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 08:04:12 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_notw.native -fma -compact -variables 4 -pipeline-latency 4 -n 64 -name n1_64 -include dft/scalar/n.h */
|
Chris@82
|
29
|
Chris@82
|
30 /*
|
Chris@82
|
31 * This function contains 912 FP additions, 392 FP multiplications,
|
Chris@82
|
32 * (or, 520 additions, 0 multiplications, 392 fused multiply/add),
|
Chris@82
|
33 * 172 stack variables, 15 constants, and 256 memory accesses
|
Chris@82
|
34 */
|
Chris@82
|
35 #include "dft/scalar/n.h"
|
Chris@82
|
36
|
Chris@82
|
37 static void n1_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@82
|
38 {
|
Chris@82
|
39 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@82
|
40 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@82
|
41 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
Chris@82
|
42 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
Chris@82
|
43 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@82
|
44 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@82
|
45 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
Chris@82
|
46 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
Chris@82
|
47 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@82
|
48 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@82
|
49 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@82
|
50 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
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 i;
|
Chris@82
|
56 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(256, is), MAKE_VOLATILE_STRIDE(256, os)) {
|
Chris@82
|
57 E T37, T7B, T8F, T5Z, Tf, Td9, TbB, TcB, T62, T7C, T2i, TdH, Tah, Tcb, T3e;
|
Chris@82
|
58 E T8G, Tu, TdI, Tak, TbC, Tan, TbD, T2x, Tda, T3m, T65, T7G, T8I, T7J, T8J;
|
Chris@82
|
59 E T3t, T64, TK, Tdd, Tas, Tce, Tav, Tcf, T2N, Tdc, T3G, T6G, T7O, T9k, T7R;
|
Chris@82
|
60 E T9l, T3N, T6H, T1L, TdA, Tbs, Tct, Tdx, Teo, T5j, T6Y, T5Q, T6V, T8y, T9z;
|
Chris@82
|
61 E Tbb, Tcw, T8n, T9C, TZ, Tdf, Taz, Tch, TaC, Tci, T32, Tdg, T3Z, T6J, T7V;
|
Chris@82
|
62 E T9n, T7Y, T9o, T46, T6K, T1g, Tdp, Tb1, Tcm, Tdm, Tej, T4q, T6R, T4X, T6O;
|
Chris@82
|
63 E T8f, T9s, TaK, Tcp, T84, T9v, T1v, Tdn, Tb4, Tcq, Tds, Tek, T4N, T6P, T50;
|
Chris@82
|
64 E T6S, T8i, T9w, TaV, Tcn, T8b, T9t, T20, Tdy, Tbv, Tcx, TdD, Tep, T5G, T6W;
|
Chris@82
|
65 E T5T, T6Z, T8B, T9D, Tbm, Tcu, T8u, T9A;
|
Chris@82
|
66 {
|
Chris@82
|
67 E T3, T35, T26, T5Y, T6, T5X, T29, T36, Ta, T39, T2d, T38, Td, T3b, T2g;
|
Chris@82
|
68 E T3c;
|
Chris@82
|
69 {
|
Chris@82
|
70 E T1, T2, T24, T25;
|
Chris@82
|
71 T1 = ri[0];
|
Chris@82
|
72 T2 = ri[WS(is, 32)];
|
Chris@82
|
73 T3 = T1 + T2;
|
Chris@82
|
74 T35 = T1 - T2;
|
Chris@82
|
75 T24 = ii[0];
|
Chris@82
|
76 T25 = ii[WS(is, 32)];
|
Chris@82
|
77 T26 = T24 + T25;
|
Chris@82
|
78 T5Y = T24 - T25;
|
Chris@82
|
79 }
|
Chris@82
|
80 {
|
Chris@82
|
81 E T4, T5, T27, T28;
|
Chris@82
|
82 T4 = ri[WS(is, 16)];
|
Chris@82
|
83 T5 = ri[WS(is, 48)];
|
Chris@82
|
84 T6 = T4 + T5;
|
Chris@82
|
85 T5X = T4 - T5;
|
Chris@82
|
86 T27 = ii[WS(is, 16)];
|
Chris@82
|
87 T28 = ii[WS(is, 48)];
|
Chris@82
|
88 T29 = T27 + T28;
|
Chris@82
|
89 T36 = T27 - T28;
|
Chris@82
|
90 }
|
Chris@82
|
91 {
|
Chris@82
|
92 E T8, T9, T2b, T2c;
|
Chris@82
|
93 T8 = ri[WS(is, 8)];
|
Chris@82
|
94 T9 = ri[WS(is, 40)];
|
Chris@82
|
95 Ta = T8 + T9;
|
Chris@82
|
96 T39 = T8 - T9;
|
Chris@82
|
97 T2b = ii[WS(is, 8)];
|
Chris@82
|
98 T2c = ii[WS(is, 40)];
|
Chris@82
|
99 T2d = T2b + T2c;
|
Chris@82
|
100 T38 = T2b - T2c;
|
Chris@82
|
101 }
|
Chris@82
|
102 {
|
Chris@82
|
103 E Tb, Tc, T2e, T2f;
|
Chris@82
|
104 Tb = ri[WS(is, 56)];
|
Chris@82
|
105 Tc = ri[WS(is, 24)];
|
Chris@82
|
106 Td = Tb + Tc;
|
Chris@82
|
107 T3b = Tb - Tc;
|
Chris@82
|
108 T2e = ii[WS(is, 56)];
|
Chris@82
|
109 T2f = ii[WS(is, 24)];
|
Chris@82
|
110 T2g = T2e + T2f;
|
Chris@82
|
111 T3c = T2e - T2f;
|
Chris@82
|
112 }
|
Chris@82
|
113 {
|
Chris@82
|
114 E T7, Te, T2a, T2h;
|
Chris@82
|
115 T37 = T35 - T36;
|
Chris@82
|
116 T7B = T35 + T36;
|
Chris@82
|
117 T8F = T5Y - T5X;
|
Chris@82
|
118 T5Z = T5X + T5Y;
|
Chris@82
|
119 T7 = T3 + T6;
|
Chris@82
|
120 Te = Ta + Td;
|
Chris@82
|
121 Tf = T7 + Te;
|
Chris@82
|
122 Td9 = T7 - Te;
|
Chris@82
|
123 {
|
Chris@82
|
124 E Tbz, TbA, T60, T61;
|
Chris@82
|
125 Tbz = Td - Ta;
|
Chris@82
|
126 TbA = T26 - T29;
|
Chris@82
|
127 TbB = Tbz + TbA;
|
Chris@82
|
128 TcB = TbA - Tbz;
|
Chris@82
|
129 T60 = T3b - T3c;
|
Chris@82
|
130 T61 = T39 + T38;
|
Chris@82
|
131 T62 = T60 - T61;
|
Chris@82
|
132 T7C = T61 + T60;
|
Chris@82
|
133 }
|
Chris@82
|
134 T2a = T26 + T29;
|
Chris@82
|
135 T2h = T2d + T2g;
|
Chris@82
|
136 T2i = T2a + T2h;
|
Chris@82
|
137 TdH = T2a - T2h;
|
Chris@82
|
138 {
|
Chris@82
|
139 E Taf, Tag, T3a, T3d;
|
Chris@82
|
140 Taf = T3 - T6;
|
Chris@82
|
141 Tag = T2d - T2g;
|
Chris@82
|
142 Tah = Taf + Tag;
|
Chris@82
|
143 Tcb = Taf - Tag;
|
Chris@82
|
144 T3a = T38 - T39;
|
Chris@82
|
145 T3d = T3b + T3c;
|
Chris@82
|
146 T3e = T3a - T3d;
|
Chris@82
|
147 T8G = T3a + T3d;
|
Chris@82
|
148 }
|
Chris@82
|
149 }
|
Chris@82
|
150 }
|
Chris@82
|
151 {
|
Chris@82
|
152 E Ti, T3j, T2l, T3h, Tl, T3g, T2o, T3k, Tp, T3q, T2s, T3o, Ts, T3n, T2v;
|
Chris@82
|
153 E T3r;
|
Chris@82
|
154 {
|
Chris@82
|
155 E Tg, Th, T2j, T2k;
|
Chris@82
|
156 Tg = ri[WS(is, 4)];
|
Chris@82
|
157 Th = ri[WS(is, 36)];
|
Chris@82
|
158 Ti = Tg + Th;
|
Chris@82
|
159 T3j = Tg - Th;
|
Chris@82
|
160 T2j = ii[WS(is, 4)];
|
Chris@82
|
161 T2k = ii[WS(is, 36)];
|
Chris@82
|
162 T2l = T2j + T2k;
|
Chris@82
|
163 T3h = T2j - T2k;
|
Chris@82
|
164 }
|
Chris@82
|
165 {
|
Chris@82
|
166 E Tj, Tk, T2m, T2n;
|
Chris@82
|
167 Tj = ri[WS(is, 20)];
|
Chris@82
|
168 Tk = ri[WS(is, 52)];
|
Chris@82
|
169 Tl = Tj + Tk;
|
Chris@82
|
170 T3g = Tj - Tk;
|
Chris@82
|
171 T2m = ii[WS(is, 20)];
|
Chris@82
|
172 T2n = ii[WS(is, 52)];
|
Chris@82
|
173 T2o = T2m + T2n;
|
Chris@82
|
174 T3k = T2m - T2n;
|
Chris@82
|
175 }
|
Chris@82
|
176 {
|
Chris@82
|
177 E Tn, To, T2q, T2r;
|
Chris@82
|
178 Tn = ri[WS(is, 60)];
|
Chris@82
|
179 To = ri[WS(is, 28)];
|
Chris@82
|
180 Tp = Tn + To;
|
Chris@82
|
181 T3q = Tn - To;
|
Chris@82
|
182 T2q = ii[WS(is, 60)];
|
Chris@82
|
183 T2r = ii[WS(is, 28)];
|
Chris@82
|
184 T2s = T2q + T2r;
|
Chris@82
|
185 T3o = T2q - T2r;
|
Chris@82
|
186 }
|
Chris@82
|
187 {
|
Chris@82
|
188 E Tq, Tr, T2t, T2u;
|
Chris@82
|
189 Tq = ri[WS(is, 12)];
|
Chris@82
|
190 Tr = ri[WS(is, 44)];
|
Chris@82
|
191 Ts = Tq + Tr;
|
Chris@82
|
192 T3n = Tq - Tr;
|
Chris@82
|
193 T2t = ii[WS(is, 12)];
|
Chris@82
|
194 T2u = ii[WS(is, 44)];
|
Chris@82
|
195 T2v = T2t + T2u;
|
Chris@82
|
196 T3r = T2t - T2u;
|
Chris@82
|
197 }
|
Chris@82
|
198 {
|
Chris@82
|
199 E Tm, Tt, Tai, Taj;
|
Chris@82
|
200 Tm = Ti + Tl;
|
Chris@82
|
201 Tt = Tp + Ts;
|
Chris@82
|
202 Tu = Tm + Tt;
|
Chris@82
|
203 TdI = Tt - Tm;
|
Chris@82
|
204 Tai = Ti - Tl;
|
Chris@82
|
205 Taj = T2l - T2o;
|
Chris@82
|
206 Tak = Tai + Taj;
|
Chris@82
|
207 TbC = Taj - Tai;
|
Chris@82
|
208 }
|
Chris@82
|
209 {
|
Chris@82
|
210 E Tal, Tam, T2p, T2w;
|
Chris@82
|
211 Tal = Tp - Ts;
|
Chris@82
|
212 Tam = T2s - T2v;
|
Chris@82
|
213 Tan = Tal - Tam;
|
Chris@82
|
214 TbD = Tal + Tam;
|
Chris@82
|
215 T2p = T2l + T2o;
|
Chris@82
|
216 T2w = T2s + T2v;
|
Chris@82
|
217 T2x = T2p + T2w;
|
Chris@82
|
218 Tda = T2p - T2w;
|
Chris@82
|
219 }
|
Chris@82
|
220 {
|
Chris@82
|
221 E T3i, T3l, T7E, T7F;
|
Chris@82
|
222 T3i = T3g + T3h;
|
Chris@82
|
223 T3l = T3j - T3k;
|
Chris@82
|
224 T3m = FMA(KP414213562, T3l, T3i);
|
Chris@82
|
225 T65 = FNMS(KP414213562, T3i, T3l);
|
Chris@82
|
226 T7E = T3j + T3k;
|
Chris@82
|
227 T7F = T3h - T3g;
|
Chris@82
|
228 T7G = FMA(KP414213562, T7F, T7E);
|
Chris@82
|
229 T8I = FNMS(KP414213562, T7E, T7F);
|
Chris@82
|
230 }
|
Chris@82
|
231 {
|
Chris@82
|
232 E T7H, T7I, T3p, T3s;
|
Chris@82
|
233 T7H = T3q + T3r;
|
Chris@82
|
234 T7I = T3o - T3n;
|
Chris@82
|
235 T7J = FNMS(KP414213562, T7I, T7H);
|
Chris@82
|
236 T8J = FMA(KP414213562, T7H, T7I);
|
Chris@82
|
237 T3p = T3n + T3o;
|
Chris@82
|
238 T3s = T3q - T3r;
|
Chris@82
|
239 T3t = FNMS(KP414213562, T3s, T3p);
|
Chris@82
|
240 T64 = FMA(KP414213562, T3p, T3s);
|
Chris@82
|
241 }
|
Chris@82
|
242 }
|
Chris@82
|
243 {
|
Chris@82
|
244 E Ty, T3H, T2B, T3x, TB, T3w, T2E, T3I, TI, T3K, T2L, T3E, TF, T3L, T2I;
|
Chris@82
|
245 E T3B;
|
Chris@82
|
246 {
|
Chris@82
|
247 E Tw, Tx, T2C, T2D;
|
Chris@82
|
248 Tw = ri[WS(is, 2)];
|
Chris@82
|
249 Tx = ri[WS(is, 34)];
|
Chris@82
|
250 Ty = Tw + Tx;
|
Chris@82
|
251 T3H = Tw - Tx;
|
Chris@82
|
252 {
|
Chris@82
|
253 E T2z, T2A, Tz, TA;
|
Chris@82
|
254 T2z = ii[WS(is, 2)];
|
Chris@82
|
255 T2A = ii[WS(is, 34)];
|
Chris@82
|
256 T2B = T2z + T2A;
|
Chris@82
|
257 T3x = T2z - T2A;
|
Chris@82
|
258 Tz = ri[WS(is, 18)];
|
Chris@82
|
259 TA = ri[WS(is, 50)];
|
Chris@82
|
260 TB = Tz + TA;
|
Chris@82
|
261 T3w = Tz - TA;
|
Chris@82
|
262 }
|
Chris@82
|
263 T2C = ii[WS(is, 18)];
|
Chris@82
|
264 T2D = ii[WS(is, 50)];
|
Chris@82
|
265 T2E = T2C + T2D;
|
Chris@82
|
266 T3I = T2C - T2D;
|
Chris@82
|
267 {
|
Chris@82
|
268 E TG, TH, T3C, T2J, T2K, T3D;
|
Chris@82
|
269 TG = ri[WS(is, 58)];
|
Chris@82
|
270 TH = ri[WS(is, 26)];
|
Chris@82
|
271 T3C = TG - TH;
|
Chris@82
|
272 T2J = ii[WS(is, 58)];
|
Chris@82
|
273 T2K = ii[WS(is, 26)];
|
Chris@82
|
274 T3D = T2J - T2K;
|
Chris@82
|
275 TI = TG + TH;
|
Chris@82
|
276 T3K = T3C + T3D;
|
Chris@82
|
277 T2L = T2J + T2K;
|
Chris@82
|
278 T3E = T3C - T3D;
|
Chris@82
|
279 }
|
Chris@82
|
280 {
|
Chris@82
|
281 E TD, TE, T3z, T2G, T2H, T3A;
|
Chris@82
|
282 TD = ri[WS(is, 10)];
|
Chris@82
|
283 TE = ri[WS(is, 42)];
|
Chris@82
|
284 T3z = TD - TE;
|
Chris@82
|
285 T2G = ii[WS(is, 10)];
|
Chris@82
|
286 T2H = ii[WS(is, 42)];
|
Chris@82
|
287 T3A = T2G - T2H;
|
Chris@82
|
288 TF = TD + TE;
|
Chris@82
|
289 T3L = T3A - T3z;
|
Chris@82
|
290 T2I = T2G + T2H;
|
Chris@82
|
291 T3B = T3z + T3A;
|
Chris@82
|
292 }
|
Chris@82
|
293 }
|
Chris@82
|
294 {
|
Chris@82
|
295 E TC, TJ, Taq, Tar;
|
Chris@82
|
296 TC = Ty + TB;
|
Chris@82
|
297 TJ = TF + TI;
|
Chris@82
|
298 TK = TC + TJ;
|
Chris@82
|
299 Tdd = TC - TJ;
|
Chris@82
|
300 Taq = TI - TF;
|
Chris@82
|
301 Tar = T2B - T2E;
|
Chris@82
|
302 Tas = Taq + Tar;
|
Chris@82
|
303 Tce = Tar - Taq;
|
Chris@82
|
304 }
|
Chris@82
|
305 {
|
Chris@82
|
306 E Tat, Tau, T2F, T2M;
|
Chris@82
|
307 Tat = Ty - TB;
|
Chris@82
|
308 Tau = T2I - T2L;
|
Chris@82
|
309 Tav = Tat + Tau;
|
Chris@82
|
310 Tcf = Tat - Tau;
|
Chris@82
|
311 T2F = T2B + T2E;
|
Chris@82
|
312 T2M = T2I + T2L;
|
Chris@82
|
313 T2N = T2F + T2M;
|
Chris@82
|
314 Tdc = T2F - T2M;
|
Chris@82
|
315 }
|
Chris@82
|
316 {
|
Chris@82
|
317 E T3y, T3F, T7M, T7N;
|
Chris@82
|
318 T3y = T3w + T3x;
|
Chris@82
|
319 T3F = T3B - T3E;
|
Chris@82
|
320 T3G = FNMS(KP707106781, T3F, T3y);
|
Chris@82
|
321 T6G = FMA(KP707106781, T3F, T3y);
|
Chris@82
|
322 T7M = T3x - T3w;
|
Chris@82
|
323 T7N = T3L + T3K;
|
Chris@82
|
324 T7O = FMA(KP707106781, T7N, T7M);
|
Chris@82
|
325 T9k = FNMS(KP707106781, T7N, T7M);
|
Chris@82
|
326 }
|
Chris@82
|
327 {
|
Chris@82
|
328 E T7P, T7Q, T3J, T3M;
|
Chris@82
|
329 T7P = T3H + T3I;
|
Chris@82
|
330 T7Q = T3B + T3E;
|
Chris@82
|
331 T7R = FMA(KP707106781, T7Q, T7P);
|
Chris@82
|
332 T9l = FNMS(KP707106781, T7Q, T7P);
|
Chris@82
|
333 T3J = T3H - T3I;
|
Chris@82
|
334 T3M = T3K - T3L;
|
Chris@82
|
335 T3N = FNMS(KP707106781, T3M, T3J);
|
Chris@82
|
336 T6H = FMA(KP707106781, T3M, T3J);
|
Chris@82
|
337 }
|
Chris@82
|
338 }
|
Chris@82
|
339 {
|
Chris@82
|
340 E T1z, T5I, T56, Tb8, T1C, T53, T5L, Tb9, T1J, Tbq, T5h, T5N, T1G, Tbp, T5c;
|
Chris@82
|
341 E T5O;
|
Chris@82
|
342 {
|
Chris@82
|
343 E T1x, T1y, T5J, T5K;
|
Chris@82
|
344 T1x = ri[WS(is, 63)];
|
Chris@82
|
345 T1y = ri[WS(is, 31)];
|
Chris@82
|
346 T1z = T1x + T1y;
|
Chris@82
|
347 T5I = T1x - T1y;
|
Chris@82
|
348 {
|
Chris@82
|
349 E T54, T55, T1A, T1B;
|
Chris@82
|
350 T54 = ii[WS(is, 63)];
|
Chris@82
|
351 T55 = ii[WS(is, 31)];
|
Chris@82
|
352 T56 = T54 - T55;
|
Chris@82
|
353 Tb8 = T54 + T55;
|
Chris@82
|
354 T1A = ri[WS(is, 15)];
|
Chris@82
|
355 T1B = ri[WS(is, 47)];
|
Chris@82
|
356 T1C = T1A + T1B;
|
Chris@82
|
357 T53 = T1A - T1B;
|
Chris@82
|
358 }
|
Chris@82
|
359 T5J = ii[WS(is, 15)];
|
Chris@82
|
360 T5K = ii[WS(is, 47)];
|
Chris@82
|
361 T5L = T5J - T5K;
|
Chris@82
|
362 Tb9 = T5J + T5K;
|
Chris@82
|
363 {
|
Chris@82
|
364 E T1H, T1I, T5d, T5e, T5f, T5g;
|
Chris@82
|
365 T1H = ri[WS(is, 55)];
|
Chris@82
|
366 T1I = ri[WS(is, 23)];
|
Chris@82
|
367 T5d = T1H - T1I;
|
Chris@82
|
368 T5e = ii[WS(is, 55)];
|
Chris@82
|
369 T5f = ii[WS(is, 23)];
|
Chris@82
|
370 T5g = T5e - T5f;
|
Chris@82
|
371 T1J = T1H + T1I;
|
Chris@82
|
372 Tbq = T5e + T5f;
|
Chris@82
|
373 T5h = T5d - T5g;
|
Chris@82
|
374 T5N = T5d + T5g;
|
Chris@82
|
375 }
|
Chris@82
|
376 {
|
Chris@82
|
377 E T1E, T1F, T58, T59, T5a, T5b;
|
Chris@82
|
378 T1E = ri[WS(is, 7)];
|
Chris@82
|
379 T1F = ri[WS(is, 39)];
|
Chris@82
|
380 T58 = T1E - T1F;
|
Chris@82
|
381 T59 = ii[WS(is, 7)];
|
Chris@82
|
382 T5a = ii[WS(is, 39)];
|
Chris@82
|
383 T5b = T59 - T5a;
|
Chris@82
|
384 T1G = T1E + T1F;
|
Chris@82
|
385 Tbp = T59 + T5a;
|
Chris@82
|
386 T5c = T58 + T5b;
|
Chris@82
|
387 T5O = T5b - T58;
|
Chris@82
|
388 }
|
Chris@82
|
389 }
|
Chris@82
|
390 {
|
Chris@82
|
391 E T1D, T1K, Tbo, Tbr;
|
Chris@82
|
392 T1D = T1z + T1C;
|
Chris@82
|
393 T1K = T1G + T1J;
|
Chris@82
|
394 T1L = T1D + T1K;
|
Chris@82
|
395 TdA = T1D - T1K;
|
Chris@82
|
396 Tbo = T1z - T1C;
|
Chris@82
|
397 Tbr = Tbp - Tbq;
|
Chris@82
|
398 Tbs = Tbo + Tbr;
|
Chris@82
|
399 Tct = Tbo - Tbr;
|
Chris@82
|
400 }
|
Chris@82
|
401 {
|
Chris@82
|
402 E Tdv, Tdw, T57, T5i;
|
Chris@82
|
403 Tdv = Tb8 + Tb9;
|
Chris@82
|
404 Tdw = Tbp + Tbq;
|
Chris@82
|
405 Tdx = Tdv - Tdw;
|
Chris@82
|
406 Teo = Tdv + Tdw;
|
Chris@82
|
407 T57 = T53 + T56;
|
Chris@82
|
408 T5i = T5c - T5h;
|
Chris@82
|
409 T5j = FNMS(KP707106781, T5i, T57);
|
Chris@82
|
410 T6Y = FMA(KP707106781, T5i, T57);
|
Chris@82
|
411 }
|
Chris@82
|
412 {
|
Chris@82
|
413 E T5M, T5P, T8w, T8x;
|
Chris@82
|
414 T5M = T5I - T5L;
|
Chris@82
|
415 T5P = T5N - T5O;
|
Chris@82
|
416 T5Q = FNMS(KP707106781, T5P, T5M);
|
Chris@82
|
417 T6V = FMA(KP707106781, T5P, T5M);
|
Chris@82
|
418 T8w = T5I + T5L;
|
Chris@82
|
419 T8x = T5c + T5h;
|
Chris@82
|
420 T8y = FMA(KP707106781, T8x, T8w);
|
Chris@82
|
421 T9z = FNMS(KP707106781, T8x, T8w);
|
Chris@82
|
422 }
|
Chris@82
|
423 {
|
Chris@82
|
424 E Tb7, Tba, T8l, T8m;
|
Chris@82
|
425 Tb7 = T1J - T1G;
|
Chris@82
|
426 Tba = Tb8 - Tb9;
|
Chris@82
|
427 Tbb = Tb7 + Tba;
|
Chris@82
|
428 Tcw = Tba - Tb7;
|
Chris@82
|
429 T8l = T56 - T53;
|
Chris@82
|
