cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:04:30 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "dft/codelet-dft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_twidsq.native -fma -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 8 -name q1_8 -include dft/scalar/q.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 528 FP additions, 288 FP multiplications,
|
cannam@167
|
32 * (or, 352 additions, 112 multiplications, 176 fused multiply/add),
|
cannam@167
|
33 * 152 stack variables, 1 constants, and 256 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "dft/scalar/q.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void q1_8(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
40 {
|
cannam@167
|
41 INT m;
|
cannam@167
|
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)) {
|
cannam@167
|
43 E T7, T1d, T1t, Tk, TD, TV, T18, TQ, T4F, T5L, T61, T4S, T5b, T5t, T5G;
|
cannam@167
|
44 E T5o, T6b, T7h, T7x, T6o, T6H, T6Z, T7c, T6U, TaJ, TbP, Tc5, TaW, Tbf, Tbx;
|
cannam@167
|
45 E TbK, Tbs, T1D, T2J, T2Z, T1Q, T29, T2r, T2E, T2m, T39, T4f, T4v, T3m, T3F;
|
cannam@167
|
46 E T3X, T4a, T3S, T7H, T8N, T93, T7U, T8d, T8v, T8I, T8q, T9d, Taj, Taz, T9q;
|
cannam@167
|
47 E T9J, Ta1, Tae, T9W, Te, T19, T1u, T1g, TE, TF, TW, Tv, TR, T4M, T5H;
|
cannam@167
|
48 E T62, T5O, T5c, T5d, T5u, T53, T5p, T6i, T7d, T7y, T7k, T6I, T6J, T70, T6z;
|
cannam@167
|
49 E T6V, TaQ, TbL, Tc6, TbS, Tbg, Tbh, Tby, Tb7, Tbt, T1K, T2F, T30, T2M, T2a;
|
cannam@167
|
50 E T2b, T2s, T21, T2n, T3g, T4b, T4w, T4i, T3G, T3H, T3Y, T3x, T3T, T7O, T8J;
|
cannam@167
|
51 E T94, T8Q, T8e, T8f, T8w, T85, T8r, T9k, Taf, TaA, Tam, T9K, T9L, Ta2, T9B;
|
cannam@167
|
52 E T9X;
|
cannam@167
|
53 {
|
cannam@167
|
54 E T3, Tz, Tj, T16, T6, Tg, TC, T17;
|
cannam@167
|
55 {
|
cannam@167
|
56 E T1, T2, Th, Ti;
|
cannam@167
|
57 T1 = rio[0];
|
cannam@167
|
58 T2 = rio[WS(rs, 4)];
|
cannam@167
|
59 T3 = T1 + T2;
|
cannam@167
|
60 Tz = T1 - T2;
|
cannam@167
|
61 Th = iio[0];
|
cannam@167
|
62 Ti = iio[WS(rs, 4)];
|
cannam@167
|
63 Tj = Th - Ti;
|
cannam@167
|
64 T16 = Th + Ti;
|
cannam@167
|
65 }
|
cannam@167
|
66 {
|
cannam@167
|
67 E T4, T5, TA, TB;
|
cannam@167
|
68 T4 = rio[WS(rs, 2)];
|
cannam@167
|
69 T5 = rio[WS(rs, 6)];
|
cannam@167
|
70 T6 = T4 + T5;
|
cannam@167
|
71 Tg = T4 - T5;
|
cannam@167
|
72 TA = iio[WS(rs, 2)];
|
cannam@167
|
73 TB = iio[WS(rs, 6)];
|
cannam@167
|
74 TC = TA - TB;
|
cannam@167
|
75 T17 = TA + TB;
|
cannam@167
|
76 }
|
cannam@167
|
77 T7 = T3 + T6;
|
cannam@167
|
78 T1d = T3 - T6;
|
cannam@167
|
79 T1t = T16 + T17;
|
cannam@167
|
80 Tk = Tg + Tj;
|
cannam@167
|
81 TD = Tz - TC;
|
cannam@167
|
82 TV = Tj - Tg;
|
cannam@167
|
83 T18 = T16 - T17;
|
cannam@167
|
84 TQ = Tz + TC;
|
cannam@167
|
85 }
|
cannam@167
|
86 {
|
cannam@167
|
87 E T4B, T57, T4R, T5E, T4E, T4O, T5a, T5F;
|
cannam@167
|
88 {
|
cannam@167
|
89 E T4z, T4A, T4P, T4Q;
|
cannam@167
|
90 T4z = rio[WS(vs, 3)];
|
cannam@167
|
91 T4A = rio[WS(vs, 3) + WS(rs, 4)];
|
cannam@167
|
92 T4B = T4z + T4A;
|
cannam@167
|
93 T57 = T4z - T4A;
|
cannam@167
|
94 T4P = iio[WS(vs, 3)];
|
cannam@167
|
95 T4Q = iio[WS(vs, 3) + WS(rs, 4)];
|
cannam@167
|
96 T4R = T4P - T4Q;
|
cannam@167
|
97 T5E = T4P + T4Q;
|
cannam@167
|
98 }
|
cannam@167
|
99 {
|
cannam@167
|
100 E T4C, T4D, T58, T59;
|
cannam@167
|
101 T4C = rio[WS(vs, 3) + WS(rs, 2)];
|
cannam@167
|
102 T4D = rio[WS(vs, 3) + WS(rs, 6)];
|
cannam@167
|
103 T4E = T4C + T4D;
|
cannam@167
|
104 T4O = T4C - T4D;
|
cannam@167
|
105 T58 = iio[WS(vs, 3) + WS(rs, 2)];
|
cannam@167
|
106 T59 = iio[WS(vs, 3) + WS(rs, 6)];
|
cannam@167
|
107 T5a = T58 - T59;
|
cannam@167
|
108 T5F = T58 + T59;
|
cannam@167
|
109 }
|
cannam@167
|
110 T4F = T4B + T4E;
|
cannam@167
|
111 T5L = T4B - T4E;
|
cannam@167
|
112 T61 = T5E + T5F;
|
cannam@167
|
113 T4S = T4O + T4R;
|
cannam@167
|
114 T5b = T57 - T5a;
|
cannam@167
|
115 T5t = T4R - T4O;
|
cannam@167
|
116 T5G = T5E - T5F;
|
cannam@167
|
117 T5o = T57 + T5a;
|
cannam@167
|
118 }
|
cannam@167
|
119 {
|
cannam@167
|
120 E T67, T6D, T6n, T7a, T6a, T6k, T6G, T7b;
|
cannam@167
|
121 {
|
cannam@167
|
122 E T65, T66, T6l, T6m;
|
cannam@167
|
123 T65 = rio[WS(vs, 4)];
|
cannam@167
|
124 T66 = rio[WS(vs, 4) + WS(rs, 4)];
|
cannam@167
|
125 T67 = T65 + T66;
|
cannam@167
|
126 T6D = T65 - T66;
|
cannam@167
|
127 T6l = iio[WS(vs, 4)];
|
cannam@167
|
128 T6m = iio[WS(vs, 4) + WS(rs, 4)];
|
cannam@167
|
129 T6n = T6l - T6m;
|
cannam@167
|
130 T7a = T6l + T6m;
|
cannam@167
|
131 }
|
cannam@167
|
132 {
|
cannam@167
|
133 E T68, T69, T6E, T6F;
|
cannam@167
|
134 T68 = rio[WS(vs, 4) + WS(rs, 2)];
|
cannam@167
|
135 T69 = rio[WS(vs, 4) + WS(rs, 6)];
|
cannam@167
|
136 T6a = T68 + T69;
|
cannam@167
|
137 T6k = T68 - T69;
|
cannam@167
|
138 T6E = iio[WS(vs, 4) + WS(rs, 2)];
|
cannam@167
|
139 T6F = iio[WS(vs, 4) + WS(rs, 6)];
|
cannam@167
|
140 T6G = T6E - T6F;
|
cannam@167
|
141 T7b = T6E + T6F;
|
cannam@167
|
142 }
|
cannam@167
|
143 T6b = T67 + T6a;
|
cannam@167
|
144 T7h = T67 - T6a;
|
cannam@167
|
145 T7x = T7a + T7b;
|
cannam@167
|
146 T6o = T6k + T6n;
|
cannam@167
|
147 T6H = T6D - T6G;
|
cannam@167
|
148 T6Z = T6n - T6k;
|
cannam@167
|
149 T7c = T7a - T7b;
|
cannam@167
|
150 T6U = T6D + T6G;
|
cannam@167
|
151 }
|
cannam@167
|
152 {
|
cannam@167
|
153 E TaF, Tbb, TaV, TbI, TaI, TaS, Tbe, TbJ;
|
cannam@167
|
154 {
|
cannam@167
|
155 E TaD, TaE, TaT, TaU;
|
cannam@167
|
156 TaD = rio[WS(vs, 7)];
|
cannam@167
|
157 TaE = rio[WS(vs, 7) + WS(rs, 4)];
|
cannam@167
|
158 TaF = TaD + TaE;
|
cannam@167
|
159 Tbb = TaD - TaE;
|
cannam@167
|
160 TaT = iio[WS(vs, 7)];
|
cannam@167
|
161 TaU = iio[WS(vs, 7) + WS(rs, 4)];
|
cannam@167
|
162 TaV = TaT - TaU;
|
cannam@167
|
163 TbI = TaT + TaU;
|
cannam@167
|
164 }
|
cannam@167
|
165 {
|
cannam@167
|
166 E TaG, TaH, Tbc, Tbd;
|
cannam@167
|
167 TaG = rio[WS(vs, 7) + WS(rs, 2)];
|
cannam@167
|
168 TaH = rio[WS(vs, 7) + WS(rs, 6)];
|
cannam@167
|
169 TaI = TaG + TaH;
|
cannam@167
|
170 TaS = TaG - TaH;
|
cannam@167
|
171 Tbc = iio[WS(vs, 7) + WS(rs, 2)];
|
cannam@167
|
172 Tbd = iio[WS(vs, 7) + WS(rs, 6)];
|
cannam@167
|
173 Tbe = Tbc - Tbd;
|
cannam@167
|
174 TbJ = Tbc + Tbd;
|
cannam@167
|
175 }
|
cannam@167
|
176 TaJ = TaF + TaI;
|
cannam@167
|
177 TbP = TaF - TaI;
|
cannam@167
|
178 Tc5 = TbI + TbJ;
|
cannam@167
|
179 TaW = TaS + TaV;
|
cannam@167
|
180 Tbf = Tbb - Tbe;
|
cannam@167
|
181 Tbx = TaV - TaS;
|
cannam@167
|
182 TbK = TbI - TbJ;
|
cannam@167
|
183 Tbs = Tbb + Tbe;
|
cannam@167
|
184 }
|
cannam@167
|
185 {
|
cannam@167
|
186 E T1z, T25, T1P, T2C, T1C, T1M, T28, T2D;
|
cannam@167
|
187 {
|
cannam@167
|
188 E T1x, T1y, T1N, T1O;
|
cannam@167
|
189 T1x = rio[WS(vs, 1)];
|
cannam@167
|
190 T1y = rio[WS(vs, 1) + WS(rs, 4)];
|
cannam@167
|
191 T1z = T1x + T1y;
|
cannam@167
|
192 T25 = T1x - T1y;
|
cannam@167
|
193 T1N = iio[WS(vs, 1)];
|
cannam@167
|
194 T1O = iio[WS(vs, 1) + WS(rs, 4)];
|
cannam@167
|
195 T1P = T1N - T1O;
|
cannam@167
|
196 T2C = T1N + T1O;
|
cannam@167
|
197 }
|
cannam@167
|
198 {
|
cannam@167
|
199 E T1A, T1B, T26, T27;
|
cannam@167
|
200 T1A = rio[WS(vs, 1) + WS(rs, 2)];
|
cannam@167
|
201 T1B = rio[WS(vs, 1) + WS(rs, 6)];
|
cannam@167
|
202 T1C = T1A + T1B;
|
cannam@167
|
203 T1M = T1A - T1B;
|
cannam@167
|
204 T26 = iio[WS(vs, 1) + WS(rs, 2)];
|
cannam@167
|
205 T27 = iio[WS(vs, 1) + WS(rs, 6)];
|
cannam@167
|
206 T28 = T26 - T27;
|
cannam@167
|
207 T2D = T26 + T27;
|
cannam@167
|
208 }
|
cannam@167
|
209 T1D = T1z + T1C;
|
cannam@167
|
210 T2J = T1z - T1C;
|
cannam@167
|
211 T2Z = T2C + T2D;
|
cannam@167
|
212 T1Q = T1M + T1P;
|
cannam@167
|
213 T29 = T25 - T28;
|
cannam@167
|
214 T2r = T1P - T1M;
|
cannam@167
|
215 T2E = T2C - T2D;
|
cannam@167
|
216 T2m = T25 + T28;
|
cannam@167
|
217 }
|
cannam@167
|
218 {
|
cannam@167
|
219 E T35, T3B, T3l, T48, T38, T3i, T3E, T49;
|
cannam@167
|
220 {
|
cannam@167
|
221 E T33, T34, T3j, T3k;
|
cannam@167
|
222 T33 = rio[WS(vs, 2)];
|
cannam@167
|
223 T34 = rio[WS(vs, 2) + WS(rs, 4)];
|
cannam@167
|
224 T35 = T33 + T34;
|
cannam@167
|
225 T3B = T33 - T34;
|
cannam@167
|
226 T3j = iio[WS(vs, 2)];
|
cannam@167
|
227 T3k = iio[WS(vs, 2) + WS(rs, 4)];
|
cannam@167
|
228 T3l = T3j - T3k;
|
cannam@167
|
229 T48 = T3j + T3k;
|
cannam@167
|
230 }
|
cannam@167
|
231 {
|
cannam@167
|
232 E T36, T37, T3C, T3D;
|
cannam@167
|
233 T36 = rio[WS(vs, 2) + WS(rs, 2)];
|
cannam@167
|
234 T37 = rio[WS(vs, 2) + WS(rs, 6)];
|
cannam@167
|
235 T38 = T36 + T37;
|
cannam@167
|
236 T3i = T36 - T37;
|
cannam@167
|
237 T3C = iio[WS(vs, 2) + WS(rs, 2)];
|
cannam@167
|
238 T3D = iio[WS(vs, 2) + WS(rs, 6)];
|
cannam@167
|
239 T3E = T3C - T3D;
|
cannam@167
|
240 T49 = T3C + T3D;
|
cannam@167
|
241 }
|
cannam@167
|
242 T39 = T35 + T38;
|
cannam@167
|
243 T4f = T35 - T38;
|
cannam@167
|
244 T4v = T48 + T49;
|
cannam@167
|
245 T3m = T3i + T3l;
|
cannam@167
|
246 T3F = T3B - T3E;
|
cannam@167
|
247 T3X = T3l - T3i;
|
cannam@167
|
248 T4a = T48 - T49;
|
cannam@167
|
249 T3S = T3B + T3E;
|
cannam@167
|
250 }
|
cannam@167
|
251 {
|
cannam@167
|
252 E T7D, T89, T7T, T8G, T7G, T7Q, T8c, T8H;
|
cannam@167
|
253 {
|
cannam@167
|
254 E T7B, T7C, T7R, T7S;
|
cannam@167
|
255 T7B = rio[WS(vs, 5)];
|
cannam@167
|
256 T7C = rio[WS(vs, 5) + WS(rs, 4)];
|
cannam@167
|
257 T7D = T7B + T7C;
|
cannam@167
|
258 T89 = T7B - T7C;
|
cannam@167
|
259 T7R = iio[WS(vs, 5)];
|
cannam@167
|
260 T7S = iio[WS(vs, 5) + WS(rs, 4)];
|
cannam@167
|
261 T7T = T7R - T7S;
|
cannam@167
|
262 T8G = T7R + T7S;
|
cannam@167
|
263 }
|
cannam@167
|
264 {
|
cannam@167
|
265 E T7E, T7F, T8a, T8b;
|
cannam@167
|
266 T7E = rio[WS(vs, 5) + WS(rs, 2)];
|
cannam@167
|
267 T7F = rio[WS(vs, 5) + WS(rs, 6)];
|
cannam@167
|
268 T7G = T7E + T7F;
|
cannam@167
|
269 T7Q = T7E - T7F;
|
cannam@167
|
270 T8a = iio[WS(vs, 5) + WS(rs, 2)];
|
cannam@167
|
271 T8b = iio[WS(vs, 5) + WS(rs, 6)];
|
cannam@167
|
272 T8c = T8a - T8b;
|
cannam@167
|
273 T8H = T8a + T8b;
|
cannam@167
|
274 }
|
cannam@167
|
275 T7H = T7D + T7G;
|
cannam@167
|
276 T8N = T7D - T7G;
|
cannam@167
|
277 T93 = T8G + T8H;
|
cannam@167
|
278 T7U = T7Q + T7T;
|
cannam@167
|
279 T8d = T89 - T8c;
|
cannam@167
|
280 T8v = T7T - T7Q;
|
cannam@167
|
281 T8I = T8G - T8H;
|
cannam@167
|
282 T8q = T89 + T8c;
|
cannam@167
|
283 }
|
cannam@167
|
284 {
|
cannam@167
|
285 E T99, T9F, T9p, Tac, T9c, T9m, T9I, Tad;
|
cannam@167
|
286 {
|
cannam@167
|
287 E T97, T98, T9n, T9o;
|
cannam@167
|
288 T97 = rio[WS(vs, 6)];
|
cannam@167
|
289 T98 = rio[WS(vs, 6) + WS(rs, 4)];
|
cannam@167
|
290 T99 = T97 + T98;
|
cannam@167
|
291 T9F = T97 - T98;
|
cannam@167
|
292 T9n = iio[WS(vs, 6)];
|
cannam@167
|
293 T9o = iio[WS(vs, 6) + WS(rs, 4)];
|
cannam@167
|
294 T9p = T9n - T9o;
|
cannam@167
|
295 Tac = T9n + T9o;
|
cannam@167
|
296 }
|
cannam@167
|
297 {
|
cannam@167
|
298 E T9a, T9b, T9G, T9H;
|
cannam@167
|
299 T9a = rio[WS(vs, 6) + WS(rs, 2)];
|
cannam@167
|
300 T9b = rio[WS(vs, 6) + WS(rs, 6)];
|
cannam@167
|
301 T9c = T9a + T9b;
|
cannam@167
|
302 T9m = T9a - T9b;
|
cannam@167
|
303 T9G = iio[WS(vs, 6) + WS(rs, 2)];
|
cannam@167
|
304 T9H = iio[WS(vs, 6) + WS(rs, 6)];
|
cannam@167
|
305 T9I = T9G - T9H;
|
cannam@167
|
306 Tad = T9G + T9H;
|
cannam@167
|
307 }
|
cannam@167
|
308 T9d = T99 + T9c;
|
cannam@167
|
309 Taj = T99 - T9c;
|
cannam@167
|
310 Taz = Tac + Tad;
|
cannam@167
|
311 T9q = T9m + T9p;
|
cannam@167
|
312 T9J = T9F - T9I;
|
cannam@167
|
313 Ta1 = T9p - T9m;
|
cannam@167
|
314 Tae = Tac - Tad;
|
cannam@167
|
315 T9W = T9F + T9I;
|
cannam@167
|
316 }
|
cannam@167
|
317 {
|
cannam@167
|
318 E Ta, Tq, Tt, T1e, Td, Tl, To, T1f, Tp, Tu;
|
cannam@167
|
319 {
|
cannam@167
|
320 E T8, T9, Tr, Ts;
|
cannam@167
|
321 T8 = rio[WS(rs, 1)];
|
cannam@167
|
322 T9 = rio[WS(rs, 5)];
|
cannam@167
|
323 Ta = T8 + T9;
|
cannam@167
|
324 Tq = T8 - T9;
|
cannam@167
|
325 Tr = iio[WS(rs, 1)];
|
cannam@167
|
326 Ts = iio[WS(rs, 5)];
|
cannam@167
|
327 Tt = Tr - Ts;
|
cannam@167
|
328 T1e = Tr + Ts;
|
cannam@167
|
329 }
|
cannam@167
|
330 {
|
cannam@167
|
331 E Tb, Tc, Tm, Tn;
|
cannam@167
|
332 Tb = rio[WS(rs, 7)];
|
cannam@167
|
