Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:36:17 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-dft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_twidsq.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 8 -name q1_8 -include q.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 528 FP additions, 288 FP multiplications,
|
Chris@10
|
32 * (or, 352 additions, 112 multiplications, 176 fused multiply/add),
|
Chris@10
|
33 * 190 stack variables, 1 constants, and 256 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "q.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void q1_8(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
40 {
|
Chris@10
|
41 INT m;
|
Chris@10
|
42 for (m = mb, W = W + (mb * 14); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 14, MAKE_VOLATILE_STRIDE(16, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
|
Chris@10
|
43 E T9C, T9N, T9l, T9E, T9D, T9O;
|
Chris@10
|
44 {
|
Chris@10
|
45 E TV, Tk, T1d, T7, T18, T1t, TQ, TD, T5t, T4S, T5L, T4F, T5G, T61, T5o;
|
Chris@10
|
46 E T5b, T6Z, T6o, T7h, T6b, T7c, T7x, T6U, T6H, Tbx, TaW, TbP, TaJ, TbK, Tc5;
|
Chris@10
|
47 E Tbs, Tbf, T2r, T1Q, T2J, T1D, T2E, T2Z, T2m, T29, T3X, T3m, T4f, T39, T4a;
|
Chris@10
|
48 E T4v, T3S, T3F, T8v, T7U, T8N, T7H, T8I, T93, T8q, T8d, Ta1, T9q, Taj, T9d;
|
Chris@10
|
49 E Tae, Taz, T9W, T9J, Te, T19, T1u, T1g, Tv, TR, TG, TW, T5H, T4M, T5O;
|
Chris@10
|
50 E T62, T5p, T53, T5u, T5e, T6i, T7d, T7y, T7k, T6z, T6V, T6K, T70, TbL, TaQ;
|
Chris@10
|
51 E TbS, Tc6, Tbt, Tb7, Tby, Tbi, T1K, T2F, T30, T2M, T21, T2n, T2c, T2s, T4b;
|
Chris@10
|
52 E T3g, T4i, T4w, T3T, T3x, T3Y, T3I, T7O, T8J, T94, T8Q, T85, T8r, T8g, T8w;
|
Chris@10
|
53 E Tak, T9r, T9K, T9A, Taf, T9k, Tal, T9u;
|
Chris@10
|
54 {
|
Chris@10
|
55 E T9a, T9F, T99, Tac, T9p, T9b, T9G, T9H;
|
Chris@10
|
56 {
|
Chris@10
|
57 E TaG, Tbb, TaF, TbI, TaV, TaH, Tbc, Tbd;
|
Chris@10
|
58 {
|
Chris@10
|
59 E T4C, T57, T4B, T5E, T4R, T4D, T58, T59;
|
Chris@10
|
60 {
|
Chris@10
|
61 E T4, Tz, T3, T16, Tj, T5, TA, TB;
|
Chris@10
|
62 {
|
Chris@10
|
63 E T1, T2, Th, Ti;
|
Chris@10
|
64 T1 = rio[0];
|
Chris@10
|
65 T2 = rio[WS(rs, 4)];
|
Chris@10
|
66 Th = iio[0];
|
Chris@10
|
67 Ti = iio[WS(rs, 4)];
|
Chris@10
|
68 T4 = rio[WS(rs, 2)];
|
Chris@10
|
69 Tz = T1 - T2;
|
Chris@10
|
70 T3 = T1 + T2;
|
Chris@10
|
71 T16 = Th + Ti;
|
Chris@10
|
72 Tj = Th - Ti;
|
Chris@10
|
73 T5 = rio[WS(rs, 6)];
|
Chris@10
|
74 TA = iio[WS(rs, 2)];
|
Chris@10
|
75 TB = iio[WS(rs, 6)];
|
Chris@10
|
76 }
|
Chris@10
|
77 {
|
Chris@10
|
78 E T4z, T4A, T4P, T4Q;
|
Chris@10
|
79 T4z = rio[WS(vs, 3)];
|
Chris@10
|
80 {
|
Chris@10
|
81 E Tg, T6, T17, TC;
|
Chris@10
|
82 Tg = T4 - T5;
|
Chris@10
|
83 T6 = T4 + T5;
|
Chris@10
|
84 T17 = TA + TB;
|
Chris@10
|
85 TC = TA - TB;
|
Chris@10
|
86 TV = Tj - Tg;
|
Chris@10
|
87 Tk = Tg + Tj;
|
Chris@10
|
88 T1d = T3 - T6;
|
Chris@10
|
89 T7 = T3 + T6;
|
Chris@10
|
90 T18 = T16 - T17;
|
Chris@10
|
91 T1t = T16 + T17;
|
Chris@10
|
92 TQ = Tz + TC;
|
Chris@10
|
93 TD = Tz - TC;
|
Chris@10
|
94 T4A = rio[WS(vs, 3) + WS(rs, 4)];
|
Chris@10
|
95 }
|
Chris@10
|
96 T4P = iio[WS(vs, 3)];
|
Chris@10
|
97 T4Q = iio[WS(vs, 3) + WS(rs, 4)];
|
Chris@10
|
98 T4C = rio[WS(vs, 3) + WS(rs, 2)];
|
Chris@10
|
99 T57 = T4z - T4A;
|
Chris@10
|
100 T4B = T4z + T4A;
|
Chris@10
|
101 T5E = T4P + T4Q;
|
Chris@10
|
102 T4R = T4P - T4Q;
|
Chris@10
|
103 T4D = rio[WS(vs, 3) + WS(rs, 6)];
|
Chris@10
|
104 T58 = iio[WS(vs, 3) + WS(rs, 2)];
|
Chris@10
|
105 T59 = iio[WS(vs, 3) + WS(rs, 6)];
|
Chris@10
|
106 }
|
Chris@10
|
107 }
|
Chris@10
|
108 {
|
Chris@10
|
109 E T68, T6D, T67, T7a, T6n, T69, T6E, T6F;
|
Chris@10
|
110 {
|
Chris@10
|
111 E T65, T66, T6l, T6m;
|
Chris@10
|
112 T65 = rio[WS(vs, 4)];
|
Chris@10
|
113 {
|
Chris@10
|
114 E T4O, T4E, T5F, T5a;
|
Chris@10
|
115 T4O = T4C - T4D;
|
Chris@10
|
116 T4E = T4C + T4D;
|
Chris@10
|
117 T5F = T58 + T59;
|
Chris@10
|
118 T5a = T58 - T59;
|
Chris@10
|
119 T5t = T4R - T4O;
|
Chris@10
|
120 T4S = T4O + T4R;
|
Chris@10
|
121 T5L = T4B - T4E;
|
Chris@10
|
122 T4F = T4B + T4E;
|
Chris@10
|
123 T5G = T5E - T5F;
|
Chris@10
|
124 T61 = T5E + T5F;
|
Chris@10
|
125 T5o = T57 + T5a;
|
Chris@10
|
126 T5b = T57 - T5a;
|
Chris@10
|
127 T66 = rio[WS(vs, 4) + WS(rs, 4)];
|
Chris@10
|
128 }
|
Chris@10
|
129 T6l = iio[WS(vs, 4)];
|
Chris@10
|
130 T6m = iio[WS(vs, 4) + WS(rs, 4)];
|
Chris@10
|
131 T68 = rio[WS(vs, 4) + WS(rs, 2)];
|
Chris@10
|
132 T6D = T65 - T66;
|
Chris@10
|
133 T67 = T65 + T66;
|
Chris@10
|
134 T7a = T6l + T6m;
|
Chris@10
|
135 T6n = T6l - T6m;
|
Chris@10
|
136 T69 = rio[WS(vs, 4) + WS(rs, 6)];
|
Chris@10
|
137 T6E = iio[WS(vs, 4) + WS(rs, 2)];
|
Chris@10
|
138 T6F = iio[WS(vs, 4) + WS(rs, 6)];
|
Chris@10
|
139 }
|
Chris@10
|
140 {
|
Chris@10
|
141 E TaD, TaE, TaT, TaU;
|
Chris@10
|
142 TaD = rio[WS(vs, 7)];
|
Chris@10
|
143 {
|
Chris@10
|
144 E T6k, T6a, T7b, T6G;
|
Chris@10
|
145 T6k = T68 - T69;
|
Chris@10
|
146 T6a = T68 + T69;
|
Chris@10
|
147 T7b = T6E + T6F;
|
Chris@10
|
148 T6G = T6E - T6F;
|
Chris@10
|
149 T6Z = T6n - T6k;
|
Chris@10
|
150 T6o = T6k + T6n;
|
Chris@10
|
151 T7h = T67 - T6a;
|
Chris@10
|
152 T6b = T67 + T6a;
|
Chris@10
|
153 T7c = T7a - T7b;
|
Chris@10
|
154 T7x = T7a + T7b;
|
Chris@10
|
155 T6U = T6D + T6G;
|
Chris@10
|
156 T6H = T6D - T6G;
|
Chris@10
|
157 TaE = rio[WS(vs, 7) + WS(rs, 4)];
|
Chris@10
|
158 }
|
Chris@10
|
159 TaT = iio[WS(vs, 7)];
|
Chris@10
|
160 TaU = iio[WS(vs, 7) + WS(rs, 4)];
|
Chris@10
|
161 TaG = rio[WS(vs, 7) + WS(rs, 2)];
|
Chris@10
|
162 Tbb = TaD - TaE;
|
Chris@10
|
163 TaF = TaD + TaE;
|
Chris@10
|
164 TbI = TaT + TaU;
|
Chris@10
|
165 TaV = TaT - TaU;
|
Chris@10
|
166 TaH = rio[WS(vs, 7) + WS(rs, 6)];
|
Chris@10
|
167 Tbc = iio[WS(vs, 7) + WS(rs, 2)];
|
Chris@10
|
168 Tbd = iio[WS(vs, 7) + WS(rs, 6)];
|
Chris@10
|
169 }
|
Chris@10
|
170 }
|
Chris@10
|
171 }
|
Chris@10
|
172 {
|
Chris@10
|
173 E T36, T3B, T35, T48, T3l, T37, T3C, T3D;
|
Chris@10
|
174 {
|
Chris@10
|
175 E T1A, T25, T1z, T2C, T1P, T1B, T26, T27;
|
Chris@10
|
176 {
|
Chris@10
|
177 E T1x, T1y, T1N, T1O;
|
Chris@10
|
178 T1x = rio[WS(vs, 1)];
|
Chris@10
|
179 {
|
Chris@10
|
180 E TaS, TaI, TbJ, Tbe;
|
Chris@10
|
181 TaS = TaG - TaH;
|
Chris@10
|
182 TaI = TaG + TaH;
|
Chris@10
|
183 TbJ = Tbc + Tbd;
|
Chris@10
|
184 Tbe = Tbc - Tbd;
|
Chris@10
|
185 Tbx = TaV - TaS;
|
Chris@10
|
186 TaW = TaS + TaV;
|
Chris@10
|
187 TbP = TaF - TaI;
|
Chris@10
|
188 TaJ = TaF + TaI;
|
Chris@10
|
189 TbK = TbI - TbJ;
|
Chris@10
|
190 Tc5 = TbI + TbJ;
|
Chris@10
|
191 Tbs = Tbb + Tbe;
|
Chris@10
|
192 Tbf = Tbb - Tbe;
|
Chris@10
|
193 T1y = rio[WS(vs, 1) + WS(rs, 4)];
|
Chris@10
|
194 }
|
Chris@10
|
195 T1N = iio[WS(vs, 1)];
|
Chris@10
|
196 T1O = iio[WS(vs, 1) + WS(rs, 4)];
|
Chris@10
|
197 T1A = rio[WS(vs, 1) + WS(rs, 2)];
|
Chris@10
|
198 T25 = T1x - T1y;
|
Chris@10
|
199 T1z = T1x + T1y;
|
Chris@10
|
200 T2C = T1N + T1O;
|
Chris@10
|
201 T1P = T1N - T1O;
|
Chris@10
|
202 T1B = rio[WS(vs, 1) + WS(rs, 6)];
|
Chris@10
|
203 T26 = iio[WS(vs, 1) + WS(rs, 2)];
|
Chris@10
|
204 T27 = iio[WS(vs, 1) + WS(rs, 6)];
|
Chris@10
|
205 }
|
Chris@10
|
206 {
|
Chris@10
|
207 E T33, T34, T3j, T3k;
|
Chris@10
|
208 T33 = rio[WS(vs, 2)];
|
Chris@10
|
209 {
|
Chris@10
|
210 E T1M, T1C, T2D, T28;
|
Chris@10
|
211 T1M = T1A - T1B;
|
Chris@10
|
212 T1C = T1A + T1B;
|
Chris@10
|
213 T2D = T26 + T27;
|
Chris@10
|
214 T28 = T26 - T27;
|
Chris@10
|
215 T2r = T1P - T1M;
|
Chris@10
|
216 T1Q = T1M + T1P;
|
Chris@10
|
217 T2J = T1z - T1C;
|
Chris@10
|
218 T1D = T1z + T1C;
|
Chris@10
|
219 T2E = T2C - T2D;
|
Chris@10
|
220 T2Z = T2C + T2D;
|
Chris@10
|
221 T2m = T25 + T28;
|
Chris@10
|
222 T29 = T25 - T28;
|
Chris@10
|
223 T34 = rio[WS(vs, 2) + WS(rs, 4)];
|
Chris@10
|
224 }
|
Chris@10
|
225 T3j = iio[WS(vs, 2)];
|
Chris@10
|
226 T3k = iio[WS(vs, 2) + WS(rs, 4)];
|
Chris@10
|
227 T36 = rio[WS(vs, 2) + WS(rs, 2)];
|
Chris@10
|
228 T3B = T33 - T34;
|
Chris@10
|
229 T35 = T33 + T34;
|
Chris@10
|
230 T48 = T3j + T3k;
|
Chris@10
|
231 T3l = T3j - T3k;
|
Chris@10
|
232 T37 = rio[WS(vs, 2) + WS(rs, 6)];
|
Chris@10
|
233 T3C = iio[WS(vs, 2) + WS(rs, 2)];
|
Chris@10
|
234 T3D = iio[WS(vs, 2) + WS(rs, 6)];
|
Chris@10
|
235 }
|
Chris@10
|
236 }
|
Chris@10
|
237 {
|
Chris@10
|
238 E T7E, T89, T7D, T8G, T7T, T7F, T8a, T8b;
|
Chris@10
|
239 {
|
Chris@10
|
240 E T7B, T7C, T7R, T7S;
|
Chris@10
|
241 T7B = rio[WS(vs, 5)];
|
Chris@10
|
242 {
|
Chris@10
|
243 E T3i, T38, T49, T3E;
|
Chris@10
|
244 T3i = T36 - T37;
|
Chris@10
|
245 T38 = T36 + T37;
|
Chris@10
|
246 T49 = T3C + T3D;
|
Chris@10
|
247 T3E = T3C - T3D;
|
Chris@10
|
248 T3X = T3l - T3i;
|
Chris@10
|
249 T3m = T3i + T3l;
|
Chris@10
|
250 T4f = T35 - T38;
|
Chris@10
|
251 T39 = T35 + T38;
|
Chris@10
|
252 T4a = T48 - T49;
|
Chris@10
|
253 T4v = T48 + T49;
|
Chris@10
|
254 T3S = T3B + T3E;
|
Chris@10
|
255 T3F = T3B - T3E;
|
Chris@10
|
256 T7C = rio[WS(vs, 5) + WS(rs, 4)];
|
Chris@10
|
257 }
|
Chris@10
|
258 T7R = iio[WS(vs, 5)];
|
Chris@10
|
259 T7S = iio[WS(vs, 5) + WS(rs, 4)];
|
Chris@10
|
260 T7E = rio[WS(vs, 5) + WS(rs, 2)];
|
Chris@10
|
261 T89 = T7B - T7C;
|
Chris@10
|
262 T7D = T7B + T7C;
|
Chris@10
|
263 T8G = T7R + T7S;
|
Chris@10
|
264 T7T = T7R - T7S;
|
Chris@10
|
265 T7F = rio[WS(vs, 5) + WS(rs, 6)];
|
Chris@10
|
266 T8a = iio[WS(vs, 5) + WS(rs, 2)];
|
Chris@10
|
267 T8b = iio[WS(vs, 5) + WS(rs, 6)];
|
Chris@10
|
268 }
|
Chris@10
|
269 {
|
Chris@10
|
270 E T97, T98, T9n, T9o;
|
Chris@10
|
271 T97 = rio[WS(vs, 6)];
|
Chris@10
|
272 {
|
Chris@10
|
273 E T7Q, T7G, T8H, T8c;
|
Chris@10
|
274 T7Q = T7E - T7F;
|
Chris@10
|
275 T7G = T7E + T7F;
|
Chris@10
|
276 T8H = T8a + T8b;
|
Chris@10
|
277 T8c = T8a - T8b;
|
Chris@10
|
278 T8v = T7T - T7Q;
|
Chris@10
|
279 T7U = T7Q + T7T;
|
Chris@10
|
280 T8N = T7D - T7G;
|
Chris@10
|
281 T7H = T7D + T7G;
|
Chris@10
|
282 T8I = T8G - T8H;
|
Chris@10
|
283 T93 = T8G + T8H;
|
Chris@10
|
284 T8q = T89 + T8c;
|
Chris@10
|
285 T8d = T89 - T8c;
|
Chris@10
|
286 T98 = rio[WS(vs, 6) + WS(rs, 4)];
|
Chris@10
|
287 }
|
Chris@10
|
288 T9n = iio[WS(vs, 6)];
|
Chris@10
|
289 T9o = iio[WS(vs, 6) + WS(rs, 4)];
|
Chris@10
|
290 T9a = rio[WS(vs, 6) + WS(rs, 2)];
|
Chris@10
|
291 T9F = T97 - T98;
|
Chris@10
|
292 T99 = T97 + T98;
|
Chris@10
|
293 Tac = T9n + T9o;
|
Chris@10
|
294 T9p = T9n - T9o;
|
Chris@10
|
295 T9b = rio[WS(vs, 6) + WS(rs, 6)];
|
Chris@10
|
296 T9G = iio[WS(vs, 6) + WS(rs, 2)];
|
Chris@10
|
297 T9H = iio[WS(vs, 6) + WS(rs, 6)];
|
Chris@10
|
298 }
|
Chris@10
|
299 }
|
Chris@10
|
300 }
|
Chris@10
|
301 }
|
Chris@10
|
302 {
|
Chris@10
|
303 E TbQ, TaX, Tbg, Tb6, TbR, Tb0;
|
Chris@10
|
304 {
|
Chris@10
|
305 E T5M, T4T, T5c, T52, T5N, T4W;
|
Chris@10
|
306 {
|
Chris@10
|
307 E Tu, TE, TF, Tp;
|
Chris@10
|
308 {
|
Chris@10
|
309 E Tb, Tq, Ta, T1e, Tt, Tc, Tm, Tn;
|
Chris@10
|
310 {
|
Chris@10
|
311 E T8, T9, Tr, Ts;
|
Chris@10
|
312 T8 = rio[WS(rs, 1)];
|
Chris@10
|
313 {
|
Chris@10
|
314 E T9m, T9c, Tad, T9I;
|
Chris@10
|
315 T9m = T9a - T9b;
|
Chris@10
|
316 T9c = T9a + T9b;
|
Chris@10
|
317 Tad = T9G + T9H;
|
Chris@10
|
318 T9I = T9G - T9H;
|
Chris@10
|
319 Ta1 = T9p - T9m;
|
Chris@10
|
320 T9q = T9m + T9p;
|
Chris@10
|
321 Taj = T99 - T9c;
|
Chris@10
|
322 T9d = T99 + T9c;
|
Chris@10
|
323 Tae = Tac - Tad;
|
Chris@10
|
324 Taz = Tac + Tad;
|
Chris@10
|
325 T9W = T9F + T9I;
|
Chris@10
|
326 T9J = T9F - T9I;
|
Chris@10
|
327 T9 = rio[WS(rs, 5)];
|
Chris@10
|
328 }
|
Chris@10
|
329 Tr = iio[WS(rs, 1)];
|
Chris@10
|
330 Ts = iio[WS(rs, 5)];
|
Chris@10
|
331 Tb = rio[WS(rs, 7)];
|
Chris@10
|
332 Tq = T8 - T9;
|
Chris@10
|
333 Ta = T8 + T9;
|
Chris@10
|
334 T1e = Tr + Ts;
|
Chris@10
|
