cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:06:31 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -compact -variables 4 -pipeline-latency 4 -n 64 -dit -name hf_64 -include rdft/scalar/hf.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 1038 FP additions, 644 FP multiplications,
|
cannam@167
|
32 * (or, 520 additions, 126 multiplications, 518 fused multiply/add),
|
cannam@167
|
33 * 190 stack variables, 15 constants, and 256 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hf.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hf_64(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
40 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
41 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
42 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
43 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
44 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
45 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
46 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@167
|
47 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@167
|
48 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@167
|
49 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@167
|
50 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
51 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
52 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
53 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
54 {
|
cannam@167
|
55 INT m;
|
cannam@167
|
56 for (m = mb, W = W + ((mb - 1) * 126); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 126, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
57 E Tm, TeM, TjR, Tkl, T7e, TcA, TiV, Tjm, T1G, TeW, TeZ, Thr, T7Q, TcI, T7X;
|
cannam@167
|
58 E TcJ, T29, Tf8, Tf5, Thw, T87, TcN, T8u, TcQ, T5K, TfS, Tgc, ThX, Taq, Tdm;
|
cannam@167
|
59 E Tbj, Tdx, TN, Tjl, TeP, TiP, T7l, TcB, T7s, TcC, T1f, TeR, TeU, Ths, T7B;
|
cannam@167
|
60 E TcF, T7I, TcG, T32, Tfj, Tfg, ThB, T8G, TcU, T93, TcX, T3X, Tfr, TfK, ThM;
|
cannam@167
|
61 E T9h, Td3, Taa, Tde, T2A, Tf6, Tfb, Thx, T8m, TcR, T8x, TcO, T3t, Tfh, Tfm;
|
cannam@167
|
62 E ThC, T8V, TcY, T96, TcV, T4o, TfL, Tfu, ThN, T9w, Tdf, Tad, Td4, T6b, Tg9;
|
cannam@167
|
63 E TfV, ThY, TaF, Tdy, Tbm, Tdn, T4Q, ThJ, TfA, TfN, T9M, Tdh, Taf, Td8, T5h;
|
cannam@167
|
64 E ThI, TfF, TfO, Ta1, Tdi, Tag, Tdb, T6D, ThU, Tg1, Tgf, TaV, TdA, Tbo, Tdr;
|
cannam@167
|
65 E T74, ThT, Tg6, Tge, Tba, TdB, Tbp, Tdu;
|
cannam@167
|
66 {
|
cannam@167
|
67 E T1, TiT, T7, TiS, Te, T7a, Tk, T7c;
|
cannam@167
|
68 T1 = cr[0];
|
cannam@167
|
69 TiT = ci[0];
|
cannam@167
|
70 {
|
cannam@167
|
71 E T3, T6, T4, TiR, T2, T5;
|
cannam@167
|
72 T3 = cr[WS(rs, 32)];
|
cannam@167
|
73 T6 = ci[WS(rs, 32)];
|
cannam@167
|
74 T2 = W[62];
|
cannam@167
|
75 T4 = T2 * T3;
|
cannam@167
|
76 TiR = T2 * T6;
|
cannam@167
|
77 T5 = W[63];
|
cannam@167
|
78 T7 = FMA(T5, T6, T4);
|
cannam@167
|
79 TiS = FNMS(T5, T3, TiR);
|
cannam@167
|
80 }
|
cannam@167
|
81 {
|
cannam@167
|
82 E Ta, Td, Tb, T79, T9, Tc;
|
cannam@167
|
83 Ta = cr[WS(rs, 16)];
|
cannam@167
|
84 Td = ci[WS(rs, 16)];
|
cannam@167
|
85 T9 = W[30];
|
cannam@167
|
86 Tb = T9 * Ta;
|
cannam@167
|
87 T79 = T9 * Td;
|
cannam@167
|
88 Tc = W[31];
|
cannam@167
|
89 Te = FMA(Tc, Td, Tb);
|
cannam@167
|
90 T7a = FNMS(Tc, Ta, T79);
|
cannam@167
|
91 }
|
cannam@167
|
92 {
|
cannam@167
|
93 E Tg, Tj, Th, T7b, Tf, Ti;
|
cannam@167
|
94 Tg = cr[WS(rs, 48)];
|
cannam@167
|
95 Tj = ci[WS(rs, 48)];
|
cannam@167
|
96 Tf = W[94];
|
cannam@167
|
97 Th = Tf * Tg;
|
cannam@167
|
98 T7b = Tf * Tj;
|
cannam@167
|
99 Ti = W[95];
|
cannam@167
|
100 Tk = FMA(Ti, Tj, Th);
|
cannam@167
|
101 T7c = FNMS(Ti, Tg, T7b);
|
cannam@167
|
102 }
|
cannam@167
|
103 {
|
cannam@167
|
104 E T8, Tl, TjP, TjQ;
|
cannam@167
|
105 T8 = T1 + T7;
|
cannam@167
|
106 Tl = Te + Tk;
|
cannam@167
|
107 Tm = T8 + Tl;
|
cannam@167
|
108 TeM = T8 - Tl;
|
cannam@167
|
109 TjP = Te - Tk;
|
cannam@167
|
110 TjQ = TiT - TiS;
|
cannam@167
|
111 TjR = TjP + TjQ;
|
cannam@167
|
112 Tkl = TjQ - TjP;
|
cannam@167
|
113 }
|
cannam@167
|
114 {
|
cannam@167
|
115 E T78, T7d, TiQ, TiU;
|
cannam@167
|
116 T78 = T1 - T7;
|
cannam@167
|
117 T7d = T7a - T7c;
|
cannam@167
|
118 T7e = T78 - T7d;
|
cannam@167
|
119 TcA = T78 + T7d;
|
cannam@167
|
120 TiQ = T7a + T7c;
|
cannam@167
|
121 TiU = TiS + TiT;
|
cannam@167
|
122 TiV = TiQ + TiU;
|
cannam@167
|
123 Tjm = TiU - TiQ;
|
cannam@167
|
124 }
|
cannam@167
|
125 }
|
cannam@167
|
126 {
|
cannam@167
|
127 E T1l, T7S, T1E, T7O, T1r, T7U, T1y, T7M;
|
cannam@167
|
128 {
|
cannam@167
|
129 E T1h, T1k, T1i, T7R, T1g, T1j;
|
cannam@167
|
130 T1h = cr[WS(rs, 60)];
|
cannam@167
|
131 T1k = ci[WS(rs, 60)];
|
cannam@167
|
132 T1g = W[118];
|
cannam@167
|
133 T1i = T1g * T1h;
|
cannam@167
|
134 T7R = T1g * T1k;
|
cannam@167
|
135 T1j = W[119];
|
cannam@167
|
136 T1l = FMA(T1j, T1k, T1i);
|
cannam@167
|
137 T7S = FNMS(T1j, T1h, T7R);
|
cannam@167
|
138 }
|
cannam@167
|
139 {
|
cannam@167
|
140 E T1A, T1D, T1B, T7N, T1z, T1C;
|
cannam@167
|
141 T1A = cr[WS(rs, 44)];
|
cannam@167
|
142 T1D = ci[WS(rs, 44)];
|
cannam@167
|
143 T1z = W[86];
|
cannam@167
|
144 T1B = T1z * T1A;
|
cannam@167
|
145 T7N = T1z * T1D;
|
cannam@167
|
146 T1C = W[87];
|
cannam@167
|
147 T1E = FMA(T1C, T1D, T1B);
|
cannam@167
|
148 T7O = FNMS(T1C, T1A, T7N);
|
cannam@167
|
149 }
|
cannam@167
|
150 {
|
cannam@167
|
151 E T1n, T1q, T1o, T7T, T1m, T1p;
|
cannam@167
|
152 T1n = cr[WS(rs, 28)];
|
cannam@167
|
153 T1q = ci[WS(rs, 28)];
|
cannam@167
|
154 T1m = W[54];
|
cannam@167
|
155 T1o = T1m * T1n;
|
cannam@167
|
156 T7T = T1m * T1q;
|
cannam@167
|
157 T1p = W[55];
|
cannam@167
|
158 T1r = FMA(T1p, T1q, T1o);
|
cannam@167
|
159 T7U = FNMS(T1p, T1n, T7T);
|
cannam@167
|
160 }
|
cannam@167
|
161 {
|
cannam@167
|
162 E T1u, T1x, T1v, T7L, T1t, T1w;
|
cannam@167
|
163 T1u = cr[WS(rs, 12)];
|
cannam@167
|
164 T1x = ci[WS(rs, 12)];
|
cannam@167
|
165 T1t = W[22];
|
cannam@167
|
166 T1v = T1t * T1u;
|
cannam@167
|
167 T7L = T1t * T1x;
|
cannam@167
|
168 T1w = W[23];
|
cannam@167
|
169 T1y = FMA(T1w, T1x, T1v);
|
cannam@167
|
170 T7M = FNMS(T1w, T1u, T7L);
|
cannam@167
|
171 }
|
cannam@167
|
172 {
|
cannam@167
|
173 E T1s, T1F, TeX, TeY;
|
cannam@167
|
174 T1s = T1l + T1r;
|
cannam@167
|
175 T1F = T1y + T1E;
|
cannam@167
|
176 T1G = T1s + T1F;
|
cannam@167
|
177 TeW = T1s - T1F;
|
cannam@167
|
178 TeX = T7S + T7U;
|
cannam@167
|
179 TeY = T7M + T7O;
|
cannam@167
|
180 TeZ = TeX - TeY;
|
cannam@167
|
181 Thr = TeX + TeY;
|
cannam@167
|
182 }
|
cannam@167
|
183 {
|
cannam@167
|
184 E T7K, T7P, T7V, T7W;
|
cannam@167
|
185 T7K = T1l - T1r;
|
cannam@167
|
186 T7P = T7M - T7O;
|
cannam@167
|
187 T7Q = T7K - T7P;
|
cannam@167
|
188 TcI = T7K + T7P;
|
cannam@167
|
189 T7V = T7S - T7U;
|
cannam@167
|
190 T7W = T1y - T1E;
|
cannam@167
|
191 T7X = T7V + T7W;
|
cannam@167
|
192 TcJ = T7V - T7W;
|
cannam@167
|
193 }
|
cannam@167
|
194 }
|
cannam@167
|
195 {
|
cannam@167
|
196 E T1O, T8p, T27, T85, T1U, T8r, T21, T83;
|
cannam@167
|
197 {
|
cannam@167
|
198 E T1K, T1N, T1L, T8o, T1J, T1M;
|
cannam@167
|
199 T1K = cr[WS(rs, 2)];
|
cannam@167
|
200 T1N = ci[WS(rs, 2)];
|
cannam@167
|
201 T1J = W[2];
|
cannam@167
|
202 T1L = T1J * T1K;
|
cannam@167
|
203 T8o = T1J * T1N;
|
cannam@167
|
204 T1M = W[3];
|
cannam@167
|
205 T1O = FMA(T1M, T1N, T1L);
|
cannam@167
|
206 T8p = FNMS(T1M, T1K, T8o);
|
cannam@167
|
207 }
|
cannam@167
|
208 {
|
cannam@167
|
209 E T23, T26, T24, T84, T22, T25;
|
cannam@167
|
210 T23 = cr[WS(rs, 50)];
|
cannam@167
|
211 T26 = ci[WS(rs, 50)];
|
cannam@167
|
212 T22 = W[98];
|
cannam@167
|
213 T24 = T22 * T23;
|
cannam@167
|
214 T84 = T22 * T26;
|
cannam@167
|
215 T25 = W[99];
|
cannam@167
|
216 T27 = FMA(T25, T26, T24);
|
cannam@167
|
217 T85 = FNMS(T25, T23, T84);
|
cannam@167
|
218 }
|
cannam@167
|
219 {
|
cannam@167
|
220 E T1Q, T1T, T1R, T8q, T1P, T1S;
|
cannam@167
|
221 T1Q = cr[WS(rs, 34)];
|
cannam@167
|
222 T1T = ci[WS(rs, 34)];
|
cannam@167
|
223 T1P = W[66];
|
cannam@167
|
224 T1R = T1P * T1Q;
|
cannam@167
|
225 T8q = T1P * T1T;
|
cannam@167
|
226 T1S = W[67];
|
cannam@167
|
227 T1U = FMA(T1S, T1T, T1R);
|
cannam@167
|
228 T8r = FNMS(T1S, T1Q, T8q);
|
cannam@167
|
229 }
|
cannam@167
|
230 {
|
cannam@167
|
231 E T1X, T20, T1Y, T82, T1W, T1Z;
|
cannam@167
|
232 T1X = cr[WS(rs, 18)];
|
cannam@167
|
233 T20 = ci[WS(rs, 18)];
|
cannam@167
|
234 T1W = W[34];
|
cannam@167
|
235 T1Y = T1W * T1X;
|
cannam@167
|
236 T82 = T1W * T20;
|
cannam@167
|
237 T1Z = W[35];
|
cannam@167
|
238 T21 = FMA(T1Z, T20, T1Y);
|
cannam@167
|
239 T83 = FNMS(T1Z, T1X, T82);
|
cannam@167
|
240 }
|
cannam@167
|
241 {
|
cannam@167
|
242 E T1V, T28, Tf3, Tf4;
|
cannam@167
|
243 T1V = T1O + T1U;
|
cannam@167
|
244 T28 = T21 + T27;
|
cannam@167
|
245 T29 = T1V + T28;
|
cannam@167
|
246 Tf8 = T1V - T28;
|
cannam@167
|
247 Tf3 = T8p + T8r;
|
cannam@167
|
248 Tf4 = T83 + T85;
|
cannam@167
|
249 Tf5 = Tf3 - Tf4;
|
cannam@167
|
250 Thw = Tf3 + Tf4;
|
cannam@167
|
251 }
|
cannam@167
|
252 {
|
cannam@167
|
253 E T81, T86, T8s, T8t;
|
cannam@167
|
254 T81 = T1O - T1U;
|
cannam@167
|
255 T86 = T83 - T85;
|
cannam@167
|
256 T87 = T81 - T86;
|
cannam@167
|
257 TcN = T81 + T86;
|
cannam@167
|
258 T8s = T8p - T8r;
|
cannam@167
|
259 T8t = T21 - T27;
|
cannam@167
|
260 T8u = T8s + T8t;
|
cannam@167
|
261 TcQ = T8s - T8t;
|
cannam@167
|
262 }
|
cannam@167
|
263 }
|
cannam@167
|
264 {
|
cannam@167
|
265 E T5p, Tbf, T5I, Tao, T5v, Tbh, T5C, Tam;
|
cannam@167
|
266 {
|
cannam@167
|
267 E T5l, T5o, T5m, Tbe, T5k, T5n;
|
cannam@167
|
268 T5l = cr[WS(rs, 63)];
|
cannam@167
|
269 T5o = ci[WS(rs, 63)];
|
cannam@167
|
270 T5k = W[124];
|
cannam@167
|
271 T5m = T5k * T5l;
|
cannam@167
|
272 Tbe = T5k * T5o;
|
cannam@167
|
273 T5n = W[125];
|
cannam@167
|
274 T5p = FMA(T5n, T5o, T5m);
|
cannam@167
|
275 Tbf = FNMS(T5n, T5l, Tbe);
|
cannam@167
|
276 }
|
cannam@167
|
277 {
|
cannam@167
|
278 E T5E, T5H, T5F, Tan, T5D, T5G;
|
cannam@167
|
279 T5E = cr[WS(rs, 47)];
|
cannam@167
|
280 T5H = ci[WS(rs, 47)];
|
cannam@167
|
281 T5D = W[92];
|
cannam@167
|
282 T5F = T5D * T5E;
|
cannam@167
|
283 Tan = T5D * T5H;
|
cannam@167
|
284 T5G = W[93];
|
cannam@167
|
285 T5I = FMA(T5G, T5H, T5F);
|
cannam@167
|
286 Tao = FNMS(T5G, T5E, Tan);
|
cannam@167
|
287 }
|
cannam@167
|
288 {
|
cannam@167
|
289 E T5r, T5u, T5s, Tbg, T5q, T5t;
|
cannam@167
|
290 T5r = cr[WS(rs, 31)];
|
cannam@167
|
291 T5u = ci[WS(rs, 31)];
|
cannam@167
|
292 T5q = W[60];
|
cannam@167
|
293 T5s = T5q * T5r;
|
cannam@167
|
294 Tbg = T5q * T5u;
|
cannam@167
|
295 T5t = W[61];
|
cannam@167
|
296 T5v = FMA(T5t, T5u, T5s);
|
cannam@167
|
297 Tbh = FNMS(T5t, T5r, Tbg);
|
cannam@167
|
298 }
|
cannam@167
|
299 {
|
cannam@167
|
300 E T5y, T5B, T5z, Tal, T5x, T5A;
|
cannam@167
|
301 T5y = cr[WS(rs, 15)];
|
cannam@167
|
302 T5B = ci[WS(rs, 15)];
|
cannam@167
|
303 T5x = W[28];
|
cannam@167
|
304 T5z = T5x * T5y;
|
cannam@167
|
305 Tal = T5x * T5B;
|
cannam@167
|
306 T5A = W[29];
|
cannam@167
|
307 T5C = FMA(T5A, T5B, T5z);
|
cannam@167
|
308 Tam = FNMS(T5A, T5y, Tal);
|
cannam@167
|
309 }
|
cannam@167
|
310 {
|
cannam@167
|
311 E T5w, T5J, Tga, Tgb;
|
cannam@167
|
312 T5w = T5p + T5v;
|
cannam@167
|
313 T5J = T5C + T5I;
|
cannam@167
|
314 T5K = T5w + T5J;
|
cannam@167
|
315 TfS = T5w - T5J;
|
cannam@167
|
316 Tga = Tbf + Tbh;
|
cannam@167
|
317 Tgb = Tam + Tao;
|
cannam@167
|
318 Tgc = Tga - Tgb;
|
cannam@167
|
319 ThX = Tga + Tgb;
|
cannam@167
|
320 }
|
cannam@167
|
321 {
|
cannam@167
|
322 E Tak, Tap, Tbd, Tbi;
|
cannam@167
|
323 Tak = T5p - T5v;
|
cannam@167
|
324 Tap = Tam - Tao;
|
cannam@167
|
325 Taq = Tak - Tap;
|
cannam@167
|
326 Tdm = Tak + Tap;
|
cannam@167
|
327 Tbd = T5I - T5C;
|
cannam@167
|
328 Tbi = Tbf - Tbh;
|
cannam@167
|
329 Tbj = Tbd - Tbi;
|
cannam@167
|
330 Tdx = Tbi + Tbd;
|
cannam@167
|
331 }
|
cannam@167
|
332 }
|
cannam@167
|
333 {
|
cannam@167
|
334 E Ts, T7h, TL, T7q, Ty, T7j, TF, T7o;
|
cannam@167
|
335 {
|
cannam@167
|
336 E To, Tr, Tp, T7g, Tn, Tq;
|
cannam@167
|
337 To = cr[WS(rs, 8)];
|
cannam@167
|
338 Tr = ci[WS(rs, 8)];
|
cannam@167
|
339 Tn = W[14];
|
cannam@167
|
340 Tp = Tn * To;
|
cannam@167
|
341 T7g = Tn * Tr;
|
cannam@167
|
342 Tq = W[15];
|
cannam@167
|
343 Ts = FMA(Tq, Tr, Tp);
|
cannam@167
|
344 T7h = FNMS(Tq, To, T7g);
|
cannam@167
|
345 }
|
cannam@167
|
346 {
|
cannam@167
|
347 E TH, TK, TI, T7p, TG, TJ;
|
cannam@167
|
348 TH = cr[WS(rs, 24)];
|
cannam@167
|
349 TK = ci[WS(rs, 24)];
|
cannam@167
|
350 TG = W[46];
|
cannam@167
|
351 TI = TG * TH;
|
cannam@167
|
352 T7p = TG * TK;
|
cannam@167
|
353 TJ = W[47];
|
cannam@167
|
354 TL = FMA(TJ, TK, TI);
|
cannam@167
|
355 T7q = FNMS(TJ, TH, T7p);
|
cannam@167
|
356 }
|
cannam@167
|
357 {
|
cannam@167
|
358 E Tu, Tx, Tv, T7i, Tt, Tw;
|
cannam@167
|
359 Tu = cr[WS(rs, 40)];
|
cannam@167
|
360 Tx = ci[WS(rs, 40)];
|
cannam@167
|
361 Tt = W[78];
|
cannam@167
|
362 Tv = Tt * Tu;
|
cannam@167
|
363 T7i = Tt * Tx;
|
cannam@167
|
364 Tw = W[79];
|
cannam@167
|
365 Ty = FMA(Tw, Tx, Tv);
|
cannam@167
|
366 T7j = FNMS(Tw, Tu, T7i);
|
cannam@167
|
367 }
|
cannam@167
|
368 {
|
cannam@167
|
369 E TB, TE, TC, T7n, TA, TD;
|
cannam@167
|
370 TB = cr[WS(rs, 56)];
|
cannam@167
|
371 TE = ci[WS(rs, 56)];
|
cannam@167
|
372 TA = W[110];
|
cannam@167
|
373 TC = TA * TB;
|
cannam@167
|
374 T7n = TA * TE;
|
cannam@167
|
375 TD = W[111];
|
cannam@167
|
376 TF = FMA(TD, TE, TC);
|
cannam@167
|
377 T7o = FNMS(TD, TB, T7n);
|
cannam@167
|
378 }
|
cannam@167
|
379 {
|
cannam@167
|
380 E Tz, TM, TeN, TeO;
|
cannam@167
|
381 Tz = Ts + Ty;
|
cannam@167
|
382 TM = TF + TL;
|
cannam@167
|
383 TN = Tz + TM;
|
cannam@167
|
384 Tjl = Tz - TM;
|
cannam@167
|
385 TeN = T7o + T7q;
|
cannam@167
|
386 TeO = T7h + T7j;
|
cannam@167
|
387 TeP = TeN - TeO;
|
cannam@167
|
388 TiP = TeO + TeN;
|
cannam@167
|
389 }
|
cannam@167
|
390 {
|
cannam@167
|
391 E T7f, T7k, T7m, T7r;
|
cannam@167
|
392 T7f = Ts - Ty;
|
cannam@167
|
393 T7k = T7h - T7j;
|
cannam@167
|
394 T7l = T7f - T7k;
|
cannam@167
|
395 TcB = T7f + T7k;
|
cannam@167
|
396 T7m = TF - TL;
|
cannam@167
|
397 T7r = T7o - T7q;
|
cannam@167
|
398 T7s = T7m + T7r;
|
cannam@167
|
399 TcC = T7m - T7r;
|
cannam@167
|
400 }
|
cannam@167
|
401 }
|
cannam@167
|
402 {
|
cannam@167
|
403 E TU, T7D, T1d, T7z, T10, T7F, T17, T7x;
|
cannam@167
|
404 {
|
cannam@167
|
405 E TQ, TT, TR, T7C, TP, TS;
|
cannam@167
|
406 TQ = cr[WS(rs, 4)];
|
cannam@167
|
407 TT = ci[WS(rs, 4)];
|
cannam@167
|
408 TP = W[6];
|
cannam@167
|
409 TR = TP * TQ;
|
cannam@167
|
410 T7C = TP * TT;
|
cannam@167
|
411 TS = W[7];
|
cannam@167
|
412 TU = FMA(TS, TT, TR);
|
cannam@167
|
413 T7D = FNMS(TS, TQ, T7C);
|
cannam@167
|
414 }
|
cannam@167
|
415 {
|
cannam@167
|
416 E T19, T1c, T1a, T7y, T18, T1b;
|
cannam@167
|
417 T19 = cr[WS(rs, 52)];
|
cannam@167
|
418 T1c = ci[WS(rs, 52)];
|
cannam@167
|
419 T18 = W[102];
|
cannam@167
|
420 T1a = T18 * T19;
|
cannam@167
|
421 T7y = T18 * T1c;
|
cannam@167
|
422 T1b = W[103];
|
cannam@167
|
423 T1d = FMA(T1b, T1c, T1a);
|
cannam@167
|
424 T7z = FNMS(T1b, T19, T7y);
|
cannam@167
|
425 }
|
cannam@167
|
426 {
|
cannam@167
|
427 E TW, TZ, TX, T7E, TV, TY;
|
cannam@167
|
428 TW = cr[WS(rs, 36)];
|
cannam@167
|
429 TZ = ci[WS(rs, 36)];
|
cannam@167
|
430 TV = W[70];
|
cannam@167
|
431 TX = TV * TW;
|
cannam@167
|
432 T7E = TV * TZ;
|
cannam@167
|
433 TY = W[71];
|
cannam@167
|
434 T10 = FMA(TY, TZ, TX);
|
cannam@167
|
435 T7F = FNMS(TY, TW, T7E);
|
cannam@167
|
436 }
|
cannam@167
|
437 {
|
cannam@167
|
438 E T13, T16, T14, T7w, T12, T15;
|
cannam@167
|
439 T13 = cr[WS(rs, 20)];
|
cannam@167
|
440 T16 = ci[WS(rs, 20)];
|
cannam@167
|
441 T12 = W[38];
|
cannam@167
|
442 T14 = T12 * T13;
|
cannam@167
|
443 T7w = T12 * T16;
|
cannam@167
|
444 T15 = W[39];
|
cannam@167
|
445 T17 = FMA(T15, T16, T14);
|
cannam@167
|
446 T7x = FNMS(T15, T13, T7w);
|
cannam@167
|
447 }
|
cannam@167
|
448 {
|
cannam@167
|
449 E T11, T1e, TeS, TeT;
|
cannam@167
|
450 T11 = TU + T10;
|
cannam@167
|
451 T1e = T17 + T1d;
|
cannam@167
|
452 T1f = T11 + T1e;
|
cannam@167
|
453 TeR = T11 - T1e;
|
cannam@167
|
454 TeS = T7D + T7F;
|
cannam@167
|
455 TeT = T7x + T7z;
|
cannam@167
|
456 TeU = TeS - TeT;
|
cannam@167
|
457 Ths = TeS + TeT;
|
cannam@167
|
458 }
|
cannam@167
|
459 {
|
cannam@167
|
460 E T7v, T7A, T7G, T7H;
|
cannam@167
|
461 T7v = TU - T10;
|
cannam@167
|
462 T7A = T7x - T7z;
|
cannam@167
|
463 T7B = T7v - T7A;
|
cannam@167
|
464 TcF = T7v + T7A;
|
cannam@167
|
465 T7G = T7D - T7F;
|
cannam@167
|
466 T7H = T17 - T1d;
|
cannam@167
|
467 T7I = T7G + T7H;
|
cannam@167
|
468 TcG = T7G - T7H;
|
cannam@167
|
469 }
|
cannam@167
|
470 }
|
cannam@167
|
471 {
|
cannam@167
|
472 E T2H, T8Y, T30, T8E, T2N, T90, T2U, T8C;
|
cannam@167
|
473 {
|
cannam@167
|
474 E T2D, T2G, T2E, T8X, T2C, T2F;
|
cannam@167
|
475 T2D = cr[WS(rs, 62)];
|
cannam@167
|
476 T2G = ci[WS(rs, 62)];
|
cannam@167
|
477 T2C = W[122];
|
cannam@167
|
478 T2E = T2C * T2D;
|
cannam@167
|
479 T8X = T2C * T2G;
|
cannam@167
|
480 T2F = W[123];
|
cannam@167
|
481 T2H = FMA(T2F, T2G, T2E);
|
cannam@167
|
482 T8Y = FNMS(T2F, T2D, T8X);
|
cannam@167
|
483 }
|
cannam@167
|
484 {
|
cannam@167
|
485 E T2W, T2Z, T2X, T8D, T2V, T2Y;
|
cannam@167
|
486 T2W = cr[WS(rs, 46)];
|
cannam@167
|
487 T2Z = ci[WS(rs, 46)];
|
cannam@167
|
488 T2V = W[90];
|
cannam@167
|
489 T2X = T2V * T2W;
|
cannam@167
|
490 T8D = T2V * T2Z;
|
cannam@167
|
491 T2Y = W[91];
|
cannam@167
|
492 T30 = FMA(T2Y, T2Z, T2X);
|
cannam@167
|
493 T8E = FNMS(T2Y, T2W, T8D);
|
cannam@167
|
494 }
|
cannam@167
|
495 {
|
cannam@167
|
496 E T2J, T2M, T2K, T8Z, T2I, T2L;
|
cannam@167
|
497 T2J = cr[WS(rs, 30)];
|
cannam@167
|
498 T2M = ci[WS(rs, 30)];
|
cannam@167
|
499 T2I = W[58];
|
cannam@167
|
500 T2K = T2I * T2J;
|
cannam@167
|
501 T8Z = T2I * T2M;
|
cannam@167
|
502 T2L = W[59];
|
cannam@167
|
503 T2N = FMA(T2L, T2M, T2K);
|
cannam@167
|
504 T90 = FNMS(T2L, T2J, T8Z);
|
cannam@167
|
505 }
|
cannam@167
|
506 {
|
cannam@167
|
507 E T2Q, T2T, T2R, T8B, T2P, T2S;
|
cannam@167
|
508 T2Q = cr[WS(rs, 14)];
|
cannam@167
|
509 T2T = ci[WS(rs, 14)];
|
cannam@167
|
510 T2P = W[26];
|
cannam@167
|
511 T2R = T2P * T2Q;
|
cannam@167
|
512 T8B = T2P * T2T;
|
cannam@167
|
513 T2S = W[27];
|
cannam@167
|
514 T2U = FMA(T2S, T2T, T2R);
|
cannam@167
|
515 T8C = FNMS(T2S, T2Q, T8B);
|
cannam@167
|
516 }
|
cannam@167
|
517 {
|
cannam@167
|
518 E T2O, T31, Tfe, Tff;
|
cannam@167
|
519 T2O = T2H + T2N;
|
cannam@167
|
520 T31 = T2U + T30;
|
cannam@167
|
521 T32 = T2O + T31;
|
cannam@167
|
522 Tfj = T2O - T31;
|
cannam@167
|
523 Tfe = T8Y + T90;
|
cannam@167
|
524 Tff = T8C + T8E;
|
cannam@167
|
525 Tfg = Tfe - Tff;
|
cannam@167
|
526 ThB = Tfe + Tff;
|
cannam@167
|
527 }
|
cannam@167
|
528 {
|
cannam@167
|
529 E T8A, T8F, T91, T92;
|
cannam@167
|
530 T8A = T2H - T2N;
|
cannam@167
|
531 T8F = T8C - T8E;
|
cannam@167
|
532 T8G = T8A - T8F;
|
cannam@167
|
533 TcU = T8A + T8F;
|
cannam@167
|
534 T91 = T8Y - T90;
|
cannam@167
|
535 T92 = T2U - T30;
|
cannam@167
|
536 T93 = T91 + T92;
|
cannam@167
|
537 TcX = T91 - T92;
|
cannam@167
|
538 }
|
cannam@167
|
539 }
|
cannam@167
|
540 {
|
cannam@167
|
541 E T3C, Ta5, T3V, T9f, T3I, Ta7, T3P, T9d;
|
cannam@167
|
542 {
|
cannam@167
|
543 E T3y, T3B, T3z, Ta4, T3x, T3A;
|
cannam@167
|
544 T3y = cr[WS(rs, 1)];
|
cannam@167
|
545 T3B = ci[WS(rs, 1)];
|
cannam@167
|
546 T3x = W[0];
|
cannam@167
|
547 T3z = T3x * T3y;
|
cannam@167
|
548 Ta4 = T3x * T3B;
|
cannam@167
|
549 T3A = W[1];
|
cannam@167
|
550 T3C = FMA(T3A, T3B, T3z);
|
cannam@167
|
551 Ta5 = FNMS(T3A, T3y, Ta4);
|
cannam@167
|
552 }
|
cannam@167
|
553 {
|
cannam@167
|
554 E T3R, T3U, T3S, T9e, T3Q, T3T;
|
cannam@167
|
555 T3R = cr[WS(rs, 49)];
|
cannam@167
|
556 T3U = ci[WS(rs, 49)];
|
cannam@167
|
557 T3Q = W[96];
|
cannam@167
|
558 T3S = T3Q * T3R;
|
cannam@167
|
559 T9e = T3Q * T3U;
|
cannam@167
|
560 T3T = W[97];
|
cannam@167
|
561 T3V = FMA(T3T, T3U, T3S);
|
cannam@167
|
562 T9f = FNMS(T3T, T3R, T9e);
|
cannam@167
|
563 }
|
cannam@167
|
564 {
|
cannam@167
|
565 E T3E, T3H, T3F, Ta6, T3D, T3G;
|
cannam@167
|
566 T3E = cr[WS(rs, 33)];
|
cannam@167
|
567 T3H = ci[WS(rs, 33)];
|
cannam@167
|