333 Tc = rio[WS(rs, 3)];
|
cannam@167
|
334 Td = Tb + Tc;
|
cannam@167
|
335 Tl = Tb - Tc;
|
cannam@167
|
336 Tm = iio[WS(rs, 7)];
|
cannam@167
|
337 Tn = iio[WS(rs, 3)];
|
cannam@167
|
338 To = Tm - Tn;
|
cannam@167
|
339 T1f = Tm + Tn;
|
cannam@167
|
340 }
|
cannam@167
|
341 Te = Ta + Td;
|
cannam@167
|
342 T19 = Td - Ta;
|
cannam@167
|
343 T1u = T1e + T1f;
|
cannam@167
|
344 T1g = T1e - T1f;
|
cannam@167
|
345 TE = Tt - Tq;
|
cannam@167
|
346 TF = Tl + To;
|
cannam@167
|
347 TW = TE + TF;
|
cannam@167
|
348 Tp = Tl - To;
|
cannam@167
|
349 Tu = Tq + Tt;
|
cannam@167
|
350 Tv = Tp - Tu;
|
cannam@167
|
351 TR = Tu + Tp;
|
cannam@167
|
352 }
|
cannam@167
|
353 {
|
cannam@167
|
354 E T4I, T4Y, T51, T5M, T4L, T4T, T4W, T5N, T4X, T52;
|
cannam@167
|
355 {
|
cannam@167
|
356 E T4G, T4H, T4Z, T50;
|
cannam@167
|
357 T4G = rio[WS(vs, 3) + WS(rs, 1)];
|
cannam@167
|
358 T4H = rio[WS(vs, 3) + WS(rs, 5)];
|
cannam@167
|
359 T4I = T4G + T4H;
|
cannam@167
|
360 T4Y = T4G - T4H;
|
cannam@167
|
361 T4Z = iio[WS(vs, 3) + WS(rs, 1)];
|
cannam@167
|
362 T50 = iio[WS(vs, 3) + WS(rs, 5)];
|
cannam@167
|
363 T51 = T4Z - T50;
|
cannam@167
|
364 T5M = T4Z + T50;
|
cannam@167
|
365 }
|
cannam@167
|
366 {
|
cannam@167
|
367 E T4J, T4K, T4U, T4V;
|
cannam@167
|
368 T4J = rio[WS(vs, 3) + WS(rs, 7)];
|
cannam@167
|
369 T4K = rio[WS(vs, 3) + WS(rs, 3)];
|
cannam@167
|
370 T4L = T4J + T4K;
|
cannam@167
|
371 T4T = T4J - T4K;
|
cannam@167
|
372 T4U = iio[WS(vs, 3) + WS(rs, 7)];
|
cannam@167
|
373 T4V = iio[WS(vs, 3) + WS(rs, 3)];
|
cannam@167
|
374 T4W = T4U - T4V;
|
cannam@167
|
375 T5N = T4U + T4V;
|
cannam@167
|
376 }
|
cannam@167
|
377 T4M = T4I + T4L;
|
cannam@167
|
378 T5H = T4L - T4I;
|
cannam@167
|
379 T62 = T5M + T5N;
|
cannam@167
|
380 T5O = T5M - T5N;
|
cannam@167
|
381 T5c = T51 - T4Y;
|
cannam@167
|
382 T5d = T4T + T4W;
|
cannam@167
|
383 T5u = T5c + T5d;
|
cannam@167
|
384 T4X = T4T - T4W;
|
cannam@167
|
385 T52 = T4Y + T51;
|
cannam@167
|
386 T53 = T4X - T52;
|
cannam@167
|
387 T5p = T52 + T4X;
|
cannam@167
|
388 }
|
cannam@167
|
389 {
|
cannam@167
|
390 E T6e, T6u, T6x, T7i, T6h, T6p, T6s, T7j, T6t, T6y;
|
cannam@167
|
391 {
|
cannam@167
|
392 E T6c, T6d, T6v, T6w;
|
cannam@167
|
393 T6c = rio[WS(vs, 4) + WS(rs, 1)];
|
cannam@167
|
394 T6d = rio[WS(vs, 4) + WS(rs, 5)];
|
cannam@167
|
395 T6e = T6c + T6d;
|
cannam@167
|
396 T6u = T6c - T6d;
|
cannam@167
|
397 T6v = iio[WS(vs, 4) + WS(rs, 1)];
|
cannam@167
|
398 T6w = iio[WS(vs, 4) + WS(rs, 5)];
|
cannam@167
|
399 T6x = T6v - T6w;
|
cannam@167
|
400 T7i = T6v + T6w;
|
cannam@167
|
401 }
|
cannam@167
|
402 {
|
cannam@167
|
403 E T6f, T6g, T6q, T6r;
|
cannam@167
|
404 T6f = rio[WS(vs, 4) + WS(rs, 7)];
|
cannam@167
|
405 T6g = rio[WS(vs, 4) + WS(rs, 3)];
|
cannam@167
|
406 T6h = T6f + T6g;
|
cannam@167
|
407 T6p = T6f - T6g;
|
cannam@167
|
408 T6q = iio[WS(vs, 4) + WS(rs, 7)];
|
cannam@167
|
409 T6r = iio[WS(vs, 4) + WS(rs, 3)];
|
cannam@167
|
410 T6s = T6q - T6r;
|
cannam@167
|
411 T7j = T6q + T6r;
|
cannam@167
|
412 }
|
cannam@167
|
413 T6i = T6e + T6h;
|
cannam@167
|
414 T7d = T6h - T6e;
|
cannam@167
|
415 T7y = T7i + T7j;
|
cannam@167
|
416 T7k = T7i - T7j;
|
cannam@167
|
417 T6I = T6x - T6u;
|
cannam@167
|
418 T6J = T6p + T6s;
|
cannam@167
|
419 T70 = T6I + T6J;
|
cannam@167
|
420 T6t = T6p - T6s;
|
cannam@167
|
421 T6y = T6u + T6x;
|
cannam@167
|
422 T6z = T6t - T6y;
|
cannam@167
|
423 T6V = T6y + T6t;
|
cannam@167
|
424 }
|
cannam@167
|
425 {
|
cannam@167
|
426 E TaM, Tb2, Tb5, TbQ, TaP, TaX, Tb0, TbR, Tb1, Tb6;
|
cannam@167
|
427 {
|
cannam@167
|
428 E TaK, TaL, Tb3, Tb4;
|
cannam@167
|
429 TaK = rio[WS(vs, 7) + WS(rs, 1)];
|
cannam@167
|
430 TaL = rio[WS(vs, 7) + WS(rs, 5)];
|
cannam@167
|
431 TaM = TaK + TaL;
|
cannam@167
|
432 Tb2 = TaK - TaL;
|
cannam@167
|
433 Tb3 = iio[WS(vs, 7) + WS(rs, 1)];
|
cannam@167
|
434 Tb4 = iio[WS(vs, 7) + WS(rs, 5)];
|
cannam@167
|
435 Tb5 = Tb3 - Tb4;
|
cannam@167
|
436 TbQ = Tb3 + Tb4;
|
cannam@167
|
437 }
|
cannam@167
|
438 {
|
cannam@167
|
439 E TaN, TaO, TaY, TaZ;
|
cannam@167
|
440 TaN = rio[WS(vs, 7) + WS(rs, 7)];
|
cannam@167
|
441 TaO = rio[WS(vs, 7) + WS(rs, 3)];
|
cannam@167
|
442 TaP = TaN + TaO;
|
cannam@167
|
443 TaX = TaN - TaO;
|
cannam@167
|
444 TaY = iio[WS(vs, 7) + WS(rs, 7)];
|
cannam@167
|
445 TaZ = iio[WS(vs, 7) + WS(rs, 3)];
|
cannam@167
|
446 Tb0 = TaY - TaZ;
|
cannam@167
|
447 TbR = TaY + TaZ;
|
cannam@167
|
448 }
|
cannam@167
|
449 TaQ = TaM + TaP;
|
cannam@167
|
450 TbL = TaP - TaM;
|
cannam@167
|
451 Tc6 = TbQ + TbR;
|
cannam@167
|
452 TbS = TbQ - TbR;
|
cannam@167
|
453 Tbg = Tb5 - Tb2;
|
cannam@167
|
454 Tbh = TaX + Tb0;
|
cannam@167
|
455 Tby = Tbg + Tbh;
|
cannam@167
|
456 Tb1 = TaX - Tb0;
|
cannam@167
|
457 Tb6 = Tb2 + Tb5;
|
cannam@167
|
458 Tb7 = Tb1 - Tb6;
|
cannam@167
|
459 Tbt = Tb6 + Tb1;
|
cannam@167
|
460 }
|
cannam@167
|
461 {
|
cannam@167
|
462 E T1G, T1W, T1Z, T2K, T1J, T1R, T1U, T2L, T1V, T20;
|
cannam@167
|
463 {
|
cannam@167
|
464 E T1E, T1F, T1X, T1Y;
|
cannam@167
|
465 T1E = rio[WS(vs, 1) + WS(rs, 1)];
|
cannam@167
|
466 T1F = rio[WS(vs, 1) + WS(rs, 5)];
|
cannam@167
|
467 T1G = T1E + T1F;
|
cannam@167
|
468 T1W = T1E - T1F;
|
cannam@167
|
469 T1X = iio[WS(vs, 1) + WS(rs, 1)];
|
cannam@167
|
470 T1Y = iio[WS(vs, 1) + WS(rs, 5)];
|
cannam@167
|
471 T1Z = T1X - T1Y;
|
cannam@167
|
472 T2K = T1X + T1Y;
|
cannam@167
|
473 }
|
cannam@167
|
474 {
|
cannam@167
|
475 E T1H, T1I, T1S, T1T;
|
cannam@167
|
476 T1H = rio[WS(vs, 1) + WS(rs, 7)];
|
cannam@167
|
477 T1I = rio[WS(vs, 1) + WS(rs, 3)];
|
cannam@167
|
478 T1J = T1H + T1I;
|
cannam@167
|
479 T1R = T1H - T1I;
|
cannam@167
|
480 T1S = iio[WS(vs, 1) + WS(rs, 7)];
|
cannam@167
|
481 T1T = iio[WS(vs, 1) + WS(rs, 3)];
|
cannam@167
|
482 T1U = T1S - T1T;
|
cannam@167
|
483 T2L = T1S + T1T;
|
cannam@167
|
484 }
|
cannam@167
|
485 T1K = T1G + T1J;
|
cannam@167
|
486 T2F = T1J - T1G;
|
cannam@167
|
487 T30 = T2K + T2L;
|
cannam@167
|
488 T2M = T2K - T2L;
|
cannam@167
|
489 T2a = T1Z - T1W;
|
cannam@167
|
490 T2b = T1R + T1U;
|
cannam@167
|
491 T2s = T2a + T2b;
|
cannam@167
|
492 T1V = T1R - T1U;
|
cannam@167
|
493 T20 = T1W + T1Z;
|
cannam@167
|
494 T21 = T1V - T20;
|
cannam@167
|
495 T2n = T20 + T1V;
|
cannam@167
|
496 }
|
cannam@167
|
497 {
|
cannam@167
|
498 E T3c, T3s, T3v, T4g, T3f, T3n, T3q, T4h, T3r, T3w;
|
cannam@167
|
499 {
|
cannam@167
|
500 E T3a, T3b, T3t, T3u;
|
cannam@167
|
501 T3a = rio[WS(vs, 2) + WS(rs, 1)];
|
cannam@167
|
502 T3b = rio[WS(vs, 2) + WS(rs, 5)];
|
cannam@167
|
503 T3c = T3a + T3b;
|
cannam@167
|
504 T3s = T3a - T3b;
|
cannam@167
|
505 T3t = iio[WS(vs, 2) + WS(rs, 1)];
|
cannam@167
|
506 T3u = iio[WS(vs, 2) + WS(rs, 5)];
|
cannam@167
|
507 T3v = T3t - T3u;
|
cannam@167
|
508 T4g = T3t + T3u;
|
cannam@167
|
509 }
|
cannam@167
|
510 {
|
cannam@167
|
511 E T3d, T3e, T3o, T3p;
|
cannam@167
|
512 T3d = rio[WS(vs, 2) + WS(rs, 7)];
|
cannam@167
|
513 T3e = rio[WS(vs, 2) + WS(rs, 3)];
|
cannam@167
|
514 T3f = T3d + T3e;
|
cannam@167
|
515 T3n = T3d - T3e;
|
cannam@167
|
516 T3o = iio[WS(vs, 2) + WS(rs, 7)];
|
cannam@167
|
517 T3p = iio[WS(vs, 2) + WS(rs, 3)];
|
cannam@167
|
518 T3q = T3o - T3p;
|
cannam@167
|
519 T4h = T3o + T3p;
|
cannam@167
|
520 }
|
cannam@167
|
521 T3g = T3c + T3f;
|
cannam@167
|
522 T4b = T3f - T3c;
|
cannam@167
|
523 T4w = T4g + T4h;
|
cannam@167
|
524 T4i = T4g - T4h;
|
cannam@167
|
525 T3G = T3v - T3s;
|
cannam@167
|
526 T3H = T3n + T3q;
|
cannam@167
|
527 T3Y = T3G + T3H;
|
cannam@167
|
528 T3r = T3n - T3q;
|
cannam@167
|
529 T3w = T3s + T3v;
|
cannam@167
|
530 T3x = T3r - T3w;
|
cannam@167
|
531 T3T = T3w + T3r;
|
cannam@167
|
532 }
|
cannam@167
|
533 {
|
cannam@167
|
534 E T7K, T80, T83, T8O, T7N, T7V, T7Y, T8P, T7Z, T84;
|
cannam@167
|
535 {
|
cannam@167
|
536 E T7I, T7J, T81, T82;
|
cannam@167
|
537 T7I = rio[WS(vs, 5) + WS(rs, 1)];
|
cannam@167
|
538 T7J = rio[WS(vs, 5) + WS(rs, 5)];
|
cannam@167
|
539 T7K = T7I + T7J;
|
cannam@167
|
540 T80 = T7I - T7J;
|
cannam@167
|
541 T81 = iio[WS(vs, 5) + WS(rs, 1)];
|
cannam@167
|
542 T82 = iio[WS(vs, 5) + WS(rs, 5)];
|
cannam@167
|
543 T83 = T81 - T82;
|
cannam@167
|
544 T8O = T81 + T82;
|
cannam@167
|
545 }
|
cannam@167
|
546 {
|
cannam@167
|
547 E T7L, T7M, T7W, T7X;
|
cannam@167
|
548 T7L = rio[WS(vs, 5) + WS(rs, 7)];
|
cannam@167
|
549 T7M = rio[WS(vs, 5) + WS(rs, 3)];
|
cannam@167
|
550 T7N = T7L + T7M;
|
cannam@167
|
551 T7V = T7L - T7M;
|
cannam@167
|
552 T7W = iio[WS(vs, 5) + WS(rs, 7)];
|
cannam@167
|
553 T7X = iio[WS(vs, 5) + WS(rs, 3)];
|
cannam@167
|
554 T7Y = T7W - T7X;
|
cannam@167
|
555 T8P = T7W + T7X;
|
cannam@167
|
556 }
|
cannam@167
|
557 T7O = T7K + T7N;
|
cannam@167
|
558 T8J = T7N - T7K;
|
cannam@167
|
559 T94 = T8O + T8P;
|
cannam@167
|
560 T8Q = T8O - T8P;
|
cannam@167
|
561 T8e = T83 - T80;
|
cannam@167
|
562 T8f = T7V + T7Y;
|
cannam@167
|
563 T8w = T8e + T8f;
|
cannam@167
|
564 T7Z = T7V - T7Y;
|
cannam@167
|
565 T84 = T80 + T83;
|
cannam@167
|
566 T85 = T7Z - T84;
|
cannam@167
|
567 T8r = T84 + T7Z;
|
cannam@167
|
568 }
|
cannam@167
|
569 {
|
cannam@167
|
570 E T9g, T9w, T9z, Tak, T9j, T9r, T9u, Tal, T9v, T9A;
|
cannam@167
|
571 {
|
cannam@167
|
572 E T9e, T9f, T9x, T9y;
|
cannam@167
|
573 T9e = rio[WS(vs, 6) + WS(rs, 1)];
|
cannam@167
|
574 T9f = rio[WS(vs, 6) + WS(rs, 5)];
|
cannam@167
|
575 T9g = T9e + T9f;
|
cannam@167
|
576 T9w = T9e - T9f;
|
cannam@167
|
577 T9x = iio[WS(vs, 6) + WS(rs, 1)];
|
cannam@167
|
578 T9y = iio[WS(vs, 6) + WS(rs, 5)];
|
cannam@167
|
579 T9z = T9x - T9y;
|
cannam@167
|
580 Tak = T9x + T9y;
|
cannam@167
|
581 }
|
cannam@167
|
582 {
|
cannam@167
|
583 E T9h, T9i, T9s, T9t;
|
cannam@167
|
584 T9h = rio[WS(vs, 6) + WS(rs, 7)];
|
cannam@167
|
585 T9i = rio[WS(vs, 6) + WS(rs, 3)];
|
cannam@167
|
586 T9j = T9h + T9i;
|
cannam@167
|
587 T9r = T9h - T9i;
|
cannam@167
|
588 T9s = iio[WS(vs, 6) + WS(rs, 7)];
|
cannam@167
|
589 T9t = iio[WS(vs, 6) + WS(rs, 3)];
|
cannam@167
|
590 T9u = T9s - T9t;
|
cannam@167
|
591 Tal = T9s + T9t;
|
cannam@167
|
592 }
|
cannam@167
|
593 T9k = T9g + T9j;
|
cannam@167
|
594 Taf = T9j - T9g;
|
cannam@167
|
595 TaA = Tak + Tal;
|
cannam@167
|
596 Tam = Tak - Tal;
|
cannam@167
|
597 T9K = T9z - T9w;
|
cannam@167
|
598 T9L = T9r + T9u;
|
cannam@167
|
599 Ta2 = T9K + T9L;
|
cannam@167
|
600 T9v = T9r - T9u;
|
cannam@167
|
601 T9A = T9w + T9z;
|
cannam@167
|
602 T9B = T9v - T9A;
|
cannam@167
|
603 T9X = T9A + T9v;
|
cannam@167
|
604 }
|
cannam@167
|
605 rio[0] = T7 + Te;
|
cannam@167
|
606 iio[0] = T1t + T1u;
|
cannam@167
|
607 rio[WS(rs, 1)] = T1D + T1K;
|
cannam@167
|
608 iio[WS(rs, 1)] = T2Z + T30;
|
cannam@167
|
609 rio[WS(rs, 2)] = T39 + T3g;
|
cannam@167
|
610 iio[WS(rs, 2)] = T4v + T4w;
|
cannam@167
|
611 rio[WS(rs, 3)] = T4F + T4M;
|
cannam@167
|
612 iio[WS(rs, 3)] = T61 + T62;
|
cannam@167
|
613 rio[WS(rs, 4)] = T6b + T6i;
|
cannam@167
|
614 iio[WS(rs, 4)] = T7x + T7y;
|
cannam@167
|
615 rio[WS(rs, 5)] = T7H + T7O;
|
cannam@167
|
616 iio[WS(rs, 5)] = T93 + T94;
|
cannam@167
|
617 rio[WS(rs, 6)] = T9d + T9k;
|
cannam@167
|
618 iio[WS(rs, 6)] = Taz + TaA;
|
cannam@167
|
619 rio[WS(rs, 7)] = TaJ + TaQ;
|
cannam@167
|
620 iio[WS(rs, 7)] = Tc5 + Tc6;
|
cannam@167
|
621 {
|
cannam@167
|
622 E TS, TX, TT, TY, TP, TU;
|
cannam@167
|
623 TS = FNMS(KP707106781, TR, TQ);
|
cannam@167
|
624 TX = FNMS(KP707106781, TW, TV);
|
cannam@167
|
625 TP = W[8];
|
cannam@167
|
626 TT = TP * TS;
|
cannam@167
|
627 TY = TP * TX;
|
cannam@167
|
628 TU = W[9];
|
cannam@167
|
629 rio[WS(vs, 5)] = FMA(TU, TX, TT);
|
cannam@167
|
630 iio[WS(vs, 5)] = FNMS(TU, TS, TY);
|
cannam@167
|
631 }
|
cannam@167
|
632 {
|
cannam@167
|
633 E T2N, T2B, T2H, T2I, T2O, T2G;
|
cannam@167
|
634 T2N = T2J - T2M;
|
cannam@167
|
635 T2G = T2E - T2F;
|
cannam@167
|
636 T2B = W[10];
|
cannam@167
|
637 T2H = T2B * T2G;
|
cannam@167
|
638 T2I = W[11];
|
cannam@167
|
639 T2O = T2I * T2G;
|
cannam@167
|
640 iio[WS(vs, 6) + WS(rs, 1)] = FNMS(T2I, T2N, T2H);
|
cannam@167
|
641 rio[WS(vs, 6) + WS(rs, 1)] = FMA(T2B, T2N, T2O);
|
cannam@167
|
642 }
|
cannam@167
|
643 {
|
cannam@167
|
644 E T1n, T1j, T1l, T1m, T1o, T1k;
|
