Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:41:25 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-rdft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 32 -dif -name hb2_32 -include hb.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 488 FP additions, 350 FP multiplications,
|
Chris@10
|
32 * (or, 236 additions, 98 multiplications, 252 fused multiply/add),
|
Chris@10
|
33 * 204 stack variables, 7 constants, and 128 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hb.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hb2_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
40 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@10
|
41 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
42 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@10
|
43 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
44 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@10
|
45 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
46 {
|
Chris@10
|
47 INT m;
|
Chris@10
|
48 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 8, MAKE_VOLATILE_STRIDE(64, rs)) {
|
Chris@10
|
49 E T5u, T6b, T6e, T5I, T66, T60, T5U, T5R, T67, T5L, T61, T5x, T5A, T5D, T5O;
|
Chris@10
|
50 E T62, T5V, T5P;
|
Chris@10
|
51 {
|
Chris@10
|
52 E T11, T14, T12, T37, T17, T1b, T39, T15, T7C, T8P, T8S, T7I, T98, T7e, T78;
|
Chris@10
|
53 E T8V, T3d, T3x, T3a, T3v, T9s, T3G, T4p, T5X, T16, T9m, T3y, T4b, T3C, T4g;
|
Chris@10
|
54 E T5Z, T1a, T4r, T3J, T2O, T1c, T4W, T4s, T3Y, T3K, T3l, T3e, T3i, T3q, T8K;
|
Chris@10
|
55 E T8E, T8m, T7S, T5k, T5e;
|
Chris@10
|
56 {
|
Chris@10
|
57 E T13, T3c, T38, T3F, T7B, T9l, T77, T7d, T9r, T7H;
|
Chris@10
|
58 T11 = W[2];
|
Chris@10
|
59 T14 = W[3];
|
Chris@10
|
60 T12 = W[4];
|
Chris@10
|
61 T37 = W[0];
|
Chris@10
|
62 T17 = W[6];
|
Chris@10
|
63 T1b = W[7];
|
Chris@10
|
64 T13 = T11 * T12;
|
Chris@10
|
65 T3c = T37 * T14;
|
Chris@10
|
66 T38 = T37 * T11;
|
Chris@10
|
67 T3F = T37 * T12;
|
Chris@10
|
68 T7B = T11 * T17;
|
Chris@10
|
69 T9l = T12 * T17;
|
Chris@10
|
70 T77 = T37 * T17;
|
Chris@10
|
71 T7d = T37 * T1b;
|
Chris@10
|
72 T9r = T12 * T1b;
|
Chris@10
|
73 T7H = T11 * T1b;
|
Chris@10
|
74 T39 = W[1];
|
Chris@10
|
75 T15 = W[5];
|
Chris@10
|
76 {
|
Chris@10
|
77 E T3I, T19, T5d, T3b, T18, T2N;
|
Chris@10
|
78 T7C = FMA(T14, T1b, T7B);
|
Chris@10
|
79 T8P = FNMS(T14, T1b, T7B);
|
Chris@10
|
80 T8S = FMA(T14, T17, T7H);
|
Chris@10
|
81 T7I = FNMS(T14, T17, T7H);
|
Chris@10
|
82 T98 = FNMS(T39, T17, T7d);
|
Chris@10
|
83 T7e = FMA(T39, T17, T7d);
|
Chris@10
|
84 T78 = FNMS(T39, T1b, T77);
|
Chris@10
|
85 T8V = FMA(T39, T1b, T77);
|
Chris@10
|
86 T3d = FMA(T39, T11, T3c);
|
Chris@10
|
87 T3x = FNMS(T39, T11, T3c);
|
Chris@10
|
88 T3a = FNMS(T39, T14, T38);
|
Chris@10
|
89 T3v = FMA(T39, T14, T38);
|
Chris@10
|
90 T9s = FNMS(T15, T17, T9r);
|
Chris@10
|
91 T3G = FNMS(T39, T15, T3F);
|
Chris@10
|
92 T4p = FMA(T39, T15, T3F);
|
Chris@10
|
93 T5X = FNMS(T14, T15, T13);
|
Chris@10
|
94 T16 = FMA(T14, T15, T13);
|
Chris@10
|
95 T3I = T37 * T15;
|
Chris@10
|
96 T19 = T11 * T15;
|
Chris@10
|
97 T5d = T3v * T12;
|
Chris@10
|
98 T3b = T3a * T12;
|
Chris@10
|
99 T9m = FMA(T15, T1b, T9l);
|
Chris@10
|
100 {
|
Chris@10
|
101 E T3w, T3B, T5t, T5H;
|
Chris@10
|
102 T3w = T3v * T17;
|
Chris@10
|
103 T3B = T3v * T1b;
|
Chris@10
|
104 T5t = T3a * T17;
|
Chris@10
|
105 T5H = T3a * T1b;
|
Chris@10
|
106 T3y = FNMS(T3x, T1b, T3w);
|
Chris@10
|
107 T4b = FMA(T3x, T1b, T3w);
|
Chris@10
|
108 T3C = FMA(T3x, T17, T3B);
|
Chris@10
|
109 T4g = FNMS(T3x, T17, T3B);
|
Chris@10
|
110 T5u = FMA(T3d, T1b, T5t);
|
Chris@10
|
111 T6b = FNMS(T3d, T1b, T5t);
|
Chris@10
|
112 T6e = FMA(T3d, T17, T5H);
|
Chris@10
|
113 T5I = FNMS(T3d, T17, T5H);
|
Chris@10
|
114 T18 = T16 * T17;
|
Chris@10
|
115 T2N = T16 * T1b;
|
Chris@10
|
116 T5Z = FMA(T14, T12, T19);
|
Chris@10
|
117 T1a = FNMS(T14, T12, T19);
|
Chris@10
|
118 }
|
Chris@10
|
119 {
|
Chris@10
|
120 E T3H, T3X, T4q, T4V, T5Y, T65;
|
Chris@10
|
121 T4q = T4p * T17;
|
Chris@10
|
122 T4V = T4p * T1b;
|
Chris@10
|
123 T4r = FNMS(T39, T12, T3I);
|
Chris@10
|
124 T3J = FMA(T39, T12, T3I);
|
Chris@10
|
125 T2O = FNMS(T1a, T17, T2N);
|
Chris@10
|
126 T1c = FMA(T1a, T1b, T18);
|
Chris@10
|
127 T3H = T3G * T17;
|
Chris@10
|
128 T4W = FNMS(T4r, T17, T4V);
|
Chris@10
|
129 T4s = FMA(T4r, T1b, T4q);
|
Chris@10
|
130 T3X = T3G * T1b;
|
Chris@10
|
131 T5Y = T5X * T17;
|
Chris@10
|
132 T65 = T5X * T1b;
|
Chris@10
|
133 T3Y = FNMS(T3J, T17, T3X);
|
Chris@10
|
134 T3K = FMA(T3J, T1b, T3H);
|
Chris@10
|
135 {
|
Chris@10
|
136 E T8J, T8D, T3h, T5j, T8l, T7R;
|
Chris@10
|
137 T3h = T3a * T15;
|
Chris@10
|
138 T66 = FNMS(T5Z, T17, T65);
|
Chris@10
|
139 T60 = FMA(T5Z, T1b, T5Y);
|
Chris@10
|
140 T3l = FNMS(T3d, T15, T3b);
|
Chris@10
|
141 T3e = FMA(T3d, T15, T3b);
|
Chris@10
|
142 T3i = FNMS(T3d, T12, T3h);
|
Chris@10
|
143 T3q = FMA(T3d, T12, T3h);
|
Chris@10
|
144 T8J = T3l * T1b;
|
Chris@10
|
145 T8D = T3l * T17;
|
Chris@10
|
146 T5j = T3v * T15;
|
Chris@10
|
147 T8l = T3e * T1b;
|
Chris@10
|
148 T7R = T3e * T17;
|
Chris@10
|
149 T8K = FNMS(T3q, T17, T8J);
|
Chris@10
|
150 T8E = FMA(T3q, T1b, T8D);
|
Chris@10
|
151 T8m = FNMS(T3i, T17, T8l);
|
Chris@10
|
152 T7S = FMA(T3i, T1b, T7R);
|
Chris@10
|
153 T5U = FNMS(T3x, T12, T5j);
|
Chris@10
|
154 T5k = FMA(T3x, T12, T5j);
|
Chris@10
|
155 T5e = FNMS(T3x, T15, T5d);
|
Chris@10
|
156 T5R = FMA(T3x, T15, T5d);
|
Chris@10
|
157 }
|
Chris@10
|
158 }
|
Chris@10
|
159 }
|
Chris@10
|
160 }
|
Chris@10
|
161 {
|
Chris@10
|
162 E T6O, T6i, T7s, T7o, T6j, Tf, T8W, T7V, T99, T8p, T3L, T1t, T3Z, T2X, T5J;
|
Chris@10
|
163 E T4Z, T7t, T6W, T5v, T4v, TZ, T7x, T91, T9d, T28, T3S, T3R, T2h, T5B, T4Q;
|
Chris@10
|
164 E T8v, T8a, T5C, T4N, T6Z, T6J, TK, T7w, T2z, T3P, T94, T9c, T3O, T2I, T5y;
|
Chris@10
|
165 E T4J, T8u, T8h, T5z, T4G, T6Y, T6A, T6p, T6m, T6P, Tu, T9a, T82, T8X, T8s;
|
Chris@10
|
166 E T40, T1Q, T4y, T4B, T3M, T30, T5w, T52;
|
Chris@10
|
167 {
|
Chris@10
|
168 E T6B, T6I, T4L, T4M, T4t, T4u;
|
Chris@10
|
169 {
|
Chris@10
|
170 E T1d, T3, T2P, T6, T6Q, T2S, T6R, T1g, Td, T6U, T1i, Ta, T2V, T1r, T6T;
|
Chris@10
|
171 E T1l;
|
Chris@10
|
172 {
|
Chris@10
|
173 E T4, T5, T2Q, T2R, T1, T2, T1e, T1f;
|
Chris@10
|
174 T1 = cr[0];
|
Chris@10
|
175 T2 = ci[WS(rs, 15)];
|
Chris@10
|
176 {
|
Chris@10
|
177 E T6N, T6h, T7r, T7n;
|
Chris@10
|
178 T6N = T5R * T1b;
|
Chris@10
|
179 T6h = T5R * T17;
|
Chris@10
|
180 T7r = T5e * T1b;
|
Chris@10
|
181 T7n = T5e * T17;
|
Chris@10
|
182 T6O = FNMS(T5U, T17, T6N);
|
Chris@10
|
183 T6i = FMA(T5U, T1b, T6h);
|
Chris@10
|
184 T7s = FNMS(T5k, T17, T7r);
|
Chris@10
|
185 T7o = FMA(T5k, T1b, T7n);
|
Chris@10
|
186 T1d = T1 - T2;
|
Chris@10
|
187 T3 = T1 + T2;
|
Chris@10
|
188 }
|
Chris@10
|
189 T4 = cr[WS(rs, 8)];
|
Chris@10
|
190 T5 = ci[WS(rs, 7)];
|
Chris@10
|
191 T2Q = ci[WS(rs, 31)];
|
Chris@10
|
192 T2R = cr[WS(rs, 16)];
|
Chris@10
|
193 T1e = ci[WS(rs, 23)];
|
Chris@10
|
194 T2P = T4 - T5;
|
Chris@10
|
195 T6 = T4 + T5;
|
Chris@10
|
196 T6Q = T2Q - T2R;
|
Chris@10
|
197 T2S = T2Q + T2R;
|
Chris@10
|
198 T1f = cr[WS(rs, 24)];
|
Chris@10
|
199 {
|
Chris@10
|
200 E T1o, T1n, T1p, Tb, Tc;
|
Chris@10
|
201 Tb = ci[WS(rs, 3)];
|
Chris@10
|
202 Tc = cr[WS(rs, 12)];
|
Chris@10
|
203 T1o = ci[WS(rs, 19)];
|
Chris@10
|
204 T6R = T1e - T1f;
|
Chris@10
|
205 T1g = T1e + T1f;
|
Chris@10
|
206 T1n = Tb - Tc;
|
Chris@10
|
207 Td = Tb + Tc;
|
Chris@10
|
208 T1p = cr[WS(rs, 28)];
|
Chris@10
|
209 {
|
Chris@10
|
210 E T1j, T1k, T8, T9, T1q;
|
Chris@10
|
211 T8 = cr[WS(rs, 4)];
|
Chris@10
|
212 T9 = ci[WS(rs, 11)];
|
Chris@10
|
213 T1q = T1o + T1p;
|
Chris@10
|
214 T6U = T1o - T1p;
|
Chris@10
|
215 T1j = ci[WS(rs, 27)];
|
Chris@10
|
216 T1i = T8 - T9;
|
Chris@10
|
217 Ta = T8 + T9;
|
Chris@10
|
218 T1k = cr[WS(rs, 20)];
|
Chris@10
|
219 T2V = T1n + T1q;
|
Chris@10
|
220 T1r = T1n - T1q;
|
Chris@10
|
221 T6T = T1j - T1k;
|
Chris@10
|
222 T1l = T1j + T1k;
|
Chris@10
|
223 }
|
Chris@10
|
224 }
|
Chris@10
|
225 }
|
Chris@10
|
226 {
|
Chris@10
|
227 E T2U, T6V, T6S, T1h, T1s, T4Y, T4X, T2T, T2W;
|
Chris@10
|
228 {
|
Chris@10
|
229 E T7T, T8o, T1m, T7U, T7, Te, T8n;
|
Chris@10
|
230 T7T = T3 - T6;
|
Chris@10
|
231 T7 = T3 + T6;
|
Chris@10
|
232 Te = Ta + Td;
|
Chris@10
|
233 T8o = Ta - Td;
|
Chris@10
|
234 T1m = T1i - T1l;
|
Chris@10
|
235 T2U = T1i + T1l;
|
Chris@10
|
236 T6j = T7 - Te;
|
Chris@10
|
237 Tf = T7 + Te;
|
Chris@10
|
238 T7U = T6U - T6T;
|
Chris@10
|
239 T6V = T6T + T6U;
|
Chris@10
|
240 T6S = T6Q + T6R;
|
Chris@10
|
241 T8n = T6Q - T6R;
|
Chris@10
|
242 T4t = T1d + T1g;
|
Chris@10
|
243 T1h = T1d - T1g;
|
Chris@10
|
244 T8W = T7T + T7U;
|
Chris@10
|
245 T7V = T7T - T7U;
|
Chris@10