335 Tt = Tr - Ts;
|
Chris@10
|
336 Tc = rio[WS(rs, 3)];
|
Chris@10
|
337 Tm = iio[WS(rs, 7)];
|
Chris@10
|
338 Tn = iio[WS(rs, 3)];
|
Chris@10
|
339 }
|
Chris@10
|
340 {
|
Chris@10
|
341 E Tl, Td, T1f, To;
|
Chris@10
|
342 Tu = Tq + Tt;
|
Chris@10
|
343 TE = Tt - Tq;
|
Chris@10
|
344 Tl = Tb - Tc;
|
Chris@10
|
345 Td = Tb + Tc;
|
Chris@10
|
346 T1f = Tm + Tn;
|
Chris@10
|
347 To = Tm - Tn;
|
Chris@10
|
348 Te = Ta + Td;
|
Chris@10
|
349 T19 = Td - Ta;
|
Chris@10
|
350 T1u = T1e + T1f;
|
Chris@10
|
351 T1g = T1e - T1f;
|
Chris@10
|
352 TF = Tl + To;
|
Chris@10
|
353 Tp = Tl - To;
|
Chris@10
|
354 }
|
Chris@10
|
355 }
|
Chris@10
|
356 {
|
Chris@10
|
357 E T4I, T4Y, T4U, T51, T4L, T4V;
|
Chris@10
|
358 {
|
Chris@10
|
359 E T4Z, T50, T4G, T4H, T4J, T4K;
|
Chris@10
|
360 T4G = rio[WS(vs, 3) + WS(rs, 1)];
|
Chris@10
|
361 T4H = rio[WS(vs, 3) + WS(rs, 5)];
|
Chris@10
|
362 Tv = Tp - Tu;
|
Chris@10
|
363 TR = Tu + Tp;
|
Chris@10
|
364 TG = TE - TF;
|
Chris@10
|
365 TW = TE + TF;
|
Chris@10
|
366 T4I = T4G + T4H;
|
Chris@10
|
367 T4Y = T4G - T4H;
|
Chris@10
|
368 T4Z = iio[WS(vs, 3) + WS(rs, 1)];
|
Chris@10
|
369 T50 = iio[WS(vs, 3) + WS(rs, 5)];
|
Chris@10
|
370 T4J = rio[WS(vs, 3) + WS(rs, 7)];
|
Chris@10
|
371 T4K = rio[WS(vs, 3) + WS(rs, 3)];
|
Chris@10
|
372 T4U = iio[WS(vs, 3) + WS(rs, 7)];
|
Chris@10
|
373 T51 = T4Z - T50;
|
Chris@10
|
374 T5M = T4Z + T50;
|
Chris@10
|
375 T4L = T4J + T4K;
|
Chris@10
|
376 T4T = T4J - T4K;
|
Chris@10
|
377 T4V = iio[WS(vs, 3) + WS(rs, 3)];
|
Chris@10
|
378 }
|
Chris@10
|
379 T5c = T51 - T4Y;
|
Chris@10
|
380 T52 = T4Y + T51;
|
Chris@10
|
381 T5H = T4L - T4I;
|
Chris@10
|
382 T4M = T4I + T4L;
|
Chris@10
|
383 T5N = T4U + T4V;
|
Chris@10
|
384 T4W = T4U - T4V;
|
Chris@10
|
385 }
|
Chris@10
|
386 }
|
Chris@10
|
387 {
|
Chris@10
|
388 E T7i, T6p, T6y, T6I, T6s, T7j;
|
Chris@10
|
389 {
|
Chris@10
|
390 E T6e, T6u, T6q, T6x, T6h, T6r;
|
Chris@10
|
391 {
|
Chris@10
|
392 E T6v, T6w, T6f, T6g;
|
Chris@10
|
393 {
|
Chris@10
|
394 E T4X, T5d, T6c, T6d;
|
Chris@10
|
395 T6c = rio[WS(vs, 4) + WS(rs, 1)];
|
Chris@10
|
396 T6d = rio[WS(vs, 4) + WS(rs, 5)];
|
Chris@10
|
397 T5O = T5M - T5N;
|
Chris@10
|
398 T62 = T5M + T5N;
|
Chris@10
|
399 T4X = T4T - T4W;
|
Chris@10
|
400 T5d = T4T + T4W;
|
Chris@10
|
401 T6e = T6c + T6d;
|
Chris@10
|
402 T6u = T6c - T6d;
|
Chris@10
|
403 T5p = T52 + T4X;
|
Chris@10
|
404 T53 = T4X - T52;
|
Chris@10
|
405 T5u = T5c + T5d;
|
Chris@10
|
406 T5e = T5c - T5d;
|
Chris@10
|
407 T6v = iio[WS(vs, 4) + WS(rs, 1)];
|
Chris@10
|
408 T6w = iio[WS(vs, 4) + WS(rs, 5)];
|
Chris@10
|
409 }
|
Chris@10
|
410 T6f = rio[WS(vs, 4) + WS(rs, 7)];
|
Chris@10
|
411 T6g = rio[WS(vs, 4) + WS(rs, 3)];
|
Chris@10
|
412 T6q = iio[WS(vs, 4) + WS(rs, 7)];
|
Chris@10
|
413 T7i = T6v + T6w;
|
Chris@10
|
414 T6x = T6v - T6w;
|
Chris@10
|
415 T6p = T6f - T6g;
|
Chris@10
|
416 T6h = T6f + T6g;
|
Chris@10
|
417 T6r = iio[WS(vs, 4) + WS(rs, 3)];
|
Chris@10
|
418 }
|
Chris@10
|
419 T6y = T6u + T6x;
|
Chris@10
|
420 T6I = T6x - T6u;
|
Chris@10
|
421 T6i = T6e + T6h;
|
Chris@10
|
422 T7d = T6h - T6e;
|
Chris@10
|
423 T6s = T6q - T6r;
|
Chris@10
|
424 T7j = T6q + T6r;
|
Chris@10
|
425 }
|
Chris@10
|
426 {
|
Chris@10
|
427 E Tb2, TaM, TaY, Tb5, TaP, TaZ;
|
Chris@10
|
428 {
|
Chris@10
|
429 E Tb3, Tb4, TaN, TaO;
|
Chris@10
|
430 {
|
Chris@10
|
431 E T6J, T6t, TaK, TaL;
|
Chris@10
|
432 TaK = rio[WS(vs, 7) + WS(rs, 1)];
|
Chris@10
|
433 TaL = rio[WS(vs, 7) + WS(rs, 5)];
|
Chris@10
|
434 T7y = T7i + T7j;
|
Chris@10
|
435 T7k = T7i - T7j;
|
Chris@10
|
436 T6J = T6p + T6s;
|
Chris@10
|
437 T6t = T6p - T6s;
|
Chris@10
|
438 Tb2 = TaK - TaL;
|
Chris@10
|
439 TaM = TaK + TaL;
|
Chris@10
|
440 T6z = T6t - T6y;
|
Chris@10
|
441 T6V = T6y + T6t;
|
Chris@10
|
442 T6K = T6I - T6J;
|
Chris@10
|
443 T70 = T6I + T6J;
|
Chris@10
|
444 Tb3 = iio[WS(vs, 7) + WS(rs, 1)];
|
Chris@10
|
445 Tb4 = iio[WS(vs, 7) + WS(rs, 5)];
|
Chris@10
|
446 }
|
Chris@10
|
447 TaN = rio[WS(vs, 7) + WS(rs, 7)];
|
Chris@10
|
448 TaO = rio[WS(vs, 7) + WS(rs, 3)];
|
Chris@10
|
449 TaY = iio[WS(vs, 7) + WS(rs, 7)];
|
Chris@10
|
450 Tb5 = Tb3 - Tb4;
|
Chris@10
|
451 TbQ = Tb3 + Tb4;
|
Chris@10
|
452 TaP = TaN + TaO;
|
Chris@10
|
453 TaX = TaN - TaO;
|
Chris@10
|
454 TaZ = iio[WS(vs, 7) + WS(rs, 3)];
|
Chris@10
|
455 }
|
Chris@10
|
456 Tbg = Tb5 - Tb2;
|
Chris@10
|
457 Tb6 = Tb2 + Tb5;
|
Chris@10
|
458 TbL = TaP - TaM;
|
Chris@10
|
459 TaQ = TaM + TaP;
|
Chris@10
|
460 TbR = TaY + TaZ;
|
Chris@10
|
461 Tb0 = TaY - TaZ;
|
Chris@10
|
462 }
|
Chris@10
|
463 }
|
Chris@10
|
464 }
|
Chris@10
|
465 {
|
Chris@10
|
466 E T4g, T3n, T3G, T3w, T4h, T3q;
|
Chris@10
|
467 {
|
Chris@10
|
468 E T2K, T1R, T20, T2a, T1U, T2L;
|
Chris@10
|
469 {
|
Chris@10
|
470 E T1G, T1W, T1S, T1Z, T1J, T1T;
|
Chris@10
|
471 {
|
Chris@10
|
472 E T1X, T1Y, T1H, T1I;
|
Chris@10
|
473 {
|
Chris@10
|
474 E Tb1, Tbh, T1E, T1F;
|
Chris@10
|
475 T1E = rio[WS(vs, 1) + WS(rs, 1)];
|
Chris@10
|
476 T1F = rio[WS(vs, 1) + WS(rs, 5)];
|
Chris@10
|
477 TbS = TbQ - TbR;
|
Chris@10
|
478 Tc6 = TbQ + TbR;
|
Chris@10
|
479 Tb1 = TaX - Tb0;
|
Chris@10
|
480 Tbh = TaX + Tb0;
|
Chris@10
|
481 T1G = T1E + T1F;
|
Chris@10
|
482 T1W = T1E - T1F;
|
Chris@10
|
483 Tbt = Tb6 + Tb1;
|
Chris@10
|
484 Tb7 = Tb1 - Tb6;
|
Chris@10
|
485 Tby = Tbg + Tbh;
|
Chris@10
|
486 Tbi = Tbg - Tbh;
|
Chris@10
|
487 T1X = iio[WS(vs, 1) + WS(rs, 1)];
|
Chris@10
|
488 T1Y = iio[WS(vs, 1) + WS(rs, 5)];
|
Chris@10
|
489 }
|
Chris@10
|
490 T1H = rio[WS(vs, 1) + WS(rs, 7)];
|
Chris@10
|
491 T1I = rio[WS(vs, 1) + WS(rs, 3)];
|
Chris@10
|
492 T1S = iio[WS(vs, 1) + WS(rs, 7)];
|
Chris@10
|
493 T2K = T1X + T1Y;
|
Chris@10
|
494 T1Z = T1X - T1Y;
|
Chris@10
|
495 T1R = T1H - T1I;
|
Chris@10
|
496 T1J = T1H + T1I;
|
Chris@10
|
497 T1T = iio[WS(vs, 1) + WS(rs, 3)];
|
Chris@10
|
498 }
|
Chris@10
|
499 T20 = T1W + T1Z;
|
Chris@10
|
500 T2a = T1Z - T1W;
|
Chris@10
|
501 T1K = T1G + T1J;
|
Chris@10
|
502 T2F = T1J - T1G;
|
Chris@10
|
503 T1U = T1S - T1T;
|
Chris@10
|
504 T2L = T1S + T1T;
|
Chris@10
|
505 }
|
Chris@10
|
506 {
|
Chris@10
|
507 E T3s, T3c, T3o, T3v, T3f, T3p;
|
Chris@10
|
508 {
|
Chris@10
|
509 E T3t, T3u, T3d, T3e;
|
Chris@10
|
510 {
|
Chris@10
|
511 E T2b, T1V, T3a, T3b;
|
Chris@10
|
512 T3a = rio[WS(vs, 2) + WS(rs, 1)];
|
Chris@10
|
513 T3b = rio[WS(vs, 2) + WS(rs, 5)];
|
Chris@10
|
514 T30 = T2K + T2L;
|
Chris@10
|
515 T2M = T2K - T2L;
|
Chris@10
|
516 T2b = T1R + T1U;
|
Chris@10
|
517 T1V = T1R - T1U;
|
Chris@10
|
518 T3s = T3a - T3b;
|
Chris@10
|
519 T3c = T3a + T3b;
|
Chris@10
|
520 T21 = T1V - T20;
|
Chris@10
|
521 T2n = T20 + T1V;
|
Chris@10
|
522 T2c = T2a - T2b;
|
Chris@10
|
523 T2s = T2a + T2b;
|
Chris@10
|
524 T3t = iio[WS(vs, 2) + WS(rs, 1)];
|
Chris@10
|
525 T3u = iio[WS(vs, 2) + WS(rs, 5)];
|
Chris@10
|
526 }
|
Chris@10
|
527 T3d = rio[WS(vs, 2) + WS(rs, 7)];
|
Chris@10
|
528 T3e = rio[WS(vs, 2) + WS(rs, 3)];
|
Chris@10
|
529 T3o = iio[WS(vs, 2) + WS(rs, 7)];
|
Chris@10
|
530 T3v = T3t - T3u;
|
Chris@10
|
531 T4g = T3t + T3u;
|
Chris@10
|
532 T3f = T3d + T3e;
|
Chris@10
|
533 T3n = T3d - T3e;
|
Chris@10
|
534 T3p = iio[WS(vs, 2) + WS(rs, 3)];
|
Chris@10
|
535 }
|
Chris@10
|
536 T3G = T3v - T3s;
|
Chris@10
|
537 T3w = T3s + T3v;
|
Chris@10
|
538 T4b = T3f - T3c;
|
Chris@10
|
539 T3g = T3c + T3f;
|
Chris@10
|
540 T4h = T3o + T3p;
|
Chris@10
|
541 T3q = T3o - T3p;
|
Chris@10
|
542 }
|
Chris@10
|
543 }
|
Chris@10
|
544 {
|
Chris@10
|
545 E T8O, T7V, T84, T8e, T7Y, T8P;
|
Chris@10
|
546 {
|
Chris@10
|
547 E T7K, T80, T7W, T83, T7N, T7X;
|
Chris@10
|
548 {
|
Chris@10
|
549 E T81, T82, T7L, T7M;
|
Chris@10
|
550 {
|
Chris@10
|
551 E T3r, T3H, T7I, T7J;
|
Chris@10
|
552 T7I = rio[WS(vs, 5) + WS(rs, 1)];
|
Chris@10
|
553 T7J = rio[WS(vs, 5) + WS(rs, 5)];
|
Chris@10
|
554 T4i = T4g - T4h;
|
Chris@10
|
555 T4w = T4g + T4h;
|
Chris@10
|
556 T3r = T3n - T3q;
|
Chris@10
|
557 T3H = T3n + T3q;
|
Chris@10
|
558 T7K = T7I + T7J;
|
Chris@10
|
559 T80 = T7I - T7J;
|
Chris@10
|
560 T3T = T3w + T3r;
|
Chris@10
|
561 T3x = T3r - T3w;
|
Chris@10
|
562 T3Y = T3G + T3H;
|
Chris@10
|
563 T3I = T3G - T3H;
|
Chris@10
|
564 T81 = iio[WS(vs, 5) + WS(rs, 1)];
|
Chris@10
|
565 T82 = iio[WS(vs, 5) + WS(rs, 5)];
|
Chris@10
|
566 }
|
Chris@10
|
567 T7L = rio[WS(vs, 5) + WS(rs, 7)];
|
Chris@10
|
568 T7M = rio[WS(vs, 5) + WS(rs, 3)];
|
Chris@10
|
569 T7W = iio[WS(vs, 5) + WS(rs, 7)];
|
Chris@10
|
570 T8O = T81 + T82;
|
Chris@10
|
571 T83 = T81 - T82;
|
Chris@10
|
572 T7V = T7L - T7M;
|
Chris@10
|
573 T7N = T7L + T7M;
|
Chris@10
|
574 T7X = iio[WS(vs, 5) + WS(rs, 3)];
|
Chris@10
|
575 }
|
Chris@10
|
576 T84 = T80 + T83;
|
Chris@10
|
577 T8e = T83 - T80;
|
Chris@10
|
578 T7O = T7K + T7N;
|
Chris@10
|
579 T8J = T7N - T7K;
|
Chris@10
|
580 T7Y = T7W - T7X;
|
Chris@10
|
581 T8P = T7W + T7X;
|
Chris@10
|
582 }
|
Chris@10
|
583 {
|
Chris@10
|
584 E T9w, T9g, T9s, T9z, T9j, T9t;
|
Chris@10
|
585 {
|
Chris@10
|
586 E T9x, T9y, T9h, T9i;
|
Chris@10
|
587 {
|
Chris@10
|
588 E T8f, T7Z, T9e, T9f;
|
Chris@10
|
589 T9e = rio[WS(vs, 6) + WS(rs, 1)];
|
Chris@10
|
590 T9f = rio[WS(vs, 6) + WS(rs, 5)];
|
Chris@10
|
591 T94 = T8O + T8P;
|
Chris@10
|
592 T8Q = T8O - T8P;
|
Chris@10
|
593 T8f = T7V + T7Y;
|
Chris@10
|
594 T7Z = T7V - T7Y;
|
Chris@10
|
595 T9w = T9e - T9f;
|
Chris@10
|
596 T9g = T9e + T9f;
|
Chris@10
|
597 T85 = T7Z - T84;
|
Chris@10
|
598 T8r = T84 + T7Z;
|
Chris@10
|
599 T8g = T8e - T8f;
|
Chris@10
|
600 T8w = T8e + T8f;
|
Chris@10
|
601 T9x = iio[WS(vs, 6) + WS(rs, 1)];
|
Chris@10
|
602 T9y = iio[WS(vs, 6) + WS(rs, 5)];
|
Chris@10
|
603 }
|
Chris@10
|
604 T9h = rio[WS(vs, 6) + WS(rs, 7)];
|
Chris@10
|
605 T9i = rio[WS(vs, 6) + WS(rs, 3)];
|
Chris@10
|
606 T9s = iio[WS(vs, 6) + WS(rs, 7)];
|
Chris@10
|
607 T9z = T9x - T9y;
|
Chris@10
|
608 Tak = T9x + T9y;
|
Chris@10
|
609 T9j = T9h + T9i;
|
Chris@10
|
610 T9r = T9h - T9i;
|
Chris@10
|
611 T9t = iio[WS(vs, 6) + WS(rs, 3)];
|
Chris@10
|
612 }
|
Chris@10
|
613 T9K = T9z - T9w;
|
Chris@10
|
614 T9A = T9w + T9z;
|
Chris@10
|
615 Taf = T9j - T9g;
|
Chris@10
|
616 T9k = T9g + T9j;
|
Chris@10
|
617 Tal = T9s + T9t;
|
Chris@10
|
618 T9u = T9s - T9t;
|
Chris@10
|
619 }
|
Chris@10
|
620 }
|
Chris@10
|
621 }
|
Chris@10
|
622 }
|
Chris@10
|
623 }
|
Chris@10
|
624 {
|
Chris@10
|
625 E T9X, T9B, Ta2, T9M, T2T, T2Q, TbT, TbH, TbO, TbN, TbU;
|
Chris@10
|
626 {
|
Chris@10
|
627 E Tam, TaA, T9v, T9L;
|
Chris@10
|
628 rio[0] = T7 + Te;
|
Chris@10
|
629 iio[0] = T1t + T1u;
|
Chris@10
|
630 Tam = Tak - Tal;
|
Chris@10
|
631 TaA = Tak + Tal;
|
Chris@10
|
632 T9v = T9r - T9u;
|
Chris@10
|
633 T9L = T9r + T9u;
|
Chris@10
|
634 rio[WS(rs, 1)] = T1D + T1K;
|
Chris@10
|
635 iio[WS(rs, 1)] = T2Z + T30;
|
Chris@10
|
636 T9X = T9A + T9v;
|
Chris@10
|
637 T9B = T9v - T9A;
|
Chris@10
|
638 Ta2 = T9K + T9L;
|
Chris@10
|
639 T9M = T9K - T9L;
|
Chris@10
|
640 rio[WS(rs, 2)] = T39 + T3g;
|
Chris@10
|
641 iio[WS(rs, 2)] = T4v + T4w;
|
Chris@10
|
642 rio[WS(rs, 3)] = T4F + T4M;
|
Chris@10
|
643 iio[WS(rs, 3)] = T61 + T62;
|
Chris@10
|
644 rio[WS(rs, 4)] = T6b + T6i;
|
Chris@10
|
645 iio[WS(rs, 4)] = T7x + T7y;
|
Chris@10
|
646 rio[WS(rs, 5)] = T7H + T7O;
|
Chris@10
|
647 iio[WS(rs, 5)] = T93 + T94;
|
Chris@10
|
648 rio[WS(rs, 6)] = T9d + T9k;
|
Chris@10
|
649 iio[WS(rs, 6)] = Taz + TaA;
|
Chris@10
|
650 rio[WS(rs, 7)] = TaJ + TaQ;
|
Chris@10
|
651 iio[WS(rs, 7)] = Tc5 + Tc6;
|
Chris@10
|
652 {
|
Chris@10
|
653 E T10, T13, T1h, T1a, Tat, Taq, TbC, TbF, TbE, TbG, TbD;
|
Chris@10
|
654 {
|
Chris@10
|
655 E T1q, T1v, T1s, T1w, T1r;
|
Chris@10
|
656 {
|
Chris@10
|
657 E T2N, T2B, T2I, T2H, T2O;
|
Chris@10
|
658 {
|
Chris@10
|
659 E TS, TX, TP, TU, T2G, TY, TT;
|
Chris@10
|
660 T10 = FMA(KP707106781, TR, TQ);
