
|
Chris@10
|
2522 T7s = KP707106781 * (T7m + T7r);
|
Chris@10
|
2523 TaU = KP707106781 * (T7r - T7m);
|
Chris@10
|
2524 T7z = T7i + T7l;
|
Chris@10
|
2525 T7A = T7q - T7n;
|
Chris@10
|
2526 T7B = KP707106781 * (T7z + T7A);
|
Chris@10
|
2527 TaX = KP707106781 * (T7z - T7A);
|
Chris@10
|
2528 }
|
Chris@10
|
2529 {
|
Chris@10
|
2530 E Tdf, Tdg, T7w, T7x;
|
Chris@10
|
2531 Tdf = Tdd - Tde;
|
Chris@10
|
2532 Tdg = T2M - T2X;
|
Chris@10
|
2533 Tdh = Tdf - Tdg;
|
Chris@10
|
2534 TeM = Tdf + Tdg;
|
Chris@10
|
2535 T7w = T7u - T7v;
|
Chris@10
|
2536 T7x = T2u - T2z;
|
Chris@10
|
2537 T7y = T7w + T7x;
|
Chris@10
|
2538 TaT = T7w - T7x;
|
Chris@10
|
2539 }
|
Chris@10
|
2540 }
|
Chris@10
|
2541 {
|
Chris@10
|
2542 E T4D, T9e, T4I, T9f, T4J, TdR, T4O, T8A, T4T, T8B, T4U, TdS, T56, Tea, T8E;
|
Chris@10
|
2543 E T8H, T5h, Te9, T8J, T8M;
|
Chris@10
|
2544 {
|
Chris@10
|
2545 E T4A, T4C, T4z, T4B;
|
Chris@10
|
2546 T4A = cr[WS(rs, 63)];
|
Chris@10
|
2547 T4C = ci[WS(rs, 63)];
|
Chris@10
|
2548 T4z = W[124];
|
Chris@10
|
2549 T4B = W[125];
|
Chris@10
|
2550 T4D = FMA(T4z, T4A, T4B * T4C);
|
Chris@10
|
2551 T9e = FNMS(T4B, T4A, T4z * T4C);
|
Chris@10
|
2552 }
|
Chris@10
|
2553 {
|
Chris@10
|
2554 E T4F, T4H, T4E, T4G;
|
Chris@10
|
2555 T4F = cr[WS(rs, 31)];
|
Chris@10
|
2556 T4H = ci[WS(rs, 31)];
|
Chris@10
|
2557 T4E = W[60];
|
Chris@10
|
2558 T4G = W[61];
|
Chris@10
|
2559 T4I = FMA(T4E, T4F, T4G * T4H);
|
Chris@10
|
2560 T9f = FNMS(T4G, T4F, T4E * T4H);
|
Chris@10
|
2561 }
|
Chris@10
|
2562 T4J = T4D + T4I;
|
Chris@10
|
2563 TdR = T9e + T9f;
|
Chris@10
|
2564 {
|
Chris@10
|
2565 E T4L, T4N, T4K, T4M;
|
Chris@10
|
2566 T4L = cr[WS(rs, 15)];
|
Chris@10
|
2567 T4N = ci[WS(rs, 15)];
|
Chris@10
|
2568 T4K = W[28];
|
Chris@10
|
2569 T4M = W[29];
|
Chris@10
|
2570 T4O = FMA(T4K, T4L, T4M * T4N);
|
Chris@10
|
2571 T8A = FNMS(T4M, T4L, T4K * T4N);
|
Chris@10
|
2572 }
|
Chris@10
|
2573 {
|
Chris@10
|
2574 E T4Q, T4S, T4P, T4R;
|
Chris@10
|
2575 T4Q = cr[WS(rs, 47)];
|
Chris@10
|
2576 T4S = ci[WS(rs, 47)];
|
Chris@10
|
2577 T4P = W[92];
|
Chris@10
|
2578 T4R = W[93];
|
Chris@10
|
2579 T4T = FMA(T4P, T4Q, T4R * T4S);
|
Chris@10
|
2580 T8B = FNMS(T4R, T4Q, T4P * T4S);
|
Chris@10
|
2581 }
|
Chris@10
|
2582 T4U = T4O + T4T;
|
Chris@10
|
2583 TdS = T8A + T8B;
|
Chris@10
|
2584 {
|
Chris@10
|
2585 E T50, T8F, T55, T8G;
|
Chris@10
|
2586 {
|
Chris@10
|
2587 E T4X, T4Z, T4W, T4Y;
|
Chris@10
|
2588 T4X = cr[WS(rs, 7)];
|
Chris@10
|
2589 T4Z = ci[WS(rs, 7)];
|
Chris@10
|
2590 T4W = W[12];
|
Chris@10
|
2591 T4Y = W[13];
|
Chris@10
|
2592 T50 = FMA(T4W, T4X, T4Y * T4Z);
|
Chris@10
|
2593 T8F = FNMS(T4Y, T4X, T4W * T4Z);
|
Chris@10
|
2594 }
|
Chris@10
|
2595 {
|
Chris@10
|
2596 E T52, T54, T51, T53;
|
Chris@10
|
2597 T52 = cr[WS(rs, 39)];
|
Chris@10
|
2598 T54 = ci[WS(rs, 39)];
|
Chris@10
|
2599 T51 = W[76];
|
Chris@10
|
2600 T53 = W[77];
|
Chris@10
|
2601 T55 = FMA(T51, T52, T53 * T54);
|
Chris@10
|
2602 T8G = FNMS(T53, T52, T51 * T54);
|
Chris@10
|
2603 }
|
Chris@10
|
2604 T56 = T50 + T55;
|
Chris@10
|
2605 Tea = T8F + T8G;
|
Chris@10
|
2606 T8E = T50 - T55;
|
Chris@10
|
2607 T8H = T8F - T8G;
|
Chris@10
|
2608 }
|
Chris@10
|
2609 {
|
Chris@10
|
2610 E T5b, T8K, T5g, T8L;
|
Chris@10
|
2611 {
|
Chris@10
|
2612 E T58, T5a, T57, T59;
|
Chris@10
|
2613 T58 = cr[WS(rs, 55)];
|
Chris@10
|
2614 T5a = ci[WS(rs, 55)];
|
Chris@10
|
2615 T57 = W[108];
|
Chris@10
|
2616 T59 = W[109];
|
Chris@10
|
2617 T5b = FMA(T57, T58, T59 * T5a);
|
Chris@10
|
2618 T8K = FNMS(T59, T58, T57 * T5a);
|
Chris@10
|
2619 }
|
Chris@10
|
2620 {
|
Chris@10
|
2621 E T5d, T5f, T5c, T5e;
|
Chris@10
|
2622 T5d = cr[WS(rs, 23)];
|
Chris@10
|
2623 T5f = ci[WS(rs, 23)];
|
Chris@10
|
2624 T5c = W[44];
|
Chris@10
|
2625 T5e = W[45];
|
Chris@10
|
2626 T5g = FMA(T5c, T5d, T5e * T5f);
|
Chris@10
|
2627 T8L = FNMS(T5e, T5d, T5c * T5f);
|
Chris@10
|
2628 }
|
Chris@10
|
2629 T5h = T5b + T5g;
|
Chris@10
|
2630 Te9 = T8K + T8L;
|
Chris@10
|
2631 T8J = T5b - T5g;
|
Chris@10
|
2632 T8M = T8K - T8L;
|
Chris@10
|
2633 }
|
Chris@10
|
2634 {
|
Chris@10
|
2635 E T4V, T5i, Te8, Teb;
|
Chris@10
|
2636 T4V = T4J + T4U;
|
Chris@10
|
2637 T5i = T56 + T5h;
|
Chris@10
|
2638 T5j = T4V + T5i;
|
Chris@10
|
2639 TfR = T4V - T5i;
|
Chris@10
|
2640 Te8 = T4J - T4U;
|
Chris@10
|
2641 Teb = Te9 - Tea;
|
Chris@10
|
2642 Tec = Te8 - Teb;
|
Chris@10
|
2643 TeX = Te8 + Teb;
|
Chris@10
|
2644 }
|
Chris@10
|
2645 {
|
Chris@10
|
2646 E TfW, TfX, T8z, T8C;
|
Chris@10
|
2647 TfW = TdR + TdS;
|
Chris@10
|
2648 TfX = Tea + Te9;
|
Chris@10
|
2649 TfY = TfW - TfX;
|
Chris@10
|
2650 Tgy = TfW + TfX;
|
Chris@10
|
2651 T8z = T4D - T4I;
|
Chris@10
|
2652 T8C = T8A - T8B;
|
Chris@10
|
2653 T8D = T8z - T8C;
|
Chris@10
|
2654 Tbl = T8z + T8C;
|
Chris@10
|
2655 }
|
Chris@10
|
2656 {
|
Chris@10
|
2657 E T8I, T8N, T9j, T9k;
|
Chris@10
|
2658 T8I = T8E - T8H;
|
Chris@10
|
2659 T8N = T8J + T8M;
|
Chris@10
|
2660 T8O = KP707106781 * (T8I + T8N);
|
Chris@10
|
2661 Tbx = KP707106781 * (T8N - T8I);
|
Chris@10
|
2662 T9j = T8E + T8H;
|
Chris@10
|
2663 T9k = T8M - T8J;
|
Chris@10
|
2664 T9l = KP707106781 * (T9j + T9k);
|
Chris@10
|
2665 Tbm = KP707106781 * (T9j - T9k);
|
Chris@10
|
2666 }
|
Chris@10
|
2667 {
|
Chris@10
|
2668 E TdT, TdU, T9g, T9h;
|
Chris@10
|
2669 TdT = TdR - TdS;
|
Chris@10
|
2670 TdU = T56 - T5h;
|
Chris@10
|
2671 TdV = TdT - TdU;
|
Chris@10
|
2672 Tf0 = TdT + TdU;
|
Chris@10
|
2673 T9g = T9e - T9f;
|
Chris@10
|
2674 T9h = T4O - T4T;
|
Chris@10
|
2675 T9i = T9g + T9h;
|
Chris@10
|
2676 Tbw = T9g - T9h;
|
Chris@10
|
2677 }
|
Chris@10
|
2678 }
|
Chris@10
|
2679 {
|
Chris@10
|
2680 E T36, T7G, T3b, T7H, T3c, TdH, T3h, T8m, T3m, T8n, T3n, TdI, T3z, Tds, T7L;
|
Chris@10
|
2681 E T7O, T3K, Tdr, T7S, T7T;
|
Chris@10
|
2682 {
|
Chris@10
|
2683 E T33, T35, T32, T34;
|
Chris@10
|
2684 T33 = cr[WS(rs, 1)];
|
Chris@10
|
2685 T35 = ci[WS(rs, 1)];
|
Chris@10
|
2686 T32 = W[0];
|
Chris@10
|
2687 T34 = W[1];
|
Chris@10
|
2688 T36 = FMA(T32, T33, T34 * T35);
|
Chris@10
|
2689 T7G = FNMS(T34, T33, T32 * T35);
|
Chris@10
|
2690 }
|
Chris@10
|
2691 {
|
Chris@10
|
2692 E T38, T3a, T37, T39;
|
Chris@10
|
2693 T38 = cr[WS(rs, 33)];
|
Chris@10
|
2694 T3a = ci[WS(rs, 33)];
|
Chris@10
|
2695 T37 = W[64];
|
Chris@10
|
2696 T39 = W[65];
|
Chris@10
|
2697 T3b = FMA(T37, T38, T39 * T3a);
|
Chris@10
|
2698 T7H = FNMS(T39, T38, T37 * T3a);
|
Chris@10
|
2699 }
|
Chris@10
|
2700 T3c = T36 + T3b;
|
Chris@10
|
2701 TdH = T7G + T7H;
|
Chris@10
|
2702 {
|
Chris@10
|
2703 E T3e, T3g, T3d, T3f;
|
Chris@10
|
2704 T3e = cr[WS(rs, 17)];
|
Chris@10
|
2705 T3g = ci[WS(rs, 17)];
|
Chris@10
|
2706 T3d = W[32];
|
Chris@10
|
2707 T3f = W[33];
