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 07:35:52 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-dft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 64 -name t1_64 -include t.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 * 228 stack variables, 15 constants, and 256 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "t.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void t1_64(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@10
|
40 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@10
|
41 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@10
|
42 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@10
|
43 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
Chris@10
|
44 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
Chris@10
|
45 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
Chris@10
|
46 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
Chris@10
|
47 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
48 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
49 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@10
|
50 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@10
|
51 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
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 * 126); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 126, MAKE_VOLATILE_STRIDE(128, rs)) {
|
Chris@10
|
57 E TeI, Tkk, Tkj, TeL;
|
Chris@10
|
58 {
|
Chris@10
|
59 E TiV, Tjm, T7e, TcA, TjR, Tkl, Tm, TeM, TeZ, Ths, T7Q, TcJ, T1G, TeW, TcI;
|
Chris@10
|
60 E T7X, Tf5, Thv, T87, TcN, T29, Tf8, TcQ, T8u, TfU, ThS, Taq, Tdm, T5K, Tg9;
|
Chris@10
|
61 E Tdx, Tbj, TcB, T7l, TiP, TeP, Tjl, TN, TcC, T7s, T7I, TcF, TeU, Thr, T7B;
|
Chris@10
|
62 E TcG, T1f, TeR, Tfg, ThB, T8G, TcU, T32, Tfj, TcX, T93, Tft, ThH, T9h, Td3;
|
Chris@10
|
63 E T3X, TfI, Tde, Taa, Thw, Tfb, Tf6, T2A, T8x, TcO, T8m, TcR, Tfm, ThC, T3t;
|
Chris@10
|
64 E Tfh, T96, TcV, T8V, TcY, ThI, TfL, Tfu, T4o, Tad, Td4, T9w, Tdf, Tgc, ThT;
|
Chris@10
|
65 E T6b, TfV, Tbm, Tdn, TaF, Tdy, ThN, T4Q, TfN, TfA, Taf, Ta1, Td8, Tdh, ThO;
|
Chris@10
|
66 E T5h, TfO, TfF, Tag, T9M, Tdb, Tdi, ThY, T6D, Tge, Tg1, Tbo, Tba, Tdr, TdA;
|
Chris@10
|
67 E TaN, Tdt, Tg5, ThZ, Tg2, T74, Tds, TaU;
|
Chris@10
|
68 {
|
Chris@10
|
69 E T7a, Te, T78, T8, TjP, TiU, T7c, Tk;
|
Chris@10
|
70 {
|
Chris@10
|
71 E T1, TiT, TiS, T7, Tg, Tj, Tf, Ti, T7b, Th;
|
Chris@10
|
72 T1 = ri[0];
|
Chris@10
|
73 TiT = ii[0];
|
Chris@10
|
74 {
|
Chris@10
|
75 E T3, T6, T2, T5;
|
Chris@10
|
76 T3 = ri[WS(rs, 32)];
|
Chris@10
|
77 T6 = ii[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 = ri[WS(rs, 16)];
|
Chris@10
|
83 Td = ii[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 = ri[WS(rs, 48)];
|
Chris@10
|
93 Tj = ii[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 TjP = 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 T7L, T1l, T7V, T1E, T1u, T1x, T1w, T7N, T1r, T7S, T1v;
|
Chris@10
|
111 {
|
Chris@10
|
112 E T1A, T1D, T1C, T7U, T1B;
|
Chris@10
|
113 {
|
Chris@10
|
114 E T1h, T1k, T1g, T1j, T7K, T1i, T1z;
|
Chris@10
|
115 T1h = ri[WS(rs, 60)];
|
Chris@10
|
116 T1k = ii[WS(rs, 60)];
|
Chris@10
|
117 {
|
Chris@10
|
118 E T7d, TiQ, Tl, TjQ;
|
Chris@10
|
119 T7d = T7a - T7c;
|
Chris@10
|
120 TiQ = T7a + T7c;
|
Chris@10
|
121 Tl = Te + Tk;
|
Chris@10
|
122 TjQ = 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 = ri[WS(rs, 44)];
|
Chris@10
|
135 T1D = ii[WS(rs, 44)];
|
Chris@10
|
136 T7K = T1g * T1k;
|
Chris@10
|
137 T1i = T1g * T1h;
|
Chris@10
|
138 T1z = W[86];
|
Chris@10
|
139 T1C = W[87];
|
Chris@10
|
140 T7L = FNMS(T1j, T1h, T7K);
|
Chris@10
|
141 T1l = FMA(T1j, T1k, T1i);
|
Chris@10
|
142 T7U = T1z * T1D;
|
Chris@10
|
143 T1B = T1z * T1A;
|
Chris@10
|
144 }
|
Chris@10
|
145 {
|
Chris@10
|
146 E T1n, T1q, T1m, T1p, T7M, T1o, T1t;
|
Chris@10
|
147 T1n = ri[WS(rs, 28)];
|
Chris@10
|
148 T1q = ii[WS(rs, 28)];
|
Chris@10
|
149 T7V = FNMS(T1C, T1A, T7U);
|
Chris@10
|
150 T1E = FMA(T1C, T1D, T1B);
|
Chris@10
|
151 T1m = W[54];
|
Chris@10
|
152 T1p = W[55];
|
Chris@10
|
153 T1u = ri[WS(rs, 12)];
|
Chris@10
|
154 T1x = ii[WS(rs, 12)];
|
Chris@10
|
155 T7M = T1m * T1q;
|
Chris@10
|
156 T1o = T1m * T1n;
|
Chris@10
|
157 T1t = W[22];
|
Chris@10
|
158 T1w = W[23];
|
Chris@10
|
159 T7N = FNMS(T1p, T1n, T7M);
|
Chris@10
|
160 T1r = FMA(T1p, T1q, T1o);
|
Chris@10
|
161 T7S = T1t * T1x;
|
Chris@10
|
162 T1v = T1t * T1u;
|
Chris@10
|
163 }
|
Chris@10
|
164 }
|
Chris@10
|
165 {
|
Chris@10
|
166 E T7O, TeX, T1s, T7R, T7T, T1y;
|
Chris@10
|
167 T7O = T7L - T7N;
|
Chris@10
|
168 TeX = T7L + T7N;
|
Chris@10
|
169 T1s = T1l + T1r;
|
Chris@10
|
170 T7R = T1l - T1r;
|
Chris@10
|
171 T7T = FNMS(T1w, T1u, T7S);
|
Chris@10
|
172 T1y = FMA(T1w, T1x, T1v);
|
Chris@10
|
173 {
|
Chris@10
|
174 E T7W, TeY, T7P, T1F;
|
Chris@10
|
175 T7W = T7T - T7V;
|
Chris@10
|
176 TeY = T7T + T7V;
|
Chris@10
|
177 T7P = T1y - T1E;
|
Chris@10
|
178 T1F = T1y + T1E;
|
Chris@10
|
179 TeZ = TeX - TeY;
|
Chris@10
|
180 Ths = TeX + TeY;
|
Chris@10
|
181 T7Q = T7O + T7P;
|
Chris@10
|
182 TcJ = T7O - T7P;
|
Chris@10
|
183 T1G = T1s + T1F;
|
Chris@10
|
184 TeW = T1s - T1F;
|
Chris@10
|
185 TcI = T7R + T7W;
|
Chris@10
|
186 T7X = T7R - T7W;
|
Chris@10
|
187 }
|
Chris@10
|
188 }
|
Chris@10
|
189 }
|
Chris@10
|
190 }
|
Chris@10
|
191 {
|
Chris@10
|
192 E T82, T1O, T8s, T27, T1X, T20, T1Z, T84, T1U, T8p, T1Y;
|
Chris@10
|
193 {
|
Chris@10
|
194 E T23, T26, T25, T8r, T24;
|
Chris@10
|
195 {
|
Chris@10
|
196 E T1K, T1N, T1J, T1M, T81, T1L, T22;
|
Chris@10
|
197 T1K = ri[WS(rs, 2)];
|
Chris@10
|
198 T1N = ii[WS(rs, 2)];
|
Chris@10
|
199 T1J = W[2];
|
Chris@10
|
200 T1M = W[3];
|
Chris@10
|
201 T23 = ri[WS(rs, 50)];
|
Chris@10
|
202 T26 = ii[WS(rs, 50)];
|
Chris@10
|
203 T81 = T1J * T1N;
|
Chris@10
|
204 T1L = T1J * T1K;
|
Chris@10
|
205 T22 = W[98];
|
Chris@10
|
206 T25 = W[99];
|
Chris@10
|
207 T82 = FNMS(T1M, T1K, T81);
|
Chris@10
|
208 T1O = FMA(T1M, T1N, T1L);
|
Chris@10
|
209 T8r = T22 * T26;
|
Chris@10
|
210 T24 = T22 * T23;
|
Chris@10
|
211 }
|
Chris@10
|
212 {
|
Chris@10
|
213 E T1Q, T1T, T1P, T1S, T83, T1R, T1W;
|
Chris@10
|
214 T1Q = ri[WS(rs, 34)];
|
Chris@10
|
215 T1T = ii[WS(rs, 34)];
|
Chris@10
|
216 T8s = FNMS(T25, T23, T8r);
|
Chris@10
|
217 T27 = FMA(T25, T26, T24);
|
Chris@10
|
218 T1P = W[66];
|
Chris@10
|
219 T1S = W[67];
|
Chris@10
|
220 T1X = ri[WS(rs, 18)];
|
Chris@10
|
221 T20 = ii[WS(rs, 18)];
|
Chris@10
|
222 T83 = T1P * T1T;
|
Chris@10
|
223 T1R = T1P * T1Q;
|
Chris@10
|
224 T1W = W[34];
|
Chris@10
|
225 T1Z = W[35];
|
Chris@10
|
226 T84 = FNMS(T1S, T1Q, T83);
|
Chris@10
|
227 T1U = FMA(T1S, T1T, T1R);
|
Chris@10
|
228 T8p = T1W * T20;
|
Chris@10
|
229 T1Y = T1W * T1X;
|
Chris@10
|
230 }
|
Chris@10
|
231 }
|
Chris@10
|
232 {
|
Chris@10
|
233 E T85, Tf3, T1V, T8o, T8q, T21;
|
Chris@10
|
234 T85 = T82 - T84;
|
Chris@10
|
235 Tf3 = T82 + T84;
|
Chris@10
|
236 T1V = T1O + T1U;
|
Chris@10
|
237 T8o = T1O - T1U;
|
Chris@10
|
238 T8q = FNMS(T1Z, T1X, T8p);
|
Chris@10
|
239 T21 = FMA(T1Z, T20, T1Y);
|
Chris@10
|
240 {
|
Chris@10
|
241 E T8t, Tf4, T86, T28;
|
Chris@10
|
242 T8t = T8q - T8s;
|
Chris@10
|
243 Tf4 = T8q + T8s;
|
Chris@10
|
244 T86 = T21 - T27;
|
Chris@10
|
245 T28 = T21 + T27;
|
Chris@10
|
246 Tf5 = Tf3 - Tf4;
|
Chris@10
|
247 Thv = Tf3 + Tf4;
|
Chris@10
|
248 T87 = T85 + T86;
|
Chris@10
|
249 TcN = T85 - T86;
|
Chris@10
|
250 T29 = T1V + T28;
|
Chris@10
|
251 Tf8 = T1V - T28;
|
Chris@10
|
252 TcQ = T8o + T8t;
|
Chris@10
|
253 T8u = T8o - T8t;
|
Chris@10
|
254 }
|
Chris@10
|
255 }
|
Chris@10
|
256 }
|
Chris@10
|
257 {
|
Chris@10
|
258 E Tal, T5p, Tbh, T5I, T5y, T5B, T5A, Tan, T5v, Tbe, T5z;
|
Chris@10
|
259 {
|
Chris@10
|
260 E T5E, T5H, T5G, Tbg, T5F;
|
Chris@10
|
261 {
|
Chris@10
|
262 E T5l, T5o, T5k, T5n, Tak, T5m, T5D;
|
Chris@10
|
263 T5l = ri[WS(rs, 63)];
|
Chris@10
|
264 T5o = ii[WS(rs, 63)];
|
Chris@10
|
265 T5k = W[124];
|
Chris@10
|
266 T5n = W[125];
|
Chris@10
|
267 T5E = ri[WS(rs, 47)];
|
Chris@10
|
268 T5H = ii[WS(rs, 47)];
|
Chris@10
|
269 Tak = T5k * T5o;
|
Chris@10
|
270 T5m = T5k * T5l;
|
Chris@10
|
271 T5D = W[92];
|
Chris@10
|
272 T5G = W[93];
|
Chris@10
|
273 Tal = FNMS(T5n, T5l, Tak);
|
Chris@10
|
274 T5p = FMA(T5n, T5o, T5m);
|
Chris@10
|
275 Tbg = T5D * T5H;
|
Chris@10
|
276 T5F = T5D * T5E;
|
Chris@10
|
277 }
|
Chris@10
|
278 {
|
Chris@10
|
279 E T5r, T5u, T5q, T5t, Tam, T5s, T5x;
|
Chris@10
|
280 T5r = ri[WS(rs, 31)];
|
Chris@10
|
281 T5u = ii[WS(rs, 31)];
|
Chris@10
|
282 Tbh = FNMS(T5G, T5E, Tbg);
|
Chris@10
|
283 T5I = FMA(T5G, T5H, T5F);
|
Chris@10
|
284 T5q = W[60];
|
Chris@10
|
285 T5t = W[61];
|
Chris@10
|
286 T5y = ri[WS(rs, 15)];
|
Chris@10
|
287 T5B = ii[WS(rs, 15)];
|
Chris@10
|
288 Tam = T5q * T5u;
|
Chris@10
|
289 T5s = T5q * T5r;
|
Chris@10
|
290 T5x = W[28];
|
Chris@10
|
291 T5A = W[29];
|
Chris@10
|
292 Tan = FNMS(T5t, T5r, Tam);
|
Chris@10
|
293 T5v = FMA(T5t, T5u, T5s);
|
Chris@10
|
294 Tbe = T5x * T5B;
|
Chris@10
|
295 T5z = T5x * T5y;
|
Chris@10
|
296 }
|
Chris@10
|
297 }
|
Chris@10
|
298 {
|
Chris@10
|
299 E Tao, TfS, T5w, Tbd, Tbf, T5C;
|
Chris@10
|
300 Tao = Tal - Tan;
|
Chris@10
|
301 TfS = Tal + Tan;
|
Chris@10
|
302 T5w = T5p + T5v;
|
Chris@10
|
303 Tbd = T5p - T5v;
|
Chris@10
|
304 Tbf = FNMS(T5A, T5y, Tbe);
|
Chris@10
|
305 T5C = FMA(T5A, T5B, T5z);
|
Chris@10
|
306 {
|
Chris@10
|
307 E Tbi, TfT, Tap, T5J;
|
Chris@10
|
308 Tbi = Tbf - Tbh;
|
Chris@10
|
309 TfT = Tbf + Tbh;
|
Chris@10
|
310 Tap = T5C - T5I;
|
Chris@10
|
311 T5J = T5C + T5I;
|
Chris@10
|
312 TfU = TfS - TfT;
|
Chris@10
|
313 ThS = TfS + TfT;
|
Chris@10
|
314 Taq = Tao + Tap;
|
Chris@10
|
315 Tdm = Tao - Tap;
|
Chris@10
|
316 T5K = T5w + T5J;
|
Chris@10
|
317 Tg9 = T5w - T5J;
|
Chris@10
|
318 Tdx = Tbd + Tbi;
|
Chris@10
|
319 Tbj = Tbd - Tbi;
|
Chris@10
|
320 }
|
Chris@10
|
321 }
|
Chris@10
|
322 }
|
Chris@10
|
323 {
|
Chris@10
|
324 E T7G, T1d, T7z, TeS, T11, T7C, T7E, T17, T7r, T7m;
|
Chris@10
|
325 {
|
Chris@10
|
326 E T7g, Ts, T7q, TL, TB, TE, TD, T7i, 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, T7f, Tp, TG;
|
Chris@10
|
331 To = ri[WS(rs, 8)];
|
Chris@10
|
332 Tr = ii[WS(rs, 8)];
|
Chris@10
|
333 Tn = W[14];
|
Chris@10
|
334 Tq = W[15];
|
Chris@10
|
335 TH = ri[WS(rs, 24)];
|
Chris@10
|
336 TK = ii[WS(rs, 24)];
|
Chris@10
|
337 T7f = Tn * Tr;
|
Chris@10
|
338 Tp = Tn * To;
|
Chris@10
|
339 TG = W[46];
|
Chris@10
|
340 TJ = W[47];
|
Chris@10
|
341 T7g = FNMS(Tq, To, T7f);
|
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, T7h, Tv, TA;
|
Chris@10
|
348 Tu = ri[WS(rs, 40)];
|
Chris@10
|
349 Tx = ii[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 = ri[WS(rs, 56)];
|
Chris@10
|
355 TE = ii[WS(rs, 56)];
|
Chris@10
|
356 T7h = Tt * Tx;
|
Chris@10
|
357 Tv = Tt * Tu;
|
Chris@10
|
358 TA = W[110];
|
Chris@10
|
359 TD = W[111];
|
Chris@10
|
360 T7i = FNMS(Tw, Tu, T7h);
|
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 T7j, TeN, Tz, T7k, T7o, TF, TeO, TM;
|
Chris@10
|
368 T7j = T7g - T7i;
|
Chris@10
|
369 TeN = T7g + T7i;
|
Chris@10
|
370 Tz = Ts + Ty;
|
Chris@10
|
371 T7k = 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 TeO = T7o + T7q;
|
Chris@10
|
376 TM = TF + TL;
|
Chris@10
|
377 T7m = TF - TL;
|
Chris@10
|
378 TcB = T7k + T7j;
|
Chris@10
|
379 T7l = T7j - T7k;
|
Chris@10
|
380 TiP = TeN + TeO;
|
Chris@10
|
381 TeP = TeN - TeO;
|
Chris@10
|
382 Tjl = TM - Tz;
|
Chris@10
|
383 TN = Tz + TM;
|
Chris@10
|
384 }
|
Chris@10
|
385 }
|
Chris@10
|
386 {
|
Chris@10
|
387 E T7w, TU, T13, T16, T7y, T10, T12, T15, T7D, T14;
|
Chris@10
|
388 {
|
Chris@10
|
389 E T19, T1c, T18, T1b;
|
Chris@10
|
390 {
|
Chris@10
|
391 E TQ, TT, TS, T7v, TR, TP;
|
Chris@10
|
392 TQ = ri[WS(rs, 4)];
|
Chris@10
|
393 TT = ii[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 T7v = TP * TT;
|
Chris@10
|
399 TR = TP * TQ;
|
Chris@10
|
400 T19 = ri[WS(rs, 52)];
|
Chris@10
|
401 T1c = ii[WS(rs, 52)];
|
Chris@10
|
402 T7w = FNMS(TS, TQ, T7v);
|
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, T7x, TX, T7F, T1a, TV;
|
Chris@10
|
409 TW = ri[WS(rs, 36)];
|
Chris@10
|
410 TZ = ii[WS(rs, 36)];
|
Chris@10
|
411 T7F = T18 * T1c;
|
Chris@10
|
412 T1a = T18 * T19;
|
Chris@10
|
413 TV = W[70];
|
Chris@10
|
414 TY = W[71];
|
Chris@10
|
415 T7G = FNMS(T1b, T19, T7F);
|
Chris@10
|
416 T1d = FMA(T1b, T1c, T1a);
|
Chris@10
|
417 T7x = TV * TZ;
|
Chris@10
|
418 TX = TV * TW;
|
Chris@10
|
419 T13 = ri[WS(rs, 20)];
|
Chris@10
|
420 T16 = ii[WS(rs, 20)];
|
Chris@10
|
421 T7y = FNMS(TY, TW, T7x);
|
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 T7z = T7w - T7y;
|
Chris@10
|
428 TeS = T7w + T7y;
|
Chris@10
|
429 T11 = TU + T10;
|
Chris@10
|
430 T7C = TU - T10;
|
Chris@10
|
431 T7D = T12 * T16;
|
Chris@10
|
432 T14 = T12 * T13;
|
Chris@10
|
433 T7E = FNMS(T15, T13, T7D);
|
Chris@10
|
434 T17 = FMA(T15, T16, T14);
|
Chris@10
|
435 }
|
Chris@10
|
436 {
|
Chris@10
|
437 E T8B, T2H, T91, T30, T2Q, T2T, T2S, T8D, T2N, T8Y, T2R;
|
Chris@10
|
438 {
|
Chris@10
|
439 E T2W, T2Z, T2Y, T90, T2X;
|
Chris@10
|
440 {
|
Chris@10
|
441 E T2D, T2G, T2C, T2F, T8A, T2E, T2V;
|
Chris@10
|
442 T2D = ri[WS(rs, 62)];
|
Chris@10
|
443 T2G = ii[WS(rs, 62)];
|
Chris@10
|
444 {
|
Chris@10
|
445 E TeT, T7H, T1e, T7A;
|
Chris@10
|
446 TeT = T7E + T7G;
|
Chris@10
|
447 T7H = T7E - T7G;
|
Chris@10
|
448 T1e = T17 + T1d;
|
Chris@10
|
449 T7A = T17 - T1d;
|
Chris@10
|
450 T7I = T7C - T7H;
|
Chris@10
|
451 TcF = T7C + T7H;
|
Chris@10
|
452 TeU = TeS - TeT;
|
Chris@10
|
453 Thr = TeS + TeT;
|
Chris@10
|
454 T7B = T7z + T7A;
|
Chris@10
|
455 TcG = T7z - T7A;
|
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 = ri[WS(rs, 46)];
|
Chris@10
|
462 T2Z = ii[WS(rs, 46)];
|
Chris@10
|
463 T8A = T2C * T2G;
|
Chris@10
|
464 T2E = T2C * T2D;
|
Chris@10
|
465 T2V = W[90];
|
Chris@10
|
466 T2Y = W[91];
|
Chris@10
|
467 T8B = FNMS(T2F, T2D, T8A);
|
Chris@10
|
468 T2H = FMA(T2F, T2G, T2E);
|
Chris@10
|
469 T90 = T2V * T2Z;
|
Chris@10
|
470 T2X = T2V * T2W;
|
Chris@10
|
471 }
|
Chris@10
|
472 {
|
Chris@10
|
473 E T2J, T2M, T2I, T2L, T8C, T2K, T2P;
|
Chris@10
|
474 T2J = ri[WS(rs, 30)];
|
Chris@10
|
475 T2M = ii[WS(rs, 30)];
|
Chris@10
|
476 T91 = FNMS(T2Y, T2W, T90);
|
Chris@10
|
477 T30 = FMA(T2Y, T2Z, T2X);
|
Chris@10
|
478 T2I = W[58];
|
Chris@10
|
479 T2L = W[59];
|
Chris@10
|
480 T2Q = ri[WS(rs, 14)];
|
Chris@10
|
481 T2T = ii[WS(rs, 14)];
|
Chris@10
|
482 T8C = T2I * T2M;
|
Chris@10
|
483 T2K = T2I * T2J;
|
Chris@10
|
484 T2P = W[26];
|
Chris@10
|
485 T2S = W[27];
|
Chris@10
|
486 T8D = FNMS(T2L, T2J, T8C);
|
Chris@10
|
487 T2N = FMA(T2L, T2M, T2K);
|
Chris@10
|
488 T8Y = T2P * T2T;
|
Chris@10
|
489 T2R = T2P * T2Q;
|
Chris@10
|
490 }
|
Chris@10
|
491 }
|
Chris@10
|
492 {
|
Chris@10
|
493 E T8E, Tfe, T2O, T8X, T8Z, T2U;
|
Chris@10
|
494 T8E = T8B - T8D;
|
Chris@10
|
495 Tfe = T8B + T8D;
|
Chris@10
|
496 T2O = T2H + T2N;
|
Chris@10
|
497 T8X = T2H - T2N;
|
Chris@10
|
498 T8Z = FNMS(T2S, T2Q, T8Y);
|
Chris@10
|
499 T2U = FMA(T2S, T2T, T2R);
|
Chris@10
|
500 {
|
Chris@10
|
501 E T92, Tff, T8F, T31;
|
Chris@10
|
502 T92 = T8Z - T91;
|
Chris@10
|
503 Tff = T8Z + T91;
|
Chris@10
|
504 T8F = T2U - T30;
|
Chris@10
|
505 T31 = T2U + T30;
|
Chris@10
|
506 Tfg = Tfe - Tff;
|
Chris@10
|
507 ThB = Tfe + Tff;
|
Chris@10
|
508 T8G = T8E + T8F;
|
Chris@10
|
509 TcU = T8E - T8F;
|
Chris@10
|
510 T32 = T2O + T31;
|
Chris@10
|
511 Tfj = T2O - T31;
|
Chris@10
|
512 TcX = T8X + T92;
|
Chris@10
|
513 T93 = T8X - T92;
|
Chris@10
|
514 }
|
Chris@10
|
515 }
|
Chris@10
|
516 }
|
Chris@10
|
517 {
|
Chris@10
|
518 E T9c, T3C, Ta8, T3V, T3L, T3O, T3N, T9e, T3I, Ta5, T3M;
|
Chris@10
|
519 {
|
Chris@10
|
520 E T3R, T3U, T3T, Ta7, T3S;
|
Chris@10
|
521 {
|
Chris@10
|
522 E T3y, T3B, T3x, T3A, T9b, T3z, T3Q;
|
Chris@10
|
523 T3y = ri[WS(rs, 1)];
|
Chris@10
|
524 T3B = ii[WS(rs, 1)];
|
Chris@10
|
525 T3x = W[0];
|
Chris@10
|
526 T3A = W[1];
|
Chris@10
|
527 T3R = ri[WS(rs, 49)];
|
Chris@10
|
528 T3U = ii[WS(rs, 49)];
|
Chris@10
|
529 T9b = T3x * T3B;
|
Chris@10
|
530 T3z = T3x * T3y;
|
Chris@10
|
531 T3Q = W[96];
|
Chris@10
|
532 T3T = W[97];
|
Chris@10
|
533 T9c = FNMS(T3A, T3y, T9b);
|
Chris@10
|
534 T3C = FMA(T3A, T3B, T3z);
|
Chris@10
|
535 Ta7 = T3Q * T3U;
|
Chris@10
|
536 T3S = T3Q * T3R;
|
Chris@10
|
537 }
|
Chris@10
|
538 {
|
Chris@10
|
539 E T3E, T3H, T3D, T3G, T9d, T3F, T3K;
|
Chris@10
|
540 T3E = ri[WS(rs, 33)];
|
Chris@10
|
541 T3H = ii[WS(rs, 33)];
|
Chris@10
|
542 Ta8 = FNMS(T3T, T3R, Ta7);
|
Chris@10
|
543 T3V = FMA(T3T, T3U, T3S);
|
Chris@10
|
544 T3D = W[64];
|
Chris@10
|
545 T3G = W[65];
|
Chris@10
|
546 T3L = ri[WS(rs, 17)];
|
Chris@10
|
547 T3O = ii[WS(rs, 17)];
|
Chris@10
|
548 T9d = T3D * T3H;
|
Chris@10
|
549 T3F = T3D * T3E;
|
Chris@10
|
550 T3K = W[32];
|
Chris@10
|
551 T3N = W[33];
|
Chris@10
|
552 T9e = FNMS(T3G, T3E, T9d);
|
Chris@10
|
553 T3I = FMA(T3G, T3H, T3F);
|
Chris@10
|
554 Ta5 = T3K * T3O;
|
Chris@10
|
555 T3M = T3K * T3L;
|
Chris@10
|
556 }
|
Chris@10
|
557 }
|
Chris@10
|
558 {
|
Chris@10
|
559 E T9f, Tfr, T3J, Ta4, Ta6, T3P;
|
Chris@10
|
560 T9f = T9c - T9e;
|
Chris@10
|
561 Tfr = T9c + T9e;
|
Chris@10
|
562 T3J = T3C + T3I;
|
Chris@10
|
563 Ta4 = T3C - T3I;
|
Chris@10
|
564 Ta6 = FNMS(T3N, T3L, Ta5);
|
Chris@10
|
565 T3P = FMA(T3N, T3O, T3M);
|
Chris@10
|
566 {
|
Chris@10
|
567 E Ta9, Tfs, T9g, T3W;
|
Chris@10
|
568 Ta9 = Ta6 - Ta8;
|
Chris@10
|
569 Tfs = Ta6 + Ta8;
|
Chris@10
|
570 T9g = T3P - T3V;
|
Chris@10
|
571 T3W = T3P + T3V;
|
Chris@10
|
572 Tft = Tfr - Tfs;
|
Chris@10
|
573 ThH = Tfr + Tfs;
|
Chris@10
|
574 T9h = T9f + T9g;
|
Chris@10
|
575 Td3 = T9f - T9g;
|
Chris@10
|
576 T3X = T3J + T3W;
|
Chris@10
|
577 TfI = T3J - T3W;
|
Chris@10
|
578 Tde = Ta4 + Ta9;
|
Chris@10
|
579 Taa = Ta4 - Ta9;
|
Chris@10
|
580 }
|
Chris@10
|
581 }
|
Chris@10
|
582 }
|
Chris@10
|
583 }
|
Chris@10
|
584 {
|
Chris@10
|
585 E TaC, T69, Taw, Tga, T5X, Tar, TaA, T63;
|
Chris@10
|
586 {
|
Chris@10
|
587 E T8S, T3r, T8M, Tfk, T3f, T8H, T8Q, T3l;