cannam@167
|
645 T1n = T1d + T1g;
|
cannam@167
|
646 T1k = T19 + T18;
|
cannam@167
|
647 T1j = W[2];
|
cannam@167
|
648 T1l = T1j * T1k;
|
cannam@167
|
649 T1m = W[3];
|
cannam@167
|
650 T1o = T1m * T1k;
|
cannam@167
|
651 iio[WS(vs, 2)] = FNMS(T1m, T1n, T1l);
|
cannam@167
|
652 rio[WS(vs, 2)] = FMA(T1j, T1n, T1o);
|
cannam@167
|
653 }
|
cannam@167
|
654 {
|
cannam@167
|
655 E T1q, T1v, T1r, T1w, T1p, T1s;
|
cannam@167
|
656 T1q = T7 - Te;
|
cannam@167
|
657 T1v = T1t - T1u;
|
cannam@167
|
658 T1p = W[6];
|
cannam@167
|
659 T1r = T1p * T1q;
|
cannam@167
|
660 T1w = T1p * T1v;
|
cannam@167
|
661 T1s = W[7];
|
cannam@167
|
662 rio[WS(vs, 4)] = FMA(T1s, T1v, T1r);
|
cannam@167
|
663 iio[WS(vs, 4)] = FNMS(T1s, T1q, T1w);
|
cannam@167
|
664 }
|
cannam@167
|
665 {
|
cannam@167
|
666 E Tan, Tab, Tah, Tai, Tao, Tag;
|
cannam@167
|
667 Tan = Taj - Tam;
|
cannam@167
|
668 Tag = Tae - Taf;
|
cannam@167
|
669 Tab = W[10];
|
cannam@167
|
670 Tah = Tab * Tag;
|
cannam@167
|
671 Tai = W[11];
|
cannam@167
|
672 Tao = Tai * Tag;
|
cannam@167
|
673 iio[WS(vs, 6) + WS(rs, 6)] = FNMS(Tai, Tan, Tah);
|
cannam@167
|
674 rio[WS(vs, 6) + WS(rs, 6)] = FMA(Tab, Tan, Tao);
|
cannam@167
|
675 }
|
cannam@167
|
676 {
|
cannam@167
|
677 E Tc2, Tc7, Tc3, Tc8, Tc1, Tc4;
|
cannam@167
|
678 Tc2 = TaJ - TaQ;
|
cannam@167
|
679 Tc7 = Tc5 - Tc6;
|
cannam@167
|
680 Tc1 = W[6];
|
cannam@167
|
681 Tc3 = Tc1 * Tc2;
|
cannam@167
|
682 Tc8 = Tc1 * Tc7;
|
cannam@167
|
683 Tc4 = W[7];
|
cannam@167
|
684 rio[WS(vs, 4) + WS(rs, 7)] = FMA(Tc4, Tc7, Tc3);
|
cannam@167
|
685 iio[WS(vs, 4) + WS(rs, 7)] = FNMS(Tc4, Tc2, Tc8);
|
cannam@167
|
686 }
|
cannam@167
|
687 {
|
cannam@167
|
688 E Tbu, Tbz, Tbv, TbA, Tbr, Tbw;
|
cannam@167
|
689 Tbu = FNMS(KP707106781, Tbt, Tbs);
|
cannam@167
|
690 Tbz = FNMS(KP707106781, Tby, Tbx);
|
cannam@167
|
691 Tbr = W[8];
|
cannam@167
|
692 Tbv = Tbr * Tbu;
|
cannam@167
|
693 TbA = Tbr * Tbz;
|
cannam@167
|
694 Tbw = W[9];
|
cannam@167
|
695 rio[WS(vs, 5) + WS(rs, 7)] = FMA(Tbw, Tbz, Tbv);
|
cannam@167
|
696 iio[WS(vs, 5) + WS(rs, 7)] = FNMS(Tbw, Tbu, TbA);
|
cannam@167
|
697 }
|
cannam@167
|
698 {
|
cannam@167
|
699 E TbC, TbF, TbD, TbG, TbB, TbE;
|
cannam@167
|
700 TbC = FMA(KP707106781, Tbt, Tbs);
|
cannam@167
|
701 TbF = FMA(KP707106781, Tby, Tbx);
|
cannam@167
|
702 TbB = W[0];
|
cannam@167
|
703 TbD = TbB * TbC;
|
cannam@167
|
704 TbG = TbB * TbF;
|
cannam@167
|
705 TbE = W[1];
|
cannam@167
|
706 rio[WS(vs, 1) + WS(rs, 7)] = FMA(TbE, TbF, TbD);
|
cannam@167
|
707 iio[WS(vs, 1) + WS(rs, 7)] = FNMS(TbE, TbC, TbG);
|
cannam@167
|
708 }
|
cannam@167
|
709 {
|
cannam@167
|
710 E T10, T13, T11, T14, TZ, T12;
|
cannam@167
|
711 T10 = FMA(KP707106781, TR, TQ);
|
cannam@167
|
712 T13 = FMA(KP707106781, TW, TV);
|
cannam@167
|
713 TZ = W[0];
|
cannam@167
|
714 T11 = TZ * T10;
|
cannam@167
|
715 T14 = TZ * T13;
|
cannam@167
|
716 T12 = W[1];
|
cannam@167
|
717 rio[WS(vs, 1)] = FMA(T12, T13, T11);
|
cannam@167
|
718 iio[WS(vs, 1)] = FNMS(T12, T10, T14);
|
cannam@167
|
719 }
|
cannam@167
|
720 {
|
cannam@167
|
721 E T2w, T2z, T2x, T2A, T2v, T2y;
|
cannam@167
|
722 T2w = FMA(KP707106781, T2n, T2m);
|
cannam@167
|
723 T2z = FMA(KP707106781, T2s, T2r);
|
cannam@167
|
724 T2v = W[0];
|
cannam@167
|
725 T2x = T2v * T2w;
|
cannam@167
|
726 T2A = T2v * T2z;
|
cannam@167
|
727 T2y = W[1];
|
cannam@167
|
728 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T2y, T2z, T2x);
|
cannam@167
|
729 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T2y, T2w, T2A);
|
cannam@167
|
730 }
|
cannam@167
|
731 {
|
cannam@167
|
732 E T1h, T15, T1b, T1c, T1i, T1a;
|
cannam@167
|
733 T1h = T1d - T1g;
|
cannam@167
|
734 T1a = T18 - T19;
|
cannam@167
|
735 T15 = W[10];
|
cannam@167
|
736 T1b = T15 * T1a;
|
cannam@167
|
737 T1c = W[11];
|
cannam@167
|
738 T1i = T1c * T1a;
|
cannam@167
|
739 iio[WS(vs, 6)] = FNMS(T1c, T1h, T1b);
|
cannam@167
|
740 rio[WS(vs, 6)] = FMA(T15, T1h, T1i);
|
cannam@167
|
741 }
|
cannam@167
|
742 {
|
cannam@167
|
743 E T2o, T2t, T2p, T2u, T2l, T2q;
|
cannam@167
|
744 T2o = FNMS(KP707106781, T2n, T2m);
|
cannam@167
|
745 T2t = FNMS(KP707106781, T2s, T2r);
|
cannam@167
|
746 T2l = W[8];
|
cannam@167
|
747 T2p = T2l * T2o;
|
cannam@167
|
748 T2u = T2l * T2t;
|
cannam@167
|
749 T2q = W[9];
|
cannam@167
|
750 rio[WS(vs, 5) + WS(rs, 1)] = FMA(T2q, T2t, T2p);
|
cannam@167
|
751 iio[WS(vs, 5) + WS(rs, 1)] = FNMS(T2q, T2o, T2u);
|
cannam@167
|
752 }
|
cannam@167
|
753 {
|
cannam@167
|
754 E Tat, Tap, Tar, Tas, Tau, Taq;
|
cannam@167
|
755 Tat = Taj + Tam;
|
cannam@167
|
756 Taq = Taf + Tae;
|
cannam@167
|
757 Tap = W[2];
|
cannam@167
|
758 Tar = Tap * Taq;
|
cannam@167
|
759 Tas = W[3];
|
cannam@167
|
760 Tau = Tas * Taq;
|
cannam@167
|
761 iio[WS(vs, 2) + WS(rs, 6)] = FNMS(Tas, Tat, Tar);
|
cannam@167
|
762 rio[WS(vs, 2) + WS(rs, 6)] = FMA(Tap, Tat, Tau);
|
cannam@167
|
763 }
|
cannam@167
|
764 {
|
cannam@167
|
765 E TbZ, TbV, TbX, TbY, Tc0, TbW;
|
cannam@167
|
766 TbZ = TbP + TbS;
|
cannam@167
|
767 TbW = TbL + TbK;
|
cannam@167
|
768 TbV = W[2];
|
cannam@167
|
769 TbX = TbV * TbW;
|
cannam@167
|
770 TbY = W[3];
|
cannam@167
|
771 Tc0 = TbY * TbW;
|
cannam@167
|
772 iio[WS(vs, 2) + WS(rs, 7)] = FNMS(TbY, TbZ, TbX);
|
cannam@167
|
773 rio[WS(vs, 2) + WS(rs, 7)] = FMA(TbV, TbZ, Tc0);
|
cannam@167
|
774 }
|
cannam@167
|
775 {
|
cannam@167
|
776 E Taw, TaB, Tax, TaC, Tav, Tay;
|
cannam@167
|
777 Taw = T9d - T9k;
|
cannam@167
|
778 TaB = Taz - TaA;
|
cannam@167
|
779 Tav = W[6];
|
cannam@167
|
780 Tax = Tav * Taw;
|
cannam@167
|
781 TaC = Tav * TaB;
|
cannam@167
|
782 Tay = W[7];
|
cannam@167
|
783 rio[WS(vs, 4) + WS(rs, 6)] = FMA(Tay, TaB, Tax);
|
cannam@167
|
784 iio[WS(vs, 4) + WS(rs, 6)] = FNMS(Tay, Taw, TaC);
|
cannam@167
|
785 }
|
cannam@167
|
786 {
|
cannam@167
|
787 E TbT, TbH, TbN, TbO, TbU, TbM;
|
cannam@167
|
788 TbT = TbP - TbS;
|
cannam@167
|
789 TbM = TbK - TbL;
|
cannam@167
|
790 TbH = W[10];
|
cannam@167
|
791 TbN = TbH * TbM;
|
cannam@167
|
792 TbO = W[11];
|
cannam@167
|
793 TbU = TbO * TbM;
|
cannam@167
|
794 iio[WS(vs, 6) + WS(rs, 7)] = FNMS(TbO, TbT, TbN);
|
cannam@167
|
795 rio[WS(vs, 6) + WS(rs, 7)] = FMA(TbH, TbT, TbU);
|
cannam@167
|
796 }
|
cannam@167
|
797 {
|
cannam@167
|
798 E T2T, T2P, T2R, T2S, T2U, T2Q;
|
cannam@167
|
799 T2T = T2J + T2M;
|
cannam@167
|
800 T2Q = T2F + T2E;
|
cannam@167
|
801 T2P = W[2];
|
cannam@167
|
802 T2R = T2P * T2Q;
|
cannam@167
|
803 T2S = W[3];
|
cannam@167
|
804 T2U = T2S * T2Q;
|
cannam@167
|
805 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T2S, T2T, T2R);
|
cannam@167
|
806 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T2P, T2T, T2U);
|
cannam@167
|
807 }
|
cannam@167
|
808 {
|
cannam@167
|
809 E T5Y, T63, T5Z, T64, T5X, T60;
|
cannam@167
|
810 T5Y = T4F - T4M;
|
cannam@167
|
811 T63 = T61 - T62;
|
cannam@167
|
812 T5X = W[6];
|
cannam@167
|
813 T5Z = T5X * T5Y;
|
cannam@167
|
814 T64 = T5X * T63;
|
cannam@167
|
815 T60 = W[7];
|
cannam@167
|
816 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T60, T63, T5Z);
|
cannam@167
|
817 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T60, T5Y, T64);
|
cannam@167
|
818 }
|
cannam@167
|
819 {
|
cannam@167
|
820 E T42, T45, T43, T46, T41, T44;
|
cannam@167
|
821 T42 = FMA(KP707106781, T3T, T3S);
|
cannam@167
|
822 T45 = FMA(KP707106781, T3Y, T3X);
|
cannam@167
|
823 T41 = W[0];
|
cannam@167
|
824 T43 = T41 * T42;
|
cannam@167
|
825 T46 = T41 * T45;
|
cannam@167
|
826 T44 = W[1];
|
cannam@167
|
827 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T44, T45, T43);
|
cannam@167
|
828 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T44, T42, T46);
|
cannam@167
|
829 }
|
cannam@167
|
830 {
|
cannam@167
|
831 E T5y, T5B, T5z, T5C, T5x, T5A;
|
cannam@167
|
832 T5y = FMA(KP707106781, T5p, T5o);
|
cannam@167
|
833 T5B = FMA(KP707106781, T5u, T5t);
|
cannam@167
|
834 T5x = W[0];
|
cannam@167
|
835 T5z = T5x * T5y;
|
cannam@167
|
836 T5C = T5x * T5B;
|
cannam@167
|
837 T5A = W[1];
|
cannam@167
|
838 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T5A, T5B, T5z);
|
cannam@167
|
839 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T5A, T5y, T5C);
|
cannam@167
|
840 }
|
cannam@167
|
841 {
|
cannam@167
|
842 E T6W, T71, T6X, T72, T6T, T6Y;
|
cannam@167
|
843 T6W = FNMS(KP707106781, T6V, T6U);
|
cannam@167
|
844 T71 = FNMS(KP707106781, T70, T6Z);
|
cannam@167
|
845 T6T = W[8];
|
cannam@167
|
846 T6X = T6T * T6W;
|
cannam@167
|
847 T72 = T6T * T71;
|
cannam@167
|
848 T6Y = W[9];
|
cannam@167
|
849 rio[WS(vs, 5) + WS(rs, 4)] = FMA(T6Y, T71, T6X);
|
cannam@167
|
850 iio[WS(vs, 5) + WS(rs, 4)] = FNMS(T6Y, T6W, T72);
|
cannam@167
|
851 }
|
cannam@167
|
852 {
|
cannam@167
|
853 E Ta6, Ta9, Ta7, Taa, Ta5, Ta8;
|
cannam@167
|
854 Ta6 = FMA(KP707106781, T9X, T9W);
|
cannam@167
|
855 Ta9 = FMA(KP707106781, Ta2, Ta1);
|
cannam@167
|
856 Ta5 = W[0];
|
cannam@167
|
857 Ta7 = Ta5 * Ta6;
|
cannam@167
|
858 Taa = Ta5 * Ta9;
|
cannam@167
|
859 Ta8 = W[1];
|
cannam@167
|
860 rio[WS(vs, 1) + WS(rs, 6)] = FMA(Ta8, Ta9, Ta7);
|
cannam@167
|
861 iio[WS(vs, 1) + WS(rs, 6)] = FNMS(Ta8, Ta6, Taa);
|
cannam@167
|
862 }
|
cannam@167
|
863 {
|
cannam@167
|
864 E T7r, T7n, T7p, T7q, T7s, T7o;
|
cannam@167
|
865 T7r = T7h + T7k;
|
cannam@167
|
866 T7o = T7d + T7c;
|
cannam@167
|
867 T7n = W[2];
|
cannam@167
|
868 T7p = T7n * T7o;
|
cannam@167
|
869 T7q = W[3];
|
cannam@167
|
870 T7s = T7q * T7o;
|
cannam@167
|
871 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T7q, T7r, T7p);
|
cannam@167
|
872 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T7n, T7r, T7s);
|
cannam@167
|
873 }
|
cannam@167
|
874 {
|
cannam@167
|
875 E T8X, T8T, T8V, T8W, T8Y, T8U;
|
cannam@167
|
876 T8X = T8N + T8Q;
|
cannam@167
|
877 T8U = T8J + T8I;
|
cannam@167
|
878 T8T = W[2];
|
cannam@167
|
879 T8V = T8T * T8U;
|
cannam@167
|
880 T8W = W[3];
|
cannam@167
|
881 T8Y = T8W * T8U;
|
cannam@167
|
882 iio[WS(vs, 2) + WS(rs, 5)] = FNMS(T8W, T8X, T8V);
|
cannam@167
|
883 rio[WS(vs, 2) + WS(rs, 5)] = FMA(T8T, T8X, T8Y);
|
cannam@167
|
884 }
|
cannam@167
|
885 {
|
cannam@167
|
886 E T2W, T31, T2X, T32, T2V, T2Y;
|
cannam@167
|
887 T2W = T1D - T1K;
|
cannam@167
|
888 T31 = T2Z - T30;
|
cannam@167
|
889 T2V = W[6];
|
cannam@167
|
890 T2X = T2V * T2W;
|
cannam@167
|
891 T32 = T2V * T31;
|
cannam@167
|
892 T2Y = W[7];
|
cannam@167
|
893 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T2Y, T31, T2X);
|
cannam@167
|
894 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T2Y, T2W, T32);
|
cannam@167
|
895 }
|
cannam@167
|
896 {
|
cannam@167
|
897 E T5V, T5R, T5T, T5U, T5W, T5S;
|
cannam@167
|
898 T5V = T5L + T5O;
|
cannam@167
|
899 T5S = T5H + T5G;
|
cannam@167
|
900 T5R = W[2];
|
cannam@167
|
901 T5T = T5R * T5S;
|
cannam@167
|
902 T5U = W[3];
|
cannam@167
|
903 T5W = T5U * T5S;
|
cannam@167
|
904 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T5U, T5V, T5T);
|
cannam@167
|
905 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T5R, T5V, T5W);
|
cannam@167
|
906 }
|
cannam@167
|
907 {
|
cannam@167
|
908 E T3U, T3Z, T3V, T40, T3R, T3W;
|
cannam@167
|
909 T3U = FNMS(KP707106781, T3T, T3S);
|
cannam@167
|
910 T3Z = FNMS(KP707106781, T3Y, T3X);
|
cannam@167
|
911 T3R = W[8];
|
cannam@167
|
912 T3V = T3R * T3U;
|
cannam@167
|
913 T40 = T3R * T3Z;
|
cannam@167
|
914 T3W = W[9];
|
cannam@167
|
915 rio[WS(vs, 5) + WS(rs, 2)] = FMA(T3W, T3Z, T3V);
|
cannam@167
|
916 iio[WS(vs, 5) + WS(rs, 2)] = FNMS(T3W, T3U, T40);
|
cannam@167
|
917 }
|
cannam@167
|
918 {
|
cannam@167
|
919 E T5P, T5D, T5J, T5K, T5Q, T5I;
|
cannam@167
|
920 T5P = T5L - T5O;
|
cannam@167
|
921 T5I = T5G - T5H;
|
cannam@167
|
922 T5D = W[10];
|
cannam@167
|
923 T5J = T5D * T5I;
|
cannam@167
|
924 T5K = W[11];
|
cannam@167
|
925 T5Q = T5K * T5I;
|
cannam@167
|
926 iio[WS(vs, 6) + WS(rs, 3)] = FNMS(T5K, T5P, T5J);
|
cannam@167
|
927 rio[WS(vs, 6) + WS(rs, 3)] = FMA(T5D, T5P, T5Q);
|
cannam@167
|
928 }
|
cannam@167
|
929 {
|
cannam@167
|
930 E T74, T77, T75, T78, T73, T76;
|
cannam@167
|
931 T74 = FMA(KP707106781, T6V, T6U);
|
cannam@167
|
932 T77 = FMA(KP707106781, T70, T6Z);
|
cannam@167
|
933 T73 = W[0];