568 T3D = W[64];
|
cannam@167
|
569 T3F = T3D * T3E;
|
cannam@167
|
570 Ta6 = T3D * T3H;
|
cannam@167
|
571 T3G = W[65];
|
cannam@167
|
572 T3I = FMA(T3G, T3H, T3F);
|
cannam@167
|
573 Ta7 = FNMS(T3G, T3E, Ta6);
|
cannam@167
|
574 }
|
cannam@167
|
575 {
|
cannam@167
|
576 E T3L, T3O, T3M, T9c, T3K, T3N;
|
cannam@167
|
577 T3L = cr[WS(rs, 17)];
|
cannam@167
|
578 T3O = ci[WS(rs, 17)];
|
cannam@167
|
579 T3K = W[32];
|
cannam@167
|
580 T3M = T3K * T3L;
|
cannam@167
|
581 T9c = T3K * T3O;
|
cannam@167
|
582 T3N = W[33];
|
cannam@167
|
583 T3P = FMA(T3N, T3O, T3M);
|
cannam@167
|
584 T9d = FNMS(T3N, T3L, T9c);
|
cannam@167
|
585 }
|
cannam@167
|
586 {
|
cannam@167
|
587 E T3J, T3W, TfI, TfJ;
|
cannam@167
|
588 T3J = T3C + T3I;
|
cannam@167
|
589 T3W = T3P + T3V;
|
cannam@167
|
590 T3X = T3J + T3W;
|
cannam@167
|
591 Tfr = T3J - T3W;
|
cannam@167
|
592 TfI = Ta5 + Ta7;
|
cannam@167
|
593 TfJ = T9d + T9f;
|
cannam@167
|
594 TfK = TfI - TfJ;
|
cannam@167
|
595 ThM = TfI + TfJ;
|
cannam@167
|
596 }
|
cannam@167
|
597 {
|
cannam@167
|
598 E T9b, T9g, Ta8, Ta9;
|
cannam@167
|
599 T9b = T3C - T3I;
|
cannam@167
|
600 T9g = T9d - T9f;
|
cannam@167
|
601 T9h = T9b - T9g;
|
cannam@167
|
602 Td3 = T9b + T9g;
|
cannam@167
|
603 Ta8 = Ta5 - Ta7;
|
cannam@167
|
604 Ta9 = T3P - T3V;
|
cannam@167
|
605 Taa = Ta8 + Ta9;
|
cannam@167
|
606 Tde = Ta8 - Ta9;
|
cannam@167
|
607 }
|
cannam@167
|
608 }
|
cannam@167
|
609 {
|
cannam@167
|
610 E T2f, T8a, T2y, T8j, T2l, T8c, T2s, T8h;
|
cannam@167
|
611 {
|
cannam@167
|
612 E T2b, T2e, T2c, T89, T2a, T2d;
|
cannam@167
|
613 T2b = cr[WS(rs, 10)];
|
cannam@167
|
614 T2e = ci[WS(rs, 10)];
|
cannam@167
|
615 T2a = W[18];
|
cannam@167
|
616 T2c = T2a * T2b;
|
cannam@167
|
617 T89 = T2a * T2e;
|
cannam@167
|
618 T2d = W[19];
|
cannam@167
|
619 T2f = FMA(T2d, T2e, T2c);
|
cannam@167
|
620 T8a = FNMS(T2d, T2b, T89);
|
cannam@167
|
621 }
|
cannam@167
|
622 {
|
cannam@167
|
623 E T2u, T2x, T2v, T8i, T2t, T2w;
|
cannam@167
|
624 T2u = cr[WS(rs, 26)];
|
cannam@167
|
625 T2x = ci[WS(rs, 26)];
|
cannam@167
|
626 T2t = W[50];
|
cannam@167
|
627 T2v = T2t * T2u;
|
cannam@167
|
628 T8i = T2t * T2x;
|
cannam@167
|
629 T2w = W[51];
|
cannam@167
|
630 T2y = FMA(T2w, T2x, T2v);
|
cannam@167
|
631 T8j = FNMS(T2w, T2u, T8i);
|
cannam@167
|
632 }
|
cannam@167
|
633 {
|
cannam@167
|
634 E T2h, T2k, T2i, T8b, T2g, T2j;
|
cannam@167
|
635 T2h = cr[WS(rs, 42)];
|
cannam@167
|
636 T2k = ci[WS(rs, 42)];
|
cannam@167
|
637 T2g = W[82];
|
cannam@167
|
638 T2i = T2g * T2h;
|
cannam@167
|
639 T8b = T2g * T2k;
|
cannam@167
|
640 T2j = W[83];
|
cannam@167
|
641 T2l = FMA(T2j, T2k, T2i);
|
cannam@167
|
642 T8c = FNMS(T2j, T2h, T8b);
|
cannam@167
|
643 }
|
cannam@167
|
644 {
|
cannam@167
|
645 E T2o, T2r, T2p, T8g, T2n, T2q;
|
cannam@167
|
646 T2o = cr[WS(rs, 58)];
|
cannam@167
|
647 T2r = ci[WS(rs, 58)];
|
cannam@167
|
648 T2n = W[114];
|
cannam@167
|
649 T2p = T2n * T2o;
|
cannam@167
|
650 T8g = T2n * T2r;
|
cannam@167
|
651 T2q = W[115];
|
cannam@167
|
652 T2s = FMA(T2q, T2r, T2p);
|
cannam@167
|
653 T8h = FNMS(T2q, T2o, T8g);
|
cannam@167
|
654 }
|
cannam@167
|
655 {
|
cannam@167
|
656 E T2m, T2z, Tf9, Tfa;
|
cannam@167
|
657 T2m = T2f + T2l;
|
cannam@167
|
658 T2z = T2s + T2y;
|
cannam@167
|
659 T2A = T2m + T2z;
|
cannam@167
|
660 Tf6 = T2m - T2z;
|
cannam@167
|
661 Tf9 = T8h + T8j;
|
cannam@167
|
662 Tfa = T8a + T8c;
|
cannam@167
|
663 Tfb = Tf9 - Tfa;
|
cannam@167
|
664 Thx = Tfa + Tf9;
|
cannam@167
|
665 {
|
cannam@167
|
666 E T8e, T8v, T8l, T8w;
|
cannam@167
|
667 {
|
cannam@167
|
668 E T88, T8d, T8f, T8k;
|
cannam@167
|
669 T88 = T2f - T2l;
|
cannam@167
|
670 T8d = T8a - T8c;
|
cannam@167
|
671 T8e = T88 - T8d;
|
cannam@167
|
672 T8v = T88 + T8d;
|
cannam@167
|
673 T8f = T2s - T2y;
|
cannam@167
|
674 T8k = T8h - T8j;
|
cannam@167
|
675 T8l = T8f + T8k;
|
cannam@167
|
676 T8w = T8k - T8f;
|
cannam@167
|
677 }
|
cannam@167
|
678 T8m = T8e + T8l;
|
cannam@167
|
679 TcR = T8l - T8e;
|
cannam@167
|
680 T8x = T8v + T8w;
|
cannam@167
|
681 TcO = T8v - T8w;
|
cannam@167
|
682 }
|
cannam@167
|
683 }
|
cannam@167
|
684 }
|
cannam@167
|
685 {
|
cannam@167
|
686 E T38, T8J, T3r, T8S, T3e, T8L, T3l, T8Q;
|
cannam@167
|
687 {
|
cannam@167
|
688 E T34, T37, T35, T8I, T33, T36;
|
cannam@167
|
689 T34 = cr[WS(rs, 6)];
|
cannam@167
|
690 T37 = ci[WS(rs, 6)];
|
cannam@167
|
691 T33 = W[10];
|
cannam@167
|
692 T35 = T33 * T34;
|
cannam@167
|
693 T8I = T33 * T37;
|
cannam@167
|
694 T36 = W[11];
|
cannam@167
|
695 T38 = FMA(T36, T37, T35);
|
cannam@167
|
696 T8J = FNMS(T36, T34, T8I);
|
cannam@167
|
697 }
|
cannam@167
|
698 {
|
cannam@167
|
699 E T3n, T3q, T3o, T8R, T3m, T3p;
|
cannam@167
|
700 T3n = cr[WS(rs, 22)];
|
cannam@167
|
701 T3q = ci[WS(rs, 22)];
|
cannam@167
|
702 T3m = W[42];
|
cannam@167
|
703 T3o = T3m * T3n;
|
cannam@167
|
704 T8R = T3m * T3q;
|
cannam@167
|
705 T3p = W[43];
|
cannam@167
|
706 T3r = FMA(T3p, T3q, T3o);
|
cannam@167
|
707 T8S = FNMS(T3p, T3n, T8R);
|
cannam@167
|
708 }
|
cannam@167
|
709 {
|
cannam@167
|
710 E T3a, T3d, T3b, T8K, T39, T3c;
|
cannam@167
|
711 T3a = cr[WS(rs, 38)];
|
cannam@167
|
712 T3d = ci[WS(rs, 38)];
|
cannam@167
|
713 T39 = W[74];
|
cannam@167
|
714 T3b = T39 * T3a;
|
cannam@167
|
715 T8K = T39 * T3d;
|
cannam@167
|
716 T3c = W[75];
|
cannam@167
|
717 T3e = FMA(T3c, T3d, T3b);
|
cannam@167
|
718 T8L = FNMS(T3c, T3a, T8K);
|
cannam@167
|
719 }
|
cannam@167
|
720 {
|
cannam@167
|
721 E T3h, T3k, T3i, T8P, T3g, T3j;
|
cannam@167
|
722 T3h = cr[WS(rs, 54)];
|
cannam@167
|
723 T3k = ci[WS(rs, 54)];
|
cannam@167
|
724 T3g = W[106];
|
cannam@167
|
725 T3i = T3g * T3h;
|
cannam@167
|
726 T8P = T3g * T3k;
|
cannam@167
|
727 T3j = W[107];
|
cannam@167
|
728 T3l = FMA(T3j, T3k, T3i);
|
cannam@167
|
729 T8Q = FNMS(T3j, T3h, T8P);
|
cannam@167
|
730 }
|
cannam@167
|
731 {
|
cannam@167
|
732 E T3f, T3s, Tfk, Tfl;
|
cannam@167
|
733 T3f = T38 + T3e;
|
cannam@167
|
734 T3s = T3l + T3r;
|
cannam@167
|
735 T3t = T3f + T3s;
|
cannam@167
|
736 Tfh = T3f - T3s;
|
cannam@167
|
737 Tfk = T8Q + T8S;
|
cannam@167
|
738 Tfl = T8J + T8L;
|
cannam@167
|
739 Tfm = Tfk - Tfl;
|
cannam@167
|
740 ThC = Tfl + Tfk;
|
cannam@167
|
741 {
|
cannam@167
|
742 E T8N, T94, T8U, T95;
|
cannam@167
|
743 {
|
cannam@167
|
744 E T8H, T8M, T8O, T8T;
|
cannam@167
|
745 T8H = T38 - T3e;
|
cannam@167
|
746 T8M = T8J - T8L;
|
cannam@167
|
747 T8N = T8H - T8M;
|
cannam@167
|
748 T94 = T8H + T8M;
|
cannam@167
|
749 T8O = T3l - T3r;
|
cannam@167
|
750 T8T = T8Q - T8S;
|
cannam@167
|
751 T8U = T8O + T8T;
|
cannam@167
|
752 T95 = T8T - T8O;
|
cannam@167
|
753 }
|
cannam@167
|
754 T8V = T8N + T8U;
|
cannam@167
|
755 TcY = T8U - T8N;
|
cannam@167
|
756 T96 = T94 + T95;
|
cannam@167
|
757 TcV = T94 - T95;
|
cannam@167
|
758 }
|
cannam@167
|
759 }
|
cannam@167
|
760 }
|
cannam@167
|
761 {
|
cannam@167
|
762 E T43, T9k, T4m, T9t, T49, T9m, T4g, T9r;
|
cannam@167
|
763 {
|
cannam@167
|
764 E T3Z, T42, T40, T9j, T3Y, T41;
|
cannam@167
|
765 T3Z = cr[WS(rs, 9)];
|
cannam@167
|
766 T42 = ci[WS(rs, 9)];
|
cannam@167
|
767 T3Y = W[16];
|
cannam@167
|
768 T40 = T3Y * T3Z;
|
cannam@167
|
769 T9j = T3Y * T42;
|
cannam@167
|
770 T41 = W[17];
|
cannam@167
|
771 T43 = FMA(T41, T42, T40);
|
cannam@167
|
772 T9k = FNMS(T41, T3Z, T9j);
|
cannam@167
|
773 }
|
cannam@167
|
774 {
|
cannam@167
|
775 E T4i, T4l, T4j, T9s, T4h, T4k;
|
cannam@167
|
776 T4i = cr[WS(rs, 25)];
|
cannam@167
|
777 T4l = ci[WS(rs, 25)];
|
cannam@167
|
778 T4h = W[48];
|
cannam@167
|
779 T4j = T4h * T4i;
|
cannam@167
|
780 T9s = T4h * T4l;
|
cannam@167
|
781 T4k = W[49];
|
cannam@167
|
782 T4m = FMA(T4k, T4l, T4j);
|
cannam@167
|
783 T9t = FNMS(T4k, T4i, T9s);
|
cannam@167
|
784 }
|
cannam@167
|
785 {
|
cannam@167
|
786 E T45, T48, T46, T9l, T44, T47;
|
cannam@167
|
787 T45 = cr[WS(rs, 41)];
|
cannam@167
|
788 T48 = ci[WS(rs, 41)];
|
cannam@167
|
789 T44 = W[80];
|
cannam@167
|
790 T46 = T44 * T45;
|
cannam@167
|
791 T9l = T44 * T48;
|
cannam@167
|
792 T47 = W[81];
|
cannam@167
|
793 T49 = FMA(T47, T48, T46);
|
cannam@167
|
794 T9m = FNMS(T47, T45, T9l);
|
cannam@167
|
795 }
|
cannam@167
|
796 {
|
cannam@167
|
797 E T4c, T4f, T4d, T9q, T4b, T4e;
|
cannam@167
|
798 T4c = cr[WS(rs, 57)];
|
cannam@167
|
799 T4f = ci[WS(rs, 57)];
|
cannam@167
|
800 T4b = W[112];
|
cannam@167
|
801 T4d = T4b * T4c;
|
cannam@167
|
802 T9q = T4b * T4f;
|
cannam@167
|
803 T4e = W[113];
|
cannam@167
|
804 T4g = FMA(T4e, T4f, T4d);
|
cannam@167
|
805 T9r = FNMS(T4e, T4c, T9q);
|
cannam@167
|
806 }
|
cannam@167
|
807 {
|
cannam@167
|
808 E T4a, T4n, Tfs, Tft;
|
cannam@167
|
809 T4a = T43 + T49;
|
cannam@167
|
810 T4n = T4g + T4m;
|
cannam@167
|
811 T4o = T4a + T4n;
|
cannam@167
|
812 TfL = T4a - T4n;
|
cannam@167
|
813 Tfs = T9r + T9t;
|
cannam@167
|
814 Tft = T9k + T9m;
|
cannam@167
|
815 Tfu = Tfs - Tft;
|
cannam@167
|
816 ThN = Tft + Tfs;
|
cannam@167
|
817 {
|
cannam@167
|
818 E T9o, Tab, T9v, Tac;
|
cannam@167
|
819 {
|
cannam@167
|
820 E T9i, T9n, T9p, T9u;
|
cannam@167
|
821 T9i = T43 - T49;
|
cannam@167
|
822 T9n = T9k - T9m;
|
cannam@167
|
823 T9o = T9i - T9n;
|
cannam@167
|
824 Tab = T9i + T9n;
|
cannam@167
|
825 T9p = T4g - T4m;
|
cannam@167
|
826 T9u = T9r - T9t;
|
cannam@167
|
827 T9v = T9p + T9u;
|
cannam@167
|
828 Tac = T9u - T9p;
|
cannam@167
|
829 }
|
cannam@167
|
830 T9w = T9o + T9v;
|
cannam@167
|
831 Tdf = T9v - T9o;
|
cannam@167
|
832 Tad = Tab + Tac;
|
cannam@167
|
833 Td4 = Tab - Tac;
|
cannam@167
|
834 }
|
cannam@167
|
835 }
|
cannam@167
|
836 }
|
cannam@167
|
837 {
|
cannam@167
|
838 E T5Q, Tat, T69, TaC, T5W, Tav, T63, TaA;
|
cannam@167
|
839 {
|
cannam@167
|
840 E T5M, T5P, T5N, Tas, T5L, T5O;
|
cannam@167
|
841 T5M = cr[WS(rs, 7)];
|
cannam@167
|
842 T5P = ci[WS(rs, 7)];
|
cannam@167
|
843 T5L = W[12];
|
cannam@167
|
844 T5N = T5L * T5M;
|
cannam@167
|
845 Tas = T5L * T5P;
|
cannam@167
|
846 T5O = W[13];
|
cannam@167
|
847 T5Q = FMA(T5O, T5P, T5N);
|
cannam@167
|
848 Tat = FNMS(T5O, T5M, Tas);
|
cannam@167
|
849 }
|
cannam@167
|
850 {
|
cannam@167
|
851 E T65, T68, T66, TaB, T64, T67;
|
cannam@167
|
852 T65 = cr[WS(rs, 23)];
|
cannam@167
|
853 T68 = ci[WS(rs, 23)];
|
cannam@167
|
854 T64 = W[44];
|
cannam@167
|
855 T66 = T64 * T65;
|
cannam@167
|
856 TaB = T64 * T68;
|
cannam@167
|
857 T67 = W[45];
|
cannam@167
|
858 T69 = FMA(T67, T68, T66);
|
cannam@167
|
859 TaC = FNMS(T67, T65, TaB);
|
cannam@167
|
860 }
|
cannam@167
|
861 {
|
cannam@167
|
862 E T5S, T5V, T5T, Tau, T5R, T5U;
|
cannam@167
|
863 T5S = cr[WS(rs, 39)];
|
cannam@167
|
864 T5V = ci[WS(rs, 39)];
|
cannam@167
|
865 T5R = W[76];
|
cannam@167
|
866 T5T = T5R * T5S;
|
cannam@167
|
867 Tau = T5R * T5V;
|
cannam@167
|
868 T5U = W[77];
|
cannam@167
|
869 T5W = FMA(T5U, T5V, T5T);
|
cannam@167
|
870 Tav = FNMS(T5U, T5S, Tau);
|
cannam@167
|
871 }
|
cannam@167
|
872 {
|
cannam@167
|
873 E T5Z, T62, T60, Taz, T5Y, T61;
|
cannam@167
|
874 T5Z = cr[WS(rs, 55)];
|
cannam@167
|
875 T62 = ci[WS(rs, 55)];
|
cannam@167
|
876 T5Y = W[108];
|
cannam@167
|
877 T60 = T5Y * T5Z;
|
cannam@167
|
878 Taz = T5Y * T62;
|
cannam@167
|
879 T61 = W[109];
|
cannam@167
|
880 T63 = FMA(T61, T62, T60);
|
cannam@167
|
881 TaA = FNMS(T61, T5Z, Taz);
|
cannam@167
|
882 }
|
cannam@167
|
883 {
|
cannam@167
|
884 E T5X, T6a, TfT, TfU;
|
cannam@167
|
885 T5X = T5Q + T5W;
|
cannam@167
|
886 T6a = T63 + T69;
|
cannam@167
|
887 T6b = T5X + T6a;
|
cannam@167
|
888 Tg9 = T6a - T5X;
|
cannam@167
|
889 TfT = TaA + TaC;
|
cannam@167
|
890 TfU = Tat + Tav;
|
cannam@167
|
891 TfV = TfT - TfU;
|
cannam@167
|
892 ThY = TfU + TfT;
|
cannam@167
|
893 {
|
cannam@167
|
894 E Tax, Tbl, TaE, Tbk;
|
cannam@167
|
895 {
|
cannam@167
|
896 E Tar, Taw, Tay, TaD;
|
cannam@167
|
897 Tar = T5Q - T5W;
|
cannam@167
|
898 Taw = Tat - Tav;
|
cannam@167
|
899 Tax = Tar - Taw;
|
cannam@167
|
900 Tbl = Tar + Taw;
|
cannam@167
|
901 Tay = T63 - T69;
|
cannam@167
|
902 TaD = TaA - TaC;
|
cannam@167
|
903 TaE = Tay + TaD;
|
cannam@167
|
904 Tbk = Tay - TaD;
|
cannam@167
|
905 }
|
cannam@167
|
906 TaF = Tax + TaE;
|
cannam@167
|
907 Tdy = TaE - Tax;
|
cannam@167
|
908 Tbm = Tbk - Tbl;
|
cannam@167
|
909 Tdn = Tbl + Tbk;
|
cannam@167
|
910 }
|
cannam@167
|
911 }
|
cannam@167
|
912 }
|
cannam@167
|
913 {
|
cannam@167
|
914 E T4v, T9G, T4O, T9C, T4B, T9I, T4I, T9A;
|
cannam@167
|
915 {
|
cannam@167
|
916 E T4r, T4u, T4s, T9F, T4q, T4t;
|
cannam@167
|
917 T4r = cr[WS(rs, 5)];
|
cannam@167
|
918 T4u = ci[WS(rs, 5)];
|
cannam@167
|
919 T4q = W[8];
|
cannam@167
|
920 T4s = T4q * T4r;
|
cannam@167
|
921 T9F = T4q * T4u;
|
cannam@167
|
922 T4t = W[9];
|
cannam@167
|
923 T4v = FMA(T4t, T4u, T4s);
|
cannam@167
|
924 T9G = FNMS(T4t, T4r, T9F);
|
cannam@167
|
925 }
|
cannam@167
|
926 {
|
cannam@167
|
927 E T4K, T4N, T4L, T9B, T4J, T4M;
|
cannam@167
|
928 T4K = cr[WS(rs, 53)];
|
cannam@167
|
929 T4N = ci[WS(rs, 53)];
|
cannam@167
|
930 T4J = W[104];
|
cannam@167
|
931 T4L = T4J * T4K;
|
cannam@167
|
932 T9B = T4J * T4N;
|
cannam@167
|
933 T4M = W[105];
|
cannam@167
|
934 T4O = FMA(T4M, T4N, T4L);
|
cannam@167
|
935 T9C = FNMS(T4M, T4K, T9B);
|
cannam@167
|
936 }
|
cannam@167
|
937 {
|
cannam@167
|
938 E T4x, T4A, T4y, T9H, T4w, T4z;
|
cannam@167
|
939 T4x = cr[WS(rs, 37)];
|
cannam@167
|
940 T4A = ci[WS(rs, 37)];
|
cannam@167
|
941 T4w = W[72];
|
cannam@167
|
942 T4y = T4w * T4x;
|
cannam@167
|
943 T9H = T4w * T4A;
|
cannam@167
|
944 T4z = W[73];
|
cannam@167
|
945 T4B = FMA(T4z, T4A, T4y);
|
cannam@167
|
946 T9I = FNMS(T4z, T4x, T9H);
|
cannam@167
|
947 }
|
cannam@167
|
948 {
|
cannam@167
|
949 E T4E, T4H, T4F, T9z, T4D, T4G;
|
cannam@167
|
950 T4E = cr[WS(rs, 21)];
|
cannam@167
|
951 T4H = ci[WS(rs, 21)];
|
cannam@167
|
952 T4D = W[40];
|
cannam@167
|
953 T4F = T4D * T4E;
|
cannam@167
|
954 T9z = T4D * T4H;
|
cannam@167
|
955 T4G = W[41];
|
cannam@167
|
956 T4I = FMA(T4G, T4H, T4F);
|
cannam@167
|
957 T9A = FNMS(T4G, T4E, T9z);
|
cannam@167
|
958 }
|
cannam@167
|
959 {
|
cannam@167
|
960 E T4C, T4P, Tfw, Tfx, Tfy, Tfz;
|
cannam@167
|
961 T4C = T4v + T4B;
|
cannam@167
|
962 T4P = T4I + T4O;
|
cannam@167
|
963 Tfw = T4C - T4P;
|
cannam@167
|
964 Tfx = T9G + T9I;
|
cannam@167
|
965 Tfy = T9A + T9C;
|
cannam@167
|
966 Tfz = Tfx - Tfy;
|
cannam@167
|
967 T4Q = T4C + T4P;
|
cannam@167
|
968 ThJ = Tfx + Tfy;
|
cannam@167
|
969 TfA = Tfw - Tfz;
|
cannam@167
|
970 TfN = Tfw + Tfz;
|
cannam@167
|
971 }
|
cannam@167
|
972 {
|
cannam@167
|
973 E T9E, Td6, T9L, Td7;
|
cannam@167
|
974 {
|
cannam@167
|
975 E T9y, T9D, T9J, T9K;
|
cannam@167
|
976 T9y = T4v - T4B;
|
cannam@167
|
977 T9D = T9A - T9C;
|
cannam@167
|
978 T9E = T9y - T9D;
|
cannam@167
|
979 Td6 = T9y + T9D;
|
cannam@167
|
980 T9J = T9G - T9I;
|
cannam@167
|
981 T9K = T4I - T4O;
|
cannam@167
|
982 T9L = T9J + T9K;
|
cannam@167
|
983 Td7 = T9J - T9K;
|
cannam@167
|
984 }
|
cannam@167
|
985 T9M = FNMS(KP414213562, T9L, T9E);
|
cannam@167
|
986 Tdh = FNMS(KP414213562, Td6, Td7);
|
cannam@167
|
987 Taf = FMA(KP414213562, T9E, T9L);
|
cannam@167
|
988 Td8 = FMA(KP414213562, Td7, Td6);
|
cannam@167
|
989 }
|
cannam@167
|
990 }
|
cannam@167
|
991 {
|
cannam@167
|
992 E T4W, T9V, T5f, T9R, T52, T9X, T59, T9P;
|
cannam@167
|
993 {
|
cannam@167
|
994 E T4S, T4V, T4T, T9U, T4R, T4U;
|
cannam@167
|
995 T4S = cr[WS(rs, 61)];
|
cannam@167
|
996 T4V = ci[WS(rs, 61)];
|
cannam@167
|
997 T4R = W[120];
|
cannam@167
|
998 T4T = T4R * T4S;
|
cannam@167
|
999 T9U = T4R * T4V;
|
cannam@167
|
1000 T4U = W[121];
|
cannam@167
|
1001 T4W = FMA(T4U, T4V, T4T);
|
cannam@167
|
1002 T9V = FNMS(T4U, T4S, T9U);
|
cannam@167
|
1003 }
|
cannam@167
|
1004 {
|
cannam@167
|
1005 E T5b, T5e, T5c, T9Q, T5a, T5d;
|
cannam@167
|
1006 T5b = cr[WS(rs, 45)];
|
cannam@167
|
1007 T5e = ci[WS(rs, 45)];
|
cannam@167
|
1008 T5a = W[88];
|
cannam@167
|
1009 T5c = T5a * T5b;
|
cannam@167
|
1010 T9Q = T5a * T5e;
|
cannam@167
|
1011 T5d = W[89];
|
cannam@167
|
1012 T5f = FMA(T5d, T5e, T5c);
|
cannam@167
|
1013 T9R = FNMS(T5d, T5b, T9Q);
|
cannam@167
|
1014 }
|
cannam@167
|
1015 {
|
cannam@167
|
1016 E T4Y, T51, T4Z, T9W, T4X, T50;
|
cannam@167
|
1017 T4Y = cr[WS(rs, 29)];
|
cannam@167
|
1018 T51 = ci[WS(rs, 29)];
|
cannam@167
|
1019 T4X = W[56];
|
cannam@167
|
1020 T4Z = T4X * T4Y;
|
cannam@167
|
1021 T9W = T4X * T51;
|
cannam@167
|
1022 T50 = W[57];
|
cannam@167
|
1023 T52 = FMA(T50, T51, T4Z);
|
cannam@167
|
1024 T9X = FNMS(T50, T4Y, T9W);
|
cannam@167
|
1025 }
|
cannam@167
|
1026 {
|
cannam@167
|
1027 E T55, T58, T56, T9O, T54, T57;
|
cannam@167
|
1028 T55 = cr[WS(rs, 13)];
|
cannam@167
|
1029 T58 = ci[WS(rs, 13)];
|
cannam@167
|
1030 T54 = W[24];
|
cannam@167
|
1031 T56 = T54 * T55;
|
cannam@167
|
1032 T9O = T54 * T58;
|
cannam@167
|
1033 T57 = W[25];
|
cannam@167
|
1034 T59 = FMA(T57, T58, T56);
|
cannam@167
|
1035 T9P = FNMS(T57, T55, T9O);
|
cannam@167
|
1036 }
|
cannam@167
|
1037 {
|
cannam@167
|
1038 E T53, T5g, TfB, TfC, TfD, TfE;
|
cannam@167
|
1039 T53 = T4W + T52;
|
cannam@167
|
1040 T5g = T59 + T5f;
|
cannam@167
|
1041 TfB = T53 - T5g;
|
cannam@167
|
1042 TfC = T9V + T9X;
|
cannam@167
|
1043 TfD = T9P + T9R;
|
cannam@167
|
1044 TfE = TfC - TfD;
|
cannam@167
|
1045 T5h = T53 + T5g;
|
cannam@167
|
1046 ThI = TfC + TfD;
|
cannam@167
|
1047 TfF = TfB + TfE;
|
cannam@167
|
1048 TfO = TfE - TfB;
|
cannam@167
|
1049 }
|
cannam@167
|
1050 {
|
cannam@167
|
1051 E T9T, Td9, Ta0, Tda;
|
cannam@167
|
1052 {
|
cannam@167
|
1053 E T9N, T9S, T9Y, T9Z;
|
cannam@167
|
1054 T9N = T4W - T52;
|
cannam@167
|
1055 T9S = T9P - T9R;
|
cannam@167
|
1056 T9T = T9N - T9S;
|
cannam@167
|
1057 Td9 = T9N + T9S;
|
cannam@167
|
1058 T9Y = T9V - T9X;
|
cannam@167
|
1059 T9Z = T59 - T5f;
|
cannam@167
|
1060 Ta0 = T9Y + T9Z;
|
cannam@167
|
1061 Tda = T9Y - T9Z;
|
cannam@167
|
1062 }
|
cannam@167
|
1063 Ta1 = FMA(KP414213562, Ta0, T9T);
|
cannam@167
|
1064 Tdi = FMA(KP414213562, Td9, Tda);
|
cannam@167
|
1065 Tag = FNMS(KP414213562, T9T, Ta0);
|
cannam@167
|
1066 Tdb = FNMS(KP414213562, Tda, Td9);
|
cannam@167
|
1067 }
|
cannam@167
|
1068 }
|
cannam@167
|
1069 {
|
cannam@167
|
1070 E T6i, TaQ, T6B, TaL, T6o, TaS, T6v, TaJ;
|
cannam@167
|
1071 {
|
cannam@167
|
1072 E T6e, T6h, T6f, TaP, T6d, T6g;
|
cannam@167
|
1073 T6e = cr[WS(rs, 3)];
|
cannam@167
|
1074 T6h = ci[WS(rs, 3)];
|
cannam@167
|
1075 T6d = W[4];
|
cannam@167
|
1076 T6f = T6d * T6e;
|
cannam@167
|
1077 TaP = T6d * T6h;
|
cannam@167
|
1078 T6g = W[5];
|
cannam@167
|
1079 T6i = FMA(T6g, T6h, T6f);
|
cannam@167
|
1080 TaQ = FNMS(T6g, T6e, TaP);
|
cannam@167
|
1081 }
|
cannam@167
|
1082 {
|
cannam@167
|
1083 E T6x, T6A, T6y, TaK, T6w, T6z;
|
cannam@167
|
1084 T6x = cr[WS(rs, 51)];
|
cannam@167
|
1085 T6A = ci[WS(rs, 51)];
|
cannam@167
|
1086 T6w = W[100];
|
cannam@167
|
1087 T6y = T6w * T6x;
|
cannam@167
|
1088 TaK = T6w * T6A;
|
cannam@167
|
1089 T6z = W[101];
|
cannam@167
|
1090 T6B = FMA(T6z, T6A, T6y);
|
cannam@167
|
1091 TaL = FNMS(T6z, T6x, TaK);
|
cannam@167
|
1092 }
|
cannam@167
|
1093 {
|
cannam@167
|
1094 E T6k, T6n, T6l, TaR, T6j, T6m;
|
cannam@167
|
1095 T6k = cr[WS(rs, 35)];
|
cannam@167
|
1096 T6n = ci[WS(rs, 35)];
|
cannam@167
|
1097 T6j = W[68];
|
cannam@167
|
1098 T6l = T6j * T6k;
|
cannam@167
|
1099 TaR = T6j * T6n;
|
cannam@167
|
1100 T6m = W[69];
|
cannam@167
|
1101 T6o = FMA(T6m, T6n, T6l);
|
cannam@167
|
1102 TaS = FNMS(T6m, T6k, TaR);
|
cannam@167
|
1103 }
|
cannam@167
|
1104 {
|
cannam@167
|
1105 E T6r, T6u, T6s, TaI, T6q, T6t;
|
cannam@167
|
1106 T6r = cr[WS(rs, 19)];
|
cannam@167
|
1107 T6u = ci[WS(rs, 19)];
|
cannam@167
|
1108 T6q = W[36];
|
cannam@167
|
1109 T6s = T6q * T6r;
|
cannam@167
|
1110 TaI = T6q * T6u;
|
cannam@167
|
1111 T6t = W[37];
|
cannam@167
|
1112 T6v = FMA(T6t, T6u, T6s);
|
cannam@167
|
1113 TaJ = FNMS(T6t, T6r, TaI);
|
cannam@167
|
1114 }
|
cannam@167
|
1115 {
|
cannam@167
|
1116 E T6p, T6C, TfX, TfY, TfZ, Tg0;
|
cannam@167
|
1117 T6p = T6i + T6o;
|
cannam@167
|
1118 T6C = T6v + T6B;
|
cannam@167
|
1119 TfX = T6p - T6C;
|
cannam@167
|
1120 TfY = TaQ + TaS;
|
cannam@167
|
1121 TfZ = TaJ + TaL;
|
cannam@167
|
1122 Tg0 = TfY - TfZ;
|
cannam@167
|
1123 T6D = T6p + T6C;
|
cannam@167
|
1124 ThU = TfY + TfZ;
|
cannam@167
|
1125 Tg1 = TfX - Tg0;
|
cannam@167
|
1126 Tgf = TfX + Tg0;
|
cannam@167
|
1127 }
|
cannam@167
|
1128 {
|