|
Chris@10
|
588 {
|
Chris@10
|
589 E T8k, T8f, T8w, 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 = ri[WS(rs, 10)];
|
Chris@10
|
597 T2e = ii[WS(rs, 10)];
|
Chris@10
|
598 T2a = W[18];
|
Chris@10
|
599 T2d = W[19];
|
Chris@10
|
600 T2u = ri[WS(rs, 26)];
|
Chris@10
|
601 T2x = ii[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 = ri[WS(rs, 42)];
|
Chris@10
|
614 T2k = ii[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 = ri[WS(rs, 58)];
|
Chris@10
|
620 T2r = ii[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, Tf9, T2m, T88, T8h, T2s, Tfa, T2z;
|
Chris@10
|
633 T8d = T8a - T8c;
|
Chris@10
|
634 Tf9 = 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 Tfa = T8h + T8j;
|
Chris@10
|
641 T2z = T2s + T2y;
|
Chris@10
|
642 T8f = T2s - T2y;
|
Chris@10
|
643 T8w = T8d - T88;
|
Chris@10
|
644 T8e = T88 + T8d;
|
Chris@10
|
645 Thw = Tf9 + Tfa;
|
Chris@10
|
646 Tfb = Tf9 - Tfa;
|
Chris@10
|
647 Tf6 = T2z - T2m;
|
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, T8v, T8l, T36, T8I, T35;
|
Chris@10
|
657 T34 = ri[WS(rs, 6)];
|
Chris@10
|
658 T37 = ii[WS(rs, 6)];
|
Chris@10
|
659 T33 = W[10];
|
Chris@10
|
660 T8v = T8f + T8k;
|
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 = T8w + T8v;
|
Chris@10
|
667 T8m = T8e - T8l;
|
Chris@10
|
668 TcR = T8e + T8l;
|
Chris@10
|
669 T38 = FMA(T36, T37, T35);
|
Chris@10
|
670 T8J = FNMS(T36, T34, T8I);
|
Chris@10
|
671 }
|
Chris@10
|
672 T3n = ri[WS(rs, 22)];
|
Chris@10
|
673 T3q = ii[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 = ri[WS(rs, 38)];
|
Chris@10
|
679 T3d = ii[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 = ri[WS(rs, 54)];
|
Chris@10
|
689 T3k = ii[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 Tfk = 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, Tac, 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 T95, T8N, T8T, Tfl, T8O, T3s, T8U, T94;
|
Chris@10
|
716 T3Z = ri[WS(rs, 9)];
|
Chris@10
|
717 T95 = T8M - T8H;
|
Chris@10
|
718 T8N = T8H + T8M;
|
Chris@10
|
719 T8T = T8Q - T8S;
|
Chris@10
|
720 Tfl = T8Q + T8S;
|
Chris@10
|
721 T8O = T3l - T3r;
|
Chris@10
|
722 T3s = T3l + T3r;
|
Chris@10
|
723 T42 = ii[WS(rs, 9)];
|
Chris@10
|
724 Tfm = Tfk - Tfl;
|
Chris@10
|
725 ThC = Tfk + Tfl;
|
Chris@10
|
726 T8U = T8O - T8T;
|
Chris@10
|
727 T94 = T8O + T8T;
|
Chris@10
|
728 T3t = T3f + T3s;
|
Chris@10
|
729 Tfh = T3s - T3f;
|
Chris@10
|
730 T96 = T94 - T95;
|
Chris@10
|
731 TcV = T95 + T94;
|
Chris@10
|
732 T8V = T8N - T8U;
|
Chris@10
|
733 TcY = T8N + T8U;
|
Chris@10
|
734 T3Y = W[16];
|
Chris@10
|
735 }
|
Chris@10
|
736 T41 = W[17];
|
Chris@10
|
737 T4i = ri[WS(rs, 25)];
|
Chris@10
|
738 T4l = ii[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 = ri[WS(rs, 41)];
|
Chris@10
|
751 T48 = ii[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 = ri[WS(rs, 57)];
|
Chris@10
|
757 T4f = ii[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, TfJ, T4a, T9i, T9r, T4g, TfK, T4n;
|
Chris@10
|
770 T9n = T9k - T9m;
|
Chris@10
|
771 TfJ = 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 TfK = T9r + T9t;
|
Chris@10
|
778 T4n = T4g + T4m;
|
Chris@10
|
779 T9p = T4g - T4m;
|
Chris@10
|
780 Tac = T9n - T9i;
|
Chris@10
|
781 T9o = T9i + T9n;
|
Chris@10
|
782 ThI = TfJ + TfK;
|
Chris@10
|
783 TfL = TfJ - TfK;
|
Chris@10
|
784 Tfu = T4n - T4a;
|
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, Tab, T9v, T5O, Tas, T5N;
|
Chris@10
|
794 T5M = ri[WS(rs, 7)];
|
Chris@10
|
795 T5P = ii[WS(rs, 7)];
|
Chris@10
|
796 T5L = W[12];
|
Chris@10
|
797 Tab = T9p + T9u;
|
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 = Tac + Tab;
|
Chris@10
|
804 T9w = T9o - T9v;
|
Chris@10
|
805 Tdf = T9o + T9v;
|
Chris@10
|
806 T5Q = FMA(T5O, T5P, T5N);
|
Chris@10
|
807 Tat = FNMS(T5O, T5M, Tas);
|
Chris@10
|
808 }
|
Chris@10
|
809 T65 = ri[WS(rs, 23)];
|
Chris@10
|
810 T68 = ii[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 = ri[WS(rs, 39)];
|
Chris@10
|
816 T5V = ii[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 = ri[WS(rs, 55)];
|
Chris@10
|
826 T62 = ii[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 Tga = 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 T9E, Tda, TfE, TfB, Td9, T9L;
|
Chris@10
|
846 {
|
Chris@10
|
847 E T9T, Td7, Tfy, Tfz, Td6, Ta0;
|
Chris@10
|
848 {
|
Chris@10
|
849 E T9V, T4v, T9R, T4O, T4E, T4H, T4G, T9X, T4B, T9O, T4F;
|
Chris@10
|
850 {
|
Chris@10
|
851 E T4K, T4N, T4M, T9Q, T4L;
|
Chris@10
|
852 {
|
Chris@10
|
853 E T4r, T4u, T4q, T4t, T9U, T4s, T4J;
|
Chris@10
|
854 {
|
Chris@10
|
855 E Tbl, Tax, TaD, Tgb, Tay, T6a, TaE, Tbk;
|
Chris@10
|
856 T4r = ri[WS(rs, 5)];
|
Chris@10
|
857 Tbl = Taw - Tar;
|
Chris@10
|
858 Tax = Tar + Taw;
|
Chris@10
|
859 TaD = TaA - TaC;
|
Chris@10
|
860 Tgb = TaA + TaC;
|
Chris@10
|
861 Tay = T63 - T69;
|
Chris@10
|
862 T6a = T63 + T69;
|
Chris@10
|
863 T4u = ii[WS(rs, 5)];
|
Chris@10
|
864 Tgc = Tga - Tgb;
|
Chris@10
|
865 ThT = Tga + Tgb;
|
Chris@10
|
866 TaE = Tay - TaD;
|
Chris@10
|
867 Tbk = Tay + TaD;
|
Chris@10
|
868 T6b = T5X + T6a;
|
Chris@10
|
869 TfV = T6a - T5X;
|
Chris@10
|
870 Tbm = Tbk - Tbl;
|
Chris@10
|
871 Tdn = Tbl + Tbk;
|
Chris@10
|
872 TaF = Tax - TaE;
|
Chris@10
|
873 Tdy = Tax + TaE;
|
Chris@10
|
874 T4q = W[8];
|
Chris@10
|
875 }
|
Chris@10
|
876 T4t = W[9];
|
Chris@10
|
877 T4K = ri[WS(rs, 53)];
|
Chris@10
|
878 T4N = ii[WS(rs, 53)];
|
Chris@10
|
879 T9U = T4q * T4u;
|
Chris@10
|
880 T4s = T4q * T4r;
|
Chris@10
|
881 T4J = W[104];
|
Chris@10
|
882 T4M = W[105];
|
Chris@10
|
883 T9V = FNMS(T4t, T4r, T9U);
|
Chris@10
|
884 T4v = FMA(T4t, T4u, T4s);
|
Chris@10
|
885 T9Q = T4J * T4N;
|
Chris@10
|
886 T4L = T4J * T4K;
|
Chris@10
|
887 }
|
Chris@10
|
888 {
|
Chris@10
|
889 E T4x, T4A, T4w, T4z, T9W, T4y, T4D;
|
Chris@10
|
890 T4x = ri[WS(rs, 37)];
|
Chris@10
|
891 T4A = ii[WS(rs, 37)];
|
Chris@10
|
892 T9R = FNMS(T4M, T4K, T9Q);
|
Chris@10
|
893 T4O = FMA(T4M, T4N, T4L);
|
Chris@10
|
894 T4w = W[72];
|
Chris@10
|
895 T4z = W[73];
|
Chris@10
|
896 T4E = ri[WS(rs, 21)];
|
Chris@10
|
897 T4H = ii[WS(rs, 21)];
|
Chris@10
|
898 T9W = T4w * T4A;
|
Chris@10
|
899 T4y = T4w * T4x;
|
Chris@10
|
900 T4D = W[40];
|
Chris@10
|
901 T4G = W[41];
|
Chris@10
|
902 T9X = FNMS(T4z, T4x, T9W);
|
Chris@10
|
903 T4B = FMA(T4z, T4A, T4y);
|
Chris@10
|
904 T9O = T4D * T4H;
|
Chris@10
|
905 T4F = T4D * T4E;
|
Chris@10
|
906 }
|
Chris@10
|
907 }
|
Chris@10
|
908 {
|
Chris@10
|
909 E T9Y, Tfw, T4C, T9N, T9P, T4I;
|
Chris@10
|
910 T9Y = T9V - T9X;
|
Chris@10
|
911 Tfw = T9V + T9X;
|
Chris@10
|
912 T4C = T4v + T4B;
|
Chris@10
|
913 T9N = T4v - T4B;
|
Chris@10
|
914 T9P = FNMS(T4G, T4E, T9O);
|
Chris@10
|
915 T4I = FMA(T4G, T4H, T4F);
|
Chris@10
|
916 {
|
Chris@10
|
917 E Tfx, T9S, T9Z, T4P;
|
Chris@10
|
918 Tfx = T9P + T9R;
|
Chris@10
|
919 T9S = T9P - T9R;
|
Chris@10
|
920 T9Z = T4I - T4O;
|
Chris@10
|
921 T4P = T4I + T4O;
|
Chris@10
|
922 T9T = T9N - T9S;
|
Chris@10
|
923 Td7 = T9N + T9S;
|
Chris@10
|
924 Tfy = Tfw - Tfx;
|
Chris@10
|
925 ThN = Tfw + Tfx;
|
Chris@10
|
926 Tfz = T4C - T4P;
|
Chris@10
|
927 T4Q = T4C + T4P;
|
Chris@10
|
928 Td6 = T9Y - T9Z;
|
Chris@10
|
929 Ta0 = T9Y + T9Z;
|
Chris@10
|
930 }
|
Chris@10
|
931 }
|
Chris@10
|
932 }
|
Chris@10
|
933 {
|
Chris@10
|
934 E T9G, T4W, T9C, T5f, T55, T58, T57, T9I, T52, T9z, T56;
|
Chris@10
|
935 {
|
Chris@10
|
936 E T5b, T5e, T5d, T9B, T5c;
|
Chris@10
|
937 {
|
Chris@10
|
938 E T4S, T4V, T4R, T4U, T9F, T4T, T5a;
|
Chris@10
|
939 T4S = ri[WS(rs, 61)];
|
Chris@10
|
940 TfN = Tfz + Tfy;
|
Chris@10
|
941 TfA = Tfy - Tfz;
|
Chris@10
|
942 Taf = FMA(KP414213562, T9T, Ta0);
|
Chris@10
|
943 Ta1 = FNMS(KP414213562, Ta0, T9T);
|
Chris@10
|
944 Td8 = FNMS(KP414213562, Td7, Td6);
|
Chris@10
|
945 Tdh = FMA(KP414213562, Td6, Td7);
|
Chris@10
|
946 T4V = ii[WS(rs, 61)];
|
Chris@10
|
947 T4R = W[120];
|
Chris@10
|
948 T4U = W[121];
|
Chris@10
|
949 T5b = ri[WS(rs, 45)];
|
Chris@10
|
950 T5e = ii[WS(rs, 45)];
|
Chris@10
|
951 T9F = T4R * T4V;
|
Chris@10
|
952 T4T = T4R * T4S;
|
Chris@10
|
953 T5a = W[88];
|
Chris@10
|
954 T5d = W[89];
|
Chris@10
|
955 T9G = FNMS(T4U, T4S, T9F);
|
Chris@10
|
956 T4W = FMA(T4U, T4V, T4T);
|
Chris@10
|
957 T9B = T5a * T5e;
|
Chris@10
|
958 T5c = T5a * T5b;
|
Chris@10
|
959 }
|
Chris@10
|
960 {
|
Chris@10
|
961 E T4Y, T51, T4X, T50, T9H, T4Z, T54;
|
Chris@10
|
962 T4Y = ri[WS(rs, 29)];
|
Chris@10
|
963 T51 = ii[WS(rs, 29)];
|
Chris@10
|
964 T9C = FNMS(T5d, T5b, T9B);
|
Chris@10
|
965 T5f = FMA(T5d, T5e, T5c);
|
Chris@10
|
966 T4X = W[56];
|
Chris@10
|
967 T50 = W[57];
|
Chris@10
|
968 T55 = ri[WS(rs, 13)];
|
Chris@10
|
969 T58 = ii[WS(rs, 13)];
|
Chris@10
|
970 T9H = T4X * T51;
|
Chris@10
|
971 T4Z = T4X * T4Y;
|
Chris@10
|
972 T54 = W[24];
|
Chris@10
|
973 T57 = W[25];
|
Chris@10
|
974 T9I = FNMS(T50, T4Y, T9H);
|
Chris@10
|
975 T52 = FMA(T50, T51, T4Z);
|
Chris@10
|
976 T9z = T54 * T58;
|
Chris@10
|
977 T56 = T54 * T55;
|
Chris@10
|
978 }
|
Chris@10
|
979 }
|
Chris@10
|
980 {
|
Chris@10
|
981 E T9J, TfC, T53, T9y, T9A, T59;
|
Chris@10
|
982 T9J = T9G - T9I;
|
Chris@10
|
983 TfC = T9G + T9I;
|
Chris@10
|
984 T53 = T4W + T52;
|
Chris@10
|
985 T9y = T4W - T52;
|
Chris@10
|
986 T9A = FNMS(T57, T55, T9z);
|
Chris@10
|
987 T59 = FMA(T57, T58, T56);
|
Chris@10
|
988 {
|
Chris@10
|
989 E TfD, T9D, T9K, T5g;
|
Chris@10
|
990 TfD = T9A + T9C;
|
Chris@10
|
991 T9D = T9A - T9C;
|
Chris@10
|
992 T9K = T59 - T5f;
|
Chris@10
|
993 T5g = T59 + T5f;
|
Chris@10
|
994 T9E = T9y - T9D;
|
Chris@10
|
995 Tda = T9y + T9D;
|
Chris@10
|
996 TfE = TfC - TfD;
|
Chris@10
|
997 ThO = TfC + TfD;
|
Chris@10
|
998 TfB = T53 - T5g;
|
Chris@10
|
999 T5h = T53 + T5g;
|
Chris@10
|
1000 Td9 = T9J - T9K;
|
Chris@10
|
1001 T9L = T9J + T9K;
|
Chris@10
|
1002 }
|
Chris@10
|
1003 }
|
Chris@10
|
1004 }
|
Chris@10
|
1005 }
|
Chris@10
|
1006 {
|
Chris@10
|
1007 E Tb2, Tdq, TfZ, Tg0, Tdp, Tb9;
|
Chris@10
|
1008 {
|
Chris@10
|
1009 E Tb4, T6i, Tb0, T6B, T6r, T6u, T6t, Tb6, T6o, TaX, T6s;
|
Chris@10
|
1010 {
|
Chris@10
|
1011 E T6x, T6A, T6z, TaZ, T6y;
|
Chris@10
|
1012 {
|
Chris@10
|
1013 E T6e, T6h, T6d, T6g, Tb3, T6f, T6w;
|
Chris@10
|
1014 T6e = ri[WS(rs, 3)];
|
Chris@10
|
1015 TfO = TfB - TfE;
|
Chris@10
|
1016 TfF = TfB + TfE;
|
Chris@10
|
1017 Tag = FNMS(KP414213562, T9E, T9L);
|
Chris@10
|
1018 T9M = FMA(KP414213562, T9L, T9E);
|
Chris@10
|
1019 Tdb = FMA(KP414213562, Tda, Td9);
|
Chris@10
|
1020 Tdi = FNMS(KP414213562, Td9, Tda);
|
Chris@10
|
1021 T6h = ii[WS(rs, 3)];
|
Chris@10
|
1022 T6d = W[4];
|
Chris@10
|
1023 T6g = W[5];
|
Chris@10
|
1024 T6x = ri[WS(rs, 51)];
|
Chris@10
|
1025 T6A = ii[WS(rs, 51)];
|
Chris@10
|
1026 Tb3 = T6d * T6h;
|
Chris@10
|
1027 T6f = T6d * T6e;
|
Chris@10
|
1028 T6w = W[100];
|
Chris@10
|
1029 T6z = W[101];
|
Chris@10
|
1030 Tb4 = FNMS(T6g, T6e, Tb3);
|
Chris@10
|
1031 T6i = FMA(T6g, T6h, T6f);
|
Chris@10
|
1032 TaZ = T6w * T6A;
|
Chris@10
|
1033 T6y = T6w * T6x;
|
Chris@10
|
1034 }
|
Chris@10
|
1035 {
|
Chris@10
|
1036 E T6k, T6n, T6j, T6m, Tb5, T6l, T6q;
|
Chris@10
|
1037 T6k = ri[WS(rs, 35)];
|
Chris@10
|
1038 T6n = ii[WS(rs, 35)];
|
Chris@10
|
1039 Tb0 = FNMS(T6z, T6x, TaZ);
|
Chris@10
|
1040 T6B = FMA(T6z, T6A, T6y);
|
Chris@10
|
1041 T6j = W[68];
|
Chris@10
|
1042 T6m = W[69];
|
Chris@10
|
1043 T6r = ri[WS(rs, 19)];
|
Chris@10
|
1044 T6u = ii[WS(rs, 19)];
|
Chris@10
|
1045 Tb5 = T6j * T6n;
|
Chris@10
|
1046 T6l = T6j * T6k;
|
Chris@10
|
1047 T6q = W[36];
|
Chris@10
|
1048 T6t = W[37];
|
Chris@10
|
1049 Tb6 = FNMS(T6m, T6k, Tb5);
|
Chris@10
|
1050 T6o = FMA(T6m, T6n, T6l);
|
Chris@10
|
1051 TaX = T6q * T6u;
|
Chris@10
|
1052 T6s = T6q * T6r;
|
Chris@10
|
1053 }
|
Chris@10
|
1054 }
|
Chris@10
|
1055 {
|
Chris@10
|
1056 E Tb7, TfX, T6p, TaW, TaY, T6v;
|
Chris@10
|
1057 Tb7 = Tb4 - Tb6;
|
Chris@10
|
1058 TfX = Tb4 + Tb6;
|
Chris@10
|
1059 T6p = T6i + T6o;
|
Chris@10
|
1060 TaW = T6i - T6o;
|
Chris@10
|
1061 TaY = FNMS(T6t, T6r, TaX);
|
Chris@10
|
1062 T6v = FMA(T6t, T6u, T6s);
|
Chris@10
|
1063 {
|
Chris@10
|
1064 E TfY, Tb1, Tb8, T6C;
|
Chris@10
|
1065 TfY = TaY + Tb0;
|
Chris@10
|
1066 Tb1 = TaY - Tb0;
|
Chris@10
|
1067 Tb8 = T6v - T6B;
|
Chris@10
|
1068 T6C = T6v + T6B;
|
Chris@10
|
1069 Tb2 = TaW - Tb1;
|
Chris@10
|
1070 Tdq = TaW + Tb1;
|
Chris@10
|
1071 TfZ = TfX - TfY;
|
Chris@10
|
1072 ThY = TfX + TfY;
|
Chris@10
|
1073 Tg0 = T6p - T6C;
|
Chris@10
|
1074 T6D = T6p + T6C;
|
Chris@10
|
1075 Tdp = Tb7 - Tb8;
|
Chris@10
|
1076 Tb9 = Tb7 + Tb8;
|
Chris@10
|
1077 }
|
Chris@10
|
1078 }
|
Chris@10
|
1079 }
|
Chris@10
|
1080 {
|
Chris@10
|
1081 E TaP, T6J, TaL, T72, T6S, T6V, T6U, TaR, T6P, TaI, T6T;
|
Chris@10
|
1082 {
|
Chris@10
|
1083 E T6Y, T71, T70, TaK, T6Z;
|
Chris@10
|
1084 {
|
Chris@10
|
1085 E T6F, T6I, T6E, T6H, TaO, T6G, T6X;
|
Chris@10
|
1086 T6F = ri[WS(rs, 59)];
|
Chris@10
|
1087 Tge = Tg0 + TfZ;
|
Chris@10
|
1088 Tg1 = TfZ - Tg0;
|
Chris@10
|
1089 Tbo = FMA(KP414213562, Tb2, Tb9);
|
Chris@10
|
1090 Tba = FNMS(KP414213562, Tb9, Tb2);
|
Chris@10
|
1091 Tdr = FNMS(KP414213562, Tdq, Tdp);
|
Chris@10
|
1092 TdA = FMA(KP414213562, Tdp, Tdq);
|
Chris@10
|
1093 T6I = ii[WS(rs, 59)];
|
Chris@10
|
1094 T6E = W[116];
|
Chris@10
|
1095 T6H = W[117];
|
Chris@10
|
1096 T6Y = ri[WS(rs, 43)];
|
Chris@10
|
1097 T71 = ii[WS(rs, 43)];
|
Chris@10
|
1098 TaO = T6E * T6I;
|
Chris@10
|
1099 T6G = T6E * T6F;
|
Chris@10
|
1100 T6X = W[84];
|
Chris@10
|
1101 T70 = W[85];
|
Chris@10
|
1102 TaP = FNMS(T6H, T6F, TaO);
|
Chris@10
|
1103 T6J = FMA(T6H, T6I, T6G);
|
Chris@10
|
1104 TaK = T6X * T71;
|
Chris@10
|
1105 T6Z = T6X * T6Y;
|
Chris@10
|
1106 }
|
Chris@10
|
1107 {
|
Chris@10
|
1108 E T6L, T6O, T6K, T6N, TaQ, T6M, T6R;
|
Chris@10
|
1109 T6L = ri[WS(rs, 27)];
|
Chris@10
|
1110 T6O = ii[WS(rs, 27)];
|
Chris@10
|
1111 TaL = FNMS(T70, T6Y, TaK);
|
Chris@10
|
1112 T72 = FMA(T70, T71, T6Z);
|
Chris@10
|
1113 T6K = W[52];
|
Chris@10
|
1114 T6N = W[53];
|
Chris@10
|
1115 T6S = ri[WS(rs, 11)];
|
Chris@10
|
1116 T6V = ii[WS(rs, 11)];
|
Chris@10
|
1117 TaQ = T6K * T6O;
|
Chris@10
|
1118 T6M = T6K * T6L;
|
Chris@10
|
1119 T6R = W[20];
|
Chris@10
|
1120 T6U = W[21];
|
Chris@10
|
1121 TaR = FNMS(T6N, T6L, TaQ);
|
Chris@10
|
1122 T6P = FMA(T6N, T6O, T6M);
|
Chris@10
|
1123 TaI = T6R * T6V;
|
Chris@10
|
1124 T6T = T6R * T6S;
|
Chris@10
|
1125 }
|
Chris@10
|
1126 }
|
Chris@10
|
1127 {
|
Chris@10
|
1128 E TaS, Tg3, T6Q, TaH, TaJ, T6W;
|
Chris@10
|
1129 TaS = TaP - TaR;
|
Chris@10
|
1130 Tg3 = TaP + TaR;
|
Chris@10
|
1131 T6Q = T6J + T6P;
|
Chris@10
|
1132 TaH = T6J - T6P;
|
Chris@10
|
1133 TaJ = FNMS(T6U, T6S, TaI);
|
Chris@10
|
1134 T6W = FMA(T6U, T6V, T6T);
|
Chris@10
|
1135 {
|
Chris@10
|
1136 E Tg4, TaM, TaT, T73;
|
Chris@10
|
1137 Tg4 = TaJ + TaL;
|
Chris@10
|
1138 TaM = TaJ - TaL;
|
Chris@10
|
1139 TaT = T6W - T72;
|
Chris@10
|
1140 T73 = T6W + T72;
|
Chris@10
|
1141 TaN = TaH - TaM;
|
Chris@10
|
1142 Tdt = TaH + TaM;
|
Chris@10
|
1143 Tg5 = Tg3 - Tg4;
|
Chris@10
|
1144 ThZ = Tg3 + Tg4;
|
Chris@10
|
1145 Tg2 = T6Q - T73;
|
Chris@10
|
1146 T74 = T6Q + T73;
|
Chris@10
|
1147 Tds = TaS - TaT;
|
Chris@10
|
1148 TaU = TaS + TaT;
|
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 Tgf, Tg6, Tbp, TaV, Tdu, TdB, Tje, Tjd, TjO, TjN;
|
Chris@10
|
1157 {
|
Chris@10
|
1158 E Thq, Tj7, Thy, ThA, Tht, Tj8, Thx, ThD, ThX, ThV, ThU, Ti0, ThM, ThK, ThJ;
|
Chris@10
|
1159 E ThP, TiI, TiZ, TiL, Tj0;
|
Chris@10
|
1160 {
|
Chris@10
|
1161 E Tio, T1I, Tj1, T3v, Tj2, TiX, TiN, Tir, T76, TiK, TiC, TiG, T5j, Tit, Tiw;
|
Chris@10
|
1162 E TiJ;
|
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 Tgf = Tg2 - Tg5;
|
Chris@10
|
1170 Tg6 = Tg2 + Tg5;
|
Chris@10
|
1171 Tbp = FNMS(KP414213562, TaN, TaU);
|
Chris@10
|
1172 TaV = FMA(KP414213562, TaU, TaN);
|
Chris@10
|
1173 Tdu = FMA(KP414213562, Tdt, Tds);
|
Chris@10
|
1174 TdB = FNMS(KP414213562, Tds, Tdt);
|
Chris@10
|
1175 T1H = T1f + T1G;
|
Chris@10
|
1176 Tj7 = T1G - T1f;
|
Chris@10
|
1177 Thy = 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 = Thr + Ths;
|
Chris@10
|
1183 Tio = TO - T1H;
|
Chris@10
|
1184 T1I = TO + T1H;
|
Chris@10
|
1185 Tj1 = T3u - T2B;
|
Chris@10
|
1186 T3v = T2B + T3u;
|
Chris@10
|
1187 TiW = TiP + TiV;
|
Chris@10
|
1188 Tj8 = TiV - TiP;
|
Chris@10
|
1189 }
|
Chris@10
|
1190 Thx = Thv - Thw;
|
Chris@10
|
1191 Tip = Thv + Thw;
|
Chris@10
|
1192 Tiq = ThB + ThC;
|
Chris@10
|
1193 ThD = ThB - ThC;
|
Chris@10
|
1194 {
|
Chris@10
|
1195 E T6c, T75, Tiz, TiA;
|
Chris@10
|
1196 ThX = T5K - T6b;
|
Chris@10
|
1197 T6c = T5K + T6b;
|
Chris@10
|
1198 Tj2 = TiW - TiO;
|
Chris@10
|
1199 TiX = TiO + TiW;
|
Chris@10
|
1200 TiN = Tip + Tiq;
|
Chris@10
|
1201 Tir = Tip - Tiq;
|
Chris@10
|
1202 T75 = T6D + T74;
|
Chris@10
|
1203 ThV = T74 - T6D;
|
Chris@10
|
1204 ThU = ThS - ThT;
|
Chris@10
|
1205 Tiz = ThS + ThT;
|
Chris@10
|
1206 TiA = ThY + ThZ;
|
Chris@10
|
1207 Ti0 = ThY - ThZ;
|
Chris@10
|
1208 {
|
Chris@10
|
1209 E T4p, Tiy, TiB, T5i, Tiu, Tiv;
|
Chris@10
|
1210 ThM = T3X - T4o;
|
Chris@10
|
1211 T4p = T3X + T4o;
|
Chris@10
|
1212 T76 = T6c + T75;
|
Chris@10
|
1213 Tiy = T6c - T75;
|
Chris@10
|
1214 TiK = Tiz + TiA;
|
Chris@10
|
1215 TiB = Tiz - TiA;
|
Chris@10
|
1216 T5i = T4Q + T5h;
|
Chris@10
|
1217 ThK = T5h - T4Q;
|
Chris@10
|
1218 ThJ = ThH - ThI;
|
Chris@10
|
1219 Tiu = ThH + ThI;
|
Chris@10
|
1220 Tiv = ThN + ThO;
|
Chris@10
|
1221 ThP = ThN - ThO;
|
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 TiJ = Tiu + Tiv;
|
Chris@10
|
1228 }
|
Chris@10
|
1229 }
|
Chris@10
|
1230 }
|
Chris@10
|
1231 {
|
Chris@10
|
1232 E TiE, Tis, TiD, Tj6, Tj5, Tj3, Tj4, TiH;
|
Chris@10
|
1233 {
|
Chris@10
|
1234 E T3w, TiF, Tix, T77, TiM, TiY;
|
Chris@10
|
1235 TiI = T1I - T3v;
|
Chris@10
|
1236 T3w = T1I + T3v;
|
Chris@10
|
1237 TiF = Tiw - Tit;
|
Chris@10
|
1238 Tix = Tit + Tiw;
|
Chris@10
|
1239 T77 = T5j + T76;
|
Chris@10
|
1240 TiZ = T76 - T5j;
|
Chris@10
|
1241 TiL = TiJ - TiK;
|
Chris@10
|
1242 TiM = TiJ + TiK;
|
Chris@10
|
1243 TiY = TiN + TiX;
|
Chris@10
|
1244 Tj0 = TiX - TiN;
|
Chris@10
|
1245 TiE = Tio - Tir;
|