|
246 T99 = T8o + T8n;
|
Chris@10
|
247 T8p = T8n - T8o;
|
Chris@10
|
248 T1s = T1m + T1r;
|
Chris@10
|
249 T4Y = T1m - T1r;
|
Chris@10
|
250 }
|
Chris@10
|
251 T4X = T2S - T2P;
|
Chris@10
|
252 T2T = T2P + T2S;
|
Chris@10
|
253 T2W = T2U - T2V;
|
Chris@10
|
254 T4u = T2U + T2V;
|
Chris@10
|
255 T3L = FMA(KP707106781, T1s, T1h);
|
Chris@10
|
256 T1t = FNMS(KP707106781, T1s, T1h);
|
Chris@10
|
257 T3Z = FMA(KP707106781, T2W, T2T);
|
Chris@10
|
258 T2X = FNMS(KP707106781, T2W, T2T);
|
Chris@10
|
259 T5J = FNMS(KP707106781, T4Y, T4X);
|
Chris@10
|
260 T4Z = FMA(KP707106781, T4Y, T4X);
|
Chris@10
|
261 T7t = T6S + T6V;
|
Chris@10
|
262 T6W = T6S - T6V;
|
Chris@10
|
263 }
|
Chris@10
|
264 }
|
Chris@10
|
265 {
|
Chris@10
|
266 E T29, T1S, T1V, T87, TR, T2c, T84, T6E, T1X, TU, T1Y, T6G, T25, T22, TX;
|
Chris@10
|
267 E T1Z;
|
Chris@10
|
268 {
|
Chris@10
|
269 E TO, TN, TP, TL, TM, T6C, T6D;
|
Chris@10
|
270 TL = ci[0];
|
Chris@10
|
271 TM = cr[WS(rs, 15)];
|
Chris@10
|
272 TO = cr[WS(rs, 7)];
|
Chris@10
|
273 T5v = FMA(KP707106781, T4u, T4t);
|
Chris@10
|
274 T4v = FNMS(KP707106781, T4u, T4t);
|
Chris@10
|
275 TN = TL + TM;
|
Chris@10
|
276 T29 = TL - TM;
|
Chris@10
|
277 TP = ci[WS(rs, 8)];
|
Chris@10
|
278 {
|
Chris@10
|
279 E T2a, T2b, T1T, T1U, TQ;
|
Chris@10
|
280 T1T = ci[WS(rs, 16)];
|
Chris@10
|
281 T1U = cr[WS(rs, 31)];
|
Chris@10
|
282 TQ = TO + TP;
|
Chris@10
|
283 T1S = TO - TP;
|
Chris@10
|
284 T2a = ci[WS(rs, 24)];
|
Chris@10
|
285 T6C = T1T - T1U;
|
Chris@10
|
286 T1V = T1T + T1U;
|
Chris@10
|
287 T2b = cr[WS(rs, 23)];
|
Chris@10
|
288 T87 = TN - TQ;
|
Chris@10
|
289 TR = TN + TQ;
|
Chris@10
|
290 T2c = T2a + T2b;
|
Chris@10
|
291 T6D = T2a - T2b;
|
Chris@10
|
292 }
|
Chris@10
|
293 {
|
Chris@10
|
294 E T23, T24, TS, TT, TV, TW;
|
Chris@10
|
295 TS = cr[WS(rs, 3)];
|
Chris@10
|
296 TT = ci[WS(rs, 12)];
|
Chris@10
|
297 T84 = T6C - T6D;
|
Chris@10
|
298 T6E = T6C + T6D;
|
Chris@10
|
299 T23 = ci[WS(rs, 20)];
|
Chris@10
|
300 T1X = TS - TT;
|
Chris@10
|
301 TU = TS + TT;
|
Chris@10
|
302 T24 = cr[WS(rs, 27)];
|
Chris@10
|
303 TV = ci[WS(rs, 4)];
|
Chris@10
|
304 TW = cr[WS(rs, 11)];
|
Chris@10
|
305 T1Y = ci[WS(rs, 28)];
|
Chris@10
|
306 T6G = T23 - T24;
|
Chris@10
|
307 T25 = T23 + T24;
|
Chris@10
|
308 T22 = TV - TW;
|
Chris@10
|
309 TX = TV + TW;
|
Chris@10
|
310 T1Z = cr[WS(rs, 19)];
|
Chris@10
|
311 }
|
Chris@10
|
312 }
|
Chris@10
|
313 {
|
Chris@10
|
314 E T4O, T1W, T2f, T26, T8Z, T86, T2e, T21, T89, T90;
|
Chris@10
|
315 {
|
Chris@10
|
316 E T85, TY, T6F, T20, T6H, T88;
|
Chris@10
|
317 T4O = T1S + T1V;
|
Chris@10
|
318 T1W = T1S - T1V;
|
Chris@10
|
319 T2f = T22 - T25;
|
Chris@10
|
320 T26 = T22 + T25;
|
Chris@10
|
321 T85 = TU - TX;
|
Chris@10
|
322 TY = TU + TX;
|
Chris@10
|
323 T6F = T1Y - T1Z;
|
Chris@10
|
324 T20 = T1Y + T1Z;
|
Chris@10
|
325 T8Z = T85 + T84;
|
Chris@10
|
326 T86 = T84 - T85;
|
Chris@10
|
327 T6B = TR - TY;
|
Chris@10
|
328 TZ = TR + TY;
|
Chris@10
|
329 T6H = T6F + T6G;
|
Chris@10
|
330 T88 = T6G - T6F;
|
Chris@10
|
331 T2e = T1X - T20;
|
Chris@10
|
332 T21 = T1X + T20;
|
Chris@10
|
333 T7x = T6E + T6H;
|
Chris@10
|
334 T6I = T6E - T6H;
|
Chris@10
|
335 T89 = T87 - T88;
|
Chris@10
|
336 T90 = T87 + T88;
|
Chris@10
|
337 }
|
Chris@10
|
338 {
|
Chris@10
|
339 E T4P, T2d, T27, T2g;
|
Chris@10
|
340 T2d = T29 - T2c;
|
Chris@10
|
341 T4L = T29 + T2c;
|
Chris@10
|
342 T4M = T21 + T26;
|
Chris@10
|
343 T27 = T21 - T26;
|
Chris@10
|
344 T2g = T2e + T2f;
|
Chris@10
|
345 T4P = T2e - T2f;
|
Chris@10
|
346 T91 = FNMS(KP414213562, T90, T8Z);
|
Chris@10
|
347 T9d = FMA(KP414213562, T8Z, T90);
|
Chris@10
|
348 T28 = FNMS(KP707106781, T27, T1W);
|
Chris@10
|
349 T3S = FMA(KP707106781, T27, T1W);
|
Chris@10
|
350 T3R = FMA(KP707106781, T2g, T2d);
|
Chris@10
|
351 T2h = FNMS(KP707106781, T2g, T2d);
|
Chris@10
|
352 T5B = FMA(KP707106781, T4P, T4O);
|
Chris@10
|
353 T4Q = FNMS(KP707106781, T4P, T4O);
|
Chris@10
|
354 T8v = FNMS(KP414213562, T86, T89);
|
Chris@10
|
355 T8a = FMA(KP414213562, T89, T86);
|
Chris@10
|
356 }
|
Chris@10
|
357 }
|
Chris@10
|
358 }
|
Chris@10
|
359 {
|
Chris@10
|
360 E T6s, T6z, T4F, T4E;
|
Chris@10
|
361 {
|
Chris@10
|
362 E T2A, T2j, TC, T8e, T2m, T2D, T6v, T8b, TG, T2o, TF, T6x, T2w, TH, T2p;
|
Chris@10
|
363 E T2q;
|
Chris@10
|
364 {
|
Chris@10
|
365 E Tw, Tx, Tz, TA, T6t, T6u;
|
Chris@10
|
366 Tw = cr[WS(rs, 1)];
|
Chris@10
|
367 T5C = FMA(KP707106781, T4M, T4L);
|
Chris@10
|
368 T4N = FNMS(KP707106781, T4M, T4L);
|
Chris@10
|
369 T6Z = T6I - T6B;
|
Chris@10
|
370 T6J = T6B + T6I;
|
Chris@10
|
371 Tx = ci[WS(rs, 14)];
|
Chris@10
|
372 Tz = cr[WS(rs, 9)];
|
Chris@10
|
373 TA = ci[WS(rs, 6)];
|
Chris@10
|
374 {
|
Chris@10
|
375 E T2k, Ty, TB, T2l, T2B, T2C;
|
Chris@10
|
376 T2k = ci[WS(rs, 30)];
|
Chris@10
|
377 T2A = Tw - Tx;
|
Chris@10
|
378 Ty = Tw + Tx;
|
Chris@10
|
379 T2j = Tz - TA;
|
Chris@10
|
380 TB = Tz + TA;
|
Chris@10
|
381 T2l = cr[WS(rs, 17)];
|
Chris@10
|
382 T2B = ci[WS(rs, 22)];
|
Chris@10
|
383 T2C = cr[WS(rs, 25)];
|
Chris@10
|
384 TC = Ty + TB;
|
Chris@10
|
385 T8e = Ty - TB;
|
Chris@10
|
386 T2m = T2k + T2l;
|
Chris@10
|
387 T6t = T2k - T2l;
|
Chris@10
|
388 T6u = T2B - T2C;
|
Chris@10
|
389 T2D = T2B + T2C;
|
Chris@10
|
390 }
|
Chris@10
|
391 {
|
Chris@10
|
392 E TD, TE, T2u, T2v;
|
Chris@10
|
393 TD = cr[WS(rs, 5)];
|
Chris@10
|
394 T6v = T6t + T6u;
|
Chris@10
|
395 T8b = T6t - T6u;
|
Chris@10
|
396 TE = ci[WS(rs, 10)];
|
Chris@10
|
397 T2u = ci[WS(rs, 18)];
|
Chris@10
|
398 T2v = cr[WS(rs, 29)];
|
Chris@10
|
399 TG = ci[WS(rs, 2)];
|
Chris@10
|
400 T2o = TD - TE;
|
Chris@10
|
401 TF = TD + TE;
|
Chris@10
|
402 T6x = T2u - T2v;
|
Chris@10
|
403 T2w = T2u + T2v;
|
Chris@10
|
404 TH = cr[WS(rs, 13)];
|
Chris@10
|
405 T2p = ci[WS(rs, 26)];
|
Chris@10
|
406 T2q = cr[WS(rs, 21)];
|
Chris@10
|
407 }
|
Chris@10
|
408 }
|
Chris@10
|
409 {
|
Chris@10
|
410 E T4H, T2n, T2G, T2F, T92, T8d, T2y, T93, T8g, T4I, T2E, T2H;
|
Chris@10
|
411 {
|
Chris@10
|
412 E T2x, T8c, T8f, T2s, T2t, TI;
|
Chris@10
|
413 T4H = T2m - T2j;
|
Chris@10
|
414 T2n = T2j + T2m;
|
Chris@10
|
415 T2t = TG - TH;
|
Chris@10
|
416 TI = TG + TH;
|
Chris@10
|
417 {
|
Chris@10
|
418 E T6w, T2r, TJ, T6y;
|
Chris@10
|
419 T6w = T2p - T2q;
|
Chris@10
|
420 T2r = T2p + T2q;
|
Chris@10
|
421 T2G = T2t - T2w;
|
Chris@10
|
422 T2x = T2t + T2w;
|
Chris@10
|
423 T8c = TF - TI;
|
Chris@10
|
424 TJ = TF + TI;
|
Chris@10
|
425 T6y = T6w + T6x;
|
Chris@10
|
426 T8f = T6x - T6w;
|
Chris@10
|
427 T2F = T2o - T2r;
|
Chris@10
|
428 T2s = T2o + T2r;
|
Chris@10
|
429 TK = TC + TJ;
|
Chris@10
|
430 T6s = TC - TJ;
|
Chris@10
|
431 T6z = T6v - T6y;
|
Chris@10
|
432 T7w = T6v + T6y;
|
Chris@10
|
433 }
|
Chris@10
|
434 T92 = T8c + T8b;
|
Chris@10
|
435 T8d = T8b - T8c;
|
Chris@10
|
436 T4F = T2s + T2x;
|
Chris@10
|
437 T2y = T2s - T2x;
|
Chris@10
|
438 T93 = T8e + T8f;
|
Chris@10
|
439 T8g = T8e - T8f;
|
Chris@10
|
440 }
|
Chris@10
|
441 T4E = T2A + T2D;
|
Chris@10
|
442 T2E = T2A - T2D;
|
Chris@10
|
443 T2H = T2F + T2G;
|
Chris@10
|
444 T4I = T2G - T2F;
|
Chris@10
|
445 T2z = FNMS(KP707106781, T2y, T2n);
|
Chris@10
|
446 T3P = FMA(KP707106781, T2y, T2n);
|
Chris@10
|
447 T94 = FMA(KP414213562, T93, T92);
|
Chris@10
|
448 T9c = FNMS(KP414213562, T92, T93);
|
Chris@10
|
449 T3O = FMA(KP707106781, T2H, T2E);
|
Chris@10
|
450 T2I = FNMS(KP707106781, T2H, T2E);
|
Chris@10
|
451 T5y = FMA(KP707106781, T4I, T4H);
|
Chris@10
|
452 T4J = FNMS(KP707106781, T4I, T4H);
|
Chris@10
|
453 T8u = FMA(KP414213562, T8d, T8g);
|
Chris@10
|
454 T8h = FNMS(KP414213562, T8g, T8d);
|
Chris@10
|
455 }
|
Chris@10
|
456 }
|
Chris@10
|
457 {
|
Chris@10
|
458 E T4x, T1O, Tm, T7Z, T80, T4w, T1J, T4A, T1D, Tt, T7X, T7W, T4z, T1y;
|
Chris@10
|
459 {
|
Chris@10
|
460 E Tj, T1K, Ti, T6o, T1N, Tk, T1G, T1H;
|
Chris@10
|
461 {
|
Chris@10
|
462 E Tg, Th, T1L, T1M;
|
Chris@10
|
463 Tg = cr[WS(rs, 2)];
|
Chris@10
|
464 T5z = FMA(KP707106781, T4F, T4E);
|
Chris@10
|
465 T4G = FNMS(KP707106781, T4F, T4E);
|
Chris@10
|
466 T6Y = T6s + T6z;
|
Chris@10
|
467 T6A = T6s - T6z;
|
Chris@10
|
468 Th = ci[WS(rs, 13)];
|
Chris@10
|
469 T1L = ci[WS(rs, 21)];
|
Chris@10
|
470 T1M = cr[WS(rs, 26)];
|
Chris@10
|
471 Tj = cr[WS(rs, 10)];
|
Chris@10
|
472 T1K = Tg - Th;
|
Chris@10
|
473 Ti = Tg + Th;
|
Chris@10
|
474 T6o = T1L - T1M;
|
Chris@10
|
475 T1N = T1L + T1M;
|
Chris@10
|
476 Tk = ci[WS(rs, 5)];
|
Chris@10
|
477 T1G = ci[WS(rs, 29)];
|
Chris@10
|
478 T1H = cr[WS(rs, 18)];
|
Chris@10
|
479 }
|
Chris@10
|
480 {
|
Chris@10
|
481 E T1F, Tl, T6n, T1I;
|
Chris@10
|
482 T4x = T1K + T1N;
|
Chris@10
|
483 T1O = T1K - T1N;
|
Chris@10
|
484 T1F = Tj - Tk;
|
Chris@10
|
485 Tl = Tj + Tk;
|
Chris@10
|
486 T6n = T1G - T1H;
|
Chris@10
|