|
Chris@10
|
2708 T3h = FMA(T3d, T3e, T3f * T3g);
|
Chris@10
|
2709 T8m = FNMS(T3f, T3e, T3d * T3g);
|
Chris@10
|
2710 }
|
Chris@10
|
2711 {
|
Chris@10
|
2712 E T3j, T3l, T3i, T3k;
|
Chris@10
|
2713 T3j = cr[WS(rs, 49)];
|
Chris@10
|
2714 T3l = ci[WS(rs, 49)];
|
Chris@10
|
2715 T3i = W[96];
|
Chris@10
|
2716 T3k = W[97];
|
Chris@10
|
2717 T3m = FMA(T3i, T3j, T3k * T3l);
|
Chris@10
|
2718 T8n = FNMS(T3k, T3j, T3i * T3l);
|
Chris@10
|
2719 }
|
Chris@10
|
2720 T3n = T3h + T3m;
|
Chris@10
|
2721 TdI = T8m + T8n;
|
Chris@10
|
2722 {
|
Chris@10
|
2723 E T3t, T7M, T3y, T7N;
|
Chris@10
|
2724 {
|
Chris@10
|
2725 E T3q, T3s, T3p, T3r;
|
Chris@10
|
2726 T3q = cr[WS(rs, 9)];
|
Chris@10
|
2727 T3s = ci[WS(rs, 9)];
|
Chris@10
|
2728 T3p = W[16];
|
Chris@10
|
2729 T3r = W[17];
|
Chris@10
|
2730 T3t = FMA(T3p, T3q, T3r * T3s);
|
Chris@10
|
2731 T7M = FNMS(T3r, T3q, T3p * T3s);
|
Chris@10
|
2732 }
|
Chris@10
|
2733 {
|
Chris@10
|
2734 E T3v, T3x, T3u, T3w;
|
Chris@10
|
2735 T3v = cr[WS(rs, 41)];
|
Chris@10
|
2736 T3x = ci[WS(rs, 41)];
|
Chris@10
|
2737 T3u = W[80];
|
Chris@10
|
2738 T3w = W[81];
|
Chris@10
|
2739 T3y = FMA(T3u, T3v, T3w * T3x);
|
Chris@10
|
2740 T7N = FNMS(T3w, T3v, T3u * T3x);
|
Chris@10
|
2741 }
|
Chris@10
|
2742 T3z = T3t + T3y;
|
Chris@10
|
2743 Tds = T7M + T7N;
|
Chris@10
|
2744 T7L = T3t - T3y;
|
Chris@10
|
2745 T7O = T7M - T7N;
|
Chris@10
|
2746 }
|
Chris@10
|
2747 {
|
Chris@10
|
2748 E T3E, T7Q, T3J, T7R;
|
Chris@10
|
2749 {
|
Chris@10
|
2750 E T3B, T3D, T3A, T3C;
|
Chris@10
|
2751 T3B = cr[WS(rs, 57)];
|
Chris@10
|
2752 T3D = ci[WS(rs, 57)];
|
Chris@10
|
2753 T3A = W[112];
|
Chris@10
|
2754 T3C = W[113];
|
Chris@10
|
2755 T3E = FMA(T3A, T3B, T3C * T3D);
|
Chris@10
|
2756 T7Q = FNMS(T3C, T3B, T3A * T3D);
|
Chris@10
|
2757 }
|
Chris@10
|
2758 {
|
Chris@10
|
2759 E T3G, T3I, T3F, T3H;
|
Chris@10
|
2760 T3G = cr[WS(rs, 25)];
|
Chris@10
|
2761 T3I = ci[WS(rs, 25)];
|
Chris@10
|
2762 T3F = W[48];
|
Chris@10
|
2763 T3H = W[49];
|
Chris@10
|
2764 T3J = FMA(T3F, T3G, T3H * T3I);
|
Chris@10
|
2765 T7R = FNMS(T3H, T3G, T3F * T3I);
|
Chris@10
|
2766 }
|
Chris@10
|
2767 T3K = T3E + T3J;
|
Chris@10
|
2768 Tdr = T7Q + T7R;
|
Chris@10
|
2769 T7S = T7Q - T7R;
|
Chris@10
|
2770 T7T = T3E - T3J;
|
Chris@10
|
2771 }
|
Chris@10
|
2772 {
|
Chris@10
|
2773 E T3o, T3L, TdJ, TdK;
|
Chris@10
|
2774 T3o = T3c + T3n;
|
Chris@10
|
2775 T3L = T3z + T3K;
|
Chris@10
|
2776 T3M = T3o + T3L;
|
Chris@10
|
2777 TfL = T3o - T3L;
|
Chris@10
|
2778 TdJ = TdH - TdI;
|
Chris@10
|
2779 TdK = T3z - T3K;
|
Chris@10
|
2780 TdL = TdJ - TdK;
|
Chris@10
|
2781 TeT = TdJ + TdK;
|
Chris@10
|
2782 }
|
Chris@10
|
2783 {
|
Chris@10
|
2784 E TfG, TfH, T7I, T7J;
|
Chris@10
|
2785 TfG = TdH + TdI;
|
Chris@10
|
2786 TfH = Tds + Tdr;
|
Chris@10
|
2787 TfI = TfG - TfH;
|
Chris@10
|
2788 Tgt = TfG + TfH;
|
Chris@10
|
2789 T7I = T7G - T7H;
|
Chris@10
|
2790 T7J = T3h - T3m;
|
Chris@10
|
2791 T7K = T7I + T7J;
|
Chris@10
|
2792 Tbd = T7I - T7J;
|
Chris@10
|
2793 }
|
Chris@10
|
2794 {
|
Chris@10
|
2795 E T7P, T7U, T8q, T8r;
|
Chris@10
|
2796 T7P = T7L + T7O;
|
Chris@10
|
2797 T7U = T7S - T7T;
|
Chris@10
|
2798 T7V = KP707106781 * (T7P + T7U);
|
Chris@10
|
2799 Tb3 = KP707106781 * (T7P - T7U);
|
Chris@10
|
2800 T8q = T7L - T7O;
|
Chris@10
|
2801 T8r = T7T + T7S;
|
Chris@10
|
2802 T8s = KP707106781 * (T8q + T8r);
|
Chris@10
|
2803 Tbe = KP707106781 * (T8r - T8q);
|
Chris@10
|
2804 }
|
Chris@10
|
2805 {
|
Chris@10
|
2806 E Tdq, Tdt, T8l, T8o;
|
Chris@10
|
2807 Tdq = T3c - T3n;
|
Chris@10
|
2808 Tdt = Tdr - Tds;
|
Chris@10
|
2809 Tdu = Tdq - Tdt;
|
Chris@10
|
2810 TeQ = Tdq + Tdt;
|
Chris@10
|
2811 T8l = T36 - T3b;
|
Chris@10
|
2812 T8o = T8m - T8n;
|
Chris@10
|
2813 T8p = T8l - T8o;
|
Chris@10
|
2814 Tb2 = T8l + T8o;
|
Chris@10
|
2815 }
|
Chris@10
|
2816 }
|
Chris@10
|
2817 {
|
Chris@10
|
2818 E T3X, Tdw, T7Z, T82, T4v, TdB, T8b, T8g, T48, Tdx, T80, T85, T4k, TdA, T8a;
|
Chris@10
|
2819 E T8d;
|
Chris@10
|
2820 {
|
Chris@10
|
2821 E T3R, T7X, T3W, T7Y;
|
Chris@10
|
2822 {
|
Chris@10
|
2823 E T3O, T3Q, T3N, T3P;
|
Chris@10
|
2824 T3O = cr[WS(rs, 5)];
|
Chris@10
|
2825 T3Q = ci[WS(rs, 5)];
|
Chris@10
|
2826 T3N = W[8];
|
Chris@10
|
2827 T3P = W[9];
|
Chris@10
|
2828 T3R = FMA(T3N, T3O, T3P * T3Q);
|
Chris@10
|
2829 T7X = FNMS(T3P, T3O, T3N * T3Q);
|
Chris@10
|
2830 }
|
Chris@10
|
2831 {
|
Chris@10
|
2832 E T3T, T3V, T3S, T3U;
|
Chris@10
|
2833 T3T = cr[WS(rs, 37)];
|
Chris@10
|
2834 T3V = ci[WS(rs, 37)];
|
Chris@10
|
2835 T3S = W[72];
|
Chris@10
|
2836 T3U = W[73];
|
Chris@10
|
2837 T3W = FMA(T3S, T3T, T3U * T3V);
|
Chris@10
|
2838 T7Y = FNMS(T3U, T3T, T3S * T3V);
|
Chris@10
|
2839 }
|
Chris@10
|
2840 T3X = T3R + T3W;
|
Chris@10
|
2841 Tdw = T7X + T7Y;
|
Chris@10
|
2842 T7Z = T7X - T7Y;
|
Chris@10
|
2843 T82 = T3R - T3W;
|
Chris@10
|
2844 }
|
Chris@10
|
2845 {
|
Chris@10
|
2846 E T4p, T8e, T4u, T8f;
|
Chris@10
|
2847 {
|
Chris@10
|
2848 E T4m, T4o, T4l, T4n;
|
Chris@10
|
2849 T4m = cr[WS(rs, 13)];
|
Chris@10
|
2850 T4o = ci[WS(rs, 13)];
|
Chris@10
|
2851 T4l = W[24];
|
Chris@10
|
2852 T4n = W[25];
|
Chris@10
|
2853 T4p = FMA(T4l, T4m, T4n * T4o);
|
Chris@10
|
2854 T8e = FNMS(T4n, T4m, T4l * T4o);
|
Chris@10
|
2855 }
|
Chris@10
|
2856 {
|
Chris@10
|
2857 E T4r, T4t, T4q, T4s;
|
Chris@10
|
2858 T4r = cr[WS(rs, 45)];
|
Chris@10
|
2859 T4t = ci[WS(rs, 45)];
|
Chris@10
|
2860 T4q = W[88];
|
Chris@10
|
2861 T4s = W[89];
|
Chris@10
|
2862 T4u = FMA(T4q, T4r, T4s * T4t);
|
Chris@10
|
2863 T8f = FNMS(T4s, T4r, T4q * T4t);
|
Chris@10
|
2864 }
|
Chris@10
|
2865 T4v = T4p + T4u;
|
Chris@10
|
2866 TdB = T8e + T8f;
|
Chris@10
|
2867 T8b = T4p - T4u;
|
Chris@10
|
2868 T8g = T8e - T8f;
|
Chris@10
|
2869 }
|
Chris@10
|
2870 {
|
Chris@10
|
2871 E T42, T83, T47, T84;
|
Chris@10
|
2872 {
|
Chris@10
|
2873 E T3Z, T41, T3Y, T40;
|
Chris@10
|
2874 T3Z = cr[WS(rs, 21)];
|
Chris@10
|
2875 T41 = ci[WS(rs, 21)];
|
Chris@10
|
2876 T3Y = W[40];
|
Chris@10
|
2877 T40 = W[41];
|
Chris@10
|
2878 T42 = FMA(T3Y, T3Z, T40 * T41);
|
Chris@10
|
2879 T83 = FNMS(T40, T3Z, T3Y * T41);
|
Chris@10
|
2880 }
|
Chris@10
|
2881 {
|
Chris@10
|
2882 E T44, T46, T43, T45;
|
Chris@10
|
2883 T44 = cr[WS(rs, 53)];
|
Chris@10
|
2884 T46 = ci[WS(rs, 53)];
|
Chris@10
|
2885 T43 = W[104];
|
Chris@10
|
2886 T45 = W[105];
|
Chris@10
|
2887 T47 = FMA(T43, T44, T45 * T46);
|
Chris@10
|
2888 T84 = FNMS(T45, T44, T43 * T46);
|
Chris@10
|
2889 }
|
Chris@10
|
2890 T48 = T42 + T47;
|
Chris@10
|
2891 Tdx = T83 + T84;
|
Chris@10
|
2892 T80 = T42 - T47;
|
Chris@10
|
2893 T85 = T83 - T84;
|
Chris@10
|
2894 }
|
Chris@10
|
2895 {
|
Chris@10
|
2896 E T4e, T88, T4j, T89;
|
Chris@10
|
2897 {
|
Chris@10
|