Chris@10
|
1246 Tis = Tio + Tir;
|
Chris@10
|
1247 ri[0] = T3w + T77;
|
Chris@10
|
1248 ri[WS(rs, 32)] = T3w - T77;
|
Chris@10
|
1249 ii[WS(rs, 32)] = TiY - TiM;
|
Chris@10
|
1250 ii[0] = TiM + TiY;
|
Chris@10
|
1251 TiD = Tix + TiC;
|
Chris@10
|
1252 Tj6 = TiC - Tix;
|
Chris@10
|
1253 Tj5 = Tj2 - Tj1;
|
Chris@10
|
1254 Tj3 = Tj1 + Tj2;
|
Chris@10
|
1255 Tj4 = TiF + TiG;
|
Chris@10
|
1256 TiH = TiF - TiG;
|
Chris@10
|
1257 }
|
Chris@10
|
1258 ri[WS(rs, 8)] = FMA(KP707106781, TiD, Tis);
|
Chris@10
|
1259 ri[WS(rs, 40)] = FNMS(KP707106781, TiD, Tis);
|
Chris@10
|
1260 ii[WS(rs, 40)] = FNMS(KP707106781, Tj4, Tj3);
|
Chris@10
|
1261 ii[WS(rs, 8)] = FMA(KP707106781, Tj4, Tj3);
|
Chris@10
|
1262 ri[WS(rs, 24)] = FMA(KP707106781, TiH, TiE);
|
Chris@10
|
1263 ri[WS(rs, 56)] = FNMS(KP707106781, TiH, TiE);
|
Chris@10
|
1264 ii[WS(rs, 56)] = FNMS(KP707106781, Tj6, Tj5);
|
Chris@10
|
1265 ii[WS(rs, 24)] = FMA(KP707106781, Tj6, Tj5);
|
Chris@10
|
1266 }
|
Chris@10
|
1267 }
|
Chris@10
|
1268 {
|
Chris@10
|
1269 E Ti8, Thu, Tjf, Tj9, Tib, Tjg, Tja, ThF, Tih, ThW, Tif, Til, Ti5, ThR;
|
Chris@10
|
1270 ri[WS(rs, 16)] = TiI + TiL;
|
Chris@10
|
1271 ri[WS(rs, 48)] = TiI - TiL;
|
Chris@10
|
1272 ii[WS(rs, 48)] = Tj0 - TiZ;
|
Chris@10
|
1273 ii[WS(rs, 16)] = TiZ + Tj0;
|
Chris@10
|
1274 Ti8 = Thq + Tht;
|
Chris@10
|
1275 Thu = Thq - Tht;
|
Chris@10
|
1276 Tjf = Tj8 - Tj7;
|
Chris@10
|
1277 Tj9 = Tj7 + Tj8;
|
Chris@10
|
1278 {
|
Chris@10
|
1279 E Tie, ThL, Tid, ThQ;
|
Chris@10
|
1280 {
|
Chris@10
|
1281 E Ti9, Thz, Tia, ThE;
|
Chris@10
|
1282 Ti9 = Thy + Thx;
|
Chris@10
|
1283 Thz = Thx - Thy;
|
Chris@10
|
1284 Tia = ThA - ThD;
|
Chris@10
|
1285 ThE = ThA + ThD;
|
Chris@10
|
1286 Tib = Ti9 + Tia;
|
Chris@10
|
1287 Tjg = Tia - Ti9;
|
Chris@10
|
1288 Tja = Thz + ThE;
|
Chris@10
|
1289 ThF = Thz - ThE;
|
Chris@10
|
1290 Tie = ThJ + ThK;
|
Chris@10
|
1291 ThL = ThJ - ThK;
|
Chris@10
|
1292 }
|
Chris@10
|
1293 Tid = ThM + ThP;
|
Chris@10
|
1294 ThQ = ThM - ThP;
|
Chris@10
|
1295 Tih = ThU + ThV;
|
Chris@10
|
1296 ThW = ThU - ThV;
|
Chris@10
|
1297 Tif = FMA(KP414213562, Tie, Tid);
|
Chris@10
|
1298 Til = FNMS(KP414213562, Tid, Tie);
|
Chris@10
|
1299 Ti5 = FNMS(KP414213562, ThL, ThQ);
|
Chris@10
|
1300 ThR = FMA(KP414213562, ThQ, ThL);
|
Chris@10
|
1301 }
|
Chris@10
|
1302 {
|
Chris@10
|
1303 E Ti4, ThG, Tjh, Tjj, Tig, Ti1;
|
Chris@10
|
1304 Ti4 = FNMS(KP707106781, ThF, Thu);
|
Chris@10
|
1305 ThG = FMA(KP707106781, ThF, Thu);
|
Chris@10
|
1306 Tjh = FMA(KP707106781, Tjg, Tjf);
|
Chris@10
|
1307 Tjj = FNMS(KP707106781, Tjg, Tjf);
|
Chris@10
|
1308 Tig = ThX + Ti0;
|
Chris@10
|
1309 Ti1 = ThX - Ti0;
|
Chris@10
|
1310 {
|
Chris@10
|
1311 E Tik, Tjb, Tjc, Tin;
|
Chris@10
|
1312 {
|
Chris@10
|
1313 E Tic, Tim, Ti6, Ti2, Tij, Tii;
|
Chris@10
|
1314 Tik = FNMS(KP707106781, Tib, Ti8);
|
Chris@10
|
1315 Tic = FMA(KP707106781, Tib, Ti8);
|
Chris@10
|
1316 Tii = FNMS(KP414213562, Tih, Tig);
|
Chris@10
|
1317 Tim = FMA(KP414213562, Tig, Tih);
|
Chris@10
|
1318 Ti6 = FMA(KP414213562, ThW, Ti1);
|
Chris@10
|
1319 Ti2 = FNMS(KP414213562, Ti1, ThW);
|
Chris@10
|
1320 Tij = Tif + Tii;
|
Chris@10
|
1321 Tje = Tii - Tif;
|
Chris@10
|
1322 Tjd = FNMS(KP707106781, Tja, Tj9);
|
Chris@10
|
1323 Tjb = FMA(KP707106781, Tja, Tj9);
|
Chris@10
|
1324 {
|
Chris@10
|
1325 E Ti7, Tji, Tjk, Ti3;
|
Chris@10
|
1326 Ti7 = Ti5 + Ti6;
|
Chris@10
|
1327 Tji = Ti6 - Ti5;
|
Chris@10
|
1328 Tjk = ThR + Ti2;
|
Chris@10
|
1329 Ti3 = ThR - Ti2;
|
Chris@10
|
1330 ri[WS(rs, 4)] = FMA(KP923879532, Tij, Tic);
|
Chris@10
|
1331 ri[WS(rs, 36)] = FNMS(KP923879532, Tij, Tic);
|
Chris@10
|
1332 ri[WS(rs, 60)] = FMA(KP923879532, Ti7, Ti4);
|
Chris@10
|
1333 ri[WS(rs, 28)] = FNMS(KP923879532, Ti7, Ti4);
|
Chris@10
|
1334 ii[WS(rs, 44)] = FNMS(KP923879532, Tji, Tjh);
|
Chris@10
|
1335 ii[WS(rs, 12)] = FMA(KP923879532, Tji, Tjh);
|
Chris@10
|
1336 ii[WS(rs, 60)] = FMA(KP923879532, Tjk, Tjj);
|
Chris@10
|
1337 ii[WS(rs, 28)] = FNMS(KP923879532, Tjk, Tjj);
|
Chris@10
|
1338 ri[WS(rs, 12)] = FMA(KP923879532, Ti3, ThG);
|
Chris@10
|
1339 ri[WS(rs, 44)] = FNMS(KP923879532, Ti3, ThG);
|
Chris@10
|
1340 Tjc = Til + Tim;
|
Chris@10
|
1341 Tin = Til - Tim;
|
Chris@10
|
1342 }
|
Chris@10
|
1343 }
|
Chris@10
|
1344 ii[WS(rs, 36)] = FNMS(KP923879532, Tjc, Tjb);
|
Chris@10
|
1345 ii[WS(rs, 4)] = FMA(KP923879532, Tjc, Tjb);
|
Chris@10
|
1346 ri[WS(rs, 20)] = FMA(KP923879532, Tin, Tik);
|
Chris@10
|
1347 ri[WS(rs, 52)] = FNMS(KP923879532, Tin, Tik);
|
Chris@10
|
1348 }
|
Chris@10
|
1349 }
|
Chris@10
|
1350 }
|
Chris@10
|
1351 }
|
Chris@10
|
1352 {
|
Chris@10
|
1353 E TjD, TjJ, Tgo, Tf2, Tjp, Tjv, Tha, TgI, Tgd, Tgr, Tjw, Tjq, Tfp, Tgg, Thk;
|
Chris@10
|
1354 E Tho, Th8, Th4, Tgv, TgB, Tgl, TfR, TjE, Thd, TjK, TgP, Tgx, Tg8, Thh, Thn;
|
Chris@10
|
1355 E Th7, TgX;
|
Chris@10
|
1356 {
|
Chris@10
|
1357 E TgJ, TgK, TgM, TgN, Tg7, TfW, Th1, Thj, Th0, Th2;
|
Chris@10
|
1358 {
|
Chris@10
|
1359 E TgE, TeQ, TjB, Tjn, TgF, TgG, TjC, Tf1, TeV, Tf0;
|
Chris@10
|
1360 TgE = TeM - TeP;
|
Chris@10
|
1361 TeQ = TeM + TeP;
|
Chris@10
|
1362 TjB = Tjm - Tjl;
|
Chris@10
|
1363 Tjn = Tjl + Tjm;
|
Chris@10
|
1364 TgF = TeU - TeR;
|
Chris@10
|
1365 TeV = TeR + TeU;
|
Chris@10
|
1366 ii[WS(rs, 52)] = FNMS(KP923879532, Tje, Tjd);
|
Chris@10
|
1367 ii[WS(rs, 20)] = FMA(KP923879532, Tje, Tjd);
|
Chris@10
|
1368 Tf0 = TeW - TeZ;
|
Chris@10
|
1369 TgG = TeW + TeZ;
|
Chris@10
|
1370 TjC = Tf0 - TeV;
|
Chris@10
|
1371 Tf1 = TeV + Tf0;
|
Chris@10
|
1372 {
|
Chris@10
|
1373 E Tfi, Tgp, Tfd, Tfn;
|
Chris@10
|
1374 {
|
Chris@10
|
1375 E Tf7, Tjo, TgH, Tfc;
|
Chris@10
|
1376 TgJ = Tf5 - Tf6;
|
Chris@10
|
1377 Tf7 = Tf5 + Tf6;
|
Chris@10
|
1378 TjD = FMA(KP707106781, TjC, TjB);
|
Chris@10
|
1379 TjJ = FNMS(KP707106781, TjC, TjB);
|
Chris@10
|
1380 Tgo = FMA(KP707106781, Tf1, TeQ);
|
Chris@10
|
1381 Tf2 = FNMS(KP707106781, Tf1, TeQ);
|
Chris@10
|
1382 Tjo = TgF + TgG;
|
Chris@10
|
1383 TgH = TgF - TgG;
|
Chris@10
|
1384 Tfc = Tf8 + Tfb;
|
Chris@10
|
1385 TgK = Tf8 - Tfb;
|
Chris@10
|
1386 TgM = Tfg - Tfh;
|
Chris@10
|
1387 Tfi = Tfg + Tfh;
|
Chris@10
|
1388 Tjp = FMA(KP707106781, Tjo, Tjn);
|
Chris@10
|
1389 Tjv = FNMS(KP707106781, Tjo, Tjn);
|
Chris@10
|
1390 Tha = FNMS(KP707106781, TgH, TgE);
|
Chris@10
|
1391 TgI = FMA(KP707106781, TgH, TgE);
|
Chris@10
|
1392 Tgp = FMA(KP414213562, Tf7, Tfc);
|
Chris@10
|
1393 Tfd = FNMS(KP414213562, Tfc, Tf7);
|
Chris@10
|
1394 Tfn = Tfj + Tfm;
|
Chris@10
|
1395 TgN = Tfj - Tfm;
|
Chris@10
|
1396 }
|
Chris@10
|
1397 {
|
Chris@10
|
1398 E TgY, TgZ, Tgq, Tfo;
|
Chris@10
|
1399 Tgd = Tg9 + Tgc;
|
Chris@10
|
1400 TgY = Tg9 - Tgc;
|
Chris@10
|
1401 TgZ = Tg6 - Tg1;
|
Chris@10
|
1402 Tg7 = Tg1 + Tg6;
|
Chris@10
|
1403 TfW = TfU + TfV;
|
Chris@10
|
1404 Th1 = TfU - TfV;
|
Chris@10
|
1405 Tgq = FNMS(KP414213562, Tfi, Tfn);
|
Chris@10
|
1406 Tfo = FMA(KP414213562, Tfn, Tfi);
|
Chris@10
|
1407 Thj = FMA(KP707106781, TgZ, TgY);
|
Chris@10
|
1408 Th0 = FNMS(KP707106781, TgZ, TgY);
|
Chris@10
|
1409 Tgr = Tgp + Tgq;
|
Chris@10
|
1410 Tjw = Tgq - Tgp;
|
Chris@10
|
1411 Tjq = Tfd + Tfo;
|
Chris@10
|
1412 Tfp = Tfd - Tfo;
|
Chris@10
|
1413 Th2 = Tge - Tgf;
|
Chris@10
|
1414 Tgg = Tge + Tgf;
|
Chris@10
|
1415 }
|
Chris@10
|
1416 }
|
Chris@10
|
1417 }
|
Chris@10
|
1418 {
|
Chris@10
|
1419 E TgU, TgS, TgR, TgV, Thb, TgL;
|
Chris@10
|
1420 {
|
Chris@10
|
1421 E TfM, Tgu, TfH, TfP, Tgt, TfQ;
|
Chris@10
|
1422 {
|
Chris@10
|
1423 E Tfv, TfG, Thi, Th3;
|
Chris@10
|
1424 TgU = Tft - Tfu;
|
Chris@10
|
1425 Tfv = Tft + Tfu;
|
Chris@10
|
1426 TfG = TfA + TfF;
|
Chris@10
|
1427 TgS = TfF - TfA;
|
Chris@10
|
1428 TgR = TfI - TfL;
|
Chris@10
|
1429 TfM = TfI + TfL;
|
Chris@10
|
1430 Thi = FMA(KP707106781, Th2, Th1);
|
Chris@10
|
1431 Th3 = FNMS(KP707106781, Th2, Th1);
|
Chris@10
|
1432 Tgu = FMA(KP707106781, TfG, Tfv);
|
Chris@10
|
1433 TfH = FNMS(KP707106781, TfG, Tfv);
|
Chris@10
|
1434 Thk = FNMS(KP198912367, Thj, Thi);
|
Chris@10
|
1435 Tho = FMA(KP198912367, Thi, Thj);
|
Chris@10
|
1436 Th8 = FMA(KP668178637, Th0, Th3);
|
Chris@10
|
1437 Th4 = FNMS(KP668178637, Th3, Th0);
|
Chris@10
|
1438 TfP = TfN + TfO;
|
Chris@10
|
1439 TgV = TfN - TfO;
|
Chris@10
|
1440 }
|
Chris@10
|
1441 Tgt = FMA(KP707106781, TfP, TfM);
|
Chris@10
|
1442 TfQ = FNMS(KP707106781, TfP, TfM);
|
Chris@10
|
1443 Thb = FNMS(KP414213562, TgJ, TgK);
|
Chris@10
|
1444 TgL = FMA(KP414213562, TgK, TgJ);
|
Chris@10
|
1445 Tgv = FMA(KP198912367, Tgu, Tgt);
|
Chris@10
|
1446 TgB = FNMS(KP198912367, Tgt, Tgu);
|
Chris@10
|
1447 Tgl = FNMS(KP668178637, TfH, TfQ);
|
Chris@10
|
1448 TfR = FMA(KP668178637, TfQ, TfH);
|
Chris@10
|
1449 }
|
Chris@10
|
1450 {
|
Chris@10
|
1451 E Thg, TgT, Thc, TgO, Thf, TgW;
|
Chris@10
|
1452 Thc = FMA(KP414213562, TgM, TgN);
|
Chris@10
|
1453 TgO = FNMS(KP414213562, TgN, TgM);
|
Chris@10
|
1454 Thg = FMA(KP707106781, TgS, TgR);
|
Chris@10
|
1455 TgT = FNMS(KP707106781, TgS, TgR);
|
Chris@10
|
1456 TjE = Thc - Thb;
|
Chris@10
|
1457 Thd = Thb + Thc;
|
Chris@10
|
1458 TjK = TgL + TgO;
|
Chris@10
|
1459 TgP = TgL - TgO;
|
Chris@10
|
1460 Thf = FMA(KP707106781, TgV, TgU);
|
Chris@10
|
1461 TgW = FNMS(KP707106781, TgV, TgU);
|
Chris@10
|
1462 Tgx = FMA(KP707106781, Tg7, TfW);
|
Chris@10
|
1463 Tg8 = FNMS(KP707106781, Tg7, TfW);
|
Chris@10
|
1464 Thh = FMA(KP198912367, Thg, Thf);
|
Chris@10
|
1465 Thn = FNMS(KP198912367, Thf, Thg);
|
Chris@10
|
1466 Th7 = FNMS(KP668178637, TgT, TgW);
|
Chris@10
|
1467 TgX = FMA(KP668178637, TgW, TgT);
|
Chris@10
|
1468 }
|
Chris@10
|
1469 }
|
Chris@10
|
1470 }
|
Chris@10
|
1471 {
|
Chris@10
|
1472 E Tju, Tjt, TjI, TjH;
|
Chris@10
|
1473 {
|
Chris@10
|
1474 E Tgk, Tfq, Tjx, Tjz, Tgw, Tgh;
|
Chris@10
|
1475 Tgk = FNMS(KP923879532, Tfp, Tf2);
|
Chris@10
|
1476 Tfq = FMA(KP923879532, Tfp, Tf2);
|
Chris@10
|
1477 Tjx = FMA(KP923879532, Tjw, Tjv);
|
Chris@10
|
1478 Tjz = FNMS(KP923879532, Tjw, Tjv);
|
Chris@10
|
1479 Tgw = FMA(KP707106781, Tgg, Tgd);
|
Chris@10
|
1480 Tgh = FNMS(KP707106781, Tgg, Tgd);
|
Chris@10
|
1481 {
|
Chris@10
|
1482 E TgA, Tjr, Tjs, TgD;
|
Chris@10
|
1483 {
|
Chris@10
|
1484 E Tgs, TgC, Tgm, Tgi, Tgz, Tgy;
|
Chris@10
|
1485 TgA = FNMS(KP923879532, Tgr, Tgo);
|
Chris@10
|
1486 Tgs = FMA(KP923879532, Tgr, Tgo);
|
Chris@10
|
1487 Tgy = FNMS(KP198912367, Tgx, Tgw);
|
Chris@10
|
1488 TgC = FMA(KP198912367, Tgw, Tgx);
|
Chris@10
|
1489 Tgm = FMA(KP668178637, Tg8, Tgh);
|
Chris@10
|
1490 Tgi = FNMS(KP668178637, Tgh, Tg8);
|
Chris@10
|
1491 Tgz = Tgv + Tgy;
|
Chris@10
|
1492 Tju = Tgy - Tgv;
|
Chris@10
|
1493 Tjt = FNMS(KP923879532, Tjq, Tjp);
|
Chris@10
|
1494 Tjr = FMA(KP923879532, Tjq, Tjp);
|
Chris@10
|
1495 {
|
Chris@10
|
1496 E Tgn, Tjy, TjA, Tgj;
|
Chris@10
|
1497 Tgn = Tgl + Tgm;
|
Chris@10
|
1498 Tjy = Tgm - Tgl;
|
Chris@10
|
1499 TjA = TfR + Tgi;
|
Chris@10
|
1500 Tgj = TfR - Tgi;
|
Chris@10
|
1501 ri[WS(rs, 2)] = FMA(KP980785280, Tgz, Tgs);
|
Chris@10
|
1502 ri[WS(rs, 34)] = FNMS(KP980785280, Tgz, Tgs);
|
Chris@10
|
1503 ri[WS(rs, 58)] = FMA(KP831469612, Tgn, Tgk);
|
Chris@10
|
1504 ri[WS(rs, 26)] = FNMS(KP831469612, Tgn, Tgk);
|
Chris@10
|
1505 ii[WS(rs, 42)] = FNMS(KP831469612, Tjy, Tjx);
|
Chris@10
|
1506 ii[WS(rs, 10)] = FMA(KP831469612, Tjy, Tjx);
|
Chris@10
|
1507 ii[WS(rs, 58)] = FMA(KP831469612, TjA, Tjz);
|
Chris@10
|
1508 ii[WS(rs, 26)] = FNMS(KP831469612, TjA, Tjz);
|
Chris@10
|
1509 ri[WS(rs, 10)] = FMA(KP831469612, Tgj, Tfq);
|
Chris@10
|
1510 ri[WS(rs, 42)] = FNMS(KP831469612, Tgj, Tfq);
|
Chris@10
|
1511 Tjs = TgB + TgC;
|
Chris@10
|
1512 TgD = TgB - TgC;
|
Chris@10
|
1513 }
|
Chris@10
|
1514 }
|
Chris@10
|
1515 ii[WS(rs, 34)] = FNMS(KP980785280, Tjs, Tjr);
|
Chris@10
|
1516 ii[WS(rs, 2)] = FMA(KP980785280, Tjs, Tjr);
|
Chris@10
|
1517 ri[WS(rs, 18)] = FMA(KP980785280, TgD, TgA);
|
Chris@10
|
1518 ri[WS(rs, 50)] = FNMS(KP980785280, TgD, TgA);
|
Chris@10
|
1519 }
|
Chris@10
|
1520 }
|
Chris@10
|
1521 {
|
Chris@10
|
1522 E Th6, TjF, TjG, Th9, TgQ, Th5;
|
Chris@10
|
1523 Th6 = FNMS(KP923879532, TgP, TgI);
|
Chris@10
|
1524 TgQ = FMA(KP923879532, TgP, TgI);
|
Chris@10
|
1525 Th5 = TgX + Th4;
|
Chris@10
|
1526 TjI = Th4 - TgX;
|
Chris@10
|
1527 TjH = FNMS(KP923879532, TjE, TjD);
|
Chris@10
|
1528 TjF = FMA(KP923879532, TjE, TjD);
|
Chris@10
|
1529 ii[WS(rs, 50)] = FNMS(KP980785280, Tju, Tjt);
|
Chris@10
|
1530 ii[WS(rs, 18)] = FMA(KP980785280, Tju, Tjt);
|
Chris@10
|
1531 ri[WS(rs, 6)] = FMA(KP831469612, Th5, TgQ);
|
Chris@10
|
1532 ri[WS(rs, 38)] = FNMS(KP831469612, Th5, TgQ);
|
Chris@10
|
1533 TjG = Th7 + Th8;
|
Chris@10
|
1534 Th9 = Th7 - Th8;
|
Chris@10
|
1535 ii[WS(rs, 38)] = FNMS(KP831469612, TjG, TjF);
|
Chris@10
|
1536 ii[WS(rs, 6)] = FMA(KP831469612, TjG, TjF);
|
Chris@10
|
1537 ri[WS(rs, 22)] = FMA(KP831469612, Th9, Th6);
|
Chris@10
|
1538 ri[WS(rs, 54)] = FNMS(KP831469612, Th9, Th6);
|
Chris@10
|
1539 }
|
Chris@10
|
1540 {
|
Chris@10
|
1541 E Thm, TjL, TjM, Thp, The, Thl;
|
Chris@10
|
1542 Thm = FMA(KP923879532, Thd, Tha);
|
Chris@10
|
1543 The = FNMS(KP923879532, Thd, Tha);
|
Chris@10
|
1544 Thl = Thh - Thk;
|
Chris@10
|
1545 TjO = Thh + Thk;
|
Chris@10
|
1546 TjN = FMA(KP923879532, TjK, TjJ);
|
Chris@10
|
1547 TjL = FNMS(KP923879532, TjK, TjJ);
|
Chris@10
|
1548 ii[WS(rs, 54)] = FNMS(KP831469612, TjI, TjH);
|
Chris@10
|
1549 ii[WS(rs, 22)] = FMA(KP831469612, TjI, TjH);
|
Chris@10
|
1550 ri[WS(rs, 14)] = FMA(KP980785280, Thl, The);
|
Chris@10
|
1551 ri[WS(rs, 46)] = FNMS(KP980785280, Thl, The);
|
Chris@10
|
1552 TjM = Tho - Thn;
|
Chris@10
|
1553 Thp = Thn + Tho;
|
Chris@10
|
1554 ii[WS(rs, 46)] = FNMS(KP980785280, TjM, TjL);
|
Chris@10
|
1555 ii[WS(rs, 14)] = FMA(KP980785280, TjM, TjL);
|
Chris@10
|
1556 ri[WS(rs, 62)] = FMA(KP980785280, Thp, Thm);
|
Chris@10
|
1557 ri[WS(rs, 30)] = FNMS(KP980785280, Thp, Thm);
|
Chris@10
|
1558 }
|
Chris@10
|
1559 }
|
Chris@10
|
1560 }
|
Chris@10
|
1561 {
|
Chris@10
|
1562 E TjS, TcD, Tcw, TkO, TkN, Tcz;
|
Chris@10
|
1563 {
|
Chris@10
|
1564 E TbB, Tkw, Tkq, T99, TbF, TbL, Tbv, Taj, Tcu, Tcy, Tci, Tce, Tcr, Tcx, Tch;
|
Chris@10
|
1565 E Tc7, TkE, Tcn, TkK, TbZ, TbP, T7J, TbO, T7u, TkB, Tkn, TbI, TbM, Tbw, Tbs;
|
Chris@10
|
1566 E T7Y, TbQ;
|
Chris@10
|
1567 {
|
Chris@10
|
1568 E TbT, TbU, TbW, TbX, Tc4, Tc2, Tc1, Tc5, Tbn, Tbb, TaG, Tcb, Tct, Tca, Tcc;
|
Chris@10
|
1569 E Tbq, Tcl, TbV;
|
Chris@10
|
1570 {
|
Chris@10
|
1571 E T8W, Tbz, T8z, T97, T8n, T8y;
|
Chris@10
|
1572 TbT = FMA(KP707106781, T8m, T87);
|
Chris@10
|
1573 T8n = FNMS(KP707106781, T8m, T87);
|
Chris@10
|
1574 T8y = FNMS(KP707106781, T8x, T8u);
|
Chris@10
|
1575 TbU = FMA(KP707106781, T8x, T8u);
|
Chris@10
|
1576 TbW = FMA(KP707106781, T8V, T8G);
|
Chris@10
|
1577 T8W = FNMS(KP707106781, T8V, T8G);
|
Chris@10
|
1578 ii[WS(rs, 62)] = FMA(KP980785280, TjO, TjN);
|
Chris@10
|
1579 ii[WS(rs, 30)] = FNMS(KP980785280, TjO, TjN);
|
Chris@10
|
1580 Tbz = FMA(KP668178637, T8n, T8y);
|
Chris@10
|
1581 T8z = FNMS(KP668178637, T8y, T8n);
|
Chris@10
|
1582 T97 = FNMS(KP707106781, T96, T93);
|
Chris@10
|
1583 TbX = FMA(KP707106781, T96, T93);
|
Chris@10
|
1584 {
|
Chris@10
|
1585 E Tae, TbE, Ta3, Tah;
|
Chris@10
|
1586 {
|
Chris@10
|
1587 E T9x, Ta2, TbA, T98;
|
Chris@10
|
1588 Tc4 = FMA(KP707106781, T9w, T9h);
|
Chris@10
|
1589 T9x = FNMS(KP707106781, T9w, T9h);
|
Chris@10
|
1590 Ta2 = T9M - Ta1;
|
Chris@10
|
1591 Tc2 = Ta1 + T9M;
|
Chris@10
|
1592 Tc1 = FMA(KP707106781, Tad, Taa);
|
Chris@10
|
1593 Tae = FNMS(KP707106781, Tad, Taa);
|
Chris@10
|
1594 TbA = FNMS(KP668178637, T8W, T97);
|
Chris@10
|
1595 T98 = FMA(KP668178637, T97, T8W);
|
Chris@10
|
1596 TbE = FMA(KP923879532, Ta2, T9x);
|
Chris@10
|
1597 Ta3 = FNMS(KP923879532, Ta2, T9x);
|
Chris@10
|
1598 TbB = Tbz + TbA;
|
Chris@10
|
1599 Tkw = TbA - Tbz;
|
Chris@10
|
1600 Tkq = T8z + T98;
|
Chris@10
|
1601 T99 = T8z - T98;
|
Chris@10
|
1602 Tah = Taf - Tag;
|
Chris@10
|
1603 Tc5 = Taf + Tag;
|
Chris@10
|
1604 }
|
Chris@10
|
1605 {
|
Chris@10
|
1606 E Tc8, Tc9, TbD, Tai;
|
Chris@10
|
1607 Tbn = FNMS(KP707106781, Tbm, Tbj);
|
Chris@10
|
1608 Tc8 = FMA(KP707106781, Tbm, Tbj);
|
Chris@10
|
1609 Tc9 = Tba + TaV;
|
Chris@10
|
1610 Tbb = TaV - Tba;
|
Chris@10
|
1611 TaG = FNMS(KP707106781, TaF, Taq);
|
Chris@10
|
1612 Tcb = FMA(KP707106781, TaF, Taq);
|
Chris@10
|
1613 TbD = FMA(KP923879532, Tah, Tae);
|
Chris@10
|
1614 Tai = FNMS(KP923879532, Tah, Tae);
|
Chris@10
|
1615 Tct = FMA(KP923879532, Tc9, Tc8);
|
Chris@10
|
1616 Tca = FNMS(KP923879532, Tc9, Tc8);
|
Chris@10
|
1617 TbF = FMA(KP303346683, TbE, TbD);
|
Chris@10
|
1618 TbL = FNMS(KP303346683, TbD, TbE);
|
Chris@10
|
1619 Tbv = FNMS(KP534511135, Ta3, Tai);
|
Chris@10
|
1620 Taj = FMA(KP534511135, Tai, Ta3);
|
Chris@10
|
1621 Tcc = Tbo + Tbp;
|
Chris@10
|
1622 Tbq = Tbo - Tbp;
|
Chris@10
|
1623 }
|
Chris@10
|
1624 }
|
Chris@10
|
1625 }
|
Chris@10
|
1626 {
|
Chris@10
|
1627 E Tcq, Tc3, Tcs, Tcd, Tcp, Tc6;
|
Chris@10
|
1628 Tcs = FMA(KP923879532, Tcc, Tcb);
|
Chris@10
|
1629 Tcd = FNMS(KP923879532, Tcc, Tcb);
|
Chris@10
|
1630 Tcq = FMA(KP923879532, Tc2, Tc1);
|
Chris@10
|
1631 Tc3 = FNMS(KP923879532, Tc2, Tc1);
|
Chris@10
|
1632 Tcu = FNMS(KP098491403, Tct, Tcs);
|
Chris@10
|
1633 Tcy = FMA(KP098491403, Tcs, Tct);
|
Chris@10
|
1634 Tci = FMA(KP820678790, Tca, Tcd);
|
Chris@10
|
1635 Tce = FNMS(KP820678790, Tcd, Tca);
|
Chris@10
|
1636 Tcp = FMA(KP923879532, Tc5, Tc4);
|
Chris@10
|
1637 Tc6 = FNMS(KP923879532, Tc5, Tc4);
|
Chris@10
|
1638 Tcl = FNMS(KP198912367, TbT, TbU);
|
Chris@10
|
1639 TbV = FMA(KP198912367, TbU, TbT);
|
Chris@10
|
1640 Tcr = FMA(KP098491403, Tcq, Tcp);
|
Chris@10
|
1641 Tcx = FNMS(KP098491403, Tcp, Tcq);
|
Chris@10
|
1642 Tch = FNMS(KP820678790, Tc3, Tc6);
|
Chris@10
|
1643 Tc7 = FMA(KP820678790, Tc6, Tc3);
|
Chris@10
|
1644 }
|
Chris@10
|
1645 {
|
Chris@10
|
1646 E TbH, Tbc, Tcm, TbY;
|
Chris@10
|
1647 Tcm = FMA(KP198912367, TbW, TbX);
|
Chris@10
|
1648 TbY = FNMS(KP198912367, TbX, TbW);
|
Chris@10
|
1649 TbH = FMA(KP923879532, Tbb, TaG);
|
Chris@10
|
1650 Tbc = FNMS(KP923879532, Tbb, TaG);
|
Chris@10
|