|
cannam@167
|
934 T75 = T73 * T74;
|
cannam@167
|
935 T78 = T73 * T77;
|
cannam@167
|
936 T76 = W[1];
|
cannam@167
|
937 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T76, T77, T75);
|
cannam@167
|
938 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T76, T74, T78);
|
cannam@167
|
939 }
|
cannam@167
|
940 {
|
cannam@167
|
941 E T9Y, Ta3, T9Z, Ta4, T9V, Ta0;
|
cannam@167
|
942 T9Y = FNMS(KP707106781, T9X, T9W);
|
cannam@167
|
943 Ta3 = FNMS(KP707106781, Ta2, Ta1);
|
cannam@167
|
944 T9V = W[8];
|
cannam@167
|
945 T9Z = T9V * T9Y;
|
cannam@167
|
946 Ta4 = T9V * Ta3;
|
cannam@167
|
947 Ta0 = W[9];
|
cannam@167
|
948 rio[WS(vs, 5) + WS(rs, 6)] = FMA(Ta0, Ta3, T9Z);
|
cannam@167
|
949 iio[WS(vs, 5) + WS(rs, 6)] = FNMS(Ta0, T9Y, Ta4);
|
cannam@167
|
950 }
|
cannam@167
|
951 {
|
cannam@167
|
952 E T7l, T79, T7f, T7g, T7m, T7e;
|
cannam@167
|
953 T7l = T7h - T7k;
|
cannam@167
|
954 T7e = T7c - T7d;
|
cannam@167
|
955 T79 = W[10];
|
cannam@167
|
956 T7f = T79 * T7e;
|
cannam@167
|
957 T7g = W[11];
|
cannam@167
|
958 T7m = T7g * T7e;
|
cannam@167
|
959 iio[WS(vs, 6) + WS(rs, 4)] = FNMS(T7g, T7l, T7f);
|
cannam@167
|
960 rio[WS(vs, 6) + WS(rs, 4)] = FMA(T79, T7l, T7m);
|
cannam@167
|
961 }
|
cannam@167
|
962 {
|
cannam@167
|
963 E T90, T95, T91, T96, T8Z, T92;
|
cannam@167
|
964 T90 = T7H - T7O;
|
cannam@167
|
965 T95 = T93 - T94;
|
cannam@167
|
966 T8Z = W[6];
|
cannam@167
|
967 T91 = T8Z * T90;
|
cannam@167
|
968 T96 = T8Z * T95;
|
cannam@167
|
969 T92 = W[7];
|
cannam@167
|
970 rio[WS(vs, 4) + WS(rs, 5)] = FMA(T92, T95, T91);
|
cannam@167
|
971 iio[WS(vs, 4) + WS(rs, 5)] = FNMS(T92, T90, T96);
|
cannam@167
|
972 }
|
cannam@167
|
973 {
|
cannam@167
|
974 E T4j, T47, T4d, T4e, T4k, T4c;
|
cannam@167
|
975 T4j = T4f - T4i;
|
cannam@167
|
976 T4c = T4a - T4b;
|
cannam@167
|
977 T47 = W[10];
|
cannam@167
|
978 T4d = T47 * T4c;
|
cannam@167
|
979 T4e = W[11];
|
cannam@167
|
980 T4k = T4e * T4c;
|
cannam@167
|
981 iio[WS(vs, 6) + WS(rs, 2)] = FNMS(T4e, T4j, T4d);
|
cannam@167
|
982 rio[WS(vs, 6) + WS(rs, 2)] = FMA(T47, T4j, T4k);
|
cannam@167
|
983 }
|
cannam@167
|
984 {
|
cannam@167
|
985 E T5q, T5v, T5r, T5w, T5n, T5s;
|
cannam@167
|
986 T5q = FNMS(KP707106781, T5p, T5o);
|
cannam@167
|
987 T5v = FNMS(KP707106781, T5u, T5t);
|
cannam@167
|
988 T5n = W[8];
|
cannam@167
|
989 T5r = T5n * T5q;
|
cannam@167
|
990 T5w = T5n * T5v;
|
cannam@167
|
991 T5s = W[9];
|
cannam@167
|
992 rio[WS(vs, 5) + WS(rs, 3)] = FMA(T5s, T5v, T5r);
|
cannam@167
|
993 iio[WS(vs, 5) + WS(rs, 3)] = FNMS(T5s, T5q, T5w);
|
cannam@167
|
994 }
|
cannam@167
|
995 {
|
cannam@167
|
996 E T4p, T4l, T4n, T4o, T4q, T4m;
|
cannam@167
|
997 T4p = T4f + T4i;
|
cannam@167
|
998 T4m = T4b + T4a;
|
cannam@167
|
999 T4l = W[2];
|
cannam@167
|
1000 T4n = T4l * T4m;
|
cannam@167
|
1001 T4o = W[3];
|
cannam@167
|
1002 T4q = T4o * T4m;
|
cannam@167
|
1003 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T4o, T4p, T4n);
|
cannam@167
|
1004 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T4l, T4p, T4q);
|
cannam@167
|
1005 }
|
cannam@167
|
1006 {
|
cannam@167
|
1007 E T4s, T4x, T4t, T4y, T4r, T4u;
|
cannam@167
|
1008 T4s = T39 - T3g;
|
cannam@167
|
1009 T4x = T4v - T4w;
|
cannam@167
|
1010 T4r = W[6];
|
cannam@167
|
1011 T4t = T4r * T4s;
|
cannam@167
|
1012 T4y = T4r * T4x;
|
cannam@167
|
1013 T4u = W[7];
|
cannam@167
|
1014 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T4u, T4x, T4t);
|
cannam@167
|
1015 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T4u, T4s, T4y);
|
cannam@167
|
1016 }
|
cannam@167
|
1017 {
|
cannam@167
|
1018 E T7u, T7z, T7v, T7A, T7t, T7w;
|
cannam@167
|
1019 T7u = T6b - T6i;
|
cannam@167
|
1020 T7z = T7x - T7y;
|
cannam@167
|
1021 T7t = W[6];
|
cannam@167
|
1022 T7v = T7t * T7u;
|
cannam@167
|
1023 T7A = T7t * T7z;
|
cannam@167
|
1024 T7w = W[7];
|
cannam@167
|
1025 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T7w, T7z, T7v);
|
cannam@167
|
1026 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T7w, T7u, T7A);
|
cannam@167
|
1027 }
|
cannam@167
|
1028 {
|
cannam@167
|
1029 E T8R, T8F, T8L, T8M, T8S, T8K;
|
cannam@167
|
1030 T8R = T8N - T8Q;
|
cannam@167
|
1031 T8K = T8I - T8J;
|
cannam@167
|
1032 T8F = W[10];
|
cannam@167
|
1033 T8L = T8F * T8K;
|
cannam@167
|
1034 T8M = W[11];
|
cannam@167
|
1035 T8S = T8M * T8K;
|
cannam@167
|
1036 iio[WS(vs, 6) + WS(rs, 5)] = FNMS(T8M, T8R, T8L);
|
cannam@167
|
1037 rio[WS(vs, 6) + WS(rs, 5)] = FMA(T8F, T8R, T8S);
|
cannam@167
|
1038 }
|
cannam@167
|
1039 {
|
cannam@167
|
1040 E T8s, T8x, T8t, T8y, T8p, T8u;
|
cannam@167
|
1041 T8s = FNMS(KP707106781, T8r, T8q);
|
cannam@167
|
1042 T8x = FNMS(KP707106781, T8w, T8v);
|
cannam@167
|
1043 T8p = W[8];
|
cannam@167
|
1044 T8t = T8p * T8s;
|
cannam@167
|
1045 T8y = T8p * T8x;
|
cannam@167
|
1046 T8u = W[9];
|
cannam@167
|
1047 rio[WS(vs, 5) + WS(rs, 5)] = FMA(T8u, T8x, T8t);
|
cannam@167
|
1048 iio[WS(vs, 5) + WS(rs, 5)] = FNMS(T8u, T8s, T8y);
|
cannam@167
|
1049 }
|
cannam@167
|
1050 {
|
cannam@167
|
1051 E T8A, T8D, T8B, T8E, T8z, T8C;
|
cannam@167
|
1052 T8A = FMA(KP707106781, T8r, T8q);
|
cannam@167
|
1053 T8D = FMA(KP707106781, T8w, T8v);
|
cannam@167
|
1054 T8z = W[0];
|
cannam@167
|
1055 T8B = T8z * T8A;
|
cannam@167
|
1056 T8E = T8z * T8D;
|
cannam@167
|
1057 T8C = W[1];
|
cannam@167
|
1058 rio[WS(vs, 1) + WS(rs, 5)] = FMA(T8C, T8D, T8B);
|
cannam@167
|
1059 iio[WS(vs, 1) + WS(rs, 5)] = FNMS(T8C, T8A, T8E);
|
cannam@167
|
1060 }
|
cannam@167
|
1061 {
|
cannam@167
|
1062 E TH, TN, TJ, TL, TM, TO, Tf, Tx, Ty, TI, TG, TK, Tw;
|
cannam@167
|
1063 TG = TE - TF;
|
cannam@167
|
1064 TH = FNMS(KP707106781, TG, TD);
|
cannam@167
|
1065 TN = FMA(KP707106781, TG, TD);
|
cannam@167
|
1066 TK = FMA(KP707106781, Tv, Tk);
|
cannam@167
|
1067 TJ = W[4];
|
cannam@167
|
1068 TL = TJ * TK;
|
cannam@167
|
1069 TM = W[5];
|
cannam@167
|
1070 TO = TM * TK;
|
cannam@167
|
1071 Tw = FNMS(KP707106781, Tv, Tk);
|
cannam@167
|
1072 Tf = W[12];
|
cannam@167
|
1073 Tx = Tf * Tw;
|
cannam@167
|
1074 Ty = W[13];
|
cannam@167
|
1075 TI = Ty * Tw;
|
cannam@167
|
1076 iio[WS(vs, 7)] = FNMS(Ty, TH, Tx);
|
cannam@167
|
1077 rio[WS(vs, 7)] = FMA(Tf, TH, TI);
|
cannam@167
|
1078 iio[WS(vs, 3)] = FNMS(TM, TN, TL);
|
cannam@167
|
1079 rio[WS(vs, 3)] = FMA(TJ, TN, TO);
|
cannam@167
|
1080 }
|
cannam@167
|
1081 {
|
cannam@167
|
1082 E T5f, T5l, T5h, T5j, T5k, T5m, T4N, T55, T56, T5g, T5e, T5i, T54;
|
cannam@167
|
1083 T5e = T5c - T5d;
|
cannam@167
|
1084 T5f = FNMS(KP707106781, T5e, T5b);
|
cannam@167
|
1085 T5l = FMA(KP707106781, T5e, T5b);
|
cannam@167
|
1086 T5i = FMA(KP707106781, T53, T4S);
|
cannam@167
|
1087 T5h = W[4];
|
cannam@167
|
1088 T5j = T5h * T5i;
|
cannam@167
|
1089 T5k = W[5];
|
cannam@167
|
1090 T5m = T5k * T5i;
|
cannam@167
|
1091 T54 = FNMS(KP707106781, T53, T4S);
|
cannam@167
|
1092 T4N = W[12];
|
cannam@167
|
1093 T55 = T4N * T54;
|
cannam@167
|
1094 T56 = W[13];
|
cannam@167
|
1095 T5g = T56 * T54;
|
cannam@167
|
1096 iio[WS(vs, 7) + WS(rs, 3)] = FNMS(T56, T5f, T55);
|
cannam@167
|
1097 rio[WS(vs, 7) + WS(rs, 3)] = FMA(T4N, T5f, T5g);
|
cannam@167
|
1098 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T5k, T5l, T5j);
|
cannam@167
|
1099 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T5h, T5l, T5m);
|
cannam@167
|
1100 }
|
cannam@167
|
1101 {
|
cannam@167
|
1102 E T2d, T2j, T2f, T2h, T2i, T2k, T1L, T23, T24, T2e, T2c, T2g, T22;
|
cannam@167
|
1103 T2c = T2a - T2b;
|
cannam@167
|
1104 T2d = FNMS(KP707106781, T2c, T29);
|
cannam@167
|
1105 T2j = FMA(KP707106781, T2c, T29);
|
cannam@167
|
1106 T2g = FMA(KP707106781, T21, T1Q);
|
cannam@167
|
1107 T2f = W[4];
|
cannam@167
|
1108 T2h = T2f * T2g;
|
cannam@167
|
1109 T2i = W[5];
|
cannam@167
|
1110 T2k = T2i * T2g;
|
cannam@167
|
1111 T22 = FNMS(KP707106781, T21, T1Q);
|
cannam@167
|
1112 T1L = W[12];
|
cannam@167
|
1113 T23 = T1L * T22;
|
cannam@167
|
1114 T24 = W[13];
|
cannam@167
|
1115 T2e = T24 * T22;
|
cannam@167
|
1116 iio[WS(vs, 7) + WS(rs, 1)] = FNMS(T24, T2d, T23);
|
cannam@167
|
1117 rio[WS(vs, 7) + WS(rs, 1)] = FMA(T1L, T2d, T2e);
|
cannam@167
|
1118 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T2i, T2j, T2h);
|
cannam@167
|
1119 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T2f, T2j, T2k);
|
cannam@167
|
1120 }
|
cannam@167
|
1121 {
|
cannam@167
|
1122 E T3J, T3P, T3L, T3N, T3O, T3Q, T3h, T3z, T3A, T3K, T3I, T3M, T3y;
|
cannam@167
|
1123 T3I = T3G - T3H;
|
cannam@167
|
1124 T3J = FNMS(KP707106781, T3I, T3F);
|
cannam@167
|
1125 T3P = FMA(KP707106781, T3I, T3F);
|
cannam@167
|
1126 T3M = FMA(KP707106781, T3x, T3m);
|
cannam@167
|
1127 T3L = W[4];
|
cannam@167
|
1128 T3N = T3L * T3M;
|
cannam@167
|
1129 T3O = W[5];
|
cannam@167
|
1130 T3Q = T3O * T3M;
|
cannam@167
|
1131 T3y = FNMS(KP707106781, T3x, T3m);
|
cannam@167
|
1132 T3h = W[12];
|
cannam@167
|
1133 T3z = T3h * T3y;
|
cannam@167
|
1134 T3A = W[13];
|
cannam@167
|
1135 T3K = T3A * T3y;
|
cannam@167
|
1136 iio[WS(vs, 7) + WS(rs, 2)] = FNMS(T3A, T3J, T3z);
|
cannam@167
|
1137 rio[WS(vs, 7) + WS(rs, 2)] = FMA(T3h, T3J, T3K);
|
cannam@167
|
1138 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T3O, T3P, T3N);
|
cannam@167
|
1139 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T3L, T3P, T3Q);
|
cannam@167
|
1140 }
|
cannam@167
|
1141 {
|
cannam@167
|
1142 E T6L, T6R, T6N, T6P, T6Q, T6S, T6j, T6B, T6C, T6M, T6K, T6O, T6A;
|
cannam@167
|
1143 T6K = T6I - T6J;
|
cannam@167
|
1144 T6L = FNMS(KP707106781, T6K, T6H);
|
cannam@167
|
1145 T6R = FMA(KP707106781, T6K, T6H);
|
cannam@167
|
1146 T6O = FMA(KP707106781, T6z, T6o);
|
cannam@167
|
1147 T6N = W[4];
|
cannam@167
|
1148 T6P = T6N * T6O;
|
cannam@167
|
1149 T6Q = W[5];
|
cannam@167
|
1150 T6S = T6Q * T6O;
|
cannam@167
|
1151 T6A = FNMS(KP707106781, T6z, T6o);
|
cannam@167
|
1152 T6j = W[12];
|
cannam@167
|
1153 T6B = T6j * T6A;
|
cannam@167
|
1154 T6C = W[13];
|
cannam@167
|
1155 T6M = T6C * T6A;
|
cannam@167
|
1156 iio[WS(vs, 7) + WS(rs, 4)] = FNMS(T6C, T6L, T6B);
|
cannam@167
|
1157 rio[WS(vs, 7) + WS(rs, 4)] = FMA(T6j, T6L, T6M);
|
cannam@167
|
1158 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T6Q, T6R, T6P);
|
cannam@167
|
1159 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T6N, T6R, T6S);
|
cannam@167
|
1160 }
|
cannam@167
|
1161 {
|
cannam@167
|
1162 E Tbj, Tbp, Tbl, Tbn, Tbo, Tbq, TaR, Tb9, Tba, Tbk, Tbi, Tbm, Tb8;
|
cannam@167
|
1163 Tbi = Tbg - Tbh;
|
cannam@167
|
1164 Tbj = FNMS(KP707106781, Tbi, Tbf);
|
cannam@167
|
1165 Tbp = FMA(KP707106781, Tbi, Tbf);
|
cannam@167
|
1166 Tbm = FMA(KP707106781, Tb7, TaW);
|
cannam@167
|
1167 Tbl = W[4];
|
cannam@167
|
1168 Tbn = Tbl * Tbm;
|
cannam@167
|
1169 Tbo = W[5];
|
cannam@167
|
1170 Tbq = Tbo * Tbm;
|
cannam@167
|
1171 Tb8 = FNMS(KP707106781, Tb7, TaW);
|
cannam@167
|
1172 TaR = W[12];
|
cannam@167
|
1173 Tb9 = TaR * Tb8;
|
cannam@167
|
1174 Tba = W[13];
|
cannam@167
|
1175 Tbk = Tba * Tb8;
|
cannam@167
|
1176 iio[WS(vs, 7) + WS(rs, 7)] = FNMS(Tba, Tbj, Tb9);
|
cannam@167
|
1177 rio[WS(vs, 7) + WS(rs, 7)] = FMA(TaR, Tbj, Tbk);
|
cannam@167
|
1178 iio[WS(vs, 3) + WS(rs, 7)] = FNMS(Tbo, Tbp, Tbn);
|
cannam@167
|
1179 rio[WS(vs, 3) + WS(rs, 7)] = FMA(Tbl, Tbp, Tbq);
|
cannam@167
|
1180 }
|
cannam@167
|
1181 {
|
cannam@167
|
1182 E T8h, T8n, T8j, T8l, T8m, T8o, T7P, T87, T88, T8i, T8g, T8k, T86;
|
cannam@167
|
1183 T8g = T8e - T8f;
|
cannam@167
|
1184 T8h = FNMS(KP707106781, T8g, T8d);
|
cannam@167
|
1185 T8n = FMA(KP707106781, T8g, T8d);
|
cannam@167
|
1186 T8k = FMA(KP707106781, T85, T7U);
|
cannam@167
|
1187 T8j = W[4];
|
cannam@167
|
1188 T8l = T8j * T8k;
|
cannam@167
|
1189 T8m = W[5];
|
cannam@167
|
1190 T8o = T8m * T8k;
|
cannam@167
|
1191 T86 = FNMS(KP707106781, T85, T7U);
|
cannam@167
|
1192 T7P = W[12];
|
cannam@167
|
1193 T87 = T7P * T86;
|
cannam@167
|
1194 T88 = W[13];
|
cannam@167
|
1195 T8i = T88 * T86;
|
cannam@167
|
1196 iio[WS(vs, 7) + WS(rs, 5)] = FNMS(T88, T8h, T87);
|
cannam@167
|
1197 rio[WS(vs, 7) + WS(rs, 5)] = FMA(T7P, T8h, T8i);
|
cannam@167
|
1198 iio[WS(vs, 3) + WS(rs, 5)] = FNMS(T8m, T8n, T8l);
|
cannam@167
|
1199 rio[WS(vs, 3) + WS(rs, 5)] = FMA(T8j, T8n, T8o);