2898 E T4b, T4d, T4a, T4c;
|
Chris@10
|
2899 T4b = cr[WS(rs, 61)];
|
Chris@10
|
2900 T4d = ci[WS(rs, 61)];
|
Chris@10
|
2901 T4a = W[120];
|
Chris@10
|
2902 T4c = W[121];
|
Chris@10
|
2903 T4e = FMA(T4a, T4b, T4c * T4d);
|
Chris@10
|
2904 T88 = FNMS(T4c, T4b, T4a * T4d);
|
Chris@10
|
2905 }
|
Chris@10
|
2906 {
|
Chris@10
|
2907 E T4g, T4i, T4f, T4h;
|
Chris@10
|
2908 T4g = cr[WS(rs, 29)];
|
Chris@10
|
2909 T4i = ci[WS(rs, 29)];
|
Chris@10
|
2910 T4f = W[56];
|
Chris@10
|
2911 T4h = W[57];
|
Chris@10
|
2912 T4j = FMA(T4f, T4g, T4h * T4i);
|
Chris@10
|
2913 T89 = FNMS(T4h, T4g, T4f * T4i);
|
Chris@10
|
2914 }
|
Chris@10
|
2915 T4k = T4e + T4j;
|
Chris@10
|
2916 TdA = T88 + T89;
|
Chris@10
|
2917 T8a = T88 - T89;
|
Chris@10
|
2918 T8d = T4e - T4j;
|
Chris@10
|
2919 }
|
Chris@10
|
2920 {
|
Chris@10
|
2921 E T49, T4w, TdC, TdD;
|
Chris@10
|
2922 T49 = T3X + T48;
|
Chris@10
|
2923 T4w = T4k + T4v;
|
Chris@10
|
2924 T4x = T49 + T4w;
|
Chris@10
|
2925 TfJ = T49 - T4w;
|
Chris@10
|
2926 TdC = TdA - TdB;
|
Chris@10
|
2927 TdD = T4k - T4v;
|
Chris@10
|
2928 TdE = TdC - TdD;
|
Chris@10
|
2929 TdM = TdD + TdC;
|
Chris@10
|
2930 }
|
Chris@10
|
2931 {
|
Chris@10
|
2932 E TfM, TfN, T81, T86;
|
Chris@10
|
2933 TfM = TdA + TdB;
|
Chris@10
|
2934 TfN = Tdw + Tdx;
|
Chris@10
|
2935 TfO = TfM - TfN;
|
Chris@10
|
2936 Tgu = TfN + TfM;
|
Chris@10
|
2937 T81 = T7Z + T80;
|
Chris@10
|
2938 T86 = T82 - T85;
|
Chris@10
|
2939 T87 = FMA(KP923879532, T81, KP382683432 * T86);
|
Chris@10
|
2940 T8u = FNMS(KP382683432, T81, KP923879532 * T86);
|
Chris@10
|
2941 }
|
Chris@10
|
2942 {
|
Chris@10
|
2943 E T8c, T8h, Tb8, Tb9;
|
Chris@10
|
2944 T8c = T8a + T8b;
|
Chris@10
|
2945 T8h = T8d - T8g;
|
Chris@10
|
2946 T8i = FNMS(KP382683432, T8h, KP923879532 * T8c);
|
Chris@10
|
2947 T8v = FMA(KP382683432, T8c, KP923879532 * T8h);
|
Chris@10
|
2948 Tb8 = T8d + T8g;
|
Chris@10
|
2949 Tb9 = T8a - T8b;
|
Chris@10
|
2950 Tba = FNMS(KP382683432, Tb9, KP923879532 * Tb8);
|
Chris@10
|
2951 Tbh = FMA(KP923879532, Tb9, KP382683432 * Tb8);
|
Chris@10
|
2952 }
|
Chris@10
|
2953 {
|
Chris@10
|
2954 E Tdv, Tdy, Tb5, Tb6;
|
Chris@10
|
2955 Tdv = T3X - T48;
|
Chris@10
|
2956 Tdy = Tdw - Tdx;
|
Chris@10
|
2957 Tdz = Tdv + Tdy;
|
Chris@10
|
2958 TdN = Tdv - Tdy;
|
Chris@10
|
2959 Tb5 = T7Z - T80;
|
Chris@10
|
2960 Tb6 = T82 + T85;
|
Chris@10
|
2961 Tb7 = FMA(KP382683432, Tb5, KP923879532 * Tb6);
|
Chris@10
|
2962 Tbg = FNMS(KP382683432, Tb6, KP923879532 * Tb5);
|
Chris@10
|
2963 }
|
Chris@10
|
2964 }
|
Chris@10
|
2965 {
|
Chris@10
|
2966 E T5u, Te2, T8Q, T8X, T62, TdY, T94, T99, T5F, Te3, T8T, T8Y, T5R, TdX, T93;
|
Chris@10
|
2967 E T96;
|
Chris@10
|
2968 {
|
Chris@10
|
2969 E T5o, T8V, T5t, T8W;
|
Chris@10
|
2970 {
|
Chris@10
|
2971 E T5l, T5n, T5k, T5m;
|
Chris@10
|
2972 T5l = cr[WS(rs, 3)];
|
Chris@10
|
2973 T5n = ci[WS(rs, 3)];
|
Chris@10
|
2974 T5k = W[4];
|
Chris@10
|
2975 T5m = W[5];
|
Chris@10
|
2976 T5o = FMA(T5k, T5l, T5m * T5n);
|
Chris@10
|
2977 T8V = FNMS(T5m, T5l, T5k * T5n);
|
Chris@10
|
2978 }
|
Chris@10
|
2979 {
|
Chris@10
|
2980 E T5q, T5s, T5p, T5r;
|
Chris@10
|
2981 T5q = cr[WS(rs, 35)];
|
Chris@10
|
2982 T5s = ci[WS(rs, 35)];
|
Chris@10
|
2983 T5p = W[68];
|
Chris@10
|
2984 T5r = W[69];
|
Chris@10
|
2985 T5t = FMA(T5p, T5q, T5r * T5s);
|
Chris@10
|
2986 T8W = FNMS(T5r, T5q, T5p * T5s);
|
Chris@10
|
2987 }
|
Chris@10
|
2988 T5u = T5o + T5t;
|
Chris@10
|
2989 Te2 = T8V + T8W;
|
Chris@10
|
2990 T8Q = T5o - T5t;
|
Chris@10
|
2991 T8X = T8V - T8W;
|
Chris@10
|
2992 }
|
Chris@10
|
2993 {
|
Chris@10
|
2994 E T5W, T97, T61, T98;
|
Chris@10
|
2995 {
|
Chris@10
|
2996 E T5T, T5V, T5S, T5U;
|
Chris@10
|
2997 T5T = cr[WS(rs, 11)];
|
Chris@10
|
2998 T5V = ci[WS(rs, 11)];
|
Chris@10
|
2999 T5S = W[20];
|
Chris@10
|
3000 T5U = W[21];
|
Chris@10
|
3001 T5W = FMA(T5S, T5T, T5U * T5V);
|
Chris@10
|
3002 T97 = FNMS(T5U, T5T, T5S * T5V);
|
Chris@10
|
3003 }
|
Chris@10
|
3004 {
|
Chris@10
|
3005 E T5Y, T60, T5X, T5Z;
|
Chris@10
|
3006 T5Y = cr[WS(rs, 43)];
|
Chris@10
|
3007 T60 = ci[WS(rs, 43)];
|
Chris@10
|
3008 T5X = W[84];
|
Chris@10
|
3009 T5Z = W[85];
|
Chris@10
|
3010 T61 = FMA(T5X, T5Y, T5Z * T60);
|
Chris@10
|
3011 T98 = FNMS(T5Z, T5Y, T5X * T60);
|
Chris@10
|
3012 }
|
Chris@10
|
3013 T62 = T5W + T61;
|
Chris@10
|
3014 TdY = T97 + T98;
|
Chris@10
|
3015 T94 = T5W - T61;
|
Chris@10
|
3016 T99 = T97 - T98;
|
Chris@10
|
3017 }
|
Chris@10
|
3018 {
|
Chris@10
|
3019 E T5z, T8R, T5E, T8S;
|
Chris@10
|
3020 {
|
Chris@10
|
3021 E T5w, T5y, T5v, T5x;
|
Chris@10
|
3022 T5w = cr[WS(rs, 19)];
|
Chris@10
|
3023 T5y = ci[WS(rs, 19)];
|
Chris@10
|
3024 T5v = W[36];
|
Chris@10
|
3025 T5x = W[37];
|
Chris@10
|
3026 T5z = FMA(T5v, T5w, T5x * T5y);
|
Chris@10
|
3027 T8R = FNMS(T5x, T5w, T5v * T5y);
|
Chris@10
|
3028 }
|
Chris@10
|
3029 {
|
Chris@10
|
3030 E T5B, T5D, T5A, T5C;
|
Chris@10
|
3031 T5B = cr[WS(rs, 51)];
|
Chris@10
|
3032 T5D = ci[WS(rs, 51)];
|
Chris@10
|
3033 T5A = W[100];
|
Chris@10
|
3034 T5C = W[101];
|
Chris@10
|
3035 T5E = FMA(T5A, T5B, T5C * T5D);
|
Chris@10
|
3036 T8S = FNMS(T5C, T5B, T5A * T5D);
|
Chris@10
|
3037 }
|
Chris@10
|
3038 T5F = T5z + T5E;
|
Chris@10
|
3039 Te3 = T8R + T8S;
|
Chris@10
|
3040 T8T = T8R - T8S;
|
Chris@10
|
3041 T8Y = T5z - T5E;
|
Chris@10
|
3042 }
|
Chris@10
|
3043 {
|
Chris@10
|
3044 E T5L, T91, T5Q, T92;
|
Chris@10
|
3045 {
|
Chris@10
|
3046 E T5I, T5K, T5H, T5J;
|
Chris@10
|
3047 T5I = cr[WS(rs, 59)];
|
Chris@10
|
3048 T5K = ci[WS(rs, 59)];
|
Chris@10
|
3049 T5H = W[116];
|
Chris@10
|
3050 T5J = W[117];
|
Chris@10
|
3051 T5L = FMA(T5H, T5I, T5J * T5K);
|
Chris@10
|
3052 T91 = FNMS(T5J, T5I, T5H * T5K);
|
Chris@10
|
3053 }
|
Chris@10
|
3054 {
|
Chris@10
|
3055 E T5N, T5P, T5M, T5O;
|
Chris@10
|
3056 T5N = cr[WS(rs, 27)];
|
Chris@10
|
3057 T5P = ci[WS(rs, 27)];
|
Chris@10
|
3058 T5M = W[52];
|
Chris@10
|
3059 T5O = W[53];
|
Chris@10
|
3060 T5Q = FMA(T5M, T5N, T5O * T5P);
|
Chris@10
|
3061 T92 = FNMS(T5O, T5N, T5M * T5P);
|
Chris@10
|
3062 }
|
Chris@10
|
3063 T5R = T5L + T5Q;
|
Chris@10
|
3064 TdX = T91 + T92;
|
Chris@10
|
3065 T93 = T91 - T92;
|
Chris@10
|
3066 T96 = T5L - T5Q;
|
Chris@10
|
3067 }
|
Chris@10
|
3068 {
|
Chris@10
|
3069 E T5G, T63, Te1, Te4;
|
Chris@10
|
3070 T5G = T5u + T5F;
|
Chris@10
|
3071 T63 = T5R + T62;
|
Chris@10
|
3072 T64 = T5G + T63;
|
Chris@10
|
3073 TfZ = T5G - T63;
|
Chris@10
|
3074 Te1 = T5u - T5F;
|
Chris@10
|
3075 Te4 = Te2 - Te3;
|
Chris@10
|
3076 Te5 = Te1 - Te4;
|
Chris@10
|
3077 Ted = Te1 + Te4;
|
Chris@10
|
3078 }
|
Chris@10
|
3079 {
|
Chris@10
|
3080 E TfS, TfT, T8U, T8Z;
|
Chris@10
|
3081 TfS = TdX + TdY;
|
Chris@10
|
3082 TfT = Te2 + Te3;
|
Chris@10
|