1651 TkE = Tcm - Tcl;
|
Chris@10
|
1652 Tcn = Tcl + Tcm;
|
Chris@10
|
1653 TkK = TbV + TbY;
|
Chris@10
|
1654 TbZ = TbV - TbY;
|
Chris@10
|
1655 {
|
Chris@10
|
1656 E T7t, Tkm, TbG, Tbr;
|
Chris@10
|
1657 TjS = T7l + T7s;
|
Chris@10
|
1658 T7t = T7l - T7s;
|
Chris@10
|
1659 Tkm = TcC - TcB;
|
Chris@10
|
1660 TcD = TcB + TcC;
|
Chris@10
|
1661 TbP = FNMS(KP414213562, T7B, T7I);
|
Chris@10
|
1662 T7J = FMA(KP414213562, T7I, T7B);
|
Chris@10
|
1663 TbG = FMA(KP923879532, Tbq, Tbn);
|
Chris@10
|
1664 Tbr = FNMS(KP923879532, Tbq, Tbn);
|
Chris@10
|
1665 TbO = FNMS(KP707106781, T7t, T7e);
|
Chris@10
|
1666 T7u = FMA(KP707106781, T7t, T7e);
|
Chris@10
|
1667 TkB = FNMS(KP707106781, Tkm, Tkl);
|
Chris@10
|
1668 Tkn = FMA(KP707106781, Tkm, Tkl);
|
Chris@10
|
1669 TbI = FNMS(KP303346683, TbH, TbG);
|
Chris@10
|
1670 TbM = FMA(KP303346683, TbG, TbH);
|
Chris@10
|
1671 Tbw = FMA(KP534511135, Tbc, Tbr);
|
Chris@10
|
1672 Tbs = FNMS(KP534511135, Tbr, Tbc);
|
Chris@10
|
1673 T7Y = FNMS(KP414213562, T7X, T7Q);
|
Chris@10
|
1674 TbQ = FMA(KP414213562, T7Q, T7X);
|
Chris@10
|
1675 }
|
Chris@10
|
1676 }
|
Chris@10
|
1677 }
|
Chris@10
|
1678 {
|
Chris@10
|
1679 E TkJ, TkD, Tck, TbS, TbK, Tku, Tkt, TbN;
|
Chris@10
|
1680 {
|
Chris@10
|
1681 E TkA, Tby, Tkp, Tbu, Tkz, Tbx;
|
Chris@10
|
1682 {
|
Chris@10
|
1683 E Tbt, T9a, Tkx, Tky, Tkv;
|
Chris@10
|
1684 TkA = Taj + Tbs;
|
Chris@10
|
1685 Tbt = Taj - Tbs;
|
Chris@10
|
1686 {
|
Chris@10
|
1687 E TkC, T7Z, Tko, TbR, T80;
|
Chris@10
|
1688 TkC = T7J + T7Y;
|
Chris@10
|
1689 T7Z = T7J - T7Y;
|
Chris@10
|
1690 Tko = TbQ - TbP;
|
Chris@10
|
1691 TbR = TbP + TbQ;
|
Chris@10
|
1692 TkJ = FMA(KP923879532, TkC, TkB);
|
Chris@10
|
1693 TkD = FNMS(KP923879532, TkC, TkB);
|
Chris@10
|
1694 Tby = FMA(KP923879532, T7Z, T7u);
|
Chris@10
|
1695 T80 = FNMS(KP923879532, T7Z, T7u);
|
Chris@10
|
1696 Tkv = FNMS(KP923879532, Tko, Tkn);
|
Chris@10
|
1697 Tkp = FMA(KP923879532, Tko, Tkn);
|
Chris@10
|
1698 Tck = FMA(KP923879532, TbR, TbO);
|
Chris@10
|
1699 TbS = FNMS(KP923879532, TbR, TbO);
|
Chris@10
|
1700 T9a = FMA(KP831469612, T99, T80);
|
Chris@10
|
1701 Tbu = FNMS(KP831469612, T99, T80);
|
Chris@10
|
1702 }
|
Chris@10
|
1703 Tkz = FNMS(KP831469612, Tkw, Tkv);
|
Chris@10
|
1704 Tkx = FMA(KP831469612, Tkw, Tkv);
|
Chris@10
|
1705 Tky = Tbw - Tbv;
|
Chris@10
|
1706 Tbx = Tbv + Tbw;
|
Chris@10
|
1707 ri[WS(rs, 11)] = FMA(KP881921264, Tbt, T9a);
|
Chris@10
|
1708 ri[WS(rs, 43)] = FNMS(KP881921264, Tbt, T9a);
|
Chris@10
|
1709 ii[WS(rs, 43)] = FNMS(KP881921264, Tky, Tkx);
|
Chris@10
|
1710 ii[WS(rs, 11)] = FMA(KP881921264, Tky, Tkx);
|
Chris@10
|
1711 }
|
Chris@10
|
1712 {
|
Chris@10
|
1713 E TbC, TbJ, Tkr, Tks;
|
Chris@10
|
1714 TbK = FNMS(KP831469612, TbB, Tby);
|
Chris@10
|
1715 TbC = FMA(KP831469612, TbB, Tby);
|
Chris@10
|
1716 ri[WS(rs, 59)] = FMA(KP881921264, Tbx, Tbu);
|
Chris@10
|
1717 ri[WS(rs, 27)] = FNMS(KP881921264, Tbx, Tbu);
|
Chris@10
|
1718 ii[WS(rs, 59)] = FMA(KP881921264, TkA, Tkz);
|
Chris@10
|
1719 ii[WS(rs, 27)] = FNMS(KP881921264, TkA, Tkz);
|
Chris@10
|
1720 TbJ = TbF + TbI;
|
Chris@10
|
1721 Tku = TbI - TbF;
|
Chris@10
|
1722 Tkt = FNMS(KP831469612, Tkq, Tkp);
|
Chris@10
|
1723 Tkr = FMA(KP831469612, Tkq, Tkp);
|
Chris@10
|
1724 Tks = TbL + TbM;
|
Chris@10
|
1725 TbN = TbL - TbM;
|
Chris@10
|
1726 ri[WS(rs, 3)] = FMA(KP956940335, TbJ, TbC);
|
Chris@10
|
1727 ri[WS(rs, 35)] = FNMS(KP956940335, TbJ, TbC);
|
Chris@10
|
1728 ii[WS(rs, 35)] = FNMS(KP956940335, Tks, Tkr);
|
Chris@10
|
1729 ii[WS(rs, 3)] = FMA(KP956940335, Tks, Tkr);
|
Chris@10
|
1730 }
|
Chris@10
|
1731 }
|
Chris@10
|
1732 {
|
Chris@10
|
1733 E Tcg, TkI, TkH, Tcj;
|
Chris@10
|
1734 {
|
Chris@10
|
1735 E Tc0, Tcf, TkF, TkG;
|
Chris@10
|
1736 Tcg = FNMS(KP980785280, TbZ, TbS);
|
Chris@10
|
1737 Tc0 = FMA(KP980785280, TbZ, TbS);
|
Chris@10
|
1738 ri[WS(rs, 19)] = FMA(KP956940335, TbN, TbK);
|
Chris@10
|
1739 ri[WS(rs, 51)] = FNMS(KP956940335, TbN, TbK);
|
Chris@10
|
1740 ii[WS(rs, 51)] = FNMS(KP956940335, Tku, Tkt);
|
Chris@10
|
1741 ii[WS(rs, 19)] = FMA(KP956940335, Tku, Tkt);
|
Chris@10
|
1742 Tcf = Tc7 + Tce;
|
Chris@10
|
1743 TkI = Tce - Tc7;
|
Chris@10
|
1744 TkH = FNMS(KP980785280, TkE, TkD);
|
Chris@10
|
1745 TkF = FMA(KP980785280, TkE, TkD);
|
Chris@10
|
1746 TkG = Tch + Tci;
|
Chris@10
|
1747 Tcj = Tch - Tci;
|
Chris@10
|
1748 ri[WS(rs, 7)] = FMA(KP773010453, Tcf, Tc0);
|
Chris@10
|
1749 ri[WS(rs, 39)] = FNMS(KP773010453, Tcf, Tc0);
|
Chris@10
|
1750 ii[WS(rs, 39)] = FNMS(KP773010453, TkG, TkF);
|
Chris@10
|
1751 ii[WS(rs, 7)] = FMA(KP773010453, TkG, TkF);
|
Chris@10
|
1752 }
|
Chris@10
|
1753 {
|
Chris@10
|
1754 E Tco, Tcv, TkL, TkM;
|
Chris@10
|
1755 Tcw = FMA(KP980785280, Tcn, Tck);
|
Chris@10
|
1756 Tco = FNMS(KP980785280, Tcn, Tck);
|
Chris@10
|
1757 ri[WS(rs, 23)] = FMA(KP773010453, Tcj, Tcg);
|
Chris@10
|
1758 ri[WS(rs, 55)] = FNMS(KP773010453, Tcj, Tcg);
|
Chris@10
|
1759 ii[WS(rs, 55)] = FNMS(KP773010453, TkI, TkH);
|
Chris@10
|
1760 ii[WS(rs, 23)] = FMA(KP773010453, TkI, TkH);
|
Chris@10
|
1761 Tcv = Tcr - Tcu;
|
Chris@10
|
1762 TkO = Tcr + Tcu;
|
Chris@10
|
1763 TkN = FMA(KP980785280, TkK, TkJ);
|
Chris@10
|
1764 TkL = FNMS(KP980785280, TkK, TkJ);
|
Chris@10
|
1765 TkM = Tcy - Tcx;
|
Chris@10
|
1766 Tcz = Tcx + Tcy;
|
Chris@10
|
1767 ri[WS(rs, 15)] = FMA(KP995184726, Tcv, Tco);
|
Chris@10
|
1768 ri[WS(rs, 47)] = FNMS(KP995184726, Tcv, Tco);
|
Chris@10
|
1769 ii[WS(rs, 47)] = FNMS(KP995184726, TkM, TkL);
|
Chris@10
|
1770 ii[WS(rs, 15)] = FMA(KP995184726, TkM, TkL);
|
Chris@10
|
1771 }
|
Chris@10
|
1772 }
|
Chris@10
|
1773 }
|
Chris@10
|
1774 }
|
Chris@10
|
1775 {
|
Chris@10
|
1776 E TdN, Tk2, TjW, Td1, TdR, TdX, TdH, Tdl, TeG, TeK, Teu, Teq, TeD, TeJ, Tet;
|
Chris@10
|
1777 E Tej, Tka, Tez, Tkg, Teb, Te1, TcH, Te0, TcE, Tk7, TjT, TdU, TdY, TdI, TdE;
|
Chris@10
|
1778 E TcK, Te2;
|
Chris@10
|
1779 {
|
Chris@10
|
1780 E Te5, Te6, Te8, Te9, Teg, Tee, Ted, Teh, Tdz, Tdv, Tdo, Ten, TeF, Tem, Teo;
|
Chris@10
|
1781 E TdC, Tex, Te7;
|
Chris@10
|
1782 {
|
Chris@10
|
1783 E TcP, TcS, TcW, TcZ;
|
Chris@10
|
1784 Te5 = FNMS(KP707106781, TcO, TcN);
|
Chris@10
|
1785 TcP = FMA(KP707106781, TcO, TcN);
|
Chris@10
|
1786 ri[WS(rs, 63)] = FMA(KP995184726, Tcz, Tcw);
|
Chris@10
|
1787 ri[WS(rs, 31)] = FNMS(KP995184726, Tcz, Tcw);
|
Chris@10
|
1788 ii[WS(rs, 63)] = FMA(KP995184726, TkO, TkN);
|
Chris@10
|
1789 ii[WS(rs, 31)] = FNMS(KP995184726, TkO, TkN);
|
Chris@10
|
1790 TcS = FMA(KP707106781, TcR, TcQ);
|
Chris@10
|
1791 Te6 = FNMS(KP707106781, TcR, TcQ);
|
Chris@10
|
1792 Te8 = FNMS(KP707106781, TcV, TcU);
|
Chris@10
|
1793 TcW = FMA(KP707106781, TcV, TcU);
|
Chris@10
|
1794 TcZ = FMA(KP707106781, TcY, TcX);
|
Chris@10
|
1795 Te9 = FNMS(KP707106781, TcY, TcX);
|
Chris@10
|
1796 {
|
Chris@10
|
1797 E Tdg, TdQ, Tdd, Tdj;
|
Chris@10
|
1798 {
|
Chris@10
|
1799 E Td5, TdL, TcT, TdM, Td0, Tdc;
|
Chris@10
|
1800 Teg = FNMS(KP707106781, Td4, Td3);
|
Chris@10
|
1801 Td5 = FMA(KP707106781, Td4, Td3);
|
Chris@10
|
1802 TdL = FMA(KP198912367, TcP, TcS);
|
Chris@10
|
1803 TcT = FNMS(KP198912367, TcS, TcP);
|
Chris@10
|
1804 TdM = FNMS(KP198912367, TcW, TcZ);
|
Chris@10
|
1805 Td0 = FMA(KP198912367, TcZ, TcW);
|
Chris@10
|
1806 Tdc = Td8 + Tdb;
|
Chris@10
|
1807 Tee = Tdb - Td8;
|
Chris@10
|
1808 Ted = FNMS(KP707106781, Tdf, Tde);
|
Chris@10
|
1809 Tdg = FMA(KP707106781, Tdf, Tde);
|
Chris@10
|
1810 TdN = TdL + TdM;
|
Chris@10
|
1811 Tk2 = TdM - TdL;
|
Chris@10
|
1812 TjW = TcT + Td0;
|
Chris@10
|
1813 Td1 = TcT - Td0;
|
Chris@10
|
1814 TdQ = FMA(KP923879532, Tdc, Td5);
|
Chris@10
|
1815 Tdd = FNMS(KP923879532, Tdc, Td5);
|
Chris@10
|
1816 Tdj = Tdh + Tdi;
|
Chris@10
|
1817 Teh = Tdh - Tdi;
|
Chris@10
|
1818 }
|
Chris@10
|
1819 {
|
Chris@10
|
1820 E Tek, Tel, TdP, Tdk;
|
Chris@10
|
1821 Tdz = FMA(KP707106781, Tdy, Tdx);
|
Chris@10
|
1822 Tek = FNMS(KP707106781, Tdy, Tdx);
|
Chris@10
|
1823 Tel = Tdu - Tdr;
|
Chris@10
|
1824 Tdv = Tdr + Tdu;
|
Chris@10
|
1825 Tdo = FMA(KP707106781, Tdn, Tdm);
|
Chris@10
|
1826 Ten = FNMS(KP707106781, Tdn, Tdm);
|
Chris@10
|
1827 TdP = FMA(KP923879532, Tdj, Tdg);
|
Chris@10
|
1828 Tdk = FNMS(KP923879532, Tdj, Tdg);
|
Chris@10
|
1829 TeF = FMA(KP923879532, Tel, Tek);
|
Chris@10
|
1830 Tem = FNMS(KP923879532, Tel, Tek);
|
Chris@10
|
1831 TdR = FMA(KP098491403, TdQ, TdP);
|
Chris@10
|
1832 TdX = FNMS(KP098491403, TdP, TdQ);
|
Chris@10
|
1833 TdH = FNMS(KP820678790, Tdd, Tdk);
|
Chris@10
|
1834 Tdl = FMA(KP820678790, Tdk, Tdd);
|
Chris@10
|
1835 Teo = TdA - TdB;
|
Chris@10
|
1836 TdC = TdA + TdB;
|
Chris@10
|
1837 }
|
Chris@10
|
1838 }
|
Chris@10
|
1839 }
|
Chris@10
|
1840 {
|
Chris@10
|
1841 E TeC, Tef, TeE, Tep, TeB, Tei;
|
Chris@10
|
1842 TeE = FMA(KP923879532, Teo, Ten);
|
Chris@10
|
1843 Tep = FNMS(KP923879532, Teo, Ten);
|
Chris@10
|
1844 TeC = FMA(KP923879532, Tee, Ted);
|
Chris@10
|
1845 Tef = FNMS(KP923879532, Tee, Ted);
|
Chris@10
|
1846 TeG = FNMS(KP303346683, TeF, TeE);
|
Chris@10
|
1847 TeK = FMA(KP303346683, TeE, TeF);
|
Chris@10
|
1848 Teu = FMA(KP534511135, Tem, Tep);
|
Chris@10
|
1849 Teq = FNMS(KP534511135, Tep, Tem);
|
Chris@10
|
1850 TeB = FMA(KP923879532, Teh, Teg);
|
Chris@10
|
1851 Tei = FNMS(KP923879532, Teh, Teg);
|
Chris@10
|
1852 Tex = FNMS(KP668178637, Te5, Te6);
|
Chris@10
|
1853 Te7 = FMA(KP668178637, Te6, Te5);
|
Chris@10
|
1854 TeD = FMA(KP303346683, TeC, TeB);
|
Chris@10
|
1855 TeJ = FNMS(KP303346683, TeB, TeC);
|
Chris@10
|
1856 Tet = FNMS(KP534511135, Tef, Tei);
|
Chris@10
|
1857 Tej = FMA(KP534511135, Tei, Tef);
|
Chris@10
|
1858 }
|
Chris@10
|
1859 {
|
Chris@10
|
1860 E TdT, Tdw, Tey, Tea, TdS, TdD;
|
Chris@10
|
1861 Tey = FMA(KP668178637, Te8, Te9);
|
Chris@10
|
1862 Tea = FNMS(KP668178637, Te9, Te8);
|
Chris@10
|
1863 TdT = FMA(KP923879532, Tdv, Tdo);
|
Chris@10
|
1864 Tdw = FNMS(KP923879532, Tdv, Tdo);
|
Chris@10
|
1865 Tka = Tey - Tex;
|
Chris@10
|
1866 Tez = Tex + Tey;
|
Chris@10
|
1867 Tkg = Te7 + Tea;
|
Chris@10
|
1868 Teb = Te7 - Tea;
|
Chris@10
|
1869 Te1 = FNMS(KP414213562, TcF, TcG);
|
Chris@10
|
1870 TcH = FMA(KP414213562, TcG, TcF);
|
Chris@10
|
1871 TdS = FMA(KP923879532, TdC, Tdz);
|
Chris@10
|
1872 TdD = FNMS(KP923879532, TdC, Tdz);
|
Chris@10
|
1873 Te0 = FNMS(KP707106781, TcD, TcA);
|
Chris@10
|
1874 TcE = FMA(KP707106781, TcD, TcA);
|
Chris@10
|
1875 Tk7 = FNMS(KP707106781, TjS, TjR);
|
Chris@10
|
1876 TjT = FMA(KP707106781, TjS, TjR);
|
Chris@10
|
1877 TdU = FNMS(KP098491403, TdT, TdS);
|
Chris@10
|
1878 TdY = FMA(KP098491403, TdS, TdT);
|
Chris@10
|
1879 TdI = FMA(KP820678790, Tdw, TdD);
|
Chris@10
|
1880 TdE = FNMS(KP820678790, TdD, Tdw);
|
Chris@10
|
1881 TcK = FNMS(KP414213562, TcJ, TcI);
|
Chris@10
|
1882 Te2 = FMA(KP414213562, TcI, TcJ);
|
Chris@10
|
1883 }
|
Chris@10
|
1884 }
|
Chris@10
|
1885 {
|
Chris@10
|
1886 E Tkf, Tk9, Tew, Te4, TdW, Tk0, TjZ, TdZ;
|
Chris@10
|
1887 {
|
Chris@10
|
1888 E Tk6, TdK, TjV, TdG, Tk5, TdJ;
|
Chris@10
|
1889 {
|
Chris@10
|
1890 E TdF, Td2, Tk3, Tk4, Tk1;
|
Chris@10
|
1891 Tk6 = Tdl + TdE;
|
Chris@10
|
1892 TdF = Tdl - TdE;
|
Chris@10
|
1893 {
|
Chris@10
|
1894 E Tk8, TcL, TjU, Te3, TcM;
|
Chris@10
|
1895 Tk8 = TcK - TcH;
|
Chris@10
|
1896 TcL = TcH + TcK;
|
Chris@10
|
1897 TjU = Te1 + Te2;
|
Chris@10
|
1898 Te3 = Te1 - Te2;
|
Chris@10
|
1899 Tkf = FNMS(KP923879532, Tk8, Tk7);
|
Chris@10
|
1900 Tk9 = FMA(KP923879532, Tk8, Tk7);
|
Chris@10
|
1901 TdK = FMA(KP923879532, TcL, TcE);
|
Chris@10
|
1902 TcM = FNMS(KP923879532, TcL, TcE);
|
Chris@10
|
1903 Tk1 = FNMS(KP923879532, TjU, TjT);
|
Chris@10
|
1904 TjV = FMA(KP923879532, TjU, TjT);
|
Chris@10
|
1905 Tew = FNMS(KP923879532, Te3, Te0);
|
Chris@10
|
1906 Te4 = FMA(KP923879532, Te3, Te0);
|
Chris@10
|
1907 Td2 = FMA(KP980785280, Td1, TcM);
|
Chris@10
|
1908 TdG = FNMS(KP980785280, Td1, TcM);
|
Chris@10
|
1909 }
|
Chris@10
|
1910 Tk5 = FNMS(KP980785280, Tk2, Tk1);
|
Chris@10
|
1911 Tk3 = FMA(KP980785280, Tk2, Tk1);
|
Chris@10
|
1912 Tk4 = TdI - TdH;
|
Chris@10
|
1913 TdJ = TdH + TdI;
|
Chris@10
|
1914 ri[WS(rs, 9)] = FMA(KP773010453, TdF, Td2);
|
Chris@10
|
1915 ri[WS(rs, 41)] = FNMS(KP773010453, TdF, Td2);
|
Chris@10
|
1916 ii[WS(rs, 41)] = FNMS(KP773010453, Tk4, Tk3);
|
Chris@10
|
1917 ii[WS(rs, 9)] = FMA(KP773010453, Tk4, Tk3);
|
Chris@10
|
1918 }
|
Chris@10
|
1919 {
|
Chris@10
|
1920 E TdO, TdV, TjX, TjY;
|
Chris@10
|
1921 TdW = FNMS(KP980785280, TdN, TdK);
|
Chris@10
|
1922 TdO = FMA(KP980785280, TdN, TdK);
|
Chris@10
|
1923 ri[WS(rs, 57)] = FMA(KP773010453, TdJ, TdG);
|
Chris@10
|
1924 ri[WS(rs, 25)] = FNMS(KP773010453, TdJ, TdG);
|
Chris@10
|
1925 ii[WS(rs, 57)] = FMA(KP773010453, Tk6, Tk5);
|
Chris@10
|
1926 ii[WS(rs, 25)] = FNMS(KP773010453, Tk6, Tk5);
|
Chris@10
|
1927 TdV = TdR + TdU;
|
Chris@10
|
1928 Tk0 = TdU - TdR;
|
Chris@10
|
1929 TjZ = FNMS(KP980785280, TjW, TjV);
|
Chris@10
|
1930 TjX = FMA(KP980785280, TjW, TjV);
|
Chris@10
|
1931 TjY = TdX + TdY;
|
Chris@10
|
1932 TdZ = TdX - TdY;
|
Chris@10
|
1933 ri[WS(rs, 1)] = FMA(KP995184726, TdV, TdO);
|
Chris@10
|
1934 ri[WS(rs, 33)] = FNMS(KP995184726, TdV, TdO);
|
Chris@10
|
1935 ii[WS(rs, 33)] = FNMS(KP995184726, TjY, TjX);
|
Chris@10
|
1936 ii[WS(rs, 1)] = FMA(KP995184726, TjY, TjX);
|
Chris@10
|
1937 }
|
Chris@10
|
1938 }
|
Chris@10
|
1939 {
|
Chris@10
|
1940 E Tes, Tke, Tkd, Tev;
|
Chris@10
|
1941 {
|
Chris@10
|
1942 E Tec, Ter, Tkb, Tkc;
|
Chris@10
|
1943 Tes = FNMS(KP831469612, Teb, Te4);
|
Chris@10
|
1944 Tec = FMA(KP831469612, Teb, Te4);
|
Chris@10
|
1945 ri[WS(rs, 17)] = FMA(KP995184726, TdZ, TdW);
|
Chris@10
|
1946 ri[WS(rs, 49)] = FNMS(KP995184726, TdZ, TdW);
|
Chris@10
|
1947 ii[WS(rs, 49)] = FNMS(KP995184726, Tk0, TjZ);
|
Chris@10
|
1948 ii[WS(rs, 17)] = FMA(KP995184726, Tk0, TjZ);
|
Chris@10
|
1949 Ter = Tej + Teq;
|
Chris@10
|
1950 Tke = Teq - Tej;
|
Chris@10
|
1951 Tkd = FNMS(KP831469612, Tka, Tk9);
|
Chris@10
|
1952 Tkb = FMA(KP831469612, Tka, Tk9);
|
Chris@10
|
1953 Tkc = Tet + Teu;
|
Chris@10
|
1954 Tev = Tet - Teu;
|
Chris@10
|
1955 ri[WS(rs, 5)] = FMA(KP881921264, Ter, Tec);
|
Chris@10
|
1956 ri[WS(rs, 37)] = FNMS(KP881921264, Ter, Tec);
|
Chris@10
|
1957 ii[WS(rs, 37)] = FNMS(KP881921264, Tkc, Tkb);
|
Chris@10
|
1958 ii[WS(rs, 5)] = FMA(KP881921264, Tkc, Tkb);
|
Chris@10
|
1959 }
|
Chris@10
|
1960 {
|
Chris@10
|
1961 E TeA, TeH, Tkh, Tki;
|
Chris@10
|
1962 TeI = FMA(KP831469612, Tez, Tew);
|
Chris@10
|
1963 TeA = FNMS(KP831469612, Tez, Tew);
|
Chris@10
|
1964 ri[WS(rs, 21)] = FMA(KP881921264, Tev, Tes);
|
Chris@10
|
1965 ri[WS(rs, 53)] = FNMS(KP881921264, Tev, Tes);
|
Chris@10
|
1966 ii[WS(rs, 53)] = FNMS(KP881921264, Tke, Tkd);
|
Chris@10
|
1967 ii[WS(rs, 21)] = FMA(KP881921264, Tke, Tkd);
|
Chris@10
|
1968 TeH = TeD - TeG;
|
Chris@10
|
1969 Tkk = TeD + TeG;
|
Chris@10
|
1970 Tkj = FMA(KP831469612, Tkg, Tkf);
|
Chris@10
|
1971 Tkh = FNMS(KP831469612, Tkg, Tkf);
|
Chris@10
|
1972 Tki = TeK - TeJ;
|
Chris@10
|
1973 TeL = TeJ + TeK;
|
Chris@10
|
1974 ri[WS(rs, 13)] = FMA(KP956940335, TeH, TeA);
|
Chris@10
|
1975 ri[WS(rs, 45)] = FNMS(KP956940335, TeH, TeA);
|
Chris@10
|
1976 ii[WS(rs, 45)] = FNMS(KP956940335, Tki, Tkh);
|
Chris@10
|
1977 ii[WS(rs, 13)] = FMA(KP956940335, Tki, Tkh);
|
Chris@10
|
1978 }
|
Chris@10
|
1979 }
|
Chris@10
|
1980 }
|
Chris@10
|
1981 }
|
Chris@10
|
1982 }
|
Chris@10
|
1983 }
|
Chris@10
|
1984 }
|
Chris@10
|
1985 ri[WS(rs, 61)] = FMA(KP956940335, TeL, TeI);
|
Chris@10
|
1986 ri[WS(rs, 29)] = FNMS(KP956940335, TeL, TeI);
|
Chris@10
|
1987 ii[WS(rs, 61)] = FMA(KP956940335, Tkk, Tkj);
|
Chris@10
|
1988 ii[WS(rs, 29)] = FNMS(KP956940335, Tkk, Tkj);
|
Chris@10
|
1989 }
|
Chris@10
|
1990 }
|
Chris@10
|
1991 }
|
Chris@10
|
1992
|
Chris@10
|
1993 static const tw_instr twinstr[] = {
|
Chris@10
|
1994 {TW_FULL, 0, 64},
|
Chris@10
|
1995 {TW_NEXT, 1, 0}
|
Chris@10
|
1996 };
|
Chris@10
|
1997
|
Chris@10
|
1998 static const ct_desc desc = { 64, "t1_64", twinstr, &GENUS, {520, 126, 518, 0}, 0, 0, 0 };
|
Chris@10
|
1999
|
Chris@10
|
2000 void X(codelet_t1_64) (planner *p) {
|
Chris@10
|
2001 X(kdft_dit_register) (p, t1_64, &desc);
|
Chris@10
|
2002 }
|
Chris@10
|
2003 #else /* HAVE_FMA */
|
Chris@10
|
2004
|
Chris@10
|
2005 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 64 -name t1_64 -include t.h */
|
Chris@10
|
2006
|
Chris@10
|
2007 /*
|
Chris@10
|
2008 * This function contains 1038 FP additions, 500 FP multiplications,
|
Chris@10
|
2009 * (or, 808 additions, 270 multiplications, 230 fused multiply/add),
|
Chris@10
|
2010 * 176 stack variables, 15 constants, and 256 memory accesses
|
Chris@10
|
2011 */
|
Chris@10
|
2012 #include "t.h"
|
Chris@10
|
2013
|
Chris@10
|
2014 static void t1_64(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
2015 {
|
Chris@10
|
2016 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
Chris@10
|
2017 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@10
|
2018 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
Chris@10
|
2019 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@10
|
2020 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
Chris@10
|
2021 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@10
|
2022 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
Chris@10
|
2023 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@10
|
2024 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@10
|
2025 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
2026 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
2027 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@10
|
2028 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
2029 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
2030 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
2031 {
|
Chris@10
|
2032 INT m;
|
Chris@10
|
2033 for (m = mb, W = W + (mb * 126); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 126, MAKE_VOLATILE_STRIDE(128, rs)) {
|
Chris@10
|
2034 E Tj, TcL, ThT, Tin, T6b, Taz, TgT, Thn, TG, Thm, TcO, TgO, T6m, ThQ, TaC;
|
Chris@10
|
2035 E Tim, T14, Tfq, T6y, T9O, TaG, Tc0, TcU, TeE, T1r, Tfr, T6J, T9P, TaJ, Tc1;