487 T1I = T1G + T1H;
|
Chris@10
|
488 Tm = Ti + Tl;
|
Chris@10
|
489 T7Z = Ti - Tl;
|
Chris@10
|
490 T80 = T6n - T6o;
|
Chris@10
|
491 T6p = T6n + T6o;
|
Chris@10
|
492 T4w = T1I - T1F;
|
Chris@10
|
493 T1J = T1F + T1I;
|
Chris@10
|
494 }
|
Chris@10
|
495 }
|
Chris@10
|
496 {
|
Chris@10
|
497 E Tq, T1z, Tp, T6l, T1C, Tr, T1v, T1w;
|
Chris@10
|
498 {
|
Chris@10
|
499 E Tn, To, T1A, T1B;
|
Chris@10
|
500 Tn = ci[WS(rs, 1)];
|
Chris@10
|
501 To = cr[WS(rs, 14)];
|
Chris@10
|
502 T1A = ci[WS(rs, 25)];
|
Chris@10
|
503 T1B = cr[WS(rs, 22)];
|
Chris@10
|
504 Tq = cr[WS(rs, 6)];
|
Chris@10
|
505 T1z = Tn - To;
|
Chris@10
|
506 Tp = Tn + To;
|
Chris@10
|
507 T6l = T1A - T1B;
|
Chris@10
|
508 T1C = T1A + T1B;
|
Chris@10
|
509 Tr = ci[WS(rs, 9)];
|
Chris@10
|
510 T1v = ci[WS(rs, 17)];
|
Chris@10
|
511 T1w = cr[WS(rs, 30)];
|
Chris@10
|
512 }
|
Chris@10
|
513 {
|
Chris@10
|
514 E T1u, Ts, T6k, T1x;
|
Chris@10
|
515 T4A = T1z + T1C;
|
Chris@10
|
516 T1D = T1z - T1C;
|
Chris@10
|
517 T1u = Tq - Tr;
|
Chris@10
|
518 Ts = Tq + Tr;
|
Chris@10
|
519 T6k = T1v - T1w;
|
Chris@10
|
520 T1x = T1v + T1w;
|
Chris@10
|
521 Tt = Tp + Ts;
|
Chris@10
|
522 T7X = Tp - Ts;
|
Chris@10
|
523 T7W = T6k - T6l;
|
Chris@10
|
524 T6m = T6k + T6l;
|
Chris@10
|
525 T4z = T1u + T1x;
|
Chris@10
|
526 T1y = T1u - T1x;
|
Chris@10
|
527 }
|
Chris@10
|
528 }
|
Chris@10
|
529 {
|
Chris@10
|
530 E T8r, T8q, T2Z, T1E, T1P, T2Y, T7Y, T81, T50, T51;
|
Chris@10
|
531 T8r = T7X + T7W;
|
Chris@10
|
532 T7Y = T7W - T7X;
|
Chris@10
|
533 T81 = T7Z + T80;
|
Chris@10
|
534 T8q = T7Z - T80;
|
Chris@10
|
535 T6P = Tm - Tt;
|
Chris@10
|
536 Tu = Tm + Tt;
|
Chris@10
|
537 T9a = T81 + T7Y;
|
Chris@10
|
538 T82 = T7Y - T81;
|
Chris@10
|
539 T2Z = FMA(KP414213562, T1y, T1D);
|
Chris@10
|
540 T1E = FNMS(KP414213562, T1D, T1y);
|
Chris@10
|
541 T1P = FMA(KP414213562, T1O, T1J);
|
Chris@10
|
542 T2Y = FNMS(KP414213562, T1J, T1O);
|
Chris@10
|
543 T8X = T8q + T8r;
|
Chris@10
|
544 T8s = T8q - T8r;
|
Chris@10
|
545 T40 = T1P + T1E;
|
Chris@10
|
546 T1Q = T1E - T1P;
|
Chris@10
|
547 T4y = FNMS(KP414213562, T4x, T4w);
|
Chris@10
|
548 T50 = FMA(KP414213562, T4w, T4x);
|
Chris@10
|
549 T51 = FMA(KP414213562, T4z, T4A);
|
Chris@10
|
550 T4B = FNMS(KP414213562, T4A, T4z);
|
Chris@10
|
551 T3M = T2Y + T2Z;
|
Chris@10
|
552 T30 = T2Y - T2Z;
|
Chris@10
|
553 T5w = T50 + T51;
|
Chris@10
|
554 T52 = T50 - T51;
|
Chris@10
|
555 }
|
Chris@10
|
556 }
|
Chris@10
|
557 }
|
Chris@10
|
558 }
|
Chris@10
|
559 {
|
Chris@10
|
560 E T7D, T5K, T4C, T7K, T7J, T7E, T83, T8w, T8t, T8i, T6r, T70, T6X, T6K;
|
Chris@10
|
561 {
|
Chris@10
|
562 E T6q, T8Y, T9e, T9b, T95, T8L, T8Q, T8H, T8M, T8I, T8R;
|
Chris@10
|
563 {
|
Chris@10
|
564 E Tv, T10, T7v, T7y, T7u;
|
Chris@10
|
565 T7D = Tf - Tu;
|
Chris@10
|
566 Tv = Tf + Tu;
|
Chris@10
|
567 T7u = T6p + T6m;
|
Chris@10
|
568 T6q = T6m - T6p;
|
Chris@10
|
569 T5K = T4B - T4y;
|
Chris@10
|
570 T4C = T4y + T4B;
|
Chris@10
|
571 T10 = TK + TZ;
|
Chris@10
|
572 T7K = TK - TZ;
|
Chris@10
|
573 T7J = T7t - T7u;
|
Chris@10
|
574 T7v = T7t + T7u;
|
Chris@10
|
575 T7y = T7w + T7x;
|
Chris@10
|
576 T7E = T7x - T7w;
|
Chris@10
|
577 {
|
Chris@10
|
578 E T9t, T9x, T9p, T9u, T9q, T9y;
|
Chris@10
|
579 {
|
Chris@10
|
580 E T9n, T7z, T9o, T7A, T7q, T7p;
|
Chris@10
|
581 T8Y = FNMS(KP707106781, T8X, T8W);
|
Chris@10
|
582 T9n = FMA(KP707106781, T8X, T8W);
|
Chris@10
|
583 cr[0] = Tv + T10;
|
Chris@10
|
584 T7p = Tv - T10;
|
Chris@10
|
585 ci[0] = T7v + T7y;
|
Chris@10
|
586 T7z = T7v - T7y;
|
Chris@10
|
587 T9o = T9c + T9d;
|
Chris@10
|
588 T9e = T9c - T9d;
|
Chris@10
|
589 T7A = T7s * T7p;
|
Chris@10
|
590 T7q = T7o * T7p;
|
Chris@10
|
591 T9b = FNMS(KP707106781, T9a, T99);
|
Chris@10
|
592 T9t = FMA(KP707106781, T9a, T99);
|
Chris@10
|
593 T9x = FMA(KP923879532, T9o, T9n);
|
Chris@10
|
594 T9p = FNMS(KP923879532, T9o, T9n);
|
Chris@10
|
595 ci[WS(rs, 16)] = FMA(T7o, T7z, T7A);
|
Chris@10
|
596 cr[WS(rs, 16)] = FNMS(T7s, T7z, T7q);
|
Chris@10
|
597 T9u = T94 + T91;
|
Chris@10
|
598 T95 = T91 - T94;
|
Chris@10
|
599 }
|
Chris@10
|
600 T9q = T9m * T9p;
|
Chris@10
|
601 T9y = T3v * T9x;
|
Chris@10
|
602 {
|
Chris@10
|
603 E T8F, T9z, T9v, T8G, T9A, T9w;
|
Chris@10
|
604 T83 = FMA(KP707106781, T82, T7V);
|
Chris@10
|
605 T8F = FNMS(KP707106781, T82, T7V);
|
Chris@10
|
606 T9z = FMA(KP923879532, T9u, T9t);
|
Chris@10
|
607 T9v = FNMS(KP923879532, T9u, T9t);
|
Chris@10
|
608 T8G = T8u + T8v;
|
Chris@10
|
609 T8w = T8u - T8v;
|
Chris@10
|
610 T8t = FMA(KP707106781, T8s, T8p);
|
Chris@10
|
611 T8L = FNMS(KP707106781, T8s, T8p);
|
Chris@10
|
612 T9A = T3v * T9z;
|
Chris@10
|
613 cr[WS(rs, 2)] = FNMS(T3x, T9z, T9y);
|
Chris@10
|
614 T9w = T9m * T9v;
|
Chris@10
|
615 cr[WS(rs, 18)] = FNMS(T9s, T9v, T9q);
|
Chris@10
|
616 T8Q = FMA(KP923879532, T8G, T8F);
|
Chris@10
|
617 T8H = FNMS(KP923879532, T8G, T8F);
|
Chris@10
|
618 ci[WS(rs, 2)] = FMA(T3x, T9x, T9A);
|
Chris@10
|
619 ci[WS(rs, 18)] = FMA(T9s, T9p, T9w);
|
Chris@10
|
620 T8M = T8h + T8a;
|
Chris@10
|
621 T8i = T8a - T8h;
|
Chris@10
|
622 }
|
Chris@10
|
623 T8I = T8E * T8H;
|
Chris@10
|
624 T8R = T8P * T8Q;
|
Chris@10
|
625 }
|
Chris@10
|
626 }
|
Chris@10
|
627 {
|
Chris@10
|
628 E T7f, T7j, T7b, T7g, T7c, T7k;
|
Chris@10
|
629 {
|
Chris@10
|
630 E T79, T8T, T8N, T7a, T8U, T8O;
|
Chris@10
|
631 T6r = T6j + T6q;
|
Chris@10
|
632 T79 = T6j - T6q;
|
Chris@10
|
633 T8T = FMA(KP923879532, T8M, T8L);
|
Chris@10
|
634 T8N = FNMS(KP923879532, T8M, T8L);
|
Chris@10
|
635 T7a = T6Z - T6Y;
|
Chris@10
|
636 T70 = T6Y + T6Z;
|
Chris@10
|
637 T6X = T6P + T6W;
|
Chris@10
|
638 T7f = T6W - T6P;
|
Chris@10
|
639 T8U = T8P * T8T;
|
Chris@10
|
640 cr[WS(rs, 30)] = FNMS(T8S, T8T, T8R);
|
Chris@10
|
641 T8O = T8E * T8N;
|
Chris@10
|
642 cr[WS(rs, 14)] = FNMS(T8K, T8N, T8I);
|
Chris@10
|
643 T7j = FMA(KP707106781, T7a, T79);
|
Chris@10
|
644 T7b = FNMS(KP707106781, T7a, T79);
|
Chris@10
|
645 ci[WS(rs, 30)] = FMA(T8S, T8Q, T8U);
|
Chris@10
|
646 ci[WS(rs, 14)] = FMA(T8K, T8H, T8O);
|
Chris@10
|
647 T7g = T6A - T6J;
|
Chris@10
|
648 T6K = T6A + T6J;
|
Chris@10
|
649 }
|
Chris@10
|
650 T7c = T78 * T7b;
|
Chris@10
|
651 T7k = T5X * T7j;
|
Chris@10
|
652 {
|
Chris@10
|
653 E T97, T9g, T9i, T9j, T9f, T9k, T9h, T96;
|
Chris@10
|
654 {
|
Chris@10
|
655 E T7l, T7h, T7m, T7i;
|
Chris@10
|
656 T7l = FMA(KP707106781, T7g, T7f);
|
Chris@10
|
657 T7h = FNMS(KP707106781, T7g, T7f);
|
Chris@10
|
658 T7m = T5X * T7l;
|
Chris@10
|
659 cr[WS(rs, 12)] = FNMS(T5Z, T7l, T7k);
|
Chris@10
|
660 T7i = T78 * T7h;
|
Chris@10
|
661 cr[WS(rs, 28)] = FNMS(T7e, T7h, T7c);
|
Chris@10
|
662 T9h = FMA(KP923879532, T95, T8Y);
|
Chris@10
|
663 T96 = FNMS(KP923879532, T95, T8Y);
|
Chris@10
|
664 ci[WS(rs, 12)] = FMA(T5Z, T7j, T7m);
|
Chris@10
|
665 ci[WS(rs, 28)] = FMA(T7e, T7b, T7i);
|
Chris@10
|
666 }
|
Chris@10
|
667 T97 = T8V * T96;
|
Chris@10
|
668 T9g = T98 * T96;
|
Chris@10
|
669 T9i = T3G * T9h;
|
Chris@10
|
670 T9j = FMA(KP923879532, T9e, T9b);
|
Chris@10
|
671 T9f = FNMS(KP923879532, T9e, T9b);
|
Chris@10
|
672 T9k = T3J * T9h;
|
Chris@10
|
673 cr[WS(rs, 10)] = FNMS(T3J, T9j, T9i);
|
Chris@10
|
674 ci[WS(rs, 26)] = FMA(T8V, T9f, T9g);
|
Chris@10
|
675 cr[WS(rs, 26)] = FNMS(T98, T9f, T97);
|
Chris@10
|
676 ci[WS(rs, 10)] = FMA(T3G, T9j, T9k);
|
Chris@10
|
677 }
|
Chris@10
|
678 }
|
Chris@10
|
679 }
|
Chris@10
|
680 {
|
Chris@10
|
681 E T31, T3r, T1R, T3m, T33, T32, T3s, T2K, T8z, T8j;
|
Chris@10
|
682 {
|
Chris@10
|
683 E T73, T6L, T75, T71;
|
Chris@10
|
684 T73 = FMA(KP707106781, T6K, T6r);
|
Chris@10
|
685 T6L = FNMS(KP707106781, T6K, T6r);
|
Chris@10
|
686 T75 = FMA(KP707106781, T70, T6X);
|
Chris@10
|
687 T71 = FNMS(KP707106781, T70, T6X);
|
Chris@10
|
688 {
|
Chris@10
|
689 E T76, T74, T72, T6M;
|
Chris@10
|
690 T76 = T3d * T73;
|
Chris@10
|
691 T74 = T3a * T73;
|
Chris@10
|
692 T72 = T6O * T6L;
|
Chris@10
|
693 T6M = T6i * T6L;
|
Chris@10
|
694 ci[WS(rs, 4)] = FMA(T3a, T75, T76);
|
Chris@10
|
695 cr[WS(rs, 4)] = FNMS(T3d, T75, T74);
|
Chris@10
|
696 ci[WS(rs, 20)] = FMA(T6i, T71, T72);
|
Chris@10
|
697 cr[WS(rs, 20)] = FNMS(T6O, T71, T6M);
|
Chris@10
|
698 }
|
Chris@10
|
699 }
|
Chris@10
|
700 {
|
Chris@10
|
701 E T7N, T7F, T7P, T7L;
|
Chris@10
|
702 T7N = T7D + T7E;
|
Chris@10
|
703 T7F = T7D - T7E;
|
Chris@10
|
704 T7P = T7K + T7J;
|
Chris@10
|
705 T7L = T7J - T7K;
|
Chris@10
|
706 {
|
Chris@10
|
707 E T7O, T7G, T7Q, T7M;
|
Chris@10
|
708 T7O = T4p * T7N;
|
Chris@10
|
709 T7G = T7C * T7F;
|
Chris@10
|
710 T7Q = T4p * T7P;
|
Chris@10
|
711 T7M = T7C * T7L;
|
Chris@10
|
712 cr[WS(rs, 8)] = FNMS(T4r, T7P, T7O);
|
Chris@10
|
713 cr[WS(rs, 24)] = FNMS(T7I, T7L, T7G);
|
Chris@10
|
714 ci[WS(rs, 8)] = FMA(T4r, T7N, T7Q);
|
Chris@10
|