430 T8m = T5O + T5N;
|
Chris@82
|
431 T8n = FMA(KP707106781, T8m, T8l);
|
Chris@82
|
432 T9C = FNMS(KP707106781, T8m, T8l);
|
Chris@82
|
433 }
|
Chris@82
|
434 }
|
Chris@82
|
435 {
|
Chris@82
|
436 E TN, T40, T2Q, T3Q, TQ, T3P, T2T, T41, TX, T43, T30, T3X, TU, T44, T2X;
|
Chris@82
|
437 E T3U;
|
Chris@82
|
438 {
|
Chris@82
|
439 E TL, TM, T2R, T2S;
|
Chris@82
|
440 TL = ri[WS(is, 62)];
|
Chris@82
|
441 TM = ri[WS(is, 30)];
|
Chris@82
|
442 TN = TL + TM;
|
Chris@82
|
443 T40 = TL - TM;
|
Chris@82
|
444 {
|
Chris@82
|
445 E T2O, T2P, TO, TP;
|
Chris@82
|
446 T2O = ii[WS(is, 62)];
|
Chris@82
|
447 T2P = ii[WS(is, 30)];
|
Chris@82
|
448 T2Q = T2O + T2P;
|
Chris@82
|
449 T3Q = T2O - T2P;
|
Chris@82
|
450 TO = ri[WS(is, 14)];
|
Chris@82
|
451 TP = ri[WS(is, 46)];
|
Chris@82
|
452 TQ = TO + TP;
|
Chris@82
|
453 T3P = TO - TP;
|
Chris@82
|
454 }
|
Chris@82
|
455 T2R = ii[WS(is, 14)];
|
Chris@82
|
456 T2S = ii[WS(is, 46)];
|
Chris@82
|
457 T2T = T2R + T2S;
|
Chris@82
|
458 T41 = T2R - T2S;
|
Chris@82
|
459 {
|
Chris@82
|
460 E TV, TW, T3V, T2Y, T2Z, T3W;
|
Chris@82
|
461 TV = ri[WS(is, 54)];
|
Chris@82
|
462 TW = ri[WS(is, 22)];
|
Chris@82
|
463 T3V = TV - TW;
|
Chris@82
|
464 T2Y = ii[WS(is, 54)];
|
Chris@82
|
465 T2Z = ii[WS(is, 22)];
|
Chris@82
|
466 T3W = T2Y - T2Z;
|
Chris@82
|
467 TX = TV + TW;
|
Chris@82
|
468 T43 = T3V + T3W;
|
Chris@82
|
469 T30 = T2Y + T2Z;
|
Chris@82
|
470 T3X = T3V - T3W;
|
Chris@82
|
471 }
|
Chris@82
|
472 {
|
Chris@82
|
473 E TS, TT, T3S, T2V, T2W, T3T;
|
Chris@82
|
474 TS = ri[WS(is, 6)];
|
Chris@82
|
475 TT = ri[WS(is, 38)];
|
Chris@82
|
476 T3S = TS - TT;
|
Chris@82
|
477 T2V = ii[WS(is, 6)];
|
Chris@82
|
478 T2W = ii[WS(is, 38)];
|
Chris@82
|
479 T3T = T2V - T2W;
|
Chris@82
|
480 TU = TS + TT;
|
Chris@82
|
481 T44 = T3T - T3S;
|
Chris@82
|
482 T2X = T2V + T2W;
|
Chris@82
|
483 T3U = T3S + T3T;
|
Chris@82
|
484 }
|
Chris@82
|
485 }
|
Chris@82
|
486 {
|
Chris@82
|
487 E TR, TY, Tax, Tay;
|
Chris@82
|
488 TR = TN + TQ;
|
Chris@82
|
489 TY = TU + TX;
|
Chris@82
|
490 TZ = TR + TY;
|
Chris@82
|
491 Tdf = TR - TY;
|
Chris@82
|
492 Tax = TX - TU;
|
Chris@82
|
493 Tay = T2Q - T2T;
|
Chris@82
|
494 Taz = Tax + Tay;
|
Chris@82
|
495 Tch = Tay - Tax;
|
Chris@82
|
496 }
|
Chris@82
|
497 {
|
Chris@82
|
498 E TaA, TaB, T2U, T31;
|
Chris@82
|
499 TaA = TN - TQ;
|
Chris@82
|
500 TaB = T2X - T30;
|
Chris@82
|
501 TaC = TaA + TaB;
|
Chris@82
|
502 Tci = TaA - TaB;
|
Chris@82
|
503 T2U = T2Q + T2T;
|
Chris@82
|
504 T31 = T2X + T30;
|
Chris@82
|
505 T32 = T2U + T31;
|
Chris@82
|
506 Tdg = T2U - T31;
|
Chris@82
|
507 }
|
Chris@82
|
508 {
|
Chris@82
|
509 E T3R, T3Y, T7T, T7U;
|
Chris@82
|
510 T3R = T3P + T3Q;
|
Chris@82
|
511 T3Y = T3U - T3X;
|
Chris@82
|
512 T3Z = FNMS(KP707106781, T3Y, T3R);
|
Chris@82
|
513 T6J = FMA(KP707106781, T3Y, T3R);
|
Chris@82
|
514 T7T = T3Q - T3P;
|
Chris@82
|
515 T7U = T44 + T43;
|
Chris@82
|
516 T7V = FMA(KP707106781, T7U, T7T);
|
Chris@82
|
517 T9n = FNMS(KP707106781, T7U, T7T);
|
Chris@82
|
518 }
|
Chris@82
|
519 {
|
Chris@82
|
520 E T7W, T7X, T42, T45;
|
Chris@82
|
521 T7W = T40 + T41;
|
Chris@82
|
522 T7X = T3U + T3X;
|
Chris@82
|
523 T7Y = FMA(KP707106781, T7X, T7W);
|
Chris@82
|
524 T9o = FNMS(KP707106781, T7X, T7W);
|
Chris@82
|
525 T42 = T40 - T41;
|
Chris@82
|
526 T45 = T43 - T44;
|
Chris@82
|
527 T46 = FNMS(KP707106781, T45, T42);
|
Chris@82
|
528 T6K = FMA(KP707106781, T45, T42);
|
Chris@82
|
529 }
|
Chris@82
|
530 }
|
Chris@82
|
531 {
|
Chris@82
|
532 E T14, T4P, T4d, TaH, T17, T4a, T4S, TaI, T1e, TaZ, T4o, T4U, T1b, TaY, T4j;
|
Chris@82
|
533 E T4V;
|
Chris@82
|
534 {
|
Chris@82
|
535 E T12, T13, T4Q, T4R;
|
Chris@82
|
536 T12 = ri[WS(is, 1)];
|
Chris@82
|
537 T13 = ri[WS(is, 33)];
|
Chris@82
|
538 T14 = T12 + T13;
|
Chris@82
|
539 T4P = T12 - T13;
|
Chris@82
|
540 {
|
Chris@82
|
541 E T4b, T4c, T15, T16;
|
Chris@82
|
542 T4b = ii[WS(is, 1)];
|
Chris@82
|
543 T4c = ii[WS(is, 33)];
|
Chris@82
|
544 T4d = T4b - T4c;
|
Chris@82
|
545 TaH = T4b + T4c;
|
Chris@82
|
546 T15 = ri[WS(is, 17)];
|
Chris@82
|
547 T16 = ri[WS(is, 49)];
|
Chris@82
|
548 T17 = T15 + T16;
|
Chris@82
|
549 T4a = T15 - T16;
|
Chris@82
|
550 }
|
Chris@82
|
551 T4Q = ii[WS(is, 17)];
|
Chris@82
|
552 T4R = ii[WS(is, 49)];
|
Chris@82
|
553 T4S = T4Q - T4R;
|
Chris@82
|
554 TaI = T4Q + T4R;
|
Chris@82
|
555 {
|
Chris@82
|
556 E T1c, T1d, T4k, T4l, T4m, T4n;
|
Chris@82
|
557 T1c = ri[WS(is, 57)];
|
Chris@82
|
558 T1d = ri[WS(is, 25)];
|
Chris@82
|
559 T4k = T1c - T1d;
|
Chris@82
|
560 T4l = ii[WS(is, 57)];
|
Chris@82
|
561 T4m = ii[WS(is, 25)];
|
Chris@82
|
562 T4n = T4l - T4m;
|
Chris@82
|
563 T1e = T1c + T1d;
|
Chris@82
|
564 TaZ = T4l + T4m;
|
Chris@82
|
565 T4o = T4k - T4n;
|
Chris@82
|
566 T4U = T4k + T4n;
|
Chris@82
|
567 }
|
Chris@82
|
568 {
|
Chris@82
|
569 E T19, T1a, T4f, T4g, T4h, T4i;
|
Chris@82
|
570 T19 = ri[WS(is, 9)];
|
Chris@82
|
571 T1a = ri[WS(is, 41)];
|
Chris@82
|
572 T4f = T19 - T1a;
|
Chris@82
|
573 T4g = ii[WS(is, 9)];
|
Chris@82
|
574 T4h = ii[WS(is, 41)];
|
Chris@82
|
575 T4i = T4g - T4h;
|
Chris@82
|
576 T1b = T19 + T1a;
|
Chris@82
|
577 TaY = T4g + T4h;
|
Chris@82
|
578 T4j = T4f + T4i;
|
Chris@82
|
579 T4V = T4i - T4f;
|
Chris@82
|
580 }
|
Chris@82
|
581 }
|
Chris@82
|
582 {
|
Chris@82
|
583 E T18, T1f, TaX, Tb0;
|
Chris@82
|
584 T18 = T14 + T17;
|
Chris@82
|
585 T1f = T1b + T1e;
|
Chris@82
|
586 T1g = T18 + T1f;
|
Chris@82
|
587 Tdp = T18 - T1f;
|
Chris@82
|
588 TaX = T14 - T17;
|
Chris@82
|
589 Tb0 = TaY - TaZ;
|
Chris@82
|
590 Tb1 = TaX + Tb0;
|
Chris@82
|
591 Tcm = TaX - Tb0;
|
Chris@82
|
592 }
|
Chris@82
|
593 {
|
Chris@82
|
594 E Tdk, Tdl, T4e, T4p;
|
Chris@82
|
595 Tdk = TaH + TaI;
|
Chris@82
|
596 Tdl = TaY + TaZ;
|
Chris@82
|
597 Tdm = Tdk - Tdl;
|
Chris@82
|
598 Tej = Tdk + Tdl;
|
Chris@82
|
599 T4e = T4a + T4d;
|
Chris@82
|
600 T4p = T4j - T4o;
|
Chris@82
|
601 T4q = FNMS(KP707106781, T4p, T4e);
|
Chris@82
|
602 T6R = FMA(KP707106781, T4p, T4e);
|
Chris@82
|
603 }
|
Chris@82
|
604 {
|
Chris@82
|
605 E T4T, T4W, T8d, T8e;
|
Chris@82
|
606 T4T = T4P - T4S;
|
Chris@82
|
607 T4W = T4U - T4V;
|
Chris@82
|
608 T4X = FNMS(KP707106781, T4W, T4T);
|
Chris@82
|
609 T6O = FMA(KP707106781, T4W, T4T);
|
Chris@82
|
610 T8d = T4P + T4S;
|
Chris@82
|
611 T8e = T4j + T4o;
|
Chris@82
|
612 T8f = FMA(KP707106781, T8e, T8d);
|
Chris@82
|
613 T9s = FNMS(KP707106781, T8e, T8d);
|
Chris@82
|
614 }
|
Chris@82
|
615 {
|
Chris@82
|
616 E TaG, TaJ, T82, T83;
|
Chris@82
|
617 TaG = T1e - T1b;
|
Chris@82
|
618 TaJ = TaH - TaI;
|
Chris@82
|
619 TaK = TaG + TaJ;
|
Chris@82
|
620 Tcp = TaJ - TaG;
|
Chris@82
|
621 T82 = T4d - T4a;
|
Chris@82
|
622 T83 = T4V + T4U;
|
Chris@82
|
623 T84 = FMA(KP707106781, T83, T82);
|
Chris@82
|
624 T9v = FNMS(KP707106781, T83, T82);
|
Chris@82
|
625 }
|
Chris@82
|
626 }
|
Chris@82
|
627 {
|
Chris@82
|
628 E T1j, TaL, T1m, TaM, T4G, T4L, TaO, TaN, T86, T85, T1q, TaR, T1t, TaS, T4v;
|
Chris@82
|
629 E T4A, TaT, TaQ, T89, T88;
|
Chris@82
|
630 {
|
Chris@82
|
631 E T4C, T4K, T4H, T4F;
|
Chris@82
|
632 {
|
Chris@82
|
633 E T1h, T1i, T4I, T4J;
|
Chris@82
|
634 T1h = ri[WS(is, 5)];
|
Chris@82
|
635 T1i = ri[WS(is, 37)];
|
Chris@82
|
636 T1j = T1h + T1i;
|
Chris@82
|
637 T4C = T1h - T1i;
|
Chris@82
|
638 T4I = ii[WS(is, 5)];
|
Chris@82
|
639 T4J = ii[WS(is, 37)];
|
Chris@82
|
640 T4K = T4I - T4J;
|
Chris@82
|
641 TaL = T4I + T4J;
|
Chris@82
|
642 }
|
Chris@82
|
643 {
|
Chris@82
|
644 E T1k, T1l, T4D, T4E;
|
Chris@82
|
645 T1k = ri[WS(is, 21)];
|
Chris@82
|
646 T1l = ri[WS(is, 53)];
|
Chris@82
|
647 T1m = T1k + T1l;
|
Chris@82
|
648 T4H = T1k - T1l;
|
Chris@82
|
649 T4D = ii[WS(is, 21)];
|
Chris@82
|
650 T4E = ii[WS(is, 53)];
|
Chris@82
|
651 T4F = T4D - T4E;
|
Chris@82
|
652 TaM = T4D + T4E;
|
Chris@82
|
653 }
|
Chris@82
|
654 T4G = T4C - T4F;
|
Chris@82
|
655 T4L = T4H + T4K;
|
Chris@82
|
656 TaO = T1j - T1m;
|
Chris@82
|
657 TaN = TaL - TaM;
|
Chris@82
|
658 T86 = T4C + T4F;
|
Chris@82
|
659 T85 = T4K - T4H;
|
Chris@82
|
660 }
|
Chris@82
|
661 {
|
Chris@82
|
662 E T4r, T4z, T4w, T4u;
|
Chris@82
|
663 {
|
Chris@82
|
664 E T1o, T1p, T4x, T4y;
|
Chris@82
|
665 T1o = ri[WS(is, 61)];
|
Chris@82
|
666 T1p = ri[WS(is, 29)];
|
Chris@82
|
667 T1q = T1o + T1p;
|
Chris@82
|
668 T4r = T1o - T1p;
|
Chris@82
|
669 T4x = ii[WS(is, 61)];
|
Chris@82
|
670 T4y = ii[WS(is, 29)];
|
Chris@82
|
671 T4z = T4x - T4y;
|
Chris@82
|
672 TaR = T4x + T4y;
|
Chris@82
|
673 }
|
Chris@82
|
674 {
|
Chris@82
|
675 E T1r, T1s, T4s, T4t;
|
Chris@82
|
676 T1r = ri[WS(is, 13)];
|
Chris@82
|
677 T1s = ri[WS(is, 45)];
|
Chris@82
|
678 T1t = T1r + T1s;
|
Chris@82
|
679 T4w = T1r - T1s;
|
Chris@82
|
680 T4s = ii[WS(is, 13)];
|
Chris@82
|
681 T4t = ii[WS(is, 45)];
|
Chris@82
|
682 T4u = T4s - T4t;
|
Chris@82
|
683 TaS = T4s + T4t;
|
Chris@82
|
684 }
|
Chris@82
|
685 T4v = T4r - T4u;
|
Chris@82
|
686 T4A = T4w + T4z;
|
Chris@82
|
687 TaT = TaR - TaS;
|
Chris@82
|
688 TaQ = T1q - T1t;
|
Chris@82
|
689 T89 = T4r + T4u;
|
Chris@82
|
690 T88 = T4z - T4w;
|
Chris@82
|
691 }
|
Chris@82
|
692 {
|
Chris@82
|
693 E T1n, T1u, Tb2, Tb3;
|
Chris@82
|
694 T1n = T1j + T1m;
|
Chris@82
|
695 T1u = T1q + T1t;
|
Chris@82
|
696 T1v = T1n + T1u;
|
Chris@82
|
697 Tdn = T1u - T1n;
|
Chris@82
|
698 Tb2 = TaO + TaN;
|
Chris@82
|
699 Tb3 = TaQ - TaT;
|
Chris@82
|
700 Tb4 = Tb2 + Tb3;
|
Chris@82
|
701 Tcq = Tb2 - Tb3;
|
Chris@82
|
702 }
|
Chris@82
|
703 {
|
Chris@82
|
704 E Tdq, Tdr, T4B, T4M;
|
Chris@82
|
705 Tdq = TaL + TaM;
|
Chris@82
|
706 Tdr = TaR + TaS;
|
Chris@82
|
707 Tds = Tdq - Tdr;
|
Chris@82
|
708 Tek = Tdq + Tdr;
|
Chris@82
|
709 T4B = FMA(KP414213562, T4A, T4v);
|
Chris@82
|
710 T4M = FNMS(KP414213562, T4L, T4G);
|
Chris@82
|
711 T4N = T4B - T4M;
|
Chris@82
|
712 T6P = T4M + T4B;
|
Chris@82
|
713 }
|
Chris@82
|
714 {
|
Chris@82
|
715 E T4Y, T4Z, T8g, T8h;
|
Chris@82
|
716 T4Y = FMA(KP414213562, T4G, T4L);
|
Chris@82
|
717 T4Z = FNMS(KP414213562, T4v, T4A);
|
Chris@82
|
718 T50 = T4Y - T4Z;
|
Chris@82
|
719 T6S = T4Y + T4Z;
|
Chris@82
|
720 T8g = FMA(KP414213562, T85, T86);
|
Chris@82
|
721 T8h = FNMS(KP414213562, T88, T89);
|
Chris@82
|
722 T8i = T8g + T8h;
|
Chris@82
|
723 T9w = T8g - T8h;
|
Chris@82
|
724 }
|
Chris@82
|
725 {
|
Chris@82
|
726 E TaP, TaU, T87, T8a;
|
Chris@82
|
727 TaP = TaN - TaO;
|
Chris@82
|
728 TaU = TaQ + TaT;
|
Chris@82
|
729 TaV = TaP + TaU;
|
Chris@82
|
730 Tcn = TaU - TaP;
|
Chris@82
|
731 T87 = FNMS(KP414213562, T86, T85);
|
Chris@82
|
732 T8a = FMA(KP414213562, T89, T88);
|
Chris@82
|
733 T8b = T87 + T8a;
|
Chris@82
|
734 T9t = T8a - T87;
|
Chris@82
|
735 }
|
Chris@82
|
736 }
|
Chris@82
|
737 {
|
Chris@82
|
738 E T1O, Tbc, T1R, Tbd, T5z, T5E, Tbf, Tbe, T8p, T8o, T1V, Tbi, T1Y, Tbj, T5o;
|
Chris@82
|
739 E T5t, Tbk, Tbh, T8s, T8r;
|
Chris@82
|
740 {
|
Chris@82
|
741 E T5v, T5D, T5A, T5y;
|
Chris@82
|
742 {
|
Chris@82
|
743 E T1M, T1N, T5B, T5C;
|
Chris@82
|
744 T1M = ri[WS(is, 3)];
|
Chris@82
|
745 T1N = ri[WS(is, 35)];
|
Chris@82
|
746 T1O = T1M + T1N;
|
Chris@82
|
747 T5v = T1M - T1N;
|
Chris@82
|
748 T5B = ii[WS(is, 3)];
|
Chris@82
|
749 T5C = ii[WS(is, 35)];
|
Chris@82
|
750 T5D = T5B - T5C;
|
Chris@82
|
751 Tbc = T5B + T5C;
|
Chris@82
|
752 }
|
Chris@82
|
753 {
|
Chris@82
|
754 E T1P, T1Q, T5w, T5x;
|
Chris@82
|
755 T1P = ri[WS(is, 19)];
|
Chris@82
|
756 T1Q = ri[WS(is, 51)];
|
Chris@82
|
757 T1R = T1P + T1Q;
|
Chris@82
|
758 T5A = T1P - T1Q;
|
Chris@82
|
759 T5w = ii[WS(is, 19)];
|
Chris@82
|
760 T5x = ii[WS(is, 51)];
|
Chris@82
|
761 T5y = T5w - T5x;
|
Chris@82
|
762 Tbd = T5w + T5x;
|
Chris@82
|
763 }
|
Chris@82
|
764 T5z = T5v - T5y;
|
Chris@82
|
765 T5E = T5A + T5D;
|
Chris@82
|
766 Tbf = T1O - T1R;
|
Chris@82
|
767 Tbe = Tbc - Tbd;
|
Chris@82
|
768 T8p = T5v + T5y;
|
Chris@82
|
769 T8o = T5D - T5A;
|
Chris@82
|
770 }
|
Chris@82
|
771 {
|
Chris@82
|
772 E T5k, T5s, T5p, T5n;
|
Chris@82
|
773 {
|
Chris@82
|
774 E T1T, T1U, T5q, T5r;
|
Chris@82
|
775 T1T = ri[WS(is, 59)];
|
Chris@82
|
776 T1U = ri[WS(is, 27)];
|
Chris@82
|
777 T1V = T1T + T1U;
|
Chris@82
|
778 T5k = T1T - T1U;
|
Chris@82
|
779 T5q = ii[WS(is, 59)];
|
Chris@82
|
780 T5r = ii[WS(is, 27)];
|
Chris@82
|
781 T5s = T5q - T5r;
|
Chris@82
|
782 Tbi = T5q + T5r;
|
Chris@82
|
783 }
|
Chris@82
|
784 {
|
Chris@82
|
785 E T1W, T1X, T5l, T5m;
|
Chris@82
|
786 T1W = ri[WS(is, 11)];
|
Chris@82
|
787 T1X = ri[WS(is, 43)];
|
Chris@82
|
788 T1Y = T1W + T1X;
|
Chris@82
|
789 T5p = T1W - T1X;
|
Chris@82
|
790 T5l = ii[WS(is, 11)];
|
Chris@82
|
791 T5m = ii[WS(is, 43)];
|
Chris@82
|
792 T5n = T5l - T5m;
|
Chris@82
|
793 Tbj = T5l + T5m;
|
Chris@82
|
794 }
|
Chris@82
|
795 T5o = T5k - T5n;
|
Chris@82
|
796 T5t = T5p + T5s;
|
Chris@82
|
797 Tbk = Tbi - Tbj;
|
Chris@82
|
798 Tbh = T1V - T1Y;
|
Chris@82
|
799 T8s = T5k + T5n;
|
Chris@82
|
800 T8r = T5s - T5p;
|
Chris@82
|
801 }
|
Chris@82
|
802 {
|
Chris@82
|
803 E T1S, T1Z, Tbt, Tbu;
|
Chris@82
|
804 T1S = T1O + T1R;
|
Chris@82
|
805 T1Z = T1V + T1Y;
|
Chris@82
|
806 T20 = T1S + T1Z;
|
Chris@82
|
807 Tdy = T1Z - T1S;
|
Chris@82
|
808 Tbt = Tbf + Tbe;
|
Chris@82
|
809 Tbu = Tbh - Tbk;
|
Chris@82
|
810 Tbv = Tbt + Tbu;
|
Chris@82
|
811 Tcx = Tbt - Tbu;
|
Chris@82
|
812 }
|
Chris@82
|
813 {
|
Chris@82
|
814 E TdB, TdC, T5u, T5F;
|
Chris@82
|
815 TdB = Tbc + Tbd;
|
Chris@82
|
816 TdC = Tbi + Tbj;
|
Chris@82
|
817 TdD = TdB - TdC;
|
Chris@82
|
818 Tep = TdB + TdC;
|
Chris@82
|
819 T5u = FMA(KP414213562, T5t, T5o);
|
Chris@82
|
820 T5F = FNMS(KP414213562, T5E, T5z);
|
Chris@82
|
821 T5G = T5u - T5F;
|
Chris@82
|
822 T6W = T5F + T5u;
|
Chris@82
|
823 }
|
Chris@82
|
824 {
|
Chris@82
|
825 E T5R, T5S, T8z, T8A;
|
Chris@82
|
826 T5R = FMA(KP414213562, T5z, T5E);
|
Chris@82
|
827 T5S = FNMS(KP414213562, T5o, T5t);
|
Chris@82
|
828 T5T = T5R - T5S;
|
Chris@82
|
829 T6Z = T5R + T5S;
|
Chris@82
|
830 T8z = FMA(KP414213562, T8o, T8p);
|
Chris@82
|
831 T8A = FNMS(KP414213562, T8r, T8s);
|
Chris@82
|
832 T8B = T8z + T8A;
|
Chris@82
|
833 T9D = T8z - T8A;
|
Chris@82
|
834 }
|
Chris@82
|
835 {
|
Chris@82
|
836 E Tbg, Tbl, T8q, T8t;
|
Chris@82
|
837 Tbg = Tbe - Tbf;
|
Chris@82
|
838 Tbl = Tbh + Tbk;
|
Chris@82
|
839 Tbm = Tbg + Tbl;
|
Chris@82
|
840 Tcu = Tbl - Tbg;
|
Chris@82
|
841 T8q = FNMS(KP414213562, T8p, T8o);
|
Chris@82
|
842 T8t = FMA(KP414213562, T8s, T8r);
|
Chris@82
|
843 T8u = T8q + T8t;
|
Chris@82
|
844 T9A = T8t - T8q;
|
Chris@82
|
845 }
|
Chris@82
|
846 }
|
Chris@82
|
847 {
|
Chris@82
|
848 E T11, TeD, TeG, TeI, T22, T23, T34, TeH;
|
Chris@82
|
849 {
|
Chris@82
|
850 E Tv, T10, TeE, TeF;
|
Chris@82
|
851 Tv = Tf + Tu;
|
Chris@82
|
852 T10 = TK + TZ;
|
Chris@82
|
853 T11 = Tv + T10;
|
Chris@82
|
854 TeD = Tv - T10;
|
Chris@82
|
855 TeE = Tej + Tek;
|
Chris@82
|
856 TeF = Teo + Tep;
|
Chris@82
|
857 TeG = TeE - TeF;
|
Chris@82
|
858 TeI = TeE + TeF;
|
Chris@82
|
859 }
|