|
Chris@10
|
2036 E TcZ, TeF, T1Q, T2d, Tfx, Tfu, Tfv, Tfw, T6Q, TaM, Tdb, TeJ, T71, TaQ, T7a;
|
Chris@10
|
2037 E TaN, Td6, TeI, T77, TaP, T2B, T2Y, Tfz, TfA, TfB, TfC, T7h, TaW, Tdm, TeM;
|
Chris@10
|
2038 E T7s, TaU, T7B, TaX, Tdh, TeL, T7y, TaT, T5j, TfR, Tec, Tf0, TfY, Tgy, T8D;
|
Chris@10
|
2039 E Tbl, T8O, Tbx, T9l, Tbm, TdV, TeX, T9i, Tbw, T3M, TfL, TdL, TeQ, TfI, Tgt;
|
Chris@10
|
2040 E T7K, Tb2, T7V, Tbe, T8s, Tb3, Tdu, TeT, T8p, Tbd, T4x, TfJ, TdE, TdM, TfO;
|
Chris@10
|
2041 E Tgu, T87, T8v, T8i, T8u, Tba, Tbg, Tdz, TdN, Tb7, Tbh, T64, TfZ, Te5, Ted;
|
Chris@10
|
2042 E TfU, Tgz, T90, T9o, T9b, T9n, Tbt, Tbz, Te0, Tee, Tbq, TbA;
|
Chris@10
|
2043 {
|
Chris@10
|
2044 E T1, TgR, T6, TgQ, Tc, T68, Th, T69;
|
Chris@10
|
2045 T1 = ri[0];
|
Chris@10
|
2046 TgR = ii[0];
|
Chris@10
|
2047 {
|
Chris@10
|
2048 E T3, T5, T2, T4;
|
Chris@10
|
2049 T3 = ri[WS(rs, 32)];
|
Chris@10
|
2050 T5 = ii[WS(rs, 32)];
|
Chris@10
|
2051 T2 = W[62];
|
Chris@10
|
2052 T4 = W[63];
|
Chris@10
|
2053 T6 = FMA(T2, T3, T4 * T5);
|
Chris@10
|
2054 TgQ = FNMS(T4, T3, T2 * T5);
|
Chris@10
|
2055 }
|
Chris@10
|
2056 {
|
Chris@10
|
2057 E T9, Tb, T8, Ta;
|
Chris@10
|
2058 T9 = ri[WS(rs, 16)];
|
Chris@10
|
2059 Tb = ii[WS(rs, 16)];
|
Chris@10
|
2060 T8 = W[30];
|
Chris@10
|
2061 Ta = W[31];
|
Chris@10
|
2062 Tc = FMA(T8, T9, Ta * Tb);
|
Chris@10
|
2063 T68 = FNMS(Ta, T9, T8 * Tb);
|
Chris@10
|
2064 }
|
Chris@10
|
2065 {
|
Chris@10
|
2066 E Te, Tg, Td, Tf;
|
Chris@10
|
2067 Te = ri[WS(rs, 48)];
|
Chris@10
|
2068 Tg = ii[WS(rs, 48)];
|
Chris@10
|
2069 Td = W[94];
|
Chris@10
|
2070 Tf = W[95];
|
Chris@10
|
2071 Th = FMA(Td, Te, Tf * Tg);
|
Chris@10
|
2072 T69 = FNMS(Tf, Te, Td * Tg);
|
Chris@10
|
2073 }
|
Chris@10
|
2074 {
|
Chris@10
|
2075 E T7, Ti, ThR, ThS;
|
Chris@10
|
2076 T7 = T1 + T6;
|
Chris@10
|
2077 Ti = Tc + Th;
|
Chris@10
|
2078 Tj = T7 + Ti;
|
Chris@10
|
2079 TcL = T7 - Ti;
|
Chris@10
|
2080 ThR = TgR - TgQ;
|
Chris@10
|
2081 ThS = Tc - Th;
|
Chris@10
|
2082 ThT = ThR - ThS;
|
Chris@10
|
2083 Tin = ThS + ThR;
|
Chris@10
|
2084 }
|
Chris@10
|
2085 {
|
Chris@10
|
2086 E T67, T6a, TgP, TgS;
|
Chris@10
|
2087 T67 = T1 - T6;
|
Chris@10
|
2088 T6a = T68 - T69;
|
Chris@10
|
2089 T6b = T67 - T6a;
|
Chris@10
|
2090 Taz = T67 + T6a;
|
Chris@10
|
2091 TgP = T68 + T69;
|
Chris@10
|
2092 TgS = TgQ + TgR;
|
Chris@10
|
2093 TgT = TgP + TgS;
|
Chris@10
|
2094 Thn = TgS - TgP;
|
Chris@10
|
2095 }
|
Chris@10
|
2096 }
|
Chris@10
|
2097 {
|
Chris@10
|
2098 E To, T6c, Tt, T6d, T6e, T6f, Tz, T6i, TE, T6j, T6h, T6k;
|
Chris@10
|
2099 {
|
Chris@10
|
2100 E Tl, Tn, Tk, Tm;
|
Chris@10
|
2101 Tl = ri[WS(rs, 8)];
|
Chris@10
|
2102 Tn = ii[WS(rs, 8)];
|
Chris@10
|
2103 Tk = W[14];
|
Chris@10
|
2104 Tm = W[15];
|
Chris@10
|
2105 To = FMA(Tk, Tl, Tm * Tn);
|
Chris@10
|
2106 T6c = FNMS(Tm, Tl, Tk * Tn);
|
Chris@10
|
2107 }
|
Chris@10
|
2108 {
|
Chris@10
|
2109 E Tq, Ts, Tp, Tr;
|
Chris@10
|
2110 Tq = ri[WS(rs, 40)];
|
Chris@10
|
2111 Ts = ii[WS(rs, 40)];
|
Chris@10
|
2112 Tp = W[78];
|
Chris@10
|
2113 Tr = W[79];
|
Chris@10
|
2114 Tt = FMA(Tp, Tq, Tr * Ts);
|
Chris@10
|
2115 T6d = FNMS(Tr, Tq, Tp * Ts);
|
Chris@10
|
2116 }
|
Chris@10
|
2117 T6e = T6c - T6d;
|
Chris@10
|
2118 T6f = To - Tt;
|
Chris@10
|
2119 {
|
Chris@10
|
2120 E Tw, Ty, Tv, Tx;
|
Chris@10
|
2121 Tw = ri[WS(rs, 56)];
|
Chris@10
|
2122 Ty = ii[WS(rs, 56)];
|
Chris@10
|
2123 Tv = W[110];
|
Chris@10
|
2124 Tx = W[111];
|
Chris@10
|
2125 Tz = FMA(Tv, Tw, Tx * Ty);
|
Chris@10
|
2126 T6i = FNMS(Tx, Tw, Tv * Ty);
|
Chris@10
|
2127 }
|
Chris@10
|
2128 {
|
Chris@10
|
2129 E TB, TD, TA, TC;
|
Chris@10
|
2130 TB = ri[WS(rs, 24)];
|
Chris@10
|
2131 TD = ii[WS(rs, 24)];
|
Chris@10
|
2132 TA = W[46];
|
Chris@10
|
2133 TC = W[47];
|
Chris@10
|
2134 TE = FMA(TA, TB, TC * TD);
|
Chris@10
|
2135 T6j = FNMS(TC, TB, TA * TD);
|
Chris@10
|
2136 }
|
Chris@10
|
2137 T6h = Tz - TE;
|
Chris@10
|
2138 T6k = T6i - T6j;
|
Chris@10
|
2139 {
|
Chris@10
|
2140 E Tu, TF, TcM, TcN;
|
Chris@10
|
2141 Tu = To + Tt;
|
Chris@10
|
2142 TF = Tz + TE;
|
Chris@10
|
2143 TG = Tu + TF;
|
Chris@10
|
2144 Thm = TF - Tu;
|
Chris@10
|
2145 TcM = T6c + T6d;
|
Chris@10
|
2146 TcN = T6i + T6j;
|
Chris@10
|
2147 TcO = TcM - TcN;
|
Chris@10
|
2148 TgO = TcM + TcN;
|
Chris@10
|
2149 }
|
Chris@10
|
2150 {
|
Chris@10
|
2151 E T6g, T6l, TaA, TaB;
|
Chris@10
|
2152 T6g = T6e - T6f;
|
Chris@10
|
2153 T6l = T6h + T6k;
|
Chris@10
|
2154 T6m = KP707106781 * (T6g - T6l);
|
Chris@10
|
2155 ThQ = KP707106781 * (T6g + T6l);
|
Chris@10
|
2156 TaA = T6f + T6e;
|
Chris@10
|
2157 TaB = T6h - T6k;
|
Chris@10
|
2158 TaC = KP707106781 * (TaA + TaB);
|
Chris@10
|
2159 Tim = KP707106781 * (TaB - TaA);
|
Chris@10
|
2160 }
|
Chris@10
|
2161 }
|
Chris@10
|
2162 {
|
Chris@10
|
2163 E TS, TcQ, T6q, T6t, T13, TcR, T6r, T6w, T6s, T6x;
|
Chris@10
|
2164 {
|
Chris@10
|
2165 E TM, T6o, TR, T6p;
|
Chris@10
|
2166 {
|
Chris@10
|
2167 E TJ, TL, TI, TK;
|
Chris@10
|
2168 TJ = ri[WS(rs, 4)];
|
Chris@10
|
2169 TL = ii[WS(rs, 4)];
|
Chris@10
|
2170 TI = W[6];
|
Chris@10
|
2171 TK = W[7];
|
Chris@10
|
2172 TM = FMA(TI, TJ, TK * TL);
|
Chris@10
|
2173 T6o = FNMS(TK, TJ, TI * TL);
|
Chris@10
|
2174 }
|
Chris@10
|
2175 {
|
Chris@10
|
2176 E TO, TQ, TN, TP;
|
Chris@10
|
2177 TO = ri[WS(rs, 36)];
|
Chris@10
|
2178 TQ = ii[WS(rs, 36)];
|
Chris@10
|
2179 TN = W[70];
|
Chris@10
|
2180 TP = W[71];
|
Chris@10
|
2181 TR = FMA(TN, TO, TP * TQ);
|
Chris@10
|
2182 T6p = FNMS(TP, TO, TN * TQ);
|
Chris@10
|
2183 }
|
Chris@10
|
2184 TS = TM + TR;
|
Chris@10
|
2185 TcQ = T6o + T6p;
|
Chris@10
|
2186 T6q = T6o - T6p;
|
Chris@10
|
2187 T6t = TM - TR;
|
Chris@10
|
2188 }
|
Chris@10
|
2189 {
|
Chris@10
|
2190 E TX, T6u, T12, T6v;
|
Chris@10
|
2191 {
|
Chris@10
|
2192 E TU, TW, TT, TV;
|
Chris@10
|
2193 TU = ri[WS(rs, 20)];
|
Chris@10
|
2194 TW = ii[WS(rs, 20)];
|
Chris@10
|
2195 TT = W[38];
|
Chris@10
|
2196 TV = W[39];
|
Chris@10
|
2197 TX = FMA(TT, TU, TV * TW);
|
Chris@10
|
2198 T6u = FNMS(TV, TU, TT * TW);
|
Chris@10
|
2199 }
|
Chris@10
|
2200 {
|
Chris@10
|
2201 E TZ, T11, TY, T10;
|
Chris@10
|
2202 TZ = ri[WS(rs, 52)];
|
Chris@10
|
2203 T11 = ii[WS(rs, 52)];
|
Chris@10
|
2204 TY = W[102];
|
Chris@10
|
2205 T10 = W[103];
|
Chris@10
|
2206 T12 = FMA(TY, TZ, T10 * T11);
|
Chris@10
|
2207 T6v = FNMS(T10, TZ, TY * T11);
|
Chris@10
|
2208 }
|
Chris@10
|
2209 T13 = TX + T12;
|
Chris@10
|
2210 TcR = T6u + T6v;
|
Chris@10
|
2211 T6r = TX - T12;
|
Chris@10
|
2212 T6w = T6u - T6v;
|
Chris@10
|
2213 }
|
Chris@10
|
2214 T14 = TS + T13;
|
Chris@10
|
2215 Tfq = TcQ + TcR;
|
Chris@10
|
2216 T6s = T6q + T6r;
|
Chris@10
|
2217 T6x = T6t - T6w;
|
Chris@10
|
2218 T6y = FNMS(KP923879532, T6x, KP382683432 * T6s);
|
Chris@10
|
2219 T9O = FMA(KP923879532, T6s, KP382683432 * T6x);
|
Chris@10
|
2220 {
|
Chris@10
|
2221 E TaE, TaF, TcS, TcT;
|
Chris@10
|
2222 TaE = T6q - T6r;
|
Chris@10
|
2223 TaF = T6t + T6w;
|
Chris@10
|
2224 TaG = FNMS(KP382683432, TaF, KP923879532 * TaE);
|
Chris@10
|
2225 Tc0 = FMA(KP382683432, TaE, KP923879532 * TaF);
|
Chris@10
|
2226 TcS = TcQ - TcR;
|
Chris@10
|
2227 TcT = TS - T13;
|
Chris@10
|
2228 TcU = TcS - TcT;
|
Chris@10
|
2229 TeE = TcT + TcS;
|
Chris@10
|
2230 }
|
Chris@10
|
2231 }
|
Chris@10
|
2232 {
|
Chris@10
|
2233 E T1f, TcW, T6B, T6E, T1q, TcX, T6C, T6H, T6D, T6I;
|
Chris@10
|
2234 {
|
Chris@10
|
2235 E T19, T6z, T1e, T6A;
|
Chris@10
|
2236 {
|
Chris@10
|
2237 E T16, T18, T15, T17;
|
Chris@10
|
2238 T16 = ri[WS(rs, 60)];
|
Chris@10
|
2239 T18 = ii[WS(rs, 60)];
|
Chris@10
|
2240 T15 = W[118];
|
Chris@10
|
2241 T17 = W[119];
|
Chris@10
|
2242 T19 = FMA(T15, T16, T17 * T18);
|
Chris@10
|
2243 T6z = FNMS(T17, T16, T15 * T18);
|
Chris@10
|
2244 }
|
Chris@10
|
2245 {
|
Chris@10
|
2246 E T1b, T1d, T1a, T1c;
|
Chris@10
|
2247 T1b = ri[WS(rs, 28)];
|
Chris@10
|
2248 T1d = ii[WS(rs, 28)];
|
Chris@10
|
2249 T1a = W[54];
|
Chris@10
|
2250 T1c = W[55];
|
Chris@10
|
2251 T1e = FMA(T1a, T1b, T1c * T1d);
|
Chris@10
|
2252 T6A = FNMS(T1c, T1b, T1a * T1d);
|
Chris@10
|
2253 }
|
Chris@10
|
2254 T1f = T19 + T1e;
|
Chris@10
|
2255 TcW = T6z + T6A;
|
Chris@10
|
2256 T6B = T6z - T6A;
|
Chris@10
|
2257 T6E = T19 - T1e;
|
Chris@10
|
2258 }
|
Chris@10
|
2259 {
|
Chris@10
|
2260 E T1k, T6F, T1p, T6G;
|
Chris@10
|
2261 {
|
Chris@10
|
2262 E T1h, T1j, T1g, T1i;
|
Chris@10
|
2263 T1h = ri[WS(rs, 12)];
|
Chris@10
|
2264 T1j = ii[WS(rs, 12)];
|
Chris@10
|
2265 T1g = W[22];
|
Chris@10
|
2266 T1i = W[23];
|
Chris@10
|
2267 T1k = FMA(T1g, T1h, T1i * T1j);
|
Chris@10
|
2268 T6F = FNMS(T1i, T1h, T1g * T1j);
|
Chris@10
|
2269 }
|
Chris@10
|
2270 {
|
Chris@10
|
2271 E T1m, T1o, T1l, T1n;
|
Chris@10
|
2272 T1m = ri[WS(rs, 44)];
|
Chris@10
|
2273 T1o = ii[WS(rs, 44)];
|
Chris@10
|
2274 T1l = W[86];
|
Chris@10
|
2275 T1n = W[87];
|
Chris@10
|
2276 T1p = FMA(T1l, T1m, T1n * T1o);
|
Chris@10
|
2277 T6G = FNMS(T1n, T1m, T1l * T1o);
|
Chris@10
|
2278 }
|
Chris@10
|
2279 T1q = T1k + T1p;
|
Chris@10
|
2280 TcX = T6F + T6G;
|
Chris@10
|
2281 T6C = T1k - T1p;
|
Chris@10
|
2282 T6H = T6F - T6G;
|
Chris@10
|
2283 }
|
Chris@10
|
2284 T1r = T1f + T1q;
|
Chris@10
|
2285 Tfr = TcW + TcX;
|
Chris@10
|
2286 T6D = T6B + T6C;
|
Chris@10
|
2287 T6I = T6E - T6H;
|
Chris@10
|
2288 T6J = FMA(KP382683432, T6D, KP923879532 * T6I);
|
Chris@10
|
2289 T9P = FNMS(KP923879532, T6D, KP382683432 * T6I);
|
Chris@10
|
2290 {
|
Chris@10
|
2291 E TaH, TaI, TcV, TcY;
|
Chris@10
|
2292 TaH = T6B - T6C;
|
Chris@10
|
2293 TaI = T6E + T6H;
|
Chris@10
|
2294 TaJ = FMA(KP923879532, TaH, KP382683432 * TaI);
|
Chris@10
|
2295 Tc1 = FNMS(KP382683432, TaH, KP923879532 * TaI);
|
Chris@10
|
2296 TcV = T1f - T1q;
|
Chris@10
|
2297 TcY = TcW - TcX;
|
Chris@10
|
2298 TcZ = TcV + TcY;
|
Chris@10
|
2299 TeF = TcV - TcY;
|
Chris@10
|
2300 }
|
Chris@10
|
2301 }
|
Chris@10
|
2302 {
|
Chris@10
|
2303 E T1y, T6M, T1D, T6N, T1E, Td2, T1J, T74, T1O, T75, T1P, Td3, T21, Td8, T6W;
|
Chris@10
|
2304 E T6Z, T2c, Td9, T6R, T6U;
|
Chris@10
|
2305 {
|
Chris@10
|
2306 E T1v, T1x, T1u, T1w;
|
Chris@10
|
2307 T1v = ri[WS(rs, 2)];
|
Chris@10
|
2308 T1x = ii[WS(rs, 2)];
|
Chris@10
|
2309 T1u = W[2];
|
Chris@10
|
2310 T1w = W[3];
|
Chris@10
|
2311 T1y = FMA(T1u, T1v, T1w * T1x);
|
Chris@10
|
2312 T6M = FNMS(T1w, T1v, T1u * T1x);
|
Chris@10
|
2313 }
|
Chris@10
|
2314 {
|
Chris@10
|
2315 E T1A, T1C, T1z, T1B;
|
Chris@10
|
2316 T1A = ri[WS(rs, 34)];
|
Chris@10
|
2317 T1C = ii[WS(rs, 34)];
|
Chris@10
|
2318 T1z = W[66];
|
Chris@10
|
2319 T1B = W[67];
|
Chris@10
|
2320 T1D = FMA(T1z, T1A, T1B * T1C);
|
Chris@10
|
2321 T6N = FNMS(T1B, T1A, T1z * T1C);
|
Chris@10
|
2322 }
|
Chris@10
|
2323 T1E = T1y + T1D;
|
Chris@10
|
2324 Td2 = T6M + T6N;
|
Chris@10
|
2325 {
|
Chris@10
|
2326 E T1G, T1I, T1F, T1H;
|
Chris@10
|
2327 T1G = ri[WS(rs, 18)];
|
Chris@10
|
2328 T1I = ii[WS(rs, 18)];
|
Chris@10
|
2329 T1F = W[34];
|
Chris@10
|
2330 T1H = W[35];
|
Chris@10
|
2331 T1J = FMA(T1F, T1G, T1H * T1I);
|
Chris@10
|
2332 T74 = FNMS(T1H, T1G, T1F * T1I);
|
Chris@10
|
2333 }
|
Chris@10
|
2334 {
|
Chris@10
|
2335 E T1L, T1N, T1K, T1M;
|
Chris@10
|
2336 T1L = ri[WS(rs, 50)];
|
Chris@10
|
2337 T1N = ii[WS(rs, 50)];
|
Chris@10
|
2338 T1K = W[98];
|
Chris@10
|
2339 T1M = W[99];
|
Chris@10
|
2340 T1O = FMA(T1K, T1L, T1M * T1N);
|
Chris@10
|
2341 T75 = FNMS(T1M, T1L, T1K * T1N);
|
Chris@10
|
2342 }
|
Chris@10
|
2343 T1P = T1J + T1O;
|
Chris@10
|
2344 Td3 = T74 + T75;
|
Chris@10
|
2345 {
|
Chris@10
|
2346 E T1V, T6X, T20, T6Y;
|
Chris@10
|
2347 {
|
Chris@10
|
2348 E T1S, T1U, T1R, T1T;
|
Chris@10
|
2349 T1S = ri[WS(rs, 10)];
|
Chris@10
|
2350 T1U = ii[WS(rs, 10)];
|
Chris@10
|
2351 T1R = W[18];
|
Chris@10
|
2352 T1T = W[19];
|
Chris@10
|
2353 T1V = FMA(T1R, T1S, T1T * T1U);
|
Chris@10
|
2354 T6X = FNMS(T1T, T1S, T1R * T1U);
|
Chris@10
|
2355 }
|
Chris@10
|
2356 {
|
Chris@10
|
2357 E T1X, T1Z, T1W, T1Y;
|
Chris@10
|
2358 T1X = ri[WS(rs, 42)];
|
Chris@10
|
2359 T1Z = ii[WS(rs, 42)];
|
Chris@10
|
2360 T1W = W[82];
|
Chris@10
|
2361 T1Y = W[83];
|
Chris@10
|
2362 T20 = FMA(T1W, T1X, T1Y * T1Z);
|
Chris@10
|
2363 T6Y = FNMS(T1Y, T1X, T1W * T1Z);
|
Chris@10
|
2364 }
|
Chris@10
|
2365 T21 = T1V + T20;
|
Chris@10
|
2366 Td8 = T6X + T6Y;
|
Chris@10
|
2367 T6W = T1V - T20;
|
Chris@10
|
2368 T6Z = T6X - T6Y;
|
Chris@10
|
2369 }
|
Chris@10
|
2370 {
|
Chris@10
|
2371 E T26, T6S, T2b, T6T;
|
Chris@10
|
2372 {
|
Chris@10
|
2373 E T23, T25, T22, T24;
|
Chris@10
|
2374 T23 = ri[WS(rs, 58)];
|
Chris@10
|
2375 T25 = ii[WS(rs, 58)];
|
Chris@10
|
2376 T22 = W[114];
|
Chris@10
|
2377 T24 = W[115];
|
Chris@10
|
2378 T26 = FMA(T22, T23, T24 * T25);
|
Chris@10
|
2379 T6S = FNMS(T24, T23, T22 * T25);
|
Chris@10
|
2380 }
|
Chris@10
|
2381 {
|
Chris@10
|
2382 E T28, T2a, T27, T29;
|
Chris@10
|
2383 T28 = ri[WS(rs, 26)];
|
Chris@10
|
2384 T2a = ii[WS(rs, 26)];
|
Chris@10
|
2385 T27 = W[50];
|
Chris@10
|
2386 T29 = W[51];
|
Chris@10
|
2387 T2b = FMA(T27, T28, T29 * T2a);
|
Chris@10
|
2388 T6T = FNMS(T29, T28, T27 * T2a);
|
Chris@10
|
2389 }
|
Chris@10
|
2390 T2c = T26 + T2b;
|
Chris@10
|
2391 Td9 = T6S + T6T;
|
Chris@10
|
2392 T6R = T26 - T2b;
|
Chris@10
|
2393 T6U = T6S - T6T;
|
Chris@10
|
2394 }
|
Chris@10
|
2395 T1Q = T1E + T1P;
|
Chris@10
|
2396 T2d = T21 + T2c;
|
Chris@10
|
2397 Tfx = T1Q - T2d;
|
Chris@10
|
2398 Tfu = Td2 + Td3;
|
Chris@10
|
2399 Tfv = Td8 + Td9;
|
Chris@10
|
2400 Tfw = Tfu - Tfv;
|
Chris@10
|
2401 {
|
Chris@10
|
2402 E T6O, T6P, Td7, Tda;
|
Chris@10
|
2403 T6O = T6M - T6N;
|
Chris@10
|
2404 T6P = T1J - T1O;
|
Chris@10
|
2405 T6Q = T6O + T6P;
|
Chris@10
|
2406 TaM = T6O - T6P;
|
Chris@10
|
2407 Td7 = T1E - T1P;
|
Chris@10
|
2408 Tda = Td8 - Td9;
|
Chris@10
|
2409 Tdb = Td7 - Tda;
|
Chris@10
|
2410 TeJ = Td7 + Tda;
|
Chris@10
|
2411 }
|
Chris@10
|
2412 {
|
Chris@10
|
2413 E T6V, T70, T78, T79;
|
Chris@10
|
2414 T6V = T6R - T6U;
|
Chris@10
|
2415 T70 = T6W + T6Z;
|
Chris@10
|
2416 T71 = KP707106781 * (T6V - T70);
|
Chris@10
|
2417 TaQ = KP707106781 * (T70 + T6V);
|
Chris@10
|
2418 T78 = T6Z - T6W;
|
Chris@10
|
2419 T79 = T6R + T6U;
|
Chris@10
|
2420 T7a = KP707106781 * (T78 - T79);
|
Chris@10
|
2421 TaN = KP707106781 * (T78 + T79);
|
Chris@10
|
2422 }
|
Chris@10
|
2423 {
|
Chris@10
|
2424 E Td4, Td5, T73, T76;
|
Chris@10
|
2425 Td4 = Td2 - Td3;
|
Chris@10
|
2426 Td5 = T2c - T21;
|
Chris@10
|
2427 Td6 = Td4 - Td5;
|
Chris@10
|
2428 TeI = Td4 + Td5;
|
Chris@10
|
2429 T73 = T1y - T1D;
|
Chris@10
|
2430 T76 = T74 - T75;
|
Chris@10
|
2431 T77 = T73 - T76;
|
Chris@10
|
2432 TaP = T73 + T76;
|
Chris@10
|
2433 }
|
Chris@10
|
2434 }
|
Chris@10
|
2435 {
|
Chris@10
|
2436 E T2j, T7d, T2o, T7e, T2p, Tdd, T2u, T7v, T2z, T7w, T2A, Tde, T2M, Tdj, T7n;
|
Chris@10
|
2437 E T7q, T2X, Tdk, T7i, T7l;
|
Chris@10
|
2438 {
|
Chris@10
|
2439 E T2g, T2i, T2f, T2h;
|
Chris@10
|
2440 T2g = ri[WS(rs, 62)];
|
Chris@10
|
2441 T2i = ii[WS(rs, 62)];
|
Chris@10
|
2442 T2f = W[122];
|
Chris@10
|
2443 T2h = W[123];
|
Chris@10
|
2444 T2j = FMA(T2f, T2g, T2h * T2i);
|
Chris@10
|
2445 T7d = FNMS(T2h, T2g, T2f * T2i);
|
Chris@10
|
2446 }
|
Chris@10
|
2447 {
|
Chris@10
|
2448 E T2l, T2n, T2k, T2m;
|
Chris@10
|
2449 T2l = ri[WS(rs, 30)];
|
Chris@10
|
2450 T2n = ii[WS(rs, 30)];
|
Chris@10
|
2451 T2k = W[58];
|
Chris@10
|
2452 T2m = W[59];
|
Chris@10
|
2453 T2o = FMA(T2k, T2l, T2m * T2n);
|
Chris@10
|
2454 T7e = FNMS(T2m, T2l, T2k * T2n);
|
Chris@10
|
2455 }
|
Chris@10
|
2456 T2p = T2j + T2o;
|
Chris@10
|
2457 Tdd = T7d + T7e;
|
Chris@10
|
2458 {
|
Chris@10
|
2459 E T2r, T2t, T2q, T2s;
|
Chris@10
|
2460 T2r = ri[WS(rs, 14)];
|
Chris@10
|
2461 T2t = ii[WS(rs, 14)];
|
Chris@10
|
2462 T2q = W[26];
|
Chris@10
|
2463 T2s = W[27];
|
Chris@10
|
2464 T2u = FMA(T2q, T2r, T2s * T2t);
|
Chris@10
|
2465 T7v = FNMS(T2s, T2r, T2q * T2t);
|
Chris@10
|
2466 }
|
Chris@10
|
2467 {
|
Chris@10
|
2468 E T2w, T2y, T2v, T2x;
|
Chris@10
|
2469 T2w = ri[WS(rs, 46)];
|
Chris@10
|
2470 T2y = ii[WS(rs, 46)];
|
Chris@10
|
2471 T2v = W[90];
|
Chris@10
|
2472 T2x = W[91];
|
Chris@10
|
2473 T2z = FMA(T2v, T2w, T2x * T2y);
|
Chris@10
|
2474 T7w = FNMS(T2x, T2w, T2v * T2y);
|
Chris@10
|
2475 }
|
Chris@10
|
2476 T2A = T2u + T2z;
|
Chris@10
|
2477 Tde = T7v + T7w;
|
Chris@10
|
2478 {
|
Chris@10
|
2479 E T2G, T7o, T2L, T7p;
|
Chris@10
|
2480 {
|
Chris@10
|
2481 E T2D, T2F, T2C, T2E;
|
Chris@10
|
2482 T2D = ri[WS(rs, 6)];
|
Chris@10
|
2483 T2F = ii[WS(rs, 6)];
|
Chris@10
|
2484 T2C = W[10];
|
Chris@10
|
2485 T2E = W[11];
|
Chris@10
|
2486 T2G = FMA(T2C, T2D, T2E * T2F);
|
Chris@10
|
2487 T7o = FNMS(T2E, T2D, T2C * T2F);
|
Chris@10
|
2488 }
|
Chris@10
|
2489 {
|
Chris@10
|
2490 E T2I, T2K, T2H, T2J;
|
Chris@10
|
2491 T2I = ri[WS(rs, 38)];
|
Chris@10
|
2492 T2K = ii[WS(rs, 38)];
|
Chris@10
|
2493 T2H = W[74];
|
Chris@10
|
2494 T2J = W[75];
|
Chris@10
|
2495 T2L = FMA(T2H, T2I, T2J * T2K);
|
Chris@10
|
2496 T7p = FNMS(T2J, T2I, T2H * T2K);
|
Chris@10
|
2497 }
|
Chris@10
|
2498 T2M = T2G + T2L;
|
Chris@10
|
2499 Tdj = T7o + T7p;
|
Chris@10
|
2500 T7n = T2G - T2L;
|
Chris@10
|
2501 T7q = T7o - T7p;
|
Chris@10
|
2502 }
|
Chris@10
|
2503 {
|
Chris@10
|
2504 E T2R, T7j, T2W, T7k;
|
Chris@10
|
2505 {
|
Chris@10
|
2506 E T2O, T2Q, T2N, T2P;
|
Chris@10
|
2507 T2O = ri[WS(rs, 54)];
|
Chris@10
|
2508 T2Q = ii[WS(rs, 54)];
|
Chris@10
|
2509 T2N = W[106];
|
Chris@10
|
2510 T2P = W[107];
|
Chris@10
|
2511 T2R = FMA(T2N, T2O, T2P * T2Q);
|
Chris@10
|
2512 T7j = FNMS(T2P, T2O, T2N * T2Q);
|
Chris@10
|
2513 }
|
Chris@10
|
2514 {
|
Chris@10
|
2515 E T2T, T2V, T2S, T2U;
|
Chris@10
|
2516 T2T = ri[WS(rs, 22)];
|
Chris@10
|
2517 T2V = ii[WS(rs, 22)];
|
Chris@10
|
2518 T2S = W[42];
|
Chris@10
|
2519 T2U = W[43];
|
Chris@10
|
2520 T2W = FMA(T2S, T2T, T2U * T2V);
|
Chris@10
|
2521 T7k = FNMS(T2U, T2T, T2S * T2V);
|
Chris@10
|
2522 }
|
Chris@10
|
2523 T2X = T2R + T2W;
|
Chris@10
|
2524 Tdk = T7j + T7k;
|
Chris@10
|
2525 T7i = T2R - T2W;
|
Chris@10
|
2526 T7l = T7j - T7k;
|
Chris@10
|
2527 }
|
Chris@10
|
2528 T2B = T2p + T2A;
|
Chris@10
|
2529 T2Y = T2M + T2X;
|
Chris@10