3083 TfU = TfS - TfT;
|
Chris@10
|
3084 Tgz = TfT + TfS;
|
Chris@10
|
3085 T8U = T8Q - T8T;
|
Chris@10
|
3086 T8Z = T8X + T8Y;
|
Chris@10
|
3087 T90 = FNMS(KP382683432, T8Z, KP923879532 * T8U);
|
Chris@10
|
3088 T9n = FMA(KP923879532, T8Z, KP382683432 * T8U);
|
Chris@10
|
3089 }
|
Chris@10
|
3090 {
|
Chris@10
|
3091 E T95, T9a, Tbr, Tbs;
|
Chris@10
|
3092 T95 = T93 + T94;
|
Chris@10
|
3093 T9a = T96 - T99;
|
Chris@10
|
3094 T9b = FMA(KP382683432, T95, KP923879532 * T9a);
|
Chris@10
|
3095 T9o = FNMS(KP382683432, T9a, KP923879532 * T95);
|
Chris@10
|
3096 Tbr = T96 + T99;
|
Chris@10
|
3097 Tbs = T93 - T94;
|
Chris@10
|
3098 Tbt = FNMS(KP382683432, Tbs, KP923879532 * Tbr);
|
Chris@10
|
3099 TbA = FMA(KP923879532, Tbs, KP382683432 * Tbr);
|
Chris@10
|
3100 }
|
Chris@10
|
3101 {
|
Chris@10
|
3102 E TdW, TdZ, Tbo, Tbp;
|
Chris@10
|
3103 TdW = T5R - T62;
|
Chris@10
|
3104 TdZ = TdX - TdY;
|
Chris@10
|
3105 Te0 = TdW + TdZ;
|
Chris@10
|
3106 Tee = TdZ - TdW;
|
Chris@10
|
3107 Tbo = T8X - T8Y;
|
Chris@10
|
3108 Tbp = T8Q + T8T;
|
Chris@10
|
3109 Tbq = FMA(KP382683432, Tbo, KP923879532 * Tbp);
|
Chris@10
|
3110 Tbz = FNMS(KP382683432, Tbp, KP923879532 * Tbo);
|
Chris@10
|
3111 }
|
Chris@10
|
3112 }
|
Chris@10
|
3113 {
|
Chris@10
|
3114 E T1t, Tgn, TgK, TgL, TgV, Th1, T30, Th0, T66, TgX, Tgw, TgE, TgB, TgF, Tgq;
|
Chris@10
|
3115 E TgM;
|
Chris@10
|
3116 {
|
Chris@10
|
3117 E TH, T1s, TgI, TgJ;
|
Chris@10
|
3118 TH = Tj + TG;
|
Chris@10
|
3119 T1s = T14 + T1r;
|
Chris@10
|
3120 T1t = TH + T1s;
|
Chris@10
|
3121 Tgn = TH - T1s;
|
Chris@10
|
3122 TgI = Tgy + Tgz;
|
Chris@10
|
3123 TgJ = Tgt + Tgu;
|
Chris@10
|
3124 TgK = TgI - TgJ;
|
Chris@10
|
3125 TgL = TgJ + TgI;
|
Chris@10
|
3126 }
|
Chris@10
|
3127 {
|
Chris@10
|
3128 E TgN, TgU, T2e, T2Z;
|
Chris@10
|
3129 TgN = Tfr + Tfq;
|
Chris@10
|
3130 TgU = TgO + TgT;
|
Chris@10
|
3131 TgV = TgN + TgU;
|
Chris@10
|
3132 Th1 = TgU - TgN;
|
Chris@10
|
3133 T2e = T1Q + T2d;
|
Chris@10
|
3134 T2Z = T2B + T2Y;
|
Chris@10
|
3135 T30 = T2e + T2Z;
|
Chris@10
|
3136 Th0 = T2e - T2Z;
|
Chris@10
|
3137 }
|
Chris@10
|
3138 {
|
Chris@10
|
3139 E T4y, T65, Tgs, Tgv;
|
Chris@10
|
3140 T4y = T3M + T4x;
|
Chris@10
|
3141 T65 = T5j + T64;
|
Chris@10
|
3142 T66 = T4y + T65;
|
Chris@10
|
3143 TgX = T65 - T4y;
|
Chris@10
|
3144 Tgs = T3M - T4x;
|
Chris@10
|
3145 Tgv = Tgt - Tgu;
|
Chris@10
|
3146 Tgw = Tgs + Tgv;
|
Chris@10
|
3147 TgE = Tgs - Tgv;
|
Chris@10
|
3148 }
|
Chris@10
|
3149 {
|
Chris@10
|
3150 E Tgx, TgA, Tgo, Tgp;
|
Chris@10
|
3151 Tgx = T5j - T64;
|
Chris@10
|
3152 TgA = Tgy - Tgz;
|
Chris@10
|
3153 TgB = Tgx - TgA;
|
Chris@10
|
3154 TgF = Tgx + TgA;
|
Chris@10
|
3155 Tgo = TfA + TfB;
|
Chris@10
|
3156 Tgp = Tfv + Tfw;
|
Chris@10
|
3157 Tgq = Tgo - Tgp;
|
Chris@10
|
3158 TgM = Tgp + Tgo;
|
Chris@10
|
3159 }
|
Chris@10
|
3160 {
|
Chris@10
|
3161 E T31, TgW, TgY, TgH;
|
Chris@10
|
3162 T31 = T1t + T30;
|
Chris@10
|
3163 ci[WS(rs, 31)] = T31 - T66;
|
Chris@10
|
3164 cr[0] = T31 + T66;
|
Chris@10
|
3165 TgW = TgM + TgV;
|
Chris@10
|
3166 cr[WS(rs, 32)] = TgL - TgW;
|
Chris@10
|
3167 ci[WS(rs, 63)] = TgL + TgW;
|
Chris@10
|
3168 TgY = TgV - TgM;
|
Chris@10
|
3169 cr[WS(rs, 48)] = TgX - TgY;
|
Chris@10
|
3170 ci[WS(rs, 47)] = TgX + TgY;
|
Chris@10
|
3171 TgH = T1t - T30;
|
Chris@10
|
3172 cr[WS(rs, 16)] = TgH - TgK;
|
Chris@10
|
3173 ci[WS(rs, 15)] = TgH + TgK;
|
Chris@10
|
3174 }
|
Chris@10
|
3175 {
|
Chris@10
|
3176 E Tgr, TgC, TgZ, Th2;
|
Chris@10
|
3177 Tgr = Tgn - Tgq;
|
Chris@10
|
3178 TgC = KP707106781 * (Tgw + TgB);
|
Chris@10
|
3179 ci[WS(rs, 23)] = Tgr - TgC;
|
Chris@10
|
3180 cr[WS(rs, 8)] = Tgr + TgC;
|
Chris@10
|
3181 TgZ = KP707106781 * (TgB - Tgw);
|
Chris@10
|
3182 Th2 = Th0 + Th1;
|
Chris@10
|
3183 cr[WS(rs, 56)] = TgZ - Th2;
|
Chris@10
|
3184 ci[WS(rs, 39)] = TgZ + Th2;
|
Chris@10
|
3185 }
|
Chris@10
|
3186 {
|
Chris@10
|
3187 E Th3, Th4, TgD, TgG;
|
Chris@10
|
3188 Th3 = KP707106781 * (TgF - TgE);
|
Chris@10
|
3189 Th4 = Th1 - Th0;
|
Chris@10
|
3190 cr[WS(rs, 40)] = Th3 - Th4;
|
Chris@10
|
3191 ci[WS(rs, 55)] = Th3 + Th4;
|
Chris@10
|
3192 TgD = Tgn + Tgq;
|
Chris@10
|
3193 TgG = KP707106781 * (TgE + TgF);
|
Chris@10
|
3194 cr[WS(rs, 24)] = TgD - TgG;
|
Chris@10
|
3195 ci[WS(rs, 7)] = TgD + TgG;
|
Chris@10
|
3196 }
|
Chris@10
|
3197 }
|
Chris@10
|
3198 {
|
Chris@10
|
3199 E T6L, T9x, ThV, Ti1, T7E, Ti0, T9A, ThO, T8y, T9K, T9u, T9E, T9r, T9L, T9v;
|
Chris@10
|
3200 E T9H;
|
Chris@10
|
3201 {
|
Chris@10
|
3202 E T6n, T6K, ThP, ThU;
|
Chris@10
|
3203 T6n = T6b + T6m;
|
Chris@10
|
3204 T6K = T6y + T6J;
|
Chris@10
|
3205 T6L = T6n - T6K;
|
Chris@10
|
3206 T9x = T6n + T6K;
|
Chris@10
|
3207 ThP = T9O - T9P;
|
Chris@10
|
3208 ThU = ThQ + ThT;
|
Chris@10
|
3209 ThV = ThP + ThU;
|
Chris@10
|
3210 Ti1 = ThU - ThP;
|
Chris@10
|
3211 }
|
Chris@10
|
3212 {
|
Chris@10
|
3213 E T7c, T9y, T7D, T9z;
|
Chris@10
|
3214 {
|
Chris@10
|
3215 E T72, T7b, T7t, T7C;
|
Chris@10
|
3216 T72 = T6Q + T71;
|
Chris@10
|
3217 T7b = T77 + T7a;
|
Chris@10
|
3218 T7c = FMA(KP195090322, T72, KP980785280 * T7b);
|
Chris@10
|
3219 T9y = FNMS(KP195090322, T7b, KP980785280 * T72);
|
Chris@10
|
3220 T7t = T7h + T7s;
|
Chris@10
|
3221 T7C = T7y + T7B;
|
Chris@10
|
3222 T7D = FNMS(KP980785280, T7C, KP195090322 * T7t);
|
Chris@10
|
3223 T9z = FMA(KP980785280, T7t, KP195090322 * T7C);
|
Chris@10
|
3224 }
|
Chris@10
|
3225 T7E = T7c + T7D;
|
Chris@10
|
3226 Ti0 = T9z - T9y;
|
Chris@10
|
3227 T9A = T9y + T9z;
|
Chris@10
|
3228 ThO = T7c - T7D;
|
Chris@10
|
3229 }
|
Chris@10
|
3230 {
|
Chris@10
|
3231 E T8k, T9D, T8x, T9C;
|
Chris@10
|
3232 {
|
Chris@10
|
3233 E T7W, T8j, T8t, T8w;
|
Chris@10
|
3234 T7W = T7K + T7V;
|
Chris@10
|
3235 T8j = T87 + T8i;
|
Chris@10
|
3236 T8k = T7W - T8j;
|
Chris@10
|
3237 T9D = T7W + T8j;
|
Chris@10
|
3238 T8t = T8p + T8s;
|
Chris@10
|
3239 T8w = T8u + T8v;
|
Chris@10
|
3240 T8x = T8t - T8w;
|
Chris@10
|
3241 T9C = T8t + T8w;
|
Chris@10
|
3242 }
|
Chris@10
|
3243 T8y = FMA(KP634393284, T8k, KP773010453 * T8x);
|
Chris@10
|
3244 T9K = FMA(KP995184726, T9D, KP098017140 * T9C);
|
Chris@10
|
3245 T9u = FNMS(KP773010453, T8k, KP634393284 * T8x);
|
Chris@10
|
3246 T9E = FNMS(KP098017140, T9D, KP995184726 * T9C);
|
Chris@10
|
3247 }
|
Chris@10
|
3248 {
|
Chris@10
|
3249 E T9d, T9G, T9q, T9F;
|
Chris@10
|
3250 {
|
Chris@10
|
3251 E T8P, T9c, T9m, T9p;
|
Chris@10
|
3252 T8P = T8D + T8O;
|
Chris@10
|
3253 T9c = T90 + T9b;
|
Chris@10
|
3254 T9d = T8P - T9c;
|
Chris@10
|
3255 T9G = T8P + T9c;
|
Chris@10
|
3256 T9m = T9i + T9l;
|
Chris@10
|
3257 T9p = T9n + T9o;
|
Chris@10
|
3258 T9q = T9m - T9p;
|
Chris@10
|
3259 T9F = T9m + T9p;
|
Chris@10
|
3260 }
|
Chris@10
|