cannam@167
|
1129 E TaN, Tdp, TaU, Tdq;
|
cannam@167
|
1130 {
|
cannam@167
|
1131 E TaH, TaM, TaO, TaT;
|
cannam@167
|
1132 TaH = T6i - T6o;
|
cannam@167
|
1133 TaM = TaJ - TaL;
|
cannam@167
|
1134 TaN = TaH - TaM;
|
cannam@167
|
1135 Tdp = TaH + TaM;
|
cannam@167
|
1136 TaO = T6B - T6v;
|
cannam@167
|
1137 TaT = TaQ - TaS;
|
cannam@167
|
1138 TaU = TaO - TaT;
|
cannam@167
|
1139 Tdq = TaT + TaO;
|
cannam@167
|
1140 }
|
cannam@167
|
1141 TaV = FMA(KP414213562, TaU, TaN);
|
cannam@167
|
1142 TdA = FNMS(KP414213562, Tdp, Tdq);
|
cannam@167
|
1143 Tbo = FNMS(KP414213562, TaN, TaU);
|
cannam@167
|
1144 Tdr = FMA(KP414213562, Tdq, Tdp);
|
cannam@167
|
1145 }
|
cannam@167
|
1146 }
|
cannam@167
|
1147 {
|
cannam@167
|
1148 E T6J, Tb5, T72, Tb0, T6P, Tb7, T6W, TaY;
|
cannam@167
|
1149 {
|
cannam@167
|
1150 E T6F, T6I, T6G, Tb4, T6E, T6H;
|
cannam@167
|
1151 T6F = cr[WS(rs, 59)];
|
cannam@167
|
1152 T6I = ci[WS(rs, 59)];
|
cannam@167
|
1153 T6E = W[116];
|
cannam@167
|
1154 T6G = T6E * T6F;
|
cannam@167
|
1155 Tb4 = T6E * T6I;
|
cannam@167
|
1156 T6H = W[117];
|
cannam@167
|
1157 T6J = FMA(T6H, T6I, T6G);
|
cannam@167
|
1158 Tb5 = FNMS(T6H, T6F, Tb4);
|
cannam@167
|
1159 }
|
cannam@167
|
1160 {
|
cannam@167
|
1161 E T6Y, T71, T6Z, TaZ, T6X, T70;
|
cannam@167
|
1162 T6Y = cr[WS(rs, 43)];
|
cannam@167
|
1163 T71 = ci[WS(rs, 43)];
|
cannam@167
|
1164 T6X = W[84];
|
cannam@167
|
1165 T6Z = T6X * T6Y;
|
cannam@167
|
1166 TaZ = T6X * T71;
|
cannam@167
|
1167 T70 = W[85];
|
cannam@167
|
1168 T72 = FMA(T70, T71, T6Z);
|
cannam@167
|
1169 Tb0 = FNMS(T70, T6Y, TaZ);
|
cannam@167
|
1170 }
|
cannam@167
|
1171 {
|
cannam@167
|
1172 E T6L, T6O, T6M, Tb6, T6K, T6N;
|
cannam@167
|
1173 T6L = cr[WS(rs, 27)];
|
cannam@167
|
1174 T6O = ci[WS(rs, 27)];
|
cannam@167
|
1175 T6K = W[52];
|
cannam@167
|
1176 T6M = T6K * T6L;
|
cannam@167
|
1177 Tb6 = T6K * T6O;
|
cannam@167
|
1178 T6N = W[53];
|
cannam@167
|
1179 T6P = FMA(T6N, T6O, T6M);
|
cannam@167
|
1180 Tb7 = FNMS(T6N, T6L, Tb6);
|
cannam@167
|
1181 }
|
cannam@167
|
1182 {
|
cannam@167
|
1183 E T6S, T6V, T6T, TaX, T6R, T6U;
|
cannam@167
|
1184 T6S = cr[WS(rs, 11)];
|
cannam@167
|
1185 T6V = ci[WS(rs, 11)];
|
cannam@167
|
1186 T6R = W[20];
|
cannam@167
|
1187 T6T = T6R * T6S;
|
cannam@167
|
1188 TaX = T6R * T6V;
|
cannam@167
|
1189 T6U = W[21];
|
cannam@167
|
1190 T6W = FMA(T6U, T6V, T6T);
|
cannam@167
|
1191 TaY = FNMS(T6U, T6S, TaX);
|
cannam@167
|
1192 }
|
cannam@167
|
1193 {
|
cannam@167
|
1194 E T6Q, T73, Tg2, Tg3, Tg4, Tg5;
|
cannam@167
|
1195 T6Q = T6J + T6P;
|
cannam@167
|
1196 T73 = T6W + T72;
|
cannam@167
|
1197 Tg2 = T6Q - T73;
|
cannam@167
|
1198 Tg3 = Tb5 + Tb7;
|
cannam@167
|
1199 Tg4 = TaY + Tb0;
|
cannam@167
|
1200 Tg5 = Tg3 - Tg4;
|
cannam@167
|
1201 T74 = T6Q + T73;
|
cannam@167
|
1202 ThT = Tg3 + Tg4;
|
cannam@167
|
1203 Tg6 = Tg2 + Tg5;
|
cannam@167
|
1204 Tge = Tg2 - Tg5;
|
cannam@167
|
1205 }
|
cannam@167
|
1206 {
|
cannam@167
|
1207 E Tb2, Tds, Tb9, Tdt;
|
cannam@167
|
1208 {
|
cannam@167
|
1209 E TaW, Tb1, Tb3, Tb8;
|
cannam@167
|
1210 TaW = T6J - T6P;
|
cannam@167
|
1211 Tb1 = TaY - Tb0;
|
cannam@167
|
1212 Tb2 = TaW - Tb1;
|
cannam@167
|
1213 Tds = TaW + Tb1;
|
cannam@167
|
1214 Tb3 = T72 - T6W;
|
cannam@167
|
1215 Tb8 = Tb5 - Tb7;
|
cannam@167
|
1216 Tb9 = Tb3 - Tb8;
|
cannam@167
|
1217 Tdt = Tb8 + Tb3;
|
cannam@167
|
1218 }
|
cannam@167
|
1219 Tba = FNMS(KP414213562, Tb9, Tb2);
|
cannam@167
|
1220 TdB = FMA(KP414213562, Tds, Tdt);
|
cannam@167
|
1221 Tbp = FMA(KP414213562, Tb2, Tb9);
|
cannam@167
|
1222 Tdu = FNMS(KP414213562, Tdt, Tds);
|
cannam@167
|
1223 }
|
cannam@167
|
1224 }
|
cannam@167
|
1225 {
|
cannam@167
|
1226 E T1I, Tio, T3v, Tj1, TiX, Tj2, Tir, TiN, T76, TiJ, TiC, TiG, T5j, TiK, Tix;
|
cannam@167
|
1227 E TiF;
|
cannam@167
|
1228 {
|
cannam@167
|
1229 E TO, T1H, Tip, Tiq;
|
cannam@167
|
1230 TO = Tm + TN;
|
cannam@167
|
1231 T1H = T1f + T1G;
|
cannam@167
|
1232 T1I = TO + T1H;
|
cannam@167
|
1233 Tio = TO - T1H;
|
cannam@167
|
1234 {
|
cannam@167
|
1235 E T2B, T3u, TiO, TiW;
|
cannam@167
|
1236 T2B = T29 + T2A;
|
cannam@167
|
1237 T3u = T32 + T3t;
|
cannam@167
|
1238 T3v = T2B + T3u;
|
cannam@167
|
1239 Tj1 = T2B - T3u;
|
cannam@167
|
1240 TiO = Ths + Thr;
|
cannam@167
|
1241 TiW = TiP + TiV;
|
cannam@167
|
1242 TiX = TiO + TiW;
|
cannam@167
|
1243 Tj2 = TiW - TiO;
|
cannam@167
|
1244 }
|
cannam@167
|
1245 Tip = ThB + ThC;
|
cannam@167
|
1246 Tiq = Thw + Thx;
|
cannam@167
|
1247 Tir = Tip - Tiq;
|
cannam@167
|
1248 TiN = Tiq + Tip;
|
cannam@167
|
1249 {
|
cannam@167
|
1250 E T6c, T75, Tiy, Tiz, TiA, TiB;
|
cannam@167
|
1251 T6c = T5K + T6b;
|
cannam@167
|
1252 T75 = T6D + T74;
|
cannam@167
|
1253 Tiy = T6c - T75;
|
cannam@167
|
1254 Tiz = ThX + ThY;
|
cannam@167
|
1255 TiA = ThU + ThT;
|
cannam@167
|
1256 TiB = Tiz - TiA;
|
cannam@167
|
1257 T76 = T6c + T75;
|
cannam@167
|
1258 TiJ = Tiz + TiA;
|
cannam@167
|
1259 TiC = Tiy - TiB;
|
cannam@167
|
1260 TiG = Tiy + TiB;
|
cannam@167
|
1261 }
|
cannam@167
|
1262 {
|
cannam@167
|
1263 E T4p, T5i, Tit, Tiu, Tiv, Tiw;
|
cannam@167
|
1264 T4p = T3X + T4o;
|
cannam@167
|
1265 T5i = T4Q + T5h;
|
cannam@167
|
1266 Tit = T4p - T5i;
|
cannam@167
|
1267 Tiu = ThM + ThN;
|
cannam@167
|
1268 Tiv = ThJ + ThI;
|
cannam@167
|
1269 Tiw = Tiu - Tiv;
|
cannam@167
|
1270 T5j = T4p + T5i;
|
cannam@167
|
1271 TiK = Tiu + Tiv;
|
cannam@167
|
1272 Tix = Tit + Tiw;
|
cannam@167
|
1273 TiF = Tit - Tiw;
|
cannam@167
|
1274 }
|
cannam@167
|
1275 }
|
cannam@167
|
1276 {
|
cannam@167
|
1277 E T3w, T77, Tj3, Tj4;
|
cannam@167
|
1278 T3w = T1I + T3v;
|
cannam@167
|
1279 T77 = T5j + T76;
|
cannam@167
|
1280 ci[WS(rs, 31)] = T3w - T77;
|
cannam@167
|
1281 cr[0] = T3w + T77;
|
cannam@167
|
1282 Tj3 = Tj1 + Tj2;
|
cannam@167
|
1283 Tj4 = TiC - Tix;
|
cannam@167
|
1284 cr[WS(rs, 56)] = FMS(KP707106781, Tj4, Tj3);
|
cannam@167
|
1285 ci[WS(rs, 39)] = FMA(KP707106781, Tj4, Tj3);
|
cannam@167
|
1286 }
|
cannam@167
|
1287 {
|
cannam@167
|
1288 E Tj5, Tj6, Tis, TiD;
|
cannam@167
|
1289 Tj5 = Tj2 - Tj1;
|
cannam@167
|
1290 Tj6 = TiG - TiF;
|
cannam@167
|
1291 cr[WS(rs, 40)] = FMS(KP707106781, Tj6, Tj5);
|
cannam@167
|
1292 ci[WS(rs, 55)] = FMA(KP707106781, Tj6, Tj5);
|
cannam@167
|
1293 Tis = Tio - Tir;
|
cannam@167
|
1294 TiD = Tix + TiC;
|
cannam@167
|
1295 ci[WS(rs, 23)] = FNMS(KP707106781, TiD, Tis);
|
cannam@167
|
1296 cr[WS(rs, 8)] = FMA(KP707106781, TiD, Tis);
|
cannam@167
|
1297 }
|
cannam@167
|
1298 {
|
cannam@167
|
1299 E TiE, TiH, TiM, TiY;
|
cannam@167
|
1300 TiE = Tio + Tir;
|
cannam@167
|
1301 TiH = TiF + TiG;
|
cannam@167
|
1302 cr[WS(rs, 24)] = FNMS(KP707106781, TiH, TiE);
|
cannam@167
|
1303 ci[WS(rs, 7)] = FMA(KP707106781, TiH, TiE);
|
cannam@167
|
1304 TiM = TiK + TiJ;
|
cannam@167
|
1305 TiY = TiN + TiX;
|
cannam@167
|
1306 cr[WS(rs, 32)] = TiM - TiY;
|
cannam@167
|
1307 ci[WS(rs, 63)] = TiM + TiY;
|
cannam@167
|
1308 }
|
cannam@167
|
1309 {
|
cannam@167
|
1310 E TiZ, Tj0, TiI, TiL;
|
cannam@167
|
1311 TiZ = T76 - T5j;
|
cannam@167
|
1312 Tj0 = TiX - TiN;
|
cannam@167
|
1313 cr[WS(rs, 48)] = TiZ - Tj0;
|
cannam@167
|
1314 ci[WS(rs, 47)] = TiZ + Tj0;
|
cannam@167
|
1315 TiI = T1I - T3v;
|
cannam@167
|
1316 TiL = TiJ - TiK;
|
cannam@167
|
1317 cr[WS(rs, 16)] = TiI - TiL;
|
cannam@167
|
1318 ci[WS(rs, 15)] = TiI + TiL;
|
cannam@167
|
1319 }
|
cannam@167
|
1320 }
|
cannam@167
|
1321 {
|
cannam@167
|
1322 E T99, Tk2, TbB, TjW, Taj, TbL, Tbv, TbF, Tce, Tcy, Tci, Tcu, Tc7, Tcx, Tch;
|
cannam@167
|
1323 E Tcr, TbZ, Tkg, Tcn, Tka, Tbs, TbM, Tbw, TbI, T80, Tk9, Tkf, Tby, TbS, TjV;
|
cannam@167
|
1324 E Tk1, Tck;
|
cannam@167
|
1325 {
|
cannam@167
|
1326 E T8z, Tbz, T98, TbA;
|
cannam@167
|
1327 {
|
cannam@167
|
1328 E T8n, T8y, T8W, T97;
|
cannam@167
|
1329 T8n = FMA(KP707106781, T8m, T87);
|
cannam@167
|
1330 T8y = FMA(KP707106781, T8x, T8u);
|
cannam@167
|
1331 T8z = FNMS(KP198912367, T8y, T8n);
|
cannam@167
|
1332 Tbz = FMA(KP198912367, T8n, T8y);
|
cannam@167
|
1333 T8W = FMA(KP707106781, T8V, T8G);
|
cannam@167
|
1334 T97 = FMA(KP707106781, T96, T93);
|
cannam@167
|
1335 T98 = FMA(KP198912367, T97, T8W);
|
cannam@167
|
1336 TbA = FNMS(KP198912367, T8W, T97);
|
cannam@167
|
1337 }
|
cannam@167
|
1338 T99 = T8z + T98;
|
cannam@167
|
1339 Tk2 = T98 - T8z;
|
cannam@167
|
1340 TbB = Tbz - TbA;
|
cannam@167
|
1341 TjW = Tbz + TbA;
|
cannam@167
|
1342 }
|
cannam@167
|
1343 {
|
cannam@167
|
1344 E Ta3, TbD, Tai, TbE;
|
cannam@167
|
1345 {
|
cannam@167
|
1346 E T9x, Ta2, Tae, Tah;
|
cannam@167
|
1347 T9x = FMA(KP707106781, T9w, T9h);
|
cannam@167
|
1348 Ta2 = T9M + Ta1;
|
cannam@167
|
1349 Ta3 = FMA(KP923879532, Ta2, T9x);
|
cannam@167
|
1350 TbD = FNMS(KP923879532, Ta2, T9x);
|
cannam@167
|
1351 Tae = FMA(KP707106781, Tad, Taa);
|
cannam@167
|
1352 Tah = Taf + Tag;
|
cannam@167
|
1353 Tai = FMA(KP923879532, Tah, Tae);
|
cannam@167
|
1354 TbE = FNMS(KP923879532, Tah, Tae);
|
cannam@167
|
1355 }
|
cannam@167
|
1356 Taj = FNMS(KP098491403, Tai, Ta3);
|
cannam@167
|
1357 TbL = FNMS(KP820678790, TbD, TbE);
|
cannam@167
|
1358 Tbv = FMA(KP098491403, Ta3, Tai);
|
cannam@167
|
1359 TbF = FMA(KP820678790, TbE, TbD);
|
cannam@167
|
1360 }
|
cannam@167
|
1361 {
|
cannam@167
|
1362 E Tca, Tcs, Tcd, Tct;
|
cannam@167
|
1363 {
|
cannam@167
|
1364 E Tc8, Tc9, Tcb, Tcc;
|
cannam@167
|
1365 Tc8 = FNMS(KP707106781, TaF, Taq);
|
cannam@167
|
1366 Tc9 = Tbp - Tbo;
|
cannam@167
|
1367 Tca = FNMS(KP923879532, Tc9, Tc8);
|
cannam@167
|
1368 Tcs = FMA(KP923879532, Tc9, Tc8);
|
cannam@167
|
1369 Tcb = FNMS(KP707106781, Tbm, Tbj);
|
cannam@167
|
1370 Tcc = Tba - TaV;
|
cannam@167
|
1371 Tcd = FMA(KP923879532, Tcc, Tcb);
|
cannam@167
|
1372 Tct = FNMS(KP923879532, Tcc, Tcb);
|
cannam@167
|
1373 }
|
cannam@167
|
1374 Tce = FNMS(KP534511135, Tcd, Tca);
|
cannam@167
|
1375 Tcy = FNMS(KP303346683, Tcs, Tct);
|
cannam@167
|
1376 Tci = FMA(KP534511135, Tca, Tcd);
|
cannam@167
|
1377 Tcu = FMA(KP303346683, Tct, Tcs);
|
cannam@167
|
1378 }
|
cannam@167
|
1379 {
|
cannam@167
|
1380 E Tc3, Tcp, Tc6, Tcq;
|
cannam@167
|
1381 {
|
cannam@167
|
1382 E Tc1, Tc2, Tc4, Tc5;
|
cannam@167
|
1383 Tc1 = FNMS(KP707106781, T9w, T9h);
|
cannam@167
|
1384 Tc2 = Taf - Tag;
|
cannam@167
|
1385 Tc3 = FNMS(KP923879532, Tc2, Tc1);
|
cannam@167
|
1386 Tcp = FMA(KP923879532, Tc2, Tc1);
|
cannam@167
|
1387 Tc4 = FNMS(KP707106781, Tad, Taa);
|
cannam@167
|
1388 Tc5 = Ta1 - T9M;
|
cannam@167
|
1389 Tc6 = FNMS(KP923879532, Tc5, Tc4);
|
cannam@167
|
1390 Tcq = FMA(KP923879532, Tc5, Tc4);
|
cannam@167
|
1391 }
|
cannam@167
|
1392 Tc7 = FNMS(KP534511135, Tc6, Tc3);
|
cannam@167
|
1393 Tcx = FNMS(KP303346683, Tcp, Tcq);
|
cannam@167
|
1394 Tch = FMA(KP534511135, Tc3, Tc6);
|
cannam@167
|
1395 Tcr = FMA(KP303346683, Tcq, Tcp);
|
cannam@167
|
1396 }
|
cannam@167
|
1397 {
|
cannam@167
|
1398 E TbV, Tcm, TbY, Tcl;
|
cannam@167
|
1399 {
|
cannam@167
|
1400 E TbT, TbU, TbW, TbX;
|
cannam@167
|
1401 TbT = FNMS(KP707106781, T96, T93);
|
cannam@167
|
1402 TbU = FNMS(KP707106781, T8V, T8G);
|
cannam@167
|
1403 TbV = FMA(KP668178637, TbU, TbT);
|
cannam@167
|
1404 Tcm = FNMS(KP668178637, TbT, TbU);
|
cannam@167
|
1405 TbW = FNMS(KP707106781, T8x, T8u);
|
cannam@167
|
1406 TbX = FNMS(KP707106781, T8m, T87);
|
cannam@167
|
1407 TbY = FNMS(KP668178637, TbX, TbW);
|
cannam@167
|
1408 Tcl = FMA(KP668178637, TbW, TbX);
|
cannam@167
|
1409 }
|
cannam@167
|
1410 TbZ = TbV - TbY;
|
cannam@167
|
1411 Tkg = Tcl - Tcm;
|
cannam@167
|
1412 Tcn = Tcl + Tcm;
|
cannam@167
|
1413 Tka = TbY + TbV;
|
cannam@167
|
1414 }
|
cannam@167
|
1415 {
|
cannam@167
|
1416 E Tbc, TbG, Tbr, TbH;
|
cannam@167
|
1417 {
|
cannam@167
|
1418 E TaG, Tbb, Tbn, Tbq;
|
cannam@167
|
1419 TaG = FMA(KP707106781, TaF, Taq);
|
cannam@167
|
1420 Tbb = TaV + Tba;
|
cannam@167
|
1421 Tbc = FMA(KP923879532, Tbb, TaG);
|
cannam@167
|
1422 TbG = FNMS(KP923879532, Tbb, TaG);
|
cannam@167
|
1423 Tbn = FMA(KP707106781, Tbm, Tbj);
|
cannam@167
|
1424 Tbq = Tbo + Tbp;
|
cannam@167
|
1425 Tbr = FMA(KP923879532, Tbq, Tbn);
|
cannam@167
|
1426 TbH = FNMS(KP923879532, Tbq, Tbn);
|
cannam@167
|
1427 }
|
cannam@167
|
1428 Tbs = FNMS(KP098491403, Tbr, Tbc);
|
cannam@167
|
1429 TbM = FNMS(KP820678790, TbG, TbH);
|
cannam@167
|
1430 Tbw = FMA(KP098491403, Tbc, Tbr);
|
cannam@167
|
1431 TbI = FMA(KP820678790, TbH, TbG);
|
cannam@167
|
1432 }
|
cannam@167
|
1433 {
|
cannam@167
|
1434 E T7u, TbO, TjT, Tk7, T7Z, Tk8, TbR, TjU, T7t, TjS;
|
cannam@167
|
1435 T7t = T7l + T7s;
|
cannam@167
|
1436 T7u = FMA(KP707106781, T7t, T7e);
|
cannam@167
|
1437 TbO = FNMS(KP707106781, T7t, T7e);
|
cannam@167
|
1438 TjS = TcB - TcC;
|
cannam@167
|
1439 TjT = FMA(KP707106781, TjS, TjR);
|
cannam@167
|
1440 Tk7 = FNMS(KP707106781, TjS, TjR);
|
cannam@167
|
1441 {
|
cannam@167
|
1442 E T7J, T7Y, TbP, TbQ;
|
cannam@167
|
1443 T7J = FNMS(KP414213562, T7I, T7B);
|
cannam@167
|
1444 T7Y = FMA(KP414213562, T7X, T7Q);
|
cannam@167
|
1445 T7Z = T7J + T7Y;
|
cannam@167
|
1446 Tk8 = T7Y - T7J;
|
cannam@167
|
1447 TbP = FMA(KP414213562, T7B, T7I);
|
cannam@167
|
1448 TbQ = FNMS(KP414213562, T7Q, T7X);
|
cannam@167
|
1449 TbR = TbP - TbQ;
|
cannam@167
|
1450 TjU = TbP + TbQ;
|
cannam@167
|
1451 }
|
cannam@167
|
1452 T80 = FMA(KP923879532, T7Z, T7u);
|
cannam@167
|
1453 Tk9 = FMA(KP923879532, Tk8, Tk7);
|
cannam@167
|
1454 Tkf = FNMS(KP923879532, Tk8, Tk7);
|
cannam@167
|
1455 Tby = FNMS(KP923879532, T7Z, T7u);
|
cannam@167
|
1456 TbS = FNMS(KP923879532, TbR, TbO);
|
cannam@167
|
1457 TjV = FMA(KP923879532, TjU, TjT);
|
cannam@167
|
1458 Tk1 = FNMS(KP923879532, TjU, TjT);
|
cannam@167
|
1459 Tck = FMA(KP923879532, TbR, TbO);
|
cannam@167
|
1460 }
|
cannam@167
|
1461 {
|
cannam@167
|
1462 E T9a, Tbt, TbK, TbN;
|
cannam@167
|
1463 T9a = FMA(KP980785280, T99, T80);
|
cannam@167
|
1464 Tbt = Taj + Tbs;
|
cannam@167
|
1465 cr[WS(rs, 31)] = FNMS(KP995184726, Tbt, T9a);
|
cannam@167
|
1466 ci[0] = FMA(KP995184726, Tbt, T9a);
|
cannam@167
|
1467 TbK = FNMS(KP980785280, TbB, Tby);
|
cannam@167
|
1468 TbN = TbL + TbM;
|
cannam@167
|
1469 cr[WS(rs, 23)] = FMA(KP773010453, TbN, TbK);
|
cannam@167
|
1470 ci[WS(rs, 8)] = FNMS(KP773010453, TbN, TbK);
|
cannam@167
|
1471 }
|
cannam@167
|
1472 {
|
cannam@167
|
1473 E Tkb, Tkc, Tkj, Tkk;
|
cannam@167
|
1474 Tkb = FMA(KP831469612, Tka, Tk9);
|
cannam@167
|
1475 Tkc = Tcx - Tcy;
|
cannam@167
|
1476 cr[WS(rs, 35)] = FMS(KP956940335, Tkc, Tkb);
|
cannam@167
|
1477 ci[WS(rs, 60)] = FMA(KP956940335, Tkc, Tkb);
|
cannam@167
|
1478 Tkj = FNMS(KP831469612, Tkg, Tkf);
|
cannam@167
|
1479 Tkk = Tce - Tc7;
|
cannam@167
|
1480 cr[WS(rs, 43)] = FMS(KP881921264, Tkk, Tkj);
|
cannam@167
|
1481 ci[WS(rs, 52)] = FMA(KP881921264, Tkk, Tkj);
|
cannam@167
|
1482 }
|
cannam@167
|
1483 {
|
cannam@167
|
1484 E Tbu, Tbx, TbC, TbJ;
|
cannam@167
|
1485 Tbu = FNMS(KP980785280, T99, T80);
|
cannam@167
|
1486 Tbx = Tbv + Tbw;
|
cannam@167
|
1487 ci[WS(rs, 16)] = FNMS(KP995184726, Tbx, Tbu);
|
cannam@167
|
1488 cr[WS(rs, 15)] = FMA(KP995184726, Tbx, Tbu);
|
cannam@167
|
1489 TbC = FMA(KP980785280, TbB, Tby);
|
cannam@167
|
1490 TbJ = TbF + TbI;
|
cannam@167
|
1491 ci[WS(rs, 24)] = FNMS(KP773010453, TbJ, TbC);
|
cannam@167
|
1492 cr[WS(rs, 7)] = FMA(KP773010453, TbJ, TbC);
|
cannam@167
|
1493 }
|
cannam@167
|
1494 {
|
cannam@167
|
1495 E Tkd, Tke, Tkh, Tki;
|
cannam@167
|
1496 Tkd = FNMS(KP831469612, Tka, Tk9);
|
cannam@167
|
1497 Tke = Tcu - Tcr;
|
cannam@167
|
1498 cr[WS(rs, 51)] = FMS(KP956940335, Tke, Tkd);
|
cannam@167
|
1499 ci[WS(rs, 44)] = FMA(KP956940335, Tke, Tkd);
|
cannam@167
|
1500 Tkh = FMA(KP831469612, Tkg, Tkf);
|
cannam@167
|
1501 Tki = Tci - Tch;
|
cannam@167
|
1502 cr[WS(rs, 59)] = FMS(KP881921264, Tki, Tkh);
|
cannam@167
|
1503 ci[WS(rs, 36)] = FMA(KP881921264, Tki, Tkh);
|
cannam@167
|
1504 }
|
cannam@167
|
1505 {
|
cannam@167
|
1506 E Tc0, Tcf, Tcw, Tcz;
|
cannam@167
|
1507 Tc0 = FMA(KP831469612, TbZ, TbS);
|
cannam@167
|
1508 Tcf = Tc7 + Tce;
|
cannam@167
|
1509 cr[WS(rs, 27)] = FNMS(KP881921264, Tcf, Tc0);
|
cannam@167
|
1510 ci[WS(rs, 4)] = FMA(KP881921264, Tcf, Tc0);
|
cannam@167
|
1511 Tcw = FNMS(KP831469612, Tcn, Tck);
|
cannam@167
|
1512 Tcz = Tcx + Tcy;
|
cannam@167
|
1513 cr[WS(rs, 19)] = FMA(KP956940335, Tcz, Tcw);
|
cannam@167
|
1514 ci[WS(rs, 12)] = FNMS(KP956940335, Tcz, Tcw);
|
cannam@167
|
1515 }
|
cannam@167
|
1516 {
|
cannam@167
|
1517 E TjX, TjY, Tk5, Tk6;
|
cannam@167
|
1518 TjX = FMA(KP980785280, TjW, TjV);
|
cannam@167
|
1519 TjY = Tbw - Tbv;
|
cannam@167
|
1520 cr[WS(rs, 63)] = FMS(KP995184726, TjY, TjX);
|
cannam@167
|
1521 ci[WS(rs, 32)] = FMA(KP995184726, TjY, TjX);
|
cannam@167
|
1522 Tk5 = FNMS(KP980785280, Tk2, Tk1);
|
cannam@167
|
1523 Tk6 = TbI - TbF;
|
cannam@167
|
1524 cr[WS(rs, 55)] = FMS(KP773010453, Tk6, Tk5);
|
cannam@167
|
1525 ci[WS(rs, 40)] = FMA(KP773010453, Tk6, Tk5);
|
cannam@167
|
1526 }
|
cannam@167
|
1527 {
|
cannam@167
|
1528 E Tcg, Tcj, Tco, Tcv;
|
cannam@167
|
1529 Tcg = FNMS(KP831469612, TbZ, TbS);
|
cannam@167
|
1530 Tcj = Tch + Tci;
|
cannam@167
|
1531 ci[WS(rs, 20)] = FNMS(KP881921264, Tcj, Tcg);
|
cannam@167
|
1532 cr[WS(rs, 11)] = FMA(KP881921264, Tcj, Tcg);
|
cannam@167
|
1533 Tco = FMA(KP831469612, Tcn, Tck);
|
cannam@167
|
1534 Tcv = Tcr + Tcu;
|
cannam@167
|
1535 ci[WS(rs, 28)] = FNMS(KP956940335, Tcv, Tco);
|
cannam@167
|
1536 cr[WS(rs, 3)] = FMA(KP956940335, Tcv, Tco);
|
cannam@167
|
1537 }
|
cannam@167
|
1538 {
|
cannam@167
|
1539 E TjZ, Tk0, Tk3, Tk4;
|
cannam@167
|
1540 TjZ = FNMS(KP980785280, TjW, TjV);
|
cannam@167
|
1541 Tk0 = Tbs - Taj;
|
cannam@167
|
1542 cr[WS(rs, 47)] = FMS(KP995184726, Tk0, TjZ);
|
cannam@167
|
1543 ci[WS(rs, 48)] = FMA(KP995184726, Tk0, TjZ);
|
cannam@167
|
1544 Tk3 = FMA(KP980785280, Tk2, Tk1);
|
cannam@167
|
1545 Tk4 = TbL - TbM;
|
cannam@167
|
1546 cr[WS(rs, 39)] = FMS(KP773010453, Tk4, Tk3);
|
cannam@167
|
1547 ci[WS(rs, 56)] = FMA(KP773010453, Tk4, Tk3);
|
cannam@167
|
1548 }
|
cannam@167
|
1549 }
|
cannam@167
|
1550 {
|
cannam@167
|
1551 E Thu, Ti8, Tj9, Tjf, ThF, Tjg, Tib, Tja, ThR, Til, Ti6, Tif, Ti2, Tim, Ti5;
|
cannam@167
|
1552 E Tii;
|
cannam@167
|
1553 {
|
cannam@167
|
1554 E Thq, Tht, Tj7, Tj8;
|
cannam@167
|
1555 Thq = Tm - TN;
|
cannam@167
|
1556 Tht = Thr - Ths;
|
cannam@167
|
1557 Thu = Thq - Tht;
|
cannam@167
|
1558 Ti8 = Thq + Tht;
|
cannam@167
|
1559 Tj7 = T1f - T1G;
|
cannam@167
|
1560 Tj8 = TiV - TiP;
|
cannam@167
|
1561 Tj9 = Tj7 + Tj8;
|
cannam@167
|
1562 Tjf = Tj8 - Tj7;
|
cannam@167
|
1563 }
|
cannam@167
|
1564 {
|
cannam@167
|
1565 E Thz, Ti9, ThE, Tia;
|
cannam@167
|
1566 {
|
cannam@167
|
1567 E Thv, Thy, ThA, ThD;
|
cannam@167
|
1568 Thv = T29 - T2A;
|
cannam@167
|
1569 Thy = Thw - Thx;
|
cannam@167
|
1570 Thz = Thv + Thy;
|
cannam@167
|
1571 Ti9 = Thv - Thy;
|
cannam@167
|
1572 ThA = T32 - T3t;
|
cannam@167
|
1573 ThD = ThB - ThC;
|
cannam@167
|
1574 ThE = ThA - ThD;
|
cannam@167
|
1575 Tia = ThA + ThD;
|
cannam@167
|
1576 }
|
cannam@167
|
1577 ThF = Thz + ThE;
|
cannam@167
|
1578 Tjg = Tia - Ti9;
|
cannam@167
|
1579 Tib = Ti9 + Tia;
|
cannam@167
|
1580 Tja = Thz - ThE;
|
cannam@167
|
1581 }
|
cannam@167
|
1582 {
|
cannam@167
|
1583 E ThL, Tid, ThQ, Tie;
|
cannam@167
|
1584 {
|
cannam@167
|
1585 E ThH, ThK, ThO, ThP;
|
cannam@167
|
1586 ThH = T3X - T4o;
|
cannam@167
|
1587 ThK = ThI - ThJ;
|
cannam@167
|
1588 ThL = ThH - ThK;
|
cannam@167
|
1589 Tid = ThH + ThK;
|
cannam@167
|
1590 ThO = ThM - ThN;
|
cannam@167
|
1591 ThP = T4Q - T5h;
|
cannam@167
|
1592 ThQ = ThO - ThP;
|
cannam@167
|
1593 Tie = ThO + ThP;
|
cannam@167
|
1594 }
|
cannam@167
|
1595 ThR = FMA(KP414213562, ThQ, ThL);
|
cannam@167
|
1596 Til = FMA(KP414213562, Tid, Tie);
|
cannam@167
|
1597 Ti6 = FNMS(KP414213562, ThL, ThQ);
|
cannam@167
|
1598 Tif = FNMS(KP414213562, Tie, Tid);
|
cannam@167
|
1599 }
|
cannam@167
|
1600 {
|
cannam@167
|
1601 E ThW, Tig, Ti1, Tih;
|
cannam@167
|
1602 {
|
cannam@167
|
1603 E ThS, ThV, ThZ, Ti0;
|
cannam@167
|
1604 ThS = T5K - T6b;
|
cannam@167
|
1605 ThV = ThT - ThU;
|
cannam@167
|
1606 ThW = ThS - ThV;
|
cannam@167
|
1607 Tig = ThS + ThV;
|
cannam@167
|
1608 ThZ = ThX - ThY;
|
cannam@167
|
1609 Ti0 = T74 - T6D;
|
cannam@167
|
1610 Ti1 = ThZ + Ti0;
|
cannam@167
|
1611 Tih = Ti0 - ThZ;
|
cannam@167
|
1612 }
|
cannam@167
|
1613 Ti2 = FNMS(KP414213562, Ti1, ThW);