|
2530 Tfz = T2B - T2Y;
|
Chris@10
|
2531 TfA = Tdd + Tde;
|
Chris@10
|
2532 TfB = Tdj + Tdk;
|
Chris@10
|
2533 TfC = TfA - TfB;
|
Chris@10
|
2534 {
|
Chris@10
|
2535 E T7f, T7g, Tdi, Tdl;
|
Chris@10
|
2536 T7f = T7d - T7e;
|
Chris@10
|
2537 T7g = T2u - T2z;
|
Chris@10
|
2538 T7h = T7f + T7g;
|
Chris@10
|
2539 TaW = T7f - T7g;
|
Chris@10
|
2540 Tdi = T2p - T2A;
|
Chris@10
|
2541 Tdl = Tdj - Tdk;
|
Chris@10
|
2542 Tdm = Tdi - Tdl;
|
Chris@10
|
2543 TeM = Tdi + Tdl;
|
Chris@10
|
2544 }
|
Chris@10
|
2545 {
|
Chris@10
|
2546 E T7m, T7r, T7z, T7A;
|
Chris@10
|
2547 T7m = T7i - T7l;
|
Chris@10
|
2548 T7r = T7n + T7q;
|
Chris@10
|
2549 T7s = KP707106781 * (T7m - T7r);
|
Chris@10
|
2550 TaU = KP707106781 * (T7r + T7m);
|
Chris@10
|
2551 T7z = T7q - T7n;
|
Chris@10
|
2552 T7A = T7i + T7l;
|
Chris@10
|
2553 T7B = KP707106781 * (T7z - T7A);
|
Chris@10
|
2554 TaX = KP707106781 * (T7z + T7A);
|
Chris@10
|
2555 }
|
Chris@10
|
2556 {
|
Chris@10
|
2557 E Tdf, Tdg, T7u, T7x;
|
Chris@10
|
2558 Tdf = Tdd - Tde;
|
Chris@10
|
2559 Tdg = T2X - T2M;
|
Chris@10
|
2560 Tdh = Tdf - Tdg;
|
Chris@10
|
2561 TeL = Tdf + Tdg;
|
Chris@10
|
2562 T7u = T2j - T2o;
|
Chris@10
|
2563 T7x = T7v - T7w;
|
Chris@10
|
2564 T7y = T7u - T7x;
|
Chris@10
|
2565 TaT = T7u + T7x;
|
Chris@10
|
2566 }
|
Chris@10
|
2567 }
|
Chris@10
|
2568 {
|
Chris@10
|
2569 E T4D, T9e, T4I, T9f, T4J, Te8, T4O, T8A, T4T, T8B, T4U, Te9, T56, TdS, T8G;
|
Chris@10
|
2570 E T8H, T5h, TdT, T8J, T8M;
|
Chris@10
|
2571 {
|
Chris@10
|
2572 E T4A, T4C, T4z, T4B;
|
Chris@10
|
2573 T4A = ri[WS(rs, 63)];
|
Chris@10
|
2574 T4C = ii[WS(rs, 63)];
|
Chris@10
|
2575 T4z = W[124];
|
Chris@10
|
2576 T4B = W[125];
|
Chris@10
|
2577 T4D = FMA(T4z, T4A, T4B * T4C);
|
Chris@10
|
2578 T9e = FNMS(T4B, T4A, T4z * T4C);
|
Chris@10
|
2579 }
|
Chris@10
|
2580 {
|
Chris@10
|
2581 E T4F, T4H, T4E, T4G;
|
Chris@10
|
2582 T4F = ri[WS(rs, 31)];
|
Chris@10
|
2583 T4H = ii[WS(rs, 31)];
|
Chris@10
|
2584 T4E = W[60];
|
Chris@10
|
2585 T4G = W[61];
|
Chris@10
|
2586 T4I = FMA(T4E, T4F, T4G * T4H);
|
Chris@10
|
2587 T9f = FNMS(T4G, T4F, T4E * T4H);
|
Chris@10
|
2588 }
|
Chris@10
|
2589 T4J = T4D + T4I;
|
Chris@10
|
2590 Te8 = T9e + T9f;
|
Chris@10
|
2591 {
|
Chris@10
|
2592 E T4L, T4N, T4K, T4M;
|
Chris@10
|
2593 T4L = ri[WS(rs, 15)];
|
Chris@10
|
2594 T4N = ii[WS(rs, 15)];
|
Chris@10
|
2595 T4K = W[28];
|
Chris@10
|
2596 T4M = W[29];
|
Chris@10
|
2597 T4O = FMA(T4K, T4L, T4M * T4N);
|
Chris@10
|
2598 T8A = FNMS(T4M, T4L, T4K * T4N);
|
Chris@10
|
2599 }
|
Chris@10
|
2600 {
|
Chris@10
|
2601 E T4Q, T4S, T4P, T4R;
|
Chris@10
|
2602 T4Q = ri[WS(rs, 47)];
|
Chris@10
|
2603 T4S = ii[WS(rs, 47)];
|
Chris@10
|
2604 T4P = W[92];
|
Chris@10
|
2605 T4R = W[93];
|
Chris@10
|
2606 T4T = FMA(T4P, T4Q, T4R * T4S);
|
Chris@10
|
2607 T8B = FNMS(T4R, T4Q, T4P * T4S);
|
Chris@10
|
2608 }
|
Chris@10
|
2609 T4U = T4O + T4T;
|
Chris@10
|
2610 Te9 = T8A + T8B;
|
Chris@10
|
2611 {
|
Chris@10
|
2612 E T50, T8E, T55, T8F;
|
Chris@10
|
2613 {
|
Chris@10
|
2614 E T4X, T4Z, T4W, T4Y;
|
Chris@10
|
2615 T4X = ri[WS(rs, 7)];
|
Chris@10
|
2616 T4Z = ii[WS(rs, 7)];
|
Chris@10
|
2617 T4W = W[12];
|
Chris@10
|
2618 T4Y = W[13];
|
Chris@10
|
2619 T50 = FMA(T4W, T4X, T4Y * T4Z);
|
Chris@10
|
2620 T8E = FNMS(T4Y, T4X, T4W * T4Z);
|
Chris@10
|
2621 }
|
Chris@10
|
2622 {
|
Chris@10
|
2623 E T52, T54, T51, T53;
|
Chris@10
|
2624 T52 = ri[WS(rs, 39)];
|
Chris@10
|
2625 T54 = ii[WS(rs, 39)];
|
Chris@10
|
2626 T51 = W[76];
|
Chris@10
|
2627 T53 = W[77];
|
Chris@10
|
2628 T55 = FMA(T51, T52, T53 * T54);
|
Chris@10
|
2629 T8F = FNMS(T53, T52, T51 * T54);
|
Chris@10
|
2630 }
|
Chris@10
|
2631 T56 = T50 + T55;
|
Chris@10
|
2632 TdS = T8E + T8F;
|
Chris@10
|
2633 T8G = T8E - T8F;
|
Chris@10
|
2634 T8H = T50 - T55;
|
Chris@10
|
2635 }
|
Chris@10
|
2636 {
|
Chris@10
|
2637 E T5b, T8K, T5g, T8L;
|
Chris@10
|
2638 {
|
Chris@10
|
2639 E T58, T5a, T57, T59;
|
Chris@10
|
2640 T58 = ri[WS(rs, 55)];
|
Chris@10
|
2641 T5a = ii[WS(rs, 55)];
|
Chris@10
|
2642 T57 = W[108];
|
Chris@10
|
2643 T59 = W[109];
|
Chris@10
|
2644 T5b = FMA(T57, T58, T59 * T5a);
|
Chris@10
|
2645 T8K = FNMS(T59, T58, T57 * T5a);
|
Chris@10
|
2646 }
|
Chris@10
|
2647 {
|
Chris@10
|
2648 E T5d, T5f, T5c, T5e;
|
Chris@10
|
2649 T5d = ri[WS(rs, 23)];
|
Chris@10
|
2650 T5f = ii[WS(rs, 23)];
|
Chris@10
|
2651 T5c = W[44];
|
Chris@10
|
2652 T5e = W[45];
|
Chris@10
|
2653 T5g = FMA(T5c, T5d, T5e * T5f);
|
Chris@10
|
2654 T8L = FNMS(T5e, T5d, T5c * T5f);
|
Chris@10
|
2655 }
|
Chris@10
|
2656 T5h = T5b + T5g;
|
Chris@10
|
2657 TdT = T8K + T8L;
|
Chris@10
|
2658 T8J = T5b - T5g;
|
Chris@10
|
2659 T8M = T8K - T8L;
|
Chris@10
|
2660 }
|
Chris@10
|
2661 {
|
Chris@10
|
2662 E T4V, T5i, Tea, Teb;
|
Chris@10
|
2663 T4V = T4J + T4U;
|
Chris@10
|
2664 T5i = T56 + T5h;
|
Chris@10
|
2665 T5j = T4V + T5i;
|
Chris@10
|
2666 TfR = T4V - T5i;
|
Chris@10
|
2667 Tea = Te8 - Te9;
|
Chris@10
|
2668 Teb = T5h - T56;
|
Chris@10
|
2669 Tec = Tea - Teb;
|
Chris@10
|
2670 Tf0 = Tea + Teb;
|
Chris@10
|
2671 }
|
Chris@10
|
2672 {
|
Chris@10
|
2673 E TfW, TfX, T8z, T8C;
|
Chris@10
|
2674 TfW = Te8 + Te9;
|
Chris@10
|
2675 TfX = TdS + TdT;
|
Chris@10
|
2676 TfY = TfW - TfX;
|
Chris@10
|
2677 Tgy = TfW + TfX;
|
Chris@10
|
2678 T8z = T4D - T4I;
|
Chris@10
|
2679 T8C = T8A - T8B;
|
Chris@10
|
2680 T8D = T8z - T8C;
|
Chris@10
|
2681 Tbl = T8z + T8C;
|
Chris@10
|
2682 }
|
Chris@10
|
2683 {
|
Chris@10
|
2684 E T8I, T8N, T9j, T9k;
|
Chris@10
|
2685 T8I = T8G - T8H;
|
Chris@10
|
2686 T8N = T8J + T8M;
|
Chris@10
|
2687 T8O = KP707106781 * (T8I - T8N);
|
Chris@10
|
2688 Tbx = KP707106781 * (T8I + T8N);
|
Chris@10
|
2689 T9j = T8J - T8M;
|
Chris@10
|
2690 T9k = T8H + T8G;
|
Chris@10
|
2691 T9l = KP707106781 * (T9j - T9k);
|
Chris@10
|
2692 Tbm = KP707106781 * (T9k + T9j);
|
Chris@10
|
2693 }
|
Chris@10
|
2694 {
|
Chris@10
|
2695 E TdR, TdU, T9g, T9h;
|
Chris@10
|
2696 TdR = T4J - T4U;
|
Chris@10
|
2697 TdU = TdS - TdT;
|
Chris@10
|
2698 TdV = TdR - TdU;
|
Chris@10
|
2699 TeX = TdR + TdU;
|
Chris@10
|
2700 T9g = T9e - T9f;
|
Chris@10
|
2701 T9h = T4O - T4T;
|
Chris@10
|
2702 T9i = T9g + T9h;
|
Chris@10
|
2703 Tbw = T9g - T9h;
|
Chris@10
|
2704 }
|
Chris@10
|
2705 }
|
Chris@10
|
2706 {
|
Chris@10
|
2707 E T36, T7G, T3b, T7H, T3c, Tdq, T3h, T8m, T3m, T8n, T3n, Tdr, T3z, TdI, T7Q;
|
Chris@10
|
2708 E T7T, T3K, TdJ, T7L, T7O;
|
Chris@10
|
2709 {
|
Chris@10
|
2710 E T33, T35, T32, T34;
|
Chris@10
|
2711 T33 = ri[WS(rs, 1)];
|
Chris@10
|
2712 T35 = ii[WS(rs, 1)];
|
Chris@10
|
2713 T32 = W[0];
|
Chris@10
|
2714 T34 = W[1];
|
Chris@10
|
2715 T36 = FMA(T32, T33, T34 * T35);
|
Chris@10
|
2716 T7G = FNMS(T34, T33, T32 * T35);
|
Chris@10
|
2717 }
|
Chris@10
|
2718 {
|
Chris@10
|
2719 E T38, T3a, T37, T39;
|
Chris@10
|
2720 T38 = ri[WS(rs, 33)];
|
Chris@10
|
2721 T3a = ii[WS(rs, 33)];
|
Chris@10
|
2722 T37 = W[64];
|
Chris@10
|
2723 T39 = W[65];
|
Chris@10
|
2724 T3b = FMA(T37, T38, T39 * T3a);
|
Chris@10
|
2725 T7H = FNMS(T39, T38, T37 * T3a);
|
Chris@10
|
2726 }
|
Chris@10
|
2727 T3c = T36 + T3b;
|
Chris@10
|
2728 Tdq = T7G + T7H;
|
Chris@10
|
2729 {
|
Chris@10
|
2730 E T3e, T3g, T3d, T3f;
|
Chris@10
|
2731 T3e = ri[WS(rs, 17)];
|
Chris@10
|
2732 T3g = ii[WS(rs, 17)];
|
Chris@10
|
2733 T3d = W[32];
|
Chris@10
|
2734 T3f = W[33];
|
Chris@10
|
2735 T3h = FMA(T3d, T3e, T3f * T3g);
|
Chris@10
|
2736 T8m = FNMS(T3f, T3e, T3d * T3g);
|
Chris@10
|
2737 }
|
Chris@10
|
2738 {
|
Chris@10
|
2739 E T3j, T3l, T3i, T3k;
|
Chris@10
|
2740 T3j = ri[WS(rs, 49)];
|
Chris@10
|
2741 T3l = ii[WS(rs, 49)];
|
Chris@10
|
2742 T3i = W[96];
|
Chris@10
|
2743 T3k = W[97];
|
Chris@10
|
2744 T3m = FMA(T3i, T3j, T3k * T3l);
|
Chris@10
|
2745 T8n = FNMS(T3k, T3j, T3i * T3l);
|
Chris@10
|
2746 }
|
Chris@10
|
2747 T3n = T3h + T3m;
|
Chris@10
|
2748 Tdr = T8m + T8n;
|
Chris@10
|
2749 {
|
Chris@10
|
2750 E T3t, T7R, T3y, T7S;
|
Chris@10
|
2751 {
|
Chris@10
|
2752 E T3q, T3s, T3p, T3r;
|
Chris@10
|
2753 T3q = ri[WS(rs, 9)];
|
Chris@10
|
2754 T3s = ii[WS(rs, 9)];
|
Chris@10
|
2755 T3p = W[16];
|
Chris@10
|
2756 T3r = W[17];
|
Chris@10
|
2757 T3t = FMA(T3p, T3q, T3r * T3s);
|
Chris@10
|
2758 T7R = FNMS(T3r, T3q, T3p * T3s);
|
Chris@10
|
2759 }
|
Chris@10
|
2760 {
|
Chris@10
|
2761 E T3v, T3x, T3u, T3w;
|
Chris@10
|
2762 T3v = ri[WS(rs, 41)];
|
Chris@10
|
2763 T3x = ii[WS(rs, 41)];
|
Chris@10
|
2764 T3u = W[80];
|
Chris@10
|
2765 T3w = W[81];
|
Chris@10
|
2766 T3y = FMA(T3u, T3v, T3w * T3x);
|
Chris@10
|
2767 T7S = FNMS(T3w, T3v, T3u * T3x);
|
Chris@10
|
2768 }
|
Chris@10
|
2769 T3z = T3t + T3y;
|
Chris@10
|
2770 TdI = T7R + T7S;
|
Chris@10
|
2771 T7Q = T3t - T3y;
|
Chris@10
|
2772 T7T = T7R - T7S;
|
Chris@10
|
2773 }
|
Chris@10
|
2774 {
|
Chris@10
|
2775 E T3E, T7M, T3J, T7N;
|
Chris@10
|
2776 {
|
Chris@10
|
2777 E T3B, T3D, T3A, T3C;
|
Chris@10
|
2778 T3B = ri[WS(rs, 57)];
|
Chris@10
|
2779 T3D = ii[WS(rs, 57)];
|
Chris@10
|
2780 T3A = W[112];
|
Chris@10
|
2781 T3C = W[113];
|
Chris@10
|
2782 T3E = FMA(T3A, T3B, T3C * T3D);
|
Chris@10
|
2783 T7M = FNMS(T3C, T3B, T3A * T3D);
|
Chris@10
|
2784 }
|
Chris@10
|
2785 {
|
Chris@10
|
2786 E T3G, T3I, T3F, T3H;
|
Chris@10
|
2787 T3G = ri[WS(rs, 25)];
|
Chris@10
|
2788 T3I = ii[WS(rs, 25)];
|
Chris@10
|
2789 T3F = W[48];
|
Chris@10
|
2790 T3H = W[49];
|
Chris@10
|
2791 T3J = FMA(T3F, T3G, T3H * T3I);
|
Chris@10
|
2792 T7N = FNMS(T3H, T3G, T3F * T3I);
|
Chris@10
|
2793 }
|
Chris@10
|
2794 T3K = T3E + T3J;
|
Chris@10
|
2795 TdJ = T7M + T7N;
|
Chris@10
|
2796 T7L = T3E - T3J;
|
Chris@10
|
2797 T7O = T7M - T7N;
|
Chris@10
|
2798 }
|
Chris@10
|
2799 {
|
Chris@10
|
2800 E T3o, T3L, TdH, TdK;
|
Chris@10
|
2801 T3o = T3c + T3n;
|
Chris@10
|
2802 T3L = T3z + T3K;
|
Chris@10
|
2803 T3M = T3o + T3L;
|
Chris@10
|
2804 TfL = T3o - T3L;
|
Chris@10
|
2805 TdH = T3c - T3n;
|
Chris@10
|
2806 TdK = TdI - TdJ;
|
Chris@10
|
2807 TdL = TdH - TdK;
|
Chris@10
|
2808 TeQ = TdH + TdK;
|
Chris@10
|
2809 }
|
Chris@10
|
2810 {
|
Chris@10
|
2811 E TfG, TfH, T7I, T7J;
|
Chris@10
|
2812 TfG = Tdq + Tdr;
|
Chris@10
|
2813 TfH = TdI + TdJ;
|
Chris@10
|
2814 TfI = TfG - TfH;
|
Chris@10
|
2815 Tgt = TfG + TfH;
|
Chris@10
|
2816 T7I = T7G - T7H;
|
Chris@10
|
2817 T7J = T3h - T3m;
|
Chris@10
|
2818 T7K = T7I + T7J;
|
Chris@10
|
2819 Tb2 = T7I - T7J;
|
Chris@10
|
2820 }
|
Chris@10
|
2821 {
|
Chris@10
|
2822 E T7P, T7U, T8q, T8r;
|
Chris@10
|
2823 T7P = T7L - T7O;
|
Chris@10
|
2824 T7U = T7Q + T7T;
|
Chris@10
|
2825 T7V = KP707106781 * (T7P - T7U);
|
Chris@10
|
2826 Tbe = KP707106781 * (T7U + T7P);
|
Chris@10
|
2827 T8q = T7T - T7Q;
|
Chris@10
|
2828 T8r = T7L + T7O;
|
Chris@10
|
2829 T8s = KP707106781 * (T8q - T8r);
|
Chris@10
|
2830 Tb3 = KP707106781 * (T8q + T8r);
|
Chris@10
|
2831 }
|
Chris@10
|
2832 {
|
Chris@10
|
2833 E Tds, Tdt, T8l, T8o;
|
Chris@10
|
2834 Tds = Tdq - Tdr;
|
Chris@10
|
2835 Tdt = T3K - T3z;
|
Chris@10
|
2836 Tdu = Tds - Tdt;
|
Chris@10
|
2837 TeT = Tds + Tdt;
|
Chris@10
|
2838 T8l = T36 - T3b;
|
Chris@10
|
2839 T8o = T8m - T8n;
|
Chris@10
|
2840 T8p = T8l - T8o;
|
Chris@10
|
2841 Tbd = T8l + T8o;
|
Chris@10
|
2842 }
|
Chris@10
|
2843 }
|
Chris@10
|
2844 {
|
Chris@10
|
2845 E T3X, TdB, T8a, T8d, T4v, Tdx, T80, T85, T48, TdC, T8b, T8g, T4k, Tdw, T7X;
|
Chris@10
|
2846 E T84;
|
Chris@10
|
2847 {
|
Chris@10
|
2848 E T3R, T88, T3W, T89;
|
Chris@10
|
2849 {
|
Chris@10
|
2850 E T3O, T3Q, T3N, T3P;
|
Chris@10
|
2851 T3O = ri[WS(rs, 5)];
|
Chris@10
|
2852 T3Q = ii[WS(rs, 5)];
|
Chris@10
|
2853 T3N = W[8];
|
Chris@10
|
2854 T3P = W[9];
|
Chris@10
|
2855 T3R = FMA(T3N, T3O, T3P * T3Q);
|
Chris@10
|
2856 T88 = FNMS(T3P, T3O, T3N * T3Q);
|
Chris@10
|
2857 }
|
Chris@10
|
2858 {
|
Chris@10
|
2859 E T3T, T3V, T3S, T3U;
|
Chris@10
|
2860 T3T = ri[WS(rs, 37)];
|
Chris@10
|
2861 T3V = ii[WS(rs, 37)];
|
Chris@10
|
2862 T3S = W[72];
|
Chris@10
|
2863 T3U = W[73];
|
Chris@10
|
2864 T3W = FMA(T3S, T3T, T3U * T3V);
|
Chris@10
|
2865 T89 = FNMS(T3U, T3T, T3S * T3V);
|
Chris@10
|
2866 }
|
Chris@10
|
2867 T3X = T3R + T3W;
|
Chris@10
|
2868 TdB = T88 + T89;
|
Chris@10
|
2869 T8a = T88 - T89;
|
Chris@10
|
2870 T8d = T3R - T3W;
|
Chris@10
|
2871 }
|
Chris@10
|
2872 {
|
Chris@10
|
2873 E T4p, T7Y, T4u, T7Z;
|
Chris@10
|
2874 {
|
Chris@10
|
2875 E T4m, T4o, T4l, T4n;
|
Chris@10
|
2876 T4m = ri[WS(rs, 13)];
|
Chris@10
|
2877 T4o = ii[WS(rs, 13)];
|
Chris@10
|
2878 T4l = W[24];
|
Chris@10
|
2879 T4n = W[25];
|
Chris@10
|
2880 T4p = FMA(T4l, T4m, T4n * T4o);
|
Chris@10
|
2881 T7Y = FNMS(T4n, T4m, T4l * T4o);
|
Chris@10
|
2882 }
|
Chris@10
|
2883 {
|
Chris@10
|
2884 E T4r, T4t, T4q, T4s;
|
Chris@10
|
2885 T4r = ri[WS(rs, 45)];
|
Chris@10
|
2886 T4t = ii[WS(rs, 45)];
|
Chris@10
|
2887 T4q = W[88];
|
Chris@10
|
2888 T4s = W[89];
|
Chris@10
|
2889 T4u = FMA(T4q, T4r, T4s * T4t);
|
Chris@10
|
2890 T7Z = FNMS(T4s, T4r, T4q * T4t);
|
Chris@10
|
2891 }
|
Chris@10
|
2892 T4v = T4p + T4u;
|
Chris@10
|
2893 Tdx = T7Y + T7Z;
|
Chris@10
|
2894 T80 = T7Y - T7Z;
|
Chris@10
|
2895 T85 = T4p - T4u;
|
Chris@10
|
2896 }
|
Chris@10
|
2897 {
|
Chris@10
|
2898 E T42, T8e, T47, T8f;
|
Chris@10
|
2899 {
|
Chris@10
|
2900 E T3Z, T41, T3Y, T40;
|
Chris@10
|
2901 T3Z = ri[WS(rs, 21)];
|
Chris@10
|
2902 T41 = ii[WS(rs, 21)];
|
Chris@10
|
2903 T3Y = W[40];
|
Chris@10
|
2904 T40 = W[41];
|
Chris@10
|
2905 T42 = FMA(T3Y, T3Z, T40 * T41);
|
Chris@10
|
2906 T8e = FNMS(T40, T3Z, T3Y * T41);
|
Chris@10
|
2907 }
|
Chris@10
|
2908 {
|
Chris@10
|
2909 E T44, T46, T43, T45;
|
Chris@10
|
2910 T44 = ri[WS(rs, 53)];
|
Chris@10
|
2911 T46 = ii[WS(rs, 53)];
|
Chris@10
|
2912 T43 = W[104];
|
Chris@10
|
2913 T45 = W[105];
|
Chris@10
|
2914 T47 = FMA(T43, T44, T45 * T46);
|
Chris@10
|
2915 T8f = FNMS(T45, T44, T43 * T46);
|
Chris@10
|
2916 }
|
Chris@10
|
2917 T48 = T42 + T47;
|
Chris@10
|
2918 TdC = T8e + T8f;
|
Chris@10
|
2919 T8b = T42 - T47;
|
Chris@10
|
2920 T8g = T8e - T8f;
|
Chris@10
|
2921 }
|
Chris@10
|
2922 {
|
Chris@10
|
2923 E T4e, T82, T4j, T83;
|
Chris@10
|
2924 {
|
Chris@10
|
2925 E T4b, T4d, T4a, T4c;
|
Chris@10
|
2926 T4b = ri[WS(rs, 61)];
|
Chris@10
|
2927 T4d = ii[WS(rs, 61)];
|
Chris@10
|
2928 T4a = W[120];
|
Chris@10
|
2929 T4c = W[121];
|
Chris@10
|
2930 T4e = FMA(T4a, T4b, T4c * T4d);
|
Chris@10
|
2931 T82 = FNMS(T4c, T4b, T4a * T4d);
|
Chris@10
|
2932 }
|
Chris@10
|
2933 {
|
Chris@10
|
2934 E T4g, T4i, T4f, T4h;
|
Chris@10
|
2935 T4g = ri[WS(rs, 29)];
|
Chris@10
|
2936 T4i = ii[WS(rs, 29)];
|
Chris@10
|
2937 T4f = W[56];
|
Chris@10
|
2938 T4h = W[57];
|
Chris@10
|
2939 T4j = FMA(T4f, T4g, T4h * T4i);
|
Chris@10
|
2940 T83 = FNMS(T4h, T4g, T4f * T4i);
|
Chris@10
|
2941 }
|
Chris@10
|
2942 T4k = T4e + T4j;
|
Chris@10
|
2943 Tdw = T82 + T83;
|
Chris@10
|
2944 T7X = T4e - T4j;
|
Chris@10
|
2945 T84 = T82 - T83;
|
Chris@10
|
2946 }
|
Chris@10
|
2947 {
|
Chris@10
|
2948 E T49, T4w, TdA, TdD;
|
Chris@10
|
2949 T49 = T3X + T48;
|
Chris@10
|
2950 T4w = T4k + T4v;
|
Chris@10
|
2951 T4x = T49 + T4w;
|
Chris@10
|
2952 TfJ = T4w - T49;
|
Chris@10
|
2953 TdA = T3X - T48;
|
Chris@10
|
2954 TdD = TdB - TdC;
|
Chris@10
|
2955 TdE = TdA + TdD;
|
Chris@10
|
2956 TdM = TdD - TdA;
|
Chris@10
|
2957 }
|
Chris@10
|
2958 {
|
Chris@10
|
2959 E TfM, TfN, T81, T86;
|
Chris@10
|
2960 TfM = TdB + TdC;
|
Chris@10
|
2961 TfN = Tdw + Tdx;
|
Chris@10
|
2962 TfO = TfM - TfN;
|
Chris@10
|
2963 Tgu = TfM + TfN;
|
Chris@10
|
2964 T81 = T7X - T80;
|
Chris@10
|
2965 T86 = T84 + T85;
|
Chris@10
|
2966 T87 = FNMS(KP923879532, T86, KP382683432 * T81);
|
Chris@10
|
2967 T8v = FMA(KP382683432, T86, KP923879532 * T81);
|
Chris@10
|
2968 }
|
Chris@10
|
2969 {
|
Chris@10
|
2970 E T8c, T8h, Tb8, Tb9;
|
Chris@10
|
2971 T8c = T8a + T8b;
|
Chris@10
|
2972 T8h = T8d - T8g;
|
Chris@10
|
2973 T8i = FMA(KP923879532, T8c, KP382683432 * T8h);
|
Chris@10
|
2974 T8u = FNMS(KP923879532, T8h, KP382683432 * T8c);
|
Chris@10
|
2975 Tb8 = T8a - T8b;
|
Chris@10
|
2976 Tb9 = T8d + T8g;
|
Chris@10
|
2977 Tba = FMA(KP382683432, Tb8, KP923879532 * Tb9);
|
Chris@10
|
2978 Tbg = FNMS(KP382683432, Tb9, KP923879532 * Tb8);
|
Chris@10
|
2979 }
|
Chris@10
|
2980 {
|
Chris@10
|
2981 E Tdv, Tdy, Tb5, Tb6;
|
Chris@10
|
2982 Tdv = T4k - T4v;
|
Chris@10
|
2983 Tdy = Tdw - Tdx;
|
Chris@10
|
2984 Tdz = Tdv - Tdy;
|
Chris@10
|
2985 TdN = Tdv + Tdy;
|
Chris@10
|
2986 Tb5 = T7X + T80;
|
Chris@10
|
2987 Tb6 = T84 - T85;
|
Chris@10
|
2988 Tb7 = FNMS(KP382683432, Tb6, KP923879532 * Tb5);
|
Chris@10
|
2989 Tbh = FMA(KP923879532, Tb6, KP382683432 * Tb5);
|
Chris@10
|
2990 }
|
Chris@10
|
2991 }
|
Chris@10
|
2992 {
|
Chris@10
|
2993 E T5u, TdW, T8S, T8V, T62, Te3, T94, T99, T5F, TdX, T8T, T8Y, T5R, Te2, T93;
|
Chris@10
|
2994 E T96;
|
Chris@10
|
2995 {
|
Chris@10
|
2996 E T5o, T8Q, T5t, T8R;
|
Chris@10
|
2997 {
|
Chris@10
|
2998 E T5l, T5n, T5k, T5m;
|
Chris@10
|
2999 T5l = ri[WS(rs, 3)];
|
Chris@10
|
3000 T5n = ii[WS(rs, 3)];
|
Chris@10
|
3001 T5k = W[4];
|
Chris@10
|
3002 T5m = W[5];
|
Chris@10
|
3003 T5o = FMA(T5k, T5l, T5m * T5n);
|
Chris@10
|
3004 T8Q = FNMS(T5m, T5l, T5k * T5n);
|
Chris@10
|
3005 }
|
Chris@10
|
3006 {
|
Chris@10
|
3007 E T5q, T5s, T5p, T5r;
|
Chris@10
|
3008 T5q = ri[WS(rs, 35)];
|
Chris@10
|
3009 T5s = ii[WS(rs, 35)];
|
Chris@10
|
3010 T5p = W[68];
|
Chris@10
|
3011 T5r = W[69];
|
Chris@10
|
3012 T5t = FMA(T5p, T5q, T5r * T5s);
|
Chris@10
|
3013 T8R = FNMS(T5r, T5q, T5p * T5s);
|
Chris@10
|
3014 }
|
Chris@10
|
3015 T5u = T5o + T5t;
|
Chris@10
|
3016 TdW = T8Q + T8R;
|
Chris@10
|
3017 T8S = T8Q - T8R;
|
Chris@10
|
3018 T8V = T5o - T5t;
|
Chris@10
|
3019 }
|
Chris@10
|
3020 {
|
Chris@10
|
3021 E T5W, T97, T61, T98;
|
Chris@10
|
3022 {
|
Chris@10
|
3023 E T5T, T5V, T5S, T5U;
|
Chris@10
|
3024 T5T = ri[WS(rs, 11)];
|
Chris@10
|
3025 T5V = ii[WS(rs, 11)];
|
Chris@10
|
3026 T5S = W[20];
|
Chris@10
|
3027 T5U = W[21];
|
Chris@10
|
3028 T5W = FMA(T5S, T5T, T5U * T5V);
|
Chris@10
|
3029 T97 = FNMS(T5U, T5T, T5S * T5V);
|
Chris@10
|
3030 }
|
Chris@10
|
3031 {
|
Chris@10
|
3032 E T5Y, T60, T5X, T5Z;
|
Chris@10
|
3033 T5Y = ri[WS(rs, 43)];
|
Chris@10
|
