Chris@82
|
1 /*
|
Chris@82
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
Chris@82
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
Chris@82
|
4 *
|
Chris@82
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@82
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@82
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@82
|
8 * (at your option) any later version.
|
Chris@82
|
9 *
|
Chris@82
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@82
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@82
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@82
|
13 * GNU General Public License for more details.
|
Chris@82
|
14 *
|
Chris@82
|
15 * You should have received a copy of the GNU General Public License
|
Chris@82
|
16 * along with this program; if not, write to the Free Software
|
Chris@82
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@82
|
18 *
|
Chris@82
|
19 */
|
Chris@82
|
20
|
Chris@82
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@82
|
22 /* Generated on Thu May 24 08:04:15 EDT 2018 */
|
Chris@82
|
23
|
Chris@82
|
24 #include "dft/codelet-dft.h"
|
Chris@82
|
25
|
Chris@82
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
Chris@82
|
27
|
Chris@82
|
28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -compact -variables 4 -pipeline-latency 4 -n 64 -name t1_64 -include dft/scalar/t.h */
|
Chris@82
|
29
|
Chris@82
|
30 /*
|
Chris@82
|
31 * This function contains 1038 FP additions, 644 FP multiplications,
|
Chris@82
|
32 * (or, 520 additions, 126 multiplications, 518 fused multiply/add),
|
Chris@82
|
33 * 190 stack variables, 15 constants, and 256 memory accesses
|
Chris@82
|
34 */
|
Chris@82
|
35 #include "dft/scalar/t.h"
|
Chris@82
|
36
|
Chris@82
|
37 static void t1_64(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@82
|
38 {
|
Chris@82
|
39 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@82
|
40 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@82
|
41 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@82
|
42 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@82
|
43 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
Chris@82
|
44 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
Chris@82
|
45 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
Chris@82
|
46 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
Chris@82
|
47 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@82
|
48 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@82
|
49 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@82
|
50 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@82
|
51 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@82
|
52 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@82
|
53 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@82
|
54 {
|
Chris@82
|
55 INT m;
|
Chris@82
|
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@82
|
57 E Tm, TeM, TjR, Tkl, T7e, TcA, TiV, Tjm, T1G, TeW, TeZ, Ths, T7Q, TcJ, T7X;
|
Chris@82
|
58 E TcI, T29, Tf8, Tf5, Thv, T87, TcN, T8u, TcQ, T5K, Tg9, TfU, ThS, Taq, Tdm;
|
Chris@82
|
59 E Tbj, Tdx, TN, Tjl, TeP, TiP, T7l, TcB, T7s, TcC, T1f, TeR, TeU, Thr, T7B;
|
Chris@82
|
60 E TcG, T7I, TcF, T32, Tfj, Tfg, ThB, T8G, TcU, T93, TcX, T3X, TfI, Tft, ThH;
|
Chris@82
|
61 E T9h, Td3, Taa, Tde, T2A, Tf6, Tfb, Thw, T8m, TcR, T8x, TcO, T3t, Tfh, Tfm;
|
Chris@82
|
62 E ThC, T8V, TcY, T96, TcV, T4o, Tfu, TfL, ThI, T9w, Tdf, Tad, Td4, T6b, TfV;
|
Chris@82
|
63 E Tgc, ThT, TaF, Tdy, Tbm, Tdn, T4Q, ThN, TfA, TfN, Ta1, Tdh, Taf, Td8, T5h;
|
Chris@82
|
64 E ThO, TfF, TfO, T9M, Tdi, Tag, Tdb, T6D, ThY, Tg1, Tge, Tba, TdA, Tbo, Tdr;
|
Chris@82
|
65 E T74, ThZ, Tg6, Tgf, TaV, TdB, Tbp, Tdu;
|
Chris@82
|
66 {
|
Chris@82
|
67 E T1, TiT, T7, TiS, Te, T7a, Tk, T7c;
|
Chris@82
|
68 T1 = ri[0];
|
Chris@82
|
69 TiT = ii[0];
|
Chris@82
|
70 {
|
Chris@82
|
71 E T3, T6, T4, TiR, T2, T5;
|
Chris@82
|
72 T3 = ri[WS(rs, 32)];
|
Chris@82
|
73 T6 = ii[WS(rs, 32)];
|
Chris@82
|
74 T2 = W[62];
|
Chris@82
|
75 T4 = T2 * T3;
|
Chris@82
|
76 TiR = T2 * T6;
|
Chris@82
|
77 T5 = W[63];
|
Chris@82
|
78 T7 = FMA(T5, T6, T4);
|
Chris@82
|
79 TiS = FNMS(T5, T3, TiR);
|
Chris@82
|
80 }
|
Chris@82
|
81 {
|
Chris@82
|
82 E Ta, Td, Tb, T79, T9, Tc;
|
Chris@82
|
83 Ta = ri[WS(rs, 16)];
|
Chris@82
|
84 Td = ii[WS(rs, 16)];
|
Chris@82
|
85 T9 = W[30];
|
Chris@82
|
86 Tb = T9 * Ta;
|
Chris@82
|
87 T79 = T9 * Td;
|
Chris@82
|
88 Tc = W[31];
|
Chris@82
|
89 Te = FMA(Tc, Td, Tb);
|
Chris@82
|
90 T7a = FNMS(Tc, Ta, T79);
|
Chris@82
|
91 }
|
Chris@82
|
92 {
|
Chris@82
|
93 E Tg, Tj, Th, T7b, Tf, Ti;
|
Chris@82
|
94 Tg = ri[WS(rs, 48)];
|
Chris@82
|
95 Tj = ii[WS(rs, 48)];
|
Chris@82
|
96 Tf = W[94];
|
Chris@82
|
97 Th = Tf * Tg;
|
Chris@82
|
98 T7b = Tf * Tj;
|
Chris@82
|
99 Ti = W[95];
|
Chris@82
|
100 Tk = FMA(Ti, Tj, Th);
|
Chris@82
|
101 T7c = FNMS(Ti, Tg, T7b);
|
Chris@82
|
102 }
|
Chris@82
|
103 {
|
Chris@82
|
104 E T8, Tl, TjP, TjQ;
|
Chris@82
|
105 T8 = T1 + T7;
|
Chris@82
|
106 Tl = Te + Tk;
|
Chris@82
|
107 Tm = T8 + Tl;
|
Chris@82
|
108 TeM = T8 - Tl;
|
Chris@82
|
109 TjP = TiT - TiS;
|
Chris@82
|
110 TjQ = Te - Tk;
|
Chris@82
|
111 TjR = TjP - TjQ;
|
Chris@82
|
112 Tkl = TjQ + TjP;
|
Chris@82
|
113 }
|
Chris@82
|
114 {
|
Chris@82
|
115 E T78, T7d, TiQ, TiU;
|
Chris@82
|
116 T78 = T1 - T7;
|
Chris@82
|
117 T7d = T7a - T7c;
|
Chris@82
|
118 T7e = T78 - T7d;
|
Chris@82
|
119 TcA = T78 + T7d;
|
Chris@82
|
120 TiQ = T7a + T7c;
|
Chris@82
|
121 TiU = TiS + TiT;
|
Chris@82
|
122 TiV = TiQ + TiU;
|
Chris@82
|
123 Tjm = TiU - TiQ;
|
Chris@82
|
124 }
|
Chris@82
|
125 }
|
Chris@82
|
126 {
|
Chris@82
|
127 E T1l, T7L, T1E, T7V, T1r, T7N, T1y, T7T;
|
Chris@82
|
128 {
|
Chris@82
|
129 E T1h, T1k, T1i, T7K, T1g, T1j;
|
Chris@82
|
130 T1h = ri[WS(rs, 60)];
|
Chris@82
|
131 T1k = ii[WS(rs, 60)];
|
Chris@82
|
132 T1g = W[118];
|
Chris@82
|
133 T1i = T1g * T1h;
|
Chris@82
|
134 T7K = T1g * T1k;
|
Chris@82
|
135 T1j = W[119];
|
Chris@82
|
136 T1l = FMA(T1j, T1k, T1i);
|
Chris@82
|
137 T7L = FNMS(T1j, T1h, T7K);
|
Chris@82
|
138 }
|
Chris@82
|
139 {
|
Chris@82
|
140 E T1A, T1D, T1B, T7U, T1z, T1C;
|
Chris@82
|
141 T1A = ri[WS(rs, 44)];
|
Chris@82
|
142 T1D = ii[WS(rs, 44)];
|
Chris@82
|
143 T1z = W[86];
|
Chris@82
|
144 T1B = T1z * T1A;
|
Chris@82
|
145 T7U = T1z * T1D;
|
Chris@82
|
146 T1C = W[87];
|
Chris@82
|
147 T1E = FMA(T1C, T1D, T1B);
|
Chris@82
|
148 T7V = FNMS(T1C, T1A, T7U);
|
Chris@82
|
149 }
|
Chris@82
|
150 {
|
Chris@82
|
151 E T1n, T1q, T1o, T7M, T1m, T1p;
|
Chris@82
|
152 T1n = ri[WS(rs, 28)];
|
Chris@82
|
153 T1q = ii[WS(rs, 28)];
|
Chris@82
|
154 T1m = W[54];
|
Chris@82
|
155 T1o = T1m * T1n;
|
Chris@82
|
156 T7M = T1m * T1q;
|
Chris@82
|
157 T1p = W[55];
|
Chris@82
|
158 T1r = FMA(T1p, T1q, T1o);
|
Chris@82
|
159 T7N = FNMS(T1p, T1n, T7M);
|
Chris@82
|
160 }
|
Chris@82
|
161 {
|
Chris@82
|
162 E T1u, T1x, T1v, T7S, T1t, T1w;
|
Chris@82
|
163 T1u = ri[WS(rs, 12)];
|
Chris@82
|
164 T1x = ii[WS(rs, 12)];
|
Chris@82
|
165 T1t = W[22];
|
Chris@82
|
166 T1v = T1t * T1u;
|
Chris@82
|
167 T7S = T1t * T1x;
|
Chris@82
|
168 T1w = W[23];
|
Chris@82
|
169 T1y = FMA(T1w, T1x, T1v);
|
Chris@82
|
170 T7T = FNMS(T1w, T1u, T7S);
|
Chris@82
|
171 }
|
Chris@82
|
172 {
|
Chris@82
|
173 E T1s, T1F, TeX, TeY;
|
Chris@82
|
174 T1s = T1l + T1r;
|
Chris@82
|
175 T1F = T1y + T1E;
|
Chris@82
|
176 T1G = T1s + T1F;
|
Chris@82
|
177 TeW = T1s - T1F;
|
Chris@82
|
178 TeX = T7L + T7N;
|
Chris@82
|
179 TeY = T7T + T7V;
|
Chris@82
|
180 TeZ = TeX - TeY;
|
Chris@82
|
181 Ths = TeX + TeY;
|
Chris@82
|
182 }
|
Chris@82
|
183 {
|
Chris@82
|
184 E T7O, T7P, T7R, T7W;
|
Chris@82
|
185 T7O = T7L - T7N;
|
Chris@82
|
186 T7P = T1y - T1E;
|
Chris@82
|
187 T7Q = T7O + T7P;
|
Chris@82
|
188 TcJ = T7O - T7P;
|
Chris@82
|
189 T7R = T1l - T1r;
|
Chris@82
|
190 T7W = T7T - T7V;
|
Chris@82
|
191 T7X = T7R - T7W;
|
Chris@82
|
192 TcI = T7R + T7W;
|
Chris@82
|
193 }
|
Chris@82
|
194 }
|
Chris@82
|
195 {
|
Chris@82
|
196 E T1O, T82, T27, T8s, T1U, T84, T21, T8q;
|
Chris@82
|
197 {
|
Chris@82
|
198 E T1K, T1N, T1L, T81, T1J, T1M;
|
Chris@82
|
199 T1K = ri[WS(rs, 2)];
|
Chris@82
|
200 T1N = ii[WS(rs, 2)];
|
Chris@82
|
201 T1J = W[2];
|
Chris@82
|
202 T1L = T1J * T1K;
|
Chris@82
|
203 T81 = T1J * T1N;
|
Chris@82
|
204 T1M = W[3];
|
Chris@82
|
205 T1O = FMA(T1M, T1N, T1L);
|
Chris@82
|
206 T82 = FNMS(T1M, T1K, T81);
|
Chris@82
|
207 }
|
Chris@82
|
208 {
|
Chris@82
|
209 E T23, T26, T24, T8r, T22, T25;
|
Chris@82
|
210 T23 = ri[WS(rs, 50)];
|
Chris@82
|
211 T26 = ii[WS(rs, 50)];
|
Chris@82
|
212 T22 = W[98];
|
Chris@82
|
213 T24 = T22 * T23;
|
Chris@82
|
214 T8r = T22 * T26;
|
Chris@82
|
215 T25 = W[99];
|
Chris@82
|
216 T27 = FMA(T25, T26, T24);
|
Chris@82
|
217 T8s = FNMS(T25, T23, T8r);
|
Chris@82
|
218 }
|
Chris@82
|
219 {
|
Chris@82
|
220 E T1Q, T1T, T1R, T83, T1P, T1S;
|
Chris@82
|
221 T1Q = ri[WS(rs, 34)];
|
Chris@82
|
222 T1T = ii[WS(rs, 34)];
|
Chris@82
|
223 T1P = W[66];
|
Chris@82
|
224 T1R = T1P * T1Q;
|
Chris@82
|
225 T83 = T1P * T1T;
|
Chris@82
|
226 T1S = W[67];
|
Chris@82
|
227 T1U = FMA(T1S, T1T, T1R);
|
Chris@82
|
228 T84 = FNMS(T1S, T1Q, T83);
|
Chris@82
|
229 }
|
Chris@82
|
230 {
|
Chris@82
|
231 E T1X, T20, T1Y, T8p, T1W, T1Z;
|
Chris@82
|
232 T1X = ri[WS(rs, 18)];
|
Chris@82
|
233 T20 = ii[WS(rs, 18)];
|
Chris@82
|
234 T1W = W[34];
|
Chris@82
|
235 T1Y = T1W * T1X;
|
Chris@82
|
236 T8p = T1W * T20;
|
Chris@82
|
237 T1Z = W[35];
|
Chris@82
|
238 T21 = FMA(T1Z, T20, T1Y);
|
Chris@82
|
239 T8q = FNMS(T1Z, T1X, T8p);
|
Chris@82
|
240 }
|
Chris@82
|
241 {
|
Chris@82
|
242 E T1V, T28, Tf3, Tf4;
|
Chris@82
|
243 T1V = T1O + T1U;
|
Chris@82
|
244 T28 = T21 + T27;
|
Chris@82
|
245 T29 = T1V + T28;
|
Chris@82
|
246 Tf8 = T1V - T28;
|
Chris@82
|
247 Tf3 = T82 + T84;
|
Chris@82
|
248 Tf4 = T8q + T8s;
|
Chris@82
|
249 Tf5 = Tf3 - Tf4;
|
Chris@82
|
250 Thv = Tf3 + Tf4;
|
Chris@82
|
251 }
|
Chris@82
|
252 {
|
Chris@82
|
253 E T85, T86, T8o, T8t;
|
Chris@82
|
254 T85 = T82 - T84;
|
Chris@82
|
255 T86 = T21 - T27;
|
Chris@82
|
256 T87 = T85 + T86;
|
Chris@82
|
257 TcN = T85 - T86;
|
Chris@82
|
258 T8o = T1O - T1U;
|
Chris@82
|
259 T8t = T8q - T8s;
|
Chris@82
|
260 T8u = T8o - T8t;
|
Chris@82
|
261 TcQ = T8o + T8t;
|
Chris@82
|
262 }
|
Chris@82
|
263 }
|
Chris@82
|
264 {
|
Chris@82
|
265 E T5p, Tal, T5I, Tbh, T5v, Tan, T5C, Tbf;
|
Chris@82
|
266 {
|
Chris@82
|
267 E T5l, T5o, T5m, Tak, T5k, T5n;
|
Chris@82
|
268 T5l = ri[WS(rs, 63)];
|
Chris@82
|
269 T5o = ii[WS(rs, 63)];
|
Chris@82
|
270 T5k = W[124];
|
Chris@82
|
271 T5m = T5k * T5l;
|
Chris@82
|
272 Tak = T5k * T5o;
|
Chris@82
|
273 T5n = W[125];
|
Chris@82
|
274 T5p = FMA(T5n, T5o, T5m);
|
Chris@82
|
275 Tal = FNMS(T5n, T5l, Tak);
|
Chris@82
|
276 }
|
Chris@82
|
277 {
|
Chris@82
|
278 E T5E, T5H, T5F, Tbg, T5D, T5G;
|
Chris@82
|
279 T5E = ri[WS(rs, 47)];
|
Chris@82
|
280 T5H = ii[WS(rs, 47)];
|
Chris@82
|
281 T5D = W[92];
|
Chris@82
|
282 T5F = T5D * T5E;
|
Chris@82
|
283 Tbg = T5D * T5H;
|
Chris@82
|
284 T5G = W[93];
|
Chris@82
|
285 T5I = FMA(T5G, T5H, T5F);
|
Chris@82
|
286 Tbh = FNMS(T5G, T5E, Tbg);
|
Chris@82
|
287 }
|
Chris@82
|
288 {
|
Chris@82
|
289 E T5r, T5u, T5s, Tam, T5q, T5t;
|
Chris@82
|
290 T5r = ri[WS(rs, 31)];
|
Chris@82
|
291 T5u = ii[WS(rs, 31)];
|
Chris@82
|
292 T5q = W[60];
|
Chris@82
|
293 T5s = T5q * T5r;
|
Chris@82
|
294 Tam = T5q * T5u;
|
Chris@82
|
295 T5t = W[61];
|
Chris@82
|
296 T5v = FMA(T5t, T5u, T5s);
|
Chris@82
|
297 Tan = FNMS(T5t, T5r, Tam);
|
Chris@82
|
298 }
|
Chris@82
|
299 {
|
Chris@82
|
300 E T5y, T5B, T5z, Tbe, T5x, T5A;
|
Chris@82
|
301 T5y = ri[WS(rs, 15)];
|
Chris@82
|
302 T5B = ii[WS(rs, 15)];
|
Chris@82
|
303 T5x = W[28];
|
Chris@82
|
304 T5z = T5x * T5y;
|
Chris@82
|
305 Tbe = T5x * T5B;
|
Chris@82
|
306 T5A = W[29];
|
Chris@82
|
307 T5C = FMA(T5A, T5B, T5z);
|
Chris@82
|
308 Tbf = FNMS(T5A, T5y, Tbe);
|
Chris@82
|
309 }
|
Chris@82
|
310 {
|
Chris@82
|
311 E T5w, T5J, TfS, TfT;
|
Chris@82
|
312 T5w = T5p + T5v;
|
Chris@82
|
313 T5J = T5C + T5I;
|
Chris@82
|
314 T5K = T5w + T5J;
|
Chris@82
|
315 Tg9 = T5w - T5J;
|
Chris@82
|
316 TfS = Tal + Tan;
|
Chris@82
|
317 TfT = Tbf + Tbh;
|
Chris@82
|
318 TfU = TfS - TfT;
|
Chris@82
|
319 ThS = TfS + TfT;
|
Chris@82
|
320 }
|
Chris@82
|
321 {
|
Chris@82
|
322 E Tao, Tap, Tbd, Tbi;
|
Chris@82
|
323 Tao = Tal - Tan;
|
Chris@82
|
324 Tap = T5C - T5I;
|
Chris@82
|
325 Taq = Tao + Tap;
|
Chris@82
|
326 Tdm = Tao - Tap;
|
Chris@82
|
327 Tbd = T5p - T5v;
|
Chris@82
|
328 Tbi = Tbf - Tbh;
|
Chris@82
|
329 Tbj = Tbd - Tbi;
|
Chris@82
|
330 Tdx = Tbd + Tbi;
|
Chris@82
|
331 }
|
Chris@82
|
332 }
|
Chris@82
|
333 {
|
Chris@82
|
334 E Ts, T7g, TL, T7q, Ty, T7i, TF, T7o;
|
Chris@82
|
335 {
|
Chris@82
|
336 E To, Tr, Tp, T7f, Tn, Tq;
|
Chris@82
|
337 To = ri[WS(rs, 8)];
|
Chris@82
|
338 Tr = ii[WS(rs, 8)];
|
Chris@82
|
339 Tn = W[14];
|
Chris@82
|
340 Tp = Tn * To;
|
Chris@82
|
341 T7f = Tn * Tr;
|
Chris@82
|
342 Tq = W[15];
|
Chris@82
|
343 Ts = FMA(Tq, Tr, Tp);
|
Chris@82
|
344 T7g = FNMS(Tq, To, T7f);
|
Chris@82
|
345 }
|
Chris@82
|
346 {
|
Chris@82
|
347 E TH, TK, TI, T7p, TG, TJ;
|
Chris@82
|
348 TH = ri[WS(rs, 24)];
|
Chris@82
|
349 TK = ii[WS(rs, 24)];
|
Chris@82
|
350 TG = W[46];
|
Chris@82
|
351 TI = TG * TH;
|
Chris@82
|
352 T7p = TG * TK;
|
Chris@82
|
353 TJ = W[47];
|
Chris@82
|
354 TL = FMA(TJ, TK, TI);
|
Chris@82
|
355 T7q = FNMS(TJ, TH, T7p);
|
Chris@82
|
356 }
|
Chris@82
|
357 {
|
Chris@82
|
358 E Tu, Tx, Tv, T7h, Tt, Tw;
|
Chris@82
|
359 Tu = ri[WS(rs, 40)];
|
Chris@82
|
360 Tx = ii[WS(rs, 40)];
|
Chris@82
|
361 Tt = W[78];
|
Chris@82
|
362 Tv = Tt * Tu;
|
Chris@82
|
363 T7h = Tt * Tx;
|
Chris@82
|
364 Tw = W[79];
|
Chris@82
|
365 Ty = FMA(Tw, Tx, Tv);
|
Chris@82
|
366 T7i = FNMS(Tw, Tu, T7h);
|
Chris@82
|
367 }
|
Chris@82
|
368 {
|
Chris@82
|
369 E TB, TE, TC, T7n, TA, TD;
|
Chris@82
|
370 TB = ri[WS(rs, 56)];
|
Chris@82
|
371 TE = ii[WS(rs, 56)];
|
Chris@82
|
372 TA = W[110];
|
Chris@82
|
373 TC = TA * TB;
|
Chris@82
|
374 T7n = TA * TE;
|
Chris@82
|
375 TD = W[111];
|
Chris@82
|
376 TF = FMA(TD, TE, TC);
|
Chris@82
|
377 T7o = FNMS(TD, TB, T7n);
|
Chris@82
|
378 }
|
Chris@82
|
379 {
|
Chris@82
|
380 E Tz, TM, TeN, TeO;
|
Chris@82
|
381 Tz = Ts + Ty;
|
Chris@82
|
382 TM = TF + TL;
|
Chris@82
|
383 TN = Tz + TM;
|
Chris@82
|
384 Tjl = TM - Tz;
|
Chris@82
|
385 TeN = T7g + T7i;
|
Chris@82
|
386 TeO = T7o + T7q;
|
Chris@82
|
387 TeP = TeN - TeO;
|
Chris@82
|
388 TiP = TeN + TeO;
|
Chris@82
|
389 }
|
Chris@82
|
390 {
|
Chris@82
|
391 E T7j, T7k, T7m, T7r;
|
Chris@82
|
392 T7j = T7g - T7i;
|
Chris@82
|
393 T7k = Ts - Ty;
|
Chris@82
|
394 T7l = T7j - T7k;
|
Chris@82
|
395 TcB = T7k + T7j;
|
Chris@82
|
396 T7m = TF - TL;
|
Chris@82
|
397 T7r = T7o - T7q;
|
Chris@82
|
398 T7s = T7m + T7r;
|
Chris@82
|
399 TcC = T7m - T7r;
|
Chris@82
|
400 }
|
Chris@82
|
401 }
|
Chris@82
|
402 {
|
Chris@82
|
403 E TU, T7w, T1d, T7G, T10, T7y, T17, T7E;
|
Chris@82
|
404 {
|
Chris@82
|
405 E TQ, TT, TR, T7v, TP, TS;
|
Chris@82
|
406 TQ = ri[WS(rs, 4)];
|
Chris@82
|
407 TT = ii[WS(rs, 4)];
|
Chris@82
|
408 TP = W[6];
|
Chris@82
|
409 TR = TP * TQ;
|
Chris@82
|
410 T7v = TP * TT;
|
Chris@82
|
411 TS = W[7];
|
Chris@82
|
412 TU = FMA(TS, TT, TR);
|
Chris@82
|
413 T7w = FNMS(TS, TQ, T7v);
|
Chris@82
|
414 }
|
Chris@82
|
415 {
|
Chris@82
|
416 E T19, T1c, T1a, T7F, T18, T1b;
|
Chris@82
|
417 T19 = ri[WS(rs, 52)];
|
Chris@82
|
418 T1c = ii[WS(rs, 52)];
|
Chris@82
|
419 T18 = W[102];
|
Chris@82
|
420 T1a = T18 * T19;
|
Chris@82
|
421 T7F = T18 * T1c;
|
Chris@82
|
422 T1b = W[103];
|
Chris@82
|
423 T1d = FMA(T1b, T1c, T1a);
|
Chris@82
|
424 T7G = FNMS(T1b, T19, T7F);
|
Chris@82
|
425 }
|
Chris@82
|
426 {
|
Chris@82
|
427 E TW, TZ, TX, T7x, TV, TY;
|
Chris@82
|
428 TW = ri[WS(rs, 36)];
|
Chris@82
|
429 TZ = ii[WS(rs, 36)];
|
Chris@82
|
430 TV = W[70];
|
Chris@82
|
431 TX = TV * TW;
|
Chris@82
|
432 T7x = TV * TZ;
|
Chris@82
|
433 TY = W[71];
|
Chris@82
|
434 T10 = FMA(TY, TZ, TX);
|
Chris@82
|
435 T7y = FNMS(TY, TW, T7x);
|
Chris@82
|
436 }
|
Chris@82
|
437 {
|
Chris@82
|
438 E T13, T16, T14, T7D, T12, T15;
|
Chris@82
|
439 T13 = ri[WS(rs, 20)];
|
Chris@82
|
440 T16 = ii[WS(rs, 20)];
|
Chris@82
|
441 T12 = W[38];
|
Chris@82
|
442 T14 = T12 * T13;
|
Chris@82
|
443 T7D = T12 * T16;
|
Chris@82
|
444 T15 = W[39];
|
Chris@82
|
445 T17 = FMA(T15, T16, T14);
|
Chris@82
|
446 T7E = FNMS(T15, T13, T7D);
|
Chris@82
|
447 }
|
Chris@82
|
448 {
|
Chris@82
|
449 E T11, T1e, TeS, TeT;
|
Chris@82
|
450 T11 = TU + T10;
|
Chris@82
|
451 T1e = T17 + T1d;
|
Chris@82
|
452 T1f = T11 + T1e;
|
Chris@82
|
453 TeR = T11 - T1e;
|
Chris@82
|
454 TeS = T7w + T7y;
|
Chris@82
|
455 TeT = T7E + T7G;
|
Chris@82
|
456 TeU = TeS - TeT;
|
Chris@82
|
457 Thr = TeS + TeT;
|
Chris@82
|
458 }
|
Chris@82
|
459 {
|
Chris@82
|
460 E T7z, T7A, T7C, T7H;
|
Chris@82
|
461 T7z = T7w - T7y;
|
Chris@82
|
462 T7A = T17 - T1d;
|
Chris@82
|
463 T7B = T7z + T7A;
|
Chris@82
|
464 TcG = T7z - T7A;
|
Chris@82
|
465 T7C = TU - T10;
|
Chris@82
|
466 T7H = T7E - T7G;
|
Chris@82
|
467 T7I = T7C - T7H;
|
Chris@82
|
468 TcF = T7C + T7H;
|
Chris@82
|
469 }
|
Chris@82
|
470 }
|
Chris@82
|
471 {
|
Chris@82
|
472 E T2H, T8B, T30, T91, T2N, T8D, T2U, T8Z;
|
Chris@82
|
473 {
|
Chris@82
|
474 E T2D, T2G, T2E, T8A, T2C, T2F;
|
Chris@82
|
475 T2D = ri[WS(rs, 62)];
|
Chris@82
|
476 T2G = ii[WS(rs, 62)];
|
Chris@82
|
477 T2C = W[122];
|
Chris@82
|
478 T2E = T2C * T2D;
|
Chris@82
|
479 T8A = T2C * T2G;
|
Chris@82
|
480 T2F = W[123];
|
Chris@82
|
481 T2H = FMA(T2F, T2G, T2E);
|
Chris@82
|
482 T8B = FNMS(T2F, T2D, T8A);
|
Chris@82
|
483 }
|
Chris@82
|
484 {
|
Chris@82
|
485 E T2W, T2Z, T2X, T90, T2V, T2Y;
|
Chris@82
|
486 T2W = ri[WS(rs, 46)];
|
Chris@82
|
487 T2Z = ii[WS(rs, 46)];
|
Chris@82
|
488 T2V = W[90];
|
Chris@82
|
489 T2X = T2V * T2W;
|
Chris@82
|
490 T90 = T2V * T2Z;
|
Chris@82
|
491 T2Y = W[91];
|
Chris@82
|
492 T30 = FMA(T2Y, T2Z, T2X);
|
Chris@82
|
493 T91 = FNMS(T2Y, T2W, T90);
|
Chris@82
|
494 }
|
Chris@82
|
495 {
|
Chris@82
|
496 E T2J, T2M, T2K, T8C, T2I, T2L;
|
Chris@82
|
497 T2J = ri[WS(rs, 30)];
|
Chris@82
|
498 T2M = ii[WS(rs, 30)];
|
Chris@82
|
499 T2I = W[58];
|
Chris@82
|
500 T2K = T2I * T2J;
|
Chris@82
|
501 T8C = T2I * T2M;
|
Chris@82
|
502 T2L = W[59];
|
Chris@82
|
503 T2N = FMA(T2L, T2M, T2K);
|
Chris@82
|
504 T8D = FNMS(T2L, T2J, T8C);
|
Chris@82
|
505 }
|
Chris@82
|
506 {
|
Chris@82
|
507 E T2Q, T2T, T2R, T8Y, T2P, T2S;
|
Chris@82
|
508 T2Q = ri[WS(rs, 14)];
|
Chris@82
|
509 T2T = ii[WS(rs, 14)];
|
Chris@82
|
510 T2P = W[26];
|
Chris@82
|
511 T2R = T2P * T2Q;
|
Chris@82
|
512 T8Y = T2P * T2T;
|
Chris@82
|
513 T2S = W[27];
|
Chris@82
|
514 T2U = FMA(T2S, T2T, T2R);
|
Chris@82
|
515 T8Z = FNMS(T2S, T2Q, T8Y);
|
Chris@82
|
516 }
|
Chris@82
|
517 {
|
Chris@82
|
518 E T2O, T31, Tfe, Tff;
|
Chris@82
|
519 T2O = T2H + T2N;
|
Chris@82
|
520 T31 = T2U + T30;
|
Chris@82
|
521 T32 = T2O + T31;
|
Chris@82
|
522 Tfj = T2O - T31;
|
Chris@82
|
523 Tfe = T8B + T8D;
|
Chris@82
|
524 Tff = T8Z + T91;
|
Chris@82
|
525 Tfg = Tfe - Tff;
|
Chris@82
|
526 ThB = Tfe + Tff;
|
Chris@82
|
527 }
|
Chris@82
|
528 {
|
Chris@82
|
529 E T8E, T8F, T8X, T92;
|
Chris@82
|
530 T8E = T8B - T8D;
|
Chris@82
|
531 T8F = T2U - T30;
|
Chris@82
|
532 T8G = T8E + T8F;
|
Chris@82
|
533 TcU = T8E - T8F;
|
Chris@82
|
534 T8X = T2H - T2N;
|
Chris@82
|
535 T92 = T8Z - T91;
|
Chris@82
|
536 T93 = T8X - T92;