|
cannam@167
|
1614 Tim = FMA(KP414213562, Tig, Tih);
|
cannam@167
|
1615 Ti5 = FMA(KP414213562, ThW, Ti1);
|
cannam@167
|
1616 Tii = FNMS(KP414213562, Tih, Tig);
|
cannam@167
|
1617 }
|
cannam@167
|
1618 {
|
cannam@167
|
1619 E ThG, Ti3, Tjh, Tji;
|
cannam@167
|
1620 ThG = FMA(KP707106781, ThF, Thu);
|
cannam@167
|
1621 Ti3 = ThR + Ti2;
|
cannam@167
|
1622 ci[WS(rs, 27)] = FNMS(KP923879532, Ti3, ThG);
|
cannam@167
|
1623 cr[WS(rs, 4)] = FMA(KP923879532, Ti3, ThG);
|
cannam@167
|
1624 Tjh = FMA(KP707106781, Tjg, Tjf);
|
cannam@167
|
1625 Tji = Ti6 + Ti5;
|
cannam@167
|
1626 cr[WS(rs, 36)] = FMS(KP923879532, Tji, Tjh);
|
cannam@167
|
1627 ci[WS(rs, 59)] = FMA(KP923879532, Tji, Tjh);
|
cannam@167
|
1628 }
|
cannam@167
|
1629 {
|
cannam@167
|
1630 E Tjj, Tjk, Ti4, Ti7;
|
cannam@167
|
1631 Tjj = FNMS(KP707106781, Tjg, Tjf);
|
cannam@167
|
1632 Tjk = Ti2 - ThR;
|
cannam@167
|
1633 cr[WS(rs, 52)] = FMS(KP923879532, Tjk, Tjj);
|
cannam@167
|
1634 ci[WS(rs, 43)] = FMA(KP923879532, Tjk, Tjj);
|
cannam@167
|
1635 Ti4 = FNMS(KP707106781, ThF, Thu);
|
cannam@167
|
1636 Ti7 = Ti5 - Ti6;
|
cannam@167
|
1637 cr[WS(rs, 20)] = FNMS(KP923879532, Ti7, Ti4);
|
cannam@167
|
1638 ci[WS(rs, 11)] = FMA(KP923879532, Ti7, Ti4);
|
cannam@167
|
1639 }
|
cannam@167
|
1640 {
|
cannam@167
|
1641 E Tic, Tij, Tjb, Tjc;
|
cannam@167
|
1642 Tic = FMA(KP707106781, Tib, Ti8);
|
cannam@167
|
1643 Tij = Tif + Tii;
|
cannam@167
|
1644 cr[WS(rs, 28)] = FNMS(KP923879532, Tij, Tic);
|
cannam@167
|
1645 ci[WS(rs, 3)] = FMA(KP923879532, Tij, Tic);
|
cannam@167
|
1646 Tjb = FMA(KP707106781, Tja, Tj9);
|
cannam@167
|
1647 Tjc = Tim - Til;
|
cannam@167
|
1648 cr[WS(rs, 60)] = FMS(KP923879532, Tjc, Tjb);
|
cannam@167
|
1649 ci[WS(rs, 35)] = FMA(KP923879532, Tjc, Tjb);
|
cannam@167
|
1650 }
|
cannam@167
|
1651 {
|
cannam@167
|
1652 E Tjd, Tje, Tik, Tin;
|
cannam@167
|
1653 Tjd = FNMS(KP707106781, Tja, Tj9);
|
cannam@167
|
1654 Tje = Tii - Tif;
|
cannam@167
|
1655 cr[WS(rs, 44)] = FMS(KP923879532, Tje, Tjd);
|
cannam@167
|
1656 ci[WS(rs, 51)] = FMA(KP923879532, Tje, Tjd);
|
cannam@167
|
1657 Tik = FNMS(KP707106781, Tib, Ti8);
|
cannam@167
|
1658 Tin = Til + Tim;
|
cannam@167
|
1659 ci[WS(rs, 19)] = FNMS(KP923879532, Tin, Tik);
|
cannam@167
|
1660 cr[WS(rs, 12)] = FMA(KP923879532, Tin, Tik);
|
cannam@167
|
1661 }
|
cannam@167
|
1662 }
|
cannam@167
|
1663 {
|
cannam@167
|
1664 E Tf2, TjJ, Tgo, TjD, TgI, Tjv, Tha, Tjp, Tfp, Tjw, Tgr, Tjq, Th4, Tho, Th7;
|
cannam@167
|
1665 E Thk, TfR, TgB, Tgl, Tgv, TgP, TjK, Thd, TjE, TgX, Thn, Th8, Thh, Tgi, TgC;
|
cannam@167
|
1666 E Tgm, Tgy;
|
cannam@167
|
1667 {
|
cannam@167
|
1668 E TeQ, TjB, Tf1, TjC, TeV, Tf0;
|
cannam@167
|
1669 TeQ = TeM + TeP;
|
cannam@167
|
1670 TjB = Tjm - Tjl;
|
cannam@167
|
1671 TeV = TeR - TeU;
|
cannam@167
|
1672 Tf0 = TeW + TeZ;
|
cannam@167
|
1673 Tf1 = TeV + Tf0;
|
cannam@167
|
1674 TjC = Tf0 - TeV;
|
cannam@167
|
1675 Tf2 = FNMS(KP707106781, Tf1, TeQ);
|
cannam@167
|
1676 TjJ = FNMS(KP707106781, TjC, TjB);
|
cannam@167
|
1677 Tgo = FMA(KP707106781, Tf1, TeQ);
|
cannam@167
|
1678 TjD = FMA(KP707106781, TjC, TjB);
|
cannam@167
|
1679 }
|
cannam@167
|
1680 {
|
cannam@167
|
1681 E TgE, Tjn, TgH, Tjo, TgF, TgG;
|
cannam@167
|
1682 TgE = TeM - TeP;
|
cannam@167
|
1683 Tjn = Tjl + Tjm;
|
cannam@167
|
1684 TgF = TeR + TeU;
|
cannam@167
|
1685 TgG = TeW - TeZ;
|
cannam@167
|
1686 TgH = TgF + TgG;
|
cannam@167
|
1687 Tjo = TgF - TgG;
|
cannam@167
|
1688 TgI = FMA(KP707106781, TgH, TgE);
|
cannam@167
|
1689 Tjv = FNMS(KP707106781, Tjo, Tjn);
|
cannam@167
|
1690 Tha = FNMS(KP707106781, TgH, TgE);
|
cannam@167
|
1691 Tjp = FMA(KP707106781, Tjo, Tjn);
|
cannam@167
|
1692 }
|
cannam@167
|
1693 {
|
cannam@167
|
1694 E Tfd, Tgp, Tfo, Tgq;
|
cannam@167
|
1695 {
|
cannam@167
|
1696 E Tf7, Tfc, Tfi, Tfn;
|
cannam@167
|
1697 Tf7 = Tf5 + Tf6;
|
cannam@167
|
1698 Tfc = Tf8 + Tfb;
|
cannam@167
|
1699 Tfd = FMA(KP414213562, Tfc, Tf7);
|
cannam@167
|
1700 Tgp = FNMS(KP414213562, Tf7, Tfc);
|
cannam@167
|
1701 Tfi = Tfg + Tfh;
|
cannam@167
|
1702 Tfn = Tfj + Tfm;
|
cannam@167
|
1703 Tfo = FNMS(KP414213562, Tfn, Tfi);
|
cannam@167
|
1704 Tgq = FMA(KP414213562, Tfi, Tfn);
|
cannam@167
|
1705 }
|
cannam@167
|
1706 Tfp = Tfd - Tfo;
|
cannam@167
|
1707 Tjw = Tgq - Tgp;
|
cannam@167
|
1708 Tgr = Tgp + Tgq;
|
cannam@167
|
1709 Tjq = Tfd + Tfo;
|
cannam@167
|
1710 }
|
cannam@167
|
1711 {
|
cannam@167
|
1712 E Th0, Thi, Th3, Thj;
|
cannam@167
|
1713 {
|
cannam@167
|
1714 E TgY, TgZ, Th1, Th2;
|
cannam@167
|
1715 TgY = TfS - TfV;
|
cannam@167
|
1716 TgZ = Tgf + Tge;
|
cannam@167
|
1717 Th0 = FMA(KP707106781, TgZ, TgY);
|
cannam@167
|
1718 Thi = FNMS(KP707106781, TgZ, TgY);
|
cannam@167
|
1719 Th1 = Tgc + Tg9;
|
cannam@167
|
1720 Th2 = Tg6 - Tg1;
|
cannam@167
|
1721 Th3 = FMA(KP707106781, Th2, Th1);
|
cannam@167
|
1722 Thj = FNMS(KP707106781, Th2, Th1);
|
cannam@167
|
1723 }
|
cannam@167
|
1724 Th4 = FNMS(KP198912367, Th3, Th0);
|
cannam@167
|
1725 Tho = FNMS(KP668178637, Thi, Thj);
|
cannam@167
|
1726 Th7 = FMA(KP198912367, Th0, Th3);
|
cannam@167
|
1727 Thk = FMA(KP668178637, Thj, Thi);
|
cannam@167
|
1728 }
|
cannam@167
|
1729 {
|
cannam@167
|
1730 E TfH, Tgt, TfQ, Tgu;
|
cannam@167
|
1731 {
|
cannam@167
|
1732 E Tfv, TfG, TfM, TfP;
|
cannam@167
|
1733 Tfv = Tfr + Tfu;
|
cannam@167
|
1734 TfG = TfA + TfF;
|
cannam@167
|
1735 TfH = FNMS(KP707106781, TfG, Tfv);
|
cannam@167
|
1736 Tgt = FMA(KP707106781, TfG, Tfv);
|
cannam@167
|
1737 TfM = TfK + TfL;
|
cannam@167
|
1738 TfP = TfN + TfO;
|
cannam@167
|
1739 TfQ = FNMS(KP707106781, TfP, TfM);
|
cannam@167
|
1740 Tgu = FMA(KP707106781, TfP, TfM);
|
cannam@167
|
1741 }
|
cannam@167
|
1742 TfR = FMA(KP668178637, TfQ, TfH);
|
cannam@167
|
1743 TgB = FMA(KP198912367, Tgt, Tgu);
|
cannam@167
|
1744 Tgl = FNMS(KP668178637, TfH, TfQ);
|
cannam@167
|
1745 Tgv = FNMS(KP198912367, Tgu, Tgt);
|
cannam@167
|
1746 }
|
cannam@167
|
1747 {
|
cannam@167
|
1748 E TgL, Thc, TgO, Thb;
|
cannam@167
|
1749 {
|
cannam@167
|
1750 E TgJ, TgK, TgM, TgN;
|
cannam@167
|
1751 TgJ = Tf8 - Tfb;
|
cannam@167
|
1752 TgK = Tf5 - Tf6;
|
cannam@167
|
1753 TgL = FMA(KP414213562, TgK, TgJ);
|
cannam@167
|
1754 Thc = FNMS(KP414213562, TgJ, TgK);
|
cannam@167
|
1755 TgM = Tfj - Tfm;
|
cannam@167
|
1756 TgN = Tfg - Tfh;
|
cannam@167
|
1757 TgO = FNMS(KP414213562, TgN, TgM);
|
cannam@167
|
1758 Thb = FMA(KP414213562, TgM, TgN);
|
cannam@167
|
1759 }
|
cannam@167
|
1760 TgP = TgL + TgO;
|
cannam@167
|
1761 TjK = TgL - TgO;
|
cannam@167
|
1762 Thd = Thb - Thc;
|
cannam@167
|
1763 TjE = Thc + Thb;
|
cannam@167
|
1764 }
|
cannam@167
|
1765 {
|
cannam@167
|
1766 E TgT, Thf, TgW, Thg;
|
cannam@167
|
1767 {
|
cannam@167
|
1768 E TgR, TgS, TgU, TgV;
|
cannam@167
|
1769 TgR = Tfr - Tfu;
|
cannam@167
|
1770 TgS = TfN - TfO;
|
cannam@167
|
1771 TgT = FMA(KP707106781, TgS, TgR);
|
cannam@167
|
1772 Thf = FNMS(KP707106781, TgS, TgR);
|
cannam@167
|
1773 TgU = TfK - TfL;
|
cannam@167
|
1774 TgV = TfF - TfA;
|
cannam@167
|
1775 TgW = FMA(KP707106781, TgV, TgU);
|
cannam@167
|
1776 Thg = FNMS(KP707106781, TgV, TgU);
|
cannam@167
|
1777 }
|
cannam@167
|
1778 TgX = FMA(KP198912367, TgW, TgT);
|
cannam@167
|
1779 Thn = FMA(KP668178637, Thf, Thg);
|
cannam@167
|
1780 Th8 = FNMS(KP198912367, TgT, TgW);
|
cannam@167
|
1781 Thh = FNMS(KP668178637, Thg, Thf);
|
cannam@167
|
1782 }
|
cannam@167
|
1783 {
|
cannam@167
|
1784 E Tg8, Tgw, Tgh, Tgx;
|
cannam@167
|
1785 {
|
cannam@167
|
1786 E TfW, Tg7, Tgd, Tgg;
|
cannam@167
|
1787 TfW = TfS + TfV;
|
cannam@167
|
1788 Tg7 = Tg1 + Tg6;
|
cannam@167
|
1789 Tg8 = FNMS(KP707106781, Tg7, TfW);
|
cannam@167
|
1790 Tgw = FMA(KP707106781, Tg7, TfW);
|
cannam@167
|
1791 Tgd = Tg9 - Tgc;
|
cannam@167
|
1792 Tgg = Tge - Tgf;
|
cannam@167
|
1793 Tgh = FNMS(KP707106781, Tgg, Tgd);
|
cannam@167
|
1794 Tgx = FMA(KP707106781, Tgg, Tgd);
|
cannam@167
|
1795 }
|
cannam@167
|
1796 Tgi = FMA(KP668178637, Tgh, Tg8);
|
cannam@167
|
1797 TgC = FMA(KP198912367, Tgw, Tgx);
|
cannam@167
|
1798 Tgm = FNMS(KP668178637, Tg8, Tgh);
|
cannam@167
|
1799 Tgy = FNMS(KP198912367, Tgx, Tgw);
|
cannam@167
|
1800 }
|
cannam@167
|
1801 {
|
cannam@167
|
1802 E Tfq, Tgj, TgA, TgD;
|
cannam@167
|
1803 Tfq = FMA(KP923879532, Tfp, Tf2);
|
cannam@167
|
1804 Tgj = TfR + Tgi;
|
cannam@167
|
1805 ci[WS(rs, 25)] = FNMS(KP831469612, Tgj, Tfq);
|
cannam@167
|
1806 cr[WS(rs, 6)] = FMA(KP831469612, Tgj, Tfq);
|
cannam@167
|
1807 TgA = FNMS(KP923879532, Tgr, Tgo);
|
cannam@167
|
1808 TgD = TgB + TgC;
|
cannam@167
|
1809 ci[WS(rs, 17)] = FNMS(KP980785280, TgD, TgA);
|
cannam@167
|
1810 cr[WS(rs, 14)] = FMA(KP980785280, TgD, TgA);
|
cannam@167
|
1811 }
|
cannam@167
|
1812 {
|
cannam@167
|
1813 E TjF, TjG, TjN, TjO;
|
cannam@167
|
1814 TjF = FMA(KP923879532, TjE, TjD);
|
cannam@167
|
1815 TjG = Th8 + Th7;
|
cannam@167
|
1816 cr[WS(rs, 34)] = FMS(KP980785280, TjG, TjF);
|
cannam@167
|
1817 ci[WS(rs, 61)] = FMA(KP980785280, TjG, TjF);
|
cannam@167
|
1818 TjN = FNMS(KP923879532, TjK, TjJ);
|
cannam@167
|
1819 TjO = Thk - Thh;
|
cannam@167
|
1820 cr[WS(rs, 42)] = FMS(KP831469612, TjO, TjN);
|
cannam@167
|
1821 ci[WS(rs, 53)] = FMA(KP831469612, TjO, TjN);
|
cannam@167
|
1822 }
|
cannam@167
|
1823 {
|
cannam@167
|
1824 E Tgk, Tgn, Tgs, Tgz;
|
cannam@167
|
1825 Tgk = FNMS(KP923879532, Tfp, Tf2);
|
cannam@167
|
1826 Tgn = Tgl + Tgm;
|
cannam@167
|
1827 cr[WS(rs, 22)] = FMA(KP831469612, Tgn, Tgk);
|
cannam@167
|
1828 ci[WS(rs, 9)] = FNMS(KP831469612, Tgn, Tgk);
|
cannam@167
|
1829 Tgs = FMA(KP923879532, Tgr, Tgo);
|
cannam@167
|
1830 Tgz = Tgv + Tgy;
|
cannam@167
|
1831 cr[WS(rs, 30)] = FNMS(KP980785280, Tgz, Tgs);
|
cannam@167
|
1832 ci[WS(rs, 1)] = FMA(KP980785280, Tgz, Tgs);
|
cannam@167
|
1833 }
|
cannam@167
|
1834 {
|
cannam@167
|
1835 E TjH, TjI, TjL, TjM;
|
cannam@167
|
1836 TjH = FNMS(KP923879532, TjE, TjD);
|
cannam@167
|
1837 TjI = Th4 - TgX;
|
cannam@167
|
1838 cr[WS(rs, 50)] = FMS(KP980785280, TjI, TjH);
|
cannam@167
|
1839 ci[WS(rs, 45)] = FMA(KP980785280, TjI, TjH);
|
cannam@167
|
1840 TjL = FMA(KP923879532, TjK, TjJ);
|
cannam@167
|
1841 TjM = Thn + Tho;
|
cannam@167
|
1842 cr[WS(rs, 58)] = -(FMA(KP831469612, TjM, TjL));
|
cannam@167
|
1843 ci[WS(rs, 37)] = FNMS(KP831469612, TjM, TjL);
|
cannam@167
|
1844 }
|
cannam@167
|
1845 {
|
cannam@167
|
1846 E TgQ, Th5, Thm, Thp;
|
cannam@167
|
1847 TgQ = FMA(KP923879532, TgP, TgI);
|
cannam@167
|
1848 Th5 = TgX + Th4;
|
cannam@167
|
1849 ci[WS(rs, 29)] = FNMS(KP980785280, Th5, TgQ);
|
cannam@167
|
1850 cr[WS(rs, 2)] = FMA(KP980785280, Th5, TgQ);
|
cannam@167
|
1851 Thm = FNMS(KP923879532, Thd, Tha);
|
cannam@167
|
1852 Thp = Thn - Tho;
|
cannam@167
|
1853 ci[WS(rs, 21)] = FNMS(KP831469612, Thp, Thm);
|
cannam@167
|
1854 cr[WS(rs, 10)] = FMA(KP831469612, Thp, Thm);
|
cannam@167
|
1855 }
|
cannam@167
|
1856 {
|
cannam@167
|
1857 E Tjr, Tjs, Tjz, TjA;
|
cannam@167
|
1858 Tjr = FMA(KP923879532, Tjq, Tjp);
|
cannam@167
|
1859 Tjs = TgC - TgB;
|
cannam@167
|
1860 cr[WS(rs, 62)] = FMS(KP980785280, Tjs, Tjr);
|
cannam@167
|
1861 ci[WS(rs, 33)] = FMA(KP980785280, Tjs, Tjr);
|
cannam@167
|
1862 Tjz = FNMS(KP923879532, Tjw, Tjv);
|
cannam@167
|
1863 TjA = Tgi - TfR;
|
cannam@167
|
1864 cr[WS(rs, 54)] = FMS(KP831469612, TjA, Tjz);
|
cannam@167
|
1865 ci[WS(rs, 41)] = FMA(KP831469612, TjA, Tjz);
|
cannam@167
|
1866 }
|
cannam@167
|
1867 {
|
cannam@167
|
1868 E Th6, Th9, The, Thl;
|
cannam@167
|
1869 Th6 = FNMS(KP923879532, TgP, TgI);
|
cannam@167
|
1870 Th9 = Th7 - Th8;
|
cannam@167
|
1871 cr[WS(rs, 18)] = FNMS(KP980785280, Th9, Th6);
|
cannam@167
|
1872 ci[WS(rs, 13)] = FMA(KP980785280, Th9, Th6);
|
cannam@167
|
1873 The = FMA(KP923879532, Thd, Tha);
|
cannam@167
|
1874 Thl = Thh + Thk;
|
cannam@167
|
1875 cr[WS(rs, 26)] = FNMS(KP831469612, Thl, The);
|
cannam@167
|
1876 ci[WS(rs, 5)] = FMA(KP831469612, Thl, The);
|
cannam@167
|
1877 }
|
cannam@167
|
1878 {
|
cannam@167
|
1879 E Tjt, Tju, Tjx, Tjy;
|
cannam@167
|
1880 Tjt = FNMS(KP923879532, Tjq, Tjp);
|
cannam@167
|
1881 Tju = Tgy - Tgv;
|
cannam@167
|
1882 cr[WS(rs, 46)] = FMS(KP980785280, Tju, Tjt);
|
cannam@167
|
1883 ci[WS(rs, 49)] = FMA(KP980785280, Tju, Tjt);
|
cannam@167
|
1884 Tjx = FMA(KP923879532, Tjw, Tjv);
|
cannam@167
|
1885 Tjy = Tgl - Tgm;
|
cannam@167
|
1886 cr[WS(rs, 38)] = FMS(KP831469612, Tjy, Tjx);
|
cannam@167
|
1887 ci[WS(rs, 57)] = FMA(KP831469612, Tjy, Tjx);
|
cannam@167
|
1888 }
|
cannam@167
|
1889 }
|
cannam@167
|
1890 {
|
cannam@167
|
1891 E Td1, Tkw, TdN, Tkq, Tdl, TdX, TdI, TdR, Teq, TeK, Tet, TeG, Tej, TeJ, Teu;
|
cannam@167
|
1892 E TeD, Teb, TkK, Tez, TkE, TdE, TdY, TdH, TdU, TcM, TkD, TkJ, TdK, Te4, Tkp;
|
cannam@167
|
1893 E Tkv, Tew;
|
cannam@167
|
1894 {
|
cannam@167
|
1895 E TcT, TdM, Td0, TdL;
|
cannam@167
|
1896 {
|
cannam@167
|
1897 E TcP, TcS, TcW, TcZ;
|
cannam@167
|
1898 TcP = FMA(KP707106781, TcO, TcN);
|
cannam@167
|
1899 TcS = FMA(KP707106781, TcR, TcQ);
|
cannam@167
|
1900 TcT = FMA(KP198912367, TcS, TcP);
|
cannam@167
|
1901 TdM = FNMS(KP198912367, TcP, TcS);
|
cannam@167
|
1902 TcW = FMA(KP707106781, TcV, TcU);
|
cannam@167
|
1903 TcZ = FMA(KP707106781, TcY, TcX);
|
cannam@167
|
1904 Td0 = FNMS(KP198912367, TcZ, TcW);
|
cannam@167
|
1905 TdL = FMA(KP198912367, TcW, TcZ);
|
cannam@167
|
1906 }
|
cannam@167
|
1907 Td1 = TcT + Td0;
|
cannam@167
|
1908 Tkw = TcT - Td0;
|
cannam@167
|
1909 TdN = TdL - TdM;
|
cannam@167
|
1910 Tkq = TdM + TdL;
|
cannam@167
|
1911 }
|
cannam@167
|
1912 {
|
cannam@167
|
1913 E Tdd, TdP, Tdk, TdQ;
|
cannam@167
|
1914 {
|
cannam@167
|
1915 E Td5, Tdc, Tdg, Tdj;
|
cannam@167
|
1916 Td5 = FMA(KP707106781, Td4, Td3);
|
cannam@167
|
1917 Tdc = Td8 + Tdb;
|
cannam@167
|
1918 Tdd = FMA(KP923879532, Tdc, Td5);
|
cannam@167
|
1919 TdP = FNMS(KP923879532, Tdc, Td5);
|
cannam@167
|
1920 Tdg = FMA(KP707106781, Tdf, Tde);
|
cannam@167
|
1921 Tdj = Tdh + Tdi;
|
cannam@167
|
1922 Tdk = FMA(KP923879532, Tdj, Tdg);
|
cannam@167
|
1923 TdQ = FNMS(KP923879532, Tdj, Tdg);
|
cannam@167
|
1924 }
|
cannam@167
|
1925 Tdl = FMA(KP098491403, Tdk, Tdd);
|
cannam@167
|
1926 TdX = FMA(KP820678790, TdP, TdQ);
|
cannam@167
|
1927 TdI = FNMS(KP098491403, Tdd, Tdk);
|
cannam@167
|
1928 TdR = FNMS(KP820678790, TdQ, TdP);
|
cannam@167
|
1929 }
|
cannam@167
|
1930 {
|
cannam@167
|
1931 E Tem, TeE, Tep, TeF;
|
cannam@167
|
1932 {
|
cannam@167
|
1933 E Tek, Tel, Ten, Teo;
|
cannam@167
|
1934 Tek = FNMS(KP707106781, Tdn, Tdm);
|
cannam@167
|
1935 Tel = TdB - TdA;
|
cannam@167
|
1936 Tem = FNMS(KP923879532, Tel, Tek);
|
cannam@167
|
1937 TeE = FMA(KP923879532, Tel, Tek);
|
cannam@167
|
1938 Ten = FNMS(KP707106781, Tdy, Tdx);
|
cannam@167
|
1939 Teo = Tdu - Tdr;
|
cannam@167
|
1940 Tep = FMA(KP923879532, Teo, Ten);
|
cannam@167
|
1941 TeF = FNMS(KP923879532, Teo, Ten);
|
cannam@167
|
1942 }
|
cannam@167
|
1943 Teq = FNMS(KP534511135, Tep, Tem);
|
cannam@167
|
1944 TeK = FNMS(KP303346683, TeE, TeF);
|
cannam@167
|
1945 Tet = FMA(KP534511135, Tem, Tep);
|
cannam@167
|
1946 TeG = FMA(KP303346683, TeF, TeE);
|
cannam@167
|
1947 }
|
cannam@167
|
1948 {
|
cannam@167
|
1949 E Tef, TeB, Tei, TeC;
|
cannam@167
|
1950 {
|
cannam@167
|
1951 E Ted, Tee, Teg, Teh;
|
cannam@167
|
1952 Ted = FNMS(KP707106781, Td4, Td3);
|
cannam@167
|
1953 Tee = Tdi - Tdh;
|
cannam@167
|
1954 Tef = FNMS(KP923879532, Tee, Ted);
|
cannam@167
|
1955 TeB = FMA(KP923879532, Tee, Ted);
|
cannam@167
|
1956 Teg = FNMS(KP707106781, Tdf, Tde);
|
cannam@167
|
1957 Teh = Td8 - Tdb;
|
cannam@167
|
1958 Tei = FNMS(KP923879532, Teh, Teg);
|
cannam@167
|
1959 TeC = FMA(KP923879532, Teh, Teg);
|
cannam@167
|
1960 }
|
cannam@167
|
1961 Tej = FMA(KP534511135, Tei, Tef);
|
cannam@167
|
1962 TeJ = FMA(KP303346683, TeB, TeC);
|
cannam@167
|
1963 Teu = FNMS(KP534511135, Tef, Tei);
|
cannam@167
|
1964 TeD = FNMS(KP303346683, TeC, TeB);
|
cannam@167
|
1965 }
|
cannam@167
|
1966 {
|
cannam@167
|
1967 E Te7, Tex, Tea, Tey;
|
cannam@167
|
1968 {
|
cannam@167
|
1969 E Te5, Te6, Te8, Te9;
|
cannam@167
|
1970 Te5 = FNMS(KP707106781, TcR, TcQ);
|
cannam@167
|
1971 Te6 = FNMS(KP707106781, TcO, TcN);
|
cannam@167
|
1972 Te7 = FMA(KP668178637, Te6, Te5);
|
cannam@167
|
1973 Tex = FNMS(KP668178637, Te5, Te6);
|
cannam@167
|
1974 Te8 = FNMS(KP707106781, TcY, TcX);
|
cannam@167
|
1975 Te9 = FNMS(KP707106781, TcV, TcU);
|
cannam@167
|
1976 Tea = FNMS(KP668178637, Te9, Te8);
|
cannam@167
|
1977 Tey = FMA(KP668178637, Te8, Te9);
|
cannam@167
|
1978 }
|
cannam@167
|
1979 Teb = Te7 - Tea;
|
cannam@167
|
1980 TkK = Tey - Tex;
|
cannam@167
|
1981 Tez = Tex + Tey;
|
cannam@167
|
1982 TkE = Te7 + Tea;
|
cannam@167
|
1983 }
|
cannam@167
|
1984 {
|
cannam@167
|
1985 E Tdw, TdS, TdD, TdT;
|
cannam@167
|
1986 {
|
cannam@167
|
1987 E Tdo, Tdv, Tdz, TdC;
|
cannam@167
|
1988 Tdo = FMA(KP707106781, Tdn, Tdm);
|
cannam@167
|
1989 Tdv = Tdr + Tdu;
|
cannam@167
|
1990 Tdw = FMA(KP923879532, Tdv, Tdo);
|
cannam@167
|
1991 TdS = FNMS(KP923879532, Tdv, Tdo);
|
cannam@167
|
1992 Tdz = FMA(KP707106781, Tdy, Tdx);
|
cannam@167
|
1993 TdC = TdA + TdB;
|
cannam@167
|
1994 TdD = FMA(KP923879532, TdC, Tdz);
|
cannam@167
|
1995 TdT = FNMS(KP923879532, TdC, Tdz);
|
cannam@167
|
1996 }
|
cannam@167
|
1997 TdE = FNMS(KP098491403, TdD, Tdw);
|
cannam@167
|
1998 TdY = FNMS(KP820678790, TdS, TdT);
|
cannam@167
|
1999 TdH = FMA(KP098491403, Tdw, TdD);
|
cannam@167
|
2000 TdU = FMA(KP820678790, TdT, TdS);
|
cannam@167
|
2001 }
|
cannam@167
|
2002 {
|
cannam@167
|
2003 E TcE, Te0, Tkn, TkB, TcL, TkC, Te3, Tko, TcD, Tkm;
|
cannam@167
|
2004 TcD = TcB + TcC;
|
cannam@167
|
2005 TcE = FMA(KP707106781, TcD, TcA);
|
cannam@167
|
2006 Te0 = FNMS(KP707106781, TcD, TcA);
|
cannam@167
|
2007 Tkm = T7s - T7l;
|
cannam@167
|
2008 Tkn = FMA(KP707106781, Tkm, Tkl);
|
cannam@167
|
2009 TkB = FNMS(KP707106781, Tkm, Tkl);
|
cannam@167
|
2010 {
|
cannam@167
|
2011 E TcH, TcK, Te1, Te2;
|
cannam@167
|
2012 TcH = FMA(KP414213562, TcG, TcF);
|
cannam@167
|
2013 TcK = FNMS(KP414213562, TcJ, TcI);
|
cannam@167
|
2014 TcL = TcH + TcK;
|
cannam@167
|
2015 TkC = TcH - TcK;
|
cannam@167
|
2016 Te1 = FMA(KP414213562, TcI, TcJ);
|
cannam@167
|
2017 Te2 = FNMS(KP414213562, TcF, TcG);
|
cannam@167
|
2018 Te3 = Te1 - Te2;
|
cannam@167
|
2019 Tko = Te2 + Te1;
|
cannam@167
|
2020 }
|
cannam@167
|
2021 TcM = FMA(KP923879532, TcL, TcE);
|
cannam@167
|
2022 TkD = FMA(KP923879532, TkC, TkB);
|
cannam@167
|
2023 TkJ = FNMS(KP923879532, TkC, TkB);
|
cannam@167
|
2024 TdK = FNMS(KP923879532, TcL, TcE);
|
cannam@167
|
2025 Te4 = FNMS(KP923879532, Te3, Te0);
|
cannam@167
|
2026 Tkp = FMA(KP923879532, Tko, Tkn);
|
cannam@167
|
2027 Tkv = FNMS(KP923879532, Tko, Tkn);
|
cannam@167
|
2028 Tew = FMA(KP923879532, Te3, Te0);
|
cannam@167
|
2029 }
|
cannam@167
|
2030 {
|
cannam@167
|
2031 E Td2, TdF, TdW, TdZ;
|
cannam@167
|
2032 Td2 = FMA(KP980785280, Td1, TcM);
|
cannam@167
|
2033 TdF = Tdl + TdE;
|
cannam@167
|
2034 ci[WS(rs, 30)] = FNMS(KP995184726, TdF, Td2);
|
cannam@167
|
2035 cr[WS(rs, 1)] = FMA(KP995184726, TdF, Td2);
|
cannam@167
|
2036 TdW = FNMS(KP980785280, TdN, TdK);
|
cannam@167
|
2037 TdZ = TdX - TdY;
|
cannam@167
|
2038 ci[WS(rs, 22)] = FNMS(KP773010453, TdZ, TdW);
|
cannam@167
|
2039 cr[WS(rs, 9)] = FMA(KP773010453, TdZ, TdW);
|
cannam@167
|
2040 }
|
cannam@167
|
2041 {
|
cannam@167
|
2042 E TkF, TkG, TkN, TkO;
|
cannam@167
|
2043 TkF = FMA(KP831469612, TkE, TkD);
|
cannam@167
|
2044 TkG = TeJ + TeK;
|
cannam@167
|
2045 cr[WS(rs, 61)] = -(FMA(KP956940335, TkG, TkF));
|
cannam@167
|
2046 ci[WS(rs, 34)] = FNMS(KP956940335, TkG, TkF);
|
cannam@167
|
2047 TkN = FNMS(KP831469612, TkK, TkJ);
|
cannam@167
|
2048 TkO = Teq - Tej;
|
cannam@167
|
2049 cr[WS(rs, 53)] = FMS(KP881921264, TkO, TkN);
|
cannam@167
|
2050 ci[WS(rs, 42)] = FMA(KP881921264, TkO, TkN);
|
cannam@167
|
2051 }
|
cannam@167
|
2052 {
|
cannam@167
|
2053 E TdG, TdJ, TdO, TdV;
|
cannam@167
|
2054 TdG = FNMS(KP980785280, Td1, TcM);
|
cannam@167
|
2055 TdJ = TdH - TdI;
|
cannam@167
|
2056 cr[WS(rs, 17)] = FNMS(KP995184726, TdJ, TdG);
|
cannam@167
|
2057 ci[WS(rs, 14)] = FMA(KP995184726, TdJ, TdG);
|
cannam@167
|
2058 TdO = FMA(KP980785280, TdN, TdK);
|
cannam@167
|
2059 TdV = TdR + TdU;
|
cannam@167
|
2060 cr[WS(rs, 25)] = FNMS(KP773010453, TdV, TdO);
|
cannam@167
|
2061 ci[WS(rs, 6)] = FMA(KP773010453, TdV, TdO);