|
cannam@167
|
1200 }
|
cannam@167
|
1201 {
|
cannam@167
|
1202 E T9N, T9T, T9P, T9R, T9S, T9U, T9l, T9D, T9E, T9O, T9M, T9Q, T9C;
|
cannam@167
|
1203 T9M = T9K - T9L;
|
cannam@167
|
1204 T9N = FNMS(KP707106781, T9M, T9J);
|
cannam@167
|
1205 T9T = FMA(KP707106781, T9M, T9J);
|
cannam@167
|
1206 T9Q = FMA(KP707106781, T9B, T9q);
|
cannam@167
|
1207 T9P = W[4];
|
cannam@167
|
1208 T9R = T9P * T9Q;
|
cannam@167
|
1209 T9S = W[5];
|
cannam@167
|
1210 T9U = T9S * T9Q;
|
cannam@167
|
1211 T9C = FNMS(KP707106781, T9B, T9q);
|
cannam@167
|
1212 T9l = W[12];
|
cannam@167
|
1213 T9D = T9l * T9C;
|
cannam@167
|
1214 T9E = W[13];
|
cannam@167
|
1215 T9O = T9E * T9C;
|
cannam@167
|
1216 iio[WS(vs, 7) + WS(rs, 6)] = FNMS(T9E, T9N, T9D);
|
cannam@167
|
1217 rio[WS(vs, 7) + WS(rs, 6)] = FMA(T9l, T9N, T9O);
|
cannam@167
|
1218 iio[WS(vs, 3) + WS(rs, 6)] = FNMS(T9S, T9T, T9R);
|
cannam@167
|
1219 rio[WS(vs, 3) + WS(rs, 6)] = FMA(T9P, T9T, T9U);
|
cannam@167
|
1220 }
|
cannam@167
|
1221 }
|
cannam@167
|
1222 }
|
cannam@167
|
1223 }
|
cannam@167
|
1224
|
cannam@167
|
1225 static const tw_instr twinstr[] = {
|
cannam@167
|
1226 {TW_FULL, 0, 8},
|
cannam@167
|
1227 {TW_NEXT, 1, 0}
|
cannam@167
|
1228 };
|
cannam@167
|
1229
|
cannam@167
|
1230 static const ct_desc desc = { 8, "q1_8", twinstr, &GENUS, {352, 112, 176, 0}, 0, 0, 0 };
|
cannam@167
|
1231
|
cannam@167
|
1232 void X(codelet_q1_8) (planner *p) {
|
cannam@167
|
1233 X(kdft_difsq_register) (p, q1_8, &desc);
|
cannam@167
|
1234 }
|
cannam@167
|
1235 #else
|
cannam@167
|
1236
|
cannam@167
|
1237 /* Generated by: ../../../genfft/gen_twidsq.native -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 8 -name q1_8 -include dft/scalar/q.h */
|
cannam@167
|
1238
|
cannam@167
|
1239 /*
|
cannam@167
|
1240 * This function contains 528 FP additions, 256 FP multiplications,
|
cannam@167
|
1241 * (or, 416 additions, 144 multiplications, 112 fused multiply/add),
|
cannam@167
|
1242 * 142 stack variables, 1 constants, and 256 memory accesses
|
cannam@167
|
1243 */
|
cannam@167
|
1244 #include "dft/scalar/q.h"
|
cannam@167
|
1245
|
cannam@167
|
1246 static void q1_8(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
|
cannam@167
|
1247 {
|
cannam@167
|
1248 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
1249 {
|
cannam@167
|
1250 INT m;
|
cannam@167
|
1251 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)) {
|
cannam@167
|
1252 E T7, T14, T1g, Tk, TC, TQ, T10, TM, T1w, T2p, T2z, T1H, T1M, T1W, T2j;
|
cannam@167
|
1253 E T1V, T7R, T8O, T90, T84, T8m, T8A, T8K, T8w, T9g, Ta9, Taj, T9r, T9w, T9G;
|
cannam@167
|
1254 E Ta3, T9F, Te, T17, T1h, Tp, Tu, TE, T11, TD, T1p, T2m, T2y, T1C, T1U;
|
cannam@167
|
1255 E T28, T2i, T24, T7Y, T8R, T91, T89, T8e, T8o, T8L, T8n, T99, Ta6, Tai, T9m;
|
cannam@167
|
1256 E T9E, T9S, Ta2, T9O, T2H, T3E, T3Q, T2U, T3c, T3q, T3A, T3m, T46, T4Z, T59;
|
cannam@167
|
1257 E T4h, T4m, T4w, T4T, T4v, T5h, T6e, T6q, T5u, T5M, T60, T6a, T5W, T6G, T7z;
|
cannam@167
|
1258 E T7J, T6R, T6W, T76, T7t, T75, T2O, T3H, T3R, T2Z, T34, T3e, T3B, T3d, T3Z;
|
cannam@167
|
1259 E T4W, T58, T4c, T4u, T4I, T4S, T4E, T5o, T6h, T6r, T5z, T5E, T5O, T6b, T5N;
|
cannam@167
|
1260 E T6z, T7w, T7I, T6M, T74, T7i, T7s, T7e;
|
cannam@167
|
1261 {
|
cannam@167
|
1262 E T3, Ty, Tj, TY, T6, Tg, TB, TZ;
|
cannam@167
|
1263 {
|
cannam@167
|
1264 E T1, T2, Th, Ti;
|
cannam@167
|
1265 T1 = rio[0];
|
cannam@167
|
1266 T2 = rio[WS(rs, 4)];
|
cannam@167
|
1267 T3 = T1 + T2;
|
cannam@167
|
1268 Ty = T1 - T2;
|
cannam@167
|
1269 Th = iio[0];
|
cannam@167
|
1270 Ti = iio[WS(rs, 4)];
|
cannam@167
|
1271 Tj = Th - Ti;
|
cannam@167
|
1272 TY = Th + Ti;
|
cannam@167
|
1273 }
|
cannam@167
|
1274 {
|
cannam@167
|
1275 E T4, T5, Tz, TA;
|
cannam@167
|
1276 T4 = rio[WS(rs, 2)];
|
cannam@167
|
1277 T5 = rio[WS(rs, 6)];
|
cannam@167
|
1278 T6 = T4 + T5;
|
cannam@167
|
1279 Tg = T4 - T5;
|
cannam@167
|
1280 Tz = iio[WS(rs, 2)];
|
cannam@167
|
1281 TA = iio[WS(rs, 6)];
|
cannam@167
|
1282 TB = Tz - TA;
|
cannam@167
|
1283 TZ = Tz + TA;
|
cannam@167
|
1284 }
|
cannam@167
|
1285 T7 = T3 + T6;
|
cannam@167
|
1286 T14 = T3 - T6;
|
cannam@167
|
1287 T1g = TY + TZ;
|
cannam@167
|
1288 Tk = Tg + Tj;
|
cannam@167
|
1289 TC = Ty - TB;
|
cannam@167
|
1290 TQ = Tj - Tg;
|
cannam@167
|
1291 T10 = TY - TZ;
|
cannam@167
|
1292 TM = Ty + TB;
|
cannam@167
|
1293 }
|
cannam@167
|
1294 {
|
cannam@167
|
1295 E T1s, T1I, T1L, T2n, T1v, T1D, T1G, T2o;
|
cannam@167
|
1296 {
|
cannam@167
|
1297 E T1q, T1r, T1J, T1K;
|
cannam@167
|
1298 T1q = rio[WS(vs, 1) + WS(rs, 1)];
|
cannam@167
|
1299 T1r = rio[WS(vs, 1) + WS(rs, 5)];
|
cannam@167
|
1300 T1s = T1q + T1r;
|
cannam@167
|
1301 T1I = T1q - T1r;
|
cannam@167
|
1302 T1J = iio[WS(vs, 1) + WS(rs, 1)];
|
cannam@167
|
1303 T1K = iio[WS(vs, 1) + WS(rs, 5)];
|
cannam@167
|
1304 T1L = T1J - T1K;
|
cannam@167
|
1305 T2n = T1J + T1K;
|
cannam@167
|
1306 }
|
cannam@167
|
1307 {
|
cannam@167
|
1308 E T1t, T1u, T1E, T1F;
|
cannam@167
|
1309 T1t = rio[WS(vs, 1) + WS(rs, 7)];
|
cannam@167
|
1310 T1u = rio[WS(vs, 1) + WS(rs, 3)];
|
cannam@167
|
1311 T1v = T1t + T1u;
|
cannam@167
|
1312 T1D = T1t - T1u;
|
cannam@167
|
1313 T1E = iio[WS(vs, 1) + WS(rs, 7)];
|
cannam@167
|
1314 T1F = iio[WS(vs, 1) + WS(rs, 3)];
|
cannam@167
|
1315 T1G = T1E - T1F;
|
cannam@167
|
1316 T2o = T1E + T1F;
|
cannam@167
|
1317 }
|
cannam@167
|
1318 T1w = T1s + T1v;
|
cannam@167
|
1319 T2p = T2n - T2o;
|
cannam@167
|
1320 T2z = T2n + T2o;
|
cannam@167
|
1321 T1H = T1D - T1G;
|
cannam@167
|
1322 T1M = T1I + T1L;
|
cannam@167
|
1323 T1W = T1D + T1G;
|
cannam@167
|
1324 T2j = T1v - T1s;
|
cannam@167
|
1325 T1V = T1L - T1I;
|
cannam@167
|
1326 }
|
cannam@167
|
1327 {
|
cannam@167
|
1328 E T7N, T8i, T83, T8I, T7Q, T80, T8l, T8J;
|
cannam@167
|
1329 {
|
cannam@167
|
1330 E T7L, T7M, T81, T82;
|
cannam@167
|
1331 T7L = rio[WS(vs, 6)];
|
cannam@167
|
1332 T7M = rio[WS(vs, 6) + WS(rs, 4)];
|
cannam@167
|
1333 T7N = T7L + T7M;
|
cannam@167
|
1334 T8i = T7L - T7M;
|
cannam@167
|
1335 T81 = iio[WS(vs, 6)];
|
cannam@167
|
1336 T82 = iio[WS(vs, 6) + WS(rs, 4)];
|
cannam@167
|
1337 T83 = T81 - T82;
|
cannam@167
|
1338 T8I = T81 + T82;
|
cannam@167
|
1339 }
|
cannam@167
|
1340 {
|
cannam@167
|
1341 E T7O, T7P, T8j, T8k;
|
cannam@167
|
1342 T7O = rio[WS(vs, 6) + WS(rs, 2)];
|
cannam@167
|
1343 T7P = rio[WS(vs, 6) + WS(rs, 6)];
|
cannam@167
|
1344 T7Q = T7O + T7P;
|
cannam@167
|
1345 T80 = T7O - T7P;
|
cannam@167
|
1346 T8j = iio[WS(vs, 6) + WS(rs, 2)];
|
cannam@167
|
1347 T8k = iio[WS(vs, 6) + WS(rs, 6)];
|
cannam@167
|
1348 T8l = T8j - T8k;
|
cannam@167
|
1349 T8J = T8j + T8k;
|
cannam@167
|
1350 }
|
cannam@167
|
1351 T7R = T7N + T7Q;
|
cannam@167
|
1352 T8O = T7N - T7Q;
|
cannam@167
|
1353 T90 = T8I + T8J;
|
cannam@167
|
1354 T84 = T80 + T83;
|
cannam@167
|
1355 T8m = T8i - T8l;
|
cannam@167
|
1356 T8A = T83 - T80;
|
cannam@167
|
1357 T8K = T8I - T8J;
|
cannam@167
|
1358 T8w = T8i + T8l;
|
cannam@167
|
1359 }
|
cannam@167
|
1360 {
|
cannam@167
|
1361 E T9c, T9s, T9v, Ta7, T9f, T9n, T9q, Ta8;
|
cannam@167
|
1362 {
|
cannam@167
|
1363 E T9a, T9b, T9t, T9u;
|
cannam@167
|
1364 T9a = rio[WS(vs, 7) + WS(rs, 1)];
|
cannam@167
|
1365 T9b = rio[WS(vs, 7) + WS(rs, 5)];
|
cannam@167
|
1366 T9c = T9a + T9b;
|
cannam@167
|
1367 T9s = T9a - T9b;
|
cannam@167
|
1368 T9t = iio[WS(vs, 7) + WS(rs, 1)];
|
cannam@167
|
1369 T9u = iio[WS(vs, 7) + WS(rs, 5)];
|
cannam@167
|
1370 T9v = T9t - T9u;
|
cannam@167
|
1371 Ta7 = T9t + T9u;
|
cannam@167
|
1372 }
|
cannam@167
|
1373 {
|
cannam@167
|
1374 E T9d, T9e, T9o, T9p;
|
cannam@167
|
1375 T9d = rio[WS(vs, 7) + WS(rs, 7)];
|
cannam@167
|
1376 T9e = rio[WS(vs, 7) + WS(rs, 3)];
|
cannam@167
|
1377 T9f = T9d + T9e;
|
cannam@167
|
1378 T9n = T9d - T9e;
|
cannam@167
|
1379 T9o = iio[WS(vs, 7) + WS(rs, 7)];
|
cannam@167
|
1380 T9p = iio[WS(vs, 7) + WS(rs, 3)];
|
cannam@167
|
1381 T9q = T9o - T9p;
|
cannam@167
|
1382 Ta8 = T9o + T9p;
|
cannam@167
|
1383 }
|
cannam@167
|
1384 T9g = T9c + T9f;
|
cannam@167
|
1385 Ta9 = Ta7 - Ta8;
|
cannam@167
|
1386 Taj = Ta7 + Ta8;
|
cannam@167
|
1387 T9r = T9n - T9q;
|
cannam@167
|
1388 T9w = T9s + T9v;
|
cannam@167
|
1389 T9G = T9n + T9q;
|
cannam@167
|
1390 Ta3 = T9f - T9c;
|
cannam@167
|
1391 T9F = T9v - T9s;
|
cannam@167
|
1392 }
|
cannam@167
|
1393 {
|
cannam@167
|
1394 E Ta, Tq, Tt, T15, Td, Tl, To, T16;
|
cannam@167
|
1395 {
|
cannam@167
|
1396 E T8, T9, Tr, Ts;
|
cannam@167
|
1397 T8 = rio[WS(rs, 1)];
|
cannam@167
|
1398 T9 = rio[WS(rs, 5)];
|
cannam@167
|
1399 Ta = T8 + T9;
|
cannam@167
|
1400 Tq = T8 - T9;
|
cannam@167
|
1401 Tr = iio[WS(rs, 1)];
|
cannam@167
|
1402 Ts = iio[WS(rs, 5)];
|
cannam@167
|
1403 Tt = Tr - Ts;
|
cannam@167
|
1404 T15 = Tr + Ts;
|
cannam@167
|
1405 }
|
cannam@167
|
1406 {
|
cannam@167
|
1407 E Tb, Tc, Tm, Tn;
|
cannam@167
|
1408 Tb = rio[WS(rs, 7)];
|
cannam@167
|
1409 Tc = rio[WS(rs, 3)];
|
cannam@167
|
1410 Td = Tb + Tc;
|
cannam@167
|
1411 Tl = Tb - Tc;
|
cannam@167
|
1412 Tm = iio[WS(rs, 7)];
|
cannam@167
|
1413 Tn = iio[WS(rs, 3)];
|
cannam@167
|
1414 To = Tm - Tn;
|
cannam@167
|
1415 T16 = Tm + Tn;
|
cannam@167
|
1416 }
|
cannam@167
|
1417 Te = Ta + Td;
|
cannam@167
|
1418 T17 = T15 - T16;
|
cannam@167
|
1419 T1h = T15 + T16;
|
cannam@167
|
1420 Tp = Tl - To;
|
cannam@167
|
1421 Tu = Tq + Tt;
|
cannam@167
|
1422 TE = Tl + To;
|
cannam@167
|
1423 T11 = Td - Ta;
|
cannam@167
|
1424 TD = Tt - Tq;
|
cannam@167
|
1425 }
|
cannam@167
|
1426 {
|
cannam@167
|
1427 E T1l, T1Q, T1B, T2g, T1o, T1y, T1T, T2h;
|
cannam@167
|
1428 {
|
cannam@167
|
1429 E T1j, T1k, T1z, T1A;
|
cannam@167
|
1430 T1j = rio[WS(vs, 1)];
|
cannam@167
|
1431 T1k = rio[WS(vs, 1) + WS(rs, 4)];
|
cannam@167
|
1432 T1l = T1j + T1k;
|
cannam@167
|
1433 T1Q = T1j - T1k;
|
cannam@167
|
1434 T1z = iio[WS(vs, 1)];
|
cannam@167
|
1435 T1A = iio[WS(vs, 1) + WS(rs, 4)];
|
cannam@167
|
1436 T1B = T1z - T1A;
|
cannam@167
|
1437 T2g = T1z + T1A;
|
cannam@167
|
1438 }
|
cannam@167
|
1439 {
|
cannam@167
|
1440 E T1m, T1n, T1R, T1S;
|
cannam@167
|
1441 T1m = rio[WS(vs, 1) + WS(rs, 2)];
|
cannam@167
|
1442 T1n = rio[WS(vs, 1) + WS(rs, 6)];
|
cannam@167
|
1443 T1o = T1m + T1n;
|
cannam@167
|
1444 T1y = T1m - T1n;
|
cannam@167
|
1445 T1R = iio[WS(vs, 1) + WS(rs, 2)];
|
cannam@167
|
1446 T1S = iio[WS(vs, 1) + WS(rs, 6)];
|
cannam@167
|
1447 T1T = T1R - T1S;
|
cannam@167
|
1448 T2h = T1R + T1S;
|
cannam@167
|
1449 }
|
cannam@167
|
1450 T1p = T1l + T1o;
|
cannam@167
|
1451 T2m = T1l - T1o;
|
cannam@167
|
1452 T2y = T2g + T2h;
|
cannam@167
|
1453 T1C = T1y + T1B;
|
cannam@167
|
1454 T1U = T1Q - T1T;
|
cannam@167
|
1455 T28 = T1B - T1y;
|
cannam@167
|
1456 T2i = T2g - T2h;
|
cannam@167
|
1457 T24 = T1Q + T1T;
|
cannam@167
|
1458 }
|
cannam@167
|
1459 {
|
cannam@167
|
1460 E T7U, T8a, T8d, T8P, T7X, T85, T88, T8Q;
|
cannam@167
|
1461 {
|
cannam@167
|
1462 E T7S, T7T, T8b, T8c;
|
cannam@167
|
1463 T7S = rio[WS(vs, 6) + WS(rs, 1)];
|
cannam@167
|
1464 T7T = rio[WS(vs, 6) + WS(rs, 5)];
|
cannam@167
|
1465 T7U = T7S + T7T;
|
cannam@167
|
1466 T8a = T7S - T7T;
|
cannam@167
|
1467 T8b = iio[WS(vs, 6) + WS(rs, 1)];
|
cannam@167
|
1468 T8c = iio[WS(vs, 6) + WS(rs, 5)];
|
cannam@167
|
1469 T8d = T8b - T8c;
|
cannam@167
|
1470 T8P = T8b + T8c;
|
cannam@167
|
1471 }
|
cannam@167
|
1472 {
|
cannam@167
|
1473 E T7V, T7W, T86, T87;
|
cannam@167
|
1474 T7V = rio[WS(vs, 6) + WS(rs, 7)];
|
cannam@167
|
1475 T7W = rio[WS(vs, 6) + WS(rs, 3)];
|
cannam@167
|
1476 T7X = T7V + T7W;
|
cannam@167
|
1477 T85 = T7V - T7W;
|
cannam@167
|
1478 T86 = iio[WS(vs, 6) + WS(rs, 7)];
|
cannam@167
|
1479 T87 = iio[WS(vs, 6) + WS(rs, 3)];
|
cannam@167
|
1480 T88 = T86 - T87;
|
cannam@167
|
1481 T8Q = T86 + T87;
|
cannam@167
|
1482 }
|
cannam@167
|
1483 T7Y = T7U + T7X;
|
cannam@167
|
1484 T8R = T8P - T8Q;
|
cannam@167
|
1485 T91 = T8P + T8Q;
|
cannam@167
|
1486 T89 = T85 - T88;
|
cannam@167
|
1487 T8e = T8a + T8d;
|
cannam@167
|
1488 T8o = T85 + T88;
|
cannam@167
|
1489 T8L = T7X - T7U;
|
cannam@167
|
1490 T8n = T8d - T8a;
|
cannam@167
|
1491 }
|
cannam@167
|
1492 {
|
cannam@167
|
1493 E T95, T9A, T9l, Ta0, T98, T9i, T9D, Ta1;