715 ci[WS(rs, 24)] = FMA(T7I, T7F, T7M);
|
Chris@10
|
716 }
|
Chris@10
|
717 }
|
Chris@10
|
718 T31 = FMA(KP923879532, T30, T2X);
|
Chris@10
|
719 T3r = FNMS(KP923879532, T30, T2X);
|
Chris@10
|
720 T8z = FMA(KP923879532, T8i, T83);
|
Chris@10
|
721 T8j = FNMS(KP923879532, T8i, T83);
|
Chris@10
|
722 {
|
Chris@10
|
723 E T8B, T8x, T8C, T8A;
|
Chris@10
|
724 T8B = FMA(KP923879532, T8w, T8t);
|
Chris@10
|
725 T8x = FNMS(KP923879532, T8w, T8t);
|
Chris@10
|
726 T8C = T1a * T8z;
|
Chris@10
|
727 T8A = T16 * T8z;
|
Chris@10
|
728 {
|
Chris@10
|
729 E T8y, T8k, T2i, T2J;
|
Chris@10
|
730 T8y = T8m * T8j;
|
Chris@10
|
731 T8k = T7S * T8j;
|
Chris@10
|
732 ci[WS(rs, 6)] = FMA(T16, T8B, T8C);
|
Chris@10
|
733 cr[WS(rs, 6)] = FNMS(T1a, T8B, T8A);
|
Chris@10
|
734 ci[WS(rs, 22)] = FMA(T7S, T8x, T8y);
|
Chris@10
|
735 cr[WS(rs, 22)] = FNMS(T8m, T8x, T8k);
|
Chris@10
|
736 T1R = FMA(KP923879532, T1Q, T1t);
|
Chris@10
|
737 T3m = FNMS(KP923879532, T1Q, T1t);
|
Chris@10
|
738 T33 = FNMS(KP668178637, T28, T2h);
|
Chris@10
|
739 T2i = FMA(KP668178637, T2h, T28);
|
Chris@10
|
740 T2J = FNMS(KP668178637, T2I, T2z);
|
Chris@10
|
741 T32 = FMA(KP668178637, T2z, T2I);
|
Chris@10
|
742 T3s = T2J + T2i;
|
Chris@10
|
743 T2K = T2i - T2J;
|
Chris@10
|
744 }
|
Chris@10
|
745 }
|
Chris@10
|
746 {
|
Chris@10
|
747 E T5l, T53, T5f, T4D, T4K, T4R, T56, T5g;
|
Chris@10
|
748 T5l = FNMS(KP923879532, T52, T4Z);
|
Chris@10
|
749 T53 = FMA(KP923879532, T52, T4Z);
|
Chris@10
|
750 {
|
Chris@10
|
751 E T3t, T3D, T3f, T2L;
|
Chris@10
|
752 T3t = FNMS(KP831469612, T3s, T3r);
|
Chris@10
|
753 T3D = FMA(KP831469612, T3s, T3r);
|
Chris@10
|
754 T3f = FMA(KP831469612, T2K, T1R);
|
Chris@10
|
755 T2L = FNMS(KP831469612, T2K, T1R);
|
Chris@10
|
756 {
|
Chris@10
|
757 E T3n, T34, T3g, T2M;
|
Chris@10
|
758 T3n = T32 + T33;
|
Chris@10
|
759 T34 = T32 - T33;
|
Chris@10
|
760 T3g = T3e * T3f;
|
Chris@10
|
761 T2M = T1c * T2L;
|
Chris@10
|
762 {
|
Chris@10
|
763 E T3o, T3z, T3j, T35;
|
Chris@10
|
764 T3o = FNMS(KP831469612, T3n, T3m);
|
Chris@10
|
765 T3z = FMA(KP831469612, T3n, T3m);
|
Chris@10
|
766 T3j = FMA(KP831469612, T34, T31);
|
Chris@10
|
767 T35 = FNMS(KP831469612, T34, T31);
|
Chris@10
|
768 {
|
Chris@10
|
769 E T3u, T3p, T3E, T3A;
|
Chris@10
|
770 T3u = T3q * T3o;
|
Chris@10
|
771 T3p = T3l * T3o;
|
Chris@10
|
772 T3E = T3C * T3z;
|
Chris@10
|
773 T3A = T3y * T3z;
|
Chris@10
|
774 {
|
Chris@10
|
775 E T3k, T36, T54, T55;
|
Chris@10
|
776 T3k = T3e * T3j;
|
Chris@10
|
777 cr[WS(rs, 5)] = FNMS(T3i, T3j, T3g);
|
Chris@10
|
778 T36 = T1c * T35;
|
Chris@10
|
779 cr[WS(rs, 21)] = FNMS(T2O, T35, T2M);
|
Chris@10
|
780 ci[WS(rs, 13)] = FMA(T3l, T3t, T3u);
|
Chris@10
|
781 cr[WS(rs, 13)] = FNMS(T3q, T3t, T3p);
|
Chris@10
|
782 ci[WS(rs, 29)] = FMA(T3y, T3D, T3E);
|
Chris@10
|
783 cr[WS(rs, 29)] = FNMS(T3C, T3D, T3A);
|
Chris@10
|
784 ci[WS(rs, 5)] = FMA(T3i, T3f, T3k);
|
Chris@10
|
785 ci[WS(rs, 21)] = FMA(T2O, T2L, T36);
|
Chris@10
|
786 T5f = FMA(KP923879532, T4C, T4v);
|
Chris@10
|
787 T4D = FNMS(KP923879532, T4C, T4v);
|
Chris@10
|
788 T4K = FNMS(KP668178637, T4J, T4G);
|
Chris@10
|
789 T54 = FMA(KP668178637, T4G, T4J);
|
Chris@10
|
790 T55 = FMA(KP668178637, T4N, T4Q);
|
Chris@10
|
791 T4R = FNMS(KP668178637, T4Q, T4N);
|
Chris@10
|
792 T56 = T54 - T55;
|
Chris@10
|
793 T5g = T54 + T55;
|
Chris@10
|
794 }
|
Chris@10
|
795 }
|
Chris@10
|
796 }
|
Chris@10
|
797 }
|
Chris@10
|
798 }
|
Chris@10
|
799 {
|
Chris@10
|
800 E T4h, T41, T4c, T3N, T3Q, T3T, T44, T4d;
|
Chris@10
|
801 T4h = FNMS(KP923879532, T40, T3Z);
|
Chris@10
|
802 T41 = FMA(KP923879532, T40, T3Z);
|
Chris@10
|
803 {
|
Chris@10
|
804 E T57, T5b, T5h, T5p;
|
Chris@10
|
805 T57 = FNMS(KP831469612, T56, T53);
|
Chris@10
|
806 T5b = FMA(KP831469612, T56, T53);
|
Chris@10
|
807 T5h = FNMS(KP831469612, T5g, T5f);
|
Chris@10
|
808 T5p = FMA(KP831469612, T5g, T5f);
|
Chris@10
|
809 {
|
Chris@10
|
810 E T5m, T4S, T5i, T5q;
|
Chris@10
|
811 T5m = T4K - T4R;
|
Chris@10
|
812 T4S = T4K + T4R;
|
Chris@10
|
813 T5i = T5e * T5h;
|
Chris@10
|
814 T5q = T17 * T5p;
|
Chris@10
|
815 {
|
Chris@10
|
816 E T5n, T5r, T59, T4T;
|
Chris@10
|
817 T5n = FMA(KP831469612, T5m, T5l);
|
Chris@10
|
818 T5r = FNMS(KP831469612, T5m, T5l);
|
Chris@10
|
819 T59 = FMA(KP831469612, T4S, T4D);
|
Chris@10
|
820 T4T = FNMS(KP831469612, T4S, T4D);
|
Chris@10
|
821 {
|
Chris@10
|
822 E T5o, T5s, T5c, T5a;
|
Chris@10
|
823 T5o = T5e * T5n;
|
Chris@10
|
824 cr[WS(rs, 11)] = FNMS(T5k, T5n, T5i);
|
Chris@10
|
825 T5s = T17 * T5r;
|
Chris@10
|
826 cr[WS(rs, 27)] = FNMS(T1b, T5r, T5q);
|
Chris@10
|
827 T5c = T14 * T59;
|
Chris@10
|
828 T5a = T11 * T59;
|
Chris@10
|
829 {
|
Chris@10
|
830 E T58, T4U, T42, T43;
|
Chris@10
|
831 T58 = T4W * T4T;
|
Chris@10
|
832 T4U = T4s * T4T;
|
Chris@10
|
833 ci[WS(rs, 11)] = FMA(T5k, T5h, T5o);
|
Chris@10
|
834 ci[WS(rs, 27)] = FMA(T1b, T5p, T5s);
|
Chris@10
|
835 ci[WS(rs, 3)] = FMA(T11, T5b, T5c);
|
Chris@10
|
836 cr[WS(rs, 3)] = FNMS(T14, T5b, T5a);
|
Chris@10
|
837 ci[WS(rs, 19)] = FMA(T4s, T57, T58);
|
Chris@10
|
838 cr[WS(rs, 19)] = FNMS(T4W, T57, T4U);
|
Chris@10
|
839 T4c = FNMS(KP923879532, T3M, T3L);
|
Chris@10
|
840 T3N = FMA(KP923879532, T3M, T3L);
|
Chris@10
|
841 T3Q = FNMS(KP198912367, T3P, T3O);
|
Chris@10
|
842 T42 = FMA(KP198912367, T3O, T3P);
|
Chris@10
|
843 T43 = FNMS(KP198912367, T3R, T3S);
|
Chris@10
|
844 T3T = FMA(KP198912367, T3S, T3R);
|
Chris@10
|
845 T44 = T42 + T43;
|
Chris@10
|
846 T4d = T43 - T42;
|
Chris@10
|
847 }
|
Chris@10
|
848 }
|
Chris@10
|
849 }
|
Chris@10
|
850 }
|
Chris@10
|
851 }
|
Chris@10
|
852 T67 = FNMS(KP923879532, T5K, T5J);
|
Chris@10
|
853 T5L = FMA(KP923879532, T5K, T5J);
|
Chris@10
|
854 {
|
Chris@10
|
855 E T45, T49, T4e, T4l;
|
Chris@10
|
856 T45 = FNMS(KP980785280, T44, T41);
|
Chris@10
|
857 T49 = FMA(KP980785280, T44, T41);
|
Chris@10
|
858 T4e = FNMS(KP980785280, T4d, T4c);
|
Chris@10
|
859 T4l = FMA(KP980785280, T4d, T4c);
|
Chris@10
|
860 {
|
Chris@10
|
861 E T4i, T3U, T4f, T4m;
|
Chris@10
|
862 T4i = T3Q - T3T;
|
Chris@10
|
863 T3U = T3Q + T3T;
|
Chris@10
|
864 T4f = T4b * T4e;
|
Chris@10
|
865 T4m = T12 * T4l;
|
Chris@10
|
866 {
|
Chris@10
|
867 E T4j, T4n, T47, T3V;
|
Chris@10
|
868 T4j = FNMS(KP980785280, T4i, T4h);
|
Chris@10
|
869 T4n = FMA(KP980785280, T4i, T4h);
|
Chris@10
|
870 T47 = FMA(KP980785280, T3U, T3N);
|
Chris@10
|
871 T3V = FNMS(KP980785280, T3U, T3N);
|
Chris@10
|
872 {
|
Chris@10
|
873 E T4k, T4o, T4a, T48;
|
Chris@10
|
874 T4k = T4b * T4j;
|
Chris@10
|
875 cr[WS(rs, 25)] = FNMS(T4g, T4j, T4f);
|
Chris@10
|
876 T4o = T12 * T4n;
|
Chris@10
|
877 cr[WS(rs, 9)] = FNMS(T15, T4n, T4m);
|
Chris@10
|
878 T4a = T39 * T47;
|
Chris@10
|
879 T48 = T37 * T47;
|
Chris@10
|
880 {
|
Chris@10
|
881 E T46, T3W, T5M, T5N;
|
Chris@10
|
882 T46 = T3Y * T3V;
|
Chris@10
|
883 T3W = T3K * T3V;
|
Chris@10
|
884 ci[WS(rs, 25)] = FMA(T4g, T4e, T4k);
|
Chris@10
|
885 ci[WS(rs, 9)] = FMA(T15, T4l, T4o);
|
Chris@10
|
886 ci[WS(rs, 1)] = FMA(T37, T49, T4a);
|
Chris@10
|
887 cr[WS(rs, 1)] = FNMS(T39, T49, T48);
|
Chris@10
|
888 ci[WS(rs, 17)] = FMA(T3K, T45, T46);
|
Chris@10
|
889 cr[WS(rs, 17)] = FNMS(T3Y, T45, T3W);
|
Chris@10
|
890 T61 = FMA(KP923879532, T5w, T5v);
|
Chris@10
|
891 T5x = FNMS(KP923879532, T5w, T5v);
|
Chris@10
|
892 T5A = FNMS(KP198912367, T5z, T5y);
|
Chris@10
|
893 T5M = FMA(KP198912367, T5y, T5z);
|
Chris@10
|
894 T5N = FMA(KP198912367, T5B, T5C);
|
Chris@10
|
895 T5D = FNMS(KP198912367, T5C, T5B);
|
Chris@10
|
896 T5O = T5M - T5N;
|
Chris@10
|
897 T62 = T5M + T5N;
|
Chris@10
|
898 }
|
Chris@10
|
899 }
|
Chris@10
|
900 }
|
Chris@10
|
901 }
|
Chris@10
|
902 }
|
Chris@10
|
903 }
|
Chris@10
|
904 }
|
Chris@10
|
905 }
|
Chris@10
|
906 }
|
Chris@10
|
907 }
|
Chris@10
|
908 }
|
Chris@10
|
909 T5V = FMA(KP980785280, T5O, T5L);
|
Chris@10
|
910 T5P = FNMS(KP980785280, T5O, T5L);
|
Chris@10
|
911 {
|
Chris@10
|
912 E T6c, T63, T5E, T68;
|
Chris@10
|
913 T6c = FMA(KP980785280, T62, T61);
|
Chris@10
|
914 T63 = FNMS(KP980785280, T62, T61);
|
Chris@10
|
915 T5E = T5A + T5D;
|
Chris@10
|
916 T68 = T5D - T5A;
|
Chris@10
|
917 {
|
Chris@10
|
918 E T64, T6d, T6f, T69;
|
Chris@10
|
919 T64 = T60 * T63;
|
Chris@10
|
920 T6d = T6b * T6c;
|
Chris@10
|
921 T6f = FNMS(KP980785280, T68, T67);
|
Chris@10
|
922 T69 = FMA(KP980785280, T68, T67);
|
Chris@10
|
923 {
|
Chris@10
|
924 E T5F, T5S, T6a, T6g;
|
Chris@10
|
925 T5F = FMA(KP980785280, T5E, T5x);
|
Chris@10
|
926 T5S = FNMS(KP980785280, T5E, T5x);
|
Chris@10
|
927 T6a = T60 * T69;
|
Chris@10
|
928 cr[WS(rs, 15)] = FNMS(T66, T69, T64);
|
Chris@10
|
929 T6g = T6b * T6f;
|
Chris@10
|
930 cr[WS(rs, 31)] = FNMS(T6e, T6f, T6d);
|
Chris@10
|
931 {
|
Chris@10
|
932 E T5W, T5T, T5Q, T5G;
|