3261 T9r = FNMS(KP634393284, T9q, KP773010453 * T9d);
|
Chris@10
|
3262 T9L = FNMS(KP995184726, T9F, KP098017140 * T9G);
|
Chris@10
|
3263 T9v = FMA(KP773010453, T9q, KP634393284 * T9d);
|
Chris@10
|
3264 T9H = FMA(KP098017140, T9F, KP995184726 * T9G);
|
Chris@10
|
3265 }
|
Chris@10
|
3266 {
|
Chris@10
|
3267 E T7F, T9s, ThZ, Ti2;
|
Chris@10
|
3268 T7F = T6L + T7E;
|
Chris@10
|
3269 T9s = T8y + T9r;
|
Chris@10
|
3270 ci[WS(rs, 24)] = T7F - T9s;
|
Chris@10
|
3271 cr[WS(rs, 7)] = T7F + T9s;
|
Chris@10
|
3272 ThZ = T9v - T9u;
|
Chris@10
|
3273 Ti2 = Ti0 + Ti1;
|
Chris@10
|
3274 cr[WS(rs, 39)] = ThZ - Ti2;
|
Chris@10
|
3275 ci[WS(rs, 56)] = ThZ + Ti2;
|
Chris@10
|
3276 }
|
Chris@10
|
3277 {
|
Chris@10
|
3278 E Ti3, Ti4, T9t, T9w;
|
Chris@10
|
3279 Ti3 = T9r - T8y;
|
Chris@10
|
3280 Ti4 = Ti1 - Ti0;
|
Chris@10
|
3281 cr[WS(rs, 55)] = Ti3 - Ti4;
|
Chris@10
|
3282 ci[WS(rs, 40)] = Ti3 + Ti4;
|
Chris@10
|
3283 T9t = T6L - T7E;
|
Chris@10
|
3284 T9w = T9u + T9v;
|
Chris@10
|
3285 cr[WS(rs, 23)] = T9t - T9w;
|
Chris@10
|
3286 ci[WS(rs, 8)] = T9t + T9w;
|
Chris@10
|
3287 }
|
Chris@10
|
3288 {
|
Chris@10
|
3289 E T9B, T9I, ThN, ThW;
|
Chris@10
|
3290 T9B = T9x + T9A;
|
Chris@10
|
3291 T9I = T9E + T9H;
|
Chris@10
|
3292 cr[WS(rs, 31)] = T9B - T9I;
|
Chris@10
|
3293 ci[0] = T9B + T9I;
|
Chris@10
|
3294 ThN = T9L - T9K;
|
Chris@10
|
3295 ThW = ThO + ThV;
|
Chris@10
|
3296 cr[WS(rs, 63)] = ThN - ThW;
|
Chris@10
|
3297 ci[WS(rs, 32)] = ThN + ThW;
|
Chris@10
|
3298 }
|
Chris@10
|
3299 {
|
Chris@10
|
3300 E ThX, ThY, T9J, T9M;
|
Chris@10
|
3301 ThX = T9H - T9E;
|
Chris@10
|
3302 ThY = ThV - ThO;
|
Chris@10
|
3303 cr[WS(rs, 47)] = ThX - ThY;
|
Chris@10
|
3304 ci[WS(rs, 48)] = ThX + ThY;
|
Chris@10
|
3305 T9J = T9x - T9A;
|
Chris@10
|
3306 T9M = T9K + T9L;
|
Chris@10
|
3307 ci[WS(rs, 16)] = T9J - T9M;
|
Chris@10
|
3308 cr[WS(rs, 15)] = T9J + T9M;
|
Chris@10
|
3309 }
|
Chris@10
|
3310 }
|
Chris@10
|
3311 {
|
Chris@10
|
3312 E Tft, Tg7, Tgh, Tgl, Th9, Thf, TfE, Th6, TfQ, Tg4, Tga, The, Tge, Tgk, Tg1;
|
Chris@10
|
3313 E Tg5;
|
Chris@10
|
3314 {
|
Chris@10
|
3315 E Tfp, Tfs, Tgf, Tgg;
|
Chris@10
|
3316 Tfp = Tj - TG;
|
Chris@10
|
3317 Tfs = Tfq - Tfr;
|
Chris@10
|
3318 Tft = Tfp - Tfs;
|
Chris@10
|
3319 Tg7 = Tfp + Tfs;
|
Chris@10
|
3320 Tgf = TfY + TfZ;
|
Chris@10
|
3321 Tgg = TfR + TfU;
|
Chris@10
|
3322 Tgh = FMA(KP382683432, Tgf, KP923879532 * Tgg);
|
Chris@10
|
3323 Tgl = FNMS(KP923879532, Tgf, KP382683432 * Tgg);
|
Chris@10
|
3324 }
|
Chris@10
|
3325 {
|
Chris@10
|
3326 E Th7, Th8, Tfy, TfD;
|
Chris@10
|
3327 Th7 = T14 - T1r;
|
Chris@10
|
3328 Th8 = TgT - TgO;
|
Chris@10
|
3329 Th9 = Th7 + Th8;
|
Chris@10
|
3330 Thf = Th8 - Th7;
|
Chris@10
|
3331 Tfy = Tfu + Tfx;
|
Chris@10
|
3332 TfD = Tfz - TfC;
|
Chris@10
|
3333 TfE = KP707106781 * (Tfy + TfD);
|
Chris@10
|
3334 Th6 = KP707106781 * (Tfy - TfD);
|
Chris@10
|
3335 }
|
Chris@10
|
3336 {
|
Chris@10
|
3337 E TfK, TfP, Tg8, Tg9;
|
Chris@10
|
3338 TfK = TfI - TfJ;
|
Chris@10
|
3339 TfP = TfL - TfO;
|
Chris@10
|
3340 TfQ = FMA(KP382683432, TfK, KP923879532 * TfP);
|
Chris@10
|
3341 Tg4 = FNMS(KP923879532, TfK, KP382683432 * TfP);
|
Chris@10
|
3342 Tg8 = Tfu - Tfx;
|
Chris@10
|
3343 Tg9 = Tfz + TfC;
|
Chris@10
|
3344 Tga = KP707106781 * (Tg8 + Tg9);
|
Chris@10
|
3345 The = KP707106781 * (Tg9 - Tg8);
|
Chris@10
|
3346 }
|
Chris@10
|
3347 {
|
Chris@10
|
3348 E Tgc, Tgd, TfV, Tg0;
|
Chris@10
|
3349 Tgc = TfL + TfO;
|
Chris@10
|
3350 Tgd = TfI + TfJ;
|
Chris@10
|
3351 Tge = FNMS(KP382683432, Tgd, KP923879532 * Tgc);
|
Chris@10
|
3352 Tgk = FMA(KP923879532, Tgd, KP382683432 * Tgc);
|
Chris@10
|
3353 TfV = TfR - TfU;
|
Chris@10
|
3354 Tg0 = TfY - TfZ;
|
Chris@10
|
3355 Tg1 = FNMS(KP382683432, Tg0, KP923879532 * TfV);
|
Chris@10
|
3356 Tg5 = FMA(KP923879532, Tg0, KP382683432 * TfV);
|
Chris@10
|
3357 }
|
Chris@10
|
3358 {
|
Chris@10
|
3359 E TfF, Tg2, Thd, Thg;
|
Chris@10
|
3360 TfF = Tft + TfE;
|
Chris@10
|
3361 Tg2 = TfQ + Tg1;
|
Chris@10
|
3362 ci[WS(rs, 27)] = TfF - Tg2;
|
Chris@10
|
3363 cr[WS(rs, 4)] = TfF + Tg2;
|
Chris@10
|
3364 Thd = Tg5 - Tg4;
|
Chris@10
|
3365 Thg = The + Thf;
|
Chris@10
|
3366 cr[WS(rs, 36)] = Thd - Thg;
|
Chris@10
|
3367 ci[WS(rs, 59)] = Thd + Thg;
|
Chris@10
|
3368 }
|
Chris@10
|
3369 {
|
Chris@10
|
3370 E Thh, Thi, Tg3, Tg6;
|
Chris@10
|
3371 Thh = Tg1 - TfQ;
|
Chris@10
|
3372 Thi = Thf - The;
|
Chris@10
|
3373 cr[WS(rs, 52)] = Thh - Thi;
|
Chris@10
|
3374 ci[WS(rs, 43)] = Thh + Thi;
|
Chris@10
|
3375 Tg3 = Tft - TfE;
|
Chris@10
|
3376 Tg6 = Tg4 + Tg5;
|
Chris@10
|
3377 cr[WS(rs, 20)] = Tg3 - Tg6;
|
Chris@10
|
3378 ci[WS(rs, 11)] = Tg3 + Tg6;
|
Chris@10
|
3379 }
|
Chris@10
|
3380 {
|
Chris@10
|
3381 E Tgb, Tgi, Th5, Tha;
|
Chris@10
|
3382 Tgb = Tg7 + Tga;
|
Chris@10
|
3383 Tgi = Tge + Tgh;
|
Chris@10
|
3384 cr[WS(rs, 28)] = Tgb - Tgi;
|
Chris@10
|
3385 ci[WS(rs, 3)] = Tgb + Tgi;
|
Chris@10
|
3386 Th5 = Tgl - Tgk;
|
Chris@10
|
3387 Tha = Th6 + Th9;
|
Chris@10
|
3388 cr[WS(rs, 60)] = Th5 - Tha;
|
Chris@10
|
3389 ci[WS(rs, 35)] = Th5 + Tha;
|
Chris@10
|
3390 }
|
Chris@10
|
3391 {
|
Chris@10
|
3392 E Thb, Thc, Tgj, Tgm;
|
Chris@10
|
3393 Thb = Tgh - Tge;
|
Chris@10
|
3394 Thc = Th9 - Th6;
|
Chris@10
|
3395 cr[WS(rs, 44)] = Thb - Thc;
|
Chris@10
|
3396 ci[WS(rs, 51)] = Thb + Thc;
|
Chris@10
|
3397 Tgj = Tg7 - Tga;
|
Chris@10
|
3398 Tgm = Tgk + Tgl;
|
Chris@10
|
3399 ci[WS(rs, 19)] = Tgj - Tgm;
|
Chris@10
|
3400 cr[WS(rs, 12)] = Tgj + Tgm;
|
Chris@10
|
3401 }
|
Chris@10
|
3402 }
|
Chris@10
|
3403 {
|
Chris@10
|
3404 E TeH, Tf9, TeO, Thk, Thp, Thv, Tfc, Thu, Tf3, Tfn, Tf7, Tfj, TeW, Tfm, Tf6;
|
Chris@10
|
3405 E Tfg;
|
Chris@10
|
3406 {
|
Chris@10
|
3407 E TeD, TeG, Tfa, Tfb;
|
Chris@10
|
3408 TeD = TcL + TcO;
|
Chris@10
|
3409 TeG = KP707106781 * (TeE + TeF);
|
Chris@10
|
3410 TeH = TeD - TeG;
|
Chris@10
|
3411 Tf9 = TeD + TeG;
|
Chris@10
|
3412 {
|
Chris@10
|
3413 E TeK, TeN, Thl, Tho;
|
Chris@10
|
3414 TeK = FMA(KP923879532, TeI, KP382683432 * TeJ);
|
Chris@10
|
3415 TeN = FNMS(KP923879532, TeM, KP382683432 * TeL);
|
Chris@10
|
3416 TeO = TeK + TeN;
|
Chris@10
|
3417 Thk = TeK - TeN;
|
Chris@10
|
3418 Thl = KP707106781 * (TcU - TcZ);
|
Chris@10
|
3419 Tho = Thm + Thn;
|
Chris@10
|
3420 Thp = Thl + Tho;
|
Chris@10
|
3421 Thv = Tho - Thl;
|
Chris@10
|
3422 }
|
Chris@10
|
3423 Tfa = FNMS(KP382683432, TeI, KP923879532 * TeJ);
|
Chris@10
|
3424 Tfb = FMA(KP382683432, TeM, KP923879532 * TeL);
|
Chris@10
|
3425 Tfc = Tfa + Tfb;
|
Chris@10
|
3426 Thu = Tfb - Tfa;
|
Chris@10
|
3427 {
|
Chris@10
|
3428 E TeZ, Tfh, Tf2, Tfi, TeY, Tf1;
|
Chris@10
|
3429 TeY = KP707106781 * (Te5 + Te0);
|
Chris@10
|
3430 TeZ = TeX - TeY;
|
Chris@10
|
3431 Tfh = TeX + TeY;