|
Chris@82
|
537 TcX = T8X + T92;
|
Chris@82
|
538 }
|
Chris@82
|
539 }
|
Chris@82
|
540 {
|
Chris@82
|
541 E T3C, T9c, T3V, Ta8, T3I, T9e, T3P, Ta6;
|
Chris@82
|
542 {
|
Chris@82
|
543 E T3y, T3B, T3z, T9b, T3x, T3A;
|
Chris@82
|
544 T3y = ri[WS(rs, 1)];
|
Chris@82
|
545 T3B = ii[WS(rs, 1)];
|
Chris@82
|
546 T3x = W[0];
|
Chris@82
|
547 T3z = T3x * T3y;
|
Chris@82
|
548 T9b = T3x * T3B;
|
Chris@82
|
549 T3A = W[1];
|
Chris@82
|
550 T3C = FMA(T3A, T3B, T3z);
|
Chris@82
|
551 T9c = FNMS(T3A, T3y, T9b);
|
Chris@82
|
552 }
|
Chris@82
|
553 {
|
Chris@82
|
554 E T3R, T3U, T3S, Ta7, T3Q, T3T;
|
Chris@82
|
555 T3R = ri[WS(rs, 49)];
|
Chris@82
|
556 T3U = ii[WS(rs, 49)];
|
Chris@82
|
557 T3Q = W[96];
|
Chris@82
|
558 T3S = T3Q * T3R;
|
Chris@82
|
559 Ta7 = T3Q * T3U;
|
Chris@82
|
560 T3T = W[97];
|
Chris@82
|
561 T3V = FMA(T3T, T3U, T3S);
|
Chris@82
|
562 Ta8 = FNMS(T3T, T3R, Ta7);
|
Chris@82
|
563 }
|
Chris@82
|
564 {
|
Chris@82
|
565 E T3E, T3H, T3F, T9d, T3D, T3G;
|
Chris@82
|
566 T3E = ri[WS(rs, 33)];
|
Chris@82
|
567 T3H = ii[WS(rs, 33)];
|
Chris@82
|
568 T3D = W[64];
|
Chris@82
|
569 T3F = T3D * T3E;
|
Chris@82
|
570 T9d = T3D * T3H;
|
Chris@82
|
571 T3G = W[65];
|
Chris@82
|
572 T3I = FMA(T3G, T3H, T3F);
|
Chris@82
|
573 T9e = FNMS(T3G, T3E, T9d);
|
Chris@82
|
574 }
|
Chris@82
|
575 {
|
Chris@82
|
576 E T3L, T3O, T3M, Ta5, T3K, T3N;
|
Chris@82
|
577 T3L = ri[WS(rs, 17)];
|
Chris@82
|
578 T3O = ii[WS(rs, 17)];
|
Chris@82
|
579 T3K = W[32];
|
Chris@82
|
580 T3M = T3K * T3L;
|
Chris@82
|
581 Ta5 = T3K * T3O;
|
Chris@82
|
582 T3N = W[33];
|
Chris@82
|
583 T3P = FMA(T3N, T3O, T3M);
|
Chris@82
|
584 Ta6 = FNMS(T3N, T3L, Ta5);
|
Chris@82
|
585 }
|
Chris@82
|
586 {
|
Chris@82
|
587 E T3J, T3W, Tfr, Tfs;
|
Chris@82
|
588 T3J = T3C + T3I;
|
Chris@82
|
589 T3W = T3P + T3V;
|
Chris@82
|
590 T3X = T3J + T3W;
|
Chris@82
|
591 TfI = T3J - T3W;
|
Chris@82
|
592 Tfr = T9c + T9e;
|
Chris@82
|
593 Tfs = Ta6 + Ta8;
|
Chris@82
|
594 Tft = Tfr - Tfs;
|
Chris@82
|
595 ThH = Tfr + Tfs;
|
Chris@82
|
596 }
|
Chris@82
|
597 {
|
Chris@82
|
598 E T9f, T9g, Ta4, Ta9;
|
Chris@82
|
599 T9f = T9c - T9e;
|
Chris@82
|
600 T9g = T3P - T3V;
|
Chris@82
|
601 T9h = T9f + T9g;
|
Chris@82
|
602 Td3 = T9f - T9g;
|
Chris@82
|
603 Ta4 = T3C - T3I;
|
Chris@82
|
604 Ta9 = Ta6 - Ta8;
|
Chris@82
|
605 Taa = Ta4 - Ta9;
|
Chris@82
|
606 Tde = Ta4 + Ta9;
|
Chris@82
|
607 }
|
Chris@82
|
608 }
|
Chris@82
|
609 {
|
Chris@82
|
610 E T2f, T8a, T2y, T8j, T2l, T8c, T2s, T8h;
|
Chris@82
|
611 {
|
Chris@82
|
612 E T2b, T2e, T2c, T89, T2a, T2d;
|
Chris@82
|
613 T2b = ri[WS(rs, 10)];
|
Chris@82
|
614 T2e = ii[WS(rs, 10)];
|
Chris@82
|
615 T2a = W[18];
|
Chris@82
|
616 T2c = T2a * T2b;
|
Chris@82
|
617 T89 = T2a * T2e;
|
Chris@82
|
618 T2d = W[19];
|
Chris@82
|
619 T2f = FMA(T2d, T2e, T2c);
|
Chris@82
|
620 T8a = FNMS(T2d, T2b, T89);
|
Chris@82
|
621 }
|
Chris@82
|
622 {
|
Chris@82
|
623 E T2u, T2x, T2v, T8i, T2t, T2w;
|
Chris@82
|
624 T2u = ri[WS(rs, 26)];
|
Chris@82
|
625 T2x = ii[WS(rs, 26)];
|
Chris@82
|
626 T2t = W[50];
|
Chris@82
|
627 T2v = T2t * T2u;
|
Chris@82
|
628 T8i = T2t * T2x;
|
Chris@82
|
629 T2w = W[51];
|
Chris@82
|
630 T2y = FMA(T2w, T2x, T2v);
|
Chris@82
|
631 T8j = FNMS(T2w, T2u, T8i);
|
Chris@82
|
632 }
|
Chris@82
|
633 {
|
Chris@82
|
634 E T2h, T2k, T2i, T8b, T2g, T2j;
|
Chris@82
|
635 T2h = ri[WS(rs, 42)];
|
Chris@82
|
636 T2k = ii[WS(rs, 42)];
|
Chris@82
|
637 T2g = W[82];
|
Chris@82
|
638 T2i = T2g * T2h;
|
Chris@82
|
639 T8b = T2g * T2k;
|
Chris@82
|
640 T2j = W[83];
|
Chris@82
|
641 T2l = FMA(T2j, T2k, T2i);
|
Chris@82
|
642 T8c = FNMS(T2j, T2h, T8b);
|
Chris@82
|
643 }
|
Chris@82
|
644 {
|
Chris@82
|
645 E T2o, T2r, T2p, T8g, T2n, T2q;
|
Chris@82
|
646 T2o = ri[WS(rs, 58)];
|
Chris@82
|
647 T2r = ii[WS(rs, 58)];
|
Chris@82
|
648 T2n = W[114];
|
Chris@82
|
649 T2p = T2n * T2o;
|
Chris@82
|
650 T8g = T2n * T2r;
|
Chris@82
|
651 T2q = W[115];
|
Chris@82
|
652 T2s = FMA(T2q, T2r, T2p);
|
Chris@82
|
653 T8h = FNMS(T2q, T2o, T8g);
|
Chris@82
|
654 }
|
Chris@82
|
655 {
|
Chris@82
|
656 E T2m, T2z, Tf9, Tfa;
|
Chris@82
|
657 T2m = T2f + T2l;
|
Chris@82
|
658 T2z = T2s + T2y;
|
Chris@82
|
659 T2A = T2m + T2z;
|
Chris@82
|
660 Tf6 = T2z - T2m;
|
Chris@82
|
661 Tf9 = T8a + T8c;
|
Chris@82
|
662 Tfa = T8h + T8j;
|
Chris@82
|
663 Tfb = Tf9 - Tfa;
|
Chris@82
|
664 Thw = Tf9 + Tfa;
|
Chris@82
|
665 {
|
Chris@82
|
666 E T8e, T8w, T8l, T8v;
|
Chris@82
|
667 {
|
Chris@82
|
668 E T88, T8d, T8f, T8k;
|
Chris@82
|
669 T88 = T2f - T2l;
|
Chris@82
|
670 T8d = T8a - T8c;
|
Chris@82
|
671 T8e = T88 + T8d;
|
Chris@82
|
672 T8w = T8d - T88;
|
Chris@82
|
673 T8f = T2s - T2y;
|
Chris@82
|
674 T8k = T8h - T8j;
|
Chris@82
|
675 T8l = T8f - T8k;
|
Chris@82
|
676 T8v = T8f + T8k;
|
Chris@82
|
677 }
|
Chris@82
|
678 T8m = T8e - T8l;
|
Chris@82
|
679 TcR = T8e + T8l;
|
Chris@82
|
680 T8x = T8v - T8w;
|
Chris@82
|
681 TcO = T8w + T8v;
|
Chris@82
|
682 }
|
Chris@82
|
683 }
|
Chris@82
|
684 }
|
Chris@82
|
685 {
|
Chris@82
|
686 E T38, T8J, T3r, T8S, T3e, T8L, T3l, T8Q;
|
Chris@82
|
687 {
|
Chris@82
|
688 E T34, T37, T35, T8I, T33, T36;
|
Chris@82
|
689 T34 = ri[WS(rs, 6)];
|
Chris@82
|
690 T37 = ii[WS(rs, 6)];
|
Chris@82
|
691 T33 = W[10];
|
Chris@82
|
692 T35 = T33 * T34;
|
Chris@82
|
693 T8I = T33 * T37;
|
Chris@82
|
694 T36 = W[11];
|
Chris@82
|
695 T38 = FMA(T36, T37, T35);
|
Chris@82
|
696 T8J = FNMS(T36, T34, T8I);
|
Chris@82
|
697 }
|
Chris@82
|
698 {
|
Chris@82
|
699 E T3n, T3q, T3o, T8R, T3m, T3p;
|
Chris@82
|
700 T3n = ri[WS(rs, 22)];
|
Chris@82
|
701 T3q = ii[WS(rs, 22)];
|
Chris@82
|
702 T3m = W[42];
|
Chris@82
|
703 T3o = T3m * T3n;
|
Chris@82
|
704 T8R = T3m * T3q;
|
Chris@82
|
705 T3p = W[43];
|
Chris@82
|
706 T3r = FMA(T3p, T3q, T3o);
|
Chris@82
|
707 T8S = FNMS(T3p, T3n, T8R);
|
Chris@82
|
708 }
|
Chris@82
|
709 {
|
Chris@82
|
710 E T3a, T3d, T3b, T8K, T39, T3c;
|
Chris@82
|
711 T3a = ri[WS(rs, 38)];
|
Chris@82
|
712 T3d = ii[WS(rs, 38)];
|
Chris@82
|
713 T39 = W[74];
|
Chris@82
|
714 T3b = T39 * T3a;
|
Chris@82
|
715 T8K = T39 * T3d;
|
Chris@82
|
716 T3c = W[75];
|
Chris@82
|
717 T3e = FMA(T3c, T3d, T3b);
|
Chris@82
|
718 T8L = FNMS(T3c, T3a, T8K);
|
Chris@82
|
719 }
|
Chris@82
|
720 {
|
Chris@82
|
721 E T3h, T3k, T3i, T8P, T3g, T3j;
|
Chris@82
|
722 T3h = ri[WS(rs, 54)];
|
Chris@82
|
723 T3k = ii[WS(rs, 54)];
|
Chris@82
|
724 T3g = W[106];
|
Chris@82
|
725 T3i = T3g * T3h;
|
Chris@82
|
726 T8P = T3g * T3k;
|
Chris@82
|
727 T3j = W[107];
|
Chris@82
|
728 T3l = FMA(T3j, T3k, T3i);
|
Chris@82
|
729 T8Q = FNMS(T3j, T3h, T8P);
|
Chris@82
|
730 }
|
Chris@82
|
731 {
|
Chris@82
|
732 E T3f, T3s, Tfk, Tfl;
|
Chris@82
|
733 T3f = T38 + T3e;
|
Chris@82
|
734 T3s = T3l + T3r;
|
Chris@82
|
735 T3t = T3f + T3s;
|
Chris@82
|
736 Tfh = T3s - T3f;
|
Chris@82
|
737 Tfk = T8J + T8L;
|
Chris@82
|
738 Tfl = T8Q + T8S;
|
Chris@82
|
739 Tfm = Tfk - Tfl;
|
Chris@82
|
740 ThC = Tfk + Tfl;
|
Chris@82
|
741 {
|
Chris@82
|
742 E T8N, T95, T8U, T94;
|
Chris@82
|
743 {
|
Chris@82
|
744 E T8H, T8M, T8O, T8T;
|
Chris@82
|
745 T8H = T38 - T3e;
|
Chris@82
|
746 T8M = T8J - T8L;
|
Chris@82
|
747 T8N = T8H + T8M;
|
Chris@82
|
748 T95 = T8M - T8H;
|
Chris@82
|
749 T8O = T3l - T3r;
|
Chris@82
|
750 T8T = T8Q - T8S;
|
Chris@82
|
751 T8U = T8O - T8T;
|
Chris@82
|
752 T94 = T8O + T8T;
|
Chris@82
|
753 }
|
Chris@82
|
754 T8V = T8N - T8U;
|
Chris@82
|
755 TcY = T8N + T8U;
|
Chris@82
|
756 T96 = T94 - T95;
|
Chris@82
|
757 TcV = T95 + T94;
|
Chris@82
|
758 }
|
Chris@82
|
759 }
|
Chris@82
|
760 }
|
Chris@82
|
761 {
|
Chris@82
|
762 E T43, T9k, T4m, T9t, T49, T9m, T4g, T9r;
|
Chris@82
|
763 {
|
Chris@82
|
764 E T3Z, T42, T40, T9j, T3Y, T41;
|
Chris@82
|
765 T3Z = ri[WS(rs, 9)];
|
Chris@82
|
766 T42 = ii[WS(rs, 9)];
|
Chris@82
|
767 T3Y = W[16];
|
Chris@82
|
768 T40 = T3Y * T3Z;
|
Chris@82
|
769 T9j = T3Y * T42;
|
Chris@82
|
770 T41 = W[17];
|
Chris@82
|
771 T43 = FMA(T41, T42, T40);
|
Chris@82
|
772 T9k = FNMS(T41, T3Z, T9j);
|
Chris@82
|
773 }
|
Chris@82
|
774 {
|
Chris@82
|
775 E T4i, T4l, T4j, T9s, T4h, T4k;
|
Chris@82
|
776 T4i = ri[WS(rs, 25)];
|
Chris@82
|
777 T4l = ii[WS(rs, 25)];
|
Chris@82
|
778 T4h = W[48];
|
Chris@82
|
779 T4j = T4h * T4i;
|
Chris@82
|
780 T9s = T4h * T4l;
|
Chris@82
|
781 T4k = W[49];
|
Chris@82
|
782 T4m = FMA(T4k, T4l, T4j);
|
Chris@82
|
783 T9t = FNMS(T4k, T4i, T9s);
|
Chris@82
|
784 }
|
Chris@82
|
785 {
|
Chris@82
|
786 E T45, T48, T46, T9l, T44, T47;
|
Chris@82
|
787 T45 = ri[WS(rs, 41)];
|
Chris@82
|
788 T48 = ii[WS(rs, 41)];
|
Chris@82
|
789 T44 = W[80];
|
Chris@82
|
790 T46 = T44 * T45;
|
Chris@82
|
791 T9l = T44 * T48;
|
Chris@82
|
792 T47 = W[81];
|
Chris@82
|
793 T49 = FMA(T47, T48, T46);
|
Chris@82
|
794 T9m = FNMS(T47, T45, T9l);
|
Chris@82
|
795 }
|
Chris@82
|
796 {
|
Chris@82
|
797 E T4c, T4f, T4d, T9q, T4b, T4e;
|
Chris@82
|
798 T4c = ri[WS(rs, 57)];
|
Chris@82
|
799 T4f = ii[WS(rs, 57)];
|
Chris@82
|
800 T4b = W[112];
|
Chris@82
|
801 T4d = T4b * T4c;
|
Chris@82
|
802 T9q = T4b * T4f;
|
Chris@82
|
803 T4e = W[113];
|
Chris@82
|
804 T4g = FMA(T4e, T4f, T4d);
|
Chris@82
|
805 T9r = FNMS(T4e, T4c, T9q);
|
Chris@82
|
806 }
|
Chris@82
|
807 {
|
Chris@82
|
808 E T4a, T4n, TfJ, TfK;
|
Chris@82
|
809 T4a = T43 + T49;
|
Chris@82
|
810 T4n = T4g + T4m;
|
Chris@82
|
811 T4o = T4a + T4n;
|
Chris@82
|
812 Tfu = T4n - T4a;
|
Chris@82
|
813 TfJ = T9k + T9m;
|
Chris@82
|
814 TfK = T9r + T9t;
|
Chris@82
|
815 TfL = TfJ - TfK;
|
Chris@82
|
816 ThI = TfJ + TfK;
|
Chris@82
|
817 {
|
Chris@82
|
818 E T9o, Tac, T9v, Tab;
|
Chris@82
|
819 {
|
Chris@82
|
820 E T9i, T9n, T9p, T9u;
|
Chris@82
|
821 T9i = T43 - T49;
|
Chris@82
|
822 T9n = T9k - T9m;
|
Chris@82
|
823 T9o = T9i + T9n;
|
Chris@82
|
824 Tac = T9n - T9i;
|
Chris@82
|
825 T9p = T4g - T4m;
|
Chris@82
|
826 T9u = T9r - T9t;
|
Chris@82
|
827 T9v = T9p - T9u;
|
Chris@82
|
828 Tab = T9p + T9u;
|
Chris@82
|
829 }
|
Chris@82
|
830 T9w = T9o - T9v;
|
Chris@82
|
831 Tdf = T9o + T9v;
|
Chris@82
|
832 Tad = Tab - Tac;
|
Chris@82
|
833 Td4 = Tac + Tab;
|
Chris@82
|
834 }
|
Chris@82
|
835 }
|
Chris@82
|
836 }
|
Chris@82
|
837 {
|
Chris@82
|
838 E T5Q, Tat, T69, TaC, T5W, Tav, T63, TaA;
|
Chris@82
|
839 {
|
Chris@82
|
840 E T5M, T5P, T5N, Tas, T5L, T5O;
|
Chris@82
|
841 T5M = ri[WS(rs, 7)];
|
Chris@82
|
842 T5P = ii[WS(rs, 7)];
|
Chris@82
|
843 T5L = W[12];
|
Chris@82
|
844 T5N = T5L * T5M;
|
Chris@82
|
845 Tas = T5L * T5P;
|
Chris@82
|
846 T5O = W[13];
|
Chris@82
|
847 T5Q = FMA(T5O, T5P, T5N);
|
Chris@82
|
848 Tat = FNMS(T5O, T5M, Tas);
|
Chris@82
|
849 }
|
Chris@82
|
850 {
|
Chris@82
|
851 E T65, T68, T66, TaB, T64, T67;
|
Chris@82
|
852 T65 = ri[WS(rs, 23)];
|
Chris@82
|
853 T68 = ii[WS(rs, 23)];
|
Chris@82
|
854 T64 = W[44];
|
Chris@82
|
855 T66 = T64 * T65;
|
Chris@82
|
856 TaB = T64 * T68;
|
Chris@82
|
857 T67 = W[45];
|
Chris@82
|
858 T69 = FMA(T67, T68, T66);
|
Chris@82
|
859 TaC = FNMS(T67, T65, TaB);
|
Chris@82
|
860 }
|
Chris@82
|
861 {
|
Chris@82
|
862 E T5S, T5V, T5T, Tau, T5R, T5U;
|
Chris@82
|
863 T5S = ri[WS(rs, 39)];
|
Chris@82
|
864 T5V = ii[WS(rs, 39)];
|
Chris@82
|
865 T5R = W[76];
|
Chris@82
|
866 T5T = T5R * T5S;
|
Chris@82
|
867 Tau = T5R * T5V;
|
Chris@82
|
868 T5U = W[77];
|
Chris@82
|
869 T5W = FMA(T5U, T5V, T5T);
|
Chris@82
|
870 Tav = FNMS(T5U, T5S, Tau);
|
Chris@82
|
871 }
|
Chris@82
|
872 {
|
Chris@82
|
873 E T5Z, T62, T60, Taz, T5Y, T61;
|
Chris@82
|
874 T5Z = ri[WS(rs, 55)];
|
Chris@82
|
875 T62 = ii[WS(rs, 55)];
|
Chris@82
|
876 T5Y = W[108];
|
Chris@82
|
877 T60 = T5Y * T5Z;
|
Chris@82
|
878 Taz = T5Y * T62;
|
Chris@82
|
879 T61 = W[109];
|
Chris@82
|
880 T63 = FMA(T61, T62, T60);
|
Chris@82
|
881 TaA = FNMS(T61, T5Z, Taz);
|
Chris@82
|
882 }
|
Chris@82
|
883 {
|
Chris@82
|
884 E T5X, T6a, Tga, Tgb;
|
Chris@82
|
885 T5X = T5Q + T5W;
|
Chris@82
|
886 T6a = T63 + T69;
|
Chris@82
|
887 T6b = T5X + T6a;
|
Chris@82
|
888 TfV = T6a - T5X;
|
Chris@82
|
889 Tga = Tat + Tav;
|
Chris@82
|
890 Tgb = TaA + TaC;
|
Chris@82
|
891 Tgc = Tga - Tgb;
|
Chris@82
|
892 ThT = Tga + Tgb;
|
Chris@82
|
893 {
|
Chris@82
|
894 E Tax, Tbl, TaE, Tbk;
|
Chris@82
|
895 {
|
Chris@82
|
896 E Tar, Taw, Tay, TaD;
|
Chris@82
|
897 Tar = T5Q - T5W;
|
Chris@82
|
898 Taw = Tat - Tav;
|
Chris@82
|
899 Tax = Tar + Taw;
|
Chris@82
|
900 Tbl = Taw - Tar;
|
Chris@82
|
901 Tay = T63 - T69;
|
Chris@82
|
902 TaD = TaA - TaC;
|
Chris@82
|
903 TaE = Tay - TaD;
|
Chris@82
|
904 Tbk = Tay + TaD;
|
Chris@82
|
905 }
|
Chris@82
|
906 TaF = Tax - TaE;
|
Chris@82
|
907 Tdy = Tax + TaE;
|
Chris@82
|
908 Tbm = Tbk - Tbl;
|
Chris@82
|
909 Tdn = Tbl + Tbk;
|
Chris@82
|
910 }
|
Chris@82
|
911 }
|
Chris@82
|
912 }
|
Chris@82
|
913 {
|
Chris@82
|
914 E T4v, T9V, T4O, T9R, T4B, T9X, T4I, T9P;
|
Chris@82
|
915 {
|
Chris@82
|
916 E T4r, T4u, T4s, T9U, T4q, T4t;
|
Chris@82
|
917 T4r = ri[WS(rs, 5)];
|
Chris@82
|
918 T4u = ii[WS(rs, 5)];
|
Chris@82
|
919 T4q = W[8];
|
Chris@82
|
920 T4s = T4q * T4r;
|
Chris@82
|
921 T9U = T4q * T4u;
|
Chris@82
|
922 T4t = W[9];
|
Chris@82
|
923 T4v = FMA(T4t, T4u, T4s);
|
Chris@82
|
924 T9V = FNMS(T4t, T4r, T9U);
|
Chris@82
|
925 }
|
Chris@82
|
926 {
|
Chris@82
|
927 E T4K, T4N, T4L, T9Q, T4J, T4M;
|
Chris@82
|
928 T4K = ri[WS(rs, 53)];
|
Chris@82
|
929 T4N = ii[WS(rs, 53)];
|
Chris@82
|
930 T4J = W[104];
|
Chris@82
|
931 T4L = T4J * T4K;
|
Chris@82
|
932 T9Q = T4J * T4N;
|
Chris@82
|
933 T4M = W[105];
|
Chris@82
|
934 T4O = FMA(T4M, T4N, T4L);
|
Chris@82
|
935 T9R = FNMS(T4M, T4K, T9Q);
|
Chris@82
|
936 }
|
Chris@82
|
937 {
|
Chris@82
|
938 E T4x, T4A, T4y, T9W, T4w, T4z;
|
Chris@82
|
939 T4x = ri[WS(rs, 37)];
|
Chris@82
|
940 T4A = ii[WS(rs, 37)];
|
Chris@82
|
941 T4w = W[72];
|
Chris@82
|
942 T4y = T4w * T4x;
|
Chris@82
|
943 T9W = T4w * T4A;
|
Chris@82
|
944 T4z = W[73];
|
Chris@82
|
945 T4B = FMA(T4z, T4A, T4y);
|
Chris@82
|
946 T9X = FNMS(T4z, T4x, T9W);
|
Chris@82
|
947 }
|
Chris@82
|
948 {
|
Chris@82
|
949 E T4E, T4H, T4F, T9O, T4D, T4G;
|
Chris@82
|
950 T4E = ri[WS(rs, 21)];
|
Chris@82
|
951 T4H = ii[WS(rs, 21)];
|
Chris@82
|
952 T4D = W[40];
|
Chris@82
|
953 T4F = T4D * T4E;
|
Chris@82
|
954 T9O = T4D * T4H;
|
Chris@82
|
955 T4G = W[41];
|
Chris@82
|
956 T4I = FMA(T4G, T4H, T4F);
|
Chris@82
|
957 T9P = FNMS(T4G, T4E, T9O);
|
Chris@82
|
958 }
|
Chris@82
|
959 {
|
Chris@82
|
960 E T4C, T4P, Tfz, Tfw, Tfx, Tfy;
|
Chris@82
|
961 T4C = T4v + T4B;
|
Chris@82
|
962 T4P = T4I + T4O;
|
Chris@82
|
963 Tfz = T4C - T4P;
|
Chris@82
|
964 Tfw = T9V + T9X;
|
Chris@82
|
965 Tfx = T9P + T9R;
|
Chris@82
|
966 Tfy = Tfw - Tfx;
|
Chris@82
|
967 T4Q = T4C + T4P;
|
Chris@82
|
968 ThN = Tfw + Tfx;
|
Chris@82
|
969 TfA = Tfy - Tfz;
|
Chris@82
|
970 TfN = Tfz + Tfy;
|
Chris@82
|
971 }
|
Chris@82
|
972 {
|
Chris@82
|
973 E T9T, Td7, Ta0, Td6;
|
Chris@82
|
974 {
|
Chris@82
|
975 E T9N, T9S, T9Y, T9Z;
|
Chris@82
|
976 T9N = T4v - T4B;
|
Chris@82
|
977 T9S = T9P - T9R;
|
Chris@82
|
978 T9T = T9N - T9S;
|
Chris@82
|
979 Td7 = T9N + T9S;
|
Chris@82
|
980 T9Y = T9V - T9X;
|
Chris@82
|
981 T9Z = T4I - T4O;
|
Chris@82
|
982 Ta0 = T9Y + T9Z;
|
Chris@82
|
983 Td6 = T9Y - T9Z;
|
Chris@82
|
984 }
|
Chris@82
|
985 Ta1 = FNMS(KP414213562, Ta0, T9T);
|
Chris@82
|
986 Tdh = FMA(KP414213562, Td6, Td7);
|
Chris@82
|
987 Taf = FMA(KP414213562, T9T, Ta0);
|
Chris@82
|
988 Td8 = FNMS(KP414213562, Td7, Td6);
|
Chris@82
|
989 }
|
Chris@82
|
990 }
|
Chris@82
|
991 {
|
Chris@82
|
992 E T4W, T9G, T5f, T9C, T52, T9I, T59, T9A;
|
Chris@82
|
993 {
|
Chris@82
|
994 E T4S, T4V, T4T, T9F, T4R, T4U;
|
Chris@82
|
995 T4S = ri[WS(rs, 61)];
|
Chris@82
|
996 T4V = ii[WS(rs, 61)];
|
Chris@82
|
997 T4R = W[120];
|
Chris@82
|
998 T4T = T4R * T4S;
|
Chris@82
|
999 T9F = T4R * T4V;
|
Chris@82
|
1000 T4U = W[121];
|
Chris@82
|
1001 T4W = FMA(T4U, T4V, T4T);
|
Chris@82
|
1002 T9G = FNMS(T4U, T4S, T9F);
|
Chris@82
|
1003 }
|
Chris@82
|
1004 {
|
Chris@82
|
1005 E T5b, T5e, T5c, T9B, T5a, T5d;
|
Chris@82
|
1006 T5b = ri[WS(rs, 45)];
|
Chris@82
|
1007 T5e = ii[WS(rs, 45)];
|
Chris@82
|
1008 T5a = W[88];
|
Chris@82
|
1009 T5c = T5a * T5b;
|
Chris@82
|
1010 T9B = T5a * T5e;
|
Chris@82
|
1011 T5d = W[89];
|
Chris@82
|
1012 T5f = FMA(T5d, T5e, T5c);
|
Chris@82
|
1013 T9C = FNMS(T5d, T5b, T9B);
|
Chris@82
|
1014 }
|
Chris@82
|
1015 {
|
Chris@82
|
1016 E T4Y, T51, T4Z, T9H, T4X, T50;
|
Chris@82
|
1017 T4Y = ri[WS(rs, 29)];
|
Chris@82
|
1018 T51 = ii[WS(rs, 29)];
|
Chris@82
|
1019 T4X = W[56];
|
Chris@82
|
1020 T4Z = T4X * T4Y;
|
Chris@82
|
1021 T9H = T4X * T51;
|
Chris@82
|
1022 T50 = W[57];
|
Chris@82
|
1023 T52 = FMA(T50, T51, T4Z);
|
Chris@82
|
1024 T9I = FNMS(T50, T4Y, T9H);
|
Chris@82
|
1025 }
|
Chris@82
|
1026 {
|
Chris@82
|
1027 E T55, T58, T56, T9z, T54, T57;
|
Chris@82
|
1028 T55 = ri[WS(rs, 13)];
|
Chris@82
|
1029 T58 = ii[WS(rs, 13)];
|
Chris@82
|
1030 T54 = W[24];
|
Chris@82
|
1031 T56 = T54 * T55;
|
Chris@82
|
1032 T9z = T54 * T58;
|
Chris@82
|
1033 T57 = W[25];
|
Chris@82
|
1034 T59 = FMA(T57, T58, T56);
|
Chris@82
|
1035 T9A = FNMS(T57, T55, T9z);
|
Chris@82
|
1036 }
|
Chris@82
|
1037 {
|
Chris@82
|
1038 E T53, T5g, TfB, TfC, TfD, TfE;
|
Chris@82
|
1039 T53 = T4W + T52;
|
Chris@82
|
1040 T5g = T59 + T5f;
|
Chris@82
|
1041 TfB = T53 - T5g;
|
Chris@82
|
1042 TfC = T9G + T9I;
|
Chris@82
|
1043 TfD = T9A + T9C;
|
Chris@82
|
1044 TfE = TfC - TfD;
|
Chris@82
|
1045 T5h = T53 + T5g;
|
Chris@82
|
1046 ThO = TfC + TfD;
|
Chris@82
|
1047 TfF = TfB + TfE;
|
Chris@82
|
1048 TfO = TfB - TfE;
|
Chris@82
|
1049 }
|
Chris@82
|
1050 {
|
Chris@82
|
1051 E T9E, Tda, T9L, Td9;
|
Chris@82
|
1052 {
|
Chris@82
|
1053 E T9y, T9D, T9J, T9K;
|
Chris@82
|
1054 T9y = T4W - T52;
|
Chris@82
|
1055 T9D = T9A - T9C;
|
Chris@82
|
1056 T9E = T9y - T9D;
|
Chris@82
|
1057 Tda = T9y + T9D;
|
Chris@82
|
1058 T9J = T9G - T9I;
|
Chris@82
|
1059 T9K = T59 - T5f;
|
Chris@82
|
1060 T9L = T9J + T9K;
|
Chris@82
|
1061 Td9 = T9J - T9K;
|
Chris@82
|
1062 }
|
Chris@82
|
1063 T9M = FMA(KP414213562, T9L, T9E);
|
Chris@82
|
1064 Tdi = FNMS(KP414213562, Td9, Tda);
|
Chris@82
|
1065 Tag = FNMS(KP414213562, T9E, T9L);
|
Chris@82
|
1066 Tdb = FMA(KP414213562, Tda, Td9);
|
Chris@82
|
1067 }
|
Chris@82
|
1068 }
|
Chris@82
|
1069 {
|
Chris@82
|
1070 E T6i, Tb4, T6B, Tb0, T6o, Tb6, T6v, TaY;
|
Chris@82
|
1071 {
|
Chris@82
|
1072 E T6e, T6h, T6f, Tb3, T6d, T6g;
|
Chris@82
|
1073 T6e = ri[WS(rs, 3)];
|
Chris@82
|
1074 T6h = ii[WS(rs, 3)];
|
Chris@82
|
1075 T6d = W[4];
|
Chris@82
|
1076 T6f = T6d * T6e;
|
Chris@82
|
1077 Tb3 = T6d * T6h;
|
Chris@82
|
1078 T6g = W[5];
|
Chris@82
|
1079 T6i = FMA(T6g, T6h, T6f);
|
Chris@82
|
1080 Tb4 = FNMS(T6g, T6e, Tb3);
|
Chris@82
|
1081 }
|
Chris@82
|
1082 {
|
Chris@82
|
1083 E T6x, T6A, T6y, TaZ, T6w, T6z;
|
Chris@82
|
1084 T6x = ri[WS(rs, 51)];
|
Chris@82
|
1085 T6A = ii[WS(rs, 51)];
|
Chris@82
|
1086 T6w = W[100];
|
Chris@82
|
1087 T6y = T6w * T6x;
|
Chris@82
|
1088 TaZ = T6w * T6A;
|
Chris@82
|
1089 T6z = W[101];
|
Chris@82
|
1090 T6B = FMA(T6z, T6A, T6y);
|
Chris@82
|
1091 Tb0 = FNMS(T6z, T6x, TaZ);
|
Chris@82
|
1092 }
|
Chris@82
|
1093 {
|
Chris@82
|
1094 E T6k, T6n, T6l, Tb5, T6j, T6m;
|
Chris@82
|
1095 T6k = ri[WS(rs, 35)];
|
Chris@82
|
1096 T6n = ii[WS(rs, 35)];
|
Chris@82
|
1097 T6j = W[68];
|
Chris@82
|
1098 T6l = T6j * T6k;
|
Chris@82
|
1099 Tb5 = T6j * T6n;
|
Chris@82
|
1100 T6m = W[69];
|
Chris@82
|
1101 T6o = FMA(T6m, T6n, T6l);
|
Chris@82
|
1102 Tb6 = FNMS(T6m, T6k, Tb5);
|
Chris@82
|
1103 }
|
Chris@82
|
1104 {
|
Chris@82
|