|
Chris@10
|
661 TS = FNMS(KP707106781, TR, TQ);
|
Chris@10
|
662 TX = FNMS(KP707106781, TW, TV);
|
Chris@10
|
663 T13 = FMA(KP707106781, TW, TV);
|
Chris@10
|
664 TP = W[8];
|
Chris@10
|
665 TU = W[9];
|
Chris@10
|
666 T2T = T2J + T2M;
|
Chris@10
|
667 T2N = T2J - T2M;
|
Chris@10
|
668 T2G = T2E - T2F;
|
Chris@10
|
669 T2Q = T2F + T2E;
|
Chris@10
|
670 TY = TP * TX;
|
Chris@10
|
671 TT = TP * TS;
|
Chris@10
|
672 T2B = W[10];
|
Chris@10
|
673 T2I = W[11];
|
Chris@10
|
674 iio[WS(vs, 5)] = FNMS(TU, TS, TY);
|
Chris@10
|
675 rio[WS(vs, 5)] = FMA(TU, TX, TT);
|
Chris@10
|
676 T2H = T2B * T2G;
|
Chris@10
|
677 T2O = T2I * T2G;
|
Chris@10
|
678 }
|
Chris@10
|
679 {
|
Chris@10
|
680 E T1n, T1k, T1j, T1m, T1l, T1o, T1p;
|
Chris@10
|
681 T1h = T1d - T1g;
|
Chris@10
|
682 T1n = T1d + T1g;
|
Chris@10
|
683 T1k = T19 + T18;
|
Chris@10
|
684 T1a = T18 - T19;
|
Chris@10
|
685 iio[WS(vs, 6) + WS(rs, 1)] = FNMS(T2I, T2N, T2H);
|
Chris@10
|
686 rio[WS(vs, 6) + WS(rs, 1)] = FMA(T2B, T2N, T2O);
|
Chris@10
|
687 T1j = W[2];
|
Chris@10
|
688 T1m = W[3];
|
Chris@10
|
689 T1q = T7 - Te;
|
Chris@10
|
690 T1v = T1t - T1u;
|
Chris@10
|
691 T1l = T1j * T1k;
|
Chris@10
|
692 T1o = T1m * T1k;
|
Chris@10
|
693 T1p = W[6];
|
Chris@10
|
694 T1s = W[7];
|
Chris@10
|
695 iio[WS(vs, 2)] = FNMS(T1m, T1n, T1l);
|
Chris@10
|
696 rio[WS(vs, 2)] = FMA(T1j, T1n, T1o);
|
Chris@10
|
697 T1w = T1p * T1v;
|
Chris@10
|
698 T1r = T1p * T1q;
|
Chris@10
|
699 }
|
Chris@10
|
700 }
|
Chris@10
|
701 {
|
Chris@10
|
702 E Tc2, Tc7, Tc4, Tc8, Tc3;
|
Chris@10
|
703 {
|
Chris@10
|
704 E Tan, Tag, Tab, Tai, Tah, Tao, Tc1;
|
Chris@10
|
705 Tat = Taj + Tam;
|
Chris@10
|
706 Tan = Taj - Tam;
|
Chris@10
|
707 Tag = Tae - Taf;
|
Chris@10
|
708 Taq = Taf + Tae;
|
Chris@10
|
709 iio[WS(vs, 4)] = FNMS(T1s, T1q, T1w);
|
Chris@10
|
710 rio[WS(vs, 4)] = FMA(T1s, T1v, T1r);
|
Chris@10
|
711 Tab = W[10];
|
Chris@10
|
712 Tai = W[11];
|
Chris@10
|
713 Tc2 = TaJ - TaQ;
|
Chris@10
|
714 Tc7 = Tc5 - Tc6;
|
Chris@10
|
715 Tah = Tab * Tag;
|
Chris@10
|
716 Tao = Tai * Tag;
|
Chris@10
|
717 Tc1 = W[6];
|
Chris@10
|
718 Tc4 = W[7];
|
Chris@10
|
719 iio[WS(vs, 6) + WS(rs, 6)] = FNMS(Tai, Tan, Tah);
|
Chris@10
|
720 rio[WS(vs, 6) + WS(rs, 6)] = FMA(Tab, Tan, Tao);
|
Chris@10
|
721 Tc8 = Tc1 * Tc7;
|
Chris@10
|
722 Tc3 = Tc1 * Tc2;
|
Chris@10
|
723 }
|
Chris@10
|
724 {
|
Chris@10
|
725 E Tbu, Tbz, Tbr, Tbw, TbA, Tbv, TbB;
|
Chris@10
|
726 TbC = FMA(KP707106781, Tbt, Tbs);
|
Chris@10
|
727 Tbu = FNMS(KP707106781, Tbt, Tbs);
|
Chris@10
|
728 Tbz = FNMS(KP707106781, Tby, Tbx);
|
Chris@10
|
729 TbF = FMA(KP707106781, Tby, Tbx);
|
Chris@10
|
730 iio[WS(vs, 4) + WS(rs, 7)] = FNMS(Tc4, Tc2, Tc8);
|
Chris@10
|
731 rio[WS(vs, 4) + WS(rs, 7)] = FMA(Tc4, Tc7, Tc3);
|
Chris@10
|
732 Tbr = W[8];
|
Chris@10
|
733 Tbw = W[9];
|
Chris@10
|
734 TbA = Tbr * Tbz;
|
Chris@10
|
735 Tbv = Tbr * Tbu;
|
Chris@10
|
736 TbB = W[0];
|
Chris@10
|
737 TbE = W[1];
|
Chris@10
|
738 iio[WS(vs, 5) + WS(rs, 7)] = FNMS(Tbw, Tbu, TbA);
|
Chris@10
|
739 rio[WS(vs, 5) + WS(rs, 7)] = FMA(Tbw, Tbz, Tbv);
|
Chris@10
|
740 TbG = TbB * TbF;
|
Chris@10
|
741 TbD = TbB * TbC;
|
Chris@10
|
742 }
|
Chris@10
|
743 }
|
Chris@10
|
744 }
|
Chris@10
|
745 {
|
Chris@10
|
746 E T2o, T2t, T2q, T2u, T2p;
|
Chris@10
|
747 {
|
Chris@10
|
748 E T2w, T2z, T2y, T2A, T2x;
|
Chris@10
|
749 {
|
Chris@10
|
750 E TZ, T12, T14, T11, T2v;
|
Chris@10
|
751 iio[WS(vs, 1) + WS(rs, 7)] = FNMS(TbE, TbC, TbG);
|
Chris@10
|
752 rio[WS(vs, 1) + WS(rs, 7)] = FMA(TbE, TbF, TbD);
|
Chris@10
|
753 TZ = W[0];
|
Chris@10
|
754 T12 = W[1];
|
Chris@10
|
755 T2o = FNMS(KP707106781, T2n, T2m);
|
Chris@10
|
756 T2w = FMA(KP707106781, T2n, T2m);
|
Chris@10
|
757 T2z = FMA(KP707106781, T2s, T2r);
|
Chris@10
|
758 T2t = FNMS(KP707106781, T2s, T2r);
|
Chris@10
|
759 T14 = TZ * T13;
|
Chris@10
|
760 T11 = TZ * T10;
|
Chris@10
|
761 T2v = W[0];
|
Chris@10
|
762 T2y = W[1];
|
Chris@10
|
763 iio[WS(vs, 1)] = FNMS(T12, T10, T14);
|
Chris@10
|
764 rio[WS(vs, 1)] = FMA(T12, T13, T11);
|
Chris@10
|
765 T2A = T2v * T2z;
|
Chris@10
|
766 T2x = T2v * T2w;
|
Chris@10
|
767 }
|
Chris@10
|
768 {
|
Chris@10
|
769 E T15, T1c, T1b, T1i, T2l;
|
Chris@10
|
770 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T2y, T2w, T2A);
|
Chris@10
|
771 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T2y, T2z, T2x);
|
Chris@10
|
772 T15 = W[10];
|
Chris@10
|
773 T1c = W[11];
|
Chris@10
|
774 T1b = T15 * T1a;
|
Chris@10
|
775 T1i = T1c * T1a;
|
Chris@10
|
776 T2l = W[8];
|
Chris@10
|
777 T2q = W[9];
|
Chris@10
|
778 iio[WS(vs, 6)] = FNMS(T1c, T1h, T1b);
|
Chris@10
|
779 rio[WS(vs, 6)] = FMA(T15, T1h, T1i);
|
Chris@10
|
780 T2u = T2l * T2t;
|
Chris@10
|
781 T2p = T2l * T2o;
|
Chris@10
|
782 }
|
Chris@10
|
783 }
|
Chris@10
|
784 {
|
Chris@10
|
785 E TbZ, TbM, TbV, TbY, TbX, Tc0;
|
Chris@10
|
786 {
|
Chris@10
|
787 E Tap, Tas, TbW, Tar, Tau;
|
Chris@10
|
788 iio[WS(vs, 5) + WS(rs, 1)] = FNMS(T2q, T2o, T2u);
|
Chris@10
|
789 rio[WS(vs, 5) + WS(rs, 1)] = FMA(T2q, T2t, T2p);
|
Chris@10
|
790 Tap = W[2];
|
Chris@10
|
791 Tas = W[3];
|
Chris@10
|
792 TbT = TbP - TbS;
|
Chris@10
|
793 TbZ = TbP + TbS;
|
Chris@10
|
794 TbW = TbL + TbK;
|
Chris@10
|
795 TbM = TbK - TbL;
|
Chris@10
|
796 Tar = Tap * Taq;
|
Chris@10
|
797 Tau = Tas * Taq;
|
Chris@10
|
798 TbV = W[2];
|
Chris@10
|
799 TbY = W[3];
|
Chris@10
|
800 iio[WS(vs, 2) + WS(rs, 6)] = FNMS(Tas, Tat, Tar);
|
Chris@10
|
801 rio[WS(vs, 2) + WS(rs, 6)] = FMA(Tap, Tat, Tau);
|
Chris@10
|
802 TbX = TbV * TbW;
|
Chris@10
|
803 Tc0 = TbY * TbW;
|
Chris@10
|
804 }
|
Chris@10
|
805 {
|
Chris@10
|
806 E Taw, TaB, Tav, Tay, TaC, Tax;
|
Chris@10
|
807 Taw = T9d - T9k;
|
Chris@10
|
808 TaB = Taz - TaA;
|
Chris@10
|
809 iio[WS(vs, 2) + WS(rs, 7)] = FNMS(TbY, TbZ, TbX);
|
Chris@10
|
810 rio[WS(vs, 2) + WS(rs, 7)] = FMA(TbV, TbZ, Tc0);
|
Chris@10
|
811 Tav = W[6];
|
Chris@10
|
812 Tay = W[7];
|
Chris@10
|
813 TaC = Tav * TaB;
|
Chris@10
|
814 Tax = Tav * Taw;
|
Chris@10
|
815 TbH = W[10];
|
Chris@10
|
816 TbO = W[11];
|
Chris@10
|
817 iio[WS(vs, 4) + WS(rs, 6)] = FNMS(Tay, Taw, TaC);
|
Chris@10
|
818 rio[WS(vs, 4) + WS(rs, 6)] = FMA(Tay, TaB, Tax);
|
Chris@10
|
819 TbN = TbH * TbM;
|
Chris@10
|
820 TbU = TbO * TbM;
|
Chris@10
|
821 }
|
Chris@10
|
822 }
|
Chris@10
|
823 }
|
Chris@10
|
824 }
|
Chris@10
|
825 }
|
Chris@10
|
826 {
|
Chris@10
|
827 E T5q, T5v, T8R, T8K, T90, T95, T92, T96, T91;
|
Chris@10
|
828 {
|
Chris@10
|
829 E T3U, T3Z, T74, T77, T9Y, Ta3, T7l, T7e, T8X, T8T, T8W, T8V, T8Y;
|
Chris@10
|
830 {
|
Chris@10
|
831 E T5y, T5B, T5A, T5C, T5z;
|
Chris@10
|
832 {
|
Chris@10
|
833 E T5Y, T63, T60, T64, T5Z;
|
Chris@10
|
834 {
|
Chris@10
|
835 E T2P, T2S, T2R, T2U, T5X;
|
Chris@10
|
836 iio[WS(vs, 6) + WS(rs, 7)] = FNMS(TbO, TbT, TbN);
|
Chris@10
|
837 rio[WS(vs, 6) + WS(rs, 7)] = FMA(TbH, TbT, TbU);
|
Chris@10
|
838 T2P = W[2];
|
Chris@10
|
839 T2S = W[3];
|
Chris@10
|
840 T5Y = T4F - T4M;
|
Chris@10
|
841 T63 = T61 - T62;
|
Chris@10
|
842 T2R = T2P * T2Q;
|
Chris@10
|
843 T2U = T2S * T2Q;
|
Chris@10
|
844 T5X = W[6];
|
Chris@10
|
845 T60 = W[7];
|
Chris@10
|
846 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T2S, T2T, T2R);
|
Chris@10
|
847 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T2P, T2T, T2U);
|
Chris@10
|
848 T64 = T5X * T63;
|
Chris@10
|
849 T5Z = T5X * T5Y;
|
Chris@10
|
850 }
|
Chris@10
|
851 {
|
Chris@10
|
852 E T42, T45, T41, T44, T46, T43, T5x;
|
Chris@10
|
853 T3U = FNMS(KP707106781, T3T, T3S);
|
Chris@10
|
854 T42 = FMA(KP707106781, T3T, T3S);
|
Chris@10
|
855 T45 = FMA(KP707106781, T3Y, T3X);
|
Chris@10
|
856 T3Z = FNMS(KP707106781, T3Y, T3X);
|
Chris@10
|
857 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T60, T5Y, T64);
|
Chris@10
|
858 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T60, T63, T5Z);
|
Chris@10
|
859 T41 = W[0];
|
Chris@10
|
860 T44 = W[1];
|
Chris@10
|
861 T5q = FNMS(KP707106781, T5p, T5o);
|
Chris@10
|
862 T5y = FMA(KP707106781, T5p, T5o);
|
Chris@10
|
863 T5B = FMA(KP707106781, T5u, T5t);
|
Chris@10
|
864 T5v = FNMS(KP707106781, T5u, T5t);
|
Chris@10
|
865 T46 = T41 * T45;
|
Chris@10
|
866 T43 = T41 * T42;
|
Chris@10
|
867 T5x = W[0];
|
Chris@10
|
868 T5A = W[1];
|
Chris@10
|
869 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T44, T42, T46);
|
Chris@10
|
870 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T44, T45, T43);
|
Chris@10
|
871 T5C = T5x * T5B;
|
Chris@10
|
872 T5z = T5x * T5y;
|
Chris@10
|
873 }
|
Chris@10
|
874 }
|
Chris@10
|
875 {
|
Chris@10
|
876 E Ta6, Ta9, Ta8, Taa, Ta7;
|
Chris@10
|
877 {
|
Chris@10
|
878 E T6W, T71, T6T, T6Y, T72, T6X, Ta5;
|
Chris@10
|
879 T74 = FMA(KP707106781, T6V, T6U);
|
Chris@10
|
880 T6W = FNMS(KP707106781, T6V, T6U);
|
Chris@10
|
881 T71 = FNMS(KP707106781, T70, T6Z);
|
Chris@10
|
882 T77 = FMA(KP707106781, T70, T6Z);
|
Chris@10
|
883 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T5A, T5y, T5C);
|
Chris@10
|
884 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T5A, T5B, T5z);
|
Chris@10
|
885 T6T = W[8];
|
Chris@10
|
886 T6Y = W[9];
|
Chris@10
|
887 T9Y = FNMS(KP707106781, T9X, T9W);
|
Chris@10
|
888 Ta6 = FMA(KP707106781, T9X, T9W);
|
Chris@10
|
889 Ta9 = FMA(KP707106781, Ta2, Ta1);
|
Chris@10
|
890 Ta3 = FNMS(KP707106781, Ta2, Ta1);
|
Chris@10
|
891 T72 = T6T * T71;
|
Chris@10
|
892 T6X = T6T * T6W;
|
Chris@10
|
893 Ta5 = W[0];
|
Chris@10
|
894 Ta8 = W[1];
|
Chris@10
|
895 iio[WS(vs, 5) + WS(rs, 4)] = FNMS(T6Y, T6W, T72);
|
Chris@10
|
896 rio[WS(vs, 5) + WS(rs, 4)] = FMA(T6Y, T71, T6X);
|
Chris@10
|
897 Taa = Ta5 * Ta9;
|
Chris@10
|
898 Ta7 = Ta5 * Ta6;
|
Chris@10
|
899 }
|
Chris@10
|
900 {
|
Chris@10
|
901 E T7r, T7o, T7n, T7q, T8U, T7p, T7s;
|
Chris@10
|
902 T7l = T7h - T7k;
|
Chris@10
|
903 T7r = T7h + T7k;
|
Chris@10
|
904 T7o = T7d + T7c;
|
Chris@10
|
905 T7e = T7c - T7d;
|
Chris@10
|
906 iio[WS(vs, 1) + WS(rs, 6)] = FNMS(Ta8, Ta6, Taa);
|
Chris@10
|
907 rio[WS(vs, 1) + WS(rs, 6)] = FMA(Ta8, Ta9, Ta7);
|
Chris@10
|
908 T7n = W[2];
|
Chris@10
|
909 T7q = W[3];
|
Chris@10
|
910 T8R = T8N - T8Q;
|
Chris@10
|
911 T8X = T8N + T8Q;
|
Chris@10
|
912 T8U = T8J + T8I;
|
Chris@10
|
913 T8K = T8I - T8J;
|
Chris@10
|
914 T7p = T7n * T7o;
|
Chris@10
|
915 T7s = T7q * T7o;
|
Chris@10
|
916 T8T = W[2];
|
Chris@10
|
917 T8W = W[3];
|
Chris@10
|
918 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T7q, T7r, T7p);
|
Chris@10
|
919 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T7n, T7r, T7s);
|
Chris@10
|
920 T8V = T8T * T8U;
|
Chris@10
|
921 T8Y = T8W * T8U;
|
Chris@10
|
922 }
|
Chris@10
|
923 }
|
Chris@10
|
924 }
|
Chris@10
|
925 {
|
Chris@10
|
926 E T5P, T5D, T5K, T5J, T5Q, Ta0, Ta4, T9Z;
|
Chris@10
|
927 {
|
Chris@10
|
928 E T5V, T5I, T5R, T5U, T5T, T5W;
|
Chris@10
|
929 {
|
Chris@10
|
930 E T2W, T31, T2V, T2Y, T5S, T32, T2X;
|
Chris@10
|
931 T2W = T1D - T1K;
|
Chris@10
|
932 T31 = T2Z - T30;
|
Chris@10
|
933 iio[WS(vs, 2) + WS(rs, 5)] = FNMS(T8W, T8X, T8V);
|
Chris@10
|
934 rio[WS(vs, 2) + WS(rs, 5)] = FMA(T8T, T8X, T8Y);
|
Chris@10
|
935 T2V = W[6];
|
Chris@10
|
936 T2Y = W[7];
|
Chris@10
|
937 T5P = T5L - T5O;
|
Chris@10
|
938 T5V = T5L + T5O;
|
Chris@10
|
939 T5S = T5H + T5G;
|
Chris@10
|
940 T5I = T5G - T5H;
|
Chris@10
|
941 T32 = T2V * T31;
|
Chris@10
|
942 T2X = T2V * T2W;
|
Chris@10
|
943 T5R = W[2];
|
Chris@10
|
944 T5U = W[3];
|
Chris@10
|
945 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T2Y, T2W, T32);
|
Chris@10
|
946 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T2Y, T31, T2X);
|
Chris@10
|
947 T5T = T5R * T5S;
|
Chris@10
|
948 T5W = T5U * T5S;
|
Chris@10
|
949 }
|
Chris@10
|
950 {
|
Chris@10
|
951 E T3R, T3W, T40, T3V;
|
Chris@10
|