3034 T60 = ii[WS(rs, 43)];
|
Chris@10
|
3035 T5X = W[84];
|
Chris@10
|
3036 T5Z = W[85];
|
Chris@10
|
3037 T61 = FMA(T5X, T5Y, T5Z * T60);
|
Chris@10
|
3038 T98 = FNMS(T5Z, T5Y, T5X * T60);
|
Chris@10
|
3039 }
|
Chris@10
|
3040 T62 = T5W + T61;
|
Chris@10
|
3041 Te3 = T97 + T98;
|
Chris@10
|
3042 T94 = T5W - T61;
|
Chris@10
|
3043 T99 = T97 - T98;
|
Chris@10
|
3044 }
|
Chris@10
|
3045 {
|
Chris@10
|
3046 E T5z, T8W, T5E, T8X;
|
Chris@10
|
3047 {
|
Chris@10
|
3048 E T5w, T5y, T5v, T5x;
|
Chris@10
|
3049 T5w = ri[WS(rs, 19)];
|
Chris@10
|
3050 T5y = ii[WS(rs, 19)];
|
Chris@10
|
3051 T5v = W[36];
|
Chris@10
|
3052 T5x = W[37];
|
Chris@10
|
3053 T5z = FMA(T5v, T5w, T5x * T5y);
|
Chris@10
|
3054 T8W = FNMS(T5x, T5w, T5v * T5y);
|
Chris@10
|
3055 }
|
Chris@10
|
3056 {
|
Chris@10
|
3057 E T5B, T5D, T5A, T5C;
|
Chris@10
|
3058 T5B = ri[WS(rs, 51)];
|
Chris@10
|
3059 T5D = ii[WS(rs, 51)];
|
Chris@10
|
3060 T5A = W[100];
|
Chris@10
|
3061 T5C = W[101];
|
Chris@10
|
3062 T5E = FMA(T5A, T5B, T5C * T5D);
|
Chris@10
|
3063 T8X = FNMS(T5C, T5B, T5A * T5D);
|
Chris@10
|
3064 }
|
Chris@10
|
3065 T5F = T5z + T5E;
|
Chris@10
|
3066 TdX = T8W + T8X;
|
Chris@10
|
3067 T8T = T5z - T5E;
|
Chris@10
|
3068 T8Y = T8W - T8X;
|
Chris@10
|
3069 }
|
Chris@10
|
3070 {
|
Chris@10
|
3071 E T5L, T91, T5Q, T92;
|
Chris@10
|
3072 {
|
Chris@10
|
3073 E T5I, T5K, T5H, T5J;
|
Chris@10
|
3074 T5I = ri[WS(rs, 59)];
|
Chris@10
|
3075 T5K = ii[WS(rs, 59)];
|
Chris@10
|
3076 T5H = W[116];
|
Chris@10
|
3077 T5J = W[117];
|
Chris@10
|
3078 T5L = FMA(T5H, T5I, T5J * T5K);
|
Chris@10
|
3079 T91 = FNMS(T5J, T5I, T5H * T5K);
|
Chris@10
|
3080 }
|
Chris@10
|
3081 {
|
Chris@10
|
3082 E T5N, T5P, T5M, T5O;
|
Chris@10
|
3083 T5N = ri[WS(rs, 27)];
|
Chris@10
|
3084 T5P = ii[WS(rs, 27)];
|
Chris@10
|
3085 T5M = W[52];
|
Chris@10
|
3086 T5O = W[53];
|
Chris@10
|
3087 T5Q = FMA(T5M, T5N, T5O * T5P);
|
Chris@10
|
3088 T92 = FNMS(T5O, T5N, T5M * T5P);
|
Chris@10
|
3089 }
|
Chris@10
|
3090 T5R = T5L + T5Q;
|
Chris@10
|
3091 Te2 = T91 + T92;
|
Chris@10
|
3092 T93 = T91 - T92;
|
Chris@10
|
3093 T96 = T5L - T5Q;
|
Chris@10
|
3094 }
|
Chris@10
|
3095 {
|
Chris@10
|
3096 E T5G, T63, Te1, Te4;
|
Chris@10
|
3097 T5G = T5u + T5F;
|
Chris@10
|
3098 T63 = T5R + T62;
|
Chris@10
|
3099 T64 = T5G + T63;
|
Chris@10
|
3100 TfZ = T63 - T5G;
|
Chris@10
|
3101 Te1 = T5R - T62;
|
Chris@10
|
3102 Te4 = Te2 - Te3;
|
Chris@10
|
3103 Te5 = Te1 + Te4;
|
Chris@10
|
3104 Ted = Te1 - Te4;
|
Chris@10
|
3105 }
|
Chris@10
|
3106 {
|
Chris@10
|
3107 E TfS, TfT, T8U, T8Z;
|
Chris@10
|
3108 TfS = TdW + TdX;
|
Chris@10
|
3109 TfT = Te2 + Te3;
|
Chris@10
|
3110 TfU = TfS - TfT;
|
Chris@10
|
3111 Tgz = TfS + TfT;
|
Chris@10
|
3112 T8U = T8S + T8T;
|
Chris@10
|
3113 T8Z = T8V - T8Y;
|
Chris@10
|
3114 T90 = FNMS(KP923879532, T8Z, KP382683432 * T8U);
|
Chris@10
|
3115 T9o = FMA(KP923879532, T8U, KP382683432 * T8Z);
|
Chris@10
|
3116 }
|
Chris@10
|
3117 {
|
Chris@10
|
3118 E T95, T9a, Tbr, Tbs;
|
Chris@10
|
3119 T95 = T93 + T94;
|
Chris@10
|
3120 T9a = T96 - T99;
|
Chris@10
|
3121 T9b = FMA(KP382683432, T95, KP923879532 * T9a);
|
Chris@10
|
3122 T9n = FNMS(KP923879532, T95, KP382683432 * T9a);
|
Chris@10
|
3123 Tbr = T93 - T94;
|
Chris@10
|
3124 Tbs = T96 + T99;
|
Chris@10
|
3125 Tbt = FMA(KP923879532, Tbr, KP382683432 * Tbs);
|
Chris@10
|
3126 Tbz = FNMS(KP382683432, Tbr, KP923879532 * Tbs);
|
Chris@10
|
3127 }
|
Chris@10
|
3128 {
|
Chris@10
|
3129 E TdY, TdZ, Tbo, Tbp;
|
Chris@10
|
3130 TdY = TdW - TdX;
|
Chris@10
|
3131 TdZ = T5u - T5F;
|
Chris@10
|
3132 Te0 = TdY - TdZ;
|
Chris@10
|
3133 Tee = TdZ + TdY;
|
Chris@10
|
3134 Tbo = T8S - T8T;
|
Chris@10
|
3135 Tbp = T8V + T8Y;
|
Chris@10
|
3136 Tbq = FNMS(KP382683432, Tbp, KP923879532 * Tbo);
|
Chris@10
|
3137 TbA = FMA(KP382683432, Tbo, KP923879532 * Tbp);
|
Chris@10
|
3138 }
|
Chris@10
|
3139 }
|
Chris@10
|
3140 {
|
Chris@10
|
3141 E T1t, Tgn, TgK, TgL, TgV, Th1, T30, Th0, T66, TgX, Tgw, TgE, TgB, TgF, Tgq;
|
Chris@10
|
3142 E TgM;
|
Chris@10
|
3143 {
|
Chris@10
|
3144 E TH, T1s, TgI, TgJ;
|
Chris@10
|
3145 TH = Tj + TG;
|
Chris@10
|
3146 T1s = T14 + T1r;
|
Chris@10
|
3147 T1t = TH + T1s;
|
Chris@10
|
3148 Tgn = TH - T1s;
|
Chris@10
|
3149 TgI = Tgt + Tgu;
|
Chris@10
|
3150 TgJ = Tgy + Tgz;
|
Chris@10
|
3151 TgK = TgI - TgJ;
|
Chris@10
|
3152 TgL = TgI + TgJ;
|
Chris@10
|
3153 }
|
Chris@10
|
3154 {
|
Chris@10
|
3155 E TgN, TgU, T2e, T2Z;
|
Chris@10
|
3156 TgN = Tfq + Tfr;
|
Chris@10
|
3157 TgU = TgO + TgT;
|
Chris@10
|
3158 TgV = TgN + TgU;
|
Chris@10
|
3159 Th1 = TgU - TgN;
|
Chris@10
|
3160 T2e = T1Q + T2d;
|
Chris@10
|
3161 T2Z = T2B + T2Y;
|
Chris@10
|
3162 T30 = T2e + T2Z;
|
Chris@10
|
3163 Th0 = T2Z - T2e;
|
Chris@10
|
3164 }
|
Chris@10
|
3165 {
|
Chris@10
|
3166 E T4y, T65, Tgs, Tgv;
|
Chris@10
|
3167 T4y = T3M + T4x;
|
Chris@10
|
3168 T65 = T5j + T64;
|
Chris@10
|
3169 T66 = T4y + T65;
|
Chris@10
|
3170 TgX = T65 - T4y;
|
Chris@10
|
3171 Tgs = T3M - T4x;
|
Chris@10
|
3172 Tgv = Tgt - Tgu;
|
Chris@10
|
3173 Tgw = Tgs + Tgv;
|
Chris@10
|
3174 TgE = Tgv - Tgs;
|
Chris@10
|
3175 }
|
Chris@10
|
3176 {
|
Chris@10
|
3177 E Tgx, TgA, Tgo, Tgp;
|
Chris@10
|
3178 Tgx = T5j - T64;
|
Chris@10
|
3179 TgA = Tgy - Tgz;
|
Chris@10
|
3180 TgB = Tgx - TgA;
|
Chris@10
|
3181 TgF = Tgx + TgA;
|
Chris@10
|
3182 Tgo = Tfu + Tfv;
|
Chris@10
|
3183 Tgp = TfA + TfB;
|
Chris@10
|
3184 Tgq = Tgo - Tgp;
|
Chris@10
|
3185 TgM = Tgo + Tgp;
|
Chris@10
|
3186 }
|
Chris@10
|
3187 {
|
Chris@10
|
3188 E T31, TgW, TgH, TgY;
|
Chris@10
|
3189 T31 = T1t + T30;
|
Chris@10
|
3190 ri[WS(rs, 32)] = T31 - T66;
|
Chris@10
|
3191 ri[0] = T31 + T66;
|
Chris@10
|
3192 TgW = TgM + TgV;
|
Chris@10
|
3193 ii[0] = TgL + TgW;
|
Chris@10
|
3194 ii[WS(rs, 32)] = TgW - TgL;
|
Chris@10
|
3195 TgH = T1t - T30;
|
Chris@10
|
3196 ri[WS(rs, 48)] = TgH - TgK;
|
Chris@10
|
3197 ri[WS(rs, 16)] = TgH + TgK;
|
Chris@10
|
3198 TgY = TgV - TgM;
|
Chris@10
|
3199 ii[WS(rs, 16)] = TgX + TgY;
|
Chris@10
|
3200 ii[WS(rs, 48)] = TgY - TgX;
|
Chris@10
|
3201 }
|
Chris@10
|
3202 {
|
Chris@10
|
3203 E Tgr, TgC, TgZ, Th2;
|
Chris@10
|
3204 Tgr = Tgn + Tgq;
|
Chris@10
|
3205 TgC = KP707106781 * (Tgw + TgB);
|
Chris@10
|
3206 ri[WS(rs, 40)] = Tgr - TgC;
|
Chris@10
|
3207 ri[WS(rs, 8)] = Tgr + TgC;
|
Chris@10
|
3208 TgZ = KP707106781 * (TgE + TgF);
|
Chris@10
|
3209 Th2 = Th0 + Th1;
|
Chris@10
|
3210 ii[WS(rs, 8)] = TgZ + Th2;
|
Chris@10
|
3211 ii[WS(rs, 40)] = Th2 - TgZ;
|
Chris@10
|
3212 }
|
Chris@10
|
3213 {
|
Chris@10
|
3214 E TgD, TgG, Th3, Th4;
|
Chris@10
|
3215 TgD = Tgn - Tgq;
|
Chris@10
|
3216 TgG = KP707106781 * (TgE - TgF);
|
Chris@10
|
3217 ri[WS(rs, 56)] = TgD - TgG;
|
Chris@10
|
3218 ri[WS(rs, 24)] = TgD + TgG;
|
Chris@10
|
3219 Th3 = KP707106781 * (TgB - Tgw);
|
Chris@10
|
3220 Th4 = Th1 - Th0;
|
Chris@10
|
3221 ii[WS(rs, 24)] = Th3 + Th4;
|
Chris@10
|
3222 ii[WS(rs, 56)] = Th4 - Th3;
|
Chris@10
|
3223 }
|
Chris@10
|
3224 }
|
Chris@10
|
3225 {
|
Chris@10
|
3226 E Tft, Tg7, Tgh, Tgl, Th9, Thf, TfE, Th6, TfQ, Tg4, Tga, The, Tge, Tgk, Tg1;
|
Chris@10
|
3227 E Tg5;
|
Chris@10
|
3228 {
|
Chris@10
|
3229 E Tfp, Tfs, Tgf, Tgg;
|
Chris@10
|
3230 Tfp = Tj - TG;
|
Chris@10
|
3231 Tfs = Tfq - Tfr;
|
Chris@10
|
3232 Tft = Tfp - Tfs;
|
Chris@10
|
3233 Tg7 = Tfp + Tfs;
|
Chris@10
|
3234 Tgf = TfR + TfU;
|
Chris@10
|
3235 Tgg = TfY + TfZ;
|
Chris@10
|
3236 Tgh = FNMS(KP382683432, Tgg, KP923879532 * Tgf);
|
Chris@10
|
3237 Tgl = FMA(KP923879532, Tgg, KP382683432 * Tgf);
|
Chris@10
|
3238 }
|
Chris@10
|
3239 {
|
Chris@10
|
3240 E Th7, Th8, Tfy, TfD;
|
Chris@10
|
3241 Th7 = T1r - T14;
|
Chris@10
|
3242 Th8 = TgT - TgO;
|
Chris@10
|
3243 Th9 = Th7 + Th8;
|
Chris@10
|
3244 Thf = Th8 - Th7;
|
Chris@10
|
3245 Tfy = Tfw - Tfx;
|
Chris@10
|
3246 TfD = Tfz + TfC;
|
Chris@10
|
3247 TfE = KP707106781 * (Tfy - TfD);
|
Chris@10
|
3248 Th6 = KP707106781 * (Tfy + TfD);
|
Chris@10
|
3249 }
|
Chris@10
|
3250 {
|
Chris@10
|
3251 E TfK, TfP, Tg8, Tg9;
|
Chris@10
|
3252 TfK = TfI - TfJ;
|
Chris@10
|
3253 TfP = TfL - TfO;
|
Chris@10
|
3254 TfQ = FMA(KP923879532, TfK, KP382683432 * TfP);
|
Chris@10
|
3255 Tg4 = FNMS(KP923879532, TfP, KP382683432 * TfK);
|
Chris@10
|
3256 Tg8 = Tfx + Tfw;
|
Chris@10
|
3257 Tg9 = Tfz - TfC;
|
Chris@10
|
3258 Tga = KP707106781 * (Tg8 + Tg9);
|
Chris@10
|
3259 The = KP707106781 * (Tg9 - Tg8);
|
Chris@10
|
3260 }
|
Chris@10
|
3261 {
|
Chris@10
|
3262 E Tgc, Tgd, TfV, Tg0;
|
Chris@10
|
3263 Tgc = TfI + TfJ;
|
Chris@10
|
3264 Tgd = TfL + TfO;
|
Chris@10
|
3265 Tge = FMA(KP382683432, Tgc, KP923879532 * Tgd);
|
Chris@10
|
3266 Tgk = FNMS(KP382683432, Tgd, KP923879532 * Tgc);
|
Chris@10
|
3267 TfV = TfR - TfU;
|
Chris@10
|
3268 Tg0 = TfY - TfZ;
|
Chris@10
|
3269 Tg1 = FNMS(KP923879532, Tg0, KP382683432 * TfV);
|
Chris@10
|
3270 Tg5 = FMA(KP382683432, Tg0, KP923879532 * TfV);
|
Chris@10
|
3271 }
|
Chris@10
|
3272 {
|
Chris@10
|
3273 E TfF, Tg2, Thd, Thg;
|
Chris@10
|
3274 TfF = Tft + TfE;
|
Chris@10
|
3275 Tg2 = TfQ + Tg1;
|
Chris@10
|
3276 ri[WS(rs, 44)] = TfF - Tg2;
|
Chris@10
|
3277 ri[WS(rs, 12)] = TfF + Tg2;
|
Chris@10
|
3278 Thd = Tg4 + Tg5;
|
Chris@10
|
3279 Thg = The + Thf;
|
Chris@10
|
3280 ii[WS(rs, 12)] = Thd + Thg;
|
Chris@10
|
3281 ii[WS(rs, 44)] = Thg - Thd;
|
Chris@10
|
3282 }
|
Chris@10
|
3283 {
|
Chris@10
|
3284 E Tg3, Tg6, Thh, Thi;
|
Chris@10
|
3285 Tg3 = Tft - TfE;
|
Chris@10
|
3286 Tg6 = Tg4 - Tg5;
|
Chris@10
|
3287 ri[WS(rs, 60)] = Tg3 - Tg6;
|
Chris@10
|
3288 ri[WS(rs, 28)] = Tg3 + Tg6;
|
Chris@10
|
3289 Thh = Tg1 - TfQ;
|
Chris@10
|
3290 Thi = Thf - The;
|
Chris@10
|
3291 ii[WS(rs, 28)] = Thh + Thi;
|
Chris@10
|
3292 ii[WS(rs, 60)] = Thi - Thh;
|
Chris@10
|
3293 }
|
Chris@10
|
3294 {
|
Chris@10
|
3295 E Tgb, Tgi, Th5, Tha;
|
Chris@10
|
3296 Tgb = Tg7 + Tga;
|
Chris@10
|
3297 Tgi = Tge + Tgh;
|
Chris@10
|
3298 ri[WS(rs, 36)] = Tgb - Tgi;
|
Chris@10
|
3299 ri[WS(rs, 4)] = Tgb + Tgi;
|
Chris@10
|
3300 Th5 = Tgk + Tgl;
|
Chris@10
|
3301 Tha = Th6 + Th9;
|
Chris@10
|
3302 ii[WS(rs, 4)] = Th5 + Tha;
|
Chris@10
|
3303 ii[WS(rs, 36)] = Tha - Th5;
|
Chris@10
|
3304 }
|
Chris@10
|
3305 {
|
Chris@10
|
3306 E Tgj, Tgm, Thb, Thc;
|
Chris@10
|
3307 Tgj = Tg7 - Tga;
|
Chris@10
|
3308 Tgm = Tgk - Tgl;
|
Chris@10
|
3309 ri[WS(rs, 52)] = Tgj - Tgm;
|
Chris@10
|
3310 ri[WS(rs, 20)] = Tgj + Tgm;
|
Chris@10
|
3311 Thb = Tgh - Tge;
|
Chris@10
|
3312 Thc = Th9 - Th6;
|
Chris@10
|
3313 ii[WS(rs, 20)] = Thb + Thc;
|
Chris@10
|
3314 ii[WS(rs, 52)] = Thc - Thb;
|
Chris@10
|
3315 }
|
Chris@10
|
3316 }
|
Chris@10
|
3317 {
|
Chris@10
|
3318 E Td1, Ten, Tdo, ThA, ThD, ThJ, Teq, ThI, Teh, TeB, Tel, Tex, TdQ, TeA, Tek;
|
Chris@10
|
3319 E Teu;
|
Chris@10
|
3320 {
|
Chris@10
|
3321 E TcP, Td0, Teo, Tep;
|
Chris@10
|
3322 TcP = TcL - TcO;
|
Chris@10
|
3323 Td0 = KP707106781 * (TcU - TcZ);
|
Chris@10
|
3324 Td1 = TcP - Td0;
|
Chris@10
|
3325 Ten = TcP + Td0;
|
Chris@10
|
3326 {
|
Chris@10
|
3327 E Tdc, Tdn, ThB, ThC;
|
Chris@10
|
3328 Tdc = FNMS(KP923879532, Tdb, KP382683432 * Td6);
|
Chris@10
|
3329 Tdn = FMA(KP382683432, Tdh, KP923879532 * Tdm);
|
Chris@10
|
3330 Tdo = Tdc - Tdn;
|
Chris@10
|
3331 ThA = Tdc + Tdn;
|
Chris@10
|
3332 ThB = KP707106781 * (TeF - TeE);
|
Chris@10
|
3333 ThC = Thn - Thm;
|
Chris@10
|
3334 ThD = ThB + ThC;
|
Chris@10
|
3335 ThJ = ThC - ThB;
|
Chris@10
|
3336 }
|
Chris@10
|
3337 Teo = FMA(KP923879532, Td6, KP382683432 * Tdb);
|
Chris@10
|
3338 Tep = FNMS(KP923879532, Tdh, KP382683432 * Tdm);
|
Chris@10
|
3339 Teq = Teo + Tep;
|
Chris@10
|
3340 ThI = Tep - Teo;
|
Chris@10
|
3341 {
|
Chris@10
|
3342 E Te7, Tev, Teg, Tew, Te6, Tef;
|
Chris@10
|
3343 Te6 = KP707106781 * (Te0 - Te5);
|
Chris@10
|
3344 Te7 = TdV - Te6;
|
Chris@10
|
3345 Tev = TdV + Te6;
|
Chris@10
|
3346 Tef = KP707106781 * (Ted - Tee);
|
Chris@10
|
3347 Teg = Tec - Tef;
|
Chris@10
|
3348 Tew = Tec + Tef;
|
Chris@10
|
3349 Teh = FNMS(KP980785280, Teg, KP195090322 * Te7);
|
Chris@10
|
3350 TeB = FMA(KP831469612, Tew, KP555570233 * Tev);
|
Chris@10
|
3351 Tel = FMA(KP195090322, Teg, KP980785280 * Te7);
|
Chris@10
|
3352 Tex = FNMS(KP555570233, Tew, KP831469612 * Tev);
|
Chris@10
|
3353 }
|
Chris@10
|
3354 {
|
Chris@10
|
3355 E TdG, Tes, TdP, Tet, TdF, TdO;
|
Chris@10
|
3356 TdF = KP707106781 * (Tdz - TdE);
|
Chris@10
|
3357 TdG = Tdu - TdF;
|
Chris@10
|
3358 Tes = Tdu + TdF;
|
Chris@10
|
3359 TdO = KP707106781 * (TdM - TdN);
|
Chris@10
|
3360 TdP = TdL - TdO;
|
Chris@10
|
3361 Tet = TdL + TdO;
|
Chris@10
|
3362 TdQ = FMA(KP980785280, TdG, KP195090322 * TdP);
|
Chris@10
|
3363 TeA = FNMS(KP555570233, Tet, KP831469612 * Tes);
|
Chris@10
|
3364 Tek = FNMS(KP980785280, TdP, KP195090322 * TdG);
|
Chris@10
|
3365 Teu = FMA(KP555570233, Tes, KP831469612 * Tet);
|
Chris@10
|
3366 }
|
Chris@10
|
3367 }
|
Chris@10
|
3368 {
|
Chris@10
|
3369 E Tdp, Tei, ThH, ThK;
|
Chris@10
|
3370 Tdp = Td1 + Tdo;
|
Chris@10
|
3371 Tei = TdQ + Teh;
|
Chris@10
|
3372 ri[WS(rs, 46)] = Tdp - Tei;
|
Chris@10
|
3373 ri[WS(rs, 14)] = Tdp + Tei;
|
Chris@10
|
3374 ThH = Tek + Tel;
|
Chris@10
|
3375 ThK = ThI + ThJ;
|
Chris@10
|
3376 ii[WS(rs, 14)] = ThH + ThK;
|
Chris@10
|
3377 ii[WS(rs, 46)] = ThK - ThH;
|
Chris@10
|
3378 }
|
Chris@10
|
3379 {
|
Chris@10
|
3380 E Tej, Tem, ThL, ThM;
|
Chris@10
|
3381 Tej = Td1 - Tdo;
|
Chris@10
|
3382 Tem = Tek - Tel;
|
Chris@10
|
3383 ri[WS(rs, 62)] = Tej - Tem;
|
Chris@10
|
3384 ri[WS(rs, 30)] = Tej + Tem;
|
Chris@10
|
3385 ThL = Teh - TdQ;
|
Chris@10
|
3386 ThM = ThJ - ThI;
|
Chris@10
|
3387 ii[WS(rs, 30)] = ThL + ThM;
|
Chris@10
|
3388 ii[WS(rs, 62)] = ThM - ThL;
|
Chris@10
|
3389 }
|
Chris@10
|
3390 {
|
Chris@10
|
3391 E Ter, Tey, Thz, ThE;
|
Chris@10
|
3392 Ter = Ten + Teq;
|
Chris@10
|
3393 Tey = Teu + Tex;
|
Chris@10
|
3394 ri[WS(rs, 38)] = Ter - Tey;
|
Chris@10
|
3395 ri[WS(rs, 6)] = Ter + Tey;
|
Chris@10
|
3396 Thz = TeA + TeB;
|
Chris@10
|
3397 ThE = ThA + ThD;
|
Chris@10
|
3398 ii[WS(rs, 6)] = Thz + ThE;
|
Chris@10
|
3399 ii[WS(rs, 38)] = ThE - Thz;
|
Chris@10
|
3400 }
|
Chris@10
|
3401 {
|
Chris@10
|
3402 E Tez, TeC, ThF, ThG;
|
Chris@10
|
3403 Tez = Ten - Teq;
|
Chris@10
|
3404 TeC = TeA - TeB;
|
Chris@10
|
3405 ri[WS(rs, 54)] = Tez - TeC;
|
Chris@10
|
3406 ri[WS(rs, 22)] = Tez + TeC;
|
Chris@10
|
3407 ThF = Tex - Teu;
|
Chris@10
|
3408 ThG = ThD - ThA;
|
Chris@10
|
3409 ii[WS(rs, 22)] = ThF + ThG;
|
Chris@10
|
3410 ii[WS(rs, 54)] = ThG - ThF;
|
Chris@10
|
3411 }
|
Chris@10
|
3412 }
|
Chris@10
|
3413 {
|
Chris@10
|
3414 E TeH, Tf9, TeO, Thk, Thp, Thv, Tfc, Thu, Tf3, Tfn, Tf7, Tfj, TeW, Tfm, Tf6;
|
Chris@10
|
3415 E Tfg;
|
Chris@10
|
3416 {
|
Chris@10
|
3417 E TeD, TeG, Tfa, Tfb;
|
Chris@10
|
3418 TeD = TcL + TcO;
|
Chris@10
|
3419 TeG = KP707106781 * (TeE + TeF);
|
Chris@10
|
3420 TeH = TeD - TeG;
|
Chris@10
|
3421 Tf9 = TeD + TeG;
|
Chris@10
|
3422 {
|
Chris@10
|
3423 E TeK, TeN, Thl, Tho;
|
Chris@10
|
3424 TeK = FNMS(KP382683432, TeJ, KP923879532 * TeI);
|
Chris@10
|
3425 TeN = FMA(KP923879532, TeL, KP382683432 * TeM);
|
Chris@10
|
3426 TeO = TeK - TeN;
|
Chris@10
|
3427 Thk = TeK + TeN;
|
Chris@10
|
3428 Thl = KP707106781 * (TcU + TcZ);
|
Chris@10
|
3429 Tho = Thm + Thn;
|
Chris@10
|
3430 Thp = Thl + Tho;
|
Chris@10
|
3431 Thv = Tho - Thl;
|
Chris@10
|
3432 }
|
Chris@10
|
3433 Tfa = FMA(KP382683432, TeI, KP923879532 * TeJ);
|
Chris@10
|
3434 Tfb = FNMS(KP382683432, TeL, KP923879532 * TeM);
|
Chris@10
|
3435 Tfc = Tfa + Tfb;
|
Chris@10
|
3436 Thu = Tfb - Tfa;
|
Chris@10
|
3437 {
|
Chris@10
|
3438 E TeZ, Tfh, Tf2, Tfi, TeY, Tf1;
|
Chris@10
|
3439 TeY = KP707106781 * (Tee + Ted);
|
Chris@10
|
3440 TeZ = TeX - TeY;
|
Chris@10
|
3441 Tfh = TeX + TeY;
|
Chris@10
|
3442 Tf1 = KP707106781 * (Te0 + Te5);
|
Chris@10
|
3443 Tf2 = Tf0 - Tf1;
|
Chris@10
|
3444 Tfi = Tf0 + Tf1;
|
Chris@10
|
3445 Tf3 = FNMS(KP831469612, Tf2, KP555570233 * TeZ);
|
Chris@10
|
3446 Tfn = FMA(KP195090322, Tfh, KP980785280 * Tfi);
|
Chris@10
|
3447 Tf7 = FMA(KP831469612, TeZ, KP555570233 * Tf2);
|
Chris@10
|
3448 Tfj = FNMS(KP195090322, Tfi, KP980785280 * Tfh);
|
Chris@10
|
3449 }
|
Chris@10
|
3450 {
|
Chris@10
|
3451 E TeS, Tfe, TeV, Tff, TeR, TeU;
|
Chris@10
|
3452 TeR = KP707106781 * (TdE + Tdz);
|
Chris@10
|
3453 TeS = TeQ - TeR;
|
Chris@10
|
3454 Tfe = TeQ + TeR;
|
Chris@10
|
3455 TeU = KP707106781 * (TdM + TdN);
|
Chris@10
|
3456 TeV = TeT - TeU;
|
Chris@10
|
3457 Tff = TeT + TeU;
|
Chris@10
|
3458 TeW = FMA(KP555570233, TeS, KP831469612 * TeV);
|
Chris@10
|
3459 Tfm = FNMS(KP195090322, Tfe, KP980785280 * Tff);
|
Chris@10
|
3460 Tf6 = FNMS(KP831469612, TeS, KP555570233 * TeV);
|
Chris@10
|
3461 Tfg = FMA(KP980785280, Tfe, KP195090322 * Tff);
|
Chris@10
|
3462 }
|
Chris@10
|
3463 }
|
Chris@10
|
3464 {
|
Chris@10
|
3465 E TeP, Tf4, Tht, Thw;
|
Chris@10
|
3466 TeP = TeH + TeO;
|
Chris@10
|
3467 Tf4 = TeW + Tf3;
|
Chris@10
|
3468 ri[WS(rs, 42)] = TeP - Tf4;
|
Chris@10
|
3469 ri[WS(rs, 10)] = TeP + Tf4;
|
Chris@10
|
3470 Tht = Tf6 + Tf7;
|
Chris@10
|
3471 Thw = Thu + Thv;
|
Chris@10
|
3472 ii[WS(rs, 10)] = Tht + Thw;
|
Chris@10
|
3473 ii[WS(rs, 42)] = Thw - Tht;
|
Chris@10
|
3474 }
|
Chris@10
|
3475 {
|
Chris@10
|
3476 E Tf5, Tf8, Thx, Thy;
|
Chris@10
|
3477 Tf5 = TeH - TeO;
|
Chris@10
|
3478 Tf8 = Tf6 - Tf7;
|
Chris@10
|
3479 ri[WS(rs, 58)] = Tf5 - Tf8;
|
Chris@10
|
3480 ri[WS(rs, 26)] = Tf5 + Tf8;
|
Chris@10
|
3481 Thx = Tf3 - TeW;
|
Chris@10
|
3482 Thy = Thv - Thu;
|
Chris@10
|
3483 ii[WS(rs, 26)] = Thx + Thy;
|
Chris@10
|
3484 ii[WS(rs, 58)] = Thy - Thx;
|
Chris@10
|
3485 }
|
Chris@10
|
3486 {
|
Chris@10
|
3487 E Tfd, Tfk, Thj, Thq;
|
Chris@10
|
3488 Tfd = Tf9 + Tfc;
|
Chris@10
|
3489 Tfk = Tfg + Tfj;
|
Chris@10
|
3490 ri[WS(rs, 34)] = Tfd - Tfk;
|
Chris@10
|
3491 ri[WS(rs, 2)] = Tfd + Tfk;
|
Chris@10
|
3492 Thj = Tfm + Tfn;
|
Chris@10
|
3493 Thq = Thk + Thp;
|
Chris@10
|
3494 ii[WS(rs, 2)] = Thj + Thq;
|
Chris@10
|
3495 ii[WS(rs, 34)] = Thq - Thj;
|
Chris@10
|
3496 }
|
Chris@10
|
3497 {
|
Chris@10
|
3498 E Tfl, Tfo, Thr, Ths;
|
Chris@10
|
3499 Tfl = Tf9 - Tfc;
|
Chris@10
|
3500 Tfo = Tfm - Tfn;
|
Chris@10
|
3501 ri[WS(rs, 50)] = Tfl - Tfo;