1105 E T6r, T6u, T6s, TaX, T6q, T6t;
|
Chris@82
|
1106 T6r = ri[WS(rs, 19)];
|
Chris@82
|
1107 T6u = ii[WS(rs, 19)];
|
Chris@82
|
1108 T6q = W[36];
|
Chris@82
|
1109 T6s = T6q * T6r;
|
Chris@82
|
1110 TaX = T6q * T6u;
|
Chris@82
|
1111 T6t = W[37];
|
Chris@82
|
1112 T6v = FMA(T6t, T6u, T6s);
|
Chris@82
|
1113 TaY = FNMS(T6t, T6r, TaX);
|
Chris@82
|
1114 }
|
Chris@82
|
1115 {
|
Chris@82
|
1116 E T6p, T6C, Tg0, TfX, TfY, TfZ;
|
Chris@82
|
1117 T6p = T6i + T6o;
|
Chris@82
|
1118 T6C = T6v + T6B;
|
Chris@82
|
1119 Tg0 = T6p - T6C;
|
Chris@82
|
1120 TfX = Tb4 + Tb6;
|
Chris@82
|
1121 TfY = TaY + Tb0;
|
Chris@82
|
1122 TfZ = TfX - TfY;
|
Chris@82
|
1123 T6D = T6p + T6C;
|
Chris@82
|
1124 ThY = TfX + TfY;
|
Chris@82
|
1125 Tg1 = TfZ - Tg0;
|
Chris@82
|
1126 Tge = Tg0 + TfZ;
|
Chris@82
|
1127 }
|
Chris@82
|
1128 {
|
Chris@82
|
1129 E Tb2, Tdq, Tb9, Tdp;
|
Chris@82
|
1130 {
|
Chris@82
|
1131 E TaW, Tb1, Tb7, Tb8;
|
Chris@82
|
1132 TaW = T6i - T6o;
|
Chris@82
|
1133 Tb1 = TaY - Tb0;
|
Chris@82
|
1134 Tb2 = TaW - Tb1;
|
Chris@82
|
1135 Tdq = TaW + Tb1;
|
Chris@82
|
1136 Tb7 = Tb4 - Tb6;
|
Chris@82
|
1137 Tb8 = T6v - T6B;
|
Chris@82
|
1138 Tb9 = Tb7 + Tb8;
|
Chris@82
|
1139 Tdp = Tb7 - Tb8;
|
Chris@82
|
1140 }
|
Chris@82
|
1141 Tba = FNMS(KP414213562, Tb9, Tb2);
|
Chris@82
|
1142 TdA = FMA(KP414213562, Tdp, Tdq);
|
Chris@82
|
1143 Tbo = FMA(KP414213562, Tb2, Tb9);
|
Chris@82
|
1144 Tdr = FNMS(KP414213562, Tdq, Tdp);
|
Chris@82
|
1145 }
|
Chris@82
|
1146 }
|
Chris@82
|
1147 {
|
Chris@82
|
1148 E T6J, TaP, T72, TaL, T6P, TaR, T6W, TaJ;
|
Chris@82
|
1149 {
|
Chris@82
|
1150 E T6F, T6I, T6G, TaO, T6E, T6H;
|
Chris@82
|
1151 T6F = ri[WS(rs, 59)];
|
Chris@82
|
1152 T6I = ii[WS(rs, 59)];
|
Chris@82
|
1153 T6E = W[116];
|
Chris@82
|
1154 T6G = T6E * T6F;
|
Chris@82
|
1155 TaO = T6E * T6I;
|
Chris@82
|
1156 T6H = W[117];
|
Chris@82
|
1157 T6J = FMA(T6H, T6I, T6G);
|
Chris@82
|
1158 TaP = FNMS(T6H, T6F, TaO);
|
Chris@82
|
1159 }
|
Chris@82
|
1160 {
|
Chris@82
|
1161 E T6Y, T71, T6Z, TaK, T6X, T70;
|
Chris@82
|
1162 T6Y = ri[WS(rs, 43)];
|
Chris@82
|
1163 T71 = ii[WS(rs, 43)];
|
Chris@82
|
1164 T6X = W[84];
|
Chris@82
|
1165 T6Z = T6X * T6Y;
|
Chris@82
|
1166 TaK = T6X * T71;
|
Chris@82
|
1167 T70 = W[85];
|
Chris@82
|
1168 T72 = FMA(T70, T71, T6Z);
|
Chris@82
|
1169 TaL = FNMS(T70, T6Y, TaK);
|
Chris@82
|
1170 }
|
Chris@82
|
1171 {
|
Chris@82
|
1172 E T6L, T6O, T6M, TaQ, T6K, T6N;
|
Chris@82
|
1173 T6L = ri[WS(rs, 27)];
|
Chris@82
|
1174 T6O = ii[WS(rs, 27)];
|
Chris@82
|
1175 T6K = W[52];
|
Chris@82
|
1176 T6M = T6K * T6L;
|
Chris@82
|
1177 TaQ = T6K * T6O;
|
Chris@82
|
1178 T6N = W[53];
|
Chris@82
|
1179 T6P = FMA(T6N, T6O, T6M);
|
Chris@82
|
1180 TaR = FNMS(T6N, T6L, TaQ);
|
Chris@82
|
1181 }
|
Chris@82
|
1182 {
|
Chris@82
|
1183 E T6S, T6V, T6T, TaI, T6R, T6U;
|
Chris@82
|
1184 T6S = ri[WS(rs, 11)];
|
Chris@82
|
1185 T6V = ii[WS(rs, 11)];
|
Chris@82
|
1186 T6R = W[20];
|
Chris@82
|
1187 T6T = T6R * T6S;
|
Chris@82
|
1188 TaI = T6R * T6V;
|
Chris@82
|
1189 T6U = W[21];
|
Chris@82
|
1190 T6W = FMA(T6U, T6V, T6T);
|
Chris@82
|
1191 TaJ = FNMS(T6U, T6S, TaI);
|
Chris@82
|
1192 }
|
Chris@82
|
1193 {
|
Chris@82
|
1194 E T6Q, T73, Tg2, Tg3, Tg4, Tg5;
|
Chris@82
|
1195 T6Q = T6J + T6P;
|
Chris@82
|
1196 T73 = T6W + T72;
|
Chris@82
|
1197 Tg2 = T6Q - T73;
|
Chris@82
|
1198 Tg3 = TaP + TaR;
|
Chris@82
|
1199 Tg4 = TaJ + TaL;
|
Chris@82
|
1200 Tg5 = Tg3 - Tg4;
|
Chris@82
|
1201 T74 = T6Q + T73;
|
Chris@82
|
1202 ThZ = Tg3 + Tg4;
|
Chris@82
|
1203 Tg6 = Tg2 + Tg5;
|
Chris@82
|
1204 Tgf = Tg2 - Tg5;
|
Chris@82
|
1205 }
|
Chris@82
|
1206 {
|
Chris@82
|
1207 E TaN, Tdt, TaU, Tds;
|
Chris@82
|
1208 {
|
Chris@82
|
1209 E TaH, TaM, TaS, TaT;
|
Chris@82
|
1210 TaH = T6J - T6P;
|
Chris@82
|
1211 TaM = TaJ - TaL;
|
Chris@82
|
1212 TaN = TaH - TaM;
|
Chris@82
|
1213 Tdt = TaH + TaM;
|
Chris@82
|
1214 TaS = TaP - TaR;
|
Chris@82
|
1215 TaT = T6W - T72;
|
Chris@82
|
1216 TaU = TaS + TaT;
|
Chris@82
|
1217 Tds = TaS - TaT;
|
Chris@82
|
1218 }
|
Chris@82
|
1219 TaV = FMA(KP414213562, TaU, TaN);
|
Chris@82
|
1220 TdB = FNMS(KP414213562, Tds, Tdt);
|
Chris@82
|
1221 Tbp = FNMS(KP414213562, TaN, TaU);
|
Chris@82
|
1222 Tdu = FMA(KP414213562, Tdt, Tds);
|
Chris@82
|
1223 }
|
Chris@82
|
1224 }
|
Chris@82
|
1225 {
|
Chris@82
|
1226 E T1I, Tio, T3v, Tj1, TiX, Tj2, Tir, TiN, T76, TiK, TiC, TiG, T5j, TiJ, Tix;
|
Chris@82
|
1227 E TiF;
|
Chris@82
|
1228 {
|
Chris@82
|
1229 E TO, T1H, Tip, Tiq;
|
Chris@82
|
1230 TO = Tm + TN;
|
Chris@82
|
1231 T1H = T1f + T1G;
|
Chris@82
|
1232 T1I = TO + T1H;
|
Chris@82
|
1233 Tio = TO - T1H;
|
Chris@82
|
1234 {
|
Chris@82
|
1235 E T2B, T3u, TiO, TiW;
|
Chris@82
|
1236 T2B = T29 + T2A;
|
Chris@82
|
1237 T3u = T32 + T3t;
|
Chris@82
|
1238 T3v = T2B + T3u;
|
Chris@82
|
1239 Tj1 = T3u - T2B;
|
Chris@82
|
1240 TiO = Thr + Ths;
|
Chris@82
|
1241 TiW = TiP + TiV;
|
Chris@82
|
1242 TiX = TiO + TiW;
|
Chris@82
|
1243 Tj2 = TiW - TiO;
|
Chris@82
|
1244 }
|
Chris@82
|
1245 Tip = Thv + Thw;
|
Chris@82
|
1246 Tiq = ThB + ThC;
|
Chris@82
|
1247 Tir = Tip - Tiq;
|
Chris@82
|
1248 TiN = Tip + Tiq;
|
Chris@82
|
1249 {
|
Chris@82
|
1250 E T6c, T75, Tiy, Tiz, TiA, TiB;
|
Chris@82
|
1251 T6c = T5K + T6b;
|
Chris@82
|
1252 T75 = T6D + T74;
|
Chris@82
|
1253 Tiy = T6c - T75;
|
Chris@82
|
1254 Tiz = ThS + ThT;
|
Chris@82
|
1255 TiA = ThY + ThZ;
|
Chris@82
|
1256 TiB = Tiz - TiA;
|
Chris@82
|
1257 T76 = T6c + T75;
|
Chris@82
|
1258 TiK = Tiz + TiA;
|
Chris@82
|
1259 TiC = Tiy - TiB;
|
Chris@82
|
1260 TiG = Tiy + TiB;
|
Chris@82
|
1261 }
|
Chris@82
|
1262 {
|
Chris@82
|
1263 E T4p, T5i, Tit, Tiu, Tiv, Tiw;
|
Chris@82
|
1264 T4p = T3X + T4o;
|
Chris@82
|
1265 T5i = T4Q + T5h;
|
Chris@82
|
1266 Tit = T4p - T5i;
|
Chris@82
|
1267 Tiu = ThH + ThI;
|
Chris@82
|
1268 Tiv = ThN + ThO;
|
Chris@82
|
1269 Tiw = Tiu - Tiv;
|
Chris@82
|
1270 T5j = T4p + T5i;
|
Chris@82
|
1271 TiJ = Tiu + Tiv;
|
Chris@82
|
1272 Tix = Tit + Tiw;
|
Chris@82
|
1273 TiF = Tiw - Tit;
|
Chris@82
|
1274 }
|
Chris@82
|
1275 }
|
Chris@82
|
1276 {
|
Chris@82
|
1277 E T3w, T77, TiM, TiY;
|
Chris@82
|
1278 T3w = T1I + T3v;
|
Chris@82
|
1279 T77 = T5j + T76;
|
Chris@82
|
1280 ri[WS(rs, 32)] = T3w - T77;
|
Chris@82
|
1281 ri[0] = T3w + T77;
|
Chris@82
|
1282 TiM = TiJ + TiK;
|
Chris@82
|
1283 TiY = TiN + TiX;
|
Chris@82
|
1284 ii[0] = TiM + TiY;
|
Chris@82
|
1285 ii[WS(rs, 32)] = TiY - TiM;
|
Chris@82
|
1286 }
|
Chris@82
|
1287 {
|
Chris@82
|
1288 E Tis, TiD, Tj3, Tj4;
|
Chris@82
|
1289 Tis = Tio + Tir;
|
Chris@82
|
1290 TiD = Tix + TiC;
|
Chris@82
|
1291 ri[WS(rs, 40)] = FNMS(KP707106781, TiD, Tis);
|
Chris@82
|
1292 ri[WS(rs, 8)] = FMA(KP707106781, TiD, Tis);
|
Chris@82
|
1293 Tj3 = Tj1 + Tj2;
|
Chris@82
|
1294 Tj4 = TiF + TiG;
|
Chris@82
|
1295 ii[WS(rs, 8)] = FMA(KP707106781, Tj4, Tj3);
|
Chris@82
|
1296 ii[WS(rs, 40)] = FNMS(KP707106781, Tj4, Tj3);
|
Chris@82
|
1297 }
|
Chris@82
|
1298 {
|
Chris@82
|
1299 E TiE, TiH, Tj5, Tj6;
|
Chris@82
|
1300 TiE = Tio - Tir;
|
Chris@82
|
1301 TiH = TiF - TiG;
|
Chris@82
|
1302 ri[WS(rs, 56)] = FNMS(KP707106781, TiH, TiE);
|
Chris@82
|
1303 ri[WS(rs, 24)] = FMA(KP707106781, TiH, TiE);
|
Chris@82
|
1304 Tj5 = Tj2 - Tj1;
|
Chris@82
|
1305 Tj6 = TiC - Tix;
|
Chris@82
|
1306 ii[WS(rs, 24)] = FMA(KP707106781, Tj6, Tj5);
|
Chris@82
|
1307 ii[WS(rs, 56)] = FNMS(KP707106781, Tj6, Tj5);
|
Chris@82
|
1308 }
|
Chris@82
|
1309 {
|
Chris@82
|
1310 E TiI, TiL, TiZ, Tj0;
|
Chris@82
|
1311 TiI = T1I - T3v;
|
Chris@82
|
1312 TiL = TiJ - TiK;
|
Chris@82
|
1313 ri[WS(rs, 48)] = TiI - TiL;
|
Chris@82
|
1314 ri[WS(rs, 16)] = TiI + TiL;
|
Chris@82
|
1315 TiZ = T76 - T5j;
|
Chris@82
|
1316 Tj0 = TiX - TiN;
|
Chris@82
|
1317 ii[WS(rs, 16)] = TiZ + Tj0;
|
Chris@82
|
1318 ii[WS(rs, 48)] = Tj0 - TiZ;
|
Chris@82
|
1319 }
|
Chris@82
|
1320 }
|
Chris@82
|
1321 {
|
Chris@82
|
1322 E Thu, Ti8, Tj9, Tjf, ThF, Tjg, Tib, Tja, ThR, Til, Ti5, Tif, Ti2, Tim, Ti6;
|
Chris@82
|
1323 E Tii;
|
Chris@82
|
1324 {
|
Chris@82
|
1325 E Thq, Tht, Tj7, Tj8;
|
Chris@82
|
1326 Thq = Tm - TN;
|
Chris@82
|
1327 Tht = Thr - Ths;
|
Chris@82
|
1328 Thu = Thq - Tht;
|
Chris@82
|
1329 Ti8 = Thq + Tht;
|
Chris@82
|
1330 Tj7 = T1G - T1f;
|
Chris@82
|
1331 Tj8 = TiV - TiP;
|
Chris@82
|
1332 Tj9 = Tj7 + Tj8;
|
Chris@82
|
1333 Tjf = Tj8 - Tj7;
|
Chris@82
|
1334 }
|
Chris@82
|
1335 {
|
Chris@82
|
1336 E Thz, Ti9, ThE, Tia;
|
Chris@82
|
1337 {
|
Chris@82
|
1338 E Thx, Thy, ThA, ThD;
|
Chris@82
|
1339 Thx = Thv - Thw;
|
Chris@82
|
1340 Thy = T29 - T2A;
|
Chris@82
|
1341 Thz = Thx - Thy;
|
Chris@82
|
1342 Ti9 = Thy + Thx;
|
Chris@82
|
1343 ThA = T32 - T3t;
|
Chris@82
|
1344 ThD = ThB - ThC;
|
Chris@82
|
1345 ThE = ThA + ThD;
|
Chris@82
|
1346 Tia = ThA - ThD;
|
Chris@82
|
1347 }
|
Chris@82
|
1348 ThF = Thz - ThE;
|
Chris@82
|
1349 Tjg = Tia - Ti9;
|
Chris@82
|
1350 Tib = Ti9 + Tia;
|
Chris@82
|
1351 Tja = Thz + ThE;
|
Chris@82
|
1352 }
|
Chris@82
|
1353 {
|
Chris@82
|
1354 E ThL, Tie, ThQ, Tid;
|
Chris@82
|
1355 {
|
Chris@82
|
1356 E ThJ, ThK, ThM, ThP;
|
Chris@82
|
1357 ThJ = ThH - ThI;
|
Chris@82
|
1358 ThK = T5h - T4Q;
|
Chris@82
|
1359 ThL = ThJ - ThK;
|
Chris@82
|
1360 Tie = ThJ + ThK;
|
Chris@82
|
1361 ThM = T3X - T4o;
|
Chris@82
|
1362 ThP = ThN - ThO;
|
Chris@82
|
1363 ThQ = ThM - ThP;
|
Chris@82
|
1364 Tid = ThM + ThP;
|
Chris@82
|
1365 }
|
Chris@82
|
1366 ThR = FMA(KP414213562, ThQ, ThL);
|
Chris@82
|
1367 Til = FNMS(KP414213562, Tid, Tie);
|
Chris@82
|
1368 Ti5 = FNMS(KP414213562, ThL, ThQ);
|
Chris@82
|
1369 Tif = FMA(KP414213562, Tie, Tid);
|
Chris@82
|
1370 }
|
Chris@82
|
1371 {
|
Chris@82
|
1372 E ThW, Tih, Ti1, Tig;
|
Chris@82
|
1373 {
|
Chris@82
|
1374 E ThU, ThV, ThX, Ti0;
|
Chris@82
|
1375 ThU = ThS - ThT;
|
Chris@82
|
1376 ThV = T74 - T6D;
|
Chris@82
|
1377 ThW = ThU - ThV;
|
Chris@82
|
1378 Tih = ThU + ThV;
|
Chris@82
|
1379 ThX = T5K - T6b;
|
Chris@82
|
1380 Ti0 = ThY - ThZ;
|
Chris@82
|
1381 Ti1 = ThX - Ti0;
|
Chris@82
|
1382 Tig = ThX + Ti0;
|
Chris@82
|
1383 }
|
Chris@82
|
1384 Ti2 = FNMS(KP414213562, Ti1, ThW);
|
Chris@82
|
1385 Tim = FMA(KP414213562, Tig, Tih);
|
Chris@82
|
1386 Ti6 = FMA(KP414213562, ThW, Ti1);
|
Chris@82
|
1387 Tii = FNMS(KP414213562, Tih, Tig);
|
Chris@82
|
1388 }
|
Chris@82
|
1389 {
|
Chris@82
|
1390 E ThG, Ti3, Tjh, Tji;
|
Chris@82
|
1391 ThG = FMA(KP707106781, ThF, Thu);
|
Chris@82
|
1392 Ti3 = ThR - Ti2;
|
Chris@82
|
1393 ri[WS(rs, 44)] = FNMS(KP923879532, Ti3, ThG);
|
Chris@82
|
1394 ri[WS(rs, 12)] = FMA(KP923879532, Ti3, ThG);
|
Chris@82
|
1395 Tjh = FMA(KP707106781, Tjg, Tjf);
|
Chris@82
|
1396 Tji = Ti6 - Ti5;
|
Chris@82
|
1397 ii[WS(rs, 12)] = FMA(KP923879532, Tji, Tjh);
|
Chris@82
|
1398 ii[WS(rs, 44)] = FNMS(KP923879532, Tji, Tjh);
|
Chris@82
|
1399 }
|
Chris@82
|
1400 {
|
Chris@82
|
1401 E Ti4, Ti7, Tjj, Tjk;
|
Chris@82
|
1402 Ti4 = FNMS(KP707106781, ThF, Thu);
|
Chris@82
|
1403 Ti7 = Ti5 + Ti6;
|
Chris@82
|
1404 ri[WS(rs, 28)] = FNMS(KP923879532, Ti7, Ti4);
|
Chris@82
|
1405 ri[WS(rs, 60)] = FMA(KP923879532, Ti7, Ti4);
|
Chris@82
|
1406 Tjj = FNMS(KP707106781, Tjg, Tjf);
|
Chris@82
|
1407 Tjk = ThR + Ti2;
|
Chris@82
|
1408 ii[WS(rs, 28)] = FNMS(KP923879532, Tjk, Tjj);
|
Chris@82
|
1409 ii[WS(rs, 60)] = FMA(KP923879532, Tjk, Tjj);
|
Chris@82
|
1410 }
|
Chris@82
|
1411 {
|
Chris@82
|
1412 E Tic, Tij, Tjb, Tjc;
|
Chris@82
|
1413 Tic = FMA(KP707106781, Tib, Ti8);
|
Chris@82
|
1414 Tij = Tif + Tii;
|
Chris@82
|
1415 ri[WS(rs, 36)] = FNMS(KP923879532, Tij, Tic);
|
Chris@82
|
1416 ri[WS(rs, 4)] = FMA(KP923879532, Tij, Tic);
|
Chris@82
|
1417 Tjb = FMA(KP707106781, Tja, Tj9);
|
Chris@82
|
1418 Tjc = Til + Tim;
|
Chris@82
|
1419 ii[WS(rs, 4)] = FMA(KP923879532, Tjc, Tjb);
|
Chris@82
|
1420 ii[WS(rs, 36)] = FNMS(KP923879532, Tjc, Tjb);
|
Chris@82
|
1421 }
|
Chris@82
|
1422 {
|
Chris@82
|
1423 E Tik, Tin, Tjd, Tje;
|
Chris@82
|
1424 Tik = FNMS(KP707106781, Tib, Ti8);
|
Chris@82
|
1425 Tin = Til - Tim;
|
Chris@82
|
1426 ri[WS(rs, 52)] = FNMS(KP923879532, Tin, Tik);
|
Chris@82
|
1427 ri[WS(rs, 20)] = FMA(KP923879532, Tin, Tik);
|
Chris@82
|
1428 Tjd = FNMS(KP707106781, Tja, Tj9);
|
Chris@82
|
1429 Tje = Tii - Tif;
|
Chris@82
|
1430 ii[WS(rs, 20)] = FMA(KP923879532, Tje, Tjd);
|
Chris@82
|
1431 ii[WS(rs, 52)] = FNMS(KP923879532, Tje, Tjd);
|
Chris@82
|
1432 }
|
Chris@82
|
1433 }
|
Chris@82
|
1434 {
|
Chris@82
|
1435 E Tf2, TjJ, Tgo, TjD, TgI, Tjv, Tha, Tjp, Tfp, Tjw, Tgr, Tjq, Th4, Tho, Th8;
|
Chris@82
|
1436 E Thk, TfR, TgB, Tgl, Tgv, TgP, TjK, Thd, TjE, TgX, Thn, Th7, Thh, Tgi, TgC;
|
Chris@82
|
1437 E Tgm, Tgy;
|
Chris@82
|
1438 {
|
Chris@82
|
1439 E TeQ, TjB, Tf1, TjC, TeV, Tf0;
|
Chris@82
|
1440 TeQ = TeM + TeP;
|
Chris@82
|
1441 TjB = Tjm - Tjl;
|
Chris@82
|
1442 TeV = TeR + TeU;
|
Chris@82
|
1443 Tf0 = TeW - TeZ;
|
Chris@82
|
1444 Tf1 = TeV + Tf0;
|
Chris@82
|
1445 TjC = Tf0 - TeV;
|
Chris@82
|
1446 Tf2 = FNMS(KP707106781, Tf1, TeQ);
|
Chris@82
|
1447 TjJ = FNMS(KP707106781, TjC, TjB);
|
Chris@82
|
1448 Tgo = FMA(KP707106781, Tf1, TeQ);
|
Chris@82
|
1449 TjD = FMA(KP707106781, TjC, TjB);
|
Chris@82
|
1450 }
|
Chris@82
|
1451 {
|
Chris@82
|
1452 E TgE, Tjn, TgH, Tjo, TgF, TgG;
|
Chris@82
|
1453 TgE = TeM - TeP;
|
Chris@82
|
1454 Tjn = Tjl + Tjm;
|
Chris@82
|
1455 TgF = TeU - TeR;
|
Chris@82
|
1456 TgG = TeW + TeZ;
|
Chris@82
|
1457 TgH = TgF - TgG;
|
Chris@82
|
1458 Tjo = TgF + TgG;
|
Chris@82
|
1459 TgI = FMA(KP707106781, TgH, TgE);
|
Chris@82
|
1460 Tjv = FNMS(KP707106781, Tjo, Tjn);
|
Chris@82
|
1461 Tha = FNMS(KP707106781, TgH, TgE);
|
Chris@82
|
1462 Tjp = FMA(KP707106781, Tjo, Tjn);
|
Chris@82
|
1463 }
|
Chris@82
|
1464 {
|
Chris@82
|
1465 E Tfd, Tgp, Tfo, Tgq;
|
Chris@82
|
1466 {
|
Chris@82
|
1467 E Tf7, Tfc, Tfi, Tfn;
|
Chris@82
|
1468 Tf7 = Tf5 + Tf6;
|
Chris@82
|
1469 Tfc = Tf8 + Tfb;
|
Chris@82
|
1470 Tfd = FNMS(KP414213562, Tfc, Tf7);
|
Chris@82
|
1471 Tgp = FMA(KP414213562, Tf7, Tfc);
|
Chris@82
|
1472 Tfi = Tfg + Tfh;
|
Chris@82
|
1473 Tfn = Tfj + Tfm;
|
Chris@82
|
1474 Tfo = FMA(KP414213562, Tfn, Tfi);
|
Chris@82
|
1475 Tgq = FNMS(KP414213562, Tfi, Tfn);
|
Chris@82
|
1476 }
|
Chris@82
|
1477 Tfp = Tfd - Tfo;
|
Chris@82
|
1478 Tjw = Tgq - Tgp;
|
Chris@82
|
1479 Tgr = Tgp + Tgq;
|
Chris@82
|
1480 Tjq = Tfd + Tfo;
|
Chris@82
|
1481 }
|
Chris@82
|
1482 {
|
Chris@82
|
1483 E Th0, Thj, Th3, Thi;
|
Chris@82
|
1484 {
|
Chris@82
|
1485 E TgY, TgZ, Th1, Th2;
|
Chris@82
|
1486 TgY = Tg9 - Tgc;
|
Chris@82
|
1487 TgZ = Tg6 - Tg1;
|
Chris@82
|
1488 Th0 = FNMS(KP707106781, TgZ, TgY);
|
Chris@82
|
1489 Thj = FMA(KP707106781, TgZ, TgY);
|
Chris@82
|
1490 Th1 = TfU - TfV;
|
Chris@82
|
1491 Th2 = Tge - Tgf;
|
Chris@82
|
1492 Th3 = FNMS(KP707106781, Th2, Th1);
|
Chris@82
|
1493 Thi = FMA(KP707106781, Th2, Th1);
|
Chris@82
|
1494 }
|
Chris@82
|
1495 Th4 = FNMS(KP668178637, Th3, Th0);
|
Chris@82
|
1496 Tho = FMA(KP198912367, Thi, Thj);
|
Chris@82
|
1497 Th8 = FMA(KP668178637, Th0, Th3);
|
Chris@82
|
1498 Thk = FNMS(KP198912367, Thj, Thi);
|
Chris@82
|
1499 }
|
Chris@82
|
1500 {
|
Chris@82
|
1501 E TfH, Tgu, TfQ, Tgt;
|
Chris@82
|
1502 {
|
Chris@82
|
1503 E Tfv, TfG, TfM, TfP;
|
Chris@82
|
1504 Tfv = Tft + Tfu;
|
Chris@82
|
1505 TfG = TfA + TfF;
|
Chris@82
|
1506 TfH = FNMS(KP707106781, TfG, Tfv);
|
Chris@82
|
1507 Tgu = FMA(KP707106781, TfG, Tfv);
|
Chris@82
|
1508 TfM = TfI + TfL;
|
Chris@82
|
1509 TfP = TfN + TfO;
|
Chris@82
|
1510 TfQ = FNMS(KP707106781, TfP, TfM);
|
Chris@82
|
1511 Tgt = FMA(KP707106781, TfP, TfM);
|
Chris@82
|
1512 }
|
Chris@82
|
1513 TfR = FMA(KP668178637, TfQ, TfH);
|
Chris@82
|
1514 TgB = FNMS(KP198912367, Tgt, Tgu);
|
Chris@82
|
1515 Tgl = FNMS(KP668178637, TfH, TfQ);
|
Chris@82
|
1516 Tgv = FMA(KP198912367, Tgu, Tgt);
|
Chris@82
|
1517 }
|
Chris@82
|
1518 {
|
Chris@82
|
1519 E TgL, Thb, TgO, Thc;
|
Chris@82
|
1520 {
|
Chris@82
|
1521 E TgJ, TgK, TgM, TgN;
|
Chris@82
|
1522 TgJ = Tf5 - Tf6;
|
Chris@82
|
1523 TgK = Tf8 - Tfb;
|
Chris@82
|
1524 TgL = FMA(KP414213562, TgK, TgJ);
|
Chris@82
|
1525 Thb = FNMS(KP414213562, TgJ, TgK);
|
Chris@82
|
1526 TgM = Tfg - Tfh;
|
Chris@82
|
1527 TgN = Tfj - Tfm;
|
Chris@82
|
1528 TgO = FNMS(KP414213562, TgN, TgM);
|
Chris@82
|
1529 Thc = FMA(KP414213562, TgM, TgN);
|
Chris@82
|
1530 }
|
Chris@82
|
1531 TgP = TgL - TgO;
|
Chris@82
|
1532 TjK = TgL + TgO;
|
Chris@82
|
1533 Thd = Thb + Thc;
|
Chris@82
|
1534 TjE = Thc - Thb;
|
Chris@82
|
1535 }
|
Chris@82
|
1536 {
|
Chris@82
|
1537 E TgT, Thg, TgW, Thf;
|
Chris@82
|
1538 {
|
Chris@82
|
1539 E TgR, TgS, TgU, TgV;
|
Chris@82
|
1540 TgR = TfI - TfL;
|
Chris@82
|
1541 TgS = TfF - TfA;
|
Chris@82
|
1542 TgT = FNMS(KP707106781, TgS, TgR);
|
Chris@82
|
1543 Thg = FMA(KP707106781, TgS, TgR);
|
Chris@82
|
1544 TgU = Tft - Tfu;
|
Chris@82
|
1545 TgV = TfN - TfO;
|
Chris@82
|
1546 TgW = FNMS(KP707106781, TgV, TgU);
|
Chris@82
|
1547 Thf = FMA(KP707106781, TgV, TgU);
|
Chris@82
|
1548 }
|
Chris@82
|
1549 TgX = FMA(KP668178637, TgW, TgT);
|
Chris@82
|
1550 Thn = FNMS(KP198912367, Thf, Thg);
|
Chris@82
|
1551 Th7 = FNMS(KP668178637, TgT, TgW);
|
Chris@82
|
1552 Thh = FMA(KP198912367, Thg, Thf);
|
Chris@82
|
1553 }
|
Chris@82
|
1554 {
|
Chris@82
|
1555 E Tg8, Tgx, Tgh, Tgw;
|
Chris@82
|
1556 {
|
Chris@82
|
1557 E TfW, Tg7, Tgd, Tgg;
|
Chris@82
|
1558 TfW = TfU + TfV;
|
Chris@82
|
1559 Tg7 = Tg1 + Tg6;
|
Chris@82
|
1560 Tg8 = FNMS(KP707106781, Tg7, TfW);
|
Chris@82
|
1561 Tgx = FMA(KP707106781, Tg7, TfW);
|
Chris@82
|
1562 Tgd = Tg9 + Tgc;
|
Chris@82
|
1563 Tgg = Tge + Tgf;
|
Chris@82
|
1564 Tgh = FNMS(KP707106781, Tgg, Tgd);
|
Chris@82
|
1565 Tgw = FMA(KP707106781, Tgg, Tgd);
|
Chris@82
|
1566 }
|
Chris@82
|
1567 Tgi = FNMS(KP668178637, Tgh, Tg8);
|
Chris@82
|
1568 TgC = FMA(KP198912367, Tgw, Tgx);
|
Chris@82
|
1569 Tgm = FMA(KP668178637, Tg8, Tgh);
|
Chris@82
|
1570 Tgy = FNMS(KP198912367, Tgx, Tgw);
|
Chris@82
|
1571 }
|
Chris@82
|
1572 {
|
Chris@82
|
1573 E Tfq, Tgj, Tjx, Tjy;
|
Chris@82
|
1574 Tfq = FMA(KP923879532, Tfp, Tf2);
|
Chris@82
|
1575 Tgj = TfR - Tgi;
|
Chris@82
|
1576 ri[WS(rs, 42)] = FNMS(KP831469612, Tgj, Tfq);
|
Chris@82
|
1577 ri[WS(rs, 10)] = FMA(KP831469612, Tgj, Tfq);
|
Chris@82
|
1578 Tjx = FMA(KP923879532, Tjw, Tjv);
|
Chris@82
|
1579 Tjy = Tgm - Tgl;
|
Chris@82
|
1580 ii[WS(rs, 10)] = FMA(KP831469612, Tjy, Tjx);
|
Chris@82
|
1581 ii[WS(rs, 42)] = FNMS(KP831469612, Tjy, Tjx);
|
Chris@82
|
1582 }
|
Chris@82
|
1583 {
|
Chris@82
|
1584 E Tgk, Tgn, Tjz, TjA;
|
Chris@82
|
1585 Tgk = FNMS(KP923879532, Tfp, Tf2);
|
Chris@82
|
1586 Tgn = Tgl + Tgm;
|
Chris@82
|
1587 ri[WS(rs, 26)] = FNMS(KP831469612, Tgn, Tgk);
|
Chris@82
|
1588 ri[WS(rs, 58)] = FMA(KP831469612, Tgn, Tgk);
|
Chris@82
|
1589 Tjz = FNMS(KP923879532, Tjw, Tjv);
|
Chris@82
|
1590 TjA = TfR + Tgi;
|
Chris@82
|
1591 ii[WS(rs, 26)] = FNMS(KP831469612, TjA, Tjz);
|
Chris@82
|
1592 ii[WS(rs, 58)] = FMA(KP831469612, TjA, Tjz);
|
Chris@82
|
1593 }
|
Chris@82
|
1594 {
|
Chris@82
|
1595 E Tgs, Tgz, Tjr, Tjs;
|
Chris@82
|
1596 Tgs = FMA(KP923879532, Tgr, Tgo);
|
Chris@82
|
1597 Tgz = Tgv + Tgy;
|
Chris@82
|
1598 ri[WS(rs, 34)] = FNMS(KP980785280, Tgz, Tgs);
|
Chris@82
|
1599 ri[WS(rs, 2)] = FMA(KP980785280, Tgz, Tgs);
|
Chris@82
|
1600 Tjr = FMA(KP923879532, Tjq, Tjp);
|
Chris@82
|
1601 Tjs = TgB + TgC;
|
Chris@82
|
1602 ii[WS(rs, 2)] = FMA(KP980785280, Tjs, Tjr);
|
Chris@82
|
1603 ii[WS(rs, 34)] = FNMS(KP980785280, Tjs, Tjr);
|
Chris@82
|
1604 }
|
Chris@82
|
1605 {
|
Chris@82
|
1606 E TgA, TgD, Tjt, Tju;
|
Chris@82
|
1607 TgA = FNMS(KP923879532, Tgr, Tgo);
|
Chris@82
|