952 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T5U, T5V, T5T);
|
Chris@10
|
953 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T5R, T5V, T5W);
|
Chris@10
|
954 T3R = W[8];
|
Chris@10
|
955 T3W = W[9];
|
Chris@10
|
956 T40 = T3R * T3Z;
|
Chris@10
|
957 T3V = T3R * T3U;
|
Chris@10
|
958 T5D = W[10];
|
Chris@10
|
959 T5K = W[11];
|
Chris@10
|
960 iio[WS(vs, 5) + WS(rs, 2)] = FNMS(T3W, T3U, T40);
|
Chris@10
|
961 rio[WS(vs, 5) + WS(rs, 2)] = FMA(T3W, T3Z, T3V);
|
Chris@10
|
962 T5J = T5D * T5I;
|
Chris@10
|
963 T5Q = T5K * T5I;
|
Chris@10
|
964 }
|
Chris@10
|
965 }
|
Chris@10
|
966 {
|
Chris@10
|
967 E T73, T76, T78, T75, T9V;
|
Chris@10
|
968 iio[WS(vs, 6) + WS(rs, 3)] = FNMS(T5K, T5P, T5J);
|
Chris@10
|
969 rio[WS(vs, 6) + WS(rs, 3)] = FMA(T5D, T5P, T5Q);
|
Chris@10
|
970 T73 = W[0];
|
Chris@10
|
971 T76 = W[1];
|
Chris@10
|
972 T78 = T73 * T77;
|
Chris@10
|
973 T75 = T73 * T74;
|
Chris@10
|
974 T9V = W[8];
|
Chris@10
|
975 Ta0 = W[9];
|
Chris@10
|
976 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T76, T74, T78);
|
Chris@10
|
977 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T76, T77, T75);
|
Chris@10
|
978 Ta4 = T9V * Ta3;
|
Chris@10
|
979 T9Z = T9V * T9Y;
|
Chris@10
|
980 }
|
Chris@10
|
981 {
|
Chris@10
|
982 E T79, T7g, T7f, T7m, T8Z;
|
Chris@10
|
983 iio[WS(vs, 5) + WS(rs, 6)] = FNMS(Ta0, T9Y, Ta4);
|
Chris@10
|
984 rio[WS(vs, 5) + WS(rs, 6)] = FMA(Ta0, Ta3, T9Z);
|
Chris@10
|
985 T79 = W[10];
|
Chris@10
|
986 T7g = W[11];
|
Chris@10
|
987 T90 = T7H - T7O;
|
Chris@10
|
988 T95 = T93 - T94;
|
Chris@10
|
989 T7f = T79 * T7e;
|
Chris@10
|
990 T7m = T7g * T7e;
|
Chris@10
|
991 T8Z = W[6];
|
Chris@10
|
992 T92 = W[7];
|
Chris@10
|
993 iio[WS(vs, 6) + WS(rs, 4)] = FNMS(T7g, T7l, T7f);
|
Chris@10
|
994 rio[WS(vs, 6) + WS(rs, 4)] = FMA(T79, T7l, T7m);
|
Chris@10
|
995 T96 = T8Z * T95;
|
Chris@10
|
996 T91 = T8Z * T90;
|
Chris@10
|
997 }
|
Chris@10
|
998 }
|
Chris@10
|
999 }
|
Chris@10
|
1000 {
|
Chris@10
|
1001 E T8A, T8D, T8C, T8E, T8B;
|
Chris@10
|
1002 {
|
Chris@10
|
1003 E T4s, T4x, T4u, T4y, T4t;
|
Chris@10
|
1004 {
|
Chris@10
|
1005 E T4p, T4m, T5s, T5w, T5r;
|
Chris@10
|
1006 {
|
Chris@10
|
1007 E T4j, T4c, T47, T4e, T4d, T4k, T5n;
|
Chris@10
|
1008 T4p = T4f + T4i;
|
Chris@10
|
1009 T4j = T4f - T4i;
|
Chris@10
|
1010 T4c = T4a - T4b;
|
Chris@10
|
1011 T4m = T4b + T4a;
|
Chris@10
|
1012 iio[WS(vs, 4) + WS(rs, 5)] = FNMS(T92, T90, T96);
|
Chris@10
|
1013 rio[WS(vs, 4) + WS(rs, 5)] = FMA(T92, T95, T91);
|
Chris@10
|
1014 T47 = W[10];
|
Chris@10
|
1015 T4e = W[11];
|
Chris@10
|
1016 T4d = T47 * T4c;
|
Chris@10
|
1017 T4k = T4e * T4c;
|
Chris@10
|
1018 T5n = W[8];
|
Chris@10
|
1019 T5s = W[9];
|
Chris@10
|
1020 iio[WS(vs, 6) + WS(rs, 2)] = FNMS(T4e, T4j, T4d);
|
Chris@10
|
1021 rio[WS(vs, 6) + WS(rs, 2)] = FMA(T47, T4j, T4k);
|
Chris@10
|
1022 T5w = T5n * T5v;
|
Chris@10
|
1023 T5r = T5n * T5q;
|
Chris@10
|
1024 }
|
Chris@10
|
1025 {
|
Chris@10
|
1026 E T4l, T4o, T4n, T4q, T4r;
|
Chris@10
|
1027 iio[WS(vs, 5) + WS(rs, 3)] = FNMS(T5s, T5q, T5w);
|
Chris@10
|
1028 rio[WS(vs, 5) + WS(rs, 3)] = FMA(T5s, T5v, T5r);
|
Chris@10
|
1029 T4l = W[2];
|
Chris@10
|
1030 T4o = W[3];
|
Chris@10
|
1031 T4s = T39 - T3g;
|
Chris@10
|
1032 T4x = T4v - T4w;
|
Chris@10
|
1033 T4n = T4l * T4m;
|
Chris@10
|
1034 T4q = T4o * T4m;
|
Chris@10
|
1035 T4r = W[6];
|
Chris@10
|
1036 T4u = W[7];
|
Chris@10
|
1037 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T4o, T4p, T4n);
|
Chris@10
|
1038 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T4l, T4p, T4q);
|
Chris@10
|
1039 T4y = T4r * T4x;
|
Chris@10
|
1040 T4t = T4r * T4s;
|
Chris@10
|
1041 }
|
Chris@10
|
1042 }
|
Chris@10
|
1043 {
|
Chris@10
|
1044 E T8F, T8M, T8L, T8S;
|
Chris@10
|
1045 {
|
Chris@10
|
1046 E T7u, T7z, T7t, T7w, T7A, T7v;
|
Chris@10
|
1047 T7u = T6b - T6i;
|
Chris@10
|
1048 T7z = T7x - T7y;
|
Chris@10
|
1049 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T4u, T4s, T4y);
|
Chris@10
|
1050 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T4u, T4x, T4t);
|
Chris@10
|
1051 T7t = W[6];
|
Chris@10
|
1052 T7w = W[7];
|
Chris@10
|
1053 T7A = T7t * T7z;
|
Chris@10
|
1054 T7v = T7t * T7u;
|
Chris@10
|
1055 T8F = W[10];
|
Chris@10
|
1056 T8M = W[11];
|
Chris@10
|
1057 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T7w, T7u, T7A);
|
Chris@10
|
1058 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T7w, T7z, T7v);
|
Chris@10
|
1059 T8L = T8F * T8K;
|
Chris@10
|
1060 T8S = T8M * T8K;
|
Chris@10
|
1061 }
|
Chris@10
|
1062 {
|
Chris@10
|
1063 E T8s, T8x, T8p, T8u, T8y, T8t, T8z;
|
Chris@10
|
1064 T8A = FMA(KP707106781, T8r, T8q);
|
Chris@10
|
1065 T8s = FNMS(KP707106781, T8r, T8q);
|
Chris@10
|
1066 T8x = FNMS(KP707106781, T8w, T8v);
|
Chris@10
|
1067 T8D = FMA(KP707106781, T8w, T8v);
|
Chris@10
|
1068 iio[WS(vs, 6) + WS(rs, 5)] = FNMS(T8M, T8R, T8L);
|
Chris@10
|
1069 rio[WS(vs, 6) + WS(rs, 5)] = FMA(T8F, T8R, T8S);
|
Chris@10
|
1070 T8p = W[8];
|
Chris@10
|
1071 T8u = W[9];
|
Chris@10
|
1072 T8y = T8p * T8x;
|
Chris@10
|
1073 T8t = T8p * T8s;
|
Chris@10
|
1074 T8z = W[0];
|
Chris@10
|
1075 T8C = W[1];
|
Chris@10
|
1076 iio[WS(vs, 5) + WS(rs, 5)] = FNMS(T8u, T8s, T8y);
|
Chris@10
|
1077 rio[WS(vs, 5) + WS(rs, 5)] = FMA(T8u, T8x, T8t);
|
Chris@10
|
1078 T8E = T8z * T8D;
|
Chris@10
|
1079 T8B = T8z * T8A;
|
Chris@10
|
1080 }
|
Chris@10
|
1081 }
|
Chris@10
|
1082 }
|
Chris@10
|
1083 {
|
Chris@10
|
1084 E T3y, T3J, T3h, T3A, T3z, T3K;
|
Chris@10
|
1085 {
|
Chris@10
|
1086 E T54, T5f, T4N, T56, T55, T5g;
|
Chris@10
|
1087 {
|
Chris@10
|
1088 E Tw, TH, Tf, Ty, Tx, TI;
|
Chris@10
|
1089 {
|
Chris@10
|
1090 E TN, TJ, TM, TL, TO, TK;
|
Chris@10
|
1091 TK = FMA(KP707106781, Tv, Tk);
|
Chris@10
|
1092 Tw = FNMS(KP707106781, Tv, Tk);
|
Chris@10
|
1093 iio[WS(vs, 1) + WS(rs, 5)] = FNMS(T8C, T8A, T8E);
|
Chris@10
|
1094 rio[WS(vs, 1) + WS(rs, 5)] = FMA(T8C, T8D, T8B);
|
Chris@10
|
1095 TH = FNMS(KP707106781, TG, TD);
|
Chris@10
|
1096 TN = FMA(KP707106781, TG, TD);
|
Chris@10
|
1097 TJ = W[4];
|
Chris@10
|
1098 TM = W[5];
|
Chris@10
|
1099 Tf = W[12];
|
Chris@10
|
1100 TL = TJ * TK;
|
Chris@10
|
1101 TO = TM * TK;
|
Chris@10
|
1102 Ty = W[13];
|
Chris@10
|
1103 Tx = Tf * Tw;
|
Chris@10
|
1104 iio[WS(vs, 3)] = FNMS(TM, TN, TL);
|
Chris@10
|
1105 rio[WS(vs, 3)] = FMA(TJ, TN, TO);
|
Chris@10
|
1106 }
|
Chris@10
|
1107 TI = Ty * Tw;
|
Chris@10
|
1108 iio[WS(vs, 7)] = FNMS(Ty, TH, Tx);
|
Chris@10
|
1109 {
|
Chris@10
|
1110 E T5h, T5l, T5k, T5j, T5m, T5i;
|
Chris@10
|
1111 T5i = FMA(KP707106781, T53, T4S);
|
Chris@10
|
1112 T54 = FNMS(KP707106781, T53, T4S);
|
Chris@10
|
1113 rio[WS(vs, 7)] = FMA(Tf, TH, TI);
|
Chris@10
|
1114 T5h = W[4];
|
Chris@10
|
1115 T5f = FNMS(KP707106781, T5e, T5b);
|
Chris@10
|
1116 T5l = FMA(KP707106781, T5e, T5b);
|
Chris@10
|
1117 T5k = W[5];
|
Chris@10
|
1118 T5j = T5h * T5i;
|
Chris@10
|
1119 T4N = W[12];
|
Chris@10
|
1120 T5m = T5k * T5i;
|
Chris@10
|
1121 T56 = W[13];
|
Chris@10
|
1122 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T5k, T5l, T5j);
|
Chris@10
|
1123 T55 = T4N * T54;
|
Chris@10
|
1124 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T5h, T5l, T5m);
|
Chris@10
|
1125 }
|
Chris@10
|
1126 }
|
Chris@10
|
1127 T5g = T56 * T54;
|
Chris@10
|
1128 {
|
Chris@10
|
1129 E T22, T2d, T1L, T24, T23, T2e;
|
Chris@10
|
1130 {
|
Chris@10
|
1131 E T2j, T2f, T2i, T2h, T2k, T2g;
|
Chris@10
|
1132 iio[WS(vs, 7) + WS(rs, 3)] = FNMS(T56, T5f, T55);
|
Chris@10
|
1133 T22 = FNMS(KP707106781, T21, T1Q);
|
Chris@10
|
1134 T2g = FMA(KP707106781, T21, T1Q);
|
Chris@10
|
1135 rio[WS(vs, 7) + WS(rs, 3)] = FMA(T4N, T5f, T5g);
|
Chris@10
|
1136 T2d = FNMS(KP707106781, T2c, T29);
|
Chris@10
|
1137 T2j = FMA(KP707106781, T2c, T29);
|
Chris@10
|
1138 T2f = W[4];
|
Chris@10
|
1139 T2i = W[5];
|
Chris@10
|
1140 T1L = W[12];
|
Chris@10
|
1141 T2h = T2f * T2g;
|
Chris@10
|
1142 T2k = T2i * T2g;
|
Chris@10
|
1143 T24 = W[13];
|
Chris@10
|
1144 T23 = T1L * T22;
|
Chris@10
|
1145 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T2i, T2j, T2h);
|
Chris@10
|
1146 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T2f, T2j, T2k);
|
Chris@10
|
1147 }
|
Chris@10
|
1148 T2e = T24 * T22;
|
Chris@10
|
1149 iio[WS(vs, 7) + WS(rs, 1)] = FNMS(T24, T2d, T23);
|
Chris@10
|
1150 {
|
Chris@10
|
1151 E T3L, T3P, T3O, T3N, T3Q, T3M;
|
Chris@10
|
1152 T3M = FMA(KP707106781, T3x, T3m);
|
Chris@10
|
1153 T3y = FNMS(KP707106781, T3x, T3m);
|
Chris@10
|
1154 rio[WS(vs, 7) + WS(rs, 1)] = FMA(T1L, T2d, T2e);
|
Chris@10
|
1155 T3L = W[4];
|
Chris@10
|
1156 T3J = FNMS(KP707106781, T3I, T3F);
|
Chris@10
|
1157 T3P = FMA(KP707106781, T3I, T3F);
|
Chris@10
|
1158 T3O = W[5];
|
Chris@10
|
1159 T3N = T3L * T3M;
|
Chris@10
|
1160 T3h = W[12];
|
Chris@10
|
1161 T3Q = T3O * T3M;
|
Chris@10
|
1162 T3A = W[13];
|
Chris@10
|
1163 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T3O, T3P, T3N);
|
Chris@10
|
1164 T3z = T3h * T3y;
|
Chris@10
|
1165 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T3L, T3P, T3Q);
|
Chris@10
|
1166 }
|
Chris@10
|
1167 }
|
Chris@10
|
1168 }
|
Chris@10
|
1169 T3K = T3A * T3y;
|
Chris@10
|
1170 {
|
Chris@10
|
1171 E Tb8, Tbj, TaR, Tba, Tb9, Tbk;
|
Chris@10
|
1172 {
|
Chris@10
|
1173 E T6A, T6L, T6j, T6C, T6B, T6M;
|
Chris@10
|
1174 {
|
Chris@10
|
1175 E T6R, T6N, T6Q, T6P, T6S, T6O;
|
Chris@10
|
1176 iio[WS(vs, 7) + WS(rs, 2)] = FNMS(T3A, T3J, T3z);
|
Chris@10
|
1177 T6A = FNMS(KP707106781, T6z, T6o);
|
Chris@10
|
1178 T6O = FMA(KP707106781, T6z, T6o);
|
Chris@10
|
1179 rio[WS(vs, 7) + WS(rs, 2)] = FMA(T3h, T3J, T3K);
|
Chris@10
|
1180 T6L = FNMS(KP707106781, T6K, T6H);
|
Chris@10
|
1181 T6R = FMA(KP707106781, T6K, T6H);
|
Chris@10
|
1182 T6N = W[4];
|
Chris@10
|
1183 T6Q = W[5];
|
Chris@10
|
1184 T6j = W[12];
|
Chris@10
|
1185 T6P = T6N * T6O;
|
Chris@10
|
1186 T6S = T6Q * T6O;
|
Chris@10
|
1187 T6C = W[13];
|
Chris@10
|
1188 T6B = T6j * T6A;
|
Chris@10
|
1189 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T6Q, T6R, T6P);
|
Chris@10
|
1190 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T6N, T6R, T6S);
|
Chris@10
|
1191 }
|
Chris@10
|
1192 T6M = T6C * T6A;
|
Chris@10
|
1193 iio[WS(vs, 7) + WS(rs, 4)] = FNMS(T6C, T6L, T6B);
|
Chris@10
|
1194 {
|
Chris@10
|
1195 E Tbl, Tbp, Tbo, Tbn, Tbq, Tbm;
|
Chris@10
|
1196 Tbm = FMA(KP707106781, Tb7, TaW);
|
Chris@10
|
1197 Tb8 = FNMS(KP707106781, Tb7, TaW);
|
Chris@10
|
1198 rio[WS(vs, 7) + WS(rs, 4)] = FMA(T6j, T6L, T6M);
|
Chris@10
|
1199 Tbl = W[4];
|
Chris@10
|
1200 Tbj = FNMS(KP707106781, Tbi, Tbf);
|
Chris@10
|
1201 Tbp = FMA(KP707106781, Tbi, Tbf);
|
Chris@10
|
1202 Tbo = W[5];
|
Chris@10
|
1203 Tbn = Tbl * Tbm;
|
Chris@10
|
1204 TaR = W[12];
|
Chris@10
|
1205 Tbq = Tbo * Tbm;
|
Chris@10
|
1206 Tba = W[13];
|
Chris@10
|
1207 iio[WS(vs, 3) + WS(rs, 7)] = FNMS(Tbo, Tbp, Tbn);
|
Chris@10
|
1208 Tb9 = TaR * Tb8;
|
Chris@10
|
1209 rio[WS(vs, 3) + WS(rs, 7)] = FMA(Tbl, Tbp, Tbq);
|
Chris@10
|
1210 }
|
Chris@10
|
1211 }
|
Chris@10
|
1212 Tbk = Tba * Tb8;
|
Chris@10
|
1213 {
|
Chris@10
|
1214 E T86, T8h, T7P, T88, T87, T8i;
|
Chris@10
|
1215 {
|
Chris@10
|
1216 E T8n, T8j, T8m, T8l, T8o, T8k;
|
Chris@10
|
1217 iio[WS(vs, 7) + WS(rs, 7)] = FNMS(Tba, Tbj, Tb9);
|
Chris@10
|
1218 T86 = FNMS(KP707106781, T85, T7U);
|
Chris@10
|
1219 T8k = FMA(KP707106781, T85, T7U);
|
Chris@10
|
1220 rio[WS(vs, 7) + WS(rs, 7)] = FMA(TaR, Tbj, Tbk);
|
Chris@10
|
1221 T8h = FNMS(KP707106781, T8g, T8d);
|
Chris@10
|
1222 T8n = FMA(KP707106781, T8g, T8d);
|
Chris@10
|
1223 T8j = W[4];
|
Chris@10
|
1224 T8m = W[5];
|
Chris@10
|
1225 T7P = W[12];
|
Chris@10
|
1226 T8l = T8j * T8k;
|
Chris@10
|
1227 T8o = T8m * T8k;
|
Chris@10
|
1228 T88 = W[13];
|
Chris@10