|
cannam@167
|
2062 }
|
cannam@167
|
2063 {
|
cannam@167
|
2064 E TkH, TkI, TkL, TkM;
|
cannam@167
|
2065 TkH = FNMS(KP831469612, TkE, TkD);
|
cannam@167
|
2066 TkI = TeG - TeD;
|
cannam@167
|
2067 cr[WS(rs, 45)] = FMS(KP956940335, TkI, TkH);
|
cannam@167
|
2068 ci[WS(rs, 50)] = FMA(KP956940335, TkI, TkH);
|
cannam@167
|
2069 TkL = FMA(KP831469612, TkK, TkJ);
|
cannam@167
|
2070 TkM = Teu + Tet;
|
cannam@167
|
2071 cr[WS(rs, 37)] = FMS(KP881921264, TkM, TkL);
|
cannam@167
|
2072 ci[WS(rs, 58)] = FMA(KP881921264, TkM, TkL);
|
cannam@167
|
2073 }
|
cannam@167
|
2074 {
|
cannam@167
|
2075 E Tec, Ter, TeI, TeL;
|
cannam@167
|
2076 Tec = FMA(KP831469612, Teb, Te4);
|
cannam@167
|
2077 Ter = Tej + Teq;
|
cannam@167
|
2078 ci[WS(rs, 26)] = FNMS(KP881921264, Ter, Tec);
|
cannam@167
|
2079 cr[WS(rs, 5)] = FMA(KP881921264, Ter, Tec);
|
cannam@167
|
2080 TeI = FNMS(KP831469612, Tez, Tew);
|
cannam@167
|
2081 TeL = TeJ - TeK;
|
cannam@167
|
2082 ci[WS(rs, 18)] = FNMS(KP956940335, TeL, TeI);
|
cannam@167
|
2083 cr[WS(rs, 13)] = FMA(KP956940335, TeL, TeI);
|
cannam@167
|
2084 }
|
cannam@167
|
2085 {
|
cannam@167
|
2086 E Tkr, Tks, Tkz, TkA;
|
cannam@167
|
2087 Tkr = FMA(KP980785280, Tkq, Tkp);
|
cannam@167
|
2088 Tks = TdI + TdH;
|
cannam@167
|
2089 cr[WS(rs, 33)] = FMS(KP995184726, Tks, Tkr);
|
cannam@167
|
2090 ci[WS(rs, 62)] = FMA(KP995184726, Tks, Tkr);
|
cannam@167
|
2091 Tkz = FNMS(KP980785280, Tkw, Tkv);
|
cannam@167
|
2092 TkA = TdU - TdR;
|
cannam@167
|
2093 cr[WS(rs, 41)] = FMS(KP773010453, TkA, Tkz);
|
cannam@167
|
2094 ci[WS(rs, 54)] = FMA(KP773010453, TkA, Tkz);
|
cannam@167
|
2095 }
|
cannam@167
|
2096 {
|
cannam@167
|
2097 E Tes, Tev, TeA, TeH;
|
cannam@167
|
2098 Tes = FNMS(KP831469612, Teb, Te4);
|
cannam@167
|
2099 Tev = Tet - Teu;
|
cannam@167
|
2100 cr[WS(rs, 21)] = FNMS(KP881921264, Tev, Tes);
|
cannam@167
|
2101 ci[WS(rs, 10)] = FMA(KP881921264, Tev, Tes);
|
cannam@167
|
2102 TeA = FMA(KP831469612, Tez, Tew);
|
cannam@167
|
2103 TeH = TeD + TeG;
|
cannam@167
|
2104 cr[WS(rs, 29)] = FNMS(KP956940335, TeH, TeA);
|
cannam@167
|
2105 ci[WS(rs, 2)] = FMA(KP956940335, TeH, TeA);
|
cannam@167
|
2106 }
|
cannam@167
|
2107 {
|
cannam@167
|
2108 E Tkt, Tku, Tkx, Tky;
|
cannam@167
|
2109 Tkt = FNMS(KP980785280, Tkq, Tkp);
|
cannam@167
|
2110 Tku = TdE - Tdl;
|
cannam@167
|
2111 cr[WS(rs, 49)] = FMS(KP995184726, Tku, Tkt);
|
cannam@167
|
2112 ci[WS(rs, 46)] = FMA(KP995184726, Tku, Tkt);
|
cannam@167
|
2113 Tkx = FMA(KP980785280, Tkw, Tkv);
|
cannam@167
|
2114 Tky = TdX + TdY;
|
cannam@167
|
2115 cr[WS(rs, 57)] = -(FMA(KP773010453, Tky, Tkx));
|
cannam@167
|
2116 ci[WS(rs, 38)] = FNMS(KP773010453, Tky, Tkx);
|
cannam@167
|
2117 }
|
cannam@167
|
2118 }
|
cannam@167
|
2119 }
|
cannam@167
|
2120 }
|
cannam@167
|
2121 }
|
cannam@167
|
2122
|
cannam@167
|
2123 static const tw_instr twinstr[] = {
|
cannam@167
|
2124 {TW_FULL, 1, 64},
|
cannam@167
|
2125 {TW_NEXT, 1, 0}
|
cannam@167
|
2126 };
|
cannam@167
|
2127
|
cannam@167
|
2128 static const hc2hc_desc desc = { 64, "hf_64", twinstr, &GENUS, {520, 126, 518, 0} };
|
cannam@167
|
2129
|
cannam@167
|
2130 void X(codelet_hf_64) (planner *p) {
|
cannam@167
|
2131 X(khc2hc_register) (p, hf_64, &desc);
|
cannam@167
|
2132 }
|
cannam@167
|
2133 #else
|
cannam@167
|
2134
|
cannam@167
|
2135 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 64 -dit -name hf_64 -include rdft/scalar/hf.h */
|
cannam@167
|
2136
|
cannam@167
|
2137 /*
|
cannam@167
|
2138 * This function contains 1038 FP additions, 500 FP multiplications,
|
cannam@167
|
2139 * (or, 808 additions, 270 multiplications, 230 fused multiply/add),
|
cannam@167
|
2140 * 176 stack variables, 15 constants, and 256 memory accesses
|
cannam@167
|
2141 */
|
cannam@167
|
2142 #include "rdft/scalar/hf.h"
|
cannam@167
|
2143
|
cannam@167
|
2144 static void hf_64(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
2145 {
|
cannam@167
|
2146 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
cannam@167
|
2147 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
2148 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
2149 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
cannam@167
|
2150 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
2151 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
2152 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
cannam@167
|
2153 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
2154 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
2155 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
cannam@167
|
2156 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
2157 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
2158 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
2159 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
2160 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
2161 {
|
cannam@167
|
2162 INT m;
|
cannam@167
|
2163 for (m = mb, W = W + ((mb - 1) * 126); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 126, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
2164 E Tj, TcL, ThT, Tin, T6b, Taz, TgT, Thn, TG, Thm, TcO, TgO, T6m, Tim, TaC;
|
cannam@167
|
2165 E ThQ, T14, Tfr, T6y, T9O, TaG, Tc0, TcU, TeE, T1r, Tfq, T6J, T9P, TaJ, Tc1;
|
cannam@167
|
2166 E TcZ, TeF, T1Q, T2d, Tfu, Tfv, Tfw, Tfx, T6Q, TaM, Tdb, TeI, T71, TaQ, T7a;
|
cannam@167
|
2167 E TaN, Td6, TeJ, T77, TaP, T2B, T2Y, Tfz, TfA, TfB, TfC, T7h, TaW, Tdm, TeL;
|
cannam@167
|
2168 E T7s, TaU, T7B, TaX, Tdh, TeM, T7y, TaT, T5j, TfR, Tec, TeX, TfY, Tgy, T8D;
|
cannam@167
|
2169 E Tbl, T8O, Tbx, T9l, Tbm, TdV, Tf0, T9i, Tbw, T3M, TfL, TdL, TeT, TfI, Tgt;
|
cannam@167
|
2170 E T7K, Tbd, T7V, Tb3, T8s, Tbe, Tdu, TeQ, T8p, Tb2, T4x, TfJ, TdE, TdM, TfO;
|
cannam@167
|
2171 E Tgu, T87, T8u, T8i, T8v, Tba, Tbh, Tdz, TdN, Tb7, Tbg, T64, TfZ, Te5, Ted;
|
cannam@167
|
2172 E TfU, Tgz, T90, T9n, T9b, T9o, Tbt, TbA, Te0, Tee, Tbq, Tbz;
|
cannam@167
|
2173 {
|
cannam@167
|
2174 E T1, TgR, T6, TgQ, Tc, T68, Th, T69;
|
cannam@167
|
2175 T1 = cr[0];
|
cannam@167
|
2176 TgR = ci[0];
|
cannam@167
|
2177 {
|
cannam@167
|
2178 E T3, T5, T2, T4;
|
cannam@167
|
2179 T3 = cr[WS(rs, 32)];
|
cannam@167
|
2180 T5 = ci[WS(rs, 32)];
|
cannam@167
|
2181 T2 = W[62];
|
cannam@167
|
2182 T4 = W[63];
|
cannam@167
|
2183 T6 = FMA(T2, T3, T4 * T5);
|
cannam@167
|
2184 TgQ = FNMS(T4, T3, T2 * T5);
|
cannam@167
|
2185 }
|
cannam@167
|
2186 {
|
cannam@167
|
2187 E T9, Tb, T8, Ta;
|
cannam@167
|
2188 T9 = cr[WS(rs, 16)];
|
cannam@167
|
2189 Tb = ci[WS(rs, 16)];
|
cannam@167
|
2190 T8 = W[30];
|
cannam@167
|
2191 Ta = W[31];
|
cannam@167
|
2192 Tc = FMA(T8, T9, Ta * Tb);
|
cannam@167
|
2193 T68 = FNMS(Ta, T9, T8 * Tb);
|
cannam@167
|
2194 }
|
cannam@167
|
2195 {
|
cannam@167
|
2196 E Te, Tg, Td, Tf;
|
cannam@167
|
2197 Te = cr[WS(rs, 48)];
|
cannam@167
|
2198 Tg = ci[WS(rs, 48)];
|
cannam@167
|
2199 Td = W[94];
|
cannam@167
|
2200 Tf = W[95];
|
cannam@167
|
2201 Th = FMA(Td, Te, Tf * Tg);
|
cannam@167
|
2202 T69 = FNMS(Tf, Te, Td * Tg);
|
cannam@167
|
2203 }
|
cannam@167
|
2204 {
|
cannam@167
|
2205 E T7, Ti, ThR, ThS;
|
cannam@167
|
2206 T7 = T1 + T6;
|
cannam@167
|
2207 Ti = Tc + Th;
|
cannam@167
|
2208 Tj = T7 + Ti;
|
cannam@167
|
2209 TcL = T7 - Ti;
|
cannam@167
|
2210 ThR = Tc - Th;
|
cannam@167
|
2211 ThS = TgR - TgQ;
|
cannam@167
|
2212 ThT = ThR + ThS;
|
cannam@167
|
2213 Tin = ThS - ThR;
|
cannam@167
|
2214 }
|
cannam@167
|
2215 {
|
cannam@167
|
2216 E T67, T6a, TgP, TgS;
|
cannam@167
|
2217 T67 = T1 - T6;
|
cannam@167
|
2218 T6a = T68 - T69;
|
cannam@167
|
2219 T6b = T67 - T6a;
|
cannam@167
|
2220 Taz = T67 + T6a;
|
cannam@167
|
2221 TgP = T68 + T69;
|
cannam@167
|
2222 TgS = TgQ + TgR;
|
cannam@167
|
2223 TgT = TgP + TgS;
|
cannam@167
|
2224 Thn = TgS - TgP;
|
cannam@167
|
2225 }
|
cannam@167
|
2226 }
|
cannam@167
|
2227 {
|
cannam@167
|
2228 E To, T6d, Tt, T6e, T6c, T6f, Tz, T6i, TE, T6j, T6h, T6k;
|
cannam@167
|
2229 {
|
cannam@167
|
2230 E Tl, Tn, Tk, Tm;
|
cannam@167
|
2231 Tl = cr[WS(rs, 8)];
|
cannam@167
|
2232 Tn = ci[WS(rs, 8)];
|
cannam@167
|
2233 Tk = W[14];
|
cannam@167
|
2234 Tm = W[15];
|
cannam@167
|
2235 To = FMA(Tk, Tl, Tm * Tn);
|
cannam@167
|
2236 T6d = FNMS(Tm, Tl, Tk * Tn);
|
cannam@167
|
2237 }
|
cannam@167
|
2238 {
|
cannam@167
|
2239 E Tq, Ts, Tp, Tr;
|
cannam@167
|
2240 Tq = cr[WS(rs, 40)];
|
cannam@167
|
2241 Ts = ci[WS(rs, 40)];
|
cannam@167
|
2242 Tp = W[78];
|
cannam@167
|
2243 Tr = W[79];
|
cannam@167
|
2244 Tt = FMA(Tp, Tq, Tr * Ts);
|
cannam@167
|
2245 T6e = FNMS(Tr, Tq, Tp * Ts);
|
cannam@167
|
2246 }
|
cannam@167
|
2247 T6c = To - Tt;
|
cannam@167
|
2248 T6f = T6d - T6e;
|
cannam@167
|
2249 {
|
cannam@167
|
2250 E Tw, Ty, Tv, Tx;
|
cannam@167
|
2251 Tw = cr[WS(rs, 56)];
|
cannam@167
|
2252 Ty = ci[WS(rs, 56)];
|
cannam@167
|
2253 Tv = W[110];
|
cannam@167
|
2254 Tx = W[111];
|
cannam@167
|
2255 Tz = FMA(Tv, Tw, Tx * Ty);
|
cannam@167
|
2256 T6i = FNMS(Tx, Tw, Tv * Ty);
|
cannam@167
|
2257 }
|
cannam@167
|
2258 {
|
cannam@167
|
2259 E TB, TD, TA, TC;
|
cannam@167
|
2260 TB = cr[WS(rs, 24)];
|
cannam@167
|
2261 TD = ci[WS(rs, 24)];
|
cannam@167
|
2262 TA = W[46];
|
cannam@167
|
2263 TC = W[47];
|
cannam@167
|
2264 TE = FMA(TA, TB, TC * TD);
|
cannam@167
|
2265 T6j = FNMS(TC, TB, TA * TD);
|
cannam@167
|
2266 }
|
cannam@167
|
2267 T6h = Tz - TE;
|
cannam@167
|
2268 T6k = T6i - T6j;
|
cannam@167
|
2269 {
|
cannam@167
|
2270 E Tu, TF, TcM, TcN;
|
cannam@167
|
2271 Tu = To + Tt;
|
cannam@167
|
2272 TF = Tz + TE;
|
cannam@167
|
2273 TG = Tu + TF;
|
cannam@167
|
2274 Thm = Tu - TF;
|
cannam@167
|
2275 TcM = T6i + T6j;
|
cannam@167
|
2276 TcN = T6d + T6e;
|
cannam@167
|
2277 TcO = TcM - TcN;
|
cannam@167
|
2278 TgO = TcN + TcM;
|
cannam@167
|
2279 }
|
cannam@167
|
2280 {
|
cannam@167
|
2281 E T6g, T6l, TaA, TaB;
|
cannam@167
|
2282 T6g = T6c - T6f;
|
cannam@167
|
2283 T6l = T6h + T6k;
|
cannam@167
|
2284 T6m = KP707106781 * (T6g + T6l);
|
cannam@167
|
2285 Tim = KP707106781 * (T6l - T6g);
|
cannam@167
|
2286 TaA = T6c + T6f;
|
cannam@167
|
2287 TaB = T6h - T6k;
|
cannam@167
|
2288 TaC = KP707106781 * (TaA + TaB);
|
cannam@167
|
2289 ThQ = KP707106781 * (TaA - TaB);
|
cannam@167
|
2290 }
|
cannam@167
|
2291 }
|
cannam@167
|
2292 {
|
cannam@167
|
2293 E TS, TcR, T6o, T6v, T13, TcS, T6r, T6w, T6s, T6x;
|
cannam@167
|
2294 {
|
cannam@167
|
2295 E TM, T6t, TR, T6u;
|
cannam@167
|
2296 {
|
cannam@167
|
2297 E TJ, TL, TI, TK;
|
cannam@167
|
2298 TJ = cr[WS(rs, 4)];
|
cannam@167
|
2299 TL = ci[WS(rs, 4)];
|
cannam@167
|
2300 TI = W[6];
|
cannam@167
|
2301 TK = W[7];
|
cannam@167
|
2302 TM = FMA(TI, TJ, TK * TL);
|
cannam@167
|
2303 T6t = FNMS(TK, TJ, TI * TL);
|
cannam@167
|
2304 }
|
cannam@167
|
2305 {
|
cannam@167
|
2306 E TO, TQ, TN, TP;
|
cannam@167
|
2307 TO = cr[WS(rs, 36)];
|
cannam@167
|
2308 TQ = ci[WS(rs, 36)];
|
cannam@167
|
2309 TN = W[70];
|
cannam@167
|
2310 TP = W[71];
|
cannam@167
|
2311 TR = FMA(TN, TO, TP * TQ);
|
cannam@167
|
2312 T6u = FNMS(TP, TO, TN * TQ);
|
cannam@167
|
2313 }
|
cannam@167
|
2314 TS = TM + TR;
|
cannam@167
|
2315 TcR = T6t + T6u;
|
cannam@167
|
2316 T6o = TM - TR;
|
cannam@167
|
2317 T6v = T6t - T6u;
|
cannam@167
|
2318 }
|
cannam@167
|
2319 {
|
cannam@167
|
2320 E TX, T6p, T12, T6q;
|
cannam@167
|
2321 {
|
cannam@167
|
2322 E TU, TW, TT, TV;
|
cannam@167
|
2323 TU = cr[WS(rs, 20)];
|
cannam@167
|
2324 TW = ci[WS(rs, 20)];
|
cannam@167
|
2325 TT = W[38];
|
cannam@167
|
2326 TV = W[39];
|
cannam@167
|
2327 TX = FMA(TT, TU, TV * TW);
|
cannam@167
|
2328 T6p = FNMS(TV, TU, TT * TW);
|
cannam@167
|
2329 }
|
cannam@167
|
2330 {
|
cannam@167
|
2331 E TZ, T11, TY, T10;
|
cannam@167
|
2332 TZ = cr[WS(rs, 52)];
|
cannam@167
|
2333 T11 = ci[WS(rs, 52)];
|
cannam@167
|
2334 TY = W[102];
|
cannam@167
|
2335 T10 = W[103];
|
cannam@167
|
2336 T12 = FMA(TY, TZ, T10 * T11);
|
cannam@167
|
2337 T6q = FNMS(T10, TZ, TY * T11);
|
cannam@167
|
2338 }
|
cannam@167
|
2339 T13 = TX + T12;
|
cannam@167
|
2340 TcS = T6p + T6q;
|
cannam@167
|
2341 T6r = T6p - T6q;
|
cannam@167
|
2342 T6w = TX - T12;
|
cannam@167
|
2343 }
|
cannam@167
|
2344 T14 = TS + T13;
|
cannam@167
|
2345 Tfr = TcR + TcS;
|
cannam@167
|
2346 T6s = T6o - T6r;
|
cannam@167
|
2347 T6x = T6v + T6w;
|
cannam@167
|
2348 T6y = FNMS(KP382683432, T6x, KP923879532 * T6s);
|
cannam@167
|
2349 T9O = FMA(KP923879532, T6x, KP382683432 * T6s);
|
cannam@167
|
2350 {
|
cannam@167
|
2351 E TaE, TaF, TcQ, TcT;
|
cannam@167
|
2352 TaE = T6v - T6w;
|
cannam@167
|
2353 TaF = T6o + T6r;
|
cannam@167
|
2354 TaG = FMA(KP382683432, TaE, KP923879532 * TaF);
|
cannam@167
|
2355 Tc0 = FNMS(KP923879532, TaE, KP382683432 * TaF);
|
cannam@167
|
2356 TcQ = TS - T13;
|
cannam@167
|
2357 TcT = TcR - TcS;
|
cannam@167
|
2358 TcU = TcQ + TcT;
|
cannam@167
|
2359 TeE = TcQ - TcT;
|
cannam@167
|
2360 }
|
cannam@167
|
2361 }
|
cannam@167
|
2362 {
|
cannam@167
|
2363 E T1f, TcW, T6B, T6E, T1q, TcX, T6C, T6H, T6D, T6I;
|
cannam@167
|
2364 {
|
cannam@167
|
2365 E T19, T6z, T1e, T6A;
|
cannam@167
|
2366 {
|
cannam@167
|
2367 E T16, T18, T15, T17;
|
cannam@167
|
2368 T16 = cr[WS(rs, 60)];
|
cannam@167
|
2369 T18 = ci[WS(rs, 60)];
|
cannam@167
|
2370 T15 = W[118];
|
cannam@167
|
2371 T17 = W[119];
|
cannam@167
|
2372 T19 = FMA(T15, T16, T17 * T18);
|
cannam@167
|
2373 T6z = FNMS(T17, T16, T15 * T18);
|
cannam@167
|
2374 }
|
cannam@167
|
2375 {
|
cannam@167
|
2376 E T1b, T1d, T1a, T1c;
|
cannam@167
|
2377 T1b = cr[WS(rs, 28)];
|
cannam@167
|
2378 T1d = ci[WS(rs, 28)];
|
cannam@167
|
2379 T1a = W[54];
|
cannam@167
|
2380 T1c = W[55];
|
cannam@167
|
2381 T1e = FMA(T1a, T1b, T1c * T1d);
|
cannam@167
|
2382 T6A = FNMS(T1c, T1b, T1a * T1d);
|
cannam@167
|
2383 }
|
cannam@167
|
2384 T1f = T19 + T1e;
|
cannam@167
|
2385 TcW = T6z + T6A;
|
cannam@167
|
2386 T6B = T6z - T6A;
|
cannam@167
|
2387 T6E = T19 - T1e;
|
cannam@167
|
2388 }
|
cannam@167
|
2389 {
|
cannam@167
|
2390 E T1k, T6F, T1p, T6G;
|
cannam@167
|
2391 {
|
cannam@167
|
2392 E T1h, T1j, T1g, T1i;
|
cannam@167
|
2393 T1h = cr[WS(rs, 12)];
|
cannam@167
|
2394 T1j = ci[WS(rs, 12)];
|
cannam@167
|
2395 T1g = W[22];
|
cannam@167
|
2396 T1i = W[23];
|
cannam@167
|
2397 T1k = FMA(T1g, T1h, T1i * T1j);
|
cannam@167
|
2398 T6F = FNMS(T1i, T1h, T1g * T1j);
|
cannam@167
|
2399 }
|
cannam@167
|
2400 {
|
cannam@167
|
2401 E T1m, T1o, T1l, T1n;
|
cannam@167
|
2402 T1m = cr[WS(rs, 44)];
|
cannam@167
|
2403 T1o = ci[WS(rs, 44)];
|
cannam@167
|
2404 T1l = W[86];
|
cannam@167
|
2405 T1n = W[87];
|
cannam@167
|
2406 T1p = FMA(T1l, T1m, T1n * T1o);
|
cannam@167
|
2407 T6G = FNMS(T1n, T1m, T1l * T1o);
|
cannam@167
|
2408 }
|
cannam@167
|
2409 T1q = T1k + T1p;
|
cannam@167
|
2410 TcX = T6F + T6G;
|
cannam@167
|
2411 T6C = T1k - T1p;
|
cannam@167
|
2412 T6H = T6F - T6G;
|
cannam@167
|
2413 }
|
cannam@167
|
2414 T1r = T1f + T1q;
|
cannam@167
|
2415 Tfq = TcW + TcX;
|
cannam@167
|
2416 T6D = T6B + T6C;
|
cannam@167
|
2417 T6I = T6E - T6H;
|
cannam@167
|
2418 T6J = FMA(KP382683432, T6D, KP923879532 * T6I);
|
cannam@167
|
2419 T9P = FNMS(KP923879532, T6D, KP382683432 * T6I);
|
cannam@167
|
2420 {
|
cannam@167
|
2421 E TaH, TaI, TcV, TcY;
|
cannam@167
|
2422 TaH = T6E + T6H;
|
cannam@167
|
2423 TaI = T6B - T6C;
|
cannam@167
|
2424 TaJ = FNMS(KP382683432, TaI, KP923879532 * TaH);
|
cannam@167
|
2425 Tc1 = FMA(KP923879532, TaI, KP382683432 * TaH);
|
cannam@167
|
2426 TcV = T1f - T1q;
|
cannam@167
|
2427 TcY = TcW - TcX;
|
cannam@167
|
2428 TcZ = TcV - TcY;
|
cannam@167
|
2429 TeF = TcV + TcY;
|
cannam@167
|
2430 }
|
cannam@167
|
2431 }
|
cannam@167
|
2432 {
|
cannam@167
|
2433 E T1y, T73, T1D, T74, T1E, Td7, T1J, T6N, T1O, T6O, T1P, Td8, T21, Td4, T6R;
|
cannam@167
|
2434 E T6U, T2c, Td3, T6W, T6Z;
|
cannam@167
|
2435 {
|
cannam@167
|
2436 E T1v, T1x, T1u, T1w;
|
cannam@167
|
2437 T1v = cr[WS(rs, 2)];
|
cannam@167
|
2438 T1x = ci[WS(rs, 2)];
|
cannam@167
|
2439 T1u = W[2];
|
cannam@167
|
2440 T1w = W[3];
|
cannam@167
|
2441 T1y = FMA(T1u, T1v, T1w * T1x);
|
cannam@167
|
2442 T73 = FNMS(T1w, T1v, T1u * T1x);
|
cannam@167
|
2443 }
|
cannam@167
|
2444 {
|
cannam@167
|
2445 E T1A, T1C, T1z, T1B;
|
cannam@167
|
2446 T1A = cr[WS(rs, 34)];
|
cannam@167
|
2447 T1C = ci[WS(rs, 34)];
|
cannam@167
|
2448 T1z = W[66];
|
cannam@167
|
2449 T1B = W[67];
|
cannam@167
|
2450 T1D = FMA(T1z, T1A, T1B * T1C);
|
cannam@167
|
2451 T74 = FNMS(T1B, T1A, T1z * T1C);
|
cannam@167
|
2452 }
|
cannam@167
|
2453 T1E = T1y + T1D;
|
cannam@167
|
2454 Td7 = T73 + T74;
|
cannam@167
|
2455 {
|
cannam@167
|
2456 E T1G, T1I, T1F, T1H;
|
cannam@167
|
2457 T1G = cr[WS(rs, 18)];
|
cannam@167
|
2458 T1I = ci[WS(rs, 18)];
|
cannam@167
|
2459 T1F = W[34];
|
cannam@167
|
2460 T1H = W[35];
|
cannam@167
|
2461 T1J = FMA(T1F, T1G, T1H * T1I);
|
cannam@167
|
2462 T6N = FNMS(T1H, T1G, T1F * T1I);
|
cannam@167
|
2463 }
|
cannam@167
|
2464 {
|
cannam@167
|
2465 E T1L, T1N, T1K, T1M;
|
cannam@167
|
2466 T1L = cr[WS(rs, 50)];
|
cannam@167
|
2467 T1N = ci[WS(rs, 50)];
|
cannam@167
|
2468 T1K = W[98];
|
cannam@167
|
2469 T1M = W[99];
|
cannam@167
|
2470 T1O = FMA(T1K, T1L, T1M * T1N);
|
cannam@167
|
2471 T6O = FNMS(T1M, T1L, T1K * T1N);
|
cannam@167
|
2472 }
|
cannam@167
|
2473 T1P = T1J + T1O;
|
cannam@167
|
2474 Td8 = T6N + T6O;
|
cannam@167
|
2475 {
|
cannam@167
|
2476 E T1V, T6S, T20, T6T;
|
cannam@167
|
2477 {
|
cannam@167
|
2478 E T1S, T1U, T1R, T1T;
|
cannam@167
|
2479 T1S = cr[WS(rs, 10)];
|
cannam@167
|
2480 T1U = ci[WS(rs, 10)];
|
cannam@167
|
2481 T1R = W[18];
|
cannam@167
|
2482 T1T = W[19];
|
cannam@167
|
2483 T1V = FMA(T1R, T1S, T1T * T1U);
|
cannam@167
|
2484 T6S = FNMS(T1T, T1S, T1R * T1U);
|
cannam@167
|
2485 }
|
cannam@167
|
2486 {
|
cannam@167
|
2487 E T1X, T1Z, T1W, T1Y;
|
cannam@167
|
2488 T1X = cr[WS(rs, 42)];
|
cannam@167
|
2489 T1Z = ci[WS(rs, 42)];
|
cannam@167
|
2490 T1W = W[82];
|
cannam@167
|
2491 T1Y = W[83];
|
cannam@167
|
2492 T20 = FMA(T1W, T1X, T1Y * T1Z);
|
cannam@167
|
2493 T6T = FNMS(T1Y, T1X, T1W * T1Z);
|
cannam@167
|
2494 }
|
cannam@167
|
2495 T21 = T1V + T20;
|
cannam@167
|
2496 Td4 = T6S + T6T;
|
cannam@167
|
2497 T6R = T1V - T20;
|
cannam@167
|
2498 T6U = T6S - T6T;
|
cannam@167
|
2499 }
|
cannam@167
|
2500 {
|
cannam@167
|
2501 E T26, T6X, T2b, T6Y;
|
cannam@167
|
2502 {
|
cannam@167
|
2503 E T23, T25, T22, T24;
|
cannam@167
|
2504 T23 = cr[WS(rs, 58)];
|
cannam@167
|
2505 T25 = ci[WS(rs, 58)];
|
cannam@167
|
2506 T22 = W[114];
|
cannam@167
|
2507 T24 = W[115];
|
cannam@167
|
2508 T26 = FMA(T22, T23, T24 * T25);
|
cannam@167
|
2509 T6X = FNMS(T24, T23, T22 * T25);
|
cannam@167
|
2510 }
|
cannam@167
|
2511 {
|
cannam@167
|
2512 E T28, T2a, T27, T29;
|
cannam@167
|
2513 T28 = cr[WS(rs, 26)];
|
cannam@167
|
2514 T2a = ci[WS(rs, 26)];
|
cannam@167
|
2515 T27 = W[50];
|
cannam@167
|
2516 T29 = W[51];
|
cannam@167
|
2517 T2b = FMA(T27, T28, T29 * T2a);
|
cannam@167
|
2518 T6Y = FNMS(T29, T28, T27 * T2a);
|
cannam@167
|
2519 }
|
cannam@167
|
2520 T2c = T26 + T2b;
|
cannam@167
|
2521 Td3 = T6X + T6Y;
|
cannam@167
|
2522 T6W = T26 - T2b;
|
cannam@167
|
2523 T6Z = T6X - T6Y;
|
cannam@167
|
2524 }
|
cannam@167
|
2525 T1Q = T1E + T1P;
|
cannam@167
|
2526 T2d = T21 + T2c;
|
cannam@167
|
2527 Tfu = T1Q - T2d;
|
cannam@167
|
2528 Tfv = Td7 + Td8;
|
cannam@167
|
2529 Tfw = Td4 + Td3;
|
cannam@167
|
2530 Tfx = Tfv - Tfw;
|
cannam@167
|
2531 {
|
cannam@167
|
2532 E T6M, T6P, Td9, Tda;
|
cannam@167
|
2533 T6M = T1y - T1D;
|
cannam@167
|
2534 T6P = T6N - T6O;
|
cannam@167
|
2535 T6Q = T6M - T6P;
|
cannam@167
|
2536 TaM = T6M + T6P;
|
cannam@167
|
2537 Td9 = Td7 - Td8;
|
cannam@167
|
2538 Tda = T21 - T2c;
|
cannam@167
|
2539 Tdb = Td9 - Tda;
|
cannam@167
|
2540 TeI = Td9 + Tda;
|
cannam@167
|
2541 }
|
cannam@167
|
2542 {
|
cannam@167
|
2543 E T6V, T70, T78, T79;
|
cannam@167
|
2544 T6V = T6R - T6U;
|
cannam@167
|
2545 T70 = T6W + T6Z;
|
cannam@167
|
2546 T71 = KP707106781 * (T6V + T70);
|
cannam@167
|
2547 TaQ = KP707106781 * (T70 - T6V);
|
cannam@167
|
2548 T78 = T6R + T6U;
|
cannam@167
|
2549 T79 = T6Z - T6W;
|
cannam@167
|
2550 T7a = KP707106781 * (T78 + T79);
|
cannam@167
|
2551 TaN = KP707106781 * (T78 - T79);
|
cannam@167
|
2552 }
|
cannam@167
|
2553 {
|
cannam@167
|
2554 E Td2, Td5, T75, T76;
|
cannam@167