1608 TgD = TgB - TgC;
|
Chris@82
|
1609 ri[WS(rs, 50)] = FNMS(KP980785280, TgD, TgA);
|
Chris@82
|
1610 ri[WS(rs, 18)] = FMA(KP980785280, TgD, TgA);
|
Chris@82
|
1611 Tjt = FNMS(KP923879532, Tjq, Tjp);
|
Chris@82
|
1612 Tju = Tgy - Tgv;
|
Chris@82
|
1613 ii[WS(rs, 18)] = FMA(KP980785280, Tju, Tjt);
|
Chris@82
|
1614 ii[WS(rs, 50)] = FNMS(KP980785280, Tju, Tjt);
|
Chris@82
|
1615 }
|
Chris@82
|
1616 {
|
Chris@82
|
1617 E TgQ, Th5, TjF, TjG;
|
Chris@82
|
1618 TgQ = FMA(KP923879532, TgP, TgI);
|
Chris@82
|
1619 Th5 = TgX + Th4;
|
Chris@82
|
1620 ri[WS(rs, 38)] = FNMS(KP831469612, Th5, TgQ);
|
Chris@82
|
1621 ri[WS(rs, 6)] = FMA(KP831469612, Th5, TgQ);
|
Chris@82
|
1622 TjF = FMA(KP923879532, TjE, TjD);
|
Chris@82
|
1623 TjG = Th7 + Th8;
|
Chris@82
|
1624 ii[WS(rs, 6)] = FMA(KP831469612, TjG, TjF);
|
Chris@82
|
1625 ii[WS(rs, 38)] = FNMS(KP831469612, TjG, TjF);
|
Chris@82
|
1626 }
|
Chris@82
|
1627 {
|
Chris@82
|
1628 E Th6, Th9, TjH, TjI;
|
Chris@82
|
1629 Th6 = FNMS(KP923879532, TgP, TgI);
|
Chris@82
|
1630 Th9 = Th7 - Th8;
|
Chris@82
|
1631 ri[WS(rs, 54)] = FNMS(KP831469612, Th9, Th6);
|
Chris@82
|
1632 ri[WS(rs, 22)] = FMA(KP831469612, Th9, Th6);
|
Chris@82
|
1633 TjH = FNMS(KP923879532, TjE, TjD);
|
Chris@82
|
1634 TjI = Th4 - TgX;
|
Chris@82
|
1635 ii[WS(rs, 22)] = FMA(KP831469612, TjI, TjH);
|
Chris@82
|
1636 ii[WS(rs, 54)] = FNMS(KP831469612, TjI, TjH);
|
Chris@82
|
1637 }
|
Chris@82
|
1638 {
|
Chris@82
|
1639 E The, Thl, TjL, TjM;
|
Chris@82
|
1640 The = FNMS(KP923879532, Thd, Tha);
|
Chris@82
|
1641 Thl = Thh - Thk;
|
Chris@82
|
1642 ri[WS(rs, 46)] = FNMS(KP980785280, Thl, The);
|
Chris@82
|
1643 ri[WS(rs, 14)] = FMA(KP980785280, Thl, The);
|
Chris@82
|
1644 TjL = FNMS(KP923879532, TjK, TjJ);
|
Chris@82
|
1645 TjM = Tho - Thn;
|
Chris@82
|
1646 ii[WS(rs, 14)] = FMA(KP980785280, TjM, TjL);
|
Chris@82
|
1647 ii[WS(rs, 46)] = FNMS(KP980785280, TjM, TjL);
|
Chris@82
|
1648 }
|
Chris@82
|
1649 {
|
Chris@82
|
1650 E Thm, Thp, TjN, TjO;
|
Chris@82
|
1651 Thm = FMA(KP923879532, Thd, Tha);
|
Chris@82
|
1652 Thp = Thn + Tho;
|
Chris@82
|
1653 ri[WS(rs, 30)] = FNMS(KP980785280, Thp, Thm);
|
Chris@82
|
1654 ri[WS(rs, 62)] = FMA(KP980785280, Thp, Thm);
|
Chris@82
|
1655 TjN = FMA(KP923879532, TjK, TjJ);
|
Chris@82
|
1656 TjO = Thh + Thk;
|
Chris@82
|
1657 ii[WS(rs, 30)] = FNMS(KP980785280, TjO, TjN);
|
Chris@82
|
1658 ii[WS(rs, 62)] = FMA(KP980785280, TjO, TjN);
|
Chris@82
|
1659 }
|
Chris@82
|
1660 }
|
Chris@82
|
1661 {
|
Chris@82
|
1662 E T99, Tkw, TbB, Tkq, Taj, TbL, Tbv, TbF, Tce, Tcy, Tci, Tcu, Tc7, Tcx, Tch;
|
Chris@82
|
1663 E Tcr, TbZ, TkK, Tcn, TkE, Tbs, TbM, Tbw, TbI, T80, TkD, TkJ, Tby, TbS, Tkp;
|
Chris@82
|
1664 E Tkv, Tck;
|
Chris@82
|
1665 {
|
Chris@82
|
1666 E T8z, Tbz, T98, TbA;
|
Chris@82
|
1667 {
|
Chris@82
|
1668 E T8n, T8y, T8W, T97;
|
Chris@82
|
1669 T8n = FNMS(KP707106781, T8m, T87);
|
Chris@82
|
1670 T8y = FNMS(KP707106781, T8x, T8u);
|
Chris@82
|
1671 T8z = FNMS(KP668178637, T8y, T8n);
|
Chris@82
|
1672 Tbz = FMA(KP668178637, T8n, T8y);
|
Chris@82
|
1673 T8W = FNMS(KP707106781, T8V, T8G);
|
Chris@82
|
1674 T97 = FNMS(KP707106781, T96, T93);
|
Chris@82
|
1675 T98 = FMA(KP668178637, T97, T8W);
|
Chris@82
|
1676 TbA = FNMS(KP668178637, T8W, T97);
|
Chris@82
|
1677 }
|
Chris@82
|
1678 T99 = T8z - T98;
|
Chris@82
|
1679 Tkw = TbA - Tbz;
|
Chris@82
|
1680 TbB = Tbz + TbA;
|
Chris@82
|
1681 Tkq = T8z + T98;
|
Chris@82
|
1682 }
|
Chris@82
|
1683 {
|
Chris@82
|
1684 E Ta3, TbE, Tai, TbD;
|
Chris@82
|
1685 {
|
Chris@82
|
1686 E T9x, Ta2, Tae, Tah;
|
Chris@82
|
1687 T9x = FNMS(KP707106781, T9w, T9h);
|
Chris@82
|
1688 Ta2 = T9M - Ta1;
|
Chris@82
|
1689 Ta3 = FNMS(KP923879532, Ta2, T9x);
|
Chris@82
|
1690 TbE = FMA(KP923879532, Ta2, T9x);
|
Chris@82
|
1691 Tae = FNMS(KP707106781, Tad, Taa);
|
Chris@82
|
1692 Tah = Taf - Tag;
|
Chris@82
|
1693 Tai = FNMS(KP923879532, Tah, Tae);
|
Chris@82
|
1694 TbD = FMA(KP923879532, Tah, Tae);
|
Chris@82
|
1695 }
|
Chris@82
|
1696 Taj = FMA(KP534511135, Tai, Ta3);
|
Chris@82
|
1697 TbL = FNMS(KP303346683, TbD, TbE);
|
Chris@82
|
1698 Tbv = FNMS(KP534511135, Ta3, Tai);
|
Chris@82
|
1699 TbF = FMA(KP303346683, TbE, TbD);
|
Chris@82
|
1700 }
|
Chris@82
|
1701 {
|
Chris@82
|
1702 E Tca, Tct, Tcd, Tcs;
|
Chris@82
|
1703 {
|
Chris@82
|
1704 E Tc8, Tc9, Tcb, Tcc;
|
Chris@82
|
1705 Tc8 = FMA(KP707106781, Tbm, Tbj);
|
Chris@82
|
1706 Tc9 = Tba + TaV;
|
Chris@82
|
1707 Tca = FNMS(KP923879532, Tc9, Tc8);
|
Chris@82
|
1708 Tct = FMA(KP923879532, Tc9, Tc8);
|
Chris@82
|
1709 Tcb = FMA(KP707106781, TaF, Taq);
|
Chris@82
|
1710 Tcc = Tbo + Tbp;
|
Chris@82
|
1711 Tcd = FNMS(KP923879532, Tcc, Tcb);
|
Chris@82
|
1712 Tcs = FMA(KP923879532, Tcc, Tcb);
|
Chris@82
|
1713 }
|
Chris@82
|
1714 Tce = FNMS(KP820678790, Tcd, Tca);
|
Chris@82
|
1715 Tcy = FMA(KP098491403, Tcs, Tct);
|
Chris@82
|
1716 Tci = FMA(KP820678790, Tca, Tcd);
|
Chris@82
|
1717 Tcu = FNMS(KP098491403, Tct, Tcs);
|
Chris@82
|
1718 }
|
Chris@82
|
1719 {
|
Chris@82
|
1720 E Tc3, Tcq, Tc6, Tcp;
|
Chris@82
|
1721 {
|
Chris@82
|
1722 E Tc1, Tc2, Tc4, Tc5;
|
Chris@82
|
1723 Tc1 = FMA(KP707106781, Tad, Taa);
|
Chris@82
|
1724 Tc2 = Ta1 + T9M;
|
Chris@82
|
1725 Tc3 = FNMS(KP923879532, Tc2, Tc1);
|
Chris@82
|
1726 Tcq = FMA(KP923879532, Tc2, Tc1);
|
Chris@82
|
1727 Tc4 = FMA(KP707106781, T9w, T9h);
|
Chris@82
|
1728 Tc5 = Taf + Tag;
|
Chris@82
|
1729 Tc6 = FNMS(KP923879532, Tc5, Tc4);
|
Chris@82
|
1730 Tcp = FMA(KP923879532, Tc5, Tc4);
|
Chris@82
|
1731 }
|
Chris@82
|
1732 Tc7 = FMA(KP820678790, Tc6, Tc3);
|
Chris@82
|
1733 Tcx = FNMS(KP098491403, Tcp, Tcq);
|
Chris@82
|
1734 Tch = FNMS(KP820678790, Tc3, Tc6);
|
Chris@82
|
1735 Tcr = FMA(KP098491403, Tcq, Tcp);
|
Chris@82
|
1736 }
|
Chris@82
|
1737 {
|
Chris@82
|
1738 E TbV, Tcl, TbY, Tcm;
|
Chris@82
|
1739 {
|
Chris@82
|
1740 E TbT, TbU, TbW, TbX;
|
Chris@82
|
1741 TbT = FMA(KP707106781, T8m, T87);
|
Chris@82
|
1742 TbU = FMA(KP707106781, T8x, T8u);
|
Chris@82
|
1743 TbV = FMA(KP198912367, TbU, TbT);
|
Chris@82
|
1744 Tcl = FNMS(KP198912367, TbT, TbU);
|
Chris@82
|
1745 TbW = FMA(KP707106781, T8V, T8G);
|
Chris@82
|
1746 TbX = FMA(KP707106781, T96, T93);
|
Chris@82
|
1747 TbY = FNMS(KP198912367, TbX, TbW);
|
Chris@82
|
1748 Tcm = FMA(KP198912367, TbW, TbX);
|
Chris@82
|
1749 }
|
Chris@82
|
1750 TbZ = TbV - TbY;
|
Chris@82
|
1751 TkK = TbV + TbY;
|
Chris@82
|
1752 Tcn = Tcl + Tcm;
|
Chris@82
|
1753 TkE = Tcm - Tcl;
|
Chris@82
|
1754 }
|
Chris@82
|
1755 {
|
Chris@82
|
1756 E Tbc, TbH, Tbr, TbG;
|
Chris@82
|
1757 {
|
Chris@82
|
1758 E TaG, Tbb, Tbn, Tbq;
|
Chris@82
|
1759 TaG = FNMS(KP707106781, TaF, Taq);
|
Chris@82
|
1760 Tbb = TaV - Tba;
|
Chris@82
|
1761 Tbc = FNMS(KP923879532, Tbb, TaG);
|
Chris@82
|
1762 TbH = FMA(KP923879532, Tbb, TaG);
|
Chris@82
|
1763 Tbn = FNMS(KP707106781, Tbm, Tbj);
|
Chris@82
|
1764 Tbq = Tbo - Tbp;
|
Chris@82
|
1765 Tbr = FNMS(KP923879532, Tbq, Tbn);
|
Chris@82
|
1766 TbG = FMA(KP923879532, Tbq, Tbn);
|
Chris@82
|
1767 }
|
Chris@82
|
1768 Tbs = FNMS(KP534511135, Tbr, Tbc);
|
Chris@82
|
1769 TbM = FMA(KP303346683, TbG, TbH);
|
Chris@82
|
1770 Tbw = FMA(KP534511135, Tbc, Tbr);
|
Chris@82
|
1771 TbI = FNMS(KP303346683, TbH, TbG);
|
Chris@82
|
1772 }
|
Chris@82
|
1773 {
|
Chris@82
|
1774 E T7u, TbO, Tkn, TkB, T7Z, TkC, TbR, Tko, T7t, Tkm;
|
Chris@82
|
1775 T7t = T7l - T7s;
|
Chris@82
|
1776 T7u = FMA(KP707106781, T7t, T7e);
|
Chris@82
|
1777 TbO = FNMS(KP707106781, T7t, T7e);
|
Chris@82
|
1778 Tkm = TcC - TcB;
|
Chris@82
|
1779 Tkn = FMA(KP707106781, Tkm, Tkl);
|
Chris@82
|
1780 TkB = FNMS(KP707106781, Tkm, Tkl);
|
Chris@82
|
1781 {
|
Chris@82
|
1782 E T7J, T7Y, TbP, TbQ;
|
Chris@82
|
1783 T7J = FMA(KP414213562, T7I, T7B);
|
Chris@82
|
1784 T7Y = FNMS(KP414213562, T7X, T7Q);
|
Chris@82
|
1785 T7Z = T7J - T7Y;
|
Chris@82
|
1786 TkC = T7J + T7Y;
|
Chris@82
|
1787 TbP = FNMS(KP414213562, T7B, T7I);
|
Chris@82
|
1788 TbQ = FMA(KP414213562, T7Q, T7X);
|
Chris@82
|
1789 TbR = TbP + TbQ;
|
Chris@82
|
1790 Tko = TbQ - TbP;
|
Chris@82
|
1791 }
|
Chris@82
|
1792 T80 = FNMS(KP923879532, T7Z, T7u);
|
Chris@82
|
1793 TkD = FNMS(KP923879532, TkC, TkB);
|
Chris@82
|
1794 TkJ = FMA(KP923879532, TkC, TkB);
|
Chris@82
|
1795 Tby = FMA(KP923879532, T7Z, T7u);
|
Chris@82
|
1796 TbS = FNMS(KP923879532, TbR, TbO);
|
Chris@82
|
1797 Tkp = FMA(KP923879532, Tko, Tkn);
|
Chris@82
|
1798 Tkv = FNMS(KP923879532, Tko, Tkn);
|
Chris@82
|
1799 Tck = FMA(KP923879532, TbR, TbO);
|
Chris@82
|
1800 }
|
Chris@82
|
1801 {
|
Chris@82
|
1802 E T9a, Tbt, Tkx, Tky;
|
Chris@82
|
1803 T9a = FMA(KP831469612, T99, T80);
|
Chris@82
|
1804 Tbt = Taj - Tbs;
|
Chris@82
|
1805 ri[WS(rs, 43)] = FNMS(KP881921264, Tbt, T9a);
|
Chris@82
|
1806 ri[WS(rs, 11)] = FMA(KP881921264, Tbt, T9a);
|
Chris@82
|
1807 Tkx = FMA(KP831469612, Tkw, Tkv);
|
Chris@82
|
1808 Tky = Tbw - Tbv;
|
Chris@82
|
1809 ii[WS(rs, 11)] = FMA(KP881921264, Tky, Tkx);
|
Chris@82
|
1810 ii[WS(rs, 43)] = FNMS(KP881921264, Tky, Tkx);
|
Chris@82
|
1811 }
|
Chris@82
|
1812 {
|
Chris@82
|
1813 E Tbu, Tbx, Tkz, TkA;
|
Chris@82
|
1814 Tbu = FNMS(KP831469612, T99, T80);
|
Chris@82
|
1815 Tbx = Tbv + Tbw;
|
Chris@82
|
1816 ri[WS(rs, 27)] = FNMS(KP881921264, Tbx, Tbu);
|
Chris@82
|
1817 ri[WS(rs, 59)] = FMA(KP881921264, Tbx, Tbu);
|
Chris@82
|
1818 Tkz = FNMS(KP831469612, Tkw, Tkv);
|
Chris@82
|
1819 TkA = Taj + Tbs;
|
Chris@82
|
1820 ii[WS(rs, 27)] = FNMS(KP881921264, TkA, Tkz);
|
Chris@82
|
1821 ii[WS(rs, 59)] = FMA(KP881921264, TkA, Tkz);
|
Chris@82
|
1822 }
|
Chris@82
|
1823 {
|
Chris@82
|
1824 E TbC, TbJ, Tkr, Tks;
|
Chris@82
|
1825 TbC = FMA(KP831469612, TbB, Tby);
|
Chris@82
|
1826 TbJ = TbF + TbI;
|
Chris@82
|
1827 ri[WS(rs, 35)] = FNMS(KP956940335, TbJ, TbC);
|
Chris@82
|
1828 ri[WS(rs, 3)] = FMA(KP956940335, TbJ, TbC);
|
Chris@82
|
1829 Tkr = FMA(KP831469612, Tkq, Tkp);
|
Chris@82
|
1830 Tks = TbL + TbM;
|
Chris@82
|
1831 ii[WS(rs, 3)] = FMA(KP956940335, Tks, Tkr);
|
Chris@82
|
1832 ii[WS(rs, 35)] = FNMS(KP956940335, Tks, Tkr);
|
Chris@82
|
1833 }
|
Chris@82
|
1834 {
|
Chris@82
|
1835 E TbK, TbN, Tkt, Tku;
|
Chris@82
|
1836 TbK = FNMS(KP831469612, TbB, Tby);
|
Chris@82
|
1837 TbN = TbL - TbM;
|
Chris@82
|
1838 ri[WS(rs, 51)] = FNMS(KP956940335, TbN, TbK);
|
Chris@82
|
1839 ri[WS(rs, 19)] = FMA(KP956940335, TbN, TbK);
|
Chris@82
|
1840 Tkt = FNMS(KP831469612, Tkq, Tkp);
|
Chris@82
|
1841 Tku = TbI - TbF;
|
Chris@82
|
1842 ii[WS(rs, 19)] = FMA(KP956940335, Tku, Tkt);
|
Chris@82
|
1843 ii[WS(rs, 51)] = FNMS(KP956940335, Tku, Tkt);
|
Chris@82
|
1844 }
|
Chris@82
|
1845 {
|
Chris@82
|
1846 E Tc0, Tcf, TkF, TkG;
|
Chris@82
|
1847 Tc0 = FMA(KP980785280, TbZ, TbS);
|
Chris@82
|
1848 Tcf = Tc7 + Tce;
|
Chris@82
|
1849 ri[WS(rs, 39)] = FNMS(KP773010453, Tcf, Tc0);
|
Chris@82
|
1850 ri[WS(rs, 7)] = FMA(KP773010453, Tcf, Tc0);
|
Chris@82
|
1851 TkF = FMA(KP980785280, TkE, TkD);
|
Chris@82
|
1852 TkG = Tch + Tci;
|
Chris@82
|
1853 ii[WS(rs, 7)] = FMA(KP773010453, TkG, TkF);
|
Chris@82
|
1854 ii[WS(rs, 39)] = FNMS(KP773010453, TkG, TkF);
|
Chris@82
|
1855 }
|
Chris@82
|
1856 {
|
Chris@82
|
1857 E Tcg, Tcj, TkH, TkI;
|
Chris@82
|
1858 Tcg = FNMS(KP980785280, TbZ, TbS);
|
Chris@82
|
1859 Tcj = Tch - Tci;
|
Chris@82
|
1860 ri[WS(rs, 55)] = FNMS(KP773010453, Tcj, Tcg);
|
Chris@82
|
1861 ri[WS(rs, 23)] = FMA(KP773010453, Tcj, Tcg);
|
Chris@82
|
1862 TkH = FNMS(KP980785280, TkE, TkD);
|
Chris@82
|
1863 TkI = Tce - Tc7;
|
Chris@82
|
1864 ii[WS(rs, 23)] = FMA(KP773010453, TkI, TkH);
|
Chris@82
|
1865 ii[WS(rs, 55)] = FNMS(KP773010453, TkI, TkH);
|
Chris@82
|
1866 }
|
Chris@82
|
1867 {
|
Chris@82
|
1868 E Tco, Tcv, TkL, TkM;
|
Chris@82
|
1869 Tco = FNMS(KP980785280, Tcn, Tck);
|
Chris@82
|
1870 Tcv = Tcr - Tcu;
|
Chris@82
|
1871 ri[WS(rs, 47)] = FNMS(KP995184726, Tcv, Tco);
|
Chris@82
|
1872 ri[WS(rs, 15)] = FMA(KP995184726, Tcv, Tco);
|
Chris@82
|
1873 TkL = FNMS(KP980785280, TkK, TkJ);
|
Chris@82
|
1874 TkM = Tcy - Tcx;
|
Chris@82
|
1875 ii[WS(rs, 15)] = FMA(KP995184726, TkM, TkL);
|
Chris@82
|
1876 ii[WS(rs, 47)] = FNMS(KP995184726, TkM, TkL);
|
Chris@82
|
1877 }
|
Chris@82
|
1878 {
|
Chris@82
|
1879 E Tcw, Tcz, TkN, TkO;
|
Chris@82
|
1880 Tcw = FMA(KP980785280, Tcn, Tck);
|
Chris@82
|
1881 Tcz = Tcx + Tcy;
|
Chris@82
|
1882 ri[WS(rs, 31)] = FNMS(KP995184726, Tcz, Tcw);
|
Chris@82
|
1883 ri[WS(rs, 63)] = FMA(KP995184726, Tcz, Tcw);
|
Chris@82
|
1884 TkN = FMA(KP980785280, TkK, TkJ);
|
Chris@82
|
1885 TkO = Tcr + Tcu;
|
Chris@82
|
1886 ii[WS(rs, 31)] = FNMS(KP995184726, TkO, TkN);
|
Chris@82
|
1887 ii[WS(rs, 63)] = FMA(KP995184726, TkO, TkN);
|
Chris@82
|
1888 }
|
Chris@82
|
1889 }
|
Chris@82
|
1890 {
|
Chris@82
|
1891 E Td1, Tk2, TdN, TjW, Tdl, TdX, TdH, TdR, Teq, TeK, Teu, TeG, Tej, TeJ, Tet;
|
Chris@82
|
1892 E TeD, Teb, Tkg, Tez, Tka, TdE, TdY, TdI, TdU, TcM, Tk9, Tkf, TdK, Te4, TjV;
|
Chris@82
|
1893 E Tk1, Tew;
|
Chris@82
|
1894 {
|
Chris@82
|
1895 E TcT, TdL, Td0, TdM;
|
Chris@82
|
1896 {
|
Chris@82
|
1897 E TcP, TcS, TcW, TcZ;
|
Chris@82
|
1898 TcP = FMA(KP707106781, TcO, TcN);
|
Chris@82
|
1899 TcS = FMA(KP707106781, TcR, TcQ);
|
Chris@82
|
1900 TcT = FNMS(KP198912367, TcS, TcP);
|
Chris@82
|
1901 TdL = FMA(KP198912367, TcP, TcS);
|
Chris@82
|
1902 TcW = FMA(KP707106781, TcV, TcU);
|
Chris@82
|
1903 TcZ = FMA(KP707106781, TcY, TcX);
|
Chris@82
|
1904 Td0 = FMA(KP198912367, TcZ, TcW);
|
Chris@82
|
1905 TdM = FNMS(KP198912367, TcW, TcZ);
|
Chris@82
|
1906 }
|
Chris@82
|
1907 Td1 = TcT - Td0;
|
Chris@82
|
1908 Tk2 = TdM - TdL;
|
Chris@82
|
1909 TdN = TdL + TdM;
|
Chris@82
|
1910 TjW = TcT + Td0;
|
Chris@82
|
1911 }
|
Chris@82
|
1912 {
|
Chris@82
|
1913 E Tdd, TdQ, Tdk, TdP;
|
Chris@82
|
1914 {
|
Chris@82
|
1915 E Td5, Tdc, Tdg, Tdj;
|
Chris@82
|
1916 Td5 = FMA(KP707106781, Td4, Td3);
|
Chris@82
|
1917 Tdc = Td8 + Tdb;
|
Chris@82
|
1918 Tdd = FNMS(KP923879532, Tdc, Td5);
|
Chris@82
|
1919 TdQ = FMA(KP923879532, Tdc, Td5);
|
Chris@82
|
1920 Tdg = FMA(KP707106781, Tdf, Tde);
|
Chris@82
|
1921 Tdj = Tdh + Tdi;
|
Chris@82
|
1922 Tdk = FNMS(KP923879532, Tdj, Tdg);
|
Chris@82
|
1923 TdP = FMA(KP923879532, Tdj, Tdg);
|
Chris@82
|
1924 }
|
Chris@82
|
1925 Tdl = FMA(KP820678790, Tdk, Tdd);
|
Chris@82
|
1926 TdX = FNMS(KP098491403, TdP, TdQ);
|
Chris@82
|
1927 TdH = FNMS(KP820678790, Tdd, Tdk);
|
Chris@82
|
1928 TdR = FMA(KP098491403, TdQ, TdP);
|
Chris@82
|
1929 }
|
Chris@82
|
1930 {
|
Chris@82
|
1931 E Tem, TeF, Tep, TeE;
|
Chris@82
|
1932 {
|
Chris@82
|
1933 E Tek, Tel, Ten, Teo;
|
Chris@82
|
1934 Tek = FNMS(KP707106781, Tdy, Tdx);
|
Chris@82
|
1935 Tel = Tdu - Tdr;
|
Chris@82
|
1936 Tem = FNMS(KP923879532, Tel, Tek);
|
Chris@82
|
1937 TeF = FMA(KP923879532, Tel, Tek);
|
Chris@82
|
1938 Ten = FNMS(KP707106781, Tdn, Tdm);
|
Chris@82
|
1939 Teo = TdA - TdB;
|
Chris@82
|
1940 Tep = FNMS(KP923879532, Teo, Ten);
|
Chris@82
|
1941 TeE = FMA(KP923879532, Teo, Ten);
|
Chris@82
|
1942 }
|
Chris@82
|
1943 Teq = FNMS(KP534511135, Tep, Tem);
|
Chris@82
|
1944 TeK = FMA(KP303346683, TeE, TeF);
|
Chris@82
|
1945 Teu = FMA(KP534511135, Tem, Tep);
|
Chris@82
|
1946 TeG = FNMS(KP303346683, TeF, TeE);
|
Chris@82
|
1947 }
|
Chris@82
|
1948 {
|
Chris@82
|
1949 E Tef, TeC, Tei, TeB;
|
Chris@82
|
1950 {
|
Chris@82
|
1951 E Ted, Tee, Teg, Teh;
|
Chris@82
|
1952 Ted = FNMS(KP707106781, Tdf, Tde);
|
Chris@82
|
1953 Tee = Tdb - Td8;
|
Chris@82
|
1954 Tef = FNMS(KP923879532, Tee, Ted);
|
Chris@82
|
1955 TeC = FMA(KP923879532, Tee, Ted);
|
Chris@82
|
1956 Teg = FNMS(KP707106781, Td4, Td3);
|
Chris@82
|
1957 Teh = Tdh - Tdi;
|
Chris@82
|
1958 Tei = FNMS(KP923879532, Teh, Teg);
|
Chris@82
|
1959 TeB = FMA(KP923879532, Teh, Teg);
|
Chris@82
|
1960 }
|
Chris@82
|
1961 Tej = FMA(KP534511135, Tei, Tef);
|
Chris@82
|
1962 TeJ = FNMS(KP303346683, TeB, TeC);
|
Chris@82
|
1963 Tet = FNMS(KP534511135, Tef, Tei);
|
Chris@82
|
1964 TeD = FMA(KP303346683, TeC, TeB);
|
Chris@82
|
1965 }
|
Chris@82
|
1966 {
|
Chris@82
|
1967 E Te7, Tex, Tea, Tey;
|
Chris@82
|
1968 {
|
Chris@82
|
1969 E Te5, Te6, Te8, Te9;
|
Chris@82
|
1970 Te5 = FNMS(KP707106781, TcO, TcN);
|
Chris@82
|
1971 Te6 = FNMS(KP707106781, TcR, TcQ);
|
Chris@82
|
1972 Te7 = FMA(KP668178637, Te6, Te5);
|
Chris@82
|
1973 Tex = FNMS(KP668178637, Te5, Te6);
|
Chris@82
|
1974 Te8 = FNMS(KP707106781, TcV, TcU);
|
Chris@82
|
1975 Te9 = FNMS(KP707106781, TcY, TcX);
|
Chris@82
|
1976 Tea = FNMS(KP668178637, Te9, Te8);
|
Chris@82
|
1977 Tey = FMA(KP668178637, Te8, Te9);
|
Chris@82
|
1978 }
|
Chris@82
|
1979 Teb = Te7 - Tea;
|
Chris@82
|
1980 Tkg = Te7 + Tea;
|
Chris@82
|
1981 Tez = Tex + Tey;
|
Chris@82
|
1982 Tka = Tey - Tex;
|
Chris@82
|
1983 }
|
Chris@82
|
1984 {
|
Chris@82
|
1985 E Tdw, TdT, TdD, TdS;
|
Chris@82
|
1986 {
|
Chris@82
|
1987 E Tdo, Tdv, Tdz, TdC;
|
Chris@82
|
1988 Tdo = FMA(KP707106781, Tdn, Tdm);
|
Chris@82
|
1989 Tdv = Tdr + Tdu;
|
Chris@82
|
1990 Tdw = FNMS(KP923879532, Tdv, Tdo);
|
Chris@82
|
1991 TdT = FMA(KP923879532, Tdv, Tdo);
|
Chris@82
|
1992 Tdz = FMA(KP707106781, Tdy, Tdx);
|
Chris@82
|
1993 TdC = TdA + TdB;
|
Chris@82
|
1994 TdD = FNMS(KP923879532, TdC, Tdz);
|
Chris@82
|
1995 TdS = FMA(KP923879532, TdC, Tdz);
|
Chris@82
|
1996 }
|
Chris@82
|
1997 TdE = FNMS(KP820678790, TdD, Tdw);
|
Chris@82
|
1998 TdY = FMA(KP098491403, TdS, TdT);
|
Chris@82
|
1999 TdI = FMA(KP820678790, Tdw, TdD);
|
Chris@82
|
2000 TdU = FNMS(KP098491403, TdT, TdS);
|
Chris@82
|
2001 }
|
Chris@82
|
2002 {
|
Chris@82
|
2003 E TcE, Te0, TjT, Tk7, TcL, Tk8, Te3, TjU, TcD, TjS;
|
Chris@82
|
2004 TcD = TcB + TcC;
|
Chris@82
|
2005 TcE = FMA(KP707106781, TcD, TcA);
|
Chris@82
|
2006 Te0 = FNMS(KP707106781, TcD, TcA);
|
Chris@82
|
2007 TjS = T7l + T7s;
|
Chris@82
|
2008 TjT = FMA(KP707106781, TjS, TjR);
|
Chris@82
|
2009 Tk7 = FNMS(KP707106781, TjS, TjR);
|
Chris@82
|
2010 {
|
Chris@82
|
2011 E TcH, TcK, Te1, Te2;
|
Chris@82
|
2012 TcH = FMA(KP414213562, TcG, TcF);
|
Chris@82
|
2013 TcK = FNMS(KP414213562, TcJ, TcI);
|
Chris@82
|
2014 TcL = TcH + TcK;
|
Chris@82
|
2015 Tk8 = TcK - TcH;
|
Chris@82
|
2016 Te1 = FNMS(KP414213562, TcF, TcG);
|
Chris@82
|
2017 Te2 = FMA(KP414213562, TcI, TcJ);
|
Chris@82
|
2018 Te3 = Te1 - Te2;
|
Chris@82
|
2019 TjU = Te1 + Te2;
|
Chris@82
|
2020 }
|
Chris@82
|
2021 TcM = FNMS(KP923879532, TcL, TcE);
|
Chris@82
|
2022 Tk9 = FMA(KP923879532, Tk8, Tk7);
|
Chris@82
|
2023 Tkf = FNMS(KP923879532, Tk8, Tk7);
|
Chris@82
|
2024 TdK = FMA(KP923879532, TcL, TcE);
|
Chris@82
|
2025 Te4 = FMA(KP923879532, Te3, Te0);
|
Chris@82
|
2026 TjV = FMA(KP923879532, TjU, TjT);
|
Chris@82
|
2027 Tk1 = FNMS(KP923879532, TjU, TjT);
|
Chris@82
|
2028 Tew = FNMS(KP923879532, Te3, Te0);
|
Chris@82
|
2029 }
|
Chris@82
|
2030 {
|
Chris@82
|
2031 E Td2, TdF, Tk3, Tk4;
|
Chris@82
|
2032 Td2 = FMA(KP980785280, Td1, TcM);
|
Chris@82
|
2033 TdF = Tdl - TdE;
|
Chris@82
|
2034 ri[WS(rs, 41)] = FNMS(KP773010453, TdF, Td2);
|
Chris@82
|
2035 ri[WS(rs, 9)] = FMA(KP773010453, TdF, Td2);
|
Chris@82
|
2036 Tk3 = FMA(KP980785280, Tk2, Tk1);
|
Chris@82
|
2037 Tk4 = TdI - TdH;
|
Chris@82
|
2038 ii[WS(rs, 9)] = FMA(KP773010453, Tk4, Tk3);
|
Chris@82
|
2039 ii[WS(rs, 41)] = FNMS(KP773010453, Tk4, Tk3);
|
Chris@82
|
2040 }
|
Chris@82
|
2041 {
|
Chris@82
|
2042 E TdG, TdJ, Tk5, Tk6;
|
Chris@82
|
2043 TdG = FNMS(KP980785280, Td1, TcM);
|
Chris@82
|
2044 TdJ = TdH + TdI;
|
Chris@82
|
2045 ri[WS(rs, 25)] = FNMS(KP773010453, TdJ, TdG);
|
Chris@82
|
2046 ri[WS(rs, 57)] = FMA(KP773010453, TdJ, TdG);
|
Chris@82
|