|
1229 T87 = T7P * T86;
|
Chris@10
|
1230 iio[WS(vs, 3) + WS(rs, 5)] = FNMS(T8m, T8n, T8l);
|
Chris@10
|
1231 rio[WS(vs, 3) + WS(rs, 5)] = FMA(T8j, T8n, T8o);
|
Chris@10
|
1232 }
|
Chris@10
|
1233 T8i = T88 * T86;
|
Chris@10
|
1234 iio[WS(vs, 7) + WS(rs, 5)] = FNMS(T88, T8h, T87);
|
Chris@10
|
1235 {
|
Chris@10
|
1236 E T9P, T9T, T9S, T9R, T9U, T9Q;
|
Chris@10
|
1237 T9Q = FMA(KP707106781, T9B, T9q);
|
Chris@10
|
1238 T9C = FNMS(KP707106781, T9B, T9q);
|
Chris@10
|
1239 rio[WS(vs, 7) + WS(rs, 5)] = FMA(T7P, T8h, T8i);
|
Chris@10
|
1240 T9P = W[4];
|
Chris@10
|
1241 T9N = FNMS(KP707106781, T9M, T9J);
|
Chris@10
|
1242 T9T = FMA(KP707106781, T9M, T9J);
|
Chris@10
|
1243 T9S = W[5];
|
Chris@10
|
1244 T9R = T9P * T9Q;
|
Chris@10
|
1245 T9l = W[12];
|
Chris@10
|
1246 T9U = T9S * T9Q;
|
Chris@10
|
1247 T9E = W[13];
|
Chris@10
|
1248 iio[WS(vs, 3) + WS(rs, 6)] = FNMS(T9S, T9T, T9R);
|
Chris@10
|
1249 T9D = T9l * T9C;
|
Chris@10
|
1250 rio[WS(vs, 3) + WS(rs, 6)] = FMA(T9P, T9T, T9U);
|
Chris@10
|
1251 }
|
Chris@10
|
1252 }
|
Chris@10
|
1253 }
|
Chris@10
|
1254 }
|
Chris@10
|
1255 }
|
Chris@10
|
1256 }
|
Chris@10
|
1257 }
|
Chris@10
|
1258 }
|
Chris@10
|
1259 T9O = T9E * T9C;
|
Chris@10
|
1260 iio[WS(vs, 7) + WS(rs, 6)] = FNMS(T9E, T9N, T9D);
|
Chris@10
|
1261 rio[WS(vs, 7) + WS(rs, 6)] = FMA(T9l, T9N, T9O);
|
Chris@10
|
1262 }
|
Chris@10
|
1263 }
|
Chris@10
|
1264 }
|
Chris@10
|
1265
|
Chris@10
|
1266 static const tw_instr twinstr[] = {
|
Chris@10
|
1267 {TW_FULL, 0, 8},
|
Chris@10
|
1268 {TW_NEXT, 1, 0}
|
Chris@10
|
1269 };
|
Chris@10
|
1270
|
Chris@10
|
1271 static const ct_desc desc = { 8, "q1_8", twinstr, &GENUS, {352, 112, 176, 0}, 0, 0, 0 };
|
Chris@10
|
1272
|
Chris@10
|
1273 void X(codelet_q1_8) (planner *p) {
|
Chris@10
|
1274 X(kdft_difsq_register) (p, q1_8, &desc);
|
Chris@10
|
1275 }
|
Chris@10
|
1276 #else /* HAVE_FMA */
|
Chris@10
|
1277
|
Chris@10
|
1278 /* Generated by: ../../../genfft/gen_twidsq.native -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 8 -name q1_8 -include q.h */
|
Chris@10
|
1279
|
Chris@10
|
1280 /*
|
Chris@10
|
1281 * This function contains 528 FP additions, 256 FP multiplications,
|
Chris@10
|
1282 * (or, 416 additions, 144 multiplications, 112 fused multiply/add),
|
Chris@10
|
1283 * 142 stack variables, 1 constants, and 256 memory accesses
|
Chris@10
|
1284 */
|
Chris@10
|
1285 #include "q.h"
|
Chris@10
|
1286
|
Chris@10
|
1287 static void q1_8(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
|
Chris@10
|
1288 {
|
Chris@10
|
1289 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
1290 {
|
Chris@10
|
1291 INT m;
|
Chris@10
|
1292 for (m = mb, W = W + (mb * 14); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 14, MAKE_VOLATILE_STRIDE(16, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
|
Chris@10
|
1293 E T7, T14, T1g, Tk, TC, TQ, T10, TM, T1w, T2p, T2z, T1H, T1M, T1W, T2j;
|
Chris@10
|
1294 E T1V, T7R, T8O, T90, T84, T8m, T8A, T8K, T8w, T9g, Ta9, Taj, T9r, T9w, T9G;
|
Chris@10
|
1295 E Ta3, T9F, Te, T17, T1h, Tp, Tu, TE, T11, TD, T1p, T2m, T2y, T1C, T1U;
|
Chris@10
|
1296 E T28, T2i, T24, T7Y, T8R, T91, T89, T8e, T8o, T8L, T8n, T99, Ta6, Tai, T9m;
|
Chris@10
|
1297 E T9E, T9S, Ta2, T9O, T2H, T3E, T3Q, T2U, T3c, T3q, T3A, T3m, T46, T4Z, T59;
|
Chris@10
|
1298 E T4h, T4m, T4w, T4T, T4v, T5h, T6e, T6q, T5u, T5M, T60, T6a, T5W, T6G, T7z;
|
Chris@10
|
1299 E T7J, T6R, T6W, T76, T7t, T75, T2O, T3H, T3R, T2Z, T34, T3e, T3B, T3d, T3Z;
|
Chris@10
|
1300 E T4W, T58, T4c, T4u, T4I, T4S, T4E, T5o, T6h, T6r, T5z, T5E, T5O, T6b, T5N;
|
Chris@10
|
1301 E T6z, T7w, T7I, T6M, T74, T7i, T7s, T7e;
|
Chris@10
|
1302 {
|
Chris@10
|
1303 E T3, Ty, Tj, TY, T6, Tg, TB, TZ;
|
Chris@10
|
1304 {
|
Chris@10
|
1305 E T1, T2, Th, Ti;
|
Chris@10
|
1306 T1 = rio[0];
|
Chris@10
|
1307 T2 = rio[WS(rs, 4)];
|
Chris@10
|
1308 T3 = T1 + T2;
|
Chris@10
|
1309 Ty = T1 - T2;
|
Chris@10
|
1310 Th = iio[0];
|
Chris@10
|
1311 Ti = iio[WS(rs, 4)];
|
Chris@10
|
1312 Tj = Th - Ti;
|
Chris@10
|
1313 TY = Th + Ti;
|
Chris@10
|
1314 }
|
Chris@10
|
1315 {
|
Chris@10
|
1316 E T4, T5, Tz, TA;
|
Chris@10
|
1317 T4 = rio[WS(rs, 2)];
|
Chris@10
|
1318 T5 = rio[WS(rs, 6)];
|
Chris@10
|
1319 T6 = T4 + T5;
|
Chris@10
|
1320 Tg = T4 - T5;
|
Chris@10
|
1321 Tz = iio[WS(rs, 2)];
|
Chris@10
|
1322 TA = iio[WS(rs, 6)];
|
Chris@10
|
1323 TB = Tz - TA;
|
Chris@10
|
1324 TZ = Tz + TA;
|
Chris@10
|
1325 }
|
Chris@10
|
1326 T7 = T3 + T6;
|
Chris@10
|
1327 T14 = T3 - T6;
|
Chris@10
|
1328 T1g = TY + TZ;
|
Chris@10
|
1329 Tk = Tg + Tj;
|
Chris@10
|
1330 TC = Ty - TB;
|
Chris@10
|
1331 TQ = Tj - Tg;
|
Chris@10
|
1332 T10 = TY - TZ;
|
Chris@10
|
1333 TM = Ty + TB;
|
Chris@10
|
1334 }
|
Chris@10
|
1335 {
|
Chris@10
|
1336 E T1s, T1I, T1L, T2n, T1v, T1D, T1G, T2o;
|
Chris@10
|
1337 {
|
Chris@10
|
1338 E T1q, T1r, T1J, T1K;
|
Chris@10
|
1339 T1q = rio[WS(vs, 1) + WS(rs, 1)];
|
Chris@10
|
1340 T1r = rio[WS(vs, 1) + WS(rs, 5)];
|
Chris@10
|
1341 T1s = T1q + T1r;
|
Chris@10
|
1342 T1I = T1q - T1r;
|
Chris@10
|
1343 T1J = iio[WS(vs, 1) + WS(rs, 1)];
|
Chris@10
|
1344 T1K = iio[WS(vs, 1) + WS(rs, 5)];
|
Chris@10
|
1345 T1L = T1J - T1K;
|
Chris@10
|
1346 T2n = T1J + T1K;
|
Chris@10
|
1347 }
|
Chris@10
|
1348 {
|
Chris@10
|
1349 E T1t, T1u, T1E, T1F;
|
Chris@10
|
1350 T1t = rio[WS(vs, 1) + WS(rs, 7)];
|
Chris@10
|
1351 T1u = rio[WS(vs, 1) + WS(rs, 3)];
|
Chris@10
|
1352 T1v = T1t + T1u;
|
Chris@10
|
1353 T1D = T1t - T1u;
|
Chris@10
|
1354 T1E = iio[WS(vs, 1) + WS(rs, 7)];
|
Chris@10
|
1355 T1F = iio[WS(vs, 1) + WS(rs, 3)];
|
Chris@10
|
1356 T1G = T1E - T1F;
|
Chris@10
|
1357 T2o = T1E + T1F;
|
Chris@10
|
1358 }
|
Chris@10
|
1359 T1w = T1s + T1v;
|
Chris@10
|
1360 T2p = T2n - T2o;
|
Chris@10
|
1361 T2z = T2n + T2o;
|
Chris@10
|
1362 T1H = T1D - T1G;
|
Chris@10
|
1363 T1M = T1I + T1L;
|
Chris@10
|
1364 T1W = T1D + T1G;
|
Chris@10
|
1365 T2j = T1v - T1s;
|
Chris@10
|
1366 T1V = T1L - T1I;
|
Chris@10
|
1367 }
|
Chris@10
|
1368 {
|
Chris@10
|
1369 E T7N, T8i, T83, T8I, T7Q, T80, T8l, T8J;
|
Chris@10
|
1370 {
|
Chris@10
|
1371 E T7L, T7M, T81, T82;
|
Chris@10
|
1372 T7L = rio[WS(vs, 6)];
|
Chris@10
|
1373 T7M = rio[WS(vs, 6) + WS(rs, 4)];
|
Chris@10
|
1374 T7N = T7L + T7M;
|
Chris@10
|
1375 T8i = T7L - T7M;
|
Chris@10
|
1376 T81 = iio[WS(vs, 6)];
|
Chris@10
|
1377 T82 = iio[WS(vs, 6) + WS(rs, 4)];
|
Chris@10
|
1378 T83 = T81 - T82;
|
Chris@10
|
1379 T8I = T81 + T82;
|
Chris@10
|
1380 }
|
Chris@10
|
1381 {
|
Chris@10
|
1382 E T7O, T7P, T8j, T8k;
|
Chris@10
|
1383 T7O = rio[WS(vs, 6) + WS(rs, 2)];
|
Chris@10
|
1384 T7P = rio[WS(vs, 6) + WS(rs, 6)];
|
Chris@10
|
1385 T7Q = T7O + T7P;
|
Chris@10
|
1386 T80 = T7O - T7P;
|
Chris@10
|
1387 T8j = iio[WS(vs, 6) + WS(rs, 2)];
|
Chris@10
|
1388 T8k = iio[WS(vs, 6) + WS(rs, 6)];
|
Chris@10
|
1389 T8l = T8j - T8k;
|
Chris@10
|
1390 T8J = T8j + T8k;
|
Chris@10
|
1391 }
|
Chris@10
|
1392 T7R = T7N + T7Q;
|
Chris@10
|
1393 T8O = T7N - T7Q;
|
Chris@10
|
1394 T90 = T8I + T8J;
|
Chris@10
|
1395 T84 = T80 + T83;
|
Chris@10
|
1396 T8m = T8i - T8l;
|
Chris@10
|
1397 T8A = T83 - T80;
|
Chris@10
|
1398 T8K = T8I - T8J;
|
Chris@10
|
1399 T8w = T8i + T8l;
|
Chris@10
|
1400 }
|
Chris@10
|
1401 {
|
Chris@10
|
1402 E T9c, T9s, T9v, Ta7, T9f, T9n, T9q, Ta8;
|
Chris@10
|
1403 {
|
Chris@10
|
1404 E T9a, T9b, T9t, T9u;
|
Chris@10
|
1405 T9a = rio[WS(vs, 7) + WS(rs, 1)];
|
Chris@10
|
1406 T9b = rio[WS(vs, 7) + WS(rs, 5)];
|
Chris@10
|
1407 T9c = T9a + T9b;
|
Chris@10
|
1408 T9s = T9a - T9b;
|
Chris@10
|
1409 T9t = iio[WS(vs, 7) + WS(rs, 1)];
|
Chris@10
|
1410 T9u = iio[WS(vs, 7) + WS(rs, 5)];
|
Chris@10
|
1411 T9v = T9t - T9u;
|
Chris@10
|
1412 Ta7 = T9t + T9u;
|
Chris@10
|
1413 }
|
Chris@10
|
1414 {
|
Chris@10
|
1415 E T9d, T9e, T9o, T9p;
|
Chris@10
|
1416 T9d = rio[WS(vs, 7) + WS(rs, 7)];
|
Chris@10
|
1417 T9e = rio[WS(vs, 7) + WS(rs, 3)];
|
Chris@10
|
1418 T9f = T9d + T9e;
|
Chris@10
|
1419 T9n = T9d - T9e;
|
Chris@10
|
1420 T9o = iio[WS(vs, 7) + WS(rs, 7)];
|
Chris@10
|
1421 T9p = iio[WS(vs, 7) + WS(rs, 3)];
|
Chris@10
|
1422 T9q = T9o - T9p;
|
Chris@10
|
1423 Ta8 = T9o + T9p;
|
Chris@10
|
1424 }
|
Chris@10
|
1425 T9g = T9c + T9f;
|
Chris@10
|
1426 Ta9 = Ta7 - Ta8;
|
Chris@10
|
1427 Taj = Ta7 + Ta8;
|
Chris@10
|
1428 T9r = T9n - T9q;
|
Chris@10
|
1429 T9w = T9s + T9v;
|
Chris@10
|
1430 T9G = T9n + T9q;
|
Chris@10
|
1431 Ta3 = T9f - T9c;
|
Chris@10
|
1432 T9F = T9v - T9s;
|
Chris@10
|
1433 }
|
Chris@10
|
1434 {
|
Chris@10
|
1435 E Ta, Tq, Tt, T15, Td, Tl, To, T16;
|
Chris@10
|
1436 {
|
Chris@10
|
1437 E T8, T9, Tr, Ts;
|
Chris@10
|
1438 T8 = rio[WS(rs, 1)];
|
Chris@10
|
1439 T9 = rio[WS(rs, 5)];
|
Chris@10
|
1440 Ta = T8 + T9;
|
Chris@10
|
1441 Tq = T8 - T9;
|
Chris@10
|
1442 Tr = iio[WS(rs, 1)];
|
Chris@10
|
1443 Ts = iio[WS(rs, 5)];
|
Chris@10
|
1444 Tt = Tr - Ts;
|
Chris@10
|
1445 T15 = Tr + Ts;
|
Chris@10
|
1446 }
|
Chris@10
|
1447 {
|
Chris@10
|
1448 E Tb, Tc, Tm, Tn;
|
Chris@10
|
1449 Tb = rio[WS(rs, 7)];
|
Chris@10
|
1450 Tc = rio[WS(rs, 3)];
|
Chris@10
|
1451 Td = Tb + Tc;
|
Chris@10
|
1452 Tl = Tb - Tc;
|
Chris@10
|
1453 Tm = iio[WS(rs, 7)];
|
Chris@10
|
1454 Tn = iio[WS(rs, 3)];
|
Chris@10
|
1455 To = Tm - Tn;
|
Chris@10
|
1456 T16 = Tm + Tn;
|
Chris@10
|
1457 }
|
Chris@10
|
1458 Te = Ta + Td;
|
Chris@10
|
1459 T17 = T15 - T16;
|
Chris@10
|
1460 T1h = T15 + T16;
|
Chris@10
|
1461 Tp = Tl - To;
|
Chris@10
|
1462 Tu = Tq + Tt;
|
Chris@10
|
1463 TE = Tl + To;
|
Chris@10
|
1464 T11 = Td - Ta;
|
Chris@10
|
1465 TD = Tt - Tq;
|
Chris@10
|
1466 }
|
Chris@10
|
1467 {
|
Chris@10
|
1468 E T1l, T1Q, T1B, T2g, T1o, T1y, T1T, T2h;
|
Chris@10
|
1469 {
|
Chris@10
|
1470 E T1j, T1k, T1z, T1A;
|
Chris@10
|
1471 T1j = rio[WS(vs, 1)];
|
Chris@10
|
1472 T1k = rio[WS(vs, 1) + WS(rs, 4)];
|
Chris@10
|
1473 T1l = T1j + T1k;
|
Chris@10
|
1474 T1Q = T1j - T1k;
|
Chris@10
|
1475 T1z = iio[WS(vs, 1)];
|
Chris@10
|
1476 T1A = iio[WS(vs, 1) + WS(rs, 4)];
|
Chris@10
|
1477 T1B = T1z - T1A;
|
Chris@10
|
1478 T2g = T1z + T1A;
|
Chris@10
|
1479 }
|
Chris@10
|
1480 {
|
Chris@10
|
1481 E T1m, T1n, T1R, T1S;
|
Chris@10
|
1482 T1m = rio[WS(vs, 1) + WS(rs, 2)];
|
Chris@10
|
1483 T1n = rio[WS(vs, 1) + WS(rs, 6)];
|
Chris@10
|
1484 T1o = T1m + T1n;
|
Chris@10
|
1485 T1y = T1m - T1n;
|
Chris@10
|
1486 T1R = iio[WS(vs, 1) + WS(rs, 2)];
|
Chris@10
|
1487 T1S = iio[WS(vs, 1) + WS(rs, 6)];
|
Chris@10
|
1488 T1T = T1R - T1S;
|
Chris@10
|
1489 T2h = T1R + T1S;
|
Chris@10
|
1490 }
|
Chris@10
|
1491 T1p = T1l + T1o;
|
Chris@10
|
1492 T2m = T1l - T1o;
|
Chris@10
|
1493 T2y = T2g + T2h;
|
Chris@10
|
1494 T1C = T1y + T1B;
|
Chris@10
|
1495 T1U = T1Q - T1T;
|
Chris@10
|
1496 T28 = T1B - T1y;
|
Chris@10
|
1497 T2i = T2g - T2h;
|
Chris@10
|
1498 T24 = T1Q + T1T;
|
Chris@10
|
1499 }
|
Chris@10
|
1500 {
|
Chris@10
|
1501 E T7U, T8a, T8d, T8P, T7X, T85, T88, T8Q;
|
Chris@10
|
1502 {
|
Chris@10
|
1503 E T7S, T7T, T8b, T8c;
|
Chris@10
|
1504 T7S = rio[WS(vs, 6) + WS(rs, 1)];
|
Chris@10
|
1505 T7T = rio[WS(vs, 6) + WS(rs, 5)];
|
Chris@10
|
1506 T7U = T7S + T7T;
|
Chris@10
|
1507 T8a = T7S - T7T;
|
Chris@10
|
1508 T8b = iio[WS(vs, 6) + WS(rs, 1)];
|
Chris@10
|
1509 T8c = iio[WS(vs, 6) + WS(rs, 5)];
|
Chris@10
|
1510 T8d = T8b - T8c;
|
Chris@10
|
1511 T8P = T8b + T8c;
|
Chris@10
|
1512 }
|
Chris@10
|
1513 {
|
Chris@10
|
1514 E T7V, T7W, T86, T87;
|
Chris@10
|
1515 T7V = rio[WS(vs, 6) + WS(rs, 7)];
|
Chris@10
|
1516 T7W = rio[WS(vs, 6) + WS(rs, 3)];
|
Chris@10
|
1517 T7X = T7V + T7W;
|
Chris@10
|
1518 T85 = T7V - T7W;
|
Chris@10
|
1519 T86 = iio[WS(vs, 6) + WS(rs, 7)];
|
Chris@10
|
1520 T87 = iio[WS(vs, 6) + WS(rs, 3)];
|
Chris@10
|
1521 T88 = T86 - T87;
|
Chris@10
|
1522 T8Q = T86 + T87;
|
Chris@10
|
1523 }
|
Chris@10