|
2555 Td2 = T1E - T1P;
|
cannam@167
|
2556 Td5 = Td3 - Td4;
|
cannam@167
|
2557 Td6 = Td2 - Td5;
|
cannam@167
|
2558 TeJ = Td2 + Td5;
|
cannam@167
|
2559 T75 = T73 - T74;
|
cannam@167
|
2560 T76 = T1J - T1O;
|
cannam@167
|
2561 T77 = T75 + T76;
|
cannam@167
|
2562 TaP = T75 - T76;
|
cannam@167
|
2563 }
|
cannam@167
|
2564 }
|
cannam@167
|
2565 {
|
cannam@167
|
2566 E T2j, T7u, T2o, T7v, T2p, Tdd, T2u, T7e, T2z, T7f, T2A, Tde, T2M, Tdk, T7i;
|
cannam@167
|
2567 E T7l, T2X, Tdj, T7n, T7q;
|
cannam@167
|
2568 {
|
cannam@167
|
2569 E T2g, T2i, T2f, T2h;
|
cannam@167
|
2570 T2g = cr[WS(rs, 62)];
|
cannam@167
|
2571 T2i = ci[WS(rs, 62)];
|
cannam@167
|
2572 T2f = W[122];
|
cannam@167
|
2573 T2h = W[123];
|
cannam@167
|
2574 T2j = FMA(T2f, T2g, T2h * T2i);
|
cannam@167
|
2575 T7u = FNMS(T2h, T2g, T2f * T2i);
|
cannam@167
|
2576 }
|
cannam@167
|
2577 {
|
cannam@167
|
2578 E T2l, T2n, T2k, T2m;
|
cannam@167
|
2579 T2l = cr[WS(rs, 30)];
|
cannam@167
|
2580 T2n = ci[WS(rs, 30)];
|
cannam@167
|
2581 T2k = W[58];
|
cannam@167
|
2582 T2m = W[59];
|
cannam@167
|
2583 T2o = FMA(T2k, T2l, T2m * T2n);
|
cannam@167
|
2584 T7v = FNMS(T2m, T2l, T2k * T2n);
|
cannam@167
|
2585 }
|
cannam@167
|
2586 T2p = T2j + T2o;
|
cannam@167
|
2587 Tdd = T7u + T7v;
|
cannam@167
|
2588 {
|
cannam@167
|
2589 E T2r, T2t, T2q, T2s;
|
cannam@167
|
2590 T2r = cr[WS(rs, 14)];
|
cannam@167
|
2591 T2t = ci[WS(rs, 14)];
|
cannam@167
|
2592 T2q = W[26];
|
cannam@167
|
2593 T2s = W[27];
|
cannam@167
|
2594 T2u = FMA(T2q, T2r, T2s * T2t);
|
cannam@167
|
2595 T7e = FNMS(T2s, T2r, T2q * T2t);
|
cannam@167
|
2596 }
|
cannam@167
|
2597 {
|
cannam@167
|
2598 E T2w, T2y, T2v, T2x;
|
cannam@167
|
2599 T2w = cr[WS(rs, 46)];
|
cannam@167
|
2600 T2y = ci[WS(rs, 46)];
|
cannam@167
|
2601 T2v = W[90];
|
cannam@167
|
2602 T2x = W[91];
|
cannam@167
|
2603 T2z = FMA(T2v, T2w, T2x * T2y);
|
cannam@167
|
2604 T7f = FNMS(T2x, T2w, T2v * T2y);
|
cannam@167
|
2605 }
|
cannam@167
|
2606 T2A = T2u + T2z;
|
cannam@167
|
2607 Tde = T7e + T7f;
|
cannam@167
|
2608 {
|
cannam@167
|
2609 E T2G, T7j, T2L, T7k;
|
cannam@167
|
2610 {
|
cannam@167
|
2611 E T2D, T2F, T2C, T2E;
|
cannam@167
|
2612 T2D = cr[WS(rs, 6)];
|
cannam@167
|
2613 T2F = ci[WS(rs, 6)];
|
cannam@167
|
2614 T2C = W[10];
|
cannam@167
|
2615 T2E = W[11];
|
cannam@167
|
2616 T2G = FMA(T2C, T2D, T2E * T2F);
|
cannam@167
|
2617 T7j = FNMS(T2E, T2D, T2C * T2F);
|
cannam@167
|
2618 }
|
cannam@167
|
2619 {
|
cannam@167
|
2620 E T2I, T2K, T2H, T2J;
|
cannam@167
|
2621 T2I = cr[WS(rs, 38)];
|
cannam@167
|
2622 T2K = ci[WS(rs, 38)];
|
cannam@167
|
2623 T2H = W[74];
|
cannam@167
|
2624 T2J = W[75];
|
cannam@167
|
2625 T2L = FMA(T2H, T2I, T2J * T2K);
|
cannam@167
|
2626 T7k = FNMS(T2J, T2I, T2H * T2K);
|
cannam@167
|
2627 }
|
cannam@167
|
2628 T2M = T2G + T2L;
|
cannam@167
|
2629 Tdk = T7j + T7k;
|
cannam@167
|
2630 T7i = T2G - T2L;
|
cannam@167
|
2631 T7l = T7j - T7k;
|
cannam@167
|
2632 }
|
cannam@167
|
2633 {
|
cannam@167
|
2634 E T2R, T7o, T2W, T7p;
|
cannam@167
|
2635 {
|
cannam@167
|
2636 E T2O, T2Q, T2N, T2P;
|
cannam@167
|
2637 T2O = cr[WS(rs, 54)];
|
cannam@167
|
2638 T2Q = ci[WS(rs, 54)];
|
cannam@167
|
2639 T2N = W[106];
|
cannam@167
|
2640 T2P = W[107];
|
cannam@167
|
2641 T2R = FMA(T2N, T2O, T2P * T2Q);
|
cannam@167
|
2642 T7o = FNMS(T2P, T2O, T2N * T2Q);
|
cannam@167
|
2643 }
|
cannam@167
|
2644 {
|
cannam@167
|
2645 E T2T, T2V, T2S, T2U;
|
cannam@167
|
2646 T2T = cr[WS(rs, 22)];
|
cannam@167
|
2647 T2V = ci[WS(rs, 22)];
|
cannam@167
|
2648 T2S = W[42];
|
cannam@167
|
2649 T2U = W[43];
|
cannam@167
|
2650 T2W = FMA(T2S, T2T, T2U * T2V);
|
cannam@167
|
2651 T7p = FNMS(T2U, T2T, T2S * T2V);
|
cannam@167
|
2652 }
|
cannam@167
|
2653 T2X = T2R + T2W;
|
cannam@167
|
2654 Tdj = T7o + T7p;
|
cannam@167
|
2655 T7n = T2R - T2W;
|
cannam@167
|
2656 T7q = T7o - T7p;
|
cannam@167
|
2657 }
|
cannam@167
|
2658 T2B = T2p + T2A;
|
cannam@167
|
2659 T2Y = T2M + T2X;
|
cannam@167
|
2660 Tfz = T2B - T2Y;
|
cannam@167
|
2661 TfA = Tdd + Tde;
|
cannam@167
|
2662 TfB = Tdk + Tdj;
|
cannam@167
|
2663 TfC = TfA - TfB;
|
cannam@167
|
2664 {
|
cannam@167
|
2665 E T7d, T7g, Tdi, Tdl;
|
cannam@167
|
2666 T7d = T2j - T2o;
|
cannam@167
|
2667 T7g = T7e - T7f;
|
cannam@167
|
2668 T7h = T7d - T7g;
|
cannam@167
|
2669 TaW = T7d + T7g;
|
cannam@167
|
2670 Tdi = T2p - T2A;
|
cannam@167
|
2671 Tdl = Tdj - Tdk;
|
cannam@167
|
2672 Tdm = Tdi - Tdl;
|
cannam@167
|
2673 TeL = Tdi + Tdl;
|
cannam@167
|
2674 }
|
cannam@167
|
2675 {
|
cannam@167
|
2676 E T7m, T7r, T7z, T7A;
|
cannam@167
|
2677 T7m = T7i - T7l;
|
cannam@167
|
2678 T7r = T7n + T7q;
|
cannam@167
|
2679 T7s = KP707106781 * (T7m + T7r);
|
cannam@167
|
2680 TaU = KP707106781 * (T7r - T7m);
|
cannam@167
|
2681 T7z = T7i + T7l;
|
cannam@167
|
2682 T7A = T7q - T7n;
|
cannam@167
|
2683 T7B = KP707106781 * (T7z + T7A);
|
cannam@167
|
2684 TaX = KP707106781 * (T7z - T7A);
|
cannam@167
|
2685 }
|
cannam@167
|
2686 {
|
cannam@167
|
2687 E Tdf, Tdg, T7w, T7x;
|
cannam@167
|
2688 Tdf = Tdd - Tde;
|
cannam@167
|
2689 Tdg = T2M - T2X;
|
cannam@167
|
2690 Tdh = Tdf - Tdg;
|
cannam@167
|
2691 TeM = Tdf + Tdg;
|
cannam@167
|
2692 T7w = T7u - T7v;
|
cannam@167
|
2693 T7x = T2u - T2z;
|
cannam@167
|
2694 T7y = T7w + T7x;
|
cannam@167
|
2695 TaT = T7w - T7x;
|
cannam@167
|
2696 }
|
cannam@167
|
2697 }
|
cannam@167
|
2698 {
|
cannam@167
|
2699 E T4D, T9e, T4I, T9f, T4J, TdR, T4O, T8A, T4T, T8B, T4U, TdS, T56, Tea, T8E;
|
cannam@167
|
2700 E T8H, T5h, Te9, T8J, T8M;
|
cannam@167
|
2701 {
|
cannam@167
|
2702 E T4A, T4C, T4z, T4B;
|
cannam@167
|
2703 T4A = cr[WS(rs, 63)];
|
cannam@167
|
2704 T4C = ci[WS(rs, 63)];
|
cannam@167
|
2705 T4z = W[124];
|
cannam@167
|
2706 T4B = W[125];
|
cannam@167
|
2707 T4D = FMA(T4z, T4A, T4B * T4C);
|
cannam@167
|
2708 T9e = FNMS(T4B, T4A, T4z * T4C);
|
cannam@167
|
2709 }
|
cannam@167
|
2710 {
|
cannam@167
|
2711 E T4F, T4H, T4E, T4G;
|
cannam@167
|
2712 T4F = cr[WS(rs, 31)];
|
cannam@167
|
2713 T4H = ci[WS(rs, 31)];
|
cannam@167
|
2714 T4E = W[60];
|
cannam@167
|
2715 T4G = W[61];
|
cannam@167
|
2716 T4I = FMA(T4E, T4F, T4G * T4H);
|
cannam@167
|
2717 T9f = FNMS(T4G, T4F, T4E * T4H);
|
cannam@167
|
2718 }
|
cannam@167
|
2719 T4J = T4D + T4I;
|
cannam@167
|
2720 TdR = T9e + T9f;
|
cannam@167
|
2721 {
|
cannam@167
|
2722 E T4L, T4N, T4K, T4M;
|
cannam@167
|
2723 T4L = cr[WS(rs, 15)];
|
cannam@167
|
2724 T4N = ci[WS(rs, 15)];
|
cannam@167
|
2725 T4K = W[28];
|
cannam@167
|
2726 T4M = W[29];
|
cannam@167
|
2727 T4O = FMA(T4K, T4L, T4M * T4N);
|
cannam@167
|
2728 T8A = FNMS(T4M, T4L, T4K * T4N);
|
cannam@167
|
2729 }
|
cannam@167
|
2730 {
|
cannam@167
|
2731 E T4Q, T4S, T4P, T4R;
|
cannam@167
|
2732 T4Q = cr[WS(rs, 47)];
|
cannam@167
|
2733 T4S = ci[WS(rs, 47)];
|
cannam@167
|
2734 T4P = W[92];
|
cannam@167
|
2735 T4R = W[93];
|
cannam@167
|
2736 T4T = FMA(T4P, T4Q, T4R * T4S);
|
cannam@167
|
2737 T8B = FNMS(T4R, T4Q, T4P * T4S);
|
cannam@167
|
2738 }
|
cannam@167
|
2739 T4U = T4O + T4T;
|
cannam@167
|
2740 TdS = T8A + T8B;
|
cannam@167
|
2741 {
|
cannam@167
|
2742 E T50, T8F, T55, T8G;
|
cannam@167
|
2743 {
|
cannam@167
|
2744 E T4X, T4Z, T4W, T4Y;
|
cannam@167
|
2745 T4X = cr[WS(rs, 7)];
|
cannam@167
|
2746 T4Z = ci[WS(rs, 7)];
|
cannam@167
|
2747 T4W = W[12];
|
cannam@167
|
2748 T4Y = W[13];
|
cannam@167
|
2749 T50 = FMA(T4W, T4X, T4Y * T4Z);
|
cannam@167
|
2750 T8F = FNMS(T4Y, T4X, T4W * T4Z);
|
cannam@167
|
2751 }
|
cannam@167
|
2752 {
|
cannam@167
|
2753 E T52, T54, T51, T53;
|
cannam@167
|
2754 T52 = cr[WS(rs, 39)];
|
cannam@167
|
2755 T54 = ci[WS(rs, 39)];
|
cannam@167
|
2756 T51 = W[76];
|
cannam@167
|
2757 T53 = W[77];
|
cannam@167
|
2758 T55 = FMA(T51, T52, T53 * T54);
|
cannam@167
|
2759 T8G = FNMS(T53, T52, T51 * T54);
|
cannam@167
|
2760 }
|
cannam@167
|
2761 T56 = T50 + T55;
|
cannam@167
|
2762 Tea = T8F + T8G;
|
cannam@167
|
2763 T8E = T50 - T55;
|
cannam@167
|
2764 T8H = T8F - T8G;
|
cannam@167
|
2765 }
|
cannam@167
|
2766 {
|
cannam@167
|
2767 E T5b, T8K, T5g, T8L;
|
cannam@167
|
2768 {
|
cannam@167
|
2769 E T58, T5a, T57, T59;
|
cannam@167
|
2770 T58 = cr[WS(rs, 55)];
|
cannam@167
|
2771 T5a = ci[WS(rs, 55)];
|
cannam@167
|
2772 T57 = W[108];
|
cannam@167
|
2773 T59 = W[109];
|
cannam@167
|
2774 T5b = FMA(T57, T58, T59 * T5a);
|
cannam@167
|
2775 T8K = FNMS(T59, T58, T57 * T5a);
|
cannam@167
|
2776 }
|
cannam@167
|
2777 {
|
cannam@167
|
2778 E T5d, T5f, T5c, T5e;
|
cannam@167
|
2779 T5d = cr[WS(rs, 23)];
|
cannam@167
|
2780 T5f = ci[WS(rs, 23)];
|
cannam@167
|
2781 T5c = W[44];
|
cannam@167
|
2782 T5e = W[45];
|
cannam@167
|
2783 T5g = FMA(T5c, T5d, T5e * T5f);
|
cannam@167
|
2784 T8L = FNMS(T5e, T5d, T5c * T5f);
|
cannam@167
|
2785 }
|
cannam@167
|
2786 T5h = T5b + T5g;
|
cannam@167
|
2787 Te9 = T8K + T8L;
|
cannam@167
|
2788 T8J = T5b - T5g;
|
cannam@167
|
2789 T8M = T8K - T8L;
|
cannam@167
|
2790 }
|
cannam@167
|
2791 {
|
cannam@167
|
2792 E T4V, T5i, Te8, Teb;
|
cannam@167
|
2793 T4V = T4J + T4U;
|
cannam@167
|
2794 T5i = T56 + T5h;
|
cannam@167
|
2795 T5j = T4V + T5i;
|
cannam@167
|
2796 TfR = T4V - T5i;
|
cannam@167
|
2797 Te8 = T4J - T4U;
|
cannam@167
|
2798 Teb = Te9 - Tea;
|
cannam@167
|
2799 Tec = Te8 - Teb;
|
cannam@167
|
2800 TeX = Te8 + Teb;
|
cannam@167
|
2801 }
|
cannam@167
|
2802 {
|
cannam@167
|
2803 E TfW, TfX, T8z, T8C;
|
cannam@167
|
2804 TfW = TdR + TdS;
|
cannam@167
|
2805 TfX = Tea + Te9;
|
cannam@167
|
2806 TfY = TfW - TfX;
|
cannam@167
|
2807 Tgy = TfW + TfX;
|
cannam@167
|
2808 T8z = T4D - T4I;
|
cannam@167
|
2809 T8C = T8A - T8B;
|
cannam@167
|
2810 T8D = T8z - T8C;
|
cannam@167
|
2811 Tbl = T8z + T8C;
|
cannam@167
|
2812 }
|
cannam@167
|
2813 {
|
cannam@167
|
2814 E T8I, T8N, T9j, T9k;
|
cannam@167
|
2815 T8I = T8E - T8H;
|
cannam@167
|
2816 T8N = T8J + T8M;
|
cannam@167
|
2817 T8O = KP707106781 * (T8I + T8N);
|
cannam@167
|
2818 Tbx = KP707106781 * (T8N - T8I);
|
cannam@167
|
2819 T9j = T8E + T8H;
|
cannam@167
|
2820 T9k = T8M - T8J;
|
cannam@167
|
2821 T9l = KP707106781 * (T9j + T9k);
|
cannam@167
|
2822 Tbm = KP707106781 * (T9j - T9k);
|
cannam@167
|
2823 }
|
cannam@167
|
2824 {
|
cannam@167
|
2825 E TdT, TdU, T9g, T9h;
|
cannam@167
|
2826 TdT = TdR - TdS;
|
cannam@167
|
2827 TdU = T56 - T5h;
|
cannam@167
|
2828 TdV = TdT - TdU;
|
cannam@167
|
2829 Tf0 = TdT + TdU;
|
cannam@167
|
2830 T9g = T9e - T9f;
|
cannam@167
|
2831 T9h = T4O - T4T;
|
cannam@167
|
2832 T9i = T9g + T9h;
|
cannam@167
|
2833 Tbw = T9g - T9h;
|
cannam@167
|
2834 }
|
cannam@167
|
2835 }
|
cannam@167
|
2836 {
|
cannam@167
|
2837 E T36, T7G, T3b, T7H, T3c, TdH, T3h, T8m, T3m, T8n, T3n, TdI, T3z, Tds, T7L;
|
cannam@167
|
2838 E T7O, T3K, Tdr, T7S, T7T;
|
cannam@167
|
2839 {
|
cannam@167
|
2840 E T33, T35, T32, T34;
|
cannam@167
|
2841 T33 = cr[WS(rs, 1)];
|
cannam@167
|
2842 T35 = ci[WS(rs, 1)];
|
cannam@167
|
2843 T32 = W[0];
|
cannam@167
|
2844 T34 = W[1];
|
cannam@167
|
2845 T36 = FMA(T32, T33, T34 * T35);
|
cannam@167
|
2846 T7G = FNMS(T34, T33, T32 * T35);
|
cannam@167
|
2847 }
|
cannam@167
|
2848 {
|
cannam@167
|
2849 E T38, T3a, T37, T39;
|
cannam@167
|
2850 T38 = cr[WS(rs, 33)];
|
cannam@167
|
2851 T3a = ci[WS(rs, 33)];
|
cannam@167
|
2852 T37 = W[64];
|
cannam@167
|
2853 T39 = W[65];
|
cannam@167
|
2854 T3b = FMA(T37, T38, T39 * T3a);
|
cannam@167
|
2855 T7H = FNMS(T39, T38, T37 * T3a);
|
cannam@167
|
2856 }
|
cannam@167
|
2857 T3c = T36 + T3b;
|
cannam@167
|
2858 TdH = T7G + T7H;
|
cannam@167
|
2859 {
|
cannam@167
|
2860 E T3e, T3g, T3d, T3f;
|
cannam@167
|
2861 T3e = cr[WS(rs, 17)];
|
cannam@167
|
2862 T3g = ci[WS(rs, 17)];
|
cannam@167
|
2863 T3d = W[32];
|
cannam@167
|
2864 T3f = W[33];
|
cannam@167
|
2865 T3h = FMA(T3d, T3e, T3f * T3g);
|
cannam@167
|
2866 T8m = FNMS(T3f, T3e, T3d * T3g);
|
cannam@167
|
2867 }
|
cannam@167
|
2868 {
|
cannam@167
|
2869 E T3j, T3l, T3i, T3k;
|
cannam@167
|
2870 T3j = cr[WS(rs, 49)];
|
cannam@167
|
2871 T3l = ci[WS(rs, 49)];
|
cannam@167
|
2872 T3i = W[96];
|
cannam@167
|
2873 T3k = W[97];
|
cannam@167
|
2874 T3m = FMA(T3i, T3j, T3k * T3l);
|
cannam@167
|
2875 T8n = FNMS(T3k, T3j, T3i * T3l);
|
cannam@167
|
2876 }
|
cannam@167
|
2877 T3n = T3h + T3m;
|
cannam@167
|
2878 TdI = T8m + T8n;
|
cannam@167
|
2879 {
|
cannam@167
|
2880 E T3t, T7M, T3y, T7N;
|
cannam@167
|
2881 {
|
cannam@167
|
2882 E T3q, T3s, T3p, T3r;
|
cannam@167
|
2883 T3q = cr[WS(rs, 9)];
|
cannam@167
|
2884 T3s = ci[WS(rs, 9)];
|
cannam@167
|
2885 T3p = W[16];
|
cannam@167
|
2886 T3r = W[17];
|
cannam@167
|
2887 T3t = FMA(T3p, T3q, T3r * T3s);
|
cannam@167
|
2888 T7M = FNMS(T3r, T3q, T3p * T3s);
|
cannam@167
|
2889 }
|
cannam@167
|
2890 {
|
cannam@167
|
2891 E T3v, T3x, T3u, T3w;
|
cannam@167
|
2892 T3v = cr[WS(rs, 41)];
|
cannam@167
|
2893 T3x = ci[WS(rs, 41)];
|
cannam@167
|
2894 T3u = W[80];
|
cannam@167
|
2895 T3w = W[81];
|
cannam@167
|
2896 T3y = FMA(T3u, T3v, T3w * T3x);
|
cannam@167
|
2897 T7N = FNMS(T3w, T3v, T3u * T3x);
|
cannam@167
|
2898 }
|
cannam@167
|
2899 T3z = T3t + T3y;
|
cannam@167
|
2900 Tds = T7M + T7N;
|
cannam@167
|
2901 T7L = T3t - T3y;
|
cannam@167
|
2902 T7O = T7M - T7N;
|
cannam@167
|
2903 }
|
cannam@167
|
2904 {
|
cannam@167
|
2905 E T3E, T7Q, T3J, T7R;
|
cannam@167
|
2906 {
|
cannam@167
|
2907 E T3B, T3D, T3A, T3C;
|
cannam@167
|
2908 T3B = cr[WS(rs, 57)];
|
cannam@167
|
2909 T3D = ci[WS(rs, 57)];
|
cannam@167
|
2910 T3A = W[112];
|
cannam@167
|
2911 T3C = W[113];
|
cannam@167
|
2912 T3E = FMA(T3A, T3B, T3C * T3D);
|
cannam@167
|
2913 T7Q = FNMS(T3C, T3B, T3A * T3D);
|
cannam@167
|
2914 }
|
cannam@167
|
2915 {
|
cannam@167
|
2916 E T3G, T3I, T3F, T3H;
|
cannam@167
|
2917 T3G = cr[WS(rs, 25)];
|
cannam@167
|
2918 T3I = ci[WS(rs, 25)];
|
cannam@167
|
2919 T3F = W[48];
|
cannam@167
|
2920 T3H = W[49];
|
cannam@167
|
2921 T3J = FMA(T3F, T3G, T3H * T3I);
|
cannam@167
|
2922 T7R = FNMS(T3H, T3G, T3F * T3I);
|
cannam@167
|
2923 }
|
cannam@167
|
2924 T3K = T3E + T3J;
|
cannam@167
|
2925 Tdr = T7Q + T7R;
|
cannam@167
|
2926 T7S = T7Q - T7R;
|
cannam@167
|
2927 T7T = T3E - T3J;
|
cannam@167
|
2928 }
|
cannam@167
|
2929 {
|
cannam@167
|
2930 E T3o, T3L, TdJ, TdK;
|
cannam@167
|
2931 T3o = T3c + T3n;
|
cannam@167
|
2932 T3L = T3z + T3K;
|
cannam@167
|
2933 T3M = T3o + T3L;
|
cannam@167
|
2934 TfL = T3o - T3L;
|
cannam@167
|
2935 TdJ = TdH - TdI;
|
cannam@167
|
2936 TdK = T3z - T3K;
|
cannam@167
|
2937 TdL = TdJ - TdK;
|
cannam@167
|
2938 TeT = TdJ + TdK;
|
cannam@167
|
2939 }
|
cannam@167
|
2940 {
|
cannam@167
|
2941 E TfG, TfH, T7I, T7J;
|
cannam@167
|
2942 TfG = TdH + TdI;
|
cannam@167
|
2943 TfH = Tds + Tdr;
|
cannam@167
|
2944 TfI = TfG - TfH;
|
cannam@167
|
2945 Tgt = TfG + TfH;
|
cannam@167
|
2946 T7I = T7G - T7H;
|
cannam@167
|
2947 T7J = T3h - T3m;
|
cannam@167
|
2948 T7K = T7I + T7J;
|
cannam@167
|
2949 Tbd = T7I - T7J;
|
cannam@167
|
2950 }
|
cannam@167
|
2951 {
|
cannam@167
|
2952 E T7P, T7U, T8q, T8r;
|
cannam@167
|
2953 T7P = T7L + T7O;
|
cannam@167
|
2954 T7U = T7S - T7T;
|
cannam@167
|
2955 T7V = KP707106781 * (T7P + T7U);
|
cannam@167
|
2956 Tb3 = KP707106781 * (T7P - T7U);
|
cannam@167
|
2957 T8q = T7L - T7O;
|
cannam@167
|
2958 T8r = T7T + T7S;
|
cannam@167
|
2959 T8s = KP707106781 * (T8q + T8r);
|
cannam@167
|
2960 Tbe = KP707106781 * (T8r - T8q);
|
cannam@167
|
2961 }
|
cannam@167
|
2962 {
|
cannam@167
|
2963 E Tdq, Tdt, T8l, T8o;
|
cannam@167
|
2964 Tdq = T3c - T3n;
|
cannam@167
|
2965 Tdt = Tdr - Tds;
|
cannam@167
|
2966 Tdu = Tdq - Tdt;
|
cannam@167
|
2967 TeQ = Tdq + Tdt;
|
cannam@167
|
2968 T8l = T36 - T3b;
|
cannam@167
|
2969 T8o = T8m - T8n;
|
cannam@167
|
2970 T8p = T8l - T8o;
|
cannam@167
|
2971 Tb2 = T8l + T8o;
|
cannam@167
|
2972 }
|
cannam@167
|
2973 }
|
cannam@167
|
2974 {
|
cannam@167
|
2975 E T3X, Tdw, T7Z, T82, T4v, TdB, T8b, T8g, T48, Tdx, T80, T85, T4k, TdA, T8a;
|
cannam@167
|
2976 E T8d;
|
cannam@167
|
2977 {
|
cannam@167
|
2978 E T3R, T7X, T3W, T7Y;
|
cannam@167
|
2979 {
|
cannam@167
|
2980 E T3O, T3Q, T3N, T3P;
|
cannam@167
|
2981 T3O = cr[WS(rs, 5)];
|
cannam@167
|
2982 T3Q = ci[WS(rs, 5)];
|
cannam@167
|
2983 T3N = W[8];
|
cannam@167
|
2984 T3P = W[9];
|
cannam@167
|
2985 T3R = FMA(T3N, T3O, T3P * T3Q);
|
cannam@167
|
2986 T7X = FNMS(T3P, T3O, T3N * T3Q);
|
cannam@167
|
2987 }
|
cannam@167
|
2988 {
|
cannam@167
|
2989 E T3T, T3V, T3S, T3U;
|
cannam@167
|
2990 T3T = cr[WS(rs, 37)];
|
cannam@167
|
2991 T3V = ci[WS(rs, 37)];
|
cannam@167
|
2992 T3S = W[72];
|
cannam@167
|
2993 T3U = W[73];
|
cannam@167
|
2994 T3W = FMA(T3S, T3T, T3U * T3V);
|
cannam@167
|
2995 T7Y = FNMS(T3U, T3T, T3S * T3V);
|
cannam@167
|
2996 }
|
cannam@167
|
2997 T3X = T3R + T3W;
|
cannam@167
|
2998 Tdw = T7X + T7Y;
|
cannam@167
|
2999 T7Z = T7X - T7Y;
|
cannam@167
|
3000 T82 = T3R - T3W;
|
cannam@167
|
3001 }
|
cannam@167
|
3002 {
|
cannam@167
|
3003 E T4p, T8e, T4u, T8f;
|
cannam@167
|
3004 {
|
cannam@167
|
3005 E T4m, T4o, T4l, T4n;
|
cannam@167
|
3006 T4m = cr[WS(rs, 13)];
|
cannam@167
|
3007 T4o = ci[WS(rs, 13)];
|
cannam@167
|
3008 T4l = W[24];
|
cannam@167
|
3009 T4n = W[25];
|
cannam@167
|
3010 T4p = FMA(T4l, T4m, T4n * T4o);
|
cannam@167
|
3011 T8e = FNMS(T4n, T4m, T4l * T4o);
|
cannam@167
|
3012 }
|
cannam@167
|
3013 {
|
cannam@167
|
3014 E T4r, T4t, T4q, T4s;
|
cannam@167
|
3015 T4r = cr[WS(rs, 45)];
|
cannam@167
|
3016 T4t = ci[WS(rs, 45)];
|
cannam@167
|
3017 T4q = W[88];
|
cannam@167
|
3018 T4s = W[89];
|
cannam@167
|
3019 T4u = FMA(T4q, T4r, T4s * T4t);
|
cannam@167
|
3020 T8f = FNMS(T4s, T4r, T4q * T4t);
|
cannam@167
|
3021 }
|
cannam@167
|
3022 T4v = T4p + T4u;
|
cannam@167
|
3023 TdB = T8e + T8f;
|
cannam@167
|
3024 T8b = T4p - T4u;
|
cannam@167
|
3025 T8g = T8e - T8f;
|
cannam@167
|
3026 }
|
cannam@167
|
3027 {
|
cannam@167
|
3028 E T42, T83, T47, T84;
|
cannam@167
|
3029 {
|
cannam@167
|
3030 E T3Z, T41, T3Y, T40;
|
cannam@167
|
3031 T3Z = cr[WS(rs, 21)];
|
cannam@167
|
3032 T41 = ci[WS(rs, 21)];
|
cannam@167
|
3033 T3Y = W[40];
|
cannam@167
|
3034 T40 = W[41];
|
cannam@167
|
3035 T42 = FMA(T3Y, T3Z, T40 * T41);
|
cannam@167
|
3036 T83 = FNMS(T40, T3Z, T3Y * T41);
|
cannam@167
|
3037 }
|
cannam@167
|
3038 {
|
cannam@167
|
3039 E T44, T46, T43, T45;
|
cannam@167
|
3040 T44 = cr[WS(rs, 53)];
|
cannam@167
|
3041 T46 = ci[WS(rs, 53)];
|
cannam@167
|
3042 T43 = W[104];
|
cannam@167
|
3043 T45 = W[105];
|
cannam@167
|
3044 T47 = FMA(T43, T44, T45 * T46);
|
cannam@167
|
3045 T84 = FNMS(T45, T44, T43 * T46);
|
cannam@167
|
3046 }
|
cannam@167
|
3047 T48 = T42 + T47;
|
cannam@167
|
3048 Tdx = T83 + T84;
|
cannam@167
|
3049 T80 = T42 - T47;
|
cannam@167
|
3050 T85 = T83 - T84;
|
cannam@167
|
3051 }
|
cannam@167
|
3052 {
|
cannam@167
|
3053 E T4e, T88, T4j, T89;
|
cannam@167
|
3054 {
|
cannam@167
|
3055 E T4b, T4d, T4a, T4c;
|
cannam@167
|
3056 T4b = cr[WS(rs, 61)];
|
cannam@167
|
3057 T4d = ci[WS(rs, 61)];
|
cannam@167
|
3058 T4a = W[120];
|
cannam@167
|
3059 T4c = W[121];
|
cannam@167
|
3060 T4e = FMA(T4a, T4b, T4c * T4d);
|
cannam@167
|
3061 T88 = FNMS(T4c, T4b, T4a * T4d);
|
cannam@167
|
3062 }
|
cannam@167
|
3063 {
|
cannam@167
|
3064 E T4g, T4i, T4f, T4h;
|
cannam@167
|
3065 T4g = cr[WS(rs, 29)];
|
cannam@167
|
3066 T4i = ci[WS(rs, 29)];
|
cannam@167
|
3067 T4f = W[56];
|
cannam@167
|
3068 T4h = W[57];
|
cannam@167
|
3069 T4j = FMA(T4f, T4g, T4h * T4i);
|
cannam@167
|
3070 T89 = FNMS(T4h, T4g, T4f * T4i);
|
cannam@167
|
3071 }
|
cannam@167
|
3072 T4k = T4e + T4j;
|
cannam@167
|
3073 TdA = T88 + T89;
|
cannam@167
|
3074 T8a = T88 - T89;
|
cannam@167
|
3075 T8d = T4e - T4j;
|
cannam@167
|
3076 }
|
cannam@167
|
3077 {
|
cannam@167
|
3078 E T49, T4w, TdC, TdD;
|
cannam@167
|
3079 T49 = T3X + T48;
|
cannam@167
|
3080 T4w = T4k + T4v;
|
cannam@167
|
3081 T4x = T49 + T4w;
|
cannam@167
|
3082 TfJ = T49 - T4w;
|
cannam@167
|
3083 TdC = TdA - TdB;
|
cannam@167
|
3084 TdD = T4k - T4v;
|
cannam@167
|
3085 TdE = TdC - TdD;
|
cannam@167
|
3086 TdM = TdD + TdC;
|
cannam@167
|
3087 }
|
cannam@167
|
3088 {
|
cannam@167
|
3089 E TfM, TfN, T81, T86;
|
cannam@167
|
3090 TfM = TdA + TdB;
|
cannam@167
|
3091 TfN = Tdw + Tdx;
|
cannam@167
|
3092 TfO = TfM - TfN;
|
cannam@167
|
3093 Tgu = TfN + TfM;
|
cannam@167
|
3094 T81 = T7Z + T80;
|
cannam@167
|
3095 T86 = T82 - T85;
|
cannam@167
|
3096 T87 = FMA(KP923879532, T81, KP382683432 * T86);