|
cannam@167
|
1494 {
|
cannam@167
|
1495 E T93, T94, T9j, T9k;
|
cannam@167
|
1496 T93 = rio[WS(vs, 7)];
|
cannam@167
|
1497 T94 = rio[WS(vs, 7) + WS(rs, 4)];
|
cannam@167
|
1498 T95 = T93 + T94;
|
cannam@167
|
1499 T9A = T93 - T94;
|
cannam@167
|
1500 T9j = iio[WS(vs, 7)];
|
cannam@167
|
1501 T9k = iio[WS(vs, 7) + WS(rs, 4)];
|
cannam@167
|
1502 T9l = T9j - T9k;
|
cannam@167
|
1503 Ta0 = T9j + T9k;
|
cannam@167
|
1504 }
|
cannam@167
|
1505 {
|
cannam@167
|
1506 E T96, T97, T9B, T9C;
|
cannam@167
|
1507 T96 = rio[WS(vs, 7) + WS(rs, 2)];
|
cannam@167
|
1508 T97 = rio[WS(vs, 7) + WS(rs, 6)];
|
cannam@167
|
1509 T98 = T96 + T97;
|
cannam@167
|
1510 T9i = T96 - T97;
|
cannam@167
|
1511 T9B = iio[WS(vs, 7) + WS(rs, 2)];
|
cannam@167
|
1512 T9C = iio[WS(vs, 7) + WS(rs, 6)];
|
cannam@167
|
1513 T9D = T9B - T9C;
|
cannam@167
|
1514 Ta1 = T9B + T9C;
|
cannam@167
|
1515 }
|
cannam@167
|
1516 T99 = T95 + T98;
|
cannam@167
|
1517 Ta6 = T95 - T98;
|
cannam@167
|
1518 Tai = Ta0 + Ta1;
|
cannam@167
|
1519 T9m = T9i + T9l;
|
cannam@167
|
1520 T9E = T9A - T9D;
|
cannam@167
|
1521 T9S = T9l - T9i;
|
cannam@167
|
1522 Ta2 = Ta0 - Ta1;
|
cannam@167
|
1523 T9O = T9A + T9D;
|
cannam@167
|
1524 }
|
cannam@167
|
1525 {
|
cannam@167
|
1526 E T2D, T38, T2T, T3y, T2G, T2Q, T3b, T3z;
|
cannam@167
|
1527 {
|
cannam@167
|
1528 E T2B, T2C, T2R, T2S;
|
cannam@167
|
1529 T2B = rio[WS(vs, 2)];
|
cannam@167
|
1530 T2C = rio[WS(vs, 2) + WS(rs, 4)];
|
cannam@167
|
1531 T2D = T2B + T2C;
|
cannam@167
|
1532 T38 = T2B - T2C;
|
cannam@167
|
1533 T2R = iio[WS(vs, 2)];
|
cannam@167
|
1534 T2S = iio[WS(vs, 2) + WS(rs, 4)];
|
cannam@167
|
1535 T2T = T2R - T2S;
|
cannam@167
|
1536 T3y = T2R + T2S;
|
cannam@167
|
1537 }
|
cannam@167
|
1538 {
|
cannam@167
|
1539 E T2E, T2F, T39, T3a;
|
cannam@167
|
1540 T2E = rio[WS(vs, 2) + WS(rs, 2)];
|
cannam@167
|
1541 T2F = rio[WS(vs, 2) + WS(rs, 6)];
|
cannam@167
|
1542 T2G = T2E + T2F;
|
cannam@167
|
1543 T2Q = T2E - T2F;
|
cannam@167
|
1544 T39 = iio[WS(vs, 2) + WS(rs, 2)];
|
cannam@167
|
1545 T3a = iio[WS(vs, 2) + WS(rs, 6)];
|
cannam@167
|
1546 T3b = T39 - T3a;
|
cannam@167
|
1547 T3z = T39 + T3a;
|
cannam@167
|
1548 }
|
cannam@167
|
1549 T2H = T2D + T2G;
|
cannam@167
|
1550 T3E = T2D - T2G;
|
cannam@167
|
1551 T3Q = T3y + T3z;
|
cannam@167
|
1552 T2U = T2Q + T2T;
|
cannam@167
|
1553 T3c = T38 - T3b;
|
cannam@167
|
1554 T3q = T2T - T2Q;
|
cannam@167
|
1555 T3A = T3y - T3z;
|
cannam@167
|
1556 T3m = T38 + T3b;
|
cannam@167
|
1557 }
|
cannam@167
|
1558 {
|
cannam@167
|
1559 E T42, T4i, T4l, T4X, T45, T4d, T4g, T4Y;
|
cannam@167
|
1560 {
|
cannam@167
|
1561 E T40, T41, T4j, T4k;
|
cannam@167
|
1562 T40 = rio[WS(vs, 3) + WS(rs, 1)];
|
cannam@167
|
1563 T41 = rio[WS(vs, 3) + WS(rs, 5)];
|
cannam@167
|
1564 T42 = T40 + T41;
|
cannam@167
|
1565 T4i = T40 - T41;
|
cannam@167
|
1566 T4j = iio[WS(vs, 3) + WS(rs, 1)];
|
cannam@167
|
1567 T4k = iio[WS(vs, 3) + WS(rs, 5)];
|
cannam@167
|
1568 T4l = T4j - T4k;
|
cannam@167
|
1569 T4X = T4j + T4k;
|
cannam@167
|
1570 }
|
cannam@167
|
1571 {
|
cannam@167
|
1572 E T43, T44, T4e, T4f;
|
cannam@167
|
1573 T43 = rio[WS(vs, 3) + WS(rs, 7)];
|
cannam@167
|
1574 T44 = rio[WS(vs, 3) + WS(rs, 3)];
|
cannam@167
|
1575 T45 = T43 + T44;
|
cannam@167
|
1576 T4d = T43 - T44;
|
cannam@167
|
1577 T4e = iio[WS(vs, 3) + WS(rs, 7)];
|
cannam@167
|
1578 T4f = iio[WS(vs, 3) + WS(rs, 3)];
|
cannam@167
|
1579 T4g = T4e - T4f;
|
cannam@167
|
1580 T4Y = T4e + T4f;
|
cannam@167
|
1581 }
|
cannam@167
|
1582 T46 = T42 + T45;
|
cannam@167
|
1583 T4Z = T4X - T4Y;
|
cannam@167
|
1584 T59 = T4X + T4Y;
|
cannam@167
|
1585 T4h = T4d - T4g;
|
cannam@167
|
1586 T4m = T4i + T4l;
|
cannam@167
|
1587 T4w = T4d + T4g;
|
cannam@167
|
1588 T4T = T45 - T42;
|
cannam@167
|
1589 T4v = T4l - T4i;
|
cannam@167
|
1590 }
|
cannam@167
|
1591 {
|
cannam@167
|
1592 E T5d, T5I, T5t, T68, T5g, T5q, T5L, T69;
|
cannam@167
|
1593 {
|
cannam@167
|
1594 E T5b, T5c, T5r, T5s;
|
cannam@167
|
1595 T5b = rio[WS(vs, 4)];
|
cannam@167
|
1596 T5c = rio[WS(vs, 4) + WS(rs, 4)];
|
cannam@167
|
1597 T5d = T5b + T5c;
|
cannam@167
|
1598 T5I = T5b - T5c;
|
cannam@167
|
1599 T5r = iio[WS(vs, 4)];
|
cannam@167
|
1600 T5s = iio[WS(vs, 4) + WS(rs, 4)];
|
cannam@167
|
1601 T5t = T5r - T5s;
|
cannam@167
|
1602 T68 = T5r + T5s;
|
cannam@167
|
1603 }
|
cannam@167
|
1604 {
|
cannam@167
|
1605 E T5e, T5f, T5J, T5K;
|
cannam@167
|
1606 T5e = rio[WS(vs, 4) + WS(rs, 2)];
|
cannam@167
|
1607 T5f = rio[WS(vs, 4) + WS(rs, 6)];
|
cannam@167
|
1608 T5g = T5e + T5f;
|
cannam@167
|
1609 T5q = T5e - T5f;
|
cannam@167
|
1610 T5J = iio[WS(vs, 4) + WS(rs, 2)];
|
cannam@167
|
1611 T5K = iio[WS(vs, 4) + WS(rs, 6)];
|
cannam@167
|
1612 T5L = T5J - T5K;
|
cannam@167
|
1613 T69 = T5J + T5K;
|
cannam@167
|
1614 }
|
cannam@167
|
1615 T5h = T5d + T5g;
|
cannam@167
|
1616 T6e = T5d - T5g;
|
cannam@167
|
1617 T6q = T68 + T69;
|
cannam@167
|
1618 T5u = T5q + T5t;
|
cannam@167
|
1619 T5M = T5I - T5L;
|
cannam@167
|
1620 T60 = T5t - T5q;
|
cannam@167
|
1621 T6a = T68 - T69;
|
cannam@167
|
1622 T5W = T5I + T5L;
|
cannam@167
|
1623 }
|
cannam@167
|
1624 {
|
cannam@167
|
1625 E T6C, T6S, T6V, T7x, T6F, T6N, T6Q, T7y;
|
cannam@167
|
1626 {
|
cannam@167
|
1627 E T6A, T6B, T6T, T6U;
|
cannam@167
|
1628 T6A = rio[WS(vs, 5) + WS(rs, 1)];
|
cannam@167
|
1629 T6B = rio[WS(vs, 5) + WS(rs, 5)];
|
cannam@167
|
1630 T6C = T6A + T6B;
|
cannam@167
|
1631 T6S = T6A - T6B;
|
cannam@167
|
1632 T6T = iio[WS(vs, 5) + WS(rs, 1)];
|
cannam@167
|
1633 T6U = iio[WS(vs, 5) + WS(rs, 5)];
|
cannam@167
|
1634 T6V = T6T - T6U;
|
cannam@167
|
1635 T7x = T6T + T6U;
|
cannam@167
|
1636 }
|
cannam@167
|
1637 {
|
cannam@167
|
1638 E T6D, T6E, T6O, T6P;
|
cannam@167
|
1639 T6D = rio[WS(vs, 5) + WS(rs, 7)];
|
cannam@167
|
1640 T6E = rio[WS(vs, 5) + WS(rs, 3)];
|
cannam@167
|
1641 T6F = T6D + T6E;
|
cannam@167
|
1642 T6N = T6D - T6E;
|
cannam@167
|
1643 T6O = iio[WS(vs, 5) + WS(rs, 7)];
|
cannam@167
|
1644 T6P = iio[WS(vs, 5) + WS(rs, 3)];
|
cannam@167
|
1645 T6Q = T6O - T6P;
|
cannam@167
|
1646 T7y = T6O + T6P;
|
cannam@167
|
1647 }
|
cannam@167
|
1648 T6G = T6C + T6F;
|
cannam@167
|
1649 T7z = T7x - T7y;
|
cannam@167
|
1650 T7J = T7x + T7y;
|
cannam@167
|
1651 T6R = T6N - T6Q;
|
cannam@167
|
1652 T6W = T6S + T6V;
|
cannam@167
|
1653 T76 = T6N + T6Q;
|
cannam@167
|
1654 T7t = T6F - T6C;
|
cannam@167
|
1655 T75 = T6V - T6S;
|
cannam@167
|
1656 }
|
cannam@167
|
1657 {
|
cannam@167
|
1658 E T2K, T30, T33, T3F, T2N, T2V, T2Y, T3G;
|
cannam@167
|
1659 {
|
cannam@167
|
1660 E T2I, T2J, T31, T32;
|
cannam@167
|
1661 T2I = rio[WS(vs, 2) + WS(rs, 1)];
|
cannam@167
|
1662 T2J = rio[WS(vs, 2) + WS(rs, 5)];
|
cannam@167
|
1663 T2K = T2I + T2J;
|
cannam@167
|
1664 T30 = T2I - T2J;
|
cannam@167
|
1665 T31 = iio[WS(vs, 2) + WS(rs, 1)];
|
cannam@167
|
1666 T32 = iio[WS(vs, 2) + WS(rs, 5)];
|
cannam@167
|
1667 T33 = T31 - T32;
|
cannam@167
|
1668 T3F = T31 + T32;
|
cannam@167
|
1669 }
|
cannam@167
|
1670 {
|
cannam@167
|
1671 E T2L, T2M, T2W, T2X;
|
cannam@167
|
1672 T2L = rio[WS(vs, 2) + WS(rs, 7)];
|
cannam@167
|
1673 T2M = rio[WS(vs, 2) + WS(rs, 3)];
|
cannam@167
|
1674 T2N = T2L + T2M;
|
cannam@167
|
1675 T2V = T2L - T2M;
|
cannam@167
|
1676 T2W = iio[WS(vs, 2) + WS(rs, 7)];
|
cannam@167
|
1677 T2X = iio[WS(vs, 2) + WS(rs, 3)];
|
cannam@167
|
1678 T2Y = T2W - T2X;
|
cannam@167
|
1679 T3G = T2W + T2X;
|
cannam@167
|
1680 }
|
cannam@167
|
1681 T2O = T2K + T2N;
|
cannam@167
|
1682 T3H = T3F - T3G;
|
cannam@167
|
1683 T3R = T3F + T3G;
|
cannam@167
|
1684 T2Z = T2V - T2Y;
|
cannam@167
|
1685 T34 = T30 + T33;
|
cannam@167
|
1686 T3e = T2V + T2Y;
|
cannam@167
|
1687 T3B = T2N - T2K;
|
cannam@167
|
1688 T3d = T33 - T30;
|
cannam@167
|
1689 }
|
cannam@167
|
1690 {
|
cannam@167
|
1691 E T3V, T4q, T4b, T4Q, T3Y, T48, T4t, T4R;
|
cannam@167
|
1692 {
|
cannam@167
|
1693 E T3T, T3U, T49, T4a;
|
cannam@167
|
1694 T3T = rio[WS(vs, 3)];
|
cannam@167
|
1695 T3U = rio[WS(vs, 3) + WS(rs, 4)];
|
cannam@167
|
1696 T3V = T3T + T3U;
|
cannam@167
|
1697 T4q = T3T - T3U;
|
cannam@167
|
1698 T49 = iio[WS(vs, 3)];
|
cannam@167
|
1699 T4a = iio[WS(vs, 3) + WS(rs, 4)];
|
cannam@167
|
1700 T4b = T49 - T4a;
|
cannam@167
|
1701 T4Q = T49 + T4a;
|
cannam@167
|
1702 }
|
cannam@167
|
1703 {
|
cannam@167
|
1704 E T3W, T3X, T4r, T4s;
|
cannam@167
|
1705 T3W = rio[WS(vs, 3) + WS(rs, 2)];
|
cannam@167
|
1706 T3X = rio[WS(vs, 3) + WS(rs, 6)];
|
cannam@167
|
1707 T3Y = T3W + T3X;
|
cannam@167
|
1708 T48 = T3W - T3X;
|
cannam@167
|
1709 T4r = iio[WS(vs, 3) + WS(rs, 2)];
|
cannam@167
|
1710 T4s = iio[WS(vs, 3) + WS(rs, 6)];
|
cannam@167
|
1711 T4t = T4r - T4s;
|
cannam@167
|
1712 T4R = T4r + T4s;
|
cannam@167
|
1713 }
|
cannam@167
|
1714 T3Z = T3V + T3Y;
|
cannam@167
|
1715 T4W = T3V - T3Y;
|
cannam@167
|
1716 T58 = T4Q + T4R;
|
cannam@167
|
1717 T4c = T48 + T4b;
|
cannam@167
|
1718 T4u = T4q - T4t;
|
cannam@167
|
1719 T4I = T4b - T48;
|
cannam@167
|
1720 T4S = T4Q - T4R;
|
cannam@167
|
1721 T4E = T4q + T4t;
|
cannam@167
|
1722 }
|
cannam@167
|
1723 {
|
cannam@167
|
1724 E T5k, T5A, T5D, T6f, T5n, T5v, T5y, T6g;
|
cannam@167
|
1725 {
|
cannam@167
|
1726 E T5i, T5j, T5B, T5C;
|
cannam@167
|
1727 T5i = rio[WS(vs, 4) + WS(rs, 1)];
|
cannam@167
|
1728 T5j = rio[WS(vs, 4) + WS(rs, 5)];
|
cannam@167
|
1729 T5k = T5i + T5j;
|
cannam@167
|
1730 T5A = T5i - T5j;
|
cannam@167
|
1731 T5B = iio[WS(vs, 4) + WS(rs, 1)];
|
cannam@167
|
1732 T5C = iio[WS(vs, 4) + WS(rs, 5)];
|
cannam@167
|
1733 T5D = T5B - T5C;
|
cannam@167
|
1734 T6f = T5B + T5C;
|
cannam@167
|
1735 }
|
cannam@167
|
1736 {
|
cannam@167
|
1737 E T5l, T5m, T5w, T5x;
|
cannam@167
|
1738 T5l = rio[WS(vs, 4) + WS(rs, 7)];
|
cannam@167
|
1739 T5m = rio[WS(vs, 4) + WS(rs, 3)];
|
cannam@167
|
1740 T5n = T5l + T5m;
|
cannam@167
|
1741 T5v = T5l - T5m;
|
cannam@167
|
1742 T5w = iio[WS(vs, 4) + WS(rs, 7)];
|
cannam@167
|
1743 T5x = iio[WS(vs, 4) + WS(rs, 3)];
|
cannam@167
|
1744 T5y = T5w - T5x;
|
cannam@167
|
1745 T6g = T5w + T5x;
|
cannam@167
|
1746 }
|
cannam@167
|
1747 T5o = T5k + T5n;
|
cannam@167
|
1748 T6h = T6f - T6g;
|
cannam@167
|
1749 T6r = T6f + T6g;
|
cannam@167
|
1750 T5z = T5v - T5y;
|
cannam@167
|
1751 T5E = T5A + T5D;
|
cannam@167
|
1752 T5O = T5v + T5y;
|
cannam@167
|
1753 T6b = T5n - T5k;
|
cannam@167
|
1754 T5N = T5D - T5A;
|
cannam@167
|
1755 }
|
cannam@167
|
1756 {
|
cannam@167
|
1757 E T6v, T70, T6L, T7q, T6y, T6I, T73, T7r;
|
cannam@167
|
1758 {
|
cannam@167
|
1759 E T6t, T6u, T6J, T6K;
|
cannam@167
|
1760 T6t = rio[WS(vs, 5)];
|
cannam@167
|
1761 T6u = rio[WS(vs, 5) + WS(rs, 4)];
|
cannam@167
|
1762 T6v = T6t + T6u;
|
cannam@167
|
1763 T70 = T6t - T6u;
|
cannam@167
|
1764 T6J = iio[WS(vs, 5)];
|
cannam@167
|
1765 T6K = iio[WS(vs, 5) + WS(rs, 4)];
|
cannam@167
|
1766 T6L = T6J - T6K;
|
cannam@167
|
1767 T7q = T6J + T6K;
|
cannam@167
|
1768 }
|
cannam@167
|
1769 {
|
cannam@167
|
1770 E T6w, T6x, T71, T72;
|
cannam@167
|
1771 T6w = rio[WS(vs, 5) + WS(rs, 2)];
|
cannam@167
|
1772 T6x = rio[WS(vs, 5) + WS(rs, 6)];
|
cannam@167
|
1773 T6y = T6w + T6x;
|
cannam@167
|
1774 T6I = T6w - T6x;
|
cannam@167
|
1775 T71 = iio[WS(vs, 5) + WS(rs, 2)];
|
cannam@167
|
1776 T72 = iio[WS(vs, 5) + WS(rs, 6)];
|
cannam@167
|
1777 T73 = T71 - T72;
|
cannam@167
|
1778 T7r = T71 + T72;
|
cannam@167
|
1779 }
|
cannam@167
|
1780 T6z = T6v + T6y;
|
cannam@167
|
1781 T7w = T6v - T6y;
|
cannam@167
|
1782 T7I = T7q + T7r;
|
cannam@167
|
1783 T6M = T6I + T6L;
|
cannam@167
|
1784 T74 = T70 - T73;
|
cannam@167
|
1785 T7i = T6L - T6I;
|
cannam@167
|
1786 T7s = T7q - T7r;
|
cannam@167
|
1787 T7e = T70 + T73;
|
cannam@167
|
1788 }
|
cannam@167
|
1789 rio[0] = T7 + Te;
|
cannam@167
|
1790 iio[0] = T1g + T1h;
|
cannam@167
|
1791 rio[WS(rs, 1)] = T1p + T1w;
|
cannam@167
|
1792 iio[WS(rs, 1)] = T2y + T2z;
|
cannam@167
|
1793 rio[WS(rs, 3)] = T3Z + T46;
|
cannam@167
|
1794 rio[WS(rs, 2)] = T2H + T2O;
|
cannam@167
|
1795 iio[WS(rs, 2)] = T3Q + T3R;
|
cannam@167
|
1796 iio[WS(rs, 3)] = T58 + T59;
|
cannam@167
|
1797 rio[WS(rs, 6)] = T7R + T7Y;
|
cannam@167
|
1798 iio[WS(rs, 6)] = T90 + T91;
|
cannam@167
|