|
1524 T7Y = T7U + T7X;
|
Chris@10
|
1525 T8R = T8P - T8Q;
|
Chris@10
|
1526 T91 = T8P + T8Q;
|
Chris@10
|
1527 T89 = T85 - T88;
|
Chris@10
|
1528 T8e = T8a + T8d;
|
Chris@10
|
1529 T8o = T85 + T88;
|
Chris@10
|
1530 T8L = T7X - T7U;
|
Chris@10
|
1531 T8n = T8d - T8a;
|
Chris@10
|
1532 }
|
Chris@10
|
1533 {
|
Chris@10
|
1534 E T95, T9A, T9l, Ta0, T98, T9i, T9D, Ta1;
|
Chris@10
|
1535 {
|
Chris@10
|
1536 E T93, T94, T9j, T9k;
|
Chris@10
|
1537 T93 = rio[WS(vs, 7)];
|
Chris@10
|
1538 T94 = rio[WS(vs, 7) + WS(rs, 4)];
|
Chris@10
|
1539 T95 = T93 + T94;
|
Chris@10
|
1540 T9A = T93 - T94;
|
Chris@10
|
1541 T9j = iio[WS(vs, 7)];
|
Chris@10
|
1542 T9k = iio[WS(vs, 7) + WS(rs, 4)];
|
Chris@10
|
1543 T9l = T9j - T9k;
|
Chris@10
|
1544 Ta0 = T9j + T9k;
|
Chris@10
|
1545 }
|
Chris@10
|
1546 {
|
Chris@10
|
1547 E T96, T97, T9B, T9C;
|
Chris@10
|
1548 T96 = rio[WS(vs, 7) + WS(rs, 2)];
|
Chris@10
|
1549 T97 = rio[WS(vs, 7) + WS(rs, 6)];
|
Chris@10
|
1550 T98 = T96 + T97;
|
Chris@10
|
1551 T9i = T96 - T97;
|
Chris@10
|
1552 T9B = iio[WS(vs, 7) + WS(rs, 2)];
|
Chris@10
|
1553 T9C = iio[WS(vs, 7) + WS(rs, 6)];
|
Chris@10
|
1554 T9D = T9B - T9C;
|
Chris@10
|
1555 Ta1 = T9B + T9C;
|
Chris@10
|
1556 }
|
Chris@10
|
1557 T99 = T95 + T98;
|
Chris@10
|
1558 Ta6 = T95 - T98;
|
Chris@10
|
1559 Tai = Ta0 + Ta1;
|
Chris@10
|
1560 T9m = T9i + T9l;
|
Chris@10
|
1561 T9E = T9A - T9D;
|
Chris@10
|
1562 T9S = T9l - T9i;
|
Chris@10
|
1563 Ta2 = Ta0 - Ta1;
|
Chris@10
|
1564 T9O = T9A + T9D;
|
Chris@10
|
1565 }
|
Chris@10
|
1566 {
|
Chris@10
|
1567 E T2D, T38, T2T, T3y, T2G, T2Q, T3b, T3z;
|
Chris@10
|
1568 {
|
Chris@10
|
1569 E T2B, T2C, T2R, T2S;
|
Chris@10
|
1570 T2B = rio[WS(vs, 2)];
|
Chris@10
|
1571 T2C = rio[WS(vs, 2) + WS(rs, 4)];
|
Chris@10
|
1572 T2D = T2B + T2C;
|
Chris@10
|
1573 T38 = T2B - T2C;
|
Chris@10
|
1574 T2R = iio[WS(vs, 2)];
|
Chris@10
|
1575 T2S = iio[WS(vs, 2) + WS(rs, 4)];
|
Chris@10
|
1576 T2T = T2R - T2S;
|
Chris@10
|
1577 T3y = T2R + T2S;
|
Chris@10
|
1578 }
|
Chris@10
|
1579 {
|
Chris@10
|
1580 E T2E, T2F, T39, T3a;
|
Chris@10
|
1581 T2E = rio[WS(vs, 2) + WS(rs, 2)];
|
Chris@10
|
1582 T2F = rio[WS(vs, 2) + WS(rs, 6)];
|
Chris@10
|
1583 T2G = T2E + T2F;
|
Chris@10
|
1584 T2Q = T2E - T2F;
|
Chris@10
|
1585 T39 = iio[WS(vs, 2) + WS(rs, 2)];
|
Chris@10
|
1586 T3a = iio[WS(vs, 2) + WS(rs, 6)];
|
Chris@10
|
1587 T3b = T39 - T3a;
|
Chris@10
|
1588 T3z = T39 + T3a;
|
Chris@10
|
1589 }
|
Chris@10
|
1590 T2H = T2D + T2G;
|
Chris@10
|
1591 T3E = T2D - T2G;
|
Chris@10
|
1592 T3Q = T3y + T3z;
|
Chris@10
|
1593 T2U = T2Q + T2T;
|
Chris@10
|
1594 T3c = T38 - T3b;
|
Chris@10
|
1595 T3q = T2T - T2Q;
|
Chris@10
|
1596 T3A = T3y - T3z;
|
Chris@10
|
1597 T3m = T38 + T3b;
|
Chris@10
|
1598 }
|
Chris@10
|
1599 {
|
Chris@10
|
1600 E T42, T4i, T4l, T4X, T45, T4d, T4g, T4Y;
|
Chris@10
|
1601 {
|
Chris@10
|
1602 E T40, T41, T4j, T4k;
|
Chris@10
|
1603 T40 = rio[WS(vs, 3) + WS(rs, 1)];
|
Chris@10
|
1604 T41 = rio[WS(vs, 3) + WS(rs, 5)];
|
Chris@10
|
1605 T42 = T40 + T41;
|
Chris@10
|
1606 T4i = T40 - T41;
|
Chris@10
|
1607 T4j = iio[WS(vs, 3) + WS(rs, 1)];
|
Chris@10
|
1608 T4k = iio[WS(vs, 3) + WS(rs, 5)];
|
Chris@10
|
1609 T4l = T4j - T4k;
|
Chris@10
|
1610 T4X = T4j + T4k;
|
Chris@10
|
1611 }
|
Chris@10
|
1612 {
|
Chris@10
|
1613 E T43, T44, T4e, T4f;
|
Chris@10
|
1614 T43 = rio[WS(vs, 3) + WS(rs, 7)];
|
Chris@10
|
1615 T44 = rio[WS(vs, 3) + WS(rs, 3)];
|
Chris@10
|
1616 T45 = T43 + T44;
|
Chris@10
|
1617 T4d = T43 - T44;
|
Chris@10
|
1618 T4e = iio[WS(vs, 3) + WS(rs, 7)];
|
Chris@10
|
1619 T4f = iio[WS(vs, 3) + WS(rs, 3)];
|
Chris@10
|
1620 T4g = T4e - T4f;
|
Chris@10
|
1621 T4Y = T4e + T4f;
|
Chris@10
|
1622 }
|
Chris@10
|
1623 T46 = T42 + T45;
|
Chris@10
|
1624 T4Z = T4X - T4Y;
|
Chris@10
|
1625 T59 = T4X + T4Y;
|
Chris@10
|
1626 T4h = T4d - T4g;
|
Chris@10
|
1627 T4m = T4i + T4l;
|
Chris@10
|
1628 T4w = T4d + T4g;
|
Chris@10
|
1629 T4T = T45 - T42;
|
Chris@10
|
1630 T4v = T4l - T4i;
|
Chris@10
|
1631 }
|
Chris@10
|
1632 {
|
Chris@10
|
1633 E T5d, T5I, T5t, T68, T5g, T5q, T5L, T69;
|
Chris@10
|
1634 {
|
Chris@10
|
1635 E T5b, T5c, T5r, T5s;
|
Chris@10
|
1636 T5b = rio[WS(vs, 4)];
|
Chris@10
|
1637 T5c = rio[WS(vs, 4) + WS(rs, 4)];
|
Chris@10
|
1638 T5d = T5b + T5c;
|
Chris@10
|
1639 T5I = T5b - T5c;
|
Chris@10
|
1640 T5r = iio[WS(vs, 4)];
|
Chris@10
|
1641 T5s = iio[WS(vs, 4) + WS(rs, 4)];
|
Chris@10
|
1642 T5t = T5r - T5s;
|
Chris@10
|
1643 T68 = T5r + T5s;
|
Chris@10
|
1644 }
|
Chris@10
|
1645 {
|
Chris@10
|
1646 E T5e, T5f, T5J, T5K;
|
Chris@10
|
1647 T5e = rio[WS(vs, 4) + WS(rs, 2)];
|
Chris@10
|
1648 T5f = rio[WS(vs, 4) + WS(rs, 6)];
|
Chris@10
|
1649 T5g = T5e + T5f;
|
Chris@10
|
1650 T5q = T5e - T5f;
|
Chris@10
|
1651 T5J = iio[WS(vs, 4) + WS(rs, 2)];
|
Chris@10
|
1652 T5K = iio[WS(vs, 4) + WS(rs, 6)];
|
Chris@10
|
1653 T5L = T5J - T5K;
|
Chris@10
|
1654 T69 = T5J + T5K;
|
Chris@10
|
1655 }
|
Chris@10
|
1656 T5h = T5d + T5g;
|
Chris@10
|
1657 T6e = T5d - T5g;
|
Chris@10
|
1658 T6q = T68 + T69;
|
Chris@10
|
1659 T5u = T5q + T5t;
|
Chris@10
|
1660 T5M = T5I - T5L;
|
Chris@10
|
1661 T60 = T5t - T5q;
|
Chris@10
|
1662 T6a = T68 - T69;
|
Chris@10
|
1663 T5W = T5I + T5L;
|
Chris@10
|
1664 }
|
Chris@10
|
1665 {
|
Chris@10
|
1666 E T6C, T6S, T6V, T7x, T6F, T6N, T6Q, T7y;
|
Chris@10
|
1667 {
|
Chris@10
|
1668 E T6A, T6B, T6T, T6U;
|
Chris@10
|
1669 T6A = rio[WS(vs, 5) + WS(rs, 1)];
|
Chris@10
|
1670 T6B = rio[WS(vs, 5) + WS(rs, 5)];
|
Chris@10
|
1671 T6C = T6A + T6B;
|
Chris@10
|
1672 T6S = T6A - T6B;
|
Chris@10
|
1673 T6T = iio[WS(vs, 5) + WS(rs, 1)];
|
Chris@10
|
1674 T6U = iio[WS(vs, 5) + WS(rs, 5)];
|
Chris@10
|
1675 T6V = T6T - T6U;
|
Chris@10
|
1676 T7x = T6T + T6U;
|
Chris@10
|
1677 }
|
Chris@10
|
1678 {
|
Chris@10
|
1679 E T6D, T6E, T6O, T6P;
|
Chris@10
|
1680 T6D = rio[WS(vs, 5) + WS(rs, 7)];
|
Chris@10
|
1681 T6E = rio[WS(vs, 5) + WS(rs, 3)];
|
Chris@10
|
1682 T6F = T6D + T6E;
|
Chris@10
|
1683 T6N = T6D - T6E;
|
Chris@10
|
1684 T6O = iio[WS(vs, 5) + WS(rs, 7)];
|
Chris@10
|
1685 T6P = iio[WS(vs, 5) + WS(rs, 3)];
|
Chris@10
|
1686 T6Q = T6O - T6P;
|
Chris@10
|
1687 T7y = T6O + T6P;
|
Chris@10
|
1688 }
|
Chris@10
|
1689 T6G = T6C + T6F;
|
Chris@10
|
1690 T7z = T7x - T7y;
|
Chris@10
|
1691 T7J = T7x + T7y;
|
Chris@10
|
1692 T6R = T6N - T6Q;
|
Chris@10
|
1693 T6W = T6S + T6V;
|
Chris@10
|
1694 T76 = T6N + T6Q;
|
Chris@10
|
1695 T7t = T6F - T6C;
|
Chris@10
|
1696 T75 = T6V - T6S;
|
Chris@10
|
1697 }
|
Chris@10
|
1698 {
|
Chris@10
|
1699 E T2K, T30, T33, T3F, T2N, T2V, T2Y, T3G;
|
Chris@10
|
1700 {
|
Chris@10
|
1701 E T2I, T2J, T31, T32;
|
Chris@10
|
1702 T2I = rio[WS(vs, 2) + WS(rs, 1)];
|
Chris@10
|
1703 T2J = rio[WS(vs, 2) + WS(rs, 5)];
|
Chris@10
|
1704 T2K = T2I + T2J;
|
Chris@10
|
1705 T30 = T2I - T2J;
|
Chris@10
|
1706 T31 = iio[WS(vs, 2) + WS(rs, 1)];
|
Chris@10
|
1707 T32 = iio[WS(vs, 2) + WS(rs, 5)];
|
Chris@10
|
1708 T33 = T31 - T32;
|
Chris@10
|
1709 T3F = T31 + T32;
|
Chris@10
|
1710 }
|
Chris@10
|
1711 {
|
Chris@10
|
1712 E T2L, T2M, T2W, T2X;
|
Chris@10
|
1713 T2L = rio[WS(vs, 2) + WS(rs, 7)];
|
Chris@10
|
1714 T2M = rio[WS(vs, 2) + WS(rs, 3)];
|
Chris@10
|
1715 T2N = T2L + T2M;
|
Chris@10
|
1716 T2V = T2L - T2M;
|
Chris@10
|
1717 T2W = iio[WS(vs, 2) + WS(rs, 7)];
|
Chris@10
|
1718 T2X = iio[WS(vs, 2) + WS(rs, 3)];
|
Chris@10
|
1719 T2Y = T2W - T2X;
|
Chris@10
|
1720 T3G = T2W + T2X;
|
Chris@10
|
1721 }
|
Chris@10
|
1722 T2O = T2K + T2N;
|
Chris@10
|
1723 T3H = T3F - T3G;
|
Chris@10
|
1724 T3R = T3F + T3G;
|
Chris@10
|
1725 T2Z = T2V - T2Y;
|
Chris@10
|
1726 T34 = T30 + T33;
|
Chris@10
|
1727 T3e = T2V + T2Y;
|
Chris@10
|
1728 T3B = T2N - T2K;
|
Chris@10
|
1729 T3d = T33 - T30;
|
Chris@10
|
1730 }
|
Chris@10
|
1731 {
|
Chris@10
|
1732 E T3V, T4q, T4b, T4Q, T3Y, T48, T4t, T4R;
|
Chris@10
|
1733 {
|
Chris@10
|
1734 E T3T, T3U, T49, T4a;
|
Chris@10
|
1735 T3T = rio[WS(vs, 3)];
|
Chris@10
|
1736 T3U = rio[WS(vs, 3) + WS(rs, 4)];
|
Chris@10
|
1737 T3V = T3T + T3U;
|
Chris@10
|
1738 T4q = T3T - T3U;
|
Chris@10
|
1739 T49 = iio[WS(vs, 3)];
|
Chris@10
|
1740 T4a = iio[WS(vs, 3) + WS(rs, 4)];
|
Chris@10
|
1741 T4b = T49 - T4a;
|
Chris@10
|
1742 T4Q = T49 + T4a;
|
Chris@10
|
1743 }
|
Chris@10
|
1744 {
|
Chris@10
|
1745 E T3W, T3X, T4r, T4s;
|
Chris@10
|
1746 T3W = rio[WS(vs, 3) + WS(rs, 2)];
|
Chris@10
|
1747 T3X = rio[WS(vs, 3) + WS(rs, 6)];
|
Chris@10
|
1748 T3Y = T3W + T3X;
|
Chris@10
|
1749 T48 = T3W - T3X;
|
Chris@10
|
1750 T4r = iio[WS(vs, 3) + WS(rs, 2)];
|
Chris@10
|
1751 T4s = iio[WS(vs, 3) + WS(rs, 6)];
|
Chris@10
|
1752 T4t = T4r - T4s;
|
Chris@10
|
1753 T4R = T4r + T4s;
|
Chris@10
|
1754 }
|
Chris@10
|
1755 T3Z = T3V + T3Y;
|
Chris@10
|
1756 T4W = T3V - T3Y;
|
Chris@10
|
1757 T58 = T4Q + T4R;
|
Chris@10
|
1758 T4c = T48 + T4b;
|
Chris@10
|
1759 T4u = T4q - T4t;
|
Chris@10
|
1760 T4I = T4b - T48;
|
Chris@10
|
1761 T4S = T4Q - T4R;
|
Chris@10
|
1762 T4E = T4q + T4t;
|
Chris@10
|
1763 }
|
Chris@10
|
1764 {
|
Chris@10
|
1765 E T5k, T5A, T5D, T6f, T5n, T5v, T5y, T6g;
|
Chris@10
|
1766 {
|
Chris@10
|
1767 E T5i, T5j, T5B, T5C;
|
Chris@10
|
1768 T5i = rio[WS(vs, 4) + WS(rs, 1)];
|
Chris@10
|
1769 T5j = rio[WS(vs, 4) + WS(rs, 5)];
|
Chris@10
|
1770 T5k = T5i + T5j;
|
Chris@10
|
1771 T5A = T5i - T5j;
|
Chris@10
|
1772 T5B = iio[WS(vs, 4) + WS(rs, 1)];
|
Chris@10
|
1773 T5C = iio[WS(vs, 4) + WS(rs, 5)];
|
Chris@10
|
1774 T5D = T5B - T5C;
|
Chris@10
|
1775 T6f = T5B + T5C;
|
Chris@10
|
1776 }
|
Chris@10
|
1777 {
|
Chris@10
|
1778 E T5l, T5m, T5w, T5x;
|
Chris@10
|
1779 T5l = rio[WS(vs, 4) + WS(rs, 7)];
|
Chris@10
|
1780 T5m = rio[WS(vs, 4) + WS(rs, 3)];
|
Chris@10
|
1781 T5n = T5l + T5m;
|
Chris@10
|
1782 T5v = T5l - T5m;
|
Chris@10
|
1783 T5w = iio[WS(vs, 4) + WS(rs, 7)];
|
Chris@10
|
1784 T5x = iio[WS(vs, 4) + WS(rs, 3)];
|
Chris@10
|
1785 T5y = T5w - T5x;
|
Chris@10
|
1786 T6g = T5w + T5x;
|
Chris@10
|
1787 }
|
Chris@10
|
1788 T5o = T5k + T5n;
|
Chris@10
|
1789 T6h = T6f - T6g;
|
Chris@10
|
1790 T6r = T6f + T6g;
|
Chris@10
|
1791 T5z = T5v - T5y;
|
Chris@10
|
1792 T5E = T5A + T5D;
|
Chris@10
|
1793 T5O = T5v + T5y;
|
Chris@10
|
1794 T6b = T5n - T5k;
|
Chris@10
|
1795 T5N = T5D - T5A;
|
Chris@10
|
1796 }
|
Chris@10
|
1797 {
|
Chris@10
|
1798 E T6v, T70, T6L, T7q, T6y, T6I, T73, T7r;
|
Chris@10
|
1799 {
|
Chris@10
|
1800 E T6t, T6u, T6J, T6K;
|
Chris@10
|
1801 T6t = rio[WS(vs, 5)];
|
Chris@10
|
1802 T6u = rio[WS(vs, 5) + WS(rs, 4)];
|
Chris@10
|
1803 T6v = T6t + T6u;
|
Chris@10
|
1804 T70 = T6t - T6u;
|
Chris@10
|
1805 T6J = iio[WS(vs, 5)];
|
Chris@10
|
1806 T6K = iio[WS(vs, 5) + WS(rs, 4)];
|
Chris@10
|
1807 T6L = T6J - T6K;
|
Chris@10
|
1808 T7q = T6J + T6K;
|
Chris@10
|
1809 }
|
Chris@10
|
1810 {
|
Chris@10
|
1811 E T6w, T6x, T71, T72;
|
Chris@10
|
1812 T6w = rio[WS(vs, 5) + WS(rs, 2)];
|
Chris@10
|
1813 T6x = rio[WS(vs, 5) + WS(rs, 6)];
|
Chris@10
|
1814 T6y = T6w + T6x;
|
Chris@10
|
1815 T6I = T6w - T6x;
|
Chris@10
|
1816 T71 = iio[WS(vs, 5) + WS(rs, 2)];
|
Chris@10
|
1817 T72 = iio[WS(vs, 5) + WS(rs, 6)];
|
Chris@10
|
1818 T73 = T71 - T72;
|
Chris@10
|
1819 T7r = T71 + T72;
|
Chris@10
|
1820 }
|
Chris@10
|
1821 T6z = T6v + T6y;
|
Chris@10
|
1822 T7w = T6v - T6y;
|
Chris@10
|
1823 T7I = T7q + T7r;
|
Chris@10
|
1824 T6M = T6I + T6L;
|
Chris@10
|
1825 T74 = T70 - T73;
|
Chris@10
|
1826 T7i = T6L - T6I;
|
Chris@10
|
1827 T7s = T7q - T7r;
|
Chris@10
|
1828 T7e = T70 + T73;
|
Chris@10
|
1829 }
|
Chris@10
|
1830 rio[0] = T7 + Te;
|
Chris@10
|
1831 iio[0] = T1g + T1h;
|
Chris@10
|
1832 rio[WS(rs, 1)] = T1p + T1w;
|
Chris@10
|
1833 iio[WS(rs, 1)] = T2y + T2z;
|
Chris@10
|
1834 rio[WS(rs, 3)] = T3Z + T46;
|
Chris@10
|
1835 rio[WS(rs, 2)] = T2H + T2O;
|
Chris@10
|