|
Chris@10
|
3432 Tf1 = KP707106781 * (Ted + Tee);
|
Chris@10
|
3433 Tf2 = Tf0 - Tf1;
|
Chris@10
|
3434 Tfi = Tf0 + Tf1;
|
Chris@10
|
3435 Tf3 = FNMS(KP555570233, Tf2, KP831469612 * TeZ);
|
Chris@10
|
3436 Tfn = FMA(KP980785280, Tfh, KP195090322 * Tfi);
|
Chris@10
|
3437 Tf7 = FMA(KP555570233, TeZ, KP831469612 * Tf2);
|
Chris@10
|
3438 Tfj = FNMS(KP980785280, Tfi, KP195090322 * Tfh);
|
Chris@10
|
3439 }
|
Chris@10
|
3440 {
|
Chris@10
|
3441 E TeS, Tfe, TeV, Tff, TeR, TeU;
|
Chris@10
|
3442 TeR = KP707106781 * (TdN + TdM);
|
Chris@10
|
3443 TeS = TeQ - TeR;
|
Chris@10
|
3444 Tfe = TeQ + TeR;
|
Chris@10
|
3445 TeU = KP707106781 * (Tdz + TdE);
|
Chris@10
|
3446 TeV = TeT - TeU;
|
Chris@10
|
3447 Tff = TeT + TeU;
|
Chris@10
|
3448 TeW = FMA(KP831469612, TeS, KP555570233 * TeV);
|
Chris@10
|
3449 Tfm = FNMS(KP195090322, Tff, KP980785280 * Tfe);
|
Chris@10
|
3450 Tf6 = FNMS(KP831469612, TeV, KP555570233 * TeS);
|
Chris@10
|
3451 Tfg = FMA(KP195090322, Tfe, KP980785280 * Tff);
|
Chris@10
|
3452 }
|
Chris@10
|
3453 }
|
Chris@10
|
3454 {
|
Chris@10
|
3455 E TeP, Tf4, Tht, Thw;
|
Chris@10
|
3456 TeP = TeH + TeO;
|
Chris@10
|
3457 Tf4 = TeW + Tf3;
|
Chris@10
|
3458 ci[WS(rs, 25)] = TeP - Tf4;
|
Chris@10
|
3459 cr[WS(rs, 6)] = TeP + Tf4;
|
Chris@10
|
3460 Tht = Tf7 - Tf6;
|
Chris@10
|
3461 Thw = Thu + Thv;
|
Chris@10
|
3462 cr[WS(rs, 38)] = Tht - Thw;
|
Chris@10
|
3463 ci[WS(rs, 57)] = Tht + Thw;
|
Chris@10
|
3464 }
|
Chris@10
|
3465 {
|
Chris@10
|
3466 E Thx, Thy, Tf5, Tf8;
|
Chris@10
|
3467 Thx = Tf3 - TeW;
|
Chris@10
|
3468 Thy = Thv - Thu;
|
Chris@10
|
3469 cr[WS(rs, 54)] = Thx - Thy;
|
Chris@10
|
3470 ci[WS(rs, 41)] = Thx + Thy;
|
Chris@10
|
3471 Tf5 = TeH - TeO;
|
Chris@10
|
3472 Tf8 = Tf6 + Tf7;
|
Chris@10
|
3473 cr[WS(rs, 22)] = Tf5 - Tf8;
|
Chris@10
|
3474 ci[WS(rs, 9)] = Tf5 + Tf8;
|
Chris@10
|
3475 }
|
Chris@10
|
3476 {
|
Chris@10
|
3477 E Tfd, Tfk, Thj, Thq;
|
Chris@10
|
3478 Tfd = Tf9 - Tfc;
|
Chris@10
|
3479 Tfk = Tfg + Tfj;
|
Chris@10
|
3480 ci[WS(rs, 17)] = Tfd - Tfk;
|
Chris@10
|
3481 cr[WS(rs, 14)] = Tfd + Tfk;
|
Chris@10
|
3482 Thj = Tfj - Tfg;
|
Chris@10
|
3483 Thq = Thk + Thp;
|
Chris@10
|
3484 cr[WS(rs, 62)] = Thj - Thq;
|
Chris@10
|
3485 ci[WS(rs, 33)] = Thj + Thq;
|
Chris@10
|
3486 }
|
Chris@10
|
3487 {
|
Chris@10
|
3488 E Thr, Ths, Tfl, Tfo;
|
Chris@10
|
3489 Thr = Tfn - Tfm;
|
Chris@10
|
3490 Ths = Thp - Thk;
|
Chris@10
|
3491 cr[WS(rs, 46)] = Thr - Ths;
|
Chris@10
|
3492 ci[WS(rs, 49)] = Thr + Ths;
|
Chris@10
|
3493 Tfl = Tf9 + Tfc;
|
Chris@10
|
3494 Tfo = Tfm + Tfn;
|
Chris@10
|
3495 cr[WS(rs, 30)] = Tfl - Tfo;
|
Chris@10
|
3496 ci[WS(rs, 1)] = Tfl + Tfo;
|
Chris@10
|
3497 }
|
Chris@10
|
3498 }
|
Chris@10
|
3499 {
|
Chris@10
|
3500 E Td1, Ten, Tdo, ThA, ThD, ThJ, Teq, ThI, Teh, TeB, Tel, Tex, TdQ, TeA, Tek;
|
Chris@10
|
3501 E Teu;
|
Chris@10
|
3502 {
|
Chris@10
|
3503 E TcP, Td0, Teo, Tep;
|
Chris@10
|
3504 TcP = TcL - TcO;
|
Chris@10
|
3505 Td0 = KP707106781 * (TcU + TcZ);
|
Chris@10
|
3506 Td1 = TcP - Td0;
|
Chris@10
|
3507 Ten = TcP + Td0;
|
Chris@10
|
3508 {
|
Chris@10
|
3509 E Tdc, Tdn, ThB, ThC;
|
Chris@10
|
3510 Tdc = FNMS(KP923879532, Tdb, KP382683432 * Td6);
|
Chris@10
|
3511 Tdn = FMA(KP923879532, Tdh, KP382683432 * Tdm);
|
Chris@10
|
3512 Tdo = Tdc + Tdn;
|
Chris@10
|
3513 ThA = Tdn - Tdc;
|
Chris@10
|
3514 ThB = KP707106781 * (TeF - TeE);
|
Chris@10
|
3515 ThC = Thn - Thm;
|
Chris@10
|
3516 ThD = ThB + ThC;
|
Chris@10
|
3517 ThJ = ThC - ThB;
|
Chris@10
|
3518 }
|
Chris@10
|
3519 Teo = FMA(KP382683432, Tdb, KP923879532 * Td6);
|
Chris@10
|
3520 Tep = FNMS(KP382683432, Tdh, KP923879532 * Tdm);
|
Chris@10
|
3521 Teq = Teo + Tep;
|
Chris@10
|
3522 ThI = Teo - Tep;
|
Chris@10
|
3523 {
|
Chris@10
|
3524 E Te7, Tew, Teg, Tev, Te6, Tef;
|
Chris@10
|
3525 Te6 = KP707106781 * (Te0 - Te5);
|
Chris@10
|
3526 Te7 = TdV - Te6;
|
Chris@10
|
3527 Tew = TdV + Te6;
|
Chris@10
|
3528 Tef = KP707106781 * (Ted - Tee);
|
Chris@10
|
3529 Teg = Tec - Tef;
|
Chris@10
|
3530 Tev = Tec + Tef;
|
Chris@10
|
3531 Teh = FMA(KP555570233, Te7, KP831469612 * Teg);
|
Chris@10
|
3532 TeB = FMA(KP980785280, Tew, KP195090322 * Tev);
|
Chris@10
|
3533 Tel = FNMS(KP831469612, Te7, KP555570233 * Teg);
|
Chris@10
|
3534 Tex = FNMS(KP195090322, Tew, KP980785280 * Tev);
|
Chris@10
|
3535 }
|
Chris@10
|
3536 {
|
Chris@10
|
3537 E TdG, Tet, TdP, Tes, TdF, TdO;
|
Chris@10
|
3538 TdF = KP707106781 * (Tdz - TdE);
|
Chris@10
|
3539 TdG = Tdu - TdF;
|
Chris@10
|
3540 Tet = Tdu + TdF;
|
Chris@10
|
3541 TdO = KP707106781 * (TdM - TdN);
|
Chris@10
|
3542 TdP = TdL - TdO;
|
Chris@10
|
3543 Tes = TdL + TdO;
|
Chris@10
|
3544 TdQ = FNMS(KP555570233, TdP, KP831469612 * TdG);
|
Chris@10
|
3545 TeA = FNMS(KP980785280, Tes, KP195090322 * Tet);
|
Chris@10
|
3546 Tek = FMA(KP831469612, TdP, KP555570233 * TdG);
|
Chris@10
|
3547 Teu = FMA(KP195090322, Tes, KP980785280 * Tet);
|
Chris@10
|
3548 }
|
Chris@10
|
3549 }
|
Chris@10
|
3550 {
|
Chris@10
|
3551 E Tdp, Tei, ThH, ThK;
|
Chris@10
|
3552 Tdp = Td1 + Tdo;
|
Chris@10
|
3553 Tei = TdQ + Teh;
|
Chris@10
|
3554 cr[WS(rs, 26)] = Tdp - Tei;
|
Chris@10
|
3555 ci[WS(rs, 5)] = Tdp + Tei;
|
Chris@10
|
3556 ThH = Tel - Tek;
|
Chris@10
|
3557 ThK = ThI + ThJ;
|
Chris@10
|
3558 cr[WS(rs, 58)] = ThH - ThK;
|
Chris@10
|
3559 ci[WS(rs, 37)] = ThH + ThK;
|
Chris@10
|
3560 }
|
Chris@10
|
3561 {
|
Chris@10
|
3562 E ThL, ThM, Tej, Tem;
|
Chris@10
|
3563 ThL = Teh - TdQ;
|
Chris@10
|
3564 ThM = ThJ - ThI;
|
Chris@10
|
3565 cr[WS(rs, 42)] = ThL - ThM;
|
Chris@10
|
3566 ci[WS(rs, 53)] = ThL + ThM;
|
Chris@10
|
3567 Tej = Td1 - Tdo;
|
Chris@10
|
3568 Tem = Tek + Tel;
|
Chris@10
|
3569 ci[WS(rs, 21)] = Tej - Tem;
|
Chris@10
|
3570 cr[WS(rs, 10)] = Tej + Tem;
|
Chris@10
|
3571 }
|
Chris@10
|
3572 {
|
Chris@10
|
3573 E Ter, Tey, Thz, ThE;
|
Chris@10
|
3574 Ter = Ten + Teq;
|
Chris@10
|
3575 Tey = Teu + Tex;
|
Chris@10
|
3576 ci[WS(rs, 29)] = Ter - Tey;
|
Chris@10
|
3577 cr[WS(rs, 2)] = Ter + Tey;
|
Chris@10
|
3578 Thz = TeB - TeA;
|
Chris@10
|
3579 ThE = ThA + ThD;
|
Chris@10
|
3580 cr[WS(rs, 34)] = Thz - ThE;
|
Chris@10
|
3581 ci[WS(rs, 61)] = Thz + ThE;
|
Chris@10
|
3582 }
|
Chris@10
|
3583 {
|
Chris@10
|
3584 E ThF, ThG, Tez, TeC;
|
Chris@10
|
3585 ThF = Tex - Teu;
|
Chris@10
|
3586 ThG = ThD - ThA;
|
Chris@10
|
3587 cr[WS(rs, 50)] = ThF - ThG;
|
Chris@10
|
3588 ci[WS(rs, 45)] = ThF + ThG;
|
Chris@10
|
3589 Tez = Ten - Teq;
|
Chris@10
|
3590 TeC = TeA + TeB;
|
Chris@10
|
3591 cr[WS(rs, 18)] = Tez - TeC;
|
Chris@10
|
3592 ci[WS(rs, 13)] = Tez + TeC;
|
Chris@10
|
3593 }
|
Chris@10
|
3594 }
|
Chris@10
|
3595 {
|
Chris@10
|
3596 E Tc3, Tcv, TiD, TiJ, Tca, TiI, Tcy, TiA, Tci, TcI, Tcs, TcC, Tcp, TcJ, Tct;