2047 Tk5 = FNMS(KP980785280, Tk2, Tk1);
|
Chris@82
|
2048 Tk6 = Tdl + TdE;
|
Chris@82
|
2049 ii[WS(rs, 25)] = FNMS(KP773010453, Tk6, Tk5);
|
Chris@82
|
2050 ii[WS(rs, 57)] = FMA(KP773010453, Tk6, Tk5);
|
Chris@82
|
2051 }
|
Chris@82
|
2052 {
|
Chris@82
|
2053 E TdO, TdV, TjX, TjY;
|
Chris@82
|
2054 TdO = FMA(KP980785280, TdN, TdK);
|
Chris@82
|
2055 TdV = TdR + TdU;
|
Chris@82
|
2056 ri[WS(rs, 33)] = FNMS(KP995184726, TdV, TdO);
|
Chris@82
|
2057 ri[WS(rs, 1)] = FMA(KP995184726, TdV, TdO);
|
Chris@82
|
2058 TjX = FMA(KP980785280, TjW, TjV);
|
Chris@82
|
2059 TjY = TdX + TdY;
|
Chris@82
|
2060 ii[WS(rs, 1)] = FMA(KP995184726, TjY, TjX);
|
Chris@82
|
2061 ii[WS(rs, 33)] = FNMS(KP995184726, TjY, TjX);
|
Chris@82
|
2062 }
|
Chris@82
|
2063 {
|
Chris@82
|
2064 E TdW, TdZ, TjZ, Tk0;
|
Chris@82
|
2065 TdW = FNMS(KP980785280, TdN, TdK);
|
Chris@82
|
2066 TdZ = TdX - TdY;
|
Chris@82
|
2067 ri[WS(rs, 49)] = FNMS(KP995184726, TdZ, TdW);
|
Chris@82
|
2068 ri[WS(rs, 17)] = FMA(KP995184726, TdZ, TdW);
|
Chris@82
|
2069 TjZ = FNMS(KP980785280, TjW, TjV);
|
Chris@82
|
2070 Tk0 = TdU - TdR;
|
Chris@82
|
2071 ii[WS(rs, 17)] = FMA(KP995184726, Tk0, TjZ);
|
Chris@82
|
2072 ii[WS(rs, 49)] = FNMS(KP995184726, Tk0, TjZ);
|
Chris@82
|
2073 }
|
Chris@82
|
2074 {
|
Chris@82
|
2075 E Tec, Ter, Tkb, Tkc;
|
Chris@82
|
2076 Tec = FMA(KP831469612, Teb, Te4);
|
Chris@82
|
2077 Ter = Tej + Teq;
|
Chris@82
|
2078 ri[WS(rs, 37)] = FNMS(KP881921264, Ter, Tec);
|
Chris@82
|
2079 ri[WS(rs, 5)] = FMA(KP881921264, Ter, Tec);
|
Chris@82
|
2080 Tkb = FMA(KP831469612, Tka, Tk9);
|
Chris@82
|
2081 Tkc = Tet + Teu;
|
Chris@82
|
2082 ii[WS(rs, 5)] = FMA(KP881921264, Tkc, Tkb);
|
Chris@82
|
2083 ii[WS(rs, 37)] = FNMS(KP881921264, Tkc, Tkb);
|
Chris@82
|
2084 }
|
Chris@82
|
2085 {
|
Chris@82
|
2086 E Tes, Tev, Tkd, Tke;
|
Chris@82
|
2087 Tes = FNMS(KP831469612, Teb, Te4);
|
Chris@82
|
2088 Tev = Tet - Teu;
|
Chris@82
|
2089 ri[WS(rs, 53)] = FNMS(KP881921264, Tev, Tes);
|
Chris@82
|
2090 ri[WS(rs, 21)] = FMA(KP881921264, Tev, Tes);
|
Chris@82
|
2091 Tkd = FNMS(KP831469612, Tka, Tk9);
|
Chris@82
|
2092 Tke = Teq - Tej;
|
Chris@82
|
2093 ii[WS(rs, 21)] = FMA(KP881921264, Tke, Tkd);
|
Chris@82
|
2094 ii[WS(rs, 53)] = FNMS(KP881921264, Tke, Tkd);
|
Chris@82
|
2095 }
|
Chris@82
|
2096 {
|
Chris@82
|
2097 E TeA, TeH, Tkh, Tki;
|
Chris@82
|
2098 TeA = FNMS(KP831469612, Tez, Tew);
|
Chris@82
|
2099 TeH = TeD - TeG;
|
Chris@82
|
2100 ri[WS(rs, 45)] = FNMS(KP956940335, TeH, TeA);
|
Chris@82
|
2101 ri[WS(rs, 13)] = FMA(KP956940335, TeH, TeA);
|
Chris@82
|
2102 Tkh = FNMS(KP831469612, Tkg, Tkf);
|
Chris@82
|
2103 Tki = TeK - TeJ;
|
Chris@82
|
2104 ii[WS(rs, 13)] = FMA(KP956940335, Tki, Tkh);
|
Chris@82
|
2105 ii[WS(rs, 45)] = FNMS(KP956940335, Tki, Tkh);
|
Chris@82
|
2106 }
|
Chris@82
|
2107 {
|
Chris@82
|
2108 E TeI, TeL, Tkj, Tkk;
|
Chris@82
|
2109 TeI = FMA(KP831469612, Tez, Tew);
|
Chris@82
|
2110 TeL = TeJ + TeK;
|
Chris@82
|
2111 ri[WS(rs, 29)] = FNMS(KP956940335, TeL, TeI);
|
Chris@82
|
2112 ri[WS(rs, 61)] = FMA(KP956940335, TeL, TeI);
|
Chris@82
|
2113 Tkj = FMA(KP831469612, Tkg, Tkf);
|
Chris@82
|
2114 Tkk = TeD + TeG;
|
Chris@82
|
2115 ii[WS(rs, 29)] = FNMS(KP956940335, Tkk, Tkj);
|
Chris@82
|
2116 ii[WS(rs, 61)] = FMA(KP956940335, Tkk, Tkj);
|
Chris@82
|
2117 }
|
Chris@82
|
2118 }
|
Chris@82
|
2119 }
|
Chris@82
|
2120 }
|
Chris@82
|
2121 }
|
Chris@82
|
2122
|
Chris@82
|
2123 static const tw_instr twinstr[] = {
|
Chris@82
|
2124 {TW_FULL, 0, 64},
|
Chris@82
|
2125 {TW_NEXT, 1, 0}
|
Chris@82
|
2126 };
|
Chris@82
|
2127
|
Chris@82
|
2128 static const ct_desc desc = { 64, "t1_64", twinstr, &GENUS, {520, 126, 518, 0}, 0, 0, 0 };
|
Chris@82
|
2129
|
Chris@82
|
2130 void X(codelet_t1_64) (planner *p) {
|
Chris@82
|
2131 X(kdft_dit_register) (p, t1_64, &desc);
|
Chris@82
|
2132 }
|
Chris@82
|
2133 #else
|
Chris@82
|
2134
|
Chris@82
|
2135 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 64 -name t1_64 -include dft/scalar/t.h */
|
Chris@82
|
2136
|
Chris@82
|
2137 /*
|
Chris@82
|
2138 * This function contains 1038 FP additions, 500 FP multiplications,
|
Chris@82
|
2139 * (or, 808 additions, 270 multiplications, 230 fused multiply/add),
|
Chris@82
|
2140 * 176 stack variables, 15 constants, and 256 memory accesses
|
Chris@82
|
2141 */
|
Chris@82
|
2142 #include "dft/scalar/t.h"
|
Chris@82
|
2143
|
Chris@82
|
2144 static void t1_64(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@82
|
2145 {
|
Chris@82
|
2146 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
Chris@82
|
2147 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@82
|
2148 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
Chris@82
|
2149 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@82
|
2150 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
Chris@82
|
2151 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@82
|
2152 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
Chris@82
|
2153 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@82
|
2154 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@82
|
2155 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@82
|
2156 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@82
|
2157 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@82
|
2158 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@82
|
2159 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@82
|
2160 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@82
|
2161 {
|
Chris@82
|
2162 INT m;
|
Chris@82
|
2163 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@82
|
2164 E Tj, TcL, ThT, Tin, T6b, Taz, TgT, Thn, TG, Thm, TcO, TgO, T6m, ThQ, TaC;
|
Chris@82
|
2165 E Tim, T14, Tfq, T6y, T9O, TaG, Tc0, TcU, TeE, T1r, Tfr, T6J, T9P, TaJ, Tc1;
|
Chris@82
|
2166 E TcZ, TeF, T1Q, T2d, Tfx, Tfu, Tfv, Tfw, T6Q, TaM, Tdb, TeJ, T71, TaQ, T7a;
|
Chris@82
|
2167 E TaN, Td6, TeI, T77, TaP, T2B, T2Y, Tfz, TfA, TfB, TfC, T7h, TaW, Tdm, TeM;
|
Chris@82
|
2168 E T7s, TaU, T7B, TaX, Tdh, TeL, T7y, TaT, T5j, TfR, Tec, Tf0, TfY, Tgy, T8D;
|
Chris@82
|
2169 E Tbl, T8O, Tbx, T9l, Tbm, TdV, TeX, T9i, Tbw, T3M, TfL, TdL, TeQ, TfI, Tgt;
|
Chris@82
|
2170 E T7K, Tb2, T7V, Tbe, T8s, Tb3, Tdu, TeT, T8p, Tbd, T4x, TfJ, TdE, TdM, TfO;
|
Chris@82
|
2171 E Tgu, T87, T8v, T8i, T8u, Tba, Tbg, Tdz, TdN, Tb7, Tbh, T64, TfZ, Te5, Ted;
|
Chris@82
|
2172 E TfU, Tgz, T90, T9o, T9b, T9n, Tbt, Tbz, Te0, Tee, Tbq, TbA;
|
Chris@82
|
2173 {
|
Chris@82
|
2174 E T1, TgR, T6, TgQ, Tc, T68, Th, T69;
|
Chris@82
|
2175 T1 = ri[0];
|
Chris@82
|
2176 TgR = ii[0];
|
Chris@82
|
2177 {
|
Chris@82
|
2178 E T3, T5, T2, T4;
|
Chris@82
|
2179 T3 = ri[WS(rs, 32)];
|
Chris@82
|
2180 T5 = ii[WS(rs, 32)];
|
Chris@82
|
2181 T2 = W[62];
|
Chris@82
|
2182 T4 = W[63];
|
Chris@82
|
2183 T6 = FMA(T2, T3, T4 * T5);
|
Chris@82
|
2184 TgQ = FNMS(T4, T3, T2 * T5);
|
Chris@82
|
2185 }
|
Chris@82
|
2186 {
|
Chris@82
|
2187 E T9, Tb, T8, Ta;
|
Chris@82
|
2188 T9 = ri[WS(rs, 16)];
|
Chris@82
|
2189 Tb = ii[WS(rs, 16)];
|
Chris@82
|
2190 T8 = W[30];
|
Chris@82
|
2191 Ta = W[31];
|
Chris@82
|
2192 Tc = FMA(T8, T9, Ta * Tb);
|
Chris@82
|
2193 T68 = FNMS(Ta, T9, T8 * Tb);
|
Chris@82
|
2194 }
|
Chris@82
|
2195 {
|
Chris@82
|
2196 E Te, Tg, Td, Tf;
|
Chris@82
|
2197 Te = ri[WS(rs, 48)];
|
Chris@82
|
2198 Tg = ii[WS(rs, 48)];
|
Chris@82
|
2199 Td = W[94];
|
Chris@82
|
2200 Tf = W[95];
|
Chris@82
|
2201 Th = FMA(Td, Te, Tf * Tg);
|
Chris@82
|
2202 T69 = FNMS(Tf, Te, Td * Tg);
|
Chris@82
|
2203 }
|
Chris@82
|
2204 {
|
Chris@82
|
2205 E T7, Ti, ThR, ThS;
|
Chris@82
|
2206 T7 = T1 + T6;
|
Chris@82
|
2207 Ti = Tc + Th;
|
Chris@82
|
2208 Tj = T7 + Ti;
|
Chris@82
|
2209 TcL = T7 - Ti;
|
Chris@82
|
2210 ThR = TgR - TgQ;
|
Chris@82
|
2211 ThS = Tc - Th;
|
Chris@82
|
2212 ThT = ThR - ThS;
|
Chris@82
|
2213 Tin = ThS + ThR;
|
Chris@82
|
2214 }
|
Chris@82
|
2215 {
|
Chris@82
|
2216 E T67, T6a, TgP, TgS;
|
Chris@82
|
2217 T67 = T1 - T6;
|
Chris@82
|
2218 T6a = T68 - T69;
|
Chris@82
|
2219 T6b = T67 - T6a;
|
Chris@82
|
2220 Taz = T67 + T6a;
|
Chris@82
|
2221 TgP = T68 + T69;
|
Chris@82
|
2222 TgS = TgQ + TgR;
|
Chris@82
|
2223 TgT = TgP + TgS;
|
Chris@82
|
2224 Thn = TgS - TgP;
|
Chris@82
|
2225 }
|
Chris@82
|
2226 }
|
Chris@82
|
2227 {
|
Chris@82
|
2228 E To, T6c, Tt, T6d, T6e, T6f, Tz, T6i, TE, T6j, T6h, T6k;
|
Chris@82
|
2229 {
|
Chris@82
|
2230 E Tl, Tn, Tk, Tm;
|
Chris@82
|
2231 Tl = ri[WS(rs, 8)];
|
Chris@82
|
2232 Tn = ii[WS(rs, 8)];
|
Chris@82
|
2233 Tk = W[14];
|
Chris@82
|
2234 Tm = W[15];
|
Chris@82
|
2235 To = FMA(Tk, Tl, Tm * Tn);
|
Chris@82
|
2236 T6c = FNMS(Tm, Tl, Tk * Tn);
|
Chris@82
|
2237 }
|
Chris@82
|
2238 {
|
Chris@82
|
2239 E Tq, Ts, Tp, Tr;
|
Chris@82
|
2240 Tq = ri[WS(rs, 40)];
|
Chris@82
|
2241 Ts = ii[WS(rs, 40)];
|
Chris@82
|
2242 Tp = W[78];
|
Chris@82
|
2243 Tr = W[79];
|
Chris@82
|
2244 Tt = FMA(Tp, Tq, Tr * Ts);
|
Chris@82
|
2245 T6d = FNMS(Tr, Tq, Tp * Ts);
|
Chris@82
|
2246 }
|
Chris@82
|
2247 T6e = T6c - T6d;
|
Chris@82
|
2248 T6f = To - Tt;
|
Chris@82
|
2249 {
|
Chris@82
|
2250 E Tw, Ty, Tv, Tx;
|
Chris@82
|
2251 Tw = ri[WS(rs, 56)];
|
Chris@82
|
2252 Ty = ii[WS(rs, 56)];
|
Chris@82
|
2253 Tv = W[110];
|
Chris@82
|
2254 Tx = W[111];
|
Chris@82
|
2255 Tz = FMA(Tv, Tw, Tx * Ty);
|
Chris@82
|
2256 T6i = FNMS(Tx, Tw, Tv * Ty);
|
Chris@82
|
2257 }
|
Chris@82
|
2258 {
|
Chris@82
|
2259 E TB, TD, TA, TC;
|
Chris@82
|
2260 TB = ri[WS(rs, 24)];
|
Chris@82
|
2261 TD = ii[WS(rs, 24)];
|
Chris@82
|
2262 TA = W[46];
|
Chris@82
|
2263 TC = W[47];
|
Chris@82
|
2264 TE = FMA(TA, TB, TC * TD);
|
Chris@82
|
2265 T6j = FNMS(TC, TB, TA * TD);
|
Chris@82
|
2266 }
|
Chris@82
|
2267 T6h = Tz - TE;
|
Chris@82
|
2268 T6k = T6i - T6j;
|
Chris@82
|
2269 {
|
Chris@82
|
2270 E Tu, TF, TcM, TcN;
|
Chris@82
|
2271 Tu = To + Tt;
|
Chris@82
|
2272 TF = Tz + TE;
|
Chris@82
|
2273 TG = Tu + TF;
|
Chris@82
|
2274 Thm = TF - Tu;
|
Chris@82
|
2275 TcM = T6c + T6d;
|
Chris@82
|
2276 TcN = T6i + T6j;
|
Chris@82
|
2277 TcO = TcM - TcN;
|
Chris@82
|
2278 TgO = TcM + TcN;
|
Chris@82
|
2279 }
|
Chris@82
|
2280 {
|
Chris@82
|
2281 E T6g, T6l, TaA, TaB;
|
Chris@82
|
2282 T6g = T6e - T6f;
|
Chris@82
|
2283 T6l = T6h + T6k;
|
Chris@82
|
2284 T6m = KP707106781 * (T6g - T6l);
|
Chris@82
|
2285 ThQ = KP707106781 * (T6g + T6l);
|
Chris@82
|
2286 TaA = T6f + T6e;
|
Chris@82
|
2287 TaB = T6h - T6k;
|
Chris@82
|
2288 TaC = KP707106781 * (TaA + TaB);
|
Chris@82
|
2289 Tim = KP707106781 * (TaB - TaA);
|
Chris@82
|
2290 }
|
Chris@82
|
2291 }
|
Chris@82
|
2292 {
|
Chris@82
|
2293 E TS, TcQ, T6q, T6t, T13, TcR, T6r, T6w, T6s, T6x;
|
Chris@82
|
2294 {
|
Chris@82
|
2295 E TM, T6o, TR, T6p;
|
Chris@82
|
2296 {
|
Chris@82
|
2297 E TJ, TL, TI, TK;
|
Chris@82
|
2298 TJ = ri[WS(rs, 4)];
|
Chris@82
|
2299 TL = ii[WS(rs, 4)];
|
Chris@82
|
2300 TI = W[6];
|
Chris@82
|
2301 TK = W[7];
|
Chris@82
|
2302 TM = FMA(TI, TJ, TK * TL);
|
Chris@82
|
2303 T6o = FNMS(TK, TJ, TI * TL);
|
Chris@82
|
2304 }
|
Chris@82
|
2305 {
|
Chris@82
|
2306 E TO, TQ, TN, TP;
|
Chris@82
|
2307 TO = ri[WS(rs, 36)];
|
Chris@82
|
2308 TQ = ii[WS(rs, 36)];
|
Chris@82
|
2309 TN = W[70];
|
Chris@82
|
2310 TP = W[71];
|
Chris@82
|
2311 TR = FMA(TN, TO, TP * TQ);
|
Chris@82
|
2312 T6p = FNMS(TP, TO, TN * TQ);
|
Chris@82
|
2313 }
|
Chris@82
|
2314 TS = TM + TR;
|
Chris@82
|
2315 TcQ = T6o + T6p;
|
Chris@82
|
2316 T6q = T6o - T6p;
|
Chris@82
|
2317 T6t = TM - TR;
|
Chris@82
|
2318 }
|
Chris@82
|
2319 {
|
Chris@82
|
2320 E TX, T6u, T12, T6v;
|
Chris@82
|
2321 {
|
Chris@82
|
2322 E TU, TW, TT, TV;
|
Chris@82
|
2323 TU = ri[WS(rs, 20)];
|
Chris@82
|
2324 TW = ii[WS(rs, 20)];
|
Chris@82
|
2325 TT = W[38];
|
Chris@82
|
2326 TV = W[39];
|
Chris@82
|
2327 TX = FMA(TT, TU, TV * TW);
|
Chris@82
|
2328 T6u = FNMS(TV, TU, TT * TW);
|
Chris@82
|
2329 }
|
Chris@82
|
2330 {
|
Chris@82
|
2331 E TZ, T11, TY, T10;
|
Chris@82
|
2332 TZ = ri[WS(rs, 52)];
|
Chris@82
|
2333 T11 = ii[WS(rs, 52)];
|
Chris@82
|
2334 TY = W[102];
|
Chris@82
|
2335 T10 = W[103];
|
Chris@82
|
2336 T12 = FMA(TY, TZ, T10 * T11);
|
Chris@82
|
2337 T6v = FNMS(T10, TZ, TY * T11);
|
Chris@82
|
2338 }
|
Chris@82
|
2339 T13 = TX + T12;
|
Chris@82
|
2340 TcR = T6u + T6v;
|
Chris@82
|
2341 T6r = TX - T12;
|
Chris@82
|
2342 T6w = T6u - T6v;
|
Chris@82
|
2343 }
|
Chris@82
|
2344 T14 = TS + T13;
|
Chris@82
|
2345 Tfq = TcQ + TcR;
|
Chris@82
|
2346 T6s = T6q + T6r;
|
Chris@82
|
2347 T6x = T6t - T6w;
|
Chris@82
|
2348 T6y = FNMS(KP923879532, T6x, KP382683432 * T6s);
|
Chris@82
|
2349 T9O = FMA(KP923879532, T6s, KP382683432 * T6x);
|
Chris@82
|
2350 {
|
Chris@82
|
2351 E TaE, TaF, TcS, TcT;
|
Chris@82
|
2352 TaE = T6q - T6r;
|
Chris@82
|
2353 TaF = T6t + T6w;
|
Chris@82
|
2354 TaG = FNMS(KP382683432, TaF, KP923879532 * TaE);
|
Chris@82
|
2355 Tc0 = FMA(KP382683432, TaE, KP923879532 * TaF);
|
Chris@82
|
2356 TcS = TcQ - TcR;
|
Chris@82
|
2357 TcT = TS - T13;
|
Chris@82
|
2358 TcU = TcS - TcT;
|
Chris@82
|
2359 TeE = TcT + TcS;
|
Chris@82
|
2360 }
|
Chris@82
|
2361 }
|
Chris@82
|
2362 {
|
Chris@82
|
2363 E T1f, TcW, T6B, T6E, T1q, TcX, T6C, T6H, T6D, T6I;
|
Chris@82
|
2364 {
|
Chris@82
|
2365 E T19, T6z, T1e, T6A;
|
Chris@82
|
2366 {
|
Chris@82
|
2367 E T16, T18, T15, T17;
|
Chris@82
|
2368 T16 = ri[WS(rs, 60)];
|
Chris@82
|
2369 T18 = ii[WS(rs, 60)];
|
Chris@82
|
2370 T15 = W[118];
|
Chris@82
|
2371 T17 = W[119];
|
Chris@82
|
2372 T19 = FMA(T15, T16, T17 * T18);
|
Chris@82
|
2373 T6z = FNMS(T17, T16, T15 * T18);
|
Chris@82
|
2374 }
|
Chris@82
|
2375 {
|
Chris@82
|
2376 E T1b, T1d, T1a, T1c;
|
Chris@82
|
2377 T1b = ri[WS(rs, 28)];
|
Chris@82
|
2378 T1d = ii[WS(rs, 28)];
|
Chris@82
|
2379 T1a = W[54];
|
Chris@82
|
2380 T1c = W[55];
|
Chris@82
|
2381 T1e = FMA(T1a, T1b, T1c * T1d);
|
Chris@82
|
2382 T6A = FNMS(T1c, T1b, T1a * T1d);
|
Chris@82
|
2383 }
|
Chris@82
|
2384 T1f = T19 + T1e;
|
Chris@82
|
2385 TcW = T6z + T6A;
|
Chris@82
|
2386 T6B = T6z - T6A;
|
Chris@82
|
2387 T6E = T19 - T1e;
|
Chris@82
|
2388 }
|
Chris@82
|
2389 {
|
Chris@82
|
2390 E T1k, T6F, T1p, T6G;
|
Chris@82
|
2391 {
|
Chris@82
|
2392 E T1h, T1j, T1g, T1i;
|
Chris@82
|
2393 T1h = ri[WS(rs, 12)];
|
Chris@82
|
2394 T1j = ii[WS(rs, 12)];
|
Chris@82
|
2395 T1g = W[22];
|
Chris@82
|
2396 T1i = W[23];
|
Chris@82
|
2397 T1k = FMA(T1g, T1h, T1i * T1j);
|
Chris@82
|
2398 T6F = FNMS(T1i, T1h, T1g * T1j);
|
Chris@82
|
2399 }
|
Chris@82
|
2400 {
|
Chris@82
|
2401 E T1m, T1o, T1l, T1n;
|
Chris@82
|
2402 T1m = ri[WS(rs, 44)];
|
Chris@82
|
2403 T1o = ii[WS(rs, 44)];
|
Chris@82
|
2404 T1l = W[86];
|
Chris@82
|
2405 T1n = W[87];
|
Chris@82
|
2406 T1p = FMA(T1l, T1m, T1n * T1o);
|
Chris@82
|
2407 T6G = FNMS(T1n, T1m, T1l * T1o);
|
Chris@82
|
2408 }
|
Chris@82
|
2409 T1q = T1k + T1p;
|
Chris@82
|
2410 TcX = T6F + T6G;
|
Chris@82
|
2411 T6C = T1k - T1p;
|
Chris@82
|
2412 T6H = T6F - T6G;
|
Chris@82
|
2413 }
|
Chris@82
|
2414 T1r = T1f + T1q;
|
Chris@82
|
2415 Tfr = TcW + TcX;
|
Chris@82
|
2416 T6D = T6B + T6C;
|
Chris@82
|
2417 T6I = T6E - T6H;
|
Chris@82
|
2418 T6J = FMA(KP382683432, T6D, KP923879532 * T6I);
|
Chris@82
|
2419 T9P = FNMS(KP923879532, T6D, KP382683432 * T6I);
|
Chris@82
|
2420 {
|
Chris@82
|
2421 E TaH, TaI, TcV, TcY;
|
Chris@82
|
2422 TaH = T6B - T6C;
|
Chris@82
|
2423 TaI = T6E + T6H;
|
Chris@82
|
2424 TaJ = FMA(KP923879532, TaH, KP382683432 * TaI);
|
Chris@82
|
2425 Tc1 = FNMS(KP382683432, TaH, KP923879532 * TaI);
|
Chris@82
|
2426 TcV = T1f - T1q;
|
Chris@82
|
2427 TcY = TcW - TcX;
|
Chris@82
|
2428 TcZ = TcV + TcY;
|
Chris@82
|
2429 TeF = TcV - TcY;
|
Chris@82
|
2430 }
|
Chris@82
|
2431 }
|
Chris@82
|
2432 {
|
Chris@82
|
2433 E T1y, T6M, T1D, T6N, T1E, Td2, T1J, T74, T1O, T75, T1P, Td3, T21, Td8, T6W;
|
Chris@82
|
2434 E T6Z, T2c, Td9, T6R, T6U;
|
Chris@82
|
2435 {
|
Chris@82
|
2436 E T1v, T1x, T1u, T1w;
|
Chris@82
|
2437 T1v = ri[WS(rs, 2)];
|
Chris@82
|
2438 T1x = ii[WS(rs, 2)];
|
Chris@82
|
2439 T1u = W[2];
|
Chris@82
|
2440 T1w = W[3];
|
Chris@82
|
2441 T1y = FMA(T1u, T1v, T1w * T1x);
|
Chris@82
|
2442 T6M = FNMS(T1w, T1v, T1u * T1x);
|
Chris@82
|
2443 }
|
Chris@82
|
2444 {
|
Chris@82
|
2445 E T1A, T1C, T1z, T1B;
|
Chris@82
|
2446 T1A = ri[WS(rs, 34)];
|
Chris@82
|
2447 T1C = ii[WS(rs, 34)];
|
Chris@82
|
2448 T1z = W[66];
|
Chris@82
|
2449 T1B = W[67];
|
Chris@82
|
2450 T1D = FMA(T1z, T1A, T1B * T1C);
|
Chris@82
|
2451 T6N = FNMS(T1B, T1A, T1z * T1C);
|
Chris@82
|
2452 }
|
Chris@82
|
2453 T1E = T1y + T1D;
|
Chris@82
|
2454 Td2 = T6M + T6N;
|
Chris@82
|
2455 {
|
Chris@82
|
2456 E T1G, T1I, T1F, T1H;
|
Chris@82
|
2457 T1G = ri[WS(rs, 18)];
|
Chris@82
|
2458 T1I = ii[WS(rs, 18)];
|
Chris@82
|
2459 T1F = W[34];
|
Chris@82
|
2460 T1H = W[35];
|
Chris@82
|
2461 T1J = FMA(T1F, T1G, T1H * T1I);
|
Chris@82
|
2462 T74 = FNMS(T1H, T1G, T1F * T1I);
|
Chris@82
|
2463 }
|
Chris@82
|
2464 {
|
Chris@82
|
2465 E T1L, T1N, T1K, T1M;
|
Chris@82
|
2466 T1L = ri[WS(rs, 50)];
|
Chris@82
|
2467 T1N = ii[WS(rs, 50)];
|
Chris@82
|
2468 T1K = W[98];
|
Chris@82
|
2469 T1M = W[99];
|
Chris@82
|
2470 T1O = FMA(T1K, T1L, T1M * T1N);
|
Chris@82
|
2471 T75 = FNMS(T1M, T1L, T1K * T1N);
|
Chris@82
|
2472 }
|
Chris@82
|
2473 T1P = T1J + T1O;
|
Chris@82
|
2474 Td3 = T74 + T75;
|
Chris@82
|
2475 {
|
Chris@82
|
2476 E T1V, T6X, T20, T6Y;
|
Chris@82
|
2477 {
|
Chris@82
|
2478 E T1S, T1U, T1R, T1T;
|
Chris@82
|
2479 T1S = ri[WS(rs, 10)];
|
Chris@82
|
2480 T1U = ii[WS(rs, 10)];
|
Chris@82
|
2481 T1R = W[18];
|
Chris@82
|
2482 T1T = W[19];
|
Chris@82
|
2483 T1V = FMA(T1R, T1S, T1T * T1U);
|
Chris@82
|
2484 T6X = FNMS(T1T, T1S, T1R * T1U);
|
Chris@82
|
2485 }
|
Chris@82
|
2486 {
|
Chris@82
|
2487 E T1X, T1Z, T1W, T1Y;
|
Chris@82
|
2488 T1X = ri[WS(rs, 42)];
|
Chris@82
|
2489 T1Z = ii[WS(rs, 42)];
|
Chris@82
|
2490 T1W = W[82];
|
Chris@82
|
2491 T1Y = W[83];
|
Chris@82
|
2492 T20 = FMA(T1W, T1X, T1Y * T1Z);
|
Chris@82
|
2493 T6Y = FNMS(T1Y, T1X, T1W * T1Z);
|
Chris@82
|
2494 }
|
Chris@82
|
2495 T21 = T1V + T20;
|
Chris@82
|
2496 Td8 = T6X + T6Y;
|
Chris@82
|
2497 T6W = T1V - T20;
|
Chris@82
|
2498 T6Z = T6X - T6Y;
|
Chris@82
|
2499 }
|
Chris@82
|
2500 {
|
Chris@82
|
2501 E T26, T6S, T2b, T6T;
|
Chris@82
|
2502 {
|
Chris@82
|
2503 E T23, T25, T22, T24;
|
Chris@82
|
2504 T23 = ri[WS(rs, 58)];
|
Chris@82
|
2505 T25 = ii[WS(rs, 58)];
|
Chris@82
|
2506 T22 = W[114];
|
Chris@82
|
2507 T24 = W[115];
|
Chris@82
|
2508 T26 = FMA(T22, T23, T24 * T25);
|
Chris@82
|
2509 T6S = FNMS(T24, T23, T22 * T25);
|
Chris@82
|
2510 }
|
Chris@82
|
2511 {
|
Chris@82
|
2512 E T28, T2a, T27, T29;
|
Chris@82
|
2513 T28 = ri[WS(rs, 26)];
|
Chris@82
|
2514 T2a = ii[WS(rs, 26)];
|
Chris@82
|
2515 T27 = W[50];
|
Chris@82
|
2516 T29 = W[51];
|
Chris@82
|
2517 T2b = FMA(T27, T28, T29 * T2a);
|
Chris@82
|
2518 T6T = FNMS(T29, T28, T27 * T2a);
|
Chris@82
|
2519 }
|
Chris@82
|
2520 T2c = T26 + T2b;
|
Chris@82
|
2521 Td9 = T6S + T6T;
|
Chris@82
|
2522 T6R = T26 - T2b;
|
Chris@82
|
2523 T6U = T6S - T6T;
|
Chris@82
|
2524 }
|
Chris@82
|
2525 T1Q = T1E + T1P;
|
Chris@82
|
2526 T2d = T21 + T2c;
|
Chris@82
|
2527 Tfx = T1Q - T2d;
|
Chris@82
|
2528 Tfu = Td2 + Td3;
|
Chris@82
|
2529 Tfv = Td8 + Td9;
|
Chris@82
|
2530 Tfw = Tfu - Tfv;
|
Chris@82
|
2531 {
|
Chris@82
|
2532 E T6O, T6P, Td7, Tda;
|
Chris@82
|
2533 T6O = T6M - T6N;
|
Chris@82
|
2534 T6P = T1J - T1O;
|
Chris@82
|
2535 T6Q = T6O + T6P;
|
Chris@82
|
2536 TaM = T6O - T6P;
|
Chris@82
|
2537 Td7 = T1E - T1P;
|
Chris@82
|
2538 Tda = Td8 - Td9;
|
Chris@82
|
2539 Tdb = Td7 - Tda;
|
Chris@82
|