Chris@82
|
860 {
|
Chris@82
|
861 E T1w, T21, T2y, T33;
|
Chris@82
|
862 T1w = T1g + T1v;
|
Chris@82
|
863 T21 = T1L + T20;
|
Chris@82
|
864 T22 = T1w + T21;
|
Chris@82
|
865 T23 = T21 - T1w;
|
Chris@82
|
866 T2y = T2i + T2x;
|
Chris@82
|
867 T33 = T2N + T32;
|
Chris@82
|
868 T34 = T2y - T33;
|
Chris@82
|
869 TeH = T2y + T33;
|
Chris@82
|
870 }
|
Chris@82
|
871 ro[WS(os, 32)] = T11 - T22;
|
Chris@82
|
872 io[WS(os, 32)] = TeH - TeI;
|
Chris@82
|
873 ro[0] = T11 + T22;
|
Chris@82
|
874 io[0] = TeH + TeI;
|
Chris@82
|
875 io[WS(os, 16)] = T23 + T34;
|
Chris@82
|
876 ro[WS(os, 16)] = TeD + TeG;
|
Chris@82
|
877 io[WS(os, 48)] = T34 - T23;
|
Chris@82
|
878 ro[WS(os, 48)] = TeD - TeG;
|
Chris@82
|
879 }
|
Chris@82
|
880 {
|
Chris@82
|
881 E Teh, Tex, Tev, TeB, Tem, Tey, Ter, Tez;
|
Chris@82
|
882 {
|
Chris@82
|
883 E Tef, Teg, Tet, Teu;
|
Chris@82
|
884 Tef = Tf - Tu;
|
Chris@82
|
885 Teg = T2N - T32;
|
Chris@82
|
886 Teh = Tef + Teg;
|
Chris@82
|
887 Tex = Tef - Teg;
|
Chris@82
|
888 Tet = T2i - T2x;
|
Chris@82
|
889 Teu = TZ - TK;
|
Chris@82
|
890 Tev = Tet - Teu;
|
Chris@82
|
891 TeB = Teu + Tet;
|
Chris@82
|
892 }
|
Chris@82
|
893 {
|
Chris@82
|
894 E Tei, Tel, Ten, Teq;
|
Chris@82
|
895 Tei = T1g - T1v;
|
Chris@82
|
896 Tel = Tej - Tek;
|
Chris@82
|
897 Tem = Tei + Tel;
|
Chris@82
|
898 Tey = Tel - Tei;
|
Chris@82
|
899 Ten = T1L - T20;
|
Chris@82
|
900 Teq = Teo - Tep;
|
Chris@82
|
901 Ter = Ten - Teq;
|
Chris@82
|
902 Tez = Ten + Teq;
|
Chris@82
|
903 }
|
Chris@82
|
904 {
|
Chris@82
|
905 E Tes, TeC, Tew, TeA;
|
Chris@82
|
906 Tes = Tem + Ter;
|
Chris@82
|
907 ro[WS(os, 40)] = FNMS(KP707106781, Tes, Teh);
|
Chris@82
|
908 ro[WS(os, 8)] = FMA(KP707106781, Tes, Teh);
|
Chris@82
|
909 TeC = Tey + Tez;
|
Chris@82
|
910 io[WS(os, 40)] = FNMS(KP707106781, TeC, TeB);
|
Chris@82
|
911 io[WS(os, 8)] = FMA(KP707106781, TeC, TeB);
|
Chris@82
|
912 Tew = Ter - Tem;
|
Chris@82
|
913 io[WS(os, 56)] = FNMS(KP707106781, Tew, Tev);
|
Chris@82
|
914 io[WS(os, 24)] = FMA(KP707106781, Tew, Tev);
|
Chris@82
|
915 TeA = Tey - Tez;
|
Chris@82
|
916 ro[WS(os, 56)] = FNMS(KP707106781, TeA, Tex);
|
Chris@82
|
917 ro[WS(os, 24)] = FMA(KP707106781, TeA, Tex);
|
Chris@82
|
918 }
|
Chris@82
|
919 }
|
Chris@82
|
920 {
|
Chris@82
|
921 E Tdb, TdV, Te5, TdJ, Tdi, Te6, Te3, Teb, TdM, TdW, Tdu, TdR, Te0, Tea, TdF;
|
Chris@82
|
922 E TdQ;
|
Chris@82
|
923 {
|
Chris@82
|
924 E Tde, Tdh, Tdo, Tdt;
|
Chris@82
|
925 Tdb = Td9 - Tda;
|
Chris@82
|
926 TdV = Td9 + Tda;
|
Chris@82
|
927 Te5 = TdI + TdH;
|
Chris@82
|
928 TdJ = TdH - TdI;
|
Chris@82
|
929 Tde = Tdc - Tdd;
|
Chris@82
|
930 Tdh = Tdf + Tdg;
|
Chris@82
|
931 Tdi = Tde - Tdh;
|
Chris@82
|
932 Te6 = Tde + Tdh;
|
Chris@82
|
933 {
|
Chris@82
|
934 E Te1, Te2, TdK, TdL;
|
Chris@82
|
935 Te1 = TdA + TdD;
|
Chris@82
|
936 Te2 = Tdy + Tdx;
|
Chris@82
|
937 Te3 = FNMS(KP414213562, Te2, Te1);
|
Chris@82
|
938 Teb = FMA(KP414213562, Te1, Te2);
|
Chris@82
|
939 TdK = Tdf - Tdg;
|
Chris@82
|
940 TdL = Tdd + Tdc;
|
Chris@82
|
941 TdM = TdK - TdL;
|
Chris@82
|
942 TdW = TdL + TdK;
|
Chris@82
|
943 }
|
Chris@82
|
944 Tdo = Tdm - Tdn;
|
Chris@82
|
945 Tdt = Tdp - Tds;
|
Chris@82
|
946 Tdu = FMA(KP414213562, Tdt, Tdo);
|
Chris@82
|
947 TdR = FNMS(KP414213562, Tdo, Tdt);
|
Chris@82
|
948 {
|
Chris@82
|
949 E TdY, TdZ, Tdz, TdE;
|
Chris@82
|
950 TdY = Tdp + Tds;
|
Chris@82
|
951 TdZ = Tdn + Tdm;
|
Chris@82
|
952 Te0 = FMA(KP414213562, TdZ, TdY);
|
Chris@82
|
953 Tea = FNMS(KP414213562, TdY, TdZ);
|
Chris@82
|
954 Tdz = Tdx - Tdy;
|
Chris@82
|
955 TdE = TdA - TdD;
|
Chris@82
|
956 TdF = FNMS(KP414213562, TdE, Tdz);
|
Chris@82
|
957 TdQ = FMA(KP414213562, Tdz, TdE);
|
Chris@82
|
958 }
|
Chris@82
|
959 }
|
Chris@82
|
960 {
|
Chris@82
|
961 E Tdj, TdG, TdP, TdS;
|
Chris@82
|
962 Tdj = FMA(KP707106781, Tdi, Tdb);
|
Chris@82
|
963 TdG = Tdu - TdF;
|
Chris@82
|
964 ro[WS(os, 44)] = FNMS(KP923879532, TdG, Tdj);
|
Chris@82
|
965 ro[WS(os, 12)] = FMA(KP923879532, TdG, Tdj);
|
Chris@82
|
966 TdP = FMA(KP707106781, TdM, TdJ);
|
Chris@82
|
967 TdS = TdQ - TdR;
|
Chris@82
|
968 io[WS(os, 44)] = FNMS(KP923879532, TdS, TdP);
|
Chris@82
|
969 io[WS(os, 12)] = FMA(KP923879532, TdS, TdP);
|
Chris@82
|
970 }
|
Chris@82
|
971 {
|
Chris@82
|
972 E TdN, TdO, TdT, TdU;
|
Chris@82
|
973 TdN = FNMS(KP707106781, TdM, TdJ);
|
Chris@82
|
974 TdO = Tdu + TdF;
|
Chris@82
|
975 io[WS(os, 28)] = FNMS(KP923879532, TdO, TdN);
|
Chris@82
|
976 io[WS(os, 60)] = FMA(KP923879532, TdO, TdN);
|
Chris@82
|
977 TdT = FNMS(KP707106781, Tdi, Tdb);
|
Chris@82
|
978 TdU = TdR + TdQ;
|
Chris@82
|
979 ro[WS(os, 28)] = FNMS(KP923879532, TdU, TdT);
|
Chris@82
|
980 ro[WS(os, 60)] = FMA(KP923879532, TdU, TdT);
|
Chris@82
|
981 }
|
Chris@82
|
982 {
|
Chris@82
|
983 E TdX, Te4, Ted, Tee;
|
Chris@82
|
984 TdX = FMA(KP707106781, TdW, TdV);
|
Chris@82
|
985 Te4 = Te0 + Te3;
|
Chris@82
|
986 ro[WS(os, 36)] = FNMS(KP923879532, Te4, TdX);
|
Chris@82
|
987 ro[WS(os, 4)] = FMA(KP923879532, Te4, TdX);
|
Chris@82
|
988 Ted = FMA(KP707106781, Te6, Te5);
|
Chris@82
|
989 Tee = Tea + Teb;
|
Chris@82
|
990 io[WS(os, 36)] = FNMS(KP923879532, Tee, Ted);
|
Chris@82
|
991 io[WS(os, 4)] = FMA(KP923879532, Tee, Ted);
|
Chris@82
|
992 }
|
Chris@82
|
993 {
|
Chris@82
|
994 E Te7, Te8, Te9, Tec;
|
Chris@82
|
995 Te7 = FNMS(KP707106781, Te6, Te5);
|
Chris@82
|
996 Te8 = Te3 - Te0;
|
Chris@82
|
997 io[WS(os, 52)] = FNMS(KP923879532, Te8, Te7);
|
Chris@82
|
998 io[WS(os, 20)] = FMA(KP923879532, Te8, Te7);
|
Chris@82
|
999 Te9 = FNMS(KP707106781, TdW, TdV);
|
Chris@82
|
1000 Tec = Tea - Teb;
|
Chris@82
|
1001 ro[WS(os, 52)] = FNMS(KP923879532, Tec, Te9);
|
Chris@82
|
1002 ro[WS(os, 20)] = FMA(KP923879532, Tec, Te9);
|
Chris@82
|
1003 }
|
Chris@82
|
1004 }
|
Chris@82
|
1005 {
|
Chris@82
|
1006 E Tcd, TcP, TcD, TcZ, Tck, Td0, TcX, Td4, Tcs, TcK, TcG, TcQ, TcU, Td5, Tcz;
|
Chris@82
|
1007 E TcL, Tcc, TcC;
|
Chris@82
|
1008 Tcc = TbC - TbD;
|
Chris@82
|
1009 Tcd = FMA(KP707106781, Tcc, Tcb);
|
Chris@82
|
1010 TcP = FNMS(KP707106781, Tcc, Tcb);
|
Chris@82
|
1011 TcC = Tan - Tak;
|
Chris@82
|
1012 TcD = FMA(KP707106781, TcC, TcB);
|
Chris@82
|
1013 TcZ = FNMS(KP707106781, TcC, TcB);
|
Chris@82
|
1014 {
|
Chris@82
|
1015 E Tcg, Tcj, TcV, TcW;
|
Chris@82
|
1016 Tcg = FMA(KP414213562, Tcf, Tce);
|
Chris@82
|
1017 Tcj = FNMS(KP414213562, Tci, Tch);
|
Chris@82
|
1018 Tck = Tcg - Tcj;
|
Chris@82
|
1019 Td0 = Tcg + Tcj;
|
Chris@82
|
1020 TcV = FMA(KP707106781, Tcx, Tcw);
|
Chris@82
|
1021 TcW = FMA(KP707106781, Tcu, Tct);
|
Chris@82
|
1022 TcX = FNMS(KP198912367, TcW, TcV);
|
Chris@82
|
1023 Td4 = FMA(KP198912367, TcV, TcW);
|
Chris@82
|
1024 }
|
Chris@82
|
1025 {
|
Chris@82
|
1026 E Tco, Tcr, TcE, TcF;
|
Chris@82
|
1027 Tco = FNMS(KP707106781, Tcn, Tcm);
|
Chris@82
|
1028 Tcr = FNMS(KP707106781, Tcq, Tcp);
|
Chris@82
|
1029 Tcs = FMA(KP668178637, Tcr, Tco);
|
Chris@82
|
1030 TcK = FNMS(KP668178637, Tco, Tcr);
|
Chris@82
|
1031 TcE = FMA(KP414213562, Tch, Tci);
|
Chris@82
|
1032 TcF = FNMS(KP414213562, Tce, Tcf);
|
Chris@82
|
1033 TcG = TcE - TcF;
|
Chris@82
|
1034 TcQ = TcF + TcE;
|
Chris@82
|
1035 }
|
Chris@82
|
1036 {
|
Chris@82
|
1037 E TcS, TcT, Tcv, Tcy;
|
Chris@82
|
1038 TcS = FMA(KP707106781, Tcq, Tcp);
|
Chris@82
|
1039 TcT = FMA(KP707106781, Tcn, Tcm);
|
Chris@82
|
1040 TcU = FMA(KP198912367, TcT, TcS);
|
Chris@82
|
1041 Td5 = FNMS(KP198912367, TcS, TcT);
|
Chris@82
|
1042 Tcv = FNMS(KP707106781, Tcu, Tct);
|
Chris@82
|
1043 Tcy = FNMS(KP707106781, Tcx, Tcw);
|
Chris@82
|
1044 Tcz = FNMS(KP668178637, Tcy, Tcv);
|
Chris@82
|
1045 TcL = FMA(KP668178637, Tcv, Tcy);
|
Chris@82
|
1046 }
|
Chris@82
|
1047 {
|
Chris@82
|
1048 E Tcl, TcA, TcN, TcO;
|
Chris@82
|
1049 Tcl = FMA(KP923879532, Tck, Tcd);
|
Chris@82
|
1050 TcA = Tcs + Tcz;
|
Chris@82
|
1051 ro[WS(os, 38)] = FNMS(KP831469612, TcA, Tcl);
|
Chris@82
|
1052 ro[WS(os, 6)] = FMA(KP831469612, TcA, Tcl);
|
Chris@82
|
1053 TcN = FMA(KP923879532, TcG, TcD);
|
Chris@82
|
1054 TcO = TcK + TcL;
|
Chris@82
|
1055 io[WS(os, 38)] = FNMS(KP831469612, TcO, TcN);
|
Chris@82
|
1056 io[WS(os, 6)] = FMA(KP831469612, TcO, TcN);
|
Chris@82
|
1057 }
|
Chris@82
|
1058 {
|
Chris@82
|
1059 E TcH, TcI, TcJ, TcM;
|
Chris@82
|
1060 TcH = FNMS(KP923879532, TcG, TcD);
|
Chris@82
|
1061 TcI = Tcz - Tcs;
|
Chris@82
|
1062 io[WS(os, 54)] = FNMS(KP831469612, TcI, TcH);
|
Chris@82
|
1063 io[WS(os, 22)] = FMA(KP831469612, TcI, TcH);
|
Chris@82
|
1064 TcJ = FNMS(KP923879532, Tck, Tcd);
|
Chris@82
|
1065 TcM = TcK - TcL;
|
Chris@82
|
1066 ro[WS(os, 54)] = FNMS(KP831469612, TcM, TcJ);
|
Chris@82
|
1067 ro[WS(os, 22)] = FMA(KP831469612, TcM, TcJ);
|
Chris@82
|
1068 }
|
Chris@82
|
1069 {
|
Chris@82
|
1070 E TcR, TcY, Td3, Td6;
|
Chris@82
|
1071 TcR = FNMS(KP923879532, TcQ, TcP);
|
Chris@82
|
1072 TcY = TcU - TcX;
|
Chris@82
|
1073 ro[WS(os, 46)] = FNMS(KP980785280, TcY, TcR);
|
Chris@82
|
1074 ro[WS(os, 14)] = FMA(KP980785280, TcY, TcR);
|
Chris@82
|
1075 Td3 = FNMS(KP923879532, Td0, TcZ);
|
Chris@82
|
1076 Td6 = Td4 - Td5;
|
Chris@82
|
1077 io[WS(os, 46)] = FNMS(KP980785280, Td6, Td3);
|
Chris@82
|
1078 io[WS(os, 14)] = FMA(KP980785280, Td6, Td3);
|
Chris@82
|
1079 }
|
Chris@82
|
1080 {
|
Chris@82
|
1081 E Td1, Td2, Td7, Td8;
|
Chris@82
|
1082 Td1 = FMA(KP923879532, Td0, TcZ);
|
Chris@82
|
1083 Td2 = TcU + TcX;
|
Chris@82
|
1084 io[WS(os, 30)] = FNMS(KP980785280, Td2, Td1);
|
Chris@82
|
1085 io[WS(os, 62)] = FMA(KP980785280, Td2, Td1);
|
Chris@82
|
1086 Td7 = FMA(KP923879532, TcQ, TcP);
|
Chris@82
|
1087 Td8 = Td5 + Td4;
|
Chris@82
|
1088 ro[WS(os, 30)] = FNMS(KP980785280, Td8, Td7);
|
Chris@82
|
1089 ro[WS(os, 62)] = FMA(KP980785280, Td8, Td7);
|
Chris@82
|
1090 }
|
Chris@82
|
1091 }
|
Chris@82
|
1092 {
|
Chris@82
|
1093 E Tap, TbR, TbF, Tc1, TaE, Tc2, TbZ, Tc7, Tb6, TbN, TbI, TbS, TbW, Tc6, Tbx;
|
Chris@82
|
1094 E TbM, Tao, TbE;
|
Chris@82
|
1095 Tao = Tak + Tan;
|
Chris@82
|
1096 Tap = FNMS(KP707106781, Tao, Tah);
|
Chris@82
|
1097 TbR = FMA(KP707106781, Tao, Tah);
|
Chris@82
|
1098 TbE = TbC + TbD;
|
Chris@82
|
1099 TbF = FNMS(KP707106781, TbE, TbB);
|
Chris@82
|
1100 Tc1 = FMA(KP707106781, TbE, TbB);
|
Chris@82
|
1101 {
|
Chris@82
|
1102 E Taw, TaD, TbX, TbY;
|
Chris@82
|
1103 Taw = FNMS(KP414213562, Tav, Tas);
|
Chris@82
|
1104 TaD = FMA(KP414213562, TaC, Taz);
|
Chris@82
|
1105 TaE = Taw - TaD;
|
Chris@82
|
1106 Tc2 = Taw + TaD;
|
Chris@82
|
1107 TbX = FMA(KP707106781, Tbv, Tbs);
|
Chris@82
|
1108 TbY = FMA(KP707106781, Tbm, Tbb);
|
Chris@82
|
1109 TbZ = FNMS(KP198912367, TbY, TbX);
|
Chris@82
|
1110 Tc7 = FMA(KP198912367, TbX, TbY);
|
Chris@82
|
1111 }
|
Chris@82
|
1112 {
|
Chris@82
|
1113 E TaW, Tb5, TbG, TbH;
|
Chris@82
|
1114 TaW = FNMS(KP707106781, TaV, TaK);
|
Chris@82
|
1115 Tb5 = FNMS(KP707106781, Tb4, Tb1);
|
Chris@82
|
1116 Tb6 = FMA(KP668178637, Tb5, TaW);
|
Chris@82
|
1117 TbN = FNMS(KP668178637, TaW, Tb5);
|
Chris@82
|
1118 TbG = FNMS(KP414213562, Taz, TaC);
|
Chris@82
|
1119 TbH = FMA(KP414213562, Tas, Tav);
|
Chris@82
|
1120 TbI = TbG - TbH;
|
Chris@82
|
1121 TbS = TbH + TbG;
|
Chris@82
|
1122 }
|
Chris@82
|
1123 {
|
Chris@82
|
1124 E TbU, TbV, Tbn, Tbw;
|
Chris@82
|
1125 TbU = FMA(KP707106781, Tb4, Tb1);
|
Chris@82
|
1126 TbV = FMA(KP707106781, TaV, TaK);
|
Chris@82
|
1127 TbW = FMA(KP198912367, TbV, TbU);
|
Chris@82
|
1128 Tc6 = FNMS(KP198912367, TbU, TbV);
|
Chris@82
|
1129 Tbn = FNMS(KP707106781, Tbm, Tbb);
|
Chris@82
|
1130 Tbw = FNMS(KP707106781, Tbv, Tbs);
|
Chris@82
|
1131 Tbx = FNMS(KP668178637, Tbw, Tbn);
|
Chris@82
|
1132 TbM = FMA(KP668178637, Tbn, Tbw);
|
Chris@82
|
1133 }
|
Chris@82
|
1134 {
|
Chris@82
|
1135 E TaF, Tby, TbL, TbO;
|
Chris@82
|
1136 TaF = FMA(KP923879532, TaE, Tap);
|
Chris@82
|
1137 Tby = Tb6 - Tbx;
|
Chris@82
|
1138 ro[WS(os, 42)] = FNMS(KP831469612, Tby, TaF);
|
Chris@82
|
1139 ro[WS(os, 10)] = FMA(KP831469612, Tby, TaF);
|
Chris@82
|
1140 TbL = FMA(KP923879532, TbI, TbF);
|
Chris@82
|
1141 TbO = TbM - TbN;
|
Chris@82
|
1142 io[WS(os, 42)] = FNMS(KP831469612, TbO, TbL);
|
Chris@82
|
1143 io[WS(os, 10)] = FMA(KP831469612, TbO, TbL);
|
Chris@82
|
1144 }
|
Chris@82
|
1145 {
|
Chris@82
|
1146 E TbJ, TbK, TbP, TbQ;
|
Chris@82
|
1147 TbJ = FNMS(KP923879532, TbI, TbF);
|
Chris@82
|
1148 TbK = Tb6 + Tbx;
|
Chris@82
|
1149 io[WS(os, 26)] = FNMS(KP831469612, TbK, TbJ);
|
Chris@82
|
1150 io[WS(os, 58)] = FMA(KP831469612, TbK, TbJ);
|
Chris@82
|
1151 TbP = FNMS(KP923879532, TaE, Tap);
|
Chris@82
|
1152 TbQ = TbN + TbM;
|
Chris@82
|
1153 ro[WS(os, 26)] = FNMS(KP831469612, TbQ, TbP);
|
Chris@82
|
1154 ro[WS(os, 58)] = FMA(KP831469612, TbQ, TbP);
|
Chris@82
|
1155 }
|
Chris@82
|
1156 {
|
Chris@82
|
1157 E TbT, Tc0, Tc9, Tca;
|
Chris@82
|
1158 TbT = FMA(KP923879532, TbS, TbR);
|
Chris@82
|
1159 Tc0 = TbW + TbZ;
|
Chris@82
|
1160 ro[WS(os, 34)] = FNMS(KP980785280, Tc0, TbT);
|
Chris@82
|
1161 ro[WS(os, 2)] = FMA(KP980785280, Tc0, TbT);
|
Chris@82
|
1162 Tc9 = FMA(KP923879532, Tc2, Tc1);
|
Chris@82
|
1163 Tca = Tc6 + Tc7;
|
Chris@82
|
1164 io[WS(os, 34)] = FNMS(KP980785280, Tca, Tc9);
|
Chris@82
|
1165 io[WS(os, 2)] = FMA(KP980785280, Tca, Tc9);
|
Chris@82
|
1166 }
|
Chris@82
|
1167 {
|
Chris@82
|
1168 E Tc3, Tc4, Tc5, Tc8;
|
Chris@82
|
1169 Tc3 = FNMS(KP923879532, Tc2, Tc1);
|
Chris@82
|
1170 Tc4 = TbZ - TbW;
|
Chris@82
|
1171 io[WS(os, 50)] = FNMS(KP980785280, Tc4, Tc3);
|
Chris@82
|
1172 io[WS(os, 18)] = FMA(KP980785280, Tc4, Tc3);
|
Chris@82
|
1173 Tc5 = FNMS(KP923879532, TbS, TbR);
|
Chris@82
|
1174 Tc8 = Tc6 - Tc7;
|
Chris@82
|
1175 ro[WS(os, 50)] = FNMS(KP980785280, Tc8, Tc5);
|
Chris@82
|
1176 ro[WS(os, 18)] = FMA(KP980785280, Tc8, Tc5);
|
Chris@82
|
1177 }
|
Chris@82
|
1178 }
|
Chris@82
|
1179 {
|
Chris@82
|
1180 E T6F, T7h, T7m, T7x, T7p, T7w, T6M, T7s, T6U, T7c, T75, T7r, T78, T7i, T71;
|
Chris@82
|
1181 E T7d;
|
Chris@82
|
1182 {
|
Chris@82
|
1183 E T6D, T6E, T7k, T7l;
|
Chris@82
|
1184 T6D = FNMS(KP707106781, T3e, T37);
|
Chris@82
|
1185 T6E = T65 + T64;
|
Chris@82
|
1186 T6F = FNMS(KP923879532, T6E, T6D);
|
Chris@82
|
1187 T7h = FMA(KP923879532, T6E, T6D);
|
Chris@82
|
1188 T7k = FMA(KP923879532, T6S, T6R);
|
Chris@82
|
1189 T7l = FMA(KP923879532, T6P, T6O);
|
Chris@82
|
1190 T7m = FMA(KP098491403, T7l, T7k);
|
Chris@82
|
1191 T7x = FNMS(KP098491403, T7k, T7l);
|
Chris@82
|
1192 }
|
Chris@82
|
1193 {
|
Chris@82
|
1194 E T7n, T7o, T6I, T6L;
|
Chris@82
|
1195 T7n = FMA(KP923879532, T6Z, T6Y);
|
Chris@82
|
1196 T7o = FMA(KP923879532, T6W, T6V);
|
Chris@82
|
1197 T7p = FNMS(KP098491403, T7o, T7n);
|
Chris@82
|
1198 T7w = FMA(KP098491403, T7n, T7o);
|
Chris@82
|
1199 T6I = FMA(KP198912367, T6H, T6G);
|
Chris@82
|
1200 T6L = FNMS(KP198912367, T6K, T6J);
|
Chris@82
|
1201 T6M = T6I - T6L;
|
Chris@82
|
1202 T7s = T6I + T6L;
|
Chris@82
|
1203 }
|
Chris@82
|
1204 {
|
Chris@82
|
1205 E T6Q, T6T, T73, T74;
|