1836 iio[WS(rs, 2)] = T3Q + T3R;
|
Chris@10
|
1837 iio[WS(rs, 3)] = T58 + T59;
|
Chris@10
|
1838 rio[WS(rs, 6)] = T7R + T7Y;
|
Chris@10
|
1839 iio[WS(rs, 6)] = T90 + T91;
|
Chris@10
|
1840 iio[WS(rs, 5)] = T7I + T7J;
|
Chris@10
|
1841 rio[WS(rs, 5)] = T6z + T6G;
|
Chris@10
|
1842 iio[WS(rs, 4)] = T6q + T6r;
|
Chris@10
|
1843 rio[WS(rs, 4)] = T5h + T5o;
|
Chris@10
|
1844 rio[WS(rs, 7)] = T99 + T9g;
|
Chris@10
|
1845 iio[WS(rs, 7)] = Tai + Taj;
|
Chris@10
|
1846 {
|
Chris@10
|
1847 E T12, T18, TX, T13;
|
Chris@10
|
1848 T12 = T10 - T11;
|
Chris@10
|
1849 T18 = T14 - T17;
|
Chris@10
|
1850 TX = W[10];
|
Chris@10
|
1851 T13 = W[11];
|
Chris@10
|
1852 iio[WS(vs, 6)] = FNMS(T13, T18, TX * T12);
|
Chris@10
|
1853 rio[WS(vs, 6)] = FMA(T13, T12, TX * T18);
|
Chris@10
|
1854 }
|
Chris@10
|
1855 {
|
Chris@10
|
1856 E Tag, Tak, Taf, Tah;
|
Chris@10
|
1857 Tag = T99 - T9g;
|
Chris@10
|
1858 Tak = Tai - Taj;
|
Chris@10
|
1859 Taf = W[6];
|
Chris@10
|
1860 Tah = W[7];
|
Chris@10
|
1861 rio[WS(vs, 4) + WS(rs, 7)] = FMA(Taf, Tag, Tah * Tak);
|
Chris@10
|
1862 iio[WS(vs, 4) + WS(rs, 7)] = FNMS(Tah, Tag, Taf * Tak);
|
Chris@10
|
1863 }
|
Chris@10
|
1864 {
|
Chris@10
|
1865 E T8M, T8S, T8H, T8N;
|
Chris@10
|
1866 T8M = T8K - T8L;
|
Chris@10
|
1867 T8S = T8O - T8R;
|
Chris@10
|
1868 T8H = W[10];
|
Chris@10
|
1869 T8N = W[11];
|
Chris@10
|
1870 iio[WS(vs, 6) + WS(rs, 6)] = FNMS(T8N, T8S, T8H * T8M);
|
Chris@10
|
1871 rio[WS(vs, 6) + WS(rs, 6)] = FMA(T8N, T8M, T8H * T8S);
|
Chris@10
|
1872 }
|
Chris@10
|
1873 {
|
Chris@10
|
1874 E T2k, T2q, T2f, T2l;
|
Chris@10
|
1875 T2k = T2i - T2j;
|
Chris@10
|
1876 T2q = T2m - T2p;
|
Chris@10
|
1877 T2f = W[10];
|
Chris@10
|
1878 T2l = W[11];
|
Chris@10
|
1879 iio[WS(vs, 6) + WS(rs, 1)] = FNMS(T2l, T2q, T2f * T2k);
|
Chris@10
|
1880 rio[WS(vs, 6) + WS(rs, 1)] = FMA(T2l, T2k, T2f * T2q);
|
Chris@10
|
1881 }
|
Chris@10
|
1882 {
|
Chris@10
|
1883 E Ta4, Taa, T9Z, Ta5;
|
Chris@10
|
1884 Ta4 = Ta2 - Ta3;
|
Chris@10
|
1885 Taa = Ta6 - Ta9;
|
Chris@10
|
1886 T9Z = W[10];
|
Chris@10
|
1887 Ta5 = W[11];
|
Chris@10
|
1888 iio[WS(vs, 6) + WS(rs, 7)] = FNMS(Ta5, Taa, T9Z * Ta4);
|
Chris@10
|
1889 rio[WS(vs, 6) + WS(rs, 7)] = FMA(Ta5, Ta4, T9Z * Taa);
|
Chris@10
|
1890 }
|
Chris@10
|
1891 {
|
Chris@10
|
1892 E T8Y, T92, T8X, T8Z;
|
Chris@10
|
1893 T8Y = T7R - T7Y;
|
Chris@10
|
1894 T92 = T90 - T91;
|
Chris@10
|
1895 T8X = W[6];
|
Chris@10
|
1896 T8Z = W[7];
|
Chris@10
|
1897 rio[WS(vs, 4) + WS(rs, 6)] = FMA(T8X, T8Y, T8Z * T92);
|
Chris@10
|
1898 iio[WS(vs, 4) + WS(rs, 6)] = FNMS(T8Z, T8Y, T8X * T92);
|
Chris@10
|
1899 }
|
Chris@10
|
1900 {
|
Chris@10
|
1901 E T2w, T2A, T2v, T2x;
|
Chris@10
|
1902 T2w = T1p - T1w;
|
Chris@10
|
1903 T2A = T2y - T2z;
|
Chris@10
|
1904 T2v = W[6];
|
Chris@10
|
1905 T2x = W[7];
|
Chris@10
|
1906 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T2v, T2w, T2x * T2A);
|
Chris@10
|
1907 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T2x, T2w, T2v * T2A);
|
Chris@10
|
1908 }
|
Chris@10
|
1909 {
|
Chris@10
|
1910 E Tac, Tae, Tab, Tad;
|
Chris@10
|
1911 Tac = Ta3 + Ta2;
|
Chris@10
|
1912 Tae = Ta6 + Ta9;
|
Chris@10
|
1913 Tab = W[2];
|
Chris@10
|
1914 Tad = W[3];
|
Chris@10
|
1915 iio[WS(vs, 2) + WS(rs, 7)] = FNMS(Tad, Tae, Tab * Tac);
|
Chris@10
|
1916 rio[WS(vs, 2) + WS(rs, 7)] = FMA(Tad, Tac, Tab * Tae);
|
Chris@10
|
1917 }
|
Chris@10
|
1918 {
|
Chris@10
|
1919 E T8U, T8W, T8T, T8V;
|
Chris@10
|
1920 T8U = T8L + T8K;
|
Chris@10
|
1921 T8W = T8O + T8R;
|
Chris@10
|
1922 T8T = W[2];
|
Chris@10
|
1923 T8V = W[3];
|
Chris@10
|
1924 iio[WS(vs, 2) + WS(rs, 6)] = FNMS(T8V, T8W, T8T * T8U);
|
Chris@10
|
1925 rio[WS(vs, 2) + WS(rs, 6)] = FMA(T8V, T8U, T8T * T8W);
|
Chris@10
|
1926 }
|
Chris@10
|
1927 {
|
Chris@10
|
1928 E T1a, T1c, T19, T1b;
|
Chris@10
|
1929 T1a = T11 + T10;
|
Chris@10
|
1930 T1c = T14 + T17;
|
Chris@10
|
1931 T19 = W[2];
|
Chris@10
|
1932 T1b = W[3];
|
Chris@10
|
1933 iio[WS(vs, 2)] = FNMS(T1b, T1c, T19 * T1a);
|
Chris@10
|
1934 rio[WS(vs, 2)] = FMA(T1b, T1a, T19 * T1c);
|
Chris@10
|
1935 }
|
Chris@10
|
1936 {
|
Chris@10
|
1937 E T1e, T1i, T1d, T1f;
|
Chris@10
|
1938 T1e = T7 - Te;
|
Chris@10
|
1939 T1i = T1g - T1h;
|
Chris@10
|
1940 T1d = W[6];
|
Chris@10
|
1941 T1f = W[7];
|
Chris@10
|
1942 rio[WS(vs, 4)] = FMA(T1d, T1e, T1f * T1i);
|
Chris@10
|
1943 iio[WS(vs, 4)] = FNMS(T1f, T1e, T1d * T1i);
|
Chris@10
|
1944 }
|
Chris@10
|
1945 {
|
Chris@10
|
1946 E T2s, T2u, T2r, T2t;
|
Chris@10
|
1947 T2s = T2j + T2i;
|
Chris@10
|
1948 T2u = T2m + T2p;
|
Chris@10
|
1949 T2r = W[2];
|
Chris@10
|
1950 T2t = W[3];
|
Chris@10
|
1951 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T2t, T2u, T2r * T2s);
|
Chris@10
|
1952 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T2t, T2s, T2r * T2u);
|
Chris@10
|
1953 }
|
Chris@10
|
1954 {
|
Chris@10
|
1955 E T3C, T3I, T3x, T3D;
|
Chris@10
|
1956 T3C = T3A - T3B;
|
Chris@10
|
1957 T3I = T3E - T3H;
|
Chris@10
|
1958 T3x = W[10];
|
Chris@10
|
1959 T3D = W[11];
|
Chris@10
|
1960 iio[WS(vs, 6) + WS(rs, 2)] = FNMS(T3D, T3I, T3x * T3C);
|
Chris@10
|
1961 rio[WS(vs, 6) + WS(rs, 2)] = FMA(T3D, T3C, T3x * T3I);
|
Chris@10
|
1962 }
|
Chris@10
|
1963 {
|
Chris@10
|
1964 E T4U, T50, T4P, T4V;
|
Chris@10
|
1965 T4U = T4S - T4T;
|
Chris@10
|
1966 T50 = T4W - T4Z;
|
Chris@10
|
1967 T4P = W[10];
|
Chris@10
|
1968 T4V = W[11];
|
Chris@10
|
1969 iio[WS(vs, 6) + WS(rs, 3)] = FNMS(T4V, T50, T4P * T4U);
|
Chris@10
|
1970 rio[WS(vs, 6) + WS(rs, 3)] = FMA(T4V, T4U, T4P * T50);
|
Chris@10
|
1971 }
|
Chris@10
|
1972 {
|
Chris@10
|
1973 E T56, T5a, T55, T57;
|
Chris@10
|
1974 T56 = T3Z - T46;
|
Chris@10
|
1975 T5a = T58 - T59;
|
Chris@10
|
1976 T55 = W[6];
|
Chris@10
|
1977 T57 = W[7];
|
Chris@10
|
1978 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T55, T56, T57 * T5a);
|
Chris@10
|
1979 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T57, T56, T55 * T5a);
|
Chris@10
|
1980 }
|
Chris@10
|
1981 {
|
Chris@10
|
1982 E T6o, T6s, T6n, T6p;
|
Chris@10
|
1983 T6o = T5h - T5o;
|
Chris@10
|
1984 T6s = T6q - T6r;
|
Chris@10
|
1985 T6n = W[6];
|
Chris@10
|
1986 T6p = W[7];
|
Chris@10
|
1987 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T6n, T6o, T6p * T6s);
|
Chris@10
|
1988 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T6p, T6o, T6n * T6s);
|
Chris@10
|
1989 }
|
Chris@10
|
1990 {
|
Chris@10
|
1991 E T7u, T7A, T7p, T7v;
|
Chris@10
|
1992 T7u = T7s - T7t;
|
Chris@10
|
1993 T7A = T7w - T7z;
|
Chris@10
|
1994 T7p = W[10];
|
Chris@10
|
1995 T7v = W[11];
|
Chris@10
|
1996 iio[WS(vs, 6) + WS(rs, 5)] = FNMS(T7v, T7A, T7p * T7u);
|
Chris@10
|
1997 rio[WS(vs, 6) + WS(rs, 5)] = FMA(T7v, T7u, T7p * T7A);
|
Chris@10
|
1998 }
|
Chris@10
|
1999 {
|
Chris@10
|
2000 E T6c, T6i, T67, T6d;
|
Chris@10
|
2001 T6c = T6a - T6b;
|
Chris@10
|
2002 T6i = T6e - T6h;
|
Chris@10
|
2003 T67 = W[10];
|
Chris@10
|
2004 T6d = W[11];
|
Chris@10
|
2005 iio[WS(vs, 6) + WS(rs, 4)] = FNMS(T6d, T6i, T67 * T6c);
|
Chris@10
|
2006 rio[WS(vs, 6) + WS(rs, 4)] = FMA(T6d, T6c, T67 * T6i);
|
Chris@10
|
2007 }
|
Chris@10
|
2008 {
|
Chris@10
|
2009 E T7G, T7K, T7F, T7H;
|
Chris@10
|
2010 T7G = T6z - T6G;
|
Chris@10
|
2011 T7K = T7I - T7J;
|
Chris@10
|
2012 T7F = W[6];
|
Chris@10
|
2013 T7H = W[7];
|
Chris@10
|
2014 rio[WS(vs, 4) + WS(rs, 5)] = FMA(T7F, T7G, T7H * T7K);
|
Chris@10
|
2015 iio[WS(vs, 4) + WS(rs, 5)] = FNMS(T7H, T7G, T7F * T7K);
|
Chris@10
|
2016 }
|
Chris@10
|
2017 {
|
Chris@10
|
2018 E T3O, T3S, T3N, T3P;
|
Chris@10
|
2019 T3O = T2H - T2O;
|
Chris@10
|
2020 T3S = T3Q - T3R;
|
Chris@10
|
2021 T3N = W[6];
|
Chris@10
|
2022 T3P = W[7];
|
Chris@10
|
2023 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T3N, T3O, T3P * T3S);
|
Chris@10
|
2024 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T3P, T3O, T3N * T3S);
|
Chris@10
|
2025 }
|
Chris@10
|
2026 {
|
Chris@10
|
2027 E T3K, T3M, T3J, T3L;
|
Chris@10
|
2028 T3K = T3B + T3A;
|
Chris@10
|
2029 T3M = T3E + T3H;
|
Chris@10
|
2030 T3J = W[2];
|
Chris@10
|
2031 T3L = W[3];
|
Chris@10
|
2032 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T3L, T3M, T3J * T3K);
|
Chris@10
|
2033 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T3L, T3K, T3J * T3M);
|
Chris@10
|
2034 }
|
Chris@10
|
2035 {
|
Chris@10
|
2036 E T7C, T7E, T7B, T7D;
|
Chris@10
|
2037 T7C = T7t + T7s;
|
Chris@10
|
2038 T7E = T7w + T7z;
|
Chris@10
|
2039 T7B = W[2];
|
Chris@10
|
2040 T7D = W[3];
|
Chris@10
|
2041 iio[WS(vs, 2) + WS(rs, 5)] = FNMS(T7D, T7E, T7B * T7C);
|
Chris@10
|
2042 rio[WS(vs, 2) + WS(rs, 5)] = FMA(T7D, T7C, T7B * T7E);
|
Chris@10
|
2043 }
|
Chris@10
|
2044 {
|
Chris@10
|
2045 E T6k, T6m, T6j, T6l;
|
Chris@10
|
2046 T6k = T6b + T6a;
|
Chris@10
|
2047 T6m = T6e + T6h;
|
Chris@10
|
2048 T6j = W[2];
|
Chris@10
|
2049 T6l = W[3];
|
Chris@10
|
2050 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T6l, T6m, T6j * T6k);
|
Chris@10
|
2051 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T6l, T6k, T6j * T6m);
|
Chris@10
|
2052 }
|
Chris@10
|
2053 {
|
Chris@10
|
2054 E T52, T54, T51, T53;
|
Chris@10
|
2055 T52 = T4T + T4S;
|
Chris@10
|
2056 T54 = T4W + T4Z;
|
Chris@10
|
2057 T51 = W[2];
|
Chris@10
|
2058 T53 = W[3];
|
Chris@10
|
2059 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T53, T54, T51 * T52);
|
Chris@10
|
2060 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T53, T52, T51 * T54);
|
Chris@10
|
2061 }
|
Chris@10
|
2062 {
|
Chris@10
|
2063 E T5G, T5S, T5Q, T5U, T5F, T5P;
|
Chris@10
|
2064 T5F = KP707106781 * (T5z - T5E);
|
Chris@10
|
2065 T5G = T5u - T5F;
|
Chris@10
|
2066 T5S = T5u + T5F;
|
Chris@10
|
2067 T5P = KP707106781 * (T5N - T5O);
|
Chris@10
|
2068 T5Q = T5M - T5P;
|
Chris@10
|
2069 T5U = T5M + T5P;
|
Chris@10
|
2070 {
|
Chris@10
|
2071 E T5p, T5H, T5R, T5T;
|
Chris@10
|
2072 T5p = W[12];
|
Chris@10
|
2073 T5H = W[13];
|
Chris@10
|
2074 iio[WS(vs, 7) + WS(rs, 4)] = FNMS(T5H, T5Q, T5p * T5G);
|
Chris@10
|
2075 rio[WS(vs, 7) + WS(rs, 4)] = FMA(T5H, T5G, T5p * T5Q);
|
Chris@10
|
2076 T5R = W[4];
|
Chris@10
|
2077 T5T = W[5];
|
Chris@10
|
2078 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T5T, T5U, T5R * T5S);
|
Chris@10
|
2079 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T5T, T5S, T5R * T5U);
|
Chris@10
|
2080 }
|
Chris@10
|
2081 }
|
Chris@10
|
2082 {
|
Chris@10
|
2083 E Tw, TI, TG, TK, Tv, TF;
|
Chris@10
|
2084 Tv = KP707106781 * (Tp - Tu);
|
Chris@10
|
2085 Tw = Tk - Tv;
|
Chris@10
|
2086 TI = Tk + Tv;
|
Chris@10
|
2087 TF = KP707106781 * (TD - TE);
|
Chris@10
|
2088 TG = TC - TF;
|
Chris@10
|
2089 TK = TC + TF;
|
Chris@10
|
2090 {
|
Chris@10
|
2091 E Tf, Tx, TH, TJ;
|
Chris@10
|
2092 Tf = W[12];
|
Chris@10
|
2093 Tx = W[13];
|
Chris@10
|
2094 iio[WS(vs, 7)] = FNMS(Tx, TG, Tf * Tw);
|
Chris@10
|
2095 rio[WS(vs, 7)] = FMA(Tx, Tw, Tf * TG);
|
Chris@10
|
2096 TH = W[4];
|
Chris@10
|
2097 TJ = W[5];
|
Chris@10
|
2098 iio[WS(vs, 3)] = FNMS(TJ, TK, TH * TI);
|
Chris@10
|
2099 rio[WS(vs, 3)] = FMA(TJ, TI, TH * TK);
|
Chris@10
|
2100 }
|
Chris@10
|
2101 }
|
Chris@10
|
2102 {
|
Chris@10
|
2103 E T9Q, T9W, T9U, T9Y, T9P, T9T;
|
Chris@10
|
2104 T9P = KP707106781 * (T9w + T9r);
|
Chris@10
|
2105 T9Q = T9O - T9P;
|
Chris@10
|
2106 T9W = T9O + T9P;
|
Chris@10
|
2107 T9T = KP707106781 * (T9F + T9G);
|
Chris@10
|
2108 T9U = T9S - T9T;
|
Chris@10
|
2109 T9Y = T9S + T9T;
|
Chris@10
|
2110 {
|
Chris@10
|
2111 E T9N, T9R, T9V, T9X;
|
Chris@10
|
2112 T9N = W[8];
|
Chris@10
|
2113 T9R = W[9];
|
Chris@10
|
2114 rio[WS(vs, 5) + WS(rs, 7)] = FMA(T9N, T9Q, T9R * T9U);
|
Chris@10
|
2115 iio[WS(vs, 5) + WS(rs, 7)] = FNMS(T9R, T9Q, T9N * T9U);
|
Chris@10
|
2116 T9V = W[0];
|
Chris@10
|
2117 T9X = W[1];
|
Chris@10
|
2118 rio[WS(vs, 1) + WS(rs, 7)] = FMA(T9V, T9W, T9X * T9Y);
|