|
Chris@10
|
3597 E TcF;
|
Chris@10
|
3598 {
|
Chris@10
|
3599 E TbZ, Tc2, TiB, TiC;
|
Chris@10
|
3600 TbZ = Taz - TaC;
|
Chris@10
|
3601 Tc2 = Tc0 + Tc1;
|
Chris@10
|
3602 Tc3 = TbZ - Tc2;
|
Chris@10
|
3603 Tcv = TbZ + Tc2;
|
Chris@10
|
3604 TiB = TaG - TaJ;
|
Chris@10
|
3605 TiC = Tin - Tim;
|
Chris@10
|
3606 TiD = TiB + TiC;
|
Chris@10
|
3607 TiJ = TiC - TiB;
|
Chris@10
|
3608 }
|
Chris@10
|
3609 {
|
Chris@10
|
3610 E Tc6, Tcw, Tc9, Tcx;
|
Chris@10
|
3611 {
|
Chris@10
|
3612 E Tc4, Tc5, Tc7, Tc8;
|
Chris@10
|
3613 Tc4 = TaP - TaQ;
|
Chris@10
|
3614 Tc5 = TaM - TaN;
|
Chris@10
|
3615 Tc6 = FMA(KP831469612, Tc4, KP555570233 * Tc5);
|
Chris@10
|
3616 Tcw = FNMS(KP555570233, Tc4, KP831469612 * Tc5);
|
Chris@10
|
3617 Tc7 = TaW - TaX;
|
Chris@10
|
3618 Tc8 = TaT - TaU;
|
Chris@10
|
3619 Tc9 = FNMS(KP831469612, Tc8, KP555570233 * Tc7);
|
Chris@10
|
3620 Tcx = FMA(KP555570233, Tc8, KP831469612 * Tc7);
|
Chris@10
|
3621 }
|
Chris@10
|
3622 Tca = Tc6 + Tc9;
|
Chris@10
|
3623 TiI = Tcx - Tcw;
|
Chris@10
|
3624 Tcy = Tcw + Tcx;
|
Chris@10
|
3625 TiA = Tc6 - Tc9;
|
Chris@10
|
3626 }
|
Chris@10
|
3627 {
|
Chris@10
|
3628 E Tce, TcB, Tch, TcA;
|
Chris@10
|
3629 {
|
Chris@10
|
3630 E Tcc, Tcd, Tcf, Tcg;
|
Chris@10
|
3631 Tcc = Tbd - Tbe;
|
Chris@10
|
3632 Tcd = Tb7 - Tba;
|
Chris@10
|
3633 Tce = Tcc - Tcd;
|
Chris@10
|
3634 TcB = Tcc + Tcd;
|
Chris@10
|
3635 Tcf = Tb2 - Tb3;
|
Chris@10
|
3636 Tcg = Tbh - Tbg;
|
Chris@10
|
3637 Tch = Tcf - Tcg;
|
Chris@10
|
3638 TcA = Tcf + Tcg;
|
Chris@10
|
3639 }
|
Chris@10
|
3640 Tci = FMA(KP471396736, Tce, KP881921264 * Tch);
|
Chris@10
|
3641 TcI = FMA(KP956940335, TcB, KP290284677 * TcA);
|
Chris@10
|
3642 Tcs = FNMS(KP881921264, Tce, KP471396736 * Tch);
|
Chris@10
|
3643 TcC = FNMS(KP290284677, TcB, KP956940335 * TcA);
|
Chris@10
|
3644 }
|
Chris@10
|
3645 {
|
Chris@10
|
3646 E Tcl, TcE, Tco, TcD;
|
Chris@10
|
3647 {
|
Chris@10
|
3648 E Tcj, Tck, Tcm, Tcn;
|
Chris@10
|
3649 Tcj = Tbl - Tbm;
|
Chris@10
|
3650 Tck = TbA - Tbz;
|
Chris@10
|
3651 Tcl = Tcj - Tck;
|
Chris@10
|
3652 TcE = Tcj + Tck;
|
Chris@10
|
3653 Tcm = Tbw - Tbx;
|
Chris@10
|
3654 Tcn = Tbq - Tbt;
|
Chris@10
|
3655 Tco = Tcm - Tcn;
|
Chris@10
|
3656 TcD = Tcm + Tcn;
|
Chris@10
|
3657 }
|
Chris@10
|
3658 Tcp = FNMS(KP471396736, Tco, KP881921264 * Tcl);
|
Chris@10
|
3659 TcJ = FNMS(KP956940335, TcD, KP290284677 * TcE);
|
Chris@10
|
3660 Tct = FMA(KP881921264, Tco, KP471396736 * Tcl);
|
Chris@10
|
3661 TcF = FMA(KP290284677, TcD, KP956940335 * TcE);
|
Chris@10
|
3662 }
|
Chris@10
|
3663 {
|
Chris@10
|
3664 E Tcb, Tcq, TiH, TiK;
|
Chris@10
|
3665 Tcb = Tc3 + Tca;
|
Chris@10
|
3666 Tcq = Tci + Tcp;
|
Chris@10
|
3667 ci[WS(rs, 26)] = Tcb - Tcq;
|
Chris@10
|
3668 cr[WS(rs, 5)] = Tcb + Tcq;
|
Chris@10
|
3669 TiH = Tct - Tcs;
|
Chris@10
|
3670 TiK = TiI + TiJ;
|
Chris@10
|
3671 cr[WS(rs, 37)] = TiH - TiK;
|
Chris@10
|
3672 ci[WS(rs, 58)] = TiH + TiK;
|
Chris@10
|
3673 }
|
Chris@10
|
3674 {
|
Chris@10
|
3675 E TiL, TiM, Tcr, Tcu;
|
Chris@10
|
3676 TiL = Tcp - Tci;
|
Chris@10
|
3677 TiM = TiJ - TiI;
|
Chris@10
|
3678 cr[WS(rs, 53)] = TiL - TiM;
|
Chris@10
|
3679 ci[WS(rs, 42)] = TiL + TiM;
|
Chris@10
|
3680 Tcr = Tc3 - Tca;
|
Chris@10
|
3681 Tcu = Tcs + Tct;
|
Chris@10
|
3682 cr[WS(rs, 21)] = Tcr - Tcu;
|
Chris@10
|
3683 ci[WS(rs, 10)] = Tcr + Tcu;
|
Chris@10
|
3684 }
|
Chris@10
|
3685 {
|
Chris@10
|
3686 E Tcz, TcG, Tiz, TiE;
|
Chris@10
|
3687 Tcz = Tcv + Tcy;
|
Chris@10
|
3688 TcG = TcC + TcF;
|
Chris@10
|
3689 cr[WS(rs, 29)] = Tcz - TcG;
|
Chris@10
|
3690 ci[WS(rs, 2)] = Tcz + TcG;
|
Chris@10
|
3691 Tiz = TcJ - TcI;
|
Chris@10
|
3692 TiE = TiA + TiD;
|
Chris@10
|
3693 cr[WS(rs, 61)] = Tiz - TiE;
|
Chris@10
|
3694 ci[WS(rs, 34)] = Tiz + TiE;
|
Chris@10
|
3695 }
|
Chris@10
|
3696 {
|
Chris@10
|
3697 E TiF, TiG, TcH, TcK;
|
Chris@10
|
3698 TiF = TcF - TcC;
|
Chris@10
|
3699 TiG = TiD - TiA;
|
Chris@10
|
3700 cr[WS(rs, 45)] = TiF - TiG;
|
Chris@10
|
3701 ci[WS(rs, 50)] = TiF + TiG;
|
Chris@10
|
3702 TcH = Tcv - Tcy;
|
Chris@10
|
3703 TcK = TcI + TcJ;
|
Chris@10
|
3704 ci[WS(rs, 18)] = TcH - TcK;
|
Chris@10
|
3705 cr[WS(rs, 13)] = TcH + TcK;
|
Chris@10
|
3706 }
|
Chris@10
|
3707 }
|
Chris@10
|
3708 {
|
Chris@10
|
3709 E TaL, TbJ, Tip, Tiv, Tb0, Tiu, TbM, Tik, Tbk, TbW, TbG, TbQ, TbD, TbX, TbH;
|
Chris@10
|
3710 E TbT;
|
Chris@10
|
3711 {
|
Chris@10
|
3712 E TaD, TaK, Til, Tio;
|
Chris@10
|
3713 TaD = Taz + TaC;
|
Chris@10
|
3714 TaK = TaG + TaJ;
|
Chris@10
|
3715 TaL = TaD - TaK;
|
Chris@10
|
3716 TbJ = TaD + TaK;
|
Chris@10
|
3717 Til = Tc1 - Tc0;
|
Chris@10
|
3718 Tio = Tim + Tin;
|
Chris@10
|
3719 Tip = Til + Tio;
|
Chris@10
|
3720 Tiv = Tio - Til;
|
Chris@10
|
3721 }
|
Chris@10
|
3722 {
|
Chris@10
|
3723 E TaS, TbK, TaZ, TbL;
|
Chris@10
|
3724 {
|
Chris@10
|
3725 E TaO, TaR, TaV, TaY;
|
Chris@10
|
3726 TaO = TaM + TaN;
|
Chris@10
|
3727 TaR = TaP + TaQ;
|
Chris@10
|
3728 TaS = FNMS(KP980785280, TaR, KP195090322 * TaO);
|
Chris@10
|
3729 TbK = FMA(KP195090322, TaR, KP980785280 * TaO);
|
Chris@10
|
3730 TaV = TaT + TaU;
|
Chris@10
|
3731 TaY = TaW + TaX;
|
Chris@10
|
3732 TaZ = FMA(KP980785280, TaV, KP195090322 * TaY);
|
Chris@10
|
3733 TbL = FNMS(KP195090322, TaV, KP980785280 * TaY);
|
Chris@10
|
3734 }
|
Chris@10
|
3735 Tb0 = TaS + TaZ;
|
Chris@10
|
3736 Tiu = TbK - TbL;
|
Chris@10
|
3737 TbM = TbK + TbL;
|
Chris@10
|
3738 Tik = TaZ - TaS;
|
Chris@10
|
3739 }
|
Chris@10
|
3740 {
|
Chris@10
|
3741 E Tbc, TbO, Tbj, TbP;
|
Chris@10
|
3742 {
|
Chris@10
|
3743 E Tb4, Tbb, Tbf, Tbi;
|
Chris@10
|
3744 Tb4 = Tb2 + Tb3;
|
Chris@10
|
3745 Tbb = Tb7 + Tba;
|
Chris@10
|
3746 Tbc = Tb4 - Tbb;
|
Chris@10
|
3747 TbO = Tb4 + Tbb;
|
Chris@10
|
3748 Tbf = Tbd + Tbe;
|
Chris@10
|
3749 Tbi = Tbg + Tbh;
|
Chris@10
|
3750 Tbj = Tbf - Tbi;
|
Chris@10
|
3751 TbP = Tbf + Tbi;
|
Chris@10
|
3752 }
|
Chris@10
|
3753 Tbk = FMA(KP634393284, Tbc, KP773010453 * Tbj);
|
Chris@10
|
3754 TbW = FNMS(KP995184726, TbP, KP098017140 * TbO);
|
Chris@10
|
3755 TbG = FNMS(KP634393284, Tbj, KP773010453 * Tbc);
|
Chris@10
|
3756 TbQ = FMA(KP995184726, TbO, KP098017140 * TbP);
|
Chris@10
|
3757 }
|
Chris@10
|
3758 {
|
Chris@10
|
3759 E Tbv, TbR, TbC, TbS;
|
Chris@10
|
3760 {
|
Chris@10
|
3761 E Tbn, Tbu, Tby, TbB;
|
Chris@10
|
3762 Tbn = Tbl + Tbm;
|
Chris@10
|
3763 Tbu = Tbq + Tbt;
|
Chris@10
|
3764 Tbv = Tbn - Tbu;
|
Chris@10
|
3765 TbR = Tbn + Tbu;
|
Chris@10
|
3766 Tby = Tbw + Tbx;
|
Chris@10
|
3767 TbB = Tbz + TbA;
|
Chris@10
|
3768 TbC = Tby - TbB;
|
Chris@10
|
3769 TbS = Tby + TbB;
|
Chris@10
|
3770 }
|
Chris@10
|
3771 TbD = FNMS(KP773010453, TbC, KP634393284 * Tbv);
|
Chris@10
|