Chris@82
|
1206 T6Q = FNMS(KP923879532, T6P, T6O);
|
Chris@82
|
1207 T6T = FNMS(KP923879532, T6S, T6R);
|
Chris@82
|
1208 T6U = FMA(KP820678790, T6T, T6Q);
|
Chris@82
|
1209 T7c = FNMS(KP820678790, T6Q, T6T);
|
Chris@82
|
1210 T73 = FNMS(KP707106781, T62, T5Z);
|
Chris@82
|
1211 T74 = T3m + T3t;
|
Chris@82
|
1212 T75 = FNMS(KP923879532, T74, T73);
|
Chris@82
|
1213 T7r = FMA(KP923879532, T74, T73);
|
Chris@82
|
1214 }
|
Chris@82
|
1215 {
|
Chris@82
|
1216 E T76, T77, T6X, T70;
|
Chris@82
|
1217 T76 = FMA(KP198912367, T6J, T6K);
|
Chris@82
|
1218 T77 = FNMS(KP198912367, T6G, T6H);
|
Chris@82
|
1219 T78 = T76 - T77;
|
Chris@82
|
1220 T7i = T77 + T76;
|
Chris@82
|
1221 T6X = FNMS(KP923879532, T6W, T6V);
|
Chris@82
|
1222 T70 = FNMS(KP923879532, T6Z, T6Y);
|
Chris@82
|
1223 T71 = FNMS(KP820678790, T70, T6X);
|
Chris@82
|
1224 T7d = FMA(KP820678790, T6X, T70);
|
Chris@82
|
1225 }
|
Chris@82
|
1226 {
|
Chris@82
|
1227 E T6N, T72, T7f, T7g;
|
Chris@82
|
1228 T6N = FMA(KP980785280, T6M, T6F);
|
Chris@82
|
1229 T72 = T6U + T71;
|
Chris@82
|
1230 ro[WS(os, 39)] = FNMS(KP773010453, T72, T6N);
|
Chris@82
|
1231 ro[WS(os, 7)] = FMA(KP773010453, T72, T6N);
|
Chris@82
|
1232 T7f = FMA(KP980785280, T78, T75);
|
Chris@82
|
1233 T7g = T7c + T7d;
|
Chris@82
|
1234 io[WS(os, 39)] = FNMS(KP773010453, T7g, T7f);
|
Chris@82
|
1235 io[WS(os, 7)] = FMA(KP773010453, T7g, T7f);
|
Chris@82
|
1236 }
|
Chris@82
|
1237 {
|
Chris@82
|
1238 E T79, T7a, T7b, T7e;
|
Chris@82
|
1239 T79 = FNMS(KP980785280, T78, T75);
|
Chris@82
|
1240 T7a = T71 - T6U;
|
Chris@82
|
1241 io[WS(os, 55)] = FNMS(KP773010453, T7a, T79);
|
Chris@82
|
1242 io[WS(os, 23)] = FMA(KP773010453, T7a, T79);
|
Chris@82
|
1243 T7b = FNMS(KP980785280, T6M, T6F);
|
Chris@82
|
1244 T7e = T7c - T7d;
|
Chris@82
|
1245 ro[WS(os, 55)] = FNMS(KP773010453, T7e, T7b);
|
Chris@82
|
1246 ro[WS(os, 23)] = FMA(KP773010453, T7e, T7b);
|
Chris@82
|
1247 }
|
Chris@82
|
1248 {
|
Chris@82
|
1249 E T7j, T7q, T7v, T7y;
|
Chris@82
|
1250 T7j = FNMS(KP980785280, T7i, T7h);
|
Chris@82
|
1251 T7q = T7m - T7p;
|
Chris@82
|
1252 ro[WS(os, 47)] = FNMS(KP995184726, T7q, T7j);
|
Chris@82
|
1253 ro[WS(os, 15)] = FMA(KP995184726, T7q, T7j);
|
Chris@82
|
1254 T7v = FNMS(KP980785280, T7s, T7r);
|
Chris@82
|
1255 T7y = T7w - T7x;
|
Chris@82
|
1256 io[WS(os, 47)] = FNMS(KP995184726, T7y, T7v);
|
Chris@82
|
1257 io[WS(os, 15)] = FMA(KP995184726, T7y, T7v);
|
Chris@82
|
1258 }
|
Chris@82
|
1259 {
|
Chris@82
|
1260 E T7t, T7u, T7z, T7A;
|
Chris@82
|
1261 T7t = FMA(KP980785280, T7s, T7r);
|
Chris@82
|
1262 T7u = T7m + T7p;
|
Chris@82
|
1263 io[WS(os, 31)] = FNMS(KP995184726, T7u, T7t);
|
Chris@82
|
1264 io[WS(os, 63)] = FMA(KP995184726, T7u, T7t);
|
Chris@82
|
1265 T7z = FMA(KP980785280, T7i, T7h);
|
Chris@82
|
1266 T7A = T7x + T7w;
|
Chris@82
|
1267 ro[WS(os, 31)] = FNMS(KP995184726, T7A, T7z);
|
Chris@82
|
1268 ro[WS(os, 63)] = FMA(KP995184726, T7A, T7z);
|
Chris@82
|
1269 }
|
Chris@82
|
1270 }
|
Chris@82
|
1271 {
|
Chris@82
|
1272 E T9j, T9V, Ta0, Tab, Ta3, Taa, T9q, Ta6, T9y, T9Q, T9J, Ta5, T9M, T9W, T9F;
|
Chris@82
|
1273 E T9R;
|
Chris@82
|
1274 {
|
Chris@82
|
1275 E T9h, T9i, T9Y, T9Z;
|
Chris@82
|
1276 T9h = FNMS(KP707106781, T7C, T7B);
|
Chris@82
|
1277 T9i = T8I - T8J;
|
Chris@82
|
1278 T9j = FMA(KP923879532, T9i, T9h);
|
Chris@82
|
1279 T9V = FNMS(KP923879532, T9i, T9h);
|
Chris@82
|
1280 T9Y = FMA(KP923879532, T9w, T9v);
|
Chris@82
|
1281 T9Z = FMA(KP923879532, T9t, T9s);
|
Chris@82
|
1282 Ta0 = FMA(KP303346683, T9Z, T9Y);
|
Chris@82
|
1283 Tab = FNMS(KP303346683, T9Y, T9Z);
|
Chris@82
|
1284 }
|
Chris@82
|
1285 {
|
Chris@82
|
1286 E Ta1, Ta2, T9m, T9p;
|
Chris@82
|
1287 Ta1 = FMA(KP923879532, T9D, T9C);
|
Chris@82
|
1288 Ta2 = FMA(KP923879532, T9A, T9z);
|
Chris@82
|
1289 Ta3 = FNMS(KP303346683, Ta2, Ta1);
|
Chris@82
|
1290 Taa = FMA(KP303346683, Ta1, Ta2);
|
Chris@82
|
1291 T9m = FMA(KP668178637, T9l, T9k);
|
Chris@82
|
1292 T9p = FNMS(KP668178637, T9o, T9n);
|
Chris@82
|
1293 T9q = T9m - T9p;
|
Chris@82
|
1294 Ta6 = T9m + T9p;
|
Chris@82
|
1295 }
|
Chris@82
|
1296 {
|
Chris@82
|
1297 E T9u, T9x, T9H, T9I;
|
Chris@82
|
1298 T9u = FNMS(KP923879532, T9t, T9s);
|
Chris@82
|
1299 T9x = FNMS(KP923879532, T9w, T9v);
|
Chris@82
|
1300 T9y = FMA(KP534511135, T9x, T9u);
|
Chris@82
|
1301 T9Q = FNMS(KP534511135, T9u, T9x);
|
Chris@82
|
1302 T9H = FNMS(KP707106781, T8G, T8F);
|
Chris@82
|
1303 T9I = T7J - T7G;
|
Chris@82
|
1304 T9J = FMA(KP923879532, T9I, T9H);
|
Chris@82
|
1305 Ta5 = FNMS(KP923879532, T9I, T9H);
|
Chris@82
|
1306 }
|
Chris@82
|
1307 {
|
Chris@82
|
1308 E T9K, T9L, T9B, T9E;
|
Chris@82
|
1309 T9K = FMA(KP668178637, T9n, T9o);
|
Chris@82
|
1310 T9L = FNMS(KP668178637, T9k, T9l);
|
Chris@82
|
1311 T9M = T9K - T9L;
|
Chris@82
|
1312 T9W = T9L + T9K;
|
Chris@82
|
1313 T9B = FNMS(KP923879532, T9A, T9z);
|
Chris@82
|
1314 T9E = FNMS(KP923879532, T9D, T9C);
|
Chris@82
|
1315 T9F = FNMS(KP534511135, T9E, T9B);
|
Chris@82
|
1316 T9R = FMA(KP534511135, T9B, T9E);
|
Chris@82
|
1317 }
|
Chris@82
|
1318 {
|
Chris@82
|
1319 E T9r, T9G, T9T, T9U;
|
Chris@82
|
1320 T9r = FMA(KP831469612, T9q, T9j);
|
Chris@82
|
1321 T9G = T9y + T9F;
|
Chris@82
|
1322 ro[WS(os, 37)] = FNMS(KP881921264, T9G, T9r);
|
Chris@82
|
1323 ro[WS(os, 5)] = FMA(KP881921264, T9G, T9r);
|
Chris@82
|
1324 T9T = FMA(KP831469612, T9M, T9J);
|
Chris@82
|
1325 T9U = T9Q + T9R;
|
Chris@82
|
1326 io[WS(os, 37)] = FNMS(KP881921264, T9U, T9T);
|
Chris@82
|
1327 io[WS(os, 5)] = FMA(KP881921264, T9U, T9T);
|
Chris@82
|
1328 }
|
Chris@82
|
1329 {
|
Chris@82
|
1330 E T9N, T9O, T9P, T9S;
|
Chris@82
|
1331 T9N = FNMS(KP831469612, T9M, T9J);
|
Chris@82
|
1332 T9O = T9F - T9y;
|
Chris@82
|
1333 io[WS(os, 53)] = FNMS(KP881921264, T9O, T9N);
|
Chris@82
|
1334 io[WS(os, 21)] = FMA(KP881921264, T9O, T9N);
|
Chris@82
|
1335 T9P = FNMS(KP831469612, T9q, T9j);
|
Chris@82
|
1336 T9S = T9Q - T9R;
|
Chris@82
|
1337 ro[WS(os, 53)] = FNMS(KP881921264, T9S, T9P);
|
Chris@82
|
1338 ro[WS(os, 21)] = FMA(KP881921264, T9S, T9P);
|
Chris@82
|
1339 }
|
Chris@82
|
1340 {
|
Chris@82
|
1341 E T9X, Ta4, Ta9, Tac;
|
Chris@82
|
1342 T9X = FNMS(KP831469612, T9W, T9V);
|
Chris@82
|
1343 Ta4 = Ta0 - Ta3;
|
Chris@82
|
1344 ro[WS(os, 45)] = FNMS(KP956940335, Ta4, T9X);
|
Chris@82
|
1345 ro[WS(os, 13)] = FMA(KP956940335, Ta4, T9X);
|
Chris@82
|
1346 Ta9 = FNMS(KP831469612, Ta6, Ta5);
|
Chris@82
|
1347 Tac = Taa - Tab;
|
Chris@82
|
1348 io[WS(os, 45)] = FNMS(KP956940335, Tac, Ta9);
|
Chris@82
|
1349 io[WS(os, 13)] = FMA(KP956940335, Tac, Ta9);
|
Chris@82
|
1350 }
|
Chris@82
|
1351 {
|
Chris@82
|
1352 E Ta7, Ta8, Tad, Tae;
|
Chris@82
|
1353 Ta7 = FMA(KP831469612, Ta6, Ta5);
|
Chris@82
|
1354 Ta8 = Ta0 + Ta3;
|
Chris@82
|
1355 io[WS(os, 29)] = FNMS(KP956940335, Ta8, Ta7);
|
Chris@82
|
1356 io[WS(os, 61)] = FMA(KP956940335, Ta8, Ta7);
|
Chris@82
|
1357 Tad = FMA(KP831469612, T9W, T9V);
|
Chris@82
|
1358 Tae = Tab + Taa;
|
Chris@82
|
1359 ro[WS(os, 29)] = FNMS(KP956940335, Tae, Tad);
|
Chris@82
|
1360 ro[WS(os, 61)] = FMA(KP956940335, Tae, Tad);
|
Chris@82
|
1361 }
|
Chris@82
|
1362 }
|
Chris@82
|
1363 {
|
Chris@82
|
1364 E T3v, T6j, T6o, T6y, T6r, T6z, T48, T6u, T52, T6f, T67, T6t, T6a, T6k, T5V;
|
Chris@82
|
1365 E T6e;
|
Chris@82
|
1366 {
|
Chris@82
|
1367 E T3f, T3u, T6m, T6n;
|
Chris@82
|
1368 T3f = FMA(KP707106781, T3e, T37);
|
Chris@82
|
1369 T3u = T3m - T3t;
|
Chris@82
|
1370 T3v = FNMS(KP923879532, T3u, T3f);
|
Chris@82
|
1371 T6j = FMA(KP923879532, T3u, T3f);
|
Chris@82
|
1372 T6m = FMA(KP923879532, T50, T4X);
|
Chris@82
|
1373 T6n = FMA(KP923879532, T4N, T4q);
|
Chris@82
|
1374 T6o = FMA(KP303346683, T6n, T6m);
|
Chris@82
|
1375 T6y = FNMS(KP303346683, T6m, T6n);
|
Chris@82
|
1376 }
|
Chris@82
|
1377 {
|
Chris@82
|
1378 E T6p, T6q, T3O, T47;
|
Chris@82
|
1379 T6p = FMA(KP923879532, T5T, T5Q);
|
Chris@82
|
1380 T6q = FMA(KP923879532, T5G, T5j);
|
Chris@82
|
1381 T6r = FNMS(KP303346683, T6q, T6p);
|
Chris@82
|
1382 T6z = FMA(KP303346683, T6p, T6q);
|
Chris@82
|
1383 T3O = FNMS(KP668178637, T3N, T3G);
|
Chris@82
|
1384 T47 = FMA(KP668178637, T46, T3Z);
|
Chris@82
|
1385 T48 = T3O - T47;
|
Chris@82
|
1386 T6u = T3O + T47;
|
Chris@82
|
1387 }
|
Chris@82
|
1388 {
|
Chris@82
|
1389 E T4O, T51, T63, T66;
|
Chris@82
|
1390 T4O = FNMS(KP923879532, T4N, T4q);
|
Chris@82
|
1391 T51 = FNMS(KP923879532, T50, T4X);
|
Chris@82
|
1392 T52 = FMA(KP534511135, T51, T4O);
|
Chris@82
|
1393 T6f = FNMS(KP534511135, T4O, T51);
|
Chris@82
|
1394 T63 = FMA(KP707106781, T62, T5Z);
|
Chris@82
|
1395 T66 = T64 - T65;
|
Chris@82
|
1396 T67 = FNMS(KP923879532, T66, T63);
|
Chris@82
|
1397 T6t = FMA(KP923879532, T66, T63);
|
Chris@82
|
1398 }
|
Chris@82
|
1399 {
|
Chris@82
|
1400 E T68, T69, T5H, T5U;
|
Chris@82
|
1401 T68 = FNMS(KP668178637, T3Z, T46);
|
Chris@82
|
1402 T69 = FMA(KP668178637, T3G, T3N);
|
Chris@82
|
1403 T6a = T68 - T69;
|
Chris@82
|
1404 T6k = T69 + T68;
|
Chris@82
|
1405 T5H = FNMS(KP923879532, T5G, T5j);
|
Chris@82
|
1406 T5U = FNMS(KP923879532, T5T, T5Q);
|
Chris@82
|
1407 T5V = FNMS(KP534511135, T5U, T5H);
|
Chris@82
|
1408 T6e = FMA(KP534511135, T5H, T5U);
|
Chris@82
|
1409 }
|
Chris@82
|
1410 {
|
Chris@82
|
1411 E T49, T5W, T6d, T6g;
|
Chris@82
|
1412 T49 = FMA(KP831469612, T48, T3v);
|
Chris@82
|
1413 T5W = T52 - T5V;
|
Chris@82
|
1414 ro[WS(os, 43)] = FNMS(KP881921264, T5W, T49);
|
Chris@82
|
1415 ro[WS(os, 11)] = FMA(KP881921264, T5W, T49);
|
Chris@82
|
1416 T6d = FMA(KP831469612, T6a, T67);
|
Chris@82
|
1417 T6g = T6e - T6f;
|
Chris@82
|
1418 io[WS(os, 43)] = FNMS(KP881921264, T6g, T6d);
|
Chris@82
|
1419 io[WS(os, 11)] = FMA(KP881921264, T6g, T6d);
|
Chris@82
|
1420 }
|
Chris@82
|
1421 {
|
Chris@82
|
1422 E T6b, T6c, T6h, T6i;
|
Chris@82
|
1423 T6b = FNMS(KP831469612, T6a, T67);
|
Chris@82
|
1424 T6c = T52 + T5V;
|
Chris@82
|
1425 io[WS(os, 27)] = FNMS(KP881921264, T6c, T6b);
|
Chris@82
|
1426 io[WS(os, 59)] = FMA(KP881921264, T6c, T6b);
|
Chris@82
|
1427 T6h = FNMS(KP831469612, T48, T3v);
|
Chris@82
|
1428 T6i = T6f + T6e;
|
Chris@82
|
1429 ro[WS(os, 27)] = FNMS(KP881921264, T6i, T6h);
|
Chris@82
|
1430 ro[WS(os, 59)] = FMA(KP881921264, T6i, T6h);
|
Chris@82
|
1431 }
|
Chris@82
|
1432 {
|
Chris@82
|
1433 E T6l, T6s, T6B, T6C;
|
Chris@82
|
1434 T6l = FMA(KP831469612, T6k, T6j);
|
Chris@82
|
1435 T6s = T6o + T6r;
|
Chris@82
|
1436 ro[WS(os, 35)] = FNMS(KP956940335, T6s, T6l);
|
Chris@82
|
1437 ro[WS(os, 3)] = FMA(KP956940335, T6s, T6l);
|
Chris@82
|
1438 T6B = FMA(KP831469612, T6u, T6t);
|
Chris@82
|
1439 T6C = T6y + T6z;
|
Chris@82
|
1440 io[WS(os, 35)] = FNMS(KP956940335, T6C, T6B);
|
Chris@82
|
1441 io[WS(os, 3)] = FMA(KP956940335, T6C, T6B);
|
Chris@82
|
1442 }
|
Chris@82
|
1443 {
|
Chris@82
|
1444 E T6v, T6w, T6x, T6A;
|
Chris@82
|
1445 T6v = FNMS(KP831469612, T6u, T6t);
|
Chris@82
|
1446 T6w = T6r - T6o;
|
Chris@82
|
1447 io[WS(os, 51)] = FNMS(KP956940335, T6w, T6v);
|
Chris@82
|
1448 io[WS(os, 19)] = FMA(KP956940335, T6w, T6v);
|
Chris@82
|
1449 T6x = FNMS(KP831469612, T6k, T6j);
|
Chris@82
|
1450 T6A = T6y - T6z;
|
Chris@82
|
1451 ro[WS(os, 51)] = FNMS(KP956940335, T6A, T6x);
|
Chris@82
|
1452 ro[WS(os, 19)] = FMA(KP956940335, T6A, T6x);
|
Chris@82
|
1453 }
|
Chris@82
|
1454 }
|
Chris@82
|
1455 {
|
Chris@82
|
1456 E T7L, T8X, T92, T9c, T95, T9d, T80, T98, T8k, T8T, T8L, T97, T8O, T8Y, T8D;
|
Chris@82
|
1457 E T8S;
|
Chris@82
|
1458 {
|
Chris@82
|
1459 E T7D, T7K, T90, T91;
|
Chris@82
|
1460 T7D = FMA(KP707106781, T7C, T7B);
|
Chris@82
|
1461 T7K = T7G + T7J;
|
Chris@82
|
1462 T7L = FNMS(KP923879532, T7K, T7D);
|
Chris@82
|
1463 T8X = FMA(KP923879532, T7K, T7D);
|
Chris@82
|
1464 T90 = FMA(KP923879532, T8i, T8f);
|
Chris@82
|
1465 T91 = FMA(KP923879532, T8b, T84);
|
Chris@82
|
1466 T92 = FMA(KP098491403, T91, T90);
|
Chris@82
|
1467 T9c = FNMS(KP098491403, T90, T91);
|
Chris@82
|
1468 }
|
Chris@82
|
1469 {
|
Chris@82
|
1470 E T93, T94, T7S, T7Z;
|
Chris@82
|
1471 T93 = FMA(KP923879532, T8B, T8y);
|
Chris@82
|
1472 T94 = FMA(KP923879532, T8u, T8n);
|
Chris@82
|
1473 T95 = FNMS(KP098491403, T94, T93);
|
Chris@82
|
1474 T9d = FMA(KP098491403, T93, T94);
|
Chris@82
|
1475 T7S = FNMS(KP198912367, T7R, T7O);
|
Chris@82
|
1476 T7Z = FMA(KP198912367, T7Y, T7V);
|
Chris@82
|
1477 T80 = T7S - T7Z;
|
Chris@82
|
1478 T98 = T7S + T7Z;
|
Chris@82
|
1479 }
|
Chris@82
|
1480 {
|
Chris@82
|
1481 E T8c, T8j, T8H, T8K;
|
Chris@82
|
1482 T8c = FNMS(KP923879532, T8b, T84);
|
Chris@82
|
1483 T8j = FNMS(KP923879532, T8i, T8f);
|
Chris@82
|
1484 T8k = FMA(KP820678790, T8j, T8c);
|
Chris@82
|
1485 T8T = FNMS(KP820678790, T8c, T8j);
|
Chris@82
|
1486 T8H = FMA(KP707106781, T8G, T8F);
|
Chris@82
|
1487 T8K = T8I + T8J;
|
Chris@82
|
1488 T8L = FNMS(KP923879532, T8K, T8H);
|
Chris@82
|
1489 T97 = FMA(KP923879532, T8K, T8H);
|
Chris@82
|
1490 }
|
Chris@82
|
1491 {
|
Chris@82
|
1492 E T8M, T8N, T8v, T8C;
|
Chris@82
|
1493 T8M = FNMS(KP198912367, T7V, T7Y);
|
Chris@82
|
1494 T8N = FMA(KP198912367, T7O, T7R);
|
Chris@82
|
1495 T8O = T8M - T8N;
|
Chris@82
|
1496 T8Y = T8N + T8M;
|
Chris@82
|
1497 T8v = FNMS(KP923879532, T8u, T8n);
|
Chris@82
|
1498 T8C = FNMS(KP923879532, T8B, T8y);
|
Chris@82
|
1499 T8D = FNMS(KP820678790, T8C, T8v);
|
Chris@82
|
1500 T8S = FMA(KP820678790, T8v, T8C);
|
Chris@82
|
1501 }
|
Chris@82
|
1502 {
|
Chris@82
|
1503 E T81, T8E, T8R, T8U;
|
Chris@82
|
1504 T81 = FMA(KP980785280, T80, T7L);
|
Chris@82
|
1505 T8E = T8k - T8D;
|
Chris@82
|
1506 ro[WS(os, 41)] = FNMS(KP773010453, T8E, T81);
|
Chris@82
|
1507 ro[WS(os, 9)] = FMA(KP773010453, T8E, T81);
|
Chris@82
|
1508 T8R = FMA(KP980785280, T8O, T8L);
|
Chris@82
|
1509 T8U = T8S - T8T;
|
Chris@82
|
1510 io[WS(os, 41)] = FNMS(KP773010453, T8U, T8R);
|
Chris@82
|
1511 io[WS(os, 9)] = FMA(KP773010453, T8U, T8R);
|
Chris@82
|
1512 }
|
Chris@82
|
1513 {
|
Chris@82
|
1514 E T8P, T8Q, T8V, T8W;
|
Chris@82
|
1515 T8P = FNMS(KP980785280, T8O, T8L);
|
Chris@82
|
1516 T8Q = T8k + T8D;
|
Chris@82
|
1517 io[WS(os, 25)] = FNMS(KP773010453, T8Q, T8P);
|
Chris@82
|
1518 io[WS(os, 57)] = FMA(KP773010453, T8Q, T8P);
|
Chris@82
|
1519 T8V = FNMS(KP980785280, T80, T7L);
|
Chris@82
|
1520 T8W = T8T + T8S;
|
Chris@82
|
1521 ro[WS(os, 25)] = FNMS(KP773010453, T8W, T8V);
|
Chris@82
|
1522 ro[WS(os, 57)] = FMA(KP773010453, T8W, T8V);
|
Chris@82
|
1523 }
|
Chris@82
|
1524 {
|
Chris@82
|
1525 E T8Z, T96, T9f, T9g;
|
Chris@82
|
1526 T8Z = FMA(KP980785280, T8Y, T8X);
|
Chris@82
|
1527 T96 = T92 + T95;
|
Chris@82
|
1528 ro[WS(os, 33)] = FNMS(KP995184726, T96, T8Z);
|
Chris@82
|
1529 ro[WS(os, 1)] = FMA(KP995184726, T96, T8Z);
|
Chris@82
|
1530 T9f = FMA(KP980785280, T98, T97);
|
Chris@82
|
1531 T9g = T9c + T9d;
|
Chris@82
|
1532 io[WS(os, 33)] = FNMS(KP995184726, T9g, T9f);
|
Chris@82
|
1533 io[WS(os, 1)] = FMA(KP995184726, T9g, T9f);
|
Chris@82