|
cannam@167
|
3097 T8u = FNMS(KP382683432, T81, KP923879532 * T86);
|
cannam@167
|
3098 }
|
cannam@167
|
3099 {
|
cannam@167
|
3100 E T8c, T8h, Tb8, Tb9;
|
cannam@167
|
3101 T8c = T8a + T8b;
|
cannam@167
|
3102 T8h = T8d - T8g;
|
cannam@167
|
3103 T8i = FNMS(KP382683432, T8h, KP923879532 * T8c);
|
cannam@167
|
3104 T8v = FMA(KP382683432, T8c, KP923879532 * T8h);
|
cannam@167
|
3105 Tb8 = T8d + T8g;
|
cannam@167
|
3106 Tb9 = T8a - T8b;
|
cannam@167
|
3107 Tba = FNMS(KP382683432, Tb9, KP923879532 * Tb8);
|
cannam@167
|
3108 Tbh = FMA(KP923879532, Tb9, KP382683432 * Tb8);
|
cannam@167
|
3109 }
|
cannam@167
|
3110 {
|
cannam@167
|
3111 E Tdv, Tdy, Tb5, Tb6;
|
cannam@167
|
3112 Tdv = T3X - T48;
|
cannam@167
|
3113 Tdy = Tdw - Tdx;
|
cannam@167
|
3114 Tdz = Tdv + Tdy;
|
cannam@167
|
3115 TdN = Tdv - Tdy;
|
cannam@167
|
3116 Tb5 = T7Z - T80;
|
cannam@167
|
3117 Tb6 = T82 + T85;
|
cannam@167
|
3118 Tb7 = FMA(KP382683432, Tb5, KP923879532 * Tb6);
|
cannam@167
|
3119 Tbg = FNMS(KP382683432, Tb6, KP923879532 * Tb5);
|
cannam@167
|
3120 }
|
cannam@167
|
3121 }
|
cannam@167
|
3122 {
|
cannam@167
|
3123 E T5u, Te2, T8Q, T8X, T62, TdY, T94, T99, T5F, Te3, T8T, T8Y, T5R, TdX, T93;
|
cannam@167
|
3124 E T96;
|
cannam@167
|
3125 {
|
cannam@167
|
3126 E T5o, T8V, T5t, T8W;
|
cannam@167
|
3127 {
|
cannam@167
|
3128 E T5l, T5n, T5k, T5m;
|
cannam@167
|
3129 T5l = cr[WS(rs, 3)];
|
cannam@167
|
3130 T5n = ci[WS(rs, 3)];
|
cannam@167
|
3131 T5k = W[4];
|
cannam@167
|
3132 T5m = W[5];
|
cannam@167
|
3133 T5o = FMA(T5k, T5l, T5m * T5n);
|
cannam@167
|
3134 T8V = FNMS(T5m, T5l, T5k * T5n);
|
cannam@167
|
3135 }
|
cannam@167
|
3136 {
|
cannam@167
|
3137 E T5q, T5s, T5p, T5r;
|
cannam@167
|
3138 T5q = cr[WS(rs, 35)];
|
cannam@167
|
3139 T5s = ci[WS(rs, 35)];
|
cannam@167
|
3140 T5p = W[68];
|
cannam@167
|
3141 T5r = W[69];
|
cannam@167
|
3142 T5t = FMA(T5p, T5q, T5r * T5s);
|
cannam@167
|
3143 T8W = FNMS(T5r, T5q, T5p * T5s);
|
cannam@167
|
3144 }
|
cannam@167
|
3145 T5u = T5o + T5t;
|
cannam@167
|
3146 Te2 = T8V + T8W;
|
cannam@167
|
3147 T8Q = T5o - T5t;
|
cannam@167
|
3148 T8X = T8V - T8W;
|
cannam@167
|
3149 }
|
cannam@167
|
3150 {
|
cannam@167
|
3151 E T5W, T97, T61, T98;
|
cannam@167
|
3152 {
|
cannam@167
|
3153 E T5T, T5V, T5S, T5U;
|
cannam@167
|
3154 T5T = cr[WS(rs, 11)];
|
cannam@167
|
3155 T5V = ci[WS(rs, 11)];
|
cannam@167
|
3156 T5S = W[20];
|
cannam@167
|
3157 T5U = W[21];
|
cannam@167
|
3158 T5W = FMA(T5S, T5T, T5U * T5V);
|
cannam@167
|
3159 T97 = FNMS(T5U, T5T, T5S * T5V);
|
cannam@167
|
3160 }
|
cannam@167
|
3161 {
|
cannam@167
|
3162 E T5Y, T60, T5X, T5Z;
|
cannam@167
|
3163 T5Y = cr[WS(rs, 43)];
|
cannam@167
|
3164 T60 = ci[WS(rs, 43)];
|
cannam@167
|
3165 T5X = W[84];
|
cannam@167
|
3166 T5Z = W[85];
|
cannam@167
|
3167 T61 = FMA(T5X, T5Y, T5Z * T60);
|
cannam@167
|
3168 T98 = FNMS(T5Z, T5Y, T5X * T60);
|
cannam@167
|
3169 }
|
cannam@167
|
3170 T62 = T5W + T61;
|
cannam@167
|
3171 TdY = T97 + T98;
|
cannam@167
|
3172 T94 = T5W - T61;
|
cannam@167
|
3173 T99 = T97 - T98;
|
cannam@167
|
3174 }
|
cannam@167
|
3175 {
|
cannam@167
|
3176 E T5z, T8R, T5E, T8S;
|
cannam@167
|
3177 {
|
cannam@167
|
3178 E T5w, T5y, T5v, T5x;
|
cannam@167
|
3179 T5w = cr[WS(rs, 19)];
|
cannam@167
|
3180 T5y = ci[WS(rs, 19)];
|
cannam@167
|
3181 T5v = W[36];
|
cannam@167
|
3182 T5x = W[37];
|
cannam@167
|
3183 T5z = FMA(T5v, T5w, T5x * T5y);
|
cannam@167
|
3184 T8R = FNMS(T5x, T5w, T5v * T5y);
|
cannam@167
|
3185 }
|
cannam@167
|
3186 {
|
cannam@167
|
3187 E T5B, T5D, T5A, T5C;
|
cannam@167
|
3188 T5B = cr[WS(rs, 51)];
|
cannam@167
|
3189 T5D = ci[WS(rs, 51)];
|
cannam@167
|
3190 T5A = W[100];
|
cannam@167
|
3191 T5C = W[101];
|
cannam@167
|
3192 T5E = FMA(T5A, T5B, T5C * T5D);
|
cannam@167
|
3193 T8S = FNMS(T5C, T5B, T5A * T5D);
|
cannam@167
|
3194 }
|
cannam@167
|
3195 T5F = T5z + T5E;
|
cannam@167
|
3196 Te3 = T8R + T8S;
|
cannam@167
|
3197 T8T = T8R - T8S;
|
cannam@167
|
3198 T8Y = T5z - T5E;
|
cannam@167
|
3199 }
|
cannam@167
|
3200 {
|
cannam@167
|
3201 E T5L, T91, T5Q, T92;
|
cannam@167
|
3202 {
|
cannam@167
|
3203 E T5I, T5K, T5H, T5J;
|
cannam@167
|
3204 T5I = cr[WS(rs, 59)];
|
cannam@167
|
3205 T5K = ci[WS(rs, 59)];
|
cannam@167
|
3206 T5H = W[116];
|
cannam@167
|
3207 T5J = W[117];
|
cannam@167
|
3208 T5L = FMA(T5H, T5I, T5J * T5K);
|
cannam@167
|
3209 T91 = FNMS(T5J, T5I, T5H * T5K);
|
cannam@167
|
3210 }
|
cannam@167
|
3211 {
|
cannam@167
|
3212 E T5N, T5P, T5M, T5O;
|
cannam@167
|
3213 T5N = cr[WS(rs, 27)];
|
cannam@167
|
3214 T5P = ci[WS(rs, 27)];
|
cannam@167
|
3215 T5M = W[52];
|
cannam@167
|
3216 T5O = W[53];
|
cannam@167
|
3217 T5Q = FMA(T5M, T5N, T5O * T5P);
|
cannam@167
|
3218 T92 = FNMS(T5O, T5N, T5M * T5P);
|
cannam@167
|
3219 }
|
cannam@167
|
3220 T5R = T5L + T5Q;
|
cannam@167
|
3221 TdX = T91 + T92;
|
cannam@167
|
3222 T93 = T91 - T92;
|
cannam@167
|
3223 T96 = T5L - T5Q;
|
cannam@167
|
3224 }
|
cannam@167
|
3225 {
|
cannam@167
|
3226 E T5G, T63, Te1, Te4;
|
cannam@167
|
3227 T5G = T5u + T5F;
|
cannam@167
|
3228 T63 = T5R + T62;
|
cannam@167
|
3229 T64 = T5G + T63;
|
cannam@167
|
3230 TfZ = T5G - T63;
|
cannam@167
|
3231 Te1 = T5u - T5F;
|
cannam@167
|
3232 Te4 = Te2 - Te3;
|
cannam@167
|
3233 Te5 = Te1 - Te4;
|
cannam@167
|
3234 Ted = Te1 + Te4;
|
cannam@167
|
3235 }
|
cannam@167
|
3236 {
|
cannam@167
|
3237 E TfS, TfT, T8U, T8Z;
|
cannam@167
|
3238 TfS = TdX + TdY;
|
cannam@167
|
3239 TfT = Te2 + Te3;
|
cannam@167
|
3240 TfU = TfS - TfT;
|
cannam@167
|
3241 Tgz = TfT + TfS;
|
cannam@167
|
3242 T8U = T8Q - T8T;
|
cannam@167
|
3243 T8Z = T8X + T8Y;
|
cannam@167
|
3244 T90 = FNMS(KP382683432, T8Z, KP923879532 * T8U);
|
cannam@167
|
3245 T9n = FMA(KP923879532, T8Z, KP382683432 * T8U);
|
cannam@167
|
3246 }
|
cannam@167
|
3247 {
|
cannam@167
|
3248 E T95, T9a, Tbr, Tbs;
|
cannam@167
|
3249 T95 = T93 + T94;
|
cannam@167
|
3250 T9a = T96 - T99;
|
cannam@167
|
3251 T9b = FMA(KP382683432, T95, KP923879532 * T9a);
|
cannam@167
|
3252 T9o = FNMS(KP382683432, T9a, KP923879532 * T95);
|
cannam@167
|
3253 Tbr = T96 + T99;
|
cannam@167
|
3254 Tbs = T93 - T94;
|
cannam@167
|
3255 Tbt = FNMS(KP382683432, Tbs, KP923879532 * Tbr);
|
cannam@167
|
3256 TbA = FMA(KP923879532, Tbs, KP382683432 * Tbr);
|
cannam@167
|
3257 }
|
cannam@167
|
3258 {
|
cannam@167
|
3259 E TdW, TdZ, Tbo, Tbp;
|
cannam@167
|
3260 TdW = T5R - T62;
|
cannam@167
|
3261 TdZ = TdX - TdY;
|
cannam@167
|
3262 Te0 = TdW + TdZ;
|
cannam@167
|
3263 Tee = TdZ - TdW;
|
cannam@167
|
3264 Tbo = T8X - T8Y;
|
cannam@167
|
3265 Tbp = T8Q + T8T;
|
cannam@167
|
3266 Tbq = FMA(KP382683432, Tbo, KP923879532 * Tbp);
|
cannam@167
|
3267 Tbz = FNMS(KP382683432, Tbp, KP923879532 * Tbo);
|
cannam@167
|
3268 }
|
cannam@167
|
3269 }
|
cannam@167
|
3270 {
|
cannam@167
|
3271 E T1t, Tgn, TgK, TgL, TgV, Th1, T30, Th0, T66, TgX, Tgw, TgE, TgB, TgF, Tgq;
|
cannam@167
|
3272 E TgM;
|
cannam@167
|
3273 {
|
cannam@167
|
3274 E TH, T1s, TgI, TgJ;
|
cannam@167
|
3275 TH = Tj + TG;
|
cannam@167
|
3276 T1s = T14 + T1r;
|
cannam@167
|
3277 T1t = TH + T1s;
|
cannam@167
|
3278 Tgn = TH - T1s;
|
cannam@167
|
3279 TgI = Tgy + Tgz;
|
cannam@167
|
3280 TgJ = Tgt + Tgu;
|
cannam@167
|
3281 TgK = TgI - TgJ;
|
cannam@167
|
3282 TgL = TgJ + TgI;
|
cannam@167
|
3283 }
|
cannam@167
|
3284 {
|
cannam@167
|
3285 E TgN, TgU, T2e, T2Z;
|
cannam@167
|
3286 TgN = Tfr + Tfq;
|
cannam@167
|
3287 TgU = TgO + TgT;
|
cannam@167
|
3288 TgV = TgN + TgU;
|
cannam@167
|
3289 Th1 = TgU - TgN;
|
cannam@167
|
3290 T2e = T1Q + T2d;
|
cannam@167
|
3291 T2Z = T2B + T2Y;
|
cannam@167
|
3292 T30 = T2e + T2Z;
|
cannam@167
|
3293 Th0 = T2e - T2Z;
|
cannam@167
|
3294 }
|
cannam@167
|
3295 {
|
cannam@167
|
3296 E T4y, T65, Tgs, Tgv;
|
cannam@167
|
3297 T4y = T3M + T4x;
|
cannam@167
|
3298 T65 = T5j + T64;
|
cannam@167
|
3299 T66 = T4y + T65;
|
cannam@167
|
3300 TgX = T65 - T4y;
|
cannam@167
|
3301 Tgs = T3M - T4x;
|
cannam@167
|
3302 Tgv = Tgt - Tgu;
|
cannam@167
|
3303 Tgw = Tgs + Tgv;
|
cannam@167
|
3304 TgE = Tgs - Tgv;
|
cannam@167
|
3305 }
|
cannam@167
|
3306 {
|
cannam@167
|
3307 E Tgx, TgA, Tgo, Tgp;
|
cannam@167
|
3308 Tgx = T5j - T64;
|
cannam@167
|
3309 TgA = Tgy - Tgz;
|
cannam@167
|
3310 TgB = Tgx - TgA;
|
cannam@167
|
3311 TgF = Tgx + TgA;
|
cannam@167
|
3312 Tgo = TfA + TfB;
|
cannam@167
|
3313 Tgp = Tfv + Tfw;
|
cannam@167
|
3314 Tgq = Tgo - Tgp;
|
cannam@167
|
3315 TgM = Tgp + Tgo;
|
cannam@167
|
3316 }
|
cannam@167
|
3317 {
|
cannam@167
|
3318 E T31, TgW, TgY, TgH;
|
cannam@167
|
3319 T31 = T1t + T30;
|
cannam@167
|
3320 ci[WS(rs, 31)] = T31 - T66;
|
cannam@167
|
3321 cr[0] = T31 + T66;
|
cannam@167
|
3322 TgW = TgM + TgV;
|
cannam@167
|
3323 cr[WS(rs, 32)] = TgL - TgW;
|
cannam@167
|
3324 ci[WS(rs, 63)] = TgL + TgW;
|
cannam@167
|
3325 TgY = TgV - TgM;
|
cannam@167
|
3326 cr[WS(rs, 48)] = TgX - TgY;
|
cannam@167
|
3327 ci[WS(rs, 47)] = TgX + TgY;
|
cannam@167
|
3328 TgH = T1t - T30;
|
cannam@167
|
3329 cr[WS(rs, 16)] = TgH - TgK;
|
cannam@167
|
3330 ci[WS(rs, 15)] = TgH + TgK;
|
cannam@167
|
3331 }
|
cannam@167
|
3332 {
|
cannam@167
|
3333 E Tgr, TgC, TgZ, Th2;
|
cannam@167
|
3334 Tgr = Tgn - Tgq;
|
cannam@167
|
3335 TgC = KP707106781 * (Tgw + TgB);
|
cannam@167
|
3336 ci[WS(rs, 23)] = Tgr - TgC;
|
cannam@167
|
3337 cr[WS(rs, 8)] = Tgr + TgC;
|
cannam@167
|
3338 TgZ = KP707106781 * (TgB - Tgw);
|
cannam@167
|
3339 Th2 = Th0 + Th1;
|
cannam@167
|
3340 cr[WS(rs, 56)] = TgZ - Th2;
|
cannam@167
|
3341 ci[WS(rs, 39)] = TgZ + Th2;
|
cannam@167
|
3342 }
|
cannam@167
|
3343 {
|
cannam@167
|
3344 E Th3, Th4, TgD, TgG;
|
cannam@167
|
3345 Th3 = KP707106781 * (TgF - TgE);
|
cannam@167
|
3346 Th4 = Th1 - Th0;
|
cannam@167
|
3347 cr[WS(rs, 40)] = Th3 - Th4;
|
cannam@167
|
3348 ci[WS(rs, 55)] = Th3 + Th4;
|
cannam@167
|
3349 TgD = Tgn + Tgq;
|
cannam@167
|
3350 TgG = KP707106781 * (TgE + TgF);
|
cannam@167
|
3351 cr[WS(rs, 24)] = TgD - TgG;
|
cannam@167
|
3352 ci[WS(rs, 7)] = TgD + TgG;
|
cannam@167
|
3353 }
|
cannam@167
|
3354 }
|
cannam@167
|
3355 {
|
cannam@167
|
3356 E T6L, T9x, ThV, Ti1, T7E, Ti0, T9A, ThO, T8y, T9K, T9u, T9E, T9r, T9L, T9v;
|
cannam@167
|
3357 E T9H;
|
cannam@167
|
3358 {
|
cannam@167
|
3359 E T6n, T6K, ThP, ThU;
|
cannam@167
|
3360 T6n = T6b + T6m;
|
cannam@167
|
3361 T6K = T6y + T6J;
|
cannam@167
|
3362 T6L = T6n - T6K;
|
cannam@167
|
3363 T9x = T6n + T6K;
|
cannam@167
|
3364 ThP = T9O - T9P;
|
cannam@167
|
3365 ThU = ThQ + ThT;
|
cannam@167
|
3366 ThV = ThP + ThU;
|
cannam@167
|
3367 Ti1 = ThU - ThP;
|
cannam@167
|
3368 }
|
cannam@167
|
3369 {
|
cannam@167
|
3370 E T7c, T9y, T7D, T9z;
|
cannam@167
|
3371 {
|
cannam@167
|
3372 E T72, T7b, T7t, T7C;
|
cannam@167
|
3373 T72 = T6Q + T71;
|
cannam@167
|
3374 T7b = T77 + T7a;
|
cannam@167
|
3375 T7c = FMA(KP195090322, T72, KP980785280 * T7b);
|
cannam@167
|
3376 T9y = FNMS(KP195090322, T7b, KP980785280 * T72);
|
cannam@167
|
3377 T7t = T7h + T7s;
|
cannam@167
|
3378 T7C = T7y + T7B;
|
cannam@167
|
3379 T7D = FNMS(KP980785280, T7C, KP195090322 * T7t);
|
cannam@167
|
3380 T9z = FMA(KP980785280, T7t, KP195090322 * T7C);
|
cannam@167
|
3381 }
|
cannam@167
|
3382 T7E = T7c + T7D;
|
cannam@167
|
3383 Ti0 = T9z - T9y;
|
cannam@167
|
3384 T9A = T9y + T9z;
|
cannam@167
|
3385 ThO = T7c - T7D;
|
cannam@167
|
3386 }
|
cannam@167
|
3387 {
|
cannam@167
|
3388 E T8k, T9D, T8x, T9C;
|
cannam@167
|
3389 {
|
cannam@167
|
3390 E T7W, T8j, T8t, T8w;
|
cannam@167
|
3391 T7W = T7K + T7V;
|
cannam@167
|
3392 T8j = T87 + T8i;
|
cannam@167
|
3393 T8k = T7W - T8j;
|
cannam@167
|
3394 T9D = T7W + T8j;
|
cannam@167
|
3395 T8t = T8p + T8s;
|
cannam@167
|
3396 T8w = T8u + T8v;
|
cannam@167
|
3397 T8x = T8t - T8w;
|
cannam@167
|
3398 T9C = T8t + T8w;
|
cannam@167
|
3399 }
|
cannam@167
|
3400 T8y = FMA(KP634393284, T8k, KP773010453 * T8x);
|
cannam@167
|
3401 T9K = FMA(KP995184726, T9D, KP098017140 * T9C);
|
cannam@167
|
3402 T9u = FNMS(KP773010453, T8k, KP634393284 * T8x);
|
cannam@167
|
3403 T9E = FNMS(KP098017140, T9D, KP995184726 * T9C);
|
cannam@167
|
3404 }
|
cannam@167
|
3405 {
|
cannam@167
|
3406 E T9d, T9G, T9q, T9F;
|
cannam@167
|
3407 {
|
cannam@167
|
3408 E T8P, T9c, T9m, T9p;
|
cannam@167
|
3409 T8P = T8D + T8O;
|
cannam@167
|
3410 T9c = T90 + T9b;
|
cannam@167
|
3411 T9d = T8P - T9c;
|
cannam@167
|
3412 T9G = T8P + T9c;
|
cannam@167
|
3413 T9m = T9i + T9l;
|
cannam@167
|
3414 T9p = T9n + T9o;
|
cannam@167
|
3415 T9q = T9m - T9p;
|
cannam@167
|
3416 T9F = T9m + T9p;
|
cannam@167
|
3417 }
|
cannam@167
|
3418 T9r = FNMS(KP634393284, T9q, KP773010453 * T9d);
|
cannam@167
|
3419 T9L = FNMS(KP995184726, T9F, KP098017140 * T9G);
|
cannam@167
|
3420 T9v = FMA(KP773010453, T9q, KP634393284 * T9d);
|
cannam@167
|
3421 T9H = FMA(KP098017140, T9F, KP995184726 * T9G);
|
cannam@167
|
3422 }
|
cannam@167
|
3423 {
|
cannam@167
|
3424 E T7F, T9s, ThZ, Ti2;
|
cannam@167
|
3425 T7F = T6L + T7E;
|
cannam@167
|
3426 T9s = T8y + T9r;
|
cannam@167
|
3427 ci[WS(rs, 24)] = T7F - T9s;
|
cannam@167
|
3428 cr[WS(rs, 7)] = T7F + T9s;
|
cannam@167
|
3429 ThZ = T9v - T9u;
|
cannam@167
|
3430 Ti2 = Ti0 + Ti1;
|
cannam@167
|
3431 cr[WS(rs, 39)] = ThZ - Ti2;
|
cannam@167
|
3432 ci[WS(rs, 56)] = ThZ + Ti2;
|
cannam@167
|
3433 }
|
cannam@167
|
3434 {
|
cannam@167
|
3435 E Ti3, Ti4, T9t, T9w;
|
cannam@167
|
3436 Ti3 = T9r - T8y;
|
cannam@167
|
3437 Ti4 = Ti1 - Ti0;
|
cannam@167
|
3438 cr[WS(rs, 55)] = Ti3 - Ti4;
|
cannam@167
|
3439 ci[WS(rs, 40)] = Ti3 + Ti4;
|
cannam@167
|
3440 T9t = T6L - T7E;
|
cannam@167
|
3441 T9w = T9u + T9v;
|
cannam@167
|
3442 cr[WS(rs, 23)] = T9t - T9w;
|
cannam@167
|
3443 ci[WS(rs, 8)] = T9t + T9w;
|
cannam@167
|
3444 }
|
cannam@167
|
3445 {
|
cannam@167
|
3446 E T9B, T9I, ThN, ThW;
|
cannam@167
|
3447 T9B = T9x + T9A;
|
cannam@167
|
3448 T9I = T9E + T9H;
|
cannam@167
|
3449 cr[WS(rs, 31)] = T9B - T9I;
|
cannam@167
|
3450 ci[0] = T9B + T9I;
|
cannam@167
|
3451 ThN = T9L - T9K;
|
cannam@167
|
3452 ThW = ThO + ThV;
|
cannam@167
|
3453 cr[WS(rs, 63)] = ThN - ThW;
|
cannam@167
|
3454 ci[WS(rs, 32)] = ThN + ThW;
|
cannam@167
|
3455 }
|
cannam@167
|
3456 {
|
cannam@167
|
3457 E ThX, ThY, T9J, T9M;
|
cannam@167
|
3458 ThX = T9H - T9E;
|
cannam@167
|
3459 ThY = ThV - ThO;
|
cannam@167
|
3460 cr[WS(rs, 47)] = ThX - ThY;
|
cannam@167
|
3461 ci[WS(rs, 48)] = ThX + ThY;
|
cannam@167
|
3462 T9J = T9x - T9A;
|
cannam@167
|
3463 T9M = T9K + T9L;
|
cannam@167
|
3464 ci[WS(rs, 16)] = T9J - T9M;
|
cannam@167
|
3465 cr[WS(rs, 15)] = T9J + T9M;
|
cannam@167
|
3466 }
|
cannam@167
|
3467 }
|
cannam@167
|
3468 {
|
cannam@167
|
3469 E Tft, Tg7, Tgh, Tgl, Th9, Thf, TfE, Th6, TfQ, Tg4, Tga, The, Tge, Tgk, Tg1;
|
cannam@167
|
3470 E Tg5;
|
cannam@167
|
3471 {
|
cannam@167
|
3472 E Tfp, Tfs, Tgf, Tgg;
|
cannam@167
|
3473 Tfp = Tj - TG;
|
cannam@167
|
3474 Tfs = Tfq - Tfr;
|
cannam@167
|
3475 Tft = Tfp - Tfs;
|
cannam@167
|
3476 Tg7 = Tfp + Tfs;
|
cannam@167
|
3477 Tgf = TfY + TfZ;
|
cannam@167
|
3478 Tgg = TfR + TfU;
|
cannam@167
|
3479 Tgh = FMA(KP382683432, Tgf, KP923879532 * Tgg);
|
cannam@167
|
3480 Tgl = FNMS(KP923879532, Tgf, KP382683432 * Tgg);
|
cannam@167
|
3481 }
|
cannam@167
|
3482 {
|
cannam@167
|
3483 E Th7, Th8, Tfy, TfD;
|
cannam@167
|
3484 Th7 = T14 - T1r;
|
cannam@167
|
3485 Th8 = TgT - TgO;
|
cannam@167
|
3486 Th9 = Th7 + Th8;
|
cannam@167
|
3487 Thf = Th8 - Th7;
|
cannam@167
|
3488 Tfy = Tfu + Tfx;
|
cannam@167
|
3489 TfD = Tfz - TfC;
|
cannam@167
|
3490 TfE = KP707106781 * (Tfy + TfD);
|
cannam@167
|
3491 Th6 = KP707106781 * (Tfy - TfD);
|
cannam@167
|
3492 }
|
cannam@167
|
3493 {
|
cannam@167
|
3494 E TfK, TfP, Tg8, Tg9;
|
cannam@167
|
3495 TfK = TfI - TfJ;
|
cannam@167
|
3496 TfP = TfL - TfO;
|
cannam@167
|
3497 TfQ = FMA(KP382683432, TfK, KP923879532 * TfP);
|
cannam@167
|
3498 Tg4 = FNMS(KP923879532, TfK, KP382683432 * TfP);
|
cannam@167
|
3499 Tg8 = Tfu - Tfx;
|
cannam@167
|
3500 Tg9 = Tfz + TfC;
|
cannam@167
|
3501 Tga = KP707106781 * (Tg8 + Tg9);
|
cannam@167
|
3502 The = KP707106781 * (Tg9 - Tg8);
|
cannam@167
|
3503 }
|
cannam@167
|
3504 {
|
cannam@167
|
3505 E Tgc, Tgd, TfV, Tg0;
|
cannam@167
|
3506 Tgc = TfL + TfO;
|
cannam@167
|
3507 Tgd = TfI + TfJ;
|
cannam@167
|
3508 Tge = FNMS(KP382683432, Tgd, KP923879532 * Tgc);
|
cannam@167
|
3509 Tgk = FMA(KP923879532, Tgd, KP382683432 * Tgc);
|
cannam@167
|
3510 TfV = TfR - TfU;
|
cannam@167
|
3511 Tg0 = TfY - TfZ;
|
cannam@167
|
3512 Tg1 = FNMS(KP382683432, Tg0, KP923879532 * TfV);
|
cannam@167
|
3513 Tg5 = FMA(KP923879532, Tg0, KP382683432 * TfV);
|
cannam@167
|
3514 }
|
cannam@167
|
3515 {
|
cannam@167
|
3516 E TfF, Tg2, Thd, Thg;
|
cannam@167
|
3517 TfF = Tft + TfE;
|
cannam@167
|
3518 Tg2 = TfQ + Tg1;
|
cannam@167
|
3519 ci[WS(rs, 27)] = TfF - Tg2;
|
cannam@167
|
3520 cr[WS(rs, 4)] = TfF + Tg2;
|
cannam@167
|
3521 Thd = Tg5 - Tg4;
|
cannam@167
|
3522 Thg = The + Thf;
|
cannam@167
|
3523 cr[WS(rs, 36)] = Thd - Thg;
|
cannam@167
|
3524 ci[WS(rs, 59)] = Thd + Thg;
|
cannam@167
|
3525 }
|
cannam@167
|
3526 {
|
cannam@167
|
3527 E Thh, Thi, Tg3, Tg6;
|
cannam@167
|
3528 Thh = Tg1 - TfQ;
|
cannam@167
|
3529 Thi = Thf - The;
|
cannam@167
|
3530 cr[WS(rs, 52)] = Thh - Thi;
|
cannam@167
|
3531 ci[WS(rs, 43)] = Thh + Thi;
|
cannam@167
|
3532 Tg3 = Tft - TfE;
|
cannam@167
|
3533 Tg6 = Tg4 + Tg5;
|
cannam@167
|
3534 cr[WS(rs, 20)] = Tg3 - Tg6;
|
cannam@167
|
3535 ci[WS(rs, 11)] = Tg3 + Tg6;
|
cannam@167
|
3536 }
|
cannam@167
|
3537 {
|
cannam@167
|
3538 E Tgb, Tgi, Th5, Tha;
|
cannam@167
|
3539 Tgb = Tg7 + Tga;
|
cannam@167
|
3540 Tgi = Tge + Tgh;
|
cannam@167
|
3541 cr[WS(rs, 28)] = Tgb - Tgi;
|
cannam@167
|
3542 ci[WS(rs, 3)] = Tgb + Tgi;
|
cannam@167
|
3543 Th5 = Tgl - Tgk;
|
cannam@167
|
3544 Tha = Th6 + Th9;
|
cannam@167
|
3545 cr[WS(rs, 60)] = Th5 - Tha;
|
cannam@167
|
3546 ci[WS(rs, 35)] = Th5 + Tha;
|
cannam@167
|
3547 }
|
cannam@167
|
3548 {
|
cannam@167
|
3549 E Thb, Thc, Tgj, Tgm;
|
cannam@167
|
3550 Thb = Tgh - Tge;
|
cannam@167
|
3551 Thc = Th9 - Th6;
|
cannam@167
|
3552 cr[WS(rs, 44)] = Thb - Thc;
|
cannam@167
|
3553 ci[WS(rs, 51)] = Thb + Thc;
|
cannam@167
|
3554 Tgj = Tg7 - Tga;
|
cannam@167
|
3555 Tgm = Tgk + Tgl;
|
cannam@167
|
3556 ci[WS(rs, 19)] = Tgj - Tgm;
|
cannam@167
|
3557 cr[WS(rs, 12)] = Tgj + Tgm;
|
cannam@167
|
3558 }
|
cannam@167
|
3559 }
|
cannam@167
|
3560 {
|
cannam@167
|
3561 E TeH, Tf9, TeO, Thk, Thp, Thv, Tfc, Thu, Tf3, Tfn, Tf7, Tfj, TeW, Tfm, Tf6;
|
cannam@167
|
3562 E Tfg;
|
cannam@167
|
3563 {
|
cannam@167
|
3564 E TeD, TeG, Tfa, Tfb;
|
cannam@167
|
3565 TeD = TcL + TcO;
|
cannam@167
|
3566 TeG = KP707106781 * (TeE + TeF);
|
cannam@167
|
3567 TeH = TeD - TeG;
|
cannam@167
|
3568 Tf9 = TeD + TeG;
|
cannam@167
|
3569 {
|
cannam@167
|
3570 E TeK, TeN, Thl, Tho;
|
cannam@167
|
3571 TeK = FMA(KP923879532, TeI, KP382683432 * TeJ);
|
cannam@167
|
3572 TeN = FNMS(KP923879532, TeM, KP382683432 * TeL);
|
cannam@167
|
3573 TeO = TeK + TeN;
|
cannam@167
|
3574 Thk = TeK - TeN;
|
cannam@167
|
3575 Thl = KP707106781 * (TcU - TcZ);
|
cannam@167
|
3576 Tho = Thm + Thn;
|
cannam@167
|
3577 Thp = Thl + Tho;
|
cannam@167
|
3578 Thv = Tho - Thl;
|
cannam@167
|
3579 }
|
cannam@167
|
3580 Tfa = FNMS(KP382683432, TeI, KP923879532 * TeJ);
|
cannam@167
|
3581 Tfb = FMA(KP382683432, TeM, KP923879532 * TeL);
|
cannam@167
|
3582 Tfc = Tfa + Tfb;
|
cannam@167
|
3583 Thu = Tfb - Tfa;
|
cannam@167
|
3584 {
|
cannam@167
|
3585 E TeZ, Tfh, Tf2, Tfi, TeY, Tf1;
|
cannam@167
|
3586 TeY = KP707106781 * (Te5 + Te0);
|
cannam@167
|
3587 TeZ = TeX - TeY;
|
cannam@167
|
3588 Tfh = TeX + TeY;
|
cannam@167
|
3589 Tf1 = KP707106781 * (Ted + Tee);
|
cannam@167
|
3590 Tf2 = Tf0 - Tf1;
|
cannam@167
|
3591 Tfi = Tf0 + Tf1;
|
cannam@167
|
3592 Tf3 = FNMS(KP555570233, Tf2, KP831469612 * TeZ);
|
cannam@167
|
3593 Tfn = FMA(KP980785280, Tfh, KP195090322 * Tfi);
|
cannam@167
|
3594 Tf7 = FMA(KP555570233, TeZ, KP831469612 * Tf2);
|
cannam@167
|
3595 Tfj = FNMS(KP980785280, Tfi, KP195090322 * Tfh);
|
cannam@167
|
3596 }
|
cannam@167
|
3597 {
|
cannam@167
|
3598 E TeS, Tfe, TeV, Tff, TeR, TeU;
|
cannam@167
|
3599 TeR = KP707106781 * (TdN + TdM);
|
cannam@167
|
3600 TeS = TeQ - TeR;
|
cannam@167
|
3601 Tfe = TeQ + TeR;
|
cannam@167
|
3602 TeU = KP707106781 * (Tdz + TdE);
|
cannam@167
|
3603 TeV = TeT - TeU;
|
cannam@167
|
3604 Tff = TeT + TeU;