Chris@10
|
2119 iio[WS(vs, 1) + WS(rs, 7)] = FNMS(T9X, T9W, T9V * T9Y);
|
Chris@10
|
2120 }
|
Chris@10
|
2121 }
|
Chris@10
|
2122 {
|
Chris@10
|
2123 E T36, T3i, T3g, T3k, T35, T3f;
|
Chris@10
|
2124 T35 = KP707106781 * (T2Z - T34);
|
Chris@10
|
2125 T36 = T2U - T35;
|
Chris@10
|
2126 T3i = T2U + T35;
|
Chris@10
|
2127 T3f = KP707106781 * (T3d - T3e);
|
Chris@10
|
2128 T3g = T3c - T3f;
|
Chris@10
|
2129 T3k = T3c + T3f;
|
Chris@10
|
2130 {
|
Chris@10
|
2131 E T2P, T37, T3h, T3j;
|
Chris@10
|
2132 T2P = W[12];
|
Chris@10
|
2133 T37 = W[13];
|
Chris@10
|
2134 iio[WS(vs, 7) + WS(rs, 2)] = FNMS(T37, T3g, T2P * T36);
|
Chris@10
|
2135 rio[WS(vs, 7) + WS(rs, 2)] = FMA(T37, T36, T2P * T3g);
|
Chris@10
|
2136 T3h = W[4];
|
Chris@10
|
2137 T3j = W[5];
|
Chris@10
|
2138 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T3j, T3k, T3h * T3i);
|
Chris@10
|
2139 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T3j, T3i, T3h * T3k);
|
Chris@10
|
2140 }
|
Chris@10
|
2141 }
|
Chris@10
|
2142 {
|
Chris@10
|
2143 E T5Y, T64, T62, T66, T5X, T61;
|
Chris@10
|
2144 T5X = KP707106781 * (T5E + T5z);
|
Chris@10
|
2145 T5Y = T5W - T5X;
|
Chris@10
|
2146 T64 = T5W + T5X;
|
Chris@10
|
2147 T61 = KP707106781 * (T5N + T5O);
|
Chris@10
|
2148 T62 = T60 - T61;
|
Chris@10
|
2149 T66 = T60 + T61;
|
Chris@10
|
2150 {
|
Chris@10
|
2151 E T5V, T5Z, T63, T65;
|
Chris@10
|
2152 T5V = W[8];
|
Chris@10
|
2153 T5Z = W[9];
|
Chris@10
|
2154 rio[WS(vs, 5) + WS(rs, 4)] = FMA(T5V, T5Y, T5Z * T62);
|
Chris@10
|
2155 iio[WS(vs, 5) + WS(rs, 4)] = FNMS(T5Z, T5Y, T5V * T62);
|
Chris@10
|
2156 T63 = W[0];
|
Chris@10
|
2157 T65 = W[1];
|
Chris@10
|
2158 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T63, T64, T65 * T66);
|
Chris@10
|
2159 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T65, T64, T63 * T66);
|
Chris@10
|
2160 }
|
Chris@10
|
2161 }
|
Chris@10
|
2162 {
|
Chris@10
|
2163 E T7g, T7m, T7k, T7o, T7f, T7j;
|
Chris@10
|
2164 T7f = KP707106781 * (T6W + T6R);
|
Chris@10
|
2165 T7g = T7e - T7f;
|
Chris@10
|
2166 T7m = T7e + T7f;
|
Chris@10
|
2167 T7j = KP707106781 * (T75 + T76);
|
Chris@10
|
2168 T7k = T7i - T7j;
|
Chris@10
|
2169 T7o = T7i + T7j;
|
Chris@10
|
2170 {
|
Chris@10
|
2171 E T7d, T7h, T7l, T7n;
|
Chris@10
|
2172 T7d = W[8];
|
Chris@10
|
2173 T7h = W[9];
|
Chris@10
|
2174 rio[WS(vs, 5) + WS(rs, 5)] = FMA(T7d, T7g, T7h * T7k);
|
Chris@10
|
2175 iio[WS(vs, 5) + WS(rs, 5)] = FNMS(T7h, T7g, T7d * T7k);
|
Chris@10
|
2176 T7l = W[0];
|
Chris@10
|
2177 T7n = W[1];
|
Chris@10
|
2178 rio[WS(vs, 1) + WS(rs, 5)] = FMA(T7l, T7m, T7n * T7o);
|
Chris@10
|
2179 iio[WS(vs, 1) + WS(rs, 5)] = FNMS(T7n, T7m, T7l * T7o);
|
Chris@10
|
2180 }
|
Chris@10
|
2181 }
|
Chris@10
|
2182 {
|
Chris@10
|
2183 E T8g, T8s, T8q, T8u, T8f, T8p;
|
Chris@10
|
2184 T8f = KP707106781 * (T89 - T8e);
|
Chris@10
|
2185 T8g = T84 - T8f;
|
Chris@10
|
2186 T8s = T84 + T8f;
|
Chris@10
|
2187 T8p = KP707106781 * (T8n - T8o);
|
Chris@10
|
2188 T8q = T8m - T8p;
|
Chris@10
|
2189 T8u = T8m + T8p;
|
Chris@10
|
2190 {
|
Chris@10
|
2191 E T7Z, T8h, T8r, T8t;
|
Chris@10
|
2192 T7Z = W[12];
|
Chris@10
|
2193 T8h = W[13];
|
Chris@10
|
2194 iio[WS(vs, 7) + WS(rs, 6)] = FNMS(T8h, T8q, T7Z * T8g);
|
Chris@10
|
2195 rio[WS(vs, 7) + WS(rs, 6)] = FMA(T8h, T8g, T7Z * T8q);
|
Chris@10
|
2196 T8r = W[4];
|
Chris@10
|
2197 T8t = W[5];
|
Chris@10
|
2198 iio[WS(vs, 3) + WS(rs, 6)] = FNMS(T8t, T8u, T8r * T8s);
|
Chris@10
|
2199 rio[WS(vs, 3) + WS(rs, 6)] = FMA(T8t, T8s, T8r * T8u);
|
Chris@10
|
2200 }
|
Chris@10
|
2201 }
|
Chris@10
|
2202 {
|
Chris@10
|
2203 E T4G, T4M, T4K, T4O, T4F, T4J;
|
Chris@10
|
2204 T4F = KP707106781 * (T4m + T4h);
|
Chris@10
|
2205 T4G = T4E - T4F;
|
Chris@10
|
2206 T4M = T4E + T4F;
|
Chris@10
|
2207 T4J = KP707106781 * (T4v + T4w);
|
Chris@10
|
2208 T4K = T4I - T4J;
|
Chris@10
|
2209 T4O = T4I + T4J;
|
Chris@10
|
2210 {
|
Chris@10
|
2211 E T4D, T4H, T4L, T4N;
|
Chris@10
|
2212 T4D = W[8];
|
Chris@10
|
2213 T4H = W[9];
|
Chris@10
|
2214 rio[WS(vs, 5) + WS(rs, 3)] = FMA(T4D, T4G, T4H * T4K);
|
Chris@10
|
2215 iio[WS(vs, 5) + WS(rs, 3)] = FNMS(T4H, T4G, T4D * T4K);
|
Chris@10
|
2216 T4L = W[0];
|
Chris@10
|
2217 T4N = W[1];
|
Chris@10
|
2218 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T4L, T4M, T4N * T4O);
|
Chris@10
|
2219 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T4N, T4M, T4L * T4O);
|
Chris@10
|
2220 }
|
Chris@10
|
2221 }
|
Chris@10
|
2222 {
|
Chris@10
|
2223 E TO, TU, TS, TW, TN, TR;
|
Chris@10
|
2224 TN = KP707106781 * (Tu + Tp);
|
Chris@10
|
2225 TO = TM - TN;
|
Chris@10
|
2226 TU = TM + TN;
|
Chris@10
|
2227 TR = KP707106781 * (TD + TE);
|
Chris@10
|
2228 TS = TQ - TR;
|
Chris@10
|
2229 TW = TQ + TR;
|
Chris@10
|
2230 {
|
Chris@10
|
2231 E TL, TP, TT, TV;
|
Chris@10
|
2232 TL = W[8];
|
Chris@10
|
2233 TP = W[9];
|
Chris@10
|
2234 rio[WS(vs, 5)] = FMA(TL, TO, TP * TS);
|
Chris@10
|
2235 iio[WS(vs, 5)] = FNMS(TP, TO, TL * TS);
|
Chris@10
|
2236 TT = W[0];
|
Chris@10
|
2237 TV = W[1];
|
Chris@10
|
2238 rio[WS(vs, 1)] = FMA(TT, TU, TV * TW);
|
Chris@10
|
2239 iio[WS(vs, 1)] = FNMS(TV, TU, TT * TW);
|
Chris@10
|
2240 }
|
Chris@10
|
2241 }
|
Chris@10
|
2242 {
|
Chris@10
|
2243 E T26, T2c, T2a, T2e, T25, T29;
|
Chris@10
|
2244 T25 = KP707106781 * (T1M + T1H);
|
Chris@10
|
2245 T26 = T24 - T25;
|
Chris@10
|
2246 T2c = T24 + T25;
|
Chris@10
|
2247 T29 = KP707106781 * (T1V + T1W);
|
Chris@10
|
2248 T2a = T28 - T29;
|
Chris@10
|
2249 T2e = T28 + T29;
|
Chris@10
|
2250 {
|
Chris@10
|
2251 E T23, T27, T2b, T2d;
|
Chris@10
|
2252 T23 = W[8];
|
Chris@10
|
2253 T27 = W[9];
|
Chris@10
|
2254 rio[WS(vs, 5) + WS(rs, 1)] = FMA(T23, T26, T27 * T2a);
|
Chris@10
|
2255 iio[WS(vs, 5) + WS(rs, 1)] = FNMS(T27, T26, T23 * T2a);
|
Chris@10
|
2256 T2b = W[0];
|
Chris@10
|
2257 T2d = W[1];
|
Chris@10
|
2258 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T2b, T2c, T2d * T2e);
|
Chris@10
|
2259 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T2d, T2c, T2b * T2e);
|
Chris@10
|
2260 }
|
Chris@10
|
2261 }
|
Chris@10
|
2262 {
|
Chris@10
|
2263 E T9y, T9K, T9I, T9M, T9x, T9H;
|
Chris@10
|
2264 T9x = KP707106781 * (T9r - T9w);
|
Chris@10
|
2265 T9y = T9m - T9x;
|
Chris@10
|
2266 T9K = T9m + T9x;
|
Chris@10
|
2267 T9H = KP707106781 * (T9F - T9G);
|
Chris@10
|
2268 T9I = T9E - T9H;
|
Chris@10
|
2269 T9M = T9E + T9H;
|
Chris@10
|
2270 {
|
Chris@10
|
2271 E T9h, T9z, T9J, T9L;
|
Chris@10
|
2272 T9h = W[12];
|
Chris@10
|
2273 T9z = W[13];
|
Chris@10
|
2274 iio[WS(vs, 7) + WS(rs, 7)] = FNMS(T9z, T9I, T9h * T9y);
|
Chris@10
|
2275 rio[WS(vs, 7) + WS(rs, 7)] = FMA(T9z, T9y, T9h * T9I);
|
Chris@10
|
2276 T9J = W[4];
|
Chris@10
|
2277 T9L = W[5];
|
Chris@10
|
2278 iio[WS(vs, 3) + WS(rs, 7)] = FNMS(T9L, T9M, T9J * T9K);
|
Chris@10
|
2279 rio[WS(vs, 3) + WS(rs, 7)] = FMA(T9L, T9K, T9J * T9M);
|
Chris@10
|
2280 }
|
Chris@10
|
2281 }
|
Chris@10
|
2282 {
|
Chris@10
|
2283 E T6Y, T7a, T78, T7c, T6X, T77;
|
Chris@10
|
2284 T6X = KP707106781 * (T6R - T6W);
|
Chris@10
|
2285 T6Y = T6M - T6X;
|
Chris@10
|
2286 T7a = T6M + T6X;
|
Chris@10
|
2287 T77 = KP707106781 * (T75 - T76);
|
Chris@10
|
2288 T78 = T74 - T77;
|
Chris@10
|
2289 T7c = T74 + T77;
|
Chris@10
|
2290 {
|
Chris@10
|
2291 E T6H, T6Z, T79, T7b;
|
Chris@10
|
2292 T6H = W[12];
|
Chris@10
|
2293 T6Z = W[13];
|
Chris@10
|
2294 iio[WS(vs, 7) + WS(rs, 5)] = FNMS(T6Z, T78, T6H * T6Y);
|
Chris@10
|
2295 rio[WS(vs, 7) + WS(rs, 5)] = FMA(T6Z, T6Y, T6H * T78);
|
Chris@10
|
2296 T79 = W[4];
|
Chris@10
|
2297 T7b = W[5];
|
Chris@10
|
2298 iio[WS(vs, 3) + WS(rs, 5)] = FNMS(T7b, T7c, T79 * T7a);
|
Chris@10
|
2299 rio[WS(vs, 3) + WS(rs, 5)] = FMA(T7b, T7a, T79 * T7c);
|
Chris@10
|
2300 }
|
Chris@10
|
2301 }
|
Chris@10
|
2302 {
|
Chris@10
|
2303 E T1O, T20, T1Y, T22, T1N, T1X;
|
Chris@10
|
2304 T1N = KP707106781 * (T1H - T1M);
|
Chris@10
|
2305 T1O = T1C - T1N;
|
Chris@10
|
2306 T20 = T1C + T1N;
|
Chris@10
|
2307 T1X = KP707106781 * (T1V - T1W);
|
Chris@10
|
2308 T1Y = T1U - T1X;
|
Chris@10
|
2309 T22 = T1U + T1X;
|
Chris@10
|
2310 {
|
Chris@10
|
2311 E T1x, T1P, T1Z, T21;
|
Chris@10
|
2312 T1x = W[12];
|
Chris@10
|
2313 T1P = W[13];
|
Chris@10
|
2314 iio[WS(vs, 7) + WS(rs, 1)] = FNMS(T1P, T1Y, T1x * T1O);
|
Chris@10
|
2315 rio[WS(vs, 7) + WS(rs, 1)] = FMA(T1P, T1O, T1x * T1Y);
|
Chris@10
|
2316 T1Z = W[4];
|
Chris@10
|
2317 T21 = W[5];
|
Chris@10
|
2318 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T21, T22, T1Z * T20);
|
Chris@10
|
2319 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T21, T20, T1Z * T22);
|
Chris@10
|
2320 }
|
Chris@10
|
2321 }
|
Chris@10
|
2322 {
|
Chris@10
|
2323 E T4o, T4A, T4y, T4C, T4n, T4x;
|
Chris@10
|
2324 T4n = KP707106781 * (T4h - T4m);
|
Chris@10
|
2325 T4o = T4c - T4n;
|
Chris@10
|
2326 T4A = T4c + T4n;
|
Chris@10
|
2327 T4x = KP707106781 * (T4v - T4w);
|
Chris@10
|
2328 T4y = T4u - T4x;
|
Chris@10
|
2329 T4C = T4u + T4x;
|
Chris@10
|
2330 {
|
Chris@10
|
2331 E T47, T4p, T4z, T4B;
|
Chris@10
|
2332 T47 = W[12];
|
Chris@10
|
2333 T4p = W[13];
|
Chris@10
|
2334 iio[WS(vs, 7) + WS(rs, 3)] = FNMS(T4p, T4y, T47 * T4o);
|
Chris@10
|
2335 rio[WS(vs, 7) + WS(rs, 3)] = FMA(T4p, T4o, T47 * T4y);
|
Chris@10
|
2336 T4z = W[4];
|
Chris@10
|
2337 T4B = W[5];
|
Chris@10
|
2338 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T4B, T4C, T4z * T4A);
|
Chris@10
|
2339 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T4B, T4A, T4z * T4C);
|
Chris@10
|
2340 }
|
Chris@10
|
2341 }
|
Chris@10
|
2342 {
|
Chris@10
|
2343 E T3o, T3u, T3s, T3w, T3n, T3r;
|
Chris@10
|
2344 T3n = KP707106781 * (T34 + T2Z);
|
Chris@10
|
2345 T3o = T3m - T3n;
|
Chris@10
|
2346 T3u = T3m + T3n;
|
Chris@10
|
2347 T3r = KP707106781 * (T3d + T3e);
|
Chris@10
|
2348 T3s = T3q - T3r;
|
Chris@10
|
2349 T3w = T3q + T3r;
|
Chris@10
|
2350 {
|
Chris@10
|
2351 E T3l, T3p, T3t, T3v;
|
Chris@10
|
2352 T3l = W[8];
|
Chris@10
|
2353 T3p = W[9];
|
Chris@10
|
2354 rio[WS(vs, 5) + WS(rs, 2)] = FMA(T3l, T3o, T3p * T3s);
|
Chris@10
|
2355 iio[WS(vs, 5) + WS(rs, 2)] = FNMS(T3p, T3o, T3l * T3s);
|
Chris@10
|
2356 T3t = W[0];
|
Chris@10
|
2357 T3v = W[1];
|
Chris@10
|
2358 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T3t, T3u, T3v * T3w);
|
Chris@10
|
2359 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T3v, T3u, T3t * T3w);
|
Chris@10
|
2360 }
|
Chris@10
|
2361 }
|
Chris@10
|
2362 {
|
Chris@10
|
2363 E T8y, T8E, T8C, T8G, T8x, T8B;
|
Chris@10
|
2364 T8x = KP707106781 * (T8e + T89);
|
Chris@10
|
2365 T8y = T8w - T8x;
|
Chris@10
|
2366 T8E = T8w + T8x;
|
Chris@10
|
2367 T8B = KP707106781 * (T8n + T8o);
|
Chris@10
|
2368 T8C = T8A - T8B;
|
Chris@10
|
2369 T8G = T8A + T8B;
|
Chris@10
|
2370 {
|
Chris@10
|
2371 E T8v, T8z, T8D, T8F;
|
Chris@10
|
2372 T8v = W[8];
|
Chris@10
|
2373 T8z = W[9];
|
Chris@10
|
2374 rio[WS(vs, 5) + WS(rs, 6)] = FMA(T8v, T8y, T8z * T8C);
|
Chris@10
|
2375 iio[WS(vs, 5) + WS(rs, 6)] = FNMS(T8z, T8y, T8v * T8C);
|
Chris@10
|
2376 T8D = W[0];
|
Chris@10
|
2377 T8F = W[1];
|
Chris@10
|
2378 rio[WS(vs, 1) + WS(rs, 6)] = FMA(T8D, T8E, T8F * T8G);
|
Chris@10
|
2379 iio[WS(vs, 1) + WS(rs, 6)] = FNMS(T8F, T8E, T8D * T8G);
|
Chris@10
|
2380 }
|
Chris@10
|
2381 }
|
Chris@10
|
2382 }
|
Chris@10
|
2383 }
|
Chris@10
|
2384 }
|
Chris@10
|
2385
|
Chris@10
|
2386 static const tw_instr twinstr[] = {
|
Chris@10
|
2387 {TW_FULL, 0, 8},
|
Chris@10
|
2388 {TW_NEXT, 1, 0}
|
Chris@10
|
2389 };
|
Chris@10
|
2390
|
Chris@10
|
2391 static const ct_desc desc = { 8, "q1_8", twinstr, &GENUS, {416, 144, 112, 0}, 0, 0, 0 };
|
Chris@10
|
2392
|
Chris@10
|
2393 void X(codelet_q1_8) (planner *p) {
|
Chris@10
|
2394 X(kdft_difsq_register) (p, q1_8, &desc);
|
Chris@10
|
2395 }
|
Chris@10
|
2396 #endif /* HAVE_FMA */
|