Chris@10
|
933 T5W = T5U * T5S;
|
Chris@10
|
934 T5T = T5R * T5S;
|
Chris@10
|
935 T5Q = T5I * T5F;
|
Chris@10
|
936 T5G = T5u * T5F;
|
Chris@10
|
937 ci[WS(rs, 15)] = FMA(T66, T63, T6a);
|
Chris@10
|
938 ci[WS(rs, 31)] = FMA(T6e, T6c, T6g);
|
Chris@10
|
939 ci[WS(rs, 7)] = FMA(T5R, T5V, T5W);
|
Chris@10
|
940 cr[WS(rs, 7)] = FNMS(T5U, T5V, T5T);
|
Chris@10
|
941 ci[WS(rs, 23)] = FMA(T5u, T5P, T5Q);
|
Chris@10
|
942 cr[WS(rs, 23)] = FNMS(T5I, T5P, T5G);
|
Chris@10
|
943 }
|
Chris@10
|
944 }
|
Chris@10
|
945 }
|
Chris@10
|
946 }
|
Chris@10
|
947 }
|
Chris@10
|
948 }
|
Chris@10
|
949 }
|
Chris@10
|
950
|
Chris@10
|
951 static const tw_instr twinstr[] = {
|
Chris@10
|
952 {TW_CEXP, 1, 1},
|
Chris@10
|
953 {TW_CEXP, 1, 3},
|
Chris@10
|
954 {TW_CEXP, 1, 9},
|
Chris@10
|
955 {TW_CEXP, 1, 27},
|
Chris@10
|
956 {TW_NEXT, 1, 0}
|
Chris@10
|
957 };
|
Chris@10
|
958
|
Chris@10
|
959 static const hc2hc_desc desc = { 32, "hb2_32", twinstr, &GENUS, {236, 98, 252, 0} };
|
Chris@10
|
960
|
Chris@10
|
961 void X(codelet_hb2_32) (planner *p) {
|
Chris@10
|
962 X(khc2hc_register) (p, hb2_32, &desc);
|
Chris@10
|
963 }
|
Chris@10
|
964 #else /* HAVE_FMA */
|
Chris@10
|
965
|
Chris@10
|
966 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 32 -dif -name hb2_32 -include hb.h */
|
Chris@10
|
967
|
Chris@10
|
968 /*
|
Chris@10
|
969 * This function contains 488 FP additions, 280 FP multiplications,
|
Chris@10
|
970 * (or, 376 additions, 168 multiplications, 112 fused multiply/add),
|
Chris@10
|
971 * 160 stack variables, 7 constants, and 128 memory accesses
|
Chris@10
|
972 */
|
Chris@10
|
973 #include "hb.h"
|
Chris@10
|
974
|
Chris@10
|
975 static void hb2_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
976 {
|
Chris@10
|
977 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@10
|
978 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
979 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
980 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@10
|
981 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
982 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
983 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
984 {
|
Chris@10
|
985 INT m;
|
Chris@10
|
986 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 8, MAKE_VOLATILE_STRIDE(64, rs)) {
|
Chris@10
|
987 E T11, T14, T12, T15, T17, T2z, T2B, T1c, T18, T1d, T1g, T1k, T2F, T2L, T3t;
|
Chris@10
|
988 E T4H, T3h, T3V, T3b, T4v, T4T, T4X, T6t, T71, T6z, T75, T81, T8x, T8f, T8z;
|
Chris@10
|
989 E T2R, T2V, T8p, T8t, T4r, T4t, T53, T69, T3n, T3r, T7P, T7T, T4P, T4R, T6F;
|
Chris@10
|
990 E T6R, T1f, T2X, T1j, T2Y, T1l, T31, T2d, T2Z, T49, T4h, T4c, T4i, T4d, T4n;
|
Chris@10
|
991 E T4f, T4j;
|
Chris@10
|
992 {
|
Chris@10
|
993 E T2P, T3q, T2U, T3l, T2Q, T3p, T2T, T3m, T2D, T3g, T2K, T39, T2E, T3f, T2J;
|
Chris@10
|
994 E T3a;
|
Chris@10
|
995 {
|
Chris@10
|
996 E T13, T1b, T16, T1a;
|
Chris@10
|
997 T11 = W[0];
|
Chris@10
|
998 T14 = W[1];
|
Chris@10
|
999 T12 = W[2];
|
Chris@10
|
1000 T15 = W[3];
|
Chris@10
|
1001 T13 = T11 * T12;
|
Chris@10
|
1002 T1b = T14 * T12;
|
Chris@10
|
1003 T16 = T14 * T15;
|
Chris@10
|
1004 T1a = T11 * T15;
|
Chris@10
|
1005 T17 = T13 + T16;
|
Chris@10
|
1006 T2z = T13 - T16;
|
Chris@10
|
1007 T2B = T1a + T1b;
|
Chris@10
|
1008 T1c = T1a - T1b;
|
Chris@10
|
1009 T18 = W[4];
|
Chris@10
|
1010 T2P = T12 * T18;
|
Chris@10
|
1011 T3q = T14 * T18;
|
Chris@10
|
1012 T2U = T15 * T18;
|
Chris@10
|
1013 T3l = T11 * T18;
|
Chris@10
|
1014 T1d = W[5];
|
Chris@10
|
1015 T2Q = T15 * T1d;
|
Chris@10
|
1016 T3p = T11 * T1d;
|
Chris@10
|
1017 T2T = T12 * T1d;
|
Chris@10
|
1018 T3m = T14 * T1d;
|
Chris@10
|
1019 T1g = W[6];
|
Chris@10
|
1020 T2D = T11 * T1g;
|
Chris@10
|
1021 T3g = T15 * T1g;
|
Chris@10
|
1022 T2K = T14 * T1g;
|
Chris@10
|
1023 T39 = T12 * T1g;
|
Chris@10
|
1024 T1k = W[7];
|
Chris@10
|
1025 T2E = T14 * T1k;
|
Chris@10
|
1026 T3f = T12 * T1k;
|
Chris@10
|
1027 T2J = T11 * T1k;
|
Chris@10
|
1028 T3a = T15 * T1k;
|
Chris@10
|
1029 }
|
Chris@10
|
1030 T2F = T2D - T2E;
|
Chris@10
|
1031 T2L = T2J + T2K;
|
Chris@10
|
1032 T3t = T39 - T3a;
|
Chris@10
|
1033 T4H = T2J - T2K;
|
Chris@10
|
1034 T3h = T3f - T3g;
|
Chris@10
|
1035 T3V = T3f + T3g;
|
Chris@10
|
1036 T3b = T39 + T3a;
|
Chris@10
|
1037 T4v = T2D + T2E;
|
Chris@10
|
1038 T4T = FMA(T18, T1g, T1d * T1k);
|
Chris@10
|
1039 T4X = FNMS(T1d, T1g, T18 * T1k);
|
Chris@10
|
1040 {
|
Chris@10
|
1041 E T6r, T6s, T6x, T6y;
|
Chris@10
|
1042 T6r = T17 * T1g;
|
Chris@10
|
1043 T6s = T1c * T1k;
|
Chris@10
|
1044 T6t = T6r - T6s;
|
Chris@10
|
1045 T71 = T6r + T6s;
|
Chris@10
|
1046 T6x = T17 * T1k;
|
Chris@10
|
1047 T6y = T1c * T1g;
|
Chris@10
|
1048 T6z = T6x + T6y;
|
Chris@10
|
1049 T75 = T6x - T6y;
|
Chris@10
|
1050 }
|
Chris@10
|
1051 {
|
Chris@10
|
1052 E T7Z, T80, T8d, T8e;
|
Chris@10
|
1053 T7Z = T2z * T1g;
|
Chris@10
|
1054 T80 = T2B * T1k;
|
Chris@10
|
1055 T81 = T7Z + T80;
|
Chris@10
|
1056 T8x = T7Z - T80;
|
Chris@10
|
1057 T8d = T2z * T1k;
|
Chris@10
|
1058 T8e = T2B * T1g;
|
Chris@10
|
1059 T8f = T8d - T8e;
|
Chris@10
|
1060 T8z = T8d + T8e;
|
Chris@10
|
1061 T2R = T2P - T2Q;
|
Chris@10
|
1062 T2V = T2T + T2U;
|
Chris@10
|
1063 T8p = FMA(T2R, T1g, T2V * T1k);
|
Chris@10
|
1064 T8t = FNMS(T2V, T1g, T2R * T1k);
|
Chris@10
|
1065 }
|
Chris@10
|
1066 T4r = T2P + T2Q;
|
Chris@10
|
1067 T4t = T2T - T2U;
|
Chris@10
|
1068 T53 = FMA(T4r, T1g, T4t * T1k);
|
Chris@10
|
1069 T69 = FNMS(T4t, T1g, T4r * T1k);
|
Chris@10
|
1070 T3n = T3l + T3m;
|
Chris@10
|
1071 T3r = T3p - T3q;
|
Chris@10
|
1072 T7P = FMA(T3n, T1g, T3r * T1k);
|
Chris@10
|
1073 T7T = FNMS(T3r, T1g, T3n * T1k);
|
Chris@10
|
1074 T4P = T3l - T3m;
|
Chris@10
|
1075 T4R = T3p + T3q;
|
Chris@10
|
1076 T6F = FMA(T4P, T1g, T4R * T1k);
|
Chris@10
|
1077 T6R = FNMS(T4R, T1g, T4P * T1k);
|
Chris@10
|
1078 {
|
Chris@10
|
1079 E T19, T1e, T1h, T1i;
|
Chris@10
|
1080 T19 = T17 * T18;
|
Chris@10
|
1081 T1e = T1c * T1d;
|
Chris@10
|
1082 T1f = T19 + T1e;
|
Chris@10
|
1083 T2X = T19 - T1e;
|
Chris@10
|
1084 T1h = T17 * T1d;
|
Chris@10
|
1085 T1i = T1c * T18;
|
Chris@10
|
1086 T1j = T1h - T1i;
|
Chris@10
|
1087 T2Y = T1h + T1i;
|
Chris@10
|
1088 }
|
Chris@10
|
1089 T1l = FMA(T1f, T1g, T1j * T1k);
|
Chris@10
|
1090 T31 = FNMS(T2Y, T1g, T2X * T1k);
|
Chris@10
|
1091 T2d = FNMS(T1j, T1g, T1f * T1k);
|
Chris@10
|
1092 T2Z = FMA(T2X, T1g, T2Y * T1k);
|
Chris@10
|
1093 {
|
Chris@10
|
1094 E T47, T48, T4a, T4b;
|
Chris@10
|
1095 T47 = T2z * T18;
|
Chris@10
|
1096 T48 = T2B * T1d;
|
Chris@10
|
1097 T49 = T47 - T48;
|
Chris@10
|
1098 T4h = T47 + T48;
|
Chris@10
|
1099 T4a = T2z * T1d;
|
Chris@10
|
1100 T4b = T2B * T18;
|
Chris@10
|
1101 T4c = T4a + T4b;
|
Chris@10
|
1102 T4i = T4a - T4b;
|
Chris@10
|
1103 }
|
Chris@10
|
1104 T4d = FMA(T49, T1g, T4c * T1k);
|
Chris@10
|
1105 T4n = FNMS(T4i, T1g, T4h * T1k);
|
Chris@10
|
1106 T4f = FNMS(T4c, T1g, T49 * T1k);
|
Chris@10
|
1107 T4j = FMA(T4h, T1g, T4i * T1k);
|
Chris@10
|
1108 }
|
Chris@10
|
1109 {
|
Chris@10
|
1110 E T56, T7b, T7C, T6c, Tf, T1m, T6f, T7c, T3Y, T4I, T2t, T32, T5d, T7D, T3w;
|
Chris@10
|
1111 E T4w, Tu, T2e, T7g, T7F, T7j, T7G, T1B, T33, T3z, T40, T5l, T6i, T5s, T6h;
|
Chris@10
|
1112 E T3C, T3Z, TK, T1D, T7v, T86, T7y, T85, T1S, T35, T3O, T4C, T5F, T6J, T5M;
|
Chris@10
|
1113 E T6K, T3R, T4D, TZ, T1U, T7o, T89, T7r, T88, T29, T36, T3H, T4z, T5Y, T6M;
|
Chris@10
|
1114 E T65, T6N, T3K, T4A;
|
Chris@10
|
1115 {
|
Chris@10
|
1116 E T3, T54, T2o, T58, T2r, T5b, T6, T6a, Ta, T57, T2h, T6b, T2k, T55, Td;
|
Chris@10
|
1117 E T5a;
|
Chris@10
|
1118 {
|
Chris@10
|
1119 E T1, T2, T2m, T2n;
|
Chris@10
|
1120 T1 = cr[0];
|
Chris@10
|
1121 T2 = ci[WS(rs, 15)];
|
Chris@10
|
1122 T3 = T1 + T2;
|
Chris@10
|
1123 T54 = T1 - T2;
|
Chris@10
|
1124 T2m = ci[WS(rs, 27)];
|
Chris@10
|
1125 T2n = cr[WS(rs, 20)];
|
Chris@10
|
1126 T2o = T2m - T2n;
|
Chris@10
|
1127 T58 = T2m + T2n;
|
Chris@10
|
1128 }
|
Chris@10
|
1129 {
|
Chris@10
|
1130 E T2p, T2q, T4, T5;
|
Chris@10
|
1131 T2p = ci[WS(rs, 19)];
|
Chris@10
|
1132 T2q = cr[WS(rs, 28)];
|
Chris@10
|
1133 T2r = T2p - T2q;
|
Chris@10
|
1134 T5b = T2p + T2q;
|
Chris@10
|
1135 T4 = cr[WS(rs, 8)];
|
Chris@10
|
1136 T5 = ci[WS(rs, 7)];
|
Chris@10
|
1137 T6 = T4 + T5;
|
Chris@10
|
1138 T6a = T4 - T5;
|
Chris@10
|
1139 }
|
Chris@10
|
1140 {
|
Chris@10
|
1141 E T8, T9, T2f, T2g;
|
Chris@10
|
1142 T8 = cr[WS(rs, 4)];
|
Chris@10
|
1143 T9 = ci[WS(rs, 11)];
|
Chris@10
|
1144 Ta = T8 + T9;
|
Chris@10
|
1145 T57 = T8 - T9;
|
Chris@10
|
1146 T2f = ci[WS(rs, 31)];
|
Chris@10
|
1147 T2g = cr[WS(rs, 16)];
|
Chris@10
|
1148 T2h = T2f - T2g;
|
Chris@10
|
1149 T6b = T2f + T2g;
|
Chris@10
|
1150 }
|
Chris@10
|
1151 {
|
Chris@10
|
1152 E T2i, T2j, Tb, Tc;