|
1534 }
|
Chris@82
|
1535 {
|
Chris@82
|
1536 E T99, T9a, T9b, T9e;
|
Chris@82
|
1537 T99 = FNMS(KP980785280, T98, T97);
|
Chris@82
|
1538 T9a = T95 - T92;
|
Chris@82
|
1539 io[WS(os, 49)] = FNMS(KP995184726, T9a, T99);
|
Chris@82
|
1540 io[WS(os, 17)] = FMA(KP995184726, T9a, T99);
|
Chris@82
|
1541 T9b = FNMS(KP980785280, T8Y, T8X);
|
Chris@82
|
1542 T9e = T9c - T9d;
|
Chris@82
|
1543 ro[WS(os, 49)] = FNMS(KP995184726, T9e, T9b);
|
Chris@82
|
1544 ro[WS(os, 17)] = FMA(KP995184726, T9e, T9b);
|
Chris@82
|
1545 }
|
Chris@82
|
1546 }
|
Chris@82
|
1547 }
|
Chris@82
|
1548 }
|
Chris@82
|
1549 }
|
Chris@82
|
1550
|
Chris@82
|
1551 static const kdft_desc desc = { 64, "n1_64", {520, 0, 392, 0}, &GENUS, 0, 0, 0, 0 };
|
Chris@82
|
1552
|
Chris@82
|
1553 void X(codelet_n1_64) (planner *p) {
|
Chris@82
|
1554 X(kdft_register) (p, n1_64, &desc);
|
Chris@82
|
1555 }
|
Chris@82
|
1556
|
Chris@82
|
1557 #else
|
Chris@82
|
1558
|
Chris@82
|
1559 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 64 -name n1_64 -include dft/scalar/n.h */
|
Chris@82
|
1560
|
Chris@82
|
1561 /*
|
Chris@82
|
1562 * This function contains 912 FP additions, 248 FP multiplications,
|
Chris@82
|
1563 * (or, 808 additions, 144 multiplications, 104 fused multiply/add),
|
Chris@82
|
1564 * 172 stack variables, 15 constants, and 256 memory accesses
|
Chris@82
|
1565 */
|
Chris@82
|
1566 #include "dft/scalar/n.h"
|
Chris@82
|
1567
|
Chris@82
|
1568 static void n1_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@82
|
1569 {
|
Chris@82
|
1570 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@82
|
1571 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
Chris@82
|
1572 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
Chris@82
|
1573 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@82
|
1574 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@82
|
1575 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
Chris@82
|
1576 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
Chris@82
|
1577 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@82
|
1578 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@82
|
1579 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@82
|
1580 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@82
|
1581 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@82
|
1582 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@82
|
1583 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@82
|
1584 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@82
|
1585 {
|
Chris@82
|
1586 INT i;
|
Chris@82
|
1587 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(256, is), MAKE_VOLATILE_STRIDE(256, os)) {
|
Chris@82
|
1588 E T37, T7B, T8F, T5Z, Tf, Td9, TbB, TcB, T62, T7C, T2i, TdH, Tah, Tcb, T3e;
|
Chris@82
|
1589 E T8G, Tu, TdI, Tak, TbD, Tan, TbC, T2x, Tda, T3m, T65, T7G, T8J, T7J, T8I;
|
Chris@82
|
1590 E T3t, T64, TK, Tdd, Tas, Tce, Tav, Tcf, T2N, Tdc, T3G, T6G, T7O, T9k, T7R;
|
Chris@82
|
1591 E T9l, T3N, T6H, T1L, Tdv, Tbs, Tcw, TdC, Teo, T5j, T6V, T5Q, T6Y, T8y, T9C;
|
Chris@82
|
1592 E Tbb, Tct, T8n, T9z, TZ, Tdf, Taz, Tch, TaC, Tci, T32, Tdg, T3Z, T6J, T7V;
|
Chris@82
|
1593 E T9n, T7Y, T9o, T46, T6K, T1g, Tdp, Tb1, Tcm, Tdm, Tej, T4q, T6R, T4X, T6O;
|
Chris@82
|
1594 E T8f, T9s, TaK, Tcp, T84, T9v, T1v, Tdn, Tb4, Tcq, Tds, Tek, T4N, T6P, T50;
|
Chris@82
|
1595 E T6S, T8i, T9w, TaV, Tcn, T8b, T9t, T20, TdD, Tbv, Tcu, Tdy, Tep, T5G, T6Z;
|
Chris@82
|
1596 E T5T, T6W, T8B, T9A, Tbm, Tcx, T8u, T9D;
|
Chris@82
|
1597 {
|
Chris@82
|
1598 E T3, T35, T26, T5Y, T6, T5X, T29, T36, Ta, T39, T2d, T38, Td, T3b, T2g;
|
Chris@82
|
1599 E T3c;
|
Chris@82
|
1600 {
|
Chris@82
|
1601 E T1, T2, T24, T25;
|
Chris@82
|
1602 T1 = ri[0];
|
Chris@82
|
1603 T2 = ri[WS(is, 32)];
|
Chris@82
|
1604 T3 = T1 + T2;
|
Chris@82
|
1605 T35 = T1 - T2;
|
Chris@82
|
1606 T24 = ii[0];
|
Chris@82
|
1607 T25 = ii[WS(is, 32)];
|
Chris@82
|
1608 T26 = T24 + T25;
|
Chris@82
|
1609 T5Y = T24 - T25;
|
Chris@82
|
1610 }
|
Chris@82
|
1611 {
|
Chris@82
|
1612 E T4, T5, T27, T28;
|
Chris@82
|
1613 T4 = ri[WS(is, 16)];
|
Chris@82
|
1614 T5 = ri[WS(is, 48)];
|
Chris@82
|
1615 T6 = T4 + T5;
|
Chris@82
|
1616 T5X = T4 - T5;
|
Chris@82
|
1617 T27 = ii[WS(is, 16)];
|
Chris@82
|
1618 T28 = ii[WS(is, 48)];
|
Chris@82
|
1619 T29 = T27 + T28;
|
Chris@82
|
1620 T36 = T27 - T28;
|
Chris@82
|
1621 }
|
Chris@82
|
1622 {
|
Chris@82
|
1623 E T8, T9, T2b, T2c;
|
Chris@82
|
1624 T8 = ri[WS(is, 8)];
|
Chris@82
|
1625 T9 = ri[WS(is, 40)];
|
Chris@82
|
1626 Ta = T8 + T9;
|
Chris@82
|
1627 T39 = T8 - T9;
|
Chris@82
|
1628 T2b = ii[WS(is, 8)];
|
Chris@82
|
1629 T2c = ii[WS(is, 40)];
|
Chris@82
|
1630 T2d = T2b + T2c;
|
Chris@82
|
1631 T38 = T2b - T2c;
|
Chris@82
|
1632 }
|
Chris@82
|
1633 {
|
Chris@82
|
1634 E Tb, Tc, T2e, T2f;
|
Chris@82
|
1635 Tb = ri[WS(is, 56)];
|
Chris@82
|
1636 Tc = ri[WS(is, 24)];
|
Chris@82
|
1637 Td = Tb + Tc;
|
Chris@82
|
1638 T3b = Tb - Tc;
|
Chris@82
|
1639 T2e = ii[WS(is, 56)];
|
Chris@82
|
1640 T2f = ii[WS(is, 24)];
|
Chris@82
|
1641 T2g = T2e + T2f;
|
Chris@82
|
1642 T3c = T2e - T2f;
|
Chris@82
|
1643 }
|
Chris@82
|
1644 {
|
Chris@82
|
1645 E T7, Te, T2a, T2h;
|
Chris@82
|
1646 T37 = T35 - T36;
|
Chris@82
|
1647 T7B = T35 + T36;
|
Chris@82
|
1648 T8F = T5Y - T5X;
|
Chris@82
|
1649 T5Z = T5X + T5Y;
|
Chris@82
|
1650 T7 = T3 + T6;
|
Chris@82
|
1651 Te = Ta + Td;
|
Chris@82
|
1652 Tf = T7 + Te;
|
Chris@82
|
1653 Td9 = T7 - Te;
|
Chris@82
|
1654 {
|
Chris@82
|
1655 E Tbz, TbA, T60, T61;
|
Chris@82
|
1656 Tbz = T26 - T29;
|
Chris@82
|
1657 TbA = Td - Ta;
|
Chris@82
|
1658 TbB = Tbz - TbA;
|
Chris@82
|
1659 TcB = TbA + Tbz;
|
Chris@82
|
1660 T60 = T3b - T3c;
|
Chris@82
|
1661 T61 = T39 + T38;
|
Chris@82
|
1662 T62 = KP707106781 * (T60 - T61);
|
Chris@82
|
1663 T7C = KP707106781 * (T61 + T60);
|
Chris@82
|
1664 }
|
Chris@82
|
1665 T2a = T26 + T29;
|
Chris@82
|
1666 T2h = T2d + T2g;
|
Chris@82
|
1667 T2i = T2a + T2h;
|
Chris@82
|
1668 TdH = T2a - T2h;
|
Chris@82
|
1669 {
|
Chris@82
|
1670 E Taf, Tag, T3a, T3d;
|
Chris@82
|
1671 Taf = T3 - T6;
|
Chris@82
|
1672 Tag = T2d - T2g;
|
Chris@82
|
1673 Tah = Taf - Tag;
|
Chris@82
|
1674 Tcb = Taf + Tag;
|
Chris@82
|
1675 T3a = T38 - T39;
|
Chris@82
|
1676 T3d = T3b + T3c;
|
Chris@82
|
1677 T3e = KP707106781 * (T3a - T3d);
|
Chris@82
|
1678 T8G = KP707106781 * (T3a + T3d);
|
Chris@82
|
1679 }
|
Chris@82
|
1680 }
|
Chris@82
|
1681 }
|
Chris@82
|
1682 {
|
Chris@82
|
1683 E Ti, T3j, T2l, T3h, Tl, T3g, T2o, T3k, Tp, T3q, T2s, T3o, Ts, T3n, T2v;
|
Chris@82
|
1684 E T3r;
|
Chris@82
|
1685 {
|
Chris@82
|
1686 E Tg, Th, T2j, T2k;
|
Chris@82
|
1687 Tg = ri[WS(is, 4)];
|
Chris@82
|
1688 Th = ri[WS(is, 36)];
|
Chris@82
|
1689 Ti = Tg + Th;
|
Chris@82
|
1690 T3j = Tg - Th;
|
Chris@82
|
1691 T2j = ii[WS(is, 4)];
|
Chris@82
|
1692 T2k = ii[WS(is, 36)];
|
Chris@82
|
1693 T2l = T2j + T2k;
|
Chris@82
|
1694 T3h = T2j - T2k;
|
Chris@82
|
1695 }
|
Chris@82
|
1696 {
|
Chris@82
|
1697 E Tj, Tk, T2m, T2n;
|
Chris@82
|
1698 Tj = ri[WS(is, 20)];
|
Chris@82
|
1699 Tk = ri[WS(is, 52)];
|
Chris@82
|
1700 Tl = Tj + Tk;
|
Chris@82
|
1701 T3g = Tj - Tk;
|
Chris@82
|
1702 T2m = ii[WS(is, 20)];
|
Chris@82
|
1703 T2n = ii[WS(is, 52)];
|
Chris@82
|
1704 T2o = T2m + T2n;
|
Chris@82
|
1705 T3k = T2m - T2n;
|
Chris@82
|
1706 }
|
Chris@82
|
1707 {
|
Chris@82
|
1708 E Tn, To, T2q, T2r;
|
Chris@82
|
1709 Tn = ri[WS(is, 60)];
|
Chris@82
|
1710 To = ri[WS(is, 28)];
|
Chris@82
|
1711 Tp = Tn + To;
|
Chris@82
|
1712 T3q = Tn - To;
|
Chris@82
|
1713 T2q = ii[WS(is, 60)];
|
Chris@82
|
1714 T2r = ii[WS(is, 28)];
|
Chris@82
|
1715 T2s = T2q + T2r;
|
Chris@82
|
1716 T3o = T2q - T2r;
|
Chris@82
|
1717 }
|
Chris@82
|
1718 {
|
Chris@82
|
1719 E Tq, Tr, T2t, T2u;
|
Chris@82
|
1720 Tq = ri[WS(is, 12)];
|
Chris@82
|
1721 Tr = ri[WS(is, 44)];
|
Chris@82
|
1722 Ts = Tq + Tr;
|
Chris@82
|
1723 T3n = Tq - Tr;
|
Chris@82
|
1724 T2t = ii[WS(is, 12)];
|
Chris@82
|
1725 T2u = ii[WS(is, 44)];
|
Chris@82
|
1726 T2v = T2t + T2u;
|
Chris@82
|
1727 T3r = T2t - T2u;
|
Chris@82
|
1728 }
|
Chris@82
|
1729 {
|
Chris@82
|
1730 E Tm, Tt, Tai, Taj;
|
Chris@82
|
1731 Tm = Ti + Tl;
|
Chris@82
|
1732 Tt = Tp + Ts;
|
Chris@82
|
1733 Tu = Tm + Tt;
|
Chris@82
|
1734 TdI = Tt - Tm;
|
Chris@82
|
1735 Tai = T2l - T2o;
|
Chris@82
|
1736 Taj = Ti - Tl;
|
Chris@82
|
1737 Tak = Tai - Taj;
|
Chris@82
|
1738 TbD = Taj + Tai;
|
Chris@82
|
1739 }
|
Chris@82
|
1740 {
|
Chris@82
|
1741 E Tal, Tam, T2p, T2w;
|
Chris@82
|
1742 Tal = Tp - Ts;
|
Chris@82
|
1743 Tam = T2s - T2v;
|
Chris@82
|
1744 Tan = Tal + Tam;
|
Chris@82
|
1745 TbC = Tal - Tam;
|
Chris@82
|
1746 T2p = T2l + T2o;
|
Chris@82
|
1747 T2w = T2s + T2v;
|
Chris@82
|
1748 T2x = T2p + T2w;
|
Chris@82
|
1749 Tda = T2p - T2w;
|
Chris@82
|
1750 }
|
Chris@82
|
1751 {
|
Chris@82
|
1752 E T3i, T3l, T7E, T7F;
|
Chris@82
|
1753 T3i = T3g + T3h;
|
Chris@82
|
1754 T3l = T3j - T3k;
|
Chris@82
|
1755 T3m = FNMS(KP923879532, T3l, KP382683432 * T3i);
|
Chris@82
|
1756 T65 = FMA(KP923879532, T3i, KP382683432 * T3l);
|
Chris@82
|
1757 T7E = T3h - T3g;
|
Chris@82
|
1758 T7F = T3j + T3k;
|
Chris@82
|
1759 T7G = FNMS(KP382683432, T7F, KP923879532 * T7E);
|
Chris@82
|
1760 T8J = FMA(KP382683432, T7E, KP923879532 * T7F);
|
Chris@82
|
1761 }
|
Chris@82
|
1762 {
|
Chris@82
|
1763 E T7H, T7I, T3p, T3s;
|
Chris@82
|
1764 T7H = T3o - T3n;
|
Chris@82
|
1765 T7I = T3q + T3r;
|
Chris@82
|
1766 T7J = FMA(KP923879532, T7H, KP382683432 * T7I);
|
Chris@82
|
1767 T8I = FNMS(KP382683432, T7H, KP923879532 * T7I);
|
Chris@82
|
1768 T3p = T3n + T3o;
|
Chris@82
|
1769 T3s = T3q - T3r;
|
Chris@82
|
1770 T3t = FMA(KP382683432, T3p, KP923879532 * T3s);
|
Chris@82
|
1771 T64 = FNMS(KP923879532, T3p, KP382683432 * T3s);
|
Chris@82
|
1772 }
|
Chris@82
|
1773 }
|
Chris@82
|
1774 {
|
Chris@82
|
1775 E Ty, T3H, T2B, T3x, TB, T3w, T2E, T3I, TI, T3L, T2L, T3B, TF, T3K, T2I;
|
Chris@82
|
1776 E T3E;
|
Chris@82
|
1777 {
|
Chris@82
|
1778 E Tw, Tx, T2C, T2D;
|
Chris@82
|
1779 Tw = ri[WS(is, 2)];
|
Chris@82
|
1780 Tx = ri[WS(is, 34)];
|
Chris@82
|
1781 Ty = Tw + Tx;
|
Chris@82
|
1782 T3H = Tw - Tx;
|
Chris@82
|
1783 {
|
Chris@82
|
1784 E T2z, T2A, Tz, TA;
|
Chris@82
|
1785 T2z = ii[WS(is, 2)];
|
Chris@82
|
1786 T2A = ii[WS(is, 34)];
|
Chris@82
|
1787 T2B = T2z + T2A;
|
Chris@82
|
1788 T3x = T2z - T2A;
|
Chris@82
|
1789 Tz = ri[WS(is, 18)];
|
Chris@82
|
1790 TA = ri[WS(is, 50)];
|
Chris@82
|
1791 TB = Tz + TA;
|
Chris@82
|
1792 T3w = Tz - TA;
|
Chris@82
|
1793 }
|
Chris@82
|
1794 T2C = ii[WS(is, 18)];
|
Chris@82
|
1795 T2D = ii[WS(is, 50)];
|
Chris@82
|
1796 T2E = T2C + T2D;
|
Chris@82
|
1797 T3I = T2C - T2D;
|
Chris@82
|
1798 {
|
Chris@82
|
1799 E TG, TH, T3z, T2J, T2K, T3A;
|
Chris@82
|
1800 TG = ri[WS(is, 58)];
|
Chris@82
|
1801 TH = ri[WS(is, 26)];
|
Chris@82
|
1802 T3z = TG - TH;
|
Chris@82
|
1803 T2J = ii[WS(is, 58)];
|
Chris@82
|
1804 T2K = ii[WS(is, 26)];
|
Chris@82
|
1805 T3A = T2J - T2K;
|
Chris@82
|
1806 TI = TG + TH;
|
Chris@82
|
1807 T3L = T3z + T3A;
|
Chris@82
|
1808 T2L = T2J + T2K;
|
Chris@82
|
1809 T3B = T3z - T3A;
|
Chris@82
|
1810 }
|
Chris@82
|
1811 {
|
Chris@82
|
1812 E TD, TE, T3C, T2G, T2H, T3D;
|
Chris@82
|
1813 TD = ri[WS(is, 10)];
|
Chris@82
|
1814 TE = ri[WS(is, 42)];
|
Chris@82
|
1815 T3C = TD - TE;
|
Chris@82
|
1816 T2G = ii[WS(is, 10)];
|
Chris@82
|
1817 T2H = ii[WS(is, 42)];
|
Chris@82
|
1818 T3D = T2G - T2H;
|
Chris@82
|
1819 TF = TD + TE;
|
Chris@82
|
1820 T3K = T3D - T3C;
|
Chris@82
|
1821 T2I = T2G + T2H;
|
Chris@82
|
1822 T3E = T3C + T3D;
|
Chris@82
|
1823 }
|
Chris@82
|
1824 }
|
Chris@82
|
1825 {
|
Chris@82
|
1826 E TC, TJ, Taq, Tar;
|
Chris@82
|
1827 TC = Ty + TB;
|
Chris@82
|
1828 TJ = TF + TI;
|
Chris@82
|
1829 TK = TC + TJ;
|
Chris@82
|
1830 Tdd = TC - TJ;
|
Chris@82
|
1831 Taq = T2B - T2E;
|
Chris@82
|
1832 Tar = TI - TF;
|
Chris@82
|
1833 Tas = Taq - Tar;
|
Chris@82
|
1834 Tce = Tar + Taq;
|
Chris@82
|
1835 }
|
Chris@82
|
1836 {
|
Chris@82
|
1837 E Tat, Tau, T2F, T2M;
|
Chris@82
|
1838 Tat = Ty - TB;
|
Chris@82
|
1839 Tau = T2I - T2L;
|
Chris@82
|
1840 Tav = Tat - Tau;
|
Chris@82
|
1841 Tcf = Tat + Tau;
|
Chris@82
|
1842 T2F = T2B + T2E;
|
Chris@82
|
1843 T2M = T2I + T2L;
|
Chris@82
|
1844 T2N = T2F + T2M;
|
Chris@82
|
1845 Tdc = T2F - T2M;
|
Chris@82
|
1846 }
|
Chris@82
|
1847 {
|
Chris@82
|
1848 E T3y, T3F, T7M, T7N;
|
Chris@82
|
1849 T3y = T3w + T3x;
|
Chris@82
|
1850 T3F = KP707106781 * (T3B - T3E);
|
Chris@82
|
1851 T3G = T3y - T3F;
|
Chris@82
|
1852 T6G = T3y + T3F;
|
Chris@82
|
1853 T7M = T3x - T3w;
|
Chris@82
|
1854 T7N = KP707106781 * (T3K + T3L);
|
Chris@82
|
1855 T7O = T7M - T7N;
|
Chris@82
|
1856 T9k = T7M + T7N;
|
Chris@82
|
1857 }
|
Chris@82
|
1858 {
|
Chris@82
|
1859 E T7P, T7Q, T3J, T3M;
|
Chris@82
|
1860 T7P = T3H + T3I;
|
Chris@82
|
1861 T7Q = KP707106781 * (T3E + T3B);
|
Chris@82
|
1862 T7R = T7P - T7Q;
|
Chris@82
|
1863 T9l = T7P + T7Q;
|
Chris@82
|
1864 T3J = T3H - T3I;
|
Chris@82
|
1865 T3M = KP707106781 * (T3K - T3L);
|
Chris@82
|
1866 T3N = T3J - T3M;
|
Chris@82
|
1867 T6H = T3J + T3M;
|
Chris@82
|
1868 }
|
Chris@82
|
1869 }
|
Chris@82
|
1870 {
|
Chris@82
|
1871 E T1z, T53, T5L, Tbo, T1C, T5I, T56, Tbp, T1J, Tb9, T5h, T5N, T1G, Tb8, T5c;
|
Chris@82
|
1872 E T5O;
|
Chris@82
|
1873 {
|
Chris@82
|
1874 E T1x, T1y, T54, T55;
|
Chris@82
|
1875 T1x = ri[WS(is, 63)];
|
Chris@82
|
1876 T1y = ri[WS(is, 31)];
|
Chris@82
|
1877 T1z = T1x + T1y;
|
Chris@82
|
1878 T53 = T1x - T1y;
|
Chris@82
|
1879 {
|
Chris@82
|
1880 E T5J, T5K, T1A, T1B;
|
Chris@82
|
1881 T5J = ii[WS(is, 63)];
|
Chris@82
|
1882 T5K = ii[WS(is, 31)];
|
Chris@82
|
1883 T5L = T5J - T5K;
|
Chris@82
|
1884 Tbo = T5J + T5K;
|
Chris@82
|
1885 T1A = ri[WS(is, 15)];
|
Chris@82
|
1886 T1B = ri[WS(is, 47)];
|
Chris@82
|
1887 T1C = T1A + T1B;
|
Chris@82
|
1888 T5I = T1A - T1B;
|
Chris@82
|
1889 }
|
Chris@82
|
1890 T54 = ii[WS(is, 15)];
|
Chris@82
|
1891 T55 = ii[WS(is, 47)];
|
Chris@82
|
1892 T56 = T54 - T55;
|
Chris@82
|
1893 Tbp = T54 + T55;
|
Chris@82
|
1894 {
|
Chris@82
|
1895 E T1H, T1I, T5d, T5e, T5f, T5g;
|
Chris@82
|
1896 T1H = ri[WS(is, 55)];
|
Chris@82
|
1897 T1I = ri[WS(is, 23)];
|
Chris@82
|
1898 T5d = T1H - T1I;
|
Chris@82
|
1899 T5e = ii[WS(is, 55)];
|
Chris@82
|
1900 T5f = ii[WS(is, 23)];
|
Chris@82
|
1901 T5g = T5e - T5f;
|
Chris@82
|
1902 T1J = T1H + T1I;
|
Chris@82
|
1903 Tb9 = T5e + T5f;
|
Chris@82
|
1904 T5h = T5d + T5g;
|
Chris@82
|
1905 T5N = T5d - T5g;
|
Chris@82
|
1906 }
|
Chris@82
|
1907 {
|
Chris@82
|
1908 E T1E, T1F, T5b, T58, T59, T5a;
|
Chris@82
|
1909 T1E = ri[WS(is, 7)];
|
Chris@82
|
1910 T1F = ri[WS(is, 39)];
|
Chris@82
|
1911 T5b = T1E - T1F;
|
Chris@82
|
1912 T58 = ii[WS(is, 7)];
|
Chris@82
|
1913 T59 = ii[WS(is, 39)];
|
Chris@82
|
1914 T5a = T58 - T59;
|
Chris@82
|
1915 T1G = T1E + T1F;
|
Chris@82
|
1916 Tb8 = T58 + T59;