1799 iio[WS(rs, 5)] = T7I + T7J;
|
cannam@167
|
1800 rio[WS(rs, 5)] = T6z + T6G;
|
cannam@167
|
1801 iio[WS(rs, 4)] = T6q + T6r;
|
cannam@167
|
1802 rio[WS(rs, 4)] = T5h + T5o;
|
cannam@167
|
1803 rio[WS(rs, 7)] = T99 + T9g;
|
cannam@167
|
1804 iio[WS(rs, 7)] = Tai + Taj;
|
cannam@167
|
1805 {
|
cannam@167
|
1806 E T12, T18, TX, T13;
|
cannam@167
|
1807 T12 = T10 - T11;
|
cannam@167
|
1808 T18 = T14 - T17;
|
cannam@167
|
1809 TX = W[10];
|
cannam@167
|
1810 T13 = W[11];
|
cannam@167
|
1811 iio[WS(vs, 6)] = FNMS(T13, T18, TX * T12);
|
cannam@167
|
1812 rio[WS(vs, 6)] = FMA(T13, T12, TX * T18);
|
cannam@167
|
1813 }
|
cannam@167
|
1814 {
|
cannam@167
|
1815 E Tag, Tak, Taf, Tah;
|
cannam@167
|
1816 Tag = T99 - T9g;
|
cannam@167
|
1817 Tak = Tai - Taj;
|
cannam@167
|
1818 Taf = W[6];
|
cannam@167
|
1819 Tah = W[7];
|
cannam@167
|
1820 rio[WS(vs, 4) + WS(rs, 7)] = FMA(Taf, Tag, Tah * Tak);
|
cannam@167
|
1821 iio[WS(vs, 4) + WS(rs, 7)] = FNMS(Tah, Tag, Taf * Tak);
|
cannam@167
|
1822 }
|
cannam@167
|
1823 {
|
cannam@167
|
1824 E T8M, T8S, T8H, T8N;
|
cannam@167
|
1825 T8M = T8K - T8L;
|
cannam@167
|
1826 T8S = T8O - T8R;
|
cannam@167
|
1827 T8H = W[10];
|
cannam@167
|
1828 T8N = W[11];
|
cannam@167
|
1829 iio[WS(vs, 6) + WS(rs, 6)] = FNMS(T8N, T8S, T8H * T8M);
|
cannam@167
|
1830 rio[WS(vs, 6) + WS(rs, 6)] = FMA(T8N, T8M, T8H * T8S);
|
cannam@167
|
1831 }
|
cannam@167
|
1832 {
|
cannam@167
|
1833 E T2k, T2q, T2f, T2l;
|
cannam@167
|
1834 T2k = T2i - T2j;
|
cannam@167
|
1835 T2q = T2m - T2p;
|
cannam@167
|
1836 T2f = W[10];
|
cannam@167
|
1837 T2l = W[11];
|
cannam@167
|
1838 iio[WS(vs, 6) + WS(rs, 1)] = FNMS(T2l, T2q, T2f * T2k);
|
cannam@167
|
1839 rio[WS(vs, 6) + WS(rs, 1)] = FMA(T2l, T2k, T2f * T2q);
|
cannam@167
|
1840 }
|
cannam@167
|
1841 {
|
cannam@167
|
1842 E Ta4, Taa, T9Z, Ta5;
|
cannam@167
|
1843 Ta4 = Ta2 - Ta3;
|
cannam@167
|
1844 Taa = Ta6 - Ta9;
|
cannam@167
|
1845 T9Z = W[10];
|
cannam@167
|
1846 Ta5 = W[11];
|
cannam@167
|
1847 iio[WS(vs, 6) + WS(rs, 7)] = FNMS(Ta5, Taa, T9Z * Ta4);
|
cannam@167
|
1848 rio[WS(vs, 6) + WS(rs, 7)] = FMA(Ta5, Ta4, T9Z * Taa);
|
cannam@167
|
1849 }
|
cannam@167
|
1850 {
|
cannam@167
|
1851 E T8Y, T92, T8X, T8Z;
|
cannam@167
|
1852 T8Y = T7R - T7Y;
|
cannam@167
|
1853 T92 = T90 - T91;
|
cannam@167
|
1854 T8X = W[6];
|
cannam@167
|
1855 T8Z = W[7];
|
cannam@167
|
1856 rio[WS(vs, 4) + WS(rs, 6)] = FMA(T8X, T8Y, T8Z * T92);
|
cannam@167
|
1857 iio[WS(vs, 4) + WS(rs, 6)] = FNMS(T8Z, T8Y, T8X * T92);
|
cannam@167
|
1858 }
|
cannam@167
|
1859 {
|
cannam@167
|
1860 E T2w, T2A, T2v, T2x;
|
cannam@167
|
1861 T2w = T1p - T1w;
|
cannam@167
|
1862 T2A = T2y - T2z;
|
cannam@167
|
1863 T2v = W[6];
|
cannam@167
|
1864 T2x = W[7];
|
cannam@167
|
1865 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T2v, T2w, T2x * T2A);
|
cannam@167
|
1866 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T2x, T2w, T2v * T2A);
|
cannam@167
|
1867 }
|
cannam@167
|
1868 {
|
cannam@167
|
1869 E Tac, Tae, Tab, Tad;
|
cannam@167
|
1870 Tac = Ta3 + Ta2;
|
cannam@167
|
1871 Tae = Ta6 + Ta9;
|
cannam@167
|
1872 Tab = W[2];
|
cannam@167
|
1873 Tad = W[3];
|
cannam@167
|
1874 iio[WS(vs, 2) + WS(rs, 7)] = FNMS(Tad, Tae, Tab * Tac);
|
cannam@167
|
1875 rio[WS(vs, 2) + WS(rs, 7)] = FMA(Tad, Tac, Tab * Tae);
|
cannam@167
|
1876 }
|
cannam@167
|
1877 {
|
cannam@167
|
1878 E T8U, T8W, T8T, T8V;
|
cannam@167
|
1879 T8U = T8L + T8K;
|
cannam@167
|
1880 T8W = T8O + T8R;
|
cannam@167
|
1881 T8T = W[2];
|
cannam@167
|
1882 T8V = W[3];
|
cannam@167
|
1883 iio[WS(vs, 2) + WS(rs, 6)] = FNMS(T8V, T8W, T8T * T8U);
|
cannam@167
|
1884 rio[WS(vs, 2) + WS(rs, 6)] = FMA(T8V, T8U, T8T * T8W);
|
cannam@167
|
1885 }
|
cannam@167
|
1886 {
|
cannam@167
|
1887 E T1a, T1c, T19, T1b;
|
cannam@167
|
1888 T1a = T11 + T10;
|
cannam@167
|
1889 T1c = T14 + T17;
|
cannam@167
|
1890 T19 = W[2];
|
cannam@167
|
1891 T1b = W[3];
|
cannam@167
|
1892 iio[WS(vs, 2)] = FNMS(T1b, T1c, T19 * T1a);
|
cannam@167
|
1893 rio[WS(vs, 2)] = FMA(T1b, T1a, T19 * T1c);
|
cannam@167
|
1894 }
|
cannam@167
|
1895 {
|
cannam@167
|
1896 E T1e, T1i, T1d, T1f;
|
cannam@167
|
1897 T1e = T7 - Te;
|
cannam@167
|
1898 T1i = T1g - T1h;
|
cannam@167
|
1899 T1d = W[6];
|
cannam@167
|
1900 T1f = W[7];
|
cannam@167
|
1901 rio[WS(vs, 4)] = FMA(T1d, T1e, T1f * T1i);
|
cannam@167
|
1902 iio[WS(vs, 4)] = FNMS(T1f, T1e, T1d * T1i);
|
cannam@167
|
1903 }
|
cannam@167
|
1904 {
|
cannam@167
|
1905 E T2s, T2u, T2r, T2t;
|
cannam@167
|
1906 T2s = T2j + T2i;
|
cannam@167
|
1907 T2u = T2m + T2p;
|
cannam@167
|
1908 T2r = W[2];
|
cannam@167
|
1909 T2t = W[3];
|
cannam@167
|
1910 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T2t, T2u, T2r * T2s);
|
cannam@167
|
1911 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T2t, T2s, T2r * T2u);
|
cannam@167
|
1912 }
|
cannam@167
|
1913 {
|
cannam@167
|
1914 E T3C, T3I, T3x, T3D;
|
cannam@167
|
1915 T3C = T3A - T3B;
|
cannam@167
|
1916 T3I = T3E - T3H;
|
cannam@167
|
1917 T3x = W[10];
|
cannam@167
|
1918 T3D = W[11];
|
cannam@167
|
1919 iio[WS(vs, 6) + WS(rs, 2)] = FNMS(T3D, T3I, T3x * T3C);
|
cannam@167
|
1920 rio[WS(vs, 6) + WS(rs, 2)] = FMA(T3D, T3C, T3x * T3I);
|
cannam@167
|
1921 }
|
cannam@167
|
1922 {
|
cannam@167
|
1923 E T4U, T50, T4P, T4V;
|
cannam@167
|
1924 T4U = T4S - T4T;
|
cannam@167
|
1925 T50 = T4W - T4Z;
|
cannam@167
|
1926 T4P = W[10];
|
cannam@167
|
1927 T4V = W[11];
|
cannam@167
|
1928 iio[WS(vs, 6) + WS(rs, 3)] = FNMS(T4V, T50, T4P * T4U);
|
cannam@167
|
1929 rio[WS(vs, 6) + WS(rs, 3)] = FMA(T4V, T4U, T4P * T50);
|
cannam@167
|
1930 }
|
cannam@167
|
1931 {
|
cannam@167
|
1932 E T56, T5a, T55, T57;
|
cannam@167
|
1933 T56 = T3Z - T46;
|
cannam@167
|
1934 T5a = T58 - T59;
|
cannam@167
|
1935 T55 = W[6];
|
cannam@167
|
1936 T57 = W[7];
|
cannam@167
|
1937 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T55, T56, T57 * T5a);
|
cannam@167
|
1938 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T57, T56, T55 * T5a);
|
cannam@167
|
1939 }
|
cannam@167
|
1940 {
|
cannam@167
|
1941 E T6o, T6s, T6n, T6p;
|
cannam@167
|
1942 T6o = T5h - T5o;
|
cannam@167
|
1943 T6s = T6q - T6r;
|
cannam@167
|
1944 T6n = W[6];
|
cannam@167
|
1945 T6p = W[7];
|
cannam@167
|
1946 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T6n, T6o, T6p * T6s);
|
cannam@167
|
1947 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T6p, T6o, T6n * T6s);
|
cannam@167
|
1948 }
|
cannam@167
|
1949 {
|
cannam@167
|
1950 E T7u, T7A, T7p, T7v;
|
cannam@167
|
1951 T7u = T7s - T7t;
|
cannam@167
|
1952 T7A = T7w - T7z;
|
cannam@167
|
1953 T7p = W[10];
|
cannam@167
|
1954 T7v = W[11];
|
cannam@167
|
1955 iio[WS(vs, 6) + WS(rs, 5)] = FNMS(T7v, T7A, T7p * T7u);
|
cannam@167
|
1956 rio[WS(vs, 6) + WS(rs, 5)] = FMA(T7v, T7u, T7p * T7A);
|
cannam@167
|
1957 }
|
cannam@167
|
1958 {
|
cannam@167
|
1959 E T6c, T6i, T67, T6d;
|
cannam@167
|
1960 T6c = T6a - T6b;
|
cannam@167
|
1961 T6i = T6e - T6h;
|
cannam@167
|
1962 T67 = W[10];
|
cannam@167
|
1963 T6d = W[11];
|
cannam@167
|
1964 iio[WS(vs, 6) + WS(rs, 4)] = FNMS(T6d, T6i, T67 * T6c);
|
cannam@167
|
1965 rio[WS(vs, 6) + WS(rs, 4)] = FMA(T6d, T6c, T67 * T6i);
|
cannam@167
|
1966 }
|
cannam@167
|
1967 {
|
cannam@167
|
1968 E T7G, T7K, T7F, T7H;
|
cannam@167
|
1969 T7G = T6z - T6G;
|
cannam@167
|
1970 T7K = T7I - T7J;
|
cannam@167
|
1971 T7F = W[6];
|
cannam@167
|
1972 T7H = W[7];
|
cannam@167
|
1973 rio[WS(vs, 4) + WS(rs, 5)] = FMA(T7F, T7G, T7H * T7K);
|
cannam@167
|
1974 iio[WS(vs, 4) + WS(rs, 5)] = FNMS(T7H, T7G, T7F * T7K);
|
cannam@167
|
1975 }
|
cannam@167
|
1976 {
|
cannam@167
|
1977 E T3O, T3S, T3N, T3P;
|
cannam@167
|
1978 T3O = T2H - T2O;
|
cannam@167
|
1979 T3S = T3Q - T3R;
|
cannam@167
|
1980 T3N = W[6];
|
cannam@167
|
1981 T3P = W[7];
|
cannam@167
|
1982 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T3N, T3O, T3P * T3S);
|
cannam@167
|
1983 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T3P, T3O, T3N * T3S);
|
cannam@167
|
1984 }
|
cannam@167
|
1985 {
|
cannam@167
|
1986 E T3K, T3M, T3J, T3L;
|
cannam@167
|
1987 T3K = T3B + T3A;
|
cannam@167
|
1988 T3M = T3E + T3H;
|
cannam@167
|
1989 T3J = W[2];
|
cannam@167
|
1990 T3L = W[3];
|
cannam@167
|
1991 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T3L, T3M, T3J * T3K);
|
cannam@167
|
1992 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T3L, T3K, T3J * T3M);
|
cannam@167
|
1993 }
|
cannam@167
|
1994 {
|
cannam@167
|
1995 E T7C, T7E, T7B, T7D;
|
cannam@167
|
1996 T7C = T7t + T7s;
|
cannam@167
|
1997 T7E = T7w + T7z;
|
cannam@167
|
1998 T7B = W[2];
|
cannam@167
|
1999 T7D = W[3];
|
cannam@167
|
2000 iio[WS(vs, 2) + WS(rs, 5)] = FNMS(T7D, T7E, T7B * T7C);
|
cannam@167
|
2001 rio[WS(vs, 2) + WS(rs, 5)] = FMA(T7D, T7C, T7B * T7E);
|
cannam@167
|
2002 }
|
cannam@167
|
2003 {
|
cannam@167
|
2004 E T6k, T6m, T6j, T6l;
|
cannam@167
|
2005 T6k = T6b + T6a;
|
cannam@167
|
2006 T6m = T6e + T6h;
|
cannam@167
|
2007 T6j = W[2];
|
cannam@167
|
2008 T6l = W[3];
|
cannam@167
|
2009 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T6l, T6m, T6j * T6k);
|
cannam@167
|
2010 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T6l, T6k, T6j * T6m);
|
cannam@167
|
2011 }
|
cannam@167
|
2012 {
|
cannam@167
|
2013 E T52, T54, T51, T53;
|
cannam@167
|
2014 T52 = T4T + T4S;
|
cannam@167
|
2015 T54 = T4W + T4Z;
|
cannam@167
|
2016 T51 = W[2];
|
cannam@167
|
2017 T53 = W[3];
|
cannam@167
|
2018 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T53, T54, T51 * T52);
|
cannam@167
|
2019 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T53, T52, T51 * T54);
|
cannam@167
|
2020 }
|
cannam@167
|
2021 {
|
cannam@167
|
2022 E T5G, T5S, T5Q, T5U, T5F, T5P;
|
cannam@167
|
2023 T5F = KP707106781 * (T5z - T5E);
|
cannam@167
|
2024 T5G = T5u - T5F;
|
cannam@167
|
2025 T5S = T5u + T5F;
|
cannam@167
|
2026 T5P = KP707106781 * (T5N - T5O);
|
cannam@167
|
2027 T5Q = T5M - T5P;
|
cannam@167
|
2028 T5U = T5M + T5P;
|
cannam@167
|
2029 {
|
cannam@167
|
2030 E T5p, T5H, T5R, T5T;
|
cannam@167
|
2031 T5p = W[12];
|
cannam@167
|
2032 T5H = W[13];
|
cannam@167
|
2033 iio[WS(vs, 7) + WS(rs, 4)] = FNMS(T5H, T5Q, T5p * T5G);
|
cannam@167
|
2034 rio[WS(vs, 7) + WS(rs, 4)] = FMA(T5H, T5G, T5p * T5Q);
|
cannam@167
|
2035 T5R = W[4];
|
cannam@167
|
2036 T5T = W[5];
|
cannam@167
|
2037 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T5T, T5U, T5R * T5S);
|
cannam@167
|
2038 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T5T, T5S, T5R * T5U);
|
cannam@167
|
2039 }
|
cannam@167
|
2040 }
|
cannam@167
|
2041 {
|
cannam@167
|
2042 E Tw, TI, TG, TK, Tv, TF;
|
cannam@167
|
2043 Tv = KP707106781 * (Tp - Tu);
|
cannam@167
|
2044 Tw = Tk - Tv;
|
cannam@167
|
2045 TI = Tk + Tv;
|
cannam@167
|
2046 TF = KP707106781 * (TD - TE);
|
cannam@167
|
2047 TG = TC - TF;
|
cannam@167
|
2048 TK = TC + TF;
|
cannam@167
|
2049 {
|
cannam@167
|
2050 E Tf, Tx, TH, TJ;
|
cannam@167
|
2051 Tf = W[12];
|
cannam@167
|
2052 Tx = W[13];
|
cannam@167
|
2053 iio[WS(vs, 7)] = FNMS(Tx, TG, Tf * Tw);
|
cannam@167
|
2054 rio[WS(vs, 7)] = FMA(Tx, Tw, Tf * TG);
|
cannam@167
|
2055 TH = W[4];
|
cannam@167
|
2056 TJ = W[5];
|
cannam@167
|
2057 iio[WS(vs, 3)] = FNMS(TJ, TK, TH * TI);
|
cannam@167
|
2058 rio[WS(vs, 3)] = FMA(TJ, TI, TH * TK);
|
cannam@167
|
2059 }
|
cannam@167
|
2060 }
|
cannam@167
|
2061 {
|
cannam@167
|
2062 E T9Q, T9W, T9U, T9Y, T9P, T9T;
|
cannam@167
|
2063 T9P = KP707106781 * (T9w + T9r);
|
cannam@167
|
2064 T9Q = T9O - T9P;
|
cannam@167
|
2065 T9W = T9O + T9P;
|
cannam@167
|
2066 T9T = KP707106781 * (T9F + T9G);
|
cannam@167
|
2067 T9U = T9S - T9T;
|
cannam@167
|
2068 T9Y = T9S + T9T;
|
cannam@167
|
2069 {
|
cannam@167
|
2070 E T9N, T9R, T9V, T9X;
|
cannam@167
|
2071 T9N = W[8];
|
cannam@167
|
2072 T9R = W[9];
|
cannam@167
|
2073 rio[WS(vs, 5) + WS(rs, 7)] = FMA(T9N, T9Q, T9R * T9U);
|
cannam@167
|
2074 iio[WS(vs, 5) + WS(rs, 7)] = FNMS(T9R, T9Q, T9N * T9U);
|
cannam@167
|
2075 T9V = W[0];
|
cannam@167
|
2076 T9X = W[1];
|
cannam@167
|
2077 rio[WS(vs, 1) + WS(rs, 7)] = FMA(T9V, T9W, T9X * T9Y);
|
cannam@167
|
2078 iio[WS(vs, 1) + WS(rs, 7)] = FNMS(T9X, T9W, T9V * T9Y);
|
cannam@167
|