|
Chris@10
|
3502 ri[WS(rs, 18)] = Tfl + Tfo;
|
Chris@10
|
3503 Thr = Tfj - Tfg;
|
Chris@10
|
3504 Ths = Thp - Thk;
|
Chris@10
|
3505 ii[WS(rs, 18)] = Thr + Ths;
|
Chris@10
|
3506 ii[WS(rs, 50)] = Ths - Thr;
|
Chris@10
|
3507 }
|
Chris@10
|
3508 }
|
Chris@10
|
3509 {
|
Chris@10
|
3510 E T6L, T9x, TiD, TiJ, T7E, TiI, T9A, TiA, T8y, T9K, T9u, T9E, T9r, T9L, T9v;
|
Chris@10
|
3511 E T9H;
|
Chris@10
|
3512 {
|
Chris@10
|
3513 E T6n, T6K, TiB, TiC;
|
Chris@10
|
3514 T6n = T6b - T6m;
|
Chris@10
|
3515 T6K = T6y - T6J;
|
Chris@10
|
3516 T6L = T6n - T6K;
|
Chris@10
|
3517 T9x = T6n + T6K;
|
Chris@10
|
3518 TiB = T9P - T9O;
|
Chris@10
|
3519 TiC = Tin - Tim;
|
Chris@10
|
3520 TiD = TiB + TiC;
|
Chris@10
|
3521 TiJ = TiC - TiB;
|
Chris@10
|
3522 }
|
Chris@10
|
3523 {
|
Chris@10
|
3524 E T7c, T9y, T7D, T9z;
|
Chris@10
|
3525 {
|
Chris@10
|
3526 E T72, T7b, T7t, T7C;
|
Chris@10
|
3527 T72 = T6Q - T71;
|
Chris@10
|
3528 T7b = T77 - T7a;
|
Chris@10
|
3529 T7c = FNMS(KP980785280, T7b, KP195090322 * T72);
|
Chris@10
|
3530 T9y = FMA(KP980785280, T72, KP195090322 * T7b);
|
Chris@10
|
3531 T7t = T7h - T7s;
|
Chris@10
|
3532 T7C = T7y - T7B;
|
Chris@10
|
3533 T7D = FMA(KP195090322, T7t, KP980785280 * T7C);
|
Chris@10
|
3534 T9z = FNMS(KP980785280, T7t, KP195090322 * T7C);
|
Chris@10
|
3535 }
|
Chris@10
|
3536 T7E = T7c - T7D;
|
Chris@10
|
3537 TiI = T9z - T9y;
|
Chris@10
|
3538 T9A = T9y + T9z;
|
Chris@10
|
3539 TiA = T7c + T7D;
|
Chris@10
|
3540 }
|
Chris@10
|
3541 {
|
Chris@10
|
3542 E T8k, T9C, T8x, T9D;
|
Chris@10
|
3543 {
|
Chris@10
|
3544 E T7W, T8j, T8t, T8w;
|
Chris@10
|
3545 T7W = T7K - T7V;
|
Chris@10
|
3546 T8j = T87 - T8i;
|
Chris@10
|
3547 T8k = T7W - T8j;
|
Chris@10
|
3548 T9C = T7W + T8j;
|
Chris@10
|
3549 T8t = T8p - T8s;
|
Chris@10
|
3550 T8w = T8u - T8v;
|
Chris@10
|
3551 T8x = T8t - T8w;
|
Chris@10
|
3552 T9D = T8t + T8w;
|
Chris@10
|
3553 }
|
Chris@10
|
3554 T8y = FMA(KP995184726, T8k, KP098017140 * T8x);
|
Chris@10
|
3555 T9K = FNMS(KP634393284, T9D, KP773010453 * T9C);
|
Chris@10
|
3556 T9u = FNMS(KP995184726, T8x, KP098017140 * T8k);
|
Chris@10
|
3557 T9E = FMA(KP634393284, T9C, KP773010453 * T9D);
|
Chris@10
|
3558 }
|
Chris@10
|
3559 {
|
Chris@10
|
3560 E T9d, T9F, T9q, T9G;
|
Chris@10
|
3561 {
|
Chris@10
|
3562 E T8P, T9c, T9m, T9p;
|
Chris@10
|
3563 T8P = T8D - T8O;
|
Chris@10
|
3564 T9c = T90 - T9b;
|
Chris@10
|
3565 T9d = T8P - T9c;
|
Chris@10
|
3566 T9F = T8P + T9c;
|
Chris@10
|
3567 T9m = T9i - T9l;
|
Chris@10
|
3568 T9p = T9n - T9o;
|
Chris@10
|
3569 T9q = T9m - T9p;
|
Chris@10
|
3570 T9G = T9m + T9p;
|
Chris@10
|
3571 }
|
Chris@10
|
3572 T9r = FNMS(KP995184726, T9q, KP098017140 * T9d);
|
Chris@10
|
3573 T9L = FMA(KP773010453, T9G, KP634393284 * T9F);
|
Chris@10
|
3574 T9v = FMA(KP098017140, T9q, KP995184726 * T9d);
|
Chris@10
|
3575 T9H = FNMS(KP634393284, T9G, KP773010453 * T9F);
|
Chris@10
|
3576 }
|
Chris@10
|
3577 {
|
Chris@10
|
3578 E T7F, T9s, TiH, TiK;
|
Chris@10
|
3579 T7F = T6L + T7E;
|
Chris@10
|
3580 T9s = T8y + T9r;
|
Chris@10
|
3581 ri[WS(rs, 47)] = T7F - T9s;
|
Chris@10
|
3582 ri[WS(rs, 15)] = T7F + T9s;
|
Chris@10
|
3583 TiH = T9u + T9v;
|
Chris@10
|
3584 TiK = TiI + TiJ;
|
Chris@10
|
3585 ii[WS(rs, 15)] = TiH + TiK;
|
Chris@10
|
3586 ii[WS(rs, 47)] = TiK - TiH;
|
Chris@10
|
3587 }
|
Chris@10
|
3588 {
|
Chris@10
|
3589 E T9t, T9w, TiL, TiM;
|
Chris@10
|
3590 T9t = T6L - T7E;
|
Chris@10
|
3591 T9w = T9u - T9v;
|
Chris@10
|
3592 ri[WS(rs, 63)] = T9t - T9w;
|
Chris@10
|
3593 ri[WS(rs, 31)] = T9t + T9w;
|
Chris@10
|
3594 TiL = T9r - T8y;
|
Chris@10
|
3595 TiM = TiJ - TiI;
|
Chris@10
|
3596 ii[WS(rs, 31)] = TiL + TiM;
|
Chris@10
|
3597 ii[WS(rs, 63)] = TiM - TiL;
|
Chris@10
|
3598 }
|
Chris@10
|
3599 {
|
Chris@10
|
3600 E T9B, T9I, Tiz, TiE;
|
Chris@10
|
3601 T9B = T9x + T9A;
|
Chris@10
|
3602 T9I = T9E + T9H;
|
Chris@10
|
3603 ri[WS(rs, 39)] = T9B - T9I;
|
Chris@10
|
3604 ri[WS(rs, 7)] = T9B + T9I;
|
Chris@10
|
3605 Tiz = T9K + T9L;
|
Chris@10
|
3606 TiE = TiA + TiD;
|
Chris@10
|
3607 ii[WS(rs, 7)] = Tiz + TiE;
|
Chris@10
|
3608 ii[WS(rs, 39)] = TiE - Tiz;
|
Chris@10
|
3609 }
|
Chris@10
|
3610 {
|
Chris@10
|
3611 E T9J, T9M, TiF, TiG;
|
Chris@10
|
3612 T9J = T9x - T9A;
|
Chris@10
|
3613 T9M = T9K - T9L;
|
Chris@10
|
3614 ri[WS(rs, 55)] = T9J - T9M;
|
Chris@10
|
3615 ri[WS(rs, 23)] = T9J + T9M;
|
Chris@10
|
3616 TiF = T9H - T9E;
|
Chris@10
|
3617 TiG = TiD - TiA;
|
Chris@10
|
3618 ii[WS(rs, 23)] = TiF + TiG;
|
Chris@10
|
3619 ii[WS(rs, 55)] = TiG - TiF;
|
Chris@10
|
3620 }
|
Chris@10
|
3621 }
|
Chris@10
|
3622 {
|
Chris@10
|
3623 E TaL, TbJ, Ti9, Tif, Tb0, Tie, TbM, Ti6, Tbk, TbW, TbG, TbQ, TbD, TbX, TbH;
|
Chris@10
|
3624 E TbT;
|
Chris@10
|
3625 {
|
Chris@10
|
3626 E TaD, TaK, Ti7, Ti8;
|
Chris@10
|
3627 TaD = Taz - TaC;
|
Chris@10
|
3628 TaK = TaG - TaJ;
|
Chris@10
|
3629 TaL = TaD - TaK;
|
Chris@10
|
3630 TbJ = TaD + TaK;
|
Chris@10
|
3631 Ti7 = Tc1 - Tc0;
|
Chris@10
|
3632 Ti8 = ThT - ThQ;
|
Chris@10
|
3633 Ti9 = Ti7 + Ti8;
|
Chris@10
|
3634 Tif = Ti8 - Ti7;
|
Chris@10
|
3635 }
|
Chris@10
|
3636 {
|
Chris@10
|
3637 E TaS, TbK, TaZ, TbL;
|
Chris@10
|
3638 {
|
Chris@10
|
3639 E TaO, TaR, TaV, TaY;
|
Chris@10
|
3640 TaO = TaM - TaN;
|
Chris@10
|
3641 TaR = TaP - TaQ;
|
Chris@10
|
3642 TaS = FNMS(KP831469612, TaR, KP555570233 * TaO);
|
Chris@10
|
3643 TbK = FMA(KP555570233, TaR, KP831469612 * TaO);
|
Chris@10
|
3644 TaV = TaT - TaU;
|
Chris@10
|
3645 TaY = TaW - TaX;
|
Chris@10
|
3646 TaZ = FMA(KP831469612, TaV, KP555570233 * TaY);
|
Chris@10
|
3647 TbL = FNMS(KP831469612, TaY, KP555570233 * TaV);
|
Chris@10
|
3648 }
|
Chris@10
|
3649 Tb0 = TaS - TaZ;
|
Chris@10
|
3650 Tie = TbL - TbK;
|
Chris@10
|
3651 TbM = TbK + TbL;
|
Chris@10
|
3652 Ti6 = TaS + TaZ;
|
Chris@10
|
3653 }
|
Chris@10
|
3654 {
|
Chris@10
|
3655 E Tbc, TbO, Tbj, TbP;
|
Chris@10
|
3656 {
|
Chris@10
|
3657 E Tb4, Tbb, Tbf, Tbi;
|
Chris@10
|
3658 Tb4 = Tb2 - Tb3;
|
Chris@10
|
3659 Tbb = Tb7 - Tba;
|
Chris@10
|
3660 Tbc = Tb4 - Tbb;
|
Chris@10
|
3661 TbO = Tb4 + Tbb;
|
Chris@10
|
3662 Tbf = Tbd - Tbe;
|
Chris@10
|
3663 Tbi = Tbg - Tbh;
|
Chris@10
|
3664 Tbj = Tbf - Tbi;
|
Chris@10
|
3665 TbP = Tbf + Tbi;
|
Chris@10
|
3666 }
|
Chris@10
|
3667 Tbk = FMA(KP956940335, Tbc, KP290284677 * Tbj);
|
Chris@10
|
3668 TbW = FNMS(KP471396736, TbP, KP881921264 * TbO);
|
Chris@10
|
3669 TbG = FNMS(KP956940335, Tbj, KP290284677 * Tbc);
|
Chris@10
|
3670 TbQ = FMA(KP471396736, TbO, KP881921264 * TbP);
|
Chris@10
|
3671 }
|
Chris@10
|
3672 {
|
Chris@10
|
3673 E Tbv, TbR, TbC, TbS;
|
Chris@10
|
3674 {
|
Chris@10
|
3675 E Tbn, Tbu, Tby, TbB;
|
Chris@10
|
3676 Tbn = Tbl - Tbm;
|
Chris@10
|
3677 Tbu = Tbq - Tbt;
|
Chris@10
|
3678 Tbv = Tbn - Tbu;
|
Chris@10
|
3679 TbR = Tbn + Tbu;
|
Chris@10
|
3680 Tby = Tbw - Tbx;
|
Chris@10
|
3681 TbB = Tbz - TbA;
|
Chris@10
|
3682 TbC = Tby - TbB;
|
Chris@10
|
3683 TbS = Tby + TbB;
|
Chris@10
|
3684 }
|
Chris@10
|
3685 TbD = FNMS(KP956940335, TbC, KP290284677 * Tbv);
|
Chris@10
|
3686 TbX = FMA(KP881921264, TbS, KP471396736 * TbR);
|
Chris@10
|
3687 TbH = FMA(KP290284677, TbC, KP956940335 * Tbv);
|
Chris@10
|
3688 TbT = FNMS(KP471396736, TbS, KP881921264 * TbR);
|
Chris@10
|
3689 }
|
Chris@10
|
3690 {
|
Chris@10
|
3691 E Tb1, TbE, Tid, Tig;
|
Chris@10
|
3692 Tb1 = TaL + Tb0;
|
Chris@10
|
3693 TbE = Tbk + TbD;
|
Chris@10
|
3694 ri[WS(rs, 45)] = Tb1 - TbE;
|
Chris@10
|
3695 ri[WS(rs, 13)] = Tb1 + TbE;
|
Chris@10
|
3696 Tid = TbG + TbH;
|
Chris@10
|
3697 Tig = Tie + Tif;
|
Chris@10
|
3698 ii[WS(rs, 13)] = Tid + Tig;
|
Chris@10
|
3699 ii[WS(rs, 45)] = Tig - Tid;
|
Chris@10
|
3700 }
|
Chris@10
|
3701 {
|
Chris@10
|
3702 E TbF, TbI, Tih, Tii;
|
Chris@10
|
3703 TbF = TaL - Tb0;
|
Chris@10
|
3704 TbI = TbG - TbH;
|
Chris@10
|
3705 ri[WS(rs, 61)] = TbF - TbI;
|
Chris@10
|
3706 ri[WS(rs, 29)] = TbF + TbI;
|
Chris@10
|
3707 Tih = TbD - Tbk;
|
Chris@10
|
3708 Tii = Tif - Tie;
|
Chris@10
|
3709 ii[WS(rs, 29)] = Tih + Tii;
|
Chris@10
|
3710 ii[WS(rs, 61)] = Tii - Tih;
|
Chris@10
|
3711 }
|
Chris@10
|
3712 {
|
Chris@10
|
3713 E TbN, TbU, Ti5, Tia;
|
Chris@10
|
3714 TbN = TbJ + TbM;
|
Chris@10
|
3715 TbU = TbQ + TbT;
|
Chris@10
|
3716 ri[WS(rs, 37)] = TbN - TbU;
|
Chris@10
|
3717 ri[WS(rs, 5)] = TbN + TbU;
|
Chris@10
|
3718 Ti5 = TbW + TbX;
|
Chris@10
|
3719 Tia = Ti6 + Ti9;
|
Chris@10
|
3720 ii[WS(rs, 5)] = Ti5 + Tia;
|
Chris@10
|
3721 ii[WS(rs, 37)] = Tia - Ti5;
|
Chris@10
|
3722 }
|
Chris@10
|
3723 {
|
Chris@10
|
3724 E TbV, TbY, Tib, Tic;
|
Chris@10
|
3725 TbV = TbJ - TbM;
|
Chris@10
|
3726 TbY = TbW - TbX;
|
Chris@10
|
3727 ri[WS(rs, 53)] = TbV - TbY;
|
Chris@10
|
3728 ri[WS(rs, 21)] = TbV + TbY;
|
Chris@10
|
3729 Tib = TbT - TbQ;
|
Chris@10
|
3730 Tic = Ti9 - Ti6;
|
Chris@10
|
3731 ii[WS(rs, 21)] = Tib + Tic;
|
Chris@10
|
3732 ii[WS(rs, 53)] = Tic - Tib;
|
Chris@10
|
3733 }
|
Chris@10
|
3734 }
|
Chris@10
|
3735 {
|
Chris@10
|
3736 E Tc3, Tcv, ThV, Ti1, Tca, Ti0, Tcy, ThO, Tci, TcI, Tcs, TcC, Tcp, TcJ, Tct;
|
Chris@10
|
3737 E TcF;
|
Chris@10
|
3738 {
|
Chris@10
|
3739 E TbZ, Tc2, ThP, ThU;
|
Chris@10
|
3740 TbZ = Taz + TaC;
|
Chris@10
|
3741 Tc2 = Tc0 + Tc1;
|
Chris@10
|
3742 Tc3 = TbZ - Tc2;
|
Chris@10
|
3743 Tcv = TbZ + Tc2;
|
Chris@10
|
3744 ThP = TaG + TaJ;
|
Chris@10
|
3745 ThU = ThQ + ThT;
|
Chris@10
|
3746 ThV = ThP + ThU;
|
Chris@10
|
3747 Ti1 = ThU - ThP;
|
Chris@10
|
3748 }
|
Chris@10
|
3749 {
|
Chris@10
|
3750 E Tc6, Tcw, Tc9, Tcx;
|
Chris@10
|
3751 {
|
Chris@10
|
3752 E Tc4, Tc5, Tc7, Tc8;
|
Chris@10
|
3753 Tc4 = TaM + TaN;
|
Chris@10
|
3754 Tc5 = TaP + TaQ;
|
Chris@10
|
3755 Tc6 = FNMS(KP195090322, Tc5, KP980785280 * Tc4);
|
Chris@10
|
3756 Tcw = FMA(KP980785280, Tc5, KP195090322 * Tc4);
|
Chris@10
|
3757 Tc7 = TaT + TaU;
|
Chris@10
|
3758 Tc8 = TaW + TaX;
|
Chris@10
|
3759 Tc9 = FMA(KP195090322, Tc7, KP980785280 * Tc8);
|
Chris@10
|
3760 Tcx = FNMS(KP195090322, Tc8, KP980785280 * Tc7);
|
Chris@10
|
3761 }
|
Chris@10
|
3762 Tca = Tc6 - Tc9;
|
Chris@10
|
3763 Ti0 = Tcx - Tcw;
|
Chris@10
|
3764 Tcy = Tcw + Tcx;
|
Chris@10
|
3765 ThO = Tc6 + Tc9;
|
Chris@10
|
3766 }
|
Chris@10
|
3767 {
|
Chris@10
|
3768 E Tce, TcA, Tch, TcB;
|
Chris@10
|
3769 {
|
Chris@10
|
3770 E Tcc, Tcd, Tcf, Tcg;
|
Chris@10
|
3771 Tcc = Tbd + Tbe;
|
Chris@10
|
3772 Tcd = Tba + Tb7;
|
Chris@10
|
3773 Tce = Tcc - Tcd;
|
Chris@10
|
3774 TcA = Tcc + Tcd;
|
Chris@10
|
3775 Tcf = Tb2 + Tb3;
|
Chris@10
|
3776 Tcg = Tbg + Tbh;
|
Chris@10
|
3777 Tch = Tcf - Tcg;
|
Chris@10
|
3778 TcB = Tcf + Tcg;
|
Chris@10
|
3779 }
|
Chris@10
|
3780 Tci = FMA(KP634393284, Tce, KP773010453 * Tch);
|
Chris@10
|
3781 TcI = FNMS(KP098017140, TcA, KP995184726 * TcB);
|
Chris@10
|
3782 Tcs = FNMS(KP773010453, Tce, KP634393284 * Tch);
|
Chris@10
|
3783 TcC = FMA(KP995184726, TcA, KP098017140 * TcB);
|
Chris@10
|
3784 }
|
Chris@10
|
3785 {
|
Chris@10
|
3786 E Tcl, TcD, Tco, TcE;
|
Chris@10
|
3787 {
|
Chris@10
|
3788 E Tcj, Tck, Tcm, Tcn;
|
Chris@10
|
3789 Tcj = Tbl + Tbm;
|
Chris@10
|
3790 Tck = TbA + Tbz;
|
Chris@10
|
3791 Tcl = Tcj - Tck;
|
Chris@10
|
3792 TcD = Tcj + Tck;
|
Chris@10
|
3793 Tcm = Tbw + Tbx;
|
Chris@10
|
3794 Tcn = Tbq + Tbt;
|
Chris@10
|
3795 Tco = Tcm - Tcn;
|
Chris@10
|
3796 TcE = Tcm + Tcn;
|
Chris@10
|
3797 }
|
Chris@10
|
3798 Tcp = FNMS(KP773010453, Tco, KP634393284 * Tcl);
|
Chris@10
|
3799 TcJ = FMA(KP098017140, TcD, KP995184726 * TcE);
|
Chris@10
|
3800 Tct = FMA(KP773010453, Tcl, KP634393284 * Tco);
|
Chris@10
|
3801 TcF = FNMS(KP098017140, TcE, KP995184726 * TcD);
|
Chris@10
|
3802 }
|
Chris@10
|
3803 {
|
Chris@10
|
3804 E Tcb, Tcq, ThZ, Ti2;
|
Chris@10
|
3805 Tcb = Tc3 + Tca;
|
Chris@10
|
3806 Tcq = Tci + Tcp;
|
Chris@10
|
3807 ri[WS(rs, 41)] = Tcb - Tcq;
|
Chris@10
|
3808 ri[WS(rs, 9)] = Tcb + Tcq;
|
Chris@10
|
3809 ThZ = Tcs + Tct;
|
Chris@10
|
3810 Ti2 = Ti0 + Ti1;
|
Chris@10
|
3811 ii[WS(rs, 9)] = ThZ + Ti2;
|
Chris@10
|
3812 ii[WS(rs, 41)] = Ti2 - ThZ;
|
Chris@10
|
3813 }
|
Chris@10
|
3814 {
|
Chris@10
|
3815 E Tcr, Tcu, Ti3, Ti4;
|
Chris@10
|
3816 Tcr = Tc3 - Tca;
|
Chris@10
|
3817 Tcu = Tcs - Tct;
|
Chris@10
|
3818 ri[WS(rs, 57)] = Tcr - Tcu;
|
Chris@10
|
3819 ri[WS(rs, 25)] = Tcr + Tcu;
|
Chris@10
|
3820 Ti3 = Tcp - Tci;
|
Chris@10
|
3821 Ti4 = Ti1 - Ti0;
|
Chris@10
|
3822 ii[WS(rs, 25)] = Ti3 + Ti4;
|
Chris@10
|
3823 ii[WS(rs, 57)] = Ti4 - Ti3;
|
Chris@10
|
3824 }
|
Chris@10
|
3825 {
|
Chris@10
|
3826 E Tcz, TcG, ThN, ThW;
|
Chris@10
|
3827 Tcz = Tcv + Tcy;
|
Chris@10
|
3828 TcG = TcC + TcF;
|
Chris@10
|
3829 ri[WS(rs, 33)] = Tcz - TcG;
|
Chris@10
|
3830 ri[WS(rs, 1)] = Tcz + TcG;
|
Chris@10
|
3831 ThN = TcI + TcJ;
|
Chris@10
|
3832 ThW = ThO + ThV;
|
Chris@10
|
3833 ii[WS(rs, 1)] = ThN + ThW;
|
Chris@10
|
3834 ii[WS(rs, 33)] = ThW - ThN;
|
Chris@10
|
3835 }
|
Chris@10
|
3836 {
|
Chris@10
|
3837 E TcH, TcK, ThX, ThY;
|
Chris@10
|
3838 TcH = Tcv - Tcy;
|
Chris@10
|
3839 TcK = TcI - TcJ;
|
Chris@10
|
3840 ri[WS(rs, 49)] = TcH - TcK;
|
Chris@10
|
3841 ri[WS(rs, 17)] = TcH + TcK;
|
Chris@10
|
3842 ThX = TcF - TcC;
|
Chris@10
|
3843 ThY = ThV - ThO;
|
Chris@10
|
3844 ii[WS(rs, 17)] = ThX + ThY;
|
Chris@10
|
3845 ii[WS(rs, 49)] = ThY - ThX;
|
Chris@10
|
3846 }
|
Chris@10
|
3847 }
|
Chris@10
|
3848 {
|
Chris@10
|
3849 E T9R, Taj, Tip, Tiv, T9Y, Tiu, Tam, Tik, Ta6, Taw, Tag, Taq, Tad, Tax, Tah;
|
Chris@10
|
3850 E Tat;
|
Chris@10
|
3851 {
|
Chris@10
|
3852 E T9N, T9Q, Til, Tio;
|
Chris@10
|
3853 T9N = T6b + T6m;
|
Chris@10
|
3854 T9Q = T9O + T9P;
|
Chris@10
|
3855 T9R = T9N - T9Q;
|
Chris@10
|
3856 Taj = T9N + T9Q;
|
Chris@10
|
3857 Til = T6y + T6J;
|
Chris@10
|
3858 Tio = Tim + Tin;
|
Chris@10
|
3859 Tip = Til + Tio;
|
Chris@10
|
3860 Tiv = Tio - Til;
|
Chris@10
|
3861 }
|
Chris@10
|
3862 {
|
Chris@10
|
3863 E T9U, Tak, T9X, Tal;
|
Chris@10
|
3864 {
|
Chris@10
|
3865 E T9S, T9T, T9V, T9W;
|
Chris@10
|
3866 T9S = T6Q + T71;
|
Chris@10
|
3867 T9T = T77 + T7a;
|
Chris@10
|
3868 T9U = FNMS(KP555570233, T9T, KP831469612 * T9S);
|
Chris@10
|
3869 Tak = FMA(KP555570233, T9S, KP831469612 * T9T);
|
Chris@10
|
3870 T9V = T7h + T7s;
|
Chris@10
|
3871 T9W = T7y + T7B;
|
Chris@10
|
3872 T9X = FMA(KP831469612, T9V, KP555570233 * T9W);
|
Chris@10
|
3873 Tal = FNMS(KP555570233, T9V, KP831469612 * T9W);
|
Chris@10
|
3874 }
|
Chris@10
|
3875 T9Y = T9U - T9X;
|
Chris@10
|
3876 Tiu = Tal - Tak;
|
Chris@10
|
3877 Tam = Tak + Tal;
|
Chris@10
|
3878 Tik = T9U + T9X;
|
Chris@10
|
3879 }
|
Chris@10
|
3880 {
|
Chris@10
|
3881 E Ta2, Tao, Ta5, Tap;
|
Chris@10
|
3882 {
|
Chris@10
|
3883 E Ta0, Ta1, Ta3, Ta4;
|
Chris@10
|
3884 Ta0 = T8p + T8s;
|
Chris@10
|
3885 Ta1 = T8i + T87;
|
Chris@10
|
3886 Ta2 = Ta0 - Ta1;
|
Chris@10
|
3887 Tao = Ta0 + Ta1;
|
Chris@10
|
3888 Ta3 = T7K + T7V;
|
Chris@10
|
3889 Ta4 = T8u + T8v;
|
Chris@10
|
3890 Ta5 = Ta3 - Ta4;
|
Chris@10
|
3891 Tap = Ta3 + Ta4;
|
Chris@10
|
3892 }
|
Chris@10
|
3893 Ta6 = FMA(KP471396736, Ta2, KP881921264 * Ta5);
|
Chris@10
|
3894 Taw = FNMS(KP290284677, Tao, KP956940335 * Tap);
|
Chris@10
|
3895 Tag = FNMS(KP881921264, Ta2, KP471396736 * Ta5);
|
Chris@10
|
3896 Taq = FMA(KP956940335, Tao, KP290284677 * Tap);
|
Chris@10
|
3897 }
|
Chris@10
|
3898 {
|
Chris@10
|
3899 E Ta9, Tar, Tac, Tas;
|
Chris@10
|
3900 {
|
Chris@10
|
3901 E Ta7, Ta8, Taa, Tab;
|
Chris@10
|
3902 Ta7 = T8D + T8O;
|
Chris@10
|
3903 Ta8 = T9o + T9n;
|
Chris@10
|
3904 Ta9 = Ta7 - Ta8;
|
Chris@10
|
3905 Tar = Ta7 + Ta8;
|
Chris@10
|
3906 Taa = T9i + T9l;
|
Chris@10
|
3907 Tab = T90 + T9b;
|
Chris@10
|
3908 Tac = Taa - Tab;
|
Chris@10
|
3909 Tas = Taa + Tab;
|
Chris@10
|
3910 }
|
Chris@10
|
3911 Tad = FNMS(KP881921264, Tac, KP471396736 * Ta9);
|
Chris@10
|
3912 Tax = FMA(KP290284677, Tar, KP956940335 * Tas);
|
Chris@10
|
3913 Tah = FMA(KP881921264, Ta9, KP471396736 * Tac);
|
Chris@10
|
3914 Tat = FNMS(KP290284677, Tas, KP956940335 * Tar);
|
Chris@10
|
3915 }
|
Chris@10
|
3916 {
|
Chris@10
|
3917 E T9Z, Tae, Tit, Tiw;
|
Chris@10
|
3918 T9Z = T9R + T9Y;
|
Chris@10
|
3919 Tae = Ta6 + Tad;
|
Chris@10
|
3920 ri[WS(rs, 43)] = T9Z - Tae;
|
Chris@10
|
3921 ri[WS(rs, 11)] = T9Z + Tae;
|
Chris@10
|
3922 Tit = Tag + Tah;
|
Chris@10
|
3923 Tiw = Tiu + Tiv;
|
Chris@10
|
3924 ii[WS(rs, 11)] = Tit + Tiw;
|
Chris@10
|
3925 ii[WS(rs, 43)] = Tiw - Tit;
|
Chris@10
|
3926 }
|
Chris@10
|
3927 {
|
Chris@10
|
3928 E Taf, Tai, Tix, Tiy;
|
Chris@10
|
3929 Taf = T9R - T9Y;
|
Chris@10
|
3930 Tai = Tag - Tah;
|
Chris@10
|
3931 ri[WS(rs, 59)] = Taf - Tai;
|
Chris@10
|
3932 ri[WS(rs, 27)] = Taf + Tai;
|
Chris@10
|
3933 Tix = Tad - Ta6;
|
Chris@10
|
3934 Tiy = Tiv - Tiu;
|
Chris@10
|
3935 ii[WS(rs, 27)] = Tix + Tiy;
|
Chris@10
|
3936 ii[WS(rs, 59)] = Tiy - Tix;
|
Chris@10
|
3937 }
|
Chris@10
|
3938 {
|
Chris@10
|
3939 E Tan, Tau, Tij, Tiq;
|
Chris@10
|
3940 Tan = Taj + Tam;
|
Chris@10
|
3941 Tau = Taq + Tat;
|
Chris@10
|
3942 ri[WS(rs, 35)] = Tan - Tau;
|
Chris@10
|
3943 ri[WS(rs, 3)] = Tan + Tau;
|
Chris@10
|
3944 Tij = Taw + Tax;
|
Chris@10
|
3945 Tiq = Tik + Tip;
|
Chris@10
|
3946 ii[WS(rs, 3)] = Tij + Tiq;
|
Chris@10
|
3947 ii[WS(rs, 35)] = Tiq - Tij;
|
Chris@10
|
3948 }
|
Chris@10
|
3949 {
|
Chris@10
|
3950 E Tav, Tay, Tir, Tis;
|
Chris@10
|
3951 Tav = Taj - Tam;
|
Chris@10
|
3952 Tay = Taw - Tax;
|
Chris@10
|
3953 ri[WS(rs, 51)] = Tav - Tay;
|
Chris@10
|
3954 ri[WS(rs, 19)] = Tav + Tay;
|
Chris@10
|
3955 Tir = Tat - Taq;
|
Chris@10
|
3956 Tis = Tip - Tik;
|
Chris@10
|
3957 ii[WS(rs, 19)] = Tir + Tis;
|
Chris@10
|
3958 ii[WS(rs, 51)] = Tis - Tir;
|
Chris@10
|
3959 }
|
Chris@10
|
3960 }
|
Chris@10
|
3961 }
|
Chris@10
|
3962 }
|
Chris@10
|
3963 }
|
Chris@10
|
3964
|
Chris@10
|
3965 static const tw_instr twinstr[] = {
|
Chris@10
|
3966 {TW_FULL, 0, 64},
|
Chris@10
|
3967 {TW_NEXT, 1, 0}
|
Chris@10
|
3968 };
|
Chris@10
|
3969
|
Chris@10
|
3970 static const ct_desc desc = { 64, "t1_64", twinstr, &GENUS, {808, 270, 230, 0}, 0, 0, 0 };
|
Chris@10
|
3971
|
Chris@10
|
3972 void X(codelet_t1_64) (planner *p) {
|
Chris@10
|
3973 X(kdft_dit_register) (p, t1_64, &desc);
|
Chris@10
|
3974 }
|
Chris@10
|
3975 #endif /* HAVE_FMA */
|