|
Chris@82
|
1917 T5c = T5a - T5b;
|
Chris@82
|
1918 T5O = T5b + T5a;
|
Chris@82
|
1919 }
|
Chris@82
|
1920 }
|
Chris@82
|
1921 {
|
Chris@82
|
1922 E T1D, T1K, Tbq, Tbr;
|
Chris@82
|
1923 T1D = T1z + T1C;
|
Chris@82
|
1924 T1K = T1G + T1J;
|
Chris@82
|
1925 T1L = T1D + T1K;
|
Chris@82
|
1926 Tdv = T1D - T1K;
|
Chris@82
|
1927 Tbq = Tbo - Tbp;
|
Chris@82
|
1928 Tbr = T1J - T1G;
|
Chris@82
|
1929 Tbs = Tbq - Tbr;
|
Chris@82
|
1930 Tcw = Tbr + Tbq;
|
Chris@82
|
1931 }
|
Chris@82
|
1932 {
|
Chris@82
|
1933 E TdA, TdB, T57, T5i;
|
Chris@82
|
1934 TdA = Tbo + Tbp;
|
Chris@82
|
1935 TdB = Tb8 + Tb9;
|
Chris@82
|
1936 TdC = TdA - TdB;
|
Chris@82
|
1937 Teo = TdA + TdB;
|
Chris@82
|
1938 T57 = T53 - T56;
|
Chris@82
|
1939 T5i = KP707106781 * (T5c - T5h);
|
Chris@82
|
1940 T5j = T57 - T5i;
|
Chris@82
|
1941 T6V = T57 + T5i;
|
Chris@82
|
1942 }
|
Chris@82
|
1943 {
|
Chris@82
|
1944 E T5M, T5P, T8w, T8x;
|
Chris@82
|
1945 T5M = T5I + T5L;
|
Chris@82
|
1946 T5P = KP707106781 * (T5N - T5O);
|
Chris@82
|
1947 T5Q = T5M - T5P;
|
Chris@82
|
1948 T6Y = T5M + T5P;
|
Chris@82
|
1949 T8w = T5L - T5I;
|
Chris@82
|
1950 T8x = KP707106781 * (T5c + T5h);
|
Chris@82
|
1951 T8y = T8w - T8x;
|
Chris@82
|
1952 T9C = T8w + T8x;
|
Chris@82
|
1953 }
|
Chris@82
|
1954 {
|
Chris@82
|
1955 E Tb7, Tba, T8l, T8m;
|
Chris@82
|
1956 Tb7 = T1z - T1C;
|
Chris@82
|
1957 Tba = Tb8 - Tb9;
|
Chris@82
|
1958 Tbb = Tb7 - Tba;
|
Chris@82
|
1959 Tct = Tb7 + Tba;
|
Chris@82
|
1960 T8l = T53 + T56;
|
Chris@82
|
1961 T8m = KP707106781 * (T5O + T5N);
|
Chris@82
|
1962 T8n = T8l - T8m;
|
Chris@82
|
1963 T9z = T8l + T8m;
|
Chris@82
|
1964 }
|
Chris@82
|
1965 }
|
Chris@82
|
1966 {
|
Chris@82
|
1967 E TN, T40, T2Q, T3Q, TQ, T3P, T2T, T41, TX, T44, T30, T3U, TU, T43, T2X;
|
Chris@82
|
1968 E T3X;
|
Chris@82
|
1969 {
|
Chris@82
|
1970 E TL, TM, T2R, T2S;
|
Chris@82
|
1971 TL = ri[WS(is, 62)];
|
Chris@82
|
1972 TM = ri[WS(is, 30)];
|
Chris@82
|
1973 TN = TL + TM;
|
Chris@82
|
1974 T40 = TL - TM;
|
Chris@82
|
1975 {
|
Chris@82
|
1976 E T2O, T2P, TO, TP;
|
Chris@82
|
1977 T2O = ii[WS(is, 62)];
|
Chris@82
|
1978 T2P = ii[WS(is, 30)];
|
Chris@82
|
1979 T2Q = T2O + T2P;
|
Chris@82
|
1980 T3Q = T2O - T2P;
|
Chris@82
|
1981 TO = ri[WS(is, 14)];
|
Chris@82
|
1982 TP = ri[WS(is, 46)];
|
Chris@82
|
1983 TQ = TO + TP;
|
Chris@82
|
1984 T3P = TO - TP;
|
Chris@82
|
1985 }
|
Chris@82
|
1986 T2R = ii[WS(is, 14)];
|
Chris@82
|
1987 T2S = ii[WS(is, 46)];
|
Chris@82
|
1988 T2T = T2R + T2S;
|
Chris@82
|
1989 T41 = T2R - T2S;
|
Chris@82
|
1990 {
|
Chris@82
|
1991 E TV, TW, T3S, T2Y, T2Z, T3T;
|
Chris@82
|
1992 TV = ri[WS(is, 54)];
|
Chris@82
|
1993 TW = ri[WS(is, 22)];
|
Chris@82
|
1994 T3S = TV - TW;
|
Chris@82
|
1995 T2Y = ii[WS(is, 54)];
|
Chris@82
|
1996 T2Z = ii[WS(is, 22)];
|
Chris@82
|
1997 T3T = T2Y - T2Z;
|
Chris@82
|
1998 TX = TV + TW;
|
Chris@82
|
1999 T44 = T3S + T3T;
|
Chris@82
|
2000 T30 = T2Y + T2Z;
|
Chris@82
|
2001 T3U = T3S - T3T;
|
Chris@82
|
2002 }
|
Chris@82
|
2003 {
|
Chris@82
|
2004 E TS, TT, T3V, T2V, T2W, T3W;
|
Chris@82
|
2005 TS = ri[WS(is, 6)];
|
Chris@82
|
2006 TT = ri[WS(is, 38)];
|
Chris@82
|
2007 T3V = TS - TT;
|
Chris@82
|
2008 T2V = ii[WS(is, 6)];
|
Chris@82
|
2009 T2W = ii[WS(is, 38)];
|
Chris@82
|
2010 T3W = T2V - T2W;
|
Chris@82
|
2011 TU = TS + TT;
|
Chris@82
|
2012 T43 = T3W - T3V;
|
Chris@82
|
2013 T2X = T2V + T2W;
|
Chris@82
|
2014 T3X = T3V + T3W;
|
Chris@82
|
2015 }
|
Chris@82
|
2016 }
|
Chris@82
|
2017 {
|
Chris@82
|
2018 E TR, TY, Tax, Tay;
|
Chris@82
|
2019 TR = TN + TQ;
|
Chris@82
|
2020 TY = TU + TX;
|
Chris@82
|
2021 TZ = TR + TY;
|
Chris@82
|
2022 Tdf = TR - TY;
|
Chris@82
|
2023 Tax = T2Q - T2T;
|
Chris@82
|
2024 Tay = TX - TU;
|
Chris@82
|
2025 Taz = Tax - Tay;
|
Chris@82
|
2026 Tch = Tay + Tax;
|
Chris@82
|
2027 }
|
Chris@82
|
2028 {
|
Chris@82
|
2029 E TaA, TaB, T2U, T31;
|
Chris@82
|
2030 TaA = TN - TQ;
|
Chris@82
|
2031 TaB = T2X - T30;
|
Chris@82
|
2032 TaC = TaA - TaB;
|
Chris@82
|
2033 Tci = TaA + TaB;
|
Chris@82
|
2034 T2U = T2Q + T2T;
|
Chris@82
|
2035 T31 = T2X + T30;
|
Chris@82
|
2036 T32 = T2U + T31;
|
Chris@82
|
2037 Tdg = T2U - T31;
|
Chris@82
|
2038 }
|
Chris@82
|
2039 {
|
Chris@82
|
2040 E T3R, T3Y, T7T, T7U;
|
Chris@82
|
2041 T3R = T3P + T3Q;
|
Chris@82
|
2042 T3Y = KP707106781 * (T3U - T3X);
|
Chris@82
|
2043 T3Z = T3R - T3Y;
|
Chris@82
|
2044 T6J = T3R + T3Y;
|
Chris@82
|
2045 T7T = T40 + T41;
|
Chris@82
|
2046 T7U = KP707106781 * (T3X + T3U);
|
Chris@82
|
2047 T7V = T7T - T7U;
|
Chris@82
|
2048 T9n = T7T + T7U;
|
Chris@82
|
2049 }
|
Chris@82
|
2050 {
|
Chris@82
|
2051 E T7W, T7X, T42, T45;
|
Chris@82
|
2052 T7W = T3Q - T3P;
|
Chris@82
|
2053 T7X = KP707106781 * (T43 + T44);
|
Chris@82
|
2054 T7Y = T7W - T7X;
|
Chris@82
|
2055 T9o = T7W + T7X;
|
Chris@82
|
2056 T42 = T40 - T41;
|
Chris@82
|
2057 T45 = KP707106781 * (T43 - T44);
|
Chris@82
|
2058 T46 = T42 - T45;
|
Chris@82
|
2059 T6K = T42 + T45;
|
Chris@82
|
2060 }
|
Chris@82
|
2061 }
|
Chris@82
|
2062 {
|
Chris@82
|
2063 E T14, T4P, T4d, TaG, T17, T4a, T4S, TaH, T1e, TaZ, T4j, T4V, T1b, TaY, T4o;
|
Chris@82
|
2064 E T4U;
|
Chris@82
|
2065 {
|
Chris@82
|
2066 E T12, T13, T4Q, T4R;
|
Chris@82
|
2067 T12 = ri[WS(is, 1)];
|
Chris@82
|
2068 T13 = ri[WS(is, 33)];
|
Chris@82
|
2069 T14 = T12 + T13;
|
Chris@82
|
2070 T4P = T12 - T13;
|
Chris@82
|
2071 {
|
Chris@82
|
2072 E T4b, T4c, T15, T16;
|
Chris@82
|
2073 T4b = ii[WS(is, 1)];
|
Chris@82
|
2074 T4c = ii[WS(is, 33)];
|
Chris@82
|
2075 T4d = T4b - T4c;
|
Chris@82
|
2076 TaG = T4b + T4c;
|
Chris@82
|
2077 T15 = ri[WS(is, 17)];
|
Chris@82
|
2078 T16 = ri[WS(is, 49)];
|
Chris@82
|
2079 T17 = T15 + T16;
|
Chris@82
|
2080 T4a = T15 - T16;
|
Chris@82
|
2081 }
|
Chris@82
|
2082 T4Q = ii[WS(is, 17)];
|
Chris@82
|
2083 T4R = ii[WS(is, 49)];
|
Chris@82
|
2084 T4S = T4Q - T4R;
|
Chris@82
|
2085 TaH = T4Q + T4R;
|
Chris@82
|
2086 {
|
Chris@82
|
2087 E T1c, T1d, T4f, T4g, T4h, T4i;
|
Chris@82
|
2088 T1c = ri[WS(is, 57)];
|
Chris@82
|
2089 T1d = ri[WS(is, 25)];
|
Chris@82
|
2090 T4f = T1c - T1d;
|
Chris@82
|
2091 T4g = ii[WS(is, 57)];
|
Chris@82
|
2092 T4h = ii[WS(is, 25)];
|
Chris@82
|
2093 T4i = T4g - T4h;
|
Chris@82
|
2094 T1e = T1c + T1d;
|
Chris@82
|
2095 TaZ = T4g + T4h;
|
Chris@82
|
2096 T4j = T4f - T4i;
|
Chris@82
|
2097 T4V = T4f + T4i;
|
Chris@82
|
2098 }
|
Chris@82
|
2099 {
|
Chris@82
|
2100 E T19, T1a, T4k, T4l, T4m, T4n;
|
Chris@82
|
2101 T19 = ri[WS(is, 9)];
|
Chris@82
|
2102 T1a = ri[WS(is, 41)];
|
Chris@82
|
2103 T4k = T19 - T1a;
|
Chris@82
|
2104 T4l = ii[WS(is, 9)];
|
Chris@82
|
2105 T4m = ii[WS(is, 41)];
|
Chris@82
|
2106 T4n = T4l - T4m;
|
Chris@82
|
2107 T1b = T19 + T1a;
|
Chris@82
|
2108 TaY = T4l + T4m;
|
Chris@82
|
2109 T4o = T4k + T4n;
|
Chris@82
|
2110 T4U = T4n - T4k;
|
Chris@82
|
2111 }
|
Chris@82
|
2112 }
|
Chris@82
|
2113 {
|
Chris@82
|
2114 E T18, T1f, TaX, Tb0;
|
Chris@82
|
2115 T18 = T14 + T17;
|
Chris@82
|
2116 T1f = T1b + T1e;
|
Chris@82
|
2117 T1g = T18 + T1f;
|
Chris@82
|
2118 Tdp = T18 - T1f;
|
Chris@82
|
2119 TaX = T14 - T17;
|
Chris@82
|
2120 Tb0 = TaY - TaZ;
|
Chris@82
|
2121 Tb1 = TaX - Tb0;
|
Chris@82
|
2122 Tcm = TaX + Tb0;
|
Chris@82
|
2123 }
|
Chris@82
|
2124 {
|
Chris@82
|
2125 E Tdk, Tdl, T4e, T4p;
|
Chris@82
|
2126 Tdk = TaG + TaH;
|
Chris@82
|
2127 Tdl = TaY + TaZ;
|
Chris@82
|
2128 Tdm = Tdk - Tdl;
|
Chris@82
|
2129 Tej = Tdk + Tdl;
|
Chris@82
|
2130 T4e = T4a + T4d;
|
Chris@82
|
2131 T4p = KP707106781 * (T4j - T4o);
|
Chris@82
|
2132 T4q = T4e - T4p;
|
Chris@82
|
2133 T6R = T4e + T4p;
|
Chris@82
|
2134 }
|
Chris@82
|
2135 {
|
Chris@82
|
2136 E T4T, T4W, T8d, T8e;
|
Chris@82
|
2137 T4T = T4P - T4S;
|
Chris@82
|
2138 T4W = KP707106781 * (T4U - T4V);
|
Chris@82
|
2139 T4X = T4T - T4W;
|
Chris@82
|
2140 T6O = T4T + T4W;
|
Chris@82
|
2141 T8d = T4P + T4S;
|
Chris@82
|
2142 T8e = KP707106781 * (T4o + T4j);
|
Chris@82
|
2143 T8f = T8d - T8e;
|
Chris@82
|
2144 T9s = T8d + T8e;
|
Chris@82
|
2145 }
|
Chris@82
|
2146 {
|
Chris@82
|
2147 E TaI, TaJ, T82, T83;
|
Chris@82
|
2148 TaI = TaG - TaH;
|
Chris@82
|
2149 TaJ = T1e - T1b;
|
Chris@82
|
2150 TaK = TaI - TaJ;
|
Chris@82
|
2151 Tcp = TaJ + TaI;
|
Chris@82
|
2152 T82 = T4d - T4a;
|
Chris@82
|
2153 T83 = KP707106781 * (T4U + T4V);
|
Chris@82
|
2154 T84 = T82 - T83;
|
Chris@82
|
2155 T9v = T82 + T83;
|
Chris@82
|
2156 }
|
Chris@82
|
2157 }
|
Chris@82
|
2158 {
|
Chris@82
|
2159 E T1j, TaR, T1m, TaS, T4G, T4L, TaT, TaQ, T89, T88, T1q, TaM, T1t, TaN, T4v;
|
Chris@82
|
2160 E T4A, TaO, TaL, T86, T85;
|
Chris@82
|
2161 {
|
Chris@82
|
2162 E T4H, T4F, T4C, T4K;
|
Chris@82
|
2163 {
|
Chris@82
|
2164 E T1h, T1i, T4D, T4E;
|
Chris@82
|
2165 T1h = ri[WS(is, 5)];
|
Chris@82
|
2166 T1i = ri[WS(is, 37)];
|
Chris@82
|
2167 T1j = T1h + T1i;
|
Chris@82
|
2168 T4H = T1h - T1i;
|
Chris@82
|
2169 T4D = ii[WS(is, 5)];
|
Chris@82
|
2170 T4E = ii[WS(is, 37)];
|
Chris@82
|
2171 T4F = T4D - T4E;
|
Chris@82
|
2172 TaR = T4D + T4E;
|
Chris@82
|
2173 }
|
Chris@82
|
2174 {
|
Chris@82
|
2175 E T1k, T1l, T4I, T4J;
|
Chris@82
|
2176 T1k = ri[WS(is, 21)];
|
Chris@82
|
2177 T1l = ri[WS(is, 53)];
|
Chris@82
|
2178 T1m = T1k + T1l;
|
Chris@82
|
2179 T4C = T1k - T1l;
|
Chris@82
|
2180 T4I = ii[WS(is, 21)];
|
Chris@82
|
2181 T4J = ii[WS(is, 53)];
|
Chris@82
|
2182 T4K = T4I - T4J;
|
Chris@82
|
2183 TaS = T4I + T4J;
|
Chris@82
|
2184 }
|
Chris@82
|
2185 T4G = T4C + T4F;
|
Chris@82
|
2186 T4L = T4H - T4K;
|
Chris@82
|
2187 TaT = TaR - TaS;
|
Chris@82
|
2188 TaQ = T1j - T1m;
|
Chris@82
|
2189 T89 = T4H + T4K;
|
Chris@82
|
2190 T88 = T4F - T4C;
|
Chris@82
|
2191 }
|
Chris@82
|
2192 {
|
Chris@82
|
2193 E T4r, T4z, T4w, T4u;
|
Chris@82
|
2194 {
|
Chris@82
|
2195 E T1o, T1p, T4x, T4y;
|
Chris@82
|
2196 T1o = ri[WS(is, 61)];
|
Chris@82
|
2197 T1p = ri[WS(is, 29)];
|
Chris@82
|
2198 T1q = T1o + T1p;
|
Chris@82
|
2199 T4r = T1o - T1p;
|
Chris@82
|
2200 T4x = ii[WS(is, 61)];
|
Chris@82
|
2201 T4y = ii[WS(is, 29)];
|
Chris@82
|
2202 T4z = T4x - T4y;
|
Chris@82
|
2203 TaM = T4x + T4y;
|
Chris@82
|
2204 }
|
Chris@82
|
2205 {
|
Chris@82
|
2206 E T1r, T1s, T4s, T4t;
|
Chris@82
|
2207 T1r = ri[WS(is, 13)];
|
Chris@82
|
2208 T1s = ri[WS(is, 45)];
|
Chris@82
|
2209 T1t = T1r + T1s;
|
Chris@82
|
2210 T4w = T1r - T1s;
|
Chris@82
|
2211 T4s = ii[WS(is, 13)];
|
Chris@82
|
2212 T4t = ii[WS(is, 45)];
|
Chris@82
|
2213 T4u = T4s - T4t;
|
Chris@82
|
2214 TaN = T4s + T4t;
|
Chris@82
|
2215 }
|
Chris@82
|
2216 T4v = T4r - T4u;
|
Chris@82
|
2217 T4A = T4w + T4z;
|
Chris@82
|
2218 TaO = TaM - TaN;
|
Chris@82
|
2219 TaL = T1q - T1t;
|
Chris@82
|
2220 T86 = T4z - T4w;
|
Chris@82
|
2221 T85 = T4r + T4u;
|
Chris@82
|
2222 }
|
Chris@82
|
2223 {
|
Chris@82
|
2224 E T1n, T1u, Tb2, Tb3;
|
Chris@82
|
2225 T1n = T1j + T1m;
|
Chris@82
|
2226 T1u = T1q + T1t;
|
Chris@82
|
2227 T1v = T1n + T1u;
|
Chris@82
|
2228 Tdn = T1u - T1n;
|
Chris@82
|
2229 Tb2 = TaT - TaQ;
|
Chris@82
|
2230 Tb3 = TaL + TaO;
|
Chris@82
|
2231 Tb4 = KP707106781 * (Tb2 - Tb3);
|
Chris@82
|
2232 Tcq = KP707106781 * (Tb2 + Tb3);
|
Chris@82
|
2233 }
|
Chris@82
|
2234 {
|
Chris@82
|
2235 E Tdq, Tdr, T4B, T4M;
|
Chris@82
|
2236 Tdq = TaR + TaS;
|
Chris@82
|
2237 Tdr = TaM + TaN;
|
Chris@82
|
2238 Tds = Tdq - Tdr;
|
Chris@82
|
2239 Tek = Tdq + Tdr;
|
Chris@82
|
2240 T4B = FNMS(KP923879532, T4A, KP382683432 * T4v);
|
Chris@82
|
2241 T4M = FMA(KP923879532, T4G, KP382683432 * T4L);
|
Chris@82
|
2242 T4N = T4B - T4M;
|
Chris@82
|
2243 T6P = T4M + T4B;
|
Chris@82
|
2244 }
|
Chris@82
|
2245 {
|
Chris@82
|
2246 E T4Y, T4Z, T8g, T8h;
|
Chris@82
|
2247 T4Y = FNMS(KP923879532, T4L, KP382683432 * T4G);
|
Chris@82
|
2248 T4Z = FMA(KP382683432, T4A, KP923879532 * T4v);
|
Chris@82
|
2249 T50 = T4Y - T4Z;
|
Chris@82
|
2250 T6S = T4Y + T4Z;
|
Chris@82
|
2251 T8g = FNMS(KP382683432, T89, KP923879532 * T88);
|
Chris@82
|
2252 T8h = FMA(KP923879532, T86, KP382683432 * T85);
|
Chris@82
|
2253 T8i = T8g - T8h;
|
Chris@82
|
2254 T9w = T8g + T8h;
|
Chris@82
|
2255 }
|
Chris@82
|
2256 {
|
Chris@82
|
2257 E TaP, TaU, T87, T8a;
|
Chris@82
|
2258 TaP = TaL - TaO;
|
Chris@82
|
2259 TaU = TaQ + TaT;
|
Chris@82
|
2260 TaV = KP707106781 * (TaP - TaU);
|
Chris@82
|
2261 Tcn = KP707106781 * (TaU + TaP);
|
Chris@82
|
2262 T87 = FNMS(KP382683432, T86, KP923879532 * T85);
|
Chris@82
|
2263 T8a = FMA(KP382683432, T88, KP923879532 * T89);
|
Chris@82
|
2264 T8b = T87 - T8a;
|
Chris@82
|
2265 T9t = T8a + T87;
|
Chris@82
|
2266 }
|
Chris@82
|
2267 }
|
Chris@82
|
2268 {
|
Chris@82
|
2269 E T1O, Tbc, T1R, Tbd, T5o, T5t, Tbf, Tbe, T8p, T8o, T1V, Tbi, T1Y, Tbj, T5z;
|
Chris@82
|
2270 E T5E, Tbk, Tbh, T8s, T8r;
|
Chris@82
|
2271 {
|
Chris@82
|
2272 E T5p, T5n, T5k, T5s;
|
Chris@82
|
2273 {
|
Chris@82
|
2274 E T1M, T1N, T5l, T5m;
|
Chris@82
|
2275 T1M = ri[WS(is, 3)];
|
Chris@82
|
2276 T1N = ri[WS(is, 35)];
|
Chris@82
|
2277 T1O = T1M + T1N;
|
Chris@82
|
2278 T5p = T1M - T1N;
|
Chris@82
|
2279 T5l = ii[WS(is, 3)];
|
Chris@82
|
2280 T5m = ii[WS(is, 35)];
|
Chris@82
|
2281 T5n = T5l - T5m;
|
Chris@82
|
2282 Tbc = T5l + T5m;
|
Chris@82
|
2283 }
|
Chris@82
|
2284 {
|
Chris@82
|
2285 E T1P, T1Q, T5q, T5r;
|
Chris@82
|
2286 T1P = ri[WS(is, 19)];
|
Chris@82
|
2287 T1Q = ri[WS(is, 51)];
|
Chris@82
|
2288 T1R = T1P + T1Q;
|
Chris@82
|
2289 T5k = T1P - T1Q;
|
Chris@82
|
2290 T5q = ii[WS(is, 19)];
|
Chris@82
|
2291 T5r = ii[WS(is, 51)];
|
Chris@82
|
2292 T5s = T5q - T5r;
|
Chris@82
|
2293 Tbd = T5q + T5r;
|
Chris@82
|
2294 }
|
Chris@82
|
2295 T5o = T5k + T5n;
|
Chris@82
|
2296 T5t = T5p - T5s;
|
Chris@82
|
2297 Tbf = T1O - T1R;
|
Chris@82
|
2298 Tbe = Tbc - Tbd;
|
Chris@82
|
2299 T8p = T5p + T5s;
|
Chris@82
|
2300 T8o = T5n - T5k;
|
Chris@82
|
2301 }
|
Chris@82
|
2302 {
|
Chris@82
|
2303 E T5A, T5y, T5v, T5D;
|
Chris@82
|
2304 {
|
Chris@82
|
2305 E T1T, T1U, T5w, T5x;
|
Chris@82
|
2306 T1T = ri[WS(is, 59)];
|
Chris@82
|
2307 T1U = ri[WS(is, 27)];
|
Chris@82
|
2308 T1V = T1T + T1U;
|
Chris@82
|
2309 T5A = T1T - T1U;
|
Chris@82
|
2310 T5w = ii[WS(is, 59)];
|
Chris@82
|
2311 T5x = ii[WS(is, 27)];
|
Chris@82
|
2312 T5y = T5w - T5x;
|
Chris@82
|
2313 Tbi = T5w + T5x;
|
Chris@82
|
2314 }
|
Chris@82
|
2315 {
|
Chris@82
|
2316 E T1W, T1X, T5B, T5C;
|
Chris@82
|
2317 T1W = ri[WS(is, 11)];
|
Chris@82
|
2318 T1X = ri[WS(is, 43)];
|
Chris@82
|
2319 T1Y = T1W + T1X;
|
Chris@82
|
2320 T5v = T1W - T1X;
|
Chris@82
|
2321 T5B = ii[WS(is, 11)];
|
Chris@82
|
2322 T5C = ii[WS(is, 43)];
|
Chris@82
|
2323 T5D = T5B - T5C;
|
Chris@82
|
2324 Tbj = T5B + T5C;
|
Chris@82
|
2325 }
|
Chris@82
|
2326 T5z = T5v + T5y;
|
Chris@82
|
2327 T5E = T5A - T5D;
|
Chris@82
|
2328 Tbk = Tbi - Tbj;
|
Chris@82
|
2329 Tbh = T1V - T1Y;
|
Chris@82
|
2330 T8s = T5A + T5D;
|
Chris@82
|
2331 T8r = T5y - T5v;
|
Chris@82
|
2332 }
|
Chris@82
|
2333 {
|
Chris@82
|
2334 E T1S, T1Z, Tbt, Tbu;
|
Chris@82
|
2335 T1S = T1O + T1R;
|
Chris@82