2540 TeJ = Td7 + Tda;
|
Chris@82
|
2541 }
|
Chris@82
|
2542 {
|
Chris@82
|
2543 E T6V, T70, T78, T79;
|
Chris@82
|
2544 T6V = T6R - T6U;
|
Chris@82
|
2545 T70 = T6W + T6Z;
|
Chris@82
|
2546 T71 = KP707106781 * (T6V - T70);
|
Chris@82
|
2547 TaQ = KP707106781 * (T70 + T6V);
|
Chris@82
|
2548 T78 = T6Z - T6W;
|
Chris@82
|
2549 T79 = T6R + T6U;
|
Chris@82
|
2550 T7a = KP707106781 * (T78 - T79);
|
Chris@82
|
2551 TaN = KP707106781 * (T78 + T79);
|
Chris@82
|
2552 }
|
Chris@82
|
2553 {
|
Chris@82
|
2554 E Td4, Td5, T73, T76;
|
Chris@82
|
2555 Td4 = Td2 - Td3;
|
Chris@82
|
2556 Td5 = T2c - T21;
|
Chris@82
|
2557 Td6 = Td4 - Td5;
|
Chris@82
|
2558 TeI = Td4 + Td5;
|
Chris@82
|
2559 T73 = T1y - T1D;
|
Chris@82
|
2560 T76 = T74 - T75;
|
Chris@82
|
2561 T77 = T73 - T76;
|
Chris@82
|
2562 TaP = T73 + T76;
|
Chris@82
|
2563 }
|
Chris@82
|
2564 }
|
Chris@82
|
2565 {
|
Chris@82
|
2566 E T2j, T7d, T2o, T7e, T2p, Tdd, T2u, T7v, T2z, T7w, T2A, Tde, T2M, Tdj, T7n;
|
Chris@82
|
2567 E T7q, T2X, Tdk, T7i, T7l;
|
Chris@82
|
2568 {
|
Chris@82
|
2569 E T2g, T2i, T2f, T2h;
|
Chris@82
|
2570 T2g = ri[WS(rs, 62)];
|
Chris@82
|
2571 T2i = ii[WS(rs, 62)];
|
Chris@82
|
2572 T2f = W[122];
|
Chris@82
|
2573 T2h = W[123];
|
Chris@82
|
2574 T2j = FMA(T2f, T2g, T2h * T2i);
|
Chris@82
|
2575 T7d = FNMS(T2h, T2g, T2f * T2i);
|
Chris@82
|
2576 }
|
Chris@82
|
2577 {
|
Chris@82
|
2578 E T2l, T2n, T2k, T2m;
|
Chris@82
|
2579 T2l = ri[WS(rs, 30)];
|
Chris@82
|
2580 T2n = ii[WS(rs, 30)];
|
Chris@82
|
2581 T2k = W[58];
|
Chris@82
|
2582 T2m = W[59];
|
Chris@82
|
2583 T2o = FMA(T2k, T2l, T2m * T2n);
|
Chris@82
|
2584 T7e = FNMS(T2m, T2l, T2k * T2n);
|
Chris@82
|
2585 }
|
Chris@82
|
2586 T2p = T2j + T2o;
|
Chris@82
|
2587 Tdd = T7d + T7e;
|
Chris@82
|
2588 {
|
Chris@82
|
2589 E T2r, T2t, T2q, T2s;
|
Chris@82
|
2590 T2r = ri[WS(rs, 14)];
|
Chris@82
|
2591 T2t = ii[WS(rs, 14)];
|
Chris@82
|
2592 T2q = W[26];
|
Chris@82
|
2593 T2s = W[27];
|
Chris@82
|
2594 T2u = FMA(T2q, T2r, T2s * T2t);
|
Chris@82
|
2595 T7v = FNMS(T2s, T2r, T2q * T2t);
|
Chris@82
|
2596 }
|
Chris@82
|
2597 {
|
Chris@82
|
2598 E T2w, T2y, T2v, T2x;
|
Chris@82
|
2599 T2w = ri[WS(rs, 46)];
|
Chris@82
|
2600 T2y = ii[WS(rs, 46)];
|
Chris@82
|
2601 T2v = W[90];
|
Chris@82
|
2602 T2x = W[91];
|
Chris@82
|
2603 T2z = FMA(T2v, T2w, T2x * T2y);
|
Chris@82
|
2604 T7w = FNMS(T2x, T2w, T2v * T2y);
|
Chris@82
|
2605 }
|
Chris@82
|
2606 T2A = T2u + T2z;
|
Chris@82
|
2607 Tde = T7v + T7w;
|
Chris@82
|
2608 {
|
Chris@82
|
2609 E T2G, T7o, T2L, T7p;
|
Chris@82
|
2610 {
|
Chris@82
|
2611 E T2D, T2F, T2C, T2E;
|
Chris@82
|
2612 T2D = ri[WS(rs, 6)];
|
Chris@82
|
2613 T2F = ii[WS(rs, 6)];
|
Chris@82
|
2614 T2C = W[10];
|
Chris@82
|
2615 T2E = W[11];
|
Chris@82
|
2616 T2G = FMA(T2C, T2D, T2E * T2F);
|
Chris@82
|
2617 T7o = FNMS(T2E, T2D, T2C * T2F);
|
Chris@82
|
2618 }
|
Chris@82
|
2619 {
|
Chris@82
|
2620 E T2I, T2K, T2H, T2J;
|
Chris@82
|
2621 T2I = ri[WS(rs, 38)];
|
Chris@82
|
2622 T2K = ii[WS(rs, 38)];
|
Chris@82
|
2623 T2H = W[74];
|
Chris@82
|
2624 T2J = W[75];
|
Chris@82
|
2625 T2L = FMA(T2H, T2I, T2J * T2K);
|
Chris@82
|
2626 T7p = FNMS(T2J, T2I, T2H * T2K);
|
Chris@82
|
2627 }
|
Chris@82
|
2628 T2M = T2G + T2L;
|
Chris@82
|
2629 Tdj = T7o + T7p;
|
Chris@82
|
2630 T7n = T2G - T2L;
|
Chris@82
|
2631 T7q = T7o - T7p;
|
Chris@82
|
2632 }
|
Chris@82
|
2633 {
|
Chris@82
|
2634 E T2R, T7j, T2W, T7k;
|
Chris@82
|
2635 {
|
Chris@82
|
2636 E T2O, T2Q, T2N, T2P;
|
Chris@82
|
2637 T2O = ri[WS(rs, 54)];
|
Chris@82
|
2638 T2Q = ii[WS(rs, 54)];
|
Chris@82
|
2639 T2N = W[106];
|
Chris@82
|
2640 T2P = W[107];
|
Chris@82
|
2641 T2R = FMA(T2N, T2O, T2P * T2Q);
|
Chris@82
|
2642 T7j = FNMS(T2P, T2O, T2N * T2Q);
|
Chris@82
|
2643 }
|
Chris@82
|
2644 {
|
Chris@82
|
2645 E T2T, T2V, T2S, T2U;
|
Chris@82
|
2646 T2T = ri[WS(rs, 22)];
|
Chris@82
|
2647 T2V = ii[WS(rs, 22)];
|
Chris@82
|
2648 T2S = W[42];
|
Chris@82
|
2649 T2U = W[43];
|
Chris@82
|
2650 T2W = FMA(T2S, T2T, T2U * T2V);
|
Chris@82
|
2651 T7k = FNMS(T2U, T2T, T2S * T2V);
|
Chris@82
|
2652 }
|
Chris@82
|
2653 T2X = T2R + T2W;
|
Chris@82
|
2654 Tdk = T7j + T7k;
|
Chris@82
|
2655 T7i = T2R - T2W;
|
Chris@82
|
2656 T7l = T7j - T7k;
|
Chris@82
|
2657 }
|
Chris@82
|
2658 T2B = T2p + T2A;
|
Chris@82
|
2659 T2Y = T2M + T2X;
|
Chris@82
|
2660 Tfz = T2B - T2Y;
|
Chris@82
|
2661 TfA = Tdd + Tde;
|
Chris@82
|
2662 TfB = Tdj + Tdk;
|
Chris@82
|
2663 TfC = TfA - TfB;
|
Chris@82
|
2664 {
|
Chris@82
|
2665 E T7f, T7g, Tdi, Tdl;
|
Chris@82
|
2666 T7f = T7d - T7e;
|
Chris@82
|
2667 T7g = T2u - T2z;
|
Chris@82
|
2668 T7h = T7f + T7g;
|
Chris@82
|
2669 TaW = T7f - T7g;
|
Chris@82
|
2670 Tdi = T2p - T2A;
|
Chris@82
|
2671 Tdl = Tdj - Tdk;
|
Chris@82
|
2672 Tdm = Tdi - Tdl;
|
Chris@82
|
2673 TeM = Tdi + Tdl;
|
Chris@82
|
2674 }
|
Chris@82
|
2675 {
|
Chris@82
|
2676 E T7m, T7r, T7z, T7A;
|
Chris@82
|
2677 T7m = T7i - T7l;
|
Chris@82
|
2678 T7r = T7n + T7q;
|
Chris@82
|
2679 T7s = KP707106781 * (T7m - T7r);
|
Chris@82
|
2680 TaU = KP707106781 * (T7r + T7m);
|
Chris@82
|
2681 T7z = T7q - T7n;
|
Chris@82
|
2682 T7A = T7i + T7l;
|
Chris@82
|
2683 T7B = KP707106781 * (T7z - T7A);
|
Chris@82
|
2684 TaX = KP707106781 * (T7z + T7A);
|
Chris@82
|
2685 }
|
Chris@82
|
2686 {
|
Chris@82
|
2687 E Tdf, Tdg, T7u, T7x;
|
Chris@82
|
2688 Tdf = Tdd - Tde;
|
Chris@82
|
2689 Tdg = T2X - T2M;
|
Chris@82
|
2690 Tdh = Tdf - Tdg;
|
Chris@82
|
2691 TeL = Tdf + Tdg;
|
Chris@82
|
2692 T7u = T2j - T2o;
|
Chris@82
|
2693 T7x = T7v - T7w;
|
Chris@82
|
2694 T7y = T7u - T7x;
|
Chris@82
|
2695 TaT = T7u + T7x;
|
Chris@82
|
2696 }
|
Chris@82
|
2697 }
|
Chris@82
|
2698 {
|
Chris@82
|
2699 E T4D, T9e, T4I, T9f, T4J, Te8, T4O, T8A, T4T, T8B, T4U, Te9, T56, TdS, T8G;
|
Chris@82
|
2700 E T8H, T5h, TdT, T8J, T8M;
|
Chris@82
|
2701 {
|
Chris@82
|
2702 E T4A, T4C, T4z, T4B;
|
Chris@82
|
2703 T4A = ri[WS(rs, 63)];
|
Chris@82
|
2704 T4C = ii[WS(rs, 63)];
|
Chris@82
|
2705 T4z = W[124];
|
Chris@82
|
2706 T4B = W[125];
|
Chris@82
|
2707 T4D = FMA(T4z, T4A, T4B * T4C);
|
Chris@82
|
2708 T9e = FNMS(T4B, T4A, T4z * T4C);
|
Chris@82
|
2709 }
|
Chris@82
|
2710 {
|
Chris@82
|
2711 E T4F, T4H, T4E, T4G;
|
Chris@82
|
2712 T4F = ri[WS(rs, 31)];
|
Chris@82
|
2713 T4H = ii[WS(rs, 31)];
|
Chris@82
|
2714 T4E = W[60];
|
Chris@82
|
2715 T4G = W[61];
|
Chris@82
|
2716 T4I = FMA(T4E, T4F, T4G * T4H);
|
Chris@82
|
2717 T9f = FNMS(T4G, T4F, T4E * T4H);
|
Chris@82
|
2718 }
|
Chris@82
|
2719 T4J = T4D + T4I;
|
Chris@82
|
2720 Te8 = T9e + T9f;
|
Chris@82
|
2721 {
|
Chris@82
|
2722 E T4L, T4N, T4K, T4M;
|
Chris@82
|
2723 T4L = ri[WS(rs, 15)];
|
Chris@82
|
2724 T4N = ii[WS(rs, 15)];
|
Chris@82
|
2725 T4K = W[28];
|
Chris@82
|
2726 T4M = W[29];
|
Chris@82
|
2727 T4O = FMA(T4K, T4L, T4M * T4N);
|
Chris@82
|
2728 T8A = FNMS(T4M, T4L, T4K * T4N);
|
Chris@82
|
2729 }
|
Chris@82
|
2730 {
|
Chris@82
|
2731 E T4Q, T4S, T4P, T4R;
|
Chris@82
|
2732 T4Q = ri[WS(rs, 47)];
|
Chris@82
|
2733 T4S = ii[WS(rs, 47)];
|
Chris@82
|
2734 T4P = W[92];
|
Chris@82
|
2735 T4R = W[93];
|
Chris@82
|
2736 T4T = FMA(T4P, T4Q, T4R * T4S);
|
Chris@82
|
2737 T8B = FNMS(T4R, T4Q, T4P * T4S);
|
Chris@82
|
2738 }
|
Chris@82
|
2739 T4U = T4O + T4T;
|
Chris@82
|
2740 Te9 = T8A + T8B;
|
Chris@82
|
2741 {
|
Chris@82
|
2742 E T50, T8E, T55, T8F;
|
Chris@82
|
2743 {
|
Chris@82
|
2744 E T4X, T4Z, T4W, T4Y;
|
Chris@82
|
2745 T4X = ri[WS(rs, 7)];
|
Chris@82
|
2746 T4Z = ii[WS(rs, 7)];
|
Chris@82
|
2747 T4W = W[12];
|
Chris@82
|
2748 T4Y = W[13];
|
Chris@82
|
2749 T50 = FMA(T4W, T4X, T4Y * T4Z);
|
Chris@82
|
2750 T8E = FNMS(T4Y, T4X, T4W * T4Z);
|
Chris@82
|
2751 }
|
Chris@82
|
2752 {
|
Chris@82
|
2753 E T52, T54, T51, T53;
|
Chris@82
|
2754 T52 = ri[WS(rs, 39)];
|
Chris@82
|
2755 T54 = ii[WS(rs, 39)];
|
Chris@82
|
2756 T51 = W[76];
|
Chris@82
|
2757 T53 = W[77];
|
Chris@82
|
2758 T55 = FMA(T51, T52, T53 * T54);
|
Chris@82
|
2759 T8F = FNMS(T53, T52, T51 * T54);
|
Chris@82
|
2760 }
|
Chris@82
|
2761 T56 = T50 + T55;
|
Chris@82
|
2762 TdS = T8E + T8F;
|
Chris@82
|
2763 T8G = T8E - T8F;
|
Chris@82
|
2764 T8H = T50 - T55;
|
Chris@82
|
2765 }
|
Chris@82
|
2766 {
|
Chris@82
|
2767 E T5b, T8K, T5g, T8L;
|
Chris@82
|
2768 {
|
Chris@82
|
2769 E T58, T5a, T57, T59;
|
Chris@82
|
2770 T58 = ri[WS(rs, 55)];
|
Chris@82
|
2771 T5a = ii[WS(rs, 55)];
|
Chris@82
|
2772 T57 = W[108];
|
Chris@82
|
2773 T59 = W[109];
|
Chris@82
|
2774 T5b = FMA(T57, T58, T59 * T5a);
|
Chris@82
|
2775 T8K = FNMS(T59, T58, T57 * T5a);
|
Chris@82
|
2776 }
|
Chris@82
|
2777 {
|
Chris@82
|
2778 E T5d, T5f, T5c, T5e;
|
Chris@82
|
2779 T5d = ri[WS(rs, 23)];
|
Chris@82
|
2780 T5f = ii[WS(rs, 23)];
|
Chris@82
|
2781 T5c = W[44];
|
Chris@82
|
2782 T5e = W[45];
|
Chris@82
|
2783 T5g = FMA(T5c, T5d, T5e * T5f);
|
Chris@82
|
2784 T8L = FNMS(T5e, T5d, T5c * T5f);
|
Chris@82
|
2785 }
|
Chris@82
|
2786 T5h = T5b + T5g;
|
Chris@82
|
2787 TdT = T8K + T8L;
|
Chris@82
|
2788 T8J = T5b - T5g;
|
Chris@82
|
2789 T8M = T8K - T8L;
|
Chris@82
|
2790 }
|
Chris@82
|
2791 {
|
Chris@82
|
2792 E T4V, T5i, Tea, Teb;
|
Chris@82
|
2793 T4V = T4J + T4U;
|
Chris@82
|
2794 T5i = T56 + T5h;
|
Chris@82
|
2795 T5j = T4V + T5i;
|
Chris@82
|
2796 TfR = T4V - T5i;
|
Chris@82
|
2797 Tea = Te8 - Te9;
|
Chris@82
|
2798 Teb = T5h - T56;
|
Chris@82
|
2799 Tec = Tea - Teb;
|
Chris@82
|
2800 Tf0 = Tea + Teb;
|
Chris@82
|
2801 }
|
Chris@82
|
2802 {
|
Chris@82
|
2803 E TfW, TfX, T8z, T8C;
|
Chris@82
|
2804 TfW = Te8 + Te9;
|
Chris@82
|
2805 TfX = TdS + TdT;
|
Chris@82
|
2806 TfY = TfW - TfX;
|
Chris@82
|
2807 Tgy = TfW + TfX;
|
Chris@82
|
2808 T8z = T4D - T4I;
|
Chris@82
|
2809 T8C = T8A - T8B;
|
Chris@82
|
2810 T8D = T8z - T8C;
|
Chris@82
|
2811 Tbl = T8z + T8C;
|
Chris@82
|
2812 }
|
Chris@82
|
2813 {
|
Chris@82
|
2814 E T8I, T8N, T9j, T9k;
|
Chris@82
|
2815 T8I = T8G - T8H;
|
Chris@82
|
2816 T8N = T8J + T8M;
|
Chris@82
|
2817 T8O = KP707106781 * (T8I - T8N);
|
Chris@82
|
2818 Tbx = KP707106781 * (T8I + T8N);
|
Chris@82
|
2819 T9j = T8J - T8M;
|
Chris@82
|
2820 T9k = T8H + T8G;
|
Chris@82
|
2821 T9l = KP707106781 * (T9j - T9k);
|
Chris@82
|
2822 Tbm = KP707106781 * (T9k + T9j);
|
Chris@82
|
2823 }
|
Chris@82
|
2824 {
|
Chris@82
|
2825 E TdR, TdU, T9g, T9h;
|
Chris@82
|
2826 TdR = T4J - T4U;
|
Chris@82
|
2827 TdU = TdS - TdT;
|
Chris@82
|
2828 TdV = TdR - TdU;
|
Chris@82
|
2829 TeX = TdR + TdU;
|
Chris@82
|
2830 T9g = T9e - T9f;
|
Chris@82
|
2831 T9h = T4O - T4T;
|
Chris@82
|
2832 T9i = T9g + T9h;
|
Chris@82
|
2833 Tbw = T9g - T9h;
|
Chris@82
|
2834 }
|
Chris@82
|
2835 }
|
Chris@82
|
2836 {
|
Chris@82
|
2837 E T36, T7G, T3b, T7H, T3c, Tdq, T3h, T8m, T3m, T8n, T3n, Tdr, T3z, TdI, T7Q;
|
Chris@82
|
2838 E T7T, T3K, TdJ, T7L, T7O;
|
Chris@82
|
2839 {
|
Chris@82
|
2840 E T33, T35, T32, T34;
|
Chris@82
|
2841 T33 = ri[WS(rs, 1)];
|
Chris@82
|
2842 T35 = ii[WS(rs, 1)];
|
Chris@82
|
2843 T32 = W[0];
|
Chris@82
|
2844 T34 = W[1];
|
Chris@82
|
2845 T36 = FMA(T32, T33, T34 * T35);
|
Chris@82
|
2846 T7G = FNMS(T34, T33, T32 * T35);
|
Chris@82
|
2847 }
|
Chris@82
|
2848 {
|
Chris@82
|
2849 E T38, T3a, T37, T39;
|
Chris@82
|
2850 T38 = ri[WS(rs, 33)];
|
Chris@82
|
2851 T3a = ii[WS(rs, 33)];
|
Chris@82
|
2852 T37 = W[64];
|
Chris@82
|
2853 T39 = W[65];
|
Chris@82
|
2854 T3b = FMA(T37, T38, T39 * T3a);
|
Chris@82
|
2855 T7H = FNMS(T39, T38, T37 * T3a);
|
Chris@82
|
2856 }
|
Chris@82
|
2857 T3c = T36 + T3b;
|
Chris@82
|
2858 Tdq = T7G + T7H;
|
Chris@82
|
2859 {
|
Chris@82
|
2860 E T3e, T3g, T3d, T3f;
|
Chris@82
|
2861 T3e = ri[WS(rs, 17)];
|
Chris@82
|
2862 T3g = ii[WS(rs, 17)];
|
Chris@82
|
2863 T3d = W[32];
|
Chris@82
|
2864 T3f = W[33];
|
Chris@82
|
2865 T3h = FMA(T3d, T3e, T3f * T3g);
|
Chris@82
|
2866 T8m = FNMS(T3f, T3e, T3d * T3g);
|
Chris@82
|
2867 }
|
Chris@82
|
2868 {
|
Chris@82
|
2869 E T3j, T3l, T3i, T3k;
|
Chris@82
|
2870 T3j = ri[WS(rs, 49)];
|
Chris@82
|
2871 T3l = ii[WS(rs, 49)];
|
Chris@82
|
2872 T3i = W[96];
|
Chris@82
|
2873 T3k = W[97];
|
Chris@82
|
2874 T3m = FMA(T3i, T3j, T3k * T3l);
|
Chris@82
|
2875 T8n = FNMS(T3k, T3j, T3i * T3l);
|
Chris@82
|
2876 }
|
Chris@82
|
2877 T3n = T3h + T3m;
|
Chris@82
|
2878 Tdr = T8m + T8n;
|
Chris@82
|
2879 {
|
Chris@82
|
2880 E T3t, T7R, T3y, T7S;
|
Chris@82
|
2881 {
|
Chris@82
|
2882 E T3q, T3s, T3p, T3r;
|
Chris@82
|
2883 T3q = ri[WS(rs, 9)];
|
Chris@82
|
2884 T3s = ii[WS(rs, 9)];
|
Chris@82
|
2885 T3p = W[16];
|
Chris@82
|
2886 T3r = W[17];
|
Chris@82
|
2887 T3t = FMA(T3p, T3q, T3r * T3s);
|
Chris@82
|
2888 T7R = FNMS(T3r, T3q, T3p * T3s);
|
Chris@82
|
2889 }
|
Chris@82
|
2890 {
|
Chris@82
|
2891 E T3v, T3x, T3u, T3w;
|
Chris@82
|
2892 T3v = ri[WS(rs, 41)];
|
Chris@82
|
2893 T3x = ii[WS(rs, 41)];
|
Chris@82
|
2894 T3u = W[80];
|
Chris@82
|
2895 T3w = W[81];
|
Chris@82
|
2896 T3y = FMA(T3u, T3v, T3w * T3x);
|
Chris@82
|
2897 T7S = FNMS(T3w, T3v, T3u * T3x);
|
Chris@82
|
2898 }
|
Chris@82
|
2899 T3z = T3t + T3y;
|
Chris@82
|
2900 TdI = T7R + T7S;
|
Chris@82
|
2901 T7Q = T3t - T3y;
|
Chris@82
|
2902 T7T = T7R - T7S;
|
Chris@82
|
2903 }
|
Chris@82
|
2904 {
|
Chris@82
|
2905 E T3E, T7M, T3J, T7N;
|
Chris@82
|
2906 {
|
Chris@82
|
2907 E T3B, T3D, T3A, T3C;
|
Chris@82
|
2908 T3B = ri[WS(rs, 57)];
|
Chris@82
|
2909 T3D = ii[WS(rs, 57)];
|
Chris@82
|
2910 T3A = W[112];
|
Chris@82
|
2911 T3C = W[113];
|
Chris@82
|
2912 T3E = FMA(T3A, T3B, T3C * T3D);
|
Chris@82
|
2913 T7M = FNMS(T3C, T3B, T3A * T3D);
|
Chris@82
|
2914 }
|
Chris@82
|
2915 {
|
Chris@82
|
2916 E T3G, T3I, T3F, T3H;
|
Chris@82
|
2917 T3G = ri[WS(rs, 25)];
|
Chris@82
|
2918 T3I = ii[WS(rs, 25)];
|
Chris@82
|
2919 T3F = W[48];
|
Chris@82
|
2920 T3H = W[49];
|
Chris@82
|
2921 T3J = FMA(T3F, T3G, T3H * T3I);
|
Chris@82
|
2922 T7N = FNMS(T3H, T3G, T3F * T3I);
|
Chris@82
|
2923 }
|
Chris@82
|
2924 T3K = T3E + T3J;
|
Chris@82
|
2925 TdJ = T7M + T7N;
|
Chris@82
|
2926 T7L = T3E - T3J;
|
Chris@82
|
2927 T7O = T7M - T7N;
|
Chris@82
|
2928 }
|
Chris@82
|
2929 {
|
Chris@82
|
2930 E T3o, T3L, TdH, TdK;
|
Chris@82
|
2931 T3o = T3c + T3n;
|
Chris@82
|
2932 T3L = T3z + T3K;
|
Chris@82
|
2933 T3M = T3o + T3L;
|
Chris@82
|
2934 TfL = T3o - T3L;
|
Chris@82
|
2935 TdH = T3c - T3n;
|
Chris@82
|
2936 TdK = TdI - TdJ;
|
Chris@82
|
2937 TdL = TdH - TdK;
|
Chris@82
|
2938 TeQ = TdH + TdK;
|
Chris@82
|
2939 }
|
Chris@82
|
2940 {
|
Chris@82
|
2941 E TfG, TfH, T7I, T7J;
|
Chris@82
|
2942 TfG = Tdq + Tdr;
|
Chris@82
|
2943 TfH = TdI + TdJ;
|
Chris@82
|
2944 TfI = TfG - TfH;
|
Chris@82
|
2945 Tgt = TfG + TfH;
|
Chris@82
|
2946 T7I = T7G - T7H;
|
Chris@82
|
2947 T7J = T3h - T3m;
|
Chris@82
|
2948 T7K = T7I + T7J;
|
Chris@82
|
2949 Tb2 = T7I - T7J;
|
Chris@82
|
2950 }
|
Chris@82
|
2951 {
|
Chris@82
|
2952 E T7P, T7U, T8q, T8r;
|
Chris@82
|
2953 T7P = T7L - T7O;
|
Chris@82
|
2954 T7U = T7Q + T7T;
|
Chris@82
|
2955 T7V = KP707106781 * (T7P - T7U);
|
Chris@82
|
2956 Tbe = KP707106781 * (T7U + T7P);
|
Chris@82
|
2957 T8q = T7T - T7Q;
|
Chris@82
|
2958 T8r = T7L + T7O;
|
Chris@82
|
2959 T8s = KP707106781 * (T8q - T8r);
|
Chris@82
|
2960 Tb3 = KP707106781 * (T8q + T8r);
|
Chris@82
|
2961 }
|
Chris@82
|
2962 {
|
Chris@82
|
2963 E Tds, Tdt, T8l, T8o;
|
Chris@82
|
2964 Tds = Tdq - Tdr;
|
Chris@82
|
2965 Tdt = T3K - T3z;
|
Chris@82
|
2966 Tdu = Tds - Tdt;
|
Chris@82
|
2967 TeT = Tds + Tdt;
|
Chris@82
|
2968 T8l = T36 - T3b;
|
Chris@82
|
2969 T8o = T8m - T8n;
|
Chris@82
|
2970 T8p = T8l - T8o;
|
Chris@82
|
2971 Tbd = T8l + T8o;
|
Chris@82
|
2972 }
|
Chris@82
|
2973 }
|
Chris@82
|
2974 {
|
Chris@82
|
2975 E T3X, TdB, T8a, T8d, T4v, Tdx, T80, T85, T48, TdC, T8b, T8g, T4k, Tdw, T7X;
|
Chris@82
|
2976 E T84;
|
Chris@82
|
2977 {
|
Chris@82
|
2978 E T3R, T88, T3W, T89;
|
Chris@82
|
2979 {
|
Chris@82
|
2980 E T3O, T3Q, T3N, T3P;
|
Chris@82
|
2981 T3O = ri[WS(rs, 5)];
|
Chris@82
|
2982 T3Q = ii[WS(rs, 5)];
|
Chris@82
|
2983 T3N = W[8];
|
Chris@82
|
2984 T3P = W[9];
|
Chris@82
|
2985 T3R = FMA(T3N, T3O, T3P * T3Q);
|
Chris@82
|
2986 T88 = FNMS(T3P, T3O, T3N * T3Q);
|
Chris@82
|
2987 }
|
Chris@82
|
2988 {
|
Chris@82
|
2989 E T3T, T3V, T3S, T3U;
|
Chris@82
|
2990 T3T = ri[WS(rs, 37)];
|
Chris@82
|
2991 T3V = ii[WS(rs, 37)];
|
Chris@82
|
2992 T3S = W[72];
|
Chris@82
|
2993 T3U = W[73];
|
Chris@82
|
2994 T3W = FMA(T3S, T3T, T3U * T3V);
|
Chris@82
|
2995 T89 = FNMS(T3U, T3T, T3S * T3V);
|
Chris@82
|
2996 }
|
Chris@82
|
2997 T3X = T3R + T3W;
|
Chris@82
|
2998 TdB = T88 + T89;
|
Chris@82
|
2999 T8a = T88 - T89;
|
Chris@82
|
3000 T8d = T3R - T3W;
|
Chris@82
|
3001 }
|
Chris@82
|
3002 {
|
Chris@82
|
3003 E T4p, T7Y, T4u, T7Z;
|
Chris@82
|
3004 {
|
Chris@82
|
3005 E T4m, T4o, T4l, T4n;
|
Chris@82
|
3006 T4m = ri[WS(rs, 13)];
|
Chris@82
|
3007 T4o = ii[WS(rs, 13)];
|
Chris@82
|
3008 T4l = W[24];
|
Chris@82
|
3009 T4n = W[25];
|
Chris@82
|
3010 T4p = FMA(T4l, T4m, T4n * T4o);
|
Chris@82
|
3011 T7Y = FNMS(T4n, T4m, T4l * T4o);
|
Chris@82
|
3012 }
|
Chris@82
|
3013 {
|
Chris@82
|
3014 E T4r, T4t, T4q, T4s;
|
Chris@82
|
3015 T4r = ri[WS(rs, 45)];
|
Chris@82
|
3016 T4t = ii[WS(rs, 45)];
|
Chris@82
|
3017 T4q = W[88];
|
Chris@82
|
3018 T4s = W[89];
|
Chris@82
|
3019 T4u = FMA(T4q, T4r, T4s * T4t);
|
Chris@82
|
3020 T7Z = FNMS(T4s, T4r, T4q * T4t);
|
Chris@82
|
3021 }
|
Chris@82
|
3022 T4v = T4p + T4u;
|
Chris@82
|
3023 Tdx = T7Y + T7Z;
|
Chris@82
|
3024 T80 = T7Y - T7Z;
|
Chris@82
|
3025 T85 = T4p - T4u;
|
Chris@82
|
3026 }
|
Chris@82
|
3027 {
|
Chris@82
|
3028 E T42, T8e, T47, T8f;
|
Chris@82
|
3029 {
|
Chris@82
|
3030 E T3Z, T41, T3Y, T40;
|
Chris@82
|
3031 T3Z = ri[WS(rs, 21)];
|
Chris@82
|
3032 T41 = ii[WS(rs, 21)];
|
Chris@82
|
3033 T3Y = W[40];
|
Chris@82
|
3034 T40 = W[41];
|
Chris@82
|
3035 T42 = FMA(T3Y, T3Z, T40 * T41);
|
Chris@82
|
3036 T8e = FNMS(T40, T3Z, T3Y * T41);
|
Chris@82
|
3037 }
|
Chris@82
|
3038 {
|
Chris@82
|
3039 E T44, T46, T43, T45;
|
Chris@82
|
3040 T44 = ri[WS(rs, 53)];
|
Chris@82
|
3041 T46 = ii[WS(rs, 53)];
|
Chris@82
|
3042 T43 = W[104];
|
Chris@82
|
3043 T45 = W[105];
|
Chris@82
|
3044 T47 = FMA(T43, T44, T45 * T46);
|
Chris@82
|
3045 T8f = FNMS(T45, T44, T43 * T46);
|
Chris@82
|
3046 }
|
Chris@82
|
3047 T48 = T42 + T47;
|
Chris@82
|
3048 TdC = T8e + T8f;
|
Chris@82
|
3049 T8b = T42 - T47;
|
Chris@82
|
3050 T8g = T8e - T8f;
|
Chris@82
|
3051 }
|
Chris@82
|
3052 {
|
Chris@82
|
3053 E T4e, T82, T4j, T83;
|
Chris@82
|
3054 {
|
Chris@82
|
3055 E T4b, T4d, T4a, T4c;
|
Chris@82
|
3056 T4b = ri[WS(rs, 61)];
|
Chris@82
|
3057 T4d = ii[WS(rs, 61)];
|
Chris@82
|
3058 T4a = W[120];
|
Chris@82
|
3059 T4c = W[121];
|
Chris@82
|
3060 T4e = FMA(T4a, T4b, T4c * T4d);
|
Chris@82
|
3061 T82 = FNMS(T4c, T4b, T4a * T4d);
|
Chris@82
|
3062 }
|
Chris@82
|
3063 {
|
Chris@82
|
3064 E T4g, T4i, T4f, T4h;
|
Chris@82
|
3065 T4g = ri[WS(rs, 29)];
|
Chris@82
|
3066 T4i = ii[WS(rs, 29)];
|
Chris@82
|
3067 T4f = W[56];
|
Chris@82
|
3068 T4h = W[57];
|
Chris@82
|
3069 T4j = FMA(T4f, T4g, T4h * T4i);
|
Chris@82
|
3070 T83 = FNMS(T4h, T4g, T4f * T4i);
|
Chris@82
|
3071 }
|
Chris@82
|
3072 T4k = T4e + T4j;
|
Chris@82
|
3073 Tdw = T82 + T83;
|
Chris@82
|
3074 T7X = T4e - T4j;
|
Chris@82
|
3075 T84 = T82 - T83;
|
Chris@82
|
3076 }
|
Chris@82
|
3077 {
|
Chris@82
|
3078 E T49, T4w, TdA, TdD;
|
Chris@82
|
3079 T49 = T3X + T48;
|
Chris@82
|
3080 T4w = T4k + T4v;
|
Chris@82
|
3081 T4x = T49 + T4w;
|
Chris@82
|
3082 TfJ = T4w - T49;
|
Chris@82
|
3083 TdA = T3X - T48;
|
Chris@82
|
3084 TdD = TdB - TdC;
|
Chris@82
|
3085 TdE = TdA + TdD;
|
Chris@82
|
3086 TdM = TdD - TdA;
|
Chris@82
|
3087 }
|
Chris@82
|
3088 {