|
Chris@10
|
1153 T2i = ci[WS(rs, 23)];
|
Chris@10
|
1154 T2j = cr[WS(rs, 24)];
|
Chris@10
|
1155 T2k = T2i - T2j;
|
Chris@10
|
1156 T55 = T2i + T2j;
|
Chris@10
|
1157 Tb = ci[WS(rs, 3)];
|
Chris@10
|
1158 Tc = cr[WS(rs, 12)];
|
Chris@10
|
1159 Td = Tb + Tc;
|
Chris@10
|
1160 T5a = Tb - Tc;
|
Chris@10
|
1161 }
|
Chris@10
|
1162 {
|
Chris@10
|
1163 E T7, Te, T2l, T2s;
|
Chris@10
|
1164 T56 = T54 - T55;
|
Chris@10
|
1165 T7b = T54 + T55;
|
Chris@10
|
1166 T7C = T6b - T6a;
|
Chris@10
|
1167 T6c = T6a + T6b;
|
Chris@10
|
1168 T7 = T3 + T6;
|
Chris@10
|
1169 Te = Ta + Td;
|
Chris@10
|
1170 Tf = T7 + Te;
|
Chris@10
|
1171 T1m = T7 - Te;
|
Chris@10
|
1172 {
|
Chris@10
|
1173 E T6d, T6e, T3W, T3X;
|
Chris@10
|
1174 T6d = T57 + T58;
|
Chris@10
|
1175 T6e = T5a + T5b;
|
Chris@10
|
1176 T6f = KP707106781 * (T6d - T6e);
|
Chris@10
|
1177 T7c = KP707106781 * (T6d + T6e);
|
Chris@10
|
1178 T3W = T2h - T2k;
|
Chris@10
|
1179 T3X = Ta - Td;
|
Chris@10
|
1180 T3Y = T3W - T3X;
|
Chris@10
|
1181 T4I = T3X + T3W;
|
Chris@10
|
1182 }
|
Chris@10
|
1183 T2l = T2h + T2k;
|
Chris@10
|
1184 T2s = T2o + T2r;
|
Chris@10
|
1185 T2t = T2l - T2s;
|
Chris@10
|
1186 T32 = T2l + T2s;
|
Chris@10
|
1187 {
|
Chris@10
|
1188 E T59, T5c, T3u, T3v;
|
Chris@10
|
1189 T59 = T57 - T58;
|
Chris@10
|
1190 T5c = T5a - T5b;
|
Chris@10
|
1191 T5d = KP707106781 * (T59 + T5c);
|
Chris@10
|
1192 T7D = KP707106781 * (T59 - T5c);
|
Chris@10
|
1193 T3u = T3 - T6;
|
Chris@10
|
1194 T3v = T2r - T2o;
|
Chris@10
|
1195 T3w = T3u - T3v;
|
Chris@10
|
1196 T4w = T3u + T3v;
|
Chris@10
|
1197 }
|
Chris@10
|
1198 }
|
Chris@10
|
1199 }
|
Chris@10
|
1200 {
|
Chris@10
|
1201 E Ti, T5p, T1w, T5n, T1z, T5q, Tl, T5m, Tp, T5i, T1p, T5g, T1s, T5j, Ts;
|
Chris@10
|
1202 E T5f;
|
Chris@10
|
1203 {
|
Chris@10
|
1204 E Tg, Th, T1u, T1v;
|
Chris@10
|
1205 Tg = cr[WS(rs, 2)];
|
Chris@10
|
1206 Th = ci[WS(rs, 13)];
|
Chris@10
|
1207 Ti = Tg + Th;
|
Chris@10
|
1208 T5p = Tg - Th;
|
Chris@10
|
1209 T1u = ci[WS(rs, 29)];
|
Chris@10
|
1210 T1v = cr[WS(rs, 18)];
|
Chris@10
|
1211 T1w = T1u - T1v;
|
Chris@10
|
1212 T5n = T1u + T1v;
|
Chris@10
|
1213 }
|
Chris@10
|
1214 {
|
Chris@10
|
1215 E T1x, T1y, Tj, Tk;
|
Chris@10
|
1216 T1x = ci[WS(rs, 21)];
|
Chris@10
|
1217 T1y = cr[WS(rs, 26)];
|
Chris@10
|
1218 T1z = T1x - T1y;
|
Chris@10
|
1219 T5q = T1x + T1y;
|
Chris@10
|
1220 Tj = cr[WS(rs, 10)];
|
Chris@10
|
1221 Tk = ci[WS(rs, 5)];
|
Chris@10
|
1222 Tl = Tj + Tk;
|
Chris@10
|
1223 T5m = Tj - Tk;
|
Chris@10
|
1224 }
|
Chris@10
|
1225 {
|
Chris@10
|
1226 E Tn, To, T1n, T1o;
|
Chris@10
|
1227 Tn = ci[WS(rs, 1)];
|
Chris@10
|
1228 To = cr[WS(rs, 14)];
|
Chris@10
|
1229 Tp = Tn + To;
|
Chris@10
|
1230 T5i = Tn - To;
|
Chris@10
|
1231 T1n = ci[WS(rs, 17)];
|
Chris@10
|
1232 T1o = cr[WS(rs, 30)];
|
Chris@10
|
1233 T1p = T1n - T1o;
|
Chris@10
|
1234 T5g = T1n + T1o;
|
Chris@10
|
1235 }
|
Chris@10
|
1236 {
|
Chris@10
|
1237 E T1q, T1r, Tq, Tr;
|
Chris@10
|
1238 T1q = ci[WS(rs, 25)];
|
Chris@10
|
1239 T1r = cr[WS(rs, 22)];
|
Chris@10
|
1240 T1s = T1q - T1r;
|
Chris@10
|
1241 T5j = T1q + T1r;
|
Chris@10
|
1242 Tq = cr[WS(rs, 6)];
|
Chris@10
|
1243 Tr = ci[WS(rs, 9)];
|
Chris@10
|
1244 Ts = Tq + Tr;
|
Chris@10
|
1245 T5f = Tq - Tr;
|
Chris@10
|
1246 }
|
Chris@10
|
1247 {
|
Chris@10
|
1248 E Tm, Tt, T7e, T7f;
|
Chris@10
|
1249 Tm = Ti + Tl;
|
Chris@10
|
1250 Tt = Tp + Ts;
|
Chris@10
|
1251 Tu = Tm + Tt;
|
Chris@10
|
1252 T2e = Tm - Tt;
|
Chris@10
|
1253 T7e = T5p + T5q;
|
Chris@10
|
1254 T7f = T5n - T5m;
|
Chris@10
|
1255 T7g = FNMS(KP923879532, T7f, KP382683432 * T7e);
|
Chris@10
|
1256 T7F = FMA(KP382683432, T7f, KP923879532 * T7e);
|
Chris@10
|
1257 }
|
Chris@10
|
1258 {
|
Chris@10
|
1259 E T7h, T7i, T1t, T1A;
|
Chris@10
|
1260 T7h = T5i + T5j;
|
Chris@10
|
1261 T7i = T5f + T5g;
|
Chris@10
|
1262 T7j = FNMS(KP923879532, T7i, KP382683432 * T7h);
|
Chris@10
|
1263 T7G = FMA(KP382683432, T7i, KP923879532 * T7h);
|
Chris@10
|
1264 T1t = T1p + T1s;
|
Chris@10
|
1265 T1A = T1w + T1z;
|
Chris@10
|
1266 T1B = T1t - T1A;
|
Chris@10
|
1267 T33 = T1A + T1t;
|
Chris@10
|
1268 }
|
Chris@10
|
1269 {
|
Chris@10
|
1270 E T3x, T3y, T5h, T5k;
|
Chris@10
|
1271 T3x = T1p - T1s;
|
Chris@10
|
1272 T3y = Tp - Ts;
|
Chris@10
|
1273 T3z = T3x - T3y;
|
Chris@10
|
1274 T40 = T3y + T3x;
|
Chris@10
|
1275 T5h = T5f - T5g;
|
Chris@10
|
1276 T5k = T5i - T5j;
|
Chris@10
|
1277 T5l = FNMS(KP382683432, T5k, KP923879532 * T5h);
|
Chris@10
|
1278 T6i = FMA(KP382683432, T5h, KP923879532 * T5k);
|
Chris@10
|
1279 }
|
Chris@10
|
1280 {
|
Chris@10
|
1281 E T5o, T5r, T3A, T3B;
|
Chris@10
|
1282 T5o = T5m + T5n;
|
Chris@10
|
1283 T5r = T5p - T5q;
|
Chris@10
|
1284 T5s = FMA(KP923879532, T5o, KP382683432 * T5r);
|
Chris@10
|
1285 T6h = FNMS(KP382683432, T5o, KP923879532 * T5r);
|
Chris@10
|
1286 T3A = Ti - Tl;
|
Chris@10
|
1287 T3B = T1w - T1z;
|
Chris@10
|
1288 T3C = T3A + T3B;
|
Chris@10
|
1289 T3Z = T3A - T3B;
|
Chris@10
|
1290 }
|
Chris@10
|
1291 }
|
Chris@10
|
1292 {
|
Chris@10
|
1293 E Ty, T5v, TB, T5G, T1J, T5w, T1G, T5H, TI, T5K, T1Q, T5D, TF, T5J, T1N;
|
Chris@10
|
1294 E T5A;
|
Chris@10
|
1295 {
|
Chris@10
|
1296 E Tw, Tx, T1E, T1F;
|
Chris@10
|
1297 Tw = cr[WS(rs, 1)];
|
Chris@10
|
1298 Tx = ci[WS(rs, 14)];
|
Chris@10
|
1299 Ty = Tw + Tx;
|
Chris@10
|
1300 T5v = Tw - Tx;
|
Chris@10
|
1301 {
|
Chris@10
|
1302 E Tz, TA, T1H, T1I;
|
Chris@10
|
1303 Tz = cr[WS(rs, 9)];
|
Chris@10
|
1304 TA = ci[WS(rs, 6)];
|
Chris@10
|
1305 TB = Tz + TA;
|
Chris@10
|
1306 T5G = Tz - TA;
|
Chris@10
|
1307 T1H = ci[WS(rs, 22)];
|
Chris@10
|
1308 T1I = cr[WS(rs, 25)];
|
Chris@10
|
1309 T1J = T1H - T1I;
|
Chris@10
|
1310 T5w = T1H + T1I;
|
Chris@10
|
1311 }
|
Chris@10
|
1312 T1E = ci[WS(rs, 30)];
|
Chris@10
|
1313 T1F = cr[WS(rs, 17)];
|
Chris@10
|
1314 T1G = T1E - T1F;
|
Chris@10
|
1315 T5H = T1E + T1F;
|
Chris@10
|
1316 {
|
Chris@10
|
1317 E TG, TH, T5B, T1O, T1P, T5C;
|
Chris@10
|
1318 TG = ci[WS(rs, 2)];
|
Chris@10
|
1319 TH = cr[WS(rs, 13)];
|
Chris@10
|
1320 T5B = TG - TH;
|
Chris@10
|
1321 T1O = ci[WS(rs, 18)];
|
Chris@10
|
1322 T1P = cr[WS(rs, 29)];
|
Chris@10
|
1323 T5C = T1O + T1P;
|
Chris@10
|
1324 TI = TG + TH;
|
Chris@10
|
1325 T5K = T5B + T5C;
|
Chris@10
|
1326 T1Q = T1O - T1P;
|
Chris@10
|
1327 T5D = T5B - T5C;
|
Chris@10
|
1328 }
|
Chris@10
|
1329 {
|
Chris@10
|
1330 E TD, TE, T5y, T1L, T1M, T5z;
|
Chris@10
|
1331 TD = cr[WS(rs, 5)];
|
Chris@10
|
1332 TE = ci[WS(rs, 10)];
|
Chris@10
|
1333 T5y = TD - TE;
|
Chris@10
|
1334 T1L = ci[WS(rs, 26)];
|
Chris@10
|
1335 T1M = cr[WS(rs, 21)];
|
Chris@10
|
1336 T5z = T1L + T1M;
|
Chris@10
|
1337 TF = TD + TE;
|
Chris@10
|
1338 T5J = T5y + T5z;
|
Chris@10
|
1339 T1N = T1L - T1M;
|
Chris@10
|
1340 T5A = T5y - T5z;
|
Chris@10
|
1341 }
|
Chris@10
|
1342 }
|
Chris@10
|
1343 {
|
Chris@10
|
1344 E TC, TJ, T7t, T7u;
|
Chris@10
|
1345 TC = Ty + TB;
|
Chris@10
|
1346 TJ = TF + TI;
|
Chris@10
|
1347 TK = TC + TJ;
|
Chris@10
|
1348 T1D = TC - TJ;
|
Chris@10
|
1349 T7t = T5H - T5G;
|
Chris@10
|
1350 T7u = KP707106781 * (T5A - T5D);
|
Chris@10
|
1351 T7v = T7t + T7u;
|
Chris@10
|
1352 T86 = T7t - T7u;
|
Chris@10
|
1353 }
|
Chris@10
|
1354 {
|
Chris@10
|
1355 E T7w, T7x, T1K, T1R;
|
Chris@10
|
1356 T7w = T5v + T5w;
|
Chris@10
|
1357 T7x = KP707106781 * (T5J + T5K);
|
Chris@10
|
1358 T7y = T7w - T7x;
|
Chris@10
|
1359 T85 = T7w + T7x;
|
Chris@10
|
1360 T1K = T1G + T1J;
|
Chris@10
|
1361 T1R = T1N + T1Q;
|
Chris@10
|
1362 T1S = T1K - T1R;
|
Chris@10
|
1363 T35 = T1K + T1R;
|
Chris@10
|
1364 }
|
Chris@10
|
1365 {
|
Chris@10
|
1366 E T3M, T3N, T5x, T5E;
|
Chris@10
|
1367 T3M = T1G - T1J;
|
Chris@10
|
1368 T3N = TF - TI;
|
Chris@10
|
1369 T3O = T3M - T3N;
|
Chris@10
|
1370 T4C = T3N + T3M;
|
Chris@10
|
1371 T5x = T5v - T5w;
|
Chris@10
|
1372 T5E = KP707106781 * (T5A + T5D);
|
Chris@10
|
1373 T5F = T5x - T5E;
|
Chris@10
|
1374 T6J = T5x + T5E;
|
Chris@10
|
1375 }
|
Chris@10
|
1376 {
|
Chris@10
|
1377 E T5I, T5L, T3P, T3Q;
|
Chris@10
|
1378 T5I = T5G + T5H;
|
Chris@10
|
1379 T5L = KP707106781 * (T5J - T5K);
|
Chris@10
|
1380 T5M = T5I - T5L;
|
Chris@10
|
1381 T6K = T5I + T5L;
|
Chris@10
|
1382 T3P = Ty - TB;
|
Chris@10
|
1383 T3Q = T1Q - T1N;
|
Chris@10
|
1384 T3R = T3P - T3Q;
|
Chris@10
|
1385 T4D = T3P + T3Q;
|
Chris@10
|
1386 }
|
Chris@10
|
1387 }
|
Chris@10
|
1388 {
|
Chris@10
|
1389 E TN, T5O, TQ, T5Z, T20, T5P, T1X, T60, TX, T63, T27, T5W, TU, T62, T24;
|
Chris@10
|
1390 E T5T;
|
Chris@10
|
1391 {
|
Chris@10
|
1392 E TL, TM, T1V, T1W;
|
Chris@10
|
1393 TL = ci[0];
|
Chris@10
|
1394 TM = cr[WS(rs, 15)];
|
Chris@10
|
1395 TN = TL + TM;
|
Chris@10
|
1396 T5O = TL - TM;
|
Chris@10
|
1397 {
|
Chris@10
|
1398 E TO, TP, T1Y, T1Z;
|
Chris@10
|
1399 TO = cr[WS(rs, 7)];
|
Chris@10
|
1400 TP = ci[WS(rs, 8)];