|
2336 T1Z = T1V + T1Y;
|
Chris@82
|
2337 T20 = T1S + T1Z;
|
Chris@82
|
2338 TdD = T1Z - T1S;
|
Chris@82
|
2339 Tbt = Tbh - Tbk;
|
Chris@82
|
2340 Tbu = Tbf + Tbe;
|
Chris@82
|
2341 Tbv = KP707106781 * (Tbt - Tbu);
|
Chris@82
|
2342 Tcu = KP707106781 * (Tbu + Tbt);
|
Chris@82
|
2343 }
|
Chris@82
|
2344 {
|
Chris@82
|
2345 E Tdw, Tdx, T5u, T5F;
|
Chris@82
|
2346 Tdw = Tbc + Tbd;
|
Chris@82
|
2347 Tdx = Tbi + Tbj;
|
Chris@82
|
2348 Tdy = Tdw - Tdx;
|
Chris@82
|
2349 Tep = Tdw + Tdx;
|
Chris@82
|
2350 T5u = FNMS(KP923879532, T5t, KP382683432 * T5o);
|
Chris@82
|
2351 T5F = FMA(KP382683432, T5z, KP923879532 * T5E);
|
Chris@82
|
2352 T5G = T5u - T5F;
|
Chris@82
|
2353 T6Z = T5u + T5F;
|
Chris@82
|
2354 }
|
Chris@82
|
2355 {
|
Chris@82
|
2356 E T5R, T5S, T8z, T8A;
|
Chris@82
|
2357 T5R = FNMS(KP923879532, T5z, KP382683432 * T5E);
|
Chris@82
|
2358 T5S = FMA(KP923879532, T5o, KP382683432 * T5t);
|
Chris@82
|
2359 T5T = T5R - T5S;
|
Chris@82
|
2360 T6W = T5S + T5R;
|
Chris@82
|
2361 T8z = FNMS(KP382683432, T8r, KP923879532 * T8s);
|
Chris@82
|
2362 T8A = FMA(KP382683432, T8o, KP923879532 * T8p);
|
Chris@82
|
2363 T8B = T8z - T8A;
|
Chris@82
|
2364 T9A = T8A + T8z;
|
Chris@82
|
2365 }
|
Chris@82
|
2366 {
|
Chris@82
|
2367 E Tbg, Tbl, T8q, T8t;
|
Chris@82
|
2368 Tbg = Tbe - Tbf;
|
Chris@82
|
2369 Tbl = Tbh + Tbk;
|
Chris@82
|
2370 Tbm = KP707106781 * (Tbg - Tbl);
|
Chris@82
|
2371 Tcx = KP707106781 * (Tbg + Tbl);
|
Chris@82
|
2372 T8q = FNMS(KP382683432, T8p, KP923879532 * T8o);
|
Chris@82
|
2373 T8t = FMA(KP923879532, T8r, KP382683432 * T8s);
|
Chris@82
|
2374 T8u = T8q - T8t;
|
Chris@82
|
2375 T9D = T8q + T8t;
|
Chris@82
|
2376 }
|
Chris@82
|
2377 }
|
Chris@82
|
2378 {
|
Chris@82
|
2379 E T11, TeD, TeG, TeI, T22, T23, T34, TeH;
|
Chris@82
|
2380 {
|
Chris@82
|
2381 E Tv, T10, TeE, TeF;
|
Chris@82
|
2382 Tv = Tf + Tu;
|
Chris@82
|
2383 T10 = TK + TZ;
|
Chris@82
|
2384 T11 = Tv + T10;
|
Chris@82
|
2385 TeD = Tv - T10;
|
Chris@82
|
2386 TeE = Tej + Tek;
|
Chris@82
|
2387 TeF = Teo + Tep;
|
Chris@82
|
2388 TeG = TeE - TeF;
|
Chris@82
|
2389 TeI = TeE + TeF;
|
Chris@82
|
2390 }
|
Chris@82
|
2391 {
|
Chris@82
|
2392 E T1w, T21, T2y, T33;
|
Chris@82
|
2393 T1w = T1g + T1v;
|
Chris@82
|
2394 T21 = T1L + T20;
|
Chris@82
|
2395 T22 = T1w + T21;
|
Chris@82
|
2396 T23 = T21 - T1w;
|
Chris@82
|
2397 T2y = T2i + T2x;
|
Chris@82
|
2398 T33 = T2N + T32;
|
Chris@82
|
2399 T34 = T2y - T33;
|
Chris@82
|
2400 TeH = T2y + T33;
|
Chris@82
|
2401 }
|
Chris@82
|
2402 ro[WS(os, 32)] = T11 - T22;
|
Chris@82
|
2403 io[WS(os, 32)] = TeH - TeI;
|
Chris@82
|
2404 ro[0] = T11 + T22;
|
Chris@82
|
2405 io[0] = TeH + TeI;
|
Chris@82
|
2406 io[WS(os, 16)] = T23 + T34;
|
Chris@82
|
2407 ro[WS(os, 16)] = TeD + TeG;
|
Chris@82
|
2408 io[WS(os, 48)] = T34 - T23;
|
Chris@82
|
2409 ro[WS(os, 48)] = TeD - TeG;
|
Chris@82
|
2410 }
|
Chris@82
|
2411 {
|
Chris@82
|
2412 E Teh, Tex, Tev, TeB, Tem, Tey, Ter, Tez;
|
Chris@82
|
2413 {
|
Chris@82
|
2414 E Tef, Teg, Tet, Teu;
|
Chris@82
|
2415 Tef = Tf - Tu;
|
Chris@82
|
2416 Teg = T2N - T32;
|
Chris@82
|
2417 Teh = Tef + Teg;
|
Chris@82
|
2418 Tex = Tef - Teg;
|
Chris@82
|
2419 Tet = T2i - T2x;
|
Chris@82
|
2420 Teu = TZ - TK;
|
Chris@82
|
2421 Tev = Tet - Teu;
|
Chris@82
|
2422 TeB = Teu + Tet;
|
Chris@82
|
2423 }
|
Chris@82
|
2424 {
|
Chris@82
|
2425 E Tei, Tel, Ten, Teq;
|
Chris@82
|
2426 Tei = T1g - T1v;
|
Chris@82
|
2427 Tel = Tej - Tek;
|
Chris@82
|
2428 Tem = Tei + Tel;
|
Chris@82
|
2429 Tey = Tel - Tei;
|
Chris@82
|
2430 Ten = T1L - T20;
|
Chris@82
|
2431 Teq = Teo - Tep;
|
Chris@82
|
2432 Ter = Ten - Teq;
|
Chris@82
|
2433 Tez = Ten + Teq;
|
Chris@82
|
2434 }
|
Chris@82
|
2435 {
|
Chris@82
|
2436 E Tes, TeC, Tew, TeA;
|
Chris@82
|
2437 Tes = KP707106781 * (Tem + Ter);
|
Chris@82
|
2438 ro[WS(os, 40)] = Teh - Tes;
|
Chris@82
|
2439 ro[WS(os, 8)] = Teh + Tes;
|
Chris@82
|
2440 TeC = KP707106781 * (Tey + Tez);
|
Chris@82
|
2441 io[WS(os, 40)] = TeB - TeC;
|
Chris@82
|
2442 io[WS(os, 8)] = TeB + TeC;
|
Chris@82
|
2443 Tew = KP707106781 * (Ter - Tem);
|
Chris@82
|
2444 io[WS(os, 56)] = Tev - Tew;
|
Chris@82
|
2445 io[WS(os, 24)] = Tev + Tew;
|
Chris@82
|
2446 TeA = KP707106781 * (Tey - Tez);
|
Chris@82
|
2447 ro[WS(os, 56)] = Tex - TeA;
|
Chris@82
|
2448 ro[WS(os, 24)] = Tex + TeA;
|
Chris@82
|
2449 }
|
Chris@82
|
2450 }
|
Chris@82
|
2451 {
|
Chris@82
|
2452 E Tdb, TdV, Te5, TdJ, Tdi, Te6, Te3, Teb, TdM, TdW, Tdu, TdQ, Te0, Tea, TdF;
|
Chris@82
|
2453 E TdR;
|
Chris@82
|
2454 {
|
Chris@82
|
2455 E Tde, Tdh, Tdo, Tdt;
|
Chris@82
|
2456 Tdb = Td9 - Tda;
|
Chris@82
|
2457 TdV = Td9 + Tda;
|
Chris@82
|
2458 Te5 = TdI + TdH;
|
Chris@82
|
2459 TdJ = TdH - TdI;
|
Chris@82
|
2460 Tde = Tdc - Tdd;
|
Chris@82
|
2461 Tdh = Tdf + Tdg;
|
Chris@82
|
2462 Tdi = KP707106781 * (Tde - Tdh);
|
Chris@82
|
2463 Te6 = KP707106781 * (Tde + Tdh);
|
Chris@82
|
2464 {
|
Chris@82
|
2465 E Te1, Te2, TdK, TdL;
|
Chris@82
|
2466 Te1 = Tdv + Tdy;
|
Chris@82
|
2467 Te2 = TdD + TdC;
|
Chris@82
|
2468 Te3 = FNMS(KP382683432, Te2, KP923879532 * Te1);
|
Chris@82
|
2469 Teb = FMA(KP923879532, Te2, KP382683432 * Te1);
|
Chris@82
|
2470 TdK = Tdf - Tdg;
|
Chris@82
|
2471 TdL = Tdd + Tdc;
|
Chris@82
|
2472 TdM = KP707106781 * (TdK - TdL);
|
Chris@82
|
2473 TdW = KP707106781 * (TdL + TdK);
|
Chris@82
|
2474 }
|
Chris@82
|
2475 Tdo = Tdm - Tdn;
|
Chris@82
|
2476 Tdt = Tdp - Tds;
|
Chris@82
|
2477 Tdu = FMA(KP923879532, Tdo, KP382683432 * Tdt);
|
Chris@82
|
2478 TdQ = FNMS(KP923879532, Tdt, KP382683432 * Tdo);
|
Chris@82
|
2479 {
|
Chris@82
|
2480 E TdY, TdZ, Tdz, TdE;
|
Chris@82
|
2481 TdY = Tdn + Tdm;
|
Chris@82
|
2482 TdZ = Tdp + Tds;
|
Chris@82
|
2483 Te0 = FMA(KP382683432, TdY, KP923879532 * TdZ);
|
Chris@82
|
2484 Tea = FNMS(KP382683432, TdZ, KP923879532 * TdY);
|
Chris@82
|
2485 Tdz = Tdv - Tdy;
|
Chris@82
|
2486 TdE = TdC - TdD;
|
Chris@82
|
2487 TdF = FNMS(KP923879532, TdE, KP382683432 * Tdz);
|
Chris@82
|
2488 TdR = FMA(KP382683432, TdE, KP923879532 * Tdz);
|
Chris@82
|
2489 }
|
Chris@82
|
2490 }
|
Chris@82
|
2491 {
|
Chris@82
|
2492 E Tdj, TdG, TdT, TdU;
|
Chris@82
|
2493 Tdj = Tdb + Tdi;
|
Chris@82
|
2494 TdG = Tdu + TdF;
|
Chris@82
|
2495 ro[WS(os, 44)] = Tdj - TdG;
|
Chris@82
|
2496 ro[WS(os, 12)] = Tdj + TdG;
|
Chris@82
|
2497 TdT = TdJ + TdM;
|
Chris@82
|
2498 TdU = TdQ + TdR;
|
Chris@82
|
2499 io[WS(os, 44)] = TdT - TdU;
|
Chris@82
|
2500 io[WS(os, 12)] = TdT + TdU;
|
Chris@82
|
2501 }
|
Chris@82
|
2502 {
|
Chris@82
|
2503 E TdN, TdO, TdP, TdS;
|
Chris@82
|
2504 TdN = TdJ - TdM;
|
Chris@82
|
2505 TdO = TdF - Tdu;
|
Chris@82
|
2506 io[WS(os, 60)] = TdN - TdO;
|
Chris@82
|
2507 io[WS(os, 28)] = TdN + TdO;
|
Chris@82
|
2508 TdP = Tdb - Tdi;
|
Chris@82
|
2509 TdS = TdQ - TdR;
|
Chris@82
|
2510 ro[WS(os, 60)] = TdP - TdS;
|
Chris@82
|
2511 ro[WS(os, 28)] = TdP + TdS;
|
Chris@82
|
2512 }
|
Chris@82
|
2513 {
|
Chris@82
|
2514 E TdX, Te4, Ted, Tee;
|
Chris@82
|
2515 TdX = TdV + TdW;
|
Chris@82
|
2516 Te4 = Te0 + Te3;
|
Chris@82
|
2517 ro[WS(os, 36)] = TdX - Te4;
|
Chris@82
|
2518 ro[WS(os, 4)] = TdX + Te4;
|
Chris@82
|
2519 Ted = Te5 + Te6;
|
Chris@82
|
2520 Tee = Tea + Teb;
|
Chris@82
|
2521 io[WS(os, 36)] = Ted - Tee;
|
Chris@82
|
2522 io[WS(os, 4)] = Ted + Tee;
|
Chris@82
|
2523 }
|
Chris@82
|
2524 {
|
Chris@82
|
2525 E Te7, Te8, Te9, Tec;
|
Chris@82
|
2526 Te7 = Te5 - Te6;
|
Chris@82
|
2527 Te8 = Te3 - Te0;
|
Chris@82
|
2528 io[WS(os, 52)] = Te7 - Te8;
|
Chris@82
|
2529 io[WS(os, 20)] = Te7 + Te8;
|
Chris@82
|
2530 Te9 = TdV - TdW;
|
Chris@82
|
2531 Tec = Tea - Teb;
|
Chris@82
|
2532 ro[WS(os, 52)] = Te9 - Tec;
|
Chris@82
|
2533 ro[WS(os, 20)] = Te9 + Tec;
|
Chris@82
|
2534 }
|
Chris@82
|
2535 }
|
Chris@82
|
2536 {
|
Chris@82
|
2537 E Tcd, TcP, TcD, TcZ, Tck, Td0, TcX, Td5, Tcs, TcK, TcG, TcQ, TcU, Td4, Tcz;
|
Chris@82
|
2538 E TcL, Tcc, TcC;
|
Chris@82
|
2539 Tcc = KP707106781 * (TbD + TbC);
|
Chris@82
|
2540 Tcd = Tcb - Tcc;
|
Chris@82
|
2541 TcP = Tcb + Tcc;
|
Chris@82
|
2542 TcC = KP707106781 * (Tak + Tan);
|
Chris@82
|
2543 TcD = TcB - TcC;
|
Chris@82
|
2544 TcZ = TcB + TcC;
|
Chris@82
|
2545 {
|
Chris@82
|
2546 E Tcg, Tcj, TcV, TcW;
|
Chris@82
|
2547 Tcg = FNMS(KP382683432, Tcf, KP923879532 * Tce);
|
Chris@82
|
2548 Tcj = FMA(KP923879532, Tch, KP382683432 * Tci);
|
Chris@82
|
2549 Tck = Tcg - Tcj;
|
Chris@82
|
2550 Td0 = Tcg + Tcj;
|
Chris@82
|
2551 TcV = Tct + Tcu;
|
Chris@82
|
2552 TcW = Tcw + Tcx;
|
Chris@82
|
2553 TcX = FNMS(KP195090322, TcW, KP980785280 * TcV);
|
Chris@82
|
2554 Td5 = FMA(KP195090322, TcV, KP980785280 * TcW);
|
Chris@82
|
2555 }
|
Chris@82
|
2556 {
|
Chris@82
|
2557 E Tco, Tcr, TcE, TcF;
|
Chris@82
|
2558 Tco = Tcm - Tcn;
|
Chris@82
|
2559 Tcr = Tcp - Tcq;
|
Chris@82
|
2560 Tcs = FMA(KP555570233, Tco, KP831469612 * Tcr);
|
Chris@82
|
2561 TcK = FNMS(KP831469612, Tco, KP555570233 * Tcr);
|
Chris@82
|
2562 TcE = FNMS(KP382683432, Tch, KP923879532 * Tci);
|
Chris@82
|
2563 TcF = FMA(KP382683432, Tce, KP923879532 * Tcf);
|
Chris@82
|
2564 TcG = TcE - TcF;
|
Chris@82
|
2565 TcQ = TcF + TcE;
|
Chris@82
|
2566 }
|
Chris@82
|
2567 {
|
Chris@82
|
2568 E TcS, TcT, Tcv, Tcy;
|
Chris@82
|
2569 TcS = Tcm + Tcn;
|
Chris@82
|
2570 TcT = Tcp + Tcq;
|
Chris@82
|
2571 TcU = FMA(KP980785280, TcS, KP195090322 * TcT);
|
Chris@82
|
2572 Td4 = FNMS(KP195090322, TcS, KP980785280 * TcT);
|
Chris@82
|
2573 Tcv = Tct - Tcu;
|
Chris@82
|
2574 Tcy = Tcw - Tcx;
|
Chris@82
|
2575 Tcz = FNMS(KP831469612, Tcy, KP555570233 * Tcv);
|
Chris@82
|
2576 TcL = FMA(KP831469612, Tcv, KP555570233 * Tcy);
|
Chris@82
|
2577 }
|
Chris@82
|
2578 {
|
Chris@82
|
2579 E Tcl, TcA, TcN, TcO;
|
Chris@82
|
2580 Tcl = Tcd + Tck;
|
Chris@82
|
2581 TcA = Tcs + Tcz;
|
Chris@82
|
2582 ro[WS(os, 42)] = Tcl - TcA;
|
Chris@82
|
2583 ro[WS(os, 10)] = Tcl + TcA;
|
Chris@82
|
2584 TcN = TcD + TcG;
|
Chris@82
|
2585 TcO = TcK + TcL;
|
Chris@82
|
2586 io[WS(os, 42)] = TcN - TcO;
|
Chris@82
|
2587 io[WS(os, 10)] = TcN + TcO;
|
Chris@82
|
2588 }
|
Chris@82
|
2589 {
|
Chris@82
|
2590 E TcH, TcI, TcJ, TcM;
|
Chris@82
|
2591 TcH = TcD - TcG;
|
Chris@82
|
2592 TcI = Tcz - Tcs;
|
Chris@82
|
2593 io[WS(os, 58)] = TcH - TcI;
|
Chris@82
|
2594 io[WS(os, 26)] = TcH + TcI;
|
Chris@82
|
2595 TcJ = Tcd - Tck;
|
Chris@82
|
2596 TcM = TcK - TcL;
|
Chris@82
|
2597 ro[WS(os, 58)] = TcJ - TcM;
|
Chris@82
|
2598 ro[WS(os, 26)] = TcJ + TcM;
|
Chris@82
|
2599 }
|
Chris@82
|
2600 {
|
Chris@82
|
2601 E TcR, TcY, Td7, Td8;
|
Chris@82
|
2602 TcR = TcP + TcQ;
|
Chris@82
|
2603 TcY = TcU + TcX;
|
Chris@82
|
2604 ro[WS(os, 34)] = TcR - TcY;
|
Chris@82
|
2605 ro[WS(os, 2)] = TcR + TcY;
|
Chris@82
|
2606 Td7 = TcZ + Td0;
|
Chris@82
|
2607 Td8 = Td4 + Td5;
|
Chris@82
|
2608 io[WS(os, 34)] = Td7 - Td8;
|
Chris@82
|
2609 io[WS(os, 2)] = Td7 + Td8;
|
Chris@82
|
2610 }
|
Chris@82
|
2611 {
|
Chris@82
|
2612 E Td1, Td2, Td3, Td6;
|
Chris@82
|
2613 Td1 = TcZ - Td0;
|
Chris@82
|
2614 Td2 = TcX - TcU;
|
Chris@82
|
2615 io[WS(os, 50)] = Td1 - Td2;
|
Chris@82
|
2616 io[WS(os, 18)] = Td1 + Td2;
|
Chris@82
|
2617 Td3 = TcP - TcQ;
|
Chris@82
|
2618 Td6 = Td4 - Td5;
|
Chris@82
|
2619 ro[WS(os, 50)] = Td3 - Td6;
|
Chris@82
|
2620 ro[WS(os, 18)] = Td3 + Td6;
|
Chris@82
|
2621 }
|
Chris@82
|
2622 }
|
Chris@82
|
2623 {
|
Chris@82
|
2624 E Tap, TbR, TbF, Tc1, TaE, Tc2, TbZ, Tc7, Tb6, TbM, TbI, TbS, TbW, Tc6, Tbx;
|
Chris@82
|
2625 E TbN, Tao, TbE;
|
Chris@82
|
2626 Tao = KP707106781 * (Tak - Tan);
|
Chris@82
|
2627 Tap = Tah - Tao;
|
Chris@82
|
2628 TbR = Tah + Tao;
|
Chris@82
|
2629 TbE = KP707106781 * (TbC - TbD);
|
Chris@82
|
2630 TbF = TbB - TbE;
|
Chris@82
|
2631 Tc1 = TbB + TbE;
|
Chris@82
|
2632 {
|
Chris@82
|
2633 E Taw, TaD, TbX, TbY;
|
Chris@82
|
2634 Taw = FNMS(KP923879532, Tav, KP382683432 * Tas);
|
Chris@82
|
2635 TaD = FMA(KP382683432, Taz, KP923879532 * TaC);
|
Chris@82
|
2636 TaE = Taw - TaD;
|
Chris@82
|
2637 Tc2 = Taw + TaD;
|
Chris@82
|
2638 TbX = Tbb + Tbm;
|
Chris@82
|
2639 TbY = Tbs + Tbv;
|
Chris@82
|
2640 TbZ = FNMS(KP555570233, TbY, KP831469612 * TbX);
|
Chris@82
|
2641 Tc7 = FMA(KP831469612, TbY, KP555570233 * TbX);
|
Chris@82
|
2642 }
|
Chris@82
|
2643 {
|
Chris@82
|
2644 E TaW, Tb5, TbG, TbH;
|
Chris@82
|
2645 TaW = TaK - TaV;
|
Chris@82
|
2646 Tb5 = Tb1 - Tb4;
|
Chris@82
|
2647 Tb6 = FMA(KP980785280, TaW, KP195090322 * Tb5);
|
Chris@82
|
2648 TbM = FNMS(KP980785280, Tb5, KP195090322 * TaW);
|
Chris@82
|
2649 TbG = FNMS(KP923879532, Taz, KP382683432 * TaC);
|
Chris@82
|
2650 TbH = FMA(KP923879532, Tas, KP382683432 * Tav);
|
Chris@82
|
2651 TbI = TbG - TbH;
|
Chris@82
|
2652 TbS = TbH + TbG;
|
Chris@82
|
2653 }
|
Chris@82
|
2654 {
|
Chris@82
|
2655 E TbU, TbV, Tbn, Tbw;
|
Chris@82
|
2656 TbU = TaK + TaV;
|
Chris@82
|
2657 TbV = Tb1 + Tb4;
|
Chris@82
|
2658 TbW = FMA(KP555570233, TbU, KP831469612 * TbV);
|
Chris@82
|
2659 Tc6 = FNMS(KP555570233, TbV, KP831469612 * TbU);
|
Chris@82
|
2660 Tbn = Tbb - Tbm;
|
Chris@82
|
2661 Tbw = Tbs - Tbv;
|
Chris@82
|
2662 Tbx = FNMS(KP980785280, Tbw, KP195090322 * Tbn);
|
Chris@82
|
2663 TbN = FMA(KP195090322, Tbw, KP980785280 * Tbn);
|
Chris@82
|
2664 }
|
Chris@82
|
2665 {
|
Chris@82
|
2666 E TaF, Tby, TbP, TbQ;
|
Chris@82
|
2667 TaF = Tap + TaE;
|
Chris@82
|
2668 Tby = Tb6 + Tbx;
|
Chris@82
|
2669 ro[WS(os, 46)] = TaF - Tby;
|
Chris@82
|
2670 ro[WS(os, 14)] = TaF + Tby;
|
Chris@82
|
2671 TbP = TbF + TbI;
|
Chris@82
|
2672 TbQ = TbM + TbN;
|
Chris@82
|
2673 io[WS(os, 46)] = TbP - TbQ;
|
Chris@82
|
2674 io[WS(os, 14)] = TbP + TbQ;
|
Chris@82
|
2675 }
|
Chris@82
|
2676 {
|
Chris@82
|
2677 E TbJ, TbK, TbL, TbO;
|
Chris@82
|
2678 TbJ = TbF - TbI;
|
Chris@82
|
2679 TbK = Tbx - Tb6;
|
Chris@82
|
2680 io[WS(os, 62)] = TbJ - TbK;
|
Chris@82
|
2681 io[WS(os, 30)] = TbJ + TbK;
|
Chris@82
|
2682 TbL = Tap - TaE;
|
Chris@82
|
2683 TbO = TbM - TbN;
|
Chris@82
|
2684 ro[WS(os, 62)] = TbL - TbO;
|
Chris@82
|
2685 ro[WS(os, 30)] = TbL + TbO;
|
Chris@82
|
2686 }
|
Chris@82
|
2687 {
|
Chris@82
|
2688 E TbT, Tc0, Tc9, Tca;
|
Chris@82
|
2689 TbT = TbR + TbS;
|
Chris@82
|
2690 Tc0 = TbW + TbZ;
|
Chris@82
|
2691 ro[WS(os, 38)] = TbT - Tc0;
|
Chris@82
|
2692 ro[WS(os, 6)] = TbT + Tc0;
|
Chris@82
|
2693 Tc9 = Tc1 + Tc2;
|
Chris@82
|
2694 Tca = Tc6 + Tc7;
|
Chris@82
|
2695 io[WS(os, 38)] = Tc9 - Tca;
|
Chris@82
|
2696 io[WS(os, 6)] = Tc9 + Tca;
|
Chris@82
|
2697 }
|
Chris@82
|
2698 {
|
Chris@82
|
2699 E Tc3, Tc4, Tc5, Tc8;
|
Chris@82
|
2700 Tc3 = Tc1 - Tc2;
|
Chris@82
|
2701 Tc4 = TbZ - TbW;
|
Chris@82
|
2702 io[WS(os, 54)] = Tc3 - Tc4;
|
Chris@82
|
2703 io[WS(os, 22)] = Tc3 + Tc4;
|
Chris@82
|
2704 Tc5 = TbR - TbS;
|
Chris@82
|
2705 Tc8 = Tc6 - Tc7;
|
Chris@82
|
2706 ro[WS(os, 54)] = Tc5 - Tc8;
|
Chris@82
|
2707 ro[WS(os, 22)] = Tc5 + Tc8;
|
Chris@82
|
2708 }
|
Chris@82
|
2709 }
|
Chris@82
|
2710 {
|
Chris@82
|
2711 E T6F, T7h, T7m, T7w, T7p, T7x, T6M, T7s, T6U, T7c, T75, T7r, T78, T7i, T71;