|
Chris@82
|
3089 E TfM, TfN, T81, T86;
|
Chris@82
|
3090 TfM = TdB + TdC;
|
Chris@82
|
3091 TfN = Tdw + Tdx;
|
Chris@82
|
3092 TfO = TfM - TfN;
|
Chris@82
|
3093 Tgu = TfM + TfN;
|
Chris@82
|
3094 T81 = T7X - T80;
|
Chris@82
|
3095 T86 = T84 + T85;
|
Chris@82
|
3096 T87 = FNMS(KP923879532, T86, KP382683432 * T81);
|
Chris@82
|
3097 T8v = FMA(KP382683432, T86, KP923879532 * T81);
|
Chris@82
|
3098 }
|
Chris@82
|
3099 {
|
Chris@82
|
3100 E T8c, T8h, Tb8, Tb9;
|
Chris@82
|
3101 T8c = T8a + T8b;
|
Chris@82
|
3102 T8h = T8d - T8g;
|
Chris@82
|
3103 T8i = FMA(KP923879532, T8c, KP382683432 * T8h);
|
Chris@82
|
3104 T8u = FNMS(KP923879532, T8h, KP382683432 * T8c);
|
Chris@82
|
3105 Tb8 = T8a - T8b;
|
Chris@82
|
3106 Tb9 = T8d + T8g;
|
Chris@82
|
3107 Tba = FMA(KP382683432, Tb8, KP923879532 * Tb9);
|
Chris@82
|
3108 Tbg = FNMS(KP382683432, Tb9, KP923879532 * Tb8);
|
Chris@82
|
3109 }
|
Chris@82
|
3110 {
|
Chris@82
|
3111 E Tdv, Tdy, Tb5, Tb6;
|
Chris@82
|
3112 Tdv = T4k - T4v;
|
Chris@82
|
3113 Tdy = Tdw - Tdx;
|
Chris@82
|
3114 Tdz = Tdv - Tdy;
|
Chris@82
|
3115 TdN = Tdv + Tdy;
|
Chris@82
|
3116 Tb5 = T7X + T80;
|
Chris@82
|
3117 Tb6 = T84 - T85;
|
Chris@82
|
3118 Tb7 = FNMS(KP382683432, Tb6, KP923879532 * Tb5);
|
Chris@82
|
3119 Tbh = FMA(KP923879532, Tb6, KP382683432 * Tb5);
|
Chris@82
|
3120 }
|
Chris@82
|
3121 }
|
Chris@82
|
3122 {
|
Chris@82
|
3123 E T5u, TdW, T8S, T8V, T62, Te3, T94, T99, T5F, TdX, T8T, T8Y, T5R, Te2, T93;
|
Chris@82
|
3124 E T96;
|
Chris@82
|
3125 {
|
Chris@82
|
3126 E T5o, T8Q, T5t, T8R;
|
Chris@82
|
3127 {
|
Chris@82
|
3128 E T5l, T5n, T5k, T5m;
|
Chris@82
|
3129 T5l = ri[WS(rs, 3)];
|
Chris@82
|
3130 T5n = ii[WS(rs, 3)];
|
Chris@82
|
3131 T5k = W[4];
|
Chris@82
|
3132 T5m = W[5];
|
Chris@82
|
3133 T5o = FMA(T5k, T5l, T5m * T5n);
|
Chris@82
|
3134 T8Q = FNMS(T5m, T5l, T5k * T5n);
|
Chris@82
|
3135 }
|
Chris@82
|
3136 {
|
Chris@82
|
3137 E T5q, T5s, T5p, T5r;
|
Chris@82
|
3138 T5q = ri[WS(rs, 35)];
|
Chris@82
|
3139 T5s = ii[WS(rs, 35)];
|
Chris@82
|
3140 T5p = W[68];
|
Chris@82
|
3141 T5r = W[69];
|
Chris@82
|
3142 T5t = FMA(T5p, T5q, T5r * T5s);
|
Chris@82
|
3143 T8R = FNMS(T5r, T5q, T5p * T5s);
|
Chris@82
|
3144 }
|
Chris@82
|
3145 T5u = T5o + T5t;
|
Chris@82
|
3146 TdW = T8Q + T8R;
|
Chris@82
|
3147 T8S = T8Q - T8R;
|
Chris@82
|
3148 T8V = T5o - T5t;
|
Chris@82
|
3149 }
|
Chris@82
|
3150 {
|
Chris@82
|
3151 E T5W, T97, T61, T98;
|
Chris@82
|
3152 {
|
Chris@82
|
3153 E T5T, T5V, T5S, T5U;
|
Chris@82
|
3154 T5T = ri[WS(rs, 11)];
|
Chris@82
|
3155 T5V = ii[WS(rs, 11)];
|
Chris@82
|
3156 T5S = W[20];
|
Chris@82
|
3157 T5U = W[21];
|
Chris@82
|
3158 T5W = FMA(T5S, T5T, T5U * T5V);
|
Chris@82
|
3159 T97 = FNMS(T5U, T5T, T5S * T5V);
|
Chris@82
|
3160 }
|
Chris@82
|
3161 {
|
Chris@82
|
3162 E T5Y, T60, T5X, T5Z;
|
Chris@82
|
3163 T5Y = ri[WS(rs, 43)];
|
Chris@82
|
3164 T60 = ii[WS(rs, 43)];
|
Chris@82
|
3165 T5X = W[84];
|
Chris@82
|
3166 T5Z = W[85];
|
Chris@82
|
3167 T61 = FMA(T5X, T5Y, T5Z * T60);
|
Chris@82
|
3168 T98 = FNMS(T5Z, T5Y, T5X * T60);
|
Chris@82
|
3169 }
|
Chris@82
|
3170 T62 = T5W + T61;
|
Chris@82
|
3171 Te3 = T97 + T98;
|
Chris@82
|
3172 T94 = T5W - T61;
|
Chris@82
|
3173 T99 = T97 - T98;
|
Chris@82
|
3174 }
|
Chris@82
|
3175 {
|
Chris@82
|
3176 E T5z, T8W, T5E, T8X;
|
Chris@82
|
3177 {
|
Chris@82
|
3178 E T5w, T5y, T5v, T5x;
|
Chris@82
|
3179 T5w = ri[WS(rs, 19)];
|
Chris@82
|
3180 T5y = ii[WS(rs, 19)];
|
Chris@82
|
3181 T5v = W[36];
|
Chris@82
|
3182 T5x = W[37];
|
Chris@82
|
3183 T5z = FMA(T5v, T5w, T5x * T5y);
|
Chris@82
|
3184 T8W = FNMS(T5x, T5w, T5v * T5y);
|
Chris@82
|
3185 }
|
Chris@82
|
3186 {
|
Chris@82
|
3187 E T5B, T5D, T5A, T5C;
|
Chris@82
|
3188 T5B = ri[WS(rs, 51)];
|
Chris@82
|
3189 T5D = ii[WS(rs, 51)];
|
Chris@82
|
3190 T5A = W[100];
|
Chris@82
|
3191 T5C = W[101];
|
Chris@82
|
3192 T5E = FMA(T5A, T5B, T5C * T5D);
|
Chris@82
|
3193 T8X = FNMS(T5C, T5B, T5A * T5D);
|
Chris@82
|
3194 }
|
Chris@82
|
3195 T5F = T5z + T5E;
|
Chris@82
|
3196 TdX = T8W + T8X;
|
Chris@82
|
3197 T8T = T5z - T5E;
|
Chris@82
|
3198 T8Y = T8W - T8X;
|
Chris@82
|
3199 }
|
Chris@82
|
3200 {
|
Chris@82
|
3201 E T5L, T91, T5Q, T92;
|
Chris@82
|
3202 {
|
Chris@82
|
3203 E T5I, T5K, T5H, T5J;
|
Chris@82
|
3204 T5I = ri[WS(rs, 59)];
|
Chris@82
|
3205 T5K = ii[WS(rs, 59)];
|
Chris@82
|
3206 T5H = W[116];
|
Chris@82
|
3207 T5J = W[117];
|
Chris@82
|
3208 T5L = FMA(T5H, T5I, T5J * T5K);
|
Chris@82
|
3209 T91 = FNMS(T5J, T5I, T5H * T5K);
|
Chris@82
|
3210 }
|
Chris@82
|
3211 {
|
Chris@82
|
3212 E T5N, T5P, T5M, T5O;
|
Chris@82
|
3213 T5N = ri[WS(rs, 27)];
|
Chris@82
|
3214 T5P = ii[WS(rs, 27)];
|
Chris@82
|
3215 T5M = W[52];
|
Chris@82
|
3216 T5O = W[53];
|
Chris@82
|
3217 T5Q = FMA(T5M, T5N, T5O * T5P);
|
Chris@82
|
3218 T92 = FNMS(T5O, T5N, T5M * T5P);
|
Chris@82
|
3219 }
|
Chris@82
|
3220 T5R = T5L + T5Q;
|
Chris@82
|
3221 Te2 = T91 + T92;
|
Chris@82
|
3222 T93 = T91 - T92;
|
Chris@82
|
3223 T96 = T5L - T5Q;
|
Chris@82
|
3224 }
|
Chris@82
|
3225 {
|
Chris@82
|
3226 E T5G, T63, Te1, Te4;
|
Chris@82
|
3227 T5G = T5u + T5F;
|
Chris@82
|
3228 T63 = T5R + T62;
|
Chris@82
|
3229 T64 = T5G + T63;
|
Chris@82
|
3230 TfZ = T63 - T5G;
|
Chris@82
|
3231 Te1 = T5R - T62;
|
Chris@82
|
3232 Te4 = Te2 - Te3;
|
Chris@82
|
3233 Te5 = Te1 + Te4;
|
Chris@82
|
3234 Ted = Te1 - Te4;
|
Chris@82
|
3235 }
|
Chris@82
|
3236 {
|
Chris@82
|
3237 E TfS, TfT, T8U, T8Z;
|
Chris@82
|
3238 TfS = TdW + TdX;
|
Chris@82
|
3239 TfT = Te2 + Te3;
|
Chris@82
|
3240 TfU = TfS - TfT;
|
Chris@82
|
3241 Tgz = TfS + TfT;
|
Chris@82
|
3242 T8U = T8S + T8T;
|
Chris@82
|
3243 T8Z = T8V - T8Y;
|
Chris@82
|
3244 T90 = FNMS(KP923879532, T8Z, KP382683432 * T8U);
|
Chris@82
|
3245 T9o = FMA(KP923879532, T8U, KP382683432 * T8Z);
|
Chris@82
|
3246 }
|
Chris@82
|
3247 {
|
Chris@82
|
3248 E T95, T9a, Tbr, Tbs;
|
Chris@82
|
3249 T95 = T93 + T94;
|
Chris@82
|
3250 T9a = T96 - T99;
|
Chris@82
|
3251 T9b = FMA(KP382683432, T95, KP923879532 * T9a);
|
Chris@82
|
3252 T9n = FNMS(KP923879532, T95, KP382683432 * T9a);
|
Chris@82
|
3253 Tbr = T93 - T94;
|
Chris@82
|
3254 Tbs = T96 + T99;
|
Chris@82
|
3255 Tbt = FMA(KP923879532, Tbr, KP382683432 * Tbs);
|
Chris@82
|
3256 Tbz = FNMS(KP382683432, Tbr, KP923879532 * Tbs);
|
Chris@82
|
3257 }
|
Chris@82
|
3258 {
|
Chris@82
|
3259 E TdY, TdZ, Tbo, Tbp;
|
Chris@82
|
3260 TdY = TdW - TdX;
|
Chris@82
|
3261 TdZ = T5u - T5F;
|
Chris@82
|
3262 Te0 = TdY - TdZ;
|
Chris@82
|
3263 Tee = TdZ + TdY;
|
Chris@82
|
3264 Tbo = T8S - T8T;
|
Chris@82
|
3265 Tbp = T8V + T8Y;
|
Chris@82
|
3266 Tbq = FNMS(KP382683432, Tbp, KP923879532 * Tbo);
|
Chris@82
|
3267 TbA = FMA(KP382683432, Tbo, KP923879532 * Tbp);
|
Chris@82
|
3268 }
|
Chris@82
|
3269 }
|
Chris@82
|
3270 {
|
Chris@82
|
3271 E T1t, Tgn, TgK, TgL, TgV, Th1, T30, Th0, T66, TgX, Tgw, TgE, TgB, TgF, Tgq;
|
Chris@82
|
3272 E TgM;
|
Chris@82
|
3273 {
|
Chris@82
|
3274 E TH, T1s, TgI, TgJ;
|
Chris@82
|
3275 TH = Tj + TG;
|
Chris@82
|
3276 T1s = T14 + T1r;
|
Chris@82
|
3277 T1t = TH + T1s;
|
Chris@82
|
3278 Tgn = TH - T1s;
|
Chris@82
|
3279 TgI = Tgt + Tgu;
|
Chris@82
|
3280 TgJ = Tgy + Tgz;
|
Chris@82
|
3281 TgK = TgI - TgJ;
|
Chris@82
|
3282 TgL = TgI + TgJ;
|
Chris@82
|
3283 }
|
Chris@82
|
3284 {
|
Chris@82
|
3285 E TgN, TgU, T2e, T2Z;
|
Chris@82
|
3286 TgN = Tfq + Tfr;
|
Chris@82
|
3287 TgU = TgO + TgT;
|
Chris@82
|
3288 TgV = TgN + TgU;
|
Chris@82
|
3289 Th1 = TgU - TgN;
|
Chris@82
|
3290 T2e = T1Q + T2d;
|
Chris@82
|
3291 T2Z = T2B + T2Y;
|
Chris@82
|
3292 T30 = T2e + T2Z;
|
Chris@82
|
3293 Th0 = T2Z - T2e;
|
Chris@82
|
3294 }
|
Chris@82
|
3295 {
|
Chris@82
|
3296 E T4y, T65, Tgs, Tgv;
|
Chris@82
|
3297 T4y = T3M + T4x;
|
Chris@82
|
3298 T65 = T5j + T64;
|
Chris@82
|
3299 T66 = T4y + T65;
|
Chris@82
|
3300 TgX = T65 - T4y;
|
Chris@82
|
3301 Tgs = T3M - T4x;
|
Chris@82
|
3302 Tgv = Tgt - Tgu;
|
Chris@82
|
3303 Tgw = Tgs + Tgv;
|
Chris@82
|
3304 TgE = Tgv - Tgs;
|
Chris@82
|
3305 }
|
Chris@82
|
3306 {
|
Chris@82
|
3307 E Tgx, TgA, Tgo, Tgp;
|
Chris@82
|
3308 Tgx = T5j - T64;
|
Chris@82
|
3309 TgA = Tgy - Tgz;
|
Chris@82
|
3310 TgB = Tgx - TgA;
|
Chris@82
|
3311 TgF = Tgx + TgA;
|
Chris@82
|
3312 Tgo = Tfu + Tfv;
|
Chris@82
|
3313 Tgp = TfA + TfB;
|
Chris@82
|
3314 Tgq = Tgo - Tgp;
|
Chris@82
|
3315 TgM = Tgo + Tgp;
|
Chris@82
|
3316 }
|
Chris@82
|
3317 {
|
Chris@82
|
3318 E T31, TgW, TgH, TgY;
|
Chris@82
|
3319 T31 = T1t + T30;
|
Chris@82
|
3320 ri[WS(rs, 32)] = T31 - T66;
|
Chris@82
|
3321 ri[0] = T31 + T66;
|
Chris@82
|
3322 TgW = TgM + TgV;
|
Chris@82
|
3323 ii[0] = TgL + TgW;
|
Chris@82
|
3324 ii[WS(rs, 32)] = TgW - TgL;
|
Chris@82
|
3325 TgH = T1t - T30;
|
Chris@82
|
3326 ri[WS(rs, 48)] = TgH - TgK;
|
Chris@82
|
3327 ri[WS(rs, 16)] = TgH + TgK;
|
Chris@82
|
3328 TgY = TgV - TgM;
|
Chris@82
|
3329 ii[WS(rs, 16)] = TgX + TgY;
|
Chris@82
|
3330 ii[WS(rs, 48)] = TgY - TgX;
|
Chris@82
|
3331 }
|
Chris@82
|
3332 {
|
Chris@82
|
3333 E Tgr, TgC, TgZ, Th2;
|
Chris@82
|
3334 Tgr = Tgn + Tgq;
|
Chris@82
|
3335 TgC = KP707106781 * (Tgw + TgB);
|
Chris@82
|
3336 ri[WS(rs, 40)] = Tgr - TgC;
|
Chris@82
|
3337 ri[WS(rs, 8)] = Tgr + TgC;
|
Chris@82
|
3338 TgZ = KP707106781 * (TgE + TgF);
|
Chris@82
|
3339 Th2 = Th0 + Th1;
|
Chris@82
|
3340 ii[WS(rs, 8)] = TgZ + Th2;
|
Chris@82
|
3341 ii[WS(rs, 40)] = Th2 - TgZ;
|
Chris@82
|
3342 }
|
Chris@82
|
3343 {
|
Chris@82
|
3344 E TgD, TgG, Th3, Th4;
|
Chris@82
|
3345 TgD = Tgn - Tgq;
|
Chris@82
|
3346 TgG = KP707106781 * (TgE - TgF);
|
Chris@82
|
3347 ri[WS(rs, 56)] = TgD - TgG;
|
Chris@82
|
3348 ri[WS(rs, 24)] = TgD + TgG;
|
Chris@82
|
3349 Th3 = KP707106781 * (TgB - Tgw);
|
Chris@82
|
3350 Th4 = Th1 - Th0;
|
Chris@82
|
3351 ii[WS(rs, 24)] = Th3 + Th4;
|
Chris@82
|
3352 ii[WS(rs, 56)] = Th4 - Th3;
|
Chris@82
|
3353 }
|
Chris@82
|
3354 }
|
Chris@82
|
3355 {
|
Chris@82
|
3356 E Tft, Tg7, Tgh, Tgl, Th9, Thf, TfE, Th6, TfQ, Tg4, Tga, The, Tge, Tgk, Tg1;
|
Chris@82
|
3357 E Tg5;
|
Chris@82
|
3358 {
|
Chris@82
|
3359 E Tfp, Tfs, Tgf, Tgg;
|
Chris@82
|
3360 Tfp = Tj - TG;
|
Chris@82
|
3361 Tfs = Tfq - Tfr;
|
Chris@82
|
3362 Tft = Tfp - Tfs;
|
Chris@82
|
3363 Tg7 = Tfp + Tfs;
|
Chris@82
|
3364 Tgf = TfR + TfU;
|
Chris@82
|
3365 Tgg = TfY + TfZ;
|
Chris@82
|
3366 Tgh = FNMS(KP382683432, Tgg, KP923879532 * Tgf);
|
Chris@82
|
3367 Tgl = FMA(KP923879532, Tgg, KP382683432 * Tgf);
|
Chris@82
|
3368 }
|
Chris@82
|
3369 {
|
Chris@82
|
3370 E Th7, Th8, Tfy, TfD;
|
Chris@82
|
3371 Th7 = T1r - T14;
|
Chris@82
|
3372 Th8 = TgT - TgO;
|
Chris@82
|
3373 Th9 = Th7 + Th8;
|
Chris@82
|
3374 Thf = Th8 - Th7;
|
Chris@82
|
3375 Tfy = Tfw - Tfx;
|
Chris@82
|
3376 TfD = Tfz + TfC;
|
Chris@82
|
3377 TfE = KP707106781 * (Tfy - TfD);
|
Chris@82
|
3378 Th6 = KP707106781 * (Tfy + TfD);
|
Chris@82
|
3379 }
|
Chris@82
|
3380 {
|
Chris@82
|
3381 E TfK, TfP, Tg8, Tg9;
|
Chris@82
|
3382 TfK = TfI - TfJ;
|
Chris@82
|
3383 TfP = TfL - TfO;
|
Chris@82
|
3384 TfQ = FMA(KP923879532, TfK, KP382683432 * TfP);
|
Chris@82
|
3385 Tg4 = FNMS(KP923879532, TfP, KP382683432 * TfK);
|
Chris@82
|
3386 Tg8 = Tfx + Tfw;
|
Chris@82
|
3387 Tg9 = Tfz - TfC;
|
Chris@82
|
3388 Tga = KP707106781 * (Tg8 + Tg9);
|
Chris@82
|
3389 The = KP707106781 * (Tg9 - Tg8);
|
Chris@82
|
3390 }
|
Chris@82
|
3391 {
|
Chris@82
|
3392 E Tgc, Tgd, TfV, Tg0;
|
Chris@82
|
3393 Tgc = TfI + TfJ;
|
Chris@82
|
3394 Tgd = TfL + TfO;
|
Chris@82
|
3395 Tge = FMA(KP382683432, Tgc, KP923879532 * Tgd);
|
Chris@82
|
3396 Tgk = FNMS(KP382683432, Tgd, KP923879532 * Tgc);
|
Chris@82
|
3397 TfV = TfR - TfU;
|
Chris@82
|
3398 Tg0 = TfY - TfZ;
|
Chris@82
|
3399 Tg1 = FNMS(KP923879532, Tg0, KP382683432 * TfV);
|
Chris@82
|
3400 Tg5 = FMA(KP382683432, Tg0, KP923879532 * TfV);
|
Chris@82
|
3401 }
|
Chris@82
|
3402 {
|
Chris@82
|
3403 E TfF, Tg2, Thd, Thg;
|
Chris@82
|
3404 TfF = Tft + TfE;
|
Chris@82
|
3405 Tg2 = TfQ + Tg1;
|
Chris@82
|
3406 ri[WS(rs, 44)] = TfF - Tg2;
|
Chris@82
|
3407 ri[WS(rs, 12)] = TfF + Tg2;
|
Chris@82
|
3408 Thd = Tg4 + Tg5;
|
Chris@82
|
3409 Thg = The + Thf;
|
Chris@82
|
3410 ii[WS(rs, 12)] = Thd + Thg;
|
Chris@82
|
3411 ii[WS(rs, 44)] = Thg - Thd;
|
Chris@82
|
3412 }
|
Chris@82
|
3413 {
|
Chris@82
|
3414 E Tg3, Tg6, Thh, Thi;
|
Chris@82
|
3415 Tg3 = Tft - TfE;
|
Chris@82
|
3416 Tg6 = Tg4 - Tg5;
|
Chris@82
|
3417 ri[WS(rs, 60)] = Tg3 - Tg6;
|
Chris@82
|
3418 ri[WS(rs, 28)] = Tg3 + Tg6;
|
Chris@82
|
3419 Thh = Tg1 - TfQ;
|
Chris@82
|
3420 Thi = Thf - The;
|
Chris@82
|
3421 ii[WS(rs, 28)] = Thh + Thi;
|
Chris@82
|
3422 ii[WS(rs, 60)] = Thi - Thh;
|
Chris@82
|
3423 }
|
Chris@82
|
3424 {
|
Chris@82
|
3425 E Tgb, Tgi, Th5, Tha;
|
Chris@82
|
3426 Tgb = Tg7 + Tga;
|
Chris@82
|
3427 Tgi = Tge + Tgh;
|
Chris@82
|
3428 ri[WS(rs, 36)] = Tgb - Tgi;
|
Chris@82
|
3429 ri[WS(rs, 4)] = Tgb + Tgi;
|
Chris@82
|
3430 Th5 = Tgk + Tgl;
|
Chris@82
|
3431 Tha = Th6 + Th9;
|
Chris@82
|
3432 ii[WS(rs, 4)] = Th5 + Tha;
|
Chris@82
|
3433 ii[WS(rs, 36)] = Tha - Th5;
|
Chris@82
|
3434 }
|
Chris@82
|
3435 {
|
Chris@82
|
3436 E Tgj, Tgm, Thb, Thc;
|
Chris@82
|
3437 Tgj = Tg7 - Tga;
|
Chris@82
|
3438 Tgm = Tgk - Tgl;
|
Chris@82
|
3439 ri[WS(rs, 52)] = Tgj - Tgm;
|
Chris@82
|
3440 ri[WS(rs, 20)] = Tgj + Tgm;
|
Chris@82
|
3441 Thb = Tgh - Tge;
|
Chris@82
|
3442 Thc = Th9 - Th6;
|
Chris@82
|
3443 ii[WS(rs, 20)] = Thb + Thc;
|
Chris@82
|
3444 ii[WS(rs, 52)] = Thc - Thb;
|
Chris@82
|
3445 }
|
Chris@82
|
3446 }
|
Chris@82
|
3447 {
|
Chris@82
|
3448 E Td1, Ten, Tdo, ThA, ThD, ThJ, Teq, ThI, Teh, TeB, Tel, Tex, TdQ, TeA, Tek;
|
Chris@82
|
3449 E Teu;
|
Chris@82
|
3450 {
|
Chris@82
|
3451 E TcP, Td0, Teo, Tep;
|
Chris@82
|
3452 TcP = TcL - TcO;
|
Chris@82
|
3453 Td0 = KP707106781 * (TcU - TcZ);
|
Chris@82
|
3454 Td1 = TcP - Td0;
|
Chris@82
|
3455 Ten = TcP + Td0;
|
Chris@82
|
3456 {
|
Chris@82
|
3457 E Tdc, Tdn, ThB, ThC;
|
Chris@82
|
3458 Tdc = FNMS(KP923879532, Tdb, KP382683432 * Td6);
|
Chris@82
|
3459 Tdn = FMA(KP382683432, Tdh, KP923879532 * Tdm);
|
Chris@82
|
3460 Tdo = Tdc - Tdn;
|
Chris@82
|
3461 ThA = Tdc + Tdn;
|
Chris@82
|
3462 ThB = KP707106781 * (TeF - TeE);
|
Chris@82
|
3463 ThC = Thn - Thm;
|
Chris@82
|
3464 ThD = ThB + ThC;
|
Chris@82
|
3465 ThJ = ThC - ThB;
|
Chris@82
|
3466 }
|
Chris@82
|
3467 Teo = FMA(KP923879532, Td6, KP382683432 * Tdb);
|
Chris@82
|
3468 Tep = FNMS(KP923879532, Tdh, KP382683432 * Tdm);
|
Chris@82
|
3469 Teq = Teo + Tep;
|
Chris@82
|
3470 ThI = Tep - Teo;
|
Chris@82
|
3471 {
|
Chris@82
|
3472 E Te7, Tev, Teg, Tew, Te6, Tef;
|
Chris@82
|
3473 Te6 = KP707106781 * (Te0 - Te5);
|
Chris@82
|
3474 Te7 = TdV - Te6;
|
Chris@82
|
3475 Tev = TdV + Te6;
|
Chris@82
|
3476 Tef = KP707106781 * (Ted - Tee);
|
Chris@82
|
3477 Teg = Tec - Tef;
|
Chris@82
|
3478 Tew = Tec + Tef;
|
Chris@82
|
3479 Teh = FNMS(KP980785280, Teg, KP195090322 * Te7);
|
Chris@82
|
3480 TeB = FMA(KP831469612, Tew, KP555570233 * Tev);
|
Chris@82
|
3481 Tel = FMA(KP195090322, Teg, KP980785280 * Te7);
|
Chris@82
|
3482 Tex = FNMS(KP555570233, Tew, KP831469612 * Tev);
|
Chris@82
|
3483 }
|
Chris@82
|
3484 {
|
Chris@82
|
3485 E TdG, Tes, TdP, Tet, TdF, TdO;
|
Chris@82
|
3486 TdF = KP707106781 * (Tdz - TdE);
|
Chris@82
|
3487 TdG = Tdu - TdF;
|
Chris@82
|
3488 Tes = Tdu + TdF;
|
Chris@82
|
3489 TdO = KP707106781 * (TdM - TdN);
|
Chris@82
|
3490 TdP = TdL - TdO;
|
Chris@82
|
3491 Tet = TdL + TdO;
|
Chris@82
|
3492 TdQ = FMA(KP980785280, TdG, KP195090322 * TdP);
|
Chris@82
|
3493 TeA = FNMS(KP555570233, Tet, KP831469612 * Tes);
|
Chris@82
|
3494 Tek = FNMS(KP980785280, TdP, KP195090322 * TdG);
|
Chris@82
|
3495 Teu = FMA(KP555570233, Tes, KP831469612 * Tet);
|
Chris@82
|
3496 }
|
Chris@82
|
3497 }
|
Chris@82
|
3498 {
|
Chris@82
|
3499 E Tdp, Tei, ThH, ThK;
|
Chris@82
|
3500 Tdp = Td1 + Tdo;
|
Chris@82
|
3501 Tei = TdQ + Teh;
|
Chris@82
|
3502 ri[WS(rs, 46)] = Tdp - Tei;
|
Chris@82
|
3503 ri[WS(rs, 14)] = Tdp + Tei;
|
Chris@82
|
3504 ThH = Tek + Tel;
|
Chris@82
|
3505 ThK = ThI + ThJ;
|
Chris@82
|
3506 ii[WS(rs, 14)] = ThH + ThK;
|
Chris@82
|
3507 ii[WS(rs, 46)] = ThK - ThH;
|
Chris@82
|
3508 }
|
Chris@82
|
3509 {
|
Chris@82
|
3510 E Tej, Tem, ThL, ThM;
|
Chris@82
|
3511 Tej = Td1 - Tdo;
|
Chris@82
|
3512 Tem = Tek - Tel;
|
Chris@82
|
3513 ri[WS(rs, 62)] = Tej - Tem;
|
Chris@82
|
3514 ri[WS(rs, 30)] = Tej + Tem;
|
Chris@82
|
3515 ThL = Teh - TdQ;
|
Chris@82
|
3516 ThM = ThJ - ThI;
|
Chris@82
|
3517 ii[WS(rs, 30)] = ThL + ThM;
|
Chris@82
|
3518 ii[WS(rs, 62)] = ThM - ThL;
|
Chris@82
|
3519 }
|
Chris@82
|
3520 {
|
Chris@82
|
3521 E Ter, Tey, Thz, ThE;
|
Chris@82
|
3522 Ter = Ten + Teq;
|
Chris@82
|
3523 Tey = Teu + Tex;
|
Chris@82
|
3524 ri[WS(rs, 38)] = Ter - Tey;
|
Chris@82
|
3525 ri[WS(rs, 6)] = Ter + Tey;
|
Chris@82
|
3526 Thz = TeA + TeB;
|
Chris@82
|
3527 ThE = ThA + ThD;
|
Chris@82
|
3528 ii[WS(rs, 6)] = Thz + ThE;
|
Chris@82
|
3529 ii[WS(rs, 38)] = ThE - Thz;
|
Chris@82
|
3530 }
|
Chris@82
|
3531 {
|
Chris@82
|
3532 E Tez, TeC, ThF, ThG;
|
Chris@82
|
3533 Tez = Ten - Teq;
|
Chris@82
|
3534 TeC = TeA - TeB;
|
Chris@82
|
3535 ri[WS(rs, 54)] = Tez - TeC;
|
Chris@82
|
3536 ri[WS(rs, 22)] = Tez + TeC;
|
Chris@82
|
3537 ThF = Tex - Teu;
|
Chris@82
|
3538 ThG = ThD - ThA;
|
Chris@82
|
3539 ii[WS(rs, 22)] = ThF + ThG;
|
Chris@82
|
3540 ii[WS(rs, 54)] = ThG - ThF;
|
Chris@82
|
3541 }
|
Chris@82
|
3542 }
|
Chris@82
|
3543 {
|
Chris@82
|
3544 E TeH, Tf9, TeO, Thk, Thp, Thv, Tfc, Thu, Tf3, Tfn, Tf7, Tfj, TeW, Tfm, Tf6;
|
Chris@82
|
3545 E Tfg;
|
Chris@82
|
3546 {
|
Chris@82
|
3547 E TeD, TeG, Tfa, Tfb;
|
Chris@82
|
3548 TeD = TcL + TcO;
|
Chris@82
|
3549 TeG = KP707106781 * (TeE + TeF);
|
Chris@82
|
3550 TeH = TeD - TeG;
|
Chris@82
|
3551 Tf9 = TeD + TeG;
|
Chris@82
|
3552 {
|
Chris@82
|
3553 E TeK, TeN, Thl, Tho;
|
Chris@82
|
3554 TeK = FNMS(KP382683432, TeJ, KP923879532 * TeI);
|
Chris@82
|
3555 TeN = FMA(KP923879532, TeL, KP382683432 * TeM);
|
Chris@82
|
3556 TeO = TeK - TeN;
|
Chris@82
|
3557 Thk = TeK + TeN;
|
Chris@82
|
3558 Thl = KP707106781 * (TcU + TcZ);
|
Chris@82
|
3559 Tho = Thm + Thn;
|
Chris@82
|
3560 Thp = Thl + Tho;
|
Chris@82
|
3561 Thv = Tho - Thl;
|
Chris@82
|
3562 }
|
Chris@82
|
3563 Tfa = FMA(KP382683432, TeI, KP923879532 * TeJ);
|
Chris@82
|
3564 Tfb = FNMS(KP382683432, TeL, KP923879532 * TeM);
|
Chris@82
|
3565 Tfc = Tfa + Tfb;
|
Chris@82
|
3566 Thu = Tfb - Tfa;
|
Chris@82
|
3567 {
|
Chris@82
|
3568 E TeZ, Tfh, Tf2, Tfi, TeY, Tf1;
|
Chris@82
|
3569 TeY = KP707106781 * (Tee + Ted);
|
Chris@82
|
3570 TeZ = TeX - TeY;
|
Chris@82
|
3571 Tfh = TeX + TeY;
|
Chris@82
|
3572 Tf1 = KP707106781 * (Te0 + Te5);
|
Chris@82
|
3573 Tf2 = Tf0 - Tf1;
|
Chris@82
|
3574 Tfi = Tf0 + Tf1;
|
Chris@82
|
3575 Tf3 = FNMS(KP831469612, Tf2, KP555570233 * TeZ);
|
Chris@82
|
3576 Tfn = FMA(KP195090322, Tfh, KP980785280 * Tfi);
|
Chris@82
|
3577 Tf7 = FMA(KP831469612, TeZ, KP555570233 * Tf2);
|
Chris@82
|
3578 Tfj = FNMS(KP195090322, Tfi, KP980785280 * Tfh);
|
Chris@82
|
3579 }
|
Chris@82
|
3580 {
|
Chris@82
|
3581 E TeS, Tfe, TeV, Tff, TeR, TeU;
|
Chris@82
|
3582 TeR = KP707106781 * (TdE + Tdz);
|
Chris@82
|
3583 TeS = TeQ - TeR;
|
Chris@82
|
3584 Tfe = TeQ + TeR;
|
Chris@82
|
3585 TeU = KP707106781 * (TdM + TdN);
|
Chris@82
|
3586 TeV = TeT - TeU;
|
Chris@82
|
3587 Tff = TeT + TeU;
|
Chris@82
|
3588 TeW = FMA(KP555570233, TeS, KP831469612 * TeV);
|
Chris@82
|
3589 Tfm = FNMS(KP195090322, Tfe, KP980785280 * Tff);
|
Chris@82
|
3590 Tf6 = FNMS(KP831469612, TeS, KP555570233 * TeV);
|
Chris@82
|
3591 Tfg = FMA(KP980785280, Tfe, KP195090322 * Tff);