3772 TbX = FMA(KP098017140, TbR, KP995184726 * TbS);
|
Chris@10
|
3773 TbH = FMA(KP773010453, Tbv, KP634393284 * TbC);
|
Chris@10
|
3774 TbT = FNMS(KP098017140, TbS, KP995184726 * TbR);
|
Chris@10
|
3775 }
|
Chris@10
|
3776 {
|
Chris@10
|
3777 E Tb1, TbE, Tit, Tiw;
|
Chris@10
|
3778 Tb1 = TaL - Tb0;
|
Chris@10
|
3779 TbE = Tbk + TbD;
|
Chris@10
|
3780 ci[WS(rs, 22)] = Tb1 - TbE;
|
Chris@10
|
3781 cr[WS(rs, 9)] = Tb1 + TbE;
|
Chris@10
|
3782 Tit = TbD - Tbk;
|
Chris@10
|
3783 Tiw = Tiu + Tiv;
|
Chris@10
|
3784 cr[WS(rs, 57)] = Tit - Tiw;
|
Chris@10
|
3785 ci[WS(rs, 38)] = Tit + Tiw;
|
Chris@10
|
3786 }
|
Chris@10
|
3787 {
|
Chris@10
|
3788 E Tix, Tiy, TbF, TbI;
|
Chris@10
|
3789 Tix = TbH - TbG;
|
Chris@10
|
3790 Tiy = Tiv - Tiu;
|
Chris@10
|
3791 cr[WS(rs, 41)] = Tix - Tiy;
|
Chris@10
|
3792 ci[WS(rs, 54)] = Tix + Tiy;
|
Chris@10
|
3793 TbF = TaL + Tb0;
|
Chris@10
|
3794 TbI = TbG + TbH;
|
Chris@10
|
3795 cr[WS(rs, 25)] = TbF - TbI;
|
Chris@10
|
3796 ci[WS(rs, 6)] = TbF + TbI;
|
Chris@10
|
3797 }
|
Chris@10
|
3798 {
|
Chris@10
|
3799 E TbN, TbU, Tij, Tiq;
|
Chris@10
|
3800 TbN = TbJ + TbM;
|
Chris@10
|
3801 TbU = TbQ + TbT;
|
Chris@10
|
3802 ci[WS(rs, 30)] = TbN - TbU;
|
Chris@10
|
3803 cr[WS(rs, 1)] = TbN + TbU;
|
Chris@10
|
3804 Tij = TbX - TbW;
|
Chris@10
|
3805 Tiq = Tik + Tip;
|
Chris@10
|
3806 cr[WS(rs, 33)] = Tij - Tiq;
|
Chris@10
|
3807 ci[WS(rs, 62)] = Tij + Tiq;
|
Chris@10
|
3808 }
|
Chris@10
|
3809 {
|
Chris@10
|
3810 E Tir, Tis, TbV, TbY;
|
Chris@10
|
3811 Tir = TbT - TbQ;
|
Chris@10
|
3812 Tis = Tip - Tik;
|
Chris@10
|
3813 cr[WS(rs, 49)] = Tir - Tis;
|
Chris@10
|
3814 ci[WS(rs, 46)] = Tir + Tis;
|
Chris@10
|
3815 TbV = TbJ - TbM;
|
Chris@10
|
3816 TbY = TbW + TbX;
|
Chris@10
|
3817 cr[WS(rs, 17)] = TbV - TbY;
|
Chris@10
|
3818 ci[WS(rs, 14)] = TbV + TbY;
|
Chris@10
|
3819 }
|
Chris@10
|
3820 }
|
Chris@10
|
3821 {
|
Chris@10
|
3822 E T9R, Taj, Ti9, Tif, T9Y, Tie, Tam, Ti6, Ta6, Taw, Tag, Taq, Tad, Tax, Tah;
|
Chris@10
|
3823 E Tat;
|
Chris@10
|
3824 {
|
Chris@10
|
3825 E T9N, T9Q, Ti7, Ti8;
|
Chris@10
|
3826 T9N = T6b - T6m;
|
Chris@10
|
3827 T9Q = T9O + T9P;
|
Chris@10
|
3828 T9R = T9N - T9Q;
|
Chris@10
|
3829 Taj = T9N + T9Q;
|
Chris@10
|
3830 Ti7 = T6J - T6y;
|
Chris@10
|
3831 Ti8 = ThT - ThQ;
|
Chris@10
|
3832 Ti9 = Ti7 + Ti8;
|
Chris@10
|
3833 Tif = Ti8 - Ti7;
|
Chris@10
|
3834 }
|
Chris@10
|
3835 {
|
Chris@10
|
3836 E T9U, Tak, T9X, Tal;
|
Chris@10
|
3837 {
|
Chris@10
|
3838 E T9S, T9T, T9V, T9W;
|
Chris@10
|
3839 T9S = T6Q - T71;
|
Chris@10
|
3840 T9T = T77 - T7a;
|
Chris@10
|
3841 T9U = FNMS(KP831469612, T9T, KP555570233 * T9S);
|
Chris@10
|
3842 Tak = FMA(KP831469612, T9S, KP555570233 * T9T);
|
Chris@10
|
3843 T9V = T7h - T7s;
|
Chris@10
|
3844 T9W = T7y - T7B;
|
Chris@10
|
3845 T9X = FMA(KP555570233, T9V, KP831469612 * T9W);
|
Chris@10
|
3846 Tal = FNMS(KP555570233, T9W, KP831469612 * T9V);
|
Chris@10
|
3847 }
|
Chris@10
|
3848 T9Y = T9U + T9X;
|
Chris@10
|
3849 Tie = Tak - Tal;
|
Chris@10
|
3850 Tam = Tak + Tal;
|
Chris@10
|
3851 Ti6 = T9X - T9U;
|
Chris@10
|
3852 }
|
Chris@10
|
3853 {
|
Chris@10
|
3854 E Ta2, Tao, Ta5, Tap;
|
Chris@10
|
3855 {
|
Chris@10
|
3856 E Ta0, Ta1, Ta3, Ta4;
|
Chris@10
|
3857 Ta0 = T8p - T8s;
|
Chris@10
|
3858 Ta1 = T87 - T8i;
|
Chris@10
|
3859 Ta2 = Ta0 - Ta1;
|
Chris@10
|
3860 Tao = Ta0 + Ta1;
|
Chris@10
|
3861 Ta3 = T7K - T7V;
|
Chris@10
|
3862 Ta4 = T8v - T8u;
|
Chris@10
|
3863 Ta5 = Ta3 - Ta4;
|
Chris@10
|
3864 Tap = Ta3 + Ta4;
|
Chris@10
|
3865 }
|
Chris@10
|
3866 Ta6 = FMA(KP471396736, Ta2, KP881921264 * Ta5);
|
Chris@10
|
3867 Taw = FNMS(KP956940335, Tap, KP290284677 * Tao);
|
Chris@10
|
3868 Tag = FNMS(KP471396736, Ta5, KP881921264 * Ta2);
|
Chris@10
|
3869 Taq = FMA(KP956940335, Tao, KP290284677 * Tap);
|
Chris@10
|
3870 }
|
Chris@10
|
3871 {
|
Chris@10
|
3872 E Ta9, Tar, Tac, Tas;
|
Chris@10
|
3873 {
|
Chris@10
|
3874 E Ta7, Ta8, Taa, Tab;
|
Chris@10
|
3875 Ta7 = T8D - T8O;
|
Chris@10
|
3876 Ta8 = T9n - T9o;
|
Chris@10
|
3877 Ta9 = Ta7 - Ta8;
|
Chris@10
|
3878 Tar = Ta7 + Ta8;
|
Chris@10
|
3879 Taa = T9i - T9l;
|
Chris@10
|
3880 Tab = T9b - T90;
|
Chris@10
|
3881 Tac = Taa - Tab;
|
Chris@10
|
3882 Tas = Taa + Tab;
|
Chris@10
|
3883 }
|
Chris@10
|
3884 Tad = FNMS(KP881921264, Tac, KP471396736 * Ta9);
|
Chris@10
|
3885 Tax = FMA(KP290284677, Tar, KP956940335 * Tas);
|
Chris@10
|
3886 Tah = FMA(KP881921264, Ta9, KP471396736 * Tac);
|
Chris@10
|
3887 Tat = FNMS(KP290284677, Tas, KP956940335 * Tar);
|
Chris@10
|
3888 }
|
Chris@10
|
3889 {
|
Chris@10
|
3890 E T9Z, Tae, Tid, Tig;
|
Chris@10
|
3891 T9Z = T9R - T9Y;
|
Chris@10
|
3892 Tae = Ta6 + Tad;
|
Chris@10
|
3893 ci[WS(rs, 20)] = T9Z - Tae;
|
Chris@10
|
3894 cr[WS(rs, 11)] = T9Z + Tae;
|
Chris@10
|
3895 Tid = Tad - Ta6;
|
Chris@10
|
3896 Tig = Tie + Tif;
|
Chris@10
|
3897 cr[WS(rs, 59)] = Tid - Tig;
|
Chris@10
|
3898 ci[WS(rs, 36)] = Tid + Tig;
|
Chris@10
|
3899 }
|
Chris@10
|
3900 {
|
Chris@10
|
3901 E Tih, Tii, Taf, Tai;
|
Chris@10
|
3902 Tih = Tah - Tag;
|
Chris@10
|
3903 Tii = Tif - Tie;
|
Chris@10
|
3904 cr[WS(rs, 43)] = Tih - Tii;
|
Chris@10
|
3905 ci[WS(rs, 52)] = Tih + Tii;
|
Chris@10
|
3906 Taf = T9R + T9Y;
|
Chris@10
|
3907 Tai = Tag + Tah;
|
Chris@10
|
3908 cr[WS(rs, 27)] = Taf - Tai;
|
Chris@10
|
3909 ci[WS(rs, 4)] = Taf + Tai;
|
Chris@10
|
3910 }
|
Chris@10
|
3911 {
|
Chris@10
|
3912 E Tan, Tau, Ti5, Tia;
|
Chris@10
|
3913 Tan = Taj + Tam;
|
Chris@10
|
3914 Tau = Taq + Tat;
|
Chris@10
|
3915 ci[WS(rs, 28)] = Tan - Tau;
|
Chris@10
|
3916 cr[WS(rs, 3)] = Tan + Tau;
|
Chris@10
|
3917 Ti5 = Tax - Taw;
|
Chris@10
|
3918 Tia = Ti6 + Ti9;
|
Chris@10
|
3919 cr[WS(rs, 35)] = Ti5 - Tia;
|
Chris@10
|
3920 ci[WS(rs, 60)] = Ti5 + Tia;
|
Chris@10
|
3921 }
|
Chris@10
|
3922 {
|
Chris@10
|
3923 E Tib, Tic, Tav, Tay;
|
Chris@10
|
3924 Tib = Tat - Taq;
|
Chris@10
|
3925 Tic = Ti9 - Ti6;
|
Chris@10
|
3926 cr[WS(rs, 51)] = Tib - Tic;
|
Chris@10
|
3927 ci[WS(rs, 44)] = Tib + Tic;
|
Chris@10
|
3928 Tav = Taj - Tam;
|
Chris@10
|
3929 Tay = Taw + Tax;
|
Chris@10
|
3930 cr[WS(rs, 19)] = Tav - Tay;
|
Chris@10
|
3931 ci[WS(rs, 12)] = Tav + Tay;
|
Chris@10
|
3932 }
|
Chris@10
|
3933 }
|
Chris@10
|
3934 }
|
Chris@10
|
3935 }
|
Chris@10
|
3936 }
|
Chris@10
|
3937
|
Chris@10
|
3938 static const tw_instr twinstr[] = {
|
Chris@10
|
3939 {TW_FULL, 1, 64},
|
Chris@10
|
3940 {TW_NEXT, 1, 0}
|
Chris@10
|
3941 };
|
Chris@10
|
3942
|
Chris@10
|
3943 static const hc2hc_desc desc = { 64, "hf_64", twinstr, &GENUS, {808, 270, 230, 0} };
|
Chris@10
|
3944
|
Chris@10
|
3945 void X(codelet_hf_64) (planner *p) {
|
Chris@10
|
3946 X(khc2hc_register) (p, hf_64, &desc);
|
Chris@10
|
3947 }
|
Chris@10
|
3948 #endif /* HAVE_FMA */
|