2079 }
|
cannam@167
|
2080 }
|
cannam@167
|
2081 {
|
cannam@167
|
2082 E T36, T3i, T3g, T3k, T35, T3f;
|
cannam@167
|
2083 T35 = KP707106781 * (T2Z - T34);
|
cannam@167
|
2084 T36 = T2U - T35;
|
cannam@167
|
2085 T3i = T2U + T35;
|
cannam@167
|
2086 T3f = KP707106781 * (T3d - T3e);
|
cannam@167
|
2087 T3g = T3c - T3f;
|
cannam@167
|
2088 T3k = T3c + T3f;
|
cannam@167
|
2089 {
|
cannam@167
|
2090 E T2P, T37, T3h, T3j;
|
cannam@167
|
2091 T2P = W[12];
|
cannam@167
|
2092 T37 = W[13];
|
cannam@167
|
2093 iio[WS(vs, 7) + WS(rs, 2)] = FNMS(T37, T3g, T2P * T36);
|
cannam@167
|
2094 rio[WS(vs, 7) + WS(rs, 2)] = FMA(T37, T36, T2P * T3g);
|
cannam@167
|
2095 T3h = W[4];
|
cannam@167
|
2096 T3j = W[5];
|
cannam@167
|
2097 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T3j, T3k, T3h * T3i);
|
cannam@167
|
2098 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T3j, T3i, T3h * T3k);
|
cannam@167
|
2099 }
|
cannam@167
|
2100 }
|
cannam@167
|
2101 {
|
cannam@167
|
2102 E T5Y, T64, T62, T66, T5X, T61;
|
cannam@167
|
2103 T5X = KP707106781 * (T5E + T5z);
|
cannam@167
|
2104 T5Y = T5W - T5X;
|
cannam@167
|
2105 T64 = T5W + T5X;
|
cannam@167
|
2106 T61 = KP707106781 * (T5N + T5O);
|
cannam@167
|
2107 T62 = T60 - T61;
|
cannam@167
|
2108 T66 = T60 + T61;
|
cannam@167
|
2109 {
|
cannam@167
|
2110 E T5V, T5Z, T63, T65;
|
cannam@167
|
2111 T5V = W[8];
|
cannam@167
|
2112 T5Z = W[9];
|
cannam@167
|
2113 rio[WS(vs, 5) + WS(rs, 4)] = FMA(T5V, T5Y, T5Z * T62);
|
cannam@167
|
2114 iio[WS(vs, 5) + WS(rs, 4)] = FNMS(T5Z, T5Y, T5V * T62);
|
cannam@167
|
2115 T63 = W[0];
|
cannam@167
|
2116 T65 = W[1];
|
cannam@167
|
2117 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T63, T64, T65 * T66);
|
cannam@167
|
2118 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T65, T64, T63 * T66);
|
cannam@167
|
2119 }
|
cannam@167
|
2120 }
|
cannam@167
|
2121 {
|
cannam@167
|
2122 E T7g, T7m, T7k, T7o, T7f, T7j;
|
cannam@167
|
2123 T7f = KP707106781 * (T6W + T6R);
|
cannam@167
|
2124 T7g = T7e - T7f;
|
cannam@167
|
2125 T7m = T7e + T7f;
|
cannam@167
|
2126 T7j = KP707106781 * (T75 + T76);
|
cannam@167
|
2127 T7k = T7i - T7j;
|
cannam@167
|
2128 T7o = T7i + T7j;
|
cannam@167
|
2129 {
|
cannam@167
|
2130 E T7d, T7h, T7l, T7n;
|
cannam@167
|
2131 T7d = W[8];
|
cannam@167
|
2132 T7h = W[9];
|
cannam@167
|
2133 rio[WS(vs, 5) + WS(rs, 5)] = FMA(T7d, T7g, T7h * T7k);
|
cannam@167
|
2134 iio[WS(vs, 5) + WS(rs, 5)] = FNMS(T7h, T7g, T7d * T7k);
|
cannam@167
|
2135 T7l = W[0];
|
cannam@167
|
2136 T7n = W[1];
|
cannam@167
|
2137 rio[WS(vs, 1) + WS(rs, 5)] = FMA(T7l, T7m, T7n * T7o);
|
cannam@167
|
2138 iio[WS(vs, 1) + WS(rs, 5)] = FNMS(T7n, T7m, T7l * T7o);
|
cannam@167
|
2139 }
|
cannam@167
|
2140 }
|
cannam@167
|
2141 {
|
cannam@167
|
2142 E T8g, T8s, T8q, T8u, T8f, T8p;
|
cannam@167
|
2143 T8f = KP707106781 * (T89 - T8e);
|
cannam@167
|
2144 T8g = T84 - T8f;
|
cannam@167
|
2145 T8s = T84 + T8f;
|
cannam@167
|
2146 T8p = KP707106781 * (T8n - T8o);
|
cannam@167
|
2147 T8q = T8m - T8p;
|
cannam@167
|
2148 T8u = T8m + T8p;
|
cannam@167
|
2149 {
|
cannam@167
|
2150 E T7Z, T8h, T8r, T8t;
|
cannam@167
|
2151 T7Z = W[12];
|
cannam@167
|
2152 T8h = W[13];
|
cannam@167
|
2153 iio[WS(vs, 7) + WS(rs, 6)] = FNMS(T8h, T8q, T7Z * T8g);
|
cannam@167
|
2154 rio[WS(vs, 7) + WS(rs, 6)] = FMA(T8h, T8g, T7Z * T8q);
|
cannam@167
|
2155 T8r = W[4];
|
cannam@167
|
2156 T8t = W[5];
|
cannam@167
|
2157 iio[WS(vs, 3) + WS(rs, 6)] = FNMS(T8t, T8u, T8r * T8s);
|
cannam@167
|
2158 rio[WS(vs, 3) + WS(rs, 6)] = FMA(T8t, T8s, T8r * T8u);
|
cannam@167
|
2159 }
|
cannam@167
|
2160 }
|
cannam@167
|
2161 {
|
cannam@167
|
2162 E T4G, T4M, T4K, T4O, T4F, T4J;
|
cannam@167
|
2163 T4F = KP707106781 * (T4m + T4h);
|
cannam@167
|
2164 T4G = T4E - T4F;
|
cannam@167
|
2165 T4M = T4E + T4F;
|
cannam@167
|
2166 T4J = KP707106781 * (T4v + T4w);
|
cannam@167
|
2167 T4K = T4I - T4J;
|
cannam@167
|
2168 T4O = T4I + T4J;
|
cannam@167
|
2169 {
|
cannam@167
|
2170 E T4D, T4H, T4L, T4N;
|
cannam@167
|
2171 T4D = W[8];
|
cannam@167
|
2172 T4H = W[9];
|
cannam@167
|
2173 rio[WS(vs, 5) + WS(rs, 3)] = FMA(T4D, T4G, T4H * T4K);
|
cannam@167
|
2174 iio[WS(vs, 5) + WS(rs, 3)] = FNMS(T4H, T4G, T4D * T4K);
|
cannam@167
|
2175 T4L = W[0];
|
cannam@167
|
2176 T4N = W[1];
|
cannam@167
|
2177 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T4L, T4M, T4N * T4O);
|
cannam@167
|
2178 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T4N, T4M, T4L * T4O);
|
cannam@167
|
2179 }
|
cannam@167
|
2180 }
|
cannam@167
|
2181 {
|
cannam@167
|
2182 E TO, TU, TS, TW, TN, TR;
|
cannam@167
|
2183 TN = KP707106781 * (Tu + Tp);
|
cannam@167
|
2184 TO = TM - TN;
|
cannam@167
|
2185 TU = TM + TN;
|
cannam@167
|
2186 TR = KP707106781 * (TD + TE);
|
cannam@167
|
2187 TS = TQ - TR;
|
cannam@167
|
2188 TW = TQ + TR;
|
cannam@167
|
2189 {
|
cannam@167
|
2190 E TL, TP, TT, TV;
|
cannam@167
|
2191 TL = W[8];
|
cannam@167
|
2192 TP = W[9];
|
cannam@167
|
2193 rio[WS(vs, 5)] = FMA(TL, TO, TP * TS);
|
cannam@167
|
2194 iio[WS(vs, 5)] = FNMS(TP, TO, TL * TS);
|
cannam@167
|
2195 TT = W[0];
|
cannam@167
|
2196 TV = W[1];
|
cannam@167
|
2197 rio[WS(vs, 1)] = FMA(TT, TU, TV * TW);
|
cannam@167
|
2198 iio[WS(vs, 1)] = FNMS(TV, TU, TT * TW);
|
cannam@167
|
2199 }
|
cannam@167
|
2200 }
|
cannam@167
|
2201 {
|
cannam@167
|
2202 E T26, T2c, T2a, T2e, T25, T29;
|
cannam@167
|
2203 T25 = KP707106781 * (T1M + T1H);
|
cannam@167
|
2204 T26 = T24 - T25;
|
cannam@167
|
2205 T2c = T24 + T25;
|
cannam@167
|
2206 T29 = KP707106781 * (T1V + T1W);
|
cannam@167
|
2207 T2a = T28 - T29;
|
cannam@167
|
2208 T2e = T28 + T29;
|
cannam@167
|
2209 {
|
cannam@167
|
2210 E T23, T27, T2b, T2d;
|
cannam@167
|
2211 T23 = W[8];
|
cannam@167
|
2212 T27 = W[9];
|
cannam@167
|
2213 rio[WS(vs, 5) + WS(rs, 1)] = FMA(T23, T26, T27 * T2a);
|
cannam@167
|
2214 iio[WS(vs, 5) + WS(rs, 1)] = FNMS(T27, T26, T23 * T2a);
|
cannam@167
|
2215 T2b = W[0];
|
cannam@167
|
2216 T2d = W[1];
|
cannam@167
|
2217 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T2b, T2c, T2d * T2e);
|
cannam@167
|
2218 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T2d, T2c, T2b * T2e);
|
cannam@167
|
2219 }
|
cannam@167
|
2220 }
|
cannam@167
|
2221 {
|
cannam@167
|
2222 E T9y, T9K, T9I, T9M, T9x, T9H;
|
cannam@167
|
2223 T9x = KP707106781 * (T9r - T9w);
|
cannam@167
|
2224 T9y = T9m - T9x;
|
cannam@167
|
2225 T9K = T9m + T9x;
|
cannam@167
|
2226 T9H = KP707106781 * (T9F - T9G);
|
cannam@167
|
2227 T9I = T9E - T9H;
|
cannam@167
|
2228 T9M = T9E + T9H;
|
cannam@167
|
2229 {
|
cannam@167
|
2230 E T9h, T9z, T9J, T9L;
|
cannam@167
|
2231 T9h = W[12];
|
cannam@167
|
2232 T9z = W[13];
|
cannam@167
|
2233 iio[WS(vs, 7) + WS(rs, 7)] = FNMS(T9z, T9I, T9h * T9y);
|
cannam@167
|
2234 rio[WS(vs, 7) + WS(rs, 7)] = FMA(T9z, T9y, T9h * T9I);
|
cannam@167
|
2235 T9J = W[4];
|
cannam@167
|
2236 T9L = W[5];
|
cannam@167
|
2237 iio[WS(vs, 3) + WS(rs, 7)] = FNMS(T9L, T9M, T9J * T9K);
|
cannam@167
|
2238 rio[WS(vs, 3) + WS(rs, 7)] = FMA(T9L, T9K, T9J * T9M);
|
cannam@167
|
2239 }
|
cannam@167
|
2240 }
|
cannam@167
|
2241 {
|
cannam@167
|
2242 E T6Y, T7a, T78, T7c, T6X, T77;
|
cannam@167
|
2243 T6X = KP707106781 * (T6R - T6W);
|
cannam@167
|
2244 T6Y = T6M - T6X;
|
cannam@167
|
2245 T7a = T6M + T6X;
|
cannam@167
|
2246 T77 = KP707106781 * (T75 - T76);
|
cannam@167
|
2247 T78 = T74 - T77;
|
cannam@167
|
2248 T7c = T74 + T77;
|
cannam@167
|
2249 {
|
cannam@167
|
2250 E T6H, T6Z, T79, T7b;
|
cannam@167
|
2251 T6H = W[12];
|
cannam@167
|
2252 T6Z = W[13];
|
cannam@167
|
2253 iio[WS(vs, 7) + WS(rs, 5)] = FNMS(T6Z, T78, T6H * T6Y);
|
cannam@167
|
2254 rio[WS(vs, 7) + WS(rs, 5)] = FMA(T6Z, T6Y, T6H * T78);
|
cannam@167
|
2255 T79 = W[4];
|
cannam@167
|
2256 T7b = W[5];
|
cannam@167
|
2257 iio[WS(vs, 3) + WS(rs, 5)] = FNMS(T7b, T7c, T79 * T7a);
|
cannam@167
|
2258 rio[WS(vs, 3) + WS(rs, 5)] = FMA(T7b, T7a, T79 * T7c);
|
cannam@167
|
2259 }
|
cannam@167
|
2260 }
|
cannam@167
|
2261 {
|
cannam@167
|
2262 E T1O, T20, T1Y, T22, T1N, T1X;
|
cannam@167
|
2263 T1N = KP707106781 * (T1H - T1M);
|
cannam@167
|
2264 T1O = T1C - T1N;
|
cannam@167
|
2265 T20 = T1C + T1N;
|
cannam@167
|
2266 T1X = KP707106781 * (T1V - T1W);
|
cannam@167
|
2267 T1Y = T1U - T1X;
|
cannam@167
|
2268 T22 = T1U + T1X;
|
cannam@167
|
2269 {
|
cannam@167
|
2270 E T1x, T1P, T1Z, T21;
|
cannam@167
|
2271 T1x = W[12];
|
cannam@167
|
2272 T1P = W[13];
|
cannam@167
|
2273 iio[WS(vs, 7) + WS(rs, 1)] = FNMS(T1P, T1Y, T1x * T1O);
|
cannam@167
|
2274 rio[WS(vs, 7) + WS(rs, 1)] = FMA(T1P, T1O, T1x * T1Y);
|
cannam@167
|
2275 T1Z = W[4];
|
cannam@167
|
2276 T21 = W[5];
|
cannam@167
|
2277 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T21, T22, T1Z * T20);
|
cannam@167
|
2278 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T21, T20, T1Z * T22);
|
cannam@167
|
2279 }
|
cannam@167
|
2280 }
|
cannam@167
|
2281 {
|
cannam@167
|
2282 E T4o, T4A, T4y, T4C, T4n, T4x;
|
cannam@167
|
2283 T4n = KP707106781 * (T4h - T4m);
|
cannam@167
|
2284 T4o = T4c - T4n;
|
cannam@167
|
2285 T4A = T4c + T4n;
|
cannam@167
|
2286 T4x = KP707106781 * (T4v - T4w);
|
cannam@167
|
2287 T4y = T4u - T4x;
|
cannam@167
|
2288 T4C = T4u + T4x;
|
cannam@167
|
2289 {
|
cannam@167
|
2290 E T47, T4p, T4z, T4B;
|
cannam@167
|
2291 T47 = W[12];
|
cannam@167
|
2292 T4p = W[13];
|
cannam@167
|
2293 iio[WS(vs, 7) + WS(rs, 3)] = FNMS(T4p, T4y, T47 * T4o);
|
cannam@167
|
2294 rio[WS(vs, 7) + WS(rs, 3)] = FMA(T4p, T4o, T47 * T4y);
|
cannam@167
|
2295 T4z = W[4];
|
cannam@167
|
2296 T4B = W[5];
|
cannam@167
|
2297 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T4B, T4C, T4z * T4A);
|
cannam@167
|
2298 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T4B, T4A, T4z * T4C);
|
cannam@167
|
2299 }
|
cannam@167
|
2300 }
|
cannam@167
|
2301 {
|
cannam@167
|
2302 E T3o, T3u, T3s, T3w, T3n, T3r;
|
cannam@167
|
2303 T3n = KP707106781 * (T34 + T2Z);
|
cannam@167
|
2304 T3o = T3m - T3n;
|
cannam@167
|
2305 T3u = T3m + T3n;
|
cannam@167
|
2306 T3r = KP707106781 * (T3d + T3e);
|
cannam@167
|
2307 T3s = T3q - T3r;
|
cannam@167
|
2308 T3w = T3q + T3r;
|
cannam@167
|
2309 {
|
cannam@167
|
2310 E T3l, T3p, T3t, T3v;
|
cannam@167
|
2311 T3l = W[8];
|
cannam@167
|
2312 T3p = W[9];
|
cannam@167
|
2313 rio[WS(vs, 5) + WS(rs, 2)] = FMA(T3l, T3o, T3p * T3s);
|
cannam@167
|
2314 iio[WS(vs, 5) + WS(rs, 2)] = FNMS(T3p, T3o, T3l * T3s);
|
cannam@167
|
2315 T3t = W[0];
|
cannam@167
|
2316 T3v = W[1];
|
cannam@167
|
2317 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T3t, T3u, T3v * T3w);
|
cannam@167
|
2318 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T3v, T3u, T3t * T3w);
|
cannam@167
|
2319 }
|
cannam@167
|
2320 }
|
cannam@167
|
2321 {
|
cannam@167
|
2322 E T8y, T8E, T8C, T8G, T8x, T8B;
|
cannam@167
|
2323 T8x = KP707106781 * (T8e + T89);
|
cannam@167
|
2324 T8y = T8w - T8x;
|
cannam@167
|
2325 T8E = T8w + T8x;
|
cannam@167
|
2326 T8B = KP707106781 * (T8n + T8o);
|
cannam@167
|
2327 T8C = T8A - T8B;
|
cannam@167
|
2328 T8G = T8A + T8B;
|
cannam@167
|
2329 {
|
cannam@167
|
2330 E T8v, T8z, T8D, T8F;
|
cannam@167
|
2331 T8v = W[8];
|
cannam@167
|
2332 T8z = W[9];
|
cannam@167
|
2333 rio[WS(vs, 5) + WS(rs, 6)] = FMA(T8v, T8y, T8z * T8C);
|
cannam@167
|
2334 iio[WS(vs, 5) + WS(rs, 6)] = FNMS(T8z, T8y, T8v * T8C);
|
cannam@167
|
2335 T8D = W[0];
|
cannam@167
|
2336 T8F = W[1];
|
cannam@167
|
2337 rio[WS(vs, 1) + WS(rs, 6)] = FMA(T8D, T8E, T8F * T8G);
|
cannam@167
|
2338 iio[WS(vs, 1) + WS(rs, 6)] = FNMS(T8F, T8E, T8D * T8G);
|
cannam@167
|
2339 }
|
cannam@167
|
2340 }
|
cannam@167
|
2341 }
|
cannam@167
|
2342 }
|
cannam@167
|
2343 }
|
cannam@167
|
2344
|
cannam@167
|
2345 static const tw_instr twinstr[] = {
|
cannam@167
|
2346 {TW_FULL, 0, 8},
|
cannam@167
|
2347 {TW_NEXT, 1, 0}
|
cannam@167
|
2348 };
|
cannam@167
|
2349
|
cannam@167
|
2350 static const ct_desc desc = { 8, "q1_8", twinstr, &GENUS, {416, 144, 112, 0}, 0, 0, 0 };
|
cannam@167
|
2351
|
cannam@167
|
2352 void X(codelet_q1_8) (planner *p) {
|
cannam@167
|
2353 X(kdft_difsq_register) (p, q1_8, &desc);
|
cannam@167
|
2354 }
|
cannam@167
|
2355 #endif
|