|
Chris@82
|
2712 E T7d;
|
Chris@82
|
2713 {
|
Chris@82
|
2714 E T6D, T6E, T7k, T7l;
|
Chris@82
|
2715 T6D = T37 + T3e;
|
Chris@82
|
2716 T6E = T65 + T64;
|
Chris@82
|
2717 T6F = T6D - T6E;
|
Chris@82
|
2718 T7h = T6D + T6E;
|
Chris@82
|
2719 T7k = T6O + T6P;
|
Chris@82
|
2720 T7l = T6R + T6S;
|
Chris@82
|
2721 T7m = FMA(KP956940335, T7k, KP290284677 * T7l);
|
Chris@82
|
2722 T7w = FNMS(KP290284677, T7k, KP956940335 * T7l);
|
Chris@82
|
2723 }
|
Chris@82
|
2724 {
|
Chris@82
|
2725 E T7n, T7o, T6I, T6L;
|
Chris@82
|
2726 T7n = T6V + T6W;
|
Chris@82
|
2727 T7o = T6Y + T6Z;
|
Chris@82
|
2728 T7p = FNMS(KP290284677, T7o, KP956940335 * T7n);
|
Chris@82
|
2729 T7x = FMA(KP290284677, T7n, KP956940335 * T7o);
|
Chris@82
|
2730 T6I = FNMS(KP555570233, T6H, KP831469612 * T6G);
|
Chris@82
|
2731 T6L = FMA(KP831469612, T6J, KP555570233 * T6K);
|
Chris@82
|
2732 T6M = T6I - T6L;
|
Chris@82
|
2733 T7s = T6I + T6L;
|
Chris@82
|
2734 }
|
Chris@82
|
2735 {
|
Chris@82
|
2736 E T6Q, T6T, T73, T74;
|
Chris@82
|
2737 T6Q = T6O - T6P;
|
Chris@82
|
2738 T6T = T6R - T6S;
|
Chris@82
|
2739 T6U = FMA(KP471396736, T6Q, KP881921264 * T6T);
|
Chris@82
|
2740 T7c = FNMS(KP881921264, T6Q, KP471396736 * T6T);
|
Chris@82
|
2741 T73 = T5Z + T62;
|
Chris@82
|
2742 T74 = T3m + T3t;
|
Chris@82
|
2743 T75 = T73 - T74;
|
Chris@82
|
2744 T7r = T73 + T74;
|
Chris@82
|
2745 }
|
Chris@82
|
2746 {
|
Chris@82
|
2747 E T76, T77, T6X, T70;
|
Chris@82
|
2748 T76 = FNMS(KP555570233, T6J, KP831469612 * T6K);
|
Chris@82
|
2749 T77 = FMA(KP555570233, T6G, KP831469612 * T6H);
|
Chris@82
|
2750 T78 = T76 - T77;
|
Chris@82
|
2751 T7i = T77 + T76;
|
Chris@82
|
2752 T6X = T6V - T6W;
|
Chris@82
|
2753 T70 = T6Y - T6Z;
|
Chris@82
|
2754 T71 = FNMS(KP881921264, T70, KP471396736 * T6X);
|
Chris@82
|
2755 T7d = FMA(KP881921264, T6X, KP471396736 * T70);
|
Chris@82
|
2756 }
|
Chris@82
|
2757 {
|
Chris@82
|
2758 E T6N, T72, T7f, T7g;
|
Chris@82
|
2759 T6N = T6F + T6M;
|
Chris@82
|
2760 T72 = T6U + T71;
|
Chris@82
|
2761 ro[WS(os, 43)] = T6N - T72;
|
Chris@82
|
2762 ro[WS(os, 11)] = T6N + T72;
|
Chris@82
|
2763 T7f = T75 + T78;
|
Chris@82
|
2764 T7g = T7c + T7d;
|
Chris@82
|
2765 io[WS(os, 43)] = T7f - T7g;
|
Chris@82
|
2766 io[WS(os, 11)] = T7f + T7g;
|
Chris@82
|
2767 }
|
Chris@82
|
2768 {
|
Chris@82
|
2769 E T79, T7a, T7b, T7e;
|
Chris@82
|
2770 T79 = T75 - T78;
|
Chris@82
|
2771 T7a = T71 - T6U;
|
Chris@82
|
2772 io[WS(os, 59)] = T79 - T7a;
|
Chris@82
|
2773 io[WS(os, 27)] = T79 + T7a;
|
Chris@82
|
2774 T7b = T6F - T6M;
|
Chris@82
|
2775 T7e = T7c - T7d;
|
Chris@82
|
2776 ro[WS(os, 59)] = T7b - T7e;
|
Chris@82
|
2777 ro[WS(os, 27)] = T7b + T7e;
|
Chris@82
|
2778 }
|
Chris@82
|
2779 {
|
Chris@82
|
2780 E T7j, T7q, T7z, T7A;
|
Chris@82
|
2781 T7j = T7h + T7i;
|
Chris@82
|
2782 T7q = T7m + T7p;
|
Chris@82
|
2783 ro[WS(os, 35)] = T7j - T7q;
|
Chris@82
|
2784 ro[WS(os, 3)] = T7j + T7q;
|
Chris@82
|
2785 T7z = T7r + T7s;
|
Chris@82
|
2786 T7A = T7w + T7x;
|
Chris@82
|
2787 io[WS(os, 35)] = T7z - T7A;
|
Chris@82
|
2788 io[WS(os, 3)] = T7z + T7A;
|
Chris@82
|
2789 }
|
Chris@82
|
2790 {
|
Chris@82
|
2791 E T7t, T7u, T7v, T7y;
|
Chris@82
|
2792 T7t = T7r - T7s;
|
Chris@82
|
2793 T7u = T7p - T7m;
|
Chris@82
|
2794 io[WS(os, 51)] = T7t - T7u;
|
Chris@82
|
2795 io[WS(os, 19)] = T7t + T7u;
|
Chris@82
|
2796 T7v = T7h - T7i;
|
Chris@82
|
2797 T7y = T7w - T7x;
|
Chris@82
|
2798 ro[WS(os, 51)] = T7v - T7y;
|
Chris@82
|
2799 ro[WS(os, 19)] = T7v + T7y;
|
Chris@82
|
2800 }
|
Chris@82
|
2801 }
|
Chris@82
|
2802 {
|
Chris@82
|
2803 E T9j, T9V, Ta0, Taa, Ta3, Tab, T9q, Ta6, T9y, T9Q, T9J, Ta5, T9M, T9W, T9F;
|
Chris@82
|
2804 E T9R;
|
Chris@82
|
2805 {
|
Chris@82
|
2806 E T9h, T9i, T9Y, T9Z;
|
Chris@82
|
2807 T9h = T7B + T7C;
|
Chris@82
|
2808 T9i = T8J + T8I;
|
Chris@82
|
2809 T9j = T9h - T9i;
|
Chris@82
|
2810 T9V = T9h + T9i;
|
Chris@82
|
2811 T9Y = T9s + T9t;
|
Chris@82
|
2812 T9Z = T9v + T9w;
|
Chris@82
|
2813 Ta0 = FMA(KP995184726, T9Y, KP098017140 * T9Z);
|
Chris@82
|
2814 Taa = FNMS(KP098017140, T9Y, KP995184726 * T9Z);
|
Chris@82
|
2815 }
|
Chris@82
|
2816 {
|
Chris@82
|
2817 E Ta1, Ta2, T9m, T9p;
|
Chris@82
|
2818 Ta1 = T9z + T9A;
|
Chris@82
|
2819 Ta2 = T9C + T9D;
|
Chris@82
|
2820 Ta3 = FNMS(KP098017140, Ta2, KP995184726 * Ta1);
|
Chris@82
|
2821 Tab = FMA(KP098017140, Ta1, KP995184726 * Ta2);
|
Chris@82
|
2822 T9m = FNMS(KP195090322, T9l, KP980785280 * T9k);
|
Chris@82
|
2823 T9p = FMA(KP195090322, T9n, KP980785280 * T9o);
|
Chris@82
|
2824 T9q = T9m - T9p;
|
Chris@82
|
2825 Ta6 = T9m + T9p;
|
Chris@82
|
2826 }
|
Chris@82
|
2827 {
|
Chris@82
|
2828 E T9u, T9x, T9H, T9I;
|
Chris@82
|
2829 T9u = T9s - T9t;
|
Chris@82
|
2830 T9x = T9v - T9w;
|
Chris@82
|
2831 T9y = FMA(KP634393284, T9u, KP773010453 * T9x);
|
Chris@82
|
2832 T9Q = FNMS(KP773010453, T9u, KP634393284 * T9x);
|
Chris@82
|
2833 T9H = T8F + T8G;
|
Chris@82
|
2834 T9I = T7G + T7J;
|
Chris@82
|
2835 T9J = T9H - T9I;
|
Chris@82
|
2836 Ta5 = T9H + T9I;
|
Chris@82
|
2837 }
|
Chris@82
|
2838 {
|
Chris@82
|
2839 E T9K, T9L, T9B, T9E;
|
Chris@82
|
2840 T9K = FNMS(KP195090322, T9o, KP980785280 * T9n);
|
Chris@82
|
2841 T9L = FMA(KP980785280, T9l, KP195090322 * T9k);
|
Chris@82
|
2842 T9M = T9K - T9L;
|
Chris@82
|
2843 T9W = T9L + T9K;
|
Chris@82
|
2844 T9B = T9z - T9A;
|
Chris@82
|
2845 T9E = T9C - T9D;
|
Chris@82
|
2846 T9F = FNMS(KP773010453, T9E, KP634393284 * T9B);
|
Chris@82
|
2847 T9R = FMA(KP773010453, T9B, KP634393284 * T9E);
|
Chris@82
|
2848 }
|
Chris@82
|
2849 {
|
Chris@82
|
2850 E T9r, T9G, T9T, T9U;
|
Chris@82
|
2851 T9r = T9j + T9q;
|
Chris@82
|
2852 T9G = T9y + T9F;
|
Chris@82
|
2853 ro[WS(os, 41)] = T9r - T9G;
|
Chris@82
|
2854 ro[WS(os, 9)] = T9r + T9G;
|
Chris@82
|
2855 T9T = T9J + T9M;
|
Chris@82
|
2856 T9U = T9Q + T9R;
|
Chris@82
|
2857 io[WS(os, 41)] = T9T - T9U;
|
Chris@82
|
2858 io[WS(os, 9)] = T9T + T9U;
|
Chris@82
|
2859 }
|
Chris@82
|
2860 {
|
Chris@82
|
2861 E T9N, T9O, T9P, T9S;
|
Chris@82
|
2862 T9N = T9J - T9M;
|
Chris@82
|
2863 T9O = T9F - T9y;
|
Chris@82
|
2864 io[WS(os, 57)] = T9N - T9O;
|
Chris@82
|
2865 io[WS(os, 25)] = T9N + T9O;
|
Chris@82
|
2866 T9P = T9j - T9q;
|
Chris@82
|
2867 T9S = T9Q - T9R;
|
Chris@82
|
2868 ro[WS(os, 57)] = T9P - T9S;
|
Chris@82
|
2869 ro[WS(os, 25)] = T9P + T9S;
|
Chris@82
|
2870 }
|
Chris@82
|
2871 {
|
Chris@82
|
2872 E T9X, Ta4, Tad, Tae;
|
Chris@82
|
2873 T9X = T9V + T9W;
|
Chris@82
|
2874 Ta4 = Ta0 + Ta3;
|
Chris@82
|
2875 ro[WS(os, 33)] = T9X - Ta4;
|
Chris@82
|
2876 ro[WS(os, 1)] = T9X + Ta4;
|
Chris@82
|
2877 Tad = Ta5 + Ta6;
|
Chris@82
|
2878 Tae = Taa + Tab;
|
Chris@82
|
2879 io[WS(os, 33)] = Tad - Tae;
|
Chris@82
|
2880 io[WS(os, 1)] = Tad + Tae;
|
Chris@82
|
2881 }
|
Chris@82
|
2882 {
|
Chris@82
|
2883 E Ta7, Ta8, Ta9, Tac;
|
Chris@82
|
2884 Ta7 = Ta5 - Ta6;
|
Chris@82
|
2885 Ta8 = Ta3 - Ta0;
|
Chris@82
|
2886 io[WS(os, 49)] = Ta7 - Ta8;
|
Chris@82
|
2887 io[WS(os, 17)] = Ta7 + Ta8;
|
Chris@82
|
2888 Ta9 = T9V - T9W;
|
Chris@82
|
2889 Tac = Taa - Tab;
|
Chris@82
|
2890 ro[WS(os, 49)] = Ta9 - Tac;
|
Chris@82
|
2891 ro[WS(os, 17)] = Ta9 + Tac;
|
Chris@82
|
2892 }
|
Chris@82
|
2893 }
|
Chris@82
|
2894 {
|
Chris@82
|
2895 E T3v, T6j, T6o, T6y, T6r, T6z, T48, T6u, T52, T6e, T67, T6t, T6a, T6k, T5V;
|
Chris@82
|
2896 E T6f;
|
Chris@82
|
2897 {
|
Chris@82
|
2898 E T3f, T3u, T6m, T6n;
|
Chris@82
|
2899 T3f = T37 - T3e;
|
Chris@82
|
2900 T3u = T3m - T3t;
|
Chris@82
|
2901 T3v = T3f - T3u;
|
Chris@82
|
2902 T6j = T3f + T3u;
|
Chris@82
|
2903 T6m = T4q + T4N;
|
Chris@82
|
2904 T6n = T4X + T50;
|
Chris@82
|
2905 T6o = FMA(KP634393284, T6m, KP773010453 * T6n);
|
Chris@82
|
2906 T6y = FNMS(KP634393284, T6n, KP773010453 * T6m);
|
Chris@82
|
2907 }
|
Chris@82
|
2908 {
|
Chris@82
|
2909 E T6p, T6q, T3O, T47;
|
Chris@82
|
2910 T6p = T5j + T5G;
|
Chris@82
|
2911 T6q = T5Q + T5T;
|
Chris@82
|
2912 T6r = FNMS(KP634393284, T6q, KP773010453 * T6p);
|
Chris@82
|
2913 T6z = FMA(KP773010453, T6q, KP634393284 * T6p);
|
Chris@82
|
2914 T3O = FNMS(KP980785280, T3N, KP195090322 * T3G);
|
Chris@82
|
2915 T47 = FMA(KP195090322, T3Z, KP980785280 * T46);
|
Chris@82
|
2916 T48 = T3O - T47;
|
Chris@82
|
2917 T6u = T3O + T47;
|
Chris@82
|
2918 }
|
Chris@82
|
2919 {
|
Chris@82
|
2920 E T4O, T51, T63, T66;
|
Chris@82
|
2921 T4O = T4q - T4N;
|
Chris@82
|
2922 T51 = T4X - T50;
|
Chris@82
|
2923 T52 = FMA(KP995184726, T4O, KP098017140 * T51);
|
Chris@82
|
2924 T6e = FNMS(KP995184726, T51, KP098017140 * T4O);
|
Chris@82
|
2925 T63 = T5Z - T62;
|
Chris@82
|
2926 T66 = T64 - T65;
|
Chris@82
|
2927 T67 = T63 - T66;
|
Chris@82
|
2928 T6t = T63 + T66;
|
Chris@82
|
2929 }
|
Chris@82
|
2930 {
|
Chris@82
|
2931 E T68, T69, T5H, T5U;
|
Chris@82
|
2932 T68 = FNMS(KP980785280, T3Z, KP195090322 * T46);
|
Chris@82
|
2933 T69 = FMA(KP980785280, T3G, KP195090322 * T3N);
|
Chris@82
|
2934 T6a = T68 - T69;
|
Chris@82
|
2935 T6k = T69 + T68;
|
Chris@82
|
2936 T5H = T5j - T5G;
|
Chris@82
|
2937 T5U = T5Q - T5T;
|
Chris@82
|
2938 T5V = FNMS(KP995184726, T5U, KP098017140 * T5H);
|
Chris@82
|
2939 T6f = FMA(KP098017140, T5U, KP995184726 * T5H);
|
Chris@82
|
2940 }
|
Chris@82
|
2941 {
|
Chris@82
|
2942 E T49, T5W, T6h, T6i;
|
Chris@82
|
2943 T49 = T3v + T48;
|
Chris@82
|
2944 T5W = T52 + T5V;
|
Chris@82
|
2945 ro[WS(os, 47)] = T49 - T5W;
|
Chris@82
|
2946 ro[WS(os, 15)] = T49 + T5W;
|
Chris@82
|
2947 T6h = T67 + T6a;
|
Chris@82
|
2948 T6i = T6e + T6f;
|
Chris@82
|
2949 io[WS(os, 47)] = T6h - T6i;
|
Chris@82
|
2950 io[WS(os, 15)] = T6h + T6i;
|
Chris@82
|
2951 }
|
Chris@82
|
2952 {
|
Chris@82
|
2953 E T6b, T6c, T6d, T6g;
|
Chris@82
|
2954 T6b = T67 - T6a;
|
Chris@82
|
2955 T6c = T5V - T52;
|
Chris@82
|
2956 io[WS(os, 63)] = T6b - T6c;
|
Chris@82
|
2957 io[WS(os, 31)] = T6b + T6c;
|
Chris@82
|
2958 T6d = T3v - T48;
|
Chris@82
|
2959 T6g = T6e - T6f;
|
Chris@82
|
2960 ro[WS(os, 63)] = T6d - T6g;
|
Chris@82
|
2961 ro[WS(os, 31)] = T6d + T6g;
|
Chris@82
|
2962 }
|
Chris@82
|
2963 {
|
Chris@82
|
2964 E T6l, T6s, T6B, T6C;
|
Chris@82
|
2965 T6l = T6j + T6k;
|
Chris@82
|
2966 T6s = T6o + T6r;
|
Chris@82
|
2967 ro[WS(os, 39)] = T6l - T6s;
|
Chris@82
|
2968 ro[WS(os, 7)] = T6l + T6s;
|
Chris@82
|
2969 T6B = T6t + T6u;
|
Chris@82
|
2970 T6C = T6y + T6z;
|
Chris@82
|
2971 io[WS(os, 39)] = T6B - T6C;
|
Chris@82
|
2972 io[WS(os, 7)] = T6B + T6C;
|
Chris@82
|
2973 }
|
Chris@82
|
2974 {
|
Chris@82
|
2975 E T6v, T6w, T6x, T6A;
|
Chris@82
|
2976 T6v = T6t - T6u;
|
Chris@82
|
2977 T6w = T6r - T6o;
|
Chris@82
|
2978 io[WS(os, 55)] = T6v - T6w;
|
Chris@82
|
2979 io[WS(os, 23)] = T6v + T6w;
|
Chris@82
|
2980 T6x = T6j - T6k;
|
Chris@82
|
2981 T6A = T6y - T6z;
|
Chris@82
|
2982 ro[WS(os, 55)] = T6x - T6A;
|
Chris@82
|
2983 ro[WS(os, 23)] = T6x + T6A;
|
Chris@82
|
2984 }
|
Chris@82
|
2985 }
|
Chris@82
|
2986 {
|
Chris@82
|
2987 E T7L, T8X, T92, T9c, T95, T9d, T80, T98, T8k, T8S, T8L, T97, T8O, T8Y, T8D;
|
Chris@82
|
2988 E T8T;
|
Chris@82
|
2989 {
|
Chris@82
|
2990 E T7D, T7K, T90, T91;
|
Chris@82
|
2991 T7D = T7B - T7C;
|
Chris@82
|
2992 T7K = T7G - T7J;
|
Chris@82
|
2993 T7L = T7D - T7K;
|
Chris@82
|
2994 T8X = T7D + T7K;
|
Chris@82
|
2995 T90 = T84 + T8b;
|
Chris@82
|
2996 T91 = T8f + T8i;
|
Chris@82
|
2997 T92 = FMA(KP471396736, T90, KP881921264 * T91);
|
Chris@82
|
2998 T9c = FNMS(KP471396736, T91, KP881921264 * T90);
|
Chris@82
|
2999 }
|
Chris@82
|
3000 {
|
Chris@82
|
3001 E T93, T94, T7S, T7Z;
|
Chris@82
|
3002 T93 = T8n + T8u;
|
Chris@82
|
3003 T94 = T8y + T8B;
|
Chris@82
|
3004 T95 = FNMS(KP471396736, T94, KP881921264 * T93);
|
Chris@82
|
3005 T9d = FMA(KP881921264, T94, KP471396736 * T93);
|
Chris@82
|
3006 T7S = FNMS(KP831469612, T7R, KP555570233 * T7O);
|
Chris@82
|
3007 T7Z = FMA(KP831469612, T7V, KP555570233 * T7Y);
|
Chris@82
|
3008 T80 = T7S - T7Z;
|
Chris@82
|
3009 T98 = T7S + T7Z;
|
Chris@82
|
3010 }
|
Chris@82
|
3011 {
|
Chris@82
|
3012 E T8c, T8j, T8H, T8K;
|
Chris@82
|
3013 T8c = T84 - T8b;
|
Chris@82
|
3014 T8j = T8f - T8i;
|
Chris@82
|
3015 T8k = FMA(KP956940335, T8c, KP290284677 * T8j);
|
Chris@82
|
3016 T8S = FNMS(KP956940335, T8j, KP290284677 * T8c);
|
Chris@82
|
3017 T8H = T8F - T8G;
|
Chris@82
|
3018 T8K = T8I - T8J;
|
Chris@82
|
3019 T8L = T8H - T8K;
|
Chris@82
|
3020 T97 = T8H + T8K;
|
Chris@82
|
3021 }
|
Chris@82
|
3022 {
|
Chris@82
|
3023 E T8M, T8N, T8v, T8C;
|
Chris@82
|
3024 T8M = FNMS(KP831469612, T7Y, KP555570233 * T7V);
|
Chris@82
|
3025 T8N = FMA(KP555570233, T7R, KP831469612 * T7O);
|
Chris@82
|
3026 T8O = T8M - T8N;
|
Chris@82
|
3027 T8Y = T8N + T8M;
|
Chris@82
|
3028 T8v = T8n - T8u;
|
Chris@82
|
3029 T8C = T8y - T8B;
|
Chris@82
|
3030 T8D = FNMS(KP956940335, T8C, KP290284677 * T8v);
|
Chris@82
|
3031 T8T = FMA(KP290284677, T8C, KP956940335 * T8v);
|
Chris@82
|
3032 }
|
Chris@82
|
3033 {
|
Chris@82
|
3034 E T81, T8E, T8V, T8W;
|
Chris@82
|
3035 T81 = T7L + T80;
|
Chris@82
|
3036 T8E = T8k + T8D;
|
Chris@82
|
3037 ro[WS(os, 45)] = T81 - T8E;
|
Chris@82
|
3038 ro[WS(os, 13)] = T81 + T8E;
|
Chris@82
|
3039 T8V = T8L + T8O;
|
Chris@82
|
3040 T8W = T8S + T8T;
|
Chris@82
|
3041 io[WS(os, 45)] = T8V - T8W;
|
Chris@82
|
3042 io[WS(os, 13)] = T8V + T8W;
|
Chris@82
|
3043 }
|
Chris@82
|
3044 {
|
Chris@82
|
3045 E T8P, T8Q, T8R, T8U;
|
Chris@82
|
3046 T8P = T8L - T8O;
|
Chris@82
|
3047 T8Q = T8D - T8k;
|
Chris@82
|
3048 io[WS(os, 61)] = T8P - T8Q;
|
Chris@82
|
3049 io[WS(os, 29)] = T8P + T8Q;
|
Chris@82
|
3050 T8R = T7L - T80;
|
Chris@82
|
3051 T8U = T8S - T8T;
|
Chris@82
|
3052 ro[WS(os, 61)] = T8R - T8U;
|
Chris@82
|
3053 ro[WS(os, 29)] = T8R + T8U;
|
Chris@82
|
3054 }
|
Chris@82
|
3055 {
|
Chris@82
|
3056 E T8Z, T96, T9f, T9g;
|
Chris@82
|
3057 T8Z = T8X + T8Y;
|
Chris@82
|
3058 T96 = T92 + T95;
|
Chris@82
|
3059 ro[WS(os, 37)] = T8Z - T96;
|
Chris@82
|
3060 ro[WS(os, 5)] = T8Z + T96;
|
Chris@82
|
3061 T9f = T97 + T98;
|
Chris@82
|
3062 T9g = T9c + T9d;
|
Chris@82
|
3063 io[WS(os, 37)] = T9f - T9g;
|
Chris@82
|
3064 io[WS(os, 5)] = T9f + T9g;
|
Chris@82
|
3065 }
|
Chris@82
|
3066 {
|
Chris@82
|
3067 E T99, T9a, T9b, T9e;
|
Chris@82
|
3068 T99 = T97 - T98;
|
Chris@82
|
3069 T9a = T95 - T92;
|
Chris@82
|
3070 io[WS(os, 53)] = T99 - T9a;
|
Chris@82
|
3071 io[WS(os, 21)] = T99 + T9a;
|
Chris@82
|
3072 T9b = T8X - T8Y;
|
Chris@82
|
3073 T9e = T9c - T9d;
|
Chris@82
|
3074 ro[WS(os, 53)] = T9b - T9e;
|
Chris@82
|
3075 ro[WS(os, 21)] = T9b + T9e;
|
Chris@82
|
3076 }
|
Chris@82
|
3077 }
|
Chris@82
|
3078 }
|
Chris@82
|
3079 }
|
Chris@82
|
3080 }
|
Chris@82
|
3081
|
Chris@82
|
3082 static const kdft_desc desc = { 64, "n1_64", {808, 144, 104, 0}, &GENUS, 0, 0, 0, 0 };
|
Chris@82
|
3083
|
Chris@82
|
3084 void X(codelet_n1_64) (planner *p) {
|
Chris@82
|
3085 X(kdft_register) (p, n1_64, &desc);
|
Chris@82
|
3086 }
|
Chris@82
|
3087
|
Chris@82
|
3088 #endif
|