|
Chris@82
|
3592 }
|
Chris@82
|
3593 }
|
Chris@82
|
3594 {
|
Chris@82
|
3595 E TeP, Tf4, Tht, Thw;
|
Chris@82
|
3596 TeP = TeH + TeO;
|
Chris@82
|
3597 Tf4 = TeW + Tf3;
|
Chris@82
|
3598 ri[WS(rs, 42)] = TeP - Tf4;
|
Chris@82
|
3599 ri[WS(rs, 10)] = TeP + Tf4;
|
Chris@82
|
3600 Tht = Tf6 + Tf7;
|
Chris@82
|
3601 Thw = Thu + Thv;
|
Chris@82
|
3602 ii[WS(rs, 10)] = Tht + Thw;
|
Chris@82
|
3603 ii[WS(rs, 42)] = Thw - Tht;
|
Chris@82
|
3604 }
|
Chris@82
|
3605 {
|
Chris@82
|
3606 E Tf5, Tf8, Thx, Thy;
|
Chris@82
|
3607 Tf5 = TeH - TeO;
|
Chris@82
|
3608 Tf8 = Tf6 - Tf7;
|
Chris@82
|
3609 ri[WS(rs, 58)] = Tf5 - Tf8;
|
Chris@82
|
3610 ri[WS(rs, 26)] = Tf5 + Tf8;
|
Chris@82
|
3611 Thx = Tf3 - TeW;
|
Chris@82
|
3612 Thy = Thv - Thu;
|
Chris@82
|
3613 ii[WS(rs, 26)] = Thx + Thy;
|
Chris@82
|
3614 ii[WS(rs, 58)] = Thy - Thx;
|
Chris@82
|
3615 }
|
Chris@82
|
3616 {
|
Chris@82
|
3617 E Tfd, Tfk, Thj, Thq;
|
Chris@82
|
3618 Tfd = Tf9 + Tfc;
|
Chris@82
|
3619 Tfk = Tfg + Tfj;
|
Chris@82
|
3620 ri[WS(rs, 34)] = Tfd - Tfk;
|
Chris@82
|
3621 ri[WS(rs, 2)] = Tfd + Tfk;
|
Chris@82
|
3622 Thj = Tfm + Tfn;
|
Chris@82
|
3623 Thq = Thk + Thp;
|
Chris@82
|
3624 ii[WS(rs, 2)] = Thj + Thq;
|
Chris@82
|
3625 ii[WS(rs, 34)] = Thq - Thj;
|
Chris@82
|
3626 }
|
Chris@82
|
3627 {
|
Chris@82
|
3628 E Tfl, Tfo, Thr, Ths;
|
Chris@82
|
3629 Tfl = Tf9 - Tfc;
|
Chris@82
|
3630 Tfo = Tfm - Tfn;
|
Chris@82
|
3631 ri[WS(rs, 50)] = Tfl - Tfo;
|
Chris@82
|
3632 ri[WS(rs, 18)] = Tfl + Tfo;
|
Chris@82
|
3633 Thr = Tfj - Tfg;
|
Chris@82
|
3634 Ths = Thp - Thk;
|
Chris@82
|
3635 ii[WS(rs, 18)] = Thr + Ths;
|
Chris@82
|
3636 ii[WS(rs, 50)] = Ths - Thr;
|
Chris@82
|
3637 }
|
Chris@82
|
3638 }
|
Chris@82
|
3639 {
|
Chris@82
|
3640 E T6L, T9x, TiD, TiJ, T7E, TiI, T9A, TiA, T8y, T9K, T9u, T9E, T9r, T9L, T9v;
|
Chris@82
|
3641 E T9H;
|
Chris@82
|
3642 {
|
Chris@82
|
3643 E T6n, T6K, TiB, TiC;
|
Chris@82
|
3644 T6n = T6b - T6m;
|
Chris@82
|
3645 T6K = T6y - T6J;
|
Chris@82
|
3646 T6L = T6n - T6K;
|
Chris@82
|
3647 T9x = T6n + T6K;
|
Chris@82
|
3648 TiB = T9P - T9O;
|
Chris@82
|
3649 TiC = Tin - Tim;
|
Chris@82
|
3650 TiD = TiB + TiC;
|
Chris@82
|
3651 TiJ = TiC - TiB;
|
Chris@82
|
3652 }
|
Chris@82
|
3653 {
|
Chris@82
|
3654 E T7c, T9y, T7D, T9z;
|
Chris@82
|
3655 {
|
Chris@82
|
3656 E T72, T7b, T7t, T7C;
|
Chris@82
|
3657 T72 = T6Q - T71;
|
Chris@82
|
3658 T7b = T77 - T7a;
|
Chris@82
|
3659 T7c = FNMS(KP980785280, T7b, KP195090322 * T72);
|
Chris@82
|
3660 T9y = FMA(KP980785280, T72, KP195090322 * T7b);
|
Chris@82
|
3661 T7t = T7h - T7s;
|
Chris@82
|
3662 T7C = T7y - T7B;
|
Chris@82
|
3663 T7D = FMA(KP195090322, T7t, KP980785280 * T7C);
|
Chris@82
|
3664 T9z = FNMS(KP980785280, T7t, KP195090322 * T7C);
|
Chris@82
|
3665 }
|
Chris@82
|
3666 T7E = T7c - T7D;
|
Chris@82
|
3667 TiI = T9z - T9y;
|
Chris@82
|
3668 T9A = T9y + T9z;
|
Chris@82
|
3669 TiA = T7c + T7D;
|
Chris@82
|
3670 }
|
Chris@82
|
3671 {
|
Chris@82
|
3672 E T8k, T9C, T8x, T9D;
|
Chris@82
|
3673 {
|
Chris@82
|
3674 E T7W, T8j, T8t, T8w;
|
Chris@82
|
3675 T7W = T7K - T7V;
|
Chris@82
|
3676 T8j = T87 - T8i;
|
Chris@82
|
3677 T8k = T7W - T8j;
|
Chris@82
|
3678 T9C = T7W + T8j;
|
Chris@82
|
3679 T8t = T8p - T8s;
|
Chris@82
|
3680 T8w = T8u - T8v;
|
Chris@82
|
3681 T8x = T8t - T8w;
|
Chris@82
|
3682 T9D = T8t + T8w;
|
Chris@82
|
3683 }
|
Chris@82
|
3684 T8y = FMA(KP995184726, T8k, KP098017140 * T8x);
|
Chris@82
|
3685 T9K = FNMS(KP634393284, T9D, KP773010453 * T9C);
|
Chris@82
|
3686 T9u = FNMS(KP995184726, T8x, KP098017140 * T8k);
|
Chris@82
|
3687 T9E = FMA(KP634393284, T9C, KP773010453 * T9D);
|
Chris@82
|
3688 }
|
Chris@82
|
3689 {
|
Chris@82
|
3690 E T9d, T9F, T9q, T9G;
|
Chris@82
|
3691 {
|
Chris@82
|
3692 E T8P, T9c, T9m, T9p;
|
Chris@82
|
3693 T8P = T8D - T8O;
|
Chris@82
|
3694 T9c = T90 - T9b;
|
Chris@82
|
3695 T9d = T8P - T9c;
|
Chris@82
|
3696 T9F = T8P + T9c;
|
Chris@82
|
3697 T9m = T9i - T9l;
|
Chris@82
|
3698 T9p = T9n - T9o;
|
Chris@82
|
3699 T9q = T9m - T9p;
|
Chris@82
|
3700 T9G = T9m + T9p;
|
Chris@82
|
3701 }
|
Chris@82
|
3702 T9r = FNMS(KP995184726, T9q, KP098017140 * T9d);
|
Chris@82
|
3703 T9L = FMA(KP773010453, T9G, KP634393284 * T9F);
|
Chris@82
|
3704 T9v = FMA(KP098017140, T9q, KP995184726 * T9d);
|
Chris@82
|
3705 T9H = FNMS(KP634393284, T9G, KP773010453 * T9F);
|
Chris@82
|
3706 }
|
Chris@82
|
3707 {
|
Chris@82
|
3708 E T7F, T9s, TiH, TiK;
|
Chris@82
|
3709 T7F = T6L + T7E;
|
Chris@82
|
3710 T9s = T8y + T9r;
|
Chris@82
|
3711 ri[WS(rs, 47)] = T7F - T9s;
|
Chris@82
|
3712 ri[WS(rs, 15)] = T7F + T9s;
|
Chris@82
|
3713 TiH = T9u + T9v;
|
Chris@82
|
3714 TiK = TiI + TiJ;
|
Chris@82
|
3715 ii[WS(rs, 15)] = TiH + TiK;
|
Chris@82
|
3716 ii[WS(rs, 47)] = TiK - TiH;
|
Chris@82
|
3717 }
|
Chris@82
|
3718 {
|
Chris@82
|
3719 E T9t, T9w, TiL, TiM;
|
Chris@82
|
3720 T9t = T6L - T7E;
|
Chris@82
|
3721 T9w = T9u - T9v;
|
Chris@82
|
3722 ri[WS(rs, 63)] = T9t - T9w;
|
Chris@82
|
3723 ri[WS(rs, 31)] = T9t + T9w;
|
Chris@82
|
3724 TiL = T9r - T8y;
|
Chris@82
|
3725 TiM = TiJ - TiI;
|
Chris@82
|
3726 ii[WS(rs, 31)] = TiL + TiM;
|
Chris@82
|
3727 ii[WS(rs, 63)] = TiM - TiL;
|
Chris@82
|
3728 }
|
Chris@82
|
3729 {
|
Chris@82
|
3730 E T9B, T9I, Tiz, TiE;
|
Chris@82
|
3731 T9B = T9x + T9A;
|
Chris@82
|
3732 T9I = T9E + T9H;
|
Chris@82
|
3733 ri[WS(rs, 39)] = T9B - T9I;
|
Chris@82
|
3734 ri[WS(rs, 7)] = T9B + T9I;
|
Chris@82
|
3735 Tiz = T9K + T9L;
|
Chris@82
|
3736 TiE = TiA + TiD;
|
Chris@82
|
3737 ii[WS(rs, 7)] = Tiz + TiE;
|
Chris@82
|
3738 ii[WS(rs, 39)] = TiE - Tiz;
|
Chris@82
|
3739 }
|
Chris@82
|
3740 {
|
Chris@82
|
3741 E T9J, T9M, TiF, TiG;
|
Chris@82
|
3742 T9J = T9x - T9A;
|
Chris@82
|
3743 T9M = T9K - T9L;
|
Chris@82
|
3744 ri[WS(rs, 55)] = T9J - T9M;
|
Chris@82
|
3745 ri[WS(rs, 23)] = T9J + T9M;
|
Chris@82
|
3746 TiF = T9H - T9E;
|
Chris@82
|
3747 TiG = TiD - TiA;
|
Chris@82
|
3748 ii[WS(rs, 23)] = TiF + TiG;
|
Chris@82
|
3749 ii[WS(rs, 55)] = TiG - TiF;
|
Chris@82
|
3750 }
|
Chris@82
|
3751 }
|
Chris@82
|
3752 {
|
Chris@82
|
3753 E TaL, TbJ, Ti9, Tif, Tb0, Tie, TbM, Ti6, Tbk, TbW, TbG, TbQ, TbD, TbX, TbH;
|
Chris@82
|
3754 E TbT;
|
Chris@82
|
3755 {
|
Chris@82
|
3756 E TaD, TaK, Ti7, Ti8;
|
Chris@82
|
3757 TaD = Taz - TaC;
|
Chris@82
|
3758 TaK = TaG - TaJ;
|
Chris@82
|
3759 TaL = TaD - TaK;
|
Chris@82
|
3760 TbJ = TaD + TaK;
|
Chris@82
|
3761 Ti7 = Tc1 - Tc0;
|
Chris@82
|
3762 Ti8 = ThT - ThQ;
|
Chris@82
|
3763 Ti9 = Ti7 + Ti8;
|
Chris@82
|
3764 Tif = Ti8 - Ti7;
|
Chris@82
|
3765 }
|
Chris@82
|
3766 {
|
Chris@82
|
3767 E TaS, TbK, TaZ, TbL;
|
Chris@82
|
3768 {
|
Chris@82
|
3769 E TaO, TaR, TaV, TaY;
|
Chris@82
|
3770 TaO = TaM - TaN;
|
Chris@82
|
3771 TaR = TaP - TaQ;
|
Chris@82
|
3772 TaS = FNMS(KP831469612, TaR, KP555570233 * TaO);
|
Chris@82
|
3773 TbK = FMA(KP555570233, TaR, KP831469612 * TaO);
|
Chris@82
|
3774 TaV = TaT - TaU;
|
Chris@82
|
3775 TaY = TaW - TaX;
|
Chris@82
|
3776 TaZ = FMA(KP831469612, TaV, KP555570233 * TaY);
|
Chris@82
|
3777 TbL = FNMS(KP831469612, TaY, KP555570233 * TaV);
|
Chris@82
|
3778 }
|
Chris@82
|
3779 Tb0 = TaS - TaZ;
|
Chris@82
|
3780 Tie = TbL - TbK;
|
Chris@82
|
3781 TbM = TbK + TbL;
|
Chris@82
|
3782 Ti6 = TaS + TaZ;
|
Chris@82
|
3783 }
|
Chris@82
|
3784 {
|
Chris@82
|
3785 E Tbc, TbO, Tbj, TbP;
|
Chris@82
|
3786 {
|
Chris@82
|
3787 E Tb4, Tbb, Tbf, Tbi;
|
Chris@82
|
3788 Tb4 = Tb2 - Tb3;
|
Chris@82
|
3789 Tbb = Tb7 - Tba;
|
Chris@82
|
3790 Tbc = Tb4 - Tbb;
|
Chris@82
|
3791 TbO = Tb4 + Tbb;
|
Chris@82
|
3792 Tbf = Tbd - Tbe;
|
Chris@82
|
3793 Tbi = Tbg - Tbh;
|
Chris@82
|
3794 Tbj = Tbf - Tbi;
|
Chris@82
|
3795 TbP = Tbf + Tbi;
|
Chris@82
|
3796 }
|
Chris@82
|
3797 Tbk = FMA(KP956940335, Tbc, KP290284677 * Tbj);
|
Chris@82
|
3798 TbW = FNMS(KP471396736, TbP, KP881921264 * TbO);
|
Chris@82
|
3799 TbG = FNMS(KP956940335, Tbj, KP290284677 * Tbc);
|
Chris@82
|
3800 TbQ = FMA(KP471396736, TbO, KP881921264 * TbP);
|
Chris@82
|
3801 }
|
Chris@82
|
3802 {
|
Chris@82
|
3803 E Tbv, TbR, TbC, TbS;
|
Chris@82
|
3804 {
|
Chris@82
|
3805 E Tbn, Tbu, Tby, TbB;
|
Chris@82
|
3806 Tbn = Tbl - Tbm;
|
Chris@82
|
3807 Tbu = Tbq - Tbt;
|
Chris@82
|
3808 Tbv = Tbn - Tbu;
|
Chris@82
|
3809 TbR = Tbn + Tbu;
|
Chris@82
|
3810 Tby = Tbw - Tbx;
|
Chris@82
|
3811 TbB = Tbz - TbA;
|
Chris@82
|
3812 TbC = Tby - TbB;
|
Chris@82
|
3813 TbS = Tby + TbB;
|
Chris@82
|
3814 }
|
Chris@82
|
3815 TbD = FNMS(KP956940335, TbC, KP290284677 * Tbv);
|
Chris@82
|
3816 TbX = FMA(KP881921264, TbS, KP471396736 * TbR);
|
Chris@82
|
3817 TbH = FMA(KP290284677, TbC, KP956940335 * Tbv);
|
Chris@82
|
3818 TbT = FNMS(KP471396736, TbS, KP881921264 * TbR);
|
Chris@82
|
3819 }
|
Chris@82
|
3820 {
|
Chris@82
|
3821 E Tb1, TbE, Tid, Tig;
|
Chris@82
|
3822 Tb1 = TaL + Tb0;
|
Chris@82
|
3823 TbE = Tbk + TbD;
|
Chris@82
|
3824 ri[WS(rs, 45)] = Tb1 - TbE;
|
Chris@82
|
3825 ri[WS(rs, 13)] = Tb1 + TbE;
|
Chris@82
|
3826 Tid = TbG + TbH;
|
Chris@82
|
3827 Tig = Tie + Tif;
|
Chris@82
|
3828 ii[WS(rs, 13)] = Tid + Tig;
|
Chris@82
|
3829 ii[WS(rs, 45)] = Tig - Tid;
|
Chris@82
|
3830 }
|
Chris@82
|
3831 {
|
Chris@82
|
3832 E TbF, TbI, Tih, Tii;
|
Chris@82
|
3833 TbF = TaL - Tb0;
|
Chris@82
|
3834 TbI = TbG - TbH;
|
Chris@82
|
3835 ri[WS(rs, 61)] = TbF - TbI;
|
Chris@82
|
3836 ri[WS(rs, 29)] = TbF + TbI;
|
Chris@82
|
3837 Tih = TbD - Tbk;
|
Chris@82
|
3838 Tii = Tif - Tie;
|
Chris@82
|
3839 ii[WS(rs, 29)] = Tih + Tii;
|
Chris@82
|
3840 ii[WS(rs, 61)] = Tii - Tih;
|
Chris@82
|
3841 }
|
Chris@82
|
3842 {
|
Chris@82
|
3843 E TbN, TbU, Ti5, Tia;
|
Chris@82
|
3844 TbN = TbJ + TbM;
|
Chris@82
|
3845 TbU = TbQ + TbT;
|
Chris@82
|
3846 ri[WS(rs, 37)] = TbN - TbU;
|
Chris@82
|
3847 ri[WS(rs, 5)] = TbN + TbU;
|
Chris@82
|
3848 Ti5 = TbW + TbX;
|
Chris@82
|
3849 Tia = Ti6 + Ti9;
|
Chris@82
|
3850 ii[WS(rs, 5)] = Ti5 + Tia;
|
Chris@82
|
3851 ii[WS(rs, 37)] = Tia - Ti5;
|
Chris@82
|
3852 }
|
Chris@82
|
3853 {
|
Chris@82
|
3854 E TbV, TbY, Tib, Tic;
|
Chris@82
|
3855 TbV = TbJ - TbM;
|
Chris@82
|
3856 TbY = TbW - TbX;
|
Chris@82
|
3857 ri[WS(rs, 53)] = TbV - TbY;
|
Chris@82
|
3858 ri[WS(rs, 21)] = TbV + TbY;
|
Chris@82
|
3859 Tib = TbT - TbQ;
|
Chris@82
|
3860 Tic = Ti9 - Ti6;
|
Chris@82
|
3861 ii[WS(rs, 21)] = Tib + Tic;
|
Chris@82
|
3862 ii[WS(rs, 53)] = Tic - Tib;
|
Chris@82
|
3863 }
|
Chris@82
|
3864 }
|
Chris@82
|
3865 {
|
Chris@82
|
3866 E Tc3, Tcv, ThV, Ti1, Tca, Ti0, Tcy, ThO, Tci, TcI, Tcs, TcC, Tcp, TcJ, Tct;
|
Chris@82
|
3867 E TcF;
|
Chris@82
|
3868 {
|
Chris@82
|
3869 E TbZ, Tc2, ThP, ThU;
|
Chris@82
|
3870 TbZ = Taz + TaC;
|
Chris@82
|
3871 Tc2 = Tc0 + Tc1;
|
Chris@82
|
3872 Tc3 = TbZ - Tc2;
|
Chris@82
|
3873 Tcv = TbZ + Tc2;
|
Chris@82
|
3874 ThP = TaG + TaJ;
|
Chris@82
|
3875 ThU = ThQ + ThT;
|
Chris@82
|
3876 ThV = ThP + ThU;
|
Chris@82
|
3877 Ti1 = ThU - ThP;
|
Chris@82
|
3878 }
|
Chris@82
|
3879 {
|
Chris@82
|
3880 E Tc6, Tcw, Tc9, Tcx;
|
Chris@82
|
3881 {
|
Chris@82
|
3882 E Tc4, Tc5, Tc7, Tc8;
|
Chris@82
|
3883 Tc4 = TaM + TaN;
|
Chris@82
|
3884 Tc5 = TaP + TaQ;
|
Chris@82
|
3885 Tc6 = FNMS(KP195090322, Tc5, KP980785280 * Tc4);
|
Chris@82
|
3886 Tcw = FMA(KP980785280, Tc5, KP195090322 * Tc4);
|
Chris@82
|
3887 Tc7 = TaT + TaU;
|
Chris@82
|
3888 Tc8 = TaW + TaX;
|
Chris@82
|
3889 Tc9 = FMA(KP195090322, Tc7, KP980785280 * Tc8);
|
Chris@82
|
3890 Tcx = FNMS(KP195090322, Tc8, KP980785280 * Tc7);
|
Chris@82
|
3891 }
|
Chris@82
|
3892 Tca = Tc6 - Tc9;
|
Chris@82
|
3893 Ti0 = Tcx - Tcw;
|
Chris@82
|
3894 Tcy = Tcw + Tcx;
|
Chris@82
|
3895 ThO = Tc6 + Tc9;
|
Chris@82
|
3896 }
|
Chris@82
|
3897 {
|
Chris@82
|
3898 E Tce, TcA, Tch, TcB;
|
Chris@82
|
3899 {
|
Chris@82
|
3900 E Tcc, Tcd, Tcf, Tcg;
|
Chris@82
|
3901 Tcc = Tbd + Tbe;
|
Chris@82
|
3902 Tcd = Tba + Tb7;
|
Chris@82
|
3903 Tce = Tcc - Tcd;
|
Chris@82
|
3904 TcA = Tcc + Tcd;
|
Chris@82
|
3905 Tcf = Tb2 + Tb3;
|
Chris@82
|
3906 Tcg = Tbg + Tbh;
|
Chris@82
|
3907 Tch = Tcf - Tcg;
|
Chris@82
|
3908 TcB = Tcf + Tcg;
|
Chris@82
|
3909 }
|
Chris@82
|
3910 Tci = FMA(KP634393284, Tce, KP773010453 * Tch);
|
Chris@82
|
3911 TcI = FNMS(KP098017140, TcA, KP995184726 * TcB);
|
Chris@82
|
3912 Tcs = FNMS(KP773010453, Tce, KP634393284 * Tch);
|
Chris@82
|
3913 TcC = FMA(KP995184726, TcA, KP098017140 * TcB);
|
Chris@82
|
3914 }
|
Chris@82
|
3915 {
|
Chris@82
|
3916 E Tcl, TcD, Tco, TcE;
|
Chris@82
|
3917 {
|
Chris@82
|
3918 E Tcj, Tck, Tcm, Tcn;
|
Chris@82
|
3919 Tcj = Tbl + Tbm;
|
Chris@82
|
3920 Tck = TbA + Tbz;
|
Chris@82
|
3921 Tcl = Tcj - Tck;
|
Chris@82
|
3922 TcD = Tcj + Tck;
|
Chris@82
|
3923 Tcm = Tbw + Tbx;
|
Chris@82
|
3924 Tcn = Tbq + Tbt;
|
Chris@82
|
3925 Tco = Tcm - Tcn;
|
Chris@82
|
3926 TcE = Tcm + Tcn;
|
Chris@82
|
3927 }
|
Chris@82
|
3928 Tcp = FNMS(KP773010453, Tco, KP634393284 * Tcl);
|
Chris@82
|
3929 TcJ = FMA(KP098017140, TcD, KP995184726 * TcE);
|
Chris@82
|
3930 Tct = FMA(KP773010453, Tcl, KP634393284 * Tco);
|
Chris@82
|
3931 TcF = FNMS(KP098017140, TcE, KP995184726 * TcD);
|
Chris@82
|
3932 }
|
Chris@82
|
3933 {
|
Chris@82
|
3934 E Tcb, Tcq, ThZ, Ti2;
|
Chris@82
|
3935 Tcb = Tc3 + Tca;
|
Chris@82
|
3936 Tcq = Tci + Tcp;
|
Chris@82
|
3937 ri[WS(rs, 41)] = Tcb - Tcq;
|
Chris@82
|
3938 ri[WS(rs, 9)] = Tcb + Tcq;
|
Chris@82
|
3939 ThZ = Tcs + Tct;
|
Chris@82
|
3940 Ti2 = Ti0 + Ti1;
|
Chris@82
|
3941 ii[WS(rs, 9)] = ThZ + Ti2;
|
Chris@82
|
3942 ii[WS(rs, 41)] = Ti2 - ThZ;
|
Chris@82
|
3943 }
|
Chris@82
|
3944 {
|
Chris@82
|
3945 E Tcr, Tcu, Ti3, Ti4;
|
Chris@82
|
3946 Tcr = Tc3 - Tca;
|
Chris@82
|
3947 Tcu = Tcs - Tct;
|
Chris@82
|
3948 ri[WS(rs, 57)] = Tcr - Tcu;
|
Chris@82
|
3949 ri[WS(rs, 25)] = Tcr + Tcu;
|
Chris@82
|
3950 Ti3 = Tcp - Tci;
|
Chris@82
|
3951 Ti4 = Ti1 - Ti0;
|
Chris@82
|
3952 ii[WS(rs, 25)] = Ti3 + Ti4;
|
Chris@82
|
3953 ii[WS(rs, 57)] = Ti4 - Ti3;
|
Chris@82
|
3954 }
|
Chris@82
|
3955 {
|
Chris@82
|
3956 E Tcz, TcG, ThN, ThW;
|
Chris@82
|
3957 Tcz = Tcv + Tcy;
|
Chris@82
|
3958 TcG = TcC + TcF;
|
Chris@82
|
3959 ri[WS(rs, 33)] = Tcz - TcG;
|
Chris@82
|
3960 ri[WS(rs, 1)] = Tcz + TcG;
|
Chris@82
|
3961 ThN = TcI + TcJ;
|
Chris@82
|
3962 ThW = ThO + ThV;
|
Chris@82
|
3963 ii[WS(rs, 1)] = ThN + ThW;
|
Chris@82
|
3964 ii[WS(rs, 33)] = ThW - ThN;
|
Chris@82
|
3965 }
|
Chris@82
|
3966 {
|
Chris@82
|
3967 E TcH, TcK, ThX, ThY;
|
Chris@82
|
3968 TcH = Tcv - Tcy;
|
Chris@82
|
3969 TcK = TcI - TcJ;
|
Chris@82
|
3970 ri[WS(rs, 49)] = TcH - TcK;
|
Chris@82
|
3971 ri[WS(rs, 17)] = TcH + TcK;
|
Chris@82
|
3972 ThX = TcF - TcC;
|
Chris@82
|
3973 ThY = ThV - ThO;
|
Chris@82
|
3974 ii[WS(rs, 17)] = ThX + ThY;
|
Chris@82
|
3975 ii[WS(rs, 49)] = ThY - ThX;
|
Chris@82
|
3976 }
|
Chris@82
|
3977 }
|
Chris@82
|
3978 {
|
Chris@82
|
3979 E T9R, Taj, Tip, Tiv, T9Y, Tiu, Tam, Tik, Ta6, Taw, Tag, Taq, Tad, Tax, Tah;
|
Chris@82
|
3980 E Tat;
|
Chris@82
|
3981 {
|
Chris@82
|
3982 E T9N, T9Q, Til, Tio;
|
Chris@82
|
3983 T9N = T6b + T6m;
|
Chris@82
|
3984 T9Q = T9O + T9P;
|
Chris@82
|
3985 T9R = T9N - T9Q;
|
Chris@82
|
3986 Taj = T9N + T9Q;
|
Chris@82
|
3987 Til = T6y + T6J;
|
Chris@82
|
3988 Tio = Tim + Tin;
|
Chris@82
|
3989 Tip = Til + Tio;
|
Chris@82
|
3990 Tiv = Tio - Til;
|
Chris@82
|
3991 }
|
Chris@82
|
3992 {
|
Chris@82
|
3993 E T9U, Tak, T9X, Tal;
|
Chris@82
|
3994 {
|
Chris@82
|
3995 E T9S, T9T, T9V, T9W;
|
Chris@82
|
3996 T9S = T6Q + T71;
|
Chris@82
|
3997 T9T = T77 + T7a;
|
Chris@82
|
3998 T9U = FNMS(KP555570233, T9T, KP831469612 * T9S);
|
Chris@82
|
3999 Tak = FMA(KP555570233, T9S, KP831469612 * T9T);
|
Chris@82
|
4000 T9V = T7h + T7s;
|
Chris@82
|
4001 T9W = T7y + T7B;
|
Chris@82
|
4002 T9X = FMA(KP831469612, T9V, KP555570233 * T9W);
|
Chris@82
|
4003 Tal = FNMS(KP555570233, T9V, KP831469612 * T9W);
|
Chris@82
|
4004 }
|
Chris@82
|
4005 T9Y = T9U - T9X;
|
Chris@82
|
4006 Tiu = Tal - Tak;
|
Chris@82
|
4007 Tam = Tak + Tal;
|
Chris@82
|
4008 Tik = T9U + T9X;
|
Chris@82
|
4009 }
|
Chris@82
|
4010 {
|
Chris@82
|
4011 E Ta2, Tao, Ta5, Tap;
|
Chris@82
|
4012 {
|
Chris@82
|
4013 E Ta0, Ta1, Ta3, Ta4;
|
Chris@82
|
4014 Ta0 = T8p + T8s;
|
Chris@82
|
4015 Ta1 = T8i + T87;
|
Chris@82
|
4016 Ta2 = Ta0 - Ta1;
|
Chris@82
|
4017 Tao = Ta0 + Ta1;
|
Chris@82
|
4018 Ta3 = T7K + T7V;
|
Chris@82
|
4019 Ta4 = T8u + T8v;
|
Chris@82
|
4020 Ta5 = Ta3 - Ta4;
|
Chris@82
|
4021 Tap = Ta3 + Ta4;
|
Chris@82
|
4022 }
|
Chris@82
|
4023 Ta6 = FMA(KP471396736, Ta2, KP881921264 * Ta5);
|
Chris@82
|
4024 Taw = FNMS(KP290284677, Tao, KP956940335 * Tap);
|
Chris@82
|
4025 Tag = FNMS(KP881921264, Ta2, KP471396736 * Ta5);
|
Chris@82
|
4026 Taq = FMA(KP956940335, Tao, KP290284677 * Tap);
|
Chris@82
|
4027 }
|
Chris@82
|
4028 {
|
Chris@82
|
4029 E Ta9, Tar, Tac, Tas;
|
Chris@82
|
4030 {
|
Chris@82
|
4031 E Ta7, Ta8, Taa, Tab;
|
Chris@82
|
4032 Ta7 = T8D + T8O;
|
Chris@82
|
4033 Ta8 = T9o + T9n;
|
Chris@82
|
4034 Ta9 = Ta7 - Ta8;
|
Chris@82
|
4035 Tar = Ta7 + Ta8;
|
Chris@82
|
4036 Taa = T9i + T9l;
|
Chris@82
|
4037 Tab = T90 + T9b;
|
Chris@82
|
4038 Tac = Taa - Tab;
|
Chris@82
|
4039 Tas = Taa + Tab;
|
Chris@82
|
4040 }
|
Chris@82
|
4041 Tad = FNMS(KP881921264, Tac, KP471396736 * Ta9);
|
Chris@82
|
4042 Tax = FMA(KP290284677, Tar, KP956940335 * Tas);
|
Chris@82
|
4043 Tah = FMA(KP881921264, Ta9, KP471396736 * Tac);
|
Chris@82
|
4044 Tat = FNMS(KP290284677, Tas, KP956940335 * Tar);
|
Chris@82
|
4045 }
|
Chris@82
|
4046 {
|
Chris@82
|
4047 E T9Z, Tae, Tit, Tiw;
|
Chris@82
|
4048 T9Z = T9R + T9Y;
|
Chris@82
|
4049 Tae = Ta6 + Tad;
|
Chris@82
|
4050 ri[WS(rs, 43)] = T9Z - Tae;
|
Chris@82
|
4051 ri[WS(rs, 11)] = T9Z + Tae;
|
Chris@82
|
4052 Tit = Tag + Tah;
|
Chris@82
|
4053 Tiw = Tiu + Tiv;
|
Chris@82
|
4054 ii[WS(rs, 11)] = Tit + Tiw;
|
Chris@82
|
4055 ii[WS(rs, 43)] = Tiw - Tit;
|
Chris@82
|
4056 }
|
Chris@82
|
4057 {
|
Chris@82
|
4058 E Taf, Tai, Tix, Tiy;
|
Chris@82
|
4059 Taf = T9R - T9Y;
|
Chris@82
|
4060 Tai = Tag - Tah;
|
Chris@82
|
4061 ri[WS(rs, 59)] = Taf - Tai;
|
Chris@82
|
4062 ri[WS(rs, 27)] = Taf + Tai;
|
Chris@82
|
4063 Tix = Tad - Ta6;
|
Chris@82
|
4064 Tiy = Tiv - Tiu;
|
Chris@82
|
4065 ii[WS(rs, 27)] = Tix + Tiy;
|
Chris@82
|
4066 ii[WS(rs, 59)] = Tiy - Tix;
|
Chris@82
|
4067 }
|
Chris@82
|
4068 {
|
Chris@82
|
4069 E Tan, Tau, Tij, Tiq;
|
Chris@82
|
4070 Tan = Taj + Tam;
|
Chris@82
|
4071 Tau = Taq + Tat;
|
Chris@82
|
4072 ri[WS(rs, 35)] = Tan - Tau;
|
Chris@82
|
4073 ri[WS(rs, 3)] = Tan + Tau;
|
Chris@82
|
4074 Tij = Taw + Tax;
|
Chris@82
|
4075 Tiq = Tik + Tip;
|
Chris@82
|
4076 ii[WS(rs, 3)] = Tij + Tiq;
|
Chris@82
|
4077 ii[WS(rs, 35)] = Tiq - Tij;
|
Chris@82
|
4078 }
|
Chris@82
|
4079 {
|
Chris@82
|
4080 E Tav, Tay, Tir, Tis;
|
Chris@82
|
4081 Tav = Taj - Tam;
|
Chris@82
|
4082 Tay = Taw - Tax;
|
Chris@82
|
4083 ri[WS(rs, 51)] = Tav - Tay;
|
Chris@82
|
4084 ri[WS(rs, 19)] = Tav + Tay;
|
Chris@82
|
4085 Tir = Tat - Taq;
|
Chris@82
|
4086 Tis = Tip - Tik;
|
Chris@82
|
4087 ii[WS(rs, 19)] = Tir + Tis;
|
Chris@82
|
4088 ii[WS(rs, 51)] = Tis - Tir;
|
Chris@82
|
4089 }
|
Chris@82
|
4090 }
|
Chris@82
|
4091 }
|
Chris@82
|
4092 }
|
Chris@82
|
4093 }
|
Chris@82
|
4094
|
Chris@82
|
4095 static const tw_instr twinstr[] = {
|
Chris@82
|
4096 {TW_FULL, 0, 64},
|
Chris@82
|
4097 {TW_NEXT, 1, 0}
|
Chris@82
|
4098 };
|
Chris@82
|
4099
|
Chris@82
|
4100 static const ct_desc desc = { 64, "t1_64", twinstr, &GENUS, {808, 270, 230, 0}, 0, 0, 0 };
|
Chris@82
|
4101
|
Chris@82
|
4102 void X(codelet_t1_64) (planner *p) {
|
Chris@82
|
4103 X(kdft_dit_register) (p, t1_64, &desc);
|
Chris@82
|
4104 }
|
Chris@82
|
4105 #endif
|