|
cannam@167
|
3605 TeW = FMA(KP831469612, TeS, KP555570233 * TeV);
|
cannam@167
|
3606 Tfm = FNMS(KP195090322, Tff, KP980785280 * Tfe);
|
cannam@167
|
3607 Tf6 = FNMS(KP831469612, TeV, KP555570233 * TeS);
|
cannam@167
|
3608 Tfg = FMA(KP195090322, Tfe, KP980785280 * Tff);
|
cannam@167
|
3609 }
|
cannam@167
|
3610 }
|
cannam@167
|
3611 {
|
cannam@167
|
3612 E TeP, Tf4, Tht, Thw;
|
cannam@167
|
3613 TeP = TeH + TeO;
|
cannam@167
|
3614 Tf4 = TeW + Tf3;
|
cannam@167
|
3615 ci[WS(rs, 25)] = TeP - Tf4;
|
cannam@167
|
3616 cr[WS(rs, 6)] = TeP + Tf4;
|
cannam@167
|
3617 Tht = Tf7 - Tf6;
|
cannam@167
|
3618 Thw = Thu + Thv;
|
cannam@167
|
3619 cr[WS(rs, 38)] = Tht - Thw;
|
cannam@167
|
3620 ci[WS(rs, 57)] = Tht + Thw;
|
cannam@167
|
3621 }
|
cannam@167
|
3622 {
|
cannam@167
|
3623 E Thx, Thy, Tf5, Tf8;
|
cannam@167
|
3624 Thx = Tf3 - TeW;
|
cannam@167
|
3625 Thy = Thv - Thu;
|
cannam@167
|
3626 cr[WS(rs, 54)] = Thx - Thy;
|
cannam@167
|
3627 ci[WS(rs, 41)] = Thx + Thy;
|
cannam@167
|
3628 Tf5 = TeH - TeO;
|
cannam@167
|
3629 Tf8 = Tf6 + Tf7;
|
cannam@167
|
3630 cr[WS(rs, 22)] = Tf5 - Tf8;
|
cannam@167
|
3631 ci[WS(rs, 9)] = Tf5 + Tf8;
|
cannam@167
|
3632 }
|
cannam@167
|
3633 {
|
cannam@167
|
3634 E Tfd, Tfk, Thj, Thq;
|
cannam@167
|
3635 Tfd = Tf9 - Tfc;
|
cannam@167
|
3636 Tfk = Tfg + Tfj;
|
cannam@167
|
3637 ci[WS(rs, 17)] = Tfd - Tfk;
|
cannam@167
|
3638 cr[WS(rs, 14)] = Tfd + Tfk;
|
cannam@167
|
3639 Thj = Tfj - Tfg;
|
cannam@167
|
3640 Thq = Thk + Thp;
|
cannam@167
|
3641 cr[WS(rs, 62)] = Thj - Thq;
|
cannam@167
|
3642 ci[WS(rs, 33)] = Thj + Thq;
|
cannam@167
|
3643 }
|
cannam@167
|
3644 {
|
cannam@167
|
3645 E Thr, Ths, Tfl, Tfo;
|
cannam@167
|
3646 Thr = Tfn - Tfm;
|
cannam@167
|
3647 Ths = Thp - Thk;
|
cannam@167
|
3648 cr[WS(rs, 46)] = Thr - Ths;
|
cannam@167
|
3649 ci[WS(rs, 49)] = Thr + Ths;
|
cannam@167
|
3650 Tfl = Tf9 + Tfc;
|
cannam@167
|
3651 Tfo = Tfm + Tfn;
|
cannam@167
|
3652 cr[WS(rs, 30)] = Tfl - Tfo;
|
cannam@167
|
3653 ci[WS(rs, 1)] = Tfl + Tfo;
|
cannam@167
|
3654 }
|
cannam@167
|
3655 }
|
cannam@167
|
3656 {
|
cannam@167
|
3657 E Td1, Ten, Tdo, ThA, ThD, ThJ, Teq, ThI, Teh, TeB, Tel, Tex, TdQ, TeA, Tek;
|
cannam@167
|
3658 E Teu;
|
cannam@167
|
3659 {
|
cannam@167
|
3660 E TcP, Td0, Teo, Tep;
|
cannam@167
|
3661 TcP = TcL - TcO;
|
cannam@167
|
3662 Td0 = KP707106781 * (TcU + TcZ);
|
cannam@167
|
3663 Td1 = TcP - Td0;
|
cannam@167
|
3664 Ten = TcP + Td0;
|
cannam@167
|
3665 {
|
cannam@167
|
3666 E Tdc, Tdn, ThB, ThC;
|
cannam@167
|
3667 Tdc = FNMS(KP923879532, Tdb, KP382683432 * Td6);
|
cannam@167
|
3668 Tdn = FMA(KP923879532, Tdh, KP382683432 * Tdm);
|
cannam@167
|
3669 Tdo = Tdc + Tdn;
|
cannam@167
|
3670 ThA = Tdn - Tdc;
|
cannam@167
|
3671 ThB = KP707106781 * (TeF - TeE);
|
cannam@167
|
3672 ThC = Thn - Thm;
|
cannam@167
|
3673 ThD = ThB + ThC;
|
cannam@167
|
3674 ThJ = ThC - ThB;
|
cannam@167
|
3675 }
|
cannam@167
|
3676 Teo = FMA(KP382683432, Tdb, KP923879532 * Td6);
|
cannam@167
|
3677 Tep = FNMS(KP382683432, Tdh, KP923879532 * Tdm);
|
cannam@167
|
3678 Teq = Teo + Tep;
|
cannam@167
|
3679 ThI = Teo - Tep;
|
cannam@167
|
3680 {
|
cannam@167
|
3681 E Te7, Tew, Teg, Tev, Te6, Tef;
|
cannam@167
|
3682 Te6 = KP707106781 * (Te0 - Te5);
|
cannam@167
|
3683 Te7 = TdV - Te6;
|
cannam@167
|
3684 Tew = TdV + Te6;
|
cannam@167
|
3685 Tef = KP707106781 * (Ted - Tee);
|
cannam@167
|
3686 Teg = Tec - Tef;
|
cannam@167
|
3687 Tev = Tec + Tef;
|
cannam@167
|
3688 Teh = FMA(KP555570233, Te7, KP831469612 * Teg);
|
cannam@167
|
3689 TeB = FMA(KP980785280, Tew, KP195090322 * Tev);
|
cannam@167
|
3690 Tel = FNMS(KP831469612, Te7, KP555570233 * Teg);
|
cannam@167
|
3691 Tex = FNMS(KP195090322, Tew, KP980785280 * Tev);
|
cannam@167
|
3692 }
|
cannam@167
|
3693 {
|
cannam@167
|
3694 E TdG, Tet, TdP, Tes, TdF, TdO;
|
cannam@167
|
3695 TdF = KP707106781 * (Tdz - TdE);
|
cannam@167
|
3696 TdG = Tdu - TdF;
|
cannam@167
|
3697 Tet = Tdu + TdF;
|
cannam@167
|
3698 TdO = KP707106781 * (TdM - TdN);
|
cannam@167
|
3699 TdP = TdL - TdO;
|
cannam@167
|
3700 Tes = TdL + TdO;
|
cannam@167
|
3701 TdQ = FNMS(KP555570233, TdP, KP831469612 * TdG);
|
cannam@167
|
3702 TeA = FNMS(KP980785280, Tes, KP195090322 * Tet);
|
cannam@167
|
3703 Tek = FMA(KP831469612, TdP, KP555570233 * TdG);
|
cannam@167
|
3704 Teu = FMA(KP195090322, Tes, KP980785280 * Tet);
|
cannam@167
|
3705 }
|
cannam@167
|
3706 }
|
cannam@167
|
3707 {
|
cannam@167
|
3708 E Tdp, Tei, ThH, ThK;
|
cannam@167
|
3709 Tdp = Td1 + Tdo;
|
cannam@167
|
3710 Tei = TdQ + Teh;
|
cannam@167
|
3711 cr[WS(rs, 26)] = Tdp - Tei;
|
cannam@167
|
3712 ci[WS(rs, 5)] = Tdp + Tei;
|
cannam@167
|
3713 ThH = Tel - Tek;
|
cannam@167
|
3714 ThK = ThI + ThJ;
|
cannam@167
|
3715 cr[WS(rs, 58)] = ThH - ThK;
|
cannam@167
|
3716 ci[WS(rs, 37)] = ThH + ThK;
|
cannam@167
|
3717 }
|
cannam@167
|
3718 {
|
cannam@167
|
3719 E ThL, ThM, Tej, Tem;
|
cannam@167
|
3720 ThL = Teh - TdQ;
|
cannam@167
|
3721 ThM = ThJ - ThI;
|
cannam@167
|
3722 cr[WS(rs, 42)] = ThL - ThM;
|
cannam@167
|
3723 ci[WS(rs, 53)] = ThL + ThM;
|
cannam@167
|
3724 Tej = Td1 - Tdo;
|
cannam@167
|
3725 Tem = Tek + Tel;
|
cannam@167
|
3726 ci[WS(rs, 21)] = Tej - Tem;
|
cannam@167
|
3727 cr[WS(rs, 10)] = Tej + Tem;
|
cannam@167
|
3728 }
|
cannam@167
|
3729 {
|
cannam@167
|
3730 E Ter, Tey, Thz, ThE;
|
cannam@167
|
3731 Ter = Ten + Teq;
|
cannam@167
|
3732 Tey = Teu + Tex;
|
cannam@167
|
3733 ci[WS(rs, 29)] = Ter - Tey;
|
cannam@167
|
3734 cr[WS(rs, 2)] = Ter + Tey;
|
cannam@167
|
3735 Thz = TeB - TeA;
|
cannam@167
|
3736 ThE = ThA + ThD;
|
cannam@167
|
3737 cr[WS(rs, 34)] = Thz - ThE;
|
cannam@167
|
3738 ci[WS(rs, 61)] = Thz + ThE;
|
cannam@167
|
3739 }
|
cannam@167
|
3740 {
|
cannam@167
|
3741 E ThF, ThG, Tez, TeC;
|
cannam@167
|
3742 ThF = Tex - Teu;
|
cannam@167
|
3743 ThG = ThD - ThA;
|
cannam@167
|
3744 cr[WS(rs, 50)] = ThF - ThG;
|
cannam@167
|
3745 ci[WS(rs, 45)] = ThF + ThG;
|
cannam@167
|
3746 Tez = Ten - Teq;
|
cannam@167
|
3747 TeC = TeA + TeB;
|
cannam@167
|
3748 cr[WS(rs, 18)] = Tez - TeC;
|
cannam@167
|
3749 ci[WS(rs, 13)] = Tez + TeC;
|
cannam@167
|
3750 }
|
cannam@167
|
3751 }
|
cannam@167
|
3752 {
|
cannam@167
|
3753 E Tc3, Tcv, TiD, TiJ, Tca, TiI, Tcy, TiA, Tci, TcI, Tcs, TcC, Tcp, TcJ, Tct;
|
cannam@167
|
3754 E TcF;
|
cannam@167
|
3755 {
|
cannam@167
|
3756 E TbZ, Tc2, TiB, TiC;
|
cannam@167
|
3757 TbZ = Taz - TaC;
|
cannam@167
|
3758 Tc2 = Tc0 + Tc1;
|
cannam@167
|
3759 Tc3 = TbZ - Tc2;
|
cannam@167
|
3760 Tcv = TbZ + Tc2;
|
cannam@167
|
3761 TiB = TaG - TaJ;
|
cannam@167
|
3762 TiC = Tin - Tim;
|
cannam@167
|
3763 TiD = TiB + TiC;
|
cannam@167
|
3764 TiJ = TiC - TiB;
|
cannam@167
|
3765 }
|
cannam@167
|
3766 {
|
cannam@167
|
3767 E Tc6, Tcw, Tc9, Tcx;
|
cannam@167
|
3768 {
|
cannam@167
|
3769 E Tc4, Tc5, Tc7, Tc8;
|
cannam@167
|
3770 Tc4 = TaP - TaQ;
|
cannam@167
|
3771 Tc5 = TaM - TaN;
|
cannam@167
|
3772 Tc6 = FMA(KP831469612, Tc4, KP555570233 * Tc5);
|
cannam@167
|
3773 Tcw = FNMS(KP555570233, Tc4, KP831469612 * Tc5);
|
cannam@167
|
3774 Tc7 = TaW - TaX;
|
cannam@167
|
3775 Tc8 = TaT - TaU;
|
cannam@167
|
3776 Tc9 = FNMS(KP831469612, Tc8, KP555570233 * Tc7);
|
cannam@167
|
3777 Tcx = FMA(KP555570233, Tc8, KP831469612 * Tc7);
|
cannam@167
|
3778 }
|
cannam@167
|
3779 Tca = Tc6 + Tc9;
|
cannam@167
|
3780 TiI = Tcx - Tcw;
|
cannam@167
|
3781 Tcy = Tcw + Tcx;
|
cannam@167
|
3782 TiA = Tc6 - Tc9;
|
cannam@167
|
3783 }
|
cannam@167
|
3784 {
|
cannam@167
|
3785 E Tce, TcB, Tch, TcA;
|
cannam@167
|
3786 {
|
cannam@167
|
3787 E Tcc, Tcd, Tcf, Tcg;
|
cannam@167
|
3788 Tcc = Tbd - Tbe;
|
cannam@167
|
3789 Tcd = Tb7 - Tba;
|
cannam@167
|
3790 Tce = Tcc - Tcd;
|
cannam@167
|
3791 TcB = Tcc + Tcd;
|
cannam@167
|
3792 Tcf = Tb2 - Tb3;
|
cannam@167
|
3793 Tcg = Tbh - Tbg;
|
cannam@167
|
3794 Tch = Tcf - Tcg;
|
cannam@167
|
3795 TcA = Tcf + Tcg;
|
cannam@167
|
3796 }
|
cannam@167
|
3797 Tci = FMA(KP471396736, Tce, KP881921264 * Tch);
|
cannam@167
|
3798 TcI = FMA(KP956940335, TcB, KP290284677 * TcA);
|
cannam@167
|
3799 Tcs = FNMS(KP881921264, Tce, KP471396736 * Tch);
|
cannam@167
|
3800 TcC = FNMS(KP290284677, TcB, KP956940335 * TcA);
|
cannam@167
|
3801 }
|
cannam@167
|
3802 {
|
cannam@167
|
3803 E Tcl, TcE, Tco, TcD;
|
cannam@167
|
3804 {
|
cannam@167
|
3805 E Tcj, Tck, Tcm, Tcn;
|
cannam@167
|
3806 Tcj = Tbl - Tbm;
|
cannam@167
|
3807 Tck = TbA - Tbz;
|
cannam@167
|
3808 Tcl = Tcj - Tck;
|
cannam@167
|
3809 TcE = Tcj + Tck;
|
cannam@167
|
3810 Tcm = Tbw - Tbx;
|
cannam@167
|
3811 Tcn = Tbq - Tbt;
|
cannam@167
|
3812 Tco = Tcm - Tcn;
|
cannam@167
|
3813 TcD = Tcm + Tcn;
|
cannam@167
|
3814 }
|
cannam@167
|
3815 Tcp = FNMS(KP471396736, Tco, KP881921264 * Tcl);
|
cannam@167
|
3816 TcJ = FNMS(KP956940335, TcD, KP290284677 * TcE);
|
cannam@167
|
3817 Tct = FMA(KP881921264, Tco, KP471396736 * Tcl);
|
cannam@167
|
3818 TcF = FMA(KP290284677, TcD, KP956940335 * TcE);
|
cannam@167
|
3819 }
|
cannam@167
|
3820 {
|
cannam@167
|
3821 E Tcb, Tcq, TiH, TiK;
|
cannam@167
|
3822 Tcb = Tc3 + Tca;
|
cannam@167
|
3823 Tcq = Tci + Tcp;
|
cannam@167
|
3824 ci[WS(rs, 26)] = Tcb - Tcq;
|
cannam@167
|
3825 cr[WS(rs, 5)] = Tcb + Tcq;
|
cannam@167
|
3826 TiH = Tct - Tcs;
|
cannam@167
|
3827 TiK = TiI + TiJ;
|
cannam@167
|
3828 cr[WS(rs, 37)] = TiH - TiK;
|
cannam@167
|
3829 ci[WS(rs, 58)] = TiH + TiK;
|
cannam@167
|
3830 }
|
cannam@167
|
3831 {
|
cannam@167
|
3832 E TiL, TiM, Tcr, Tcu;
|
cannam@167
|
3833 TiL = Tcp - Tci;
|
cannam@167
|
3834 TiM = TiJ - TiI;
|
cannam@167
|
3835 cr[WS(rs, 53)] = TiL - TiM;
|
cannam@167
|
3836 ci[WS(rs, 42)] = TiL + TiM;
|
cannam@167
|
3837 Tcr = Tc3 - Tca;
|
cannam@167
|
3838 Tcu = Tcs + Tct;
|
cannam@167
|
3839 cr[WS(rs, 21)] = Tcr - Tcu;
|
cannam@167
|
3840 ci[WS(rs, 10)] = Tcr + Tcu;
|
cannam@167
|
3841 }
|
cannam@167
|
3842 {
|
cannam@167
|
3843 E Tcz, TcG, Tiz, TiE;
|
cannam@167
|
3844 Tcz = Tcv + Tcy;
|
cannam@167
|
3845 TcG = TcC + TcF;
|
cannam@167
|
3846 cr[WS(rs, 29)] = Tcz - TcG;
|
cannam@167
|
3847 ci[WS(rs, 2)] = Tcz + TcG;
|
cannam@167
|
3848 Tiz = TcJ - TcI;
|
cannam@167
|
3849 TiE = TiA + TiD;
|
cannam@167
|
3850 cr[WS(rs, 61)] = Tiz - TiE;
|
cannam@167
|
3851 ci[WS(rs, 34)] = Tiz + TiE;
|
cannam@167
|
3852 }
|
cannam@167
|
3853 {
|
cannam@167
|
3854 E TiF, TiG, TcH, TcK;
|
cannam@167
|
3855 TiF = TcF - TcC;
|
cannam@167
|
3856 TiG = TiD - TiA;
|
cannam@167
|
3857 cr[WS(rs, 45)] = TiF - TiG;
|
cannam@167
|
3858 ci[WS(rs, 50)] = TiF + TiG;
|
cannam@167
|
3859 TcH = Tcv - Tcy;
|
cannam@167
|
3860 TcK = TcI + TcJ;
|
cannam@167
|
3861 ci[WS(rs, 18)] = TcH - TcK;
|
cannam@167
|
3862 cr[WS(rs, 13)] = TcH + TcK;
|
cannam@167
|
3863 }
|
cannam@167
|
3864 }
|
cannam@167
|
3865 {
|
cannam@167
|
3866 E TaL, TbJ, Tip, Tiv, Tb0, Tiu, TbM, Tik, Tbk, TbW, TbG, TbQ, TbD, TbX, TbH;
|
cannam@167
|
3867 E TbT;
|
cannam@167
|
3868 {
|
cannam@167
|
3869 E TaD, TaK, Til, Tio;
|
cannam@167
|
3870 TaD = Taz + TaC;
|
cannam@167
|
3871 TaK = TaG + TaJ;
|
cannam@167
|
3872 TaL = TaD - TaK;
|
cannam@167
|
3873 TbJ = TaD + TaK;
|
cannam@167
|
3874 Til = Tc1 - Tc0;
|
cannam@167
|
3875 Tio = Tim + Tin;
|
cannam@167
|
3876 Tip = Til + Tio;
|
cannam@167
|
3877 Tiv = Tio - Til;
|
cannam@167
|
3878 }
|
cannam@167
|
3879 {
|
cannam@167
|
3880 E TaS, TbK, TaZ, TbL;
|
cannam@167
|
3881 {
|
cannam@167
|
3882 E TaO, TaR, TaV, TaY;
|
cannam@167
|
3883 TaO = TaM + TaN;
|
cannam@167
|
3884 TaR = TaP + TaQ;
|
cannam@167
|
3885 TaS = FNMS(KP980785280, TaR, KP195090322 * TaO);
|
cannam@167
|
3886 TbK = FMA(KP195090322, TaR, KP980785280 * TaO);
|
cannam@167
|
3887 TaV = TaT + TaU;
|
cannam@167
|
3888 TaY = TaW + TaX;
|
cannam@167
|
3889 TaZ = FMA(KP980785280, TaV, KP195090322 * TaY);
|
cannam@167
|
3890 TbL = FNMS(KP195090322, TaV, KP980785280 * TaY);
|
cannam@167
|
3891 }
|
cannam@167
|
3892 Tb0 = TaS + TaZ;
|
cannam@167
|
3893 Tiu = TbK - TbL;
|
cannam@167
|
3894 TbM = TbK + TbL;
|
cannam@167
|
3895 Tik = TaZ - TaS;
|
cannam@167
|
3896 }
|
cannam@167
|
3897 {
|
cannam@167
|
3898 E Tbc, TbO, Tbj, TbP;
|
cannam@167
|
3899 {
|
cannam@167
|
3900 E Tb4, Tbb, Tbf, Tbi;
|
cannam@167
|
3901 Tb4 = Tb2 + Tb3;
|
cannam@167
|
3902 Tbb = Tb7 + Tba;
|
cannam@167
|
3903 Tbc = Tb4 - Tbb;
|
cannam@167
|
3904 TbO = Tb4 + Tbb;
|
cannam@167
|
3905 Tbf = Tbd + Tbe;
|
cannam@167
|
3906 Tbi = Tbg + Tbh;
|
cannam@167
|
3907 Tbj = Tbf - Tbi;
|
cannam@167
|
3908 TbP = Tbf + Tbi;
|
cannam@167
|
3909 }
|
cannam@167
|
3910 Tbk = FMA(KP634393284, Tbc, KP773010453 * Tbj);
|
cannam@167
|
3911 TbW = FNMS(KP995184726, TbP, KP098017140 * TbO);
|
cannam@167
|
3912 TbG = FNMS(KP634393284, Tbj, KP773010453 * Tbc);
|
cannam@167
|
3913 TbQ = FMA(KP995184726, TbO, KP098017140 * TbP);
|
cannam@167
|
3914 }
|
cannam@167
|
3915 {
|
cannam@167
|
3916 E Tbv, TbR, TbC, TbS;
|
cannam@167
|
3917 {
|
cannam@167
|
3918 E Tbn, Tbu, Tby, TbB;
|
cannam@167
|
3919 Tbn = Tbl + Tbm;
|
cannam@167
|
3920 Tbu = Tbq + Tbt;
|
cannam@167
|
3921 Tbv = Tbn - Tbu;
|
cannam@167
|
3922 TbR = Tbn + Tbu;
|
cannam@167
|
3923 Tby = Tbw + Tbx;
|
cannam@167
|
3924 TbB = Tbz + TbA;
|
cannam@167
|
3925 TbC = Tby - TbB;
|
cannam@167
|
3926 TbS = Tby + TbB;
|
cannam@167
|
3927 }
|
cannam@167
|
3928 TbD = FNMS(KP773010453, TbC, KP634393284 * Tbv);
|
cannam@167
|
3929 TbX = FMA(KP098017140, TbR, KP995184726 * TbS);
|
cannam@167
|
3930 TbH = FMA(KP773010453, Tbv, KP634393284 * TbC);
|
cannam@167
|
3931 TbT = FNMS(KP098017140, TbS, KP995184726 * TbR);
|
cannam@167
|
3932 }
|
cannam@167
|
3933 {
|
cannam@167
|
3934 E Tb1, TbE, Tit, Tiw;
|
cannam@167
|
3935 Tb1 = TaL - Tb0;
|
cannam@167
|
3936 TbE = Tbk + TbD;
|
cannam@167
|
3937 ci[WS(rs, 22)] = Tb1 - TbE;
|
cannam@167
|
3938 cr[WS(rs, 9)] = Tb1 + TbE;
|
cannam@167
|
3939 Tit = TbD - Tbk;
|
cannam@167
|
3940 Tiw = Tiu + Tiv;
|
cannam@167
|
3941 cr[WS(rs, 57)] = Tit - Tiw;
|
cannam@167
|
3942 ci[WS(rs, 38)] = Tit + Tiw;
|
cannam@167
|
3943 }
|
cannam@167
|
3944 {
|
cannam@167
|
3945 E Tix, Tiy, TbF, TbI;
|
cannam@167
|
3946 Tix = TbH - TbG;
|
cannam@167
|
3947 Tiy = Tiv - Tiu;
|
cannam@167
|
3948 cr[WS(rs, 41)] = Tix - Tiy;
|
cannam@167
|
3949 ci[WS(rs, 54)] = Tix + Tiy;
|
cannam@167
|
3950 TbF = TaL + Tb0;
|
cannam@167
|
3951 TbI = TbG + TbH;
|
cannam@167
|
3952 cr[WS(rs, 25)] = TbF - TbI;
|
cannam@167
|
3953 ci[WS(rs, 6)] = TbF + TbI;
|
cannam@167
|
3954 }
|
cannam@167
|
3955 {
|
cannam@167
|
3956 E TbN, TbU, Tij, Tiq;
|
cannam@167
|
3957 TbN = TbJ + TbM;
|
cannam@167
|
3958 TbU = TbQ + TbT;
|
cannam@167
|
3959 ci[WS(rs, 30)] = TbN - TbU;
|
cannam@167
|
3960 cr[WS(rs, 1)] = TbN + TbU;
|
cannam@167
|
3961 Tij = TbX - TbW;
|
cannam@167
|
3962 Tiq = Tik + Tip;
|
cannam@167
|
3963 cr[WS(rs, 33)] = Tij - Tiq;
|
cannam@167
|
3964 ci[WS(rs, 62)] = Tij + Tiq;
|
cannam@167
|
3965 }
|
cannam@167
|
3966 {
|
cannam@167
|
3967 E Tir, Tis, TbV, TbY;
|
cannam@167
|
3968 Tir = TbT - TbQ;
|
cannam@167
|
3969 Tis = Tip - Tik;
|
cannam@167
|
3970 cr[WS(rs, 49)] = Tir - Tis;
|
cannam@167
|
3971 ci[WS(rs, 46)] = Tir + Tis;
|
cannam@167
|
3972 TbV = TbJ - TbM;
|
cannam@167
|
3973 TbY = TbW + TbX;
|
cannam@167
|
3974 cr[WS(rs, 17)] = TbV - TbY;
|
cannam@167
|
3975 ci[WS(rs, 14)] = TbV + TbY;
|
cannam@167
|
3976 }
|
cannam@167
|
3977 }
|
cannam@167
|
3978 {
|
cannam@167
|
3979 E T9R, Taj, Ti9, Tif, T9Y, Tie, Tam, Ti6, Ta6, Taw, Tag, Taq, Tad, Tax, Tah;
|
cannam@167
|
3980 E Tat;
|
cannam@167
|
3981 {
|
cannam@167
|
3982 E T9N, T9Q, Ti7, Ti8;
|
cannam@167
|
3983 T9N = T6b - T6m;
|
cannam@167
|
3984 T9Q = T9O + T9P;
|
cannam@167
|
3985 T9R = T9N - T9Q;
|
cannam@167
|
3986 Taj = T9N + T9Q;
|
cannam@167
|
3987 Ti7 = T6J - T6y;
|
cannam@167
|
3988 Ti8 = ThT - ThQ;
|
cannam@167
|
3989 Ti9 = Ti7 + Ti8;
|
cannam@167
|
3990 Tif = Ti8 - Ti7;
|
cannam@167
|
3991 }
|
cannam@167
|
3992 {
|
cannam@167
|
3993 E T9U, Tak, T9X, Tal;
|
cannam@167
|
3994 {
|
cannam@167
|
3995 E T9S, T9T, T9V, T9W;
|
cannam@167
|
3996 T9S = T6Q - T71;
|
cannam@167
|
3997 T9T = T77 - T7a;
|
cannam@167
|
3998 T9U = FNMS(KP831469612, T9T, KP555570233 * T9S);
|
cannam@167
|
3999 Tak = FMA(KP831469612, T9S, KP555570233 * T9T);
|
cannam@167
|
4000 T9V = T7h - T7s;
|
cannam@167
|
4001 T9W = T7y - T7B;
|
cannam@167
|
4002 T9X = FMA(KP555570233, T9V, KP831469612 * T9W);
|
cannam@167
|
4003 Tal = FNMS(KP555570233, T9W, KP831469612 * T9V);
|
cannam@167
|
4004 }
|
cannam@167
|
4005 T9Y = T9U + T9X;
|
cannam@167
|
4006 Tie = Tak - Tal;
|
cannam@167
|
4007 Tam = Tak + Tal;
|
cannam@167
|
4008 Ti6 = T9X - T9U;
|
cannam@167
|
4009 }
|
cannam@167
|
4010 {
|
cannam@167
|
4011 E Ta2, Tao, Ta5, Tap;
|
cannam@167
|
4012 {
|
cannam@167
|
4013 E Ta0, Ta1, Ta3, Ta4;
|
cannam@167
|
4014 Ta0 = T8p - T8s;
|
cannam@167
|
4015 Ta1 = T87 - T8i;
|
cannam@167
|
4016 Ta2 = Ta0 - Ta1;
|
cannam@167
|
4017 Tao = Ta0 + Ta1;
|
cannam@167
|
4018 Ta3 = T7K - T7V;
|
cannam@167
|
4019 Ta4 = T8v - T8u;
|
cannam@167
|
4020 Ta5 = Ta3 - Ta4;
|
cannam@167
|
4021 Tap = Ta3 + Ta4;
|
cannam@167
|
4022 }
|
cannam@167
|
4023 Ta6 = FMA(KP471396736, Ta2, KP881921264 * Ta5);
|
cannam@167
|
4024 Taw = FNMS(KP956940335, Tap, KP290284677 * Tao);
|
cannam@167
|
4025 Tag = FNMS(KP471396736, Ta5, KP881921264 * Ta2);
|
cannam@167
|
4026 Taq = FMA(KP956940335, Tao, KP290284677 * Tap);
|
cannam@167
|
4027 }
|
cannam@167
|
4028 {
|
cannam@167
|
4029 E Ta9, Tar, Tac, Tas;
|
cannam@167
|
4030 {
|
cannam@167
|
4031 E Ta7, Ta8, Taa, Tab;
|
cannam@167
|
4032 Ta7 = T8D - T8O;
|
cannam@167
|
4033 Ta8 = T9n - T9o;
|
cannam@167
|
4034 Ta9 = Ta7 - Ta8;
|
cannam@167
|
4035 Tar = Ta7 + Ta8;
|
cannam@167
|
4036 Taa = T9i - T9l;
|
cannam@167
|
4037 Tab = T9b - T90;
|
cannam@167
|
4038 Tac = Taa - Tab;
|
cannam@167
|
4039 Tas = Taa + Tab;
|
cannam@167
|
4040 }
|
cannam@167
|
4041 Tad = FNMS(KP881921264, Tac, KP471396736 * Ta9);
|
cannam@167
|
4042 Tax = FMA(KP290284677, Tar, KP956940335 * Tas);
|
cannam@167
|
4043 Tah = FMA(KP881921264, Ta9, KP471396736 * Tac);
|
cannam@167
|
4044 Tat = FNMS(KP290284677, Tas, KP956940335 * Tar);
|
cannam@167
|
4045 }
|
cannam@167
|
4046 {
|
cannam@167
|
4047 E T9Z, Tae, Tid, Tig;
|
cannam@167
|
4048 T9Z = T9R - T9Y;
|
cannam@167
|
4049 Tae = Ta6 + Tad;
|
cannam@167
|
4050 ci[WS(rs, 20)] = T9Z - Tae;
|
cannam@167
|
4051 cr[WS(rs, 11)] = T9Z + Tae;
|
cannam@167
|
4052 Tid = Tad - Ta6;
|
cannam@167
|
4053 Tig = Tie + Tif;
|
cannam@167
|
4054 cr[WS(rs, 59)] = Tid - Tig;
|
cannam@167
|
4055 ci[WS(rs, 36)] = Tid + Tig;
|
cannam@167
|
4056 }
|
cannam@167
|
4057 {
|
cannam@167
|
4058 E Tih, Tii, Taf, Tai;
|
cannam@167
|
4059 Tih = Tah - Tag;
|
cannam@167
|
4060 Tii = Tif - Tie;
|
cannam@167
|
4061 cr[WS(rs, 43)] = Tih - Tii;
|
cannam@167
|
4062 ci[WS(rs, 52)] = Tih + Tii;
|
cannam@167
|
4063 Taf = T9R + T9Y;
|
cannam@167
|
4064 Tai = Tag + Tah;
|
cannam@167
|
4065 cr[WS(rs, 27)] = Taf - Tai;
|
cannam@167
|
4066 ci[WS(rs, 4)] = Taf + Tai;
|
cannam@167
|
4067 }
|
cannam@167
|
4068 {
|
cannam@167
|
4069 E Tan, Tau, Ti5, Tia;
|
cannam@167
|
4070 Tan = Taj + Tam;
|
cannam@167
|
4071 Tau = Taq + Tat;
|
cannam@167
|
4072 ci[WS(rs, 28)] = Tan - Tau;
|
cannam@167
|
4073 cr[WS(rs, 3)] = Tan + Tau;
|
cannam@167
|
4074 Ti5 = Tax - Taw;
|
cannam@167
|
4075 Tia = Ti6 + Ti9;
|
cannam@167
|
4076 cr[WS(rs, 35)] = Ti5 - Tia;
|
cannam@167
|
4077 ci[WS(rs, 60)] = Ti5 + Tia;
|
cannam@167
|
4078 }
|
cannam@167
|
4079 {
|
cannam@167
|
4080 E Tib, Tic, Tav, Tay;
|
cannam@167
|
4081 Tib = Tat - Taq;
|
cannam@167
|
4082 Tic = Ti9 - Ti6;
|
cannam@167
|
4083 cr[WS(rs, 51)] = Tib - Tic;
|
cannam@167
|
4084 ci[WS(rs, 44)] = Tib + Tic;
|
cannam@167
|
4085 Tav = Taj - Tam;
|
cannam@167
|
4086 Tay = Taw + Tax;
|
cannam@167
|
4087 cr[WS(rs, 19)] = Tav - Tay;
|
cannam@167
|
4088 ci[WS(rs, 12)] = Tav + Tay;
|
cannam@167
|
4089 }
|
cannam@167
|
4090 }
|
cannam@167
|
4091 }
|
cannam@167
|
4092 }
|
cannam@167
|
4093 }
|
cannam@167
|
4094
|
cannam@167
|
4095 static const tw_instr twinstr[] = {
|
cannam@167
|
4096 {TW_FULL, 1, 64},
|
cannam@167
|
4097 {TW_NEXT, 1, 0}
|
cannam@167
|
4098 };
|
cannam@167
|
4099
|
cannam@167
|
4100 static const hc2hc_desc desc = { 64, "hf_64", twinstr, &GENUS, {808, 270, 230, 0} };
|
cannam@167
|
4101
|
cannam@167
|
4102 void X(codelet_hf_64) (planner *p) {
|
cannam@167
|
4103 X(khc2hc_register) (p, hf_64, &desc);
|
cannam@167
|
4104 }
|
cannam@167
|
4105 #endif
|