|
Chris@10
|
1401 TQ = TO + TP;
|
Chris@10
|
1402 T5Z = TO - TP;
|
Chris@10
|
1403 T1Y = ci[WS(rs, 24)];
|
Chris@10
|
1404 T1Z = cr[WS(rs, 23)];
|
Chris@10
|
1405 T20 = T1Y - T1Z;
|
Chris@10
|
1406 T5P = T1Y + T1Z;
|
Chris@10
|
1407 }
|
Chris@10
|
1408 T1V = ci[WS(rs, 16)];
|
Chris@10
|
1409 T1W = cr[WS(rs, 31)];
|
Chris@10
|
1410 T1X = T1V - T1W;
|
Chris@10
|
1411 T60 = T1V + T1W;
|
Chris@10
|
1412 {
|
Chris@10
|
1413 E TV, TW, T5U, T25, T26, T5V;
|
Chris@10
|
1414 TV = ci[WS(rs, 4)];
|
Chris@10
|
1415 TW = cr[WS(rs, 11)];
|
Chris@10
|
1416 T5U = TV - TW;
|
Chris@10
|
1417 T25 = ci[WS(rs, 20)];
|
Chris@10
|
1418 T26 = cr[WS(rs, 27)];
|
Chris@10
|
1419 T5V = T25 + T26;
|
Chris@10
|
1420 TX = TV + TW;
|
Chris@10
|
1421 T63 = T5U + T5V;
|
Chris@10
|
1422 T27 = T25 - T26;
|
Chris@10
|
1423 T5W = T5U - T5V;
|
Chris@10
|
1424 }
|
Chris@10
|
1425 {
|
Chris@10
|
1426 E TS, TT, T5R, T22, T23, T5S;
|
Chris@10
|
1427 TS = cr[WS(rs, 3)];
|
Chris@10
|
1428 TT = ci[WS(rs, 12)];
|
Chris@10
|
1429 T5R = TS - TT;
|
Chris@10
|
1430 T22 = ci[WS(rs, 28)];
|
Chris@10
|
1431 T23 = cr[WS(rs, 19)];
|
Chris@10
|
1432 T5S = T22 + T23;
|
Chris@10
|
1433 TU = TS + TT;
|
Chris@10
|
1434 T62 = T5R + T5S;
|
Chris@10
|
1435 T24 = T22 - T23;
|
Chris@10
|
1436 T5T = T5R - T5S;
|
Chris@10
|
1437 }
|
Chris@10
|
1438 }
|
Chris@10
|
1439 {
|
Chris@10
|
1440 E TR, TY, T7m, T7n;
|
Chris@10
|
1441 TR = TN + TQ;
|
Chris@10
|
1442 TY = TU + TX;
|
Chris@10
|
1443 TZ = TR + TY;
|
Chris@10
|
1444 T1U = TR - TY;
|
Chris@10
|
1445 T7m = KP707106781 * (T5T - T5W);
|
Chris@10
|
1446 T7n = T5Z + T60;
|
Chris@10
|
1447 T7o = T7m - T7n;
|
Chris@10
|
1448 T89 = T7n + T7m;
|
Chris@10
|
1449 }
|
Chris@10
|
1450 {
|
Chris@10
|
1451 E T7p, T7q, T21, T28;
|
Chris@10
|
1452 T7p = T5O + T5P;
|
Chris@10
|
1453 T7q = KP707106781 * (T62 + T63);
|
Chris@10
|
1454 T7r = T7p - T7q;
|
Chris@10
|
1455 T88 = T7p + T7q;
|
Chris@10
|
1456 T21 = T1X + T20;
|
Chris@10
|
1457 T28 = T24 + T27;
|
Chris@10
|
1458 T29 = T21 - T28;
|
Chris@10
|
1459 T36 = T21 + T28;
|
Chris@10
|
1460 }
|
Chris@10
|
1461 {
|
Chris@10
|
1462 E T3F, T3G, T5Q, T5X;
|
Chris@10
|
1463 T3F = T1X - T20;
|
Chris@10
|
1464 T3G = TU - TX;
|
Chris@10
|
1465 T3H = T3F - T3G;
|
Chris@10
|
1466 T4z = T3G + T3F;
|
Chris@10
|
1467 T5Q = T5O - T5P;
|
Chris@10
|
1468 T5X = KP707106781 * (T5T + T5W);
|
Chris@10
|
1469 T5Y = T5Q - T5X;
|
Chris@10
|
1470 T6M = T5Q + T5X;
|
Chris@10
|
1471 }
|
Chris@10
|
1472 {
|
Chris@10
|
1473 E T61, T64, T3I, T3J;
|
Chris@10
|
1474 T61 = T5Z - T60;
|
Chris@10
|
1475 T64 = KP707106781 * (T62 - T63);
|
Chris@10
|
1476 T65 = T61 - T64;
|
Chris@10
|
1477 T6N = T61 + T64;
|
Chris@10
|
1478 T3I = TN - TQ;
|
Chris@10
|
1479 T3J = T27 - T24;
|
Chris@10
|
1480 T3K = T3I - T3J;
|
Chris@10
|
1481 T4A = T3I + T3J;
|
Chris@10
|
1482 }
|
Chris@10
|
1483 }
|
Chris@10
|
1484 {
|
Chris@10
|
1485 E Tv, T10, T30, T34, T37, T38;
|
Chris@10
|
1486 Tv = Tf + Tu;
|
Chris@10
|
1487 T10 = TK + TZ;
|
Chris@10
|
1488 T30 = Tv - T10;
|
Chris@10
|
1489 T34 = T32 + T33;
|
Chris@10
|
1490 T37 = T35 + T36;
|
Chris@10
|
1491 T38 = T34 - T37;
|
Chris@10
|
1492 cr[0] = Tv + T10;
|
Chris@10
|
1493 ci[0] = T34 + T37;
|
Chris@10
|
1494 cr[WS(rs, 16)] = FNMS(T31, T38, T2Z * T30);
|
Chris@10
|
1495 ci[WS(rs, 16)] = FMA(T31, T30, T2Z * T38);
|
Chris@10
|
1496 }
|
Chris@10
|
1497 {
|
Chris@10
|
1498 E T3e, T3o, T3k, T3s;
|
Chris@10
|
1499 {
|
Chris@10
|
1500 E T3c, T3d, T3i, T3j;
|
Chris@10
|
1501 T3c = Tf - Tu;
|
Chris@10
|
1502 T3d = T36 - T35;
|
Chris@10
|
1503 T3e = T3c - T3d;
|
Chris@10
|
1504 T3o = T3c + T3d;
|
Chris@10
|
1505 T3i = T32 - T33;
|
Chris@10
|
1506 T3j = TK - TZ;
|
Chris@10
|
1507 T3k = T3i - T3j;
|
Chris@10
|
1508 T3s = T3j + T3i;
|
Chris@10
|
1509 }
|
Chris@10
|
1510 cr[WS(rs, 24)] = FNMS(T3h, T3k, T3b * T3e);
|
Chris@10
|
1511 ci[WS(rs, 24)] = FMA(T3b, T3k, T3h * T3e);
|
Chris@10
|
1512 cr[WS(rs, 8)] = FNMS(T3r, T3s, T3n * T3o);
|
Chris@10
|
1513 ci[WS(rs, 8)] = FMA(T3n, T3s, T3r * T3o);
|
Chris@10
|
1514 }
|
Chris@10
|
1515 {
|
Chris@10
|
1516 E T1C, T2u, T2M, T2G, T2x, T2H, T2b, T2N;
|
Chris@10
|
1517 T1C = T1m + T1B;
|
Chris@10
|
1518 T2u = T2e + T2t;
|
Chris@10
|
1519 T2M = T2t - T2e;
|
Chris@10
|
1520 T2G = T1m - T1B;
|
Chris@10
|
1521 {
|
Chris@10
|
1522 E T2v, T2w, T1T, T2a;
|
Chris@10
|
1523 T2v = T1D + T1S;
|
Chris@10
|
1524 T2w = T29 - T1U;
|
Chris@10
|
1525 T2x = KP707106781 * (T2v + T2w);
|
Chris@10
|
1526 T2H = KP707106781 * (T2w - T2v);
|
Chris@10
|
1527 T1T = T1D - T1S;
|
Chris@10
|
1528 T2a = T1U + T29;
|
Chris@10
|
1529 T2b = KP707106781 * (T1T + T2a);
|
Chris@10
|
1530 T2N = KP707106781 * (T1T - T2a);
|
Chris@10
|
1531 }
|
Chris@10
|
1532 {
|
Chris@10
|
1533 E T2c, T2y, T2S, T2W;
|
Chris@10
|
1534 T2c = T1C - T2b;
|
Chris@10
|
1535 T2y = T2u - T2x;
|
Chris@10
|
1536 cr[WS(rs, 20)] = FNMS(T2d, T2y, T1l * T2c);
|
Chris@10
|
1537 ci[WS(rs, 20)] = FMA(T2d, T2c, T1l * T2y);
|
Chris@10
|
1538 T2S = T2G + T2H;
|
Chris@10
|
1539 T2W = T2M + T2N;
|
Chris@10
|
1540 cr[WS(rs, 12)] = FNMS(T2V, T2W, T2R * T2S);
|
Chris@10
|
1541 ci[WS(rs, 12)] = FMA(T2R, T2W, T2V * T2S);
|
Chris@10
|
1542 }
|
Chris@10
|
1543 {
|
Chris@10
|
1544 E T2A, T2C, T2I, T2O;
|
Chris@10
|
1545 T2A = T1C + T2b;
|
Chris@10
|
1546 T2C = T2u + T2x;
|
Chris@10
|
1547 cr[WS(rs, 4)] = FNMS(T2B, T2C, T2z * T2A);
|
Chris@10
|
1548 ci[WS(rs, 4)] = FMA(T2B, T2A, T2z * T2C);
|
Chris@10
|
1549 T2I = T2G - T2H;
|
Chris@10
|
1550 T2O = T2M - T2N;
|
Chris@10
|
1551 cr[WS(rs, 28)] = FNMS(T2L, T2O, T2F * T2I);
|
Chris@10
|
1552 ci[WS(rs, 28)] = FMA(T2F, T2O, T2L * T2I);
|
Chris@10
|
1553 }
|
Chris@10
|
1554 }
|
Chris@10
|
1555 {
|
Chris@10
|
1556 E T4y, T4U, T4K, T4Y, T4F, T4Z, T4N, T4V, T4x, T4J;
|
Chris@10
|
1557 T4x = KP707106781 * (T3Z + T40);
|
Chris@10
|
1558 T4y = T4w - T4x;
|
Chris@10
|
1559 T4U = T4w + T4x;
|
Chris@10
|
1560 T4J = KP707106781 * (T3C + T3z);
|
Chris@10
|
1561 T4K = T4I - T4J;
|
Chris@10
|
1562 T4Y = T4I + T4J;
|
Chris@10
|
1563 {
|
Chris@10
|
1564 E T4B, T4E, T4L, T4M;
|
Chris@10
|
1565 T4B = FNMS(KP382683432, T4A, KP923879532 * T4z);
|
Chris@10
|
1566 T4E = FMA(KP923879532, T4C, KP382683432 * T4D);
|
Chris@10
|
1567 T4F = T4B - T4E;
|
Chris@10
|
1568 T4Z = T4E + T4B;
|
Chris@10
|
1569 T4L = FNMS(KP382683432, T4C, KP923879532 * T4D);
|
Chris@10
|
1570 T4M = FMA(KP382683432, T4z, KP923879532 * T4A);
|
Chris@10
|
1571 T4N = T4L - T4M;
|
Chris@10
|
1572 T4V = T4L + T4M;
|
Chris@10
|
1573 }
|
Chris@10
|
1574 {
|
Chris@10
|
1575 E T4G, T4O, T51, T52;
|
Chris@10
|
1576 T4G = T4y - T4F;
|
Chris@10
|
1577 T4O = T4K - T4N;
|
Chris@10
|
1578 cr[WS(rs, 26)] = FNMS(T4H, T4O, T4v * T4G);
|
Chris@10
|
1579 ci[WS(rs, 26)] = FMA(T4H, T4G, T4v * T4O);
|
Chris@10
|
1580 T51 = T4U + T4V;
|
Chris@10
|
1581 T52 = T4Y + T4Z;
|
Chris@10
|
1582 cr[WS(rs, 2)] = FNMS(T1c, T52, T17 * T51);
|
Chris@10
|
1583 ci[WS(rs, 2)] = FMA(T17, T52, T1c * T51);
|
Chris@10
|
1584 }
|
Chris@10
|
1585 {
|
Chris@10
|
1586 E T4Q, T4S, T4W, T50;
|
Chris@10
|
1587 T4Q = T4y + T4F;
|
Chris@10
|
1588 T4S = T4K + T4N;
|
Chris@10
|
1589 cr[WS(rs, 10)] = FNMS(T4R, T4S, T4P * T4Q);
|
Chris@10
|
1590 ci[WS(rs, 10)] = FMA(T4R, T4Q, T4P * T4S);
|
Chris@10
|
1591 T4W = T4U - T4V;
|
Chris@10
|
1592 T50 = T4Y - T4Z;
|
Chris@10
|
1593 cr[WS(rs, 18)] = FNMS(T4X, T50, T4T * T4W);
|
Chris@10
|
1594 ci[WS(rs, 18)] = FMA(T4T, T50, T4X * T4W);
|
Chris@10
|
1595 }
|
Chris@10
|
1596 }
|
Chris@10
|
1597 {
|
Chris@10
|
1598 E T3E, T4k, T42, T4o, T3T, T4p, T45, T4l, T3D, T41;
|
Chris@10
|
1599 T3D = KP707106781 * (T3z - T3C);
|
Chris@10
|
1600 T3E = T3w - T3D;
|
Chris@10
|
1601 T4k = T3w + T3D;
|
Chris@10
|
1602 T41 = KP707106781 * (T3Z - T40);
|
Chris@10
|
1603 T42 = T3Y - T41;
|
Chris@10
|
1604 T4o = T3Y + T41;
|
Chris@10
|
1605 {
|
Chris@10
|
1606 E T3L, T3S, T43, T44;
|
Chris@10
|
1607 T3L = FNMS(KP923879532, T3K, KP382683432 * T3H);
|
Chris@10
|
1608 T3S = FMA(KP382683432, T3O, KP923879532 * T3R);
|
Chris@10
|
1609 T3T = T3L - T3S;
|
Chris@10
|
1610 T4p = T3S + T3L;
|
Chris@10
|
1611 T43 = FNMS(KP923879532, T3O, KP382683432 * T3R);
|
Chris@10
|
1612 T44 = FMA(KP923879532, T3H, KP382683432 * T3K);
|
Chris@10
|
1613 T45 = T43 - T44;
|
Chris@10
|
1614 T4l = T43 + T44;
|
Chris@10
|
1615 }
|
Chris@10
|
1616 {
|
Chris@10
|
1617 E T3U, T46, T4s, T4u;
|
Chris@10
|
1618 T3U = T3E - T3T;
|
Chris@10
|
1619 T46 = T42 - T45;
|
Chris@10
|
1620 cr[WS(rs, 30)] = FNMS(T3V, T46, T3t * T3U);
|
Chris@10
|
1621 ci[WS(rs, 30)] = FMA(T3V, T3U, T3t * T46);
|
Chris@10
|
1622 T4s = T4k + T4l;
|
Chris@10
|
1623 T4u = T4o + T4p;
|
Chris@10
|
1624 cr[WS(rs, 6)] = FNMS(T4t, T4u, T4r * T4s);
|
Chris@10
|
1625 ci[WS(rs, 6)] = FMA(T4r, T4u, T4t * T4s);
|
Chris@10
|
1626 }
|
Chris@10
|
1627 {
|
Chris@10
|
1628 E T4e, T4g, T4m, T4q;
|
Chris@10
|
1629 T4e = T3E + T3T;
|
Chris@10
|
1630 T4g = T42 + T45;
|
Chris@10
|
1631 cr[WS(rs, 14)] = FNMS(T4f, T4g, T4d * T4e);
|
Chris@10
|
1632 ci[WS(rs, 14)] = FMA(T4f, T4e, T4d * T4g);
|
Chris@10
|
1633 T4m = T4k - T4l;
|
Chris@10
|
1634 T4q = T4o - T4p;
|
Chris@10
|
1635 cr[WS(rs, 22)] = FNMS(T4n, T4q, T4j * T4m);
|
Chris@10
|
1636 ci[WS(rs, 22)] = FMA(T4j, T4q, T4n * T4m);
|
Chris@10
|
1637 }
|
Chris@10
|
1638 }
|
Chris@10
|
1639 {
|
Chris@10
|
1640 E T6I, T72, T6X, T73, T6P, T77, T6U, T76;
|
Chris@10
|
1641 {
|
Chris@10
|
1642 E T6G, T6H, T6V, T6W;
|
Chris@10
|
1643 T6G = T56 + T5d;
|
Chris@10
|
1644 T6H = T6h + T6i;
|
Chris@10
|
1645 T6I = T6G + T6H;
|
Chris@10
|
1646 T72 = T6G - T6H;
|
Chris@10
|
1647 T6V = FMA(KP195090322, T6J, KP980785280 * T6K);
|
Chris@10
|
1648 T6W = FNMS(KP195090322, T6M, KP980785280 * T6N);
|
Chris@10
|
1649 T6X = T6V + T6W;
|
Chris@10
|
1650 T73 = T6W - T6V;
|
Chris@10
|
1651 }
|
Chris@10
|
1652 {
|
Chris@10
|
1653 E T6L, T6O, T6S, T6T;
|
Chris@10
|
1654 T6L = FNMS(KP195090322, T6K, KP980785280 * T6J);
|
Chris@10
|
1655 T6O = FMA(KP980785280, T6M, KP195090322 * T6N);
|
Chris@10
|
1656 T6P = T6L + T6O;
|
Chris@10
|
1657 T77 = T6L - T6O;
|
Chris@10
|
1658 T6S = T6c + T6f;
|
Chris@10
|
1659 T6T = T5s + T5l;
|
Chris@10
|
1660 T6U = T6S + T6T;
|
Chris@10
|
1661 T76 = T6S - T6T;
|
Chris@10
|
1662 }
|
Chris@10
|
1663 {
|
Chris@10
|
1664 E T6Q, T6Y, T79, T7a;
|
Chris@10
|
1665 T6Q = T6I - T6P;
|
Chris@10
|
1666 T6Y = T6U - T6X;
|
Chris@10
|
1667 cr[WS(rs, 17)] = FNMS(T6R, T6Y, T6F * T6Q);
|
Chris@10
|
1668 ci[WS(rs, 17)] = FMA(T6R, T6Q, T6F * T6Y);
|
Chris@10
|
1669 T79 = T72 + T73;
|
Chris@10
|
1670 T7a = T76 + T77;
|
Chris@10
|
1671 cr[WS(rs, 9)] = FNMS(T1d, T7a, T18 * T79);
|
Chris@10
|
1672 ci[WS(rs, 9)] = FMA(T18, T7a, T1d * T79);
|
Chris@10
|
1673 }
|
Chris@10
|
1674 {
|
Chris@10
|
1675 E T6Z, T70, T74, T78;
|
Chris@10
|
1676 T6Z = T6I + T6P;
|
Chris@10
|
1677 T70 = T6U + T6X;
|
Chris@10
|
1678 cr[WS(rs, 1)] = FNMS(T14, T70, T11 * T6Z);
|
Chris@10
|
1679 ci[WS(rs, 1)] = FMA(T14, T6Z, T11 * T70);
|
Chris@10
|
1680 T74 = T72 - T73;
|
Chris@10
|
1681 T78 = T76 - T77;
|
Chris@10
|
1682 cr[WS(rs, 25)] = FNMS(T75, T78, T71 * T74);
|
Chris@10
|
1683 ci[WS(rs, 25)] = FMA(T71, T78, T75 * T74);
|
Chris@10
|
1684 }
|
Chris@10
|
1685 }
|
Chris@10
|
1686 {
|
Chris@10
|
1687 E T84, T8q, T8l, T8r, T8b, T8v, T8i, T8u;
|
Chris@10
|
1688 {
|
Chris@10
|
1689 E T82, T83, T8j, T8k;
|
Chris@10
|
1690 T82 = T7b + T7c;
|
Chris@10
|
1691 T83 = T7F + T7G;
|
Chris@10
|
1692 T84 = T82 - T83;
|
Chris@10
|
1693 T8q = T82 + T83;
|
Chris@10
|
1694 T8j = FMA(KP195090322, T86, KP980785280 * T85);
|
Chris@10
|
1695 T8k = FMA(KP195090322, T89, KP980785280 * T88);
|
Chris@10
|
1696 T8l = T8j - T8k;
|
Chris@10
|
1697 T8r = T8j + T8k;
|
Chris@10
|
1698 }
|
Chris@10
|
1699 {
|
Chris@10
|
1700 E T87, T8a, T8g, T8h;
|
Chris@10
|
1701 T87 = FNMS(KP980785280, T86, KP195090322 * T85);
|
Chris@10
|
1702 T8a = FNMS(KP980785280, T89, KP195090322 * T88);
|
Chris@10
|
1703 T8b = T87 + T8a;
|
Chris@10
|
1704 T8v = T87 - T8a;
|
Chris@10
|
1705 T8g = T7C - T7D;
|
Chris@10
|
1706 T8h = T7g - T7j;
|
Chris@10
|
1707 T8i = T8g + T8h;
|
Chris@10
|
1708 T8u = T8g - T8h;
|
Chris@10
|
1709 }
|
Chris@10
|
1710 {
|
Chris@10
|
1711 E T8c, T8m, T8y, T8A;
|
Chris@10
|
1712 T8c = T84 - T8b;
|
Chris@10
|
1713 T8m = T8i - T8l;
|
Chris@10
|
1714 cr[WS(rs, 23)] = FNMS(T8f, T8m, T81 * T8c);
|
Chris@10
|
1715 ci[WS(rs, 23)] = FMA(T8f, T8c, T81 * T8m);
|
Chris@10
|
1716 T8y = T8q + T8r;
|
Chris@10
|
1717 T8A = T8u - T8v;
|
Chris@10
|
1718 cr[WS(rs, 31)] = FNMS(T8z, T8A, T8x * T8y);
|
Chris@10
|
1719 ci[WS(rs, 31)] = FMA(T8x, T8A, T8z * T8y);
|
Chris@10
|
1720 }
|
Chris@10
|
1721 {
|
Chris@10
|
1722 E T8n, T8o, T8s, T8w;
|
Chris@10
|
1723 T8n = T84 + T8b;
|
Chris@10
|
1724 T8o = T8i + T8l;
|
Chris@10
|
1725 cr[WS(rs, 7)] = FNMS(T1j, T8o, T1f * T8n);
|
Chris@10
|
1726 ci[WS(rs, 7)] = FMA(T1j, T8n, T1f * T8o);
|
Chris@10
|
1727 T8s = T8q - T8r;
|
Chris@10
|
1728 T8w = T8u + T8v;
|
Chris@10
|
1729 cr[WS(rs, 15)] = FNMS(T8t, T8w, T8p * T8s);
|
Chris@10
|
1730 ci[WS(rs, 15)] = FMA(T8p, T8w, T8t * T8s);
|
Chris@10
|
1731 }
|
Chris@10
|
1732 }
|
Chris@10
|
1733 {
|
Chris@10
|
1734 E T5u, T6u, T6n, T6v, T67, T6B, T6k, T6A;
|
Chris@10
|
1735 {
|
Chris@10
|
1736 E T5e, T5t, T6l, T6m;
|
Chris@10
|
1737 T5e = T56 - T5d;
|
Chris@10
|
1738 T5t = T5l - T5s;
|
Chris@10
|
1739 T5u = T5e + T5t;
|
Chris@10
|
1740 T6u = T5e - T5t;
|
Chris@10
|
1741 T6l = FMA(KP831469612, T5F, KP555570233 * T5M);
|
Chris@10
|
1742 T6m = FNMS(KP831469612, T5Y, KP555570233 * T65);
|
Chris@10
|
1743 T6n = T6l + T6m;
|
Chris@10
|
1744 T6v = T6m - T6l;
|
Chris@10
|
1745 }
|
Chris@10
|
1746 {
|
Chris@10
|
1747 E T5N, T66, T6g, T6j;
|
Chris@10
|
1748 T5N = FNMS(KP831469612, T5M, KP555570233 * T5F);
|
Chris@10
|
1749 T66 = FMA(KP555570233, T5Y, KP831469612 * T65);
|
Chris@10
|
1750 T67 = T5N + T66;
|
Chris@10
|
1751 T6B = T5N - T66;
|
Chris@10
|
1752 T6g = T6c - T6f;
|
Chris@10
|
1753 T6j = T6h - T6i;
|
Chris@10
|
1754 T6k = T6g + T6j;
|
Chris@10
|
1755 T6A = T6g - T6j;
|
Chris@10
|
1756 }
|
Chris@10
|
1757 {
|
Chris@10
|
1758 E T68, T6o, T6D, T6E;
|
Chris@10
|
1759 T68 = T5u - T67;
|
Chris@10
|
1760 T6o = T6k - T6n;
|
Chris@10
|
1761 cr[WS(rs, 21)] = FNMS(T69, T6o, T53 * T68);
|
Chris@10
|
1762 ci[WS(rs, 21)] = FMA(T69, T68, T53 * T6o);
|
Chris@10
|
1763 T6D = T6u + T6v;
|
Chris@10
|
1764 T6E = T6A + T6B;
|
Chris@10
|
1765 cr[WS(rs, 13)] = FNMS(T4c, T6E, T49 * T6D);
|
Chris@10
|
1766 ci[WS(rs, 13)] = FMA(T49, T6E, T4c * T6D);
|
Chris@10
|
1767 }
|
Chris@10
|
1768 {
|
Chris@10
|
1769 E T6p, T6q, T6w, T6C;
|
Chris@10
|
1770 T6p = T5u + T67;
|
Chris@10
|
1771 T6q = T6k + T6n;
|
Chris@10
|
1772 cr[WS(rs, 5)] = FNMS(T4i, T6q, T4h * T6p);
|
Chris@10
|
1773 ci[WS(rs, 5)] = FMA(T4i, T6p, T4h * T6q);
|
Chris@10
|
1774 T6w = T6u - T6v;
|
Chris@10
|
1775 T6C = T6A - T6B;
|
Chris@10
|
1776 cr[WS(rs, 29)] = FNMS(T6z, T6C, T6t * T6w);
|
Chris@10
|
1777 ci[WS(rs, 29)] = FMA(T6t, T6C, T6z * T6w);
|
Chris@10
|
1778 }
|
Chris@10
|
1779 }
|
Chris@10
|
1780 {
|
Chris@10
|
1781 E T7l, T7Q, T7L, T7R, T7A, T7V, T7I, T7U;
|
Chris@10
|
1782 {
|
Chris@10
|
1783 E T7d, T7k, T7J, T7K;
|
Chris@10
|
1784 T7d = T7b - T7c;
|
Chris@10
|
1785 T7k = T7g + T7j;
|
Chris@10
|
1786 T7l = T7d - T7k;
|
Chris@10
|
1787 T7Q = T7d + T7k;
|
Chris@10
|
1788 T7J = FNMS(KP555570233, T7v, KP831469612 * T7y);
|
Chris@10
|
1789 T7K = FMA(KP555570233, T7o, KP831469612 * T7r);
|
Chris@10
|
1790 T7L = T7J - T7K;
|
Chris@10
|
1791 T7R = T7J + T7K;
|
Chris@10
|
1792 }
|
Chris@10
|
1793 {
|
Chris@10
|
1794 E T7s, T7z, T7E, T7H;
|
Chris@10
|
1795 T7s = FNMS(KP555570233, T7r, KP831469612 * T7o);
|
Chris@10
|
1796 T7z = FMA(KP831469612, T7v, KP555570233 * T7y);
|
Chris@10
|
1797 T7A = T7s - T7z;
|
Chris@10
|
1798 T7V = T7z + T7s;
|
Chris@10
|
1799 T7E = T7C + T7D;
|
Chris@10
|
1800 T7H = T7F - T7G;
|
Chris@10
|
1801 T7I = T7E - T7H;
|
Chris@10
|
1802 T7U = T7E + T7H;
|
Chris@10
|
1803 }
|
Chris@10
|
1804 {
|
Chris@10
|
1805 E T7B, T7M, T7X, T7Y;
|
Chris@10
|
1806 T7B = T7l - T7A;
|
Chris@10
|
1807 T7M = T7I - T7L;
|
Chris@10
|
1808 cr[WS(rs, 27)] = FNMS(T1k, T7M, T1g * T7B);
|
Chris@10
|
1809 ci[WS(rs, 27)] = FMA(T1k, T7B, T1g * T7M);
|
Chris@10
|
1810 T7X = T7Q + T7R;
|
Chris@10
|
1811 T7Y = T7U + T7V;
|
Chris@10
|
1812 cr[WS(rs, 3)] = FNMS(T15, T7Y, T12 * T7X);
|
Chris@10
|
1813 ci[WS(rs, 3)] = FMA(T12, T7Y, T15 * T7X);
|
Chris@10
|
1814 }
|
Chris@10
|
1815 {
|
Chris@10
|
1816 E T7N, T7O, T7S, T7W;
|
Chris@10
|
1817 T7N = T7l + T7A;
|
Chris@10
|
1818 T7O = T7I + T7L;
|
Chris@10
|
1819 cr[WS(rs, 11)] = FNMS(T2Y, T7O, T2X * T7N);
|
Chris@10
|
1820 ci[WS(rs, 11)] = FMA(T2Y, T7N, T2X * T7O);
|
Chris@10
|
1821 T7S = T7Q - T7R;
|
Chris@10
|
1822 T7W = T7U - T7V;
|
Chris@10
|
1823 cr[WS(rs, 19)] = FNMS(T7T, T7W, T7P * T7S);
|
Chris@10
|
1824 ci[WS(rs, 19)] = FMA(T7P, T7W, T7T * T7S);
|
Chris@10
|
1825 }
|
Chris@10
|
1826 }
|
Chris@10
|
1827 }
|
Chris@10
|
1828 }
|
Chris@10
|
1829 }
|
Chris@10
|
1830 }
|
Chris@10
|
1831
|
Chris@10
|
1832 static const tw_instr twinstr[] = {
|
Chris@10
|
1833 {TW_CEXP, 1, 1},
|
Chris@10
|
1834 {TW_CEXP, 1, 3},
|
Chris@10
|
1835 {TW_CEXP, 1, 9},
|
Chris@10
|
1836 {TW_CEXP, 1, 27},
|
Chris@10
|
1837 {TW_NEXT, 1, 0}
|
Chris@10
|
1838 };
|
Chris@10
|
1839
|
Chris@10
|
1840 static const hc2hc_desc desc = { 32, "hb2_32", twinstr, &GENUS, {376, 168, 112, 0} };
|
Chris@10
|
1841
|
Chris@10
|
1842 void X(codelet_hb2_32) (planner *p) {
|
Chris@10
|
1843 X(khc2hc_register) (p, hb2_32, &desc);
|
Chris@10
|
1844 }
|
Chris@10
|
1845 #endif /* HAVE_FMA */
|